scsi: lpfc: Fix MRQ > 1 context list handling

Various oops including cpu LOCKUPs were seen.

For asynchronously received ius where the driver must assign exchange
resources, the resources were on a single get (free) list and put list
(finished, waiting to be put on get list). As all cpus are sharing the
lists, an interrupt for a receive frame may have to wait for all the
other cpus to place their done work onto the put list before it can
acquire the lock to pull from the list.

Fix by breaking the resource lists into per-cpu lists or at least more
than 1 list with cpu's sharing the lists). A cpu would allocate from the
free list for its own cpu, and put its done work on the its own put list
- avoiding the contention. As cpu load may vary, when empty, a cpu may
grab from another cpu, thereby changing resource distribution.  But
searching for a resource only occurs on 1 or a few cpus until a single
resource can be allocated. if the condition reoccurs, it starts looking
at a different cpu.

Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

1 files changed, 202 insertions, 62 deletions

diff --git a/drivers/scsi/lpfc/lpfc_nvmet.c b/drivers/scsi/lpfc/lpfc_nvmet.c
index 474c93a85f41..41abdef27909 100644
--- a/drivers/scsi/lpfc/lpfc_nvmet.c
+++ b/drivers/scsi/lpfc/lpfc_nvmet.c
@@ -170,8 +170,10 @@ lpfc_nvmet_ctxbuf_post(struct lpfc_hba *phba, struct lpfc_nvmet_ctxbuf *ctx_buf)
 	struct lpfc_nvmet_tgtport *tgtp;
 	struct fc_frame_header *fc_hdr;
 	struct rqb_dmabuf *nvmebuf;
+	struct lpfc_nvmet_ctx_info *infop;
 	uint32_t *payload;
 	uint32_t size, oxid, sid, rc;
+	int cpu;
 	unsigned long iflag;
 
 	if (ctxp->txrdy) {
@@ -267,11 +269,16 @@ lpfc_nvmet_ctxbuf_post(struct lpfc_hba *phba, struct lpfc_nvmet_ctxbuf *ctx_buf)
 	}
 	spin_unlock_irqrestore(&phba->sli4_hba.nvmet_io_wait_lock, iflag);
 
-	spin_lock_irqsave(&phba->sli4_hba.nvmet_ctx_put_lock, iflag);
-	list_add_tail(&ctx_buf->list,
-		      &phba->sli4_hba.lpfc_nvmet_ctx_put_list);
-	phba->sli4_hba.nvmet_ctx_put_cnt++;
-	spin_unlock_irqrestore(&phba->sli4_hba.nvmet_ctx_put_lock, iflag);
+	/*
+	 * Use the CPU context list, from the MRQ the IO was received on
+	 * (ctxp->idx), to save context structure.
+	 */
+	cpu = smp_processor_id();
+	infop = lpfc_get_ctx_list(phba, cpu, ctxp->idx);
+	spin_lock_irqsave(&infop->nvmet_ctx_list_lock, iflag);
+	list_add_tail(&ctx_buf->list, &infop->nvmet_ctx_list);
+	infop->nvmet_ctx_list_cnt++;
+	spin_unlock_irqrestore(&infop->nvmet_ctx_list_lock, iflag);
 #endif
 }
 
@@ -860,51 +867,54 @@ static struct nvmet_fc_target_template lpfc_tgttemplate = {
 };
 
 static void
-lpfc_nvmet_cleanup_io_context(struct lpfc_hba *phba)
+__lpfc_nvmet_clean_io_for_cpu(struct lpfc_hba *phba,
+		struct lpfc_nvmet_ctx_info *infop)
 {
 	struct lpfc_nvmet_ctxbuf *ctx_buf, *next_ctx_buf;
 	unsigned long flags;
 
-	spin_lock_irqsave(&phba->sli4_hba.nvmet_ctx_get_lock, flags);
-	spin_lock(&phba->sli4_hba.nvmet_ctx_put_lock);
+	spin_lock_irqsave(&infop->nvmet_ctx_list_lock, flags);
 	list_for_each_entry_safe(ctx_buf, next_ctx_buf,
-			&phba->sli4_hba.lpfc_nvmet_ctx_get_list, list) {
+				&infop->nvmet_ctx_list, list) {
 		spin_lock(&phba->sli4_hba.abts_nvme_buf_list_lock);
 		list_del_init(&ctx_buf->list);
 		spin_unlock(&phba->sli4_hba.abts_nvme_buf_list_lock);
-		__lpfc_clear_active_sglq(phba,
-					 ctx_buf->sglq->sli4_lxritag);
+
+		__lpfc_clear_active_sglq(phba, ctx_buf->sglq->sli4_lxritag);
 		ctx_buf->sglq->state = SGL_FREED;
 		ctx_buf->sglq->ndlp = NULL;
 
