RDMA/cxgb4: Fix QP flush logic

This patch makes following fixes in QP flush logic:

- correctly flushes unsignaled WRs followed by a signaled WR
- supports for flushing a CQ bound to multiple QPs
- resets cidx_flush if a active queue starts getting HW CQEs again
- marks WQ in error when we leave RTS. This was only being done for
  user queues, but we need it for kernel queues too so that
  post_send/post_recv will start returning the appropriate error
  synchronously
- eats unsignaled read resp CQEs. HW always inserts CQEs so we must
  silently discard them if the read work request was unsignaled.
- handles QP flushes with pending SW CQEs. The flush and out of order
  completion logic has a bug where if out of order completions are
  flushed but not yet polled by the consumer and the qp is then
  flushed then we end up inserting duplicate completions.
- c4iw_flush_sq() should only flush wrs that have not already been
  flushed.  Since we already track where in the SQ we've flushed via
  sq.cidx_flush, just start at that point and flush any remaining.
  This bug only caused a problem in the presence of unsignaled work
  requests.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Vipul Pandya <vipul@chelsio.com>

[ Fixed sparse warning due to htonl/ntohl confusion.  - Roland ]

Signed-off-by: Roland Dreier <roland@purestorage.com>

4 files changed, 254 insertions, 136 deletions

diff --git a/drivers/infiniband/hw/cxgb4/cq.c b/drivers/infiniband/hw/cxgb4/cq.c
index 0f1607c..6657390 100644
--- a/drivers/infiniband/hw/cxgb4/cq.c
+++ b/drivers/infiniband/hw/cxgb4/cq.c
@@ -225,43 +225,186 @@ static void insert_sq_cqe(struct t4_wq *wq, struct t4_cq *cq,
 	t4_swcq_produce(cq);
 }
 
-int c4iw_flush_sq(struct t4_wq *wq, struct t4_cq *cq, int count)
+static void advance_oldest_read(struct t4_wq *wq);
+
+int c4iw_flush_sq(struct c4iw_qp *qhp)
 {
 	int flushed = 0;
-	struct t4_swsqe *swsqe = &wq->sq.sw_sq[wq->sq.cidx + count];
-	int in_use = wq->sq.in_use - count;
-
-	BUG_ON(in_use < 0);
-	while (in_use--) {
-		swsqe->signaled = 0;
-		insert_sq_cqe(wq, cq, swsqe);
-		swsqe++;
-		if (swsqe == (wq->sq.sw_sq + wq->sq.size))
-			swsqe = wq->sq.sw_sq;
-		flushed++;
+	struct t4_wq *wq = &qhp->wq;
+	struct c4iw_cq *chp = to_c4iw_cq(qhp->ibqp.send_cq);
+	struct t4_cq *cq = &chp->cq;
+	int idx;
+	struct t4_swsqe *swsqe;
+	int error = (qhp->attr.state != C4IW_QP_STATE_CLOSING &&
+			qhp->attr.state != C4IW_QP_STATE_IDLE);
+
+	if (wq->sq.flush_cidx == -1)
+		wq->sq.flush_cidx = wq->sq.cidx;
+	idx = wq->sq.flush_cidx;
+	BUG_ON(idx >= wq->sq.size);
+	while (idx != wq->sq.pidx) {
+		if (error) {
+			swsqe = &wq->sq.sw_sq[idx];
+			BUG_ON(swsqe->flushed);
+			swsqe->flushed = 1;
+			insert_sq_cqe(wq, cq, swsqe);
+			if (wq->sq.oldest_read == swsqe) {
+				BUG_ON(swsqe->opcode != FW_RI_READ_REQ);
+				advance_oldest_read(wq);
+			}
+			flushed++;
+		} else {
+			t4_sq_consume(wq);
+		}
+		if (++idx == wq->sq.size)
+			idx = 0;
 	}
+	wq->sq.flush_cidx += flushed;
+	if (wq->sq.flush_cidx >= wq->sq.size)
+		wq->sq.flush_cidx -= wq->sq.size;
 	return flushed;
 }
 
