scsi: lpfc: Limit tracking of tgt queue depth in fast path

Performance is affected when target queue depth is tracked.  An atomic
counter is incremented on the submission path which competes with it being
decremented on the completion path.  In addition, multiple CPUs can
simultaniously be manipulating this counter for the same ndlp.

Reduce the overhead by only performing the target increment/decrement when
the target queue depth is less than the overall adapter depth, thus is
actually meaningful.

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, 24 insertions, 17 deletions

diff --git a/drivers/scsi/lpfc/lpfc_nvme.c b/drivers/scsi/lpfc/lpfc_nvme.c
index 32053424d546..028462e5994d 100644
--- a/drivers/scsi/lpfc/lpfc_nvme.c
+++ b/drivers/scsi/lpfc/lpfc_nvme.c
@@ -1135,9 +1135,6 @@ out_err:
 	else
 		lpfc_ncmd->flags &= ~LPFC_SBUF_XBUSY;
 
-	if (ndlp && NLP_CHK_NODE_ACT(ndlp))
-		atomic_dec(&ndlp->cmd_pending);
-
 	/* Update stats and complete the IO.  There is
 	 * no need for dma unprep because the nvme_transport
 	 * owns the dma address.
@@ -1546,17 +1543,19 @@ lpfc_nvme_fcp_io_submit(struct nvme_fc_local_port *pnvme_lport,
 	/* The node is shared with FCP IO, make sure the IO pending count does
 	 * not exceed the programmed depth.
 	 */
-	if ((atomic_read(&ndlp->cmd_pending) >= ndlp->cmd_qdepth) &&
-	    !expedite) {
-		lpfc_printf_vlog(vport, KERN_INFO, LOG_NVME_IOERR,
-				 "6174 Fail IO, ndlp qdepth exceeded: "
-				 "idx %d DID %x pend %d qdepth %d\n",
-				 lpfc_queue_info->index, ndlp->nlp_DID,
-				 atomic_read(&ndlp->cmd_pending),
-				 ndlp->cmd_qdepth);
-		atomic_inc(&lport->xmt_fcp_qdepth);
-		ret = -EBUSY;
-		goto out_fail;
+	if (lpfc_ndlp_check_qdepth(phba, ndlp)) {
+		if ((atomic_read(&ndlp->cmd_pending) >= ndlp->cmd_qdepth) &&
+		    !expedite) {
+			lpfc_printf_vlog(vport, KERN_INFO, LOG_NVME_IOERR,
+					 "6174 Fail IO, ndlp qdepth exceeded: "
+					 "idx %d DID %x pend %d qdepth %d\n",
+					 lpfc_queue_info->index, ndlp->nlp_DID,
+					 atomic_read(&ndlp->cmd_pending),
+					 ndlp->cmd_qdepth);
+			atomic_inc(&lport->xmt_fcp_qdepth);
+			ret = -EBUSY;
+			goto out_fail;
+		}
 	}
 
 	lpfc_ncmd = lpfc_get_nvme_buf(phba, ndlp, expedite);
@@ -1614,8 +1613,6 @@ lpfc_nvme_fcp_io_submit(struct nvme_fc_local_port *pnvme_lport,
 		goto out_free_nvme_buf;
 	}
 
-	atomic_inc(&ndlp->cmd_pending);
-
 	lpfc_nvmeio_data(phba, "NVME FCP XMIT: xri x%x idx %d to %06x\n",
 			 lpfc_ncmd->cur_iocbq.sli4_xritag,
 			 lpfc_queue_info->index, ndlp->nlp_DID);
@@ -1623,7 +1620,6 @@ lpfc_nvme_fcp_io_submit(struct nvme_fc_local_port *pnvme_lport,
 	ret = lpfc_sli4_issue_wqe(phba, LPFC_FCP_RING, &lpfc_ncmd->cur_iocbq);
 	if (ret) {
 		atomic_inc(&lport->xmt_fcp_wqerr);
-		atomic_dec(&ndlp->cmd_pending);
 		lpfc_printf_vlog(vport, KERN_INFO, LOG_NVME_IOERR,
 				 "6113 Fail IO, Could not issue WQE err %x "
 				 "sid: x%x did: x%x oxid: x%x\n",
@@ -2378,6 +2374,11 @@ lpfc_get_nvme_buf(struct lpfc_hba *phba, struct lpfc_nodelist *ndlp,
 			lpfc_ncmd = lpfc_nvme_buf(phba);
 	}
 	spin_unlock_irqrestore(&phba->nvme_buf_list_get_lock, iflag);
+
+	if (lpfc_ndlp_check_qdepth(phba, ndlp) && lpfc_ncmd) {
+		atomic_inc(&ndlp->cmd_pending);
+		lpfc_ncmd->flags |= LPFC_BUMP_QDEPTH;
+	}
 	return  lpfc_ncmd;
 }
 
@@ -2396,7 +2397,13 @@ lpfc_release_nvme_buf(struct lpfc_hba *phba, struct lpfc_nvme_buf *lpfc_ncmd)
 {
 	unsigned long iflag = 0;
 
+	if ((lpfc_ncmd->flags & LPFC_BUMP_QDEPTH) && lpfc_ncmd->ndlp)
+		atomic_dec(&lpfc_ncmd->ndlp->cmd_pending);
+
 	lpfc_ncmd->nonsg_phys = 0;
+	lpfc_ncmd->ndlp = NULL;
+	lpfc_ncmd->flags &= ~LPFC_BUMP_QDEPTH;
+
 	if (lpfc_ncmd->flags & LPFC_SBUF_XBUSY) {
 		lpfc_printf_log(phba, KERN_INFO, LOG_NVME_ABTS,
 				"6310 XB release deferred for "