 		spin_lock(&phba->sli4_hba.sgl_list_lock);
 		list_add_tail(&ctx_buf->sglq->list,
-			      &phba->sli4_hba.lpfc_nvmet_sgl_list);
+				&phba->sli4_hba.lpfc_nvmet_sgl_list);
 		spin_unlock(&phba->sli4_hba.sgl_list_lock);
 
 		lpfc_sli_release_iocbq(phba, ctx_buf->iocbq);
 		kfree(ctx_buf->context);
 	}
-	list_for_each_entry_safe(ctx_buf, next_ctx_buf,
-			&phba->sli4_hba.lpfc_nvmet_ctx_put_list, list) {
-		spin_lock(&phba->sli4_hba.abts_nvme_buf_list_lock);
-		list_del_init(&ctx_buf->list);
-		spin_unlock(&phba->sli4_hba.abts_nvme_buf_list_lock);
-		__lpfc_clear_active_sglq(phba,
-					 ctx_buf->sglq->sli4_lxritag);
-		ctx_buf->sglq->state = SGL_FREED;
-		ctx_buf->sglq->ndlp = NULL;
+	spin_unlock_irqrestore(&infop->nvmet_ctx_list_lock, flags);
+}
 
-		spin_lock(&phba->sli4_hba.sgl_list_lock);
-		list_add_tail(&ctx_buf->sglq->list,
-			      &phba->sli4_hba.lpfc_nvmet_sgl_list);
-		spin_unlock(&phba->sli4_hba.sgl_list_lock);
+static void
+lpfc_nvmet_cleanup_io_context(struct lpfc_hba *phba)
+{
+	struct lpfc_nvmet_ctx_info *infop;
+	int i, j;
 
-		lpfc_sli_release_iocbq(phba, ctx_buf->iocbq);
-		kfree(ctx_buf->context);
+	/* The first context list, MRQ 0 CPU 0 */
+	infop = phba->sli4_hba.nvmet_ctx_info;
+	if (!infop)
+		return;
+
+	/* Cycle the the entire CPU context list for every MRQ */
+	for (i = 0; i < phba->cfg_nvmet_mrq; i++) {
+		for (j = 0; j < phba->sli4_hba.num_present_cpu; j++) {
+			__lpfc_nvmet_clean_io_for_cpu(phba, infop);
+			infop++; /* next */
+		}
 	}
-	spin_unlock(&phba->sli4_hba.nvmet_ctx_put_lock);
-	spin_unlock_irqrestore(&phba->sli4_hba.nvmet_ctx_get_lock, flags);
+	kfree(phba->sli4_hba.nvmet_ctx_info);
+	phba->sli4_hba.nvmet_ctx_info = NULL;
 }
 
 static int
@@ -913,15 +923,71 @@ lpfc_nvmet_setup_io_context(struct lpfc_hba *phba)
 	struct lpfc_nvmet_ctxbuf *ctx_buf;
 	struct lpfc_iocbq *nvmewqe;
 	union lpfc_wqe128 *wqe;
-	int i;
+	struct lpfc_nvmet_ctx_info *last_infop;
+	struct lpfc_nvmet_ctx_info *infop;
+	int i, j, idx;
 
 	lpfc_printf_log(phba, KERN_INFO, LOG_NVME,
 			"6403 Allocate NVMET resources for %d XRIs\n",
 			phba->sli4_hba.nvmet_xri_cnt);
 