+static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
+{
+	struct t4_swsqe *swsqe;
+	int cidx;
+
+	if (wq->sq.flush_cidx == -1)
+		wq->sq.flush_cidx = wq->sq.cidx;
+	cidx = wq->sq.flush_cidx;
+	BUG_ON(cidx > wq->sq.size);
+
+	while (cidx != wq->sq.pidx) {
+		swsqe = &wq->sq.sw_sq[cidx];
+		if (!swsqe->signaled) {
+			if (++cidx == wq->sq.size)
+				cidx = 0;
+		} else if (swsqe->complete) {
+
+			BUG_ON(swsqe->flushed);
+
+			/*
+			 * Insert this completed cqe into the swcq.
+			 */
+			PDBG("%s moving cqe into swcq sq idx %u cq idx %u\n",
+					__func__, cidx, cq->sw_pidx);
+			swsqe->cqe.header |= htonl(V_CQE_SWCQE(1));
+			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
+			t4_swcq_produce(cq);
+			swsqe->flushed = 1;
+			if (++cidx == wq->sq.size)
+				cidx = 0;
+			wq->sq.flush_cidx = cidx;
+		} else
+			break;
+	}
+}
+
+static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
+		struct t4_cqe *read_cqe)
+{
+	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
+	read_cqe->len = htonl(wq->sq.oldest_read->read_len);
+	read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(hw_cqe)) |
+			V_CQE_SWCQE(SW_CQE(hw_cqe)) |
+			V_CQE_OPCODE(FW_RI_READ_REQ) |
+			V_CQE_TYPE(1));
+	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
+}
+
+static void advance_oldest_read(struct t4_wq *wq)
+{
+
+	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
+
+	if (rptr == wq->sq.size)
+		rptr = 0;
+	while (rptr != wq->sq.pidx) {
+		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
+
+		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
+			return;
+		if (++rptr == wq->sq.size)
+			rptr = 0;
+	}
+	wq->sq.oldest_read = NULL;
+}
+
 /*
  * Move all CQEs from the HWCQ into the SWCQ.
+ * Deal with out-of-order and/or completions that complete
+ * prior unsignalled WRs.
  */
-void c4iw_flush_hw_cq(struct t4_cq *cq)
+void c4iw_flush_hw_cq(struct c4iw_cq *chp)
 {
-	struct t4_cqe *cqe = NULL, *swcqe;
+	struct t4_cqe *hw_cqe, *swcqe, read_cqe;
+	struct c4iw_qp *qhp;
+	struct t4_swsqe *swsqe;
 	int ret;
 
-	PDBG("%s cq %p cqid 0x%x\n", __func__, cq, cq->cqid);
-	ret = t4_next_hw_cqe(cq, &cqe);
+	PDBG("%s  cqid 0x%x\n", __func__, chp->cq.cqid);
+	ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
+
+	/*
+	 * This logic is similar to poll_cq(), but not quite the same
+	 * unfortunately.  Need to move pertinent HW CQEs to the SW CQ but
+	 * also do any translation magic that poll_cq() normally does.
+	 */
 	while (!ret) {
-		PDBG("%s flushing hwcq cidx 0x%x swcq pidx 0x%x\n",
-		     __func__, cq->cidx, cq->sw_pidx);
-		swcqe = &cq->sw_queue[cq->sw_pidx];
-		*swcqe = *cqe;
-		swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
-		t4_swcq_produce(cq);
-		t4_hwcq_consume(cq);
-		ret = t4_next_hw_cqe(cq, &cqe);
+		qhp = get_qhp(chp->rhp, CQE_QPID(hw_cqe));
+
+		/*
+		 * drop CQEs with no associated QP
+		 */
+		if (qhp == NULL)
+			goto next_cqe;
+
+		if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE)
+			goto next_cqe;
+
+		if (CQE_OPCODE(hw_cqe) == FW_RI_READ_RESP) {
+
+			/*
+			 * drop peer2peer RTR reads.
+			 */
+			if (CQE_WRID_STAG(hw_cqe) == 1)
+				goto next_cqe;
+
+			/*
+			 * Eat completions for unsignaled read WRs.
+			 */
+			if (!qhp->wq.sq.oldest_read->signaled) {
+				advance_oldest_read(&qhp->wq);
+				goto next_cqe;
+			}
+
+			/*
+			 * Don't write to the HWCQ, create a new read req CQE
+			 * in local memory and move it into the swcq.
+			 */
+			create_read_req_cqe(&qhp->wq, hw_cqe, &read_cqe);
+			hw_cqe = &read_cqe;
+			advance_oldest_read(&qhp->wq);
+		}
+
+		/* if its a SQ completion, then do the magic to move all the
+		 * unsignaled and now in-order completions into the swcq.
+		 */
+		if (SQ_TYPE(hw_cqe)) {
+			swsqe = &qhp->wq.sq.sw_sq[CQE_WRID_SQ_IDX(hw_cqe)];
+			swsqe->cqe = *hw_cqe;
+			swsqe->complete = 1;
+			flush_completed_wrs(&qhp->wq, &chp->cq);
+		} else {
+			swcqe = &chp->cq.sw_queue[chp->cq.sw_pidx];
+			*swcqe = *hw_cqe;
+			swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
+			t4_swcq_produce(&chp->cq);
+		}
+next_cqe:
+		t4_hwcq_consume(&chp->cq);
+		ret = t4_next_hw_cqe(&chp->cq, &hw_cqe);
 	}
 }
 
