From d1257d098a5a38753a0736a50db0a26a62377ad7 Mon Sep 17 00:00:00 2001 From: Dan Williams Date: Fri, 16 Jun 2023 18:24:28 -0700 Subject: cxl/region: Move cache invalidation before region teardown, and before setup Vikram raised a concern with the theoretical case of a CPU sending MemClnEvict to a device that is not prepared to receive. MemClnEvict is a message that is sent after a CPU has taken ownership of a cacheline from accelerator memory (HDM-DB). In the case of hotplug or HDM decoder reconfiguration it is possible that the CPU is holding old contents for a new device that has taken over the physical address range being cached by the CPU. To avoid this scenario, invalidate caches prior to tearing down an HDM decoder configuration. Now, this poses another problem that it is possible for something to speculate into that space while the decode configuration is still up, so to close that gap also invalidate prior to establish new contents behind a given physical address range. With this change the cache invalidation is now explicit and need not be checked in cxl_region_probe(), and that obviates the need for CXL_REGION_F_INCOHERENT. Cc: Jonathan Cameron Fixes: d18bc74aced6 ("cxl/region: Manage CPU caches relative to DPA invalidation events") Reported-by: Vikram Sethi Closes: http://lore.kernel.org/r/BYAPR12MB33364B5EB908BF7239BB996BBD53A@BYAPR12MB3336.namprd12.prod.outlook.com Reviewed-by: Jonathan Cameron Reviewed-by: Dave Jiang Link: https://lore.kernel.org/r/168696506886.3590522.4597053660991916591.stgit@dwillia2-xfh.jf.intel.com Signed-off-by: Dan Williams --- drivers/cxl/core/region.c | 66 ++++++++++++++++++++++++++--------------------- drivers/cxl/cxl.h | 8 +----- 2 files changed, 38 insertions(+), 36 deletions(-) (limited to 'drivers/cxl') diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c index f822de44bee0..594ce3c2565d 100644 --- a/drivers/cxl/core/region.c +++ b/drivers/cxl/core/region.c @@ -125,10 +125,38 @@ static struct cxl_region_ref *cxl_rr_load(struct cxl_port *port, return xa_load(&port->regions, (unsigned long)cxlr); } +static int cxl_region_invalidate_memregion(struct cxl_region *cxlr) +{ + if (!cpu_cache_has_invalidate_memregion()) { + if (IS_ENABLED(CONFIG_CXL_REGION_INVALIDATION_TEST)) { + dev_warn_once( + &cxlr->dev, + "Bypassing cpu_cache_invalidate_memregion() for testing!\n"); + return 0; + } else { + dev_err(&cxlr->dev, + "Failed to synchronize CPU cache state\n"); + return -ENXIO; + } + } + + cpu_cache_invalidate_memregion(IORES_DESC_CXL); + return 0; +} + static int cxl_region_decode_reset(struct cxl_region *cxlr, int count) { struct cxl_region_params *p = &cxlr->params; - int i; + int i, rc = 0; + + /* + * Before region teardown attempt to flush, and if the flush + * fails cancel the region teardown for data consistency + * concerns + */ + rc = cxl_region_invalidate_memregion(cxlr); + if (rc) + return rc; for (i = count - 1; i >= 0; i--) { struct cxl_endpoint_decoder *cxled = p->targets[i]; @@ -136,7 +164,6 @@ static int cxl_region_decode_reset(struct cxl_region *cxlr, int count) struct cxl_port *iter = cxled_to_port(cxled); struct cxl_dev_state *cxlds = cxlmd->cxlds; struct cxl_ep *ep; - int rc = 0; if (cxlds->rcd) goto endpoint_reset; @@ -256,6 +283,14 @@ static ssize_t commit_store(struct device *dev, struct device_attribute *attr, goto out; } + /* + * Invalidate caches before region setup to drop any speculative + * consumption of this address space + */ + rc = cxl_region_invalidate_memregion(cxlr); + if (rc) + return rc; + if (commit) rc = cxl_region_decode_commit(cxlr); else { @@ -1674,7 +1709,6 @@ static int cxl_region_attach(struct cxl_region *cxlr, if (rc) goto err_decrement; p->state = CXL_CONFIG_ACTIVE; - set_bit(CXL_REGION_F_INCOHERENT, &cxlr->flags); } cxled->cxld.interleave_ways = p->interleave_ways; @@ -2803,30 +2837,6 @@ out: } EXPORT_SYMBOL_NS_GPL(cxl_add_to_region, CXL); -static int cxl_region_invalidate_memregion(struct cxl_region *cxlr) -{ - if (!test_bit(CXL_REGION_F_INCOHERENT, &cxlr->flags)) - return 0; - - if (!cpu_cache_has_invalidate_memregion()) { - if (IS_ENABLED(CONFIG_CXL_REGION_INVALIDATION_TEST)) { - dev_warn_once( - &cxlr->dev, - "Bypassing cpu_cache_invalidate_memregion() for testing!\n"); - clear_bit(CXL_REGION_F_INCOHERENT, &cxlr->flags); - return 0; - } else { - dev_err(&cxlr->dev, - "Failed to synchronize CPU cache state\n"); - return -ENXIO; - } - } - - cpu_cache_invalidate_memregion(IORES_DESC_CXL); - clear_bit(CXL_REGION_F_INCOHERENT, &cxlr->flags); - return 0; -} - static int is_system_ram(struct resource *res, void *arg) { struct cxl_region *cxlr = arg; @@ -2854,8 +2864,6 @@ static int cxl_region_probe(struct device *dev) goto out; } - rc = cxl_region_invalidate_memregion(cxlr); - /* * From this point on any path that changes the region's state away from * CXL_CONFIG_COMMIT is also responsible for releasing the driver. diff --git a/drivers/cxl/cxl.h b/drivers/cxl/cxl.h index f93a28538962..21ee135d91b6 100644 --- a/drivers/cxl/cxl.h +++ b/drivers/cxl/cxl.h @@ -462,18 +462,12 @@ struct cxl_region_params { int nr_targets; }; -/* - * Flag whether this region needs to have its HPA span synchronized with - * CPU cache state at region activation time. - */ -#define CXL_REGION_F_INCOHERENT 0 - /* * Indicate whether this region has been assembled by autodetection or * userspace assembly. Prevent endpoint decoders outside of automatic * detection from being added to the region. */ -#define CXL_REGION_F_AUTO 1 +#define CXL_REGION_F_AUTO 0 /** * struct cxl_region - CXL region -- cgit v1.2.3 From 2ab47045ac96a605e3037d479a7d5854570ee5bf Mon Sep 17 00:00:00 2001 From: Dan Williams Date: Fri, 16 Jun 2023 18:24:34 -0700 Subject: cxl/region: Flag partially torn down regions as unusable cxl_region_decode_reset() walks all the decoders associated with a given region and disables them. Due to decoder ordering rules it is possible that a switch in the topology notices that a given decoder can not be shutdown before another region with a higher HPA is shutdown first. That can leave the region in a partially committed state. Capture that state in a new CXL_REGION_F_NEEDS_RESET flag and require that a successful cxl_region_decode_reset() attempt must be completed before cxl_region_probe() accepts the region. This is a corollary for the bug that Jonathan identified in "CXL/region : commit reset of out of order region appears to succeed." [1]. Cc: Jonathan Cameron Link: http://lore.kernel.org/r/20230316171441.0000205b@Huawei.com [1] Fixes: 176baefb2eb5 ("cxl/hdm: Commit decoder state to hardware") Reviewed-by: Dave Jiang Link: https://lore.kernel.org/r/168696507423.3590522.16254212607926684429.stgit@dwillia2-xfh.jf.intel.com Signed-off-by: Dan Williams --- drivers/cxl/core/region.c | 12 ++++++++++++ drivers/cxl/cxl.h | 8 ++++++++ 2 files changed, 20 insertions(+) (limited to 'drivers/cxl') diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c index 594ce3c2565d..fa29bd2ec322 100644 --- a/drivers/cxl/core/region.c +++ b/drivers/cxl/core/region.c @@ -182,14 +182,19 @@ static int cxl_region_decode_reset(struct cxl_region *cxlr, int count) rc = cxld->reset(cxld); if (rc) return