From 435af0b919bf9eb78f4e05e8596ebed9ca7885b7 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Mon, 24 Apr 2023 14:46:16 -0400 Subject: drm/amdkfd: Optimize svm range map to GPU with XNACK on With XNACK on if svm_range_set_attr set the range access or access_in_place attribute, we don't call svm_range_validate_and_map to update GPU mapping. This avoids prefaulting the range pages on system memory if the range is not prefetch to VRAM and not mapped to GPUs. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 96a138a39515..c02430537e9c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -735,7 +735,9 @@ svm_range_apply_attrs(struct kfd_process *p, struct svm_range *prange, case KFD_IOCTL_SVM_ATTR_ACCESS: case KFD_IOCTL_SVM_ATTR_ACCESS_IN_PLACE: case KFD_IOCTL_SVM_ATTR_NO_ACCESS: - *update_mapping = true; + if (!p->xnack_enabled) + *update_mapping = true; + gpuidx = kfd_process_gpuidx_from_gpuid(p, attrs[i].value); if (attrs[i].type == KFD_IOCTL_SVM_ATTR_NO_ACCESS) { -- cgit v1.2.3 From cabbdea1f1861098991768d7bbf5a49ed1608213 Mon Sep 17 00:00:00 2001 From: Daniil Dulov Date: Thu, 11 May 2023 04:23:14 -0700 Subject: drm/amdkfd: Fix potential deallocation of previously deallocated memory. Pointer mqd_mem_obj can be deallocated in kfd_gtt_sa_allocate(). The function then returns non-zero value, which causes the second deallocation. Found by Linux Verification Center (linuxtesting.org) with SVACE. Fixes: d1f8f0d17d40 ("drm/amdkfd: Move non-sdma mqd allocation out of init_mqd") Signed-off-by: Daniil Dulov Signed-off-by: Felix Kuehling Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index fdbfd725841f..51b53110341b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -115,18 +115,19 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, &(mqd_mem_obj->gtt_mem), &(mqd_mem_obj->gpu_addr), (void *)&(mqd_mem_obj->cpu_ptr), true); + + if (retval) { + kfree(mqd_mem_obj); + return NULL; + } } else { retval = kfd_gtt_sa_allocate(kfd, sizeof(struct v9_mqd), &mqd_mem_obj); - } - - if (retval) { - kfree(mqd_mem_obj); - return NULL; + if (retval) + return NULL; } return mqd_mem_obj; - } static void init_mqd(struct mqd_manager *mm, void **mqd, -- cgit v1.2.3 From f544afac3f34124088b981c63843a3cc48f4ee3e Mon Sep 17 00:00:00 2001 From: Amber Lin Date: Tue, 17 May 2022 23:41:01 +0800 Subject: drm/amdgpu: Add kgd2kfd for GC 9.4.3 New GC (v9.4.3) and ATHUB (v1.8.0) versions are used. Add kgd_gfx_v9_4_3_* functions if registers in use of kgd_gfx_v9_* functions are changed or have different offset. Signed-off-by: Amber Lin Acked-by: Felix Kuehling Reviewed-by: Mukul Joshi Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/Makefile | 1 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c | 183 +++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 30 ++-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 5 + drivers/gpu/drm/amd/amdkfd/kfd_device.c | 3 +- 5 files changed, 206 insertions(+), 16 deletions(-) create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/Makefile b/drivers/gpu/drm/amd/amdgpu/Makefile index 415a7fa395c4..7fb09000efc0 100644 --- a/drivers/gpu/drm/amd/amdgpu/Makefile +++ b/drivers/gpu/drm/amd/amdgpu/Makefile @@ -228,6 +228,7 @@ amdgpu-y += \ amdgpu_amdkfd_gfx_v9.o \ amdgpu_amdkfd_arcturus.o \ amdgpu_amdkfd_aldebaran.o \ + amdgpu_amdkfd_gc_9_4_3.o \ amdgpu_amdkfd_gfx_v10.o \ amdgpu_amdkfd_gfx_v10_3.o \ amdgpu_amdkfd_gfx_v11.o diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c new file mode 100644 index 000000000000..562e1a04160f --- /dev/null +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c @@ -0,0 +1,183 @@ +/* + * Copyright 2021 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + */ +#include "amdgpu.h" +#include "amdgpu_amdkfd.h" +#include "amdgpu_amdkfd_arcturus.h" +#include "amdgpu_amdkfd_gfx_v9.h" +#include "gc/gc_9_4_3_offset.h" +#include "gc/gc_9_4_3_sh_mask.h" +#include "athub/athub_1_8_0_offset.h" +#include "athub/athub_1_8_0_sh_mask.h" +#include "oss/osssys_4_0_offset.h" +#include "oss/osssys_4_0_sh_mask.h" +#include "v9_structs.h" +#include "soc15.h" + +static int kgd_gfx_v9_4_3_set_pasid_vmid_mapping(struct amdgpu_device *adev, + u32 pasid, unsigned int vmid) +{ + unsigned long timeout; + + /* + * We have to assume that there is no outstanding mapping. + * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because + * a mapping is in progress or because a mapping finished + * and the SW cleared it. + * So the protocol is to always wait & clear. + */ + uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid | + ATC_VMID0_PASID_MAPPING__VALID_MASK; + + WREG32(SOC15_REG_OFFSET(ATHUB, 0, + regATC_VMID0_PASID_MAPPING) + vmid, pasid_mapping); + + timeout = jiffies + msecs_to_jiffies(10); + while (!(RREG32(SOC15_REG_OFFSET(ATHUB, 0, + regATC_VMID_PASID_MAPPING_UPDATE_STATUS)) & + (1U << vmid))) { + if (time_after(jiffies, timeout)) { + pr_err("Fail to program VMID-PASID mapping\n"); + return -ETIME; + } + cpu_relax(); + } + + WREG32(SOC15_REG_OFFSET(ATHUB, 0, + regATC_VMID_PASID_MAPPING_UPDATE_STATUS), + 1U << vmid); + + WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid, + pasid_mapping); + WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid, + pasid_mapping); + + return 0; +} + +static inline struct v9_mqd *get_mqd(void *mqd) +{ + return (struct v9_mqd *)mqd; +} + +static int kgd_gfx_v9_4_3_hqd_load(struct amdgpu_device *adev, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + uint32_t __user *wptr, uint32_t wptr_shift, + uint32_t wptr_mask, struct mm_struct *mm) +{ + struct v9_mqd *m; + uint32_t *mqd_hqd; + uint32_t reg, hqd_base, hqd_end, data; + + m = get_mqd(mqd); + + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); + + /* HQD registers extend to CP_HQD_AQL_DISPATCH_ID_HI */ + mqd_hqd = &m->cp_mqd_base_addr_lo; + hqd_base = SOC15_REG_OFFSET(GC, 0, regCP_MQD_BASE_ADDR); + hqd_end = SOC15_REG_OFFSET(GC, 0, regCP_HQD_AQL_DISPATCH_ID_HI); + + for (reg = hqd_base; reg <= hqd_end; reg++) + WREG32_RLC(reg, mqd_hqd[reg - hqd_base]); + + + /* Activate doorbell logic before triggering WPTR poll. */ + data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control, + CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1); + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL), + data); + + if (wptr) { + /* Don't read wptr with get_user because the user + * context may not be accessible (if this function + * runs in a work queue). Instead trigger a one-shot + * polling read from memory in the CP. This assumes + * that wptr is GPU-accessible in the queue's VMID via + * ATC or SVM. WPTR==RPTR before starting the poll so + * the CP starts fetching new commands from the right + * place. + * + * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit + * tricky. Assume that the queue didn't overflow. The + * number of valid bits in the 32-bit RPTR depends on + * the queue size. The remaining bits are taken from + * the saved 64-bit WPTR. If the WPTR wrapped, add the + * queue size. + */ + uint32_t queue_size = + 2 << REG_GET_FIELD(m->cp_hqd_pq_control, + CP_HQD_PQ_CONTROL, QUEUE_SIZE); + uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1); + + if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr) + guessed_wptr += queue_size; + guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1); + guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32; + + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_LO), + lower_32_bits(guessed_wptr)); + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_HI), + upper_32_bits(guessed_wptr)); + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR), + lower_32_bits((uintptr_t)wptr)); + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, + regCP_HQD_PQ_WPTR_POLL_ADDR_HI), + upper_32_bits((uintptr_t)wptr)); + WREG32(SOC15_REG_OFFSET(GC, 0, regCP_PQ_WPTR_POLL_CNTL1), + (uint32_t)kgd_gfx_v9_get_queue_mask(adev, pipe_id, + queue_id)); + } + + /* Start the EOP fetcher */ + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_EOP_RPTR), + REG_SET_FIELD(m->cp_hqd_eop_rptr, + CP_HQD_EOP_RPTR, INIT_FETCHER, 1)); + + data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1); + WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_ACTIVE), data); + + kgd_gfx_v9_release_queue(adev); + + return 0; +} + +const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = { + .program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings, + .set_pasid_vmid_mapping = kgd_gfx_v9_4_3_set_pasid_vmid_mapping, + .init_interrupts = kgd_gfx_v9_init_interrupts, + .hqd_load = kgd_gfx_v9_4_3_hqd_load, + .hiq_mqd_load = kgd_gfx_v9_hiq_mqd_load, + .hqd_sdma_load = kgd_arcturus_hqd_sdma_load, + .hqd_dump = kgd_gfx_v9_hqd_dump, + .hqd_sdma_dump = kgd_arcturus_hqd_sdma_dump, + .hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied, + .hqd_sdma_is_occupied = kgd_arcturus_hqd_sdma_is_occupied, + .hqd_destroy = kgd_gfx_v9_hqd_destroy, + .hqd_sdma_destroy = kgd_arcturus_hqd_sdma_destroy, + .wave_control_execute = kgd_gfx_v9_wave_control_execute, + .get_atc_vmid_pasid_mapping_info = + kgd_gfx_v9_get_atc_vmid_pasid_mapping_info, + .set_vm_context_page_table_base = + kgd_gfx_v9_set_vm_context_page_table_base, + .program_trap_handler_settings = + kgd_gfx_v9_program_trap_handler_settings +}; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index ae06d1f2af93..d36219ecd3dd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -59,7 +59,7 @@ static void unlock_srbm(struct amdgpu_device *adev) mutex_unlock(&adev->srbm_mutex); } -static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id, +void kgd_gfx_v9_acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id) { uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; @@ -68,7 +68,7 @@ static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id, lock_srbm(adev, mec, pipe, queue_id, 0); } -static uint64_t get_queue_mask(struct amdgpu_device *adev, +uint64_t kgd_gfx_v9_get_queue_mask(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id) { unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe + @@ -77,7 +77,7 @@ static uint64_t get_queue_mask(struct amdgpu_device *adev, return 1ull << bit; } -static void release_queue(struct amdgpu_device *adev) +void kgd_gfx_v9_release_queue(struct amdgpu_device *adev) { unlock_srbm(adev); } @@ -228,7 +228,7 @@ int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, m = get_mqd(mqd); - acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */ mqd_hqd = &m->cp_mqd_base_addr_lo; @@ -280,7 +280,7 @@ int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI), upper_32_bits((uintptr_t)wptr)); WREG32_SOC15(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1, - (uint32_t)get_queue_mask(adev, pipe_id, queue_id)); + (uint32_t)kgd_gfx_v9_get_queue_mask(adev, pipe_id, queue_id)); } /* Start the EOP fetcher */ @@ -291,7 +291,7 @@ int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1); WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data); - release_queue(adev); + kgd_gfx_v9_release_queue(adev); return 0; } @@ -307,7 +307,7 @@ int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, m = get_mqd(mqd); - acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec); @@ -343,7 +343,7 @@ int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, out_unlock: spin_unlock(&adev->gfx.kiq[0].ring_lock); - release_queue(adev); + kgd_gfx_v9_release_queue(adev); return r; } @@ -365,13 +365,13 @@ int kgd_gfx_v9_hqd_dump(struct amdgpu_device *adev, if (*dump == NULL) return -ENOMEM; - acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR); reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++) DUMP_REG(reg); - release_queue(adev); + kgd_gfx_v9_release_queue(adev); WARN_ON_ONCE(i != HQD_N_REGS); *n_regs = i; @@ -487,7 +487,7 @@ bool kgd_gfx_v9_hqd_is_occupied(struct amdgpu_device *adev, bool retval = false; uint32_t low, high; - acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); act = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE); if (act) { low = lower_32_bits(queue_address >> 8); @@ -497,7 +497,7 @@ bool kgd_gfx_v9_hqd_is_occupied(struct amdgpu_device *adev, high == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI)) retval = true; } - release_queue(adev); + kgd_gfx_v9_release_queue(adev); return retval; } @@ -532,7 +532,7 @@ int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd, if (amdgpu_in_reset(adev)) return -EIO; - acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); if (m->cp_hqd_vmid == 0) WREG32_FIELD15_RLC(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0); @@ -561,13 +561,13 @@ int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd, break; if (time_after(jiffies, end_jiffies)) { pr_err("cp queue preemption time out.\n"); - release_queue(adev); + kgd_gfx_v9_release_queue(adev); return -ETIME; } usleep_range(500, 1000); } - release_queue(adev); + kgd_gfx_v9_release_queue(adev); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index c7ed3bc9053c..491273a02e30 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -58,3 +58,8 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, int *pasid_wave_cnt, int *max_waves_per_cu); void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev, uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr); +void kgd_gfx_v9_acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t queue_id); +uint64_t kgd_gfx_v9_get_queue_mask(struct amdgpu_device *adev, + uint32_t pipe_id, uint32_t queue_id); +void kgd_gfx_v9_release_queue(struct amdgpu_device *adev); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 00f528eb9812..1510041a6ee1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -51,6 +51,7 @@ extern const struct kfd2kgd_calls gfx_v8_kfd2kgd; extern const struct kfd2kgd_calls gfx_v9_kfd2kgd; extern const struct kfd2kgd_calls arcturus_kfd2kgd; extern const struct kfd2kgd_calls aldebaran_kfd2kgd; +extern const struct kfd2kgd_calls gc_9_4_3_kfd2kgd; extern const struct kfd2kgd_calls gfx_v10_kfd2kgd; extern const struct kfd2kgd_calls gfx_v10_3_kfd2kgd; extern const struct kfd2kgd_calls gfx_v11_kfd2kgd; @@ -328,7 +329,7 @@ struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) break; case IP_VERSION(9, 4, 3): gfx_target_version = 90400; - f2g = &aldebaran_kfd2kgd; + f2g = &gc_9_4_3_kfd2kgd; break; /* Navi10 */ case IP_VERSION(10, 1, 10): -- cgit v1.2.3 From bfb44eacb0e2de63bc7824cc590ede51a02a7ded Mon Sep 17 00:00:00 2001 From: Amber Lin Date: Thu, 27 Jan 2022 09:02:16 -0500 Subject: drm/amdkfd: Set F8_MODE for gc_v9_4_3 Set F8_MODE for GC 9.4.3 as optimal/non-IEEE. Also update gc_v9_0 to gc_v9_4_3 to include more definitions such as the F8_MODE bit, and remove unused header files. v2: fix IP version check (Alex) Signed-off-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c index 8b2dd2670ab7..914d94679d73 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c @@ -24,9 +24,7 @@ #include "kfd_device_queue_manager.h" #include "vega10_enum.h" -#include "gc/gc_9_0_offset.h" -#include "gc/gc_9_0_sh_mask.h" -#include "sdma0/sdma0_4_0_sh_mask.h" +#include "gc/gc_9_4_3_sh_mask.h" static int update_qpd_v9(struct device_queue_manager *dqm, struct qcm_process_device *qpd); @@ -65,6 +63,10 @@ static int update_qpd_v9(struct device_queue_manager *dqm, if (dqm->dev->noretry && !dqm->dev->use_iommu_v2) qpd->sh_mem_config |= 1 << SH_MEM_CONFIG__RETRY_DISABLE__SHIFT; + if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 3)) + qpd->sh_mem_config |= + (1 << SH_MEM_CONFIG__F8_MODE__SHIFT); + qpd->sh_mem_ape1_limit = 0; qpd->sh_mem_ape1_base = 0; } -- cgit v1.2.3 From 8dc1db3172ae2f17ae71e33b608a33411ce8a1aa Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Wed, 14 Sep 2022 16:39:48 +0800 Subject: drm/amdkfd: Introduce kfd_node struct (v5) Introduce a new structure, kfd_node, which will now represent a compute node. kfd_node is carved out of kfd_dev structure. kfd_dev struct now will become the parent of kfd_node, and will store common resources such as doorbells, GTT sub-alloctor etc. kfd_node struct will store all resources specific to a compute node, such as device queue manager, interrupt handling etc. This is the first step in adding compute partition support in KFD. v2: introduce kfd_node struct to gc v11 (Hawking) v3: make reference to kfd_dev struct through kfd_node (Morris) v4: use kfd_node instead for kfd isr/mqd functions (Morris) v5: rebase (Alex) Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Hawking Zhang Signed-off-by: Morris Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 1 + drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 43 ++-- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 28 +-- drivers/gpu/drm/amd/amdkfd/kfd_crat.h | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 258 +++++++++++++-------- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 100 ++++---- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 4 +- .../drm/amd/amdkfd/kfd_device_queue_manager_v9.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c | 14 +- drivers/gpu/drm/amd/amdkfd/kfd_events.c | 12 +- drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c | 12 +- drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c | 10 +- drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c | 64 ++--- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 22 +- drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c | 18 +- drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 18 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 24 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 10 +- drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 168 ++++++++------ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 54 ++--- .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 20 +- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 40 ++-- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 56 ++--- drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 8 +- 38 files changed, 573 insertions(+), 495 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index de6ba0d4b860..af37f2ef4438 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -35,6 +35,7 @@ #include "amdgpu_dma_buf.h" #include #include "amdgpu_xgmi.h" +#include "kfd_priv.h" #include "kfd_smi_events.h" #include diff --git a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c index 5c8023cba196..4ebfff6b6c55 100644 --- a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +++ b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c @@ -26,7 +26,7 @@ #include "amdgpu_amdkfd.h" #include "kfd_smi_events.h" -static bool cik_event_interrupt_isr(struct kfd_dev *dev, +static bool cik_event_interrupt_isr(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *patched_flag) @@ -85,7 +85,7 @@ static bool cik_event_interrupt_isr(struct kfd_dev *dev, !amdgpu_no_queue_eviction_on_vm_fault); } -static void cik_event_interrupt_wq(struct kfd_dev *dev, +static void cik_event_interrupt_wq(struct kfd_node *dev, const uint32_t *ih_ring_entry) { const struct cik_ih_ring_entry *ihre = diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 81d07ecf666d..eb0b0b38f10e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -293,7 +293,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, void *data) { struct kfd_ioctl_create_queue_args *args = data; - struct kfd_dev *dev; + struct kfd_node *dev; int err = 0; unsigned int queue_id; struct kfd_process_device *pdd; @@ -328,7 +328,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, } if (!pdd->doorbell_index && - kfd_alloc_process_doorbells(dev, &pdd->doorbell_index) < 0) { + kfd_alloc_process_doorbells(dev->kfd, &pdd->doorbell_index) < 0) { err = -ENOMEM; goto err_alloc_doorbells; } @@ -336,7 +336,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, /* Starting with GFX11, wptr BOs must be mapped to GART for MES to determine work * on unmapped queues for usermode queue oversubscription (no aggregated doorbell) */ - if (dev->shared_resources.enable_mes && + if (dev->kfd->shared_resources.enable_mes && ((dev->adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >> AMDGPU_MES_API_VERSION_SHIFT) >= 2) { struct amdgpu_bo_va_mapping *wptr_mapping; @@ -887,7 +887,7 @@ static int kfd_ioctl_set_scratch_backing_va(struct file *filep, { struct kfd_ioctl_set_scratch_backing_va_args *args = data; struct kfd_process_device *pdd; - struct kfd_dev *dev; + struct kfd_node *dev; long err; mutex_lock(&p->mutex); @@ -1006,18 +1006,18 @@ err_drm_file: return ret; } -bool kfd_dev_is_large_bar(struct kfd_dev *dev) +bool kfd_dev_is_large_bar(struct kfd_node *dev) { if (debug_largebar) { pr_debug("Simulate large-bar allocation on non large-bar machine\n"); return true; } - if (dev->use_iommu_v2) + if (dev->kfd->use_iommu_v2) return false; - if (dev->local_mem_info.local_mem_size_private == 0 && - dev->local_mem_info.local_mem_size_public > 0) + if (dev->kfd->local_mem_info.local_mem_size_private == 0 && + dev->kfd->local_mem_info.local_mem_size_public > 0) return true; return false; } @@ -1041,7 +1041,7 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, struct kfd_ioctl_alloc_memory_of_gpu_args *args = data; struct kfd_process_device *pdd; void *mem; - struct kfd_dev *dev; + struct kfd_node *dev; int idr_handle; long err; uint64_t offset = args->mmap_offset; @@ -1105,7 +1105,7 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, } if (flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) { - if (args->size != kfd_doorbell_process_slice(dev)) { + if (args->size != kfd_doorbell_process_slice(dev->kfd)) { err = -EINVAL; goto err_unlock; } @@ -1231,7 +1231,7 @@ static int kfd_ioctl_map_memory_to_gpu(struct file *filep, struct kfd_ioctl_map_memory_to_gpu_args *args = data; struct kfd_process_device *pdd, *peer_pdd; void *mem; - struct kfd_dev *dev; + struct kfd_node *dev; long err = 0; int i; uint32_t *devices_arr = NULL; @@ -1405,7 +1405,7 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, args->n_success = i+1; } - flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev); + flush_tlb = kfd_flush_tlb_after_unmap(pdd->dev->kfd); if (flush_tlb) { err = amdgpu_amdkfd_gpuvm_sync_memory(pdd->dev->adev, (struct kgd_mem *) mem, true); @@ -1445,7 +1445,7 @@ static int kfd_ioctl_alloc_queue_gws(struct file *filep, int retval; struct kfd_ioctl_alloc_queue_gws_args *args = data; struct queue *q; - struct kfd_dev *dev; + struct kfd_node *dev; mutex_lock(&p->mutex); q = pqm_get_user_queue(&p->pqm, args->queue_id); @@ -1482,7 +1482,7 @@ static int kfd_ioctl_get_dmabuf_info(struct file *filep, struct kfd_process *p, void *data) { struct kfd_ioctl_get_dmabuf_info_args *args = data; - struct kfd_dev *dev = NULL; + struct kfd_node *dev = NULL; struct amdgpu_device *dmabuf_adev; void *metadata_buffer = NULL; uint32_t flags; @@ -1596,7 +1596,7 @@ static int kfd_ioctl_export_dmabuf(struct file *filep, struct kfd_ioctl_export_dmabuf_args *args = data; struct kfd_process_device *pdd; struct dma_buf *dmabuf; - struct kfd_dev *dev; + struct kfd_node *dev; void *mem; int ret = 0; @@ -2178,7 +2178,7 @@ static int criu_restore_devices(struct kfd_process *p, } for (i = 0; i < args->num_devices; i++) { - struct kfd_dev *dev; + struct kfd_node *dev; struct kfd_process_device *pdd; struct file *drm_file; @@ -2240,7 +2240,7 @@ static int criu_restore_devices(struct kfd_process *p, } if (!pdd->doorbell_index && - kfd_alloc_process_doorbells(pdd->dev, &pdd->doorbell_index) < 0) { + kfd_alloc_process_doorbells(pdd->dev->kfd, &pdd->doorbell_index) < 0) { ret = -ENOMEM; goto exit; } @@ -2268,7 +2268,8 @@ static int criu_restore_memory_of_gpu(struct kfd_process_device *pdd, u64 offset; if (bo_bucket->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) { - if (bo_bucket->size != kfd_doorbell_process_slice(pdd->dev)) + if (bo_bucket->size != + kfd_doorbell_process_slice(pdd->dev->kfd)) return -EINVAL; offset = kfd_get_process_doorbells(pdd); @@ -2350,7 +2351,7 @@ static int criu_restore_bo(struct kfd_process *p, /* now map these BOs to GPU/s */ for (j = 0; j < p->n_pdds; j++) { - struct kfd_dev *peer; + struct kfd_node *peer; struct kfd_process_device *peer_pdd; if (!bo_priv->mapped_gpuids[j]) @@ -2947,7 +2948,7 @@ err_i1: return retcode; } -static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process, +static int kfd_mmio_mmap(struct kfd_node *dev, struct kfd_process *process, struct vm_area_struct *vma) { phys_addr_t address; @@ -2981,7 +2982,7 @@ static int kfd_mmio_mmap(struct kfd_dev *dev, struct kfd_process *process, static int kfd_mmap(struct file *filp, struct vm_area_struct *vma) { struct kfd_process *process; - struct kfd_dev *dev = NULL; + struct kfd_node *dev = NULL; unsigned long mmap_offset; unsigned int gpu_id; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index 475e47027354..f5aebba31e88 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -1405,7 +1405,7 @@ static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev, return i; } -int kfd_get_gpu_cache_info(struct kfd_dev *kdev, struct kfd_gpu_cache_info **pcache_info) +int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info) { int num_of_cache_types = 0; @@ -1524,7 +1524,7 @@ int kfd_get_gpu_cache_info(struct kfd_dev *kdev, struct kfd_gpu_cache_info **pca case IP_VERSION(11, 0, 3): case IP_VERSION(11, 0, 4): num_of_cache_types = - kfd_fill_gpu_cache_info_from_gfx_config(kdev, *pcache_info); + kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info); break; default: *pcache_info = dummy_cache_info; @@ -1858,7 +1858,7 @@ static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size) } static int kfd_fill_gpu_memory_affinity(int *avail_size, - struct kfd_dev *kdev, uint8_t type, uint64_t size, + struct kfd_node *kdev, uint8_t type, uint64_t size, struct crat_subtype_memory *sub_type_hdr, uint32_t proximity_domain, const struct kfd_local_mem_info *local_mem_info) @@ -1887,7 +1887,7 @@ static int kfd_fill_gpu_memory_affinity(int *avail_size, } #ifdef CONFIG_ACPI_NUMA -static void kfd_find_numa_node_in_srat(struct kfd_dev *kdev) +static void kfd_find_numa_node_in_srat(struct kfd_node *kdev) { struct acpi_table_header *table_header = NULL; struct acpi_subtable_header *sub_header = NULL; @@ -1982,7 +1982,7 @@ static void kfd_find_numa_node_in_srat(struct kfd_dev *kdev) * Return 0 if successful else return -ve value */ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, - struct kfd_dev *kdev, + struct kfd_node *kdev, struct crat_subtype_iolink *sub_type_hdr, uint32_t proximity_domain) { @@ -2044,8 +2044,8 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, } static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, - struct kfd_dev *kdev, - struct kfd_dev *peer_kdev, + struct kfd_node *kdev, + struct kfd_node *peer_kdev, struct crat_subtype_iolink *sub_type_hdr, uint32_t proximity_domain_from, uint32_t proximity_domain_to) @@ -2081,7 +2081,7 @@ static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, * [OUT] actual size of data filled in crat_image */ static int kfd_create_vcrat_image_gpu(void *pcrat_image, - size_t *size, struct kfd_dev *kdev, + size_t *size, struct kfd_node *kdev, uint32_t proximity_domain) { struct crat_header *crat_table = (struct crat_header *)pcrat_image; @@ -2153,7 +2153,7 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, /* Check if this node supports IOMMU. During parsing this flag will * translate to HSA_CAP_ATS_PRESENT */ - if (!kfd_iommu_check_device(kdev)) + if (!kfd_iommu_check_device(kdev->kfd)) cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT; crat_table->length += sub_type_hdr->length; @@ -2164,7 +2164,7 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, * report the total FB size (public+private) as a single * private heap. */ - local_mem_info = kdev->local_mem_info; + local_mem_info = kdev->kfd->local_mem_info; sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + sub_type_hdr->length); @@ -2216,12 +2216,12 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, * (from other GPU to this GPU) will be added * in kfd_parse_subtype_iolink. */ - if (kdev->hive_id) { + if (kdev->kfd->hive_id) { for (nid = 0; nid < proximity_domain; ++nid) { peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid); if (!peer_dev->gpu) continue; - if (peer_dev->gpu->hive_id != kdev->hive_id) + if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id) continue; sub_type_hdr = (typeof(sub_type_hdr))( (char *)sub_type_hdr + @@ -2255,12 +2255,12 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU * -- this option is not currently implemented. * The assumption is that all AMD APUs will have CRAT - * @kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU + * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU * * Return 0 if successful else return -ve value */ int kfd_create_crat_image_virtual(void **crat_image, size_t *size, - int flags, struct kfd_dev *kdev, + int flags, struct kfd_node *kdev, uint32_t proximity_domain) { void *pcrat_image = NULL; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.h b/drivers/gpu/drm/amd/amdkfd/kfd_crat.h index 8d1e8ba58dee..3d0e533b93b9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.h @@ -293,7 +293,7 @@ struct crat_subtype_generic { #pragma pack() -struct kfd_dev; +struct kfd_node; /* Static table to describe GPU Cache information */ struct kfd_gpu_cache_info { @@ -305,14 +305,14 @@ struct kfd_gpu_cache_info { */ uint32_t num_cu_shared; }; -int kfd_get_gpu_cache_info(struct kfd_dev *kdev, struct kfd_gpu_cache_info **pcache_info); +int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info); int kfd_create_crat_image_acpi(void **crat_image, size_t *size); void kfd_destroy_crat_image(void *crat_image); int kfd_parse_crat_table(void *crat_image, struct list_head *device_list, uint32_t proximity_domain); int kfd_create_crat_image_virtual(void **crat_image, size_t *size, - int flags, struct kfd_dev *kdev, + int flags, struct kfd_node *kdev, uint32_t proximity_domain); #endif /* KFD_CRAT_H_INCLUDED */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c b/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c index ad5a40a685ac..4a5a0a4e00f2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debugfs.c @@ -43,7 +43,7 @@ static int kfd_debugfs_hang_hws_read(struct seq_file *m, void *data) static ssize_t kfd_debugfs_hang_hws_write(struct file *file, const char __user *user_buf, size_t size, loff_t *ppos) { - struct kfd_dev *dev; + struct kfd_node *dev; char tmp[16]; uint32_t gpu_id; int ret = -EINVAL; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 1510041a6ee1..23d9a7f77055 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -61,7 +61,7 @@ static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size, static void kfd_gtt_sa_fini(struct kfd_dev *kfd); static int kfd_resume_iommu(struct kfd_dev *kfd); -static int kfd_resume(struct kfd_dev *kfd); +static int kfd_resume(struct kfd_node *kfd); static void kfd_device_info_set_sdma_info(struct kfd_dev *kfd) { @@ -441,8 +441,6 @@ struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) memset(&kfd->doorbell_available_index, 0, sizeof(kfd->doorbell_available_index)); - atomic_set(&kfd->sram_ecc_flag, 0); - ida_init(&kfd->doorbell_ida); return kfd; @@ -489,41 +487,106 @@ static void kfd_cwsr_init(struct kfd_dev *kfd) } } -static int kfd_gws_init(struct kfd_dev *kfd) +static int kfd_gws_init(struct kfd_node *node) { int ret = 0; + struct kfd_dev *kfd = node->kfd; - if (kfd->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) + if (node->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) return 0; - if (hws_gws_support || (KFD_IS_SOC15(kfd) && - ((KFD_GC_VERSION(kfd) == IP_VERSION(9, 0, 1) + if (hws_gws_support || (KFD_IS_SOC15(node) && + ((KFD_GC_VERSION(node) == IP_VERSION(9, 0, 1) && kfd->mec2_fw_version >= 0x81b3) || - (KFD_GC_VERSION(kfd) <= IP_VERSION(9, 4, 0) + (KFD_GC_VERSION(node) <= IP_VERSION(9, 4, 0) && kfd->mec2_fw_version >= 0x1b3) || - (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 1) + (KFD_GC_VERSION(node) == IP_VERSION(9, 4, 1) && kfd->mec2_fw_version >= 0x30) || - (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2) + (KFD_GC_VERSION(node) == IP_VERSION(9, 4, 2) && kfd->mec2_fw_version >= 0x28) || - (KFD_GC_VERSION(kfd) >= IP_VERSION(10, 3, 0) - && KFD_GC_VERSION(kfd) < IP_VERSION(11, 0, 0) + (KFD_GC_VERSION(node) >= IP_VERSION(10, 3, 0) + && KFD_GC_VERSION(node) < IP_VERSION(11, 0, 0) && kfd->mec2_fw_version >= 0x6b)))) - ret = amdgpu_amdkfd_alloc_gws(kfd->adev, - kfd->adev->gds.gws_size, &kfd->gws); + ret = amdgpu_amdkfd_alloc_gws(node->adev, + node->adev->gds.gws_size, &node->gws); return ret; } -static void kfd_smi_init(struct kfd_dev *dev) +static void kfd_smi_init(struct kfd_node *dev) { INIT_LIST_HEAD(&dev->smi_clients); spin_lock_init(&dev->smi_lock); } +static int kfd_init_node(struct kfd_node *node) +{ + int err = -1; + + if (kfd_interrupt_init(node)) { + dev_err(kfd_device, "Error initializing interrupts\n"); + goto kfd_interrupt_error; + } + + node->dqm = device_queue_manager_init(node); + if (!node->dqm) { + dev_err(kfd_device, "Error initializing queue manager\n"); + goto device_queue_manager_error; + } + + if (kfd_gws_init(node)) { + dev_err(kfd_device, "Could not allocate %d gws\n", + node->adev->gds.gws_size); + goto gws_error; + } + + if (kfd_resume(node)) + goto kfd_resume_error; + + if (kfd_topology_add_device(node)) { + dev_err(kfd_device, "Error adding device to topology\n"); + goto kfd_topology_add_device_error; + } + + kfd_smi_init(node); + + return 0; + +kfd_topology_add_device_error: +kfd_resume_error: +gws_error: + device_queue_manager_uninit(node->dqm); +device_queue_manager_error: + kfd_interrupt_exit(node); +kfd_interrupt_error: + if (node->gws) + amdgpu_amdkfd_free_gws(node->adev, node->gws); + + /* Cleanup the node memory here */ + kfree(node); + return err; +} + +static void kfd_cleanup_node(struct kfd_dev *kfd) +{ + struct kfd_node *knode = kfd->node; + + device_queue_manager_uninit(knode->dqm); + kfd_interrupt_exit(knode); + kfd_topology_remove_device(knode); + if (knode->gws) + amdgpu_amdkfd_free_gws(knode->adev, knode->gws); + kfree(knode); + kfd->node = NULL; +} + bool kgd2kfd_device_init(struct kfd_dev *kfd, const struct kgd2kfd_shared_resources *gpu_resources) { unsigned int size, map_process_packet_size; + struct kfd_node *node; + uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; + unsigned int max_proc_per_quantum; kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, KGD_ENGINE_MEC1); @@ -533,10 +596,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, KGD_ENGINE_SDMA1); kfd->shared_resources = *gpu_resources; - kfd->vm_info.first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1; - kfd->vm_info.last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1; - kfd->vm_info.vmid_num_kfd = kfd->vm_info.last_vmid_kfd - - kfd->vm_info.first_vmid_kfd + 1; + first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1; + last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1; + vmid_num_kfd = last_vmid_kfd - first_vmid_kfd + 1; /* Allow BIF to recode atomics to PCIe 3.0 AtomicOps. * 32 and 64-bit requests are possible and must be @@ -557,9 +619,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, /* Verify module parameters regarding mapped process number*/ if (hws_max_conc_proc >= 0) - kfd->max_proc_per_quantum = min((u32)hws_max_conc_proc, kfd->vm_info.vmid_num_kfd); + max_proc_per_quantum = min((u32)hws_max_conc_proc, vmid_num_kfd); else - kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd; + max_proc_per_quantum = vmid_num_kfd; /* calculate max size of mqds needed for queues */ size = max_num_of_queues_per_device * @@ -609,26 +671,6 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd->noretry = kfd->adev->gmc.noretry; - if (kfd_interrupt_init(kfd)) { - dev_err(kfd_device, "Error initializing interrupts\n"); - goto kfd_interrupt_error; - } - - kfd->dqm = device_queue_manager_init(kfd); - if (!kfd->dqm) { - dev_err(kfd_device, "Error initializing queue manager\n"); - goto device_queue_manager_error; - } - - /* If supported on this device, allocate global GWS that is shared - * by all KFD processes - */ - if (kfd_gws_init(kfd)) { - dev_err(kfd_device, "Could not allocate %d gws\n", - kfd->adev->gds.gws_size); - goto gws_error; - } - /* If CRAT is broken, won't set iommu enabled */ kfd_double_confirm_iommu_support(kfd); @@ -642,46 +684,54 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, svm_migrate_init(kfd->adev); - if (kfd_resume_iommu(kfd)) - goto device_iommu_error; - - if (kfd_resume(kfd)) - goto kfd_resume_error; - - amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); + /* Allocate the KFD node */ + node = kzalloc(sizeof(struct kfd_node), GFP_KERNEL); + if (!node) { + dev_err(kfd_device, "Error allocating KFD node\n"); + goto node_alloc_error; + } - if (kfd_topology_add_device(kfd)) { - dev_err(kfd_device, "Error adding device to topology\n"); - goto kfd_topology_add_device_error; + node->adev = kfd->adev; + node->kfd = kfd; + node->kfd2kgd = kfd->kfd2kgd; + node->vm_info.vmid_num_kfd = vmid_num_kfd; + node->vm_info.first_vmid_kfd = first_vmid_kfd; + node->vm_info.last_vmid_kfd = last_vmid_kfd; + node->max_proc_per_quantum = max_proc_per_quantum; + atomic_set(&node->sram_ecc_flag, 0); + + /* Initialize the KFD node */ + if (kfd_init_node(node)) { + dev_err(kfd_device, "Error initializing KFD node\n"); + goto node_init_error; } + kfd->node = node; - kfd_smi_init(kfd); + if (kfd_resume_iommu(kfd)) + goto kfd_resume_iommu_error; + + amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); kfd->init_complete = true; dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor, kfd->adev->pdev->device); pr_debug("Starting kfd with the following scheduling policy %d\n", - kfd->dqm->sched_policy); + node->dqm->sched_policy); goto out; -kfd_topology_add_device_error: -kfd_resume_error: +kfd_resume_iommu_error: + kfd_cleanup_node(kfd); +node_init_error: +node_alloc_error: device_iommu_error: -gws_error: - device_queue_manager_uninit(kfd->dqm); -device_queue_manager_error: - kfd_interrupt_exit(kfd); -kfd_interrupt_error: kfd_doorbell_fini(kfd); kfd_doorbell_error: kfd_gtt_sa_fini(kfd); kfd_gtt_sa_init_error: amdgpu_amdkfd_free_gtt_mem(kfd->adev, kfd->gtt_mem); alloc_gtt_mem_failure: - if (kfd->gws) - amdgpu_amdkfd_free_gws(kfd->adev, kfd->gws); dev_err(kfd_device, "device %x:%x NOT added due to errors\n", kfd->adev->pdev->vendor, kfd->adev->pdev->device); @@ -692,15 +742,11 @@ out: void kgd2kfd_device_exit(struct kfd_dev *kfd) { if (kfd->init_complete) { - device_queue_manager_uninit(kfd->dqm); - kfd_interrupt_exit(kfd); - kfd_topology_remove_device(kfd); + kfd_cleanup_node(kfd); kfd_doorbell_fini(kfd); ida_destroy(&kfd->doorbell_ida); kfd_gtt_sa_fini(kfd); amdgpu_amdkfd_free_gtt_mem(kfd->adev, kfd->gtt_mem); - if (kfd->gws) - amdgpu_amdkfd_free_gws(kfd->adev, kfd->gws); } kfree(kfd); @@ -708,16 +754,18 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd) int kgd2kfd_pre_reset(struct kfd_dev *kfd) { + struct kfd_node *node = kfd->node; + if (!kfd->init_complete) return 0; - kfd_smi_event_update_gpu_reset(kfd, false); + kfd_smi_event_update_gpu_reset(node, false); - kfd->dqm->ops.pre_reset(kfd->dqm); + node->dqm->ops.pre_reset(node->dqm); kgd2kfd_suspend(kfd, false); - kfd_signal_reset_event(kfd); + kfd_signal_reset_event(node); return 0; } @@ -730,18 +778,19 @@ int kgd2kfd_pre_reset(struct kfd_dev *kfd) int kgd2kfd_post_reset(struct kfd_dev *kfd) { int ret; + struct kfd_node *node = kfd->node; if (!kfd->init_complete) return 0; - ret = kfd_resume(kfd); + ret = kfd_resume(node); if (ret) return ret; atomic_dec(&kfd_locked); - atomic_set(&kfd->sram_ecc_flag, 0); + atomic_set(&node->sram_ecc_flag, 0); - kfd_smi_event_update_gpu_reset(kfd, true); + kfd_smi_event_update_gpu_reset(node, true); return 0; } @@ -753,6 +802,8 @@ bool kfd_is_locked(void) void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) { + struct kfd_node *node = kfd->node; + if (!kfd->init_complete) return; @@ -763,18 +814,19 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) kfd_suspend_all_processes(); } - kfd->dqm->ops.stop(kfd->dqm); + node->dqm->ops.stop(node->dqm); kfd_iommu_suspend(kfd); } int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm) { int ret, count; + struct kfd_node *node = kfd->node; if (!kfd->init_complete) return 0; - ret = kfd_resume(kfd); + ret = kfd_resume(node); if (ret) return ret; @@ -809,15 +861,15 @@ static int kfd_resume_iommu(struct kfd_dev *kfd) return err; } -static int kfd_resume(struct kfd_dev *kfd) +static int kfd_resume(struct kfd_node *node) { int err = 0; - err = kfd->dqm->ops.start(kfd->dqm); + err = node->dqm->ops.start(node->dqm); if (err) dev_err(kfd_device, "Error starting queue manager for device %x:%x\n", - kfd->adev->pdev->vendor, kfd->adev->pdev->device); + node->adev->pdev->vendor, node->adev->pdev->device); return err; } @@ -843,6 +895,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE]; bool is_patched = false; unsigned long flags; + struct kfd_node *node = kfd->node; if (!kfd->init_complete) return; @@ -852,16 +905,16 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) return; } - spin_lock_irqsave(&kfd->interrupt_lock, flags); + spin_lock_irqsave(&node->interrupt_lock, flags); - if (kfd->interrupts_active - && interrupt_is_wanted(kfd, ih_ring_entry, + if (node->interrupts_active + && interrupt_is_wanted(node, ih_ring_entry, patched_ihre, &is_patched) - && enqueue_ih_ring_entry(kfd, + && enqueue_ih_ring_entry(node, is_patched ? patched_ihre : ih_ring_entry)) - kfd_queue_work(kfd->ih_wq, &kfd->interrupt_work); + kfd_queue_work(node->ih_wq, &node->interrupt_work); - spin_unlock_irqrestore(&kfd->interrupt_lock, flags); + spin_unlock_irqrestore(&node->interrupt_lock, flags); } int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger) @@ -999,10 +1052,11 @@ static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr, return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size); } -int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, +int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size, struct kfd_mem_obj **mem_obj) { unsigned int found, start_search, cur_size; + struct kfd_dev *kfd = node->kfd; if (size == 0) return -EINVAL; @@ -1102,8 +1156,10 @@ kfd_gtt_no_free_chunk: return -ENOMEM; } -int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) +int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj) { + struct kfd_dev *kfd = node->kfd; + /* Act like kfree when trying to free a NULL object */ if (!mem_obj) return 0; @@ -1126,28 +1182,28 @@ int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd) { if (kfd) - atomic_inc(&kfd->sram_ecc_flag); + atomic_inc(&kfd->node->sram_ecc_flag); } -void kfd_inc_compute_active(struct kfd_dev *kfd) +void kfd_inc_compute_active(struct kfd_node *node) { - if (atomic_inc_return(&kfd->compute_profile) == 1) - amdgpu_amdkfd_set_compute_idle(kfd->adev, false); + if (atomic_inc_return(&node->kfd->compute_profile) == 1) + amdgpu_amdkfd_set_compute_idle(node->adev, false); } -void kfd_dec_compute_active(struct kfd_dev *kfd) +void kfd_dec_compute_active(struct kfd_node *node) { - int count = atomic_dec_return(&kfd->compute_profile); + int count = atomic_dec_return(&node->kfd->compute_profile); if (count == 0) - amdgpu_amdkfd_set_compute_idle(kfd->adev, true); + amdgpu_amdkfd_set_compute_idle(node->adev, true); WARN_ONCE(count < 0, "Compute profile ref. count error"); } void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) { if (kfd && kfd->init_complete) - kfd_smi_event_update_thermal_throttling(kfd, throttle_bitmask); + kfd_smi_event_update_thermal_throttling(kfd->node, throttle_bitmask); } /* kfd_get_num_sdma_engines returns the number of PCIe optimized SDMA and @@ -1155,19 +1211,19 @@ void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) * When the device has more than two engines, we reserve two for PCIe to enable * full-duplex and the rest are used as XGMI. */ -unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev) +unsigned int kfd_get_num_sdma_engines(struct kfd_node *node) { /* If XGMI is not supported, all SDMA engines are PCIe */ - if (!kdev->adev->gmc.xgmi.supported) - return kdev->adev->sdma.num_instances; + if (!node->adev->gmc.xgmi.supported) + return node->adev->sdma.num_instances; - return min(kdev->adev->sdma.num_instances, 2); + return min(node->adev->sdma.num_instances, 2); } -unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev) +unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *node) { /* After reserved for PCIe, the rest of engines are XGMI */ - return kdev->adev->sdma.num_instances - kfd_get_num_sdma_engines(kdev); + return node->adev->sdma.num_instances - kfd_get_num_sdma_engines(node); } #if defined(CONFIG_DEBUG_FS) @@ -1175,7 +1231,7 @@ unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev) /* This function will send a package to HIQ to hang the HWS * which will trigger a GPU reset and bring the HWS back to normal state */ -int kfd_debugfs_hang_hws(struct kfd_dev *dev) +int kfd_debugfs_hang_hws(struct kfd_node *dev) { if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) { pr_err("HWS is not enabled"); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 7a95698d83f7..34977d89f01c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -74,31 +74,31 @@ enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type) static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe) { int i; - int pipe_offset = (mec * dqm->dev->shared_resources.num_pipe_per_mec - + pipe) * dqm->dev->shared_resources.num_queue_per_pipe; + int pipe_offset = (mec * dqm->dev->kfd->shared_resources.num_pipe_per_mec + + pipe) * dqm->dev->kfd->shared_resources.num_queue_per_pipe; /* queue is available for KFD usage if bit is 1 */ - for (i = 0; i < dqm->dev->shared_resources.num_queue_per_pipe; ++i) + for (i = 0; i < dqm->dev->kfd->shared_resources.num_queue_per_pipe; ++i) if (test_bit(pipe_offset + i, - dqm->dev->shared_resources.cp_queue_bitmap)) + dqm->dev->kfd->shared_resources.cp_queue_bitmap)) return true; return false; } unsigned int get_cp_queues_num(struct device_queue_manager *dqm) { - return bitmap_weight(dqm->dev->shared_resources.cp_queue_bitmap, + return bitmap_weight(dqm->dev->kfd->shared_resources.cp_queue_bitmap, KGD_MAX_QUEUES); } unsigned int get_queues_per_pipe(struct device_queue_manager *dqm) { - return dqm->dev->shared_resources.num_queue_per_pipe; + return dqm->dev->kfd->shared_resources.num_queue_per_pipe; } unsigned int get_pipes_per_mec(struct device_queue_manager *dqm) { - return dqm->dev->shared_resources.num_pipe_per_mec; + return dqm->dev->kfd->shared_resources.num_pipe_per_mec; } static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm) @@ -110,18 +110,18 @@ static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm) unsigned int get_num_sdma_queues(struct device_queue_manager *dqm) { return kfd_get_num_sdma_engines(dqm->dev) * - dqm->dev->device_info.num_sdma_queues_per_engine; + dqm->dev->kfd->device_info.num_sdma_queues_per_engine; } unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm) { return kfd_get_num_xgmi_sdma_engines(dqm->dev) * - dqm->dev->device_info.num_sdma_queues_per_engine; + dqm->dev->kfd->device_info.num_sdma_queues_per_engine; } static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manager *dqm) { - return dqm->dev->device_info.reserved_sdma_queues_bitmap; + return dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap; } void program_sh_mem_settings(struct device_queue_manager *dqm, @@ -330,7 +330,7 @@ static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q, uint32_t const *restore_id) { - struct kfd_dev *dev = qpd->dqm->dev; + struct kfd_node *dev = qpd->dqm->dev; if (!KFD_IS_SOC15(dev)) { /* On pre-SOC15 chips we need to use the queue ID to @@ -349,7 +349,7 @@ static int allocate_doorbell(struct qcm_process_device *qpd, * for a SDMA engine is 512. */ - uint32_t *idx_offset = dev->shared_resources.sdma_doorbell_idx; + uint32_t *idx_offset = dev->kfd->shared_resources.sdma_doorbell_idx; uint32_t valid_id = idx_offset[q->properties.sdma_engine_id] + (q->properties.sdma_queue_id & 1) * KFD_QUEUE_DOORBELL_MIRROR_OFFSET @@ -382,7 +382,7 @@ static int allocate_doorbell(struct qcm_process_device *qpd, } q->properties.doorbell_off = - kfd_get_doorbell_dw_offset_in_bar(dev, qpd_to_pdd(qpd), + kfd_get_doorbell_dw_offset_in_bar(dev->kfd, qpd_to_pdd(qpd), q->doorbell_id); return 0; } @@ -391,7 +391,7 @@ static void deallocate_doorbell(struct qcm_process_device *qpd, struct queue *q) { unsigned int old; - struct kfd_dev *dev = qpd->dqm->dev; + struct kfd_node *dev = qpd->dqm->dev; if (!KFD_IS_SOC15(dev) || q->properties.type == KFD_QUEUE_TYPE_SDMA || @@ -441,7 +441,7 @@ static int allocate_vmid(struct device_queue_manager *dqm, program_sh_mem_settings(dqm, qpd); - if (KFD_IS_SOC15(dqm->dev) && dqm->dev->cwsr_enabled) + if (KFD_IS_SOC15(dqm->dev) && dqm->dev->kfd->cwsr_enabled) program_trap_handler_settings(dqm, qpd); /* qpd->page_table_base is set earlier when register_process() @@ -460,7 +460,7 @@ static int allocate_vmid(struct device_queue_manager *dqm, return 0; } -static int flush_texture_cache_nocpsch(struct kfd_dev *kdev, +static int flush_texture_cache_nocpsch(struct kfd_node *kdev, struct qcm_process_device *qpd) { const struct packet_manager_funcs *pmf = qpd->dqm->packet_mgr.pmf; @@ -661,7 +661,7 @@ static inline void deallocate_hqd(struct device_queue_manager *dqm, #define SQ_IND_CMD_CMD_KILL 0x00000003 #define SQ_IND_CMD_MODE_BROADCAST 0x00000001 -static int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p) +static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process *p) { int status = 0; unsigned int vmid; @@ -837,7 +837,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, /* Make sure the queue is unmapped before updating the MQD */ if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false); else if (prev_active) @@ -858,7 +858,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, } retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, - (dqm->dev->cwsr_enabled ? + (dqm->dev->kfd->cwsr_enabled ? KFD_PREEMPT_TYPE_WAVEFRONT_SAVE : KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN), KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); @@ -895,7 +895,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, } if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) retval = map_queues_cpsch(dqm); else if (q->properties.is_active) retval = add_queue_mes(dqm, q, &pdd->qpd); @@ -951,7 +951,7 @@ static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, continue; retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, - (dqm->dev->cwsr_enabled ? + (dqm->dev->kfd->cwsr_enabled ? KFD_PREEMPT_TYPE_WAVEFRONT_SAVE : KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN), KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); @@ -993,7 +993,7 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, q->properties.is_active = false; decrement_queue_count(dqm, qpd, q); - if (dqm->dev->shared_resources.enable_mes) { + if (dqm->dev->kfd->shared_resources.enable_mes) { retval = remove_queue_mes(dqm, q, qpd); if (retval) { pr_err("Failed to evict queue %d\n", @@ -1003,7 +1003,7 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, } } pdd->last_evict_timestamp = get_jiffies_64(); - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, qpd->is_debug ? KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : @@ -1132,7 +1132,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, q->properties.is_active = true; increment_queue_count(dqm, &pdd->qpd, q); - if (dqm->dev->shared_resources.enable_mes) { + if (dqm->dev->kfd->shared_resources.enable_mes) { retval = add_queue_mes(dqm, q, qpd); if (retval) { pr_err("Failed to restore queue %d\n", @@ -1141,7 +1141,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, } } } - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); qpd->evicted = 0; @@ -1282,7 +1282,7 @@ static int initialize_nocpsch(struct device_queue_manager *dqm) for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) if (test_bit(pipe_offset + queue, - dqm->dev->shared_resources.cp_queue_bitmap)) + dqm->dev->kfd->shared_resources.cp_queue_bitmap)) dqm->allocated_queues[pipe] |= 1 << queue; } @@ -1426,14 +1426,14 @@ static int set_sched_resources(struct device_queue_manager *dqm) int i, mec; struct scheduling_resources res; - res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap; + res.vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; res.queue_mask = 0; for (i = 0; i < KGD_MAX_QUEUES; ++i) { - mec = (i / dqm->dev->shared_resources.num_queue_per_pipe) - / dqm->dev->shared_resources.num_pipe_per_mec; + mec = (i / dqm->dev->kfd->shared_resources.num_queue_per_pipe) + / dqm->dev->kfd->shared_resources.num_pipe_per_mec; - if (!test_bit(i, dqm->dev->shared_resources.cp_queue_bitmap)) + if (!test_bit(i, dqm->dev->kfd->shared_resources.cp_queue_bitmap)) continue; /* only acquire queues from the first MEC */ @@ -1489,7 +1489,7 @@ static int start_cpsch(struct device_queue_manager *dqm) dqm_lock(dqm); - if (!dqm->dev->shared_resources.enable_mes) { + if (!dqm->dev->kfd->shared_resources.enable_mes) { retval = pm_init(&dqm->packet_mgr, dqm); if (retval) goto fail_packet_manager_init; @@ -1516,14 +1516,14 @@ static int start_cpsch(struct device_queue_manager *dqm) dqm->is_hws_hang = false; dqm->is_resetting = false; dqm->sched_running = true; - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); dqm_unlock(dqm); return 0; fail_allocate_vidmem: fail_set_sched_resources: - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) pm_uninit(&dqm->packet_mgr, false); fail_packet_manager_init: dqm_unlock(dqm); @@ -1541,7 +1541,7 @@ static int stop_cpsch(struct device_queue_manager *dqm) } if (!dqm->is_hws_hang) { - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, false); else remove_all_queues_mes(dqm); @@ -1550,11 +1550,11 @@ static int stop_cpsch(struct device_queue_manager *dqm) hanging = dqm->is_hws_hang || dqm->is_resetting; dqm->sched_running = false; - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) pm_release_ib(&dqm->packet_mgr); kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) pm_uninit(&dqm->packet_mgr, hanging); dqm_unlock(dqm); @@ -1673,7 +1673,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, if (q->properties.is_active) { increment_queue_count(dqm, qpd, q); - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); else @@ -1893,7 +1893,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, list_del(&q->list); qpd->queue_count--; if (q->properties.is_active) { - if (!dqm->dev->shared_resources.enable_mes) { + if (!dqm->dev->kfd->shared_resources.enable_mes) { decrement_queue_count(dqm, qpd, q); retval = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); @@ -2056,7 +2056,7 @@ static int get_wave_state(struct device_queue_manager *dqm, mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP]; if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE || - q->properties.is_active || !q->device->cwsr_enabled || + q->properties.is_active || !q->device->kfd->cwsr_enabled || !mqd_mgr->get_wave_state) { dqm_unlock(dqm); return -EINVAL; @@ -2105,7 +2105,7 @@ static int checkpoint_mqd(struct device_queue_manager *dqm, dqm_lock(dqm); - if (q->properties.is_active || !q->device->cwsr_enabled) { + if (q->properties.is_active || !q->device->kfd->cwsr_enabled) { r = -EINVAL; goto dqm_unlock; } @@ -2158,7 +2158,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, if (q->properties.is_active) { decrement_queue_count(dqm, qpd, q); - if (dqm->dev->shared_resources.enable_mes) { + if (dqm->dev->kfd->shared_resources.enable_mes) { retval = remove_queue_mes(dqm, q, qpd); if (retval) pr_err("Failed to remove queue %d\n", @@ -2180,7 +2180,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, } } - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, filter, 0); if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { @@ -2242,11 +2242,11 @@ out_free: static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) { int retval; - struct kfd_dev *dev = dqm->dev; + struct kfd_node *dev = dqm->dev; struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd; uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * get_num_all_sdma_engines(dqm) * - dev->device_info.num_sdma_queues_per_engine + + dev->kfd->device_info.num_sdma_queues_per_engine + dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size, @@ -2256,7 +2256,7 @@ static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) return retval; } -struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) +struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev) { struct device_queue_manager *dqm; @@ -2373,7 +2373,7 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) if (init_mqd_managers(dqm)) goto out_free; - if (!dev->shared_resources.enable_mes && allocate_hiq_sdma_mqd(dqm)) { + if (!dev->kfd->shared_resources.enable_mes && allocate_hiq_sdma_mqd(dqm)) { pr_err("Failed to allocate hiq sdma mqd trunk buffer\n"); goto out_free; } @@ -2386,7 +2386,7 @@ out_free: return NULL; } -static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev, +static void deallocate_hiq_sdma_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd) { WARN(!mqd, "No hiq sdma mqd trunk to free"); @@ -2397,7 +2397,7 @@ static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev, void device_queue_manager_uninit(struct device_queue_manager *dqm) { dqm->ops.uninitialize(dqm); - if (!dqm->dev->shared_resources.enable_mes) + if (!dqm->dev->kfd->shared_resources.enable_mes) deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd); kfree(dqm); } @@ -2479,7 +2479,7 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { if (!test_bit(pipe_offset + queue, - dqm->dev->shared_resources.cp_queue_bitmap)) + dqm->dev->kfd->shared_resources.cp_queue_bitmap)) continue; r = dqm->dev->kfd2kgd->hqd_dump( @@ -2497,7 +2497,7 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) for (pipe = 0; pipe < get_num_all_sdma_engines(dqm); pipe++) { for (queue = 0; - queue < dqm->dev->device_info.num_sdma_queues_per_engine; + queue < dqm->dev->kfd->device_info.num_sdma_queues_per_engine; queue++) { r = dqm->dev->kfd2kgd->hqd_sdma_dump( dqm->dev->adev, pipe, queue, &dump, &n_regs); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index a537b9ef3e16..e554a48f3054 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -207,7 +207,7 @@ struct device_queue_manager_asic_ops { struct queue *q, struct qcm_process_device *qpd); struct mqd_manager * (*mqd_manager_init)(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); }; /** @@ -228,7 +228,7 @@ struct device_queue_manager { struct mqd_manager *mqd_mgrs[KFD_MQD_TYPE_MAX]; struct packet_manager packet_mgr; - struct kfd_dev *dev; + struct kfd_node *dev; struct mutex lock_hidden; /* use dqm_lock/unlock(dqm) */ struct list_head queues; unsigned int saved_flags; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c index 914d94679d73..8af643388768 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_v9.c @@ -60,10 +60,10 @@ static int update_qpd_v9(struct device_queue_manager *dqm, qpd->sh_mem_config = SH_MEM_ALIGNMENT_MODE_UNALIGNED << SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT; - if (dqm->dev->noretry && !dqm->dev->use_iommu_v2) + if (dqm->dev->kfd->noretry && !dqm->dev->kfd->use_iommu_v2) qpd->sh_mem_config |= 1 << SH_MEM_CONFIG__RETRY_DISABLE__SHIFT; - if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 3)) + if (KFD_GC_VERSION(dqm->dev->kfd) == IP_VERSION(9, 4, 3)) qpd->sh_mem_config |= (1 << SH_MEM_CONFIG__F8_MODE__SHIFT); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c index 38c9e1ca6691..6421b620388d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_doorbell.c @@ -138,7 +138,7 @@ void kfd_doorbell_fini(struct kfd_dev *kfd) iounmap(kfd->doorbell_kernel_ptr); } -int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, +int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process, struct vm_area_struct *vma) { phys_addr_t address; @@ -148,7 +148,7 @@ int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, * For simplicitly we only allow mapping of the entire doorbell * allocation of a single device & process. */ - if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev)) + if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev->kfd)) return -EINVAL; pdd = kfd_get_process_device_data(dev, process); @@ -170,13 +170,13 @@ int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, " vm_flags == 0x%04lX\n" " size == 0x%04lX\n", (unsigned long long) vma->vm_start, address, vma->vm_flags, - kfd_doorbell_process_slice(dev)); + kfd_doorbell_process_slice(dev->kfd)); return io_remap_pfn_range(vma, vma->vm_start, address >> PAGE_SHIFT, - kfd_doorbell_process_slice(dev), + kfd_doorbell_process_slice(dev->kfd), vma->vm_page_prot); } @@ -278,14 +278,14 @@ uint64_t kfd_get_number_elems(struct kfd_dev *kfd) phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd) { if (!pdd->doorbell_index) { - int r = kfd_alloc_process_doorbells(pdd->dev, + int r = kfd_alloc_process_doorbells(pdd->dev->kfd, &pdd->doorbell_index); if (r < 0) return 0; } - return pdd->dev->doorbell_base + - pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev); + return pdd->dev->kfd->doorbell_base + + pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev->kfd); } int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_events.c index c894cf8f7c50..9926186f88a6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c @@ -348,7 +348,7 @@ static int kfd_event_page_set(struct kfd_process *p, void *kernel_address, int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset) { - struct kfd_dev *kfd; + struct kfd_node *kfd; struct kfd_process_device *pdd; void *mem, *kern_addr; uint64_t size; @@ -1125,7 +1125,7 @@ static void lookup_events_by_type_and_signal(struct kfd_process *p, } #ifdef KFD_SUPPORT_IOMMU_V2 -void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid, +void kfd_signal_iommu_event(struct kfd_node *dev, u32 pasid, unsigned long address, bool is_write_requested, bool is_execute_requested) { @@ -1221,8 +1221,8 @@ void kfd_signal_hw_exception_event(u32 pasid) kfd_unref_process(p); } -void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, - struct kfd_vm_fault_info *info) +void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, + struct kfd_vm_fault_info *info) { struct kfd_event *ev; uint32_t id; @@ -1269,7 +1269,7 @@ void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, kfd_unref_process(p); } -void kfd_signal_reset_event(struct kfd_dev *dev) +void kfd_signal_reset_event(struct kfd_node *dev) { struct kfd_hsa_hw_exception_data hw_exception_data; struct kfd_hsa_memory_exception_data memory_exception_data; @@ -1325,7 +1325,7 @@ void kfd_signal_reset_event(struct kfd_dev *dev) srcu_read_unlock(&kfd_processes_srcu, idx); } -void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid) +void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid) { struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); struct kfd_hsa_memory_exception_data memory_exception_data; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c index 8aebe408c544..da2ca00d79e5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c @@ -322,21 +322,21 @@ static void kfd_init_apertures_vi(struct kfd_process_device *pdd, uint8_t id) pdd->lds_base = MAKE_LDS_APP_BASE_VI(); pdd->lds_limit = MAKE_LDS_APP_LIMIT(pdd->lds_base); - if (!pdd->dev->use_iommu_v2) { + if (!pdd->dev->kfd->use_iommu_v2) { /* dGPUs: SVM aperture starting at 0 * with small reserved space for kernel. * Set them to CANONICAL addresses. */ pdd->gpuvm_base = SVM_USER_BASE; pdd->gpuvm_limit = - pdd->dev->shared_resources.gpuvm_size - 1; + pdd->dev->kfd->shared_resources.gpuvm_size - 1; } else { /* set them to non CANONICAL addresses, and no SVM is * allocated. */ pdd->gpuvm_base = MAKE_GPUVM_APP_BASE_VI(id + 1); pdd->gpuvm_limit = MAKE_GPUVM_APP_LIMIT(pdd->gpuvm_base, - pdd->dev->shared_resources.gpuvm_size); + pdd->dev->kfd->shared_resources.gpuvm_size); } pdd->scratch_base = MAKE_SCRATCH_APP_BASE_VI(); @@ -356,7 +356,7 @@ static void kfd_init_apertures_v9(struct kfd_process_device *pdd, uint8_t id) */ pdd->gpuvm_base = SVM_USER_BASE; pdd->gpuvm_limit = - pdd->dev->shared_resources.gpuvm_size - 1; + pdd->dev->kfd->shared_resources.gpuvm_size - 1; pdd->scratch_base = MAKE_SCRATCH_APP_BASE_V9(); pdd->scratch_limit = MAKE_SCRATCH_APP_LIMIT(pdd->scratch_base); @@ -365,7 +365,7 @@ static void kfd_init_apertures_v9(struct kfd_process_device *pdd, uint8_t id) int kfd_init_apertures(struct kfd_process *process) { uint8_t id = 0; - struct kfd_dev *dev; + struct kfd_node *dev; struct kfd_process_device *pdd; /*Iterating over all devices*/ @@ -417,7 +417,7 @@ int kfd_init_apertures(struct kfd_process *process) } } - if (!dev->use_iommu_v2) { + if (!dev->kfd->use_iommu_v2) { /* dGPUs: the reserved space for kernel * before SVM */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c index 0d53f6067422..0f0fdea4cd8a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c @@ -187,7 +187,7 @@ static void print_sq_intr_info_error(uint32_t context_id0, uint32_t context_id1) REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_ERROR_CTXID1, WGP_ID)); } -static void event_interrupt_poison_consumption_v11(struct kfd_dev *dev, +static void event_interrupt_poison_consumption_v11(struct kfd_node *dev, uint16_t pasid, uint16_t source_id) { int ret = -EINVAL; @@ -225,7 +225,7 @@ static void event_interrupt_poison_consumption_v11(struct kfd_dev *dev, amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, true); } -static bool event_interrupt_isr_v11(struct kfd_dev *dev, +static bool event_interrupt_isr_v11(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *patched_flag) @@ -274,7 +274,7 @@ static bool event_interrupt_isr_v11(struct kfd_dev *dev, !amdgpu_no_queue_eviction_on_vm_fault); } -static void event_interrupt_wq_v11(struct kfd_dev *dev, +static void event_interrupt_wq_v11(struct kfd_node *dev, const uint32_t *ih_ring_entry) { uint16_t source_id, client_id, ring_id, pasid, vmid; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c index 0b75a37b689b..861bccb1e9dc 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c @@ -90,7 +90,7 @@ enum SQ_INTERRUPT_ERROR_TYPE { #define KFD_SQ_INT_DATA__ERR_TYPE_MASK 0xF00000 #define KFD_SQ_INT_DATA__ERR_TYPE__SHIFT 20 -static void event_interrupt_poison_consumption_v9(struct kfd_dev *dev, +static void event_interrupt_poison_consumption_v9(struct kfd_node *dev, uint16_t pasid, uint16_t client_id) { int old_poison, ret = -EINVAL; @@ -160,7 +160,7 @@ static bool context_id_expected(struct kfd_dev *dev) } } -static bool event_interrupt_isr_v9(struct kfd_dev *dev, +static bool event_interrupt_isr_v9(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *patched_flag) @@ -206,7 +206,7 @@ static bool event_interrupt_isr_v9(struct kfd_dev *dev, *patched_flag = true; memcpy(patched_ihre, ih_ring_entry, - dev->device_info.ih_ring_entry_size); + dev->kfd->device_info.ih_ring_entry_size); pasid = dev->dqm->vmid_pasid[vmid]; @@ -235,7 +235,7 @@ static bool event_interrupt_isr_v9(struct kfd_dev *dev, uint32_t context_id = SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); - if (context_id == 0 && context_id_expected(dev)) + if (context_id == 0 && context_id_expected(dev->kfd)) return false; } @@ -253,7 +253,7 @@ static bool event_interrupt_isr_v9(struct kfd_dev *dev, !amdgpu_no_queue_eviction_on_vm_fault); } -static void event_interrupt_wq_v9(struct kfd_dev *dev, +static void event_interrupt_wq_v9(struct kfd_node *dev, const uint32_t *ih_ring_entry) { uint16_t source_id, client_id, pasid, vmid; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c index 34772fe74296..dd3c43c1ad70 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_interrupt.c @@ -50,29 +50,29 @@ static void interrupt_wq(struct work_struct *); -int kfd_interrupt_init(struct kfd_dev *kfd) +int kfd_interrupt_init(struct kfd_node *node) { int r; - r = kfifo_alloc(&kfd->ih_fifo, - KFD_IH_NUM_ENTRIES * kfd->device_info.ih_ring_entry_size, + r = kfifo_alloc(&node->ih_fifo, + KFD_IH_NUM_ENTRIES * node->kfd->device_info.ih_ring_entry_size, GFP_KERNEL); if (r) { - dev_err(kfd->adev->dev, "Failed to allocate IH fifo\n"); + dev_err(node->adev->dev, "Failed to allocate IH fifo\n"); return r; } - kfd->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1); - if (unlikely(!kfd->ih_wq)) { - kfifo_free(&kfd->ih_fifo); - dev_err(kfd->adev->dev, "Failed to allocate KFD IH workqueue\n"); + node->ih_wq = alloc_workqueue("KFD IH", WQ_HIGHPRI, 1); + if (unlikely(!node->ih_wq)) { + kfifo_free(&node->ih_fifo); + dev_err(node->adev->dev, "Failed to allocate KFD IH workqueue\n"); return -ENOMEM; } - spin_lock_init(&kfd->interrupt_lock); + spin_lock_init(&node->interrupt_lock); - INIT_WORK(&kfd->interrupt_work, interrupt_wq); + INIT_WORK(&node->interrupt_work, interrupt_wq); - kfd->interrupts_active = true; + node->interrupts_active = true; /* * After this function returns, the interrupt will be enabled. This @@ -84,7 +84,7 @@ int kfd_interrupt_init(struct kfd_dev *kfd) return 0; } -void kfd_interrupt_exit(struct kfd_dev *kfd) +void kfd_interrupt_exit(struct kfd_node *node) { /* * Stop the interrupt handler from writing to the ring and scheduling @@ -93,31 +93,31 @@ void kfd_interrupt_exit(struct kfd_dev *kfd) */ unsigned long flags; - spin_lock_irqsave(&kfd->interrupt_lock, flags); - kfd->interrupts_active = false; - spin_unlock_irqrestore(&kfd->interrupt_lock, flags); + spin_lock_irqsave(&node->interrupt_lock, flags); + node->interrupts_active = false; + spin_unlock_irqrestore(&node->interrupt_lock, flags); /* * flush_work ensures that there are no outstanding * work-queue items that will access interrupt_ring. New work items * can't be created because we stopped interrupt handling above. */ - flush_workqueue(kfd->ih_wq); + flush_workqueue(node->ih_wq); - kfifo_free(&kfd->ih_fifo); + kfifo_free(&node->ih_fifo); } /* * Assumption: single reader/writer. This function is not re-entrant */ -bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry) +bool enqueue_ih_ring_entry(struct kfd_node *node, const void *ih_ring_entry) { int count; - count = kfifo_in(&kfd->ih_fifo, ih_ring_entry, - kfd->device_info.ih_ring_entry_size); - if (count != kfd->device_info.ih_ring_entry_size) { - dev_dbg_ratelimited(kfd->adev->dev, + count = kfifo_in(&node->ih_fifo, ih_ring_entry, + node->kfd->device_info.ih_ring_entry_size); + if (count != node->kfd->device_info.ih_ring_entry_size) { + dev_dbg_ratelimited(node->adev->dev, "Interrupt ring overflow, dropping interrupt %d\n", count); return false; @@ -129,32 +129,32 @@ bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry) /* * Assumption: single reader/writer. This function is not re-entrant */ -static bool dequeue_ih_ring_entry(struct kfd_dev *kfd, void *ih_ring_entry) +static bool dequeue_ih_ring_entry(struct kfd_node *node, void *ih_ring_entry) { int count; - count = kfifo_out(&kfd->ih_fifo, ih_ring_entry, - kfd->device_info.ih_ring_entry_size); + count = kfifo_out(&node->ih_fifo, ih_ring_entry, + node->kfd->device_info.ih_ring_entry_size); - WARN_ON(count && count != kfd->device_info.ih_ring_entry_size); + WARN_ON(count && count != node->kfd->device_info.ih_ring_entry_size); - return count == kfd->device_info.ih_ring_entry_size; + return count == node->kfd->device_info.ih_ring_entry_size; } static void interrupt_wq(struct work_struct *work) { - struct kfd_dev *dev = container_of(work, struct kfd_dev, + struct kfd_node *dev = container_of(work, struct kfd_node, interrupt_work); uint32_t ih_ring_entry[KFD_MAX_RING_ENTRY_SIZE]; unsigned long start_jiffies = jiffies; - if (dev->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) { + if (dev->kfd->device_info.ih_ring_entry_size > sizeof(ih_ring_entry)) { dev_err_once(dev->adev->dev, "Ring entry too small\n"); return; } while (dequeue_ih_ring_entry(dev, ih_ring_entry)) { - dev->device_info.event_interrupt_class->interrupt_wq(dev, + dev->kfd->device_info.event_interrupt_class->interrupt_wq(dev, ih_ring_entry); if (time_is_before_jiffies(start_jiffies + HZ)) { /* If we spent more than a second processing signals, @@ -166,14 +166,14 @@ static void interrupt_wq(struct work_struct *work) } } -bool interrupt_is_wanted(struct kfd_dev *dev, +bool interrupt_is_wanted(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *flag) { /* integer and bitwise OR so there is no boolean short-circuiting */ unsigned int wanted = 0; - wanted |= dev->device_info.event_interrupt_class->interrupt_isr(dev, + wanted |= dev->kfd->device_info.event_interrupt_class->interrupt_isr(dev, ih_ring_entry, patched_ihre, flag); return wanted != 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c b/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c index ec1bf611624e..6eee9a0944f3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c @@ -109,11 +109,11 @@ int kfd_iommu_device_init(struct kfd_dev *kfd) */ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd) { - struct kfd_dev *dev = pdd->dev; + struct kfd_node *dev = pdd->dev; struct kfd_process *p = pdd->process; int err; - if (!dev->use_iommu_v2 || pdd->bound == PDD_BOUND) + if (!dev->kfd->use_iommu_v2 || pdd->bound == PDD_BOUND) return 0; if (unlikely(pdd->bound == PDD_BOUND_SUSPENDED)) { @@ -146,7 +146,7 @@ void kfd_iommu_unbind_process(struct kfd_process *p) /* Callback for process shutdown invoked by the IOMMU driver */ static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, u32 pasid) { - struct kfd_dev *dev = kfd_device_by_pci_dev(pdev); + struct kfd_node *dev = kfd_device_by_pci_dev(pdev); struct kfd_process *p; struct kfd_process_device *pdd; @@ -182,7 +182,7 @@ static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, u32 pasid) static int iommu_invalid_ppr_cb(struct pci_dev *pdev, u32 pasid, unsigned long address, u16 flags) { - struct kfd_dev *dev; + struct kfd_node *dev; dev_warn_ratelimited(kfd_device, "Invalid PPR device %x:%x.%x pasid 0x%x address 0x%lX flags 0x%X", @@ -205,7 +205,7 @@ static int iommu_invalid_ppr_cb(struct pci_dev *pdev, u32 pasid, * Bind processes do the device that have been temporarily unbound * (PDD_BOUND_SUSPENDED) in kfd_unbind_processes_from_device. */ -static int kfd_bind_processes_to_device(struct kfd_dev *kfd) +static int kfd_bind_processes_to_device(struct kfd_node *knode) { struct kfd_process_device *pdd; struct kfd_process *p; @@ -216,14 +216,14 @@ static int kfd_bind_processes_to_device(struct kfd_dev *kfd) hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { mutex_lock(&p->mutex); - pdd = kfd_get_process_device_data(kfd, p); + pdd = kfd_get_process_device_data(knode, p); if (WARN_ON(!pdd) || pdd->bound != PDD_BOUND_SUSPENDED) { mutex_unlock(&p->mutex); continue; } - err = amd_iommu_bind_pasid(kfd->adev->pdev, p->pasid, + err = amd_iommu_bind_pasid(knode->adev->pdev, p->pasid, p->lead_thread); if (err < 0) { pr_err("Unexpected pasid 0x%x binding failure\n", @@ -246,7 +246,7 @@ static int kfd_bind_processes_to_device(struct kfd_dev *kfd) * processes will be restored to PDD_BOUND state in * kfd_bind_processes_to_device. */ -static void kfd_unbind_processes_from_device(struct kfd_dev *kfd) +static void kfd_unbind_processes_from_device(struct kfd_node *knode) { struct kfd_process_device *pdd; struct kfd_process *p; @@ -256,7 +256,7 @@ static void kfd_unbind_processes_from_device(struct kfd_dev *kfd) hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { mutex_lock(&p->mutex); - pdd = kfd_get_process_device_data(kfd, p); + pdd = kfd_get_process_device_data(knode, p); if (WARN_ON(!pdd)) { mutex_unlock(&p->mutex); @@ -281,7 +281,7 @@ void kfd_iommu_suspend(struct kfd_dev *kfd) if (!kfd->use_iommu_v2) return; - kfd_unbind_processes_from_device(kfd); + kfd_unbind_processes_from_device(kfd->node); amd_iommu_set_invalidate_ctx_cb(kfd->adev->pdev, NULL); amd_iommu_set_invalid_ppr_cb(kfd->adev->pdev, NULL); @@ -312,7 +312,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd) amd_iommu_set_invalid_ppr_cb(kfd->adev->pdev, iommu_invalid_ppr_cb); - err = kfd_bind_processes_to_device(kfd); + err = kfd_bind_processes_to_device(kfd->node); if (err) { amd_iommu_set_invalidate_ctx_cb(kfd->adev->pdev, NULL); amd_iommu_set_invalid_ppr_cb(kfd->adev->pdev, NULL); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c index bcf7bc3302c9..1bea629c49ca 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c @@ -38,7 +38,7 @@ /* Initialize a kernel queue, including allocations of GART memory * needed for the queue. */ -static bool kq_initialize(struct kernel_queue *kq, struct kfd_dev *dev, +static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev, enum kfd_queue_type type, unsigned int queue_size) { struct queue_properties prop; @@ -75,7 +75,7 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_dev *dev, if (!kq->mqd_mgr) return false; - prop.doorbell_ptr = kfd_get_kernel_doorbell(dev, &prop.doorbell_off); + prop.doorbell_ptr = kfd_get_kernel_doorbell(dev->kfd, &prop.doorbell_off); if (!prop.doorbell_ptr) { pr_err("Failed to initialize doorbell"); @@ -112,7 +112,7 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_dev *dev, kq->rptr_kernel = kq->rptr_mem->cpu_ptr; kq->rptr_gpu_addr = kq->rptr_mem->gpu_addr; - retval = kfd_gtt_sa_allocate(dev, dev->device_info.doorbell_size, + retval = kfd_gtt_sa_allocate(dev, dev->kfd->device_info.doorbell_size, &kq->wptr_mem); if (retval != 0) @@ -189,7 +189,7 @@ err_rptr_allocate_vidmem: err_eop_allocate_vidmem: kfd_gtt_sa_free(dev, kq->pq); err_pq_allocate_vidmem: - kfd_release_kernel_doorbell(dev, prop.doorbell_ptr); + kfd_release_kernel_doorbell(dev->kfd, prop.doorbell_ptr); err_get_kernel_doorbell: return false; @@ -220,7 +220,7 @@ static void kq_uninitialize(struct kernel_queue *kq, bool hanging) kfd_gtt_sa_free(kq->dev, kq->eop_mem); kfd_gtt_sa_free(kq->dev, kq->pq); - kfd_release_kernel_doorbell(kq->dev, + kfd_release_kernel_doorbell(kq->dev->kfd, kq->queue->properties.doorbell_ptr); uninit_queue(kq->queue); } @@ -298,7 +298,7 @@ void kq_submit_packet(struct kernel_queue *kq) } pr_debug("\n"); #endif - if (kq->dev->device_info.doorbell_size == 8) { + if (kq->dev->kfd->device_info.doorbell_size == 8) { *kq->wptr64_kernel = kq->pending_wptr64; write_kernel_doorbell64(kq->queue->properties.doorbell_ptr, kq->pending_wptr64); @@ -311,7 +311,7 @@ void kq_submit_packet(struct kernel_queue *kq) void kq_rollback_packet(struct kernel_queue *kq) { - if (kq->dev->device_info.doorbell_size == 8) { + if (kq->dev->kfd->device_info.doorbell_size == 8) { kq->pending_wptr64 = *kq->wptr64_kernel; kq->pending_wptr = *kq->wptr_kernel % (kq->queue->properties.queue_size / 4); @@ -320,7 +320,7 @@ void kq_rollback_packet(struct kernel_queue *kq) } } -struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, +struct kernel_queue *kernel_queue_init(struct kfd_node *dev, enum kfd_queue_type type) { struct kernel_queue *kq; @@ -345,7 +345,7 @@ void kernel_queue_uninit(struct kernel_queue *kq, bool hanging) } /* FIXME: Can this test be removed? */ -static __attribute__((unused)) void test_kq(struct kfd_dev *dev) +static __attribute__((unused)) void test_kq(struct kfd_node *dev) { struct kernel_queue *kq; uint32_t *buffer, i; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h index 383202fd1ea2..9a6244430845 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.h @@ -53,7 +53,7 @@ void kq_rollback_packet(struct kernel_queue *kq); struct kernel_queue { /* data */ - struct kfd_dev *dev; + struct kfd_node *dev; struct mqd_manager *mqd_mgr; struct queue *queue; uint64_t pending_wptr64; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 54933903bcb8..1e187677c90a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -423,7 +423,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - 0, adev->kfd.dev->id, prange->prefetch_loc, + 0, adev->kfd.dev->node->id, prange->prefetch_loc, prange->preferred_loc, trigger); r = migrate_vma_setup(&migrate); @@ -456,7 +456,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - 0, adev->kfd.dev->id, trigger); + 0, adev->kfd.dev->node->id, trigger); svm_range_dma_unmap(adev->dev, scratch, 0, npages); svm_range_free_dma_mappings(prange); @@ -701,7 +701,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - adev->kfd.dev->id, 0, prange->prefetch_loc, + adev->kfd.dev->node->id, 0, prange->prefetch_loc, prange->preferred_loc, trigger); r = migrate_vma_setup(&migrate); @@ -737,7 +737,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - adev->kfd.dev->id, 0, trigger); + adev->kfd.dev->node->id, 0, trigger); svm_range_dma_unmap(adev->dev, scratch, 0, npages); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c index 623ccd227b7d..61f6dd68c84b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c @@ -46,7 +46,7 @@ int pipe_priority_map[] = { KFD_PIPE_PRIORITY_CS_HIGH }; -struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_dev *dev, struct queue_properties *q) +struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev, struct queue_properties *q) { struct kfd_mem_obj *mqd_mem_obj = NULL; @@ -61,7 +61,7 @@ struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_dev *dev, struct queue_propertie return mqd_mem_obj; } -struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_dev *dev, +struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev, struct queue_properties *q) { struct kfd_mem_obj *mqd_mem_obj = NULL; @@ -72,7 +72,7 @@ struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_dev *dev, return NULL; offset = (q->sdma_engine_id * - dev->device_info.num_sdma_queues_per_engine + + dev->kfd->device_info.num_sdma_queues_per_engine + q->sdma_queue_id) * dev->dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h index 57f900ccaa10..46fc3f273d0d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h @@ -68,7 +68,7 @@ */ extern int pipe_priority_map[]; struct mqd_manager { - struct kfd_mem_obj* (*allocate_mqd)(struct kfd_dev *kfd, + struct kfd_mem_obj* (*allocate_mqd)(struct kfd_node *kfd, struct queue_properties *q); void (*init_mqd)(struct mqd_manager *mm, void **mqd, @@ -121,14 +121,14 @@ struct mqd_manager { uint32_t (*read_doorbell_id)(void *mqd); struct mutex mqd_mutex; - struct kfd_dev *dev; + struct kfd_node *dev; uint32_t mqd_size; }; -struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_dev *dev, +struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev, struct queue_properties *q); -struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_dev *dev, +struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev, struct queue_properties *q); void free_mqd_hiq_sdma(struct mqd_manager *mm, void *mqd, struct kfd_mem_obj *mqd_mem_obj); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c index 4889865c725c..03e04d5e5a11 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c @@ -74,7 +74,7 @@ static void set_priority(struct cik_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, +static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd, struct queue_properties *q) { struct kfd_mem_obj *mqd_mem_obj; @@ -390,7 +390,7 @@ static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; @@ -470,7 +470,7 @@ struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, } struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index d3e2b6a599a4..7a93be0ebb19 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -74,7 +74,7 @@ static void set_priority(struct v10_compute_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, +static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd, struct queue_properties *q) { struct kfd_mem_obj *mqd_mem_obj; @@ -122,7 +122,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; } - if (mm->dev->cwsr_enabled) { + if (mm->dev->kfd->cwsr_enabled) { m->cp_hqd_persistent_state |= (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); m->cp_hqd_ctx_save_base_addr_lo = @@ -210,7 +210,7 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, m->cp_hqd_pq_doorbell_control |= 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; } - if (mm->dev->cwsr_enabled) + if (mm->dev->kfd->cwsr_enabled) m->cp_hqd_ctx_save_control = 0; update_cu_mask(mm, mqd, minfo); @@ -405,7 +405,7 @@ static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) #endif struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index 5aa75f72caa1..dff171b54b5c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -81,7 +81,7 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, +static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node, struct queue_properties *q) { struct kfd_mem_obj *mqd_mem_obj; @@ -91,12 +91,12 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, * MES write to areas beyond MQD size. So allocate * 1 PAGE_SIZE memory for MQD is MES is enabled. */ - if (kfd->shared_resources.enable_mes) + if (node->kfd->shared_resources.enable_mes) size = PAGE_SIZE; else size = sizeof(struct v11_compute_mqd); - if (kfd_gtt_sa_allocate(kfd, size, &mqd_mem_obj)) + if (kfd_gtt_sa_allocate(node, size, &mqd_mem_obj)) return NULL; return mqd_mem_obj; @@ -113,7 +113,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; - if (mm->dev->shared_resources.enable_mes) + if (mm->dev->kfd->shared_resources.enable_mes) size = PAGE_SIZE; else size = sizeof(struct v11_compute_mqd); @@ -155,7 +155,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; } - if (mm->dev->cwsr_enabled) { + if (mm->dev->kfd->cwsr_enabled) { m->cp_hqd_persistent_state |= (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); m->cp_hqd_ctx_save_base_addr_lo = @@ -243,7 +243,7 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, m->cp_hqd_pq_doorbell_control |= 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; } - if (mm->dev->cwsr_enabled) + if (mm->dev->kfd->cwsr_enabled) m->cp_hqd_ctx_save_control = 0; update_cu_mask(mm, mqd, minfo); @@ -319,7 +319,7 @@ static void init_mqd_sdma(struct mqd_manager *mm, void **mqd, m = (struct v11_sdma_mqd *) mqd_mem_obj->cpu_ptr; - if (mm->dev->shared_resources.enable_mes) + if (mm->dev->kfd->shared_resources.enable_mes) size = PAGE_SIZE; else size = sizeof(struct v11_sdma_mqd); @@ -387,7 +387,7 @@ static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) #endif struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; @@ -463,7 +463,7 @@ struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type, * To allocate SDMA MQDs by generic functions * when MES is enabled. */ - if (dev->shared_resources.enable_mes) { + if (dev->kfd->shared_resources.enable_mes) { mqd->allocate_mqd = allocate_mqd; mqd->free_mqd = kfd_free_mqd_cp; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 51b53110341b..943a738e73f9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -83,7 +83,7 @@ static void set_priority(struct v9_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, +static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node, struct queue_properties *q) { int retval; @@ -105,11 +105,11 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, * pass a special bo flag AMDGPU_GEM_CREATE_CP_MQD_GFX9 to instruct * amdgpu memory functions to do so. */ - if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) { + if (node->kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) { mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); if (!mqd_mem_obj) return NULL; - retval = amdgpu_amdkfd_alloc_gtt_mem(kfd->adev, + retval = amdgpu_amdkfd_alloc_gtt_mem(node->adev, ALIGN(q->ctl_stack_size, PAGE_SIZE) + ALIGN(sizeof(struct v9_mqd), PAGE_SIZE), &(mqd_mem_obj->gtt_mem), @@ -121,7 +121,7 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, return NULL; } } else { - retval = kfd_gtt_sa_allocate(kfd, sizeof(struct v9_mqd), + retval = kfd_gtt_sa_allocate(node, sizeof(struct v9_mqd), &mqd_mem_obj); if (retval) return NULL; @@ -136,7 +136,6 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, { uint64_t addr; struct v9_mqd *m; - struct amdgpu_device *adev = (struct amdgpu_device *)mm->dev->adev; m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; @@ -169,7 +168,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, if (q->format == KFD_QUEUE_FORMAT_AQL) { m->cp_hqd_aql_control = 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; - if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) { + if (KFD_GC_VERSION(mm->dev) == IP_VERSION(9, 4, 3)) { /* On GC 9.4.3, DW 41 is re-purposed as * compute_tg_chunk_size. * TODO: review this setting when active CUs in the @@ -179,7 +178,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, } } else { /* PM4 queue */ - if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) { + if (KFD_GC_VERSION(mm->dev) == IP_VERSION(9, 4, 3)) { m->compute_static_thread_mgmt_se6 = 0; /* TODO: program pm4_target_xcc */ } @@ -190,7 +189,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, (1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT); } - if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) { + if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address) { m->cp_hqd_persistent_state |= (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); m->cp_hqd_ctx_save_base_addr_lo = @@ -225,7 +224,6 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, struct queue_properties *q, struct mqd_update_info *minfo) { - struct amdgpu_device *adev = (struct amdgpu_device *)mm->dev->adev; struct v9_mqd *m; m = get_mqd(mqd); @@ -275,13 +273,13 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT | 1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT | 1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT; - if (adev->ip_versions[GC_HWIP][0] != IP_VERSION(9, 4, 3)) + if (KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 3)) m->cp_hqd_pq_control |= - CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; + CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; m->cp_hqd_pq_doorbell_control |= 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; } - if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) + if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address) m->cp_hqd_ctx_save_control = 0; update_cu_mask(mm, mqd, minfo); @@ -487,7 +485,7 @@ static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) #endif struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c index 530ba6f5b57e..f6b4a5686dcb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c @@ -77,7 +77,7 @@ static void set_priority(struct vi_mqd *m, struct queue_properties *q) m->cp_hqd_queue_priority = q->priority; } -static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd, +static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd, struct queue_properties *q) { struct kfd_mem_obj *mqd_mem_obj; @@ -136,7 +136,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, (1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT); } - if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) { + if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address) { m->cp_hqd_persistent_state |= (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT); m->cp_hqd_ctx_save_base_addr_lo = @@ -227,7 +227,7 @@ static void __update_mqd(struct mqd_manager *mm, void *mqd, 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT; } - if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) + if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address) m->cp_hqd_ctx_save_control = atc_bit << CP_HQD_CTX_SAVE_CONTROL__ATC__SHIFT | mtype << CP_HQD_CTX_SAVE_CONTROL__MTYPE__SHIFT; @@ -446,7 +446,7 @@ static int debugfs_show_mqd_sdma(struct seq_file *m, void *data) #endif struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; @@ -528,7 +528,7 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, } struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, - struct kfd_dev *dev) + struct kfd_node *dev) { struct mqd_manager *mqd; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c index f612325241aa..2f54172e9175 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c @@ -45,7 +45,7 @@ static void pm_calc_rlib_size(struct packet_manager *pm, unsigned int process_count, queue_count, compute_queue_count, gws_queue_count; unsigned int map_queue_size; unsigned int max_proc_per_quantum = 1; - struct kfd_dev *dev = pm->dqm->dev; + struct kfd_node *dev = pm->dqm->dev; process_count = pm->dqm->processes_count; queue_count = pm->dqm->active_queue_count; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 18250845a989..54d7d4665ad2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -119,7 +119,7 @@ static int pm_runlist_v9(struct packet_manager *pm, uint32_t *buffer, struct pm4_mes_runlist *packet; int concurrent_proc_cnt = 0; - struct kfd_dev *kfd = pm->dqm->dev; + struct kfd_node *kfd = pm->dqm->dev; /* Determine the number of processes to map together to HW: * it can not exceed the number of VMIDs available to the @@ -220,7 +220,8 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, case KFD_QUEUE_TYPE_SDMA: case KFD_QUEUE_TYPE_SDMA_XGMI: use_static = false; /* no static queues under SDMA */ - if (q->properties.sdma_engine_id < 2 && !pm_use_ext_eng(q->device)) + if (q->properties.sdma_engine_id < 2 && + !pm_use_ext_eng(q->device->kfd)) packet->bitfields2.engine_sel = q->properties.sdma_engine_id + engine_sel__mes_map_queues__sdma0_vi; else { @@ -263,7 +264,8 @@ static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer, packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES, sizeof(struct pm4_mes_unmap_queues)); - packet->bitfields2.extended_engine_sel = pm_use_ext_eng(pm->dqm->dev) ? + packet->bitfields2.extended_engine_sel = + pm_use_ext_eng(pm->dqm->dev->kfd) ? extended_engine_sel__mes_unmap_queues__sdma0_to_7_sel : extended_engine_sel__mes_unmap_queues__legacy_engine_sel; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c index 4f951eaa6ee8..faf4772ed317 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c @@ -77,7 +77,7 @@ static int pm_runlist_vi(struct packet_manager *pm, uint32_t *buffer, { struct pm4_mes_runlist *packet; int concurrent_proc_cnt = 0; - struct kfd_dev *kfd = pm->dqm->dev; + struct kfd_node *kfd = pm->dqm->dev; if (WARN_ON(!ib)) return -EFAULT; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 94a438956868..fdb97e5d0c01 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -210,11 +210,13 @@ enum cache_policy { ((KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) || \ (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3))) +struct kfd_node; + struct kfd_event_interrupt_class { - bool (*interrupt_isr)(struct kfd_dev *dev, + bool (*interrupt_isr)(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *patched_flag); - void (*interrupt_wq)(struct kfd_dev *dev, + void (*interrupt_wq)(struct kfd_node *dev, const uint32_t *ih_ring_entry); }; @@ -236,8 +238,8 @@ struct kfd_device_info { uint64_t reserved_sdma_queues_bitmap; }; -unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev); -unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev); +unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev); +unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev); struct kfd_mem_obj { uint32_t range_start; @@ -253,13 +255,59 @@ struct kfd_vmid_info { uint32_t vmid_num_kfd; }; +struct kfd_dev; + +struct kfd_node { + struct amdgpu_device *adev; /* Duplicated here along with keeping + * a copy in kfd_dev to save a hop + */ + const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with + * keeping a copy in kfd_dev to + * save a hop + */ + struct kfd_vmid_info vm_info; + unsigned int id; /* topology stub index */ + /* Interrupts */ + struct kfifo ih_fifo; + struct workqueue_struct *ih_wq; + struct work_struct interrupt_work; + spinlock_t interrupt_lock; + + /* + * Interrupts of interest to KFD are copied + * from the HW ring into a SW ring. + */ + bool interrupts_active; + + /* QCM Device instance */ + struct device_queue_manager *dqm; + + /* Global GWS resource shared between processes */ + void *gws; + bool gws_debug_workaround; + + /* Clients watching SMI events */ + struct list_head smi_clients; + spinlock_t smi_lock; + uint32_t reset_seq_num; + + /* SRAM ECC flag */ + atomic_t sram_ecc_flag; + + /*spm process id */ + unsigned int spm_pasid; + + /* Maximum process number mapped to HW scheduler */ + unsigned int max_proc_per_quantum; + + struct kfd_dev *kfd; +}; + struct kfd_dev { struct amdgpu_device *adev; struct kfd_device_info device_info; - unsigned int id; /* topology stub index */ - phys_addr_t doorbell_base; /* Start of actual doorbells used by * KFD. It is aligned for mapping * into user mode @@ -274,7 +322,6 @@ struct kfd_dev { */ struct kgd2kfd_shared_resources shared_resources; - struct kfd_vmid_info vm_info; struct kfd_local_mem_info local_mem_info; const struct kfd2kgd_calls *kfd2kgd; @@ -290,30 +337,13 @@ struct kfd_dev { unsigned int gtt_sa_chunk_size; unsigned int gtt_sa_num_of_chunks; - /* Interrupts */ - struct kfifo ih_fifo; - struct workqueue_struct *ih_wq; - struct work_struct interrupt_work; - spinlock_t interrupt_lock; - - /* QCM Device instance */ - struct device_queue_manager *dqm; - bool init_complete; - /* - * Interrupts of interest to KFD are copied - * from the HW ring into a SW ring. - */ - bool interrupts_active; /* Firmware versions */ uint16_t mec_fw_version; uint16_t mec2_fw_version; uint16_t sdma_fw_version; - /* Maximum process number mapped to HW scheduler */ - unsigned int max_proc_per_quantum; - /* CWSR */ bool cwsr_enabled; const void *cwsr_isa; @@ -327,21 +357,9 @@ struct kfd_dev { /* Use IOMMU v2 flag */ bool use_iommu_v2; - /* SRAM ECC flag */ - atomic_t sram_ecc_flag; - /* Compute Profile ref. count */ atomic_t compute_profile; - /* Global GWS resource shared between processes */ - void *gws; - - /* Clients watching SMI events */ - struct list_head smi_clients; - spinlock_t smi_lock; - - uint32_t reset_seq_num; - struct ida doorbell_ida; unsigned int max_doorbell_slices; @@ -349,6 +367,8 @@ struct kfd_dev { /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ struct dev_pagemap pgmap; + + struct kfd_node *node; }; enum kfd_mempool { @@ -563,7 +583,7 @@ struct queue { unsigned int doorbell_id; struct kfd_process *process; - struct kfd_dev *device; + struct kfd_node *device; void *gws; /* procfs */ @@ -697,7 +717,7 @@ enum kfd_pdd_bound { /* Data that is per-process-per device. */ struct kfd_process_device { /* The device that owns this data. */ - struct kfd_dev *dev; + struct kfd_node *dev; /* The process that owns this kfd_process_device. */ struct kfd_process *process; @@ -925,7 +945,7 @@ struct amdkfd_ioctl_desc { unsigned int cmd_drv; const char *name; }; -bool kfd_dev_is_large_bar(struct kfd_dev *dev); +bool kfd_dev_is_large_bar(struct kfd_node *dev); int kfd_process_create_wq(void); void kfd_process_destroy_wq(void); @@ -961,16 +981,16 @@ int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id); int kfd_process_device_init_vm(struct kfd_process_device *pdd, struct file *drm_file); -struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, +struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev, struct kfd_process *p); -struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, +struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev, struct kfd_process *p); -struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, +struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev, struct kfd_process *p); bool kfd_process_xnack_mode(struct kfd_process *p, bool supported); -int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, +int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process, struct vm_area_struct *vma); /* KFD process API for creating and translating handles */ @@ -994,7 +1014,7 @@ void kfd_pasid_free(u32 pasid); size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); int kfd_doorbell_init(struct kfd_dev *kfd); void kfd_doorbell_fini(struct kfd_dev *kfd); -int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, +int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process, struct vm_area_struct *vma); void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, unsigned int *doorbell_off); @@ -1012,10 +1032,10 @@ void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index); /* GTT Sub-Allocator */ -int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, +int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size, struct kfd_mem_obj **mem_obj); -int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); +int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj); extern struct device *kfd_device; @@ -1028,25 +1048,25 @@ void kfd_procfs_del_queue(struct queue *q); /* Topology */ int kfd_topology_init(void); void kfd_topology_shutdown(void); -int kfd_topology_add_device(struct kfd_dev *gpu); -int kfd_topology_remove_device(struct kfd_dev *gpu); +int kfd_topology_add_device(struct kfd_node *gpu); +int kfd_topology_remove_device(struct kfd_node *gpu); struct kfd_topology_device *kfd_topology_device_by_proximity_domain( uint32_t proximity_domain); struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( uint32_t proximity_domain); struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); -struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); -struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); -struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev); -int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); +struct kfd_node *kfd_device_by_id(uint32_t gpu_id); +struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev); +struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev); +int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev); int kfd_numa_node_to_apic_id(int numa_node_id); void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); /* Interrupts */ -int kfd_interrupt_init(struct kfd_dev *dev); -void kfd_interrupt_exit(struct kfd_dev *dev); -bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); -bool interrupt_is_wanted(struct kfd_dev *dev, +int kfd_interrupt_init(struct kfd_node *dev); +void kfd_interrupt_exit(struct kfd_node *dev); +bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry); +bool interrupt_is_wanted(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *flag); @@ -1174,22 +1194,22 @@ void print_queue_properties(struct queue_properties *q); void print_queue(struct queue *q); struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); + struct kfd_node *dev); struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type, - struct kfd_dev *dev); -struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); + struct kfd_node *dev); +struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev); void device_queue_manager_uninit(struct device_queue_manager *dqm); -struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, +struct kernel_queue *kernel_queue_init(struct kfd_node *dev, enum kfd_queue_type type); void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid); @@ -1206,7 +1226,7 @@ void kfd_process_dequeue_from_all_devices(struct kfd_process *p); int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); void pqm_uninit(struct process_queue_manager *pqm); int pqm_create_queue(struct process_queue_manager *pqm, - struct kfd_dev *dev, + struct kfd_node *dev, struct file *f, struct queue_properties *properties, unsigned int *qid, @@ -1323,7 +1343,7 @@ int kfd_wait_on_events(struct kfd_process *p, uint32_t *wait_result); void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, uint32_t valid_id_bits); -void kfd_signal_iommu_event(struct kfd_dev *dev, +void kfd_signal_iommu_event(struct kfd_node *dev, u32 pasid, unsigned long address, bool is_write_requested, bool is_execute_requested); void kfd_signal_hw_exception_event(u32 pasid); @@ -1339,12 +1359,12 @@ int kfd_event_create(struct file *devkfd, struct kfd_process *p, int kfd_get_num_events(struct kfd_process *p); int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); -void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, +void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, struct kfd_vm_fault_info *info); -void kfd_signal_reset_event(struct kfd_dev *dev); +void kfd_signal_reset_event(struct kfd_node *dev); -void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid); +void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid); void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type); @@ -1359,12 +1379,12 @@ static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) bool kfd_is_locked(void); /* Compute profile */ -void kfd_inc_compute_active(struct kfd_dev *dev); -void kfd_dec_compute_active(struct kfd_dev *dev); +void kfd_inc_compute_active(struct kfd_node *dev); +void kfd_dec_compute_active(struct kfd_node *dev); /* Cgroup Support */ /* Check with device cgroup if @kfd device is accessible */ -static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) +static inline int kfd_devcgroup_check_permission(struct kfd_node *kfd) { #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) struct drm_device *ddev = adev_to_drm(kfd->adev); @@ -1389,7 +1409,7 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data); int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); int pm_debugfs_runlist(struct seq_file *m, void *data); -int kfd_debugfs_hang_hws(struct kfd_dev *dev); +int kfd_debugfs_hang_hws(struct kfd_node *dev); int pm_debugfs_hang_hws(struct packet_manager *pm); int dqm_debugfs_hang_hws(struct device_queue_manager *dqm); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 07a9eaf9b7d8..66e021889c64 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -269,7 +269,7 @@ static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer) int cu_cnt; int wave_cnt; int max_waves_per_cu; - struct kfd_dev *dev = NULL; + struct kfd_node *dev = NULL; struct kfd_process *proc = NULL; struct kfd_process_device *pdd = NULL; @@ -691,7 +691,7 @@ void kfd_process_destroy_wq(void) static void kfd_process_free_gpuvm(struct kgd_mem *mem, struct kfd_process_device *pdd, void **kptr) { - struct kfd_dev *dev = pdd->dev; + struct kfd_node *dev = pdd->dev; if (kptr && *kptr) { amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem); @@ -713,7 +713,7 @@ static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd, uint64_t gpu_va, uint32_t size, uint32_t flags, struct kgd_mem **mem, void **kptr) { - struct kfd_dev *kdev = pdd->dev; + struct kfd_node *kdev = pdd->dev; int err; err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size, @@ -982,7 +982,7 @@ static void kfd_process_device_free_bos(struct kfd_process_device *pdd) static void kfd_process_kunmap_signal_bo(struct kfd_process *p) { struct kfd_process_device *pdd; - struct kfd_dev *kdev; + struct kfd_node *kdev; void *mem; kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle)); @@ -1040,9 +1040,9 @@ static void kfd_process_destroy_pdds(struct kfd_process *p) bitmap_free(pdd->qpd.doorbell_bitmap); idr_destroy(&pdd->alloc_idr); - kfd_free_process_doorbells(pdd->dev, pdd->doorbell_index); + kfd_free_process_doorbells(pdd->dev->kfd, pdd->doorbell_index); - if (pdd->dev->shared_resources.enable_mes) + if (pdd->dev->kfd->shared_resources.enable_mes) amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev, pdd->proc_ctx_bo); /* @@ -1259,10 +1259,10 @@ static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) int i; for (i = 0; i < p->n_pdds; i++) { - struct kfd_dev *dev = p->pdds[i]->dev; + struct kfd_node *dev = p->pdds[i]->dev; struct qcm_process_device *qpd = &p->pdds[i]->qpd; - if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base) + if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base) continue; offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id); @@ -1279,7 +1279,7 @@ static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) return err; } - memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size); + memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size); qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET; pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n", @@ -1291,7 +1291,7 @@ static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) { - struct kfd_dev *dev = pdd->dev; + struct kfd_node *dev = pdd->dev; struct qcm_process_device *qpd = &pdd->qpd; uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE @@ -1300,7 +1300,7 @@ static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) void *kaddr; int ret; - if (!dev->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base) + if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base) return 0; /* cwsr_base is only set for dGPU */ @@ -1313,7 +1313,7 @@ static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) qpd->cwsr_kaddr = kaddr; qpd->tba_addr = qpd->cwsr_base; - memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size); + memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size); qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET; pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n", @@ -1324,10 +1324,10 @@ static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd) { - struct kfd_dev *dev = pdd->dev; + struct kfd_node *dev = pdd->dev; struct qcm_process_device *qpd = &pdd->qpd; - if (!dev->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base) + if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base) return; kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr); @@ -1371,7 +1371,7 @@ bool kfd_process_xnack_mode(struct kfd_process *p, bool supported) * support retry. */ for (i = 0; i < p->n_pdds; i++) { - struct kfd_dev *dev = p->pdds[i]->dev; + struct kfd_node *dev = p->pdds[i]->dev; /* Only consider GFXv9 and higher GPUs. Older GPUs don't * support the SVM APIs and don't need to be considered @@ -1394,7 +1394,7 @@ bool kfd_process_xnack_mode(struct kfd_process *p, bool supported) if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1)) return false; - if (dev->noretry) + if (dev->kfd->noretry) return false; } @@ -1528,7 +1528,7 @@ static int init_doorbell_bitmap(struct qcm_process_device *qpd, return 0; } -struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, +struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev, struct kfd_process *p) { int i; @@ -1540,7 +1540,7 @@ struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, return NULL; } -struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, +struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev, struct kfd_process *p) { struct kfd_process_device *pdd = NULL; @@ -1552,7 +1552,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, if (!pdd) return NULL; - if (init_doorbell_bitmap(&pdd->qpd, dev)) { + if (init_doorbell_bitmap(&pdd->qpd, dev->kfd)) { pr_err("Failed to init doorbell for process\n"); goto err_free_pdd; } @@ -1573,7 +1573,7 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, pdd->user_gpu_id = dev->id; atomic64_set(&pdd->evict_duration_counter, 0); - if (dev->shared_resources.enable_mes) { + if (dev->kfd->shared_resources.enable_mes) { retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, AMDGPU_MES_PROC_CTX_SIZE, &pdd->proc_ctx_bo, @@ -1619,7 +1619,7 @@ int kfd_process_device_init_vm(struct kfd_process_device *pdd, struct amdgpu_fpriv *drv_priv; struct amdgpu_vm *avm; struct kfd_process *p; - struct kfd_dev *dev; + struct kfd_node *dev; int ret; if (!drm_file) @@ -1679,7 +1679,7 @@ err_reserve_ib_mem: * * Assumes that the process lock is held. */ -struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, +struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev, struct kfd_process *p) { struct kfd_process_device *pdd; @@ -1811,7 +1811,7 @@ int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger) for (i = 0; i < p->n_pdds; i++) { struct kfd_process_device *pdd = p->pdds[i]; - kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid, + kfd_smi_event_queue_eviction(pdd->dev->kfd, p->lead_thread->pid, trigger); r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm, @@ -1839,7 +1839,7 @@ fail: if (n_evicted == 0) break; - kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid); + kfd_smi_event_queue_restore(pdd->dev->kfd, p->lead_thread->pid); if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm, &pdd->qpd)) @@ -1860,7 +1860,7 @@ int kfd_process_restore_queues(struct kfd_process *p) for (i = 0; i < p->n_pdds; i++) { struct kfd_process_device *pdd = p->pdds[i]; - kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid); + kfd_smi_event_queue_restore(pdd->dev->kfd, p->lead_thread->pid); r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm, &pdd->qpd); @@ -2016,7 +2016,7 @@ int kfd_resume_all_processes(void) return ret; } -int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, +int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process, struct vm_area_struct *vma) { struct kfd_process_device *pdd; @@ -2051,7 +2051,7 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) { struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm); - struct kfd_dev *dev = pdd->dev; + struct kfd_node *dev = pdd->dev; /* * It can be that we race and lose here, but that is extremely unlikely diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 4236539d9f93..5602498e713f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -81,7 +81,7 @@ static int find_available_queue_slot(struct process_queue_manager *pqm, void kfd_process_dequeue_from_device(struct kfd_process_device *pdd) { - struct kfd_dev *dev = pdd->dev; + struct kfd_node *dev = pdd->dev; if (pdd->already_dequeued) return; @@ -93,7 +93,7 @@ void kfd_process_dequeue_from_device(struct kfd_process_device *pdd) int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, void *gws) { - struct kfd_dev *dev = NULL; + struct kfd_node *dev = NULL; struct process_queue_node *pqn; struct kfd_process_device *pdd; struct kgd_mem *mem = NULL; @@ -178,7 +178,7 @@ void pqm_uninit(struct process_queue_manager *pqm) } static int init_user_queue(struct process_queue_manager *pqm, - struct kfd_dev *dev, struct queue **q, + struct kfd_node *dev, struct queue **q, struct queue_properties *q_properties, struct file *f, struct amdgpu_bo *wptr_bo, unsigned int qid) @@ -199,7 +199,7 @@ static int init_user_queue(struct process_queue_manager *pqm, (*q)->device = dev; (*q)->process = pqm->process; - if (dev->shared_resources.enable_mes) { + if (dev->kfd->shared_resources.enable_mes) { retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, AMDGPU_MES_GANG_CTX_SIZE, &(*q)->gang_ctx_bo, @@ -224,7 +224,7 @@ cleanup: } int pqm_create_queue(struct process_queue_manager *pqm, - struct kfd_dev *dev, + struct kfd_node *dev, struct file *f, struct queue_properties *properties, unsigned int *qid, @@ -258,7 +258,7 @@ int pqm_create_queue(struct process_queue_manager *pqm, * Hence we also check the type as well */ if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ)) - max_queues = dev->device_info.max_no_of_hqd/2; + max_queues = dev->kfd->device_info.max_no_of_hqd/2; if (pdd->qpd.queue_count >= max_queues) return -ENOSPC; @@ -354,7 +354,7 @@ int pqm_create_queue(struct process_queue_manager *pqm, */ *p_doorbell_offset_in_process = (q->properties.doorbell_off * sizeof(uint32_t)) & - (kfd_doorbell_process_slice(dev) - 1); + (kfd_doorbell_process_slice(dev->kfd) - 1); pr_debug("PQM After DQM create queue\n"); @@ -387,7 +387,7 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) struct process_queue_node *pqn; struct kfd_process_device *pdd; struct device_queue_manager *dqm; - struct kfd_dev *dev; + struct kfd_node *dev; int retval; dqm = NULL; @@ -439,7 +439,7 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) pdd->qpd.num_gws = 0; } - if (dev->shared_resources.enable_mes) { + if (dev->kfd->shared_resources.enable_mes) { amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->gang_ctx_bo); if (pqn->q->wptr_bo) @@ -859,7 +859,7 @@ int kfd_criu_restore_queue(struct kfd_process *p, } if (!pdd->doorbell_index && - kfd_alloc_process_doorbells(pdd->dev, &pdd->doorbell_index) < 0) { + kfd_alloc_process_doorbells(pdd->dev->kfd, &pdd->doorbell_index) < 0) { ret = -ENOMEM; goto exit; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c index 0472b56de245..a0bf6558f4ac 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c @@ -36,7 +36,7 @@ struct kfd_smi_client { wait_queue_head_t wait_queue; /* events enabled */ uint64_t events; - struct kfd_dev *dev; + struct kfd_node *dev; spinlock_t lock; struct rcu_head rcu; pid_t pid; @@ -149,7 +149,7 @@ static void kfd_smi_ev_client_free(struct rcu_head *p) static int kfd_smi_ev_release(struct inode *inode, struct file *filep) { struct kfd_smi_client *client = filep->private_data; - struct kfd_dev *dev = client->dev; + struct kfd_node *dev = client->dev; spin_lock(&dev->smi_lock); list_del_rcu(&client->list); @@ -171,7 +171,7 @@ static bool kfd_smi_ev_enabled(pid_t pid, struct kfd_smi_client *client, return events & KFD_SMI_EVENT_MASK_FROM_INDEX(event); } -static void add_event_to_kfifo(pid_t pid, struct kfd_dev *dev, +static void add_event_to_kfifo(pid_t pid, struct kfd_node *dev, unsigned int smi_event, char *event_msg, int len) { struct kfd_smi_client *client; @@ -196,7 +196,7 @@ static void add_event_to_kfifo(pid_t pid, struct kfd_dev *dev, } __printf(4, 5) -static void kfd_smi_event_add(pid_t pid, struct kfd_dev *dev, +static void kfd_smi_event_add(pid_t pid, struct kfd_node *dev, unsigned int event, char *fmt, ...) { char fifo_in[KFD_SMI_EVENT_MSG_SIZE]; @@ -215,7 +215,7 @@ static void kfd_smi_event_add(pid_t pid, struct kfd_dev *dev, add_event_to_kfifo(pid, dev, event, fifo_in, len); } -void kfd_smi_event_update_gpu_reset(struct kfd_dev *dev, bool post_reset) +void kfd_smi_event_update_gpu_reset(struct kfd_node *dev, bool post_reset) { unsigned int event; @@ -228,7 +228,7 @@ void kfd_smi_event_update_gpu_reset(struct kfd_dev *dev, bool post_reset) kfd_smi_event_add(0, dev, event, "%x\n", dev->reset_seq_num); } -void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev, +void kfd_smi_event_update_thermal_throttling(struct kfd_node *dev, uint64_t throttle_bitmask) { kfd_smi_event_add(0, dev, KFD_SMI_EVENT_THERMAL_THROTTLE, "%llx:%llx\n", @@ -236,7 +236,7 @@ void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev, amdgpu_dpm_get_thermal_throttling_counter(dev->adev)); } -void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid) +void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid) { struct amdgpu_task_info task_info; @@ -254,17 +254,17 @@ void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid, unsigned long address, bool write_fault, ktime_t ts) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_PAGE_FAULT_START, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_PAGE_FAULT_START, "%lld -%d @%lx(%x) %c\n", ktime_to_ns(ts), pid, - address, dev->id, write_fault ? 'W' : 'R'); + address, dev->node->id, write_fault ? 'W' : 'R'); } void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, unsigned long address, bool migration) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_PAGE_FAULT_END, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_PAGE_FAULT_END, "%lld -%d @%lx(%x) %c\n", ktime_get_boottime_ns(), - pid, address, dev->id, migration ? 'M' : 'U'); + pid, address, dev->node->id, migration ? 'M' : 'U'); } void kfd_smi_event_migration_start(struct kfd_dev *dev, pid_t pid, @@ -273,7 +273,7 @@ void kfd_smi_event_migration_start(struct kfd_dev *dev, pid_t pid, uint32_t prefetch_loc, uint32_t preferred_loc, uint32_t trigger) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_MIGRATE_START, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_MIGRATE_START, "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", ktime_get_boottime_ns(), pid, start, end - start, from, to, prefetch_loc, preferred_loc, trigger); @@ -283,7 +283,7 @@ void kfd_smi_event_migration_end(struct kfd_dev *dev, pid_t pid, unsigned long start, unsigned long end, uint32_t from, uint32_t to, uint32_t trigger) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_MIGRATE_END, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_MIGRATE_END, "%lld -%d @%lx(%lx) %x->%x %d\n", ktime_get_boottime_ns(), pid, start, end - start, from, to, trigger); @@ -292,16 +292,16 @@ void kfd_smi_event_migration_end(struct kfd_dev *dev, pid_t pid, void kfd_smi_event_queue_eviction(struct kfd_dev *dev, pid_t pid, uint32_t trigger) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_QUEUE_EVICTION, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_QUEUE_EVICTION, "%lld -%d %x %d\n", ktime_get_boottime_ns(), pid, - dev->id, trigger); + dev->node->id, trigger); } void kfd_smi_event_queue_restore(struct kfd_dev *dev, pid_t pid) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_QUEUE_RESTORE, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_QUEUE_RESTORE, "%lld -%d %x\n", ktime_get_boottime_ns(), pid, - dev->id); + dev->node->id); } void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm) @@ -328,12 +328,12 @@ void kfd_smi_event_unmap_from_gpu(struct kfd_dev *dev, pid_t pid, unsigned long address, unsigned long last, uint32_t trigger) { - kfd_smi_event_add(pid, dev, KFD_SMI_EVENT_UNMAP_FROM_GPU, + kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_UNMAP_FROM_GPU, "%lld -%d @%lx(%lx) %x %d\n", ktime_get_boottime_ns(), - pid, address, last - address + 1, dev->id, trigger); + pid, address, last - address + 1, dev->node->id, trigger); } -int kfd_smi_event_open(struct kfd_dev *dev, uint32_t *fd) +int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd) { struct kfd_smi_client *client; int ret; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h index 76fe4e0ec2d2..59cd089f80d1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h @@ -24,11 +24,11 @@ #ifndef KFD_SMI_EVENTS_H_INCLUDED #define KFD_SMI_EVENTS_H_INCLUDED -int kfd_smi_event_open(struct kfd_dev *dev, uint32_t *fd); -void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid); -void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev, +int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd); +void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid); +void kfd_smi_event_update_thermal_throttling(struct kfd_node *dev, uint64_t throttle_bitmask); -void kfd_smi_event_update_gpu_reset(struct kfd_dev *dev, bool post_reset); +void kfd_smi_event_update_gpu_reset(struct kfd_node *dev, bool post_reset); void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid, unsigned long address, bool write_fault, ktime_t ts); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index c02430537e9c..96ccff79902c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1266,7 +1266,7 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, return -EINVAL; } - kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid, + kfd_smi_event_unmap_from_gpu(pdd->dev->kfd, p->lead_thread->pid, start, last, trigger); r = svm_range_unmap_from_gpu(pdd->dev->adev, @@ -3083,7 +3083,7 @@ int svm_range_list_init(struct kfd_process *p) spin_lock_init(&svms->deferred_list_lock); for (i = 0; i < p->n_pdds; i++) - if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev)) + if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->kfd)) bitmap_set(svms->bitmap_supported, i, 1); return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 8e4124dcb6e4..06a11186d947 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -96,7 +96,7 @@ struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) return ret; } -struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) +struct kfd_node *kfd_device_by_id(uint32_t gpu_id) { struct kfd_topology_device *top_dev; @@ -107,10 +107,10 @@ struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) return top_dev->gpu; } -struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) +struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev) { struct kfd_topology_device *top_dev; - struct kfd_dev *device = NULL; + struct kfd_node *device = NULL; down_read(&topology_lock); @@ -125,10 +125,10 @@ struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) return device; } -struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev) +struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev) { struct kfd_topology_device *top_dev; - struct kfd_dev *device = NULL; + struct kfd_node *device = NULL; down_read(&topology_lock); @@ -526,7 +526,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, if (dev->gpu) { log_max_watch_addr = - __ilog2_u32(dev->gpu->device_info.num_of_watch_points); + __ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points); if (log_max_watch_addr) { dev->node_props.capability |= @@ -548,11 +548,11 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL); sysfs_show_32bit_prop(buffer, offs, "fw_version", - dev->gpu->mec_fw_version); + dev->gpu->kfd->mec_fw_version); sysfs_show_32bit_prop(buffer, offs, "capability", dev->node_props.capability); sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version", - dev->gpu->sdma_fw_version); + dev->gpu->kfd->sdma_fw_version); sysfs_show_64bit_prop(buffer, offs, "unique_id", dev->gpu->adev->unique_id); @@ -1157,7 +1157,7 @@ void kfd_topology_shutdown(void) up_write(&topology_lock); } -static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) +static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) { uint32_t hashout; uint32_t buf[7]; @@ -1167,8 +1167,8 @@ static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) if (!gpu) return 0; - local_mem_size = gpu->local_mem_info.local_mem_size_private + - gpu->local_mem_info.local_mem_size_public; + local_mem_size = gpu->kfd->local_mem_info.local_mem_size_private + + gpu->kfd->local_mem_info.local_mem_size_public; buf[0] = gpu->adev->pdev->devfn; buf[1] = gpu->adev->pdev->subsystem_vendor | (gpu->adev->pdev->subsystem_device << 16); @@ -1188,7 +1188,7 @@ static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) * list then return NULL. This means a new topology device has to * be created for this GPU. */ -static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) +static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu) { struct kfd_topology_device *dev; struct kfd_topology_device *out_dev = NULL; @@ -1201,7 +1201,7 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) /* Discrete GPUs need their own topology device list * entries. Don't assign them to CPU/APU nodes. */ - if (!gpu->use_iommu_v2 && + if (!gpu->kfd->use_iommu_v2 && dev->node_props.cpu_cores_count) continue; @@ -1275,7 +1275,7 @@ static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev, CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; /* set gpu (dev) flags. */ } else { - if (!dev->gpu->pci_atomic_requested || + if (!dev->gpu->kfd->pci_atomic_requested || dev->gpu->adev->asic_type == CHIP_HAWAII) link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; @@ -1569,8 +1569,8 @@ static int kfd_dev_create_p2p_links(void) if (dev == new_dev) break; if (!dev->gpu || !dev->gpu->adev || - (dev->gpu->hive_id && - dev->gpu->hive_id == new_dev->gpu->hive_id)) + (dev->gpu->kfd->hive_id && + dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id)) goto next; /* check if node(s) is/are peer accessible in one direction or bi-direction */ @@ -1590,7 +1590,6 @@ out: return ret; } - /* Helper function. See kfd_fill_gpu_cache_info for parameter description */ static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext, struct kfd_gpu_cache_info *pcache_info, @@ -1723,7 +1722,7 @@ static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext, /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info * tables */ -static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_dev *kdev) +static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev) { struct kfd_gpu_cache_info *pcache_info = NULL; int i, j, k; @@ -1805,7 +1804,7 @@ static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct pr_debug("Added [%d] GPU cache entries\n", num_of_entries); } -static int kfd_topology_add_device_locked(struct kfd_dev *gpu, uint32_t gpu_id, +static int kfd_topology_add_device_locked(struct kfd_node *gpu, uint32_t gpu_id, struct kfd_topology_device **dev) { int proximity_domain = ++topology_crat_proximity_domain; @@ -1865,7 +1864,7 @@ err: return res; } -int kfd_topology_add_device(struct kfd_dev *gpu) +int kfd_topology_add_device(struct kfd_node *gpu) { uint32_t gpu_id; struct kfd_topology_device *dev; @@ -1916,7 +1915,8 @@ int kfd_topology_add_device(struct kfd_dev *gpu) dev->node_props.simd_arrays_per_engine = cu_info.num_shader_arrays_per_engine; - dev->node_props.gfx_target_version = gpu->device_info.gfx_target_version; + dev->node_props.gfx_target_version = + gpu->kfd->device_info.gfx_target_version; dev->node_props.vendor_id = gpu->adev->pdev->vendor; dev->node_props.device_id = gpu->adev->pdev->device; dev->node_props.capability |= @@ -1929,15 +1929,15 @@ int kfd_topology_add_device(struct kfd_dev *gpu) dev->node_props.max_engine_clk_ccompute = cpufreq_quick_get_max(0) / 1000; dev->node_props.drm_render_minor = - gpu->shared_resources.drm_render_minor; + gpu->kfd->shared_resources.drm_render_minor; - dev->node_props.hive_id = gpu->hive_id; + dev->node_props.hive_id = gpu->kfd->hive_id; dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu); dev->node_props.num_sdma_xgmi_engines = kfd_get_num_xgmi_sdma_engines(gpu); dev->node_props.num_sdma_queues_per_engine = - gpu->device_info.num_sdma_queues_per_engine - - gpu->device_info.num_reserved_sdma_queues_per_engine; + gpu->kfd->device_info.num_sdma_queues_per_engine - + gpu->kfd->device_info.num_reserved_sdma_queues_per_engine; dev->node_props.num_gws = (dev->gpu->gws && dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? dev->gpu->adev->gds.gws_size : 0; @@ -1979,7 +1979,7 @@ int kfd_topology_add_device(struct kfd_dev *gpu) * Overwrite ATS capability according to needs_iommu_device to fix * potential missing corresponding bit in CRAT of BIOS. */ - if (dev->gpu->use_iommu_v2) + if (dev->gpu->kfd->use_iommu_v2) dev->node_props.capability |= HSA_CAP_ATS_PRESENT; else dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT; @@ -2079,7 +2079,7 @@ static void kfd_topology_update_io_links(int proximity_domain) } } -int kfd_topology_remove_device(struct kfd_dev *gpu) +int kfd_topology_remove_device(struct kfd_node *gpu) { struct kfd_topology_device *dev, *tmp; uint32_t gpu_id; @@ -2119,7 +2119,7 @@ int kfd_topology_remove_device(struct kfd_dev *gpu) * Return - 0: On success (@kdev will be NULL for non GPU nodes) * -1: If end of list */ -int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) +int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev) { struct kfd_topology_device *top_dev; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h index fca30d00a9bb..3b8afb6aba79 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h @@ -75,7 +75,7 @@ struct kfd_mem_properties { uint32_t flags; uint32_t width; uint32_t mem_clk_max; - struct kfd_dev *gpu; + struct kfd_node *gpu; struct kobject *kobj; struct attribute attr; }; @@ -93,7 +93,7 @@ struct kfd_cache_properties { uint32_t cache_latency; uint32_t cache_type; uint8_t sibling_map[CACHE_SIBLINGMAP_SIZE]; - struct kfd_dev *gpu; + struct kfd_node *gpu; struct kobject *kobj; struct attribute attr; uint32_t sibling_map_size; @@ -113,7 +113,7 @@ struct kfd_iolink_properties { uint32_t max_bandwidth; uint32_t rec_transfer_size; uint32_t flags; - struct kfd_dev *gpu; + struct kfd_node *gpu; struct kobject *kobj; struct attribute attr; }; @@ -135,7 +135,7 @@ struct kfd_topology_device { struct list_head io_link_props; struct list_head p2p_link_props; struct list_head perf_props; - struct kfd_dev *gpu; + struct kfd_node *gpu; struct kobject *kobj_node; struct kobject *kobj_mem; struct kobject *kobj_cache; -- cgit v1.2.3 From 74c5b85da75475c73a8f040397610fbfcc2c3e78 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 16:33:38 -0400 Subject: drm/amdkfd: Add spatial partitioning support in KFD This patch introduces multi-partition support in KFD. This patch includes: - Support for maximum 8 spatial partitions in KFD. - Initialize one HIQ per partition. - Management of VMID range depending on partition mode. - Management of doorbell aperture space between all partitions. - Each partition does its own queue management, interrupt handling, SMI event reporting. - IOMMU, if enabled with multiple partitions, will only work on first partition. - SPM is only supported on the first partition. - Currently, there is no support for resetting individual partitions. All partitions will reset together. Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 214 ++++++++++++++++----- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 13 +- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 16 +- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 24 +-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 8 +- 7 files changed, 208 insertions(+), 77 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 23d9a7f77055..37c6dc5c37bf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -567,23 +567,27 @@ kfd_interrupt_error: return err; } -static void kfd_cleanup_node(struct kfd_dev *kfd) +static void kfd_cleanup_nodes(struct kfd_dev *kfd, unsigned int num_nodes) { - struct kfd_node *knode = kfd->node; - - device_queue_manager_uninit(knode->dqm); - kfd_interrupt_exit(knode); - kfd_topology_remove_device(knode); - if (knode->gws) - amdgpu_amdkfd_free_gws(knode->adev, knode->gws); - kfree(knode); - kfd->node = NULL; + struct kfd_node *knode; + unsigned int i; + + for (i = 0; i < num_nodes; i++) { + knode = kfd->nodes[i]; + device_queue_manager_uninit(knode->dqm); + kfd_interrupt_exit(knode); + kfd_topology_remove_device(knode); + if (knode->gws) + amdgpu_amdkfd_free_gws(knode->adev, knode->gws); + kfree(knode); + kfd->nodes[i] = NULL; + } } bool kgd2kfd_device_init(struct kfd_dev *kfd, const struct kgd2kfd_shared_resources *gpu_resources) { - unsigned int size, map_process_packet_size; + unsigned int size, map_process_packet_size, i; struct kfd_node *node; uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; unsigned int max_proc_per_quantum; @@ -596,9 +600,18 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, KGD_ENGINE_SDMA1); kfd->shared_resources = *gpu_resources; - first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1; - last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1; - vmid_num_kfd = last_vmid_kfd - first_vmid_kfd + 1; + if (kfd->adev->gfx.num_xcd == 0 || kfd->adev->gfx.num_xcd == 1 || + kfd->adev->gfx.num_xcc_per_xcp == 0) + kfd->num_nodes = 1; + else + kfd->num_nodes = + kfd->adev->gfx.num_xcd/kfd->adev->gfx.num_xcc_per_xcp; + if (kfd->num_nodes == 0) { + dev_err(kfd_device, + "KFD num nodes cannot be 0, GC inst: %d, num_xcc_in_node: %d\n", + kfd->adev->gfx.num_xcd, kfd->adev->gfx.num_xcc_per_xcp); + goto out; + } /* Allow BIF to recode atomics to PCIe 3.0 AtomicOps. * 32 and 64-bit requests are possible and must be @@ -617,6 +630,26 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, return false; } + first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1; + last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1; + vmid_num_kfd = last_vmid_kfd - first_vmid_kfd + 1; + + /* For GFX9.4.3, we need special handling for VMIDs depending on + * partition mode. + * In CPX mode, the VMID range needs to be shared between XCDs. + * Additionally, there are 13 VMIDs (3-15) available for KFD. To + * divide them equally, we change starting VMID to 4 and not use + * VMID 3. + * If the VMID range changes for GFX9.4.3, then this code MUST be + * revisited. + */ + if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && + kfd->adev->gfx.partition_mode == AMDGPU_CPX_PARTITION_MODE && + kfd->num_nodes != 1) { + vmid_num_kfd /= 2; + first_vmid_kfd = last_vmid_kfd + 1 - vmid_num_kfd*2; + } + /* Verify module parameters regarding mapped process number*/ if (hws_max_conc_proc >= 0) max_proc_per_quantum = min((u32)hws_max_conc_proc, vmid_num_kfd); @@ -682,6 +715,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd_cwsr_init(kfd); + /* TODO: Needs to be updated for memory partitioning */ svm_migrate_init(kfd->adev); /* Allocate the KFD node */ @@ -700,12 +734,51 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, node->max_proc_per_quantum = max_proc_per_quantum; atomic_set(&node->sram_ecc_flag, 0); - /* Initialize the KFD node */ - if (kfd_init_node(node)) { - dev_err(kfd_device, "Error initializing KFD node\n"); - goto node_init_error; + dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n", + kfd->num_nodes); + for (i = 0; i < kfd->num_nodes; i++) { + node = kzalloc(sizeof(struct kfd_node), GFP_KERNEL); + if (!node) + goto node_alloc_error; + + node->adev = kfd->adev; + node->kfd = kfd; + node->kfd2kgd = kfd->kfd2kgd; + node->vm_info.vmid_num_kfd = vmid_num_kfd; + node->num_xcc_per_node = max(1U, kfd->adev->gfx.num_xcc_per_xcp); + node->start_xcc_id = node->num_xcc_per_node * i; + + if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && + kfd->adev->gfx.partition_mode == AMDGPU_CPX_PARTITION_MODE && + kfd->num_nodes != 1) { + /* For GFX9.4.3 and CPX mode, first XCD gets VMID range + * 4-9 and second XCD gets VMID range 10-15. + */ + + node->vm_info.first_vmid_kfd = (i%2 == 0) ? + first_vmid_kfd : + first_vmid_kfd+vmid_num_kfd; + node->vm_info.last_vmid_kfd = (i%2 == 0) ? + last_vmid_kfd-vmid_num_kfd : + last_vmid_kfd; + node->compute_vmid_bitmap = + ((0x1 << (node->vm_info.last_vmid_kfd + 1)) - 1) - + ((0x1 << (node->vm_info.first_vmid_kfd)) - 1); + } else { + node->vm_info.first_vmid_kfd = first_vmid_kfd; + node->vm_info.last_vmid_kfd = last_vmid_kfd; + node->compute_vmid_bitmap = + gpu_resources->compute_vmid_bitmap; + } + node->max_proc_per_quantum = max_proc_per_quantum; + atomic_set(&node->sram_ecc_flag, 0); + /* Initialize the KFD node */ + if (kfd_init_node(node)) { + dev_err(kfd_device, "Error initializing KFD node\n"); + goto node_init_error; + } + kfd->nodes[i] = node; } - kfd->node = node; if (kfd_resume_iommu(kfd)) goto kfd_resume_iommu_error; @@ -722,9 +795,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, goto out; kfd_resume_iommu_error: - kfd_cleanup_node(kfd); node_init_error: node_alloc_error: + kfd_cleanup_nodes(kfd, i); device_iommu_error: kfd_doorbell_fini(kfd); kfd_doorbell_error: @@ -742,7 +815,9 @@ out: void kgd2kfd_device_exit(struct kfd_dev *kfd) { if (kfd->init_complete) { - kfd_cleanup_node(kfd); + /* Cleanup KFD nodes */ + kfd_cleanup_nodes(kfd, kfd->num_nodes); + /* Cleanup common/shared resources */ kfd_doorbell_fini(kfd); ida_destroy(&kfd->doorbell_ida); kfd_gtt_sa_fini(kfd); @@ -754,18 +829,23 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd) int kgd2kfd_pre_reset(struct kfd_dev *kfd) { - struct kfd_node *node = kfd->node; + struct kfd_node *node; + int i; if (!kfd->init_complete) return 0; - kfd_smi_event_update_gpu_reset(node, false); - - node->dqm->ops.pre_reset(node->dqm); + for (i = 0; i < kfd->num_nodes; i++) { + node = kfd->nodes[i]; + kfd_smi_event_update_gpu_reset(node, false); + node->dqm->ops.pre_reset(node->dqm); + } kgd2kfd_suspend(kfd, false); - kfd_signal_reset_event(node); + for (i = 0; i < kfd->num_nodes; i++) + kfd_signal_reset_event(kfd->nodes[i]); + return 0; } @@ -778,19 +858,25 @@ int kgd2kfd_pre_reset(struct kfd_dev *kfd) int kgd2kfd_post_reset(struct kfd_dev *kfd) { int ret; - struct kfd_node *node = kfd->node; + struct kfd_node *node; + int i; if (!kfd->init_complete) return 0; - ret = kfd_resume(node); - if (ret) - return ret; - atomic_dec(&kfd_locked); + for (i = 0; i < kfd->num_nodes; i++) { + ret = kfd_resume(kfd->nodes[i]); + if (ret) + return ret; + } - atomic_set(&node->sram_ecc_flag, 0); + atomic_dec(&kfd_locked); - kfd_smi_event_update_gpu_reset(node, true); + for (i = 0; i < kfd->num_nodes; i++) { + node = kfd->nodes[i]; + atomic_set(&node->sram_ecc_flag, 0); + kfd_smi_event_update_gpu_reset(node, true); + } return 0; } @@ -802,7 +888,8 @@ bool kfd_is_locked(void) void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) { - struct kfd_node *node = kfd->node; + struct kfd_node *node; + int i; if (!kfd->init_complete) return; @@ -814,21 +901,25 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) kfd_suspend_all_processes(); } - node->dqm->ops.stop(node->dqm); + for (i = 0; i < kfd->num_nodes; i++) { + node = kfd->nodes[i]; + node->dqm->ops.stop(node->dqm); + } kfd_iommu_suspend(kfd); } int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm) { - int ret, count; - struct kfd_node *node = kfd->node; + int ret, count, i; if (!kfd->init_complete) return 0; - ret = kfd_resume(node); - if (ret) - return ret; + for (i = 0; i < kfd->num_nodes; i++) { + ret = kfd_resume(kfd->nodes[i]); + if (ret) + return ret; + } /* for runtime resume, skip unlocking kfd */ if (!run_pm) { @@ -892,10 +983,10 @@ static inline void kfd_queue_work(struct workqueue_struct *wq, /* This is called directly from KGD at ISR. */ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) { - uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE]; + uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE], i; bool is_patched = false; unsigned long flags; - struct kfd_node *node = kfd->node; + struct kfd_node *node; if (!kfd->init_complete) return; @@ -905,16 +996,22 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) return; } - spin_lock_irqsave(&node->interrupt_lock, flags); - - if (node->interrupts_active - && interrupt_is_wanted(node, ih_ring_entry, - patched_ihre, &is_patched) - && enqueue_ih_ring_entry(node, - is_patched ? patched_ihre : ih_ring_entry)) - kfd_queue_work(node->ih_wq, &node->interrupt_work); + for (i = 0; i < kfd->num_nodes; i++) { + node = kfd->nodes[i]; + spin_lock_irqsave(&node->interrupt_lock, flags); + + if (node->interrupts_active + && interrupt_is_wanted(node, ih_ring_entry, + patched_ihre, &is_patched) + && enqueue_ih_ring_entry(node, + is_patched ? patched_ihre : ih_ring_entry)) { + kfd_queue_work(node->ih_wq, &node->interrupt_work); + spin_unlock_irqrestore(&node->interrupt_lock, flags); + return; + } + spin_unlock_irqrestore(&node->interrupt_lock, flags); + } - spin_unlock_irqrestore(&node->interrupt_lock, flags); } int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger) @@ -1181,8 +1278,13 @@ int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj) void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd) { + /* + * TODO: Currently update SRAM ECC flag for first node. + * This needs to be updated later when we can + * identify SRAM ECC error on other nodes also. + */ if (kfd) - atomic_inc(&kfd->node->sram_ecc_flag); + atomic_inc(&kfd->nodes[0]->sram_ecc_flag); } void kfd_inc_compute_active(struct kfd_node *node) @@ -1202,8 +1304,14 @@ void kfd_dec_compute_active(struct kfd_node *node) void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) { + /* + * TODO: For now, raise the throttling event only on first node. + * This will need to change after we are able to determine + * which node raised the throttling event. + */ if (kfd && kfd->init_complete) - kfd_smi_event_update_thermal_throttling(kfd->node, throttle_bitmask); + kfd_smi_event_update_thermal_throttling(kfd->nodes[0], + throttle_bitmask); } /* kfd_get_num_sdma_engines returns the number of PCIe optimized SDMA and diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 34977d89f01c..6ee17100c333 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1426,7 +1426,7 @@ static int set_sched_resources(struct device_queue_manager *dqm) int i, mec; struct scheduling_resources res; - res.vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; + res.vmid_mask = dqm->dev->compute_vmid_bitmap; res.queue_mask = 0; for (i = 0; i < KGD_MAX_QUEUES; ++i) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c b/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c index 6eee9a0944f3..808ee010520a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_iommu.c @@ -121,6 +121,12 @@ int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd) return -EINVAL; } + if (!kfd_is_first_node(dev)) { + dev_warn_once(kfd_device, + "IOMMU supported only on first node\n"); + return 0; + } + err = amd_iommu_bind_pasid(dev->adev->pdev, p->pasid, p->lead_thread); if (!err) pdd->bound = PDD_BOUND; @@ -138,7 +144,8 @@ void kfd_iommu_unbind_process(struct kfd_process *p) int i; for (i = 0; i < p->n_pdds; i++) - if (p->pdds[i]->bound == PDD_BOUND) + if ((p->pdds[i]->bound == PDD_BOUND) && + (kfd_is_first_node((p->pdds[i]->dev)))) amd_iommu_unbind_pasid(p->pdds[i]->dev->adev->pdev, p->pasid); } @@ -281,7 +288,7 @@ void kfd_iommu_suspend(struct kfd_dev *kfd) if (!kfd->use_iommu_v2) return; - kfd_unbind_processes_from_device(kfd->node); + kfd_unbind_processes_from_device(kfd->nodes[0]); amd_iommu_set_invalidate_ctx_cb(kfd->adev->pdev, NULL); amd_iommu_set_invalid_ppr_cb(kfd->adev->pdev, NULL); @@ -312,7 +319,7 @@ int kfd_iommu_resume(struct kfd_dev *kfd) amd_iommu_set_invalid_ppr_cb(kfd->adev->pdev, iommu_invalid_ppr_cb); - err = kfd_bind_processes_to_device(kfd->node); + err = kfd_bind_processes_to_device(kfd->nodes[0]); if (err) { amd_iommu_set_invalidate_ctx_cb(kfd->adev->pdev, NULL); amd_iommu_set_invalid_ppr_cb(kfd->adev->pdev, NULL); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 1e187677c90a..5f4dc2a45bd0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -423,7 +423,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - 0, adev->kfd.dev->node->id, prange->prefetch_loc, + 0, adev->kfd.dev->nodes[0]->id, prange->prefetch_loc, prange->preferred_loc, trigger); r = migrate_vma_setup(&migrate); @@ -456,7 +456,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - 0, adev->kfd.dev->node->id, trigger); + 0, adev->kfd.dev->nodes[0]->id, trigger); svm_range_dma_unmap(adev->dev, scratch, 0, npages); svm_range_free_dma_mappings(prange); @@ -701,7 +701,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - adev->kfd.dev->node->id, 0, prange->prefetch_loc, + adev->kfd.dev->nodes[0]->id, 0, prange->prefetch_loc, prange->preferred_loc, trigger); r = migrate_vma_setup(&migrate); @@ -737,7 +737,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - adev->kfd.dev->node->id, 0, trigger); + adev->kfd.dev->nodes[0]->id, 0, trigger); svm_range_dma_unmap(adev->dev, scratch, 0, npages); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index fdb97e5d0c01..873b49238dc1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -255,6 +255,8 @@ struct kfd_vmid_info { uint32_t vmid_num_kfd; }; +#define MAX_KFD_NODES 8 + struct kfd_dev; struct kfd_node { @@ -267,6 +269,10 @@ struct kfd_node { */ struct kfd_vmid_info vm_info; unsigned int id; /* topology stub index */ + unsigned int num_xcc_per_node; + unsigned int start_xcc_id; /* Starting XCC instance + * number for the node + */ /* Interrupts */ struct kfifo ih_fifo; struct workqueue_struct *ih_wq; @@ -300,6 +306,8 @@ struct kfd_node { /* Maximum process number mapped to HW scheduler */ unsigned int max_proc_per_quantum; + unsigned int compute_vmid_bitmap; + struct kfd_dev *kfd; }; @@ -368,7 +376,8 @@ struct kfd_dev { /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ struct dev_pagemap pgmap; - struct kfd_node *node; + struct kfd_node *nodes[MAX_KFD_NODES]; + unsigned int num_nodes; }; enum kfd_mempool { @@ -1397,6 +1406,11 @@ static inline int kfd_devcgroup_check_permission(struct kfd_node *kfd) #endif } +static inline bool kfd_is_first_node(struct kfd_node *node) +{ + return (node == node->kfd->nodes[0]); +} + /* Debugfs */ #if defined(CONFIG_DEBUG_FS) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c index a0bf6558f4ac..b703da59e067 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c @@ -254,17 +254,17 @@ void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid, unsigned long address, bool write_fault, ktime_t ts) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_PAGE_FAULT_START, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_PAGE_FAULT_START, "%lld -%d @%lx(%x) %c\n", ktime_to_ns(ts), pid, - address, dev->node->id, write_fault ? 'W' : 'R'); + address, dev->nodes[0]->id, write_fault ? 'W' : 'R'); } void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, unsigned long address, bool migration) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_PAGE_FAULT_END, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_PAGE_FAULT_END, "%lld -%d @%lx(%x) %c\n", ktime_get_boottime_ns(), - pid, address, dev->node->id, migration ? 'M' : 'U'); + pid, address, dev->nodes[0]->id, migration ? 'M' : 'U'); } void kfd_smi_event_migration_start(struct kfd_dev *dev, pid_t pid, @@ -273,7 +273,7 @@ void kfd_smi_event_migration_start(struct kfd_dev *dev, pid_t pid, uint32_t prefetch_loc, uint32_t preferred_loc, uint32_t trigger) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_MIGRATE_START, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_MIGRATE_START, "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", ktime_get_boottime_ns(), pid, start, end - start, from, to, prefetch_loc, preferred_loc, trigger); @@ -283,7 +283,7 @@ void kfd_smi_event_migration_end(struct kfd_dev *dev, pid_t pid, unsigned long start, unsigned long end, uint32_t from, uint32_t to, uint32_t trigger) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_MIGRATE_END, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_MIGRATE_END, "%lld -%d @%lx(%lx) %x->%x %d\n", ktime_get_boottime_ns(), pid, start, end - start, from, to, trigger); @@ -292,16 +292,16 @@ void kfd_smi_event_migration_end(struct kfd_dev *dev, pid_t pid, void kfd_smi_event_queue_eviction(struct kfd_dev *dev, pid_t pid, uint32_t trigger) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_QUEUE_EVICTION, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_QUEUE_EVICTION, "%lld -%d %x %d\n", ktime_get_boottime_ns(), pid, - dev->node->id, trigger); + dev->nodes[0]->id, trigger); } void kfd_smi_event_queue_restore(struct kfd_dev *dev, pid_t pid) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_QUEUE_RESTORE, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_QUEUE_RESTORE, "%lld -%d %x\n", ktime_get_boottime_ns(), pid, - dev->node->id); + dev->nodes[0]->id); } void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm) @@ -328,9 +328,9 @@ void kfd_smi_event_unmap_from_gpu(struct kfd_dev *dev, pid_t pid, unsigned long address, unsigned long last, uint32_t trigger) { - kfd_smi_event_add(pid, dev->node, KFD_SMI_EVENT_UNMAP_FROM_GPU, + kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_UNMAP_FROM_GPU, "%lld -%d @%lx(%lx) %x %d\n", ktime_get_boottime_ns(), - pid, address, last - address + 1, dev->node->id, trigger); + pid, address, last - address + 1, dev->nodes[0]->id, trigger); } int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 06a11186d947..94af37df3ed2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -555,7 +555,8 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, dev->gpu->kfd->sdma_fw_version); sysfs_show_64bit_prop(buffer, offs, "unique_id", dev->gpu->adev->unique_id); - + sysfs_show_32bit_prop(buffer, offs, "num_xcc", + dev->gpu->num_xcc_per_node); } return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute", @@ -1160,7 +1161,7 @@ void kfd_topology_shutdown(void) static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) { uint32_t hashout; - uint32_t buf[7]; + uint32_t buf[8]; uint64_t local_mem_size; int i; @@ -1177,8 +1178,9 @@ static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) buf[4] = gpu->adev->pdev->bus->number; buf[5] = lower_32_bits(local_mem_size); buf[6] = upper_32_bits(local_mem_size); + buf[7] = gpu->start_xcc_id | (gpu->num_xcc_per_node << 16); - for (i = 0, hashout = 0; i < 7; i++) + for (i = 0, hashout = 0; i < 8; i++) hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); return hashout; -- cgit v1.2.3 From 2f77b9a242a2e01822efc80c8b63eaa31df0f8b4 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 21:45:50 -0400 Subject: drm/amdkfd: Update MQD management on multi XCC setup Update MQD management for both HIQ and user-mode compute queues on a multi XCC setup. MQDs needs to be allocated, initialized, loaded and destroyed for each XCC in the KFD node. v2: squash in fix "drm/amdkfd: Fix SDMA+HIQ HQD allocation on GFX9.4.3" Signed-off-by: Mukul Joshi Signed-off-by: Amber Lin Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 51 ++-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 28 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h | 8 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c | 3 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 3 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 292 ++++++++++++++++++--- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 3 + .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 16 +- drivers/gpu/drm/amd/include/v9_structs.h | 30 ++- 10 files changed, 380 insertions(+), 57 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index f61f07575f63..6bbe3b89aef5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -800,6 +800,41 @@ static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev, sg_free_table(ttm->sg); } +/* + * total_pages is constructed as MQD0+CtrlStack0 + MQD1+CtrlStack1 + ... + * MQDn+CtrlStackn where n is the number of XCCs per partition. + * pages_per_xcc is the size of one MQD+CtrlStack. The first page is MQD + * and uses memory type default, UC. The rest of pages_per_xcc are + * Ctrl stack and modify their memory type to NC. + */ +static void amdgpu_ttm_gart_bind_gfx9_mqd(struct amdgpu_device *adev, + struct ttm_tt *ttm, uint64_t flags) +{ + struct amdgpu_ttm_tt *gtt = (void *)ttm; + uint64_t total_pages = ttm->num_pages; + int num_xcc = max(1U, adev->gfx.num_xcc_per_xcp); + uint64_t page_idx, pages_per_xcc = total_pages / num_xcc; + int i; + uint64_t ctrl_flags = (flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | + AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC); + + for (i = 0, page_idx = 0; i < num_xcc; i++, page_idx += pages_per_xcc) { + /* MQD page: use default flags */ + amdgpu_gart_bind(adev, + gtt->offset + (page_idx << PAGE_SHIFT), + 1, >t->ttm.dma_address[page_idx], flags); + /* + * Ctrl pages - modify the memory type to NC (ctrl_flags) from + * the second page of the BO onward. + */ + amdgpu_gart_bind(adev, + gtt->offset + ((page_idx + 1) << PAGE_SHIFT), + pages_per_xcc - 1, + >t->ttm.dma_address[page_idx + 1], + ctrl_flags); + } +} + static void amdgpu_ttm_gart_bind(struct amdgpu_device *adev, struct ttm_buffer_object *tbo, uint64_t flags) @@ -812,21 +847,7 @@ static void amdgpu_ttm_gart_bind(struct amdgpu_device *adev, flags |= AMDGPU_PTE_TMZ; if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) { - uint64_t page_idx = 1; - - amdgpu_gart_bind(adev, gtt->offset, page_idx, - gtt->ttm.dma_address, flags); - - /* The memory type of the first page defaults to UC. Now - * modify the memory type to NC from the second page of - * the BO onward. - */ - flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK; - flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC); - - amdgpu_gart_bind(adev, gtt->offset + (page_idx << PAGE_SHIFT), - ttm->num_pages - page_idx, - &(gtt->ttm.dma_address[page_idx]), flags); + amdgpu_ttm_gart_bind_gfx9_mqd(adev, ttm, flags); } else { amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages, gtt->ttm.dma_address, flags); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 6ee17100c333..9afd3295ca85 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -2247,7 +2247,8 @@ static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * get_num_all_sdma_engines(dqm) * dev->kfd->device_info.num_sdma_queues_per_engine + - dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; + (dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size * + dqm->dev->num_xcc_per_node); retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size, &(mem_obj->gtt_mem), &(mem_obj->gpu_addr), diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c index 61f6dd68c84b..074f6075ccc7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c @@ -76,7 +76,8 @@ struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev, q->sdma_queue_id) * dev->dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size; - offset += dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; + offset += dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size * + dev->num_xcc_per_node; mqd_mem_obj->gtt_mem = (void *)((uint64_t)dev->dqm->hiq_sdma_mqd.gtt_mem + offset); @@ -246,3 +247,28 @@ bool kfd_is_occupied_sdma(struct mqd_manager *mm, void *mqd, { return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->adev, mqd); } + +uint64_t kfd_hiq_mqd_stride(struct kfd_node *dev) +{ + return dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; +} + +void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd_mem_obj, + uint32_t virtual_xcc_id) +{ + uint64_t offset; + + offset = kfd_hiq_mqd_stride(dev) * virtual_xcc_id; + + mqd_mem_obj->gtt_mem = (virtual_xcc_id == 0) ? + dev->dqm->hiq_sdma_mqd.gtt_mem : NULL; + mqd_mem_obj->gpu_addr = dev->dqm->hiq_sdma_mqd.gpu_addr + offset; + mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t) + dev->dqm->hiq_sdma_mqd.cpu_ptr + offset); +} + +uint64_t kfd_mqd_stride(struct mqd_manager *mm, + struct queue_properties *q) +{ + return mm->mqd_size; +} diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h index 46fc3f273d0d..eb18be74f559 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h @@ -119,6 +119,8 @@ struct mqd_manager { int (*debugfs_show_mqd)(struct seq_file *m, void *data); #endif uint32_t (*read_doorbell_id)(void *mqd); + uint64_t (*mqd_stride)(struct mqd_manager *mm, + struct queue_properties *p); struct mutex mqd_mutex; struct kfd_node *dev; @@ -164,4 +166,10 @@ bool kfd_is_occupied_sdma(struct mqd_manager *mm, void *mqd, uint64_t queue_address, uint32_t pipe_id, uint32_t queue_id); +void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, + struct kfd_mem_obj *mqd_mem_obj, uint32_t virtual_xcc_id); + +uint64_t kfd_hiq_mqd_stride(struct kfd_node *dev); +uint64_t kfd_mqd_stride(struct mqd_manager *mm, + struct queue_properties *q); #endif /* KFD_MQD_MANAGER_H_ */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c index 03e04d5e5a11..ca1966466759 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c @@ -428,6 +428,7 @@ struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct cik_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif @@ -442,6 +443,7 @@ struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct cik_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif @@ -457,6 +459,7 @@ struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, mqd->checkpoint_mqd = checkpoint_mqd_sdma; mqd->restore_mqd = restore_mqd_sdma; mqd->mqd_size = sizeof(struct cik_sdma_rlc_registers); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index 7a93be0ebb19..c9565ea99df5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -432,6 +432,7 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, mqd->get_wave_state = get_wave_state; mqd->checkpoint_mqd = checkpoint_mqd; mqd->restore_mqd = restore_mqd; + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif @@ -447,6 +448,7 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct v10_compute_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif @@ -478,6 +480,7 @@ struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, mqd->checkpoint_mqd = checkpoint_mqd_sdma; mqd->restore_mqd = restore_mqd_sdma; mqd->mqd_size = sizeof(struct v10_sdma_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 943a738e73f9..c677322057dd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -33,6 +33,21 @@ #include "sdma0/sdma0_4_0_sh_mask.h" #include "amdgpu_amdkfd.h" +static void update_mqd(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, + struct mqd_update_info *minfo); + +static uint64_t mqd_stride_v9(struct mqd_manager *mm, + struct queue_properties *q) +{ + if (mm->dev->kfd->cwsr_enabled && + q->type == KFD_QUEUE_TYPE_COMPUTE) + return ALIGN(q->ctl_stack_size, PAGE_SIZE) + + ALIGN(sizeof(struct v9_mqd), PAGE_SIZE); + + return mm->mqd_size; +} + static inline struct v9_mqd *get_mqd(void *mqd) { return (struct v9_mqd *)mqd; @@ -110,8 +125,9 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node, if (!mqd_mem_obj) return NULL; retval = amdgpu_amdkfd_alloc_gtt_mem(node->adev, - ALIGN(q->ctl_stack_size, PAGE_SIZE) + - ALIGN(sizeof(struct v9_mqd), PAGE_SIZE), + (ALIGN(q->ctl_stack_size, PAGE_SIZE) + + ALIGN(sizeof(struct v9_mqd), PAGE_SIZE)) * + node->num_xcc_per_node, &(mqd_mem_obj->gtt_mem), &(mqd_mem_obj->gpu_addr), (void *)&(mqd_mem_obj->cpu_ptr), true); @@ -165,24 +181,9 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; - if (q->format == KFD_QUEUE_FORMAT_AQL) { + if (q->format == KFD_QUEUE_FORMAT_AQL) m->cp_hqd_aql_control = 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; - if (KFD_GC_VERSION(mm->dev) == IP_VERSION(9, 4, 3)) { - /* On GC 9.4.3, DW 41 is re-purposed as - * compute_tg_chunk_size. - * TODO: review this setting when active CUs in the - * partition play a role - */ - m->compute_static_thread_mgmt_se6 = 1; - } - } else { - /* PM4 queue */ - if (KFD_GC_VERSION(mm->dev) == IP_VERSION(9, 4, 3)) { - m->compute_static_thread_mgmt_se6 = 0; - /* TODO: program pm4_target_xcc */ - } - } if (q->tba_addr) { m->compute_pgm_rsrc2 |= @@ -205,7 +206,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, *mqd = m; if (gart_addr) *gart_addr = addr; - mm->update_mqd(mm, m, q, NULL); + update_mqd(mm, m, q, NULL); } static int load_mqd(struct mqd_manager *mm, void *mqd, @@ -269,13 +270,10 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, m->cp_hqd_vmid = q->vmid; if (q->format == KFD_QUEUE_FORMAT_AQL) { - m->cp_hqd_pq_control |= + m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK | 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT | 1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT | 1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT; - if (KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 3)) - m->cp_hqd_pq_control |= - CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; m->cp_hqd_pq_doorbell_control |= 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT; } @@ -466,6 +464,224 @@ static void restore_mqd_sdma(struct mqd_manager *mm, void **mqd, qp->is_active = 0; } +static void init_mqd_hiq_v9_4_3(struct mqd_manager *mm, void **mqd, + struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, + struct queue_properties *q) +{ + struct v9_mqd *m; + int xcc = 0; + struct kfd_mem_obj xcc_mqd_mem_obj; + uint64_t xcc_gart_addr = 0; + + memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj)); + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + kfd_get_hiq_xcc_mqd(mm->dev, &xcc_mqd_mem_obj, xcc); + + init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q); + + m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK | + 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT | + 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT; + m->cp_mqd_stride_size = kfd_hiq_mqd_stride(mm->dev); + if (xcc == 0) { + /* Set no_update_rptr = 0 in Master XCC */ + m->cp_hqd_pq_control &= ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; + + /* Set the MQD pointer and gart address to XCC0 MQD */ + *mqd = m; + *gart_addr = xcc_gart_addr; + } + } +} + +static int hiq_load_mqd_kiq_v9_4_3(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + int xcc, err; + void *xcc_mqd; + uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev); + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + xcc_mqd = mqd + hiq_mqd_size * xcc; + err = mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, xcc_mqd, + pipe_id, queue_id, + p->doorbell_off); + if (err) { + pr_debug("Failed to load HIQ MQD for XCC: %d\n", xcc); + break; + } + } + + return err; +} + +static int destroy_hiq_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, + enum kfd_preempt_type type, unsigned int timeout, + uint32_t pipe_id, uint32_t queue_id) +{ + int xcc = 0, err; + void *xcc_mqd; + uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev); + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + xcc_mqd = mqd + hiq_mqd_size * xcc; + err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd, + type, timeout, pipe_id, + queue_id); + if (err) { + pr_debug("Destroy MQD failed for xcc: %d\n", xcc); + break; + } + } + + return err; +} + +static void get_xcc_mqd(struct kfd_mem_obj *mqd_mem_obj, + struct kfd_mem_obj *xcc_mqd_mem_obj, + uint64_t offset) +{ + xcc_mqd_mem_obj->gtt_mem = (offset == 0) ? + mqd_mem_obj->gtt_mem : NULL; + xcc_mqd_mem_obj->gpu_addr = mqd_mem_obj->gpu_addr + offset; + xcc_mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t)mqd_mem_obj->cpu_ptr + + offset); +} + +static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, + struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr, + struct queue_properties *q) +{ + struct v9_mqd *m; + int xcc = 0; + struct kfd_mem_obj xcc_mqd_mem_obj; + uint64_t xcc_gart_addr = 0; + uint64_t offset = mm->mqd_stride(mm, q); + + memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj)); + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + get_xcc_mqd(mqd_mem_obj, &xcc_mqd_mem_obj, offset*xcc); + + init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q); + + m->cp_mqd_stride_size = offset; + if (q->format == KFD_QUEUE_FORMAT_AQL) { + m->compute_tg_chunk_size = 1; + + switch (xcc) { + case 0: + /* Master XCC */ + m->cp_hqd_pq_control &= + ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; + m->compute_current_logic_xcc_id = + mm->dev->num_xcc_per_node - 1; + break; + default: + m->compute_current_logic_xcc_id = + xcc - 1; + break; + } + } else { + /* PM4 Queue */ + m->compute_current_logic_xcc_id = 0; + m->compute_tg_chunk_size = 0; + } + + if (xcc == 0) { + /* Set the MQD pointer and gart address to XCC0 MQD */ + *mqd = m; + *gart_addr = xcc_gart_addr; + } + } +} + +static void update_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, struct mqd_update_info *minfo) +{ + struct v9_mqd *m; + int xcc = 0; + uint64_t size = mm->mqd_stride(mm, q); + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + m = get_mqd(mqd + size * xcc); + update_mqd(mm, m, q, minfo); + + if (q->format == KFD_QUEUE_FORMAT_AQL) { + switch (xcc) { + case 0: + /* Master XCC */ + m->cp_hqd_pq_control &= + ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; + m->compute_current_logic_xcc_id = + mm->dev->num_xcc_per_node - 1; + break; + default: + m->compute_current_logic_xcc_id = + xcc - 1; + break; + } + m->compute_tg_chunk_size = 1; + } else { + /* PM4 Queue */ + m->compute_current_logic_xcc_id = 0; + m->compute_tg_chunk_size = 0; + } + } +} + +static int destroy_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, + enum kfd_preempt_type type, unsigned int timeout, + uint32_t pipe_id, uint32_t queue_id) +{ + int xcc = 0, err; + void *xcc_mqd; + struct v9_mqd *m; + uint64_t mqd_offset; + + m = get_mqd(mqd); + mqd_offset = m->cp_mqd_stride_size; + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + xcc_mqd = mqd + mqd_offset * xcc; + err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd, + type, timeout, pipe_id, + queue_id); + if (err) { + pr_debug("Destroy MQD failed for xcc: %d\n", xcc); + break; + } + } + + return err; +} + +static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + struct queue_properties *p, struct mm_struct *mms) +{ + /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ + uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); + int xcc = 0, err; + void *xcc_mqd; + uint64_t mqd_stride_size = mm->mqd_stride(mm, p); + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + xcc_mqd = mqd + mqd_stride_size * xcc; + err = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, xcc_mqd, + pipe_id, queue_id, + (uint32_t __user *)p->write_ptr, + wptr_shift, 0, mms); + if (err) { + pr_debug("Load MQD failed for xcc: %d\n", xcc); + break; + } + } + + return err; +} + #if defined(CONFIG_DEBUG_FS) static int debugfs_show_mqd(struct seq_file *m, void *data) @@ -501,34 +717,49 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, switch (type) { case KFD_MQD_TYPE_CP: mqd->allocate_mqd = allocate_mqd; - mqd->init_mqd = init_mqd; mqd->free_mqd = kfd_free_mqd_cp; - mqd->load_mqd = load_mqd; - mqd->update_mqd = update_mqd; - mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->get_wave_state = get_wave_state; mqd->get_checkpoint_info = get_checkpoint_info; mqd->checkpoint_mqd = checkpoint_mqd; mqd->restore_mqd = restore_mqd; mqd->mqd_size = sizeof(struct v9_mqd); + mqd->mqd_stride = mqd_stride_v9; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif + if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) { + mqd->init_mqd = init_mqd_v9_4_3; + mqd->load_mqd = load_mqd_v9_4_3; + mqd->update_mqd = update_mqd_v9_4_3; + mqd->destroy_mqd = destroy_mqd_v9_4_3; + } else { + mqd->init_mqd = init_mqd; + mqd->load_mqd = load_mqd; + mqd->update_mqd = update_mqd; + mqd->destroy_mqd = kfd_destroy_mqd_cp; + } break; case KFD_MQD_TYPE_HIQ: mqd->allocate_mqd = allocate_hiq_mqd; - mqd->init_mqd = init_mqd_hiq; mqd->free_mqd = free_mqd_hiq_sdma; - mqd->load_mqd = kfd_hiq_load_mqd_kiq; mqd->update_mqd = update_mqd; - mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct v9_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif mqd->read_doorbell_id = read_doorbell_id; + if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) { + mqd->init_mqd = init_mqd_hiq_v9_4_3; + mqd->load_mqd = hiq_load_mqd_kiq_v9_4_3; + mqd->destroy_mqd = destroy_hiq_mqd_v9_4_3; + } else { + mqd->init_mqd = init_mqd_hiq; + mqd->load_mqd = kfd_hiq_load_mqd_kiq; + mqd->destroy_mqd = kfd_destroy_mqd_cp; + } break; case KFD_MQD_TYPE_DIQ: mqd->allocate_mqd = allocate_mqd; @@ -554,6 +785,7 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, mqd->checkpoint_mqd = checkpoint_mqd_sdma; mqd->restore_mqd = restore_mqd_sdma; mqd->mqd_size = sizeof(struct v9_sdma_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c index f6b4a5686dcb..8736a3cdbe1e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c @@ -486,6 +486,7 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct vi_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif @@ -500,6 +501,7 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, mqd->destroy_mqd = kfd_destroy_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; mqd->mqd_size = sizeof(struct vi_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd; #endif @@ -515,6 +517,7 @@ struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, mqd->checkpoint_mqd = checkpoint_mqd_sdma; mqd->restore_mqd = restore_mqd_sdma; mqd->mqd_size = sizeof(struct vi_sdma_mqd); + mqd->mqd_stride = kfd_mqd_stride; #if defined(CONFIG_DEBUG_FS) mqd->debugfs_show_mqd = debugfs_show_mqd_sdma; #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 5602498e713f..b1fb017b2ef8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -927,7 +927,9 @@ int pqm_debugfs_mqds(struct seq_file *m, void *data) struct queue *q; enum KFD_MQD_TYPE mqd_type; struct mqd_manager *mqd_mgr; - int r = 0; + int r = 0, xcc, num_xccs = 1; + void *mqd; + uint64_t size = 0; list_for_each_entry(pqn, &pqm->queues, process_queue_list) { if (pqn->q) { @@ -943,6 +945,7 @@ int pqm_debugfs_mqds(struct seq_file *m, void *data) seq_printf(m, " Compute queue on device %x\n", q->device->id); mqd_type = KFD_MQD_TYPE_CP; + num_xccs = q->device->num_xcc_per_node; break; default: seq_printf(m, @@ -951,6 +954,8 @@ int pqm_debugfs_mqds(struct seq_file *m, void *data) continue; } mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type]; + size = mqd_mgr->mqd_stride(mqd_mgr, + &q->properties); } else if (pqn->kq) { q = pqn->kq->queue; mqd_mgr = pqn->kq->mqd_mgr; @@ -972,9 +977,12 @@ int pqm_debugfs_mqds(struct seq_file *m, void *data) continue; } - r = mqd_mgr->debugfs_show_mqd(m, q->mqd); - if (r != 0) - break; + for (xcc = 0; xcc < num_xccs; xcc++) { + mqd = q->mqd + size * xcc; + r = mqd_mgr->debugfs_show_mqd(m, mqd); + if (r != 0) + break; + } } return r; diff --git a/drivers/gpu/drm/amd/include/v9_structs.h b/drivers/gpu/drm/amd/include/v9_structs.h index a0c672889fe4..a2f81b9c38af 100644 --- a/drivers/gpu/drm/amd/include/v9_structs.h +++ b/drivers/gpu/drm/amd/include/v9_structs.h @@ -196,10 +196,20 @@ struct v9_mqd { uint32_t compute_wave_restore_addr_lo; uint32_t compute_wave_restore_addr_hi; uint32_t compute_wave_restore_control; - uint32_t compute_static_thread_mgmt_se4; - uint32_t compute_static_thread_mgmt_se5; - uint32_t compute_static_thread_mgmt_se6; - uint32_t compute_static_thread_mgmt_se7; + union { + struct { + uint32_t compute_static_thread_mgmt_se4; + uint32_t compute_static_thread_mgmt_se5; + uint32_t compute_static_thread_mgmt_se6; + uint32_t compute_static_thread_mgmt_se7; + }; + struct { + uint32_t compute_current_logic_xcc_id; // offset: 39 (0x27) + uint32_t compute_restart_cg_tg_id; // offset: 40 (0x28) + uint32_t compute_tg_chunk_size; // offset: 41 (0x29) + uint32_t compute_restore_tg_chunk_size; // offset: 42 (0x2A) + }; + }; uint32_t reserved_43; uint32_t reserved_44; uint32_t reserved_45; @@ -382,8 +392,16 @@ struct v9_mqd { uint32_t iqtimer_pkt_dw29; uint32_t iqtimer_pkt_dw30; uint32_t iqtimer_pkt_dw31; - uint32_t reserved_225; - uint32_t reserved_226; + union { + struct { + uint32_t reserved_225; + uint32_t reserved_226; + }; + struct { + uint32_t pm4_target_xcc_in_xcp; // offset: 225 (0xE1) + uint32_t cp_mqd_stride_size; // offset: 226 (0xE2) + }; + }; uint32_t reserved_227; uint32_t set_resources_header; uint32_t set_resources_dw1; -- cgit v1.2.3 From 3c8bdb51be0e895010da62dfa173bb1227ff3b6f Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 21:50:43 -0400 Subject: drm/amdkfd: Add PM4 target XCC In a device that supports multiple XCCs, unlike AQL queues, the PM4 queue will be only processed in one XCC in the partitioning. This patch re-purposes the queue percentage variable in create queue and update queue ioctl for the user space to specify the target XCC. Signed-off-by: Amber Lin Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 22 ++++++++++++++++++---- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 2 ++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 + .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 1 + 4 files changed, 22 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index eb0b0b38f10e..45e8da125f70 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -186,7 +186,12 @@ static int kfd_ioctl_get_version(struct file *filep, struct kfd_process *p, static int set_queue_properties_from_user(struct queue_properties *q_properties, struct kfd_ioctl_create_queue_args *args) { - if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) { + /* + * Repurpose queue percentage to accommodate new features: + * bit 0-7: queue percentage + * bit 8-15: pm4_target_xcc + */ + if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) { pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); return -EINVAL; } @@ -236,7 +241,9 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties, q_properties->is_interop = false; q_properties->is_gws = false; - q_properties->queue_percent = args->queue_percentage; + q_properties->queue_percent = args->queue_percentage & 0xFF; + /* bit 8-15 are repurposed to be PM4 target XCC */ + q_properties->pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF; q_properties->priority = args->queue_priority; q_properties->queue_address = args->ring_base_address; q_properties->queue_size = args->ring_size; @@ -442,7 +449,12 @@ static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p, struct kfd_ioctl_update_queue_args *args = data; struct queue_properties properties; - if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) { + /* + * Repurpose queue percentage to accommodate new features: + * bit 0-7: queue percentage + * bit 8-15: pm4_target_xcc + */ + if ((args->queue_percentage & 0xFF) > KFD_MAX_QUEUE_PERCENTAGE) { pr_err("Queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n"); return -EINVAL; } @@ -466,7 +478,9 @@ static int kfd_ioctl_update_queue(struct file *filp, struct kfd_process *p, properties.queue_address = args->ring_base_address; properties.queue_size = args->ring_size; - properties.queue_percent = args->queue_percentage; + properties.queue_percent = args->queue_percentage & 0xFF; + /* bit 8-15 are repurposed to be PM4 target XCC */ + properties.pm4_target_xcc = (args->queue_percentage >> 8) & 0xFF; properties.priority = args->queue_priority; pr_debug("Updating queue id %d for pasid 0x%x\n", diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index c677322057dd..b46c984b3a17 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -587,6 +587,7 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, /* PM4 Queue */ m->compute_current_logic_xcc_id = 0; m->compute_tg_chunk_size = 0; + m->pm4_target_xcc_in_xcp = q->pm4_target_xcc; } if (xcc == 0) { @@ -627,6 +628,7 @@ static void update_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, /* PM4 Queue */ m->compute_current_logic_xcc_id = 0; m->compute_tg_chunk_size = 0; + m->pm4_target_xcc_in_xcp = q->pm4_target_xcc; } } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 873b49238dc1..1337fcdf8958 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -509,6 +509,7 @@ struct queue_properties { bool is_evicted; bool is_active; bool is_gws; + uint32_t pm4_target_xcc; /* Not relevant for user mode queues in cp scheduling */ unsigned int vmid; /* Relevant only for sdma queues*/ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index b1fb017b2ef8..2b2ae0c9902b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -477,6 +477,7 @@ int pqm_update_queue_properties(struct process_queue_manager *pqm, pqn->q->properties.queue_size = p->queue_size; pqn->q->properties.queue_percent = p->queue_percent; pqn->q->properties.priority = p->priority; + pqn->q->properties.pm4_target_xcc = p->pm4_target_xcc; retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm, pqn->q, NULL); -- cgit v1.2.3 From e2069a7b0880ccdc6fa6530b6091e47168705425 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 22:20:52 -0400 Subject: drm/amdkfd: Add XCC instance to kgd2kfd interface (v3) Gfx 9 starts to have multiple XCC instances in one device. Add instance parameter to kgd2kfd functions where XCC instance was hard coded as 0. Also, update code to pass the correct instance number when running on a multi-XCC setup. v2: introduce the XCC instance to gfx v11 (Morris) v3: rebase (Alex) Signed-off-by: Amber Lin Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Tested-by: Amber Lin Signed-off-by: Morris Zhang Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c | 38 ++--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 22 +-- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c | 27 ++-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c | 19 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c | 17 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c | 17 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 160 +++++++++++---------- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 29 ++-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 103 ++++++++----- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 15 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +- drivers/gpu/drm/amd/include/kgd_kfd_interface.h | 25 ++-- 17 files changed, 270 insertions(+), 218 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c index 562e1a04160f..49d8087e469e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c @@ -33,7 +33,7 @@ #include "soc15.h" static int kgd_gfx_v9_4_3_set_pasid_vmid_mapping(struct amdgpu_device *adev, - u32 pasid, unsigned int vmid) + u32 pasid, unsigned int vmid, uint32_t inst) { unsigned long timeout; @@ -47,11 +47,11 @@ static int kgd_gfx_v9_4_3_set_pasid_vmid_mapping(struct amdgpu_device *adev, uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid | ATC_VMID0_PASID_MAPPING__VALID_MASK; - WREG32(SOC15_REG_OFFSET(ATHUB, 0, + WREG32(SOC15_REG_OFFSET(ATHUB, inst, regATC_VMID0_PASID_MAPPING) + vmid, pasid_mapping); timeout = jiffies + msecs_to_jiffies(10); - while (!(RREG32(SOC15_REG_OFFSET(ATHUB, 0, + while (!(RREG32(SOC15_REG_OFFSET(ATHUB, inst, regATC_VMID_PASID_MAPPING_UPDATE_STATUS)) & (1U << vmid))) { if (time_after(jiffies, timeout)) { @@ -61,13 +61,13 @@ static int kgd_gfx_v9_4_3_set_pasid_vmid_mapping(struct amdgpu_device *adev, cpu_relax(); } - WREG32(SOC15_REG_OFFSET(ATHUB, 0, + WREG32(SOC15_REG_OFFSET(ATHUB, inst, regATC_VMID_PASID_MAPPING_UPDATE_STATUS), 1U << vmid); - WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid, + WREG32(SOC15_REG_OFFSET(OSSSYS, inst, mmIH_VMID_0_LUT) + vmid, pasid_mapping); - WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid, + WREG32(SOC15_REG_OFFSET(OSSSYS, inst, mmIH_VMID_0_LUT_MM) + vmid, pasid_mapping); return 0; @@ -81,7 +81,7 @@ static inline struct v9_mqd *get_mqd(void *mqd) static int kgd_gfx_v9_4_3_hqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, - uint32_t wptr_mask, struct mm_struct *mm) + uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst) { struct v9_mqd *m; uint32_t *mqd_hqd; @@ -89,12 +89,12 @@ static int kgd_gfx_v9_4_3_hqd_load(struct amdgpu_device *adev, void *mqd, m = get_mqd(mqd); - kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst); /* HQD registers extend to CP_HQD_AQL_DISPATCH_ID_HI */ mqd_hqd = &m->cp_mqd_base_addr_lo; - hqd_base = SOC15_REG_OFFSET(GC, 0, regCP_MQD_BASE_ADDR); - hqd_end = SOC15_REG_OFFSET(GC, 0, regCP_HQD_AQL_DISPATCH_ID_HI); + hqd_base = SOC15_REG_OFFSET(GC, inst, regCP_MQD_BASE_ADDR); + hqd_end = SOC15_REG_OFFSET(GC, inst, regCP_HQD_AQL_DISPATCH_ID_HI); for (reg = hqd_base; reg <= hqd_end; reg++) WREG32_RLC(reg, mqd_hqd[reg - hqd_base]); @@ -103,7 +103,7 @@ static int kgd_gfx_v9_4_3_hqd_load(struct amdgpu_device *adev, void *mqd, /* Activate doorbell logic before triggering WPTR poll. */ data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_PQ_DOORBELL_CONTROL), data); if (wptr) { @@ -133,29 +133,29 @@ static int kgd_gfx_v9_4_3_hqd_load(struct amdgpu_device *adev, void *mqd, guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1); guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32; - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_LO), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_PQ_WPTR_LO), lower_32_bits(guessed_wptr)); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_HI), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_PQ_WPTR_HI), upper_32_bits(guessed_wptr)); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_PQ_WPTR_POLL_ADDR), lower_32_bits((uintptr_t)wptr)); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_PQ_WPTR_POLL_ADDR_HI), upper_32_bits((uintptr_t)wptr)); - WREG32(SOC15_REG_OFFSET(GC, 0, regCP_PQ_WPTR_POLL_CNTL1), + WREG32(SOC15_REG_OFFSET(GC, inst, regCP_PQ_WPTR_POLL_CNTL1), (uint32_t)kgd_gfx_v9_get_queue_mask(adev, pipe_id, queue_id)); } /* Start the EOP fetcher */ - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_EOP_RPTR), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_EOP_RPTR), REG_SET_FIELD(m->cp_hqd_eop_rptr, CP_HQD_EOP_RPTR, INIT_FETCHER, 1)); data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, regCP_HQD_ACTIVE), data); + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, regCP_HQD_ACTIVE), data); - kgd_gfx_v9_release_queue(adev); + kgd_gfx_v9_release_queue(adev, inst); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index f599e1e74fcc..7b60268d93c0 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -79,7 +79,7 @@ static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmi uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases) + uint32_t sh_mem_bases, uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); @@ -91,7 +91,7 @@ static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmi } static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, - unsigned int vmid) + unsigned int vmid, uint32_t inst) { /* * We have to assume that there is no outstanding mapping. @@ -135,7 +135,8 @@ static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, * but still works */ -static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id) +static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst) { uint32_t mec; uint32_t pipe; @@ -205,7 +206,7 @@ static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd) static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, - uint32_t wptr_mask, struct mm_struct *mm) + uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst) { struct v10_compute_mqd *m; uint32_t *mqd_hqd; @@ -286,7 +287,7 @@ static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd, static int kgd_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, - uint32_t doorbell_off) + uint32_t doorbell_off, uint32_t inst) { struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring; struct v10_compute_mqd *m; @@ -338,7 +339,7 @@ out_unlock: static int kgd_hqd_dump(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs) + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst) { uint32_t i = 0, reg; #define HQD_N_REGS 56 @@ -469,7 +470,7 @@ static int kgd_hqd_sdma_dump(struct amdgpu_device *adev, static bool kgd_hqd_is_occupied(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t act; bool retval = false; @@ -510,7 +511,7 @@ static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd) static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { enum hqd_dequeue_request_type type; unsigned long end_jiffies; @@ -673,7 +674,7 @@ static bool get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, static int kgd_wave_control_execute(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd) + uint32_t sq_cmd, uint32_t inst) { uint32_t data = 0; @@ -709,7 +710,8 @@ static void set_vm_context_page_table_base(struct amdgpu_device *adev, } static void program_trap_handler_settings(struct amdgpu_device *adev, - uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr) + uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, + uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c index 5c4152ae44da..52d0d35fb58d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c @@ -80,7 +80,7 @@ static void program_sh_mem_settings_v10_3(struct amdgpu_device *adev, uint32_t v uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases) + uint32_t sh_mem_bases, uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); @@ -93,7 +93,7 @@ static void program_sh_mem_settings_v10_3(struct amdgpu_device *adev, uint32_t v /* ATC is defeatured on Sienna_Cichlid */ static int set_pasid_vmid_mapping_v10_3(struct amdgpu_device *adev, unsigned int pasid, - unsigned int vmid) + unsigned int vmid, uint32_t inst) { uint32_t value = pasid << IH_VMID_0_LUT__PASID__SHIFT; @@ -105,7 +105,8 @@ static int set_pasid_vmid_mapping_v10_3(struct amdgpu_device *adev, unsigned int return 0; } -static int init_interrupts_v10_3(struct amdgpu_device *adev, uint32_t pipe_id) +static int init_interrupts_v10_3(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst) { uint32_t mec; uint32_t pipe; @@ -177,7 +178,7 @@ static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd) static int hqd_load_v10_3(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, - uint32_t wptr_mask, struct mm_struct *mm) + uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst) { struct v10_compute_mqd *m; uint32_t *mqd_hqd; @@ -273,7 +274,7 @@ static int hqd_load_v10_3(struct amdgpu_device *adev, void *mqd, static int hiq_mqd_load_v10_3(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, - uint32_t doorbell_off) + uint32_t doorbell_off, uint32_t inst) { struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring; struct v10_compute_mqd *m; @@ -325,7 +326,7 @@ out_unlock: static int hqd_dump_v10_3(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs) + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst) { uint32_t i = 0, reg; #define HQD_N_REGS 56 @@ -456,7 +457,7 @@ static int hqd_sdma_dump_v10_3(struct amdgpu_device *adev, static bool hqd_is_occupied_v10_3(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t act; bool retval = false; @@ -498,7 +499,7 @@ static bool hqd_sdma_is_occupied_v10_3(struct amdgpu_device *adev, static int hqd_destroy_v10_3(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { enum hqd_dequeue_request_type type; unsigned long end_jiffies; @@ -586,7 +587,7 @@ static int hqd_sdma_destroy_v10_3(struct amdgpu_device *adev, void *mqd, static int wave_control_execute_v10_3(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd) + uint32_t sq_cmd, uint32_t inst) { uint32_t data = 0; @@ -628,7 +629,8 @@ static void set_vm_context_page_table_base_v10_3(struct amdgpu_device *adev, } static void program_trap_handler_settings_v10_3(struct amdgpu_device *adev, - uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr) + uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, + uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); @@ -765,7 +767,7 @@ uint32_t set_wave_launch_mode_v10_3(struct amdgpu_device *adev, * deq_retry_wait_time -- Wait Count for Global Wave Syncs. */ void get_iq_wait_times_v10_3(struct amdgpu_device *adev, - uint32_t *wait_times) + uint32_t *wait_times, uint32_t inst) { *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2)); @@ -775,7 +777,8 @@ void build_grace_period_packet_info_v10_3(struct amdgpu_device *adev, uint32_t wait_times, uint32_t grace_period, uint32_t *reg_offset, - uint32_t *reg_data) + uint32_t *reg_data, + uint32_t inst) { *reg_data = wait_times; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c index 5cdb7289d35b..7deff8a547fb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c @@ -78,7 +78,7 @@ static void program_sh_mem_settings_v11(struct amdgpu_device *adev, uint32_t vmi uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases) + uint32_t sh_mem_bases, uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); @@ -89,7 +89,7 @@ static void program_sh_mem_settings_v11(struct amdgpu_device *adev, uint32_t vmi } static int set_pasid_vmid_mapping_v11(struct amdgpu_device *adev, unsigned int pasid, - unsigned int vmid) + unsigned int vmid, uint32_t inst) { uint32_t value = pasid << IH_VMID_0_LUT__PASID__SHIFT; @@ -101,7 +101,8 @@ static int set_pasid_vmid_mapping_v11(struct amdgpu_device *adev, unsigned int p return 0; } -static int init_interrupts_v11(struct amdgpu_device *adev, uint32_t pipe_id) +static int init_interrupts_v11(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst) { uint32_t mec; uint32_t pipe; @@ -162,7 +163,7 @@ static inline struct v11_sdma_mqd *get_sdma_mqd(void *mqd) static int hqd_load_v11(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, uint32_t wptr_mask, - struct mm_struct *mm) + struct mm_struct *mm, uint32_t inst) { struct v11_compute_mqd *m; uint32_t *mqd_hqd; @@ -258,7 +259,7 @@ static int hqd_load_v11(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, static int hiq_mqd_load_v11(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, - uint32_t doorbell_off) + uint32_t doorbell_off, uint32_t inst) { struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring; struct v11_compute_mqd *m; @@ -310,7 +311,7 @@ out_unlock: static int hqd_dump_v11(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs) + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst) { uint32_t i = 0, reg; #define HQD_N_REGS 56 @@ -445,7 +446,7 @@ static int hqd_sdma_dump_v11(struct amdgpu_device *adev, } static bool hqd_is_occupied_v11(struct amdgpu_device *adev, uint64_t queue_address, - uint32_t pipe_id, uint32_t queue_id) + uint32_t pipe_id, uint32_t queue_id, uint32_t inst) { uint32_t act; bool retval = false; @@ -486,7 +487,7 @@ static bool hqd_sdma_is_occupied_v11(struct amdgpu_device *adev, void *mqd) static int hqd_destroy_v11(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { enum hqd_dequeue_request_type type; unsigned long end_jiffies; @@ -571,7 +572,7 @@ static int hqd_sdma_destroy_v11(struct amdgpu_device *adev, void *mqd, static int wave_control_execute_v11(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd) + uint32_t sq_cmd, uint32_t inst) { uint32_t data = 0; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c index e83cb1c09610..6bf448ab3dff 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c @@ -78,7 +78,7 @@ static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmi uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases) + uint32_t sh_mem_bases, uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); @@ -91,7 +91,7 @@ static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmi } static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, - unsigned int vmid) + unsigned int vmid, uint32_t inst) { /* * We have to assume that there is no outstanding mapping. @@ -114,7 +114,8 @@ static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, return 0; } -static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id) +static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst) { uint32_t mec; uint32_t pipe; @@ -158,7 +159,7 @@ static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd) static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, - uint32_t wptr_mask, struct mm_struct *mm) + uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst) { struct cik_mqd *m; uint32_t *mqd_hqd; @@ -202,7 +203,7 @@ static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd, static int kgd_hqd_dump(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs) + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst) { uint32_t i = 0, reg; #define HQD_N_REGS (35+4) @@ -318,7 +319,7 @@ static int kgd_hqd_sdma_dump(struct amdgpu_device *adev, static bool kgd_hqd_is_occupied(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t act; bool retval = false; @@ -358,7 +359,7 @@ static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd) static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t temp; enum hqd_dequeue_request_type type; @@ -494,7 +495,7 @@ static int kgd_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd, static int kgd_wave_control_execute(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd) + uint32_t sq_cmd, uint32_t inst) { uint32_t data; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c index 870f352837fc..cd06e4a6d1da 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c @@ -72,7 +72,7 @@ static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmi uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases) + uint32_t sh_mem_bases, uint32_t inst) { lock_srbm(adev, 0, 0, 0, vmid); @@ -85,7 +85,7 @@ static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmi } static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, - unsigned int vmid) + unsigned int vmid, uint32_t inst) { /* * We have to assume that there is no outstanding mapping. @@ -109,7 +109,8 @@ static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, return 0; } -static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id) +static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst) { uint32_t mec; uint32_t pipe; @@ -153,7 +154,7 @@ static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd) static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, - uint32_t wptr_mask, struct mm_struct *mm) + uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst) { struct vi_mqd *m; uint32_t *mqd_hqd; @@ -226,7 +227,7 @@ static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd, static int kgd_hqd_dump(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs) + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst) { uint32_t i = 0, reg; #define HQD_N_REGS (54+4) @@ -350,7 +351,7 @@ static int kgd_hqd_sdma_dump(struct amdgpu_device *adev, static bool kgd_hqd_is_occupied(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t act; bool retval = false; @@ -390,7 +391,7 @@ static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd) static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t temp; enum hqd_dequeue_request_type type; @@ -540,7 +541,7 @@ static bool get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, static int kgd_wave_control_execute(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd) + uint32_t sq_cmd, uint32_t inst) { uint32_t data = 0; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index d36219ecd3dd..7918a00cbb5a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -46,26 +46,26 @@ enum hqd_dequeue_request_type { SAVE_WAVES }; -static void lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe, - uint32_t queue, uint32_t vmid) +static void kgd_gfx_v9_lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe, + uint32_t queue, uint32_t vmid, uint32_t inst) { mutex_lock(&adev->srbm_mutex); - soc15_grbm_select(adev, mec, pipe, queue, vmid, 0); + soc15_grbm_select(adev, mec, pipe, queue, vmid, inst); } -static void unlock_srbm(struct amdgpu_device *adev) +static void kgd_gfx_v9_unlock_srbm(struct amdgpu_device *adev, uint32_t inst) { - soc15_grbm_select(adev, 0, 0, 0, 0, 0); + soc15_grbm_select(adev, 0, 0, 0, 0, inst); mutex_unlock(&adev->srbm_mutex); } void kgd_gfx_v9_acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec); - lock_srbm(adev, mec, pipe, queue_id, 0); + kgd_gfx_v9_lock_srbm(adev, mec, pipe, queue_id, 0, inst); } uint64_t kgd_gfx_v9_get_queue_mask(struct amdgpu_device *adev, @@ -77,28 +77,28 @@ uint64_t kgd_gfx_v9_get_queue_mask(struct amdgpu_device *adev, return 1ull << bit; } -void kgd_gfx_v9_release_queue(struct amdgpu_device *adev) +void kgd_gfx_v9_release_queue(struct amdgpu_device *adev, uint32_t inst) { - unlock_srbm(adev); + kgd_gfx_v9_unlock_srbm(adev, inst); } void kgd_gfx_v9_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid, uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases) + uint32_t sh_mem_bases, uint32_t inst) { - lock_srbm(adev, 0, 0, 0, vmid); + kgd_gfx_v9_lock_srbm(adev, 0, 0, 0, vmid, inst); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases); + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmSH_MEM_CONFIG), sh_mem_config); + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmSH_MEM_BASES), sh_mem_bases); /* APE1 no longer exists on GFX9 */ - unlock_srbm(adev); + kgd_gfx_v9_unlock_srbm(adev, inst); } int kgd_gfx_v9_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, - unsigned int vmid) + unsigned int vmid, uint32_t inst) { /* * We have to assume that there is no outstanding mapping. @@ -156,7 +156,8 @@ int kgd_gfx_v9_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, * but still works */ -int kgd_gfx_v9_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id) +int kgd_gfx_v9_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst) { uint32_t mec; uint32_t pipe; @@ -164,13 +165,13 @@ int kgd_gfx_v9_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id) mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec); - lock_srbm(adev, mec, pipe, 0, 0); + kgd_gfx_v9_lock_srbm(adev, mec, pipe, 0, 0, inst); - WREG32_SOC15(GC, 0, mmCPC_INT_CNTL, + WREG32_SOC15(GC, inst, mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK | CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK); - unlock_srbm(adev); + kgd_gfx_v9_unlock_srbm(adev, inst); return 0; } @@ -220,7 +221,8 @@ static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd) int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, - uint32_t wptr_mask, struct mm_struct *mm) + uint32_t wptr_mask, struct mm_struct *mm, + uint32_t inst) { struct v9_mqd *m; uint32_t *mqd_hqd; @@ -228,21 +230,22 @@ int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, m = get_mqd(mqd); - kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst); /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */ mqd_hqd = &m->cp_mqd_base_addr_lo; - hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR); + hqd_base = SOC15_REG_OFFSET(GC, inst, mmCP_MQD_BASE_ADDR); for (reg = hqd_base; - reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++) + reg <= SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_WPTR_HI); reg++) WREG32_RLC(reg, mqd_hqd[reg - hqd_base]); /* Activate doorbell logic before triggering WPTR poll. */ data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data); + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_DOORBELL_CONTROL), + data); if (wptr) { /* Don't read wptr with get_user because the user @@ -271,43 +274,43 @@ int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1); guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32; - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_WPTR_LO), lower_32_bits(guessed_wptr)); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_WPTR_HI), upper_32_bits(guessed_wptr)); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_WPTR_POLL_ADDR), lower_32_bits((uintptr_t)wptr)); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI), upper_32_bits((uintptr_t)wptr)); WREG32_SOC15(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1, (uint32_t)kgd_gfx_v9_get_queue_mask(adev, pipe_id, queue_id)); } /* Start the EOP fetcher */ - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR), + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_EOP_RPTR), REG_SET_FIELD(m->cp_hqd_eop_rptr, CP_HQD_EOP_RPTR, INIT_FETCHER, 1)); data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1); - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data); + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_ACTIVE), data); - kgd_gfx_v9_release_queue(adev); + kgd_gfx_v9_release_queue(adev, inst); return 0; } int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, - uint32_t doorbell_off) + uint32_t doorbell_off, uint32_t inst) { - struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring; + struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[inst].ring; struct v9_mqd *m; uint32_t mec, pipe; int r; m = get_mqd(mqd); - kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst); mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec); @@ -315,7 +318,7 @@ int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n", mec, pipe, queue_id); - spin_lock(&adev->gfx.kiq[0].ring_lock); + spin_lock(&adev->gfx.kiq[inst].ring_lock); r = amdgpu_ring_alloc(kiq_ring, 7); if (r) { pr_err("Failed to alloc KIQ (%d).