@@ -281,25 +424,6 @@ static int cqe_completes_wr(struct t4_cqe *cqe, struct t4_wq *wq)
 	return 1;
 }
 
-void c4iw_count_scqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
-{
-	struct t4_cqe *cqe;
-	u32 ptr;
-
-	*count = 0;
-	ptr = cq->sw_cidx;
-	while (ptr != cq->sw_pidx) {
-		cqe = &cq->sw_queue[ptr];
-		if ((SQ_TYPE(cqe) || ((CQE_OPCODE(cqe) == FW_RI_READ_RESP) &&
-				      wq->sq.oldest_read)) &&
-		    (CQE_QPID(cqe) == wq->sq.qid))
-			(*count)++;
-		if (++ptr == cq->size)
-			ptr = 0;
-	}
-	PDBG("%s cq %p count %d\n", __func__, cq, *count);
-}
-
 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
 {
 	struct t4_cqe *cqe;
@@ -319,70 +443,6 @@ void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count)
 	PDBG("%s cq %p count %d\n", __func__, cq, *count);
 }
 
-static void flush_completed_wrs(struct t4_wq *wq, struct t4_cq *cq)
-{
-	struct t4_swsqe *swsqe;
-	u16 ptr = wq->sq.cidx;
-	int count = wq->sq.in_use;
-	int unsignaled = 0;
-
-	swsqe = &wq->sq.sw_sq[ptr];
-	while (count--)
-		if (!swsqe->signaled) {
-			if (++ptr == wq->sq.size)
-				ptr = 0;
-			swsqe = &wq->sq.sw_sq[ptr];
-			unsignaled++;
-		} else if (swsqe->complete) {
-
-			/*
-			 * Insert this completed cqe into the swcq.
-			 */
-			PDBG("%s moving cqe into swcq sq idx %u cq idx %u\n",
-			     __func__, ptr, cq->sw_pidx);
-			swsqe->cqe.header |= htonl(V_CQE_SWCQE(1));
-			cq->sw_queue[cq->sw_pidx] = swsqe->cqe;
-			t4_swcq_produce(cq);
-			swsqe->signaled = 0;
-			wq->sq.in_use -= unsignaled;
-			break;
-		} else
-			break;
-}
-
-static void create_read_req_cqe(struct t4_wq *wq, struct t4_cqe *hw_cqe,
-				struct t4_cqe *read_cqe)
-{
-	read_cqe->u.scqe.cidx = wq->sq.oldest_read->idx;
-	read_cqe->len = cpu_to_be32(wq->sq.oldest_read->read_len);
-	read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(hw_cqe)) |
-				 V_CQE_SWCQE(SW_CQE(hw_cqe)) |
-				 V_CQE_OPCODE(FW_RI_READ_REQ) |
-				 V_CQE_TYPE(1));
-	read_cqe->bits_type_ts = hw_cqe->bits_type_ts;
-}
-
-/*
- * Return a ptr to the next read wr in the SWSQ or NULL.
- */
-static void advance_oldest_read(struct t4_wq *wq)
-{
-
-	u32 rptr = wq->sq.oldest_read - wq->sq.sw_sq + 1;
-
-	if (rptr == wq->sq.size)
-		rptr = 0;
-	while (rptr != wq->sq.pidx) {
-		wq->sq.oldest_read = &wq->sq.sw_sq[rptr];
-
-		if (wq->sq.oldest_read->opcode == FW_RI_READ_REQ)
-			return;
-		if (++rptr == wq->sq.size)
-			rptr = 0;
-	}
-	wq->sq.oldest_read = NULL;
-}
-
 /*
  * poll_cq
  *
@@ -427,6 +487,22 @@ static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
 	}
 
