[openib-general] Re: [PATCH] mthca: fix posting long work request lists

[Roland Dreier]

Does this look OK?  I added SRQ support and also moved the doorbell[2]
declaration to the top of the functions (since I'm not convinved all
versions of gcc are smart enough to see that two copies of doorbell[]
can share stack slots).

 - R.

Index: infiniband/hw/mthca/mthca_srq.c
===================================================================
--- infiniband/hw/mthca/mthca_srq.c	(revision 3989)
+++ infiniband/hw/mthca/mthca_srq.c	(working copy)
@@ -414,6 +414,7 @@ int mthca_tavor_post_srq_recv(struct ib_
 {
 	struct mthca_dev *dev = to_mdev(ibsrq->device);
 	struct mthca_srq *srq = to_msrq(ibsrq);
+	__be32 doorbell[2];
 	unsigned long flags;
 	int err = 0;
 	int first_ind;
@@ -429,6 +430,23 @@ int mthca_tavor_post_srq_recv(struct ib_
 	first_ind = srq->first_free;
 
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
+			doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
+			doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
+
+			/*
+			 * Make sure that descriptors are written
+			 * before doorbell is rung.
+			 */
+			wmb();
+
+			mthca_write64(doorbell,
+				      dev->kar + MTHCA_RECEIVE_DOORBELL,
+				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+
+			first_ind = srq->first_free;
+		}
+
 		ind = srq->first_free;
 
 		if (ind < 0) {
@@ -491,8 +509,6 @@ int mthca_tavor_post_srq_recv(struct ib_
 	}
 
 	if (likely(nreq)) {
-		__be32 doorbell[2];
-
 		doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
 		doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
 
Index: infiniband/hw/mthca/mthca_wqe.h
===================================================================
--- infiniband/hw/mthca/mthca_wqe.h	(revision 3989)
+++ infiniband/hw/mthca/mthca_wqe.h	(working copy)
@@ -49,7 +49,8 @@ enum {
 };
 
 enum {
-	MTHCA_INVAL_LKEY = 0x100
+	MTHCA_INVAL_LKEY			= 0x100,
+	MTHCA_TAVOR_MAX_WQES_PER_RECV_DB	= 256
 };
 
 struct mthca_next_seg {
Index: infiniband/hw/mthca/mthca_qp.c
===================================================================
--- infiniband/hw/mthca/mthca_qp.c	(revision 4003)
+++ infiniband/hw/mthca/mthca_qp.c	(working copy)
@@ -1705,6 +1705,7 @@ int mthca_tavor_post_receive(struct ib_q
 {
 	struct mthca_dev *dev = to_mdev(ibqp->device);
 	struct mthca_qp *qp = to_mqp(ibqp);
+	__be32 doorbell[2];
 	unsigned long flags;
 	int err = 0;
 	int nreq;
@@ -1722,6 +1723,21 @@ int mthca_tavor_post_receive(struct ib_q
 	ind = qp->rq.next_ind;
 
 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
+		if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
+			doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
+			doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
+
+			wmb();
+
+			mthca_write64(doorbell,
+				      dev->kar + MTHCA_RECEIVE_DOORBELL,
+				      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
+
+			qp->rq.head += nreq;
+			nreq  = 0;
+			size0 = 0;
+		}
+
 		if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
 			mthca_err(dev, "RQ %06x full (%u head, %u tail,"
 					" %d max, %d nreq)\n", qp->qpn,
@@ -1779,8 +1795,6 @@ int mthca_tavor_post_receive(struct ib_q
 
 out:
 	if (likely(nreq)) {
-		__be32 doorbell[2];
-
 		doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
 		doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);