\n", r); @@ -342,15 +345,15 @@ int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, amdgpu_ring_commit(kiq_ring); out_unlock: - spin_unlock(&adev->gfx.kiq[0].ring_lock); - kgd_gfx_v9_release_queue(adev); + spin_unlock(&adev->gfx.kiq[inst].ring_lock); + kgd_gfx_v9_release_queue(adev, inst); return r; } int kgd_gfx_v9_hqd_dump(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs) + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst) { uint32_t i = 0, reg; #define HQD_N_REGS 56 @@ -365,13 +368,13 @@ int kgd_gfx_v9_hqd_dump(struct amdgpu_device *adev, if (*dump == NULL) return -ENOMEM; - kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst); - for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR); - reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++) + for (reg = SOC15_REG_OFFSET(GC, inst, mmCP_MQD_BASE_ADDR); + reg <= SOC15_REG_OFFSET(GC, inst, mmCP_HQD_PQ_WPTR_HI); reg++) DUMP_REG(reg); - kgd_gfx_v9_release_queue(adev); + kgd_gfx_v9_release_queue(adev, inst); WARN_ON_ONCE(i != HQD_N_REGS); *n_regs = i; @@ -481,23 +484,23 @@ static int kgd_hqd_sdma_dump(struct amdgpu_device *adev, bool kgd_gfx_v9_hqd_is_occupied(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { uint32_t act; bool retval = false; uint32_t low, high; - kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); - act = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst); + act = RREG32_SOC15(GC, inst, mmCP_HQD_ACTIVE); if (act) { low = lower_32_bits(queue_address >> 8); high = upper_32_bits(queue_address >> 8); - if (low == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE) && - high == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI)) + if (low == RREG32_SOC15(GC, inst, mmCP_HQD_PQ_BASE) && + high == RREG32_SOC15(GC, inst, mmCP_HQD_PQ_BASE_HI)) retval = true; } - kgd_gfx_v9_release_queue(adev); + kgd_gfx_v9_release_queue(adev, inst); return retval; } @@ -522,7 +525,7 @@ static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd) int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id) + uint32_t queue_id, uint32_t inst) { enum hqd_dequeue_request_type type; unsigned long end_jiffies; @@ -532,10 +535,10 @@ int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd, if (amdgpu_in_reset(adev)) return -EIO; - kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id); + kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst); if (m->cp_hqd_vmid == 0) - WREG32_FIELD15_RLC(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0); + WREG32_FIELD15_RLC(GC, inst, RLC_CP_SCHEDULERS, scheduler1, 0); switch (reset_type) { case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN: @@ -552,22 +555,22 @@ int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd, break; } - WREG32_RLC(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type); + WREG32_RLC(SOC15_REG_OFFSET(GC, inst, mmCP_HQD_DEQUEUE_REQUEST), type); end_jiffies = (utimeout * HZ / 1000) + jiffies; while (true) { - temp = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE); + temp = RREG32_SOC15(GC, inst, mmCP_HQD_ACTIVE); if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK)) break; if (time_after(jiffies, end_jiffies)) { pr_err("cp queue preemption time out.\n"); - kgd_gfx_v9_release_queue(adev); + kgd_gfx_v9_release_queue(adev, inst); return -ETIME; } usleep_range(500, 1000); } - kgd_gfx_v9_release_queue(adev); + kgd_gfx_v9_release_queue(adev, inst); return 0; } @@ -624,14 +627,14 @@ bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, int kgd_gfx_v9_wave_control_execute(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd) + uint32_t sq_cmd, uint32_t inst) { uint32_t data = 0; mutex_lock(&adev->grbm_idx_mutex); - WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, gfx_index_val); - WREG32_SOC15(GC, 0, mmSQ_CMD, sq_cmd); + WREG32_SOC15_RLC_SHADOW(GC, inst, mmGRBM_GFX_INDEX, gfx_index_val); + WREG32_SOC15(GC, inst, mmSQ_CMD, sq_cmd); data = REG_SET_FIELD(data, GRBM_GFX_INDEX, INSTANCE_BROADCAST_WRITES, 1); @@ -640,7 +643,7 @@ int kgd_gfx_v9_wave_control_execute(struct amdgpu_device *adev, data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1); - WREG32_SOC15_RLC_SHADOW(GC, 0, mmGRBM_GFX_INDEX, data); + WREG32_SOC15_RLC_SHADOW(GC, inst, mmGRBM_GFX_INDEX, data); mutex_unlock(&adev->grbm_idx_mutex); return 0; @@ -685,7 +688,7 @@ static void unlock_spi_csq_mutexes(struct amdgpu_device *adev) * is being collected */ static void get_wave_count(struct amdgpu_device *adev, int queue_idx, - int *wave_cnt, int *vmid) + int *wave_cnt, int *vmid, uint32_t inst) { int pipe_idx; int queue_slot; @@ -700,12 +703,12 @@ static void get_wave_count(struct amdgpu_device *adev, int queue_idx, *wave_cnt = 0; pipe_idx = queue_idx / adev->gfx.mec.num_queue_per_pipe; queue_slot = queue_idx % adev->gfx.mec.num_queue_per_pipe; - soc15_grbm_select(adev, 1, pipe_idx, queue_slot, 0, 0); - reg_val = RREG32_SOC15_IP(GC, SOC15_REG_OFFSET(GC, 0, mmSPI_CSQ_WF_ACTIVE_COUNT_0) + + soc15_grbm_select(adev, 1, pipe_idx, queue_slot, 0, inst); + reg_val = RREG32_SOC15_IP(GC, SOC15_REG_OFFSET(GC, inst, mmSPI_CSQ_WF_ACTIVE_COUNT_0) + queue_slot); *wave_cnt = reg_val & SPI_CSQ_WF_ACTIVE_COUNT_0__COUNT_MASK; if (*wave_cnt != 0) - *vmid = (RREG32_SOC15(GC, 0, mmCP_HQD_VMID) & + *vmid = (RREG32_SOC15(GC, inst, mmCP_HQD_VMID) & CP_HQD_VMID__VMID_MASK) >> CP_HQD_VMID__VMID__SHIFT; } @@ -756,7 +759,7 @@ static void get_wave_count(struct amdgpu_device *adev, int queue_idx, * Reading registers referenced above involves programming GRBM appropriately */ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, - int *pasid_wave_cnt, int *max_waves_per_cu) + int *pasid_wave_cnt, int *max_waves_per_cu, uint32_t inst) { int qidx; int vmid; @@ -772,7 +775,7 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, DECLARE_BITMAP(cp_queue_bitmap, KGD_MAX_QUEUES); lock_spi_csq_mutexes(adev); - soc15_grbm_select(adev, 1, 0, 0, 0, 0); + soc15_grbm_select(adev, 1, 0, 0, 0, inst); /* * Iterate through the shader engines and arrays of the device @@ -787,8 +790,8 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, for (se_idx = 0; se_idx < se_cnt; se_idx++) { for (sh_idx = 0; sh_idx < sh_cnt; sh_idx++) { - amdgpu_gfx_select_se_sh(adev, se_idx, sh_idx, 0xffffffff, 0); - queue_map = RREG32_SOC15(GC, 0, mmSPI_CSQ_WF_ACTIVE_STATUS); + amdgpu_gfx_select_se_sh(adev, se_idx, sh_idx, 0xffffffff, inst); + queue_map = RREG32_SOC15(GC, inst, mmSPI_CSQ_WF_ACTIVE_STATUS); /* * Assumption: queue map encodes following schema: four @@ -808,10 +811,11 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, continue; /* Get number of waves in flight and aggregate them */ - get_wave_count(adev, qidx, &wave_cnt, &vmid); + get_wave_count(adev, qidx, &wave_cnt, &vmid, + inst); if (wave_cnt != 0) { pasid_tmp = - RREG32(SOC15_REG_OFFSET(OSSSYS, 0, + RREG32(SOC15_REG_OFFSET(OSSSYS, inst, mmIH_VMID_0_LUT) + vmid); if (pasid_tmp == pasid) vmid_wave_cnt += wave_cnt; @@ -820,8 +824,8 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, } } - amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, 0); - soc15_grbm_select(adev, 0, 0, 0, 0, 0); + amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, inst); + soc15_grbm_select(adev, 0, 0, 0, 0, inst); unlock_spi_csq_mutexes(adev); /* Update the output parameters and return */ @@ -831,27 +835,27 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, } void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev, - uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr) + uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, uint32_t inst) { - lock_srbm(adev, 0, 0, 0, vmid); + kgd_gfx_v9_lock_srbm(adev, 0, 0, 0, vmid, inst); /* * Program TBA registers */ - WREG32_SOC15(GC, 0, mmSQ_SHADER_TBA_LO, + WREG32_SOC15(GC, inst, mmSQ_SHADER_TBA_LO, lower_32_bits(tba_addr >> 8)); - WREG32_SOC15(GC, 0, mmSQ_SHADER_TBA_HI, + WREG32_SOC15(GC, inst, mmSQ_SHADER_TBA_HI, upper_32_bits(tba_addr >> 8)); /* * Program TMA registers */ - WREG32_SOC15(GC, 0, mmSQ_SHADER_TMA_LO, + WREG32_SOC15(GC, inst, mmSQ_SHADER_TMA_LO, lower_32_bits(tma_addr >> 8)); - WREG32_SOC15(GC, 0, mmSQ_SHADER_TMA_HI, + WREG32_SOC15(GC, inst, mmSQ_SHADER_TMA_HI, upper_32_bits(tma_addr >> 8)); - unlock_srbm(adev); + kgd_gfx_v9_unlock_srbm(adev, inst); } const struct kfd2kgd_calls gfx_v9_kfd2kgd = { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index 491273a02e30..a241299f4fbc 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -25,41 +25,42 @@ void kgd_gfx_v9_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid, uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases); + uint32_t sh_mem_bases, uint32_t inst); int kgd_gfx_v9_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, - unsigned int vmid); -int kgd_gfx_v9_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id); + unsigned int vmid, uint32_t inst); +int kgd_gfx_v9_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst); int kgd_gfx_v9_hqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, uint32_t wptr_mask, - struct mm_struct *mm); + struct mm_struct *mm, uint32_t inst); int kgd_gfx_v9_hiq_mqd_load(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, - uint32_t doorbell_off); + uint32_t doorbell_off, uint32_t inst); int kgd_gfx_v9_hqd_dump(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs); + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst); bool kgd_gfx_v9_hqd_is_occupied(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id); + uint32_t queue_id, uint32_t inst); int kgd_gfx_v9_hqd_destroy(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id); + uint32_t queue_id, uint32_t inst); int kgd_gfx_v9_wave_control_execute(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd); + uint32_t sq_cmd, uint32_t inst); bool kgd_gfx_v9_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev, uint8_t vmid, uint16_t *p_pasid); - void kgd_gfx_v9_set_vm_context_page_table_base(struct amdgpu_device *adev, uint32_t vmid, uint64_t page_table_base); void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, - int *pasid_wave_cnt, int *max_waves_per_cu); + int *pasid_wave_cnt, int *max_waves_per_cu, uint32_t inst); void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev, - uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr); + uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, + uint32_t inst); void kgd_gfx_v9_acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id, - uint32_t queue_id); + uint32_t queue_id, uint32_t inst); uint64_t kgd_gfx_v9_get_queue_mask(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id); -void kgd_gfx_v9_release_queue(struct amdgpu_device *adev); +void kgd_gfx_v9_release_queue(struct amdgpu_device *adev, uint32_t inst); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 9afd3295ca85..33a9394f9e58 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -127,12 +127,16 @@ static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manag void program_sh_mem_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { - return dqm->dev->kfd2kgd->program_sh_mem_settings( + int xcc = 0; + + for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) + dqm->dev->kfd2kgd->program_sh_mem_settings( dqm->dev->adev, qpd->vmid, qpd->sh_mem_config, qpd->sh_mem_ape1_base, qpd->sh_mem_ape1_limit, - qpd->sh_mem_bases); + qpd->sh_mem_bases, + dqm->dev->start_xcc_id + xcc); } static void kfd_hws_hang(struct device_queue_manager *dqm) @@ -405,10 +409,14 @@ static void deallocate_doorbell(struct qcm_process_device *qpd, static void program_trap_handler_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { + int xcc = 0; + if (dqm->dev->kfd2kgd->program_trap_handler_settings) - dqm->dev->kfd2kgd->program_trap_handler_settings( + for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) + dqm->dev->kfd2kgd->program_trap_handler_settings( dqm->dev->adev, qpd->vmid, - qpd->tba_addr, qpd->tma_addr); + qpd->tba_addr, qpd->tma_addr, + dqm->dev->start_xcc_id + xcc); } static int allocate_vmid(struct device_queue_manager *dqm, @@ -671,6 +679,7 @@ static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process struct kfd_process_device *pdd; int first_vmid_to_scan = dev->vm_info.first_vmid_kfd; int last_vmid_to_scan = dev->vm_info.last_vmid_kfd; + int xcc = 0; reg_sq_cmd.u32All = 0; reg_gfx_index.u32All = 0; @@ -715,9 +724,11 @@ static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL; reg_sq_cmd.bits.vm_id = vmid; - dev->kfd2kgd->wave_control_execute(dev->adev, + for (xcc = 0; xcc < dev->num_xcc_per_node; xcc++) + dev->kfd2kgd->wave_control_execute(dev->adev, reg_gfx_index.u32All, - reg_sq_cmd.u32All); + reg_sq_cmd.u32All, + dev->start_xcc_id + xcc); return 0; } @@ -1229,17 +1240,32 @@ static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid, unsigned int vmid) { - return dqm->dev->kfd2kgd->set_pasid_vmid_mapping( - dqm->dev->adev, pasid, vmid); + int xcc = 0, ret; + + for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) { + ret = dqm->dev->kfd2kgd->set_pasid_vmid_mapping( + dqm->dev->adev, pasid, vmid, + dqm->dev->start_xcc_id + xcc); + if (ret) + break; + } + + return ret; } static void init_interrupts(struct device_queue_manager *dqm) { - unsigned int i; + unsigned int i, xcc; - for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) - if (is_pipe_enabled(dqm, 0, i)) - dqm->dev->kfd2kgd->init_interrupts(dqm->dev->adev, i); + for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) { + if (is_pipe_enabled(dqm, 0, i)) { + for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) + dqm->dev->kfd2kgd->init_interrupts( + dqm->dev->adev, i, + dqm->dev->start_xcc_id + + xcc); + } + } } static void init_sdma_bitmaps(struct device_queue_manager *dqm) @@ -2455,44 +2481,49 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) struct device_queue_manager *dqm = data; uint32_t (*dump)[2], n_regs; int pipe, queue; - int r = 0; + int r = 0, xcc; + uint32_t inst; if (!dqm->sched_running) { seq_puts(m, " Device is stopped\n"); return 0; } - r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, + for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) { + inst = dqm->dev->start_xcc_id + xcc; + r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE, - &dump, &n_regs); - if (!r) { - seq_printf(m, " HIQ on MEC %d Pipe %d Queue %d\n", - KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1, - KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm), - KFD_CIK_HIQ_QUEUE); - seq_reg_dump(m, dump, n_regs); + &dump, &n_regs, inst); + if (!r) { + seq_printf(m, + " Inst %d, HIQ on MEC %d Pipe %d Queue %d\n", + inst, KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1, + KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm), + KFD_CIK_HIQ_QUEUE); + seq_reg_dump(m, dump, n_regs); - kfree(dump); - } + kfree(dump); + } - for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { - int pipe_offset = pipe * get_queues_per_pipe(dqm); + for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { + int pipe_offset = pipe * get_queues_per_pipe(dqm); - for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { - if (!test_bit(pipe_offset + queue, + for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { + if (!test_bit(pipe_offset + queue, dqm->dev->kfd->shared_resources.cp_queue_bitmap)) - continue; + continue; - r = dqm->dev->kfd2kgd->hqd_dump( - dqm->dev->adev, pipe, queue, &dump, &n_regs); - if (r) - break; + r = dqm->dev->kfd2kgd->hqd_dump( + dqm->dev->adev, pipe, queue, &dump, &n_regs, inst); + if (r) + break; - seq_printf(m, " CP Pipe %d, Queue %d\n", - pipe, queue); - seq_reg_dump(m, dump, n_regs); + seq_printf(m, " Inst %d, CP Pipe %d, Queue %d\n", + inst, pipe, queue); + seq_reg_dump(m, dump, n_regs); - kfree(dump); + kfree(dump); + } } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c index 074f6075ccc7..d81125421aaf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c @@ -190,7 +190,7 @@ int kfd_hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd, struct queue_properties *p, struct mm_struct *mms) { return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, mqd, pipe_id, - queue_id, p->doorbell_off); + queue_id, p->doorbell_off, 0); } int kfd_destroy_mqd_cp(struct mqd_manager *mm, void *mqd, @@ -198,7 +198,7 @@ int kfd_destroy_mqd_cp(struct mqd_manager *mm, void *mqd, uint32_t pipe_id, uint32_t queue_id) { return mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, mqd, type, timeout, - pipe_id, queue_id); + pipe_id, queue_id, 0); } void kfd_free_mqd_cp(struct mqd_manager *mm, void *mqd, @@ -217,7 +217,7 @@ bool kfd_is_occupied_cp(struct mqd_manager *mm, void *mqd, uint32_t queue_id) { return mm->dev->kfd2kgd->hqd_is_occupied(mm->dev->adev, queue_address, - pipe_id, queue_id); + pipe_id, queue_id, 0); } int kfd_load_mqd_sdma(struct mqd_manager *mm, void *mqd, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c index ca1966466759..eb11940bec34 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c @@ -167,7 +167,7 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, uint32_t pipe_id, return mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id, (uint32_t __user *)p->write_ptr, - wptr_shift, wptr_mask, mms); + wptr_shift, wptr_mask, mms, 0); } static void __update_mqd(struct mqd_manager *mm, void *mqd, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index c9565ea99df5..d54c6fdebbb6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -151,7 +151,7 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, r = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id, (uint32_t __user *)p->write_ptr, - wptr_shift, 0, mms); + wptr_shift, 0, mms, 0); return r; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index dff171b54b5c..338d639c1898 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -184,7 +184,7 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, r = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id, (uint32_t __user *)p->write_ptr, - wptr_shift, 0, mms); + wptr_shift, 0, mms, 0); return r; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index b46c984b3a17..b53cd8f9620b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -218,7 +218,7 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, return mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id, (uint32_t __user *)p->write_ptr, - wptr_shift, 0, mms); + wptr_shift, 0, mms, 0); } static void update_mqd(struct mqd_manager *mm, void *mqd, @@ -501,13 +501,15 @@ static int hiq_load_mqd_kiq_v9_4_3(struct mqd_manager *mm, void *mqd, { int xcc, err; void *xcc_mqd; + uint32_t start_inst = mm->dev->start_xcc_id; uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev); for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { xcc_mqd = mqd + hiq_mqd_size * xcc; err = mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, xcc_mqd, pipe_id, queue_id, - p->doorbell_off); + p->doorbell_off, + start_inst+xcc); if (err) { pr_debug("Failed to load HIQ MQD for XCC: %d\n", xcc); break; @@ -523,13 +525,14 @@ static int destroy_hiq_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, { int xcc = 0, err; void *xcc_mqd; + uint32_t start_inst = mm->dev->start_xcc_id; uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev); for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { xcc_mqd = mqd + hiq_mqd_size * xcc; err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd, type, timeout, pipe_id, - queue_id); + queue_id, start_inst+xcc); if (err) { pr_debug("Destroy MQD failed for xcc: %d\n", xcc); break; @@ -641,6 +644,7 @@ static int destroy_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, void *xcc_mqd; struct v9_mqd *m; uint64_t mqd_offset; + uint32_t start_inst = mm->dev->start_xcc_id; m = get_mqd(mqd); mqd_offset = m->cp_mqd_stride_size; @@ -649,7 +653,7 @@ static int destroy_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, xcc_mqd = mqd + mqd_offset * xcc; err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd, type, timeout, pipe_id, - queue_id); + queue_id, start_inst+xcc); if (err) { pr_debug("Destroy MQD failed for xcc: %d\n", xcc); break; @@ -667,6 +671,7 @@ static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); int xcc = 0, err; void *xcc_mqd; + uint32_t start_inst = mm->dev->start_xcc_id; uint64_t mqd_stride_size = mm->mqd_stride(mm, p); for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { @@ -674,7 +679,7 @@ static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, err = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, xcc_mqd, pipe_id, queue_id, (uint32_t __user *)p->write_ptr, - wptr_shift, 0, mms); + wptr_shift, 0, mms, start_inst+xcc); if (err) { pr_debug("Load MQD failed for xcc: %d\n", xcc); break; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c index 8736a3cdbe1e..ebf963f42b51 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c @@ -165,7 +165,7 @@ static int load_mqd(struct mqd_manager *mm, void *mqd, return mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id, (uint32_t __user *)p->write_ptr, - wptr_shift, wptr_mask, mms); + wptr_shift, wptr_mask, mms, 0); } static void __update_mqd(struct mqd_manager *mm, void *mqd, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 66e021889c64..888590dfa646 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -290,7 +290,7 @@ static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer) wave_cnt = 0; max_waves_per_cu = 0; dev->kfd2kgd->get_cu_occupancy(dev->adev, proc->pasid, &wave_cnt, - &max_waves_per_cu); + &max_waves_per_cu, 0); /* Translate wave count to number of compute units */ cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu; diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h index 5cb3e8634739..8cb3dbcae3e4 100644 --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h @@ -230,28 +230,30 @@ struct kfd2kgd_calls { /* Register access functions */ void (*program_sh_mem_settings)(struct amdgpu_device *adev, uint32_t vmid, uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, - uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases); + uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases, + uint32_t inst); int (*set_pasid_vmid_mapping)(struct amdgpu_device *adev, u32 pasid, - unsigned int vmid); + unsigned int vmid, uint32_t inst); - int (*init_interrupts)(struct amdgpu_device *adev, uint32_t pipe_id); + int (*init_interrupts)(struct amdgpu_device *adev, uint32_t pipe_id, + uint32_t inst); int (*hqd_load)(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr, uint32_t wptr_shift, uint32_t wptr_mask, - struct mm_struct *mm); + struct mm_struct *mm, uint32_t inst); int (*hiq_mqd_load)(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id, uint32_t queue_id, - uint32_t doorbell_off); + uint32_t doorbell_off, uint32_t inst); int (*hqd_sdma_load)(struct amdgpu_device *adev, void *mqd, uint32_t __user *wptr, struct mm_struct *mm); int (*hqd_dump)(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs); + uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst); int (*hqd_sdma_dump)(struct amdgpu_device *adev, uint32_t engine_id, uint32_t queue_id, @@ -259,12 +261,12 @@ struct kfd2kgd_calls { bool (*hqd_is_occupied)(struct amdgpu_device *adev, uint64_t queue_address, uint32_t pipe_id, - uint32_t queue_id); + uint32_t queue_id, uint32_t inst); int (*hqd_destroy)(struct amdgpu_device *adev, void *mqd, enum kfd_preempt_type reset_type, unsigned int timeout, uint32_t pipe_id, - uint32_t queue_id); + uint32_t queue_id, uint32_t inst); bool (*hqd_sdma_is_occupied)(struct amdgpu_device *adev, void *mqd); @@ -273,7 +275,7 @@ struct kfd2kgd_calls { int (*wave_control_execute)(struct amdgpu_device *adev, uint32_t gfx_index_val, - uint32_t sq_cmd); + uint32_t sq_cmd, uint32_t inst); bool (*get_atc_vmid_pasid_mapping_info)(struct amdgpu_device *adev, uint8_t vmid, uint16_t *p_pasid); @@ -290,9 +292,10 @@ struct kfd2kgd_calls { uint32_t (*read_vmid_from_vmfault_reg)(struct amdgpu_device *adev); void (*get_cu_occupancy)(struct amdgpu_device *adev, int pasid, - int *wave_cnt, int *max_waves_per_cu); + int *wave_cnt, int *max_waves_per_cu, uint32_t inst); void (*program_trap_handler_settings)(struct amdgpu_device *adev, - uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr); + uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, + uint32_t inst); }; #endif /* KGD_KFD_INTERFACE_H_INCLUDED */ -- cgit v1.2.3 From f87f686482c6d2d4465245356854710b01f312c1 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 22:22:20 -0400 Subject: drm/amdgpu: Add XCC inst to PASID TLB flushing Add XCC instance to select the correct KIQ ring when flushing TLBs on a multi-XCC setup. Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 6 ++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 3 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h | 7 ++++--- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 12 ++++++------ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 7 +++++-- 9 files changed, 25 insertions(+), 18 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 5afbcc390d89..9d19c7ceda3f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -743,7 +743,9 @@ int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev, } int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev, - uint16_t pasid, enum TLB_FLUSH_TYPE flush_type) + uint16_t pasid, + enum TLB_FLUSH_TYPE flush_type, + uint32_t inst) { bool all_hub = false; @@ -751,7 +753,7 @@ int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev, adev->family == AMDGPU_FAMILY_RV) all_hub = true; - return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub); + return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub, inst); } bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 01ba3589b60a..df07e212c21e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -160,7 +160,8 @@ bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev); int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev, uint16_t vmid); int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev, - uint16_t pasid, enum TLB_FLUSH_TYPE flush_type); + uint16_t pasid, enum TLB_FLUSH_TYPE flush_type, + uint32_t inst); bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h index 6d105d7fb98b..572ef5be539f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h @@ -119,7 +119,8 @@ struct amdgpu_gmc_funcs { uint32_t vmhub, uint32_t flush_type); /* flush the vm tlb via pasid */ int (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid, - uint32_t flush_type, bool all_hub); + uint32_t flush_type, bool all_hub, + uint32_t inst); /* flush the vm tlb via ring */ uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid, uint64_t pd_addr); @@ -296,9 +297,9 @@ struct amdgpu_gmc { }; #define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type))) -#define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub) \ +#define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub, inst) \ ((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \ - ((adev), (pasid), (type), (allhub))) + ((adev), (pasid), (type), (allhub), (inst))) #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr)) #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid)) #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags)) diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c index ff96f11c2adf..d76f5c8d4977 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c @@ -419,7 +419,7 @@ error_alloc: */ static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid, uint32_t flush_type, - bool all_hub) + bool all_hub, uint32_t inst) { int vmid, i; signed long r; diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c index 3453f1c0e066..4bf807d825c0 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c @@ -324,7 +324,7 @@ static void gmc_v11_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, */ static int gmc_v11_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid, uint32_t flush_type, - bool all_hub) + bool all_hub, uint32_t inst) { int vmid, i; signed long r; diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c index 81609a2b226f..6f53049619cd 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c @@ -424,7 +424,7 @@ static int gmc_v7_0_mc_init(struct amdgpu_device *adev) */ static int gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid, uint32_t flush_type, - bool all_hub) + bool all_hub, uint32_t inst) { int vmid; unsigned int tmp; diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c index d48e33738a88..48475077ca92 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c @@ -622,7 +622,7 @@ static int gmc_v8_0_mc_init(struct amdgpu_device *adev) */ static int gmc_v8_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid, uint32_t flush_type, - bool all_hub) + bool all_hub, uint32_t inst) { int vmid; unsigned int tmp; diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 6da85365e5aa..0163a761ccf0 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -924,7 +924,7 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, */ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid, uint32_t flush_type, - bool all_hub) + bool all_hub, uint32_t inst) { int vmid, i; signed long r; @@ -932,8 +932,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t queried_pasid; bool ret; u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : adev->usec_timeout; - struct amdgpu_ring *ring = &adev->gfx.kiq[0].ring; - struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; + struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst]; if (amdgpu_in_reset(adev)) return -EIO; @@ -953,7 +953,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, if (vega20_xgmi_wa) ndw += kiq->pmf->invalidate_tlbs_size; - spin_lock(&adev->gfx.kiq[0].ring_lock); + spin_lock(&adev->gfx.kiq[inst].ring_lock); /* 2 dwords flush + 8 dwords fence */ amdgpu_ring_alloc(ring, ndw); if (vega20_xgmi_wa) @@ -964,13 +964,13 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); if (r) { amdgpu_ring_undo(ring); - spin_unlock(&adev->gfx.kiq[0].ring_lock); + spin_unlock(&adev->gfx.kiq[inst].ring_lock); up_read(&adev->reset_domain->sem); return -ETIME; } amdgpu_ring_commit(ring); - spin_unlock(&adev->gfx.kiq[0].ring_lock); + spin_unlock(&adev->gfx.kiq[inst].ring_lock); r = amdgpu_fence_wait_polling(ring, seq, usec_timeout); if (r < 1) { dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 888590dfa646..9b1e84d33cdc 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -2052,6 +2052,7 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm); struct kfd_node *dev = pdd->dev; + int xcc = 0; /* * It can be that we race and lose here, but that is extremely unlikely @@ -2069,8 +2070,10 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->adev, pdd->qpd.vmid); } else { - amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->adev, - pdd->process->pasid, type); + for (xcc = 0; xcc < dev->num_xcc_per_node; xcc++) + amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->adev, + pdd->process->pasid, type, + dev->start_xcc_id + xcc); } } -- cgit v1.2.3 From 7fe51e6fd2368b358441a1f6e0c94f4cd7e0720f Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 22:25:25 -0400 Subject: drm/amdkfd: Update context save handling for multi XCC setup (v2) Context save handling needs to be updated for a multi XCC setup: - On a multi XCC setup, KFD needs to report context save base address and size for each XCC in MQD. - Thunk will allocate a large context save area covering all XCCs which will be equal to: num_of_xccs in a partition * size of context save area for 1 XCC. However, it will report only the size of context save area for 1 XCC only in the ioctl call. - Driver then setups the MQD correctly using the size passed from Thunk and information about number of XCCs in a partition. - Update get_wave_state function to return context save area for all XCCs in the partition. v2: update the get_wave_state function for mqd manager v11 (Morris) Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Morris Zhang Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 1 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 1 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 62 +++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 1 + 6 files changed, 67 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 33a9394f9e58..01ea038cab6d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -2095,8 +2095,8 @@ static int get_wave_state(struct device_queue_manager *dqm, * and the queue should be protected against destruction by the process * lock. */ - return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack, - ctl_stack_used_size, save_area_used_size); + return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, &q->properties, + ctl_stack, ctl_stack_used_size, save_area_used_size); } static void get_queue_checkpoint_info(struct device_queue_manager *dqm, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h index eb18be74f559..23158db7da03 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h @@ -97,6 +97,7 @@ struct mqd_manager { uint32_t queue_id); int (*get_wave_state)(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, void __user *ctl_stack, u32 *ctl_stack_used_size, u32 *save_area_used_size); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index d54c6fdebbb6..772c09b5821b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -227,6 +227,7 @@ static uint32_t read_doorbell_id(void *mqd) } static int get_wave_state(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, void __user *ctl_stack, u32 *ctl_stack_used_size, u32 *save_area_used_size) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index 338d639c1898..632344b95d90 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -260,6 +260,7 @@ static uint32_t read_doorbell_id(void *mqd) } static int get_wave_state(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, void __user *ctl_stack, u32 *ctl_stack_used_size, u32 *save_area_used_size) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index b53cd8f9620b..5c9b3392758e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -295,6 +295,7 @@ static uint32_t read_doorbell_id(void *mqd) } static int get_wave_state(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, void __user *ctl_stack, u32 *ctl_stack_used_size, u32 *save_area_used_size) @@ -561,6 +562,7 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, int xcc = 0; struct kfd_mem_obj xcc_mqd_mem_obj; uint64_t xcc_gart_addr = 0; + uint64_t xcc_ctx_save_restore_area_address; uint64_t offset = mm->mqd_stride(mm, q); memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj)); @@ -570,6 +572,23 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q); m->cp_mqd_stride_size = offset; + + /* + * Update the CWSR address for each XCC if CWSR is enabled + * and CWSR area is allocated in thunk + */ + if (mm->dev->kfd->cwsr_enabled && + q->ctx_save_restore_area_address) { + xcc_ctx_save_restore_area_address = + q->ctx_save_restore_area_address + + (xcc * q->ctx_save_restore_area_size); + + m->cp_hqd_ctx_save_base_addr_lo = + lower_32_bits(xcc_ctx_save_restore_area_address); + m->cp_hqd_ctx_save_base_addr_hi = + upper_32_bits(xcc_ctx_save_restore_area_address); + } + if (q->format == KFD_QUEUE_FORMAT_AQL) { m->compute_tg_chunk_size = 1; @@ -689,6 +708,46 @@ static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, return err; } +static int get_wave_state_v9_4_3(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, + void __user *ctl_stack, + u32 *ctl_stack_used_size, + u32 *save_area_used_size) +{ + int xcc, err = 0; + void *xcc_mqd; + void __user *xcc_ctl_stack; + uint64_t mqd_stride_size = mm->mqd_stride(mm, q); + u32 tmp_ctl_stack_used_size = 0, tmp_save_area_used_size = 0; + + for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + xcc_mqd = mqd + mqd_stride_size * xcc; + xcc_ctl_stack = (void __user *)((uintptr_t)ctl_stack + + q->ctx_save_restore_area_size * xcc); + + err = get_wave_state(mm, xcc_mqd, q, xcc_ctl_stack, + &tmp_ctl_stack_used_size, + &tmp_save_area_used_size); + if (err) + break; + + /* + * Set the ctl_stack_used_size and save_area_used_size to + * ctl_stack_used_size and save_area_used_size of XCC 0 when + * passing the info the user-space. + * For multi XCC, user-space would have to look at the header + * info of each Control stack area to determine the control + * stack size and save area used. + */ + if (xcc == 0) { + *ctl_stack_used_size = tmp_ctl_stack_used_size; + *save_area_used_size = tmp_save_area_used_size; + } + } + + return err; +} + #if defined(CONFIG_DEBUG_FS) static int debugfs_show_mqd(struct seq_file *m, void *data) @@ -726,7 +785,6 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, mqd->allocate_mqd = allocate_mqd; mqd->free_mqd = kfd_free_mqd_cp; mqd->is_occupied = kfd_is_occupied_cp; - mqd->get_wave_state = get_wave_state; mqd->get_checkpoint_info = get_checkpoint_info; mqd->checkpoint_mqd = checkpoint_mqd; mqd->restore_mqd = restore_mqd; @@ -740,11 +798,13 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, mqd->load_mqd = load_mqd_v9_4_3; mqd->update_mqd = update_mqd_v9_4_3; mqd->destroy_mqd = destroy_mqd_v9_4_3; + mqd->get_wave_state = get_wave_state_v9_4_3; } else { mqd->init_mqd = init_mqd; mqd->load_mqd = load_mqd; mqd->update_mqd = update_mqd; mqd->destroy_mqd = kfd_destroy_mqd_cp; + mqd->get_wave_state = get_wave_state; } break; case KFD_MQD_TYPE_HIQ: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c index ebf963f42b51..fe69492b1bb3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c @@ -261,6 +261,7 @@ static void update_mqd_tonga(struct mqd_manager *mm, void *mqd, } static int get_wave_state(struct mqd_manager *mm, void *mqd, + struct queue_properties *q, void __user *ctl_stack, u32 *ctl_stack_used_size, u32 *save_area_used_size) -- cgit v1.2.3 From 5e4060123687c4f2c9fb855874f77b14f07526d6 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 22:34:38 -0400 Subject: drm/amdkfd: Call DQM stop during DQM uninitialize During DQM tear down, call DQM stop to unitialize HIQ and associated memory allocated during packet manager init. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 8 ++++++++ 1 file changed, 8 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 01ea038cab6d..f78c1e7aad57 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1348,9 +1348,16 @@ static int start_nocpsch(struct device_queue_manager *dqm) static int stop_nocpsch(struct device_queue_manager *dqm) { + dqm_lock(dqm); + if (!dqm->sched_running) { + dqm_unlock(dqm); + return 0; + } + if (dqm->dev->adev->asic_type == CHIP_HAWAII) pm_uninit(&dqm->packet_mgr, false); dqm->sched_running = false; + dqm_unlock(dqm); return 0; } @@ -2423,6 +2430,7 @@ static void deallocate_hiq_sdma_mqd(struct kfd_node *dev, void device_queue_manager_uninit(struct device_queue_manager *dqm) { + dqm->ops.stop(dqm); dqm->ops.uninitialize(dqm); if (!dqm->dev->kfd->shared_resources.enable_mes) deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd); -- cgit v1.2.3 From f38f147ab3121adbd7510a82e6eb0b41a356c26e Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 22:35:55 -0400 Subject: drm/amdkfd: Update sysfs node properties for multi XCC Update simd_count and array_count node properties to report values multiplied by number of XCCs in the partition. Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 94af37df3ed2..6d958bf0fe90 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -468,7 +468,8 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count", dev->node_props.cpu_cores_count); sysfs_show_32bit_prop(buffer, offs, "simd_count", - dev->gpu ? dev->node_props.simd_count : 0); + dev->gpu ? (dev->node_props.simd_count * + dev->gpu->num_xcc_per_node) : 0); sysfs_show_32bit_prop(buffer, offs, "mem_banks_count", dev->node_props.mem_banks_count); sysfs_show_32bit_prop(buffer, offs, "caches_count", @@ -492,7 +493,8 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, sysfs_show_32bit_prop(buffer, offs, "wave_front_size", dev->node_props.wave_front_size); sysfs_show_32bit_prop(buffer, offs, "array_count", - dev->node_props.array_count); + dev->gpu ? (dev->node_props.array_count * + dev->gpu->num_xcc_per_node) : 0); sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine", dev->node_props.simd_arrays_per_engine); sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array", -- cgit v1.2.3 From a805889a15315f7fa78c1c4bb2f1875c7c43f919 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 9 May 2022 22:52:39 -0400 Subject: drm/amdkfd: Update SDMA queue management for GFX9.4.3 This patch updates SDMA queue management for multi XCC in GFX9.4.3. - Allocate/deallocate SDMA queues from the correct SDMA engines based on the partition mode. - Updates the kgd2kfd interface to fetch the correct SDMA register addresses. - It also fixes dumping correct SDMA queue info in debugfs. v2: squash in fix "drm/amdkfd: Fix XGMI SDMA user-mode queue allocation" Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c | 194 ++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 8 +- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 59 +++---- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 3 + 5 files changed, 227 insertions(+), 41 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c index 49d8087e469e..e81bdca53f42 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c @@ -31,6 +31,192 @@ #include "oss/osssys_4_0_sh_mask.h" #include "v9_structs.h" #include "soc15.h" +#include "sdma/sdma_4_4_2_offset.h" +#include "sdma/sdma_4_4_2_sh_mask.h" + +static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd) +{ + return (struct v9_sdma_mqd *)mqd; +} + +static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev, + unsigned int engine_id, + unsigned int queue_id) +{ + uint32_t sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, engine_id, + regSDMA_RLC0_RB_CNTL) - + regSDMA_RLC0_RB_CNTL; + uint32_t retval = sdma_engine_reg_base + + queue_id * (regSDMA_RLC1_RB_CNTL - regSDMA_RLC0_RB_CNTL); + + pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id, + queue_id, retval); + return retval; +} + +int kgd_gfx_v9_4_3_hqd_sdma_load(struct amdgpu_device *adev, void *mqd, + uint32_t __user *wptr, struct mm_struct *mm) +{ + struct v9_sdma_mqd *m; + uint32_t sdma_rlc_reg_offset; + unsigned long end_jiffies; + uint32_t data; + uint64_t data64; + uint64_t __user *wptr64 = (uint64_t __user *)wptr; + + m = get_sdma_mqd(mqd); + sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id, + m->sdma_queue_id); + + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL, + m->sdmax_rlcx_rb_cntl & (~SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK)); + + end_jiffies = msecs_to_jiffies(2000) + jiffies; + while (true) { + data = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_CONTEXT_STATUS); + if (data & SDMA_RLC0_CONTEXT_STATUS__IDLE_MASK) + break; + if (time_after(jiffies, end_jiffies)) { + pr_err("SDMA RLC not idle in %s\n", __func__); + return -ETIME; + } + usleep_range(500, 1000); + } + + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_DOORBELL_OFFSET, + m->sdmax_rlcx_doorbell_offset); + + data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA_RLC0_DOORBELL, + ENABLE, 1); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_DOORBELL, data); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR, + m->sdmax_rlcx_rb_rptr); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_HI, + m->sdmax_rlcx_rb_rptr_hi); + + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_MINOR_PTR_UPDATE, 1); + if (read_user_wptr(mm, wptr64, data64)) { + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR, + lower_32_bits(data64)); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR_HI, + upper_32_bits(data64)); + } else { + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR, + m->sdmax_rlcx_rb_rptr); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_WPTR_HI, + m->sdmax_rlcx_rb_rptr_hi); + } + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_MINOR_PTR_UPDATE, 0); + + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_BASE, m->sdmax_rlcx_rb_base); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_BASE_HI, + m->sdmax_rlcx_rb_base_hi); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_ADDR_LO, + m->sdmax_rlcx_rb_rptr_addr_lo); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_ADDR_HI, + m->sdmax_rlcx_rb_rptr_addr_hi); + + data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA_RLC0_RB_CNTL, + RB_ENABLE, 1); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL, data); + + return 0; +} + +int kgd_gfx_v9_4_3_hqd_sdma_dump(struct amdgpu_device *adev, + uint32_t engine_id, uint32_t queue_id, + uint32_t (**dump)[2], uint32_t *n_regs) +{ + uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, + engine_id, queue_id); + uint32_t i = 0, reg; +#undef HQD_N_REGS +#define HQD_N_REGS (19+6+7+12) +#define DUMP_REG(addr) do { \ + if (WARN_ON_ONCE(i >= HQD_N_REGS)) \ + break; \ + (*dump)[i][0] = (addr) << 2; \ + (*dump)[i++][1] = RREG32(addr); \ + } while (0) + + *dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL); + if (*dump == NULL) + return -ENOMEM; + + for (reg = regSDMA_RLC0_RB_CNTL; reg <= regSDMA_RLC0_DOORBELL; reg++) + DUMP_REG(sdma_rlc_reg_offset + reg); + for (reg = regSDMA_RLC0_STATUS; reg <= regSDMA_RLC0_CSA_ADDR_HI; reg++) + DUMP_REG(sdma_rlc_reg_offset + reg); + for (reg = regSDMA_RLC0_IB_SUB_REMAIN; + reg <= regSDMA_RLC0_MINOR_PTR_UPDATE; reg++) + DUMP_REG(sdma_rlc_reg_offset + reg); + for (reg = regSDMA_RLC0_MIDCMD_DATA0; + reg <= regSDMA_RLC0_MIDCMD_CNTL; reg++) + DUMP_REG(sdma_rlc_reg_offset + reg); + + WARN_ON_ONCE(i != HQD_N_REGS); + *n_regs = i; + + return 0; +} + +bool kgd_gfx_v9_4_3_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd) +{ + struct v9_sdma_mqd *m; + uint32_t sdma_rlc_reg_offset; + uint32_t sdma_rlc_rb_cntl; + + m = get_sdma_mqd(mqd); + sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id, + m->sdma_queue_id); + + sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL); + + if (sdma_rlc_rb_cntl & SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK) + return true; + + return false; +} + +int kgd_gfx_v9_4_3_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd, + unsigned int utimeout) +{ + struct v9_sdma_mqd *m; + uint32_t sdma_rlc_reg_offset; + uint32_t temp; + unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies; + + m = get_sdma_mqd(mqd); + sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id, + m->sdma_queue_id); + + temp = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL); + temp = temp & ~SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK; + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL, temp); + + while (true) { + temp = RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_CONTEXT_STATUS); + if (temp & SDMA_RLC0_CONTEXT_STATUS__IDLE_MASK) + break; + if (time_after(jiffies, end_jiffies)) { + pr_err("SDMA RLC not idle in %s\n", __func__); + return -ETIME; + } + usleep_range(500, 1000); + } + + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_DOORBELL, 0); + WREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL, + RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_CNTL) | + SDMA_RLC0_RB_CNTL__RB_ENABLE_MASK); + + m->sdmax_rlcx_rb_rptr = + RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR); + m->sdmax_rlcx_rb_rptr_hi = + RREG32(sdma_rlc_reg_offset + regSDMA_RLC0_RB_RPTR_HI); + + return 0; +} static int kgd_gfx_v9_4_3_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid, unsigned int vmid, uint32_t inst) @@ -166,13 +352,13 @@ const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = { .init_interrupts = kgd_gfx_v9_init_interrupts, .hqd_load = kgd_gfx_v9_4_3_hqd_load, .hiq_mqd_load = kgd_gfx_v9_hiq_mqd_load, - .hqd_sdma_load = kgd_arcturus_hqd_sdma_load, + .hqd_sdma_load = kgd_gfx_v9_4_3_hqd_sdma_load, .hqd_dump = kgd_gfx_v9_hqd_dump, - .hqd_sdma_dump = kgd_arcturus_hqd_sdma_dump, + .hqd_sdma_dump = kgd_gfx_v9_4_3_hqd_sdma_dump, .hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied, - .hqd_sdma_is_occupied = kgd_arcturus_hqd_sdma_is_occupied, + .hqd_sdma_is_occupied = kgd_gfx_v9_4_3_hqd_sdma_is_occupied, .hqd_destroy = kgd_gfx_v9_hqd_destroy, - .hqd_sdma_destroy = kgd_arcturus_hqd_sdma_destroy, + .hqd_sdma_destroy = kgd_gfx_v9_4_3_hqd_sdma_destroy, .wave_control_execute = kgd_gfx_v9_wave_control_execute, .get_atc_vmid_pasid_mapping_info = kgd_gfx_v9_get_atc_vmid_pasid_mapping_info, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 37c6dc5c37bf..ec5f85ff34e5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -741,6 +741,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (!node) goto node_alloc_error; + node->node_id = i; node->adev = kfd->adev; node->kfd = kfd; node->kfd2kgd = kfd->kfd2kgd; @@ -1323,15 +1324,16 @@ unsigned int kfd_get_num_sdma_engines(struct kfd_node *node) { /* If XGMI is not supported, all SDMA engines are PCIe */ if (!node->adev->gmc.xgmi.supported) - return node->adev->sdma.num_instances; + return node->adev->sdma.num_instances/(int)node->kfd->num_nodes; - return min(node->adev->sdma.num_instances, 2); + return min(node->adev->sdma.num_instances/(int)node->kfd->num_nodes, 2); } unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *node) { /* After reserved for PCIe, the rest of engines are XGMI */ - return node->adev->sdma.num_instances - kfd_get_num_sdma_engines(node); + return node->adev->sdma.num_instances/(int)node->kfd->num_nodes - + kfd_get_num_sdma_engines(node); } #if defined(CONFIG_DEBUG_FS) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index f78c1e7aad57..69419a53a14e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -124,6 +124,15 @@ static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manag return dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap; } +static void init_sdma_bitmaps(struct device_queue_manager *dqm) +{ + bitmap_zero(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES); + bitmap_set(dqm->sdma_bitmap, 0, get_num_sdma_queues(dqm)); + + bitmap_zero(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES); + bitmap_set(dqm->xgmi_sdma_bitmap, 0, get_num_xgmi_sdma_queues(dqm)); +} + void program_sh_mem_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { @@ -1268,24 +1277,6 @@ static void init_interrupts(struct device_queue_manager *dqm) } } -static void init_sdma_bitmaps(struct device_queue_manager *dqm) -{ - unsigned int num_sdma_queues = - min_t(unsigned int, sizeof(dqm->sdma_bitmap)*8, - get_num_sdma_queues(dqm)); - unsigned int num_xgmi_sdma_queues = - min_t(unsigned int, sizeof(dqm->xgmi_sdma_bitmap)*8, - get_num_xgmi_sdma_queues(dqm)); - - if (num_sdma_queues) - dqm->sdma_bitmap = GENMASK_ULL(num_sdma_queues-1, 0); - if (num_xgmi_sdma_queues) - dqm->xgmi_sdma_bitmap = GENMASK_ULL(num_xgmi_sdma_queues-1, 0); - - dqm->sdma_bitmap &= ~get_reserved_sdma_queues_bitmap(dqm); - pr_info("sdma_bitmap: %llx\n", dqm->sdma_bitmap); -} - static int initialize_nocpsch(struct device_queue_manager *dqm) { int pipe, queue; @@ -1375,46 +1366,49 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm, int bit; if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { - if (dqm->sdma_bitmap == 0) { + if (bitmap_empty(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES)) { pr_err("No more SDMA queue to allocate\n"); return -ENOMEM; } if (restore_sdma_id) { /* Re-use existing sdma_id */ - if (!(dqm->sdma_bitmap & (1ULL << *restore_sdma_id))) { + if (!test_bit(*restore_sdma_id, dqm->sdma_bitmap)) { pr_err("SDMA queue already in use\n"); return -EBUSY; } - dqm->sdma_bitmap &= ~(1ULL << *restore_sdma_id); + clear_bit(*restore_sdma_id, dqm->sdma_bitmap); q->sdma_id = *restore_sdma_id; } else { /* Find first available sdma_id */ - bit = __ffs64(dqm->sdma_bitmap); - dqm->sdma_bitmap &= ~(1ULL << bit); + bit = find_first_bit(dqm->sdma_bitmap, + get_num_sdma_queues(dqm)); + clear_bit(bit, dqm->sdma_bitmap); q->sdma_id = bit; } - q->properties.sdma_engine_id = q->sdma_id % - kfd_get_num_sdma_engines(dqm->dev); + q->properties.sdma_engine_id = + dqm->dev->node_id * get_num_all_sdma_engines(dqm) + + q->sdma_id % kfd_get_num_sdma_engines(dqm->dev); q->properties.sdma_queue_id = q->sdma_id / kfd_get_num_sdma_engines(dqm->dev); } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { - if (dqm->xgmi_sdma_bitmap == 0) { + if (bitmap_empty(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES)) { pr_err("No more XGMI SDMA queue to allocate\n"); return -ENOMEM; } if (restore_sdma_id) { /* Re-use existing sdma_id */ - if (!(dqm->xgmi_sdma_bitmap & (1ULL << *restore_sdma_id))) { + if (!test_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap)) { pr_err("SDMA queue already in use\n"); return -EBUSY; } - dqm->xgmi_sdma_bitmap &= ~(1ULL << *restore_sdma_id); + clear_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap); q->sdma_id = *restore_sdma_id; } else { - bit = __ffs64(dqm->xgmi_sdma_bitmap); - dqm->xgmi_sdma_bitmap &= ~(1ULL << bit); + bit = find_first_bit(dqm->xgmi_sdma_bitmap, + get_num_xgmi_sdma_queues(dqm)); + clear_bit(bit, dqm->xgmi_sdma_bitmap); q->sdma_id = bit; } /* sdma_engine_id is sdma id including @@ -1424,6 +1418,7 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm, * PCIe-optimized ones */ q->properties.sdma_engine_id = + dqm->dev->node_id * get_num_all_sdma_engines(dqm) + kfd_get_num_sdma_engines(dqm->dev) + q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev); q->properties.sdma_queue_id = q->sdma_id / @@ -1442,11 +1437,11 @@ static void deallocate_sdma_queue(struct device_queue_manager *dqm, if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { if (q->sdma_id >= get_num_sdma_queues(dqm)) return; - dqm->sdma_bitmap |= (1ULL << q->sdma_id); + set_bit(q->sdma_id, dqm->sdma_bitmap); } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm)) return; - dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id); + set_bit(q->sdma_id, dqm->xgmi_sdma_bitmap); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index e554a48f3054..b11c474d4067 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -239,8 +239,8 @@ struct device_queue_manager { unsigned int total_queue_count; unsigned int next_pipe_to_allocate; unsigned int *allocated_queues; - uint64_t sdma_bitmap; - uint64_t xgmi_sdma_bitmap; + DECLARE_BITMAP(sdma_bitmap, KFD_MAX_SDMA_QUEUES); + DECLARE_BITMAP(xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES); /* the pasid mapping for each kfd vmid */ uint16_t vmid_pasid[VMID_NUM]; uint64_t pipelines_addr; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 1337fcdf8958..5cfebcc8b305 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -113,6 +113,8 @@ #define KFD_UNMAP_LATENCY_MS (4000) +#define KFD_MAX_SDMA_QUEUES 128 + /* * 512 = 0x200 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the @@ -260,6 +262,7 @@ struct kfd_vmid_info { struct kfd_dev; struct kfd_node { + unsigned int node_id; struct amdgpu_device *adev; /* Duplicated here along with keeping * a copy in kfd_dev to save a hop */ -- cgit v1.2.3 From ef75a6ef37235e211bbdb17c25e5f79c55df1750 Mon Sep 17 00:00:00 2001 From: Rajneesh Bhardwaj Date: Thu, 3 Mar 2022 10:56:05 -0500 Subject: drm/amdkfd: Update coherence settings for svm ranges Recently introduced commit "drm/amdgpu: Set cache coherency for GC 9.4.3" did not update the settings applicable for svm ranges. Add the coherence settings for svm ranges for GFX IP 9.4.3. Reviewed-by: Amber Lin Signed-off-by: Rajneesh Bhardwaj Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 17 +++++++++++++++++ include/uapi/linux/kfd_ioctl.h | 2 ++ 2 files changed, 19 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 96ccff79902c..4b4f3bf8b823 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1159,6 +1159,7 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, uint64_t pte_flags; bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT; + bool uncached = flags & KFD_IOCTL_SVM_FLAG_UNCACHED; if (domain == SVM_RANGE_VRAM_DOMAIN) bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); @@ -1198,6 +1199,22 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; } break; + case IP_VERSION(9, 4, 3): + //TODO: Need more work for handling multiple memory partitions + //e.g. NPS4. Current approch is only applicable without memory + //partitions. + snoop = true; + if (uncached) + mapping_flags |= AMDGPU_VM_MTYPE_UC; + /* local HBM region close to partition*/ + else if (bo_adev == adev) + mapping_flags |= AMDGPU_VM_MTYPE_RW; + /* local HBM region far from partition or remote XGMI GPU or + * system memory + */ + else + mapping_flags |= AMDGPU_VM_MTYPE_NC; + break; default: mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h index 2da5c3ad71bd..2a9671e1ddb5 100644 --- a/include/uapi/linux/kfd_ioctl.h +++ b/include/uapi/linux/kfd_ioctl.h @@ -623,6 +623,8 @@ enum kfd_mmio_remap { #define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020 /* Keep GPU memory mapping always valid as if XNACK is disable */ #define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040 +/* Uncached access to memory */ +#define KFD_IOCTL_SVM_FLAG_UNCACHED 0x00000080 /** * kfd_ioctl_svm_op - SVM ioctl operations -- cgit v1.2.3 From 92085240ef9c0ec60c27a60b3cc0d4f5266fa511 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Tue, 3 May 2022 10:16:46 -0400 Subject: drm/amdkfd: add gpu compute cores io links for gfx9.4.3 The PSP TA will only provide xGMI topology info for links between GPU sockets so links between partitions from different sockets will be hardcoded as 3 xGMI hops with 1 hops weighted as xGMI and 2 hops weighted with a new intra-socket weight to indicate the longest possible distance. If the link between a partition and the CPU is non-PCIe, then assume the CPU (CCDs) is located within the same socket as the partition and represent the link as an intra-socket weighted single hop XGMI link with memory bandwidth. Links between partitions within a single socket will be abstracted as single hop xGMI links weighted with the new intra-socket weight and will have memory bandwidth. Finally, use the unused function bits in the location ID to represent the coordinates of the compute partition within its socket. A follow on patch will resolve the requirement for GPU socket xGMI link representation sometime later. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 49 +++++++++++++++++++++---------- drivers/gpu/drm/amd/amdkfd/kfd_crat.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 8 +++++ drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 4 +++ 4 files changed, 47 insertions(+), 16 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index f5aebba31e88..dc93a67257e1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -1166,7 +1166,7 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink, if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) props->weight = 20; else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI) - props->weight = 15 * iolink->num_hops_xgmi; + props->weight = iolink->weight_xgmi; else props->weight = node_distance(id_from, id_to); @@ -1972,6 +1972,9 @@ static void kfd_find_numa_node_in_srat(struct kfd_node *kdev) } #endif +#define KFD_CRAT_INTRA_SOCKET_WEIGHT 13 +#define KFD_CRAT_XGMI_WEIGHT 15 + /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU * to its NUMA node * @avail_size: Available size in the memory @@ -2003,6 +2006,12 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, * TODO: Fill-in other fields of iolink subtype */ if (kdev->adev->gmc.xgmi.connected_to_cpu) { + bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3); + int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT : + KFD_CRAT_INTRA_SOCKET_WEIGHT; + uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes( + kdev->adev, NULL, true) : mem_bw; + /* * with host gpu xgmi link, host can access gpu memory whether * or not pcie bar type is large, so always create bidirectional @@ -2010,14 +2019,9 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, */ sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL; sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; - sub_type_hdr->num_hops_xgmi = 1; - if (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 2)) { - sub_type_hdr->minimum_bandwidth_mbs = - amdgpu_amdkfd_get_xgmi_bandwidth_mbytes( - kdev->adev, NULL, true); - sub_type_hdr->maximum_bandwidth_mbs = - sub_type_hdr->minimum_bandwidth_mbs; - } + sub_type_hdr->weight_xgmi = weight; + sub_type_hdr->minimum_bandwidth_mbs = bandwidth; + sub_type_hdr->maximum_bandwidth_mbs = bandwidth; } else { sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS; sub_type_hdr->minimum_bandwidth_mbs = @@ -2050,6 +2054,8 @@ static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, uint32_t proximity_domain_from, uint32_t proximity_domain_to) { + bool use_ta_info = kdev->kfd->num_nodes == 1; + *avail_size -= sizeof(struct crat_subtype_iolink); if (*avail_size < 0) return -ENOMEM; @@ -2064,12 +2070,25 @@ static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size, sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI; sub_type_hdr->proximity_domain_from = proximity_domain_from; sub_type_hdr->proximity_domain_to = proximity_domain_to; - sub_type_hdr->num_hops_xgmi = - amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev); - sub_type_hdr->maximum_bandwidth_mbs = - amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, peer_kdev->adev, false); - sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ? - amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0; + + if (use_ta_info) { + sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT * + amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev); + sub_type_hdr->maximum_bandwidth_mbs = + amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, + peer_kdev->adev, false); + sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ? + amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0; + } else { + bool is_single_hop = kdev->kfd == peer_kdev->kfd; + int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT : + (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT; + int mem_bw = 819200; + + sub_type_hdr->weight_xgmi = weight; + sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0; + sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0; + } return 0; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.h b/drivers/gpu/drm/amd/amdkfd/kfd_crat.h index 3d0e533b93b9..fc719389b5d6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.h @@ -275,7 +275,7 @@ struct crat_subtype_iolink { uint32_t maximum_bandwidth_mbs; uint32_t recommended_transfer_size; uint8_t reserved2[CRAT_IOLINK_RESERVED_LENGTH - 1]; - uint8_t num_hops_xgmi; + uint8_t weight_xgmi; }; /* diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index ec5f85ff34e5..971a3aa3294a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -702,6 +702,14 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (amdgpu_use_xgmi_p2p) kfd->hive_id = kfd->adev->gmc.xgmi.hive_id; + /* + * For GFX9.4.3, the KFD abstracts all partitions within a socket as + * xGMI connected in the topology so assign a unique hive id per + * device based on the pci device location if device is in PCIe mode. + */ + if (!kfd->hive_id && (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3)) && kfd->num_nodes > 1) + kfd->hive_id = pci_dev_id(kfd->adev->pdev); + kfd->noretry = kfd->adev->gmc.noretry; /* If CRAT is broken, won't set iommu enabled */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 6d958bf0fe90..d3e70341dfad 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1926,7 +1926,11 @@ int kfd_topology_add_device(struct kfd_node *gpu) dev->node_props.capability |= ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) & HSA_CAP_ASIC_REVISION_MASK); + dev->node_props.location_id = pci_dev_id(gpu->adev->pdev); + if (KFD_GC_VERSION(dev->gpu->kfd) == IP_VERSION(9, 4, 3)) + dev->node_props.location_id |= dev->gpu->node_id; + dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus); dev->node_props.max_engine_clk_fcompute = amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev); -- cgit v1.2.3 From 643e40d4c06f8c887af1789c7bf8d279e9c8e4cf Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 31 May 2022 16:25:16 -0400 Subject: drm/amdkfd: Fix SDMA in CPX mode When creating a user-mode SDMA queue, CP FW expects driver to use/set virtual SDMA engine id in MAP_QUEUES packet instead of using the physical SDMA engine id. Each partition node's virtual SDMA number should start from 0. However, when allocating doorbell for the queue, KFD needs to allocate the doorbell from doorbell space corresponding to the physical SDMA engine id, otherwise the hwardware will not see the doorbell press. Signed-off-by: Mukul Joshi Reviewed-by: Amber Lin Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 69419a53a14e..2b5c4b2dd242 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -363,7 +363,16 @@ static int allocate_doorbell(struct qcm_process_device *qpd, */ uint32_t *idx_offset = dev->kfd->shared_resources.sdma_doorbell_idx; - uint32_t valid_id = idx_offset[q->properties.sdma_engine_id] + + /* + * q->properties.sdma_engine_id corresponds to the virtual + * sdma engine number. However, for doorbell allocation, + * we need the physical sdma engine id in order to get the + * correct doorbell offset. + */ + uint32_t valid_id = idx_offset[qpd->dqm->dev->node_id * + get_num_all_sdma_engines(qpd->dqm) + + q->properties.sdma_engine_id] + (q->properties.sdma_queue_id & 1) * KFD_QUEUE_DOORBELL_MIRROR_OFFSET + (q->properties.sdma_queue_id >> 1); @@ -1388,7 +1397,6 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm, } q->properties.sdma_engine_id = - dqm->dev->node_id * get_num_all_sdma_engines(dqm) + q->sdma_id % kfd_get_num_sdma_engines(dqm->dev); q->properties.sdma_queue_id = q->sdma_id / kfd_get_num_sdma_engines(dqm->dev); @@ -1418,7 +1426,6 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm, * PCIe-optimized ones */ q->properties.sdma_engine_id = - dqm->dev->node_id * get_num_all_sdma_engines(dqm) + kfd_get_num_sdma_engines(dqm->dev) + q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev); q->properties.sdma_queue_id = q->sdma_id / @@ -2486,6 +2493,7 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) int pipe, queue; int r = 0, xcc; uint32_t inst; + uint32_t sdma_engine_start; if (!dqm->sched_running) { seq_puts(m, " Device is stopped\n"); @@ -2530,7 +2538,10 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) } } - for (pipe = 0; pipe < get_num_all_sdma_engines(dqm); pipe++) { + sdma_engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm); + for (pipe = sdma_engine_start; + pipe < (sdma_engine_start + get_num_all_sdma_engines(dqm)); + pipe++) { for (queue = 0; queue < dqm->dev->kfd->device_info.num_sdma_queues_per_engine; queue++) { -- cgit v1.2.3 From 1bd6dd21fcd53ac78a9018b96699ef1aa99a3e59 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 31 May 2022 16:31:28 -0400 Subject: drm/amdkfd: Add SDMA info for SDMA 4.4.2 Update SDMA queue information for SDMA 4.4.2. Signed-off-by: Mukul Joshi Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 1 + 1 file changed, 1 insertion(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 971a3aa3294a..8e5d785b8824 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -82,6 +82,7 @@ static void kfd_device_info_set_sdma_info(struct kfd_dev *kfd) case IP_VERSION(4, 2, 0):/* VEGA20 */ case IP_VERSION(4, 2, 2):/* ARCTURUS */ case IP_VERSION(4, 4, 0):/* ALDEBARAN */ + case IP_VERSION(4, 4, 2): case IP_VERSION(5, 0, 0):/* NAVI10 */ case IP_VERSION(5, 0, 1):/* CYAN_SKILLFISH */ case IP_VERSION(5, 0, 2):/* NAVI14 */ -- cgit v1.2.3 From d1d22df174ae512c57374f517e346b608f61555c Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 7 Jun 2022 14:46:18 -0400 Subject: drm/amdkfd: Populate memory info before adding GPU node to topology The local memory info needs to be fetched before the GPU node is added to topology. Without this, the sysfs is incorrectly populated and the size is reported as 0. This was causing rocr tests to fail. This issue was caused because of a bad merge. Signed-off-by: Mukul Joshi Reviewed-by: Amber Lin Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 8e5d785b8824..829e32433faf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -743,6 +743,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, node->max_proc_per_quantum = max_proc_per_quantum; atomic_set(&node->sram_ecc_flag, 0); + amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); + dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n", kfd->num_nodes); for (i = 0; i < kfd->num_nodes; i++) { @@ -793,8 +795,6 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (kfd_resume_iommu(kfd)) goto kfd_resume_iommu_error; - amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); - kfd->init_complete = true; dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor, kfd->adev->pdev->device); -- cgit v1.2.3 From 8078f1c610fdcdd8003e2c538fb04af41fa5c269 Mon Sep 17 00:00:00 2001 From: Lijo Lazar Date: Wed, 29 Jun 2022 11:41:53 +0530 Subject: drm/amdgpu: Change num_xcd to xcc_mask Instead of number of XCCs, keep a mask of XCCs for the exact XCCs available on the ASIC. XCC configuration could differ based on different ASIC configs. v2: Rename num_xcd to num_xcc (Hawking) Use smaller xcc_mask size, changed to u16 (Le) Signed-off-by: Lijo Lazar Reviewed-by: Hawking Zhang Reviewed-by: Le Ma Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 21 ++--- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h | 4 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 2 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 133 ++++++++++++++++++------------- drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c | 67 ++++++++++------ drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 10 +-- 7 files changed, 141 insertions(+), 99 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 76438f197de1..069b259f384c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -209,12 +209,12 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_queue_per_pipe, adev->gfx.num_compute_rings); - int num_xcd = (adev->gfx.num_xcd > 1) ? adev->gfx.num_xcd : 1; + int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1; if (multipipe_policy) { /* policy: make queues evenly cross all pipes on MEC1 only * for multiple xcc, just use the original policy for simplicity */ - for (j = 0; j < num_xcd; j++) { + for (j = 0; j < num_xcc; j++) { for (i = 0; i < max_queues_per_mec; i++) { pipe = i % adev->gfx.mec.num_pipe_per_mec; queue = (i / adev->gfx.mec.num_pipe_per_mec) % @@ -226,13 +226,13 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) } } else { /* policy: amdgpu owns all queues in the given pipe */ - for (j = 0; j < num_xcd; j++) { + for (j = 0; j < num_xcc; j++) { for (i = 0; i < max_queues_per_mec; ++i) set_bit(i, adev->gfx.mec_bitmap[j].queue_bitmap); } } - for (j = 0; j < num_xcd; j++) { + for (j = 0; j < num_xcc; j++) { dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec_bitmap[j].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES)); } @@ -1207,23 +1207,24 @@ static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev, struct drm_device *ddev = dev_get_drvdata(dev); struct amdgpu_device *adev = drm_to_adev(ddev); enum amdgpu_gfx_partition mode; - int ret; + int ret = 0, num_xcc; - if (adev->gfx.num_xcd % 2 != 0) + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + if (num_xcc % 2 != 0) return -EINVAL; if (!strncasecmp("SPX", buf, strlen("SPX"))) { mode = AMDGPU_SPX_PARTITION_MODE; } else if (!strncasecmp("DPX", buf, strlen("DPX"))) { - if (adev->gfx.num_xcd != 4 || adev->gfx.num_xcd != 8) + if (num_xcc != 4 || num_xcc != 8) return -EINVAL; mode = AMDGPU_DPX_PARTITION_MODE; } else if (!strncasecmp("TPX", buf, strlen("TPX"))) { - if (adev->gfx.num_xcd != 6) + if (num_xcc != 6) return -EINVAL; mode = AMDGPU_TPX_PARTITION_MODE; } else if (!strncasecmp("QPX", buf, strlen("QPX"))) { - if (adev->gfx.num_xcd != 8) + if (num_xcc != 8) return -EINVAL; mode = AMDGPU_QPX_PARTITION_MODE; } else if (!strncasecmp("CPX", buf, strlen("CPX"))) { @@ -1253,7 +1254,7 @@ static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev, char *supported_partition; /* TBD */ - switch (adev->gfx.num_xcd) { + switch (NUM_XCC(adev->gfx.xcc_mask)) { case 8: supported_partition = "SPX, DPX, QPX, CPX"; break; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h index 8df36527aee9..93f9875154db 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h @@ -64,6 +64,8 @@ enum amdgpu_gfx_partition { AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE, }; +#define NUM_XCC(x) hweight16(x) + struct amdgpu_mec { struct amdgpu_bo *hpd_eop_obj; u64 hpd_eop_gpu_addr; @@ -396,7 +398,7 @@ struct amdgpu_gfx { bool cp_gfx_shadow; /* for gfx11 */ enum amdgpu_gfx_partition partition_mode; - uint32_t num_xcd; + uint16_t xcc_mask; uint32_t num_xcc_per_xcp; struct mutex partition_mutex; }; diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c index 91814dc083c9..da69177dc76f 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c @@ -4536,7 +4536,7 @@ static int gfx_v9_0_early_init(void *handle) adev->gfx.num_gfx_rings = 0; else adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS; - adev->gfx.num_xcd = 1; + adev->gfx.xcc_mask = 1; adev->gfx.num_compute_rings = min(amdgpu_gfx_get_num_kcq(adev), AMDGPU_MAX_COMPUTE_RINGS); gfx_v9_0_set_kiq_pm4_funcs(adev); diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c index 52185b1d5d31..c776fc5884de 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c @@ -177,16 +177,19 @@ static const struct kiq_pm4_funcs gfx_v9_4_3_kiq_pm4_funcs = { static void gfx_v9_4_3_set_kiq_pm4_funcs(struct amdgpu_device *adev) { - int i; - for (i = 0; i < adev->gfx.num_xcd; i++) + int i, num_xcc; + + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) adev->gfx.kiq[i].pmf = &gfx_v9_4_3_kiq_pm4_funcs; } static void gfx_v9_4_3_init_golden_registers(struct amdgpu_device *adev) { - int i; + int i, num_xcc; - for (i = 2; i < adev->gfx.num_xcd; i++) + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 2; i < num_xcc; i++) WREG32_SOC15(GC, i, regGRBM_MCM_ADDR, 0x4); } @@ -499,7 +502,7 @@ static void gfx_v9_4_3_mec_fini(struct amdgpu_device *adev) static int gfx_v9_4_3_mec_init(struct amdgpu_device *adev) { - int r, i; + int r, i, num_xcc; u32 *hpd; const __le32 *fw_data; unsigned fw_size; @@ -508,7 +511,8 @@ static int gfx_v9_4_3_mec_init(struct amdgpu_device *adev) const struct gfx_firmware_header_v1_0 *mec_hdr; - for (i = 0; i < adev->gfx.num_xcd; i++) + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) bitmap_zero(adev->gfx.mec_bitmap[i].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); @@ -683,23 +687,24 @@ static int gfx_v9_4_3_switch_compute_partition(struct amdgpu_device *adev, enum amdgpu_gfx_partition mode) { u32 tmp = 0; - int num_xcc_per_partition, i; + int num_xcc_per_partition, i, num_xcc; if (mode == adev->gfx.partition_mode) return mode; + num_xcc = NUM_XCC(adev->gfx.xcc_mask); switch (mode) { case AMDGPU_SPX_PARTITION_MODE: - num_xcc_per_partition = adev->gfx.num_xcd; + num_xcc_per_partition = num_xcc; break; case AMDGPU_DPX_PARTITION_MODE: - num_xcc_per_partition = adev->gfx.num_xcd / 2; + num_xcc_per_partition = num_xcc / 2; break; case AMDGPU_TPX_PARTITION_MODE: - num_xcc_per_partition = adev->gfx.num_xcd / 3; + num_xcc_per_partition = num_xcc / 3; break; case AMDGPU_QPX_PARTITION_MODE: - num_xcc_per_partition = adev->gfx.num_xcd / 4; + num_xcc_per_partition = num_xcc / 4; break; case AMDGPU_CPX_PARTITION_MODE: num_xcc_per_partition = 1; @@ -712,7 +717,7 @@ static int gfx_v9_4_3_switch_compute_partition(struct amdgpu_device *adev, * Stop user queues and threads, and make sure GPU is empty of work. */ - for (i = 0; i < adev->gfx.num_xcd; i++) { + for (i = 0; i < num_xcc; i++) { tmp = REG_SET_FIELD(tmp, CP_HYP_XCP_CTL, NUM_XCC_IN_XCP, num_xcc_per_partition); tmp = REG_SET_FIELD(tmp, CP_HYP_XCP_CTL, VIRTUAL_XCC_ID, @@ -836,7 +841,7 @@ static int gfx_v9_4_3_compute_ring_init(struct amdgpu_device *adev, int ring_id, static int gfx_v9_4_3_sw_init(void *handle) { - int i, j, k, r, ring_id, xcc_id; + int i, j, k, r, ring_id, xcc_id, num_xcc; struct amdgpu_kiq *kiq; struct amdgpu_device *adev = (struct amdgpu_device *)handle; @@ -844,6 +849,8 @@ static int gfx_v9_4_3_sw_init(void *handle) adev->gfx.mec.num_pipe_per_mec = 4; adev->gfx.mec.num_queue_per_pipe = 8; + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + /* EOP Event */ r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_GRBM_CP, GFX_9_0__SRCID__CP_EOP_INTERRUPT, &adev->gfx.eop_irq); if (r) @@ -877,8 +884,7 @@ static int gfx_v9_4_3_sw_init(void *handle) /* set up the compute queues - allocate horizontally across pipes */ ring_id = 0; - for (xcc_id = 0; xcc_id < adev->gfx.num_xcd; xcc_id++) { - + for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) { for (i = 0; i < adev->gfx.mec.num_mec; ++i) { for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) { for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; @@ -930,14 +936,14 @@ static int gfx_v9_4_3_sw_init(void *handle) static int gfx_v9_4_3_sw_fini(void *handle) { - int i; + int i, num_xcc; struct amdgpu_device *adev = (struct amdgpu_device *)handle; - for (i = 0; i < adev->gfx.num_compute_rings * - adev->gfx.num_xcd; i++) + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < adev->gfx.num_compute_rings * num_xcc; i++) amdgpu_ring_fini(&adev->gfx.compute_ring[i]); - for (i = 0; i < adev->gfx.num_xcd; i++) { + for (i = 0; i < num_xcc; i++) { amdgpu_gfx_mqd_sw_fini(adev, i); amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq[i].ring); amdgpu_gfx_kiq_fini(adev, i); @@ -1050,9 +1056,10 @@ static void gfx_v9_4_3_init_gds_vmid(struct amdgpu_device *adev, int xcc_id) static void gfx_v9_4_3_constants_init(struct amdgpu_device *adev) { u32 tmp; - int i, j; + int i, j, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { WREG32_FIELD15_PREREG(GC, i, GRBM_CNTL, READ_TIMEOUT, 0xff); gfx_v9_4_3_setup_rb(adev, i); } @@ -1064,7 +1071,7 @@ static void gfx_v9_4_3_constants_init(struct amdgpu_device *adev) /* where to put LDS, scratch, GPUVM in FSA64 space */ mutex_lock(&adev->srbm_mutex); for (i = 0; i < adev->vm_manager.id_mgr[AMDGPU_GFXHUB(0)].num_ids; i++) { - for (j = 0; j < adev->gfx.num_xcd; j++) { + for (j = 0; j < num_xcc; j++) { soc15_grbm_select(adev, 0, 0, 0, i, j); /* CP and shaders */ if (i == 0) { @@ -1092,7 +1099,7 @@ static void gfx_v9_4_3_constants_init(struct amdgpu_device *adev) mutex_unlock(&adev->srbm_mutex); - for (i = 0; i < adev->gfx.num_xcd; i++) { + for (i = 0; i < num_xcc; i++) { gfx_v9_4_3_init_compute_vmid(adev, i); gfx_v9_4_3_init_gds_vmid(adev, i); } @@ -1150,8 +1157,10 @@ static void gfx_v9_4_3_disable_gpa_mode(struct amdgpu_device *adev, int xcc_id) static void gfx_v9_4_3_program_xcc_id(struct amdgpu_device *adev, int xcc_id) { uint32_t tmp = 0; + int num_xcc; - switch (adev->gfx.num_xcd) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + switch (num_xcc) { /* directly config VIRTUAL_XCC_ID to 0 for 1-XCC */ case 1: WREG32_SOC15(GC, xcc_id, regCP_HYP_XCP_CTL, 0x8); @@ -1288,9 +1297,10 @@ static void gfx_v9_4_3_enable_gui_idle_interrupt(struct amdgpu_device *adev, static void gfx_v9_4_3_rlc_stop(struct amdgpu_device *adev) { - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { WREG32_FIELD15_PREREG(GC, i, RLC_CNTL, RLC_ENABLE_F32, 0); gfx_v9_4_3_enable_gui_idle_interrupt(adev, false, i); gfx_v9_4_3_wait_for_rlc_serdes(adev, i); @@ -1299,9 +1309,10 @@ static void gfx_v9_4_3_rlc_stop(struct amdgpu_device *adev) static void gfx_v9_4_3_rlc_reset(struct amdgpu_device *adev) { - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { WREG32_FIELD15_PREREG(GC, i, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1); udelay(50); WREG32_FIELD15_PREREG(GC, i, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0); @@ -1314,9 +1325,10 @@ static void gfx_v9_4_3_rlc_start(struct amdgpu_device *adev) #ifdef AMDGPU_RLC_DEBUG_RETRY u32 rlc_ucode_ver; #endif - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { WREG32_FIELD15_PREREG(GC, i, RLC_CNTL, RLC_ENABLE_F32, 1); udelay(50); @@ -1377,11 +1389,12 @@ static int gfx_v9_4_3_rlc_load_microcode(struct amdgpu_device *adev, int xcc_id) static int gfx_v9_4_3_rlc_resume(struct amdgpu_device *adev) { - int r, i; + int r, i, num_xcc; adev->gfx.rlc.funcs->stop(adev); - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { /* disable CG */ WREG32_SOC15(GC, i, regRLC_CGCG_CGLS_CTRL, 0); @@ -1954,10 +1967,11 @@ done: static int gfx_v9_4_3_cp_resume(struct amdgpu_device *adev) { - int r, i, j; + int r, i, j, num_xcc; struct amdgpu_ring *ring; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { gfx_v9_4_3_enable_gui_idle_interrupt(adev, false, i); if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) { @@ -2021,12 +2035,13 @@ static int gfx_v9_4_3_hw_init(void *handle) static int gfx_v9_4_3_hw_fini(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; - int i; + int i, num_xcc; amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0); amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0); - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { if (amdgpu_gfx_disable_kcq(adev, i)) DRM_ERROR("XCD %d KCQ disable failed\n", i); @@ -2069,9 +2084,10 @@ static int gfx_v9_4_3_resume(void *handle) static bool gfx_v9_4_3_is_idle(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { if (REG_GET_FIELD(RREG32_SOC15(GC, i, regGRBM_STATUS), GRBM_STATUS, GUI_ACTIVE)) return false; @@ -2183,30 +2199,30 @@ static void gfx_v9_4_3_ring_emit_gds_switch(struct amdgpu_ring *ring, static int gfx_v9_4_3_early_init(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; + int num_xcc; - /* hardcode in emulation phase */ - adev->gfx.num_xcd = 1; + num_xcc = NUM_XCC(adev->gfx.xcc_mask); adev->gfx.partition_mode = amdgpu_user_partt_mode; /* calculate the num_xcc_in_xcp for the partition mode*/ switch (amdgpu_user_partt_mode) { case AMDGPU_SPX_PARTITION_MODE: - adev->gfx.num_xcc_per_xcp = adev->gfx.num_xcd; + adev->gfx.num_xcc_per_xcp = num_xcc; break; case AMDGPU_DPX_PARTITION_MODE: - adev->gfx.num_xcc_per_xcp = adev->gfx.num_xcd / 2; + adev->gfx.num_xcc_per_xcp = num_xcc / 2; break; case AMDGPU_TPX_PARTITION_MODE: - adev->gfx.num_xcc_per_xcp = adev->gfx.num_xcd / 3; + adev->gfx.num_xcc_per_xcp = num_xcc / 3; break; case AMDGPU_QPX_PARTITION_MODE: - adev->gfx.num_xcc_per_xcp = adev->gfx.num_xcd / 4; + adev->gfx.num_xcc_per_xcp = num_xcc / 4; break; case AMDGPU_CPX_PARTITION_MODE: adev->gfx.num_xcc_per_xcp = 1; break; default: - adev->gfx.num_xcc_per_xcp = adev->gfx.num_xcd; + adev->gfx.num_xcc_per_xcp = num_xcc; break; } @@ -2404,14 +2420,15 @@ static int gfx_v9_4_3_set_clockgating_state(void *handle, enum amd_clockgating_state state) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; - int i; + int i, num_xcc; if (amdgpu_sriov_vf(adev)) return 0; + num_xcc = NUM_XCC(adev->gfx.xcc_mask); switch (adev->ip_versions[GC_HWIP][0]) { case IP_VERSION(9, 4, 3): - for (i = 0; i < adev->gfx.num_xcd; i++) + for (i = 0; i < num_xcc; i++) gfx_v9_4_3_update_gfx_clock_gating(adev, state == AMD_CG_STATE_GATE, i); break; @@ -2739,12 +2756,13 @@ static int gfx_v9_4_3_set_priv_reg_fault_state(struct amdgpu_device *adev, unsigned type, enum amdgpu_interrupt_state state) { - int i; + int i, num_xcc; + num_xcc = NUM_XCC(adev->gfx.xcc_mask); switch (state) { case AMDGPU_IRQ_STATE_DISABLE: case AMDGPU_IRQ_STATE_ENABLE: - for (i = 0; i < adev->gfx.num_xcd; i++) + for (i = 0; i < num_xcc; i++) WREG32_FIELD15_PREREG(GC, i, CP_INT_CNTL_RING0, PRIV_REG_INT_ENABLE, state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); @@ -2761,12 +2779,13 @@ static int gfx_v9_4_3_set_priv_inst_fault_state(struct amdgpu_device *adev, unsigned type, enum amdgpu_interrupt_state state) { - int i; + int i, num_xcc; + num_xcc = NUM_XCC(adev->gfx.xcc_mask); switch (state) { case AMDGPU_IRQ_STATE_DISABLE: case AMDGPU_IRQ_STATE_ENABLE: - for (i = 0; i < adev->gfx.num_xcd; i++) + for (i = 0; i < num_xcc; i++) WREG32_FIELD15_PREREG(GC, i, CP_INT_CNTL_RING0, PRIV_INSTR_INT_ENABLE, state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); @@ -2783,8 +2802,10 @@ static int gfx_v9_4_3_set_eop_interrupt_state(struct amdgpu_device *adev, unsigned type, enum amdgpu_interrupt_state state) { - int i; - for (i = 0; i < adev->gfx.num_xcd; i++) { + int i, num_xcc; + + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { switch (type) { case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP: gfx_v9_4_3_set_compute_eop_interrupt_state(adev, 1, 0, state, i); @@ -2842,6 +2863,7 @@ static int gfx_v9_4_3_eop_irq(struct amdgpu_device *adev, /* Per-queue interrupt is supported for MEC starting from VI. * The interrupt can only be enabled/disabled per pipe instead of per queue. */ + if ((ring->me == me_id) && (ring->pipe == pipe_id) && (ring->queue == queue_id)) amdgpu_fence_process(ring); } @@ -3056,9 +3078,10 @@ static const struct amdgpu_ring_funcs gfx_v9_4_3_ring_funcs_kiq = { static void gfx_v9_4_3_set_ring_funcs(struct amdgpu_device *adev) { - int i, j; + int i, j, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { adev->gfx.kiq[i].ring.funcs = &gfx_v9_4_3_ring_funcs_kiq; for (j = 0; j < adev->gfx.num_compute_rings; j++) diff --git a/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c b/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c index 1bb17d95f720..e35365ab3f1f 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c +++ b/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c @@ -43,9 +43,10 @@ static void gfxhub_v1_2_setup_vm_pt_regs(struct amdgpu_device *adev, uint64_t page_table_base) { struct amdgpu_vmhub *hub; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { hub = &adev->vmhub[AMDGPU_GFXHUB(i)]; WREG32_SOC15_OFFSET(GC, i, regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32, @@ -56,13 +57,14 @@ static void gfxhub_v1_2_setup_vm_pt_regs(struct amdgpu_device *adev, regVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32, hub->ctx_addr_distance * vmid, upper_32_bits(page_table_base)); + } } static void gfxhub_v1_2_init_gart_aperture_regs(struct amdgpu_device *adev) { uint64_t pt_base; - int i; + int i, num_xcc; if (adev->gmc.pdb0_bo) pt_base = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo); @@ -74,7 +76,8 @@ static void gfxhub_v1_2_init_gart_aperture_regs(struct amdgpu_device *adev) /* If use GART for FB translation, vmid0 page table covers both * vram and system memory (gart) */ - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { if (adev->gmc.pdb0_bo) { WREG32_SOC15(GC, i, regVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32, @@ -111,9 +114,10 @@ static void gfxhub_v1_2_init_system_aperture_regs(struct amdgpu_device *adev) { uint64_t value; uint32_t tmp; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { /* Program the AGP BAR */ WREG32_SOC15_RLC(GC, i, regMC_VM_AGP_BASE, 0); WREG32_SOC15_RLC(GC, i, regMC_VM_AGP_BOT, adev->gmc.agp_start >> 24); @@ -177,9 +181,10 @@ static void gfxhub_v1_2_init_system_aperture_regs(struct amdgpu_device *adev) static void gfxhub_v1_2_init_tlb_regs(struct amdgpu_device *adev) { uint32_t tmp; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { /* Setup TLB control */ tmp = RREG32_SOC15(GC, i, regMC_VM_MX_L1_TLB_CNTL); @@ -202,9 +207,10 @@ static void gfxhub_v1_2_init_tlb_regs(struct amdgpu_device *adev) static void gfxhub_v1_2_init_cache_regs(struct amdgpu_device *adev) { uint32_t tmp; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { /* Setup L2 cache */ tmp = RREG32_SOC15(GC, i, regVM_L2_CNTL); tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); @@ -249,9 +255,10 @@ static void gfxhub_v1_2_init_cache_regs(struct amdgpu_device *adev) static void gfxhub_v1_2_enable_system_domain(struct amdgpu_device *adev) { uint32_t tmp; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { tmp = RREG32_SOC15(GC, i, regVM_CONTEXT0_CNTL); tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, @@ -266,9 +273,10 @@ static void gfxhub_v1_2_enable_system_domain(struct amdgpu_device *adev) static void gfxhub_v1_2_disable_identity_aperture(struct amdgpu_device *adev) { - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { WREG32_SOC15(GC, i, regVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR_LO32, 0XFFFFFFFF); @@ -295,7 +303,7 @@ static void gfxhub_v1_2_setup_vmid_config(struct amdgpu_device *adev) struct amdgpu_vmhub *hub; unsigned num_level, block_size; uint32_t tmp; - int i, j; + int i, j, num_xcc; num_level = adev->vm_manager.num_level; block_size = adev->vm_manager.block_size; @@ -304,7 +312,8 @@ static void gfxhub_v1_2_setup_vmid_config(struct amdgpu_device *adev) else block_size -= 9; - for (j = 0; j < adev->gfx.num_xcd; j++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (j = 0; j < num_xcc; j++) { hub = &adev->vmhub[AMDGPU_GFXHUB(j)]; for (i = 0; i <= 14; i++) { tmp = RREG32_SOC15_OFFSET(GC, j, regVM_CONTEXT1_CNTL, i); @@ -362,10 +371,12 @@ static void gfxhub_v1_2_setup_vmid_config(struct amdgpu_device *adev) static void gfxhub_v1_2_program_invalidation(struct amdgpu_device *adev) { struct amdgpu_vmhub *hub; - unsigned i, j; + unsigned i, j, num_xcc; - for (j = 0; j < adev->gfx.num_xcd; j++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (j = 0; j < num_xcc; j++) { hub = &adev->vmhub[AMDGPU_GFXHUB(j)]; + for (i = 0 ; i < 18; ++i) { WREG32_SOC15_OFFSET(GC, j, regVM_INVALIDATE_ENG0_ADDR_RANGE_LO32, i * hub->eng_addr_distance, 0xffffffff); @@ -377,9 +388,10 @@ static void gfxhub_v1_2_program_invalidation(struct amdgpu_device *adev) static int gfxhub_v1_2_gart_enable(struct amdgpu_device *adev) { - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { if (amdgpu_sriov_vf(adev)) { /* * MC_VM_FB_LOCATION_BASE/TOP is NULL for VF, becuase they are @@ -413,9 +425,10 @@ static void gfxhub_v1_2_gart_disable(struct amdgpu_device *adev) { struct amdgpu_vmhub *hub; u32 tmp; - u32 i, j; + u32 i, j, num_xcc; - for (j = 0; j < adev->gfx.num_xcd; j++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (j = 0; j < num_xcc; j++) { hub = &adev->vmhub[AMDGPU_GFXHUB(j)]; /* Disable all tables */ for (i = 0; i < 16; i++) @@ -449,9 +462,10 @@ static void gfxhub_v1_2_set_fault_enable_default(struct amdgpu_device *adev, bool value) { u32 tmp; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { tmp = RREG32_SOC15(GC, i, regVM_L2_PROTECTION_FAULT_CNTL); tmp = REG_SET_FIELD(tmp, VM_L2_PROTECTION_FAULT_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); @@ -490,9 +504,10 @@ static void gfxhub_v1_2_set_fault_enable_default(struct amdgpu_device *adev, static void gfxhub_v1_2_init(struct amdgpu_device *adev) { struct amdgpu_vmhub *hub; - int i; + int i, num_xcc; - for (i = 0; i < adev->gfx.num_xcd; i++) { + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + for (i = 0; i < num_xcc; i++) { hub = &adev->vmhub[AMDGPU_GFXHUB(i)]; hub->ctx0_ptb_addr_lo32 = diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 4b2c4ecd7253..2c322a25bf1c 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1733,7 +1733,8 @@ static int gmc_v9_0_sw_init(void *handle) adev->gmc.translate_further = adev->vm_manager.num_level > 1; break; case IP_VERSION(9, 4, 3): - bitmap_set(adev->vmhubs_mask, AMDGPU_GFXHUB(0), adev->gfx.num_xcd); + bitmap_set(adev->vmhubs_mask, AMDGPU_GFXHUB(0), + NUM_XCC(adev->gfx.xcc_mask)); bitmap_set(adev->vmhubs_mask, AMDGPU_MMHUB0(0), adev->num_aid); amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 829e32433faf..df96c4c508a0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -592,6 +592,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, struct kfd_node *node; uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; unsigned int max_proc_per_quantum; + int num_xcd; kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, KGD_ENGINE_MEC1); @@ -601,16 +602,15 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, KGD_ENGINE_SDMA1); kfd->shared_resources = *gpu_resources; - if (kfd->adev->gfx.num_xcd == 0 || kfd->adev->gfx.num_xcd == 1 || - kfd->adev->gfx.num_xcc_per_xcp == 0) + num_xcd = NUM_XCC(kfd->adev->gfx.xcc_mask); + if (num_xcd == 0 || num_xcd == 1 || kfd->adev->gfx.num_xcc_per_xcp == 0) kfd->num_nodes = 1; else - kfd->num_nodes = - kfd->adev->gfx.num_xcd/kfd->adev->gfx.num_xcc_per_xcp; + kfd->num_nodes = num_xcd / kfd->adev->gfx.num_xcc_per_xcp; if (kfd->num_nodes == 0) { dev_err(kfd_device, "KFD num nodes cannot be 0, GC inst: %d, num_xcc_in_node: %d\n", - kfd->adev->gfx.num_xcd, kfd->adev->gfx.num_xcc_per_xcp); + num_xcd, kfd->adev->gfx.num_xcc_per_xcp); goto out; } -- cgit v1.2.3 From 1526ec9a3ed90e7ad36de7ef8aa2768b60af14df Mon Sep 17 00:00:00 2001 From: David Belanger Date: Wed, 16 Feb 2022 12:07:28 -0500 Subject: drm/amdkfd: EOP Removal - Handle size 0 correctly On GC 9.4.3, we are removing the EOP buffer. If we specify 0 for the size, CP_HQD_EOP_CONTROL ends up with incorrect value as order_size_2 calculations does not handle 0. Fix it by using zero for the MQD entry for EOP size 0. v2: Reworked code with a conditional assignment and fixed style issues. Signed-off-by: David Belanger Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 5c9b3392758e..2e2a0f8586f7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -257,9 +257,14 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit * more than (EOP entry count - 1) so a queue size of 0x800 dwords * is safe, giving a maximum field value of 0xA. + * + * Also, do calculation only if EOP is used (size > 0), otherwise + * the order_base_2 calculation provides incorrect result. + * */ - m->cp_hqd_eop_control = min(0xA, - order_base_2(q->eop_ring_buffer_size / 4) - 1); + m->cp_hqd_eop_control = q->eop_ring_buffer_size ? + min(0xA, order_base_2(q->eop_ring_buffer_size / 4) - 1) : 0; + m->cp_hqd_eop_base_addr_lo = lower_32_bits(q->eop_ring_buffer_address >> 8); m->cp_hqd_eop_base_addr_hi = -- cgit v1.2.3 From fe1f05df5919c67c3add49efb55e251a8d78ee4e Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 31 May 2022 14:39:36 -0400 Subject: drm/amdkfd: Rework kfd_locked handling Currently, even if kfd_locked is set, a process is first created and then removed to work around a race condition in updating kfd_locked flag. Rework kfd_locked handling to ensure no processes is created if kfd_locked is set. This is achieved by updating kfd_locked under kfd_processes_mutex. With this there is no need for kfd_locked to be an atomic counter. Instead, it can be a regular integer. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 ------- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 21 ++++++++++++++++----- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 ++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 8 +++++++- 4 files changed, 25 insertions(+), 13 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 45e8da125f70..8b9accecf49b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -146,13 +146,6 @@ static int kfd_open(struct inode *inode, struct file *filep) if (IS_ERR(process)) return PTR_ERR(process); - if (kfd_is_locked()) { - dev_dbg(kfd_device, "kfd is locked!