+	phba->sli4_hba.nvmet_ctx_info = kcalloc(
+		phba->sli4_hba.num_present_cpu * phba->cfg_nvmet_mrq,
+		sizeof(struct lpfc_nvmet_ctx_info), GFP_KERNEL);
+	if (!phba->sli4_hba.nvmet_ctx_info) {
+		lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
+				"6419 Failed allocate memory for "
+				"nvmet context lists\n");
+		return -ENOMEM;
+	}
+
+	/*
+	 * Assuming X CPUs in the system, and Y MRQs, allocate some
+	 * lpfc_nvmet_ctx_info structures as follows:
+	 *
+	 * cpu0/mrq0 cpu1/mrq0 ... cpuX/mrq0
+	 * cpu0/mrq1 cpu1/mrq1 ... cpuX/mrq1
+	 * ...
+	 * cpuX/mrqY cpuX/mrqY ... cpuX/mrqY
+	 *
+	 * Each line represents a MRQ "silo" containing an entry for
+	 * every CPU.
+	 *
+	 * MRQ X is initially assumed to be associated with CPU X, thus
+	 * contexts are initially distributed across all MRQs using
+	 * the MRQ index (N) as follows cpuN/mrqN. When contexts are
+	 * freed, the are freed to the MRQ silo based on the CPU number
+	 * of the IO completion. Thus a context that was allocated for MRQ A
+	 * whose IO completed on CPU B will be freed to cpuB/mrqA.
+	 */
+	infop = phba->sli4_hba.nvmet_ctx_info;
+	for (i = 0; i < phba->sli4_hba.num_present_cpu; i++) {
+		for (j = 0; j < phba->cfg_nvmet_mrq; j++) {
+			INIT_LIST_HEAD(&infop->nvmet_ctx_list);
+			spin_lock_init(&infop->nvmet_ctx_list_lock);
+			infop->nvmet_ctx_list_cnt = 0;
+			infop++;
+		}
+	}
+
+	/*
+	 * Setup the next CPU context info ptr for each MRQ.
+	 * MRQ 0 will cycle thru CPUs 0 - X separately from
+	 * MRQ 1 cycling thru CPUs 0 - X, and so on.
+	 */
+	for (j = 0; j < phba->cfg_nvmet_mrq; j++) {
+		last_infop = lpfc_get_ctx_list(phba, 0, j);
+		for (i = phba->sli4_hba.num_present_cpu - 1;  i >= 0; i--) {
+			infop = lpfc_get_ctx_list(phba, i, j);
+			infop->nvmet_ctx_next_cpu = last_infop;
+			last_infop = infop;
+		}
+	}
+
 	/* For all nvmet xris, allocate resources needed to process a
 	 * received command on a per xri basis.
 	 */
+	idx = 0;
 	for (i = 0; i < phba->sli4_hba.nvmet_xri_cnt; i++) {
 		ctx_buf = kzalloc(sizeof(*ctx_buf), GFP_KERNEL);
 		if (!ctx_buf) {
@@ -976,12 +1042,35 @@ lpfc_nvmet_setup_io_context(struct lpfc_hba *phba)
 					"6407 Ran out of NVMET XRIs\n");
 			return -ENOMEM;
 		}
-		spin_lock(&phba->sli4_hba.nvmet_ctx_get_lock);
-		list_add_tail(&ctx_buf->list,
-			      &phba->sli4_hba.lpfc_nvmet_ctx_get_list);
-		spin_unlock(&phba->sli4_hba.nvmet_ctx_get_lock);
+
+		/*
+		 * Add ctx to MRQidx context list. Our initial assumption
+		 * is MRQidx will be associated with CPUidx. This association
+		 * can change on the fly.
+		 */
+		infop = lpfc_get_ctx_list(phba, idx, idx);
+		spin_lock(&infop->nvmet_ctx_list_lock);
+		list_add_tail(&ctx_buf->list, &infop->nvmet_ctx_list);
+		infop->nvmet_ctx_list_cnt++;
+		spin_unlock(&infop->nvmet_ctx_list_lock);
+
+		/* Spread ctx structures evenly across all MRQs */
+		idx++;
+		if (idx >= phba->cfg_nvmet_mrq)
+			idx = 0;
+	}
+
+	infop = phba->sli4_hba.nvmet_ctx_info;
+	for (j = 0; j < phba->cfg_nvmet_mrq; j++) {
+		for (i = 0; i < phba->sli4_hba.num_present_cpu; i++) {
+			lpfc_printf_log(phba, KERN_INFO, LOG_NVME | LOG_INIT,
+					"6408 TOTAL NVMET ctx for CPU %d "
+					"MRQ %d: cnt %d nextcpu %p\n",
+					i, j, infop->nvmet_ctx_list_cnt,
+					infop->nvmet_ctx_next_cpu);
+			infop++;
+		}
 	}
-	phba->sli4_hba.nvmet_ctx_get_cnt = phba->sli4_hba.nvmet_xri_cnt;
 	return 0;
 }
 