 	/*
+	* skip hw cqe's if the wq is flushed.
+	*/
+	if (wq->flushed && !SW_CQE(hw_cqe)) {
+		ret = -EAGAIN;
+		goto skip_cqe;
+	}
+
+	/*
+	 * skip TERMINATE cqes...
+	 */
+	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
+		ret = -EAGAIN;
+		goto skip_cqe;
+	}
+
+	/*
 	 * Gotta tweak READ completions:
 	 *	1) the cqe doesn't contain the sq_wptr from the wr.
 	 *	2) opcode not reflected from the wr.
@@ -440,7 +516,7 @@ static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
 		 * was generated by the kernel driver as part of peer-2-peer
 		 * connection setup.  So ignore the completion.
 		 */
-		if (!wq->sq.oldest_read) {
+		if (CQE_WRID_STAG(hw_cqe) == 1) {
 			if (CQE_STATUS(hw_cqe))
 				t4_set_wq_in_error(wq);
 			ret = -EAGAIN;
@@ -448,6 +524,15 @@ static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
 		}
 
 		/*
+		 * Eat completions for unsignaled read WRs.
+		 */
+		if (!wq->sq.oldest_read->signaled) {
+			advance_oldest_read(wq);
+			ret = -EAGAIN;
+			goto skip_cqe;
+		}
+
+		/*
 		 * Don't write to the HWCQ, so create a new read req CQE
 		 * in local memory.
 		 */
@@ -457,14 +542,8 @@ static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
 	}
 
 	if (CQE_STATUS(hw_cqe) || t4_wq_in_error(wq)) {
-		*cqe_flushed = t4_wq_in_error(wq);
+		*cqe_flushed = (CQE_STATUS(hw_cqe) == T4_ERR_SWFLUSH);
 		t4_set_wq_in_error(wq);
-		goto proc_cqe;
-	}
-
-	if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
-		ret = -EAGAIN;
-		goto skip_cqe;
 	}
 
 	/*
@@ -523,7 +602,21 @@ proc_cqe:
 	 * completion.
 	 */
 	if (SQ_TYPE(hw_cqe)) {
-		wq->sq.cidx = CQE_WRID_SQ_IDX(hw_cqe);
+		int idx = CQE_WRID_SQ_IDX(hw_cqe);
+		BUG_ON(idx > wq->sq.size);
+
+		/*
+		* Account for any unsignaled completions completed by
+		* this signaled completion.  In this case, cidx points
+		* to the first unsignaled one, and idx points to the
+		* signaled one.  So adjust in_use based on this delta.
+		* if this is not completing any unsigned wrs, then the
+		* delta will be 0.
+		*/
+		wq->sq.in_use -= idx - wq->sq.cidx;
+		BUG_ON(wq->sq.in_use < 0 && wq->sq.in_use < wq->sq.size);
+
+		wq->sq.cidx = (uint16_t)idx;
 		PDBG("%s completing sq idx %u\n", __func__, wq->sq.cidx);
 		*cookie = wq->sq.sw_sq[wq->sq.cidx].wr_id;
 		t4_sq_consume(wq);
@@ -532,6 +625,7 @@ proc_cqe:
 		*cookie = wq->rq.sw_rq[wq->rq.cidx].wr_id;
 		BUG_ON(t4_rq_empty(wq));
 		t4_rq_consume(wq);
+		goto skip_cqe;
 	}
 