\n" - "process %d unreferenced", process->pasid); - kfd_unref_process(process); - return -EAGAIN; - } - /* filep now owns the reference returned by kfd_create_process */ filep->private_data = process; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index df96c4c508a0..eb2b44fddf74 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -42,7 +42,7 @@ * once locked, kfd driver will stop any further GPU execution. * create process (open) will return -EAGAIN. */ -static atomic_t kfd_locked = ATOMIC_INIT(0); +static int kfd_locked; #ifdef CONFIG_DRM_AMDGPU_CIK extern const struct kfd2kgd_calls gfx_v7_kfd2kgd; @@ -880,7 +880,9 @@ int kgd2kfd_post_reset(struct kfd_dev *kfd) return ret; } - atomic_dec(&kfd_locked); + mutex_lock(&kfd_processes_mutex); + --kfd_locked; + mutex_unlock(&kfd_processes_mutex); for (i = 0; i < kfd->num_nodes; i++) { node = kfd->nodes[i]; @@ -893,21 +895,27 @@ int kgd2kfd_post_reset(struct kfd_dev *kfd) bool kfd_is_locked(void) { - return (atomic_read(&kfd_locked) > 0); + lockdep_assert_held(&kfd_processes_mutex); + return (kfd_locked > 0); } void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) { struct kfd_node *node; int i; + int count; if (!kfd->init_complete) return; /* for runtime suspend, skip locking kfd */ if (!run_pm) { + mutex_lock(&kfd_processes_mutex); + count = ++kfd_locked; + mutex_unlock(&kfd_processes_mutex); + /* For first KFD device suspend all the KFD processes */ - if (atomic_inc_return(&kfd_locked) == 1) + if (count == 1) kfd_suspend_all_processes(); } @@ -933,7 +941,10 @@ int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm) /* for runtime resume, skip unlocking kfd */ if (!run_pm) { - count = atomic_dec_return(&kfd_locked); + mutex_lock(&kfd_processes_mutex); + count = --kfd_locked; + mutex_unlock(&kfd_processes_mutex); + WARN_ONCE(count < 0, "KFD suspend / resume ref. error"); if (count == 0) ret = kfd_resume_all_processes(); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 5cfebcc8b305..400b4dcbdf05 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -201,6 +201,8 @@ extern int amdgpu_no_queue_eviction_on_vm_fault; /* Enable eviction debug messages */ extern bool debug_evictions; +extern struct mutex kfd_processes_mutex; + enum cache_policy { cache_policy_coherent, cache_policy_noncoherent diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 9b1e84d33cdc..c3d43e6e5236 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -50,7 +50,7 @@ struct mm_struct; * Unique/indexed by mm_struct* */ DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); -static DEFINE_MUTEX(kfd_processes_mutex); +DEFINE_MUTEX(kfd_processes_mutex); DEFINE_SRCU(kfd_processes_srcu); @@ -818,6 +818,12 @@ struct kfd_process *kfd_create_process(struct file *filep) */ mutex_lock(&kfd_processes_mutex); + if (kfd_is_locked()) { + mutex_unlock(&kfd_processes_mutex); + pr_debug("KFD is locked! Cannot create process"); + return ERR_PTR(-EINVAL); + } + /* A prior open of /dev/kfd could have already created the process. */ process = find_process(thread, false); if (process) { -- cgit v1.2.3 From 0c7315e7d5ef9b36ca4db32ffeb34a187cbaf231 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Fri, 10 Jun 2022 09:41:29 -0400 Subject: drm/amdkfd: Add device repartition support GFX9.4.3 will support dynamic repartitioning of the GPU through sysfs. Add device repartitioning support in KFD to repartition GPU from one mode to other. v2: squash in fix ("drm/amdkfd: Fix warning kgd2kfd_unlock_kfd defined but not used") Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 10 ++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 13 +++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 22 +++++++++++++++++++++- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 5 +---- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 21 +++++++++++++++++++++ 5 files changed, 66 insertions(+), 5 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 9d19c7ceda3f..bbbfe9ec4adf 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -773,3 +773,13 @@ bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev) else return false; } + +int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev) +{ + return kgd2kfd_check_and_lock_kfd(); +} + +void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev) +{ + kgd2kfd_unlock_kfd(); +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index df07e212c21e..d1d643a050a1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -151,6 +151,8 @@ void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev, void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev); void amdgpu_amdkfd_device_init(struct amdgpu_device *adev); void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev); +int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev); +void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev); int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev, enum kgd_engine_type engine, uint32_t vmid, uint64_t gpu_addr, @@ -373,6 +375,8 @@ int kgd2kfd_post_reset(struct kfd_dev *kfd); void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry); void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd); void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask); +int kgd2kfd_check_and_lock_kfd(void); +void kgd2kfd_unlock_kfd(void); #else static inline int kgd2kfd_init(void) { @@ -438,5 +442,14 @@ static inline void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask) { } + +static inline int kgd2kfd_check_and_lock_kfd(void) +{ + return 0; +} + +static inline void kgd2kfd_unlock_kfd(void) +{ +} #endif #endif /* AMDGPU_AMDKFD_H_INCLUDED */ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 069b259f384c..69bac5b801ce 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -1233,10 +1233,30 @@ static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev, return -EINVAL; } + if (!adev->kfd.init_complete) + return -EPERM; + mutex_lock(&adev->gfx.partition_mutex); - ret = adev->gfx.funcs->switch_partition_mode(adev, mode); + if (mode == adev->gfx.funcs->query_partition_mode(adev)) + goto out; + + ret = amdgpu_amdkfd_check_and_lock_kfd(adev); + if (ret) + goto out; + + amdgpu_amdkfd_device_fini_sw(adev); + + adev->gfx.funcs->switch_partition_mode(adev, mode); + + amdgpu_amdkfd_device_probe(adev); + amdgpu_amdkfd_device_init(adev); + /* If KFD init failed, return failure */ + if (!adev->kfd.init_complete) + ret = -EIO; + amdgpu_amdkfd_unlock_kfd(adev); +out: mutex_unlock(&adev->gfx.partition_mutex); if (ret) diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c index c776fc5884de..47d8ac64e877 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c @@ -675,7 +675,7 @@ static void gfx_v9_4_3_select_me_pipe_q(struct amdgpu_device *adev, static enum amdgpu_gfx_partition gfx_v9_4_3_query_compute_partition(struct amdgpu_device *adev) { - enum amdgpu_gfx_partition mode = AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE; + enum amdgpu_gfx_partition mode = adev->gfx.partition_mode; if (adev->nbio.funcs->get_compute_partition_mode) mode = adev->nbio.funcs->get_compute_partition_mode(adev); @@ -689,9 +689,6 @@ static int gfx_v9_4_3_switch_compute_partition(struct amdgpu_device *adev, u32 tmp = 0; int num_xcc_per_partition, i, num_xcc; - if (mode == adev->gfx.partition_mode) - return mode; - num_xcc = NUM_XCC(adev->gfx.xcc_mask); switch (mode) { case AMDGPU_SPX_PARTITION_MODE: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index eb2b44fddf74..293787290e36 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -1356,6 +1356,27 @@ unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *node) kfd_get_num_sdma_engines(node); } +int kgd2kfd_check_and_lock_kfd(void) +{ + mutex_lock(&kfd_processes_mutex); + if (!hash_empty(kfd_processes_table) || kfd_is_locked()) { + mutex_unlock(&kfd_processes_mutex); + return -EBUSY; + } + + ++kfd_locked; + mutex_unlock(&kfd_processes_mutex); + + return 0; +} + +void kgd2kfd_unlock_kfd(void) +{ + mutex_lock(&kfd_processes_mutex); + --kfd_locked; + mutex_unlock(&kfd_processes_mutex); +} + #if defined(CONFIG_DEBUG_FS) /* This function will send a package to HIQ to hang the HWS -- cgit v1.2.3 From 5fb34bd9cf9e248d7e84e431a4a6b731334ab564 Mon Sep 17 00:00:00 2001 From: Alex Sierra Date: Tue, 24 May 2022 10:22:12 -0500 Subject: drm/amdkfd: pass kfd_node ref to svm migration api This work is required for GC 9.4.3, previous to support memory partitions per node at SVM. When multiple partition is configured, every BO should be allocated inside one specific partition which corresponds to the current amdgpu_device and kfd_node. v2: squash in compilation fix (Alex) v3: squash in fix for pre-gfx 9.4.3 (Alex) v4: squash in best_loc fix (Alex) Signed-off-by: Alex Sierra Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 7 +- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 3 +- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 33 ++++--- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 45 ++++----- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 30 +++++- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 152 +++++++++++++++---------------- drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 21 +++-- 9 files changed, 166 insertions(+), 133 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index c3964c14f215..c390b2856cc9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2441,7 +2441,8 @@ void amdgpu_vm_set_task_info(struct amdgpu_vm *vm) * shouldn't be reported any more. */ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, - uint64_t addr, bool write_fault) + u32 client_id, u32 node_id, uint64_t addr, + bool write_fault) { bool is_compute_context = false; struct amdgpu_bo *root; @@ -2465,8 +2466,8 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, addr /= AMDGPU_GPU_PAGE_SIZE; - if (is_compute_context && - !svm_range_restore_pages(adev, pasid, addr, write_fault)) { + if (is_compute_context && !svm_range_restore_pages(adev, pasid, client_id, + node_id, addr, write_fault)) { amdgpu_bo_unref(&root); return true; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h index 9f5d32b0fda1..dbab31647186 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h @@ -455,7 +455,8 @@ void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev); void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid, struct amdgpu_task_info *task_info); bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, - uint64_t addr, bool write_fault); + u32 client_id, u32 node_id, uint64_t addr, + bool write_fault); void amdgpu_vm_set_task_info(struct amdgpu_vm *vm); diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c index d76f5c8d4977..01bd45651382 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c @@ -139,7 +139,7 @@ static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev, /* Try to handle the recoverable page faults by filling page * tables */ - if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault)) + if (amdgpu_vm_handle_fault(adev, entry->pasid, 0, 0, addr, write_fault)) return 1; } diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 2c322a25bf1c..c5752a349f3d 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -557,11 +557,24 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, u64 addr; uint32_t cam_index = 0; int ret; - uint32_t node_id; + uint32_t node_id = 0; addr = (u64)entry->src_data[0] << 12; addr |= ((u64)entry->src_data[1] & 0xf) << 44; + if (entry->client_id == SOC15_IH_CLIENTID_VMC) { + hub_name = "mmhub0"; + hub = &adev->vmhub[AMDGPU_MMHUB0(0)]; + } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) { + hub_name = "mmhub1"; + hub = &adev->vmhub[AMDGPU_MMHUB1(0)]; + } else { + hub_name = "gfxhub0"; + node_id = (adev->ip_versions[GC_HWIP][0] == + IP_VERSION(9, 4, 3)) ? entry->node_id : 0; + hub = &adev->vmhub[node_id/2]; + } + if (retry_fault) { if (adev->irq.retry_cam_enabled) { /* Delegate it to a different ring if the hardware hasn't @@ -574,7 +587,8 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, cam_index = entry->src_data[2] & 0x3ff; - ret = amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault); + ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->client_id, node_id, + addr, write_fault); WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index); if (ret) return 1; @@ -596,7 +610,8 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, /* Try to handle the recoverable page faults by filling page * tables */ - if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault)) + if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->client_id, node_id, + addr, write_fault)) return 1; } } @@ -604,18 +619,6 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, if (!printk_ratelimit()) return 0; - if (entry->client_id == SOC15_IH_CLIENTID_VMC) { - hub_name = "mmhub0"; - hub = &adev->vmhub[AMDGPU_MMHUB0(0)]; - } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) { - hub_name = "mmhub1"; - hub = &adev->vmhub[AMDGPU_MMHUB1(0)]; - } else { - hub_name = "gfxhub0"; - node_id = (adev->ip_versions[GC_HWIP][0] == - IP_VERSION(9, 4, 3)) ? entry->node_id : 0; - hub = &adev->vmhub[node_id/2]; - } memset(&task_info, 0, sizeof(struct amdgpu_task_info)); amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 5f4dc2a45bd0..e7e5abc32c84 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -287,11 +287,12 @@ static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate) } static int -svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange, +svm_migrate_copy_to_vram(struct kfd_node *node, struct svm_range *prange, struct migrate_vma *migrate, struct dma_fence **mfence, dma_addr_t *scratch, uint64_t ttm_res_offset) { - uint64_t npages = migrate->npages; + uint64_t npages = migrate->cpages; + struct amdgpu_device *adev = node->adev; struct device *dev = adev->dev; struct amdgpu_res_cursor cursor; dma_addr_t *src; @@ -321,7 +322,7 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange, DMA_TO_DEVICE); r = dma_mapping_error(dev, src[i]); if (r) { - dev_err(adev->dev, "%s: fail %d dma_map_page\n", + dev_err(dev, "%s: fail %d dma_map_page\n", __func__, r); goto out_free_vram_pages; } @@ -390,12 +391,13 @@ out_free_vram_pages: } static long -svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, +svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, struct vm_area_struct *vma, uint64_t start, uint64_t end, uint32_t trigger, uint64_t ttm_res_offset) { struct kfd_process *p = container_of(prange->svms, struct kfd_process, svms); uint64_t npages = (end - start) >> PAGE_SHIFT; + struct amdgpu_device *adev = node->adev; struct kfd_process_device *pdd; struct dma_fence *mfence = NULL; struct migrate_vma migrate = { 0 }; @@ -445,7 +447,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, else pr_debug("0x%lx pages migrated\n", cpages); - r = svm_migrate_copy_to_vram(adev, prange, &migrate, &mfence, scratch, ttm_res_offset); + r = svm_migrate_copy_to_vram(node, prange, &migrate, &mfence, scratch, ttm_res_offset); migrate_vma_pages(&migrate); pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n", @@ -465,7 +467,7 @@ out_free: kvfree(buf); out: if (!r && cpages) { - pdd = svm_range_get_pdd_by_adev(prange, adev); + pdd = svm_range_get_pdd_by_node(prange, node); if (pdd) WRITE_ONCE(pdd->page_in, pdd->page_in + cpages); @@ -492,8 +494,8 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, { unsigned long addr, start, end; struct vm_area_struct *vma; - struct amdgpu_device *adev; uint64_t ttm_res_offset; + struct kfd_node *node; unsigned long cpages = 0; long r = 0; @@ -503,9 +505,9 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, return 0; } - adev = svm_range_get_adev_by_id(prange, best_loc); - if (!adev) { - pr_debug("failed to get device by id 0x%x\n", best_loc); + node = svm_range_get_node_by_id(prange, best_loc); + if (!node) { + pr_debug("failed to get kfd node by id 0x%x\n", best_loc); return -ENODEV; } @@ -515,9 +517,9 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, start = prange->start << PAGE_SHIFT; end = (prange->last + 1) << PAGE_SHIFT; - r = svm_range_vram_node_new(adev, prange, true); + r = svm_range_vram_node_new(node, prange, true); if (r) { - dev_dbg(adev->dev, "fail %ld to alloc vram\n", r); + dev_dbg(node->adev->dev, "fail %ld to alloc vram\n", r); return r; } ttm_res_offset = prange->offset << PAGE_SHIFT; @@ -530,7 +532,7 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc, break; next = min(vma->vm_end, end); - r = svm_migrate_vma_to_vram(adev, prange, vma, addr, next, trigger, ttm_res_offset); + r = svm_migrate_vma_to_vram(node, prange, vma, addr, next, trigger, ttm_res_offset); if (r < 0) { pr_debug("failed %ld to migrate\n", r); break; @@ -663,7 +665,7 @@ out_oom: * positive values - partial migration, number of pages not migrated */ static long -svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, +svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, struct vm_area_struct *vma, uint64_t start, uint64_t end, uint32_t trigger, struct page *fault_page) { @@ -671,6 +673,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, uint64_t npages = (end - start) >> PAGE_SHIFT; unsigned long upages = npages; unsigned long cpages = 0; + struct amdgpu_device *adev = node->adev; struct kfd_process_device *pdd; struct dma_fence *mfence = NULL; struct migrate_vma migrate = { 0 }; @@ -745,7 +748,7 @@ out_free: kvfree(buf); out: if (!r && cpages) { - pdd = svm_range_get_pdd_by_adev(prange, adev); + pdd = svm_range_get_pdd_by_node(prange, node); if (pdd) WRITE_ONCE(pdd->page_out, pdd->page_out + cpages); } @@ -766,7 +769,7 @@ out: int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, uint32_t trigger, struct page *fault_page) { - struct amdgpu_device *adev; + struct kfd_node *node; struct vm_area_struct *vma; unsigned long addr; unsigned long start; @@ -780,13 +783,11 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, return 0; } - adev = svm_range_get_adev_by_id(prange, prange->actual_loc); - if (!adev) { - pr_debug("failed to get device by id 0x%x\n", - prange->actual_loc); + node = svm_range_get_node_by_id(prange, prange->actual_loc); + if (!node) { + pr_debug("failed to get kfd node by id 0x%x\n", prange->actual_loc); return -ENODEV; } - pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] from gpu 0x%x to ram\n", prange->svms, prange, prange->start, prange->last, prange->actual_loc); @@ -805,7 +806,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, } next = min(vma->vm_end, end); - r = svm_migrate_vma_to_ram(adev, prange, vma, addr, next, trigger, + r = svm_migrate_vma_to_ram(node, prange, vma, addr, next, trigger, fault_page); if (r < 0) { pr_debug("failed %ld to migrate prange %p\n", r, prange); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 400b4dcbdf05..df372de6b056 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -289,6 +289,7 @@ struct kfd_node { * from the HW ring into a SW ring. */ bool interrupts_active; + uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */ /* QCM Device instance */ struct device_queue_manager *dqm; @@ -971,9 +972,8 @@ struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid); struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id); -int kfd_process_gpuid_from_adev(struct kfd_process *p, - struct amdgpu_device *adev, uint32_t *gpuid, - uint32_t *gpuidx); +int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node, + uint32_t *gpuid, uint32_t *gpuidx); static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p, uint32_t gpuidx, uint32_t *gpuid) { return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL; @@ -1073,6 +1073,30 @@ struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); struct kfd_node *kfd_device_by_id(uint32_t gpu_id); struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev); struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev); +static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t client_id, + uint32_t node_id) +{ + if ((node->interrupt_bitmap & (0x1U << node_id)) || + ((node_id % 4) == 0 && + (node->interrupt_bitmap >> 16) & (0x1U << client_id))) + return true; + + return false; +} +static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev, + uint32_t client_id, uint32_t node_id) { + struct kfd_dev *dev = adev->kfd.dev; + uint32_t i; + + if (adev->ip_versions[GC_HWIP][0] != IP_VERSION(9, 4, 3)) + return dev->nodes[0]; + + for (i = 0; i < dev->num_nodes; i++) + if (kfd_irq_is_from_node(dev->nodes[i], client_id, node_id)) + return dev->nodes[i]; + + return NULL; +} int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev); int kfd_numa_node_to_apic_id(int numa_node_id); void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index c3d43e6e5236..666815b227a8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1891,13 +1891,13 @@ int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id) } int -kfd_process_gpuid_from_adev(struct kfd_process *p, struct amdgpu_device *adev, - uint32_t *gpuid, uint32_t *gpuidx) +kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node, + uint32_t *gpuid, uint32_t *gpuidx) { int i; for (i = 0; i < p->n_pdds; i++) - if (p->pdds[i] && p->pdds[i]->dev->adev == adev) { + if (p->pdds[i] && p->pdds[i]->dev == node) { *gpuid = p->pdds[i]->user_gpu_id; *gpuidx = i; return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 4b4f3bf8b823..639831fbb6ca 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -170,8 +170,7 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, page = hmm_pfn_to_page(hmm_pfns[i]); if (is_zone_device_page(page)) { - struct amdgpu_device *bo_adev = - amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); + struct amdgpu_device *bo_adev = prange->svm_bo->node->adev; addr[i] = (hmm_pfns[i] << PAGE_SHIFT) + bo_adev->vm_manager.vram_base_offset - @@ -424,10 +423,8 @@ static void svm_range_bo_unref(struct svm_range_bo *svm_bo) } static bool -svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange) +svm_range_validate_svm_bo(struct kfd_node *node, struct svm_range *prange) { - struct amdgpu_device *bo_adev; - mutex_lock(&prange->lock); if (!prange->svm_bo) { mutex_unlock(&prange->lock); @@ -440,12 +437,11 @@ svm_range_validate_svm_bo(struct amdgpu_device *adev, struct svm_range *prange) } if (svm_bo_ref_unless_zero(prange->svm_bo)) { /* - * Migrate from GPU to GPU, remove range from source bo_adev - * svm_bo range list, and return false to allocate svm_bo from - * destination adev. + * Migrate from GPU to GPU, remove range from source svm_bo->node + * range list, and return false to allocate svm_bo from destination + * node. */ - bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); - if (bo_adev != adev) { + if (prange->svm_bo->node != node) { mutex_unlock(&prange->lock); spin_lock(&prange->svm_bo->list_lock); @@ -513,7 +509,7 @@ static struct svm_range_bo *svm_range_bo_new(void) } int -svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, +svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, bool clear) { struct amdgpu_bo_param bp; @@ -528,7 +524,7 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, pr_debug("pasid: %x svms 0x%p [0x%lx 0x%lx]\n", p->pasid, prange->svms, prange->start, prange->last); - if (svm_range_validate_svm_bo(adev, prange)) + if (svm_range_validate_svm_bo(node, prange)) return 0; svm_bo = svm_range_bo_new(); @@ -542,6 +538,7 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, kfree(svm_bo); return -ESRCH; } + svm_bo->node = node; svm_bo->eviction_fence = amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1), mm, @@ -559,7 +556,10 @@ svm_range_vram_node_new(struct amdgpu_device *adev, struct svm_range *prange, bp.type = ttm_bo_type_device; bp.resv = NULL; - r = amdgpu_bo_create_user(adev, &bp, &ubo); + /* TODO: Allocate memory from the right memory partition. We can sort + * out the details later, once basic memory partitioning is working + */ + r = amdgpu_bo_create_user(node->adev, &bp, &ubo); if (r) { pr_debug("failed %d to create bo\n", r); goto create_bo_failed; @@ -617,45 +617,30 @@ void svm_range_vram_node_free(struct svm_range *prange) prange->ttm_res = NULL; } -struct amdgpu_device * -svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id) +struct kfd_node * +svm_range_get_node_by_id(struct svm_range *prange, uint32_t gpu_id) { - struct kfd_process_device *pdd; struct kfd_process *p; - int32_t gpu_idx; + struct kfd_process_device *pdd; p = container_of(prange->svms, struct kfd_process, svms); - - gpu_idx = kfd_process_gpuidx_from_gpuid(p, gpu_id); - if (gpu_idx < 0) { - pr_debug("failed to get device by id 0x%x\n", gpu_id); - return NULL; - } - pdd = kfd_process_device_from_gpuidx(p, gpu_idx); + pdd = kfd_process_device_data_by_id(p, gpu_id); if (!pdd) { - pr_debug("failed to get device by idx 0x%x\n", gpu_idx); + pr_debug("failed to get kfd process device by id 0x%x\n", gpu_id); return NULL; } - return pdd->dev->adev; + return pdd->dev; } struct kfd_process_device * -svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev) +svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node) { struct kfd_process *p; - int32_t gpu_idx, gpuid; - int r; p = container_of(prange->svms, struct kfd_process, svms); - r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpu_idx); - if (r) { - pr_debug("failed to get device id by adev %p\n", adev); - return NULL; - } - - return kfd_process_device_from_gpuidx(p, gpu_idx); + return kfd_get_process_device_data(node, p); } static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo) @@ -1148,12 +1133,18 @@ svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, } return 0; } +static bool +svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b) +{ + return (node_a->adev == node_b->adev || + amdgpu_xgmi_same_hive(node_a->adev, node_b->adev)); +} static uint64_t -svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, - int domain) +svm_range_get_pte_flags(struct kfd_node *node, + struct svm_range *prange, int domain) { - struct amdgpu_device *bo_adev; + struct kfd_node *bo_node; uint32_t flags = prange->flags; uint32_t mapping_flags = 0; uint64_t pte_flags; @@ -1162,18 +1153,18 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, bool uncached = flags & KFD_IOCTL_SVM_FLAG_UNCACHED; if (domain == SVM_RANGE_VRAM_DOMAIN) - bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); + bo_node = prange->svm_bo->node; - switch (KFD_GC_VERSION(adev->kfd.dev)) { + switch (node->adev->ip_versions[GC_HWIP][0]) { case IP_VERSION(9, 4, 1): if (domain == SVM_RANGE_VRAM_DOMAIN) { - if (bo_adev == adev) { + if (bo_node == node) { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; } else { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; - if (amdgpu_xgmi_same_hive(adev, bo_adev)) + if (svm_nodes_in_same_hive(node, bo_node)) snoop = true; } } else { @@ -1183,15 +1174,15 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, break; case IP_VERSION(9, 4, 2): if (domain == SVM_RANGE_VRAM_DOMAIN) { - if (bo_adev == adev) { + if (bo_node == node) { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; - if (adev->gmc.xgmi.connected_to_cpu) + if (node->adev->gmc.xgmi.connected_to_cpu) snoop = true; } else { mapping_flags |= coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC; - if (amdgpu_xgmi_same_hive(adev, bo_adev)) + if (svm_nodes_in_same_hive(node, bo_node)) snoop = true; } } else { @@ -1207,7 +1198,7 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, if (uncached) mapping_flags |= AMDGPU_VM_MTYPE_UC; /* local HBM region close to partition*/ - else if (bo_adev == adev) + else if (bo_node == node) mapping_flags |= AMDGPU_VM_MTYPE_RW; /* local HBM region far from partition or remote XGMI GPU or * system memory @@ -1231,7 +1222,7 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange, pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM; pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0; - pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags); + pte_flags |= amdgpu_gem_va_map_flags(node->adev, mapping_flags); return pte_flags; } @@ -1338,7 +1329,7 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n", last_start, prange->start + i, last_domain ? "GPU" : "CPU"); - pte_flags = svm_range_get_pte_flags(adev, prange, last_domain); + pte_flags = svm_range_get_pte_flags(pdd->dev, prange, last_domain); if (readonly) pte_flags &= ~AMDGPU_PTE_WRITEABLE; @@ -1347,6 +1338,9 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, pte_flags); + /* TODO: we still need to determine the vm_manager.vram_base_offset based on + * the memory partition. + */ r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL, last_start, prange->start + i, pte_flags, @@ -1384,16 +1378,14 @@ svm_range_map_to_gpus(struct svm_range *prange, unsigned long offset, unsigned long *bitmap, bool wait, bool flush_tlb) { struct kfd_process_device *pdd; - struct amdgpu_device *bo_adev; + struct amdgpu_device *bo_adev = NULL; struct kfd_process *p; struct dma_fence *fence = NULL; uint32_t gpuidx; int r = 0; if (prange->svm_bo && prange->ttm_res) - bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev); - else - bo_adev = NULL; + bo_adev = prange->svm_bo->node->adev; p = container_of(prange->svms, struct kfd_process, svms); for_each_set_bit(gpuidx, bitmap, MAX_GPU_INSTANCE) { @@ -2526,17 +2518,17 @@ svm_range_from_addr(struct svm_range_list *svms, unsigned long addr, */ static int32_t svm_range_best_restore_location(struct svm_range *prange, - struct amdgpu_device *adev, + struct kfd_node *node, int32_t *gpuidx) { - struct amdgpu_device *bo_adev, *preferred_adev; + struct kfd_node *bo_node, *preferred_node; struct kfd_process *p; uint32_t gpuid; int r; p = container_of(prange->svms, struct kfd_process, svms); - r = kfd_process_gpuid_from_adev(p, adev, &gpuid, gpuidx); + r = kfd_process_gpuid_from_node(p, node, &gpuid, gpuidx); if (r < 0) { pr_debug("failed to get gpuid from kgd\n"); return -1; @@ -2546,9 +2538,8 @@ svm_range_best_restore_location(struct svm_range *prange, prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) { return prange->preferred_loc; } else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) { - preferred_adev = svm_range_get_adev_by_id(prange, - prange->preferred_loc); - if (amdgpu_xgmi_same_hive(adev, preferred_adev)) + preferred_node = svm_range_get_node_by_id(prange, prange->preferred_loc); + if (preferred_node && svm_nodes_in_same_hive(node, preferred_node)) return prange->preferred_loc; /* fall through */ } @@ -2560,8 +2551,8 @@ svm_range_best_restore_location(struct svm_range *prange, if (!prange->actual_loc) return 0; - bo_adev = svm_range_get_adev_by_id(prange, prange->actual_loc); - if (amdgpu_xgmi_same_hive(adev, bo_adev)) + bo_node = svm_range_get_node_by_id(prange, prange->actual_loc); + if (bo_node && svm_nodes_in_same_hive(node, bo_node)) return prange->actual_loc; else return 0; @@ -2678,7 +2669,7 @@ svm_range_check_vm_userptr(struct kfd_process *p, uint64_t start, uint64_t last, } static struct -svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev, +svm_range *svm_range_create_unregistered_range(struct kfd_node *node, struct kfd_process *p, struct mm_struct *mm, int64_t addr) @@ -2713,7 +2704,7 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev, pr_debug("Failed to create prange in address [0x%llx]\n", addr); return NULL; } - if (kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx)) { + if (kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx)) { pr_debug("failed to get gpuid from kgd\n"); svm_range_free(prange, true); return NULL; @@ -2767,7 +2758,7 @@ static bool svm_range_skip_recover(struct svm_range *prange) } static void -svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p, +svm_range_count_fault(struct kfd_node *node, struct kfd_process *p, int32_t gpuidx) { struct kfd_process_device *pdd; @@ -2780,7 +2771,7 @@ svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p, uint32_t gpuid; int r; - r = kfd_process_gpuid_from_adev(p, adev, &gpuid, &gpuidx); + r = kfd_process_gpuid_from_node(p, node, &gpuid, &gpuidx); if (r < 0) return; } @@ -2808,6 +2799,7 @@ svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, + uint32_t client_id, uint32_t node_id, uint64_t addr, bool write_fault) { struct mm_struct *mm = NULL; @@ -2815,6 +2807,7 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, struct svm_range *prange; struct kfd_process *p; ktime_t timestamp = ktime_get_boottime(); + struct kfd_node *node; int32_t best_loc; int32_t gpuidx = MAX_GPU_INSTANCE; bool write_locked = false; @@ -2858,6 +2851,13 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, goto out; } + node = kfd_node_by_irq_ids(adev, node_id, client_id); + if (!node) { + pr_debug("kfd node does not exist node_id: %d, client_id: %d\n", node_id, + client_id); + r = -EFAULT; + goto out; + } mmap_read_lock(mm); retry_write_locked: mutex_lock(&svms->lock); @@ -2876,7 +2876,7 @@ retry_write_locked: write_locked = true; goto retry_write_locked; } - prange = svm_range_create_unregistered_range(adev, p, mm, addr); + prange = svm_range_create_unregistered_range(node, p, mm, addr); if (!prange) { pr_debug("failed to create unregistered range svms 0x%p address [0x%llx]\n", svms, addr); @@ -2891,7 +2891,7 @@ retry_write_locked: mutex_lock(&prange->migrate_mutex); if (svm_range_skip_recover(prange)) { - amdgpu_gmc_filter_faults_remove(adev, addr, pasid); + amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); r = 0; goto out_unlock_range; } @@ -2922,7 +2922,7 @@ retry_write_locked: goto out_unlock_range; } - best_loc = svm_range_best_restore_location(prange, adev, &gpuidx); + best_loc = svm_range_best_restore_location(prange, node, &gpuidx); if (best_loc == -1) { pr_debug("svms %p failed get best restore loc [0x%lx 0x%lx]\n", svms, prange->start, prange->last); @@ -2981,7 +2981,7 @@ out_unlock_svms: mutex_unlock(&svms->lock); mmap_read_unlock(mm); - svm_range_count_fault(adev, p, gpuidx); + svm_range_count_fault(node, p, gpuidx); mmput(mm); out: @@ -2989,7 +2989,7 @@ out: if (r == -EAGAIN) { pr_debug("recover vm fault later\n"); - amdgpu_gmc_filter_faults_remove(adev, addr, pasid); + amdgpu_gmc_filter_faults_remove(node->adev, addr, pasid); r = 0; } return r; @@ -3231,7 +3231,7 @@ svm_range_best_prefetch_location(struct svm_range *prange) DECLARE_BITMAP(bitmap, MAX_GPU_INSTANCE); uint32_t best_loc = prange->prefetch_loc; struct kfd_process_device *pdd; - struct amdgpu_device *bo_adev; + struct kfd_node *bo_node; struct kfd_process *p; uint32_t gpuidx; @@ -3240,9 +3240,9 @@ svm_range_best_prefetch_location(struct svm_range *prange) if (!best_loc || best_loc == KFD_IOCTL_SVM_LOCATION_UNDEFINED) goto out; - bo_adev = svm_range_get_adev_by_id(prange, best_loc); - if (!bo_adev) { - WARN_ONCE(1, "failed to get device by id 0x%x\n", best_loc); + bo_node = svm_range_get_node_by_id(prange, best_loc); + if (!bo_node) { + WARN_ONCE(1, "failed to get valid kfd node at id%x\n", best_loc); best_loc = 0; goto out; } @@ -3260,10 +3260,10 @@ svm_range_best_prefetch_location(struct svm_range *prange) continue; } - if (pdd->dev->adev == bo_adev) + if (pdd->dev->adev == bo_node->adev) continue; - if (!amdgpu_xgmi_same_hive(pdd->dev->adev, bo_adev)) { + if (!svm_nodes_in_same_hive(pdd->dev, bo_node)) { best_loc = 0; break; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index 7a33b93f9df6..a165c73b40b2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -48,6 +48,7 @@ struct svm_range_bo { struct work_struct eviction_work; uint32_t evicting; struct work_struct release_work; + struct kfd_node *node; }; enum svm_work_list_ops { @@ -163,16 +164,17 @@ int svm_ioctl(struct kfd_process *p, enum kfd_ioctl_svm_op op, uint64_t start, struct svm_range *svm_range_from_addr(struct svm_range_list *svms, unsigned long addr, struct svm_range **parent); -struct amdgpu_device *svm_range_get_adev_by_id(struct svm_range *prange, - uint32_t id); -int svm_range_vram_node_new(struct amdgpu_device *adev, - struct svm_range *prange, bool clear); +struct kfd_node *svm_range_get_node_by_id(struct svm_range *prange, + uint32_t gpu_id); +int svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, + bool clear); void svm_range_vram_node_free(struct svm_range *prange); int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, unsigned long addr, struct svm_range *parent, struct svm_range *prange); -int svm_range_restore_pages(struct amdgpu_device *adev, - unsigned int pasid, uint64_t addr, bool write_fault); +int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, + uint32_t client_id, uint32_t node_id, uint64_t addr, + bool write_fault); int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence); void svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange, struct mm_struct *mm, @@ -192,7 +194,7 @@ int kfd_criu_restore_svm(struct kfd_process *p, uint64_t max_priv_data_size); int kfd_criu_resume_svm(struct kfd_process *p); struct kfd_process_device * -svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev); +svm_range_get_pdd_by_node(struct svm_range *prange, struct kfd_node *node); void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_struct *mm); /* SVM API and HMM page migration work together, device memory type @@ -219,8 +221,9 @@ static inline void svm_range_list_fini(struct kfd_process *p) } static inline int svm_range_restore_pages(struct amdgpu_device *adev, - unsigned int pasid, uint64_t addr, - bool write_fault) + unsigned int pasid, + uint32_t client_id, uint32_t node_id, + uint64_t addr, bool write_fault) { return -EFAULT; } -- cgit v1.2.3 From d6e924ad85a0cebc9e39eb956a23386ce32cc9f9 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 9 Aug 2022 14:56:53 -0400 Subject: drm/amdkfd: Update SMI events for GFX9.4.3 On GFX 9.4.3, there can be multiple KFD nodes. As a result, SMI events for SVM, queue evict/restore should be raised for each node independently. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 16 ++++++------ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 6 ++--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 38 ++++++++++++++--------------- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 14 +++++------ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 6 ++--- 5 files changed, 40 insertions(+), 40 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index e7e5abc32c84..42e599912e52 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -423,9 +423,9 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, migrate.dst = migrate.src + npages; scratch = (dma_addr_t *)(migrate.dst + npages); - kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, + kfd_smi_event_migration_start(node, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - 0, adev->kfd.dev->nodes[0]->id, prange->prefetch_loc, + 0, node->id, prange->prefetch_loc, prange->preferred_loc, trigger); r = migrate_vma_setup(&migrate); @@ -456,9 +456,9 @@ svm_migrate_vma_to_vram(struct kfd_node *node, struct svm_range *prange, svm_migrate_copy_done(adev, mfence); migrate_vma_finalize(&migrate); - kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, + kfd_smi_event_migration_end(node, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - 0, adev->kfd.dev->nodes[0]->id, trigger); + 0, node->id, trigger); svm_range_dma_unmap(adev->dev, scratch, 0, npages); svm_range_free_dma_mappings(prange); @@ -702,9 +702,9 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, migrate.fault_page = fault_page; scratch = (dma_addr_t *)(migrate.dst + npages); - kfd_smi_event_migration_start(adev->kfd.dev, p->lead_thread->pid, + kfd_smi_event_migration_start(node, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - adev->kfd.dev->nodes[0]->id, 0, prange->prefetch_loc, + node->id, 0, prange->prefetch_loc, prange->preferred_loc, trigger); r = migrate_vma_setup(&migrate); @@ -738,9 +738,9 @@ svm_migrate_vma_to_ram(struct kfd_node *node, struct svm_range *prange, svm_migrate_copy_done(adev, mfence); migrate_vma_finalize(&migrate); - kfd_smi_event_migration_end(adev->kfd.dev, p->lead_thread->pid, + kfd_smi_event_migration_end(node, p->lead_thread->pid, start >> PAGE_SHIFT, end >> PAGE_SHIFT, - adev->kfd.dev->nodes[0]->id, 0, trigger); + node->id, 0, trigger); svm_range_dma_unmap(adev->dev, scratch, 0, npages); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 666815b227a8..a6ff57f11472 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1817,7 +1817,7 @@ int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger) for (i = 0; i < p->n_pdds; i++) { struct kfd_process_device *pdd = p->pdds[i]; - kfd_smi_event_queue_eviction(pdd->dev->kfd, p->lead_thread->pid, + kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid, trigger); r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm, @@ -1845,7 +1845,7 @@ fail: if (n_evicted == 0) break; - kfd_smi_event_queue_restore(pdd->dev->kfd, p->lead_thread->pid); + kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid); if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm, &pdd->qpd)) @@ -1866,7 +1866,7 @@ int kfd_process_restore_queues(struct kfd_process *p) for (i = 0; i < p->n_pdds; i++) { struct kfd_process_device *pdd = p->pdds[i]; - kfd_smi_event_queue_restore(pdd->dev->kfd, p->lead_thread->pid); + kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid); r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm, &pdd->qpd); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c index b703da59e067..d9953c2b2661 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c @@ -250,58 +250,58 @@ void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid) task_info.pid, task_info.task_name); } -void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_page_fault_start(struct kfd_node *node, pid_t pid, unsigned long address, bool write_fault, ktime_t ts) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_PAGE_FAULT_START, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_PAGE_FAULT_START, "%lld -%d @%lx(%x) %c\n", ktime_to_ns(ts), pid, - address, dev->nodes[0]->id, write_fault ? 'W' : 'R'); + address, node->id, write_fault ? 'W' : 'R'); } -void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_page_fault_end(struct kfd_node *node, pid_t pid, unsigned long address, bool migration) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_PAGE_FAULT_END, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_PAGE_FAULT_END, "%lld -%d @%lx(%x) %c\n", ktime_get_boottime_ns(), - pid, address, dev->nodes[0]->id, migration ? 'M' : 'U'); + pid, address, node->id, migration ? 'M' : 'U'); } -void kfd_smi_event_migration_start(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_migration_start(struct kfd_node *node, pid_t pid, unsigned long start, unsigned long end, uint32_t from, uint32_t to, uint32_t prefetch_loc, uint32_t preferred_loc, uint32_t trigger) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_MIGRATE_START, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_MIGRATE_START, "%lld -%d @%lx(%lx) %x->%x %x:%x %d\n", ktime_get_boottime_ns(), pid, start, end - start, from, to, prefetch_loc, preferred_loc, trigger); } -void kfd_smi_event_migration_end(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_migration_end(struct kfd_node *node, pid_t pid, unsigned long start, unsigned long end, uint32_t from, uint32_t to, uint32_t trigger) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_MIGRATE_END, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_MIGRATE_END, "%lld -%d @%lx(%lx) %x->%x %d\n", ktime_get_boottime_ns(), pid, start, end - start, from, to, trigger); } -void kfd_smi_event_queue_eviction(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_queue_eviction(struct kfd_node *node, pid_t pid, uint32_t trigger) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_QUEUE_EVICTION, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_QUEUE_EVICTION, "%lld -%d %x %d\n", ktime_get_boottime_ns(), pid, - dev->nodes[0]->id, trigger); + node->id, trigger); } -void kfd_smi_event_queue_restore(struct kfd_dev *dev, pid_t pid) +void kfd_smi_event_queue_restore(struct kfd_node *node, pid_t pid) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_QUEUE_RESTORE, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_QUEUE_RESTORE, "%lld -%d %x\n", ktime_get_boottime_ns(), pid, - dev->nodes[0]->id); + node->id); } void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm) @@ -324,13 +324,13 @@ void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm) kfd_unref_process(p); } -void kfd_smi_event_unmap_from_gpu(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_unmap_from_gpu(struct kfd_node *node, pid_t pid, unsigned long address, unsigned long last, uint32_t trigger) { - kfd_smi_event_add(pid, dev->nodes[0], KFD_SMI_EVENT_UNMAP_FROM_GPU, + kfd_smi_event_add(pid, node, KFD_SMI_EVENT_UNMAP_FROM_GPU, "%lld -%d @%lx(%lx) %x %d\n", ktime_get_boottime_ns(), - pid, address, last - address + 1, dev->nodes[0]->id, trigger); + pid, address, last - address + 1, node->id, trigger); } int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h index 59cd089f80d1..fa95c2dfd587 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h @@ -29,24 +29,24 @@ void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid); void kfd_smi_event_update_thermal_throttling(struct kfd_node *dev, uint64_t throttle_bitmask); void kfd_smi_event_update_gpu_reset(struct kfd_node *dev, bool post_reset); -void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_page_fault_start(struct kfd_node *node, pid_t pid, unsigned long address, bool write_fault, ktime_t ts); -void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_page_fault_end(struct kfd_node *node, pid_t pid, unsigned long address, bool migration); -void kfd_smi_event_migration_start(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_migration_start(struct kfd_node *node, pid_t pid, unsigned long start, unsigned long end, uint32_t from, uint32_t to, uint32_t prefetch_loc, uint32_t preferred_loc, uint32_t trigger); -void kfd_smi_event_migration_end(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_migration_end(struct kfd_node *node, pid_t pid, unsigned long start, unsigned long end, uint32_t from, uint32_t to, uint32_t trigger); -void kfd_smi_event_queue_eviction(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_queue_eviction(struct kfd_node *node, pid_t pid, uint32_t trigger); -void kfd_smi_event_queue_restore(struct kfd_dev *dev, pid_t pid); +void kfd_smi_event_queue_restore(struct kfd_node *node, pid_t pid); void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm); -void kfd_smi_event_unmap_from_gpu(struct kfd_dev *dev, pid_t pid, +void kfd_smi_event_unmap_from_gpu(struct kfd_node *node, pid_t pid, unsigned long address, unsigned long last, uint32_t trigger); #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 639831fbb6ca..0dafbbe954ca 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1274,7 +1274,7 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, return -EINVAL; } - kfd_smi_event_unmap_from_gpu(pdd->dev->kfd, p->lead_thread->pid, + kfd_smi_event_unmap_from_gpu(pdd->dev, p->lead_thread->pid, start, last, trigger); r = svm_range_unmap_from_gpu(pdd->dev->adev, @@ -2934,7 +2934,7 @@ retry_write_locked: svms, prange->start, prange->last, best_loc, prange->actual_loc); - kfd_smi_event_page_fault_start(adev->kfd.dev, p->lead_thread->pid, addr, + kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr, write_fault, timestamp); if (prange->actual_loc != best_loc) { @@ -2972,7 +2972,7 @@ retry_write_locked: pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n", r, svms, prange->start, prange->last); - kfd_smi_event_page_fault_end(adev->kfd.dev, p->lead_thread->pid, addr, + kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr, migration); out_unlock_range: -- cgit v1.2.3 From f5fe7edfd6ce62cd23fbd707e7f9fe0f56a45e94 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Fri, 30 Sep 2022 09:16:21 -0400 Subject: drm/amdkfd: Update interrupt handling for GFX9.4.3 Update interrupt handling in CPX mode for GFX9.4.3 by using the VMID space instead of SDMA client id to determine if an interrupt should be processed by a KFD node. This is especially needed for handling retry faults from MMHUB. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 7 +++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 4 ++-- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 16 ++++++---------- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 8 ++++---- drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 2 +- 6 files changed, 19 insertions(+), 20 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index c390b2856cc9..22a900f298f2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2434,6 +2434,9 @@ void amdgpu_vm_set_task_info(struct amdgpu_vm *vm) * amdgpu_vm_handle_fault - graceful handling of VM faults. * @adev: amdgpu device pointer * @pasid: PASID of the VM + * @vmid: VMID, only used for GFX 9.4.3. + * @node_id: Node_id received in IH cookie. Only applicable for + * GFX 9.4.3. * @addr: Address of the fault * @write_fault: true is write fault, false is read fault * @@ -2441,7 +2444,7 @@ void amdgpu_vm_set_task_info(struct amdgpu_vm *vm) * shouldn't be reported any more. */ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, - u32 client_id, u32 node_id, uint64_t addr, + u32 vmid, u32 node_id, uint64_t addr, bool write_fault) { bool is_compute_context = false; @@ -2466,7 +2469,7 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, addr /= AMDGPU_GPU_PAGE_SIZE; - if (is_compute_context && !svm_range_restore_pages(adev, pasid, client_id, + if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid, node_id, addr, write_fault)) { amdgpu_bo_unref(&root); return true; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h index dbab31647186..8add5f5eb92a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h @@ -455,7 +455,7 @@ void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev); void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid, struct amdgpu_task_info *task_info); bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, - u32 client_id, u32 node_id, uint64_t addr, + u32 vmid, u32 node_id, uint64_t addr, bool write_fault); void amdgpu_vm_set_task_info(struct amdgpu_vm *vm); diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index c5752a349f3d..f2814270da40 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -587,7 +587,7 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, cam_index = entry->src_data[2] & 0x3ff; - ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->client_id, node_id, + ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id, addr, write_fault); WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index); if (ret) @@ -610,7 +610,7 @@ static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev, /* Try to handle the recoverable page faults by filling page * tables */ - if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->client_id, node_id, + if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id, addr, write_fault)) return 1; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index df372de6b056..fb3cf2c51da8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1073,18 +1073,14 @@ struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); struct kfd_node *kfd_device_by_id(uint32_t gpu_id); struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev); struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev); -static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t client_id, - uint32_t node_id) +static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id, + uint32_t vmid) { - if ((node->interrupt_bitmap & (0x1U << node_id)) || - ((node_id % 4) == 0 && - (node->interrupt_bitmap >> 16) & (0x1U << client_id))) - return true; - - return false; + return (node->interrupt_bitmap & (1 << node_id)) != 0 && + (node->compute_vmid_bitmap & (1 << vmid)) != 0; } static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev, - uint32_t client_id, uint32_t node_id) { + uint32_t node_id, uint32_t vmid) { struct kfd_dev *dev = adev->kfd.dev; uint32_t i; @@ -1092,7 +1088,7 @@ static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev, return dev->nodes[0]; for (i = 0; i < dev->num_nodes; i++) - if (kfd_irq_is_from_node(dev->nodes[i], client_id, node_id)) + if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid)) return dev->nodes[i]; return NULL; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 0dafbbe954ca..5d6e02559d8e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -2799,7 +2799,7 @@ svm_fault_allowed(struct vm_area_struct *vma, bool write_fault) int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, - uint32_t client_id, uint32_t node_id, + uint32_t vmid, uint32_t node_id, uint64_t addr, bool write_fault) { struct mm_struct *mm = NULL; @@ -2851,10 +2851,10 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, goto out; } - node = kfd_node_by_irq_ids(adev, node_id, client_id); + node = kfd_node_by_irq_ids(adev, node_id, vmid); if (!node) { - pr_debug("kfd node does not exist node_id: %d, client_id: %d\n", node_id, - client_id); + pr_debug("kfd node does not exist node_id: %d, vmid: %d\n", node_id, + vmid); r = -EFAULT; goto out; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index a165c73b40b2..5116786718b6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -173,7 +173,7 @@ int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm, unsigned long addr, struct svm_range *parent, struct svm_range *prange); int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, - uint32_t client_id, uint32_t node_id, uint64_t addr, + uint32_t vmid, uint32_t node_id, uint64_t addr, bool write_fault); int svm_range_schedule_evict_svm_bo(struct amdgpu_amdkfd_fence *fence); void svm_range_add_list_work(struct svm_range_list *svms, -- cgit v1.2.3 From 8e7fd19380f9187dae3ad18a61793b1752dfa097 Mon Sep 17 00:00:00 2001 From: Lijo Lazar Date: Wed, 16 Nov 2022 17:15:47 +0530 Subject: drm/amdgpu: Switch to SOC partition funcs For GFXv9.4.3, use SOC level partition switch implementation rather than keeping them at GFX IP level. Change the exisiting implementation in GFX IP for keeping partition mode and restrict it to only GFX related switch. Signed-off-by: Lijo Lazar Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 31 ++---------- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h | 5 -- .../gpu/drm/amd/amdgpu/aqua_vanjaram_reg_init.c | 4 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 59 ++++------------------ drivers/gpu/drm/amd/amdkfd/kfd_device.c | 8 +-- 5 files changed, 20 insertions(+), 87 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 5ff49737d7c6..f895a4b8ca0d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -28,6 +28,7 @@ #include "amdgpu_gfx.h" #include "amdgpu_rlc.h" #include "amdgpu_ras.h" +#include "amdgpu_xcp.h" /* delay 0.1 second to enable gfx off feature */ #define GFX_OFF_DELAY_ENABLE msecs_to_jiffies(100) @@ -1170,10 +1171,10 @@ static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev, { struct drm_device *ddev = dev_get_drvdata(dev); struct amdgpu_device *adev = drm_to_adev(ddev); - enum amdgpu_gfx_partition mode; + int mode; char *partition_mode; - mode = adev->gfx.funcs->query_partition_mode(adev); + mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr); switch (mode) { case AMDGPU_SPX_PARTITION_MODE: @@ -1254,31 +1255,7 @@ static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev, return -EINVAL; } - if (!adev->kfd.init_complete) - return -EPERM; - - mutex_lock(&adev->gfx.partition_mutex); - - if (mode == adev->gfx.funcs->query_partition_mode(adev)) - goto out; - - ret = amdgpu_amdkfd_check_and_lock_kfd(adev); - if (ret) - goto out; - - amdgpu_amdkfd_device_fini_sw(adev); - - adev->gfx.funcs->switch_partition_mode(adev, mode); - - amdgpu_amdkfd_device_probe(adev); - amdgpu_amdkfd_device_init(adev); - /* If KFD init failed, return failure */ - if (!adev->kfd.init_complete) - ret = -EIO; - - amdgpu_amdkfd_unlock_kfd(adev); -out: - mutex_unlock(&adev->gfx.partition_mutex); + ret = amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, mode); if (ret) return ret; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h index 8be4ab50b171..2287768ed141 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h @@ -278,11 +278,7 @@ struct amdgpu_gfx_funcs { (*query_partition_mode)(struct amdgpu_device *adev); enum amdgpu_memory_partition (*query_mem_partition_mode)(struct amdgpu_device *adev); - int (*switch_partition_mode)(struct amdgpu_device *adev, - enum amdgpu_gfx_partition mode); - - int (*switch_gfx_partition_mode)(struct amdgpu_device *adev, int num_xccs_per_xcp); }; @@ -416,7 +412,6 @@ struct amdgpu_gfx { bool cp_gfx_shadow; /* for gfx11 */ - enum amdgpu_gfx_partition partition_mode; uint16_t xcc_mask; enum amdgpu_memory_partition mem_partition_mode; uint32_t num_xcc_per_xcp; diff --git a/drivers/gpu/drm/amd/amdgpu/aqua_vanjaram_reg_init.c b/drivers/gpu/drm/amd/amdgpu/aqua_vanjaram_reg_init.c index a9de229a2828..bbcdececfd2f 100644 --- a/drivers/gpu/drm/amd/amdgpu/aqua_vanjaram_reg_init.c +++ b/drivers/gpu/drm/amd/amdgpu/aqua_vanjaram_reg_init.c @@ -307,8 +307,8 @@ static int aqua_vanjaram_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, goto unlock; num_xcc_per_xcp = __aqua_vanjaram_get_xcc_per_xcp(xcp_mgr, mode); - if (adev->gfx.funcs->switch_gfx_partition_mode) - adev->gfx.funcs->switch_gfx_partition_mode(xcp_mgr->adev, + if (adev->gfx.funcs->switch_partition_mode) + adev->gfx.funcs->switch_partition_mode(xcp_mgr->adev, num_xcc_per_xcp); if (adev->nbio.funcs->set_compute_partition_mode) diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c index 93a0baa4515c..d684037a7a5d 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c @@ -38,6 +38,7 @@ #include "gc/gc_9_4_3_sh_mask.h" #include "gfx_v9_4_3.h" +#include "amdgpu_xcp.h" MODULE_FIRMWARE("amdgpu/gc_9_4_3_mec.bin"); MODULE_FIRMWARE("amdgpu/gc_9_4_3_rlc.bin"); @@ -614,61 +615,23 @@ gfx_v9_4_3_query_memory_partition(struct amdgpu_device *adev) return mode; } -static enum amdgpu_gfx_partition -gfx_v9_4_3_query_compute_partition(struct amdgpu_device *adev) -{ - enum amdgpu_gfx_partition mode = adev->gfx.partition_mode; - - if (adev->nbio.funcs->get_compute_partition_mode) - mode = adev->nbio.funcs->get_compute_partition_mode(adev); - - return mode; -} - static int gfx_v9_4_3_switch_compute_partition(struct amdgpu_device *adev, - enum amdgpu_gfx_partition mode) + int num_xccs_per_xcp) { + int i, num_xcc; u32 tmp = 0; - int num_xcc_per_partition, i, num_xcc; num_xcc = NUM_XCC(adev->gfx.xcc_mask); - switch (mode) { - case AMDGPU_SPX_PARTITION_MODE: - num_xcc_per_partition = num_xcc; - break; - case AMDGPU_DPX_PARTITION_MODE: - num_xcc_per_partition = num_xcc / 2; - break; - case AMDGPU_TPX_PARTITION_MODE: - num_xcc_per_partition = num_xcc / 3; - break; - case AMDGPU_QPX_PARTITION_MODE: - num_xcc_per_partition = num_xcc / 4; - break; - case AMDGPU_CPX_PARTITION_MODE: - num_xcc_per_partition = 1; - break; - default: - return -EINVAL; - } - - /* TODO: - * Stop user queues and threads, and make sure GPU is empty of work. - */ for (i = 0; i < num_xcc; i++) { tmp = REG_SET_FIELD(tmp, CP_HYP_XCP_CTL, NUM_XCC_IN_XCP, - num_xcc_per_partition); + num_xccs_per_xcp); tmp = REG_SET_FIELD(tmp, CP_HYP_XCP_CTL, VIRTUAL_XCC_ID, - i % num_xcc_per_partition); + i % num_xccs_per_xcp); WREG32_SOC15(GC, GET_INST(GC, i), regCP_HYP_XCP_CTL, tmp); } - if (adev->nbio.funcs->set_compute_partition_mode) - adev->nbio.funcs->set_compute_partition_mode(adev, mode); - - adev->gfx.num_xcc_per_xcp = num_xcc_per_partition; - adev->gfx.partition_mode = mode; + adev->gfx.num_xcc_per_xcp = num_xccs_per_xcp; return 0; } @@ -680,7 +643,6 @@ static const struct amdgpu_gfx_funcs gfx_v9_4_3_gfx_funcs = { .read_wave_sgprs = &gfx_v9_4_3_read_wave_sgprs, .read_wave_vgprs = &gfx_v9_4_3_read_wave_vgprs, .select_me_pipe_q = &gfx_v9_4_3_select_me_pipe_q, - .query_partition_mode = &gfx_v9_4_3_query_compute_partition, .switch_partition_mode = &gfx_v9_4_3_switch_compute_partition, .query_mem_partition_mode = &gfx_v9_4_3_query_memory_partition, }; @@ -1899,10 +1861,6 @@ static int gfx_v9_4_3_xcc_cp_resume(struct amdgpu_device *adev, int xcc_id) return r; } - if (adev->gfx.partition_mode == AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE) - gfx_v9_4_3_switch_compute_partition(adev, - amdgpu_user_partt_mode); - /* set the virtual and physical id based on partition_mode */ gfx_v9_4_3_xcc_program_xcc_id(adev, xcc_id); @@ -1931,6 +1889,9 @@ static int gfx_v9_4_3_cp_resume(struct amdgpu_device *adev) { int r, i, num_xcc; + if (amdgpu_xcp_query_partition_mode(adev->xcp_mgr) == AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE) + amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, amdgpu_user_partt_mode); + num_xcc = NUM_XCC(adev->gfx.xcc_mask); for (i = 0; i < num_xcc; i++) { r = gfx_v9_4_3_xcc_cp_resume(adev, i); @@ -2146,8 +2107,6 @@ static int gfx_v9_4_3_early_init(void *handle) num_xcc = NUM_XCC(adev->gfx.xcc_mask); - adev->gfx.partition_mode = AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE; - adev->gfx.num_compute_rings = min(amdgpu_gfx_get_num_kcq(adev), AMDGPU_MAX_COMPUTE_RINGS); gfx_v9_4_3_set_kiq_pm4_funcs(adev); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 293787290e36..7a963d0a34e2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -34,6 +34,7 @@ #include "kfd_smi_events.h" #include "kfd_migrate.h" #include "amdgpu.h" +#include "amdgpu_xcp.h" #define MQD_SIZE_ALIGNED 768 @@ -592,7 +593,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, struct kfd_node *node; uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; unsigned int max_proc_per_quantum; - int num_xcd; + int num_xcd, partition_mode; kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, KGD_ENGINE_MEC1); @@ -644,8 +645,9 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, * If the VMID range changes for GFX9.4.3, then this code MUST be * revisited. */ + partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr); if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && - kfd->adev->gfx.partition_mode == AMDGPU_CPX_PARTITION_MODE && + partition_mode == AMDGPU_CPX_PARTITION_MODE && kfd->num_nodes != 1) { vmid_num_kfd /= 2; first_vmid_kfd = last_vmid_kfd + 1 - vmid_num_kfd*2; @@ -761,7 +763,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, node->start_xcc_id = node->num_xcc_per_node * i; if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && - kfd->adev->gfx.partition_mode == AMDGPU_CPX_PARTITION_MODE && + partition_mode == AMDGPU_CPX_PARTITION_MODE && kfd->num_nodes != 1) { /* For GFX9.4.3 and CPX mode, first XCD gets VMID range * 4-9 and second XCD gets VMID range 10-15. -- cgit v1.2.3 From 1794e9d7e78cb52605234d0ddc3f46084937f4e9 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Thu, 8 Dec 2022 12:08:17 -0500 Subject: drm/amdkfd: Update packet manager for GFX9.4.3 In GFX 9.4.3, there can be more than 8 SDMA engines. As a result, extended_engine_sel and engine_sel fields in MAP_QUEUES packet need to be updated to allow correct mapping of SDMA queues to these SDMA engines. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 16 +++++++++++++--- drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h | 3 ++- 2 files changed, 15 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 54d7d4665ad2..44cf3a5f6fdb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -225,9 +225,19 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, packet->bitfields2.engine_sel = q->properties.sdma_engine_id + engine_sel__mes_map_queues__sdma0_vi; else { - packet->bitfields2.extended_engine_sel = - extended_engine_sel__mes_map_queues__sdma0_to_7_sel; - packet->bitfields2.engine_sel = q->properties.sdma_engine_id; + /* + * For GFX9.4.3, SDMA engine id can be greater than 8. + * For such cases, set extended_engine_sel to 2 and + * ensure engine_sel lies between 0-7. + */ + if (q->properties.sdma_engine_id >= 8) + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_map_queues__sdma8_to_15_sel; + else + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_map_queues__sdma0_to_7_sel; + + packet->bitfields2.engine_sel = q->properties.sdma_engine_id % 8; } break; default: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index a666710ed403..2ad708c64012 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -263,7 +263,8 @@ enum mes_map_queues_engine_sel_enum { enum mes_map_queues_extended_engine_sel_enum { extended_engine_sel__mes_map_queues__legacy_engine_sel = 0, - extended_engine_sel__mes_map_queues__sdma0_to_7_sel = 1 + extended_engine_sel__mes_map_queues__sdma0_to_7_sel = 1, + extended_engine_sel__mes_map_queues__sdma8_to_15_sel = 2 }; struct pm4_mes_map_queues { -- cgit v1.2.3 From c2d43918a14f7b0f04932f5a45728e0fe8161da0 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Fri, 9 Dec 2022 09:03:01 -0500 Subject: drm/amdkfd: Setup current_logical_xcc_id in MQD Setup rolling current_logical_xcc_id in MQD for GFX9.4.3 to ensure each queue starts at a different place and prevent hotspotting issues. Also, remove updating current_logical_xcc_id during queue update. Suggested-by: Joseph Greathouse Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 3 +++ drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 13 +++++-------- 2 files changed, 8 insertions(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index b11c474d4067..cd4383bb207f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -256,6 +256,9 @@ struct device_queue_manager { struct work_struct hw_exception_work; struct kfd_mem_obj hiq_sdma_mqd; bool sched_running; + + /* used for GFX 9.4.3 only */ + uint32_t current_logical_xcc_start; }; void device_queue_manager_init_cik( diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 2e2a0f8586f7..c781314b213c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -32,6 +32,7 @@ #include "gc/gc_9_0_sh_mask.h" #include "sdma0/sdma0_4_0_sh_mask.h" #include "amdgpu_amdkfd.h" +#include "kfd_device_queue_manager.h" static void update_mqd(struct mqd_manager *mm, void *mqd, struct queue_properties *q, @@ -569,6 +570,7 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, uint64_t xcc_gart_addr = 0; uint64_t xcc_ctx_save_restore_area_address; uint64_t offset = mm->mqd_stride(mm, q); + uint32_t local_xcc_start = mm->dev->dqm->current_logical_xcc_start++; memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj)); for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { @@ -596,18 +598,17 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, if (q->format == KFD_QUEUE_FORMAT_AQL) { m->compute_tg_chunk_size = 1; + m->compute_current_logic_xcc_id = + (local_xcc_start + xcc) % + mm->dev->num_xcc_per_node; switch (xcc) { case 0: /* Master XCC */ m->cp_hqd_pq_control &= ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; - m->compute_current_logic_xcc_id = - mm->dev->num_xcc_per_node - 1; break; default: - m->compute_current_logic_xcc_id = - xcc - 1; break; } } else { @@ -642,12 +643,8 @@ static void update_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, /* Master XCC */ m->cp_hqd_pq_control &= ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK; - m->compute_current_logic_xcc_id = - mm->dev->num_xcc_per_node - 1; break; default: - m->compute_current_logic_xcc_id = - xcc - 1; break; } m->compute_tg_chunk_size = 1; -- cgit v1.2.3 From 85b45b60722f506322393320bb6cc195378f2e4f Mon Sep 17 00:00:00 2001 From: Amber Lin Date: Mon, 28 Nov 2022 11:26:02 -0500 Subject: amd/amdgpu: Set MTYPE_UC for access over PCIe For GFX v9_4_3, set MTYPE_UC for memory access over PCIe. v4 - add missing indentation pointed out by Felix and add his reviewed-by tag. v3 - add missing logic for the svm path. v2 - add amdgpu_xgmi_same_hive to separate access over xgmi from pcie Reviewed-by: Felix Kuehling Signed-off-by: Amber Lin Signed-off-by: Rajneesh Bhardwaj Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 29 +++++++++++++++++++++-------- 1 file changed, 21 insertions(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 5d6e02559d8e..6daba0582bf3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1195,16 +1195,29 @@ svm_range_get_pte_flags(struct kfd_node *node, //e.g. NPS4. Current approch is only applicable without memory //partitions. snoop = true; - if (uncached) + if (uncached) { mapping_flags |= AMDGPU_VM_MTYPE_UC; - /* local HBM region close to partition*/ - else if (bo_node == node) - mapping_flags |= AMDGPU_VM_MTYPE_RW; - /* local HBM region far from partition or remote XGMI GPU or - * system memory - */ - else + } else if (domain == SVM_RANGE_VRAM_DOMAIN) { + /* local HBM region close to partition with a workaround + * for Endpoint systems. + */ + if (bo_node == node) + mapping_flags |= + (node->adev->flags & AMD_IS_APU) ? + AMDGPU_VM_MTYPE_RW : AMDGPU_VM_MTYPE_NC; + /* local HBM region far from partition or remote XGMI GPU */ + else if (svm_nodes_in_same_hive(bo_node, node)) + mapping_flags |= AMDGPU_VM_MTYPE_NC; + /* PCIe P2P */ + else + mapping_flags |= AMDGPU_VM_MTYPE_UC; + /* system memory accessed by the APU */ + } else if (node->adev->flags & AMD_IS_APU) { mapping_flags |= AMDGPU_VM_MTYPE_NC; + /* system memory accessed by the dGPU */ + } else { + mapping_flags |= AMDGPU_VM_MTYPE_UC; + } break; default: mapping_flags |= coherent ? -- cgit v1.2.3 From d839a158b2480814bc438f9f46f440a7b9f63cb6 Mon Sep 17 00:00:00 2001 From: Graham Sider Date: Thu, 5 Jan 2023 10:58:07 -0500 Subject: drm/amdgpu: Correct dGPU MTYPE settings for gfx943 Revert temporary dGPU VRAM MTYPE setting and align with expected coherency protocol. Signed-off-by: Graham Sider Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 15 +++++---------- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 8 ++------ 2 files changed, 7 insertions(+), 16 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index aca8489635b8..b6c500be6f70 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1238,17 +1238,12 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, mtype = MTYPE_RW; } else { /* dGPU */ - /* - if ((mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) && - bo_adev == adev) - mapping_flags |= AMDGPU_VM_MTYPE_RW; + if (is_vram && bo_adev == adev) + mtype = MTYPE_RW; + else if (is_vram) + mtype = MTYPE_NC; else - */ - /* Temporarily comment out above lines and use MTYPE_NC - * on both VRAM and system memory access until - * MTYPE_RW can properly work on VRAM access - */ - mtype = MTYPE_NC; + mtype = MTYPE_UC; } break; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 6daba0582bf3..2b79849ddd30 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1198,13 +1198,9 @@ svm_range_get_pte_flags(struct kfd_node *node, if (uncached) { mapping_flags |= AMDGPU_VM_MTYPE_UC; } else if (domain == SVM_RANGE_VRAM_DOMAIN) { - /* local HBM region close to partition with a workaround - * for Endpoint systems. - */ + /* local HBM region close to partition */ if (bo_node == node) - mapping_flags |= - (node->adev->flags & AMD_IS_APU) ? - AMDGPU_VM_MTYPE_RW : AMDGPU_VM_MTYPE_NC; + mapping_flags |= AMDGPU_VM_MTYPE_RW; /* local HBM region far from partition or remote XGMI GPU */ else if (svm_nodes_in_same_hive(bo_node, node)) mapping_flags |= AMDGPU_VM_MTYPE_NC; -- cgit v1.2.3 From b2ef2fdffed2a7fd5bf3f178a6a0427487dba5dd Mon Sep 17 00:00:00 2001 From: Rajneesh Bhardwaj Date: Thu, 5 Jan 2023 11:39:34 -0500 Subject: drm/amdkfd: Report XGMI IOLINKs for GFXIP9.4.3 GFXIP 9.4.3 could be in APU or carveout mode but we cannot use the xgmi.connected_to_cpu flag to identify the iolinks type. Use appropriate APU or Carveout mode based condition to report xgmi connection in kfd topology. Reviewed-by: Felix Kuehling Signed-off-by: Rajneesh Bhardwaj Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 5 ++++- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 5 ++--- 2 files changed, 6 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index dc93a67257e1..16475921587b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -2005,7 +2005,10 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, /* Fill in IOLINK subtype. * TODO: Fill-in other fields of iolink subtype */ - if (kdev->adev->gmc.xgmi.connected_to_cpu) { + if (kdev->adev->gmc.xgmi.connected_to_cpu || + (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) && + kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) == + AMDGPU_PKG_TYPE_APU)) { bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3); int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT : KFD_CRAT_INTRA_SOCKET_WEIGHT; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index d3e70341dfad..5373a79ac6a1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1327,9 +1327,8 @@ static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) continue; /* Include the CPU peer in GPU hive if connected over xGMI. */ - if (!peer_dev->gpu && !peer_dev->node_props.hive_id && - dev->node_props.hive_id && - dev->gpu->adev->gmc.xgmi.connected_to_cpu) + if (!peer_dev->gpu && + link->iolink_type == CRAT_IOLINK_TYPE_XGMI) peer_dev->node_props.hive_id = dev->node_props.hive_id; list_for_each_entry(inbound_link, &peer_dev->io_link_props, -- cgit v1.2.3 From 8c45a8340dd097ea0d6be6f718c4882283d9645d Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Tue, 24 Jan 2023 10:10:14 -0500 Subject: drm/amdkfd: Cleanup KFD nodes creation kfd node allocation outside kfd->num_nodes loop is not needed and causes memory leak because kfd->num_nodes is at least equal to 1. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 18 ++---------------- 1 file changed, 2 insertions(+), 16 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 7a963d0a34e2..d7cffd91f1d7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -729,26 +729,12 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, /* TODO: Needs to be updated for memory partitioning */ svm_migrate_init(kfd->adev); - /* Allocate the KFD node */ - node = kzalloc(sizeof(struct kfd_node), GFP_KERNEL); - if (!node) { - dev_err(kfd_device, "Error allocating KFD node\n"); - goto node_alloc_error; - } - - node->adev = kfd->adev; - node->kfd = kfd; - node->kfd2kgd = kfd->kfd2kgd; - node->vm_info.vmid_num_kfd = vmid_num_kfd; - node->vm_info.first_vmid_kfd = first_vmid_kfd; - node->vm_info.last_vmid_kfd = last_vmid_kfd; - node->max_proc_per_quantum = max_proc_per_quantum; - atomic_set(&node->sram_ecc_flag, 0); - amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n", kfd->num_nodes); + + /* Allocate the KFD nodes */ for (i = 0; i < kfd->num_nodes; i++) { node = kzalloc(sizeof(struct kfd_node), GFP_KERNEL); if (!node) -- cgit v1.2.3 From 1698e200e88db96aef7d16aa3d63df68a209ffbd Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 2 Feb 2023 11:10:08 -0500 Subject: drm/amdkfd: bind cpu and hiveless gpu to a hive if xgmi connected If a CPU and GPU are xGMI connected but the GPU is hiveless with respect to other GPUs, create a new CPU-GPU hive using the GPU's PCI device location ID as the new hive ID to maintain fine grain memory access usage. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 5373a79ac6a1..c7072fff778e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1328,8 +1328,15 @@ static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) /* Include the CPU peer in GPU hive if connected over xGMI. */ if (!peer_dev->gpu && - link->iolink_type == CRAT_IOLINK_TYPE_XGMI) + link->iolink_type == CRAT_IOLINK_TYPE_XGMI) { + /* + * If the GPU is not part of a GPU hive, use its pci + * device location as the hive ID to bind with the CPU. + */ + if (!dev->node_props.hive_id) + dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev); peer_dev->node_props.hive_id = dev->node_props.hive_id; + } list_for_each_entry(inbound_link, &peer_dev->io_link_props, list) { -- cgit v1.2.3 From 228ce176434b0f61451019065393040d58e1668d Mon Sep 17 00:00:00 2001 From: Rajneesh Bhardwaj Date: Fri, 27 Jan 2023 21:57:00 -0500 Subject: drm/amdgpu: Handle VRAM dependencies on GFXIP9.4.3 [For 1P NPS1 mode driver bringup] Changes required to initialize the amdgpu driver with frontdoor firmware loading and discovery=2 with the native mode SBIOS that enables CPU GPU unified interleaved memory. sudo modprobe amdgpu discovery=2 Once PSP TMR region is reported via the ACPI interface, the dependency on the ip_discovery.bin will be removed. Choice of where to allocate driver table is given to each IP version. In general, both GTT and VRAM domains will be considered. If one of the tables has a strict restriction for VRAM domain, then only VRAM domain is considered. Reviewed-by: Felix Kuehling (lijo: Modified the handling for SMU Tables) Signed-off-by: Lijo Lazar Signed-off-by: Rajneesh Bhardwaj Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c | 6 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 89 +++++++++++++--------- drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c | 7 +- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 3 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 9 ++- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++ drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 5 ++ drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c | 10 ++- .../gpu/drm/amd/pm/swsmu/smu13/smu_v13_0_6_ppt.c | 6 +- 11 files changed, 99 insertions(+), 47 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index af37f2ef4438..4e179e50de25 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -2292,8 +2292,9 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, (*mem)->dmabuf = dma_buf; (*mem)->bo = bo; (*mem)->va = va; - (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ? + (*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ? AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT; + (*mem)->mapped_to_gpu_memory = 0; (*mem)->process_info = avm->process_info; add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c index 9f0d5f02119e..f431205e1077 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c @@ -1044,7 +1044,7 @@ static const char * const amdgpu_vram_names[] = { int amdgpu_bo_init(struct amdgpu_device *adev) { /* On A+A platform, VRAM can be mapped as WB */ - if (!adev->gmc.xgmi.connected_to_cpu) { + if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) { /* reserve PAT memory space to WC for VRAM */ int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c index 863fa331e6ff..4395c53d09d8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_psp.c @@ -476,7 +476,8 @@ static int psp_sw_init(void *handle) return ret; ret = amdgpu_bo_create_kernel(adev, PSP_FENCE_BUFFER_SIZE, PAGE_SIZE, - AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_DOMAIN_VRAM | + AMDGPU_GEM_DOMAIN_GTT, &psp->fence_buf_bo, &psp->fence_buf_mc_addr, &psp->fence_buf); @@ -484,7 +485,8 @@ static int psp_sw_init(void *handle) goto failed1; ret = amdgpu_bo_create_kernel(adev, PSP_CMD_BUFFER_SIZE, PAGE_SIZE, - AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_DOMAIN_VRAM | + AMDGPU_GEM_DOMAIN_GTT, &psp->cmd_buf_bo, &psp->cmd_buf_mc_addr, (void **)&psp->cmd_buf_mem); if (ret) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 6bbe3b89aef5..bc11ae56bba5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1708,15 +1708,20 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS; } - ret = amdgpu_bo_create_kernel_at(adev, - adev->gmc.real_vram_size - adev->mman.discovery_tmr_size, - adev->mman.discovery_tmr_size, - &adev->mman.discovery_memory, - NULL); - if (ret) { - DRM_ERROR("alloc tmr failed(%d)!\n", ret); - amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); - return ret; + if (!adev->gmc.is_app_apu) { + ret = amdgpu_bo_create_kernel_at(adev, + adev->gmc.real_vram_size - + adev->mman.discovery_tmr_size, + adev->mman.discovery_tmr_size, + &adev->mman.discovery_memory, + NULL); + if (ret) { + DRM_ERROR("alloc tmr failed(%d)!\n", ret); + amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); + return ret; + } + } else { + DRM_DEBUG_DRIVER("backdoor fw loading path for PSP TMR, no reservation needed\n"); } return 0; @@ -1765,10 +1770,12 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base, adev->gmc.visible_vram_size); - else + else if (!adev->gmc.is_app_apu) #endif adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base, adev->gmc.visible_vram_size); + else + DRM_DEBUG_DRIVER("No need to ioremap when real vram size is 0\n"); #endif /* @@ -1803,23 +1810,32 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) * This is used for VGA emulation and pre-OS scanout buffers to * avoid display artifacts while transitioning between pre-OS * and driver. */ - r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size, - &adev->mman.stolen_vga_memory, - NULL); - if (r) - return r; - r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size, - adev->mman.stolen_extended_size, - &adev->mman.stolen_extended_memory, - NULL); - if (r) - return r; - r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_reserved_offset, - adev->mman.stolen_reserved_size, - &adev->mman.stolen_reserved_memory, - NULL); - if (r) - return r; + if (!adev->gmc.is_app_apu) { + r = amdgpu_bo_create_kernel_at(adev, 0, + adev->mman.stolen_vga_size, + &adev->mman.stolen_vga_memory, + NULL); + if (r) + return r; + + r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size, + adev->mman.stolen_extended_size, + &adev->mman.stolen_extended_memory, + NULL); + + if (r) + return r; + + r = amdgpu_bo_create_kernel_at(adev, + adev->mman.stolen_reserved_offset, + adev->mman.stolen_reserved_size, + &adev->mman.stolen_reserved_memory, + NULL); + if (r) + return r; + } else { + DRM_DEBUG_DRIVER("Skipped stolen memory reservation\n"); + } DRM_INFO("amdgpu: %uM of VRAM memory ready\n", (unsigned) (adev->gmc.real_vram_size / (1024 * 1024))); @@ -1866,7 +1882,6 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) DRM_ERROR("Failed initializing oa heap.\n"); return r; } - if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, &adev->mman.sdma_access_bo, NULL, @@ -1887,13 +1902,15 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev) amdgpu_ttm_training_reserve_vram_fini(adev); /* return the stolen vga memory back to VRAM */ - amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); - amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL); - /* return the IP Discovery TMR memory back to VRAM */ - amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); - if (adev->mman.stolen_reserved_size) - amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory, - NULL, NULL); + if (!adev->gmc.is_app_apu) { + amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); + amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL); + /* return the IP Discovery TMR memory back to VRAM */ + amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); + if (adev->mman.stolen_reserved_size) + amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory, + NULL, NULL); + } amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL, &adev->mman.sdma_access_ptr); amdgpu_ttm_fw_reserve_vram_fini(adev); @@ -1935,7 +1952,7 @@ void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable) int r; if (!adev->mman.initialized || amdgpu_in_reset(adev) || - adev->mman.buffer_funcs_enabled == enable) + adev->mman.buffer_funcs_enabled == enable || adev->gmc.is_app_apu) return; if (enable) { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c index df63dc3bca18..bc5d126b600b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c @@ -512,7 +512,12 @@ int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm, bp.size = amdgpu_vm_pt_size(adev, level); bp.byte_align = AMDGPU_GPU_PAGE_SIZE; - bp.domain = AMDGPU_GEM_DOMAIN_VRAM; + + if (!adev->gmc.is_app_apu) + bp.domain = AMDGPU_GEM_DOMAIN_VRAM; + else + bp.domain = AMDGPU_GEM_DOMAIN_GTT; + bp.domain = amdgpu_bo_get_preferred_domain(adev, bp.domain); bp.flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS | AMDGPU_GEM_CREATE_CPU_GTT_USWC; diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c index 1f1268cd5e09..42877c4505f1 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c @@ -459,7 +459,8 @@ static int gfx_v9_4_3_mec_init(struct amdgpu_device *adev) adev->gfx.num_compute_rings * num_xcc * GFX9_MEC_HPD_SIZE; if (mec_hpd_size) { r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE, - AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_DOMAIN_VRAM | + AMDGPU_GEM_DOMAIN_GTT, &adev->gfx.mec.hpd_eop_obj, &adev->gfx.mec.hpd_eop_gpu_addr, (void **)&hpd); diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 16634a791e10..245de27c7540 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1593,8 +1593,13 @@ static int gmc_v9_0_mc_init(struct amdgpu_device *adev) int r; /* size in MB on si */ - adev->gmc.mc_vram_size = - adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; + if (!adev->gmc.is_app_apu) { + adev->gmc.mc_vram_size = + adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; + } else { + DRM_DEBUG("Set mc_vram_size = 0 for APP APU\n"); + adev->gmc.mc_vram_size = 0; + } adev->gmc.real_vram_size = adev->gmc.mc_vram_size; if (!(adev->flags & AMD_IS_APU) && diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 8b9accecf49b..f85ac4dbc673 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1026,6 +1026,12 @@ bool kfd_dev_is_large_bar(struct kfd_node *dev) if (dev->kfd->local_mem_info.local_mem_size_private == 0 && dev->kfd->local_mem_info.local_mem_size_public > 0) return true; + + if (dev->kfd->local_mem_info.local_mem_size_public == 0 && dev->kfd->adev->gmc.is_app_apu) { + pr_debug("APP APU, Consider like a large bar system\n"); + return true; + } + return false; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index 16475921587b..1aaf933f9f48 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -30,6 +30,9 @@ #include "amdgpu.h" #include "amdgpu_amdkfd.h" +/* Fixme: Fake 32GB for 1PNPS1 mode bringup */ +#define DUMMY_VRAM_SIZE 31138512896 + /* GPU Processor ID base for dGPUs for which VCRAT needs to be created. * GPU processor ID are expressed with Bit[31]=1. * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs @@ -1053,6 +1056,8 @@ static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem, props->heap_type = heap_type; props->flags = flags; + if (size_in_bytes == 0) + size_in_bytes = DUMMY_VRAM_SIZE; /* Fixme: TBD */ props->size_in_bytes = size_in_bytes; props->width = width; diff --git a/drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c b/drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c index 2ddf5198e5c4..4dea79a0c5b5 100644 --- a/drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c +++ b/drivers/gpu/drm/amd/pm/swsmu/amdgpu_smu.c @@ -822,11 +822,20 @@ static int smu_init_fb_allocations(struct smu_context *smu) } } + driver_table->domain = AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT; /* VRAM allocation for driver table */ for (i = 0; i < SMU_TABLE_COUNT; i++) { if (tables[i].size == 0) continue; + /* If one of the tables has VRAM domain restriction, keep it in + * VRAM + */ + if ((tables[i].domain & + (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) == + AMDGPU_GEM_DOMAIN_VRAM) + driver_table->domain = AMDGPU_GEM_DOMAIN_VRAM; + if (i == SMU_TABLE_PMSTATUSLOG) continue; @@ -836,7 +845,6 @@ static int smu_init_fb_allocations(struct smu_context *smu) driver_table->size = max_table_size; driver_table->align = PAGE_SIZE; - driver_table->domain = AMDGPU_GEM_DOMAIN_VRAM; ret = amdgpu_bo_create_kernel(adev, driver_table->size, diff --git a/drivers/gpu/drm/amd/pm/swsmu/smu13/smu_v13_0_6_ppt.c b/drivers/gpu/drm/amd/pm/swsmu/smu13/smu_v13_0_6_ppt.c index ea8f3d6fb98b..8969b3ff5c8f 100644 --- a/drivers/gpu/drm/amd/pm/swsmu/smu13/smu_v13_0_6_ppt.c +++ b/drivers/gpu/drm/amd/pm/swsmu/smu13/smu_v13_0_6_ppt.c @@ -220,10 +220,12 @@ static int smu_v13_0_6_tables_init(struct smu_context *smu) PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM); SMU_TABLE_INIT(tables, SMU_TABLE_SMU_METRICS, sizeof(MetricsTable_t), - PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM); + PAGE_SIZE, + AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT); SMU_TABLE_INIT(tables, SMU_TABLE_I2C_COMMANDS, sizeof(SwI2cRequest_t), - PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM); + PAGE_SIZE, + AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT); smu_table->metrics_table = kzalloc(sizeof(MetricsTable_t), GFP_KERNEL); if (!smu_table->metrics_table) -- cgit v1.2.3 From ded7d99eb5b78931cec30dd49cd4097d0ac770e1 Mon Sep 17 00:00:00 2001 From: Lijo Lazar Date: Mon, 16 Jan 2023 10:55:38 +0530 Subject: drm/amdgpu: Add flags for partition mode query It's not required to take lock on all cases while querying partition mode. Querying partition mode during KFD init process doesn't need to take a lock. Init process after a switch will already be happening under lock. Control the behaviour by adding flags to xcp_query_partition_mode. Signed-off-by: Lijo Lazar Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 3 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c | 8 +++++--- drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h | 5 ++++- drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c | 4 +++- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +- 5 files changed, 15 insertions(+), 7 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 70c6099353b8..1487ecac2705 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -1177,7 +1177,8 @@ static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev, int mode; char *partition_mode; - mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr); + mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr, + AMDGPU_XCP_FL_NONE); switch (mode) { case AMDGPU_SPX_PARTITION_MODE: diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c index f59bc450cabe..5b999e5334bb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c @@ -170,7 +170,7 @@ out: return ret; } -int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr) +int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags) { int mode; @@ -180,7 +180,8 @@ int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr) if (!xcp_mgr->funcs || !xcp_mgr->funcs->query_partition_mode) return xcp_mgr->mode; - mutex_lock(&xcp_mgr->xcp_lock); + if (!(flags & AMDGPU_XCP_FL_LOCKED)) + mutex_lock(&xcp_mgr->xcp_lock); mode = xcp_mgr->funcs->query_partition_mode(xcp_mgr); if (mode != xcp_mgr->mode) dev_WARN( @@ -188,7 +189,8 @@ int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr) "Cached partition mode %d not matching with device mode %d", xcp_mgr->mode, mode); - mutex_unlock(&xcp_mgr->xcp_lock); + if (!(flags & AMDGPU_XCP_FL_LOCKED)) + mutex_unlock(&xcp_mgr->xcp_lock); return mode; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h index f0b973c6092f..9fa6f0ea2061 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h @@ -30,6 +30,9 @@ #define AMDGPU_XCP_MODE_NONE -1 +#define AMDGPU_XCP_FL_NONE 0 +#define AMDGPU_XCP_FL_LOCKED (1 << 0) + enum AMDGPU_XCP_IP_BLOCK { AMDGPU_XCP_GFXHUB, AMDGPU_XCP_GFX, @@ -99,7 +102,7 @@ int amdgpu_xcp_resume(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id); int amdgpu_xcp_mgr_init(struct amdgpu_device *adev, int init_mode, int init_xcps, struct amdgpu_xcp_mgr_funcs *xcp_funcs); -int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr); +int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags); int amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, int mode); int amdgpu_xcp_get_partition(struct amdgpu_xcp_mgr *xcp_mgr, enum AMDGPU_XCP_IP_BLOCK ip, int instance); diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c index 42877c4505f1..69867294117e 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c @@ -1940,7 +1940,9 @@ static int gfx_v9_4_3_cp_resume(struct amdgpu_device *adev) { int r, i, num_xcc; - if (amdgpu_xcp_query_partition_mode(adev->xcp_mgr) == AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE) + if (amdgpu_xcp_query_partition_mode(adev->xcp_mgr, + AMDGPU_XCP_FL_NONE) == + AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE) amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, amdgpu_user_partt_mode); num_xcc = NUM_XCC(adev->gfx.xcc_mask); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index d7cffd91f1d7..4293cbf9ceb0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -645,7 +645,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, * If the VMID range changes for GFX9.4.3, then this code MUST be * revisited. */ - partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr); + partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr, AMDGPU_XCP_FL_LOCKED); if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && partition_mode == AMDGPU_CPX_PARTITION_MODE && kfd->num_nodes != 1) { -- cgit v1.2.3 From f4d8b6f5c61ab5e98258bd0072d733741c76bd8d Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 31 Jan 2023 11:23:50 -0500 Subject: drm/amdkfd: Enable SVM on Native mode This patch enables SVM capability on GFX9.4.3 when run in Native mode. It also sets best_prefetch and best_restore locations to CPU as there is no VRAM. Signed-off-by: Mukul Joshi Acked-by: Rajneesh Bhardwaj Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 3 ++- 2 files changed, 10 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 2b79849ddd30..cf354f9e4285 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -2543,6 +2543,9 @@ svm_range_best_restore_location(struct svm_range *prange, return -1; } + if (node->adev->gmc.is_app_apu) + return 0; + if (prange->preferred_loc == gpuid || prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) { return prange->preferred_loc; @@ -3256,6 +3259,11 @@ svm_range_best_prefetch_location(struct svm_range *prange) goto out; } + if (bo_node->adev->gmc.is_app_apu) { + best_loc = 0; + goto out; + } + if (p->xnack_enabled) bitmap_copy(bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); else diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index 5116786718b6..7515ddade3ae 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -200,7 +200,8 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s /* SVM API and HMM page migration work together, device memory type * is initialized to not 0 when page migration register device memory. */ -#define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0) +#define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0 ||\ + (dev)->adev->gmc.is_app_apu) void svm_range_bo_unref_async(struct svm_range_bo *svm_bo); -- cgit v1.2.3 From 75dda67c7213c3e0d17244a8c42547c27ee746f8 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 9 Feb 2023 18:23:16 -0500 Subject: drm/amdkfd: Flush TLB after unmapping for GFX v9.4.3 kfd_flush_tlb_after_unmap should return true for GFX v9.4.3, to do TLB heavyweight flush after unmapping from GPU to guarantee that the GPU will not access pages after they have been unmapped. This also helps improve the mapping to GPU performance. Without this, KFD accidently flush TLB after mapping to GPU because the vm update sequence number is increased by previous unmapping. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index fb3cf2c51da8..6e1c15682c28 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1405,9 +1405,9 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type); static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) { - return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || - (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && - dev->adev->sdma.instance[0].fw_version >= 18) || + return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3) || + KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || + (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) || KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0); } -- cgit v1.2.3 From a75f2271a4936265c8a189ab06f9eb89e343b441 Mon Sep 17 00:00:00 2001 From: Lijo Lazar Date: Thu, 9 Feb 2023 14:44:13 +0530 Subject: drm/amdkfd: Add xcp reference to kfd node Fetch xcp information from xcp_mgr and also add xcc_mask to kfd node. Signed-off-by: Lijo Lazar Reviewed-by: Le Ma Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 19 +++++++++++++------ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 3 +++ 2 files changed, 16 insertions(+), 6 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 4293cbf9ceb0..647c3313c27e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -594,6 +594,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; unsigned int max_proc_per_quantum; int num_xcd, partition_mode; + int xcp_idx; kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, KGD_ENGINE_MEC1); @@ -603,11 +604,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, KGD_ENGINE_SDMA1); kfd->shared_resources = *gpu_resources; - num_xcd = NUM_XCC(kfd->adev->gfx.xcc_mask); - if (num_xcd == 0 || num_xcd == 1 || kfd->adev->gfx.num_xcc_per_xcp == 0) - kfd->num_nodes = 1; - else - kfd->num_nodes = num_xcd / kfd->adev->gfx.num_xcc_per_xcp; + kfd->num_nodes = amdgpu_xcp_get_num_xcp(kfd->adev->xcp_mgr); + if (kfd->num_nodes == 0) { dev_err(kfd_device, "KFD num nodes cannot be 0, GC inst: %d, num_xcc_in_node: %d\n", @@ -735,7 +733,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd->num_nodes); /* Allocate the KFD nodes */ - for (i = 0; i < kfd->num_nodes; i++) { + for (i = 0, xcp_idx = 0; i < kfd->num_nodes; i++) { node = kzalloc(sizeof(struct kfd_node), GFP_KERNEL); if (!node) goto node_alloc_error; @@ -745,6 +743,15 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, node->kfd = kfd; node->kfd2kgd = kfd->kfd2kgd; node->vm_info.vmid_num_kfd = vmid_num_kfd; + node->xcp = amdgpu_get_next_xcp(kfd->adev->xcp_mgr, &xcp_idx); + /* TODO : Check if error handling is needed */ + if (node->xcp) + amdgpu_xcp_get_inst_details(node->xcp, AMDGPU_XCP_GFX, + &node->xcc_mask); + else + node->xcc_mask = + (1U << NUM_XCC(kfd->adev->gfx.xcc_mask)) - 1; + node->num_xcc_per_node = max(1U, kfd->adev->gfx.num_xcc_per_xcp); node->start_xcc_id = node->num_xcc_per_node * i; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 6e1c15682c28..559ac5efdc26 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -278,6 +278,9 @@ struct kfd_node { unsigned int start_xcc_id; /* Starting XCC instance * number for the node */ + uint32_t xcc_mask; /* Instance mask of XCCs present */ + struct amdgpu_xcp *xcp; + /* Interrupts */ struct kfifo ih_fifo; struct workqueue_struct *ih_wq; -- cgit v1.2.3 From c4050ff1a43eec08498b1ed876efc6213592dba0 Mon Sep 17 00:00:00 2001 From: Lijo Lazar Date: Thu, 9 Feb 2023 16:30:53 +0530 Subject: drm/amdkfd: Use xcc mask for identifying xcc Instead of start xcc id and number of xcc per node, use the xcc mask which is the mask of logical ids of xccs belonging to a parition. Signed-off-by: Lijo Lazar Reviewed-by: Le Ma Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 9 +-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 86 +++++++++++----------- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 71 +++++++++--------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 4 - drivers/gpu/drm/amd/amdkfd/kfd_process.c | 8 +- .