@@ -1346,10 +1435,65 @@ dropit:
 #endif
 }
 
+static struct lpfc_nvmet_ctxbuf *
+lpfc_nvmet_replenish_context(struct lpfc_hba *phba,
+			     struct lpfc_nvmet_ctx_info *current_infop)
+{
+	struct lpfc_nvmet_ctxbuf *ctx_buf = NULL;
+	struct lpfc_nvmet_ctx_info *get_infop;
+	int i;
+
+	/*
+	 * The current_infop for the MRQ a NVME command IU was received
+	 * on is empty. Our goal is to replenish this MRQs context
+	 * list from a another CPUs.
+	 *
+	 * First we need to pick a context list to start looking on.
+	 * nvmet_ctx_start_cpu has available context the last time
+	 * we needed to replenish this CPU where nvmet_ctx_next_cpu
+	 * is just the next sequential CPU for this MRQ.
+	 */
+	if (current_infop->nvmet_ctx_start_cpu)
+		get_infop = current_infop->nvmet_ctx_start_cpu;
+	else
+		get_infop = current_infop->nvmet_ctx_next_cpu;
+
+	for (i = 0; i < phba->sli4_hba.num_present_cpu; i++) {
+		if (get_infop == current_infop) {
+			get_infop = get_infop->nvmet_ctx_next_cpu;
+			continue;
+		}
+		spin_lock(&get_infop->nvmet_ctx_list_lock);
+
+		/* Just take the entire context list, if there are any */
+		if (get_infop->nvmet_ctx_list_cnt) {
+			list_splice_init(&get_infop->nvmet_ctx_list,
+				    &current_infop->nvmet_ctx_list);
+			current_infop->nvmet_ctx_list_cnt =
+				get_infop->nvmet_ctx_list_cnt - 1;
+			get_infop->nvmet_ctx_list_cnt = 0;
+			spin_unlock(&get_infop->nvmet_ctx_list_lock);
+
+			current_infop->nvmet_ctx_start_cpu = get_infop;
+			list_remove_head(&current_infop->nvmet_ctx_list,
+					 ctx_buf, struct lpfc_nvmet_ctxbuf,
+					 list);
+			return ctx_buf;
+		}
+
+		/* Otherwise, move on to the next CPU for this MRQ */
+		spin_unlock(&get_infop->nvmet_ctx_list_lock);
+		get_infop = get_infop->nvmet_ctx_next_cpu;
+	}
+
+	/* Nothing found, all contexts for the MRQ are in-flight */
+	return NULL;
+}
+
 /**
  * lpfc_nvmet_unsol_fcp_buffer - Process an unsolicited event data buffer
  * @phba: pointer to lpfc hba data structure.
- * @pring: pointer to a SLI ring.
+ * @idx: relative index of MRQ vector
  * @nvmebuf: pointer to lpfc nvme command HBQ data structure.
  *
  * This routine is used for processing the WQE associated with a unsolicited
@@ -1361,7 +1505,7 @@ dropit:
  **/
 static void
 lpfc_nvmet_unsol_fcp_buffer(struct lpfc_hba *phba,
-			    struct lpfc_sli_ring *pring,
+			    uint32_t idx,
 			    struct rqb_dmabuf *nvmebuf,
 			    uint64_t isr_timestamp)
 {
@@ -1370,9 +1514,11 @@ lpfc_nvmet_unsol_fcp_buffer(struct lpfc_hba *phba,
 	struct lpfc_nvmet_tgtport *tgtp;
 	struct fc_frame_header *fc_hdr;
 	struct lpfc_nvmet_ctxbuf *ctx_buf;
+	struct lpfc_nvmet_ctx_info *current_infop;
 	uint32_t *payload;
 	uint32_t size, oxid, sid, rc, qno;
 	unsigned long iflag;
+	int current_cpu;
 #ifdef CONFIG_SCSI_LPFC_DEBUG_FS
 	uint32_t id;
 #endif
@@ -1388,31 +1534,24 @@ lpfc_nvmet_unsol_fcp_buffer(struct lpfc_hba *phba,
 		goto dropit;
 	}
 