 flush_wq:
diff --git a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
index b3cae9f..23eaeab 100644
--- a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
+++ b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
@@ -917,12 +917,11 @@ void c4iw_pblpool_free(struct c4iw_rdev *rdev, u32 addr, int size);
 u32 c4iw_ocqp_pool_alloc(struct c4iw_rdev *rdev, int size);
 void c4iw_ocqp_pool_free(struct c4iw_rdev *rdev, u32 addr, int size);
 int c4iw_ofld_send(struct c4iw_rdev *rdev, struct sk_buff *skb);
-void c4iw_flush_hw_cq(struct t4_cq *cq);
+void c4iw_flush_hw_cq(struct c4iw_cq *chp);
 void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count);
-void c4iw_count_scqes(struct t4_cq *cq, struct t4_wq *wq, int *count);
 int c4iw_ep_disconnect(struct c4iw_ep *ep, int abrupt, gfp_t gfp);
 int c4iw_flush_rq(struct t4_wq *wq, struct t4_cq *cq, int count);
-int c4iw_flush_sq(struct t4_wq *wq, struct t4_cq *cq, int count);
+int c4iw_flush_sq(struct c4iw_qp *qhp);
 int c4iw_ev_handler(struct c4iw_dev *rnicp, u32 qid);
 u16 c4iw_rqes_posted(struct c4iw_qp *qhp);
 int c4iw_post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe);
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index a4975e1..22d1f01 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -737,6 +737,7 @@ int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 		swsqe->idx = qhp->wq.sq.pidx;
 		swsqe->complete = 0;
 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
+		swsqe->flushed = 0;
 		swsqe->wr_id = wr->wr_id;
 
 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
@@ -1006,7 +1007,15 @@ static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
 	/* locking hierarchy: cq lock first, then qp lock. */
 	spin_lock_irqsave(&rchp->lock, flag);
 	spin_lock(&qhp->lock);
-	c4iw_flush_hw_cq(&rchp->cq);
+
+	if (qhp->wq.flushed) {
+		spin_unlock(&qhp->lock);
+		spin_unlock_irqrestore(&rchp->lock, flag);
+		return;
+	}
+	qhp->wq.flushed = 1;
+
+	c4iw_flush_hw_cq(rchp);
 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
 	flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
 	spin_unlock(&qhp->lock);
@@ -1020,9 +1029,9 @@ static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
 	/* locking hierarchy: cq lock first, then qp lock. */
 	spin_lock_irqsave(&schp->lock, flag);
 	spin_lock(&qhp->lock);
-	c4iw_flush_hw_cq(&schp->cq);
-	c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
-	flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
+	if (schp != rchp)
+		c4iw_flush_hw_cq(schp);
+	flushed = c4iw_flush_sq(qhp);
 	spin_unlock(&qhp->lock);
 	spin_unlock_irqrestore(&schp->lock, flag);
 	if (flushed) {
@@ -1037,11 +1046,11 @@ static void flush_qp(struct c4iw_qp *qhp)
 	struct c4iw_cq *rchp, *schp;
 	unsigned long flag;
 
-	rchp = get_chp(qhp->rhp, qhp->attr.rcq);
-	schp = get_chp(qhp->rhp, qhp->attr.scq);
+	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
+	schp = to_c4iw_cq(qhp->ibqp.send_cq);
 
+	t4_set_wq_in_error(&qhp->wq);
 	if (qhp->ibqp.uobject) {
-		t4_set_wq_in_error(&qhp->wq);
 		t4_set_cq_in_error(&rchp->cq);
 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
@@ -1330,8 +1339,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 				disconnect = 1;
 				c4iw_get_ep(&qhp->ep->com);
 			}
-			if (qhp->ibqp.uobject)
-				t4_set_wq_in_error(&qhp->wq);
+			t4_set_wq_in_error(&qhp->wq);
 			ret = rdma_fini(rhp, qhp, ep);
 			if (ret)
 				goto err;
@@ -1340,8 +1348,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
 			qhp->attr.layer_etype = attrs->layer_etype;
 			qhp->attr.ecode = attrs->ecode;
-			if (qhp->ibqp.uobject)
-				t4_set_wq_in_error(&qhp->wq);
+			t4_set_wq_in_error(&qhp->wq);
 			ep = qhp->ep;
 			if (!internal)
 				terminate = 1;
@@ -1350,8 +1357,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 			break;
 		case C4IW_QP_STATE_ERROR:
 			set_state(qhp, C4IW_QP_STATE_ERROR);
-			if (qhp->ibqp.uobject)
-				t4_set_wq_in_error(&qhp->wq);
+			t4_set_wq_in_error(&qhp->wq);
 			if (!internal) {
 				abort = 1;
 				disconnect = 1;
@@ -1552,12 +1558,12 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
 