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 8 +- 8 files changed, 95 insertions(+), 95 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 647c3313c27e..b5497d2ee984 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -745,15 +745,14 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, node->vm_info.vmid_num_kfd = vmid_num_kfd; node->xcp = amdgpu_get_next_xcp(kfd->adev->xcp_mgr, &xcp_idx); /* TODO : Check if error handling is needed */ - if (node->xcp) + if (node->xcp) { amdgpu_xcp_get_inst_details(node->xcp, AMDGPU_XCP_GFX, &node->xcc_mask); - else + ++xcp_idx; + } else { node->xcc_mask = (1U << NUM_XCC(kfd->adev->gfx.xcc_mask)) - 1; - - node->num_xcc_per_node = max(1U, kfd->adev->gfx.num_xcc_per_xcp); - node->start_xcc_id = node->num_xcc_per_node * i; + } if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && partition_mode == AMDGPU_CPX_PARTITION_MODE && diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 2b5c4b2dd242..493b4b66f180 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -136,16 +136,14 @@ static void init_sdma_bitmaps(struct device_queue_manager *dqm) void program_sh_mem_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { - int xcc = 0; + uint32_t xcc_mask = dqm->dev->xcc_mask; + int xcc_id; - for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) + for_each_inst(xcc_id, xcc_mask) dqm->dev->kfd2kgd->program_sh_mem_settings( - dqm->dev->adev, qpd->vmid, - qpd->sh_mem_config, - qpd->sh_mem_ape1_base, - qpd->sh_mem_ape1_limit, - qpd->sh_mem_bases, - dqm->dev->start_xcc_id + xcc); + dqm->dev->adev, qpd->vmid, qpd->sh_mem_config, + qpd->sh_mem_ape1_base, qpd->sh_mem_ape1_limit, + qpd->sh_mem_bases, xcc_id); } static void kfd_hws_hang(struct device_queue_manager *dqm) @@ -427,14 +425,14 @@ static void deallocate_doorbell(struct qcm_process_device *qpd, static void program_trap_handler_settings(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { - int xcc = 0; + uint32_t xcc_mask = dqm->dev->xcc_mask; + int xcc_id; if (dqm->dev->kfd2kgd->program_trap_handler_settings) - for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) + for_each_inst(xcc_id, xcc_mask) dqm->dev->kfd2kgd->program_trap_handler_settings( - dqm->dev->adev, qpd->vmid, - qpd->tba_addr, qpd->tma_addr, - dqm->dev->start_xcc_id + xcc); + dqm->dev->adev, qpd->vmid, qpd->tba_addr, + qpd->tma_addr, xcc_id); } static int allocate_vmid(struct device_queue_manager *dqm, @@ -697,7 +695,8 @@ static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process struct kfd_process_device *pdd; int first_vmid_to_scan = dev->vm_info.first_vmid_kfd; int last_vmid_to_scan = dev->vm_info.last_vmid_kfd; - int xcc = 0; + uint32_t xcc_mask = dev->xcc_mask; + int xcc_id; reg_sq_cmd.u32All = 0; reg_gfx_index.u32All = 0; @@ -742,11 +741,10 @@ static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL; reg_sq_cmd.bits.vm_id = vmid; - for (xcc = 0; xcc < dev->num_xcc_per_node; xcc++) - dev->kfd2kgd->wave_control_execute(dev->adev, - reg_gfx_index.u32All, - reg_sq_cmd.u32All, - dev->start_xcc_id + xcc); + for_each_inst(xcc_id, xcc_mask) + dev->kfd2kgd->wave_control_execute( + dev->adev, reg_gfx_index.u32All, + reg_sq_cmd.u32All, xcc_id); return 0; } @@ -1258,12 +1256,12 @@ static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid, unsigned int vmid) { - int xcc = 0, ret; + uint32_t xcc_mask = dqm->dev->xcc_mask; + int xcc_id, ret; - for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) { + for_each_inst(xcc_id, xcc_mask) { ret = dqm->dev->kfd2kgd->set_pasid_vmid_mapping( - dqm->dev->adev, pasid, vmid, - dqm->dev->start_xcc_id + xcc); + dqm->dev->adev, pasid, vmid, xcc_id); if (ret) break; } @@ -1273,15 +1271,14 @@ set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid, static void init_interrupts(struct device_queue_manager *dqm) { - unsigned int i, xcc; + uint32_t xcc_mask = dqm->dev->xcc_mask; + unsigned int i, xcc_id; for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) { if (is_pipe_enabled(dqm, 0, i)) { - for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) + for_each_inst(xcc_id, xcc_mask) dqm->dev->kfd2kgd->init_interrupts( - dqm->dev->adev, i, - dqm->dev->start_xcc_id + - xcc); + dqm->dev->adev, i, xcc_id); } } } @@ -2283,7 +2280,7 @@ static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) get_num_all_sdma_engines(dqm) * dev->kfd->device_info.num_sdma_queues_per_engine + (dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size * - dqm->dev->num_xcc_per_node); + NUM_XCC(dqm->dev->xcc_mask)); retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size, &(mem_obj->gtt_mem), &(mem_obj->gpu_addr), @@ -2489,10 +2486,10 @@ static void seq_reg_dump(struct seq_file *m, int dqm_debugfs_hqds(struct seq_file *m, void *data) { struct device_queue_manager *dqm = data; + uint32_t xcc_mask = dqm->dev->xcc_mask; uint32_t (*dump)[2], n_regs; int pipe, queue; - int r = 0, xcc; - uint32_t inst; + int r = 0, xcc_id; uint32_t sdma_engine_start; if (!dqm->sched_running) { @@ -2500,16 +2497,18 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) return 0; } - for (xcc = 0; xcc < dqm->dev->num_xcc_per_node; xcc++) { - inst = dqm->dev->start_xcc_id + xcc; + for_each_inst(xcc_id, xcc_mask) { r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, - KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE, - &dump, &n_regs, inst); + KFD_CIK_HIQ_PIPE, + KFD_CIK_HIQ_QUEUE, &dump, + &n_regs, xcc_id); if (!r) { - seq_printf(m, + seq_printf( + m, " Inst %d, HIQ on MEC %d Pipe %d Queue %d\n", - inst, KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1, - KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm), + xcc_id, + KFD_CIK_HIQ_PIPE / get_pipes_per_mec(dqm) + 1, + KFD_CIK_HIQ_PIPE % get_pipes_per_mec(dqm), KFD_CIK_HIQ_QUEUE); seq_reg_dump(m, dump, n_regs); @@ -2524,13 +2523,16 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data) dqm->dev->kfd->shared_resources.cp_queue_bitmap)) continue; - r = dqm->dev->kfd2kgd->hqd_dump( - dqm->dev->adev, pipe, queue, &dump, &n_regs, inst); + r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, + pipe, queue, + &dump, &n_regs, + xcc_id); if (r) break; - seq_printf(m, " Inst %d, CP Pipe %d, Queue %d\n", - inst, pipe, queue); + seq_printf(m, + " Inst %d, CP Pipe %d, Queue %d\n", + xcc_id, pipe, queue); seq_reg_dump(m, dump, n_regs); kfree(dump); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c index d81125421aaf..863cf060af48 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c @@ -77,7 +77,7 @@ struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev, dev->dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size; offset += dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size * - dev->num_xcc_per_node; + NUM_XCC(dev->xcc_mask); mqd_mem_obj->gtt_mem = (void *)((uint64_t)dev->dqm->hiq_sdma_mqd.gtt_mem + offset); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index c781314b213c..226132ec3714 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -128,7 +128,7 @@ static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node, retval = amdgpu_amdkfd_alloc_gtt_mem(node->adev, (ALIGN(q->ctl_stack_size, PAGE_SIZE) + ALIGN(sizeof(struct v9_mqd), PAGE_SIZE)) * - node->num_xcc_per_node, + NUM_XCC(node->xcc_mask), &(mqd_mem_obj->gtt_mem), &(mqd_mem_obj->gpu_addr), (void *)&(mqd_mem_obj->cpu_ptr), true); @@ -482,7 +482,7 @@ static void init_mqd_hiq_v9_4_3(struct mqd_manager *mm, void **mqd, memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj)); - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) { kfd_get_hiq_xcc_mqd(mm->dev, &xcc_mqd_mem_obj, xcc); init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q); @@ -506,21 +506,21 @@ static int hiq_load_mqd_kiq_v9_4_3(struct mqd_manager *mm, void *mqd, uint32_t pipe_id, uint32_t queue_id, struct queue_properties *p, struct mm_struct *mms) { - int xcc, err; + uint32_t xcc_mask = mm->dev->xcc_mask; + int xcc_id, err, inst = 0; void *xcc_mqd; - uint32_t start_inst = mm->dev->start_xcc_id; uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev); - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { - xcc_mqd = mqd + hiq_mqd_size * xcc; + for_each_inst(xcc_id, xcc_mask) { + xcc_mqd = mqd + hiq_mqd_size * inst; err = mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, xcc_mqd, pipe_id, queue_id, - p->doorbell_off, - start_inst+xcc); + p->doorbell_off, xcc_id); if (err) { - pr_debug("Failed to load HIQ MQD for XCC: %d\n", xcc); + pr_debug("Failed to load HIQ MQD for XCC: %d\n", inst); break; } + ++inst; } return err; @@ -530,20 +530,21 @@ static int destroy_hiq_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, enum kfd_preempt_type type, unsigned int timeout, uint32_t pipe_id, uint32_t queue_id) { - int xcc = 0, err; + uint32_t xcc_mask = mm->dev->xcc_mask; + int xcc_id, err, inst = 0; void *xcc_mqd; - uint32_t start_inst = mm->dev->start_xcc_id; uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev); - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { - xcc_mqd = mqd + hiq_mqd_size * xcc; + for_each_inst(xcc_id, xcc_mask) { + xcc_mqd = mqd + hiq_mqd_size * inst; err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd, type, timeout, pipe_id, - queue_id, start_inst+xcc); + queue_id, xcc_id); if (err) { - pr_debug("Destroy MQD failed for xcc: %d\n", xcc); + pr_debug("Destroy MQD failed for xcc: %d\n", inst); break; } + ++inst; } return err; @@ -573,7 +574,7 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, uint32_t local_xcc_start = mm->dev->dqm->current_logical_xcc_start++; memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj)); - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) { get_xcc_mqd(mqd_mem_obj, &xcc_mqd_mem_obj, offset*xcc); init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q); @@ -600,7 +601,7 @@ static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd, m->compute_tg_chunk_size = 1; m->compute_current_logic_xcc_id = (local_xcc_start + xcc) % - mm->dev->num_xcc_per_node; + NUM_XCC(mm->dev->xcc_mask); switch (xcc) { case 0: @@ -633,7 +634,7 @@ static void update_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, int xcc = 0; uint64_t size = mm->mqd_stride(mm, q); - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) { m = get_mqd(mqd + size * xcc); update_mqd(mm, m, q, minfo); @@ -661,24 +662,25 @@ static int destroy_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, enum kfd_preempt_type type, unsigned int timeout, uint32_t pipe_id, uint32_t queue_id) { - int xcc = 0, err; + uint32_t xcc_mask = mm->dev->xcc_mask; + int xcc_id, err, inst = 0; void *xcc_mqd; struct v9_mqd *m; uint64_t mqd_offset; - uint32_t start_inst = mm->dev->start_xcc_id; m = get_mqd(mqd); mqd_offset = m->cp_mqd_stride_size; - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { - xcc_mqd = mqd + mqd_offset * xcc; + for_each_inst(xcc_id, xcc_mask) { + xcc_mqd = mqd + mqd_offset * inst; err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd, type, timeout, pipe_id, - queue_id, start_inst+xcc); + queue_id, xcc_id); if (err) { - pr_debug("Destroy MQD failed for xcc: %d\n", xcc); + pr_debug("Destroy MQD failed for xcc: %d\n", inst); break; } + ++inst; } return err; @@ -690,21 +692,22 @@ static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd, { /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */ uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0); - int xcc = 0, err; + uint32_t xcc_mask = mm->dev->xcc_mask; + int xcc_id, err, inst = 0; void *xcc_mqd; - uint32_t start_inst = mm->dev->start_xcc_id; uint64_t mqd_stride_size = mm->mqd_stride(mm, p); - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { - xcc_mqd = mqd + mqd_stride_size * xcc; - err = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, xcc_mqd, - pipe_id, queue_id, - (uint32_t __user *)p->write_ptr, - wptr_shift, 0, mms, start_inst+xcc); + for_each_inst(xcc_id, xcc_mask) { + xcc_mqd = mqd + mqd_stride_size * inst; + err = mm->dev->kfd2kgd->hqd_load( + mm->dev->adev, xcc_mqd, pipe_id, queue_id, + (uint32_t __user *)p->write_ptr, wptr_shift, 0, mms, + xcc_id); if (err) { - pr_debug("Load MQD failed for xcc: %d\n", xcc); + pr_debug("Load MQD failed for xcc: %d\n", inst); break; } + ++inst; } return err; @@ -722,7 +725,7 @@ static int get_wave_state_v9_4_3(struct mqd_manager *mm, void *mqd, uint64_t mqd_stride_size = mm->mqd_stride(mm, q); u32 tmp_ctl_stack_used_size = 0, tmp_save_area_used_size = 0; - for (xcc = 0; xcc < mm->dev->num_xcc_per_node; xcc++) { + for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) { xcc_mqd = mqd + mqd_stride_size * xcc; xcc_ctl_stack = (void __user *)((uintptr_t)ctl_stack + q->ctx_save_restore_area_size * xcc); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 559ac5efdc26..02a90fd7f646 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -274,10 +274,6 @@ struct kfd_node { */ struct kfd_vmid_info vm_info; unsigned int id; /* topology stub index */ - unsigned int num_xcc_per_node; - unsigned int start_xcc_id; /* Starting XCC instance - * number for the node - */ uint32_t xcc_mask; /* Instance mask of XCCs present */ struct amdgpu_xcp *xcp; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index a6ff57f11472..7f7d1378a2f8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -2058,6 +2058,7 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv); uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm); struct kfd_node *dev = pdd->dev; + uint32_t xcc_mask = dev->xcc_mask; int xcc = 0; /* @@ -2076,10 +2077,9 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->adev, pdd->qpd.vmid); } else { - for (xcc = 0; xcc < dev->num_xcc_per_node; xcc++) - amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->adev, - pdd->process->pasid, type, - dev->start_xcc_id + xcc); + for_each_inst(xcc, xcc_mask) + amdgpu_amdkfd_flush_gpu_tlb_pasid( + dev->adev, pdd->process->pasid, type, xcc); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 2b2ae0c9902b..a3c23d07c7df 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -946,7 +946,7 @@ int pqm_debugfs_mqds(struct seq_file *m, void *data) seq_printf(m, " Compute queue on device %x\n", q->device->id); mqd_type = KFD_MQD_TYPE_CP; - num_xccs = q->device->num_xcc_per_node; + num_xccs = NUM_XCC(q->device->xcc_mask); break; default: seq_printf(m, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index c7072fff778e..d2a42b6b1fa8 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -469,7 +469,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, dev->node_props.cpu_cores_count); sysfs_show_32bit_prop(buffer, offs, "simd_count", dev->gpu ? (dev->node_props.simd_count * - dev->gpu->num_xcc_per_node) : 0); + NUM_XCC(dev->gpu->xcc_mask)) : 0); sysfs_show_32bit_prop(buffer, offs, "mem_banks_count", dev->node_props.mem_banks_count); sysfs_show_32bit_prop(buffer, offs, "caches_count", @@ -494,7 +494,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, dev->node_props.wave_front_size); sysfs_show_32bit_prop(buffer, offs, "array_count", dev->gpu ? (dev->node_props.array_count * - dev->gpu->num_xcc_per_node) : 0); + NUM_XCC(dev->gpu->xcc_mask)) : 0); sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine", dev->node_props.simd_arrays_per_engine); sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array", @@ -558,7 +558,7 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, sysfs_show_64bit_prop(buffer, offs, "unique_id", dev->gpu->adev->unique_id); sysfs_show_32bit_prop(buffer, offs, "num_xcc", - dev->gpu->num_xcc_per_node); + NUM_XCC(dev->gpu->xcc_mask)); } return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute", @@ -1180,7 +1180,7 @@ static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) buf[4] = gpu->adev->pdev->bus->number; buf[5] = lower_32_bits(local_mem_size); buf[6] = upper_32_bits(local_mem_size); - buf[7] = gpu->start_xcc_id | (gpu->num_xcc_per_node << 16); + buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16); for (i = 0, hashout = 0; i < 8; i++) hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); -- cgit v1.2.3 From 3697b9bd7c69910cb6543d8441211ecfb2f013ca Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Wed, 15 Mar 2023 14:04:33 -0400 Subject: drm/amdkfd: Increase queue number per process to 255 on GFX9.4.3 Increase the maximum number of queues that can be created per process to 255 on GFX 9.4.3. There is no HWS limitation restricting the number queues that can be created. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 7 +++++++ 1 file changed, 7 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index a3c23d07c7df..b100933340d2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -242,6 +242,13 @@ int pqm_create_queue(struct process_queue_manager *pqm, enum kfd_queue_type type = properties->type; unsigned int max_queues = 127; /* HWS limit */ + /* + * On GFX 9.4.3, increase the number of queues that + * can be created to 255. No HWS limit on GFX 9.4.3. + */ + if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) + max_queues = 255; + q = NULL; kq = NULL; -- cgit v1.2.3 From 610dab118ff5013d46069c828b58d576e0907b66 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Fri, 31 Mar 2023 11:13:40 -0400 Subject: drm/amdkfd: Move pgmap to amdgpu_kfd_dev structure VRAM pgmap resource is allocated every time when switching compute partitions because kfd_dev is re-initialized by post_partition_switch, As a result, it causes memory region resource leaking and system memory usage accounting unbalanced. pgmap resource should be allocated and registered only once when loading driver and freed when unloading driver, move it from kfd_dev to amdgpu_kfd_dev. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 4 ++++ drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 8 ++++---- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 3 --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 6 +++--- drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 4 ++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2 +- 6 files changed, 14 insertions(+), 13 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index d1d643a050a1..e4e1dbba060a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -30,6 +30,7 @@ #include #include #include +#include #include #include #include "amdgpu_sync.h" @@ -101,6 +102,9 @@ struct amdgpu_kfd_dev { uint64_t vram_used_aligned; bool init_complete; struct work_struct reset_work; + + /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ + struct dev_pagemap pgmap; }; enum kgd_engine_type { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 42e599912e52..199d32c7c289 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -206,7 +206,7 @@ svm_migrate_copy_done(struct amdgpu_device *adev, struct dma_fence *mfence) unsigned long svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr) { - return (addr + adev->kfd.dev->pgmap.range.start) >> PAGE_SHIFT; + return (addr + adev->kfd.pgmap.range.start) >> PAGE_SHIFT; } static void @@ -236,7 +236,7 @@ svm_migrate_addr(struct amdgpu_device *adev, struct page *page) unsigned long addr; addr = page_to_pfn(page) << PAGE_SHIFT; - return (addr - adev->kfd.dev->pgmap.range.start); + return (addr - adev->kfd.pgmap.range.start); } static struct page * @@ -990,14 +990,14 @@ static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { int svm_migrate_init(struct amdgpu_device *adev) { - struct kfd_dev *kfddev = adev->kfd.dev; + struct amdgpu_kfd_dev *kfddev = &adev->kfd; struct dev_pagemap *pgmap; struct resource *res = NULL; unsigned long size; void *r; /* Page migration works on Vega10 or newer */ - if (!KFD_IS_SOC15(kfddev)) + if (!KFD_IS_SOC15(kfddev->dev)) return -EINVAL; pgmap = &kfddev->pgmap; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 02a90fd7f646..214d950f948e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -378,9 +378,6 @@ struct kfd_dev { int noretry; - /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ - struct dev_pagemap pgmap; - struct kfd_node *nodes[MAX_KFD_NODES]; unsigned int num_nodes; }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index cf354f9e4285..2b2129dd1e4a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -174,7 +174,7 @@ svm_range_dma_map_dev(struct amdgpu_device *adev, struct svm_range *prange, addr[i] = (hmm_pfns[i] << PAGE_SHIFT) + bo_adev->vm_manager.vram_base_offset - - bo_adev->kfd.dev->pgmap.range.start; + bo_adev->kfd.pgmap.range.start; addr[i] |= SVM_RANGE_VRAM_DOMAIN; pr_debug_ratelimited("vram address: 0x%llx\n", addr[i]); continue; @@ -2827,7 +2827,7 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, bool migration = false; int r = 0; - if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) { + if (!KFD_IS_SVM_API_SUPPORTED(adev)) { pr_debug("device does not support SVM\n"); return -EFAULT; } @@ -3112,7 +3112,7 @@ int svm_range_list_init(struct kfd_process *p) spin_lock_init(&svms->deferred_list_lock); for (i = 0; i < p->n_pdds; i++) - if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->kfd)) + if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev)) bitmap_set(svms->bitmap_supported, i, 1); return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index 7515ddade3ae..021def496f5a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -200,8 +200,8 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s /* SVM API and HMM page migration work together, device memory type * is initialized to not 0 when page migration register device memory. */ -#define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0 ||\ - (dev)->adev->gmc.is_app_apu) +#define KFD_IS_SVM_API_SUPPORTED(adev) ((adev)->kfd.pgmap.type != 0 ||\ + (adev)->gmc.is_app_apu) void svm_range_bo_unref_async(struct svm_range_bo *svm_bo); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index d2a42b6b1fa8..6d6243b978e1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -2021,7 +2021,7 @@ int kfd_topology_add_device(struct kfd_node *gpu) dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ? HSA_CAP_RASEVENTNOTIFY : 0; - if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev->kfd.dev)) + if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev)) dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED; kfd_debug_print_topology(); -- cgit v1.2.3 From 895797d9193b38e759bc01268a8e3887e521f682 Mon Sep 17 00:00:00 2001 From: Graham Sider Date: Mon, 6 Feb 2023 14:04:42 -0500 Subject: drm/amdgpu/bu: Add use_mtype_cc_wa module param By default, set use_mtype_cc_wa to 1 to set PTE coherence flag MTYPE_CC instead of MTYPE_RW by default. This is required for the time being to mitigate a bug causing XCCs to hit stale data due to TCC marking fully dirty lines as exclusive. Signed-off-by: Graham Sider Reviewed-by: Joseph Greathouse Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 7 +++++++ drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 10 +++++++--- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 7 +++++-- 4 files changed, 20 insertions(+), 5 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index cb9373f8c25a..cd2a29a7e26d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -212,6 +212,7 @@ extern int amdgpu_noretry; extern int amdgpu_force_asic_type; extern int amdgpu_smartshift_bias; extern int amdgpu_use_xgmi_p2p; +extern bool amdgpu_use_mtype_cc_wa; #ifdef CONFIG_HSA_AMD extern int sched_policy; extern bool debug_evictions; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index da4e50aef95a..8bc37826a99f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -822,6 +822,13 @@ MODULE_PARM_DESC(no_queue_eviction_on_vm_fault, "No queue eviction on VM fault ( module_param_named(no_queue_eviction_on_vm_fault, amdgpu_no_queue_eviction_on_vm_fault, int, 0444); #endif +/** + * DOC: use_mtype_cc_wa (bool) + */ +bool amdgpu_use_mtype_cc_wa = true; +MODULE_PARM_DESC(use_mtype_cc_wa, "Use MTYPE_CC workaround (0 = use MTYPE_RW where applicable, 1 = use MTYPE_CC where applicable (default))"); +module_param_named(use_mtype_cc_wa, amdgpu_use_mtype_cc_wa, bool, 0444); + /** * DOC: pcie_p2p (bool) * Enable PCIe P2P (requires large-BAR). Default value: true (on) diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 2eb67b53e497..8623b93c05ee 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1187,6 +1187,7 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT; bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED; unsigned int mtype; + unsigned int mtype_default; bool snoop = false; switch (adev->ip_versions[GC_HWIP][0]) { @@ -1230,7 +1231,10 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, /* FIXME: Needs more work for handling multiple memory * partitions (> NPS1 mode) e.g. NPS4 for both APU and dGPU * modes. + * FIXME: Temporarily using MTYPE_CC instead of MTYPE_RW where applicable. + * To force use of MTYPE_RW, set use_mtype_cc_wa=0 */ + mtype_default = amdgpu_use_mtype_cc_wa ? MTYPE_CC : MTYPE_RW; snoop = true; if (uncached) { mtype = MTYPE_UC; @@ -1245,14 +1249,14 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, * socket should be treated as remote access so MTYPE_RW * cannot be used always. */ - mtype = MTYPE_RW; + mtype = mtype_default; } else if (adev->flags & AMD_IS_APU) { /* APU on carve out mode */ - mtype = MTYPE_RW; + mtype = mtype_default; } else { /* dGPU */ if (is_vram && bo_adev == adev) - mtype = MTYPE_RW; + mtype = mtype_default; else if (is_vram) mtype = MTYPE_NC; else diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 2b2129dd1e4a..477ef9294203 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1198,9 +1198,12 @@ svm_range_get_pte_flags(struct kfd_node *node, if (uncached) { mapping_flags |= AMDGPU_VM_MTYPE_UC; } else if (domain == SVM_RANGE_VRAM_DOMAIN) { - /* local HBM region close to partition */ + /* local HBM region close to partition + * FIXME: Temporarily using MTYPE_CC instead of MTYPE_RW where applicable. + * To force use of MTYPE_RW, set use_mtype_cc_wa=0 + */ if (bo_node == node) - mapping_flags |= AMDGPU_VM_MTYPE_RW; + mapping_flags |= amdgpu_use_mtype_cc_wa ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; /* local HBM region far from partition or remote XGMI GPU */ else if (svm_nodes_in_same_hive(bo_node, node)) mapping_flags |= AMDGPU_VM_MTYPE_NC; -- cgit v1.2.3 From 1e4a00334add40f609162914af7a24bc92951008 Mon Sep 17 00:00:00 2001 From: Felix Kuehling Date: Tue, 21 Feb 2023 17:31:32 -0500 Subject: drm/amdgpu: Fix per-BO MTYPE selection for GFXv9.4.3 Treat system memory on NUMA systems as remote by default. Overriding with a more efficient MTYPE per page will be implemented in the next patch. No need for a special case for APP APUs. System memory is handled the same for carve-out and native mode. And VRAM doesn't exist in native mode. Signed-off-by: Felix Kuehling Reviewed-by: Philip Yang Reviewed-and-tested-by: Rajneesh Bhardwaj Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 40 ++++++++++++++--------------------- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 24 ++++++++++++--------- 2 files changed, 30 insertions(+), 34 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 8623b93c05ee..cf976b5b7b63 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1186,9 +1186,10 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM; bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT; bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED; - unsigned int mtype; - unsigned int mtype_default; + /* TODO: memory partitions struct amdgpu_vm *vm = mapping->bo_va->base.vm;*/ + unsigned int mtype_local, mtype; bool snoop = false; + bool is_local; switch (adev->ip_versions[GC_HWIP][0]) { case IP_VERSION(9, 4, 1): @@ -1228,35 +1229,26 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, } break; case IP_VERSION(9, 4, 3): - /* FIXME: Needs more work for handling multiple memory - * partitions (> NPS1 mode) e.g. NPS4 for both APU and dGPU - * modes. - * FIXME: Temporarily using MTYPE_CC instead of MTYPE_RW where applicable. - * To force use of MTYPE_RW, set use_mtype_cc_wa=0 + /* Only local VRAM BOs or system memory on non-NUMA APUs + * can be assumed to be local in their entirety. Choose + * MTYPE_NC as safe fallback for all system memory BOs on + * NUMA systems. Their MTYPE can be overridden per-page in + * gmc_v9_0_override_vm_pte_flags. */ - mtype_default = amdgpu_use_mtype_cc_wa ? MTYPE_CC : MTYPE_RW; + mtype_local = amdgpu_use_mtype_cc_wa ? MTYPE_CC : MTYPE_RW; + is_local = (!is_vram && (adev->flags & AMD_IS_APU) && + num_possible_nodes() <= 1) || + (is_vram && adev == bo_adev /* TODO: memory partitions && + bo->mem_id == vm->mem_id*/); snoop = true; if (uncached) { mtype = MTYPE_UC; - } else if (adev->gmc.is_app_apu) { - /* FIXME: APU in native mode, NPS1 single socket only - * - * For suporting NUMA partitioned APU e.g. in NPS4 mode, - * this need to look at the NUMA node on which the - * system memory allocation was done. - * - * Memory access by a different partition within same - * socket should be treated as remote access so MTYPE_RW - * cannot be used always. - */ - mtype = mtype_default; } else if (adev->flags & AMD_IS_APU) { - /* APU on carve out mode */ - mtype = mtype_default; + mtype = is_local ? mtype_local : MTYPE_NC; } else { /* dGPU */ - if (is_vram && bo_adev == adev) - mtype = mtype_default; + if (is_local) + mtype = mtype_local; else if (is_vram) mtype = MTYPE_NC; else diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 477ef9294203..4eec75b28917 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1151,6 +1151,7 @@ svm_range_get_pte_flags(struct kfd_node *node, bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT; bool uncached = flags & KFD_IOCTL_SVM_FLAG_UNCACHED; + unsigned int mtype_local; if (domain == SVM_RANGE_VRAM_DOMAIN) bo_node = prange->svm_bo->node; @@ -1191,19 +1192,16 @@ svm_range_get_pte_flags(struct kfd_node *node, } break; case IP_VERSION(9, 4, 3): - //TODO: Need more work for handling multiple memory partitions - //e.g. NPS4. Current approch is only applicable without memory - //partitions. + mtype_local = amdgpu_use_mtype_cc_wa ? AMDGPU_VM_MTYPE_CC : + AMDGPU_VM_MTYPE_RW; snoop = true; if (uncached) { mapping_flags |= AMDGPU_VM_MTYPE_UC; } else if (domain == SVM_RANGE_VRAM_DOMAIN) { - /* local HBM region close to partition - * FIXME: Temporarily using MTYPE_CC instead of MTYPE_RW where applicable. - * To force use of MTYPE_RW, set use_mtype_cc_wa=0 - */ - if (bo_node == node) - mapping_flags |= amdgpu_use_mtype_cc_wa ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW; + /* local HBM region close to partition */ + if (bo_node->adev == node->adev /* TODO: memory partitions && + bo_node->mem_id == node->mem_id*/) + mapping_flags |= mtype_local; /* local HBM region far from partition or remote XGMI GPU */ else if (svm_nodes_in_same_hive(bo_node, node)) mapping_flags |= AMDGPU_VM_MTYPE_NC; @@ -1212,7 +1210,13 @@ svm_range_get_pte_flags(struct kfd_node *node, mapping_flags |= AMDGPU_VM_MTYPE_UC; /* system memory accessed by the APU */ } else if (node->adev->flags & AMD_IS_APU) { - mapping_flags |= AMDGPU_VM_MTYPE_NC; + /* On NUMA systems, locality is determined per-page + * in amdgpu_gmc_override_vm_pte_flags + */ + if (num_possible_nodes() <= 1) + mapping_flags |= mtype_local; + else + mapping_flags |= AMDGPU_VM_MTYPE_NC; /* system memory accessed by the dGPU */ } else { mapping_flags |= AMDGPU_VM_MTYPE_UC; -- cgit v1.2.3 From 76eb9c95a409ea820b2e7c968c220e7a38f27d76 Mon Sep 17 00:00:00 2001 From: David Francis Date: Mon, 27 Feb 2023 10:33:11 -0500 Subject: drm/amdgpu/bu: add mtype_local as a module parameter Selects the MTYPE to be used for local memory, (0 = MTYPE_CC (default), 1 = MTYPE_NC, 2 = MTYPE_RW) v2: squash in build fix (Alex) Reviewed-by: Graham Sider Signed-off-by: David Francis Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 8 ++++---- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 19 ++++++++++++++++--- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 3 +-- 4 files changed, 22 insertions(+), 10 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index cd2a29a7e26d..c2feaf2fd070 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -212,7 +212,7 @@ extern int amdgpu_noretry; extern int amdgpu_force_asic_type; extern int amdgpu_smartshift_bias; extern int amdgpu_use_xgmi_p2p; -extern bool amdgpu_use_mtype_cc_wa; +extern int amdgpu_mtype_local; #ifdef CONFIG_HSA_AMD extern int sched_policy; extern bool debug_evictions; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index 8bc37826a99f..706ba4af062f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -823,11 +823,11 @@ module_param_named(no_queue_eviction_on_vm_fault, amdgpu_no_queue_eviction_on_vm #endif /** - * DOC: use_mtype_cc_wa (bool) + * DOC: mtype_local (int) */ -bool amdgpu_use_mtype_cc_wa = true; -MODULE_PARM_DESC(use_mtype_cc_wa, "Use MTYPE_CC workaround (0 = use MTYPE_RW where applicable, 1 = use MTYPE_CC where applicable (default))"); -module_param_named(use_mtype_cc_wa, amdgpu_use_mtype_cc_wa, bool, 0444); +int amdgpu_mtype_local; +MODULE_PARM_DESC(mtype_local, "MTYPE for local memory (0 = MTYPE_CC (default), 1 = MTYPE_NC, 2 = MTYPE_RW)"); +module_param_named(mtype_local, amdgpu_mtype_local, int, 0444); /** * DOC: pcie_p2p (bool) diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index c64a69f75da2..5a1414300271 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1235,7 +1235,16 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, * NUMA systems. Their MTYPE can be overridden per-page in * gmc_v9_0_override_vm_pte_flags. */ - mtype_local = amdgpu_use_mtype_cc_wa ? MTYPE_CC : MTYPE_RW; + mtype_local = MTYPE_CC; + if (amdgpu_mtype_local == 1) { + DRM_INFO_ONCE("Using MTYPE_NC for local memory\n"); + mtype_local = MTYPE_NC; + } else if (amdgpu_mtype_local == 2) { + DRM_INFO_ONCE("Using MTYPE_RW for local memory\n"); + mtype_local = MTYPE_RW; + } else { + DRM_INFO_ONCE("Using MTYPE_CC for local memory\n"); + } is_local = (!is_vram && (adev->flags & AMD_IS_APU) && num_possible_nodes() <= 1) || (is_vram && adev == bo_adev /* TODO: memory partitions && @@ -1349,9 +1358,13 @@ static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev, dev_dbg(adev->dev, "vm->mem_id=%d, local_node=%d, nid=%d\n", /*vm->mem_id*/0, local_node, nid); if (nid == local_node) { - unsigned int mtype_local = - amdgpu_use_mtype_cc_wa ? MTYPE_CC : MTYPE_RW; uint64_t old_flags = *flags; + unsigned int mtype_local = MTYPE_CC; + + if (amdgpu_mtype_local == 1) + mtype_local = MTYPE_NC; + else if (amdgpu_mtype_local == 2) + mtype_local = MTYPE_RW; *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | AMDGPU_PTE_MTYPE_VG10(mtype_local); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 4eec75b28917..df0ed5677609 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1192,8 +1192,7 @@ svm_range_get_pte_flags(struct kfd_node *node, } break; case IP_VERSION(9, 4, 3): - mtype_local = amdgpu_use_mtype_cc_wa ? AMDGPU_VM_MTYPE_CC : - AMDGPU_VM_MTYPE_RW; + mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC : (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_RW : AMDGPU_VM_MTYPE_CC); snoop = true; if (uncached) { mapping_flags |= AMDGPU_VM_MTYPE_UC; -- cgit v1.2.3 From b9cbd51000ad3541351ca832b00600870ac08e5c Mon Sep 17 00:00:00 2001 From: Graham Sider Date: Mon, 6 Mar 2023 17:56:44 -0500 Subject: drm/amdgpu/bu: update mtype_local parameter settings Update mtype_local module parameter to use MTYPE_RW by default. 0: MTYPE_RW (default) 1: MTYPE_NC 2: MTYPE_CC Signed-off-by: Graham Sider Reviewed-by: Harish Kasiviswanathan Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 2 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 12 ++++++------ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 3 ++- 3 files changed, 9 insertions(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index 706ba4af062f..aa466a9eb956 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -826,7 +826,7 @@ module_param_named(no_queue_eviction_on_vm_fault, amdgpu_no_queue_eviction_on_vm * DOC: mtype_local (int) */ int amdgpu_mtype_local; -MODULE_PARM_DESC(mtype_local, "MTYPE for local memory (0 = MTYPE_CC (default), 1 = MTYPE_NC, 2 = MTYPE_RW)"); +MODULE_PARM_DESC(mtype_local, "MTYPE for local memory (0 = MTYPE_RW (default), 1 = MTYPE_NC, 2 = MTYPE_CC)"); module_param_named(mtype_local, amdgpu_mtype_local, int, 0444); /** diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 5a1414300271..32eb4f4f5492 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1235,15 +1235,15 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, * NUMA systems. Their MTYPE can be overridden per-page in * gmc_v9_0_override_vm_pte_flags. */ - mtype_local = MTYPE_CC; + mtype_local = MTYPE_RW; if (amdgpu_mtype_local == 1) { DRM_INFO_ONCE("Using MTYPE_NC for local memory\n"); mtype_local = MTYPE_NC; } else if (amdgpu_mtype_local == 2) { - DRM_INFO_ONCE("Using MTYPE_RW for local memory\n"); - mtype_local = MTYPE_RW; - } else { DRM_INFO_ONCE("Using MTYPE_CC for local memory\n"); + mtype_local = MTYPE_CC; + } else { + DRM_INFO_ONCE("Using MTYPE_RW for local memory\n"); } is_local = (!is_vram && (adev->flags & AMD_IS_APU) && num_possible_nodes() <= 1) || @@ -1359,12 +1359,12 @@ static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev, /*vm->mem_id*/0, local_node, nid); if (nid == local_node) { uint64_t old_flags = *flags; - unsigned int mtype_local = MTYPE_CC; + unsigned int mtype_local = MTYPE_RW; if (amdgpu_mtype_local == 1) mtype_local = MTYPE_NC; else if (amdgpu_mtype_local == 2) - mtype_local = MTYPE_RW; + mtype_local = MTYPE_CC; *flags = (*flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | AMDGPU_PTE_MTYPE_VG10(mtype_local); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index df0ed5677609..e6348d4133fd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1192,7 +1192,8 @@ svm_range_get_pte_flags(struct kfd_node *node, } break; case IP_VERSION(9, 4, 3): - mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC : (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_RW : AMDGPU_VM_MTYPE_CC); + mtype_local = amdgpu_mtype_local == 1 ? AMDGPU_VM_MTYPE_NC : + (amdgpu_mtype_local == 2 ? AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW); snoop = true; if (uncached) { mapping_flags |= AMDGPU_VM_MTYPE_UC; -- cgit v1.2.3 From a476c0c645535cc0361938becb440b4239996079 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 23 Feb 2023 11:03:37 -0500 Subject: drm/amdkfd: Store drm node minor number for kfd nodes From KFD topology, application will find kfd node with the corresponding drm device node minor number, for example if partition drm node starts from /dev/dri/renderD129, then KFD node 0 with store drm node minor number 129. Application will open drm node /dev/dri/renderD129 to create amdgpu vm for kfd node 0 with the correct vm->mem_id to indicate the memory partition. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 6d6243b978e1..a8e25aecf839 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1942,8 +1942,12 @@ int kfd_topology_add_device(struct kfd_node *gpu) amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev); dev->node_props.max_engine_clk_ccompute = cpufreq_quick_get_max(0) / 1000; - dev->node_props.drm_render_minor = - gpu->kfd->shared_resources.drm_render_minor; + + if (gpu->xcp) + dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index; + else + dev->node_props.drm_render_minor = + gpu->kfd->shared_resources.drm_render_minor; dev->node_props.hive_id = gpu->kfd->hive_id; dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu); -- cgit v1.2.3 From 4c6ce75fdd628c43aea11448ed41b52119dae42b Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 26 Jan 2023 18:11:29 -0500 Subject: drm/amdkfd: Show KFD node memory partition info Show KFD node memory partition id and size, add helper function KFD_XCP_MEMORY_SIZE to get kfd node memory size, will be used later to support memory accounting per partition. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 5 +++++ drivers/gpu/drm/amd/amdkfd/kfd_device.c | 7 ++++++- 2 files changed, 11 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index e4e1dbba060a..324cb566ca2f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -330,6 +330,11 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, uint64_t size, u32 alloc_flag); +#define KFD_XCP_MEMORY_SIZE(n) ((n)->adev->gmc.num_mem_partitions ?\ + (n)->adev->gmc.mem_partitions[(n)->xcp->mem_id].size /\ + (n)->adev->xcp_mgr->num_xcp_per_mem_partition :\ + (n)->adev->gmc.real_vram_size) + #if IS_ENABLED(CONFIG_HSA_AMD) void amdgpu_amdkfd_gpuvm_init_mem_limits(void); void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index b5497d2ee984..db5b53fcdf11 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -724,7 +724,6 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd_cwsr_init(kfd); - /* TODO: Needs to be updated for memory partitioning */ svm_migrate_init(kfd->adev); amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); @@ -754,6 +753,12 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, (1U << NUM_XCC(kfd->adev->gfx.xcc_mask)) - 1; } + if (node->xcp) { + dev_info(kfd_device, "KFD node %d partition %d size %lldM\n", + node->node_id, node->xcp->mem_id, + KFD_XCP_MEMORY_SIZE(node) >> 20); + } + if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && partition_mode == AMDGPU_CPX_PARTITION_MODE && kfd->num_nodes != 1) { -- cgit v1.2.3 From 2046ed6c8aa951e4ae83c5022bb0a7c777386097 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 26 Jan 2023 18:45:32 -0500 Subject: drm/amdkfd: SVM range allocation support memory partition Pass kfd node->xcp->mem_id to amdgpu bo create parameter mem_id_plus1 to allocate new svm_bo on the specified memory partition. This is only for dGPU mode as we don't migrate with APU mode. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index e6348d4133fd..62aa7fb2eaa5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -555,16 +555,20 @@ svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, bp.flags |= AMDGPU_GEM_CREATE_DISCARDABLE; bp.type = ttm_bo_type_device; bp.resv = NULL; + if (node->xcp) + bp.mem_id_plus1 = node->xcp->mem_id + 1; - /* TODO: Allocate memory from the right memory partition. We can sort - * out the details later, once basic memory partitioning is working - */ r = amdgpu_bo_create_user(node->adev, &bp, &ubo); if (r) { pr_debug("failed %d to create bo\n", r); goto create_bo_failed; } bo = &ubo->bo; + + pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n", + bo->tbo.resource->start << PAGE_SHIFT, bp.size, + bp.mem_id_plus1 - 1); + r = amdgpu_bo_reserve(bo, true); if (r) { pr_debug("failed %d to reserve bo\n", r); -- cgit v1.2.3 From dc12f9eddedb8b41f4dc948e5e636e5221fb4d43 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 2 Feb 2023 11:07:53 -0500 Subject: drm/amdkfd: Update MTYPE for far memory partition Use MTYPE RW/MTYPE_CC for mapping system memory or VRAM to KFD node within the same memory partition, use MTYPE_NC for mapping on KFD node from the far memory partition of the same socket or from another socket on same XGMI hive. On NPS4 or 4P system, MTYPE will be overridden per page depending on the memory NUMA node id and vm->mem_id. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 15 +++++++-------- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 9 +++++---- 2 files changed, 12 insertions(+), 12 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 32eb4f4f5492..263d17a8b433 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1186,7 +1186,7 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, bool is_vram = bo->tbo.resource->mem_type == TTM_PL_VRAM; bool coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT; bool uncached = bo->flags & AMDGPU_GEM_CREATE_UNCACHED; - /* TODO: memory partitions struct amdgpu_vm *vm = mapping->bo_va->base.vm;*/ + struct amdgpu_vm *vm = mapping->bo_va->base.vm; unsigned int mtype_local, mtype; bool snoop = false; bool is_local; @@ -1247,8 +1247,8 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, } is_local = (!is_vram && (adev->flags & AMD_IS_APU) && num_possible_nodes() <= 1) || - (is_vram && adev == bo_adev /* TODO: memory partitions && - bo->mem_id == vm->mem_id*/); + (is_vram && adev == bo_adev && + bo->mem_id == vm->mem_id); snoop = true; if (uncached) { mtype = MTYPE_UC; @@ -1335,13 +1335,12 @@ static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev, return; } - /* TODO: memory partitions. mem_id is hard-coded to 0 for now. - * FIXME: Only supported on native mode for now. For carve-out, the + /* FIXME: Only supported on native mode for now. For carve-out, the * NUMA affinity of the GPU/VM needs to come from the PCI info because * memory partitions are not associated with different NUMA nodes. */ - if (adev->gmc.is_app_apu) { - local_node = adev->gmc.mem_partitions[/*vm->mem_id*/0].numa.node; + if (adev->gmc.is_app_apu && vm->mem_id >= 0) { + local_node = adev->gmc.mem_partitions[vm->mem_id].numa.node; } else { dev_dbg(adev->dev, "Only native mode APU is supported.\n"); return; @@ -1356,7 +1355,7 @@ static void gmc_v9_0_override_vm_pte_flags(struct amdgpu_device *adev, } nid = pfn_to_nid(addr >> PAGE_SHIFT); dev_dbg(adev->dev, "vm->mem_id=%d, local_node=%d, nid=%d\n", - /*vm->mem_id*/0, local_node, nid); + vm->mem_id, local_node, nid); if (nid == local_node) { uint64_t old_flags = *flags; unsigned int mtype_local = MTYPE_RW; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 62aa7fb2eaa5..a700d9ccd054 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1203,8 +1203,8 @@ svm_range_get_pte_flags(struct kfd_node *node, mapping_flags |= AMDGPU_VM_MTYPE_UC; } else if (domain == SVM_RANGE_VRAM_DOMAIN) { /* local HBM region close to partition */ - if (bo_node->adev == node->adev /* TODO: memory partitions && - bo_node->mem_id == node->mem_id*/) + if (bo_node->adev == node->adev && + (!bo_node->xcp || !node->xcp || bo_node->xcp->mem_id == node->xcp->mem_id)) mapping_flags |= mtype_local; /* local HBM region far from partition or remote XGMI GPU */ else if (svm_nodes_in_same_hive(bo_node, node)) @@ -1358,8 +1358,9 @@ svm_range_map_to_gpu(struct kfd_process_device *pdd, struct svm_range *prange, (last_domain == SVM_RANGE_VRAM_DOMAIN) ? 1 : 0, pte_flags); - /* TODO: we still need to determine the vm_manager.vram_base_offset based on - * the memory partition. + /* For dGPU mode, we use same vm_manager to allocate VRAM for + * different memory partition based on fpfn/lpfn, we should use + * same vm_manager.vram_base_offset regardless memory partition. */ r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb, NULL, last_start, prange->start + i, -- cgit v1.2.3 From 3ebfd221c1a83e5f0edadb87d173d8fd93d1d125 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Wed, 8 Mar 2023 11:57:00 -0500 Subject: drm/amdkfd: Store xcp partition id to amdgpu bo For memory accounting per compute partition and export drm amdgpu bo and then import to KFD, we need the xcp id to account the memory usage or find the KFD node of the original amdgpu bo to create the KFD bo on the correct adev KFD node. Set xcp_id_plus1 of amdgpu_bo_param to create bo and store xcp_id to amddgpu bo. Add helper macro to get the mem_id from adev and xcp_id. v2: squash in fix ("drm/amdgpu: Fix BO creation failure on GFX 9.4.3 dGPU") Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 4 ++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 11 ++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 4 ++-- drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 15 ++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 12 ++++++++---- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 6 +++++- drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c | 5 +++-- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 ++-- 10 files changed, 42 insertions(+), 23 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 324cb566ca2f..05c54776951b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -330,6 +330,10 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, uint64_t size, u32 alloc_flag); +#define KFD_XCP_MEM_ID(adev, xcp_id) \ + ((adev)->xcp_mgr && (xcp_id) >= 0 ?\ + (adev)->xcp_mgr->xcp[(xcp_id)].mem_id : -1) + #define KFD_XCP_MEMORY_SIZE(n) ((n)->adev->gmc.num_mem_partitions ?\ (n)->adev->gmc.mem_partitions[(n)->xcp->mem_id].size /\ (n)->adev->xcp_mgr->num_xcp_per_mem_partition :\ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index c234dc0db799..8724a0be31b8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -1634,6 +1634,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( uint64_t *offset, uint32_t flags, bool criu_resume) { struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv); + struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm); enum ttm_bo_type bo_type = ttm_bo_type_device; struct sg_table *sg = NULL; uint64_t user_addr = 0; @@ -1641,7 +1642,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( struct drm_gem_object *gobj = NULL; u32 domain, alloc_domain; uint64_t aligned_size; - int8_t mem_id = -1; + int8_t xcp_id = -1; u64 alloc_flags; int ret; @@ -1660,7 +1661,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0; } - mem_id = avm->mem_id; + xcp_id = fpriv->xcp_id == ~0 ? 0 : fpriv->xcp_id; } else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) { domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT; alloc_flags = 0; @@ -1718,12 +1719,12 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( goto err_reserve_limit; } - pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s mem_id %d\n", + pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n", va, (*mem)->aql_queue ? size << 1 : size, - domain_string(alloc_domain), mem_id); + domain_string(alloc_domain), xcp_id); ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags, - bo_type, NULL, &gobj, mem_id + 1); + bo_type, NULL, &gobj, xcp_id + 1); if (ret) { pr_debug("Failed to create BO on domain %s. ret %d\n", domain_string(alloc_domain), ret); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 33ebee18b80d..7e8839cc6f58 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -98,7 +98,7 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size, int alignment, u32 initial_domain, u64 flags, enum ttm_bo_type type, struct dma_resv *resv, - struct drm_gem_object **obj, int8_t mem_id_plus1) + struct drm_gem_object **obj, int8_t xcp_id_plus1) { struct amdgpu_bo *bo; struct amdgpu_bo_user *ubo; @@ -116,7 +116,7 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size, bp.flags = flags; bp.domain = initial_domain; bp.bo_ptr_size = sizeof(struct amdgpu_bo); - bp.mem_id_plus1 = mem_id_plus1; + bp.xcp_id_plus1 = xcp_id_plus1; r = amdgpu_bo_create_user(adev, &bp, &ubo); if (r) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h index 646c4fcc8e40..f30264782ba2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.h @@ -43,7 +43,7 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size, int alignment, u32 initial_domain, u64 flags, enum ttm_bo_type type, struct dma_resv *resv, - struct drm_gem_object **obj, int8_t mem_id_plus1); + struct drm_gem_object **obj, int8_t xcp_id_plus1); int amdgpu_mode_dumb_create(struct drm_file *file_priv, struct drm_device *dev, struct drm_mode_create_dumb *args); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c index b2d11c4f39b0..42c02f48c3a8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c @@ -131,14 +131,15 @@ void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain) if (domain & AMDGPU_GEM_DOMAIN_VRAM) { unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT; + int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id); - if (adev->gmc.mem_partitions && abo->mem_id >= 0) { - places[c].fpfn = adev->gmc.mem_partitions[abo->mem_id].range.fpfn; + if (adev->gmc.mem_partitions && mem_id >= 0) { + places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn; /* * memory partition range lpfn is inclusive start + size - 1 * TTM place lpfn is exclusive start + size */ - places[c].lpfn = adev->gmc.mem_partitions[abo->mem_id].range.lpfn + 1; + places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1; } else { places[c].fpfn = 0; places[c].lpfn = 0; @@ -583,8 +584,12 @@ int amdgpu_bo_create(struct amdgpu_device *adev, bo->flags = bp->flags; - /* bo->mem_id -1 means any partition */ - bo->mem_id = bp->mem_id_plus1 - 1; + if (adev->gmc.mem_partitions) + /* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */ + bo->xcp_id = bp->xcp_id_plus1 - 1; + else + /* For GPUs without spatial partitioning */ + bo->xcp_id = 0; if (!amdgpu_bo_support_uswc(bo->flags)) bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h index eb24a66ccee5..05496b97ef93 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h @@ -56,8 +56,8 @@ struct amdgpu_bo_param { bool no_wait_gpu; struct dma_resv *resv; void (*destroy)(struct ttm_buffer_object *bo); - /* memory partition number plus 1, 0 means any partition */ - int8_t mem_id_plus1; + /* xcp partition number plus 1, 0 means any partition */ + int8_t xcp_id_plus1; }; /* bo virtual addresses in a vm */ @@ -111,8 +111,12 @@ struct amdgpu_bo { #endif struct kgd_mem *kfd_bo; - /* memory partition number, -1 means any partition */ - int8_t mem_id; + /* + * For GPUs with spatial partitioning, xcp partition number, -1 means + * any partition. For other ASICs without spatial partition, always 0 + * for memory accounting. + */ + int8_t xcp_id; }; struct amdgpu_bo_user { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 129c593cb2bd..23101c82519a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1051,6 +1051,7 @@ static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev, static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags) { + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev); struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo); struct amdgpu_ttm_tt *gtt; enum ttm_caching caching; @@ -1060,7 +1061,10 @@ static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo, return NULL; } gtt->gobj = &bo->base; - gtt->pool_id = abo->mem_id; + if (adev->gmc.mem_partitions && abo->xcp_id >= 0) + gtt->pool_id = KFD_XCP_MEM_ID(adev, abo->xcp_id); + else + gtt->pool_id = abo->xcp_id; if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) caching = ttm_write_combined; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c index 62fc7e8d326e..cc3b1b596e56 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c @@ -502,6 +502,7 @@ exit: int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm, int level, bool immediate, struct amdgpu_bo_vm **vmbo) { + struct amdgpu_fpriv *fpriv = container_of(vm, struct amdgpu_fpriv, vm); struct amdgpu_bo_param bp; struct amdgpu_bo *bo; struct dma_resv *resv; @@ -534,7 +535,7 @@ int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm, bp.type = ttm_bo_type_kernel; bp.no_wait_gpu = immediate; - bp.mem_id_plus1 = vm->mem_id + 1; + bp.xcp_id_plus1 = fpriv->xcp_id == ~0 ? 0 : fpriv->xcp_id + 1; if (vm->root.bo) bp.resv = vm->root.bo->tbo.base.resv; @@ -560,7 +561,7 @@ int amdgpu_vm_pt_create(struct amdgpu_device *adev, struct amdgpu_vm *vm, bp.type = ttm_bo_type_kernel; bp.resv = bo->tbo.base.resv; bp.bo_ptr_size = sizeof(struct amdgpu_bo); - bp.mem_id_plus1 = vm->mem_id + 1; + bp.xcp_id_plus1 = fpriv->xcp_id == ~0 ? 0 : fpriv->xcp_id + 1; r = amdgpu_bo_create(adev, &bp, &(*vmbo)->shadow); diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 263d17a8b433..7ea80bdf8e1e 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1248,7 +1248,7 @@ static void gmc_v9_0_get_coherence_flags(struct amdgpu_device *adev, is_local = (!is_vram && (adev->flags & AMD_IS_APU) && num_possible_nodes() <= 1) || (is_vram && adev == bo_adev && - bo->mem_id == vm->mem_id); + KFD_XCP_MEM_ID(adev, bo->xcp_id) == vm->mem_id); snoop = true; if (uncached) { mtype = MTYPE_UC; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index a700d9ccd054..45959892bc0f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -556,7 +556,7 @@ svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, bp.type = ttm_bo_type_device; bp.resv = NULL; if (node->xcp) - bp.mem_id_plus1 = node->xcp->mem_id + 1; + bp.xcp_id_plus1 = node->xcp->id + 1; r = amdgpu_bo_create_user(node->adev, &bp, &ubo); if (r) { @@ -567,7 +567,7 @@ svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange, pr_debug("alloc bo at offset 0x%lx size 0x%lx on partition %d\n", bo->tbo.resource->start << PAGE_SHIFT, bp.size, - bp.mem_id_plus1 - 1); + bp.xcp_id_plus1 - 1); r = amdgpu_bo_reserve(bo, true); if (r) { -- cgit v1.2.3 From 2fa9ff25de08e598af051c76b216d2f073b2ee89 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 9 Mar 2023 19:30:02 -0500 Subject: drm/amdgpu: KFD graphics interop support compute partition kfd_ioctl_get_dmabuf use the amdgpu bo xcp_id to get the gpu_id of the KFD node from the exported dmabuf_adev, and then create kfd bo on the correct adev and KFD node when importing the amdgpu bo to KFD. Remove function kfd_device_by_adev, it is not needed as it is the same result as dmabuf_adev->kfd.dev->nodes[0]->id. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 4 +++- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 14 ++++++-------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 - drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 18 ------------------ 5 files changed, 10 insertions(+), 29 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index bbbfe9ec4adf..00edb13d2124 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -498,7 +498,7 @@ int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, struct amdgpu_device **dmabuf_adev, uint64_t *bo_size, void *metadata_buffer, size_t buffer_size, uint32_t *metadata_size, - uint32_t *flags) + uint32_t *flags, int8_t *xcp_id) { struct dma_buf *dma_buf; struct drm_gem_object *obj; @@ -542,6 +542,8 @@ int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) *flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC; } + if (xcp_id) + *xcp_id = bo->xcp_id; out_put: dma_buf_put(dma_buf); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 05c54776951b..4e6221bccffe 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -241,7 +241,7 @@ int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd, struct amdgpu_device **dmabuf_adev, uint64_t *bo_size, void *metadata_buffer, size_t buffer_size, uint32_t *metadata_size, - uint32_t *flags); + uint32_t *flags, int8_t *xcp_id); uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst, struct amdgpu_device *src); int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index f85ac4dbc673..fcad90d53c9b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1499,6 +1499,7 @@ static int kfd_ioctl_get_dmabuf_info(struct file *filep, struct amdgpu_device *dmabuf_adev; void *metadata_buffer = NULL; uint32_t flags; + int8_t xcp_id; unsigned int i; int r; @@ -1519,17 +1520,14 @@ static int kfd_ioctl_get_dmabuf_info(struct file *filep, r = amdgpu_amdkfd_get_dmabuf_info(dev->adev, args->dmabuf_fd, &dmabuf_adev, &args->size, metadata_buffer, args->metadata_size, - &args->metadata_size, &flags); + &args->metadata_size, &flags, &xcp_id); if (r) goto exit; - /* Reverse-lookup gpu_id from kgd pointer */ - dev = kfd_device_by_adev(dmabuf_adev); - if (!dev) { - r = -EINVAL; - goto exit; - } - args->gpu_id = dev->id; + if (xcp_id >= 0) + args->gpu_id = dmabuf_adev->kfd.dev->nodes[xcp_id]->id; + else + args->gpu_id = dmabuf_adev->kfd.dev->nodes[0]->id; args->flags = flags; /* Copy metadata buffer to user mode */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 214d950f948e..44f4d5509db6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1068,7 +1068,6 @@ struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); struct kfd_node *kfd_device_by_id(uint32_t gpu_id); struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev); -struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev); static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id, uint32_t vmid) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index a8e25aecf839..dbb6159344b3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -125,24 +125,6 @@ struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev) return device; } -struct kfd_node *kfd_device_by_adev(const struct amdgpu_device *adev) -{ - struct kfd_topology_device *top_dev; - struct kfd_node *device = NULL; - - down_read(&topology_lock); - - list_for_each_entry(top_dev, &topology_device_list, list) - if (top_dev->gpu && top_dev->gpu->adev == adev) { - device = top_dev->gpu; - break; - } - - up_read(&topology_lock); - - return device; -} - /* Called with write topology_lock acquired */ static void kfd_release_topology_device(struct kfd_topology_device *dev) { -- cgit v1.2.3 From 315e29eca57f85107cc6f687c2d510aa532fb3f0 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 20 Mar 2023 11:21:38 -0400 Subject: drm/amdkfd: Move local_mem_info to kfd_node We need to track memory usage on a per partition basis. To do that, store the local memory information in KFD node instead of kfd device. v2: squash in fix ("amdkfd: Use mem_id to access mem_partition info") Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 17 +++++++++++++---- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 12 +++++++----- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 ++++--- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 7 +++++-- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 7 ++++--- 7 files changed, 36 insertions(+), 19 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 00edb13d2124..85df73f2c85e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -428,14 +428,23 @@ uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, } void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, - struct kfd_local_mem_info *mem_info) + struct kfd_local_mem_info *mem_info, + uint8_t xcp_id) { memset(mem_info, 0, sizeof(*mem_info)); - mem_info->local_mem_size_public = adev->gmc.visible_vram_size; - mem_info->local_mem_size_private = adev->gmc.real_vram_size - + if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) { + if (adev->gmc.real_vram_size == adev->gmc.visible_vram_size) + mem_info->local_mem_size_public = + KFD_XCP_MEMORY_SIZE(adev, xcp_id); + else + mem_info->local_mem_size_private = + KFD_XCP_MEMORY_SIZE(adev, xcp_id); + } else { + mem_info->local_mem_size_public = adev->gmc.visible_vram_size; + mem_info->local_mem_size_private = adev->gmc.real_vram_size - adev->gmc.visible_vram_size; - + } mem_info->vram_width = adev->gmc.vram_width; pr_debug("Address base: %pap public 0x%llx private 0x%llx\n", diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 4e6221bccffe..4bf6f5659568 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -231,7 +231,8 @@ int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem); uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, enum kgd_engine_type type); void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, - struct kfd_local_mem_info *mem_info); + struct kfd_local_mem_info *mem_info, + uint8_t xcp_id); uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev); uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev); @@ -334,10 +335,11 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, ((adev)->xcp_mgr && (xcp_id) >= 0 ?\ (adev)->xcp_mgr->xcp[(xcp_id)].mem_id : -1) -#define KFD_XCP_MEMORY_SIZE(n) ((n)->adev->gmc.num_mem_partitions ?\ - (n)->adev->gmc.mem_partitions[(n)->xcp->mem_id].size /\ - (n)->adev->xcp_mgr->num_xcp_per_mem_partition :\ - (n)->adev->gmc.real_vram_size) +#define KFD_XCP_MEMORY_SIZE(adev, xcp_id)\ + ((adev)->gmc.num_mem_partitions && (xcp_id) >= 0 ?\ + (adev)->gmc.mem_partitions[KFD_XCP_MEM_ID((adev), (xcp_id))].size /\ + (adev)->xcp_mgr->num_xcp_per_mem_partition :\ + (adev)->gmc.real_vram_size) #if IS_ENABLED(CONFIG_HSA_AMD) void amdgpu_amdkfd_gpuvm_init_mem_limits(void); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index fcad90d53c9b..1ae867482bc7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1023,11 +1023,12 @@ bool kfd_dev_is_large_bar(struct kfd_node *dev) if (dev->kfd->use_iommu_v2) return false; - if (dev->kfd->local_mem_info.local_mem_size_private == 0 && - dev->kfd->local_mem_info.local_mem_size_public > 0) + if (dev->local_mem_info.local_mem_size_private == 0 && + dev->local_mem_info.local_mem_size_public > 0) return true; - if (dev->kfd->local_mem_info.local_mem_size_public == 0 && dev->kfd->adev->gmc.is_app_apu) { + if (dev->local_mem_info.local_mem_size_public == 0 && + dev->kfd->adev->gmc.is_app_apu) { pr_debug("APP APU, Consider like a large bar system\n"); return true; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index 1aaf933f9f48..950af6820153 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -2191,7 +2191,7 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image, * report the total FB size (public+private) as a single * private heap. */ - local_mem_info = kdev->kfd->local_mem_info; + local_mem_info = kdev->local_mem_info; sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr + sub_type_hdr->length); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index db5b53fcdf11..d41da964d2f5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -726,7 +726,6 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, svm_migrate_init(kfd->adev); - amdgpu_amdkfd_get_local_mem_info(kfd->adev, &kfd->local_mem_info); dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n", kfd->num_nodes); @@ -756,7 +755,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (node->xcp) { dev_info(kfd_device, "KFD node %d partition %d size %lldM\n", node->node_id, node->xcp->mem_id, - KFD_XCP_MEMORY_SIZE(node) >> 20); + KFD_XCP_MEMORY_SIZE(node->adev, node->node_id) >> 20); } if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && @@ -783,6 +782,10 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, } node->max_proc_per_quantum = max_proc_per_quantum; atomic_set(&node->sram_ecc_flag, 0); + + amdgpu_amdkfd_get_local_mem_info(kfd->adev, + &node->local_mem_info, node->xcp->id); + /* Initialize the KFD node */ if (kfd_init_node(node)) { dev_err(kfd_device, "Error initializing KFD node\n"); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 44f4d5509db6..3bd222e8f6c3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -313,6 +313,8 @@ struct kfd_node { unsigned int compute_vmid_bitmap; + struct kfd_local_mem_info local_mem_info; + struct kfd_dev *kfd; }; @@ -335,7 +337,6 @@ struct kfd_dev { */ struct kgd2kfd_shared_resources shared_resources; - struct kfd_local_mem_info local_mem_info; const struct kfd2kgd_calls *kfd2kgd; struct mutex doorbell_mutex; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index dbb6159344b3..e0bacf017a40 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1152,8 +1152,8 @@ static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) if (!gpu) return 0; - local_mem_size = gpu->kfd->local_mem_info.local_mem_size_private + - gpu->kfd->local_mem_info.local_mem_size_public; + local_mem_size = gpu->local_mem_info.local_mem_size_private + + gpu->local_mem_info.local_mem_size_public; buf[0] = gpu->adev->pdev->devfn; buf[1] = gpu->adev->pdev->subsystem_vendor | (gpu->adev->pdev->subsystem_device << 16); @@ -1234,7 +1234,8 @@ static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) * for APUs - If CRAT from ACPI reports more than one bank, then * all the banks will report the same mem_clk_max information */ - amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info); + amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info, + dev->gpu->xcp->id); list_for_each_entry(mem, &dev->mem_props, list) mem->mem_clk_max = local_mem_info.mem_clk_max; -- cgit v1.2.3 From 1c77527a69d5ca19cb276e2728992d922b687f35 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Mon, 20 Mar 2023 11:22:30 -0400 Subject: drm/amdkfd: Fix memory reporting on GFX 9.4.3 This patch fixes memory reporting on the GFX 9.4.3 APU and dGPU by reporting available memory on a per partition basis. If its an APU, available and used memory calculations take into account system and TTM memory. v2: squash in fix ("drm/amdkfd: Fix array out of bound warning") squash in fix ("drm/amdgpu: Update memory reporting for GFX9.4.3") Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 12 ++-- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 81 ++++++++++++++++++------ drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 14 ++-- 5 files changed, 84 insertions(+), 31 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 4bf6f5659568..948d362adabb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -35,6 +35,7 @@ #include #include "amdgpu_sync.h" #include "amdgpu_vm.h" +#include "amdgpu_xcp.h" extern uint64_t amdgpu_amdkfd_total_mem_size; @@ -98,8 +99,8 @@ struct amdgpu_amdkfd_fence { struct amdgpu_kfd_dev { struct kfd_dev *dev; - int64_t vram_used; - uint64_t vram_used_aligned; + int64_t vram_used[MAX_XCP]; + uint64_t vram_used_aligned[MAX_XCP]; bool init_complete; struct work_struct reset_work; @@ -287,7 +288,8 @@ int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev, void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev, void *drm_priv); uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv); -size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev); +size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, + uint8_t xcp_id); int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( struct amdgpu_device *adev, uint64_t va, uint64_t size, void *drm_priv, struct kgd_mem **mem, @@ -327,9 +329,9 @@ void amdgpu_amdkfd_block_mmu_notifications(void *p); int amdgpu_amdkfd_criu_resume(void *p); bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev); int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 alloc_flag); + uint64_t size, u32 alloc_flag, int8_t xcp_id); void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 alloc_flag); + uint64_t size, u32 alloc_flag, int8_t xcp_id); #define KFD_XCP_MEM_ID(adev, xcp_id) \ ((adev)->xcp_mgr && (xcp_id) >= 0 ?\ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 8724a0be31b8..cc37f04651e2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -157,12 +157,13 @@ void amdgpu_amdkfd_reserve_system_mem(uint64_t size) * Return: returns -ENOMEM in case of error, ZERO otherwise */ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 alloc_flag) + uint64_t size, u32 alloc_flag, int8_t xcp_id) { uint64_t reserved_for_pt = ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size); size_t system_mem_needed, ttm_mem_needed, vram_needed; int ret = 0; + uint64_t vram_size = 0; system_mem_needed = 0; ttm_mem_needed = 0; @@ -177,6 +178,17 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, * 2M BO chunk. */ vram_needed = size; + /* + * For GFX 9.4.3, get the VRAM size from XCP structs + */ + if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id)) + return -EINVAL; + + vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id); + if (adev->gmc.is_app_apu) { + system_mem_needed = size; + ttm_mem_needed = size; + } } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { system_mem_needed = size; } else if (!(alloc_flag & @@ -196,8 +208,8 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) || (kfd_mem_limit.ttm_mem_used + ttm_mem_needed > kfd_mem_limit.max_ttm_mem_limit) || - (adev && adev->kfd.vram_used + vram_needed > - adev->gmc.real_vram_size - reserved_for_pt)) { + (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed > + vram_size - reserved_for_pt)) { ret = -ENOMEM; goto release; } @@ -207,9 +219,11 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, */ WARN_ONCE(vram_needed && !adev, "adev reference can't be null when vram is used"); - if (adev) { - adev->kfd.vram_used += vram_needed; - adev->kfd.vram_used_aligned += ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN); + if (adev && xcp_id >= 0) { + adev->kfd.vram_used[xcp_id] += vram_needed; + adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ? + vram_needed : + ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN); } kfd_mem_limit.system_mem_used += system_mem_needed; kfd_mem_limit.ttm_mem_used += ttm_mem_needed; @@ -220,7 +234,7 @@ release: } void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, - uint64_t size, u32 alloc_flag) + uint64_t size, u32 alloc_flag, int8_t xcp_id) { spin_lock(&kfd_mem_limit.mem_limit_lock); @@ -230,9 +244,19 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) { WARN_ONCE(!adev, "adev reference can't be null when alloc mem flags vram is set"); + if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id)) + goto release; + if (adev) { - adev->kfd.vram_used -= size; - adev->kfd.vram_used_aligned -= ALIGN(size, VRAM_AVAILABLITY_ALIGN); + adev->kfd.vram_used[xcp_id] -= size; + if (adev->gmc.is_app_apu) { + adev->kfd.vram_used_aligned[xcp_id] -= size; + kfd_mem_limit.system_mem_used -= size; + kfd_mem_limit.ttm_mem_used -= size; + } else { + adev->kfd.vram_used_aligned[xcp_id] -= + ALIGN(size, VRAM_AVAILABLITY_ALIGN); + } } } else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) { kfd_mem_limit.system_mem_used -= size; @@ -242,8 +266,8 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag); goto release; } - WARN_ONCE(adev && adev->kfd.vram_used < 0, - "KFD VRAM memory accounting unbalanced"); + WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0, + "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id); WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0, "KFD TTM memory accounting unbalanced"); WARN_ONCE(kfd_mem_limit.system_mem_used < 0, @@ -259,7 +283,8 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo) u32 alloc_flags = bo->kfd_bo->alloc_flags; u64 size = amdgpu_bo_size(bo); - amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags); + amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags, + bo->xcp_id); kfree(bo->kfd_bo); } @@ -1609,23 +1634,42 @@ out_unlock: return ret; } -size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev) +size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev, + uint8_t xcp_id) { uint64_t reserved_for_pt = ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size); ssize_t available; + uint64_t vram_available, system_mem_available, ttm_mem_available; spin_lock(&kfd_mem_limit.mem_limit_lock); - available = adev->gmc.real_vram_size - - adev->kfd.vram_used_aligned + vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id) + - adev->kfd.vram_used_aligned[xcp_id] - atomic64_read(&adev->vram_pin_size) - reserved_for_pt; + + if (adev->gmc.is_app_apu) { + system_mem_available = no_system_mem_limit ? + kfd_mem_limit.max_system_mem_limit : + kfd_mem_limit.max_system_mem_limit - + kfd_mem_limit.system_mem_used; + + ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit - + kfd_mem_limit.ttm_mem_used; + + available = min3(system_mem_available, ttm_mem_available, + vram_available); + available = ALIGN_DOWN(available, PAGE_SIZE); + } else { + available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN); + } + spin_unlock(&kfd_mem_limit.mem_limit_lock); if (available < 0) available = 0; - return ALIGN_DOWN(available, VRAM_AVAILABLITY_ALIGN); + return available; } int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( @@ -1713,7 +1757,8 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( amdgpu_sync_create(&(*mem)->sync); - ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags); + ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags, + xcp_id); if (ret) { pr_debug("Insufficient memory\n"); goto err_reserve_limit; @@ -1781,7 +1826,7 @@ err_node_allow: /* Don't unreserve system mem limit twice */ goto err_reserve_limit; err_bo_create: - amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags); + amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id); err_reserve_limit: mutex_destroy(&(*mem)->lock); if (gobj) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h index 68b63b970ce8..9c5912b9d8bd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.h @@ -24,8 +24,11 @@ #ifndef AMDGPU_XCP_H #define AMDGPU_XCP_H +#include #include +#include "amdgpu_ctx.h" + #define MAX_XCP 8 #define AMDGPU_XCP_MODE_NONE -1 @@ -34,6 +37,8 @@ #define AMDGPU_XCP_FL_NONE 0 #define AMDGPU_XCP_FL_LOCKED (1 << 0) +struct amdgpu_fpriv; + enum AMDGPU_XCP_IP_BLOCK { AMDGPU_XCP_GFXHUB, AMDGPU_XCP_GFX, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 1ae867482bc7..a9efff94390b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1044,7 +1044,8 @@ static int kfd_ioctl_get_available_memory(struct file *filep, if (!pdd) return -EINVAL; - args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev); + args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev, + pdd->dev->node_id); kfd_unlock_pdd(pdd); return 0; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 45959892bc0f..c1ab70faf36e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -280,7 +280,7 @@ static void svm_range_free(struct svm_range *prange, bool update_mem_usage) if (update_mem_usage && !p->xnack_enabled) { pr_debug("unreserve prange 0x%p size: 0x%llx\n", prange, size); amdgpu_amdkfd_unreserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); } mutex_destroy(&prange->lock); mutex_destroy(&prange->migrate_mutex); @@ -313,7 +313,7 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, p = container_of(svms, struct kfd_process, svms); if (!p->xnack_enabled && update_mem_usage && amdgpu_amdkfd_reserve_mem_limit(NULL, size << PAGE_SHIFT, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)) { + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0)) { pr_info("SVM mapping failed, exceeds resident system memory limit\n"); kfree(prange); return NULL; @@ -3037,10 +3037,10 @@ svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled) size = (pchild->last - pchild->start + 1) << PAGE_SHIFT; if (xnack_enabled) { amdgpu_amdkfd_unreserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); } else { r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); if (r) goto out_unlock; reserved_size += size; @@ -3050,10 +3050,10 @@ svm_range_switch_xnack_reserve_mem(struct kfd_process *p, bool xnack_enabled) size = (prange->last - prange->start + 1) << PAGE_SHIFT; if (xnack_enabled) { amdgpu_amdkfd_unreserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); } else { r = amdgpu_amdkfd_reserve_mem_limit(NULL, size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); if (r) goto out_unlock; reserved_size += size; @@ -3066,7 +3066,7 @@ out_unlock: if (r) amdgpu_amdkfd_unreserve_mem_limit(NULL, reserved_size, - KFD_IOC_ALLOC_MEM_FLAGS_USERPTR); + KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, 0); else /* Change xnack mode must be inside svms lock, to avoid race with * svm_range_deferred_list_work unreserve memory in parallel. -- cgit v1.2.3 From 25f50704343de1bea70100ad41621b5737a6a96b Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Thu, 23 Mar 2023 08:45:56 -0400 Subject: drm/amdkfd: APU mode set max svm range pages svm_migrate_init set the max svm range pages based on the KFD nodes partition size. APU mode don't init pgmap because there is no migration. kgd2kfd_device_init calls svm_migrate_init after KFD nodes allocation and initialization. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 5 ++--- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 7 +++++-- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 15 ++++++++++----- 3 files changed, 17 insertions(+), 10 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index d41da964d2f5..882ff86bba08 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -724,9 +724,6 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd_cwsr_init(kfd); - svm_migrate_init(kfd->adev); - - dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n", kfd->num_nodes); @@ -794,6 +791,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd->nodes[i] = node; } + svm_migrate_init(kfd->adev); + if (kfd_resume_iommu(kfd)) goto kfd_resume_iommu_error; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 199d32c7c289..2512bf681112 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -1000,6 +1000,11 @@ int svm_migrate_init(struct amdgpu_device *adev) if (!KFD_IS_SOC15(kfddev->dev)) return -EINVAL; + svm_range_set_max_pages(adev); + + if (adev->gmc.is_app_apu) + return 0; + pgmap = &kfddev->pgmap; memset(pgmap, 0, sizeof(*pgmap)); @@ -1042,8 +1047,6 @@ int svm_migrate_init(struct amdgpu_device *adev) amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); - svm_range_set_max_pages(adev); - pr_info("HMM registered %ldMB device memory\n", size >> 20); return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index c1ab70faf36e..206851c9e642 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1938,14 +1938,19 @@ void svm_range_set_max_pages(struct amdgpu_device *adev) { uint64_t max_pages; uint64_t pages, _pages; + uint64_t min_pages = 0; + int i; + + for (i = 0; i < adev->kfd.dev->num_nodes; i++) { + pages = KFD_XCP_MEMORY_SIZE(adev, adev->kfd.dev->nodes[i]->xcp->id) >> 17; + pages = clamp(pages, 1ULL << 9, 1ULL << 18); + pages = rounddown_pow_of_two(pages); + min_pages = min_not_zero(min_pages, pages); + } - /* 1/32 VRAM size in pages */ - pages = adev->gmc.real_vram_size >> 17; - pages = clamp(pages, 1ULL << 9, 1ULL << 18); - pages = rounddown_pow_of_two(pages); do { max_pages = READ_ONCE(max_svm_range_pages); - _pages = min_not_zero(max_pages, pages); + _pages = min_not_zero(max_pages, min_pages); } while (cmpxchg(&max_svm_range_pages, max_pages, _pages) != max_pages); } -- cgit v1.2.3 From 84b4dd3f84de424a68e1fda0d483530ddaa92b45 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Fri, 31 Mar 2023 11:18:12 -0400 Subject: drm/amdkfd: Refactor migrate init to support partition switch Rename smv_migrate_init to a better name kgd2kfd_init_zone_device because it setup zone devive pgmap for page migration and keep it in kfd_migrate.c to access static functions svm_migrate_pgmap_ops. Call it only once in amdgpu_device_ip_init after adev ip blocks are initialized, but before amdgpu_amdkfd_device_init initialize kfd nodes which enable SVM support based on pgmap. svm_range_set_max_pages is called by kgd2kfd_device_init everytime after switching compute partition mode. Signed-off-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 11 +++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 4 +++- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 8 +++----- drivers/gpu/drm/amd/amdkfd/kfd_migrate.h | 9 --------- drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 4 ++++ 6 files changed, 23 insertions(+), 16 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 948d362adabb..48d12dbff968 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -372,6 +372,17 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo) { } #endif + +#if IS_ENABLED(CONFIG_HSA_AMD_SVM) +int kgd2kfd_init_zone_device(struct amdgpu_device *adev); +#else +static inline +int kgd2kfd_init_zone_device(struct amdgpu_device *adev) +{ + return 0; +} +#endif + /* KGD2KFD callbacks */ int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger); int kgd2kfd_resume_mm(struct mm_struct *mm); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c index 02ee79b7b56d..f0666230b2ed 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c @@ -2633,8 +2633,10 @@ static int amdgpu_device_ip_init(struct amdgpu_device *adev) goto init_failed; /* Don't init kfd if whole hive need to be reset during init */ - if (!adev->gmc.xgmi.pending_reset) + if (!adev->gmc.xgmi.pending_reset) { + kgd2kfd_init_zone_device(adev); amdgpu_amdkfd_device_init(adev); + } amdgpu_fru_get_product_info(adev); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 882ff86bba08..bf32e547182c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -32,6 +32,7 @@ #include "kfd_iommu.h" #include "amdgpu_amdkfd.h" #include "kfd_smi_events.h" +#include "kfd_svm.h" #include "kfd_migrate.h" #include "amdgpu.h" #include "amdgpu_xcp.h" @@ -791,7 +792,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd->nodes[i] = node; } - svm_migrate_init(kfd->adev); + svm_range_set_max_pages(kfd->adev); if (kfd_resume_iommu(kfd)) goto kfd_resume_iommu_error; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 2512bf681112..35cf6558cf1b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -988,7 +988,7 @@ static const struct dev_pagemap_ops svm_migrate_pgmap_ops = { /* Each VRAM page uses sizeof(struct page) on system memory */ #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page)) -int svm_migrate_init(struct amdgpu_device *adev) +int kgd2kfd_init_zone_device(struct amdgpu_device *adev) { struct amdgpu_kfd_dev *kfddev = &adev->kfd; struct dev_pagemap *pgmap; @@ -996,12 +996,10 @@ int svm_migrate_init(struct amdgpu_device *adev) unsigned long size; void *r; - /* Page migration works on Vega10 or newer */ - if (!KFD_IS_SOC15(kfddev->dev)) + /* Page migration works on gfx9 or newer */ + if (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 0, 1)) return -EINVAL; - svm_range_set_max_pages(adev); - if (adev->gmc.is_app_apu) return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.h b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.h index a5d7e6d22264..487f26368164 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.h @@ -47,15 +47,6 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm, unsigned long svm_migrate_addr_to_pfn(struct amdgpu_device *adev, unsigned long addr); -int svm_migrate_init(struct amdgpu_device *adev); - -#else - -static inline int svm_migrate_init(struct amdgpu_device *adev) -{ - return 0; -} - #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */ #endif /* KFD_MIGRATE_H_ */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index 021def496f5a..762679835e31 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -265,6 +265,10 @@ static inline int kfd_criu_resume_svm(struct kfd_process *p) return 0; } +static inline void svm_range_set_max_pages(struct amdgpu_device *adev) +{ +} + #define KFD_IS_SVM_API_SUPPORTED(dev) false #endif /* IS_ENABLED(CONFIG_HSA_AMD_SVM) */ -- cgit v1.2.3 From 0409022c15623d91c112e51f38cb62633becd432 Mon Sep 17 00:00:00 2001 From: Hawking Zhang Date: Thu, 11 May 2023 17:01:03 +0800 Subject: drm/amdkfd: Fix null ptr access Avoid access null xcp_mgr pointer. Signed-off-by: Hawking Zhang Reviewed-by: Lijo Lazar Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index bf32e547182c..2cfef3f9456f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -644,12 +644,14 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, * If the VMID range changes for GFX9.4.3, then this code MUST be * revisited. */ - partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr, AMDGPU_XCP_FL_LOCKED); - if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) && - partition_mode == AMDGPU_CPX_PARTITION_MODE && - kfd->num_nodes != 1) { - vmid_num_kfd /= 2; - first_vmid_kfd = last_vmid_kfd + 1 - vmid_num_kfd*2; + if (kfd->adev->xcp_mgr) { + partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr, + AMDGPU_XCP_FL_LOCKED); + if (partition_mode == AMDGPU_CPX_PARTITION_MODE && + kfd->num_nodes != 1) { + vmid_num_kfd /= 2; + first_vmid_kfd = last_vmid_kfd + 1 - vmid_num_kfd*2; + } } /* Verify module parameters regarding mapped process number*/ -- cgit v1.2.3 From 9a3ce1a7a9e5372d8c275bf3fbef4456c8407145 Mon Sep 17 00:00:00 2001 From: Hawking Zhang Date: Fri, 12 May 2023 13:22:57 +0800 Subject: drm/amdgpu: Do not access members of xcp w/o check (v2) Not all the asic needs xcp. ensure check xcp availabity before accessing its member. v2: add missing change in kfd_topology.c Signed-off-by: Hawking Zhang Reviewed-by: Le Ma Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 8 ++++---- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2 +- 4 files changed, 7 insertions(+), 7 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 85df73f2c85e..739eb7c0d133 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -429,17 +429,17 @@ uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, struct kfd_local_mem_info *mem_info, - uint8_t xcp_id) + struct amdgpu_xcp *xcp) { memset(mem_info, 0, sizeof(*mem_info)); - if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) { + if (xcp) { if (adev->gmc.real_vram_size == adev->gmc.visible_vram_size) mem_info->local_mem_size_public = - KFD_XCP_MEMORY_SIZE(adev, xcp_id); + KFD_XCP_MEMORY_SIZE(adev, xcp->id); else mem_info->local_mem_size_private = - KFD_XCP_MEMORY_SIZE(adev, xcp_id); + KFD_XCP_MEMORY_SIZE(adev, xcp->id); } else { mem_info->local_mem_size_public = adev->gmc.visible_vram_size; mem_info->local_mem_size_private = adev->gmc.real_vram_size - diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 48d12dbff968..be43d71ba7ef 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -233,7 +233,7 @@ uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev, enum kgd_engine_type type); void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev, struct kfd_local_mem_info *mem_info, - uint8_t xcp_id); + struct amdgpu_xcp *xcp); uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev); uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 2cfef3f9456f..986543a000bf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -784,7 +784,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, atomic_set(&node->sram_ecc_flag, 0); amdgpu_amdkfd_get_local_mem_info(kfd->adev, - &node->local_mem_info, node->xcp->id); + &node->local_mem_info, node->xcp); /* Initialize the KFD node */ if (kfd_init_node(node)) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index e0bacf017a40..8302d8967158 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1235,7 +1235,7 @@ static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) * all the banks will report the same mem_clk_max information */ amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info, - dev->gpu->xcp->id); + dev->gpu->xcp); list_for_each_entry(mem, &dev->mem_props, list) mem->mem_clk_max = local_mem_info.mem_clk_max; -- cgit v1.2.3 From 45b3a914d40e63d2c9e3a3e02fb2014be975b9b0 Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Tue, 16 May 2023 17:16:30 -0400 Subject: drm/amdgpu/gmc9: fix 64 bit division in partition code Rework logic or use do_div() to avoid problems on 32 bit. v2: add a missing case for XCP macro v3: fix out of bounds array access v4: fix xcp handling harder Acked-by: Guchun Chen (v1) Reviewed-by: Mukul Joshi (v3) Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 15 +++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 9 ++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 5 ++++- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 11 ++++++----- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 8 ++++++-- 5 files changed, 35 insertions(+), 13 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 739eb7c0d133..5de92c9ab18f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -794,3 +794,18 @@ void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev) { kgd2kfd_unlock_kfd(); } + + +u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id) +{ + u64 tmp; + s8 mem_id = KFD_XCP_MEM_ID(adev, xcp_id); + + if (adev->gmc.num_mem_partitions && xcp_id >= 0 && mem_id >= 0) { + tmp = adev->gmc.mem_partitions[mem_id].size; + do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition); + return tmp; + } else { + return adev->gmc.real_vram_size; + } +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index be43d71ba7ef..94cc456761e5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -333,15 +333,14 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, uint64_t size, u32 alloc_flag, int8_t xcp_id); +u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id); + #define KFD_XCP_MEM_ID(adev, xcp_id) \ ((adev)->xcp_mgr && (xcp_id) >= 0 ?\ (adev)->xcp_mgr->xcp[(xcp_id)].mem_id : -1) -#define KFD_XCP_MEMORY_SIZE(adev, xcp_id)\ - ((adev)->gmc.num_mem_partitions && (xcp_id) >= 0 ?\ - (adev)->gmc.mem_partitions[KFD_XCP_MEM_ID((adev), (xcp_id))].size /\ - (adev)->xcp_mgr->num_xcp_per_mem_partition :\ - (adev)->gmc.real_vram_size) +#define KFD_XCP_MEMORY_SIZE(adev, xcp_id) amdgpu_amdkfd_xcp_memory_size((adev), (xcp_id)) + #if IS_ENABLED(CONFIG_HSA_AMD) void amdgpu_amdkfd_gpuvm_init_mem_limits(void); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 23101c82519a..902773ce41b5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -814,11 +814,14 @@ static void amdgpu_ttm_gart_bind_gfx9_mqd(struct amdgpu_device *adev, struct amdgpu_ttm_tt *gtt = (void *)ttm; uint64_t total_pages = ttm->num_pages; int num_xcc = max(1U, adev->gfx.num_xcc_per_xcp); - uint64_t page_idx, pages_per_xcc = total_pages / num_xcc; + uint64_t page_idx, pages_per_xcc; int i; uint64_t ctrl_flags = (flags & ~AMDGPU_PTE_MTYPE_VG10_MASK) | AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC); + pages_per_xcc = total_pages; + do_div(pages_per_xcc, num_xcc); + for (i = 0, page_idx = 0; i < num_xcc; i++, page_idx += pages_per_xcc) { /* MQD page: use default flags */ amdgpu_gart_bind(adev, diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c index 7ea80bdf8e1e..f70e666cecf2 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c @@ -1914,9 +1914,10 @@ gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev, adev->gmc.num_mem_partitions = num_ranges; /* If there is only partition, don't use entire size */ - if (adev->gmc.num_mem_partitions == 1) - mem_ranges[0].size = - (mem_ranges[0].size * (mem_groups - 1) / mem_groups); + if (adev->gmc.num_mem_partitions == 1) { + mem_ranges[0].size = mem_ranges[0].size * (mem_groups - 1); + do_div(mem_ranges[0].size, mem_groups); + } } static void @@ -1948,8 +1949,8 @@ gmc_v9_0_init_sw_mem_ranges(struct amdgpu_device *adev, break; } - size = (adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT) / - adev->gmc.num_mem_partitions; + size = adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT; + size /= adev->gmc.num_mem_partitions; for (i = 0; i < adev->gmc.num_mem_partitions; ++i) { mem_ranges[i].range.fpfn = start_addr; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 206851c9e642..b0f0d31bf3e6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1939,10 +1939,14 @@ void svm_range_set_max_pages(struct amdgpu_device *adev) uint64_t max_pages; uint64_t pages, _pages; uint64_t min_pages = 0; - int i; + int i, id; for (i = 0; i < adev->kfd.dev->num_nodes; i++) { - pages = KFD_XCP_MEMORY_SIZE(adev, adev->kfd.dev->nodes[i]->xcp->id) >> 17; + if (adev->kfd.dev->nodes[i]->xcp) + id = adev->kfd.dev->nodes[i]->xcp->id; + else + id = -1; + pages = KFD_XCP_MEMORY_SIZE(adev, id) >> 17; pages = clamp(pages, 1ULL << 9, 1ULL << 18); pages = rounddown_pow_of_two(pages); min_pages = min_not_zero(min_pages, pages); -- cgit v1.2.3 From f2cd6b26922e68ffafd14a9128e20630296e430d Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Wed, 17 May 2023 15:04:35 -0400 Subject: drm/amdkfd: fix stack size in svm_range_validate_and_map MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Allocate large local variable on heap to avoid exceeding the stack size: drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_svm.c: In function ‘svm_range_validate_and_map’: drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_svm.c:1690:1: warning: the frame size of 2360 bytes is larger than 2048 bytes [-Wframe-larger-than=] Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 56 +++++++++++++++++++++--------------- 1 file changed, 33 insertions(+), 23 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index b0f0d31bf3e6..ee16130ddc75 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1554,48 +1554,54 @@ static int svm_range_validate_and_map(struct mm_struct *mm, struct svm_range *prange, int32_t gpuidx, bool intr, bool wait, bool flush_tlb) { - struct svm_validate_context ctx; + struct svm_validate_context *ctx; unsigned long start, end, addr; struct kfd_process *p; void *owner; int32_t idx; int r = 0; - ctx.process = container_of(prange->svms, struct kfd_process, svms); - ctx.prange = prange; - ctx.intr = intr; + ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL); + if (!ctx) + return -ENOMEM; + ctx->process = container_of(prange->svms, struct kfd_process, svms); + ctx->prange = prange; + ctx->intr = intr; if (gpuidx < MAX_GPU_INSTANCE) { - bitmap_zero(ctx.bitmap, MAX_GPU_INSTANCE); - bitmap_set(ctx.bitmap, gpuidx, 1); - } else if (ctx.process->xnack_enabled) { - bitmap_copy(ctx.bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); + bitmap_zero(ctx->bitmap, MAX_GPU_INSTANCE); + bitmap_set(ctx->bitmap, gpuidx, 1); + } else if (ctx->process->xnack_enabled) { + bitmap_copy(ctx->bitmap, prange->bitmap_aip, MAX_GPU_INSTANCE); /* If prefetch range to GPU, or GPU retry fault migrate range to * GPU, which has ACCESS attribute to the range, create mapping * on that GPU. */ if (prange->actual_loc) { - gpuidx = kfd_process_gpuidx_from_gpuid(ctx.process, + gpuidx = kfd_process_gpuidx_from_gpuid(ctx->process, prange->actual_loc); if (gpuidx < 0) { WARN_ONCE(1, "failed get device by id 0x%x\n", prange->actual_loc); - return -EINVAL; + r = -EINVAL; + goto free_ctx; } if (test_bit(gpuidx, prange->bitmap_access)) - bitmap_set(ctx.bitmap, gpuidx, 1); + bitmap_set(ctx->bitmap, gpuidx, 1); } } else { - bitmap_or(ctx.bitmap, prange->bitmap_access, + bitmap_or(ctx->bitmap, prange->bitmap_access, prange->bitmap_aip, MAX_GPU_INSTANCE); } - if (bitmap_empty(ctx.bitmap, MAX_GPU_INSTANCE)) { - if (!prange->mapped_to_gpu) - return 0; + if (bitmap_empty(ctx->bitmap, MAX_GPU_INSTANCE)) { + if (!prange->mapped_to_gpu) { + r = 0; + goto free_ctx; + } - bitmap_copy(ctx.bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); + bitmap_copy(ctx->bitmap, prange->bitmap_access, MAX_GPU_INSTANCE); } if (prange->actual_loc && !prange->ttm_res) { @@ -1603,15 +1609,16 @@ static int svm_range_validate_and_map(struct mm_struct *mm, * svm_migrate_ram_to_vram after allocating a BO. */ WARN_ONCE(1, "VRAM BO missing during validation\n"); - return -EINVAL; + r = -EINVAL; + goto free_ctx; } - svm_range_reserve_bos(&ctx); + svm_range_reserve_bos(ctx); p = container_of(prange->svms, struct kfd_process, svms); - owner = kfd_svm_page_owner(p, find_first_bit(ctx.bitmap, + owner = kfd_svm_page_owner(p, find_first_bit(ctx->bitmap, MAX_GPU_INSTANCE)); - for_each_set_bit(idx, ctx.bitmap, MAX_GPU_INSTANCE) { + for_each_set_bit(idx, ctx->bitmap, MAX_GPU_INSTANCE) { if (kfd_svm_page_owner(p, idx) != owner) { owner = NULL; break; @@ -1648,7 +1655,7 @@ static int svm_range_validate_and_map(struct mm_struct *mm, } offset = (addr - start) >> PAGE_SHIFT; - r = svm_range_dma_map(prange, ctx.bitmap, offset, npages, + r = svm_range_dma_map(prange, ctx->bitmap, offset, npages, hmm_range->hmm_pfns); if (r) { pr_debug("failed %d to dma map range\n", r); @@ -1668,7 +1675,7 @@ static int svm_range_validate_and_map(struct mm_struct *mm, } r = svm_range_map_to_gpus(prange, offset, npages, readonly, - ctx.bitmap, wait, flush_tlb); + ctx->bitmap, wait, flush_tlb); unlock_out: svm_range_unlock(prange); @@ -1682,11 +1689,14 @@ unlock_out: } unreserve_out: - svm_range_unreserve_bos(&ctx); + svm_range_unreserve_bos(ctx); if (!r) prange->validate_timestamp = ktime_get_boottime(); +free_ctx: + kfree(ctx); + return r; } -- cgit v1.2.3 From 2ad00e753ae13b0c523a579fb04372787f77cce9 Mon Sep 17 00:00:00 2001 From: Srinivasan Shanmugam Date: Fri, 19 May 2023 10:44:41 +0530 Subject: drm/amdgpu: Fix uninitalized variable in kgd2kfd_device_init MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_device.c:613:4: error: variable 'num_xcd' is uninitialized when used here [-Werror,-Wuninitialized] num_xcd, kfd->adev->gfx.num_xcc_per_xcp); ^~~~~~~ include/linux/dev_printk.h:144:65: note: expanded from macro 'dev_err' dev_printk_index_wrap(_dev_err, KERN_ERR, dev, dev_fmt(fmt), ##__VA_ARGS__) ^~~~~~~~~~~ include/linux/dev_printk.h:110:23: note: expanded from macro 'dev_printk_index_wrap' _p_func(dev, fmt, ##__VA_ARGS__); \ ^~~~~~~~~~~ drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_device.c:597:13: note: initialize the variable 'num_xcd' to silence this warning int num_xcd, partition_mode; ^ = 0 1 error generated. Cc: Luben Tuikov Cc: Alex Deucher Cc: Christian König Cc: Felix Kuehling Cc: Mukul Joshi Signed-off-by: Srinivasan Shanmugam Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 986543a000bf..e92b93b2c14c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -594,7 +594,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, struct kfd_node *node; uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd; unsigned int max_proc_per_quantum; - int num_xcd, partition_mode; + int partition_mode; int xcp_idx; kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev, @@ -609,8 +609,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (kfd->num_nodes == 0) { dev_err(kfd_device, - "KFD num nodes cannot be 0, GC inst: %d, num_xcc_in_node: %d\n", - num_xcd, kfd->adev->gfx.num_xcc_per_xcp); + "KFD num nodes cannot be 0, num_xcc_in_node: %d\n", + kfd->adev->gfx.num_xcc_per_xcp); goto out; } -- cgit v1.2.3 From 55a6dc60b47c817c644af2b505d46815d8b9219e Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Tue, 23 May 2023 11:59:53 -0400 Subject: drm/amdkfd: Set event interrupt class for GFX 9.4.3 Fix the warning during driver load because the event interrupt class is not set for GFX9.4.3. Signed-off-by: Mukul Joshi Acked-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 1 + 1 file changed, 1 insertion(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index e92b93b2c14c..862a50f7b490 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -139,6 +139,7 @@ static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd) case IP_VERSION(9, 4, 0): /* VEGA20 */ case IP_VERSION(9, 4, 1): /* ARCTURUS */ case IP_VERSION(9, 4, 2): /* ALDEBARAN */ + case IP_VERSION(9, 4, 3): /* GC 9.4.3 */ case IP_VERSION(10, 3, 1): /* VANGOGH */ case IP_VERSION(10, 3, 3): /* YELLOW_CARP */ case IP_VERSION(10, 3, 6): /* GC 10.3.6 */ -- cgit v1.2.3 From 28ebbb4981cb1fad12e0b1227dbecc88810b1ee8 Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Wed, 24 May 2023 14:30:12 -0400 Subject: drm/amdkfd: fix gfx_target_version for certain 11.0.3 devices Certain boards with GC IP 11.0.3 need slightly different handling in the shader compiler due to board specific bounding box optimizations. Acked-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 862a50f7b490..ebc3c3f965f9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -411,8 +411,15 @@ struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) f2g = &gfx_v11_kfd2kgd; break; case IP_VERSION(11, 0, 3): - /* Note: Compiler version is 11.0.1 while HW version is 11.0.3 */ - gfx_target_version = 110001; + if ((adev->pdev->device == 0x7460 && + adev->pdev->revision == 0x00) || + (adev->pdev->device == 0x7461 && + adev->pdev->revision == 0x00)) + /* Note: Compiler version is 11.0.5 while HW version is 11.0.3 */ + gfx_target_version = 110005; + else + /* Note: Compiler version is 11.0.1 while HW version is 11.0.3 */ + gfx_target_version = 110001; f2g = &gfx_v11_kfd2kgd; break; default: -- cgit v1.2.3 From 07a1475279244cd8eea81bec44fd5f0a9d6871f8 Mon Sep 17 00:00:00 2001 From: Graham Sider Date: Mon, 3 Apr 2023 15:31:53 -0400 Subject: drm/amdkfd: Add new gfx_target_versions for GC 9.4.3 For GC 9.4.3, set gfx_target_version to 90402 for rev 1 and later (APU or dGPU), 90401 for rev 0 dGPU, and 90400 for rev 0 APU. Signed-off-by: Graham Sider Acked-by: Alex Deucher Reviewed-by: Amber Lin Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index ebc3c3f965f9..e84ad1c5ef44 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -332,7 +332,9 @@ struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf) f2g = &aldebaran_kfd2kgd; break; case IP_VERSION(9, 4, 3): - gfx_target_version = 90400; + gfx_target_version = adev->rev_id >= 1 ? 90402 + : adev->flags & AMD_IS_APU ? 90400 + : 90401; f2g = &gc_9_4_3_kfd2kgd; break; /* Navi10 */ -- cgit v1.2.3 From b695c97b580a1949d0dd96aa17b01d4de738eda3 Mon Sep 17 00:00:00 2001 From: Lijo Lazar Date: Mon, 29 May 2023 19:18:54 +0530 Subject: drm/amdkfd: Fix MEC pipe interrupt enablement for_each_inst modifies xcc_mask and therefore the loop doesn't initialize properly interrupts on all pipes. Keep looping through xcc as the outer loop to fix this issue. Fixes: c4050ff1a43e ("drm/amdkfd: Use xcc mask for identifying xcc") Signed-off-by: Lijo Lazar Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 493b4b66f180..80cddb46657f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1274,11 +1274,12 @@ static void init_interrupts(struct device_queue_manager *dqm) uint32_t xcc_mask = dqm->dev->xcc_mask; unsigned int i, xcc_id; - for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) { - if (is_pipe_enabled(dqm, 0, i)) { - for_each_inst(xcc_id, xcc_mask) + for_each_inst(xcc_id, xcc_mask) { + for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) { + if (is_pipe_enabled(dqm, 0, i)) { dqm->dev->kfd2kgd->init_interrupts( dqm->dev->adev, i, xcc_id); + } } } } -- cgit v1.2.3 From c22b044070971e474dd0ff81a9830df93751f726 Mon Sep 17 00:00:00 2001 From: Alex Sierra Date: Mon, 29 May 2023 16:01:37 -0500 Subject: drm/amdkfd: flag added to handle errors from svm validate and map If a return error is raised during validation and mapping of a prange, this flag is set. It is a rare occurrence, but it could happen when `amdgpu_hmm_range_get_pages_done` returns true. In such cases, the caller should retry. However, it is important to ensure that the prange is updated correctly during the retry. Signed-off-by: Alex Sierra Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 3 ++- drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 1 + 2 files changed, 3 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index ee16130ddc75..9c88d6e90c3b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -809,7 +809,7 @@ svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange, } } - return true; + return !prange->is_error_flag; } /** @@ -1691,6 +1691,7 @@ unlock_out: unreserve_out: svm_range_unreserve_bos(ctx); + prange->is_error_flag = !!r; if (!r) prange->validate_timestamp = ktime_get_boottime(); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index 762679835e31..21b14510882b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -134,6 +134,7 @@ struct svm_range { DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); bool validated_once; bool mapped_to_gpu; + bool is_error_flag; }; static inline void svm_range_lock(struct svm_range *prange) -- cgit v1.2.3 From b3a02e8b61c19a0380870c713bc704d7e4f9e0dd Mon Sep 17 00:00:00 2001 From: Srinivasan Shanmugam Date: Tue, 30 May 2023 04:03:16 +0530 Subject: drm/amdgpu: Fix up missing parameters kdoc in svm_migrate_vma_to_ram MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fix these warnings by adding & deleting the deviant arguments. gcc with W=1 drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_migrate.c:671: warning: Function parameter or member 'node' not described in 'svm_migrate_vma_to_ram' drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_migrate.c:671: warning: Function parameter or member 'trigger' not described in 'svm_migrate_vma_to_ram' drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_migrate.c:671: warning: Function parameter or member 'fault_page' not described in 'svm_migrate_vma_to_ram' drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_migrate.c:671: warning: Excess function parameter 'adev' description in 'svm_migrate_vma_to_ram' drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_migrate.c:771: warning: Function parameter or member 'fault_page' not described in 'svm_migrate_vram_to_ram' Cc: Christian König Cc: Alex Deucher Signed-off-by: Srinivasan Shanmugam Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 35cf6558cf1b..58d95fb99595 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -651,11 +651,13 @@ out_oom: /** * svm_migrate_vma_to_ram - migrate range inside one vma from device to system * - * @adev: amdgpu device to migrate from * @prange: svm range structure * @vma: vm_area_struct that range [start, end] belongs to * @start: range start virtual address in pages * @end: range end virtual address in pages + * @node: kfd node device to migrate from + * @trigger: reason of migration + * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback * * Context: Process context, caller hold mmap read lock, prange->migrate_mutex * @@ -760,6 +762,7 @@ out: * @prange: range structure * @mm: process mm, use current->mm if NULL * @trigger: reason of migration + * @fault_page: is from vmf->page, svm_migrate_to_ram(), this is CPU page fault callback * * Context: Process context, caller hold mmap read lock, prange->migrate_mutex * -- cgit v1.2.3 From 0df1106bfd13a9fe1eb7c33666bec091bc37c2a7 Mon Sep 17 00:00:00 2001 From: Tom Rix Date: Thu, 30 Mar 2023 11:20:40 -0400 Subject: drm/amdkfd: remove unused sq_int_priv variable clang with W=1 reports drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_int_process_v11.c:282:38: error: variable 'sq_int_priv' set but not used [-Werror,-Wunused-but-set-variable] uint8_t sq_int_enc, sq_int_errtype, sq_int_priv; ^ This variable is not used so remove it. Signed-off-by: Tom Rix Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c | 9 +-------- 1 file changed, 1 insertion(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c index 0f0fdea4cd8a..c2166bf964ef 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c @@ -279,7 +279,7 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, { uint16_t source_id, client_id, ring_id, pasid, vmid; uint32_t context_id0, context_id1; - uint8_t sq_int_enc, sq_int_errtype, sq_int_priv; + uint8_t sq_int_enc, sq_int_errtype; struct kfd_vm_fault_info info = {0}; struct kfd_hsa_memory_exception_data exception_data; @@ -348,13 +348,6 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, break; case SQ_INTERRUPT_WORD_ENCODING_INST: print_sq_intr_info_inst(context_id0, context_id1); - sq_int_priv = REG_GET_FIELD(context_id0, - SQ_INTERRUPT_WORD_WAVE_CTXID0, PRIV); - /*if (sq_int_priv && (kfd_set_dbg_ev_from_interrupt(dev, pasid, - KFD_CTXID0_DOORBELL_ID(context_id0), - KFD_CTXID0_TRAP_CODE(context_id0), - NULL, 0))) - return;*/ break; case SQ_INTERRUPT_WORD_ENCODING_ERROR: print_sq_intr_info_error(context_id0, context_id1); -- cgit v1.2.3 From 16cc3a221537bb3588ec2a568d7bd0e7972b25a8 Mon Sep 17 00:00:00 2001 From: Srinivasan Shanmugam Date: Wed, 31 May 2023 23:33:35 +0530 Subject: drm/amdgpu: Add function parameter 'event' to kdoc in svm_range_evict() MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fixes the following gcc with W=1: drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_svm.c:1841: warning: Function parameter or member 'event' not described in 'svm_range_evict' Cc: Felix Kuehling Cc: Christian König Cc: Alex Deucher Signed-off-by: Srinivasan Shanmugam Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 1 + 1 file changed, 1 insertion(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 9c88d6e90c3b..615eab3f78c9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1826,6 +1826,7 @@ out_reschedule: * @mm: current process mm_struct * @start: starting process queue number * @last: last process queue number + * @event: mmu notifier event when range is evicted or migrated * * Stop all queues of the process to ensure GPU doesn't access the memory, then * return to let CPU evict the buffer and proceed CPU pagetable update. -- cgit v1.2.3 From ba3c87fffb79311f54464288c66421d19c2c1234 Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Fri, 2 Jun 2023 12:58:05 -0400 Subject: amd/amdkfd: drop unused KFD_IOCTL_SVM_FLAG_UNCACHED flag Was leftover from GC 9.4.3 bring up and is currently unused. Drop it for now. Cc: Philip.Yang@amd.com Cc: rajneesh.bhardwaj@amd.com Cc: Felix.Kuehling@amd.com Reviewed-by: Rajneesh Bhardwaj Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 2 +- include/uapi/linux/kfd_ioctl.h | 2 -- 2 files changed, 1 insertion(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 615eab3f78c9..5ff1a5a89d96 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1154,7 +1154,7 @@ svm_range_get_pte_flags(struct kfd_node *node, uint64_t pte_flags; bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN); bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT; - bool uncached = flags & KFD_IOCTL_SVM_FLAG_UNCACHED; + bool uncached = false; /*flags & KFD_IOCTL_SVM_FLAG_UNCACHED;*/ unsigned int mtype_local; if (domain == SVM_RANGE_VRAM_DOMAIN) diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h index 2a9671e1ddb5..2da5c3ad71bd 100644 --- a/include/uapi/linux/kfd_ioctl.h +++ b/include/uapi/linux/kfd_ioctl.h @@ -623,8 +623,6 @@ enum kfd_mmio_remap { #define KFD_IOCTL_SVM_FLAG_GPU_READ_MOSTLY 0x00000020 /* Keep GPU memory mapping always valid as if XNACK is disable */ #define KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED 0x00000040 -/* Uncached access to memory */ -#define KFD_IOCTL_SVM_FLAG_UNCACHED 0x00000080 /** * kfd_ioctl_svm_op - SVM ioctl operations -- cgit v1.2.3 From 4f98cf2baf9faee5b6f2f7889dad7c0f7686a787 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Wed, 2 Mar 2022 14:30:12 -0500 Subject: drm/amdkfd: add debug and runtime enable interface Introduce the GPU debug operations interface. For ROCm-GDB to extend the GNU Debugger's ability to inspect the AMD GPU instruction set, provide the necessary interface to allow the debugger to HW debug-mode set and query exceptions per HSA queue, process or device. The runtime_enable interface coordinates exception handling with the HSA runtime. Usage is available in the kern docs at uapi/linux/kfd_ioctl.h. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 48 +++ include/uapi/linux/kfd_ioctl.h | 668 ++++++++++++++++++++++++++++++- 2 files changed, 715 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index a9efff94390b..00e34125987c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2729,6 +2729,48 @@ static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data) return ret; } +static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data) +{ + return 0; +} + +static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data) +{ + struct kfd_ioctl_dbg_trap_args *args = data; + int r = 0; + + if (sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Debugging does not support sched_policy %i", sched_policy); + return -EINVAL; + } + + switch (args->op) { + case KFD_IOC_DBG_TRAP_ENABLE: + case KFD_IOC_DBG_TRAP_DISABLE: + case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT: + case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED: + case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE: + case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE: + case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES: + case KFD_IOC_DBG_TRAP_RESUME_QUEUES: + case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH: + case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH: + case KFD_IOC_DBG_TRAP_SET_FLAGS: + case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT: + case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: + case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: + case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: + pr_warn("Debugging not supported yet\n"); + r = -EACCES; + break; + default: + pr_err("Invalid option: %i\n", args->op); + r = -EINVAL; + } + + return r; +} + #define AMDKFD_IOCTL_DEF(ioctl, _func, _flags) \ [_IOC_NR(ioctl)] = {.cmd = ioctl, .func = _func, .flags = _flags, \ .cmd_drv = 0, .name = #ioctl} @@ -2841,6 +2883,12 @@ static const struct amdkfd_ioctl_desc amdkfd_ioctls[] = { AMDKFD_IOCTL_DEF(AMDKFD_IOC_EXPORT_DMABUF, kfd_ioctl_export_dmabuf, 0), + + AMDKFD_IOCTL_DEF(AMDKFD_IOC_RUNTIME_ENABLE, + kfd_ioctl_runtime_enable, 0), + + AMDKFD_IOCTL_DEF(AMDKFD_IOC_DBG_TRAP, + kfd_ioctl_set_debug_trap, 0), }; #define AMDKFD_CORE_IOCTL_COUNT ARRAY_SIZE(amdkfd_ioctls) diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h index 2da5c3ad71bd..32913d674d38 100644 --- a/include/uapi/linux/kfd_ioctl.h +++ b/include/uapi/linux/kfd_ioctl.h @@ -110,6 +110,32 @@ struct kfd_ioctl_get_available_memory_args { __u32 pad; }; +struct kfd_dbg_device_info_entry { + __u64 exception_status; + __u64 lds_base; + __u64 lds_limit; + __u64 scratch_base; + __u64 scratch_limit; + __u64 gpuvm_base; + __u64 gpuvm_limit; + __u32 gpu_id; + __u32 location_id; + __u32 vendor_id; + __u32 device_id; + __u32 revision_id; + __u32 subsystem_vendor_id; + __u32 subsystem_device_id; + __u32 fw_version; + __u32 gfx_target_version; + __u32 simd_count; + __u32 max_waves_per_simd; + __u32 array_count; + __u32 simd_arrays_per_engine; + __u32 num_xcc; + __u32 capability; + __u32 debug_prop; +}; + /* For kfd_ioctl_set_memory_policy_args.default_policy and alternate_policy */ #define KFD_IOC_CACHE_POLICY_COHERENT 0 #define KFD_IOC_CACHE_POLICY_NONCOHERENT 1 @@ -773,6 +799,640 @@ struct kfd_ioctl_set_xnack_mode_args { __s32 xnack_enabled; }; +/* Wave launch override modes */ +enum kfd_dbg_trap_override_mode { + KFD_DBG_TRAP_OVERRIDE_OR = 0, + KFD_DBG_TRAP_OVERRIDE_REPLACE = 1 +}; + +/* Wave launch overrides */ +enum kfd_dbg_trap_mask { + KFD_DBG_TRAP_MASK_FP_INVALID = 1, + KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL = 2, + KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO = 4, + KFD_DBG_TRAP_MASK_FP_OVERFLOW = 8, + KFD_DBG_TRAP_MASK_FP_UNDERFLOW = 16, + KFD_DBG_TRAP_MASK_FP_INEXACT = 32, + KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO = 64, + KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH = 128, + KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION = 256, + KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START = (1 << 30), + KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END = (1 << 31) +}; + +/* Wave launch modes */ +enum kfd_dbg_trap_wave_launch_mode { + KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL = 0, + KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT = 1, + KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG = 3 +}; + +/* Address watch modes */ +enum kfd_dbg_trap_address_watch_mode { + KFD_DBG_TRAP_ADDRESS_WATCH_MODE_READ = 0, + KFD_DBG_TRAP_ADDRESS_WATCH_MODE_NONREAD = 1, + KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ATOMIC = 2, + KFD_DBG_TRAP_ADDRESS_WATCH_MODE_ALL = 3 +}; + +/* Additional wave settings */ +enum kfd_dbg_trap_flags { + KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP = 1, +}; + +/* Trap exceptions */ +enum kfd_dbg_trap_exception_code { + EC_NONE = 0, + /* per queue */ + EC_QUEUE_WAVE_ABORT = 1, + EC_QUEUE_WAVE_TRAP = 2, + EC_QUEUE_WAVE_MATH_ERROR = 3, + EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION = 4, + EC_QUEUE_WAVE_MEMORY_VIOLATION = 5, + EC_QUEUE_WAVE_APERTURE_VIOLATION = 6, + EC_QUEUE_PACKET_DISPATCH_DIM_INVALID = 16, + EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID = 17, + EC_QUEUE_PACKET_DISPATCH_CODE_INVALID = 18, + EC_QUEUE_PACKET_RESERVED = 19, + EC_QUEUE_PACKET_UNSUPPORTED = 20, + EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21, + EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22, + EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23, + EC_QUEUE_PREEMPTION_ERROR = 30, + EC_QUEUE_NEW = 31, + /* per device */ + EC_DEVICE_QUEUE_DELETE = 32, + EC_DEVICE_MEMORY_VIOLATION = 33, + EC_DEVICE_RAS_ERROR = 34, + EC_DEVICE_FATAL_HALT = 35, + EC_DEVICE_NEW = 36, + /* per process */ + EC_PROCESS_RUNTIME = 48, + EC_PROCESS_DEVICE_REMOVE = 49, + EC_MAX +}; + +/* Mask generated by ecode in kfd_dbg_trap_exception_code */ +#define KFD_EC_MASK(ecode) (1ULL << (ecode - 1)) + +/* Masks for exception code type checks below */ +#define KFD_EC_MASK_QUEUE (KFD_EC_MASK(EC_QUEUE_WAVE_ABORT) | \ + KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | \ + KFD_EC_MASK(EC_QUEUE_WAVE_MATH_ERROR) | \ + KFD_EC_MASK(EC_QUEUE_WAVE_ILLEGAL_INSTRUCTION) | \ + KFD_EC_MASK(EC_QUEUE_WAVE_MEMORY_VIOLATION) | \ + KFD_EC_MASK(EC_QUEUE_WAVE_APERTURE_VIOLATION) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_DIM_INVALID) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_GROUP_SEGMENT_SIZE_INVALID) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_CODE_INVALID) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_RESERVED) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_UNSUPPORTED) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \ + KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \ + KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \ + KFD_EC_MASK(EC_QUEUE_NEW)) +#define KFD_EC_MASK_DEVICE (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \ + KFD_EC_MASK(EC_DEVICE_RAS_ERROR) | \ + KFD_EC_MASK(EC_DEVICE_FATAL_HALT) | \ + KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION) | \ + KFD_EC_MASK(EC_DEVICE_NEW)) +#define KFD_EC_MASK_PROCESS (KFD_EC_MASK(EC_PROCESS_RUNTIME) | \ + KFD_EC_MASK(EC_PROCESS_DEVICE_REMOVE)) + +/* Checks for exception code types for KFD search */ +#define KFD_DBG_EC_TYPE_IS_QUEUE(ecode) \ + (!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_QUEUE)) +#define KFD_DBG_EC_TYPE_IS_DEVICE(ecode) \ + (!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_DEVICE)) +#define KFD_DBG_EC_TYPE_IS_PROCESS(ecode) \ + (!!(KFD_EC_MASK(ecode) & KFD_EC_MASK_PROCESS)) + + +/* Runtime enable states */ +enum kfd_dbg_runtime_state { + DEBUG_RUNTIME_STATE_DISABLED = 0, + DEBUG_RUNTIME_STATE_ENABLED = 1, + DEBUG_RUNTIME_STATE_ENABLED_BUSY = 2, + DEBUG_RUNTIME_STATE_ENABLED_ERROR = 3 +}; + +/* Runtime enable status */ +struct kfd_runtime_info { + __u64 r_debug; + __u32 runtime_state; + __u32 ttmp_setup; +}; + +/* Enable modes for runtime enable */ +#define KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK 1 +#define KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK 2 + +/** + * kfd_ioctl_runtime_enable_args - Arguments for runtime enable + * + * Coordinates debug exception signalling and debug device enablement with runtime. + * + * @r_debug - pointer to user struct for sharing information between ROCr and the debuggger + * @mode_mask - mask to set mode + * KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK - enable runtime for debugging, otherwise disable + * KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK - enable trap temporary setup (ignore on disable) + * @capabilities_mask - mask to notify runtime on what KFD supports + * + * Return - 0 on SUCCESS. + * - EBUSY if runtime enable call already pending. + * - EEXIST if user queues already active prior to call. + * If process is debug enabled, runtime enable will enable debug devices and + * wait for debugger process to send runtime exception EC_PROCESS_RUNTIME + * to unblock - see kfd_ioctl_dbg_trap_args. + * + */ +struct kfd_ioctl_runtime_enable_args { + __u64 r_debug; + __u32 mode_mask; + __u32 capabilities_mask; +}; + +/* Queue information */ +struct kfd_queue_snapshot_entry { + __u64 exception_status; + __u64 ring_base_address; + __u64 write_pointer_address; + __u64 read_pointer_address; + __u64 ctx_save_restore_address; + __u32 queue_id; + __u32 gpu_id; + __u32 ring_size; + __u32 queue_type; + __u32 ctx_save_restore_area_size; + __u32 reserved; +}; + +/* Queue status return for suspend/resume */ +#define KFD_DBG_QUEUE_ERROR_BIT 30 +#define KFD_DBG_QUEUE_INVALID_BIT 31 +#define KFD_DBG_QUEUE_ERROR_MASK (1 << KFD_DBG_QUEUE_ERROR_BIT) +#define KFD_DBG_QUEUE_INVALID_MASK (1 << KFD_DBG_QUEUE_INVALID_BIT) + +/* Context save area header information */ +struct kfd_context_save_area_header { + struct { + __u32 control_stack_offset; + __u32 control_stack_size; + __u32 wave_state_offset; + __u32 wave_state_size; + } wave_state; + __u32 debug_offset; + __u32 debug_size; + __u64 err_payload_addr; + __u32 err_event_id; + __u32 reserved1; +}; + +/* + * Debug operations + * + * For specifics on usage and return values, see documentation per operation + * below. Otherwise, generic error returns apply: + * - ESRCH if the process to debug does not exist. + * + * - EINVAL (with KFD_IOC_DBG_TRAP_ENABLE exempt) if operation + * KFD_IOC_DBG_TRAP_ENABLE has not succeeded prior. + * Also returns this error if GPU hardware scheduling is not supported. + * + * - EPERM (with KFD_IOC_DBG_TRAP_DISABLE exempt) if target process is not + * PTRACE_ATTACHED. KFD_IOC_DBG_TRAP_DISABLE is exempt to allow + * clean up of debug mode as long as process is debug enabled. + * + * - EACCES if any DBG_HW_OP (debug hardware operation) is requested when + * AMDKFD_IOC_RUNTIME_ENABLE has not succeeded prior. + * + * - ENODEV if any GPU does not support debugging on a DBG_HW_OP call. + * + * - Other errors may be returned when a DBG_HW_OP occurs while the GPU + * is in a fatal state. + * + */ +enum kfd_dbg_trap_operations { + KFD_IOC_DBG_TRAP_ENABLE = 0, + KFD_IOC_DBG_TRAP_DISABLE = 1, + KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT = 2, + KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED = 3, + KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE = 4, /* DBG_HW_OP */ + KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE = 5, /* DBG_HW_OP */ + KFD_IOC_DBG_TRAP_SUSPEND_QUEUES = 6, /* DBG_HW_OP */ + KFD_IOC_DBG_TRAP_RESUME_QUEUES = 7, /* DBG_HW_OP */ + KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH = 8, /* DBG_HW_OP */ + KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH = 9, /* DBG_HW_OP */ + KFD_IOC_DBG_TRAP_SET_FLAGS = 10, + KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT = 11, + KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO = 12, + KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT = 13, + KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT = 14 +}; + +/** + * kfd_ioctl_dbg_trap_enable_args + * + * Arguments for KFD_IOC_DBG_TRAP_ENABLE. + * + * Enables debug session for target process. Call @op KFD_IOC_DBG_TRAP_DISABLE in + * kfd_ioctl_dbg_trap_args to disable debug session. + * + * @exception_mask (IN) - exceptions to raise to the debugger + * @rinfo_ptr (IN) - pointer to runtime info buffer (see kfd_runtime_info) + * @rinfo_size (IN/OUT) - size of runtime info buffer in bytes + * @dbg_fd (IN) - fd the KFD will nofify the debugger with of raised + * exceptions set in exception_mask. + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * Copies KFD saved kfd_runtime_info to @rinfo_ptr on enable. + * Size of kfd_runtime saved by the KFD returned to @rinfo_size. + * - EBADF if KFD cannot get a reference to dbg_fd. + * - EFAULT if KFD cannot copy runtime info to rinfo_ptr. + * - EINVAL if target process is already debug enabled. + * + */ +struct kfd_ioctl_dbg_trap_enable_args { + __u64 exception_mask; + __u64 rinfo_ptr; + __u32 rinfo_size; + __u32 dbg_fd; +}; + +/** + * kfd_ioctl_dbg_trap_send_runtime_event_args + * + * + * Arguments for KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT. + * Raises exceptions to runtime. + * + * @exception_mask (IN) - exceptions to raise to runtime + * @gpu_id (IN) - target device id + * @queue_id (IN) - target queue id + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * - ENODEV if gpu_id not found. + * If exception_mask contains EC_PROCESS_RUNTIME, unblocks pending + * AMDKFD_IOC_RUNTIME_ENABLE call - see kfd_ioctl_runtime_enable_args. + * All other exceptions are raised to runtime through err_payload_addr. + * See kfd_context_save_area_header. + */ +struct kfd_ioctl_dbg_trap_send_runtime_event_args { + __u64 exception_mask; + __u32 gpu_id; + __u32 queue_id; +}; + +/** + * kfd_ioctl_dbg_trap_set_exceptions_enabled_args + * + * Arguments for KFD_IOC_SET_EXCEPTIONS_ENABLED + * Set new exceptions to be raised to the debugger. + * + * @exception_mask (IN) - new exceptions to raise the debugger + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + */ +struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args { + __u64 exception_mask; +}; + +/** + * kfd_ioctl_dbg_trap_set_wave_launch_override_args + * + * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE + * Enable HW exceptions to raise trap. + * + * @override_mode (IN) - see kfd_dbg_trap_override_mode + * @enable_mask (IN/OUT) - reference kfd_dbg_trap_mask. + * IN is the override modes requested to be enabled. + * OUT is referenced in Return below. + * @support_request_mask (IN/OUT) - reference kfd_dbg_trap_mask. + * IN is the override modes requested for support check. + * OUT is referenced in Return below. + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * Previous enablement is returned in @enable_mask. + * Actual override support is returned in @support_request_mask. + * - EINVAL if override mode is not supported. + * - EACCES if trap support requested is not actually supported. + * i.e. enable_mask (IN) is not a subset of support_request_mask (OUT). + * Otherwise it is considered a generic error (see kfd_dbg_trap_operations). + */ +struct kfd_ioctl_dbg_trap_set_wave_launch_override_args { + __u32 override_mode; + __u32 enable_mask; + __u32 support_request_mask; + __u32 pad; +}; + +/** + * kfd_ioctl_dbg_trap_set_wave_launch_mode_args + * + * Arguments for KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE + * Set wave launch mode. + * + * @mode (IN) - see kfd_dbg_trap_wave_launch_mode + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + */ +struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args { + __u32 launch_mode; + __u32 pad; +}; + +/** + * kfd_ioctl_dbg_trap_suspend_queues_ags + * + * Arguments for KFD_IOC_DBG_TRAP_SUSPEND_QUEUES + * Suspend queues. + * + * @exception_mask (IN) - raised exceptions to clear + * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id) + * to suspend + * @num_queues (IN) - number of queues to suspend in @queue_array_ptr + * @grace_period (IN) - wave time allowance before preemption + * per 1K GPU clock cycle unit + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Destruction of a suspended queue is blocked until the queue is + * resumed. This allows the debugger to access queue information and + * the its context save area without running into a race condition on + * queue destruction. + * Automatically copies per queue context save area header information + * into the save area base + * (see kfd_queue_snapshot_entry and kfd_context_save_area_header). + * + * Return - Number of queues suspended on SUCCESS. + * . KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK masked + * for each queue id in @queue_array_ptr array reports unsuccessful + * suspend reason. + * KFD_DBG_QUEUE_ERROR_MASK = HW failure. + * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist, is new or + * is being destroyed. + */ +struct kfd_ioctl_dbg_trap_suspend_queues_args { + __u64 exception_mask; + __u64 queue_array_ptr; + __u32 num_queues; + __u32 grace_period; +}; + +/** + * kfd_ioctl_dbg_trap_resume_queues_args + * + * Arguments for KFD_IOC_DBG_TRAP_RESUME_QUEUES + * Resume queues. + * + * @queue_array_ptr (IN) - pointer to array of queue ids (u32 per queue id) + * to resume + * @num_queues (IN) - number of queues to resume in @queue_array_ptr + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - Number of queues resumed on SUCCESS. + * KFD_DBG_QUEUE_ERROR_MASK and KFD_DBG_QUEUE_INVALID_MASK mask + * for each queue id in @queue_array_ptr array reports unsuccessful + * resume reason. + * KFD_DBG_QUEUE_ERROR_MASK = HW failure. + * KFD_DBG_QUEUE_INVALID_MASK = queue does not exist. + */ +struct kfd_ioctl_dbg_trap_resume_queues_args { + __u64 queue_array_ptr; + __u32 num_queues; + __u32 pad; +}; + +/** + * kfd_ioctl_dbg_trap_set_node_address_watch_args + * + * Arguments for KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH + * Sets address watch for device. + * + * @address (IN) - watch address to set + * @mode (IN) - see kfd_dbg_trap_address_watch_mode + * @mask (IN) - watch address mask + * @gpu_id (IN) - target gpu to set watch point + * @id (OUT) - watch id allocated + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * Allocated watch ID returned to @id. + * - ENODEV if gpu_id not found. + * - ENOMEM if watch IDs can be allocated + */ +struct kfd_ioctl_dbg_trap_set_node_address_watch_args { + __u64 address; + __u32 mode; + __u32 mask; + __u32 gpu_id; + __u32 id; +}; + +/** + * kfd_ioctl_dbg_trap_clear_node_address_watch_args + * + * Arguments for KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH + * Clear address watch for device. + * + * @gpu_id (IN) - target device to clear watch point + * @id (IN) - allocated watch id to clear + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * - ENODEV if gpu_id not found. + * - EINVAL if watch ID has not been allocated. + */ +struct kfd_ioctl_dbg_trap_clear_node_address_watch_args { + __u32 gpu_id; + __u32 id; +}; + +/** + * kfd_ioctl_dbg_trap_set_flags_args + * + * Arguments for KFD_IOC_DBG_TRAP_SET_FLAGS + * Sets flags for wave behaviour. + * + * @flags (IN/OUT) - IN = flags to enable, OUT = flags previously enabled + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * - EACCESS if any debug device does not allow flag options. + */ +struct kfd_ioctl_dbg_trap_set_flags_args { + __u32 flags; + __u32 pad; +}; + +/** + * kfd_ioctl_dbg_trap_query_debug_event_args + * + * Arguments for KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT + * + * Find one or more raised exceptions. This function can return multiple + * exceptions from a single queue or a single device with one call. To find + * all raised exceptions, this function must be called repeatedly until it + * returns -EAGAIN. Returned exceptions can optionally be cleared by + * setting the corresponding bit in the @exception_mask input parameter. + * However, clearing an exception prevents retrieving further information + * about it with KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO. + * + * @exception_mask (IN/OUT) - exception to clear (IN) and raised (OUT) + * @gpu_id (OUT) - gpu id of exceptions raised + * @queue_id (OUT) - queue id of exceptions raised + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on raised exception found + * Raised exceptions found are returned in @exception mask + * with reported source id returned in @gpu_id or @queue_id. + * - EAGAIN if no raised exception has been found + */ +struct kfd_ioctl_dbg_trap_query_debug_event_args { + __u64 exception_mask; + __u32 gpu_id; + __u32 queue_id; +}; + +/** + * kfd_ioctl_dbg_trap_query_exception_info_args + * + * Arguments KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO + * Get additional info on raised exception. + * + * @info_ptr (IN) - pointer to exception info buffer to copy to + * @info_size (IN/OUT) - exception info buffer size (bytes) + * @source_id (IN) - target gpu or queue id + * @exception_code (IN) - target exception + * @clear_exception (IN) - clear raised @exception_code exception + * (0 = false, 1 = true) + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * If @exception_code is EC_DEVICE_MEMORY_VIOLATION, copy @info_size(OUT) + * bytes of memory exception data to @info_ptr. + * If @exception_code is EC_PROCESS_RUNTIME, copy saved + * kfd_runtime_info to @info_ptr. + * Actual required @info_ptr size (bytes) is returned in @info_size. + */ +struct kfd_ioctl_dbg_trap_query_exception_info_args { + __u64 info_ptr; + __u32 info_size; + __u32 source_id; + __u32 exception_code; + __u32 clear_exception; +}; + +/** + * kfd_ioctl_dbg_trap_get_queue_snapshot_args + * + * Arguments KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT + * Get queue information. + * + * @exception_mask (IN) - exceptions raised to clear + * @snapshot_buf_ptr (IN) - queue snapshot entry buffer (see kfd_queue_snapshot_entry) + * @num_queues (IN/OUT) - number of queue snapshot entries + * The debugger specifies the size of the array allocated in @num_queues. + * KFD returns the number of queues that actually existed. If this is + * larger than the size specified by the debugger, KFD will not overflow + * the array allocated by the debugger. + * + * @entry_size (IN/OUT) - size per entry in bytes + * The debugger specifies sizeof(struct kfd_queue_snapshot_entry) in + * @entry_size. KFD returns the number of bytes actually populated per + * entry. The debugger should use the KFD_IOCTL_MINOR_VERSION to determine, + * which fields in struct kfd_queue_snapshot_entry are valid. This allows + * growing the ABI in a backwards compatible manner. + * Note that entry_size(IN) should still be used to stride the snapshot buffer in the + * event that it's larger than actual kfd_queue_snapshot_entry. + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * Copies @num_queues(IN) queue snapshot entries of size @entry_size(IN) + * into @snapshot_buf_ptr if @num_queues(IN) > 0. + * Otherwise return @num_queues(OUT) queue snapshot entries that exist. + */ +struct kfd_ioctl_dbg_trap_queue_snapshot_args { + __u64 exception_mask; + __u64 snapshot_buf_ptr; + __u32 num_queues; + __u32 entry_size; +}; + +/** + * kfd_ioctl_dbg_trap_get_device_snapshot_args + * + * Arguments for KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT + * Get device information. + * + * @exception_mask (IN) - exceptions raised to clear + * @snapshot_buf_ptr (IN) - pointer to snapshot buffer (see kfd_dbg_device_info_entry) + * @num_devices (IN/OUT) - number of debug devices to snapshot + * The debugger specifies the size of the array allocated in @num_devices. + * KFD returns the number of devices that actually existed. If this is + * larger than the size specified by the debugger, KFD will not overflow + * the array allocated by the debugger. + * + * @entry_size (IN/OUT) - size per entry in bytes + * The debugger specifies sizeof(struct kfd_dbg_device_info_entry) in + * @entry_size. KFD returns the number of bytes actually populated. The + * debugger should use KFD_IOCTL_MINOR_VERSION to determine, which fields + * in struct kfd_dbg_device_info_entry are valid. This allows growing the + * ABI in a backwards compatible manner. + * Note that entry_size(IN) should still be used to stride the snapshot buffer in the + * event that it's larger than actual kfd_dbg_device_info_entry. + * + * Generic errors apply (see kfd_dbg_trap_operations). + * Return - 0 on SUCCESS. + * Copies @num_devices(IN) device snapshot entries of size @entry_size(IN) + * into @snapshot_buf_ptr if @num_devices(IN) > 0. + * Otherwise return @num_devices(OUT) queue snapshot entries that exist. + */ +struct kfd_ioctl_dbg_trap_device_snapshot_args { + __u64 exception_mask; + __u64 snapshot_buf_ptr; + __u32 num_devices; + __u32 entry_size; +}; + +/** + * kfd_ioctl_dbg_trap_args + * + * Arguments to debug target process. + * + * @pid - target process to debug + * @op - debug operation (see kfd_dbg_trap_operations) + * + * @op determines which union struct args to use. + * Refer to kern docs for each kfd_ioctl_dbg_trap_*_args struct. + */ +struct kfd_ioctl_dbg_trap_args { + __u32 pid; + __u32 op; + + union { + struct kfd_ioctl_dbg_trap_enable_args enable; + struct kfd_ioctl_dbg_trap_send_runtime_event_args send_runtime_event; + struct kfd_ioctl_dbg_trap_set_exceptions_enabled_args set_exceptions_enabled; + struct kfd_ioctl_dbg_trap_set_wave_launch_override_args launch_override; + struct kfd_ioctl_dbg_trap_set_wave_launch_mode_args launch_mode; + struct kfd_ioctl_dbg_trap_suspend_queues_args suspend_queues; + struct kfd_ioctl_dbg_trap_resume_queues_args resume_queues; + struct kfd_ioctl_dbg_trap_set_node_address_watch_args set_node_address_watch; + struct kfd_ioctl_dbg_trap_clear_node_address_watch_args clear_node_address_watch; + struct kfd_ioctl_dbg_trap_set_flags_args set_flags; + struct kfd_ioctl_dbg_trap_query_debug_event_args query_debug_event; + struct kfd_ioctl_dbg_trap_query_exception_info_args query_exception_info; + struct kfd_ioctl_dbg_trap_queue_snapshot_args queue_snapshot; + struct kfd_ioctl_dbg_trap_device_snapshot_args device_snapshot; + }; +}; + #define AMDKFD_IOCTL_BASE 'K' #define AMDKFD_IO(nr) _IO(AMDKFD_IOCTL_BASE, nr) #define AMDKFD_IOR(nr, type) _IOR(AMDKFD_IOCTL_BASE, nr, type) @@ -887,7 +1547,13 @@ struct kfd_ioctl_set_xnack_mode_args { #define AMDKFD_IOC_EXPORT_DMABUF \ AMDKFD_IOWR(0x24, struct kfd_ioctl_export_dmabuf_args) +#define AMDKFD_IOC_RUNTIME_ENABLE \ + AMDKFD_IOWR(0x25, struct kfd_ioctl_runtime_enable_args) + +#define AMDKFD_IOC_DBG_TRAP \ + AMDKFD_IOWR(0x26, struct kfd_ioctl_dbg_trap_args) + #define AMDKFD_COMMAND_START 0x01 -#define AMDKFD_COMMAND_END 0x25 +#define AMDKFD_COMMAND_END 0x27 #endif -- cgit v1.2.3 From d230f1bfe7a1977565ce1e2804ddb7b7a3d911ff Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 25 Mar 2022 12:39:06 -0400 Subject: drm/amdkfd: display debug capabilities Expose debug capabilities in the KFD topology node's HSA capabilities and debug properties flags. Ensure correct capabilities are exposed based on firmware support. Flag definitions can be referenced in uapi/linux/kfd_sysfs.h. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 101 ++++++++++++++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 6 ++ include/uapi/linux/kfd_sysfs.h | 15 +++++ 3 files changed, 117 insertions(+), 5 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 8302d8967158..3def25b2bdbb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -535,6 +535,8 @@ static ssize_t node_show(struct kobject *kobj, struct attribute *attr, dev->gpu->kfd->mec_fw_version); sysfs_show_32bit_prop(buffer, offs, "capability", dev->node_props.capability); + sysfs_show_64bit_prop(buffer, offs, "debug_prop", + dev->node_props.debug_prop); sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version", dev->gpu->kfd->sdma_fw_version); sysfs_show_64bit_prop(buffer, offs, "unique_id", @@ -1857,6 +1859,97 @@ err: return res; } +static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev) +{ + bool firmware_supported = true; + + if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) && + KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) { + firmware_supported = + (dev->gpu->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 9; + goto out; + } + + /* + * Note: Any unlisted devices here are assumed to support exception handling. + * Add additional checks here as needed. + */ + switch (KFD_GC_VERSION(dev->gpu)) { + case IP_VERSION(9, 0, 1): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768; + break; + case IP_VERSION(9, 1, 0): + case IP_VERSION(9, 2, 1): + case IP_VERSION(9, 2, 2): + case IP_VERSION(9, 3, 0): + case IP_VERSION(9, 4, 0): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 459; + break; + case IP_VERSION(9, 4, 1): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 60; + break; + case IP_VERSION(9, 4, 2): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 51; + break; + case IP_VERSION(10, 1, 10): + case IP_VERSION(10, 1, 2): + case IP_VERSION(10, 1, 1): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 144; + break; + case IP_VERSION(10, 3, 0): + case IP_VERSION(10, 3, 2): + case IP_VERSION(10, 3, 1): + case IP_VERSION(10, 3, 4): + case IP_VERSION(10, 3, 5): + firmware_supported = dev->gpu->kfd->mec_fw_version >= 89; + break; + case IP_VERSION(10, 1, 3): + case IP_VERSION(10, 3, 3): + firmware_supported = false; + break; + default: + break; + } + +out: + if (firmware_supported) + dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED; +} + +static void kfd_topology_set_capabilities(struct kfd_topology_device *dev) +{ + dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << + HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & + HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); + + dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT | + HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED | + HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED; + + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) { + dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 | + HSA_DBG_WATCH_ADDR_MASK_HI_BIT; + + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 4, 2)) + dev->node_props.debug_prop |= + HSA_DBG_DISPATCH_INFO_ALWAYS_VALID; + else + dev->node_props.capability |= + HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED; + } else { + dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 | + HSA_DBG_WATCH_ADDR_MASK_HI_BIT; + + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(11, 0, 0)) + dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID; + else + dev->node_props.capability |= + HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED; + } + + kfd_topology_set_dbg_firmware_support(dev); +} + int kfd_topology_add_device(struct kfd_node *gpu) { uint32_t gpu_id; @@ -1967,13 +2060,11 @@ int kfd_topology_add_device(struct kfd_node *gpu) HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); break; default: - if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 0, 1)) - dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << - HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & - HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); - else + if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1)) WARN(1, "Unexpected ASIC family %u", dev->gpu->adev->asic_type); + else + kfd_topology_set_capabilities(dev); } /* diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h index 3b8afb6aba79..cba2cd5ed9d1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h @@ -31,6 +31,11 @@ #define KFD_TOPOLOGY_PUBLIC_NAME_SIZE 32 +#define HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 6 +#define HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 7 +#define HSA_DBG_WATCH_ADDR_MASK_HI_BIT \ + (29 << HSA_DBG_WATCH_ADDR_MASK_HI_BIT_SHIFT) + struct kfd_node_properties { uint64_t hive_id; uint32_t cpu_cores_count; @@ -42,6 +47,7 @@ struct kfd_node_properties { uint32_t cpu_core_id_base; uint32_t simd_id_base; uint32_t capability; + uint64_t debug_prop; uint32_t max_waves_per_simd; uint32_t lds_size_in_kb; uint32_t gds_size_in_kb; diff --git a/include/uapi/linux/kfd_sysfs.h b/include/uapi/linux/kfd_sysfs.h index 3e330f368917..a51b7331e0b4 100644 --- a/include/uapi/linux/kfd_sysfs.h +++ b/include/uapi/linux/kfd_sysfs.h @@ -43,6 +43,11 @@ #define HSA_CAP_DOORBELL_TYPE_2_0 0x2 #define HSA_CAP_AQL_QUEUE_DOUBLE_MAP 0x00004000 +#define HSA_CAP_TRAP_DEBUG_SUPPORT 0x00008000 +#define HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED 0x00010000 +#define HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED 0x00020000 +#define HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED 0x00040000 + /* Old buggy user mode depends on this being 0 */ #define HSA_CAP_RESERVED_WAS_SRAM_EDCSUPPORTED 0x00080000 @@ -53,8 +58,18 @@ #define HSA_CAP_SRAM_EDCSUPPORTED 0x04000000 #define HSA_CAP_SVMAPI_SUPPORTED 0x08000000 #define HSA_CAP_FLAGS_COHERENTHOSTACCESS 0x10000000 +#define HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED 0x20000000 #define HSA_CAP_RESERVED 0xe00f8000 +/* debug_prop bits in node properties */ +#define HSA_DBG_WATCH_ADDR_MASK_LO_BIT_MASK 0x0000000f +#define HSA_DBG_WATCH_ADDR_MASK_LO_BIT_SHIFT 0 +#define HSA_DBG_WATCH_ADDR_MASK_HI_BIT_MASK 0x000003f0 +#define HSA_DBG_WATCH_ADDR_MASK_HI_BIT_SHIFT 4 +#define HSA_DBG_DISPATCH_INFO_ALWAYS_VALID 0x00000400 +#define HSA_DBG_WATCHPOINTS_EXCLUSIVE 0x00000800 +#define HSA_DBG_RESERVED 0xfffffffffffff000ull + /* Heap types in memory properties */ #define HSA_MEM_HEAP_TYPE_SYSTEM 0 #define HSA_MEM_HEAP_TYPE_FB_PUBLIC 1 -- cgit v1.2.3 From 0ab2d7532b05a3e7c06fd3b0c8bd6b46c1dfb508 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 25 Mar 2022 14:55:30 -0400 Subject: drm/amdkfd: prepare per-process debug enable and disable The ROCm debugger will attach to a process to debug by PTRACE and will expect the KFD to prepare a process for the target PID, whether the target PID has opened the KFD device or not. This patch is to explicity handle this requirement. Further HW mode setting and runtime coordination requirements will be handled in following patches. In the case where the target process has not opened the KFD device, a new KFD process must be created for the target PID. The debugger as well as the target process for this case will have not acquired any VMs so handle process restoration to correctly account for this. To coordinate with HSA runtime, the debugger must be aware of the target process' runtime enablement status and will copy the runtime status information into the debugged KFD process for later query. On enablement, the debugger will subscribe to a set of exceptions where each exception events will notify the debugger through a pollable FIFO file descriptor that the debugger provides to the KFD to manage. Finally on process termination of either the debugger or the target, debugging must be disabled if it has not been done so. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/Makefile | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 102 ++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 80 ++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 32 +++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 26 ++++-- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 31 ++++++- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 60 ++++++++---- 7 files changed, 304 insertions(+), 30 deletions(-) create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_debug.c create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_debug.h (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/Makefile b/drivers/gpu/drm/amd/amdkfd/Makefile index e758c2a24cd0..747754428073 100644 --- a/drivers/gpu/drm/amd/amdkfd/Makefile +++ b/drivers/gpu/drm/amd/amdkfd/Makefile @@ -55,7 +55,8 @@ AMDKFD_FILES := $(AMDKFD_PATH)/kfd_module.o \ $(AMDKFD_PATH)/kfd_int_process_v9.o \ $(AMDKFD_PATH)/kfd_int_process_v11.o \ $(AMDKFD_PATH)/kfd_smi_events.o \ - $(AMDKFD_PATH)/kfd_crat.o + $(AMDKFD_PATH)/kfd_crat.o \ + $(AMDKFD_PATH)/kfd_debug.o ifneq ($(CONFIG_AMD_IOMMU_V2),) AMDKFD_FILES += $(AMDKFD_PATH)/kfd_iommu.o diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 00e34125987c..ee086a0a46df 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -44,6 +44,7 @@ #include "amdgpu_amdkfd.h" #include "kfd_smi_events.h" #include "amdgpu_dma_buf.h" +#include "kfd_debug.h" static long kfd_ioctl(struct file *, unsigned int, unsigned long); static int kfd_open(struct inode *, struct file *); @@ -142,10 +143,15 @@ static int kfd_open(struct inode *inode, struct file *filep) return -EPERM; } - process = kfd_create_process(filep); + process = kfd_create_process(current); if (IS_ERR(process)) return PTR_ERR(process); + if (kfd_process_init_cwsr_apu(process, filep)) { + kfd_unref_process(process); + return -EFAULT; + } + /* filep now owns the reference returned by kfd_create_process */ filep->private_data = process; @@ -2737,6 +2743,10 @@ static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, v static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data) { struct kfd_ioctl_dbg_trap_args *args = data; + struct task_struct *thread = NULL; + struct mm_struct *mm = NULL; + struct pid *pid = NULL; + struct kfd_process *target = NULL; int r = 0; if (sched_policy == KFD_SCHED_POLICY_NO_HWS) { @@ -2744,9 +2754,81 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v return -EINVAL; } + pid = find_get_pid(args->pid); + if (!pid) { + pr_debug("Cannot find pid info for %i\n", args->pid); + r = -ESRCH; + goto out; + } + + thread = get_pid_task(pid, PIDTYPE_PID); + if (!thread) { + r = -ESRCH; + goto out; + } + + mm = get_task_mm(thread); + if (!mm) { + r = -ESRCH; + goto out; + } + + if (args->op == KFD_IOC_DBG_TRAP_ENABLE) { + bool create_process; + + rcu_read_lock(); + create_process = thread && thread != current && ptrace_parent(thread) == current; + rcu_read_unlock(); + + target = create_process ? kfd_create_process(thread) : + kfd_lookup_process_by_pid(pid); + } else { + target = kfd_lookup_process_by_pid(pid); + } + + if (!target) { + pr_debug("Cannot find process PID %i to debug\n", args->pid); + r = -ESRCH; + goto out; + } + + /* Check if target is still PTRACED. */ + rcu_read_lock(); + if (target != p && args->op != KFD_IOC_DBG_TRAP_DISABLE + && ptrace_parent(target->lead_thread) != current) { + pr_err("PID %i is not PTRACED and cannot be debugged\n", args->pid); + r = -EPERM; + } + rcu_read_unlock(); + + if (r) + goto out; + + mutex_lock(&target->mutex); + + if (args->op != KFD_IOC_DBG_TRAP_ENABLE && !target->debug_trap_enabled) { + pr_err("PID %i not debug enabled for op %i\n", args->pid, args->op); + r = -EINVAL; + goto unlock_out; + } + switch (args->op) { case KFD_IOC_DBG_TRAP_ENABLE: + if (target != p) + target->debugger_process = p; + + r = kfd_dbg_trap_enable(target, + args->enable.dbg_fd, + (void __user *)args->enable.rinfo_ptr, + &args->enable.rinfo_size); + if (!r) + target->exception_enable_mask = args->enable.exception_mask; + + pr_warn("Debug functions limited\n"); + break; case KFD_IOC_DBG_TRAP_DISABLE: + r = kfd_dbg_trap_disable(target); + break; case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT: case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED: case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE: @@ -2760,7 +2842,7 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: - pr_warn("Debugging not supported yet\n"); + pr_warn("Debug op %i not supported yet\n", args->op); r = -EACCES; break; default: @@ -2768,6 +2850,22 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v r = -EINVAL; } +unlock_out: + mutex_unlock(&target->mutex); + +out: + if (thread) + put_task_struct(thread); + + if (mm) + mmput(mm); + + if (pid) + put_pid(pid); + + if (target) + kfd_unref_process(target); + return r; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c new file mode 100644 index 000000000000..898cc1fe3d13 --- /dev/null +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -0,0 +1,80 @@ +/* + * Copyright 2023 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + */ + +#include "kfd_debug.h" +#include + +int kfd_dbg_trap_disable(struct kfd_process *target) +{ + if (!target->debug_trap_enabled) + return 0; + + fput(target->dbg_ev_file); + target->dbg_ev_file = NULL; + + if (target->debugger_process) { + atomic_dec(&target->debugger_process->debugged_process_count); + target->debugger_process = NULL; + } + + target->debug_trap_enabled = false; + kfd_unref_process(target); + + return 0; +} + +int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, + void __user *runtime_info, uint32_t *runtime_size) +{ + struct file *f; + uint32_t copy_size; + int r = 0; + + if (target->debug_trap_enabled) + return -EALREADY; + + copy_size = min((size_t)(*runtime_size), sizeof(target->runtime_info)); + + f = fget(fd); + if (!f) { + pr_err("Failed to get file for (%i)\n", fd); + return -EBADF; + } + + target->dbg_ev_file = f; + + /* We already hold the process reference but hold another one for the + * debug session. + */ + kref_get(&target->ref); + target->debug_trap_enabled = true; + + if (target->debugger_process) + atomic_inc(&target->debugger_process->debugged_process_count); + + if (copy_to_user(runtime_info, (void *)&target->runtime_info, copy_size)) + r = -EFAULT; + + *runtime_size = sizeof(target->runtime_info); + + return r; +} diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h new file mode 100644 index 000000000000..a8abfe2a0a14 --- /dev/null +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -0,0 +1,32 @@ +/* + * Copyright 2023 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + */ + +#ifndef KFD_DEBUG_EVENTS_H_INCLUDED +#define KFD_DEBUG_EVENTS_H_INCLUDED + +#include "kfd_priv.h" + +int kfd_dbg_trap_disable(struct kfd_process *target); +int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, + void __user *runtime_info, + uint32_t *runtime_info_size); +#endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 80cddb46657f..2baa0781eafc 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1006,6 +1006,14 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, goto out; pdd = qpd_to_pdd(qpd); + + /* The debugger creates processes that temporarily have not acquired + * all VMs for all devices and has no VMs itself. + * Skip queue eviction on process eviction. + */ + if (!pdd->drm_priv) + goto out; + pr_debug_ratelimited("Evicting PASID 0x%x queues\n", pdd->process->pasid); @@ -1127,13 +1135,10 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, { struct queue *q; struct kfd_process_device *pdd; - uint64_t pd_base; uint64_t eviction_duration; int retval = 0; pdd = qpd_to_pdd(qpd); - /* Retrieve PD base */ - pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); dqm_lock(dqm); if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ @@ -1143,12 +1148,19 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, goto out; } + /* The debugger creates processes that temporarily have not acquired + * all VMs for all devices and has no VMs itself. + * Skip queue restore on process restore. + */ + if (!pdd->drm_priv) + goto vm_not_acquired; + pr_debug_ratelimited("Restoring PASID 0x%x queues\n", pdd->process->pasid); /* Update PD Base in QPD */ - qpd->page_table_base = pd_base; - pr_debug("Updated PD address to 0x%llx\n", pd_base); + qpd->page_table_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); + pr_debug("Updated PD address to 0x%llx\n", qpd->page_table_base); /* activate all active queues on the qpd */ list_for_each_entry(q, &qpd->queues_list, list) { @@ -1171,9 +1183,11 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); - qpd->evicted = 0; + eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; atomic64_add(eviction_duration, &pdd->evict_duration_counter); +vm_not_acquired: + qpd->evicted = 0; out: dqm_unlock(dqm); return retval; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 3bd222e8f6c3..1b272f879b4c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -920,11 +920,33 @@ struct kfd_process { */ unsigned long last_restore_timestamp; + /* Indicates device process is debug attached with reserved vmid. */ + bool debug_trap_enabled; + + /* per-process-per device debug event fd file */ + struct file *dbg_ev_file; + + /* If the process is a kfd debugger, we need to know so we can clean + * up at exit time. If a process enables debugging on itself, it does + * its own clean-up, so we don't set the flag here. We track this by + * counting the number of processes this process is debugging. + */ + atomic_t debugged_process_count; + + /* If the process is a debugged, this is the debugger process */ + struct kfd_process *debugger_process; + /* Kobj for our procfs */ struct kobject *kobj; struct kobject *kobj_queues; struct attribute attr_pasid; + /* Keep track cwsr init */ + bool has_cwsr; + + /* Exception code enable mask and status */ + uint64_t exception_enable_mask; + /* shared virtual memory registered by this process */ struct svm_range_list svms; @@ -933,6 +955,10 @@ struct kfd_process { atomic_t poison; /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */ bool queues_paused; + + /* Tracks runtime enable status */ + struct kfd_runtime_info runtime_info; + }; #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ @@ -963,7 +989,7 @@ bool kfd_dev_is_large_bar(struct kfd_node *dev); int kfd_process_create_wq(void); void kfd_process_destroy_wq(void); void kfd_cleanup_processes(void); -struct kfd_process *kfd_create_process(struct file *filep); +struct kfd_process *kfd_create_process(struct task_struct *thread); struct kfd_process *kfd_get_process(const struct task_struct *task); struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid); struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); @@ -1108,6 +1134,9 @@ void kfd_process_set_trap_handler(struct qcm_process_device *qpd, uint64_t tba_addr, uint64_t tma_addr); +/* CWSR initialization */ +int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep); + /* CRIU */ /* * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 7f7d1378a2f8..ef67f5e37c6f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -44,6 +44,7 @@ struct mm_struct; #include "kfd_iommu.h" #include "kfd_svm.h" #include "kfd_smi_events.h" +#include "kfd_debug.h" /* * List of struct kfd_process (field kfd_process). @@ -69,7 +70,6 @@ static struct kfd_process *find_process(const struct task_struct *thread, bool ref); static void kfd_process_ref_release(struct kref *ref); static struct kfd_process *create_process(const struct task_struct *thread); -static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep); static void evict_process_worker(struct work_struct *work); static void restore_process_worker(struct work_struct *work); @@ -798,18 +798,19 @@ static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd) kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr); } -struct kfd_process *kfd_create_process(struct file *filep) +struct kfd_process *kfd_create_process(struct task_struct *thread) { struct kfd_process *process; - struct task_struct *thread = current; int ret; - if (!thread->mm) + if (!(thread->mm && mmget_not_zero(thread->mm))) return ERR_PTR(-EINVAL); /* Only the pthreads threading model is supported. */ - if (thread->group_leader->mm != thread->mm) + if (thread->group_leader->mm != thread->mm) { + mmput(thread->mm); return ERR_PTR(-EINVAL); + } /* * take kfd processes mutex before starting of process creation @@ -833,10 +834,6 @@ struct kfd_process *kfd_create_process(struct file *filep) if (IS_ERR(process)) goto out; - ret = kfd_process_init_cwsr_apu(process, filep); - if (ret) - goto out_destroy; - if (!procfs.kobj) goto out; @@ -870,16 +867,9 @@ out: if (!IS_ERR(process)) kref_get(&process->ref); mutex_unlock(&kfd_processes_mutex); + mmput(thread->mm); return process; - -out_destroy: - hash_del_rcu(&process->kfd_processes); - mutex_unlock(&kfd_processes_mutex); - synchronize_srcu(&kfd_processes_srcu); - /* kfd_process_free_notifier will trigger the cleanup */ - mmu_notifier_put(&process->mmu_notifier); - return ERR_PTR(ret); } struct kfd_process *kfd_get_process(const struct task_struct *thread) @@ -1180,6 +1170,25 @@ static void kfd_process_notifier_release_internal(struct kfd_process *p) /* Indicate to other users that MM is no longer valid */ p->mm = NULL; + kfd_dbg_trap_disable(p); + + if (atomic_read(&p->debugged_process_count) > 0) { + struct kfd_process *target; + unsigned int temp; + int idx = srcu_read_lock(&kfd_processes_srcu); + + hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) { + if (target->debugger_process && target->debugger_process == p) { + mutex_lock_nested(&target->mutex, 1); + kfd_dbg_trap_disable(target); + mutex_unlock(&target->mutex); + if (atomic_read(&p->debugged_process_count) == 0) + break; + } + } + + srcu_read_unlock(&kfd_processes_srcu, idx); + } mmu_notifier_put(&p->mmu_notifier); } @@ -1259,11 +1268,14 @@ void kfd_cleanup_processes(void) mmu_notifier_synchronize(); } -static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) +int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) { unsigned long offset; int i; + if (p->has_cwsr) + return 0; + for (i = 0; i < p->n_pdds; i++) { struct kfd_node *dev = p->pdds[i]->dev; struct qcm_process_device *qpd = &p->pdds[i]->qpd; @@ -1292,6 +1304,8 @@ static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr); } + p->has_cwsr = true; + return 0; } @@ -1434,6 +1448,10 @@ static struct kfd_process *create_process(const struct task_struct *thread) if (err) goto err_event_init; process->is_32bit_user_mode = in_compat_syscall(); + process->debug_trap_enabled = false; + process->debugger_process = NULL; + process->exception_enable_mask = 0; + atomic_set(&process->debugged_process_count, 0); process->pasid = kfd_pasid_alloc(); if (process->pasid == 0) { @@ -1967,8 +1985,10 @@ static void restore_process_worker(struct work_struct *work) */ p->last_restore_timestamp = get_jiffies_64(); - ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info, - &p->ef); + /* VMs may not have been acquired yet during debugging. */ + if (p->kgd_process_info) + ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info, + &p->ef); if (ret) { pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n", p->pasid, PROCESS_BACK_OFF_TIME_MS); -- cgit v1.2.3 From 4504f14338cdc43586189558113faafa8acb9ffe Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 31 Mar 2022 12:05:00 -0400 Subject: drm/amdgpu: setup hw debug registers on driver initialization Add missing debug trap registers references and initialize all debug registers on boot by clearing the hardware exception overrides and the wave allocation ID index. The debugger requires that TTMPs 6 & 7 save the dispatch ID to map waves onto dispatch during compute context inspection. In order to correctly set this up, set the special reserved CP bit by default whenever the MQD is initailized. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 26 ++++++++ drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 1 + drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 30 ++++++++++ drivers/gpu/drm/amd/amdgpu/gfx_v9_4_2.c | 3 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 5 ++ .../drm/amd/include/asic_reg/gc/gc_10_1_0_offset.h | 14 +++++ .../amd/include/asic_reg/gc/gc_10_1_0_sh_mask.h | 69 ++++++++++++++++++++++ .../drm/amd/include/asic_reg/gc/gc_10_3_0_offset.h | 10 ++++ .../amd/include/asic_reg/gc/gc_10_3_0_sh_mask.h | 4 ++ .../amd/include/asic_reg/gc/gc_11_0_0_sh_mask.h | 4 ++ 12 files changed, 176 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c index f7ad883a70fa..be984f8c71c7 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c @@ -4825,6 +4825,29 @@ static u32 gfx_v10_0_init_pa_sc_tile_steering_override(struct amdgpu_device *ade #define DEFAULT_SH_MEM_BASES (0x6000) +static void gfx_v10_0_debug_trap_config_init(struct amdgpu_device *adev, + uint32_t first_vmid, + uint32_t last_vmid) +{ + uint32_t data; + uint32_t trap_config_vmid_mask = 0; + int i; + + /* Calculate trap config vmid mask */ + for (i = first_vmid; i < last_vmid; i++) + trap_config_vmid_mask |= (1 << i); + + data = REG_SET_FIELD(0, SPI_GDBG_TRAP_CONFIG, + VMID_SEL, trap_config_vmid_mask); + data = REG_SET_FIELD(data, SPI_GDBG_TRAP_CONFIG, + TRAP_EN, 1); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_CONFIG), data); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0); + + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_DATA0), 0); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_DATA1), 0); +} + static void gfx_v10_0_init_compute_vmid(struct amdgpu_device *adev) { int i; @@ -4856,6 +4879,9 @@ static void gfx_v10_0_init_compute_vmid(struct amdgpu_device *adev) WREG32_SOC15_OFFSET(GC, 0, mmGDS_GWS_VMID0, i, 0); WREG32_SOC15_OFFSET(GC, 0, mmGDS_OA_VMID0, i, 0); } + + gfx_v10_0_debug_trap_config_init(adev, adev->vm_manager.first_kfd_vmid, + AMDGPU_NUM_VMID); } static void gfx_v10_0_init_gds_vmid(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c index da21bf868080..690e121d9dda 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c @@ -1638,6 +1638,7 @@ static void gfx_v11_0_init_compute_vmid(struct amdgpu_device *adev) /* Enable trap for each kfd vmid. */ data = RREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL); data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1); + WREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL, data); } soc21_grbm_select(adev, 0, 0, 0, 0); mutex_unlock(&adev->srbm_mutex); diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c index 0189e50bd89f..7f17e0061027 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c @@ -2303,6 +2303,29 @@ static void gfx_v9_0_setup_rb(struct amdgpu_device *adev) adev->gfx.config.num_rbs = hweight32(active_rbs); } +static void gfx_v9_0_debug_trap_config_init(struct amdgpu_device *adev, + uint32_t first_vmid, + uint32_t last_vmid) +{ + uint32_t data; + uint32_t trap_config_vmid_mask = 0; + int i; + + /* Calculate trap config vmid mask */ + for (i = first_vmid; i < last_vmid; i++) + trap_config_vmid_mask |= (1 << i); + + data = REG_SET_FIELD(0, SPI_GDBG_TRAP_CONFIG, + VMID_SEL, trap_config_vmid_mask); + data = REG_SET_FIELD(data, SPI_GDBG_TRAP_CONFIG, + TRAP_EN, 1); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_CONFIG), data); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0); + + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_DATA0), 0); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_DATA1), 0); +} + #define DEFAULT_SH_MEM_BASES (0x6000) static void gfx_v9_0_init_compute_vmid(struct amdgpu_device *adev) { @@ -4602,6 +4625,13 @@ static int gfx_v9_0_late_init(void *handle) if (r) return r; + if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2)) + gfx_v9_4_2_debug_trap_config_init(adev, + adev->vm_manager.first_kfd_vmid, AMDGPU_NUM_VMID); + else + gfx_v9_0_debug_trap_config_init(adev, + adev->vm_manager.first_kfd_vmid, AMDGPU_NUM_VMID); + return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_2.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_2.c index 2cc3a7cb1f54..63f6843a069e 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_2.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_2.c @@ -771,6 +771,9 @@ void gfx_v9_4_2_debug_trap_config_init(struct amdgpu_device *adev, soc15_grbm_select(adev, 0, 0, 0, 0, 0); mutex_unlock(&adev->srbm_mutex); + + WREG32(SOC15_REG_OFFSET(GC, 0, regSPI_GDBG_TRAP_DATA0), 0); + WREG32(SOC15_REG_OFFSET(GC, 0, regSPI_GDBG_TRAP_DATA1), 0); } void gfx_v9_4_2_set_power_brake_sequence(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index 772c09b5821b..eaaa4f4ddaaa 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -117,6 +117,11 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; + /* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the + * DISPATCH_PTR. This is required for the kfd debugger + */ + m->cp_hqd_hq_scheduler0 = 1 << 14; + if (q->format == KFD_QUEUE_FORMAT_AQL) { m->cp_hqd_aql_control = 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index 632344b95d90..3a48bbc589fe 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -143,6 +143,11 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; + /* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the + * DISPATCH_PTR. This is required for the kfd debugger + */ + m->cp_hqd_hq_status0 = 1 << 14; + /* * GFX11 RS64 CPFW version >= 509 supports PCIe atomics support * acknowledgment. diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 226132ec3714..b7c95158d4a0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -182,6 +182,11 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT | 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT; + /* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the + * DISPATCH_PTR. This is required for the kfd debugger + */ + m->cp_hqd_hq_status0 = 1 << 14; + if (q->format == KFD_QUEUE_FORMAT_AQL) m->cp_hqd_aql_control = 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT; diff --git a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_offset.h b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_offset.h index 79c41004c0b6..4908044f7409 100644 --- a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_offset.h +++ b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_offset.h @@ -5194,6 +5194,20 @@ #define mmSPI_WCL_PIPE_PERCENT_CS6_BASE_IDX 0 #define mmSPI_WCL_PIPE_PERCENT_CS7 0x1f70 #define mmSPI_WCL_PIPE_PERCENT_CS7_BASE_IDX 0 +#define mmSPI_GDBG_WAVE_CNTL 0x1f71 +#define mmSPI_GDBG_WAVE_CNTL_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_CONFIG 0x1f72 +#define mmSPI_GDBG_TRAP_CONFIG_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_MASK 0x1f73 +#define mmSPI_GDBG_TRAP_MASK_BASE_IDX 0 +#define mmSPI_GDBG_WAVE_CNTL2 0x1f74 +#define mmSPI_GDBG_WAVE_CNTL2_BASE_IDX 0 +#define mmSPI_GDBG_WAVE_CNTL3 0x1f75 +#define mmSPI_GDBG_WAVE_CNTL3_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_DATA0 0x1f78 +#define mmSPI_GDBG_TRAP_DATA0_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_DATA1 0x1f79 +#define mmSPI_GDBG_TRAP_DATA1_BASE_IDX 0 #define mmSPI_COMPUTE_QUEUE_RESET 0x1f7b #define mmSPI_COMPUTE_QUEUE_RESET_BASE_IDX 0 #define mmSPI_RESOURCE_RESERVE_CU_0 0x1f7c diff --git a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_sh_mask.h b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_sh_mask.h index 52043e143067..9b7d219e7954 100644 --- a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_sh_mask.h +++ b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_1_0_sh_mask.h @@ -19700,6 +19700,75 @@ //SPI_WCL_PIPE_PERCENT_CS7 #define SPI_WCL_PIPE_PERCENT_CS7__VALUE__SHIFT 0x0 #define SPI_WCL_PIPE_PERCENT_CS7__VALUE_MASK 0x7FL +//SPI_GDBG_WAVE_CNTL +#define SPI_GDBG_WAVE_CNTL__STALL_RA__SHIFT 0x0 +#define SPI_GDBG_WAVE_CNTL__STALL_VMID__SHIFT 0x1 +#define SPI_GDBG_WAVE_CNTL__STALL_RA_MASK 0x00000001L +#define SPI_GDBG_WAVE_CNTL__STALL_VMID_MASK 0x0001FFFEL +//SPI_GDBG_TRAP_CONFIG +#define SPI_GDBG_TRAP_CONFIG__ME_SEL__SHIFT 0x0 +#define SPI_GDBG_TRAP_CONFIG__PIPE_SEL__SHIFT 0x2 +#define SPI_GDBG_TRAP_CONFIG__QUEUE_SEL__SHIFT 0x4 +#define SPI_GDBG_TRAP_CONFIG__ME_MATCH__SHIFT 0x7 +#define SPI_GDBG_TRAP_CONFIG__PIPE_MATCH__SHIFT 0x8 +#define SPI_GDBG_TRAP_CONFIG__QUEUE_MATCH__SHIFT 0x9 +#define SPI_GDBG_TRAP_CONFIG__TRAP_EN__SHIFT 0xf +#define SPI_GDBG_TRAP_CONFIG__VMID_SEL__SHIFT 0x10 +#define SPI_GDBG_TRAP_CONFIG__ME_SEL_MASK 0x00000003L +#define SPI_GDBG_TRAP_CONFIG__PIPE_SEL_MASK 0x0000000CL +#define SPI_GDBG_TRAP_CONFIG__QUEUE_SEL_MASK 0x00000070L +#define SPI_GDBG_TRAP_CONFIG__ME_MATCH_MASK 0x00000080L +#define SPI_GDBG_TRAP_CONFIG__PIPE_MATCH_MASK 0x00000100L +#define SPI_GDBG_TRAP_CONFIG__QUEUE_MATCH_MASK 0x00000200L +#define SPI_GDBG_TRAP_CONFIG__TRAP_EN_MASK 0x00008000L +#define SPI_GDBG_TRAP_CONFIG__VMID_SEL_MASK 0xFFFF0000L +//SPI_GDBG_TRAP_MASK +#define SPI_GDBG_TRAP_MASK__EXCP_EN__SHIFT 0x0 +#define SPI_GDBG_TRAP_MASK__REPLACE__SHIFT 0x9 +#define SPI_GDBG_TRAP_MASK__EXCP_EN_MASK 0x01FFL +#define SPI_GDBG_TRAP_MASK__REPLACE_MASK 0x0200L +//SPI_GDBG_WAVE_CNTL2 +#define SPI_GDBG_WAVE_CNTL2__VMID_MASK__SHIFT 0x0 +#define SPI_GDBG_WAVE_CNTL2__MODE__SHIFT 0x10 +#define SPI_GDBG_WAVE_CNTL2__VMID_MASK_MASK 0x0000FFFFL +#define SPI_GDBG_WAVE_CNTL2__MODE_MASK 0x00030000L +//SPI_GDBG_WAVE_CNTL3 +#define SPI_GDBG_WAVE_CNTL3__STALL_PS__SHIFT 0x0 +#define SPI_GDBG_WAVE_CNTL3__STALL_VS__SHIFT 0x1 +#define SPI_GDBG_WAVE_CNTL3__STALL_GS__SHIFT 0x2 +#define SPI_GDBG_WAVE_CNTL3__STALL_HS__SHIFT 0x3 +#define SPI_GDBG_WAVE_CNTL3__STALL_CSG__SHIFT 0x4 +#define SPI_GDBG_WAVE_CNTL3__STALL_CS0__SHIFT 0x5 +#define SPI_GDBG_WAVE_CNTL3__STALL_CS1__SHIFT 0x6 +#define SPI_GDBG_WAVE_CNTL3__STALL_CS2__SHIFT 0x7 +#define SPI_GDBG_WAVE_CNTL3__STALL_CS3__SHIFT 0x8 +#define SPI_GDBG_WAVE_CNTL3__STALL_CS4__SHIFT 0x9 +#define SPI_GDBG_WAVE_CNTL3__STALL_CS5__SHIFT 0xa +#define SPI_GDBG_WAVE_CNTL3__STALL_CS6__SHIFT 0xb +#define SPI_GDBG_WAVE_CNTL3__STALL_CS7__SHIFT 0xc +#define SPI_GDBG_WAVE_CNTL3__STALL_DURATION__SHIFT 0xd +#define SPI_GDBG_WAVE_CNTL3__STALL_MULT__SHIFT 0x1c +#define SPI_GDBG_WAVE_CNTL3__STALL_PS_MASK 0x00000001L +#define SPI_GDBG_WAVE_CNTL3__STALL_VS_MASK 0x00000002L +#define SPI_GDBG_WAVE_CNTL3__STALL_GS_MASK 0x00000004L +#define SPI_GDBG_WAVE_CNTL3__STALL_HS_MASK 0x00000008L +#define SPI_GDBG_WAVE_CNTL3__STALL_CSG_MASK 0x00000010L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS0_MASK 0x00000020L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS1_MASK 0x00000040L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS2_MASK 0x00000080L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS3_MASK 0x00000100L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS4_MASK 0x00000200L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS5_MASK 0x00000400L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS6_MASK 0x00000800L +#define SPI_GDBG_WAVE_CNTL3__STALL_CS7_MASK 0x00001000L +#define SPI_GDBG_WAVE_CNTL3__STALL_DURATION_MASK 0x0FFFE000L +#define SPI_GDBG_WAVE_CNTL3__STALL_MULT_MASK 0x10000000L +//SPI_GDBG_TRAP_DATA0 +#define SPI_GDBG_TRAP_DATA0__DATA__SHIFT 0x0 +#define SPI_GDBG_TRAP_DATA0__DATA_MASK 0xFFFFFFFFL +//SPI_GDBG_TRAP_DATA1 +#define SPI_GDBG_TRAP_DATA1__DATA__SHIFT 0x0 +#define SPI_GDBG_TRAP_DATA1__DATA_MASK 0xFFFFFFFFL //SPI_COMPUTE_QUEUE_RESET #define SPI_COMPUTE_QUEUE_RESET__RESET__SHIFT 0x0 #define SPI_COMPUTE_QUEUE_RESET__RESET_MASK 0x01L diff --git a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_offset.h b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_offset.h index a734abaa91a5..5e15ac14b63c 100644 --- a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_offset.h +++ b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_offset.h @@ -26,6 +26,8 @@ #define mmSQ_DEBUG_STS_GLOBAL_BASE_IDX 0 #define mmSQ_DEBUG_STS_GLOBAL2 0x10B0 #define mmSQ_DEBUG_STS_GLOBAL2_BASE_IDX 0 +#define mmSQ_DEBUG 0x10B1 +#define mmSQ_DEBUG_BASE_IDX 0 // addressBlock: gc_sdma0_sdma0dec // base address: 0x4980 @@ -4853,10 +4855,18 @@ #define mmSPI_WCL_PIPE_PERCENT_CS3_BASE_IDX 0 #define mmSPI_GDBG_WAVE_CNTL 0x1f71 #define mmSPI_GDBG_WAVE_CNTL_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_CONFIG 0x1f72 +#define mmSPI_GDBG_TRAP_CONFIG_BASE_IDX 0 #define mmSPI_GDBG_TRAP_MASK 0x1f73 #define mmSPI_GDBG_TRAP_MASK_BASE_IDX 0 #define mmSPI_GDBG_WAVE_CNTL2 0x1f74 #define mmSPI_GDBG_WAVE_CNTL2_BASE_IDX 0 +#define mmSPI_GDBG_WAVE_CNTL3 0x1f75 +#define mmSPI_GDBG_WAVE_CNTL3_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_DATA0 0x1f78 +#define mmSPI_GDBG_TRAP_DATA0_BASE_IDX 0 +#define mmSPI_GDBG_TRAP_DATA1 0x1f79 +#define mmSPI_GDBG_TRAP_DATA1_BASE_IDX 0 #define mmSPI_COMPUTE_QUEUE_RESET 0x1f7b #define mmSPI_COMPUTE_QUEUE_RESET_BASE_IDX 0 #define mmSPI_RESOURCE_RESERVE_CU_0 0x1f7c diff --git a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_sh_mask.h b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_sh_mask.h index d7a17bae2584..e4ecd6c2d20e 100644 --- a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_sh_mask.h +++ b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_10_3_0_sh_mask.h @@ -47907,6 +47907,10 @@ // addressBlock: sqind +//SQ_DEBUG +#define SQ_DEBUG__SINGLE_MEMOP_MASK 0x00000001L +#define SQ_DEBUG__SINGLE_MEMOP__SHIFT 0x00000000 + //SQ_DEBUG_STS_GLOBAL #define SQ_DEBUG_STS_GLOBAL2__FIFO_LEVEL_GFX0_MASK 0x000000ffL #define SQ_DEBUG_STS_GLOBAL2__FIFO_LEVEL_GFX0__SHIFT 0x00000000 diff --git a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_11_0_0_sh_mask.h b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_11_0_0_sh_mask.h index 4f08f90856fc..3088a4a13cb5 100644 --- a/drivers/gpu/drm/amd/include/asic_reg/gc/gc_11_0_0_sh_mask.h +++ b/drivers/gpu/drm/amd/include/asic_reg/gc/gc_11_0_0_sh_mask.h @@ -17216,11 +17216,15 @@ #define SPI_GDBG_PER_VMID_CNTL__TRAP_EN__SHIFT 0x3 #define SPI_GDBG_PER_VMID_CNTL__EXCP_EN__SHIFT 0x4 #define SPI_GDBG_PER_VMID_CNTL__EXCP_REPLACE__SHIFT 0xd +#define SPI_GDBG_PER_VMID_CNTL__TRAP_ON_START__SHIFT 0xe +#define SPI_GDBG_PER_VMID_CNTL__TRAP_ON_END__SHIFT 0xf #define SPI_GDBG_PER_VMID_CNTL__STALL_VMID_MASK 0x00000001L #define SPI_GDBG_PER_VMID_CNTL__LAUNCH_MODE_MASK 0x00000006L #define SPI_GDBG_PER_VMID_CNTL__TRAP_EN_MASK 0x00000008L #define SPI_GDBG_PER_VMID_CNTL__EXCP_EN_MASK 0x00001FF0L #define SPI_GDBG_PER_VMID_CNTL__EXCP_REPLACE_MASK 0x00002000L +#define SPI_GDBG_PER_VMID_CNTL__TRAP_ON_START_MASK 0x00004000L +#define SPI_GDBG_PER_VMID_CNTL__TRAP_ON_END_MASK 0x00008000L //SPI_COMPUTE_QUEUE_RESET #define SPI_COMPUTE_QUEUE_RESET__RESET__SHIFT 0x0 #define SPI_COMPUTE_QUEUE_RESET__RESET_MASK 0x01L -- cgit v1.2.3 From d3116d9f27b89d363dd528e42fcf4895a15e0c3c Mon Sep 17 00:00:00 2001 From: Yang Li Date: Wed, 31 May 2023 10:08:11 +0800 Subject: drm/amdkfd: clean up one inconsistent indenting drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_device.c:1036 kgd2kfd_interrupt() warn: inconsistent indenting Signed-off-by: Yang Li Signed-off-by: Felix Kuehling Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index e84ad1c5ef44..f0ed6e6416c3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -1042,7 +1042,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry) is_patched ? patched_ihre : ih_ring_entry)) { kfd_queue_work(node->ih_wq, &node->interrupt_work); spin_unlock_irqrestore(&node->interrupt_lock, flags); - return; + return; } spin_unlock_irqrestore(&node->interrupt_lock, flags); } -- cgit v1.2.3 From bb13d763f251c28b08d996671c5146a2113fc9e7 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 24 Mar 2023 16:19:27 -0400 Subject: drm/amdkfd: fix kfd_suspend_all_processes Flush delayed restore work in kfd_suspend_all_queues instead of cancelling. Cancelling the work before it runs results in the queues becoming permanently disabled. Flushing the work ensures that the queue suspend/resume state stays balanced. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index ef67f5e37c6f..d75dac92775c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -2014,7 +2014,7 @@ void kfd_suspend_all_processes(void) WARN(debug_evictions, "Evicting all processes"); hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { cancel_delayed_work_sync(&p->eviction_work); - cancel_delayed_work_sync(&p->restore_work); + flush_delayed_work(&p->restore_work); if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND)) pr_err("Failed to suspend process 0x%x\n", p->pasid); -- cgit v1.2.3 From 7cee6a6824a0429a6255abe91b5af01b9a01cd03 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 23 Mar 2023 17:17:20 -0400 Subject: drm/amdgpu: add configurable grace period for unmap queues The HWS schedule allows a grace period for wave completion prior to preemption for better performance by avoiding CWSR on waves that can potentially complete quickly. The debugger, on the other hand, will want to inspect wave status immediately after it actively triggers preemption (a suspend function to be provided). To minimize latency between preemption and debugger wave inspection, allow immediate preemption by setting the grace period to 0. Note that setting the preepmtion grace period to 0 will result in an infinite grace period being set due to a CP FW bug so set it to 1 for now. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c | 2 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 43 ++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h | 6 ++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 43 ++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 8 ++- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 63 +++++++++++++++------ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 3 + drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c | 32 +++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 39 +++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h | 65 ++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 ++ 14 files changed, 295 insertions(+), 20 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c index a6f98141c29c..b811a0985050 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c @@ -82,5 +82,7 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = { .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .enable_debug_trap = kgd_aldebaran_enable_debug_trap, .disable_debug_trap = kgd_aldebaran_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c index d2918e5c0dea..a62bd0068515 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c @@ -410,6 +410,8 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = { kgd_gfx_v9_set_vm_context_page_table_base, .enable_debug_trap = kgd_arcturus_enable_debug_trap, .disable_debug_trap = kgd_arcturus_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index 240f5006e278..98006c7021dd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -803,6 +803,47 @@ uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev, return 0; } +/* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values + * The values read are: + * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. + * atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads. + * wrm_offload_wait_time -- Wait Count for WAIT_REG_MEM Offloads. + * gws_wait_time -- Wait Count for Global Wave Syncs. + * que_sleep_wait_time -- Wait Count for Dequeue Retry. + * sch_wave_wait_time -- Wait Count for Scheduling Wave Message. + * sem_rearm_wait_time -- Wait Count for Semaphore re-arm. + * deq_retry_wait_time -- Wait Count for Global Wave Syncs. + */ +void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, + uint32_t *wait_times) + +{ + *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2)); +} + +void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data) +{ + *reg_data = wait_times; + + /* + * The CP cannont handle a 0 grace period input and will result in + * an infinite grace period being set so set to 1 to prevent this. + */ + if (grace_period == 0) + grace_period = 1; + + *reg_data = REG_SET_FIELD(*reg_data, + CP_IQ_WAIT_TIME2, + SCH_WAVE, + grace_period); + + *reg_offset = SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2); +} + static void program_trap_handler_settings(struct amdgpu_device *adev, uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, uint32_t inst) @@ -848,5 +889,7 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = { .set_vm_context_page_table_base = set_vm_context_page_table_base, .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .program_trap_handler_settings = program_trap_handler_settings, }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h index 251d61fbde07..1e993a213646 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h @@ -26,3 +26,9 @@ uint32_t kgd_gfx_v10_enable_debug_trap(struct amdgpu_device *adev, uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev, bool keep_trap_enabled, uint32_t vmid); +void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); +void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c index 8b293f3dcbd2..387bdf4823c9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c @@ -672,6 +672,8 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = { .get_atc_vmid_pasid_mapping_info = get_atc_vmid_pasid_mapping_info_v10_3, .set_vm_context_page_table_base = set_vm_context_page_table_base_v10_3, .program_trap_handler_settings = program_trap_handler_settings_v10_3, + .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, .disable_debug_trap = kgd_gfx_v10_disable_debug_trap }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index 8d7d04704b00..829ee720cc44 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -739,6 +739,24 @@ uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, return 0; } +/* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values + * The values read are: + * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. + * atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads. + * wrm_offload_wait_time -- Wait Count for WAIT_REG_MEM Offloads. + * gws_wait_time -- Wait Count for Global Wave Syncs. + * que_sleep_wait_time -- Wait Count for Dequeue Retry. + * sch_wave_wait_time -- Wait Count for Scheduling Wave Message. + * sem_rearm_wait_time -- Wait Count for Semaphore re-arm. + * deq_retry_wait_time -- Wait Count for Global Wave Syncs. + */ +void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, + uint32_t *wait_times) + +{ + *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2)); +} + void kgd_gfx_v9_set_vm_context_page_table_base(struct amdgpu_device *adev, uint32_t vmid, uint64_t page_table_base) { @@ -926,6 +944,29 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, adev->gfx.cu_info.max_waves_per_simd; } +void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data) +{ + *reg_data = wait_times; + + /* + * The CP cannont handle a 0 grace period input and will result in + * an infinite grace period being set so set to 1 to prevent this. + */ + if (grace_period == 0) + grace_period = 1; + + *reg_data = REG_SET_FIELD(*reg_data, + CP_IQ_WAIT_TIME2, + SCH_WAVE, + grace_period); + + *reg_offset = SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2); +} + void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev, uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, uint32_t inst) { @@ -969,6 +1010,8 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = { .set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base, .enable_debug_trap = kgd_gfx_v9_enable_debug_trap, .disable_debug_trap = kgd_gfx_v9_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index 9588ff055393..fed5b7f18b1a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -20,8 +20,6 @@ * OTHER DEALINGS IN THE SOFTWARE. */ - - void kgd_gfx_v9_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid, uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, @@ -73,3 +71,9 @@ uint32_t kgd_gfx_v9_enable_debug_trap(struct amdgpu_device *adev, uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, bool keep_trap_enabled, uint32_t vmid); +void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); +void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 2baa0781eafc..0b88a64e61fe 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -46,10 +46,13 @@ static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, static int execute_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param); + uint32_t filter_param, + uint32_t grace_period); static int unmap_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param, bool reset); + uint32_t filter_param, + uint32_t grace_period, + bool reset); static int map_queues_cpsch(struct device_queue_manager *dqm); @@ -866,7 +869,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { if (!dqm->dev->kfd->shared_resources.enable_mes) retval = unmap_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); else if (prev_active) retval = remove_queue_mes(dqm, q, &pdd->qpd); @@ -1042,7 +1045,8 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, retval = execute_queues_cpsch(dqm, qpd->is_debug ? KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); out: dqm_unlock(dqm); @@ -1182,8 +1186,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, } if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); - + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; atomic64_add(eviction_duration, &pdd->evict_duration_counter); vm_not_acquired: @@ -1525,6 +1528,9 @@ static int initialize_cpsch(struct device_queue_manager *dqm) init_sdma_bitmaps(dqm); + if (dqm->dev->kfd2kgd->get_iq_wait_times) + dqm->dev->kfd2kgd->get_iq_wait_times(dqm->dev->adev, + &dqm->wait_times); return 0; } @@ -1563,8 +1569,9 @@ static int start_cpsch(struct device_queue_manager *dqm) dqm->is_hws_hang = false; dqm->is_resetting = false; dqm->sched_running = true; + if (!dqm->dev->kfd->shared_resources.enable_mes) - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); dqm_unlock(dqm); return 0; @@ -1589,7 +1596,7 @@ static int stop_cpsch(struct device_queue_manager *dqm) if (!dqm->is_hws_hang) { if (!dqm->dev->kfd->shared_resources.enable_mes) - unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, false); + unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); else remove_all_queues_mes(dqm); } @@ -1631,7 +1638,8 @@ static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, list_add(&kq->list, &qpd->priv_queue_list); increment_queue_count(dqm, qpd, kq->queue); qpd->is_debug = true; - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); dqm_unlock(dqm); return 0; @@ -1645,7 +1653,8 @@ static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, list_del(&kq->list); decrement_queue_count(dqm, qpd, kq->queue); qpd->is_debug = false; - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); /* * Unconditionally decrement this counter, regardless of the queue's * type. @@ -1722,7 +1731,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); else retval = add_queue_mes(dqm, q, qpd); if (retval) @@ -1811,7 +1820,9 @@ static int map_queues_cpsch(struct device_queue_manager *dqm) /* dqm->lock mutex has to be locked before calling this function */ static int unmap_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param, bool reset) + uint32_t filter_param, + uint32_t grace_period, + bool reset) { int retval = 0; struct mqd_manager *mqd_mgr; @@ -1823,6 +1834,12 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, if (!dqm->active_runlist) return retval; + if (grace_period != USE_DEFAULT_GRACE_PERIOD) { + retval = pm_update_grace_period(&dqm->packet_mgr, grace_period); + if (retval) + return retval; + } + retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset); if (retval) return retval; @@ -1855,6 +1872,13 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, return -ETIME; } + /* We need to reset the grace period value for this device */ + if (grace_period != USE_DEFAULT_GRACE_PERIOD) { + if (pm_update_grace_period(&dqm->packet_mgr, + USE_DEFAULT_GRACE_PERIOD)) + pr_err("Failed to reset grace period\n"); + } + pm_release_ib(&dqm->packet_mgr); dqm->active_runlist = false; @@ -1870,7 +1894,7 @@ static int reset_queues_cpsch(struct device_queue_manager *dqm, dqm_lock(dqm); retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID, - pasid, true); + pasid, USE_DEFAULT_GRACE_PERIOD, true); dqm_unlock(dqm); return retval; @@ -1879,13 +1903,14 @@ static int reset_queues_cpsch(struct device_queue_manager *dqm, /* dqm->lock mutex has to be locked before calling this function */ static int execute_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param) + uint32_t filter_param, + uint32_t grace_period) { int retval; if (dqm->is_hws_hang) return -EIO; - retval = unmap_queues_cpsch(dqm, filter, filter_param, false); + retval = unmap_queues_cpsch(dqm, filter, filter_param, grace_period, false); if (retval) return retval; @@ -1943,7 +1968,8 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, if (!dqm->dev->kfd->shared_resources.enable_mes) { decrement_queue_count(dqm, qpd, q); retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); if (retval == -ETIME) qpd->reset_wavefronts = true; } else { @@ -2228,7 +2254,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, } if (!dqm->dev->kfd->shared_resources.enable_mes) - retval = execute_queues_cpsch(dqm, filter, 0); + retval = execute_queues_cpsch(dqm, filter, 0, USE_DEFAULT_GRACE_PERIOD); if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); @@ -2589,7 +2615,8 @@ int dqm_debugfs_hang_hws(struct device_queue_manager *dqm) return r; } dqm->active_runlist = true; - r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); + r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, + 0, USE_DEFAULT_GRACE_PERIOD); dqm_unlock(dqm); return r; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index cd4383bb207f..d4dd3b4acbf0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -37,6 +37,7 @@ #define KFD_MES_PROCESS_QUANTUM 100000 #define KFD_MES_GANG_QUANTUM 10000 +#define USE_DEFAULT_GRACE_PERIOD 0xffffffff struct device_process_node { struct qcm_process_device *qpd; @@ -259,6 +260,8 @@ struct device_queue_manager { /* used for GFX 9.4.3 only */ uint32_t current_logical_xcc_start; + + uint32_t wait_times; }; void device_queue_manager_init_cik( diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c index 2f54172e9175..401096c103b2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c @@ -370,6 +370,38 @@ out: return retval; } +int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period) +{ + int retval = 0; + uint32_t *buffer, size; + + size = pm->pmf->set_grace_period_size; + + mutex_lock(&pm->lock); + + if (size) { + kq_acquire_packet_buffer(pm->priv_queue, + size / sizeof(uint32_t), + (unsigned int **)&buffer); + + if (!buffer) { + pr_err("Failed to allocate buffer on kernel queue\n"); + retval = -ENOMEM; + goto out; + } + + retval = pm->pmf->set_grace_period(pm, buffer, grace_period); + if (!retval) + kq_submit_packet(pm->priv_queue); + else + kq_rollback_packet(pm->priv_queue); + } + +out: + mutex_unlock(&pm->lock); + return retval; +} + int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_unmap_queues_filter filter, uint32_t filter_param, bool reset) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 44cf3a5f6fdb..1fda6dcf84b1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -262,6 +262,41 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, return 0; } +static int pm_set_grace_period_v9(struct packet_manager *pm, + uint32_t *buffer, + uint32_t grace_period) +{ + struct pm4_mec_write_data_mmio *packet; + uint32_t reg_offset = 0; + uint32_t reg_data = 0; + + pm->dqm->dev->kfd2kgd->build_grace_period_packet_info( + pm->dqm->dev->adev, + pm->dqm->wait_times, + grace_period, + ®_offset, + ®_data); + + if (grace_period == USE_DEFAULT_GRACE_PERIOD) + reg_data = pm->dqm->wait_times; + + packet = (struct pm4_mec_write_data_mmio *)buffer; + memset(buffer, 0, sizeof(struct pm4_mec_write_data_mmio)); + + packet->header.u32All = pm_build_pm4_header(IT_WRITE_DATA, + sizeof(struct pm4_mec_write_data_mmio)); + + packet->bitfields2.dst_sel = dst_sel___write_data__mem_mapped_register; + packet->bitfields2.addr_incr = + addr_incr___write_data__do_not_increment_address; + + packet->bitfields3.dst_mmreg_addr = reg_offset; + + packet->data = reg_data; + + return 0; +} + static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer, enum kfd_unmap_queues_filter filter, uint32_t filter_param, bool reset) @@ -345,6 +380,7 @@ const struct packet_manager_funcs kfd_v9_pm_funcs = { .set_resources = pm_set_resources_v9, .map_queues = pm_map_queues_v9, .unmap_queues = pm_unmap_queues_v9, + .set_grace_period = pm_set_grace_period_v9, .query_status = pm_query_status_v9, .release_mem = NULL, .map_process_size = sizeof(struct pm4_mes_map_process), @@ -352,6 +388,7 @@ const struct packet_manager_funcs kfd_v9_pm_funcs = { .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), + .set_grace_period_size = sizeof(struct pm4_mec_write_data_mmio), .query_status_size = sizeof(struct pm4_mes_query_status), .release_mem_size = 0, }; @@ -362,6 +399,7 @@ const struct packet_manager_funcs kfd_aldebaran_pm_funcs = { .set_resources = pm_set_resources_v9, .map_queues = pm_map_queues_v9, .unmap_queues = pm_unmap_queues_v9, + .set_grace_period = pm_set_grace_period_v9, .query_status = pm_query_status_v9, .release_mem = NULL, .map_process_size = sizeof(struct pm4_mes_map_process_aldebaran), @@ -369,6 +407,7 @@ const struct packet_manager_funcs kfd_aldebaran_pm_funcs = { .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), + .set_grace_period_size = sizeof(struct pm4_mec_write_data_mmio), .query_status_size = sizeof(struct pm4_mes_query_status), .release_mem_size = 0, }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c index faf4772ed317..c1199d06d131 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c @@ -303,6 +303,7 @@ const struct packet_manager_funcs kfd_vi_pm_funcs = { .set_resources = pm_set_resources_vi, .map_queues = pm_map_queues_vi, .unmap_queues = pm_unmap_queues_vi, + .set_grace_period = NULL, .query_status = pm_query_status_vi, .release_mem = pm_release_mem_vi, .map_process_size = sizeof(struct pm4_mes_map_process), @@ -310,6 +311,7 @@ const struct packet_manager_funcs kfd_vi_pm_funcs = { .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), + .set_grace_period_size = 0, .query_status_size = sizeof(struct pm4_mes_query_status), .release_mem_size = sizeof(struct pm4_mec_release_mem) }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index 2ad708c64012..206f1960857f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -584,6 +584,71 @@ struct pm4_mec_release_mem { #endif +#ifndef PM4_MEC_WRITE_DATA_DEFINED +#define PM4_MEC_WRITE_DATA_DEFINED + +enum WRITE_DATA_dst_sel_enum { + dst_sel___write_data__mem_mapped_register = 0, + dst_sel___write_data__tc_l2 = 2, + dst_sel___write_data__gds = 3, + dst_sel___write_data__memory = 5, + dst_sel___write_data__memory_mapped_adc_persistent_state = 6, +}; + +enum WRITE_DATA_addr_incr_enum { + addr_incr___write_data__increment_address = 0, + addr_incr___write_data__do_not_increment_address = 1 +}; + +enum WRITE_DATA_wr_confirm_enum { + wr_confirm___write_data__do_not_wait_for_write_confirmation = 0, + wr_confirm___write_data__wait_for_write_confirmation = 1 +}; + +enum WRITE_DATA_cache_policy_enum { + cache_policy___write_data__lru = 0, + cache_policy___write_data__stream = 1 +}; + + +struct pm4_mec_write_data_mmio { + union { + union PM4_MES_TYPE_3_HEADER header; /*header */ + unsigned int ordinal1; + }; + + union { + struct { + unsigned int reserved1:8; + unsigned int dst_sel:4; + unsigned int reserved2:4; + unsigned int addr_incr:1; + unsigned int reserved3:2; + unsigned int resume_vf:1; + unsigned int wr_confirm:1; + unsigned int reserved4:4; + unsigned int cache_policy:2; + unsigned int reserved5:5; + } bitfields2; + unsigned int ordinal2; + }; + + union { + struct { + unsigned int dst_mmreg_addr:18; + unsigned int reserved6:14; + } bitfields3; + unsigned int ordinal3; + }; + + uint32_t reserved7; + + uint32_t data; + +}; + +#endif + enum { CACHE_FLUSH_AND_INV_TS_EVENT = 0x00000014 }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 1b272f879b4c..4c912b7735b5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1350,6 +1350,8 @@ struct packet_manager_funcs { int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, enum kfd_unmap_queues_filter mode, uint32_t filter_param, bool reset); + int (*set_grace_period)(struct packet_manager *pm, uint32_t *buffer, + uint32_t grace_period); int (*query_status)(struct packet_manager *pm, uint32_t *buffer, uint64_t fence_address, uint64_t fence_value); int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); @@ -1360,6 +1362,7 @@ struct packet_manager_funcs { int set_resources_size; int map_queues_size; int unmap_queues_size; + int set_grace_period_size; int query_status_size; int release_mem_size; }; @@ -1382,6 +1385,8 @@ int pm_send_unmap_queue(struct packet_manager *pm, void pm_release_ib(struct packet_manager *pm); +int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period); + /* Following PM funcs can be shared among VI and AI */ unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); -- cgit v1.2.3 From 97ae3c8cce96f3bebf883d0812cef5d3fdbe3e64 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 4 Apr 2022 12:27:43 -0400 Subject: drm/amdkfd: prepare map process for single process debug devices Older HW only supports debugging on a single process because the SPI debug mode setting registers are device global. The HWS has supplied a single pinned VMID (0xf) for MAP_PROCESS for debug purposes. To pin the VMID, the KFD will remove the VMID from the HWS dynamic VMID allocation via SET_RESOUCES so that a debugged process will never migrate away from its pinned VMID. The KFD is responsible for reserving and releasing this pinned VMID accordingly whenever the debugger attaches and detaches respectively. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 93 ++++++++++++++++++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 9 +++ drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h | 5 +- 4 files changed, 111 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 0b88a64e61fe..cfe5bd59070e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1525,6 +1525,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm) dqm->gws_queue_count = 0; dqm->active_runlist = false; INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception); + dqm->trap_debug_vmid = 0; init_sdma_bitmaps(dqm); @@ -2501,6 +2502,98 @@ static void kfd_process_hw_exception(struct work_struct *work) amdgpu_amdkfd_gpu_reset(dqm->dev->adev); } +int reserve_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd) +{ + int r; + int updated_vmid_mask; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + dqm_lock(dqm); + + if (dqm->trap_debug_vmid != 0) { + pr_err("Trap debug id already reserved\n"); + r = -EBUSY; + goto out_unlock; + } + + r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD, false); + if (r) + goto out_unlock; + + updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; + updated_vmid_mask &= ~(1 << dqm->dev->vm_info.last_vmid_kfd); + + dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask; + dqm->trap_debug_vmid = dqm->dev->vm_info.last_vmid_kfd; + r = set_sched_resources(dqm); + if (r) + goto out_unlock; + + r = map_queues_cpsch(dqm); + if (r) + goto out_unlock; + + pr_debug("Reserved VMID for trap debug: %i\n", dqm->trap_debug_vmid); + +out_unlock: + dqm_unlock(dqm); + return r; +} + +/* + * Releases vmid for the trap debugger + */ +int release_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd) +{ + int r; + int updated_vmid_mask; + uint32_t trap_debug_vmid; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + dqm_lock(dqm); + trap_debug_vmid = dqm->trap_debug_vmid; + if (dqm->trap_debug_vmid == 0) { + pr_err("Trap debug id is not reserved\n"); + r = -EINVAL; + goto out_unlock; + } + + r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD, false); + if (r) + goto out_unlock; + + updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; + updated_vmid_mask |= (1 << dqm->dev->vm_info.last_vmid_kfd); + + dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask; + dqm->trap_debug_vmid = 0; + r = set_sched_resources(dqm); + if (r) + goto out_unlock; + + r = map_queues_cpsch(dqm); + if (r) + goto out_unlock; + + pr_debug("Released VMID for trap debug: %i\n", trap_debug_vmid); + +out_unlock: + dqm_unlock(dqm); + return r; +} + #if defined(CONFIG_DEBUG_FS) static void seq_reg_dump(struct seq_file *m, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index d4dd3b4acbf0..bf7aa3f84182 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -250,6 +250,7 @@ struct device_queue_manager { struct kfd_mem_obj *fence_mem; bool active_runlist; int sched_policy; + uint32_t trap_debug_vmid; /* hw exception */ bool is_hws_hang; @@ -285,6 +286,10 @@ unsigned int get_queues_per_pipe(struct device_queue_manager *dqm); unsigned int get_pipes_per_mec(struct device_queue_manager *dqm); unsigned int get_num_sdma_queues(struct device_queue_manager *dqm); unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm); +int reserve_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd); +int release_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd); static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 1fda6dcf84b1..0fe73dbd28af 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -34,6 +34,9 @@ static int pm_map_process_v9(struct packet_manager *pm, { struct pm4_mes_map_process *packet; uint64_t vm_page_table_base_addr = qpd->page_table_base; + struct kfd_node *kfd = pm->dqm->dev; + struct kfd_process_device *pdd = + container_of(qpd, struct kfd_process_device, qpd); packet = (struct pm4_mes_map_process *)buffer; memset(buffer, 0, sizeof(struct pm4_mes_map_process)); @@ -49,6 +52,12 @@ static int pm_map_process_v9(struct packet_manager *pm, packet->bitfields14.sdma_enable = 1; packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; + if (kfd->dqm->trap_debug_vmid && pdd->process->debug_trap_enabled && + pdd->process->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) { + packet->bitfields2.debug_vmid = kfd->dqm->trap_debug_vmid; + packet->bitfields2.new_debug = 1; + } + packet->sh_mem_config = qpd->sh_mem_config; packet->sh_mem_bases = qpd->sh_mem_bases; if (qpd->tba_addr) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index 206f1960857f..8b6b2bd5c148 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -146,7 +146,10 @@ struct pm4_mes_map_process { union { struct { uint32_t pasid:16; - uint32_t reserved1:8; + uint32_t reserved1:2; + uint32_t debug_vmid:4; + uint32_t new_debug:1; + uint32_t reserved2:1; uint32_t diq_enable:1; uint32_t process_quantum:7; } bitfields2; -- cgit v1.2.3 From 0de4ec9a03537bd2b189b5afbf83acd6b72b0258 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 4 Apr 2022 13:38:11 -0400 Subject: drm/amdgpu: prepare map process for multi-process debug devices Unlike single process debug devices, multi-process debug devices allow debug mode setting per-VMID (non-device-global). Because the HWS manages PASID-VMID mapping, the new MAP_PROCESS API allows the KFD to forward the required SPI debug register write requests. To request a new debug mode setting change, the KFD must be able to preempt all queues then remap all queues with these new setting requests for MAP_PROCESS to take effect. Note that by default, trap enablement in non-debug mode must be disabled for performance reasons for multi-process debug devices due to setup overhead in FW. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 5 +++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 51 ++++++++++++++++++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 3 ++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 14 ++++++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 9 ++++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 5 +++ 6 files changed, 87 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index a8abfe2a0a14..db6d72e7930f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -29,4 +29,9 @@ int kfd_dbg_trap_disable(struct kfd_process *target); int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, void __user *runtime_info, uint32_t *runtime_info_size); +static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) +{ + return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2); +} + #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index cfe5bd59070e..495c9238254e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -36,6 +36,7 @@ #include "kfd_kernel_queue.h" #include "amdgpu_amdkfd.h" #include "mes_api_def.h" +#include "kfd_debug.h" /* Size of the per-pipe EOP queue */ #define CIK_HPD_EOP_BYTES_LOG2 11 @@ -2594,6 +2595,56 @@ out_unlock: return r; } +int debug_lock_and_unmap(struct device_queue_manager *dqm) +{ + int r; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + if (!kfd_dbg_is_per_vmid_supported(dqm->dev)) + return 0; + + dqm_lock(dqm); + + r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 0, false); + if (r) + dqm_unlock(dqm); + + return r; +} + +int debug_map_and_unlock(struct device_queue_manager *dqm) +{ + int r; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + if (!kfd_dbg_is_per_vmid_supported(dqm->dev)) + return 0; + + r = map_queues_cpsch(dqm); + + dqm_unlock(dqm); + + return r; +} + +int debug_refresh_runlist(struct device_queue_manager *dqm) +{ + int r = debug_lock_and_unmap(dqm); + + if (r) + return r; + + return debug_map_and_unlock(dqm); +} + #if defined(CONFIG_DEBUG_FS) static void seq_reg_dump(struct seq_file *m, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index bf7aa3f84182..bb75d93712eb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -290,6 +290,9 @@ int reserve_debug_trap_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd); int release_debug_trap_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd); +int debug_lock_and_unmap(struct device_queue_manager *dqm); +int debug_map_and_unlock(struct device_queue_manager *dqm); +int debug_refresh_runlist(struct device_queue_manager *dqm); static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 0fe73dbd28af..29a2d0499b67 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -88,6 +88,10 @@ static int pm_map_process_aldebaran(struct packet_manager *pm, { struct pm4_mes_map_process_aldebaran *packet; uint64_t vm_page_table_base_addr = qpd->page_table_base; + struct kfd_dev *kfd = pm->dqm->dev->kfd; + struct kfd_process_device *pdd = + container_of(qpd, struct kfd_process_device, qpd); + int i; packet = (struct pm4_mes_map_process_aldebaran *)buffer; memset(buffer, 0, sizeof(struct pm4_mes_map_process_aldebaran)); @@ -102,6 +106,16 @@ static int pm_map_process_aldebaran(struct packet_manager *pm, packet->bitfields14.num_oac = qpd->num_oac; packet->bitfields14.sdma_enable = 1; packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; + packet->spi_gdbg_per_vmid_cntl = pdd->spi_dbg_override | + pdd->spi_dbg_launch_mode; + + if (pdd->process->debug_trap_enabled) { + for (i = 0; i < kfd->device_info.num_of_watch_points; i++) + packet->tcp_watch_cntl[i] = pdd->watch_points[i]; + + packet->bitfields2.single_memops = + !!