diff mbox

IB/qib: Improve ipoib UD performance

Message ID 20151224161923.9544.79119.stgit@phlsvslse11.ph.intel.com (mailing list archive)
State Accepted
Headers show

Commit Message

Marciniszyn, Mike Dec. 24, 2015, 4:19 p.m. UTC 
Based on profiling, UD performance drops in case of processes
in a single client due to excess context switches when
the progress workqueue is scheduled.

This is solved by modifying the heuristic to select the
direct progress instead of the scheduling progress via
the workqueue when UD-like situations are detected in
the heuristic.

Reviewed-by: Vinit Agnihotri <vinit.abhay.agnihotri@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
---
 drivers/infiniband/hw/qib/qib_verbs.c |   11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)


--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Doug Ledford Jan. 19, 2016, 8:41 p.m. UTC | #1 
On 12/24/2015 11:19 AM, Mike Marciniszyn wrote:
> Based on profiling, UD performance drops in case of processes
> in a single client due to excess context switches when
> the progress workqueue is scheduled.
> 
> This is solved by modifying the heuristic to select the
> direct progress instead of the scheduling progress via
> the workqueue when UD-like situations are detected in
> the heuristic.
> 
> Reviewed-by: Vinit Agnihotri <vinit.abhay.agnihotri@intel.com>
> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>

Thanks, applied.
diff mbox

Patch

diff --git a/drivers/infiniband/hw/qib/qib_verbs.c b/drivers/infiniband/hw/qib/qib_verbs.c
index de6cb6f..f29f097 100644
--- a/drivers/infiniband/hw/qib/qib_verbs.c
+++ b/drivers/infiniband/hw/qib/qib_verbs.c
@@ -346,6 +346,7 @@  static int qib_post_one_send(struct qib_qp *qp, struct ib_send_wr *wr,
 	unsigned long flags;
 	struct qib_lkey_table *rkt;
 	struct qib_pd *pd;
+	int avoid_schedule = 0;
 
 	spin_lock_irqsave(&qp->s_lock, flags);
 
@@ -438,11 +439,15 @@  static int qib_post_one_send(struct qib_qp *qp, struct ib_send_wr *wr,
 	    qp->ibqp.qp_type == IB_QPT_RC) {
 		if (wqe->length > 0x80000000U)
 			goto bail_inval_free;
+		if (wqe->length <= qp->pmtu)
+			avoid_schedule = 1;
 	} else if (wqe->length > (dd_from_ibdev(qp->ibqp.device)->pport +
-				  qp->port_num - 1)->ibmtu)
+				  qp->port_num - 1)->ibmtu) {
 		goto bail_inval_free;
-	else
+	} else {
 		atomic_inc(&to_iah(ud_wr(wr)->ah)->refcount);
+		avoid_schedule = 1;
+	}
 	wqe->ssn = qp->s_ssn++;
 	qp->s_head = next;
 
@@ -458,7 +463,7 @@  bail_inval_free:
 bail_inval:
 	ret = -EINVAL;
 bail:
-	if (!ret && !wr->next &&
+	if (!ret && !wr->next && !avoid_schedule &&
 	 !qib_sdma_empty(
 	   dd_from_ibdev(qp->ibqp.device)->pport + qp->port_num - 1)) {
 		qib_schedule_send(qp);