 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
 
-
 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
 	if (!qhp)
 		return ERR_PTR(-ENOMEM);
 	qhp->wq.sq.size = sqsize;
 	qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
+	qhp->wq.sq.flush_cidx = -1;
 	qhp->wq.rq.size = rqsize;
 	qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
 
diff --git a/drivers/infiniband/hw/cxgb4/t4.h b/drivers/infiniband/hw/cxgb4/t4.h
index ebcb03b..3a6a289 100644
--- a/drivers/infiniband/hw/cxgb4/t4.h
+++ b/drivers/infiniband/hw/cxgb4/t4.h
@@ -36,9 +36,9 @@
 #include "t4_msg.h"
 #include "t4fw_ri_api.h"
 
-#define T4_MAX_NUM_QP (1<<16)
-#define T4_MAX_NUM_CQ (1<<15)
-#define T4_MAX_NUM_PD (1<<15)
+#define T4_MAX_NUM_QP 65536
+#define T4_MAX_NUM_CQ 65536
+#define T4_MAX_NUM_PD 65536
 #define T4_EQ_STATUS_ENTRIES (L1_CACHE_BYTES > 64 ? 2 : 1)
 #define T4_MAX_EQ_SIZE (65520 - T4_EQ_STATUS_ENTRIES)
 #define T4_MAX_IQ_SIZE (65520 - 1)
@@ -269,6 +269,7 @@ struct t4_swsqe {
 	int			complete;
 	int			signaled;
 	u16			idx;
+	int                     flushed;
 };
 
 static inline pgprot_t t4_pgprot_wc(pgprot_t prot)
@@ -300,6 +301,7 @@ struct t4_sq {
 	u16 pidx;
 	u16 wq_pidx;
 	u16 flags;
+	short flush_cidx;
 };
 
 struct t4_swrqe {
@@ -330,6 +332,7 @@ struct t4_wq {
 	void __iomem *db;
 	void __iomem *gts;
 	struct c4iw_rdev *rdev;
+	int flushed;
 };
 
 static inline int t4_rqes_posted(struct t4_wq *wq)
@@ -412,6 +415,9 @@ static inline void t4_sq_produce(struct t4_wq *wq, u8 len16)
 
 static inline void t4_sq_consume(struct t4_wq *wq)
 {
+	BUG_ON(wq->sq.in_use < 1);
+	if (wq->sq.cidx == wq->sq.flush_cidx)
+		wq->sq.flush_cidx = -1;
 	wq->sq.in_use--;
 	if (++wq->sq.cidx == wq->sq.size)
 		wq->sq.cidx = 0;
@@ -505,12 +511,18 @@ static inline int t4_arm_cq(struct t4_cq *cq, int se)
 static inline void t4_swcq_produce(struct t4_cq *cq)
 {
 	cq->sw_in_use++;
+	if (cq->sw_in_use == cq->size) {
+		PDBG("%s cxgb4 sw cq overflow cqid %u\n", __func__, cq->cqid);
+		cq->error = 1;
+		BUG_ON(1);
+	}
 	if (++cq->sw_pidx == cq->size)
 		cq->sw_pidx = 0;
 }
 
 static inline void t4_swcq_consume(struct t4_cq *cq)
 {
+	BUG_ON(cq->sw_in_use < 1);
 	cq->sw_in_use--;
 	if (++cq->sw_cidx == cq->size)
 		cq->sw_cidx = 0;
@@ -552,6 +564,7 @@ static inline int t4_next_hw_cqe(struct t4_cq *cq, struct t4_cqe **cqe)
 		ret = -EOVERFLOW;
 		cq->error = 1;
 		printk(KERN_ERR MOD "cq overflow cqid %u\n", cq->cqid);
+		BUG_ON(1);
 	} else if (t4_valid_cqe(cq, &cq->queue[cq->cidx])) {
 		*cqe = &cq->queue[cq->cidx];
 		ret = 0;
@@ -562,6 +575,12 @@ static inline int t4_next_hw_cqe(struct t4_cq *cq, struct t4_cqe **cqe)
 
 static inline struct t4_cqe *t4_next_sw_cqe(struct t4_cq *cq)
 {
+	if (cq->sw_in_use == cq->size) {
+		PDBG("%s cxgb4 sw cq overflow cqid %u\n", __func__, cq->cqid);
+		cq->error = 1;
+		BUG_ON(1);
+		return NULL;
+	}
 	if (cq->sw_in_use)
 		return &cq->sw_queue[cq->sw_cidx];
 	return NULL;