(pdd->process->dbg_flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP); + } packet->sh_mem_config = qpd->sh_mem_config; packet->sh_mem_bases = qpd->sh_mem_bases; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 4c912b7735b5..8fca7175daab 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -816,6 +816,12 @@ struct kfd_process_device { uint64_t faults; uint64_t page_in; uint64_t page_out; + + /* Tracks debug per-vmid request settings */ + uint32_t spi_dbg_override; + uint32_t spi_dbg_launch_mode; + uint32_t watch_points[4]; + /* * If this process has been checkpointed before, then the user * application will use the original gpu_id on the @@ -952,6 +958,9 @@ struct kfd_process { bool xnack_enabled; + /* Tracks debug per-vmid request for debug flags */ + bool dbg_flags; + atomic_t poison; /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */ bool queues_paused; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index d75dac92775c..725d936b2cc7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1612,6 +1612,11 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev, } p->pdds[p->n_pdds++] = pdd; + if (kfd_dbg_is_per_vmid_supported(pdd->dev)) + pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap( + pdd->dev->adev, + false, + 0); /* Init idr used for memory handle translation */ idr_init(&pdd->alloc_idr); -- cgit v1.2.3 From 218895820e6fccade42a7c3ab9c0a44dec0a1ebc Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Tue, 5 Apr 2022 12:34:55 -0400 Subject: drm/amdkfd: add per process hw trap enable and disable functions To enable HW debug mode per process, all devices must be debug enabled successfully. If a failure occures, rewind the enablement of debug mode on the enabled devices. A power management scenario that needs to be considered is HW debug mode setting during GFXOFF. During GFXOFF, these registers will be unreachable so we have to transiently disable GFXOFF when setting. Also, some devices don't support the RLC save restore function for these debug registers so we have to disable GFXOFF completely during a debug session. Cooperative launch also has debugging restriction based on HW/FW bugs. If such bugs exists, the debugger cannot attach to a process that uses GWS resources nor can GWS resources be requested if a process is being debugged. Multi-process debug devices can only enable trap temporaries based on certain runtime scenerios, which will be explained when the runtime enable functions are implemented in a follow up patch. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 148 ++++++++++++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 29 ++++++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 10 +++ 4 files changed, 190 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index ee086a0a46df..826a99acb6fb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1488,6 +1488,11 @@ static int kfd_ioctl_alloc_queue_gws(struct file *filep, goto out_unlock; } + if (!kfd_dbg_has_gws_support(dev) && p->debug_trap_enabled) { + retval = -EBUSY; + goto out_unlock; + } + retval = pqm_set_gws(&p->pqm, args->queue_id, args->num_gws ? dev->gws : NULL); mutex_unlock(&p->mutex); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 898cc1fe3d13..73b07b5f17f1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -21,13 +21,78 @@ */ #include "kfd_debug.h" +#include "kfd_device_queue_manager.h" #include +static int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) +{ + uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode; + uint32_t flags = pdd->process->dbg_flags; + + if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) + return 0; + + return amdgpu_mes_set_shader_debugger(pdd->dev->adev, pdd->proc_ctx_gpu_addr, spi_dbg_cntl, + pdd->watch_points, flags); +} + +/* kfd_dbg_trap_deactivate: + * target: target process + * unwind: If this is unwinding a failed kfd_dbg_trap_enable() + * unwind_count: + * If unwind == true, how far down the pdd list we need + * to unwind + * else: ignored + */ +static void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count) +{ + int i; + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + /* If this is an unwind, and we have unwound the required + * enable calls on the pdd list, we need to stop now + * otherwise we may mess up another debugger session. + */ + if (unwind && i == unwind_count) + break; + + /* GFX off is already disabled by debug activate if not RLC restore supported. */ + if (kfd_dbg_is_rlc_restore_supported(pdd->dev)) + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + pdd->spi_dbg_override = + pdd->dev->kfd2kgd->disable_debug_trap( + pdd->dev->adev, + target->runtime_info.ttmp_setup, + pdd->dev->vm_info.last_vmid_kfd); + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + + if (!kfd_dbg_is_per_vmid_supported(pdd->dev) && + release_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd)) + pr_err("Failed to release debug vmid on [%i]\n", pdd->dev->id); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + debug_refresh_runlist(pdd->dev->dqm); + else + kfd_dbg_set_mes_debug_mode(pdd); + } +} + int kfd_dbg_trap_disable(struct kfd_process *target) { if (!target->debug_trap_enabled) return 0; + /* + * Defer deactivation to runtime if runtime not enabled otherwise reset + * attached running target runtime state to enable for re-attach. + */ + if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) + kfd_dbg_trap_deactivate(target, false, 0); + else if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED) + target->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED; + fput(target->dbg_ev_file); target->dbg_ev_file = NULL; @@ -42,16 +107,89 @@ int kfd_dbg_trap_disable(struct kfd_process *target) return 0; } +static int kfd_dbg_trap_activate(struct kfd_process *target) +{ + int i, r = 0; + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) { + r = reserve_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd); + + if (r) { + target->runtime_info.runtime_state = (r == -EBUSY) ? + DEBUG_RUNTIME_STATE_ENABLED_BUSY : + DEBUG_RUNTIME_STATE_ENABLED_ERROR; + + goto unwind_err; + } + } + + /* Disable GFX OFF to prevent garbage read/writes to debug registers. + * If RLC restore of debug registers is not supported and runtime enable + * hasn't done so already on ttmp setup request, restore the trap config registers. + * + * If RLC restore of debug registers is not supported, keep gfx off disabled for + * the debug session. + */ + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + if (!(kfd_dbg_is_rlc_restore_supported(pdd->dev) || + target->runtime_info.ttmp_setup)) + pdd->dev->kfd2kgd->enable_debug_trap(pdd->dev->adev, true, + pdd->dev->vm_info.last_vmid_kfd); + + pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap( + pdd->dev->adev, + false, + pdd->dev->vm_info.last_vmid_kfd); + + if (kfd_dbg_is_rlc_restore_supported(pdd->dev)) + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + r = debug_refresh_runlist(pdd->dev->dqm); + else + r = kfd_dbg_set_mes_debug_mode(pdd); + + if (r) { + target->runtime_info.runtime_state = + DEBUG_RUNTIME_STATE_ENABLED_ERROR; + goto unwind_err; + } + } + + return 0; + +unwind_err: + /* Enabling debug failed, we need to disable on + * all GPUs so the enable is all or nothing. + */ + kfd_dbg_trap_deactivate(target, true, i); + return r; +} + int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, void __user *runtime_info, uint32_t *runtime_size) { struct file *f; uint32_t copy_size; - int r = 0; + int i, r = 0; if (target->debug_trap_enabled) return -EALREADY; + /* Enable pre-checks */ + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + if (!KFD_IS_SOC15(pdd->dev)) + return -ENODEV; + + if (!kfd_dbg_has_gws_support(pdd->dev) && pdd->qpd.num_gws) + return -EBUSY; + } + copy_size = min((size_t)(*runtime_size), sizeof(target->runtime_info)); f = fget(fd); @@ -62,6 +200,10 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, target->dbg_ev_file = f; + /* defer activation to runtime if not runtime enabled */ + if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) + kfd_dbg_trap_activate(target); + /* We already hold the process reference but hold another one for the * debug session. */ @@ -71,8 +213,10 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, if (target->debugger_process) atomic_inc(&target->debugger_process->debugged_process_count); - if (copy_to_user(runtime_info, (void *)&target->runtime_info, copy_size)) + if (copy_to_user(runtime_info, (void *)&target->runtime_info, copy_size)) { + kfd_dbg_trap_deactivate(target, false, 0); r = -EFAULT; + } *runtime_size = sizeof(target->runtime_info); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index db6d72e7930f..17481f824647 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -34,4 +34,33 @@ static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2); } +/* + * If GFX off is enabled, chips that do not support RLC restore for the debug + * registers will disable GFX off temporarily for the entire debug session. + * See disable_on_trap_action_entry and enable_on_trap_action_exit for details. + */ +static inline bool kfd_dbg_is_rlc_restore_supported(struct kfd_node *dev) +{ + return !(KFD_GC_VERSION(dev) == IP_VERSION(10, 1, 10) || + KFD_GC_VERSION(dev) == IP_VERSION(10, 1, 1)); +} + +static inline bool kfd_dbg_has_gws_support(struct kfd_node *dev) +{ + if ((KFD_GC_VERSION(dev) == IP_VERSION(9, 0, 1) + && dev->kfd->mec2_fw_version < 0x81b6) || + (KFD_GC_VERSION(dev) >= IP_VERSION(9, 1, 0) + && KFD_GC_VERSION(dev) <= IP_VERSION(9, 2, 2) + && dev->kfd->mec2_fw_version < 0x1b6) || + (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0) + && dev->kfd->mec2_fw_version < 0x1b6) || + (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) + && dev->kfd->mec2_fw_version < 0x30) || + (KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0) && + KFD_GC_VERSION(dev) < IP_VERSION(12, 0, 0))) + return false; + + /* Assume debugging and cooperative launch supported otherwise. */ + return true; +} #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 725d936b2cc7..e77cadadb09b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1165,9 +1165,19 @@ static void kfd_process_free_notifier(struct mmu_notifier *mn) static void kfd_process_notifier_release_internal(struct kfd_process *p) { + int i; + cancel_delayed_work_sync(&p->eviction_work); cancel_delayed_work_sync(&p->restore_work); + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + + /* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */ + if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup) + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + } + /* Indicate to other users that MM is no longer valid */ p->mm = NULL; kfd_dbg_trap_disable(p); -- cgit v1.2.3 From 69a8c3ae2dea84a6d571e4c1aad306f630f3ccfd Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 1 Sep 2022 11:27:15 -0400 Subject: drm/amdkfd: apply trap workaround for gfx11 Due to a HW bug, waves in only half the shader arrays can enter trap. When starting a debug session, relocate all waves to the first shader array of each shader engine and mask off the 2nd shader array as unavailable. When ending a debug session, re-enable the 2nd shader array per shader engine. User CU masking per queue cannot be guaranteed to remain functional if requested during debugging (e.g. user cu mask requests only 2nd shader array as an available resource leading to zero HW resources available) nor can runtime be alerted of any of these changes during execution. Make user CU masking and debugging mutual exclusive with respect to availability. If the debugger tries to attach to a process with a user cu masked queue, return the runtime status as enabled but busy. If the debugger tries to attach and fails to reallocate queue waves to the first shader array of each shader engine, return the runtime status as enabled but with an error. In addition, like any other mutli-process debug supported devices, disable trap temporary setup per-process to avoid performance impact from setup overhead. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 2 + drivers/gpu/drm/amd/amdgpu/mes_v11_0.c | 7 +-- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 - drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 57 ++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 3 +- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 7 +++ drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 42 +++++++++++----- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 +- .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 9 +++- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 7 ++- 14 files changed, 122 insertions(+), 31 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h index d20df0cf0d88..b5f5eed2b5ef 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h @@ -219,6 +219,8 @@ struct mes_add_queue_input { uint32_t gws_size; uint64_t tba_addr; uint64_t tma_addr; + uint32_t trap_en; + uint32_t skip_process_ctx_clear; uint32_t is_kfd_process; uint32_t is_aql_queue; uint32_t queue_size; diff --git a/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c b/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c index 861910a6662d..c4e3cb8d44de 100644 --- a/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +++ b/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c @@ -202,17 +202,14 @@ static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes, mes_add_queue_pkt.gws_size = input->gws_size; mes_add_queue_pkt.trap_handler_addr = input->tba_addr; mes_add_queue_pkt.tma_addr = input->tma_addr; + mes_add_queue_pkt.trap_en = input->trap_en; + mes_add_queue_pkt.skip_process_ctx_clear = input->skip_process_ctx_clear; mes_add_queue_pkt.is_kfd_process = input->is_kfd_process; /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */ mes_add_queue_pkt.is_aql_queue = input->is_aql_queue; mes_add_queue_pkt.gds_size = input->queue_size; - if (!(((adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 4) && - (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0)) && - (adev->ip_versions[GC_HWIP][0] <= IP_VERSION(11, 0, 3)))) - mes_add_queue_pkt.trap_en = 1; - /* For KFD, gds_size is re-used for queue size (needed in MES for AQL queues) */ mes_add_queue_pkt.is_aql_queue = input->is_aql_queue; mes_add_queue_pkt.gds_size = input->queue_size; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 826a99acb6fb..d4df424e4514 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -537,8 +537,6 @@ static int kfd_ioctl_set_cu_mask(struct file *filp, struct kfd_process *p, goto out; } - minfo.update_flag = UPDATE_FLAG_CU_MASK; - mutex_lock(&p->mutex); retval = pqm_update_mqd(&p->pqm, args->queue_id, &minfo); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 73b07b5f17f1..5e2ee2d1acc4 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -24,6 +24,57 @@ #include "kfd_device_queue_manager.h" #include +static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable) +{ + struct mqd_update_info minfo = {0}; + int err; + + if (!q) + return 0; + + if (KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) || + KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0)) + return 0; + + if (enable && q->properties.is_user_cu_masked) + return -EBUSY; + + minfo.update_flag = enable ? UPDATE_FLAG_DBG_WA_ENABLE : UPDATE_FLAG_DBG_WA_DISABLE; + + q->properties.is_dbg_wa = enable; + err = q->device->dqm->ops.update_queue(q->device->dqm, q, &minfo); + if (err) + q->properties.is_dbg_wa = false; + + return err; +} + +static int kfd_dbg_set_workaround(struct kfd_process *target, bool enable) +{ + struct process_queue_manager *pqm = &target->pqm; + struct process_queue_node *pqn; + int r = 0; + + list_for_each_entry(pqn, &pqm->queues, process_queue_list) { + r = kfd_dbg_set_queue_workaround(pqn->q, enable); + if (enable && r) + goto unwind; + } + + return 0; + +unwind: + list_for_each_entry(pqn, &pqm->queues, process_queue_list) + kfd_dbg_set_queue_workaround(pqn->q, false); + + if (enable) + target->runtime_info.runtime_state = r == -EBUSY ? + DEBUG_RUNTIME_STATE_ENABLED_BUSY : + DEBUG_RUNTIME_STATE_ENABLED_ERROR; + + return r; +} + static int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) { uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode; @@ -77,6 +128,8 @@ static void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int else kfd_dbg_set_mes_debug_mode(pdd); } + + kfd_dbg_set_workaround(target, false); } int kfd_dbg_trap_disable(struct kfd_process *target) @@ -111,6 +164,10 @@ static int kfd_dbg_trap_activate(struct kfd_process *target) { int i, r = 0; + r = kfd_dbg_set_workaround(target, true); + if (r) + return r; + for (i = 0; i < target->n_pdds; i++) { struct kfd_process_device *pdd = target->pdds[i]; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 17481f824647..3e56225f6ef6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -31,7 +31,8 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, uint32_t *runtime_info_size); static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) { - return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2); + return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || + KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0); } /* diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 495c9238254e..44d87943e40a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -226,6 +226,10 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, queue_input.paging = false; queue_input.tba_addr = qpd->tba_addr; queue_input.tma_addr = qpd->tma_addr; + queue_input.trap_en = KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) || + KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0) || + q->properties.is_dbg_wa; + queue_input.skip_process_ctx_clear = qpd->pqm->process->debug_trap_enabled; queue_type = convert_to_mes_queue_type(q->properties.type); if (queue_type < 0) { @@ -1716,6 +1720,9 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, * updates the is_evicted flag but is a no-op otherwise. */ q->properties.is_evicted = !!qpd->evicted; + q->properties.is_dbg_wa = qpd->pqm->process->debug_trap_enabled && + KFD_GC_VERSION(q->device) >= IP_VERSION(11, 0, 0) && + KFD_GC_VERSION(q->device) < IP_VERSION(12, 0, 0); if (qd) mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c index eb11940bec34..65c9f01a1f86 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c @@ -48,8 +48,7 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, struct cik_mqd *m; uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ - if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || - !minfo->cu_mask.ptr) + if (!minfo || !minfo->cu_mask.ptr) return; mqd_symmetrically_map_cu_mask(mm, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index eaaa4f4ddaaa..a0ac4f2fe6b5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -48,8 +48,7 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, struct v10_compute_mqd *m; uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ - if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || - !minfo->cu_mask.ptr) + if (!minfo || !minfo->cu_mask.ptr) return; mqd_symmetrically_map_cu_mask(mm, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index 3a48bbc589fe..9a9b4e853516 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -46,15 +46,33 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, { struct v11_compute_mqd *m; uint32_t se_mask[KFD_MAX_NUM_SE] = {0}; + bool has_wa_flag = minfo && (minfo->update_flag & (UPDATE_FLAG_DBG_WA_ENABLE | + UPDATE_FLAG_DBG_WA_DISABLE)); - if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || - !minfo->cu_mask.ptr) + if (!minfo || !(has_wa_flag || minfo->cu_mask.ptr)) return; + m = get_mqd(mqd); + + if (has_wa_flag) { + uint32_t wa_mask = minfo->update_flag == UPDATE_FLAG_DBG_WA_ENABLE ? + 0xffff : 0xffffffff; + + m->compute_static_thread_mgmt_se0 = wa_mask; + m->compute_static_thread_mgmt_se1 = wa_mask; + m->compute_static_thread_mgmt_se2 = wa_mask; + m->compute_static_thread_mgmt_se3 = wa_mask; + m->compute_static_thread_mgmt_se4 = wa_mask; + m->compute_static_thread_mgmt_se5 = wa_mask; + m->compute_static_thread_mgmt_se6 = wa_mask; + m->compute_static_thread_mgmt_se7 = wa_mask; + + return; + } + mqd_symmetrically_map_cu_mask(mm, minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask); - m = get_mqd(mqd); m->compute_static_thread_mgmt_se0 = se_mask[0]; m->compute_static_thread_mgmt_se1 = se_mask[1]; m->compute_static_thread_mgmt_se2 = se_mask[2]; @@ -109,6 +127,7 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, uint64_t addr; struct v11_compute_mqd *m; int size; + uint32_t wa_mask = q->is_dbg_wa ? 0xffff : 0xffffffff; m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr; addr = mqd_mem_obj->gpu_addr; @@ -122,14 +141,15 @@ static void init_mqd(struct mqd_manager *mm, void **mqd, m->header = 0xC0310800; m->compute_pipelinestat_enable = 1; - m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF; - m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF; + + m->compute_static_thread_mgmt_se0 = wa_mask; + m->compute_static_thread_mgmt_se1 = wa_mask; + m->compute_static_thread_mgmt_se2 = wa_mask; + m->compute_static_thread_mgmt_se3 = wa_mask; + m->compute_static_thread_mgmt_se4 = wa_mask; + m->compute_static_thread_mgmt_se5 = wa_mask; + m->compute_static_thread_mgmt_se6 = wa_mask; + m->compute_static_thread_mgmt_se7 = wa_mask; m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK | 0x55 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index b7c95158d4a0..5b87c244e909 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -65,8 +65,7 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, struct v9_mqd *m; uint32_t se_mask[KFD_MAX_NUM_SE] = {0}; - if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || - !minfo->cu_mask.ptr) + if (!minfo || !minfo->cu_mask.ptr) return; mqd_symmetrically_map_cu_mask(mm, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c index fe69492b1bb3..d1e962da51dd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c @@ -51,8 +51,7 @@ static void update_cu_mask(struct mqd_manager *mm, void *mqd, struct vi_mqd *m; uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */ - if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) || - !minfo->cu_mask.ptr) + if (!minfo || !minfo->cu_mask.ptr) return; mqd_symmetrically_map_cu_mask(mm, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 8fca7175daab..f0a45d184c8f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -513,6 +513,8 @@ struct queue_properties { bool is_active; bool is_gws; uint32_t pm4_target_xcc; + bool is_dbg_wa; + bool is_user_cu_masked; /* Not relevant for user mode queues in cp scheduling */ unsigned int vmid; /* Relevant only for sdma queues*/ @@ -535,7 +537,8 @@ struct queue_properties { !(q).is_evicted) enum mqd_update_flag { - UPDATE_FLAG_CU_MASK = 0, + UPDATE_FLAG_DBG_WA_ENABLE = 1, + UPDATE_FLAG_DBG_WA_DISABLE = 2, }; struct mqd_update_info { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index b100933340d2..43d432b5c5bc 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -506,8 +506,12 @@ int pqm_update_mqd(struct process_queue_manager *pqm, return -EFAULT; } + /* CUs are masked for debugger requirements so deny user mask */ + if (pqn->q->properties.is_dbg_wa && minfo && minfo->cu_mask.ptr) + return -EBUSY; + /* ASICs that have WGPs must enforce pairwise enabled mask checks. */ - if (minfo && minfo->update_flag == UPDATE_FLAG_CU_MASK && minfo->cu_mask.ptr && + if (minfo && minfo->cu_mask.ptr && KFD_GC_VERSION(pqn->q->device) >= IP_VERSION(10, 0, 0)) { int i; @@ -526,6 +530,9 @@ int pqm_update_mqd(struct process_queue_manager *pqm, if (retval != 0) return retval; + if (minfo && minfo->cu_mask.ptr) + pqn->q->properties.is_user_cu_masked = true; + return 0; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index 3def25b2bdbb..faa7939f35bd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1863,10 +1863,13 @@ static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *de { bool firmware_supported = true; + /* + * FIXME: GFX11 FW currently not sufficient to deal with CWSR WA. + * Updated FW with API changes coming soon. + */ if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) && KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) { - firmware_supported = - (dev->gpu->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 9; + firmware_supported = false; goto out; } -- cgit v1.2.3 From 44b87bb0836c65d1b9d21b01503eb6e9b9297771 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Wed, 6 Apr 2022 12:03:31 -0400 Subject: drm/amdkfd: add raise exception event function Exception events can be generated from interrupts or queue activitity. The raise event function will save exception status of a queue, device or process then notify the debugger of the status change by writing to a debugger polled file descriptor that the debugger provides during debug attach. For memory violation exceptions, extra exception data will be saved. The debugger will be able to query the saved exception states by query operation that will be provided by follow up patches. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 104 +++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 7 +++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 10 +++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 + 4 files changed, 123 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 5e2ee2d1acc4..dccb27fc764b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -24,6 +24,107 @@ #include "kfd_device_queue_manager.h" #include +void debug_event_write_work_handler(struct work_struct *work) +{ + struct kfd_process *process; + + static const char write_data = '.'; + loff_t pos = 0; + + process = container_of(work, + struct kfd_process, + debug_event_workarea); + + kernel_write(process->dbg_ev_file, &write_data, 1, &pos); +} + +/* update process/device/queue exception status, write to descriptor + * only if exception_status is enabled. + */ +bool kfd_dbg_ev_raise(uint64_t event_mask, + struct kfd_process *process, struct kfd_node *dev, + unsigned int source_id, bool use_worker, + void *exception_data, size_t exception_data_size) +{ + struct process_queue_manager *pqm; + struct process_queue_node *pqn; + int i; + static const char write_data = '.'; + loff_t pos = 0; + bool is_subscribed = true; + + if (!(process && process->debug_trap_enabled)) + return false; + + mutex_lock(&process->event_mutex); + + if (event_mask & KFD_EC_MASK_DEVICE) { + for (i = 0; i < process->n_pdds; i++) { + struct kfd_process_device *pdd = process->pdds[i]; + + if (pdd->dev != dev) + continue; + + pdd->exception_status |= event_mask & KFD_EC_MASK_DEVICE; + + if (event_mask & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) { + if (!pdd->vm_fault_exc_data) { + pdd->vm_fault_exc_data = kmemdup( + exception_data, + exception_data_size, + GFP_KERNEL); + if (!pdd->vm_fault_exc_data) + pr_debug("Failed to allocate exception data memory"); + } else { + pr_debug("Debugger exception data not saved\n"); + print_hex_dump_bytes("exception data: ", + DUMP_PREFIX_OFFSET, + exception_data, + exception_data_size); + } + } + break; + } + } else if (event_mask & KFD_EC_MASK_PROCESS) { + process->exception_status |= event_mask & KFD_EC_MASK_PROCESS; + } else { + pqm = &process->pqm; + list_for_each_entry(pqn, &pqm->queues, + process_queue_list) { + int target_id; + + if (!pqn->q) + continue; + + target_id = event_mask & KFD_EC_MASK(EC_QUEUE_NEW) ? + pqn->q->properties.queue_id : + pqn->q->doorbell_id; + + if (pqn->q->device != dev || target_id != source_id) + continue; + + pqn->q->properties.exception_status |= event_mask; + break; + } + } + + if (process->exception_enable_mask & event_mask) { + if (use_worker) + schedule_work(&process->debug_event_workarea); + else + kernel_write(process->dbg_ev_file, + &write_data, + 1, + &pos); + } else { + is_subscribed = false; + } + + mutex_unlock(&process->event_mutex); + + return is_subscribed; +} + static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable) { struct mqd_update_info minfo = {0}; @@ -99,6 +200,9 @@ static void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int { int i; + if (!unwind) + cancel_work_sync(&target->debug_event_workarea); + for (i = 0; i < target->n_pdds; i++) { struct kfd_process_device *pdd = target->pdds[i]; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 3e56225f6ef6..66ee7b95d08a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -25,6 +25,11 @@ #include "kfd_priv.h" +bool kfd_dbg_ev_raise(uint64_t event_mask, + struct kfd_process *process, struct kfd_node *dev, + unsigned int source_id, bool use_worker, + void *exception_data, + size_t exception_data_size); int kfd_dbg_trap_disable(struct kfd_process *target); int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, void __user *runtime_info, @@ -35,6 +40,8 @@ static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0); } +void debug_event_write_work_handler(struct work_struct *work); + /* * If GFX off is enabled, chips that do not support RLC restore for the debug * registers will disable GFX off temporarily for the entire debug session. diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index f0a45d184c8f..b18cd4bf76bf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -529,6 +529,7 @@ struct queue_properties { uint32_t ctl_stack_size; uint64_t tba_addr; uint64_t tma_addr; + uint64_t exception_status; }; #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ @@ -820,6 +821,11 @@ struct kfd_process_device { uint64_t page_in; uint64_t page_out; + /* Exception code status*/ + uint64_t exception_status; + void *vm_fault_exc_data; + size_t vm_fault_exc_data_size; + /* Tracks debug per-vmid request settings */ uint32_t spi_dbg_override; uint32_t spi_dbg_launch_mode; @@ -955,12 +961,16 @@ struct kfd_process { /* Exception code enable mask and status */ uint64_t exception_enable_mask; + uint64_t exception_status; /* shared virtual memory registered by this process */ struct svm_range_list svms; bool xnack_enabled; + /* Work area for debugger event writer worker. */ + struct work_struct debug_event_workarea; + /* Tracks debug per-vmid request for debug flags */ bool dbg_flags; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index e77cadadb09b..f904d6d6e01c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1509,6 +1509,8 @@ static struct kfd_process *create_process(const struct task_struct *thread) kfd_unref_process(process); get_task_struct(process->lead_thread); + INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler); + return process; err_register_notifier: -- cgit v1.2.3 From c2d2588c702364ff53916ddd97e2b26fd4f4a317 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 8 Apr 2022 12:49:48 -0400 Subject: drm/amdkfd: add send exception operation Add a debug operation that allows the debugger to send an exception directly to runtime through a payload address. For memory violations, normal vmfault signals will be applied to notify runtime instead after passing in the saved exception data when a memory violation was raised to the debugger. For runtime exceptions, this will unblock the runtime enable function which will be explained and implemented in a follow up patch. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 43 ++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 6 ++ drivers/gpu/drm/amd/amdkfd/kfd_events.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 7 ++- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 71 +++++++++++++++++++++++- 8 files changed, 135 insertions(+), 6 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c index 4ebfff6b6c55..795382b55e0a 100644 --- a/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c +++ b/drivers/gpu/drm/amd/amdkfd/cik_event_interrupt.c @@ -118,9 +118,9 @@ static void cik_event_interrupt_wq(struct kfd_node *dev, return; if (info.vmid == vmid) - kfd_signal_vm_fault_event(dev, pasid, &info); + kfd_signal_vm_fault_event(dev, pasid, &info, NULL); else - kfd_signal_vm_fault_event(dev, pasid, NULL); + kfd_signal_vm_fault_event(dev, pasid, NULL, NULL); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index d4df424e4514..5e57b3e96ff9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2833,6 +2833,11 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v r = kfd_dbg_trap_disable(target); break; case KFD_IOC_DBG_TRAP_SEND_RUNTIME_EVENT: + r = kfd_dbg_send_exception_to_runtime(target, + args->send_runtime_event.gpu_id, + args->send_runtime_event.queue_id, + args->send_runtime_event.exception_mask); + break; case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED: case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE: case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index dccb27fc764b..61098975bb0e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -125,6 +125,49 @@ bool kfd_dbg_ev_raise(uint64_t event_mask, return is_subscribed; } +int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, + unsigned int dev_id, + unsigned int queue_id, + uint64_t error_reason) +{ + if (error_reason & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) { + struct kfd_process_device *pdd = NULL; + struct kfd_hsa_memory_exception_data *data; + int i; + + for (i = 0; i < p->n_pdds; i++) { + if (p->pdds[i]->dev->id == dev_id) { + pdd = p->pdds[i]; + break; + } + } + + if (!pdd) + return -ENODEV; + + data = (struct kfd_hsa_memory_exception_data *) + pdd->vm_fault_exc_data; + + kfd_dqm_evict_pasid(pdd->dev->dqm, p->pasid); + kfd_signal_vm_fault_event(pdd->dev, p->pasid, NULL, data); + error_reason &= ~KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION); + } + + if (error_reason & (KFD_EC_MASK(EC_PROCESS_RUNTIME))) { + /* + * block should only happen after the debugger receives runtime + * enable notice. + */ + up(&p->runtime_enable_sema); + error_reason &= ~KFD_EC_MASK(EC_PROCESS_RUNTIME); + } + + if (error_reason) + return kfd_send_exception_to_runtime(p, queue_id, error_reason); + + return 0; +} + static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable) { struct mqd_update_info minfo = {0}; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 66ee7b95d08a..2c6866bb8850 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -34,6 +34,12 @@ int kfd_dbg_trap_disable(struct kfd_process *target); int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, void __user *runtime_info, uint32_t *runtime_info_size); + +int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, + unsigned int dev_id, + unsigned int queue_id, + uint64_t error_reason); + static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) { return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_events.c index 9926186f88a6..0a5e7b172a64 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c @@ -1222,7 +1222,8 @@ void kfd_signal_hw_exception_event(u32 pasid) } void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, - struct kfd_vm_fault_info *info) + struct kfd_vm_fault_info *info, + struct kfd_hsa_memory_exception_data *data) { struct kfd_event *ev; uint32_t id; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c index 861bccb1e9dc..8cf58be80f4e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c @@ -362,7 +362,7 @@ static void event_interrupt_wq_v9(struct kfd_node *dev, kfd_smi_event_update_vmfault(dev, pasid); kfd_dqm_evict_pasid(dev->dqm, pasid); - kfd_signal_vm_fault_event(dev, pasid, &info); + kfd_signal_vm_fault_event(dev, pasid, &info, NULL); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index b18cd4bf76bf..58b82fa59584 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -979,6 +979,7 @@ struct kfd_process { bool queues_paused; /* Tracks runtime enable status */ + struct semaphore runtime_enable_sema; struct kfd_runtime_info runtime_info; }; @@ -1447,7 +1448,8 @@ int kfd_get_num_events(struct kfd_process *p); int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, - struct kfd_vm_fault_info *info); + struct kfd_vm_fault_info *info, + struct kfd_hsa_memory_exception_data *data); void kfd_signal_reset_event(struct kfd_node *dev); @@ -1463,6 +1465,9 @@ static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0); } +int kfd_send_exception_to_runtime(struct kfd_process *p, + unsigned int queue_id, + uint64_t error_reason); bool kfd_is_locked(void); /* Compute profile */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index f904d6d6e01c..5cbfcaf08c8f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1462,6 +1462,7 @@ static struct kfd_process *create_process(const struct task_struct *thread) process->debugger_process = NULL; process->exception_enable_mask = 0; atomic_set(&process->debugged_process_count, 0); + sema_init(&process->runtime_enable_sema, 0); process->pasid = kfd_pasid_alloc(); if (process->pasid == 0) { @@ -2120,6 +2121,75 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) } } +struct send_exception_work_handler_workarea { + struct work_struct work; + struct kfd_process *p; + unsigned int queue_id; + uint64_t error_reason; +}; + +static void send_exception_work_handler(struct work_struct *work) +{ + struct send_exception_work_handler_workarea *workarea; + struct kfd_process *p; + struct queue *q; + struct mm_struct *mm; + struct kfd_context_save_area_header __user *csa_header; + uint64_t __user *err_payload_ptr; + uint64_t cur_err; + uint32_t ev_id; + + workarea = container_of(work, + struct send_exception_work_handler_workarea, + work); + p = workarea->p; + + mm = get_task_mm(p->lead_thread); + + if (!mm) + return; + + kthread_use_mm(mm); + + q = pqm_get_user_queue(&p->pqm, workarea->queue_id); + + if (!q) + goto out; + + csa_header = (void __user *)q->properties.ctx_save_restore_area_address; + + get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr); + get_user(cur_err, err_payload_ptr); + cur_err |= workarea->error_reason; + put_user(cur_err, err_payload_ptr); + get_user(ev_id, &csa_header->err_event_id); + + kfd_set_event(p, ev_id); + +out: + kthread_unuse_mm(mm); + mmput(mm); +} + +int kfd_send_exception_to_runtime(struct kfd_process *p, + unsigned int queue_id, + uint64_t error_reason) +{ + struct send_exception_work_handler_workarea worker; + + INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler); + + worker.p = p; + worker.queue_id = queue_id; + worker.error_reason = error_reason; + + schedule_work(&worker.work); + flush_work(&worker.work); + destroy_work_on_stack(&worker.work); + + return 0; +} + struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id) { int i; @@ -2179,4 +2249,3 @@ int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data) } #endif - -- cgit v1.2.3 From 455227c4642c5e1867213cea73a527e431779060 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 8 Apr 2022 13:12:24 -0400 Subject: drm/amdkfd: add runtime enable operation The debugger can attach to a process prior to HSA enablement (i.e. inferior is spawned by the debugger and attached to immediately before target process has been enabled for HSA dispatches) or it can attach to a running target that is already HSA enabled. Either way, the debugger needs to know the enablement status to know when it can inspect queues. For the scenario where the debugger spawns the target process, it will have to wait for ROCr's runtime enable request from the target. The runtime enable request will be able to see that its process has been debug attached. ROCr raises an EC_PROCESS_RUNTIME signal to the debugger then blocks the target process while waiting the debugger's response. Once the debugger has received the runtime signal, it will unblock the target process. For the scenario where the debugger attaches to a running target process, ROCr will set the target process' runtime status as enabled so that on an attach request, the debugger will be able to see this status and will continue with debug enablement as normal. A secondary requirement is to conditionally enable the trap tempories only if the user requests it (env var HSA_ENABLE_DEBUG=1) or if the debugger attaches with HSA runtime enabled. This is because setting up the trap temporaries incurs a performance overhead that is unacceptable for microbench performance in normal mode for certain customers. In the scenario where the debugger spawns the target process, when ROCr detects that the debugger has attached during the runtime enable request, it will enable the trap temporaries before it blocks the target process while waiting for the debugger to respond. In the scenario where the debugger attaches to a running target process, it will enable to trap temporaries itself. Finally, there is an additional restriction that is required to be enforced with runtime enable and HW debug mode setting. The debugger must first ensure that HW debug mode has been enabled before permitting HW debug mode operations. With single process debug devices, allowing the debugger to set debug HW modes prior to trap activation means that debug HW mode setting can occur before the KFD has reserved the debug VMID (0xf) from the hardware scheduler's VMID allocation resource pool. This can result in the hardware scheduler assigning VMID 0xf to a non-debugged process and having that process inherit debug HW mode settings intended for the debugged target process instead, which is both incorrect and potentially fatal for normal mode operation. With multi process debug devices, allowing the debugger to set debug HW modes prior to trap activation means that non-debugged processes migrating to a new VMID could inherit unintended debug settings. All debug operations that touch HW settings must require trap activation where trap activation is triggered by both debug attach and runtime enablement (target has KFD opened and is ready to dispatch work). Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 143 ++++++++++++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 4 + drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 + 4 files changed, 150 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 5e57b3e96ff9..615fa9ab36b7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2738,11 +2738,140 @@ static int kfd_ioctl_criu(struct file *filep, struct kfd_process *p, void *data) return ret; } -static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data) +static int runtime_enable(struct kfd_process *p, uint64_t r_debug, + bool enable_ttmp_setup) +{ + int i = 0, ret = 0; + + if (p->is_runtime_retry) + goto retry; + + if (p->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED) + return -EBUSY; + + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + + if (pdd->qpd.queue_count) + return -EEXIST; + } + + p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED; + p->runtime_info.r_debug = r_debug; + p->runtime_info.ttmp_setup = enable_ttmp_setup; + + if (p->runtime_info.ttmp_setup) { + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + + if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) { + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + pdd->dev->kfd2kgd->enable_debug_trap( + pdd->dev->adev, + true, + pdd->dev->vm_info.last_vmid_kfd); + } else if (kfd_dbg_is_per_vmid_supported(pdd->dev)) { + pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap( + pdd->dev->adev, + false, + 0); + } + } + } + +retry: + if (p->debug_trap_enabled) { + if (!p->is_runtime_retry) { + kfd_dbg_trap_activate(p); + kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME), + p, NULL, 0, false, NULL, 0); + } + + mutex_unlock(&p->mutex); + ret = down_interruptible(&p->runtime_enable_sema); + mutex_lock(&p->mutex); + + p->is_runtime_retry = !!ret; + } + + return ret; +} + +static int runtime_disable(struct kfd_process *p) { + int i = 0, ret; + bool was_enabled = p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED; + + p->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_DISABLED; + p->runtime_info.r_debug = 0; + + if (p->debug_trap_enabled) { + if (was_enabled) + kfd_dbg_trap_deactivate(p, false, 0); + + if (!p->is_runtime_retry) + kfd_dbg_ev_raise(KFD_EC_MASK(EC_PROCESS_RUNTIME), + p, NULL, 0, false, NULL, 0); + + mutex_unlock(&p->mutex); + ret = down_interruptible(&p->runtime_enable_sema); + mutex_lock(&p->mutex); + + p->is_runtime_retry = !!ret; + if (ret) + return ret; + } + + if (was_enabled && p->runtime_info.ttmp_setup) { + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + + if (!kfd_dbg_is_rlc_restore_supported(pdd->dev)) + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + } + } + + p->runtime_info.ttmp_setup = false; + + /* disable ttmp setup */ + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + + if (kfd_dbg_is_per_vmid_supported(pdd->dev)) { + pdd->spi_dbg_override = + pdd->dev->kfd2kgd->disable_debug_trap( + pdd->dev->adev, + false, + pdd->dev->vm_info.last_vmid_kfd); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + debug_refresh_runlist(pdd->dev->dqm); + else + kfd_dbg_set_mes_debug_mode(pdd); + } + } + return 0; } +static int kfd_ioctl_runtime_enable(struct file *filep, struct kfd_process *p, void *data) +{ + struct kfd_ioctl_runtime_enable_args *args = data; + int r; + + mutex_lock(&p->mutex); + + if (args->mode_mask & KFD_RUNTIME_ENABLE_MODE_ENABLE_MASK) + r = runtime_enable(p, args->r_debug, + !!(args->mode_mask & KFD_RUNTIME_ENABLE_MODE_TTMP_SAVE_MASK)); + else + r = runtime_disable(p); + + mutex_unlock(&p->mutex); + + return r; +} + static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, void *data) { struct kfd_ioctl_dbg_trap_args *args = data; @@ -2815,6 +2944,18 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v goto unlock_out; } + if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_ENABLED && + (args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE || + args->op == KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE || + args->op == KFD_IOC_DBG_TRAP_SUSPEND_QUEUES || + args->op == KFD_IOC_DBG_TRAP_RESUME_QUEUES || + args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH || + args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH || + args->op == KFD_IOC_DBG_TRAP_SET_FLAGS)) { + r = -EPERM; + goto unlock_out; + } + switch (args->op) { case KFD_IOC_DBG_TRAP_ENABLE: if (target != p) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 61098975bb0e..a19c21d04438 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -219,7 +219,7 @@ unwind: return r; } -static int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) +int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) { uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode; uint32_t flags = pdd->process->dbg_flags; @@ -239,7 +239,7 @@ static int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) * to unwind * else: ignored */ -static void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count) +void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count) { int i; @@ -307,7 +307,7 @@ int kfd_dbg_trap_disable(struct kfd_process *target) return 0; } -static int kfd_dbg_trap_activate(struct kfd_process *target) +int kfd_dbg_trap_activate(struct kfd_process *target) { int i, r = 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 2c6866bb8850..fca928564948 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -25,6 +25,8 @@ #include "kfd_priv.h" +void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count); +int kfd_dbg_trap_activate(struct kfd_process *target); bool kfd_dbg_ev_raise(uint64_t event_mask, struct kfd_process *process, struct kfd_node *dev, unsigned int source_id, bool use_worker, @@ -77,4 +79,6 @@ static inline bool kfd_dbg_has_gws_support(struct kfd_node *dev) /* Assume debugging and cooperative launch supported otherwise. */ return true; } + +int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd); #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 58b82fa59584..4b80f74b9de0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -980,6 +980,7 @@ struct kfd_process { /* Tracks runtime enable status */ struct semaphore runtime_enable_sema; + bool is_runtime_retry; struct kfd_runtime_info runtime_info; }; -- cgit v1.2.3 From 50cff45e274896235d371f16eab67a180e12a732 Mon Sep 17 00:00:00 2001 From: Jay Cornwall Date: Mon, 1 Mar 2021 18:34:39 -0600 Subject: drm/amdkfd: add debug trap enabled flag to tma Trap handler behavior will differ when a debugger is attached. Make the debug trap flag available in the trap handler TMA. Update it when the debug trap ioctl is invoked. Signed-off-by: Jay Cornwall Reviewed-by: Felix Kuehling Signed-off-by: Jonathan Kim Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 11 +++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 ++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 15 +++++++++++++++ 3 files changed, 28 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index a19c21d04438..17e8e9edccbf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -256,6 +256,8 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind if (unwind && i == unwind_count) break; + kfd_process_set_trap_debug_flag(&pdd->qpd, false); + /* GFX off is already disabled by debug activate if not RLC restore supported. */ if (kfd_dbg_is_rlc_restore_supported(pdd->dev)) amdgpu_gfx_off_ctrl(pdd->dev->adev, false); @@ -351,6 +353,15 @@ int kfd_dbg_trap_activate(struct kfd_process *target) if (kfd_dbg_is_rlc_restore_supported(pdd->dev)) amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + /* + * Setting the debug flag in the trap handler requires that the TMA has been + * allocated, which occurs during CWSR initialization. + * In the event that CWSR has not been initialized at this point, setting the + * flag will be called again during CWSR initialization if the target process + * is still debug enabled. + */ + kfd_process_set_trap_debug_flag(&pdd->qpd, true); + if (!pdd->dev->kfd->shared_resources.enable_mes) r = debug_refresh_runlist(pdd->dev->dqm); else diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 4b80f74b9de0..a02fb939614a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1157,6 +1157,8 @@ int kfd_init_apertures(struct kfd_process *process); void kfd_process_set_trap_handler(struct qcm_process_device *qpd, uint64_t tba_addr, uint64_t tma_addr); +void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd, + bool enabled); /* CWSR initialization */ int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 5cbfcaf08c8f..3b7f219c9d06 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1309,6 +1309,8 @@ int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size); + kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled); + qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET; pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n", qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr); @@ -1345,6 +1347,9 @@ static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size); + kfd_process_set_trap_debug_flag(&pdd->qpd, + pdd->process->debug_trap_enabled); + qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET; pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n", qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr); @@ -1431,6 +1436,16 @@ bool kfd_process_xnack_mode(struct kfd_process *p, bool supported) return true; } +void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd, + bool enabled) +{ + if (qpd->cwsr_kaddr) { + uint64_t *tma = + (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET); + tma[2] = enabled; + } +} + /* * On return the kfd_process is fully operational and will be freed when the * mm is released -- cgit v1.2.3 From 12fb1ad70d65edc3405884792d044fa79df7244f Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 22 Apr 2022 12:26:18 -0400 Subject: drm/amdkfd: update process interrupt handling for debug events The debugger must be notified by any debugger subscribed exception that comes from hardware interrupts. If a debugger session exits, any exceptions it subscribed to may still have interrupts in the interrupt ring buffer or KGD/KFD pipeline. To prevent a new session from inheriting stale interrupts, when a new queue is created, open an interrupt drain and allow the IH ring to drain from a timestamped checkpoint. Then inject a custom IV so that once the custom IV is picked up by the KFD, it's safe to close the drain and proceed with queue creation. The drain must also be on debug disable as SW interrupts may still be processed. Drain at this time and clear all the exception status. The debugger may also not be attached nor subscibed to certain exceptions so forward them directly to the runtime. GFX10 also requires its own IV processing, hence the creation of kfd_int_process_v10.c. This is because the IV from SQ interrupts are packed into a new continguous format unlike GFX9. To make this clear, a separate interrupting handling code file was created. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 16 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 + drivers/gpu/drm/amd/amdkfd/Makefile | 1 + drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 84 +++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 6 + drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_int_process_v10.c | 405 +++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c | 26 +- drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c | 98 ++++- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 12 + drivers/gpu/drm/amd/amdkfd/kfd_process.c | 47 +++ .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 4 + 12 files changed, 686 insertions(+), 19 deletions(-) create mode 100644 drivers/gpu/drm/amd/amdkfd/kfd_int_process_v10.c (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 66f80b9ab0c5..98cd52bb005f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -777,6 +777,22 @@ void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bo amdgpu_umc_poison_handler(adev, reset); } +int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev, + uint32_t *payload) +{ + int ret; + + /* Device or IH ring is not ready so bail. */ + ret = amdgpu_ih_wait_on_checkpoint_process_ts(adev, &adev->irq.ih); + if (ret) + return ret; + + /* Send payload to fence KFD interrupts */ + amdgpu_amdkfd_interrupt(adev, payload); + + return 0; +} + bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev) { if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index 94cc456761e5..dd740e64e6e1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -250,6 +250,8 @@ int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst, struct amdgpu_device *src, bool is_min); int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min); +int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev, + uint32_t *payload); /* Read user wptr from a specified user address space with page fault * disabled. The memory must be pinned and mapped to the hardware when diff --git a/drivers/gpu/drm/amd/amdkfd/Makefile b/drivers/gpu/drm/amd/amdkfd/Makefile index 747754428073..2ec8f27c5366 100644 --- a/drivers/gpu/drm/amd/amdkfd/Makefile +++ b/drivers/gpu/drm/amd/amdkfd/Makefile @@ -53,6 +53,7 @@ AMDKFD_FILES := $(AMDKFD_PATH)/kfd_module.o \ $(AMDKFD_PATH)/kfd_events.o \ $(AMDKFD_PATH)/cik_event_interrupt.o \ $(AMDKFD_PATH)/kfd_int_process_v9.o \ + $(AMDKFD_PATH)/kfd_int_process_v10.o \ $(AMDKFD_PATH)/kfd_int_process_v11.o \ $(AMDKFD_PATH)/kfd_smi_events.o \ $(AMDKFD_PATH)/kfd_crat.o \ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 17e8e9edccbf..68b657398d41 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -125,6 +125,64 @@ bool kfd_dbg_ev_raise(uint64_t event_mask, return is_subscribed; } +/* set pending event queue entry from ring entry */ +bool kfd_set_dbg_ev_from_interrupt(struct kfd_node *dev, + unsigned int pasid, + uint32_t doorbell_id, + uint64_t trap_mask, + void *exception_data, + size_t exception_data_size) +{ + struct kfd_process *p; + bool signaled_to_debugger_or_runtime = false; + + p = kfd_lookup_process_by_pasid(pasid); + + if (!p) + return false; + + if (!kfd_dbg_ev_raise(trap_mask, p, dev, doorbell_id, true, + exception_data, exception_data_size)) { + struct process_queue_manager *pqm; + struct process_queue_node *pqn; + + if (!!(trap_mask & KFD_EC_MASK_QUEUE) && + p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) { + mutex_lock(&p->mutex); + + pqm = &p->pqm; + list_for_each_entry(pqn, &pqm->queues, + process_queue_list) { + + if (!(pqn->q && pqn->q->device == dev && + pqn->q->doorbell_id == doorbell_id)) + continue; + + kfd_send_exception_to_runtime(p, pqn->q->properties.queue_id, + trap_mask); + + signaled_to_debugger_or_runtime = true; + + break; + } + + mutex_unlock(&p->mutex); + } else if (trap_mask & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) { + kfd_dqm_evict_pasid(dev->dqm, p->pasid); + kfd_signal_vm_fault_event(dev, p->pasid, NULL, + exception_data); + + signaled_to_debugger_or_runtime = true; + } + } else { + signaled_to_debugger_or_runtime = true; + } + + kfd_unref_process(p); + + return signaled_to_debugger_or_runtime; +} + int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, unsigned int dev_id, unsigned int queue_id, @@ -281,6 +339,31 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind kfd_dbg_set_workaround(target, false); } +static void kfd_dbg_clean_exception_status(struct kfd_process *target) +{ + struct process_queue_manager *pqm; + struct process_queue_node *pqn; + int i; + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + kfd_process_drain_interrupts(pdd); + + pdd->exception_status = 0; + } + + pqm = &target->pqm; + list_for_each_entry(pqn, &pqm->queues, process_queue_list) { + if (!pqn->q) + continue; + + pqn->q->properties.exception_status = 0; + } + + target->exception_status = 0; +} + int kfd_dbg_trap_disable(struct kfd_process *target) { if (!target->debug_trap_enabled) @@ -304,6 +387,7 @@ int kfd_dbg_trap_disable(struct kfd_process *target) } target->debug_trap_enabled = false; + kfd_dbg_clean_exception_status(target); kfd_unref_process(target); return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index fca928564948..5153ccbd7fd1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -27,6 +27,12 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count); int kfd_dbg_trap_activate(struct kfd_process *target); +bool kfd_set_dbg_ev_from_interrupt(struct kfd_node *dev, + unsigned int pasid, + uint32_t doorbell_id, + uint64_t trap_mask, + void *exception_data, + size_t exception_data_size); bool kfd_dbg_ev_raise(uint64_t event_mask, struct kfd_process *process, struct kfd_node *dev, unsigned int source_id, bool use_worker, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index f0ed6e6416c3..2c36bb578633 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -140,6 +140,8 @@ static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd) case IP_VERSION(9, 4, 1): /* ARCTURUS */ case IP_VERSION(9, 4, 2): /* ALDEBARAN */ case IP_VERSION(9, 4, 3): /* GC 9.4.3 */ + kfd->device_info.event_interrupt_class = &event_interrupt_class_v9; + break; case IP_VERSION(10, 3, 1): /* VANGOGH */ case IP_VERSION(10, 3, 3): /* YELLOW_CARP */ case IP_VERSION(10, 3, 6): /* GC 10.3.6 */ @@ -153,7 +155,7 @@ static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd) case IP_VERSION(10, 3, 2): /* NAVY_FLOUNDER */ case IP_VERSION(10, 3, 4): /* DIMGREY_CAVEFISH */ case IP_VERSION(10, 3, 5): /* BEIGE_GOBY */ - kfd->device_info.event_interrupt_class = &event_interrupt_class_v9; + kfd->device_info.event_interrupt_class = &event_interrupt_class_v10; break; case IP_VERSION(11, 0, 0): case IP_VERSION(11, 0, 1): diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v10.c new file mode 100644 index 000000000000..c7991e07b6be --- /dev/null +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v10.c @@ -0,0 +1,405 @@ +/* + * Copyright 2023 Advanced Micro Devices, Inc. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR + * OTHER DEALINGS IN THE SOFTWARE. + */ + +#include "kfd_events.h" +#include "kfd_debug.h" +#include "soc15_int.h" +#include "kfd_device_queue_manager.h" + +/* + * GFX10 SQ Interrupts + * + * There are 3 encoding types of interrupts sourced from SQ sent as a 44-bit + * packet to the Interrupt Handler: + * Auto - Generated by the SQG (various cmd overflows, timestamps etc) + * Wave - Generated by S_SENDMSG through a shader program + * Error - HW generated errors (Illegal instructions, Memviols, EDC etc) + * + * The 44-bit packet is mapped as {context_id1[7:0],context_id0[31:0]} plus + * 4-bits for VMID (SOC15_VMID_FROM_IH_ENTRY) as such: + * + * - context_id1[7:6] + * Encoding type (0 = Auto, 1 = Wave, 2 = Error) + * + * - context_id0[24] + * PRIV bit indicates that Wave S_SEND or error occurred within trap + * + * - context_id0[22:0] + * 23-bit data with the following layout per encoding type: + * Auto - only context_id0[8:0] is used, which reports various interrupts + * generated by SQG. The rest is 0. + * Wave - user data sent from m0 via S_SENDMSG + * Error - Error type (context_id0[22:19]), Error Details (rest of bits) + * + * The other context_id bits show coordinates (SE/SH/CU/SIMD/WGP) for wave + * S_SENDMSG and Errors. These are 0 for Auto. + */ + +enum SQ_INTERRUPT_WORD_ENCODING { + SQ_INTERRUPT_WORD_ENCODING_AUTO = 0x0, + SQ_INTERRUPT_WORD_ENCODING_INST, + SQ_INTERRUPT_WORD_ENCODING_ERROR, +}; + +enum SQ_INTERRUPT_ERROR_TYPE { + SQ_INTERRUPT_ERROR_TYPE_EDC_FUE = 0x0, + SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST, + SQ_INTERRUPT_ERROR_TYPE_MEMVIOL, + SQ_INTERRUPT_ERROR_TYPE_EDC_FED, +}; + +/* SQ_INTERRUPT_WORD_AUTO_CTXID */ +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE__SHIFT 0 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__WLT__SHIFT 1 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_BUF0_FULL__SHIFT 2 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_BUF1_FULL__SHIFT 3 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_UTC_ERROR__SHIFT 7 +#define SQ_INTERRUPT_WORD_AUTO_CTXID1__SE_ID__SHIFT 4 +#define SQ_INTERRUPT_WORD_AUTO_CTXID1__ENCODING__SHIFT 6 + +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_MASK 0x00000001 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__WLT_MASK 0x00000002 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_BUF0_FULL_MASK 0x00000004 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_BUF1_FULL_MASK 0x00000008 +#define SQ_INTERRUPT_WORD_AUTO_CTXID0__THREAD_TRACE_UTC_ERROR_MASK 0x00000080 +#define SQ_INTERRUPT_WORD_AUTO_CTXID1__SE_ID_MASK 0x030 +#define SQ_INTERRUPT_WORD_AUTO_CTXID1__ENCODING_MASK 0x0c0 + +/* SQ_INTERRUPT_WORD_WAVE_CTXID */ +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__DATA__SHIFT 0 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__SA_ID__SHIFT 23 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__PRIV__SHIFT 24 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__WAVE_ID__SHIFT 25 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__SIMD_ID__SHIFT 30 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__WGP_ID__SHIFT 0 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__SE_ID__SHIFT 4 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__ENCODING__SHIFT 6 + +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__DATA_MASK 0x000007fffff +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__SA_ID_MASK 0x0000800000 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__PRIV_MASK 0x00001000000 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__WAVE_ID_MASK 0x0003e000000 +#define SQ_INTERRUPT_WORD_WAVE_CTXID0__SIMD_ID_MASK 0x000c0000000 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__WGP_ID_MASK 0x00f +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__SE_ID_MASK 0x030 +#define SQ_INTERRUPT_WORD_WAVE_CTXID1__ENCODING_MASK 0x0c0 + +#define KFD_CTXID0__ERR_TYPE_MASK 0x780000 +#define KFD_CTXID0__ERR_TYPE__SHIFT 19 + +/* GFX10 SQ interrupt ENC type bit (context_id1[7:6]) for wave s_sendmsg */ +#define KFD_CONTEXT_ID1_ENC_TYPE_WAVE_MASK 0x40 +/* GFX10 SQ interrupt PRIV bit (context_id0[24]) for s_sendmsg inside trap */ +#define KFD_CONTEXT_ID0_PRIV_MASK 0x1000000 +/* + * The debugger will send user data(m0) with PRIV=1 to indicate it requires + * notification from the KFD with the following queue id (DOORBELL_ID) and + * trap code (TRAP_CODE). + */ +#define KFD_CONTEXT_ID0_DEBUG_DOORBELL_MASK 0x0003ff +#define KFD_CONTEXT_ID0_DEBUG_TRAP_CODE_SHIFT 10 +#define KFD_CONTEXT_ID0_DEBUG_TRAP_CODE_MASK 0x07fc00 +#define KFD_DEBUG_DOORBELL_ID(ctxid0) ((ctxid0) & \ + KFD_CONTEXT_ID0_DEBUG_DOORBELL_MASK) +#define KFD_DEBUG_TRAP_CODE(ctxid0) (((ctxid0) & \ + KFD_CONTEXT_ID0_DEBUG_TRAP_CODE_MASK) \ + >> KFD_CONTEXT_ID0_DEBUG_TRAP_CODE_SHIFT) +#define KFD_DEBUG_CP_BAD_OP_ECODE_MASK 0x3fffc00 +#define KFD_DEBUG_CP_BAD_OP_ECODE_SHIFT 10 +#define KFD_DEBUG_CP_BAD_OP_ECODE(ctxid0) (((ctxid0) & \ + KFD_DEBUG_CP_BAD_OP_ECODE_MASK) \ + >> KFD_DEBUG_CP_BAD_OP_ECODE_SHIFT) + +static void event_interrupt_poison_consumption(struct kfd_node *dev, + uint16_t pasid, uint16_t client_id) +{ + int old_poison, ret = -EINVAL; + struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); + + if (!p) + return; + + /* all queues of a process will be unmapped in one time */ + old_poison = atomic_cmpxchg(&p->poison, 0, 1); + kfd_unref_process(p); + if (old_poison) + return; + + switch (client_id) { + case SOC15_IH_CLIENTID_SE0SH: + case SOC15_IH_CLIENTID_SE1SH: + case SOC15_IH_CLIENTID_SE2SH: + case SOC15_IH_CLIENTID_SE3SH: + case SOC15_IH_CLIENTID_UTCL2: + ret = kfd_dqm_evict_pasid(dev->dqm, pasid); + break; + case SOC15_IH_CLIENTID_SDMA0: + case SOC15_IH_CLIENTID_SDMA1: + case SOC15_IH_CLIENTID_SDMA2: + case SOC15_IH_CLIENTID_SDMA3: + case SOC15_IH_CLIENTID_SDMA4: + break; + default: + break; + } + + kfd_signal_poison_consumed_event(dev, pasid); + + /* resetting queue passes, do page retirement without gpu reset + * resetting queue fails, fallback to gpu reset solution + */ + if (!ret) { + dev_warn(dev->adev->dev, + "RAS poison consumption, unmap queue flow succeeded: client id %d\n", + client_id); + amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, false); + } else { + dev_warn(dev->adev->dev, + "RAS poison consumption, fall back to gpu reset flow: client id %d\n", + client_id); + amdgpu_amdkfd_ras_poison_consumption_handler(dev->adev, true); + } +} + +static bool event_interrupt_isr_v10(struct kfd_node *dev, + const uint32_t *ih_ring_entry, + uint32_t *patched_ihre, + bool *patched_flag) +{ + uint16_t source_id, client_id, pasid, vmid; + const uint32_t *data = ih_ring_entry; + + source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); + client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); + + /* Only handle interrupts from KFD VMIDs */ + vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); + if (!KFD_IRQ_IS_FENCE(client_id, source_id) && + (vmid < dev->vm_info.first_vmid_kfd || + vmid > dev->vm_info.last_vmid_kfd)) + return false; + + pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); + + /* Only handle clients we care about */ + if (client_id != SOC15_IH_CLIENTID_GRBM_CP && + client_id != SOC15_IH_CLIENTID_SDMA0 && + client_id != SOC15_IH_CLIENTID_SDMA1 && + client_id != SOC15_IH_CLIENTID_SDMA2 && + client_id != SOC15_IH_CLIENTID_SDMA3 && + client_id != SOC15_IH_CLIENTID_SDMA4 && + client_id != SOC15_IH_CLIENTID_SDMA5 && + client_id != SOC15_IH_CLIENTID_SDMA6 && + client_id != SOC15_IH_CLIENTID_SDMA7 && + client_id != SOC15_IH_CLIENTID_VMC && + client_id != SOC15_IH_CLIENTID_VMC1 && + client_id != SOC15_IH_CLIENTID_UTCL2 && + client_id != SOC15_IH_CLIENTID_SE0SH && + client_id != SOC15_IH_CLIENTID_SE1SH && + client_id != SOC15_IH_CLIENTID_SE2SH && + client_id != SOC15_IH_CLIENTID_SE3SH) + return false; + + pr_debug("client id 0x%x, source id %d, vmid %d, pasid 0x%x. raw data:\n", + client_id, source_id, vmid, pasid); + pr_debug("%8X, %8X, %8X, %8X, %8X, %8X, %8X, %8X.\n", + data[0], data[1], data[2], data[3], + data[4], data[5], data[6], data[7]); + + /* If there is no valid PASID, it's likely a bug */ + if (WARN_ONCE(pasid == 0, "Bug: No PASID in KFD interrupt")) + return 0; + + /* Interrupt types we care about: various signals and faults. + * They will be forwarded to a work queue (see below). + */ + return source_id == SOC15_INTSRC_CP_END_OF_PIPE || + source_id == SOC15_INTSRC_SDMA_TRAP || + source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG || + source_id == SOC15_INTSRC_CP_BAD_OPCODE || + client_id == SOC15_IH_CLIENTID_VMC || + client_id == SOC15_IH_CLIENTID_VMC1 || + client_id == SOC15_IH_CLIENTID_UTCL2 || + KFD_IRQ_IS_FENCE(client_id, source_id); +} + +static void event_interrupt_wq_v10(struct kfd_node *dev, + const uint32_t *ih_ring_entry) +{ + uint16_t source_id, client_id, pasid, vmid; + uint32_t context_id0, context_id1; + uint32_t encoding, sq_intr_err_type; + + source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); + client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); + pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); + vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); + context_id0 = SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); + context_id1 = SOC15_CONTEXT_ID1_FROM_IH_ENTRY(ih_ring_entry); + + if (client_id == SOC15_IH_CLIENTID_GRBM_CP || + client_id == SOC15_IH_CLIENTID_SE0SH || + client_id == SOC15_IH_CLIENTID_SE1SH || + client_id == SOC15_IH_CLIENTID_SE2SH || + client_id == SOC15_IH_CLIENTID_SE3SH) { + if (source_id == SOC15_INTSRC_CP_END_OF_PIPE) + kfd_signal_event_interrupt(pasid, context_id0, 32); + else if (source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG) { + encoding = REG_GET_FIELD(context_id1, + SQ_INTERRUPT_WORD_WAVE_CTXID1, ENCODING); + switch (encoding) { + case SQ_INTERRUPT_WORD_ENCODING_AUTO: + pr_debug( + "sq_intr: auto, se %d, ttrace %d, wlt %d, ttrac_buf0_full %d, ttrac_buf1_full %d, ttrace_utc_err %d\n", + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_AUTO_CTXID1, + SE_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, + THREAD_TRACE), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, + WLT), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, + THREAD_TRACE_BUF0_FULL), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, + THREAD_TRACE_BUF1_FULL), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID0, + THREAD_TRACE_UTC_ERROR)); + break; + case SQ_INTERRUPT_WORD_ENCODING_INST: + pr_debug("sq_intr: inst, se %d, data 0x%x, sa %d, priv %d, wave_id %d, simd_id %d, wgp_id %d\n", + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_WAVE_CTXID1, + SE_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + DATA), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + SA_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + PRIV), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + WAVE_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + SIMD_ID), + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_WAVE_CTXID1, + WGP_ID)); + if (context_id0 & SQ_INTERRUPT_WORD_WAVE_CTXID0__PRIV_MASK) { + if (kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_DEBUG_DOORBELL_ID(context_id0), + KFD_DEBUG_TRAP_CODE(context_id0), + NULL, 0)) + return; + } + break; + case SQ_INTERRUPT_WORD_ENCODING_ERROR: + sq_intr_err_type = REG_GET_FIELD(context_id0, KFD_CTXID0, + ERR_TYPE); + pr_warn("sq_intr: error, se %d, data 0x%x, sa %d, priv %d, wave_id %d, simd_id %d, wgp_id %d, err_type %d\n", + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_WAVE_CTXID1, + SE_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + DATA), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + SA_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + PRIV), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + WAVE_ID), + REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID0, + SIMD_ID), + REG_GET_FIELD(context_id1, SQ_INTERRUPT_WORD_WAVE_CTXID1, + WGP_ID), + sq_intr_err_type); + if (sq_intr_err_type != SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST && + sq_intr_err_type != SQ_INTERRUPT_ERROR_TYPE_MEMVIOL) { + event_interrupt_poison_consumption(dev, pasid, source_id); + return; + } + break; + default: + break; + } + kfd_signal_event_interrupt(pasid, context_id0 & 0x7fffff, 23); + } else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) { + kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_DEBUG_DOORBELL_ID(context_id0), + KFD_EC_MASK(KFD_DEBUG_CP_BAD_OP_ECODE(context_id0)), + NULL, + 0); + } + } else if (client_id == SOC15_IH_CLIENTID_SDMA0 || + client_id == SOC15_IH_CLIENTID_SDMA1 || + client_id == SOC15_IH_CLIENTID_SDMA2 || + client_id == SOC15_IH_CLIENTID_SDMA3 || + (client_id == SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid && + KFD_GC_VERSION(dev) == IP_VERSION(10, 3, 0)) || + client_id == SOC15_IH_CLIENTID_SDMA4 || + client_id == SOC15_IH_CLIENTID_SDMA5 || + client_id == SOC15_IH_CLIENTID_SDMA6 || + client_id == SOC15_IH_CLIENTID_SDMA7) { + if (source_id == SOC15_INTSRC_SDMA_TRAP) { + kfd_signal_event_interrupt(pasid, context_id0 & 0xfffffff, 28); + } else if (source_id == SOC15_INTSRC_SDMA_ECC) { + event_interrupt_poison_consumption(dev, pasid, source_id); + return; + } + } else if (client_id == SOC15_IH_CLIENTID_VMC || + client_id == SOC15_IH_CLIENTID_VMC1 || + client_id == SOC15_IH_CLIENTID_UTCL2) { + struct kfd_vm_fault_info info = {0}; + uint16_t ring_id = SOC15_RING_ID_FROM_IH_ENTRY(ih_ring_entry); + struct kfd_hsa_memory_exception_data exception_data; + + if (client_id == SOC15_IH_CLIENTID_UTCL2 && + amdgpu_amdkfd_ras_query_utcl2_poison_status(dev->adev)) { + event_interrupt_poison_consumption(dev, pasid, client_id); + return; + } + + info.vmid = vmid; + info.mc_id = client_id; + info.page_addr = ih_ring_entry[4] | + (uint64_t)(ih_ring_entry[5] & 0xf) << 32; + info.prot_valid = ring_id & 0x08; + info.prot_read = ring_id & 0x10; + info.prot_write = ring_id & 0x20; + + memset(&exception_data, 0, sizeof(exception_data)); + exception_data.gpu_id = dev->id; + exception_data.va = (info.page_addr) << PAGE_SHIFT; + exception_data.failure.NotPresent = info.prot_valid ? 1 : 0; + exception_data.failure.NoExecute = info.prot_exec ? 1 : 0; + exception_data.failure.ReadOnly = info.prot_write ? 1 : 0; + exception_data.failure.imprecise = 0; + + kfd_set_dbg_ev_from_interrupt(dev, + pasid, + -1, + KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION), + &exception_data, + sizeof(exception_data)); + } else if (KFD_IRQ_IS_FENCE(client_id, source_id)) { + kfd_process_close_interrupt_drain(pasid); + } +} + +const struct kfd_event_interrupt_class event_interrupt_class_v10 = { + .interrupt_isr = event_interrupt_isr_v10, + .interrupt_wq = event_interrupt_wq_v10, +}; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c index c2166bf964ef..f933bd231fb9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c @@ -26,6 +26,7 @@ #include "kfd_device_queue_manager.h" #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" #include "kfd_smi_events.h" +#include "kfd_debug.h" /* * GFX11 SQ Interrupts @@ -238,7 +239,7 @@ static bool event_interrupt_isr_v11(struct kfd_node *dev, client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); /* Only handle interrupts from KFD VMIDs */ vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); - if (/*!KFD_IRQ_IS_FENCE(client_id, source_id) &&*/ + if (!KFD_IRQ_IS_FENCE(client_id, source_id) && (vmid < dev->vm_info.first_vmid_kfd || vmid > dev->vm_info.last_vmid_kfd)) return false; @@ -267,7 +268,7 @@ static bool event_interrupt_isr_v11(struct kfd_node *dev, source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG || source_id == SOC15_INTSRC_CP_BAD_OPCODE || source_id == SOC21_INTSRC_SDMA_TRAP || - /* KFD_IRQ_IS_FENCE(client_id, source_id) || */ + KFD_IRQ_IS_FENCE(client_id, source_id) || (((client_id == SOC21_IH_CLIENTID_VMC) || ((client_id == SOC21_IH_CLIENTID_GFX) && (source_id == UTCL2_1_0__SRCID__FAULT))) && @@ -279,7 +280,7 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, { uint16_t source_id, client_id, ring_id, pasid, vmid; uint32_t context_id0, context_id1; - uint8_t sq_int_enc, sq_int_errtype; + uint8_t sq_int_enc, sq_int_priv, sq_int_errtype; struct kfd_vm_fault_info info = {0}; struct kfd_hsa_memory_exception_data exception_data; @@ -312,9 +313,9 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, exception_data.failure.ReadOnly = info.prot_write ? 1 : 0; exception_data.failure.imprecise = 0; - /*kfd_set_dbg_ev_from_interrupt(dev, pasid, -1, + kfd_set_dbg_ev_from_interrupt(dev, pasid, -1, KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION), - &exception_data, sizeof(exception_data));*/ + &exception_data, sizeof(exception_data)); kfd_smi_event_update_vmfault(dev, pasid); /* GRBM, SDMA, SE, PMM */ @@ -324,11 +325,11 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, /* CP */ if (source_id == SOC15_INTSRC_CP_END_OF_PIPE) kfd_signal_event_interrupt(pasid, context_id0, 32); - /*else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) + else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) kfd_set_dbg_ev_from_interrupt(dev, pasid, KFD_CTXID0_DOORBELL_ID(context_id0), KFD_EC_MASK(KFD_CTXID0_CP_BAD_OP_ECODE(context_id0)), - NULL, 0);*/ + NULL, 0); /* SDMA */ else if (source_id == SOC21_INTSRC_SDMA_TRAP) @@ -348,6 +349,13 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, break; case SQ_INTERRUPT_WORD_ENCODING_INST: print_sq_intr_info_inst(context_id0, context_id1); + sq_int_priv = REG_GET_FIELD(context_id0, + SQ_INTERRUPT_WORD_WAVE_CTXID0, PRIV); + if (sq_int_priv && (kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_CTXID0_DOORBELL_ID(context_id0), + KFD_CTXID0_TRAP_CODE(context_id0), + NULL, 0))) + return; break; case SQ_INTERRUPT_WORD_ENCODING_ERROR: print_sq_intr_info_error(context_id0, context_id1); @@ -366,8 +374,8 @@ static void event_interrupt_wq_v11(struct kfd_node *dev, kfd_signal_event_interrupt(pasid, context_id0 & 0xffffff, 24); } - /*} else if (KFD_IRQ_IS_FENCE(client_id, source_id)) { - kfd_process_close_interrupt_drain(pasid);*/ + } else if (KFD_IRQ_IS_FENCE(client_id, source_id)) { + kfd_process_close_interrupt_drain(pasid); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c index 8cf58be80f4e..d5c9f30552e3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c @@ -23,10 +23,40 @@ #include "kfd_priv.h" #include "kfd_events.h" +#include "kfd_debug.h" #include "soc15_int.h" #include "kfd_device_queue_manager.h" #include "kfd_smi_events.h" +/* + * GFX9 SQ Interrupts + * + * There are 3 encoding types of interrupts sourced from SQ sent as a 44-bit + * packet to the Interrupt Handler: + * Auto - Generated by the SQG (various cmd overflows, timestamps etc) + * Wave - Generated by S_SENDMSG through a shader program + * Error - HW generated errors (Illegal instructions, Memviols, EDC etc) + * + * The 44-bit packet is mapped as {context_id1[7:0],context_id0[31:0]} plus + * 4-bits for VMID (SOC15_VMID_FROM_IH_ENTRY) as such: + * + * - context_id0[27:26] + * Encoding type (0 = Auto, 1 = Wave, 2 = Error) + * + * - context_id0[13] + * PRIV bit indicates that Wave S_SEND or error occurred within trap + * + * - {context_id1[7:0],context_id0[31:28],context_id0[11:0]} + * 24-bit data with the following layout per encoding type: + * Auto - only context_id0[8:0] is used, which reports various interrupts + * generated by SQG. The rest is 0. + * Wave - user data sent from m0 via S_SENDMSG + * Error - Error type (context_id1[7:4]), Error Details (rest of bits) + * + * The other context_id bits show coordinates (SE/SH/CU/SIMD/WAVE) for wave + * S_SENDMSG and Errors. These are 0 for Auto. + */ + enum SQ_INTERRUPT_WORD_ENCODING { SQ_INTERRUPT_WORD_ENCODING_AUTO = 0x0, SQ_INTERRUPT_WORD_ENCODING_INST, @@ -84,12 +114,32 @@ enum SQ_INTERRUPT_ERROR_TYPE { #define SQ_INTERRUPT_WORD_WAVE_CTXID__SE_ID_MASK 0x03000000 #define SQ_INTERRUPT_WORD_WAVE_CTXID__ENCODING_MASK 0x0c000000 +/* GFX9 SQ interrupt 24-bit data from context_id<0,1> */ #define KFD_CONTEXT_ID_GET_SQ_INT_DATA(ctx0, ctx1) \ ((ctx0 & 0xfff) | ((ctx0 >> 16) & 0xf000) | ((ctx1 << 16) & 0xff0000)) #define KFD_SQ_INT_DATA__ERR_TYPE_MASK 0xF00000 #define KFD_SQ_INT_DATA__ERR_TYPE__SHIFT 20 +/* + * The debugger will send user data(m0) with PRIV=1 to indicate it requires + * notification from the KFD with the following queue id (DOORBELL_ID) and + * trap code (TRAP_CODE). + */ +#define KFD_INT_DATA_DEBUG_DOORBELL_MASK 0x0003ff +#define KFD_INT_DATA_DEBUG_TRAP_CODE_SHIFT 10 +#define KFD_INT_DATA_DEBUG_TRAP_CODE_MASK 0x07fc00 +#define KFD_DEBUG_DOORBELL_ID(sq_int_data) ((sq_int_data) & \ + KFD_INT_DATA_DEBUG_DOORBELL_MASK) +#define KFD_DEBUG_TRAP_CODE(sq_int_data) (((sq_int_data) & \ + KFD_INT_DATA_DEBUG_TRAP_CODE_MASK) \ + >> KFD_INT_DATA_DEBUG_TRAP_CODE_SHIFT) +#define KFD_DEBUG_CP_BAD_OP_ECODE_MASK 0x3fffc00 +#define KFD_DEBUG_CP_BAD_OP_ECODE_SHIFT 10 +#define KFD_DEBUG_CP_BAD_OP_ECODE(ctxid0) (((ctxid0) & \ + KFD_DEBUG_CP_BAD_OP_ECODE_MASK) \ + >> KFD_DEBUG_CP_BAD_OP_ECODE_SHIFT) + static void event_interrupt_poison_consumption_v9(struct kfd_node *dev, uint16_t pasid, uint16_t client_id) { @@ -168,14 +218,16 @@ static bool event_interrupt_isr_v9(struct kfd_node *dev, uint16_t source_id, client_id, pasid, vmid; const uint32_t *data = ih_ring_entry; + source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); + client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); + /* Only handle interrupts from KFD VMIDs */ vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); - if (vmid < dev->vm_info.first_vmid_kfd || - vmid > dev->vm_info.last_vmid_kfd) + if (!KFD_IRQ_IS_FENCE(client_id, source_id) && + (vmid < dev->vm_info.first_vmid_kfd || + vmid > dev->vm_info.last_vmid_kfd)) return false; - source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); - client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); /* Only handle clients we care about */ @@ -194,7 +246,8 @@ static bool event_interrupt_isr_v9(struct kfd_node *dev, client_id != SOC15_IH_CLIENTID_SE0SH && client_id != SOC15_IH_CLIENTID_SE1SH && client_id != SOC15_IH_CLIENTID_SE2SH && - client_id != SOC15_IH_CLIENTID_SE3SH) + client_id != SOC15_IH_CLIENTID_SE3SH && + !KFD_IRQ_IS_FENCE(client_id, source_id)) return false; /* This is a known issue for gfx9. Under non HWS, pasid is not set @@ -247,6 +300,7 @@ static bool event_interrupt_isr_v9(struct kfd_node *dev, source_id == SOC15_INTSRC_SDMA_ECC || source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG || source_id == SOC15_INTSRC_CP_BAD_OPCODE || + KFD_IRQ_IS_FENCE(client_id, source_id) || ((client_id == SOC15_IH_CLIENTID_VMC || client_id == SOC15_IH_CLIENTID_VMC1 || client_id == SOC15_IH_CLIENTID_UTCL2) && @@ -302,6 +356,13 @@ static void event_interrupt_wq_v9(struct kfd_node *dev, REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SIMD_ID), REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, CU_ID), sq_int_data); + if (context_id0 & SQ_INTERRUPT_WORD_WAVE_CTXID__PRIV_MASK) { + if (kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_DEBUG_DOORBELL_ID(sq_int_data), + KFD_DEBUG_TRAP_CODE(sq_int_data), + NULL, 0)) + return; + } break; case SQ_INTERRUPT_WORD_ENCODING_ERROR: sq_intr_err = REG_GET_FIELD(sq_int_data, KFD_SQ_INT_DATA, ERR_TYPE); @@ -324,8 +385,12 @@ static void event_interrupt_wq_v9(struct kfd_node *dev, break; } kfd_signal_event_interrupt(pasid, context_id0 & 0xffffff, 24); - } else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) - kfd_signal_hw_exception_event(pasid); + } else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE) { + kfd_set_dbg_ev_from_interrupt(dev, pasid, + KFD_DEBUG_DOORBELL_ID(context_id0), + KFD_EC_MASK(KFD_DEBUG_CP_BAD_OP_ECODE(context_id0)), + NULL, 0); + } } else if (client_id == SOC15_IH_CLIENTID_SDMA0 || client_id == SOC15_IH_CLIENTID_SDMA1 || client_id == SOC15_IH_CLIENTID_SDMA2 || @@ -345,6 +410,7 @@ static void event_interrupt_wq_v9(struct kfd_node *dev, client_id == SOC15_IH_CLIENTID_UTCL2) { struct kfd_vm_fault_info info = {0}; uint16_t ring_id = SOC15_RING_ID_FROM_IH_ENTRY(ih_ring_entry); + struct kfd_hsa_memory_exception_data exception_data; if (client_id == SOC15_IH_CLIENTID_UTCL2 && amdgpu_amdkfd_ras_query_utcl2_poison_status(dev->adev)) { @@ -360,9 +426,23 @@ static void event_interrupt_wq_v9(struct kfd_node *dev, info.prot_read = ring_id & 0x10; info.prot_write = ring_id & 0x20; + memset(&exception_data, 0, sizeof(exception_data)); + exception_data.gpu_id = dev->id; + exception_data.va = (info.page_addr) << PAGE_SHIFT; + exception_data.failure.NotPresent = info.prot_valid ? 1 : 0; + exception_data.failure.NoExecute = info.prot_exec ? 1 : 0; + exception_data.failure.ReadOnly = info.prot_write ? 1 : 0; + exception_data.failure.imprecise = 0; + + kfd_set_dbg_ev_from_interrupt(dev, + pasid, + -1, + KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION), + &exception_data, + sizeof(exception_data)); kfd_smi_event_update_vmfault(dev, pasid); - kfd_dqm_evict_pasid(dev->dqm, pasid); - kfd_signal_vm_fault_event(dev, pasid, &info, NULL); + } else if (KFD_IRQ_IS_FENCE(client_id, source_id)) { + kfd_process_close_interrupt_drain(pasid); } } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index a02fb939614a..cd2d56e5cdf0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -963,6 +963,10 @@ struct kfd_process { uint64_t exception_enable_mask; uint64_t exception_status; + /* Used to drain stale interrupts */ + wait_queue_head_t wait_irq_drain; + bool irq_drain_is_open; + /* shared virtual memory registered by this process */ struct svm_range_list svms; @@ -1144,12 +1148,19 @@ int kfd_numa_node_to_apic_id(int numa_node_id); void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); /* Interrupts */ +#define KFD_IRQ_FENCE_CLIENTID 0xff +#define KFD_IRQ_FENCE_SOURCEID 0xff +#define KFD_IRQ_IS_FENCE(client, source) \ + ((client) == KFD_IRQ_FENCE_CLIENTID && \ + (source) == KFD_IRQ_FENCE_SOURCEID) int kfd_interrupt_init(struct kfd_node *dev); void kfd_interrupt_exit(struct kfd_node *dev); bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry); bool interrupt_is_wanted(struct kfd_node *dev, const uint32_t *ih_ring_entry, uint32_t *patched_ihre, bool *flag); +int kfd_process_drain_interrupts(struct kfd_process_device *pdd); +void kfd_process_close_interrupt_drain(unsigned int pasid); /* amdkfd Apertures */ int kfd_init_apertures(struct kfd_process *process); @@ -1421,6 +1432,7 @@ uint64_t kfd_get_number_elems(struct kfd_dev *kfd); /* Events */ extern const struct kfd_event_interrupt_class event_interrupt_class_cik; extern const struct kfd_event_interrupt_class event_interrupt_class_v9; +extern const struct kfd_event_interrupt_class event_interrupt_class_v10; extern const struct kfd_event_interrupt_class event_interrupt_class_v11; extern const struct kfd_device_global_init_class device_global_init_class_cik; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 3b7f219c9d06..3d3611705d41 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -862,6 +862,8 @@ struct kfd_process *kfd_create_process(struct task_struct *thread) kfd_procfs_add_sysfs_stats(process); kfd_procfs_add_sysfs_files(process); kfd_procfs_add_sysfs_counters(process); + + init_waitqueue_head(&process->wait_irq_drain); } out: if (!IS_ERR(process)) @@ -2136,6 +2138,51 @@ void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type) } } +/* assumes caller holds process lock. */ +int kfd_process_drain_interrupts(struct kfd_process_device *pdd) +{ + uint32_t irq_drain_fence[8]; + int r = 0; + + if (!KFD_IS_SOC15(pdd->dev)) + return 0; + + pdd->process->irq_drain_is_open = true; + + memset(irq_drain_fence, 0, sizeof(irq_drain_fence)); + irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) | + KFD_IRQ_FENCE_CLIENTID; + irq_drain_fence[3] = pdd->process->pasid; + + /* ensure stale irqs scheduled KFD interrupts and send drain fence. */ + if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev, + irq_drain_fence)) { + pdd->process->irq_drain_is_open = false; + return 0; + } + + r = wait_event_interruptible(pdd->process->wait_irq_drain, + !READ_ONCE(pdd->process->irq_drain_is_open)); + if (r) + pdd->process->irq_drain_is_open = false; + + return r; +} + +void kfd_process_close_interrupt_drain(unsigned int pasid) +{ + struct kfd_process *p; + + p = kfd_lookup_process_by_pasid(pasid); + + if (!p) + return; + + WRITE_ONCE(p->irq_drain_is_open, false); + wake_up_all(&p->wait_irq_drain); + kfd_unref_process(p); +} + struct send_exception_work_handler_workarea { struct work_struct work; struct kfd_process *p; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 43d432b5c5bc..70852a200d8f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -337,6 +337,10 @@ int pqm_create_queue(struct process_queue_manager *pqm, kq->queue->properties.queue_id = *qid; pqn->kq = kq; pqn->q = NULL; + retval = kfd_process_drain_interrupts(pdd); + if (retval) + break; + retval = dev->dqm->ops.create_kernel_queue(dev->dqm, kq, &pdd->qpd); break; -- cgit v1.2.3 From e90bf919f714ae2a658cdfd03238e7be9ce9185c Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Wed, 27 Apr 2022 10:24:37 -0400 Subject: drm/amdkfd: add debug set exceptions enabled operation The debugger subscibes to nofication for requested exceptions on attach. Allow the debugger to change its subsciption later on. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 3 +++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 36 ++++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 2 ++ 3 files changed, 41 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 615fa9ab36b7..dcf4b5f3886e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2980,6 +2980,9 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v args->send_runtime_event.exception_mask); break; case KFD_IOC_DBG_TRAP_SET_EXCEPTIONS_ENABLED: + kfd_dbg_set_enabled_debug_exception_mask(target, + args->set_exceptions_enabled.exception_mask); + break; case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE: case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE: case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 68b657398d41..48a4e3cc2234 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -521,3 +521,39 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, return r; } + +void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, + uint64_t exception_set_mask) +{ + uint64_t found_mask = 0; + struct process_queue_manager *pqm; + struct process_queue_node *pqn; + static const char write_data = '.'; + loff_t pos = 0; + int i; + + mutex_lock(&target->event_mutex); + + found_mask |= target->exception_status; + + pqm = &target->pqm; + list_for_each_entry(pqn, &pqm->queues, process_queue_list) { + if (!pqn) + continue; + + found_mask |= pqn->q->properties.exception_status; + } + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + found_mask |= pdd->exception_status; + } + + if (exception_set_mask & found_mask) + kernel_write(target->dbg_ev_file, &write_data, 1, &pos); + + target->exception_enable_mask = exception_set_mask; + + mutex_unlock(&target->event_mutex); +} diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 5153ccbd7fd1..6c1054a08872 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -56,6 +56,8 @@ static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) void debug_event_write_work_handler(struct work_struct *work); +void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, + uint64_t exception_set_mask); /* * If GFX off is enabled, chips that do not support RLC restore for the debug * registers will disable GFX off temporarily for the entire debug session. -- cgit v1.2.3 From 101827e13026a981e887527620fe9710adc0e481 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Wed, 27 Apr 2022 13:18:10 -0400 Subject: drm/amdkfd: add debug wave launch override operation This operation allows the debugger to override the enabled HW exceptions on the device. On debug devices that only support the debugging of a single process, the HW exceptions are global and set through the SPI_GDBG_TRAP_MASK register. Because they are global, only address watch exceptions are allowed to be enabled. In other words, the debugger must preserve all non-address watch exception states in normal mode operation by barring a full replacement override or a non-address watch override request. For multi-process debugging, all HW exception overrides are per-VMID so all exceptions can be overridden or fully replaced. In order for the debugger to know what is permissible, returned the supported override mask back to the debugger along with the previously enable overrides. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c | 47 ++++++++++++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 55 ++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h | 10 +++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c | 5 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c | 87 +++++++++++++++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 55 ++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 10 +++ drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 ++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 69 +++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 6 ++ 11 files changed, 351 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c index b811a0985050..d7881bbd828d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c @@ -25,6 +25,7 @@ #include "amdgpu_amdkfd_gfx_v9.h" #include "gc/gc_9_4_2_offset.h" #include "gc/gc_9_4_2_sh_mask.h" +#include /* * Returns TRAP_EN, EXCP_EN and EXCP_REPLACE. @@ -62,6 +63,50 @@ static uint32_t kgd_aldebaran_disable_debug_trap(struct amdgpu_device *adev, return data; } +static int kgd_aldebaran_validate_trap_override_request(struct amdgpu_device *adev, + uint32_t trap_override, + uint32_t *trap_mask_supported) +{ + *trap_mask_supported &= KFD_DBG_TRAP_MASK_FP_INVALID | + KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL | + KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO | + KFD_DBG_TRAP_MASK_FP_OVERFLOW | + KFD_DBG_TRAP_MASK_FP_UNDERFLOW | + KFD_DBG_TRAP_MASK_FP_INEXACT | + KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO | + KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH | + KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION; + + if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR && + trap_override != KFD_DBG_TRAP_OVERRIDE_REPLACE) + return -EPERM; + + return 0; +} + +/* returns TRAP_EN, EXCP_EN and EXCP_RPLACE. */ +static uint32_t kgd_aldebaran_set_wave_launch_trap_override(struct amdgpu_device *adev, + uint32_t vmid, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t kfd_dbg_trap_cntl_prev) + +{ + uint32_t data = 0; + + *trap_mask_prev = REG_GET_FIELD(kfd_dbg_trap_cntl_prev, SPI_GDBG_PER_VMID_CNTL, EXCP_EN); + trap_mask_bits = (trap_mask_bits & trap_mask_request) | + (*trap_mask_prev & ~trap_mask_request); + + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1); + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, trap_mask_bits); + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, trap_override); + + return data; +} + const struct kfd2kgd_calls aldebaran_kfd2kgd = { .program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings, .set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping, @@ -82,6 +127,8 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = { .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .enable_debug_trap = kgd_aldebaran_enable_debug_trap, .disable_debug_trap = kgd_aldebaran_disable_debug_trap, + .validate_trap_override_request = kgd_aldebaran_validate_trap_override_request, + .set_wave_launch_trap_override = kgd_aldebaran_set_wave_launch_trap_override, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c index a62bd0068515..ec2587664001 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c @@ -410,6 +410,8 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = { kgd_gfx_v9_set_vm_context_page_table_base, .enable_debug_trap = kgd_arcturus_enable_debug_trap, .disable_debug_trap = kgd_arcturus_disable_debug_trap, + .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request, + .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index 98006c7021dd..7ea0362dcab3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -32,6 +32,7 @@ #include "v10_structs.h" #include "nv.h" #include "nvd.h" +#include enum hqd_dequeue_request_type { NO_ACTION = 0, @@ -803,6 +804,58 @@ uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev, return 0; } +int kgd_gfx_v10_validate_trap_override_request(struct amdgpu_device *adev, + uint32_t trap_override, + uint32_t *trap_mask_supported) +{ + *trap_mask_supported &= KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH; + + /* The SPI_GDBG_TRAP_MASK register is global and affects all + * processes. Only allow OR-ing the address-watch bit, since + * this only affects processes under the debugger. Other bits + * should stay 0 to avoid the debugger interfering with other + * processes. + */ + if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR) + return -EINVAL; + + return 0; +} + +uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev, + uint32_t vmid, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t kfd_dbg_trap_cntl_prev) +{ + uint32_t data, wave_cntl_prev; + + mutex_lock(&adev->grbm_idx_mutex); + + wave_cntl_prev = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL)); + + kgd_gfx_v10_set_wave_launch_stall(adev, vmid, true); + + data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK)); + *trap_mask_prev = REG_GET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN); + + trap_mask_bits = (trap_mask_bits & trap_mask_request) | + (*trap_mask_prev & ~trap_mask_request); + + data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN, trap_mask_bits); + data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, REPLACE, trap_override); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), data); + + /* We need to preserve wave launch mode stall settings. */ + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), wave_cntl_prev); + + mutex_unlock(&adev->grbm_idx_mutex); + + return 0; +} + /* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values * The values read are: * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. @@ -889,6 +942,8 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = { .set_vm_context_page_table_base = set_vm_context_page_table_base, .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, + .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request, + .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override, .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .program_trap_handler_settings = program_trap_handler_settings, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h index 1e993a213646..57339fa12807 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h @@ -26,6 +26,16 @@ uint32_t kgd_gfx_v10_enable_debug_trap(struct amdgpu_device *adev, uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev, bool keep_trap_enabled, uint32_t vmid); +int kgd_gfx_v10_validate_trap_override_request(struct amdgpu_device *adev, + uint32_t trap_override, + uint32_t *trap_mask_supported); +uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev, + uint32_t vmid, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t kfd_dbg_trap_cntl_prev); void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, uint32_t wait_times, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c index 387bdf4823c9..7120927fed15 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c @@ -675,5 +675,8 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = { .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, - .disable_debug_trap = kgd_gfx_v10_disable_debug_trap + .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, + .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request, + .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override + }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c index cc954cf248ca..ae0c4707919f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c @@ -30,6 +30,7 @@ #include "soc15d.h" #include "v11_structs.h" #include "soc21.h" +#include enum hqd_dequeue_request_type { NO_ACTION = 0, @@ -643,6 +644,88 @@ static uint32_t kgd_gfx_v11_disable_debug_trap(struct amdgpu_device *adev, return data; } +static int kgd_gfx_v11_validate_trap_override_request(struct amdgpu_device *adev, + uint32_t trap_override, + uint32_t *trap_mask_supported) +{ + *trap_mask_supported &= KFD_DBG_TRAP_MASK_FP_INVALID | + KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL | + KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO | + KFD_DBG_TRAP_MASK_FP_OVERFLOW | + KFD_DBG_TRAP_MASK_FP_UNDERFLOW | + KFD_DBG_TRAP_MASK_FP_INEXACT | + KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO | + KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH | + KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION; + + if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 4)) + *trap_mask_supported |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START | + KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END; + + if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR && + trap_override != KFD_DBG_TRAP_OVERRIDE_REPLACE) + return -EPERM; + + return 0; +} + +static uint32_t trap_mask_map_sw_to_hw(uint32_t mask) +{ + uint32_t trap_on_start = (mask & KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START) ? 1 : 0; + uint32_t trap_on_end = (mask & KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END) ? 1 : 0; + uint32_t excp_en = mask & (KFD_DBG_TRAP_MASK_FP_INVALID | + KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL | + KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO | + KFD_DBG_TRAP_MASK_FP_OVERFLOW | + KFD_DBG_TRAP_MASK_FP_UNDERFLOW | + KFD_DBG_TRAP_MASK_FP_INEXACT | + KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO | + KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH | + KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION); + uint32_t ret; + + ret = REG_SET_FIELD(0, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, excp_en); + ret = REG_SET_FIELD(ret, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_START, trap_on_start); + ret = REG_SET_FIELD(ret, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_END, trap_on_end); + + return ret; +} + +static uint32_t trap_mask_map_hw_to_sw(uint32_t mask) +{ + uint32_t ret = REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, EXCP_EN); + + if (REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_START)) + ret |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START; + + if (REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_END)) + ret |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END; + + return ret; +} + +/* Returns TRAP_EN, EXCP_EN and EXCP_REPLACE. */ +static uint32_t kgd_gfx_v11_set_wave_launch_trap_override(struct amdgpu_device *adev, + uint32_t vmid, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t kfd_dbg_trap_cntl_prev) +{ + uint32_t data = 0; + + *trap_mask_prev = trap_mask_map_hw_to_sw(kfd_dbg_trap_cntl_prev); + + data = (trap_mask_bits & trap_mask_request) | (*trap_mask_prev & ~trap_mask_request); + data = trap_mask_map_sw_to_hw(data); + + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1); + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, trap_override); + + return data; +} + const struct kfd2kgd_calls gfx_v11_kfd2kgd = { .program_sh_mem_settings = program_sh_mem_settings_v11, .set_pasid_vmid_mapping = set_pasid_vmid_mapping_v11, @@ -660,5 +743,7 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = { .get_atc_vmid_pasid_mapping_info = NULL, .set_vm_context_page_table_base = set_vm_context_page_table_base_v11, .enable_debug_trap = kgd_gfx_v11_enable_debug_trap, - .disable_debug_trap = kgd_gfx_v11_disable_debug_trap + .disable_debug_trap = kgd_gfx_v11_disable_debug_trap, + .validate_trap_override_request = kgd_gfx_v11_validate_trap_override_request, + .set_wave_launch_trap_override = kgd_gfx_v11_set_wave_launch_trap_override }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index 829ee720cc44..ce5c256f0517 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -38,6 +38,7 @@ #include "soc15d.h" #include "gfx_v9_0.h" #include "amdgpu_amdkfd_gfx_v9.h" +#include enum hqd_dequeue_request_type { NO_ACTION = 0, @@ -739,6 +740,58 @@ uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, return 0; } +int kgd_gfx_v9_validate_trap_override_request(struct amdgpu_device *adev, + uint32_t trap_override, + uint32_t *trap_mask_supported) +{ + *trap_mask_supported &= KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH; + + /* The SPI_GDBG_TRAP_MASK register is global and affects all + * processes. Only allow OR-ing the address-watch bit, since + * this only affects processes under the debugger. Other bits + * should stay 0 to avoid the debugger interfering with other + * processes. + */ + if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR) + return -EINVAL; + + return 0; +} + +uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev, + uint32_t vmid, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t kfd_dbg_cntl_prev) +{ + uint32_t data, wave_cntl_prev; + + mutex_lock(&adev->grbm_idx_mutex); + + wave_cntl_prev = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL)); + + kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true); + + data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK)); + *trap_mask_prev = REG_GET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN); + + trap_mask_bits = (trap_mask_bits & trap_mask_request) | + (*trap_mask_prev & ~trap_mask_request); + + data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN, trap_mask_bits); + data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, REPLACE, trap_override); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), data); + + /* We need to preserve wave launch mode stall settings. */ + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), wave_cntl_prev); + + mutex_unlock(&adev->grbm_idx_mutex); + + return 0; +} + /* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values * The values read are: * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. @@ -1010,6 +1063,8 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = { .set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base, .enable_debug_trap = kgd_gfx_v9_enable_debug_trap, .disable_debug_trap = kgd_gfx_v9_disable_debug_trap, + .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request, + .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index fed5b7f18b1a..76812ddd35b1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -71,6 +71,16 @@ uint32_t kgd_gfx_v9_enable_debug_trap(struct amdgpu_device *adev, uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, bool keep_trap_enabled, uint32_t vmid); +int kgd_gfx_v9_validate_trap_override_request(struct amdgpu_device *adev, + uint32_t trap_override, + uint32_t *trap_mask_supported); +uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev, + uint32_t vmid, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t kfd_dbg_trap_cntl_prev); void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev, uint32_t wait_times, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index dcf4b5f3886e..7fa249807671 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2984,6 +2984,13 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v args->set_exceptions_enabled.exception_mask); break; case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE: + r = kfd_dbg_trap_set_wave_launch_override(target, + args->launch_override.override_mode, + args->launch_override.enable_mask, + args->launch_override.support_request_mask, + &args->launch_override.enable_mask, + &args->launch_override.support_request_mask); + break; case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE: case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES: case KFD_IOC_DBG_TRAP_RESUME_QUEUES: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 48a4e3cc2234..733390fb2459 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -522,6 +522,75 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, return r; } +static int kfd_dbg_validate_trap_override_request(struct kfd_process *p, + uint32_t trap_override, + uint32_t trap_mask_request, + uint32_t *trap_mask_supported) +{ + int i = 0; + + *trap_mask_supported = 0xffffffff; + + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + int err = pdd->dev->kfd2kgd->validate_trap_override_request( + pdd->dev->adev, + trap_override, + trap_mask_supported); + + if (err) + return err; + } + + if (trap_mask_request & ~*trap_mask_supported) + return -EACCES; + + return 0; +} + +int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t *trap_mask_supported) +{ + int r = 0, i; + + r = kfd_dbg_validate_trap_override_request(target, + trap_override, + trap_mask_request, + trap_mask_supported); + + if (r) + return r; + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + pdd->spi_dbg_override = pdd->dev->kfd2kgd->set_wave_launch_trap_override( + pdd->dev->adev, + pdd->dev->vm_info.last_vmid_kfd, + trap_override, + trap_mask_bits, + trap_mask_request, + trap_mask_prev, + pdd->spi_dbg_override); + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + r = debug_refresh_runlist(pdd->dev->dqm); + else + r = kfd_dbg_set_mes_debug_mode(pdd); + + if (r) + break; + } + + return r; +} + void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, uint64_t exception_set_mask) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 6c1054a08872..c9245221aa76 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -42,6 +42,12 @@ int kfd_dbg_trap_disable(struct kfd_process *target); int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, void __user *runtime_info, uint32_t *runtime_info_size); +int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target, + uint32_t trap_override, + uint32_t trap_mask_bits, + uint32_t trap_mask_request, + uint32_t *trap_mask_prev, + uint32_t *trap_mask_supported); int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, unsigned int dev_id, -- cgit v1.2.3 From aea1b4738bebd8092bd437ce0b03aa9587fc20a7 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 2 May 2022 11:45:05 -0400 Subject: drm/amdkfd: add debug wave launch mode operation Allow the debugger to set wave behaviour on to either normally operate, halt at launch, trap on every instruction, terminate immediately or stall on allocation. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c | 12 ++++++++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 25 +++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h | 3 ++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c | 14 ++++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 25 +++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 3 ++ drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 3 ++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 36 +++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 2 ++ 11 files changed, 124 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c index d7881bbd828d..774ecfc3451a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c @@ -107,6 +107,17 @@ static uint32_t kgd_aldebaran_set_wave_launch_trap_override(struct amdgpu_device return data; } +static uint32_t kgd_aldebaran_set_wave_launch_mode(struct amdgpu_device *adev, + uint8_t wave_launch_mode, + uint32_t vmid) +{ + uint32_t data = 0; + + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, LAUNCH_MODE, wave_launch_mode); + + return data; +} + const struct kfd2kgd_calls aldebaran_kfd2kgd = { .program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings, .set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping, @@ -129,6 +140,7 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = { .disable_debug_trap = kgd_aldebaran_disable_debug_trap, .validate_trap_override_request = kgd_aldebaran_validate_trap_override_request, .set_wave_launch_trap_override = kgd_aldebaran_set_wave_launch_trap_override, + .set_wave_launch_mode = kgd_aldebaran_set_wave_launch_mode, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c index ec2587664001..fbdc1b7b1e42 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c @@ -412,6 +412,7 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = { .disable_debug_trap = kgd_arcturus_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override, + .set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index 7ea0362dcab3..a7a6edda557f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -856,6 +856,30 @@ uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev, return 0; } +uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev, + uint8_t wave_launch_mode, + uint32_t vmid) +{ + uint32_t data = 0; + bool is_mode_set = !!wave_launch_mode; + + mutex_lock(&adev->grbm_idx_mutex); + + kgd_gfx_v10_set_wave_launch_stall(adev, vmid, true); + + data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2, + VMID_MASK, is_mode_set ? 1 << vmid : 0); + data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2, + MODE, is_mode_set ? wave_launch_mode : 0); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL2), data); + + kgd_gfx_v10_set_wave_launch_stall(adev, vmid, false); + + mutex_unlock(&adev->grbm_idx_mutex); + + return 0; +} + /* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values * The values read are: * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. @@ -944,6 +968,7 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = { .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override, + .set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode, .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .program_trap_handler_settings = program_trap_handler_settings, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h index 57339fa12807..3a6aca2b0eaa 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h @@ -36,6 +36,9 @@ uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev, uint32_t trap_mask_request, uint32_t *trap_mask_prev, uint32_t kfd_dbg_trap_cntl_prev); +uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev, + uint8_t wave_launch_mode, + uint32_t vmid); void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, uint32_t wait_times, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c index 7120927fed15..ed36b433a48b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c @@ -677,6 +677,7 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = { .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request, - .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override + .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override, + .set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c index ae0c4707919f..9711d5128d09 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c @@ -726,6 +726,17 @@ static uint32_t kgd_gfx_v11_set_wave_launch_trap_override(struct amdgpu_device * return data; } +static uint32_t kgd_gfx_v11_set_wave_launch_mode(struct amdgpu_device *adev, + uint8_t wave_launch_mode, + uint32_t vmid) +{ + uint32_t data = 0; + + data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, LAUNCH_MODE, wave_launch_mode); + + return data; +} + const struct kfd2kgd_calls gfx_v11_kfd2kgd = { .program_sh_mem_settings = program_sh_mem_settings_v11, .set_pasid_vmid_mapping = set_pasid_vmid_mapping_v11, @@ -745,5 +756,6 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = { .enable_debug_trap = kgd_gfx_v11_enable_debug_trap, .disable_debug_trap = kgd_gfx_v11_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v11_validate_trap_override_request, - .set_wave_launch_trap_override = kgd_gfx_v11_set_wave_launch_trap_override + .set_wave_launch_trap_override = kgd_gfx_v11_set_wave_launch_trap_override, + .set_wave_launch_mode = kgd_gfx_v11_set_wave_launch_mode }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index ce5c256f0517..0acc0c18dfe6 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -792,6 +792,30 @@ uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev, return 0; } +uint32_t kgd_gfx_v9_set_wave_launch_mode(struct amdgpu_device *adev, + uint8_t wave_launch_mode, + uint32_t vmid) +{ + uint32_t data = 0; + bool is_mode_set = !!wave_launch_mode; + + mutex_lock(&adev->grbm_idx_mutex); + + kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true); + + data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2, + VMID_MASK, is_mode_set ? 1 << vmid : 0); + data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2, + MODE, is_mode_set ? wave_launch_mode : 0); + WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL2), data); + + kgd_gfx_v9_set_wave_launch_stall(adev, vmid, false); + + mutex_unlock(&adev->grbm_idx_mutex); + + return 0; +} + /* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values * The values read are: * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. @@ -1065,6 +1089,7 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = { .disable_debug_trap = kgd_gfx_v9_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override, + .set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index 76812ddd35b1..18f4970ac8e4 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -74,6 +74,9 @@ uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, int kgd_gfx_v9_validate_trap_override_request(struct amdgpu_device *adev, uint32_t trap_override, uint32_t *trap_mask_supported); +uint32_t kgd_gfx_v9_set_wave_launch_mode(struct amdgpu_device *adev, + uint8_t wave_launch_mode, + uint32_t vmid); uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev, uint32_t vmid, uint32_t trap_override, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 7fa249807671..a6570b124b2b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2992,6 +2992,9 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v &args->launch_override.support_request_mask); break; case KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE: + r = kfd_dbg_trap_set_wave_launch_mode(target, + args->launch_mode.launch_mode); + break; case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES: case KFD_IOC_DBG_TRAP_RESUME_QUEUES: case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 733390fb2459..53c3418562d4 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -301,8 +301,10 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind { int i; - if (!unwind) + if (!unwind) { cancel_work_sync(&target->debug_event_workarea); + kfd_dbg_trap_set_wave_launch_mode(target, 0); + } for (i = 0; i < target->n_pdds; i++) { struct kfd_process_device *pdd = target->pdds[i]; @@ -591,6 +593,38 @@ int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target, return r; } +int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target, + uint8_t wave_launch_mode) +{ + int r = 0, i; + + if (wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL && + wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT && + wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG) + return -EINVAL; + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + pdd->spi_dbg_launch_mode = pdd->dev->kfd2kgd->set_wave_launch_mode( + pdd->dev->adev, + wave_launch_mode, + pdd->dev->vm_info.last_vmid_kfd); + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + r = debug_refresh_runlist(pdd->dev->dqm); + else + r = kfd_dbg_set_mes_debug_mode(pdd); + + if (r) + break; + } + + return r; +} + void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, uint64_t exception_set_mask) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index c9245221aa76..cb17869437c5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -48,6 +48,8 @@ int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target, uint32_t trap_mask_request, uint32_t *trap_mask_prev, uint32_t *trap_mask_supported); +int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target, + uint8_t wave_launch_mode); int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, unsigned int dev_id, -- cgit v1.2.3 From a70a93fa568b4f05aba548dadb673703eccf5480 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 5 May 2022 16:15:37 -0400 Subject: drm/amdkfd: add debug suspend and resume process queues operation In order to inspect waves from the saved context at any point during a debug session, the debugger must be able to preempt queues to trigger context save by suspending them. On queue suspend, the KFD will copy the context save header information so that the debugger can correctly crawl the appropriate size of the saved context. The debugger must then also be allowed to resume suspended queues. A queue that is newly created cannot be suspended because queue ids are recycled after destruction so the debugger needs to know that this has occurred. Query functions will be later added that will clear a given queue of its new queue status. A queue cannot be destroyed while it is suspended to preserve its saved context during debugger inspection. Have queue destruction block while a queue is suspended and unblocked when it is resumed. Likewise, if a queue is about to be destroyed, it cannot be suspended. Return the number of queues successfully suspended or resumed along with a per queue status array where the upper bits per queue status show that the request was invalid (new/destroyed queue suspend request, missing queue) or an error occurred (HWS in a fatal state so it can't suspend or resume queues). Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 5 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 11 + drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 7 + .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 447 ++++++++++++++++++++- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 10 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 10 + drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 15 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 14 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 +- .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 1 + 11 files changed, 512 insertions(+), 14 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 98cd52bb005f..b4fcad0e62f7 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -772,6 +772,11 @@ bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev) return adev->have_atomics_support; } +void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev) +{ + amdgpu_device_flush_hdp(adev, NULL); +} + void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bool reset) { amdgpu_umc_poison_handler(adev, reset); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index dd740e64e6e1..2d0406bff84e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -322,6 +322,7 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, uint64_t *mmap_offset); int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem, struct dma_buf **dmabuf); +void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev); int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev, struct tile_config *config); void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index a6570b124b2b..1fae97df7a1e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -410,6 +410,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, pr_debug("Write ptr address == 0x%016llX\n", args->write_pointer_address); + kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_NEW), p, dev, queue_id, false, NULL, 0); return 0; err_create_queue: @@ -2996,7 +2997,17 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v args->launch_mode.launch_mode); break; case KFD_IOC_DBG_TRAP_SUSPEND_QUEUES: + r = suspend_queues(target, + args->suspend_queues.num_queues, + args->suspend_queues.grace_period, + args->suspend_queues.exception_mask, + (uint32_t *)args->suspend_queues.queue_array_ptr); + + break; case KFD_IOC_DBG_TRAP_RESUME_QUEUES: + r = resume_queues(target, args->resume_queues.num_queues, + (uint32_t *)args->resume_queues.queue_array_ptr); + break; case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH: case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH: case KFD_IOC_DBG_TRAP_SET_FLAGS: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 53c3418562d4..f4d3dfb35cb3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -339,6 +339,13 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind } kfd_dbg_set_workaround(target, false); + + if (!unwind) { + int resume_count = resume_queues(target, 0, NULL); + + if (resume_count) + pr_debug("Resumed %d queues\n", resume_count); + } } static void kfd_dbg_clean_exception_status(struct kfd_process *target) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 44d87943e40a..bc9e81293165 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -952,6 +952,92 @@ out_unlock: return retval; } +/* suspend_single_queue does not lock the dqm like the + * evict_process_queues_cpsch or evict_process_queues_nocpsch. You should + * lock the dqm before calling, and unlock after calling. + * + * The reason we don't lock the dqm is because this function may be + * called on multiple queues in a loop, so rather than locking/unlocking + * multiple times, we will just keep the dqm locked for all of the calls. + */ +static int suspend_single_queue(struct device_queue_manager *dqm, + struct kfd_process_device *pdd, + struct queue *q) +{ + bool is_new; + + if (q->properties.is_suspended) + return 0; + + pr_debug("Suspending PASID %u queue [%i]\n", + pdd->process->pasid, + q->properties.queue_id); + + is_new = q->properties.exception_status & KFD_EC_MASK(EC_QUEUE_NEW); + + if (is_new || q->properties.is_being_destroyed) { + pr_debug("Suspend: skip %s queue id %i\n", + is_new ? "new" : "destroyed", + q->properties.queue_id); + return -EBUSY; + } + + q->properties.is_suspended = true; + if (q->properties.is_active) { + if (dqm->dev->kfd->shared_resources.enable_mes) { + int r = remove_queue_mes(dqm, q, &pdd->qpd); + + if (r) + return r; + } + + decrement_queue_count(dqm, &pdd->qpd, q); + q->properties.is_active = false; + } + + return 0; +} + +/* resume_single_queue does not lock the dqm like the functions + * restore_process_queues_cpsch or restore_process_queues_nocpsch. You should + * lock the dqm before calling, and unlock after calling. + * + * The reason we don't lock the dqm is because this function may be + * called on multiple queues in a loop, so rather than locking/unlocking + * multiple times, we will just keep the dqm locked for all of the calls. + */ +static int resume_single_queue(struct device_queue_manager *dqm, + struct qcm_process_device *qpd, + struct queue *q) +{ + struct kfd_process_device *pdd; + + if (!q->properties.is_suspended) + return 0; + + pdd = qpd_to_pdd(qpd); + + pr_debug("Restoring from suspend PASID %u queue [%i]\n", + pdd->process->pasid, + q->properties.queue_id); + + q->properties.is_suspended = false; + + if (QUEUE_IS_ACTIVE(q->properties)) { + if (dqm->dev->kfd->shared_resources.enable_mes) { + int r = add_queue_mes(dqm, q, &pdd->qpd); + + if (r) + return r; + } + + q->properties.is_active = true; + increment_queue_count(dqm, qpd, q); + } + + return 0; +} + static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, struct qcm_process_device *qpd) { @@ -1926,6 +2012,31 @@ static int execute_queues_cpsch(struct device_queue_manager *dqm, return map_queues_cpsch(dqm); } +static int wait_on_destroy_queue(struct device_queue_manager *dqm, + struct queue *q) +{ + struct kfd_process_device *pdd = kfd_get_process_device_data(q->device, + q->process); + int ret = 0; + + if (pdd->qpd.is_debug) + return ret; + + q->properties.is_being_destroyed = true; + + if (pdd->process->debug_trap_enabled && q->properties.is_suspended) { + dqm_unlock(dqm); + mutex_unlock(&q->process->mutex); + ret = wait_event_interruptible(dqm->destroy_wait, + !q->properties.is_suspended); + + mutex_lock(&q->process->mutex); + dqm_lock(dqm); + } + + return ret; +} + static int destroy_queue_cpsch(struct device_queue_manager *dqm, struct qcm_process_device *qpd, struct queue *q) @@ -1945,11 +2056,16 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, q->properties.queue_id); } - retval = 0; - /* remove queue from list to prevent rescheduling after preemption */ dqm_lock(dqm); + retval = wait_on_destroy_queue(dqm, q); + + if (retval) { + dqm_unlock(dqm); + return retval; + } + if (qpd->is_debug) { /* * error, currently we do not allow to destroy a queue @@ -1996,7 +2112,14 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, dqm_unlock(dqm); - /* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */ + /* + * Do free_mqd and raise delete event after dqm_unlock(dqm) to avoid + * circular locking + */ + kfd_dbg_ev_raise(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE), + qpd->pqm->process, q->device, + -1, false, NULL, 0); + mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); return retval; @@ -2461,8 +2584,10 @@ struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev) goto out_free; } - if (!dqm->ops.initialize(dqm)) + if (!dqm->ops.initialize(dqm)) { + init_waitqueue_head(&dqm->destroy_wait); return dqm; + } out_free: kfree(dqm); @@ -2602,6 +2727,320 @@ out_unlock: return r; } +#define QUEUE_NOT_FOUND -1 +/* invalidate queue operation in array */ +static void q_array_invalidate(uint32_t num_queues, uint32_t *queue_ids) +{ + int i; + + for (i = 0; i < num_queues; i++) + queue_ids[i] |= KFD_DBG_QUEUE_INVALID_MASK; +} + +/* find queue index in array */ +static int q_array_get_index(unsigned int queue_id, + uint32_t num_queues, + uint32_t *queue_ids) +{ + int i; + + for (i = 0; i < num_queues; i++) + if (queue_id == (queue_ids[i] & ~KFD_DBG_QUEUE_INVALID_MASK)) + return i; + + return QUEUE_NOT_FOUND; +} + +struct copy_context_work_handler_workarea { + struct work_struct copy_context_work; + struct kfd_process *p; +}; + +static void copy_context_work_handler (struct work_struct *work) +{ + struct copy_context_work_handler_workarea *workarea; + struct mqd_manager *mqd_mgr; + struct queue *q; + struct mm_struct *mm; + struct kfd_process *p; + uint32_t tmp_ctl_stack_used_size, tmp_save_area_used_size; + int i; + + workarea = container_of(work, + struct copy_context_work_handler_workarea, + copy_context_work); + + p = workarea->p; + mm = get_task_mm(p->lead_thread); + + if (!mm) + return; + + kthread_use_mm(mm); + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + struct device_queue_manager *dqm = pdd->dev->dqm; + struct qcm_process_device *qpd = &pdd->qpd; + + list_for_each_entry(q, &qpd->queues_list, list) { + mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP]; + + /* We ignore the return value from get_wave_state + * because + * i) right now, it always returns 0, and + * ii) if we hit an error, we would continue to the + * next queue anyway. + */ + mqd_mgr->get_wave_state(mqd_mgr, + q->mqd, + &q->properties, + (void __user *) q->properties.ctx_save_restore_area_address, + &tmp_ctl_stack_used_size, + &tmp_save_area_used_size); + } + } + kthread_unuse_mm(mm); + mmput(mm); +} + +static uint32_t *get_queue_ids(uint32_t num_queues, uint32_t *usr_queue_id_array) +{ + size_t array_size = num_queues * sizeof(uint32_t); + uint32_t *queue_ids = NULL; + + if (!usr_queue_id_array) + return NULL; + + queue_ids = kzalloc(array_size, GFP_KERNEL); + if (!queue_ids) + return ERR_PTR(-ENOMEM); + + if (copy_from_user(queue_ids, usr_queue_id_array, array_size)) + return ERR_PTR(-EFAULT); + + return queue_ids; +} + +int resume_queues(struct kfd_process *p, + uint32_t num_queues, + uint32_t *usr_queue_id_array) +{ + uint32_t *queue_ids = NULL; + int total_resumed = 0; + int i; + + if (usr_queue_id_array) { + queue_ids = get_queue_ids(num_queues, usr_queue_id_array); + + if (IS_ERR(queue_ids)) + return PTR_ERR(queue_ids); + + /* mask all queues as invalid. unmask per successful request */ + q_array_invalidate(num_queues, queue_ids); + } + + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + struct device_queue_manager *dqm = pdd->dev->dqm; + struct qcm_process_device *qpd = &pdd->qpd; + struct queue *q; + int r, per_device_resumed = 0; + + dqm_lock(dqm); + + /* unmask queues that resume or already resumed as valid */ + list_for_each_entry(q, &qpd->queues_list, list) { + int q_idx = QUEUE_NOT_FOUND; + + if (queue_ids) + q_idx = q_array_get_index( + q->properties.queue_id, + num_queues, + queue_ids); + + if (!queue_ids || q_idx != QUEUE_NOT_FOUND) { + int err = resume_single_queue(dqm, &pdd->qpd, q); + + if (queue_ids) { + if (!err) { + queue_ids[q_idx] &= + ~KFD_DBG_QUEUE_INVALID_MASK; + } else { + queue_ids[q_idx] |= + KFD_DBG_QUEUE_ERROR_MASK; + break; + } + } + + if (dqm->dev->kfd->shared_resources.enable_mes) { + wake_up_all(&dqm->destroy_wait); + if (!err) + total_resumed++; + } else { + per_device_resumed++; + } + } + } + + if (!per_device_resumed) { + dqm_unlock(dqm); + continue; + } + + r = execute_queues_cpsch(dqm, + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, + 0, + USE_DEFAULT_GRACE_PERIOD); + if (r) { + pr_err("Failed to resume process queues\n"); + if (queue_ids) { + list_for_each_entry(q, &qpd->queues_list, list) { + int q_idx = q_array_get_index( + q->properties.queue_id, + num_queues, + queue_ids); + + /* mask queue as error on resume fail */ + if (q_idx != QUEUE_NOT_FOUND) + queue_ids[q_idx] |= + KFD_DBG_QUEUE_ERROR_MASK; + } + } + } else { + wake_up_all(&dqm->destroy_wait); + total_resumed += per_device_resumed; + } + + dqm_unlock(dqm); + } + + if (queue_ids) { + if (copy_to_user((void __user *)usr_queue_id_array, queue_ids, + num_queues * sizeof(uint32_t))) + pr_err("copy_to_user failed on queue resume\n"); + + kfree(queue_ids); + } + + return total_resumed; +} + +int suspend_queues(struct kfd_process *p, + uint32_t num_queues, + uint32_t grace_period, + uint64_t exception_clear_mask, + uint32_t *usr_queue_id_array) +{ + uint32_t *queue_ids = get_queue_ids(num_queues, usr_queue_id_array); + int total_suspended = 0; + int i; + + if (IS_ERR(queue_ids)) + return PTR_ERR(queue_ids); + + /* mask all queues as invalid. umask on successful request */ + q_array_invalidate(num_queues, queue_ids); + + for (i = 0; i < p->n_pdds; i++) { + struct kfd_process_device *pdd = p->pdds[i]; + struct device_queue_manager *dqm = pdd->dev->dqm; + struct qcm_process_device *qpd = &pdd->qpd; + struct queue *q; + int r, per_device_suspended = 0; + + mutex_lock(&p->event_mutex); + dqm_lock(dqm); + + /* unmask queues that suspend or already suspended */ + list_for_each_entry(q, &qpd->queues_list, list) { + int q_idx = q_array_get_index(q->properties.queue_id, + num_queues, + queue_ids); + + if (q_idx != QUEUE_NOT_FOUND) { + int err = suspend_single_queue(dqm, pdd, q); + bool is_mes = dqm->dev->kfd->shared_resources.enable_mes; + + if (!err) { + queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK; + if (exception_clear_mask && is_mes) + q->properties.exception_status &= + ~exception_clear_mask; + + if (is_mes) + total_suspended++; + else + per_device_suspended++; + } else if (err != -EBUSY) { + r = err; + queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK; + break; + } + } + } + + if (!per_device_suspended) { + dqm_unlock(dqm); + mutex_unlock(&p->event_mutex); + if (total_suspended) + amdgpu_amdkfd_debug_mem_fence(dqm->dev->adev); + continue; + } + + r = execute_queues_cpsch(dqm, + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + grace_period); + + if (r) + pr_err("Failed to suspend process queues.\n"); + else + total_suspended += per_device_suspended; + + list_for_each_entry(q, &qpd->queues_list, list) { + int q_idx = q_array_get_index(q->properties.queue_id, + num_queues, queue_ids); + + if (q_idx == QUEUE_NOT_FOUND) + continue; + + /* mask queue as error on suspend fail */ + if (r) + queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK; + else if (exception_clear_mask) + q->properties.exception_status &= + ~exception_clear_mask; + } + + dqm_unlock(dqm); + mutex_unlock(&p->event_mutex); + amdgpu_device_flush_hdp(dqm->dev->adev, NULL); + } + + if (total_suspended) { + struct copy_context_work_handler_workarea copy_context_worker; + + INIT_WORK_ONSTACK( + ©_context_worker.copy_context_work, + copy_context_work_handler); + + copy_context_worker.p = p; + + schedule_work(©_context_worker.copy_context_work); + + + flush_work(©_context_worker.copy_context_work); + destroy_work_on_stack(©_context_worker.copy_context_work); + } + + if (copy_to_user((void __user *)usr_queue_id_array, queue_ids, + num_queues * sizeof(uint32_t))) + pr_err("copy_to_user failed on queue suspend\n"); + + kfree(queue_ids); + + return total_suspended; +} + int debug_lock_and_unmap(struct device_queue_manager *dqm) { int r; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index bb75d93712eb..d4e6dbffe8c2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -263,6 +263,8 @@ struct device_queue_manager { uint32_t current_logical_xcc_start; uint32_t wait_times; + + wait_queue_head_t destroy_wait; }; void device_queue_manager_init_cik( @@ -290,6 +292,14 @@ int reserve_debug_trap_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd); int release_debug_trap_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd); +int suspend_queues(struct kfd_process *p, + uint32_t num_queues, + uint32_t grace_period, + uint64_t exception_clear_mask, + uint32_t *usr_queue_id_array); +int resume_queues(struct kfd_process *p, + uint32_t num_queues, + uint32_t *usr_queue_id_array); int debug_lock_and_unmap(struct device_queue_manager *dqm); int debug_map_and_unlock(struct device_queue_manager *dqm); int debug_refresh_runlist(struct device_queue_manager *dqm); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c index a0ac4f2fe6b5..94c0fc2e57b7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c @@ -237,6 +237,7 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, u32 *save_area_used_size) { struct v10_compute_mqd *m; + struct kfd_context_save_area_header header; m = get_mqd(mqd); @@ -255,6 +256,15 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, * accessible to user mode */ + header.wave_state.control_stack_size = *ctl_stack_used_size; + header.wave_state.wave_state_size = *save_area_used_size; + + header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset; + header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset; + + if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state))) + return -EFAULT; + return 0; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c index 9a9b4e853516..31fec5e70d13 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c @@ -291,7 +291,7 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, u32 *save_area_used_size) { struct v11_compute_mqd *m; - /*struct mqd_user_context_save_area_header header;*/ + struct kfd_context_save_area_header header; m = get_mqd(mqd); @@ -309,16 +309,15 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, * it's part of the context save area that is already * accessible to user mode */ -/* - header.control_stack_size = *ctl_stack_used_size; - header.wave_state_size = *save_area_used_size; + header.wave_state.control_stack_size = *ctl_stack_used_size; + header.wave_state.wave_state_size = *save_area_used_size; - header.wave_state_offset = m->cp_hqd_wg_state_offset; - header.control_stack_offset = m->cp_hqd_cntl_stack_offset; + header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset; + header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset; - if (copy_to_user(ctl_stack, &header, sizeof(header))) + if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state))) return -EFAULT; -*/ + return 0; } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c index 5b87c244e909..601bb9f68048 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c @@ -311,6 +311,7 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, u32 *save_area_used_size) { struct v9_mqd *m; + struct kfd_context_save_area_header header; /* Control stack is located one page after MQD. */ void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE); @@ -322,7 +323,18 @@ static int get_wave_state(struct mqd_manager *mm, void *mqd, *save_area_used_size = m->cp_hqd_wg_state_offset - m->cp_hqd_cntl_stack_size; - if (copy_to_user(ctl_stack, mqd_ctl_stack, m->cp_hqd_cntl_stack_size)) + header.wave_state.control_stack_size = *ctl_stack_used_size; + header.wave_state.wave_state_size = *save_area_used_size; + + header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset; + header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset; + + if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state))) + return -EFAULT; + + if (copy_to_user(ctl_stack + m->cp_hqd_cntl_stack_offset, + mqd_ctl_stack + m->cp_hqd_cntl_stack_offset, + *ctl_stack_used_size)) return -EFAULT; return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index cd2d56e5cdf0..05da43bf233a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -510,6 +510,8 @@ struct queue_properties { uint32_t doorbell_off; bool is_interop; bool is_evicted; + bool is_suspended; + bool is_being_destroyed; bool is_active; bool is_gws; uint32_t pm4_target_xcc; @@ -535,7 +537,8 @@ struct queue_properties { #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ (q).queue_address != 0 && \ (q).queue_percent > 0 && \ - !(q).is_evicted) + !(q).is_evicted && \ + !(q).is_suspended) enum mqd_update_flag { UPDATE_FLAG_DBG_WA_ENABLE = 1, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 70852a200d8f..01ccab607a69 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -187,6 +187,7 @@ static int init_user_queue(struct process_queue_manager *pqm, /* Doorbell initialized in user space*/ q_properties->doorbell_ptr = NULL; + q_properties->exception_status = KFD_EC_MASK(EC_QUEUE_NEW); /* let DQM handle it*/ q_properties->vmid = 0; -- cgit v1.2.3 From e0f85f4690d089cc1a60337decafb1acf7eec45e Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Fri, 6 May 2022 14:58:55 -0400 Subject: drm/amdkfd: add debug set and clear address watch points operation Shader read, write and atomic memory operations can be alerted to the debugger as an address watch exception. Allow the debugger to pass in a watch point to a particular memory address per device. Note that there exists only 4 watch points per devices to date, so have the KFD keep track of what watch points are allocated or not. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c | 51 ++++++++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 78 ++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h | 8 ++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c | 5 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c | 52 +++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 77 ++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 8 ++ drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 24 ++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 136 +++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 6 +- 13 files changed, 452 insertions(+), 5 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c index 774ecfc3451a..efd6a72aab4e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c @@ -118,6 +118,55 @@ static uint32_t kgd_aldebaran_set_wave_launch_mode(struct amdgpu_device *adev, return data; } +#define TCP_WATCH_STRIDE (regTCP_WATCH1_ADDR_H - regTCP_WATCH0_ADDR_H) +static uint32_t kgd_gfx_aldebaran_set_address_watch( + struct amdgpu_device *adev, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t watch_id, + uint32_t watch_mode, + uint32_t debug_vmid) +{ + uint32_t watch_address_high; + uint32_t watch_address_low; + uint32_t watch_address_cntl; + + watch_address_cntl = 0; + watch_address_low = lower_32_bits(watch_address); + watch_address_high = upper_32_bits(watch_address) & 0xffff; + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MODE, + watch_mode); + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MASK, + watch_address_mask >> 6); + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VALID, + 1); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_H) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_high); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_L) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_low); + + return watch_address_cntl; +} + +uint32_t kgd_gfx_aldebaran_clear_address_watch(struct amdgpu_device *adev, + uint32_t watch_id) +{ + return 0; +} + const struct kfd2kgd_calls aldebaran_kfd2kgd = { .program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings, .set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping, @@ -141,6 +190,8 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = { .validate_trap_override_request = kgd_aldebaran_validate_trap_override_request, .set_wave_launch_trap_override = kgd_aldebaran_set_wave_launch_trap_override, .set_wave_launch_mode = kgd_aldebaran_set_wave_launch_mode, + .set_address_watch = kgd_gfx_aldebaran_set_address_watch, + .clear_address_watch = kgd_gfx_aldebaran_clear_address_watch, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c index fbdc1b7b1e42..6df215aba4c4 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c @@ -413,6 +413,8 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = { .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override, .set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode, + .set_address_watch = kgd_gfx_v9_set_address_watch, + .clear_address_watch = kgd_gfx_v9_clear_address_watch, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index a7a6edda557f..8ad7a7779e14 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -880,6 +880,82 @@ uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev, return 0; } +#define TCP_WATCH_STRIDE (mmTCP_WATCH1_ADDR_H - mmTCP_WATCH0_ADDR_H) +uint32_t kgd_gfx_v10_set_address_watch(struct amdgpu_device *adev, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t watch_id, + uint32_t watch_mode, + uint32_t debug_vmid) +{ + uint32_t watch_address_high; + uint32_t watch_address_low; + uint32_t watch_address_cntl; + + watch_address_cntl = 0; + + watch_address_low = lower_32_bits(watch_address); + watch_address_high = upper_32_bits(watch_address) & 0xffff; + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VMID, + debug_vmid); + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MODE, + watch_mode); + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MASK, + watch_address_mask >> 7); + + /* Turning off this watch point until we set all the registers */ + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VALID, + 0); + + WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_cntl); + + WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_high); + + WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_L) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_low); + + /* Enable the watch point */ + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VALID, + 1); + + WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_cntl); + + return 0; +} + +uint32_t kgd_gfx_v10_clear_address_watch(struct amdgpu_device *adev, + uint32_t watch_id) +{ + uint32_t watch_address_cntl; + + watch_address_cntl = 0; + + WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_cntl); + + return 0; +} + + /* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values * The values read are: * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. @@ -969,6 +1045,8 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = { .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override, .set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode, + .set_address_watch = kgd_gfx_v10_set_address_watch, + .clear_address_watch = kgd_gfx_v10_clear_address_watch, .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .program_trap_handler_settings = program_trap_handler_settings, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h index 3a6aca2b0eaa..e6b70196071a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h @@ -39,6 +39,14 @@ uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev, uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev, uint8_t wave_launch_mode, uint32_t vmid); +uint32_t kgd_gfx_v10_set_address_watch(struct amdgpu_device *adev, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t watch_id, + uint32_t watch_mode, + uint32_t debug_vmid); +uint32_t kgd_gfx_v10_clear_address_watch(struct amdgpu_device *adev, + uint32_t watch_id); void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, uint32_t wait_times, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c index ed36b433a48b..8c8437a4383f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c @@ -678,6 +678,7 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = { .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override, - .set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode - + .set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode, + .set_address_watch = kgd_gfx_v10_set_address_watch, + .clear_address_watch = kgd_gfx_v10_clear_address_watch }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c index 9711d5128d09..52efa690a3c2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c @@ -737,6 +737,54 @@ static uint32_t kgd_gfx_v11_set_wave_launch_mode(struct amdgpu_device *adev, return data; } +#define TCP_WATCH_STRIDE (regTCP_WATCH1_ADDR_H - regTCP_WATCH0_ADDR_H) +static uint32_t kgd_gfx_v11_set_address_watch(struct amdgpu_device *adev, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t watch_id, + uint32_t watch_mode, + uint32_t debug_vmid) +{ + uint32_t watch_address_high; + uint32_t watch_address_low; + uint32_t watch_address_cntl; + + watch_address_cntl = 0; + watch_address_low = lower_32_bits(watch_address); + watch_address_high = upper_32_bits(watch_address) & 0xffff; + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MODE, + watch_mode); + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MASK, + watch_address_mask >> 7); + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VALID, + 1); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_H) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_high); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_L) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_low); + + return watch_address_cntl; +} + +uint32_t kgd_gfx_v11_clear_address_watch(struct amdgpu_device *adev, + uint32_t watch_id) +{ + return 0; +} + const struct kfd2kgd_calls gfx_v11_kfd2kgd = { .program_sh_mem_settings = program_sh_mem_settings_v11, .set_pasid_vmid_mapping = set_pasid_vmid_mapping_v11, @@ -757,5 +805,7 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = { .disable_debug_trap = kgd_gfx_v11_disable_debug_trap, .validate_trap_override_request = kgd_gfx_v11_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v11_set_wave_launch_trap_override, - .set_wave_launch_mode = kgd_gfx_v11_set_wave_launch_mode + .set_wave_launch_mode = kgd_gfx_v11_set_wave_launch_mode, + .set_address_watch = kgd_gfx_v11_set_address_watch, + .clear_address_watch = kgd_gfx_v11_clear_address_watch }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index 0acc0c18dfe6..51d93fb13ea3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -816,6 +816,81 @@ uint32_t kgd_gfx_v9_set_wave_launch_mode(struct amdgpu_device *adev, return 0; } +#define TCP_WATCH_STRIDE (mmTCP_WATCH1_ADDR_H - mmTCP_WATCH0_ADDR_H) +uint32_t kgd_gfx_v9_set_address_watch(struct amdgpu_device *adev, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t watch_id, + uint32_t watch_mode, + uint32_t debug_vmid) +{ + uint32_t watch_address_high; + uint32_t watch_address_low; + uint32_t watch_address_cntl; + + watch_address_cntl = 0; + + watch_address_low = lower_32_bits(watch_address); + watch_address_high = upper_32_bits(watch_address) & 0xffff; + + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VMID, + debug_vmid); + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MODE, + watch_mode); + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + MASK, + watch_address_mask >> 6); + + /* Turning off this watch point until we set all the registers */ + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VALID, + 0); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_cntl); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_high); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_L) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_low); + + /* Enable the watch point */ + watch_address_cntl = REG_SET_FIELD(watch_address_cntl, + TCP_WATCH0_CNTL, + VALID, + 1); + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_cntl); + + return 0; +} + +uint32_t kgd_gfx_v9_clear_address_watch(struct amdgpu_device *adev, + uint32_t watch_id) +{ + uint32_t watch_address_cntl; + + watch_address_cntl = 0; + + WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) + + (watch_id * TCP_WATCH_STRIDE)), + watch_address_cntl); + + return 0; +} + /* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values * The values read are: * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. @@ -1090,6 +1165,8 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = { .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request, .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override, .set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode, + .set_address_watch = kgd_gfx_v9_set_address_watch, + .clear_address_watch = kgd_gfx_v9_clear_address_watch, .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index 18f4970ac8e4..5f54bff0db49 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -84,6 +84,14 @@ uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev, uint32_t trap_mask_request, uint32_t *trap_mask_prev, uint32_t kfd_dbg_trap_cntl_prev); +uint32_t kgd_gfx_v9_set_address_watch(struct amdgpu_device *adev, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t watch_id, + uint32_t watch_mode, + uint32_t debug_vmid); +uint32_t kgd_gfx_v9_clear_address_watch(struct amdgpu_device *adev, + uint32_t watch_id); void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev, uint32_t wait_times, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 1fae97df7a1e..016724c82928 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2880,6 +2880,7 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v struct mm_struct *mm = NULL; struct pid *pid = NULL; struct kfd_process *target = NULL; + struct kfd_process_device *pdd = NULL; int r = 0; if (sched_policy == KFD_SCHED_POLICY_NO_HWS) { @@ -2957,6 +2958,20 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v goto unlock_out; } + if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH || + args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) { + int user_gpu_id = kfd_process_get_user_gpu_id(target, + args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ? + args->set_node_address_watch.gpu_id : + args->clear_node_address_watch.gpu_id); + + pdd = kfd_process_device_data_by_id(target, user_gpu_id); + if (user_gpu_id == -EINVAL || !pdd) { + r = -ENODEV; + goto unlock_out; + } + } + switch (args->op) { case KFD_IOC_DBG_TRAP_ENABLE: if (target != p) @@ -3009,7 +3024,16 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v (uint32_t *)args->resume_queues.queue_array_ptr); break; case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH: + r = kfd_dbg_trap_set_dev_address_watch(pdd, + args->set_node_address_watch.address, + args->set_node_address_watch.mask, + &args->set_node_address_watch.id, + args->set_node_address_watch.mode); + break; case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH: + r = kfd_dbg_trap_clear_dev_address_watch(pdd, + args->clear_node_address_watch.id); + break; case KFD_IOC_DBG_TRAP_SET_FLAGS: case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT: case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index f4d3dfb35cb3..4b36cc8b5fb7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -24,6 +24,8 @@ #include "kfd_device_queue_manager.h" #include +#define MAX_WATCH_ADDRESSES 4 + void debug_event_write_work_handler(struct work_struct *work) { struct kfd_process *process; @@ -289,6 +291,139 @@ int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) pdd->watch_points, flags); } +#define KFD_DEBUGGER_INVALID_WATCH_POINT_ID -1 +static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_id) +{ + int i; + + *watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID; + + spin_lock(&pdd->dev->kfd->watch_points_lock); + + for (i = 0; i < MAX_WATCH_ADDRESSES; i++) { + /* device watchpoint in use so skip */ + if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1) + continue; + + pdd->alloc_watch_ids |= 0x1 << i; + pdd->dev->kfd->alloc_watch_ids |= 0x1 << i; + *watch_id = i; + spin_unlock(&pdd->dev->kfd->watch_points_lock); + return 0; + } + + spin_unlock(&pdd->dev->kfd->watch_points_lock); + + return -ENOMEM; +} + +static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id) +{ + spin_lock(&pdd->dev->kfd->watch_points_lock); + + /* process owns device watch point so safe to clear */ + if ((pdd->alloc_watch_ids >> watch_id) & 0x1) { + pdd->alloc_watch_ids &= ~(0x1 << watch_id); + pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id); + } + + spin_unlock(&pdd->dev->kfd->watch_points_lock); +} + +static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id) +{ + bool owns_watch_id = false; + + spin_lock(&pdd->dev->kfd->watch_points_lock); + owns_watch_id = watch_id < MAX_WATCH_ADDRESSES && + ((pdd->alloc_watch_ids >> watch_id) & 0x1); + + spin_unlock(&pdd->dev->kfd->watch_points_lock); + + return owns_watch_id; +} + +int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd, + uint32_t watch_id) +{ + int r; + + if (!kfd_dbg_owns_dev_watch_id(pdd, watch_id)) + return -EINVAL; + + if (!pdd->dev->kfd->shared_resources.enable_mes) { + r = debug_lock_and_unmap(pdd->dev->dqm); + if (r) + return r; + } + + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + pdd->watch_points[watch_id] = pdd->dev->kfd2kgd->clear_address_watch( + pdd->dev->adev, + watch_id); + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + r = debug_map_and_unlock(pdd->dev->dqm); + else + r = kfd_dbg_set_mes_debug_mode(pdd); + + kfd_dbg_clear_dev_watch_id(pdd, watch_id); + + return r; +} + +int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t *watch_id, + uint32_t watch_mode) +{ + int r = kfd_dbg_get_dev_watch_id(pdd, watch_id); + + if (r) + return r; + + if (!pdd->dev->kfd->shared_resources.enable_mes) { + r = debug_lock_and_unmap(pdd->dev->dqm); + if (r) { + kfd_dbg_clear_dev_watch_id(pdd, *watch_id); + return r; + } + } + + amdgpu_gfx_off_ctrl(pdd->dev->adev, false); + pdd->watch_points[*watch_id] = pdd->dev->kfd2kgd->set_address_watch( + pdd->dev->adev, + watch_address, + watch_address_mask, + *watch_id, + watch_mode, + pdd->dev->vm_info.last_vmid_kfd); + amdgpu_gfx_off_ctrl(pdd->dev->adev, true); + + if (!pdd->dev->kfd->shared_resources.enable_mes) + r = debug_map_and_unlock(pdd->dev->dqm); + else + r = kfd_dbg_set_mes_debug_mode(pdd); + + /* HWS is broken so no point in HW rollback but release the watchpoint anyways */ + if (r) + kfd_dbg_clear_dev_watch_id(pdd, *watch_id); + + return 0; +} + +static void kfd_dbg_clear_process_address_watch(struct kfd_process *target) +{ + int i, j; + + for (i = 0; i < target->n_pdds; i++) + for (j = 0; j < MAX_WATCH_ADDRESSES; j++) + kfd_dbg_trap_clear_dev_address_watch(target->pdds[i], j); +} + + /* kfd_dbg_trap_deactivate: * target: target process * unwind: If this is unwinding a failed kfd_dbg_trap_enable() @@ -303,6 +438,7 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind if (!unwind) { cancel_work_sync(&target->debug_event_workarea); + kfd_dbg_clear_process_address_watch(target); kfd_dbg_trap_set_wave_launch_mode(target, 0); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index cb17869437c5..7f0757c2af2c 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -50,7 +50,13 @@ int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target, uint32_t *trap_mask_supported); int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target, uint8_t wave_launch_mode); - +int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd, + uint32_t watch_id); +int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd, + uint64_t watch_address, + uint32_t watch_address_mask, + uint32_t *watch_id, + uint32_t watch_mode); int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, unsigned int dev_id, unsigned int queue_id, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 2c36bb578633..9fc9d32cb579 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -811,6 +811,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, if (kfd_resume_iommu(kfd)) goto kfd_resume_iommu_error; + spin_lock_init(&kfd->watch_points_lock); + kfd->init_complete = true; dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor, kfd->adev->pdev->device); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 05da43bf233a..8ec87bc8ba82 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -381,6 +381,10 @@ struct kfd_dev { struct kfd_node *nodes[MAX_KFD_NODES]; unsigned int num_nodes; + + /* Track per device allocated watch points */ + uint32_t alloc_watch_ids; + spinlock_t watch_points_lock; }; enum kfd_mempool { @@ -833,6 +837,7 @@ struct kfd_process_device { uint32_t spi_dbg_override; uint32_t spi_dbg_launch_mode; uint32_t watch_points[4]; + uint32_t alloc_watch_ids; /* * If this process has been checkpointed before, then the user @@ -989,7 +994,6 @@ struct kfd_process { struct semaphore runtime_enable_sema; bool is_runtime_retry; struct kfd_runtime_info runtime_info; - }; #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ -- cgit v1.2.3 From 103d5f08ff42b666c61c350be2c3e724c1646918 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 9 May 2022 10:51:56 -0400 Subject: drm/amdkfd: add debug set flags operation Allow the debugger to set single memory and single ALU operations. Some exceptions are imprecise (memory violations, address watch) in the sense that a trap occurs only when the exception interrupt occurs and not at the non-halting faulty instruction. Trap temporaries 0 & 1 save the program counter address, which means that these values will not point to the faulty instruction address but to whenever the interrupt was raised. Setting the Single Memory Operations flag will inject an automatic wait on every memory operation instruction forcing imprecise memory exceptions to become precise at the cost of performance. This setting is not permitted on debug devices that support only a global setting of this option. Return the previous set flags to the debugger as well. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 ++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 58 ++++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 1 + 3 files changed, 61 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 016724c82928..5ee38614ed9b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -3035,6 +3035,8 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v args->clear_node_address_watch.id); break; case KFD_IOC_DBG_TRAP_SET_FLAGS: + r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags); + break; case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT: case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 4b36cc8b5fb7..43c3170998d3 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -23,6 +23,7 @@ #include "kfd_debug.h" #include "kfd_device_queue_manager.h" #include +#include #define MAX_WATCH_ADDRESSES 4 @@ -423,6 +424,59 @@ static void kfd_dbg_clear_process_address_watch(struct kfd_process *target) kfd_dbg_trap_clear_dev_address_watch(target->pdds[i], j); } +int kfd_dbg_trap_set_flags(struct kfd_process *target, uint32_t *flags) +{ + uint32_t prev_flags = target->dbg_flags; + int i, r = 0, rewind_count = 0; + + for (i = 0; i < target->n_pdds; i++) { + if (!kfd_dbg_is_per_vmid_supported(target->pdds[i]->dev) && + (*flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP)) { + *flags = prev_flags; + return -EACCES; + } + } + + target->dbg_flags = *flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP; + *flags = prev_flags; + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) + continue; + + if (!pdd->dev->kfd->shared_resources.enable_mes) + r = debug_refresh_runlist(pdd->dev->dqm); + else + r = kfd_dbg_set_mes_debug_mode(pdd); + + if (r) { + target->dbg_flags = prev_flags; + break; + } + + rewind_count++; + } + + /* Rewind flags */ + if (r) { + target->dbg_flags = prev_flags; + + for (i = 0; i < rewind_count; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + + if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) + continue; + + if (!pdd->dev->kfd->shared_resources.enable_mes) + debug_refresh_runlist(pdd->dev->dqm); + else + kfd_dbg_set_mes_debug_mode(pdd); + } + } + + return r; +} /* kfd_dbg_trap_deactivate: * target: target process @@ -437,9 +491,13 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind int i; if (!unwind) { + uint32_t flags = 0; + cancel_work_sync(&target->debug_event_workarea); kfd_dbg_clear_process_address_watch(target); kfd_dbg_trap_set_wave_launch_mode(target, 0); + + kfd_dbg_trap_set_flags(target, &flags); } for (i = 0; i < target->n_pdds; i++) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 7f0757c2af2c..ef8e9f7f1716 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -57,6 +57,7 @@ int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd, uint32_t watch_address_mask, uint32_t *watch_id, uint32_t watch_mode); +int kfd_dbg_trap_set_flags(struct kfd_process *target, uint32_t *flags); int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, unsigned int dev_id, unsigned int queue_id, -- cgit v1.2.3 From 5bc20c224bcb863571e8831cdbba23cd61b10ac3 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 9 May 2022 11:10:32 -0400 Subject: drm/amdkfd: add debug query event operation Allow the debugger to query a single queue, device and process exception. The KFD should also return the GPU or Queue id of the exception. The debugger also has the option of clearing exceptions after being queried. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 +++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 64 ++++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 5 +++ 3 files changed, 75 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 5ee38614ed9b..498859259b55 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -3038,6 +3038,12 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v r = kfd_dbg_trap_set_flags(target, &args->set_flags.flags); break; case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT: + r = kfd_dbg_ev_query_debug_event(target, + &args->query_debug_event.queue_id, + &args->query_debug_event.gpu_id, + args->query_debug_event.exception_mask, + &args->query_debug_event.exception_mask); + break; case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 43c3170998d3..e9530e682e85 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -27,6 +27,70 @@ #define MAX_WATCH_ADDRESSES 4 +int kfd_dbg_ev_query_debug_event(struct kfd_process *process, + unsigned int *queue_id, + unsigned int *gpu_id, + uint64_t exception_clear_mask, + uint64_t *event_status) +{ + struct process_queue_manager *pqm; + struct process_queue_node *pqn; + int i; + + if (!(process && process->debug_trap_enabled)) + return -ENODATA; + + mutex_lock(&process->event_mutex); + *event_status = 0; + *queue_id = 0; + *gpu_id = 0; + + /* find and report queue events */ + pqm = &process->pqm; + list_for_each_entry(pqn, &pqm->queues, process_queue_list) { + uint64_t tmp = process->exception_enable_mask; + + if (!pqn->q) + continue; + + tmp &= pqn->q->properties.exception_status; + + if (!tmp) + continue; + + *event_status = pqn->q->properties.exception_status; + *queue_id = pqn->q->properties.queue_id; + *gpu_id = pqn->q->device->id; + pqn->q->properties.exception_status &= ~exception_clear_mask; + goto out; + } + + /* find and report device events */ + for (i = 0; i < process->n_pdds; i++) { + struct kfd_process_device *pdd = process->pdds[i]; + uint64_t tmp = process->exception_enable_mask + & pdd->exception_status; + + if (!tmp) + continue; + + *event_status = pdd->exception_status; + *gpu_id = pdd->dev->id; + pdd->exception_status &= ~exception_clear_mask; + goto out; + } + + /* report process events */ + if (process->exception_enable_mask & process->exception_status) { + *event_status = process->exception_status; + process->exception_status &= ~exception_clear_mask; + } + +out: + mutex_unlock(&process->event_mutex); + return *event_status ? 0 : -EAGAIN; +} + void debug_event_write_work_handler(struct work_struct *work) { struct kfd_process *process; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index ef8e9f7f1716..e78f954c0684 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -27,6 +27,11 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count); int kfd_dbg_trap_activate(struct kfd_process *target); +int kfd_dbg_ev_query_debug_event(struct kfd_process *process, + unsigned int *queue_id, + unsigned int *gpu_id, + uint64_t exception_clear_mask, + uint64_t *event_status); bool kfd_set_dbg_ev_from_interrupt(struct kfd_node *dev, unsigned int pasid, uint32_t doorbell_id, -- cgit v1.2.3 From 2b36de971d25daa2ad287114ae3ca11a8f8d49d7 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 9 May 2022 13:37:36 -0400 Subject: drm/amdkfd: add debug query exception info operation Allow the debugger to query additional info based on an exception code. For device exceptions, it's currently only memory violation information. For process exceptions, it's currently only runtime information. Queue exception only report the queue exception status. The debugger has the option of clearing the target exception on query. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 ++ drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 120 +++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 6 ++ 3 files changed, 133 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 498859259b55..b7ee79b5220a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -3045,6 +3045,13 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v &args->query_debug_event.exception_mask); break; case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO: + r = kfd_dbg_trap_query_exception_info(target, + args->query_exception_info.source_id, + args->query_exception_info.exception_code, + args->query_exception_info.clear_exception, + (void __user *)args->query_exception_info.info_ptr, + &args->query_exception_info.info_size); + break; case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: pr_warn("Debug op %i not supported yet\n", args->op); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index e9530e682e85..24e2b285448a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -890,6 +890,126 @@ int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target, return r; } +int kfd_dbg_trap_query_exception_info(struct kfd_process *target, + uint32_t source_id, + uint32_t exception_code, + bool clear_exception, + void __user *info, + uint32_t *info_size) +{ + bool found = false; + int r = 0; + uint32_t copy_size, actual_info_size = 0; + uint64_t *exception_status_ptr = NULL; + + if (!target) + return -EINVAL; + + if (!info || !info_size) + return -EINVAL; + + mutex_lock(&target->event_mutex); + + if (KFD_DBG_EC_TYPE_IS_QUEUE(exception_code)) { + /* Per queue exceptions */ + struct queue *queue = NULL; + int i; + + for (i = 0; i < target->n_pdds; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + struct qcm_process_device *qpd = &pdd->qpd; + + list_for_each_entry(queue, &qpd->queues_list, list) { + if (!found && queue->properties.queue_id == source_id) { + found = true; + break; + } + } + if (found) + break; + } + + if (!found) { + r = -EINVAL; + goto out; + } + + if (!(queue->properties.exception_status & KFD_EC_MASK(exception_code))) { + r = -ENODATA; + goto out; + } + exception_status_ptr = &queue->properties.exception_status; + } else if (KFD_DBG_EC_TYPE_IS_DEVICE(exception_code)) { + /* Per device exceptions */ + struct kfd_process_device *pdd = NULL; + int i; + + for (i = 0; i < target->n_pdds; i++) { + pdd = target->pdds[i]; + if (pdd->dev->id == source_id) { + found = true; + break; + } + } + + if (!found) { + r = -EINVAL; + goto out; + } + + if (!(pdd->exception_status & KFD_EC_MASK(exception_code))) { + r = -ENODATA; + goto out; + } + + if (exception_code == EC_DEVICE_MEMORY_VIOLATION) { + copy_size = min((size_t)(*info_size), pdd->vm_fault_exc_data_size); + + if (copy_to_user(info, pdd->vm_fault_exc_data, copy_size)) { + r = -EFAULT; + goto out; + } + actual_info_size = pdd->vm_fault_exc_data_size; + if (clear_exception) { + kfree(pdd->vm_fault_exc_data); + pdd->vm_fault_exc_data = NULL; + pdd->vm_fault_exc_data_size = 0; + } + } + exception_status_ptr = &pdd->exception_status; + } else if (KFD_DBG_EC_TYPE_IS_PROCESS(exception_code)) { + /* Per process exceptions */ + if (!(target->exception_status & KFD_EC_MASK(exception_code))) { + r = -ENODATA; + goto out; + } + + if (exception_code == EC_PROCESS_RUNTIME) { + copy_size = min((size_t)(*info_size), sizeof(target->runtime_info)); + + if (copy_to_user(info, (void *)&target->runtime_info, copy_size)) { + r = -EFAULT; + goto out; + } + + actual_info_size = sizeof(target->runtime_info); + } + + exception_status_ptr = &target->exception_status; + } else { + pr_debug("Bad exception type [%i]\n", exception_code); + r = -EINVAL; + goto out; + } + + *info_size = actual_info_size; + if (clear_exception) + *exception_status_ptr &= ~KFD_EC_MASK(exception_code); +out: + mutex_unlock(&target->event_mutex); + return r; +} + void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, uint64_t exception_set_mask) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index e78f954c0684..234e2ccafa87 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -63,6 +63,12 @@ int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd, uint32_t *watch_id, uint32_t watch_mode); int kfd_dbg_trap_set_flags(struct kfd_process *target, uint32_t *flags); +int kfd_dbg_trap_query_exception_info(struct kfd_process *target, + uint32_t source_id, + uint32_t exception_code, + bool clear_exception, + void __user *info, + uint32_t *info_size); int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, unsigned int dev_id, unsigned int queue_id, -- cgit v1.2.3 From b17bd5dbf64677682a3bca249c64521d5eabcb38 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Tue, 10 May 2022 11:15:29 -0400 Subject: drm/amdkfd: add debug queue snapshot operation Allow the debugger to get a snapshot of a specified number of queues containing various queue property information that is copied to the debugger. Since the debugger doesn't know how many queues exist at any given time, allow the debugger to pass the requested number of snapshots as 0 to get the actual number of potential snapshots to use for a subsequent snapshot request for actual information. To prevent future ABI breakage, pass in the requested entry_size. The KFD will return it's own entry_size in case the debugger still wants log the information in a core dump on sizing failure. Also allow the debugger to clear exceptions when doing a snapshot. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 6 ++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 36 +++++++++++++++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 3 ++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 +++ .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 40 ++++++++++++++++++++++ 5 files changed, 90 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index b7ee79b5220a..24066756e478 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -3053,6 +3053,12 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v &args->query_exception_info.info_size); break; case KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT: + r = pqm_get_queue_snapshot(&target->pqm, + args->queue_snapshot.exception_mask, + (void __user *)args->queue_snapshot.snapshot_buf_ptr, + &args->queue_snapshot.num_queues, + &args->queue_snapshot.entry_size); + break; case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: pr_warn("Debug op %i not supported yet\n", args->op); r = -EACCES; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index bc9e81293165..0c1be91a87c6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -3041,6 +3041,42 @@ int suspend_queues(struct kfd_process *p, return total_suspended; } +static uint32_t set_queue_type_for_user(struct queue_properties *q_props) +{ + switch (q_props->type) { + case KFD_QUEUE_TYPE_COMPUTE: + return q_props->format == KFD_QUEUE_FORMAT_PM4 + ? KFD_IOC_QUEUE_TYPE_COMPUTE + : KFD_IOC_QUEUE_TYPE_COMPUTE_AQL; + case KFD_QUEUE_TYPE_SDMA: + return KFD_IOC_QUEUE_TYPE_SDMA; + case KFD_QUEUE_TYPE_SDMA_XGMI: + return KFD_IOC_QUEUE_TYPE_SDMA_XGMI; + default: + WARN_ONCE(true, "queue type not recognized!"); + return 0xffffffff; + }; +} + +void set_queue_snapshot_entry(struct queue *q, + uint64_t exception_clear_mask, + struct kfd_queue_snapshot_entry *qss_entry) +{ + qss_entry->ring_base_address = q->properties.queue_address; + qss_entry->write_pointer_address = (uint64_t)q->properties.write_ptr; + qss_entry->read_pointer_address = (uint64_t)q->properties.read_ptr; + qss_entry->ctx_save_restore_address = + q->properties.ctx_save_restore_area_address; + qss_entry->ctx_save_restore_area_size = + q->properties.ctx_save_restore_area_size; + qss_entry->exception_status = q->properties.exception_status; + qss_entry->queue_id = q->properties.queue_id; + qss_entry->gpu_id = q->device->id; + qss_entry->ring_size = (uint32_t)q->properties.queue_size; + qss_entry->queue_type = set_queue_type_for_user(&q->properties); + q->properties.exception_status &= ~exception_clear_mask; +} + int debug_lock_and_unmap(struct device_queue_manager *dqm) { int r; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index d4e6dbffe8c2..7dd4b177219d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -300,6 +300,9 @@ int suspend_queues(struct kfd_process *p, int resume_queues(struct kfd_process *p, uint32_t num_queues, uint32_t *usr_queue_id_array); +void set_queue_snapshot_entry(struct queue *q, + uint64_t exception_clear_mask, + struct kfd_queue_snapshot_entry *qss_entry); int debug_lock_and_unmap(struct device_queue_manager *dqm); int debug_map_and_unlock(struct device_queue_manager *dqm); int debug_refresh_runlist(struct device_queue_manager *dqm); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 8ec87bc8ba82..023b17e0116b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1355,6 +1355,11 @@ int pqm_get_wave_state(struct process_queue_manager *pqm, void __user *ctl_stack, u32 *ctl_stack_used_size, u32 *save_area_used_size); +int pqm_get_queue_snapshot(struct process_queue_manager *pqm, + uint64_t exception_clear_mask, + void __user *buf, + int *num_qss_entries, + uint32_t *entry_size); int amdkfd_fence_wait_timeout(uint64_t *fence_addr, uint64_t fence_value, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c index 01ccab607a69..9ad1a2186a24 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c @@ -585,6 +585,46 @@ int pqm_get_wave_state(struct process_queue_manager *pqm, save_area_used_size); } +int pqm_get_queue_snapshot(struct process_queue_manager *pqm, + uint64_t exception_clear_mask, + void __user *buf, + int *num_qss_entries, + uint32_t *entry_size) +{ + struct process_queue_node *pqn; + struct kfd_queue_snapshot_entry src; + uint32_t tmp_entry_size = *entry_size, tmp_qss_entries = *num_qss_entries; + int r = 0; + + *num_qss_entries = 0; + if (!(*entry_size)) + return -EINVAL; + + *entry_size = min_t(size_t, *entry_size, sizeof(struct kfd_queue_snapshot_entry)); + mutex_lock(&pqm->process->event_mutex); + + memset(&src, 0, sizeof(src)); + + list_for_each_entry(pqn, &pqm->queues, process_queue_list) { + if (!pqn->q) + continue; + + if (*num_qss_entries < tmp_qss_entries) { + set_queue_snapshot_entry(pqn->q, exception_clear_mask, &src); + + if (copy_to_user(buf, &src, *entry_size)) { + r = -EFAULT; + break; + } + buf += tmp_entry_size; + } + *num_qss_entries += 1; + } + + mutex_unlock(&pqm->process->event_mutex); + return r; +} + static int get_queue_data_sizes(struct kfd_process_device *pdd, struct queue *q, uint32_t *mqd_size, -- cgit v1.2.3 From 12976e6a5ab8fc3766c0304d72f7eec81a109b55 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Tue, 10 May 2022 12:47:45 -0400 Subject: drm/amdkfd: add debug device snapshot operation Similar to queue snapshot, return an array of device information using an entry_size check and return. Unlike queue snapshots, the debugger needs to pass to correct number of devices that exist. If it fails to do so, the KFD will return the number of actual devices so that the debugger can make a subsequent successful call. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 ++- drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 73 ++++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 5 +++ 3 files changed, 83 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 24066756e478..f54ff5c3387d 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -3060,8 +3060,11 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v &args->queue_snapshot.entry_size); break; case KFD_IOC_DBG_TRAP_GET_DEVICE_SNAPSHOT: - pr_warn("Debug op %i not supported yet\n", args->op); - r = -EACCES; + r = kfd_dbg_trap_device_snapshot(target, + args->device_snapshot.exception_mask, + (void __user *)args->device_snapshot.snapshot_buf_ptr, + &args->device_snapshot.num_devices, + &args->device_snapshot.entry_size); break; default: pr_err("Invalid option: %i\n", args->op); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 24e2b285448a..125274445f43 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -22,6 +22,7 @@ #include "kfd_debug.h" #include "kfd_device_queue_manager.h" +#include "kfd_topology.h" #include #include @@ -1010,6 +1011,78 @@ out: return r; } +int kfd_dbg_trap_device_snapshot(struct kfd_process *target, + uint64_t exception_clear_mask, + void __user *user_info, + uint32_t *number_of_device_infos, + uint32_t *entry_size) +{ + struct kfd_dbg_device_info_entry device_info; + uint32_t tmp_entry_size = *entry_size, tmp_num_devices; + int i, r = 0; + + if (!(target && user_info && number_of_device_infos && entry_size)) + return -EINVAL; + + tmp_num_devices = min_t(size_t, *number_of_device_infos, target->n_pdds); + *number_of_device_infos = target->n_pdds; + *entry_size = min_t(size_t, *entry_size, sizeof(device_info)); + + if (!tmp_num_devices) + return 0; + + memset(&device_info, 0, sizeof(device_info)); + + mutex_lock(&target->event_mutex); + + /* Run over all pdd of the process */ + for (i = 0; i < tmp_num_devices; i++) { + struct kfd_process_device *pdd = target->pdds[i]; + struct kfd_topology_device *topo_dev = kfd_topology_device_by_id(pdd->dev->id); + + device_info.gpu_id = pdd->dev->id; + device_info.exception_status = pdd->exception_status; + device_info.lds_base = pdd->lds_base; + device_info.lds_limit = pdd->lds_limit; + device_info.scratch_base = pdd->scratch_base; + device_info.scratch_limit = pdd->scratch_limit; + device_info.gpuvm_base = pdd->gpuvm_base; + device_info.gpuvm_limit = pdd->gpuvm_limit; + device_info.location_id = topo_dev->node_props.location_id; + device_info.vendor_id = topo_dev->node_props.vendor_id; + device_info.device_id = topo_dev->node_props.device_id; + device_info.revision_id = pdd->dev->adev->pdev->revision; + device_info.subsystem_vendor_id = pdd->dev->adev->pdev->subsystem_vendor; + device_info.subsystem_device_id = pdd->dev->adev->pdev->subsystem_device; + device_info.fw_version = pdd->dev->kfd->mec_fw_version; + device_info.gfx_target_version = + topo_dev->node_props.gfx_target_version; + device_info.simd_count = topo_dev->node_props.simd_count; + device_info.max_waves_per_simd = + topo_dev->node_props.max_waves_per_simd; + device_info.array_count = topo_dev->node_props.array_count; + device_info.simd_arrays_per_engine = + topo_dev->node_props.simd_arrays_per_engine; + device_info.num_xcc = NUM_XCC(pdd->dev->xcc_mask); + device_info.capability = topo_dev->node_props.capability; + device_info.debug_prop = topo_dev->node_props.debug_prop; + + if (exception_clear_mask) + pdd->exception_status &= ~exception_clear_mask; + + if (copy_to_user(user_info, &device_info, *entry_size)) { + r = -EFAULT; + break; + } + + user_info += tmp_entry_size; + } + + mutex_unlock(&target->event_mutex); + + return r; +} + void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, uint64_t exception_set_mask) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index 234e2ccafa87..a289e59ceb79 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -81,6 +81,11 @@ static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) } void debug_event_write_work_handler(struct work_struct *work); +int kfd_dbg_trap_device_snapshot(struct kfd_process *target, + uint64_t exception_clear_mask, + void __user *user_info, + uint32_t *number_of_device_infos, + uint32_t *entry_size); void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, uint64_t exception_set_mask); -- cgit v1.2.3 From a159afdad2f6b97e4d18549cff2b53d17e68a412 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Tue, 10 May 2022 12:51:26 -0400 Subject: drm/amdkfd: bump kfd ioctl minor version for debug api availability Bump the minor version to declare debugging capability is now available. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 1 - include/uapi/linux/kfd_ioctl.h | 3 ++- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index f54ff5c3387d..cce2abe12e1b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2984,7 +2984,6 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v if (!r) target->exception_enable_mask = args->enable.exception_mask; - pr_warn("Debug functions limited\n"); break; case KFD_IOC_DBG_TRAP_DISABLE: r = kfd_dbg_trap_disable(target); diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h index 32913d674d38..1781e7669982 100644 --- a/include/uapi/linux/kfd_ioctl.h +++ b/include/uapi/linux/kfd_ioctl.h @@ -38,9 +38,10 @@ * - 1.10 - Add SMI profiler event log * - 1.11 - Add unified memory for ctx save/restore area * - 1.12 - Add DMA buf export ioctl + * - 1.13 - Add debugger API */ #define KFD_IOCTL_MAJOR_VERSION 1 -#define KFD_IOCTL_MINOR_VERSION 12 +#define KFD_IOCTL_MINOR_VERSION 13 struct kfd_ioctl_get_version_args { __u32 major_version; /* from KFD */ -- cgit v1.2.3 From 8be295046748432c53a2dee39c469f63c60b0ec3 Mon Sep 17 00:00:00 2001 From: Dan Carpenter Date: Tue, 6 Jun 2023 19:29:51 -0400 Subject: drm/amdkfd: potential error pointer dereference in ioctl The "target" either comes from kfd_create_process() which returns error pointers on error or kfd_lookup_process_by_pid() which returns NULL on error. So we need to check for both types of errors. Fixes: 0ab2d7532b05 ("drm/amdkfd: prepare per-process debug enable and disable") Signed-off-by: Dan Carpenter Reviewed-by: Jonathan Kim Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index cce2abe12e1b..d655c5bc951f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -2920,9 +2920,9 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v target = kfd_lookup_process_by_pid(pid); } - if (!target) { + if (IS_ERR_OR_NULL(target)) { pr_debug("Cannot find process PID %i to debug\n", args->pid); - r = -ESRCH; + r = target ? PTR_ERR(target) : -ESRCH; goto out; } -- cgit v1.2.3 From 7386f88ab1732af890a09ab3a7f400bb20adbe5a Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Wed, 7 Jun 2023 03:56:27 -0400 Subject: drm/amdkfd: fix vmfault signalling with additional data. Exception handling for vmfaults should be raised with additional data. Reported-by: Mukul Joshi Signed-off-by: Jonathan Kim Reviewed-by: Mukul Joshi Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_events.c | 34 +++++++++++++++++++-------------- 1 file changed, 20 insertions(+), 14 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c b/drivers/gpu/drm/amd/amdkfd/kfd_events.c index 0a5e7b172a64..ddca23ee4193 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c @@ -1240,19 +1240,24 @@ void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, return; } - memset(&memory_exception_data, 0, sizeof(memory_exception_data)); - memory_exception_data.gpu_id = user_gpu_id; - memory_exception_data.failure.imprecise = true; - /* Set failure reason */ - if (info) { - memory_exception_data.va = (info->page_addr) << PAGE_SHIFT; - memory_exception_data.failure.NotPresent = - info->prot_valid ? 1 : 0; - memory_exception_data.failure.NoExecute = - info->prot_exec ? 1 : 0; - memory_exception_data.failure.ReadOnly = - info->prot_write ? 1 : 0; - memory_exception_data.failure.imprecise = 0; + /* SoC15 chips and onwards will pass in data from now on. */ + if (!data) { + memset(&memory_exception_data, 0, sizeof(memory_exception_data)); + memory_exception_data.gpu_id = user_gpu_id; + memory_exception_data.failure.imprecise = true; + + /* Set failure reason */ + if (info) { + memory_exception_data.va = (info->page_addr) << + PAGE_SHIFT; + memory_exception_data.failure.NotPresent = + info->prot_valid ? 1 : 0; + memory_exception_data.failure.NoExecute = + info->prot_exec ? 1 : 0; + memory_exception_data.failure.ReadOnly = + info->prot_write ? 1 : 0; + memory_exception_data.failure.imprecise = 0; + } } rcu_read_lock(); @@ -1261,7 +1266,8 @@ void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid, idr_for_each_entry_continue(&p->event_idr, ev, id) if (ev->type == KFD_EVENT_TYPE_MEMORY) { spin_lock(&ev->lock); - ev->memory_exception_data = memory_exception_data; + ev->memory_exception_data = data ? *data : + memory_exception_data; set_event(ev); spin_unlock(&ev->lock); } -- cgit v1.2.3 From bbcc3514ab4f7ec3ae2273ad08b0a1b6b4aa9dd9 Mon Sep 17 00:00:00 2001 From: Mario Limonciello Date: Fri, 2 Jun 2023 07:18:06 -0500 Subject: drm/amd: Check that a system is a NUMA system before looking for SRAT It's pointless on laptops to look for the SRAT table as these are not NUMA. Check the number of possible nodes is > 1 to decide whether to look for SRAT. Suggested-by: Felix Kuehling Signed-off-by: Mario Limonciello Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c index 950af6820153..3dcd8f8bc98e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c @@ -2041,7 +2041,8 @@ static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size, sub_type_hdr->proximity_domain_from = proximity_domain; #ifdef CONFIG_ACPI_NUMA - if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE) + if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE && + num_possible_nodes() > 1) kfd_find_numa_node_in_srat(kdev); #endif #ifdef CONFIG_NUMA -- cgit v1.2.3 From 597364adc0fcf71617b3adbe647b6eec76e27554 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Wed, 31 May 2023 11:22:03 -0400 Subject: drm/amdkfd: Fix reserved SDMA queues handling This patch fixes a regression caused by a bad merge where the handling of reserved SDMA queues was accidentally removed. With the fix, the reserved SDMA queues are again correctly marked as unavailable for allocation. Fixes: a805889a1531 ("drm/amdkfd: Update SDMA queue management for GFX9.4.3") Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 13 ++++++------- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10 +++++----- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +- 3 files changed, 12 insertions(+), 13 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c index 9fc9d32cb579..9d4abfd8b55e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c @@ -106,20 +106,19 @@ static void kfd_device_info_set_sdma_info(struct kfd_dev *kfd) kfd->device_info.num_sdma_queues_per_engine = 8; } + bitmap_zero(kfd->device_info.reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES); + switch (sdma_version) { case IP_VERSION(6, 0, 0): + case IP_VERSION(6, 0, 1): case IP_VERSION(6, 0, 2): case IP_VERSION(6, 0, 3): /* Reserve 1 for paging and 1 for gfx */ kfd->device_info.num_reserved_sdma_queues_per_engine = 2; /* BIT(0)=engine-0 queue-0; BIT(1)=engine-1 queue-0; BIT(2)=engine-0 queue-1; ... */ - kfd->device_info.reserved_sdma_queues_bitmap = 0xFULL; - break; - case IP_VERSION(6, 0, 1): - /* Reserve 1 for paging and 1 for gfx */ - kfd->device_info.num_reserved_sdma_queues_per_engine = 2; - /* BIT(0)=engine-0 queue-0; BIT(1)=engine-0 queue-1; ... */ - kfd->device_info.reserved_sdma_queues_bitmap = 0x3ULL; + bitmap_set(kfd->device_info.reserved_sdma_queues_bitmap, 0, + kfd->adev->sdma.num_instances * + kfd->device_info.num_reserved_sdma_queues_per_engine); break; default: break; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 0c1be91a87c6..498ad7d4e7d9 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -123,11 +123,6 @@ unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm) dqm->dev->kfd->device_info.num_sdma_queues_per_engine; } -static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manager *dqm) -{ - return dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap; -} - static void init_sdma_bitmaps(struct device_queue_manager *dqm) { bitmap_zero(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES); @@ -135,6 +130,11 @@ static void init_sdma_bitmaps(struct device_queue_manager *dqm) bitmap_zero(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES); bitmap_set(dqm->xgmi_sdma_bitmap, 0, get_num_xgmi_sdma_queues(dqm)); + + /* Mask out the reserved queues */ + bitmap_andnot(dqm->sdma_bitmap, dqm->sdma_bitmap, + dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap, + KFD_MAX_SDMA_QUEUES); } void program_sh_mem_settings(struct device_queue_manager *dqm, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 023b17e0116b..7364a5d77c6e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -239,7 +239,7 @@ struct kfd_device_info { uint32_t no_atomic_fw_version; unsigned int num_sdma_queues_per_engine; unsigned int num_reserved_sdma_queues_per_engine; - uint64_t reserved_sdma_queues_bitmap; + DECLARE_BITMAP(reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES); }; unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev); -- cgit v1.2.3 From 09d49e14ea6fd125a21f89b80f888c09be32a174 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Tue, 23 May 2023 11:57:27 -0400 Subject: drm/amdkfd: fix and enable debugging for gfx11 There are a couple of fixes required to enable gfx11 debugging. First, ADD_QUEUE.trap_en is an inappropriate place to toggle a per-process register so move it to SET_SHADER_DEBUGGER.trap_en. When ADD_QUEUE.skip_process_ctx_clear is set, MES will prioritize the SET_SHADER_DEBUGGER.trap_en setting. Second, to preserve correct save/restore priviledged wave states in coordination with the trap enablement setting, resume suspended waves early in the disable call. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 7 ++++++- drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 4 +++- drivers/gpu/drm/amd/amdgpu/mes_v11_0.c | 1 + drivers/gpu/drm/amd/amdkfd/kfd_debug.c | 14 ++++++-------- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 3 +-- drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 12 +++++++----- drivers/gpu/drm/amd/include/mes_v11_api_def.h | 1 + 7 files changed, 25 insertions(+), 17 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c index 20cc3fffe921..e9091ebfe230 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c @@ -928,7 +928,8 @@ int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev, uint64_t process_context_addr, uint32_t spi_gdbg_per_vmid_cntl, const uint32_t *tcp_watch_cntl, - uint32_t flags) + uint32_t flags, + bool trap_en) { struct mes_misc_op_input op_input = {0}; int r; @@ -945,6 +946,10 @@ int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev, memcpy(op_input.set_shader_debugger.tcp_watch_cntl, tcp_watch_cntl, sizeof(op_input.set_shader_debugger.tcp_watch_cntl)); + if (((adev->mes.sched_version & AMDGPU_MES_API_VERSION_MASK) >> + AMDGPU_MES_API_VERSION_SHIFT) >= 14) + op_input.set_shader_debugger.trap_en = trap_en; + amdgpu_mes_lock(&adev->mes); r = adev->mes.funcs->misc_op(&adev->mes, &op_input); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h index b5f5eed2b5ef..2d6ac30b7135 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h @@ -294,6 +294,7 @@ struct mes_misc_op_input { } flags; uint32_t spi_gdbg_per_vmid_cntl; uint32_t tcp_watch_cntl[4]; + uint32_t trap_en; } set_shader_debugger; }; }; @@ -361,7 +362,8 @@ int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev, uint64_t process_context_addr, uint32_t spi_gdbg_per_vmid_cntl, const uint32_t *tcp_watch_cntl, - uint32_t flags); + uint32_t flags, + bool trap_en); int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id, int queue_type, int idx, diff --git a/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c b/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c index c4e3cb8d44de..1bdaa00c0b46 100644 --- a/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +++ b/drivers/gpu/drm/amd/amdgpu/mes_v11_0.c @@ -347,6 +347,7 @@ static int mes_v11_0_misc_op(struct amdgpu_mes *mes, memcpy(misc_pkt.set_shader_debugger.tcp_watch_cntl, input->set_shader_debugger.tcp_watch_cntl, sizeof(misc_pkt.set_shader_debugger.tcp_watch_cntl)); + misc_pkt.set_shader_debugger.trap_en = input->set_shader_debugger.trap_en; break; default: DRM_ERROR("unsupported misc op (%d) \n", input->op); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c index 125274445f43..cd34e7aaead4 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.c @@ -349,12 +349,13 @@ int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) { uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode; uint32_t flags = pdd->process->dbg_flags; + bool sq_trap_en = !!spi_dbg_cntl; if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) return 0; return amdgpu_mes_set_shader_debugger(pdd->dev->adev, pdd->proc_ctx_gpu_addr, spi_dbg_cntl, - pdd->watch_points, flags); + pdd->watch_points, flags, sq_trap_en); } #define KFD_DEBUGGER_INVALID_WATCH_POINT_ID -1 @@ -557,6 +558,10 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind if (!unwind) { uint32_t flags = 0; + int resume_count = resume_queues(target, 0, NULL); + + if (resume_count) + pr_debug("Resumed %d queues\n", resume_count); cancel_work_sync(&target->debug_event_workarea); kfd_dbg_clear_process_address_watch(target); @@ -598,13 +603,6 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind } kfd_dbg_set_workaround(target, false); - - if (!unwind) { - int resume_count = resume_queues(target, 0, NULL); - - if (resume_count) - pr_debug("Resumed %d queues\n", resume_count); - } } static void kfd_dbg_clean_exception_status(struct kfd_process *target) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 498ad7d4e7d9..d6b15493fffd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -227,8 +227,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, queue_input.tba_addr = qpd->tba_addr; queue_input.tma_addr = qpd->tma_addr; queue_input.trap_en = KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) || - KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0) || - q->properties.is_dbg_wa; + KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0); queue_input.skip_process_ctx_clear = qpd->pqm->process->debug_trap_enabled; queue_type = convert_to_mes_queue_type(q->properties.type); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c index faa7939f35bd..90b86a6ac7bd 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c @@ -1863,13 +1863,15 @@ static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *de { bool firmware_supported = true; - /* - * FIXME: GFX11 FW currently not sufficient to deal with CWSR WA. - * Updated FW with API changes coming soon. - */ if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) && KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) { - firmware_supported = false; + uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version & + AMDGPU_MES_API_VERSION_MASK) >> + AMDGPU_MES_API_VERSION_SHIFT; + uint32_t mes_rev = dev->gpu->adev->mes.sched_version & + AMDGPU_MES_VERSION_MASK; + + firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64); goto out; } diff --git a/drivers/gpu/drm/amd/include/mes_v11_api_def.h b/drivers/gpu/drm/amd/include/mes_v11_api_def.h index f3c15f18ddb5..0997e999416a 100644 --- a/drivers/gpu/drm/amd/include/mes_v11_api_def.h +++ b/drivers/gpu/drm/amd/include/mes_v11_api_def.h @@ -575,6 +575,7 @@ struct SET_SHADER_DEBUGGER { } flags; uint32_t spi_gdbg_per_vmid_cntl; uint32_t tcp_watch_cntl[4]; /* TCP_WATCHx_CNTL */ + uint32_t trap_en; }; union MESAPI__MISC { -- cgit v1.2.3