rc; + set_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags); } endpoint_reset: rc = cxled->cxld.reset(&cxled->cxld); if (rc) return rc; + set_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags); } + /* all decoders associated with this region have been torn down */ + clear_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags); + return 0; } @@ -2864,6 +2869,13 @@ static int cxl_region_probe(struct device *dev) goto out; } + if (test_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags)) { + dev_err(&cxlr->dev, + "failed to activate, re-commit region and retry\n"); + rc = -ENXIO; + goto out; + } + /* * From this point on any path that changes the region's state away from * CXL_CONFIG_COMMIT is also responsible for releasing the driver. diff --git a/drivers/cxl/cxl.h b/drivers/cxl/cxl.h index 21ee135d91b6..492673d04da2 100644 --- a/drivers/cxl/cxl.h +++ b/drivers/cxl/cxl.h @@ -469,6 +469,14 @@ struct cxl_region_params { */ #define CXL_REGION_F_AUTO 0 +/* + * Require that a committed region successfully complete a teardown once + * any of its associated decoders have been torn down. This maintains + * the commit state for the region since there are committed decoders, + * but blocks cxl_region_probe(). + */ +#define CXL_REGION_F_NEEDS_RESET 1 + /** * struct cxl_region - CXL region * @dev: This region's device -- cgit v1.2.3 From adfe19738b71a893da62cb2e30bd6bdb4299ea67 Mon Sep 17 00:00:00 2001 From: Dan Williams Date: Fri, 16 Jun 2023 18:24:39 -0700 Subject: cxl/region: Fix state transitions after reset failure Jonathan reports that failed attempts to reset a region (teardown its HDM decoder configuration) mistakenly advance the state of the region to "not committed". Revert to the previous state of the region on reset failure so that the reset can be re-attempted. Reported-by: Jonathan Cameron Closes: http://lore.kernel.org/r/20230316171441.0000205b@Huawei.com Fixes: 176baefb2eb5 ("cxl/hdm: Commit decoder state to hardware") Reviewed-by: Jonathan Cameron Reviewed-by: Dave Jiang Link: https://lore.kernel.org/r/168696507968.3590522.14484000711718573626.stgit@dwillia2-xfh.jf.intel.com Signed-off-by: Dan Williams --- drivers/cxl/core/region.c | 26 +++++++++++++++----------- 1 file changed, 15 insertions(+), 11 deletions(-) (limited to 'drivers/cxl') diff --git a/drivers/cxl/core/region.c b/drivers/cxl/core/region.c index fa29bd2ec322..bfdd424d6897 100644 --- a/drivers/cxl/core/region.c +++ b/drivers/cxl/core/region.c @@ -296,9 +296,11 @@ static ssize_t commit_store(struct device *dev, struct device_attribute *attr, if (rc) return rc; - if (commit) + if (commit) { rc = cxl_region_decode_commit(cxlr); - else { + if (rc == 0) + p->state = CXL_CONFIG_COMMIT; + } else { p->state = CXL_CONFIG_RESET_PENDING; up_write(&cxl_region_rwsem); device_release_driver(&cxlr->dev); @@ -308,18 +310,20 @@ static ssize_t commit_store(struct device *dev, struct device_attribute *attr, * The lock was dropped, so need to revalidate that the reset is * still pending. */ - if (p->state == CXL_CONFIG_RESET_PENDING) + if (p->state == CXL_CONFIG_RESET_PENDING) { rc = cxl_region_decode_reset(cxlr, p->interleave_ways); + /* + * Revert to committed since there may still be active + * decoders associated with this region, or move forward + * to active to mark the reset successful + */ + if (rc) + p->state = CXL_CONFIG_COMMIT; + else + p->state = CXL_CONFIG_ACTIVE; + } } - if (rc) - goto out; - - if (commit) - p->state = CXL_CONFIG_COMMIT; - else if (p->state == CXL_CONFIG_RESET_PENDING) - p->state = CXL_CONFIG_ACTIVE; - out: up_write(&cxl_region_rwsem); -- cgit v1.2.3