-	spin_lock_irqsave(&phba->sli4_hba.nvmet_ctx_get_lock, iflag);
-	if (phba->sli4_hba.nvmet_ctx_get_cnt) {
-		list_remove_head(&phba->sli4_hba.lpfc_nvmet_ctx_get_list,
+	/*
+	 * Get a pointer to the context list for this MRQ based on
+	 * the CPU this MRQ IRQ is associated with. If the CPU association
+	 * changes from our initial assumption, the context list could
+	 * be empty, thus it would need to be replenished with the
+	 * context list from another CPU for this MRQ.
+	 */
+	current_cpu = smp_processor_id();
+	current_infop = lpfc_get_ctx_list(phba, current_cpu, idx);
+	spin_lock_irqsave(&current_infop->nvmet_ctx_list_lock, iflag);
+	if (current_infop->nvmet_ctx_list_cnt) {
+		list_remove_head(&current_infop->nvmet_ctx_list,
 				 ctx_buf, struct lpfc_nvmet_ctxbuf, list);
-		phba->sli4_hba.nvmet_ctx_get_cnt--;
+		current_infop->nvmet_ctx_list_cnt--;
 	} else {
-		spin_lock(&phba->sli4_hba.nvmet_ctx_put_lock);
-		if (phba->sli4_hba.nvmet_ctx_put_cnt) {
-			list_splice(&phba->sli4_hba.lpfc_nvmet_ctx_put_list,
-				    &phba->sli4_hba.lpfc_nvmet_ctx_get_list);
-			INIT_LIST_HEAD(&phba->sli4_hba.lpfc_nvmet_ctx_put_list);
-			phba->sli4_hba.nvmet_ctx_get_cnt =
-				phba->sli4_hba.nvmet_ctx_put_cnt;
-			phba->sli4_hba.nvmet_ctx_put_cnt = 0;
-			spin_unlock(&phba->sli4_hba.nvmet_ctx_put_lock);
-
-			list_remove_head(
-				&phba->sli4_hba.lpfc_nvmet_ctx_get_list,
-				ctx_buf, struct lpfc_nvmet_ctxbuf, list);
-			phba->sli4_hba.nvmet_ctx_get_cnt--;
-		} else {
-			spin_unlock(&phba->sli4_hba.nvmet_ctx_put_lock);
-		}
+		ctx_buf = lpfc_nvmet_replenish_context(phba, current_infop);
 	}
-	spin_unlock_irqrestore(&phba->sli4_hba.nvmet_ctx_get_lock, iflag);
+	spin_unlock_irqrestore(&current_infop->nvmet_ctx_list_lock, iflag);
 
 	fc_hdr = (struct fc_frame_header *)(nvmebuf->hbuf.virt);
 	oxid = be16_to_cpu(fc_hdr->fh_ox_id);
@@ -1464,6 +1603,7 @@ lpfc_nvmet_unsol_fcp_buffer(struct lpfc_hba *phba,
 	ctxp->size = size;
 	ctxp->oxid = oxid;
 	ctxp->sid = sid;
+	ctxp->idx = idx;
 	ctxp->state = LPFC_NVMET_STE_RCV;
 	ctxp->entry_cnt = 1;
 	ctxp->flag = 0;
@@ -1561,7 +1701,7 @@ lpfc_nvmet_unsol_ls_event(struct lpfc_hba *phba, struct lpfc_sli_ring *pring,
 /**
  * lpfc_nvmet_unsol_fcp_event - Process an unsolicited event from an nvme nport
  * @phba: pointer to lpfc hba data structure.
- * @pring: pointer to a SLI ring.
+ * @idx: relative index of MRQ vector
  * @nvmebuf: pointer to received nvme data structure.
  *
  * This routine is used to process an unsolicited event received from a SLI
@@ -1572,7 +1712,7 @@ lpfc_nvmet_unsol_ls_event(struct lpfc_hba *phba, struct lpfc_sli_ring *pring,
  **/
 void
 lpfc_nvmet_unsol_fcp_event(struct lpfc_hba *phba,
-			   struct lpfc_sli_ring *pring,
+			   uint32_t idx,
 			   struct rqb_dmabuf *nvmebuf,
 			   uint64_t isr_timestamp)
 {
@@ -1580,7 +1720,7 @@ lpfc_nvmet_unsol_fcp_event(struct lpfc_hba *phba,
 		lpfc_rq_buf_free(phba, &nvmebuf->hbuf);
 		return;
 	}
-	lpfc_nvmet_unsol_fcp_buffer(phba, pring, nvmebuf,
+	lpfc_nvmet_unsol_fcp_buffer(phba, idx, nvmebuf,
 				    isr_timestamp);
 }