RDMA/siw: Fix SQ/RQ drain logic

Storage ULPs (e.g. iSER & NVMeOF) use ib_drain_qp() to drain
QP/CQ. Current SIW's own drain routines do not properly wait until all
SQ/RQ elements are completed and reaped from the CQ. This may cause touch
after free issues.  New logic relies on generic
__ib_drain_sq()/__ib_drain_rq() posting a final work request, which SIW
immediately flushes to CQ.

Fixes: 303ae1cdfdf7 ("rdma/siw: application interface")
Link: https://lore.kernel.org/r/20191004125356.20673-1-bmt@zurich.ibm.com
Signed-off-by: Krishnamraju Eraparaju <krishna2@chelsio.com>
Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

1 files changed, 122 insertions, 22 deletions

diff --git a/drivers/infiniband/sw/siw/siw_verbs.c b/drivers/infiniband/sw/siw/siw_verbs.c
index 869e02b69a01..c0574ddc98fa 100644
--- a/drivers/infiniband/sw/siw/siw_verbs.c
+++ b/drivers/infiniband/sw/siw/siw_verbs.c
@@ -687,6 +687,47 @@ static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
 	return bytes;
 }
 
+/* Complete SQ WR's without processing */
+static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
+			   const struct ib_send_wr **bad_wr)
+{
+	struct siw_sqe sqe = {};
+	int rv = 0;
+
+	while (wr) {
+		sqe.id = wr->wr_id;
+		sqe.opcode = wr->opcode;
+		rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
+		if (rv) {
+			if (bad_wr)
+				*bad_wr = wr;
+			break;
+		}
+		wr = wr->next;
+	}
+	return rv;
+}
+
+/* Complete RQ WR's without processing */
+static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
+			   const struct ib_recv_wr **bad_wr)
+{
+	struct siw_rqe rqe = {};
+	int rv = 0;
+
+	while (wr) {
+		rqe.id = wr->wr_id;
+		rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
+		if (rv) {
+			if (bad_wr)
+				*bad_wr = wr;
+			break;
+		}
+		wr = wr->next;
+	}
+	return rv;
+}
+
 /*
  * siw_post_send()
  *
@@ -705,26 +746,54 @@ int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
 	unsigned long flags;
 	int rv = 0;
 
+	if (wr && !qp->kernel_verbs) {
+		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
+		*bad_wr = wr;
+		return -EINVAL;
+	}
+
 	/*
 	 * Try to acquire QP state lock. Must be non-blocking
 	 * to accommodate kernel clients needs.
 	 */
 	if (!down_read_trylock(&qp->state_lock)) {
-		*bad_wr = wr;
-		siw_dbg_qp(qp, "QP locked, state %d\n", qp->attrs.state);
-		return -ENOTCONN;
+		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
+			/*
+			 * ERROR state is final, so we can be sure
+			 * this state will not change as long as the QP
+			 * exists.
+			 *
+			 * This handles an ib_drain_sq() call with
+			 * a concurrent request to set the QP state
+			 * to ERROR.
+			 */
+			rv = siw_sq_flush_wr(qp, wr, bad_wr);
+		} else {
+			siw_dbg_qp(qp, "QP locked, state %d\n",
+				   qp->attrs.state);
+			*bad_wr = wr;
+			rv = -ENOTCONN;
+		}
+		return rv;
 	}
 	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
+		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
+			/*
+			 * Immediately flush this WR to CQ, if QP
+			 * is in ERROR state. SQ is guaranteed to
+			 * be empty, so WR complets in-order.
+			 *
+			 * Typically triggered by ib_drain_sq().
+			 */
+			rv = siw_sq_flush_wr(qp, wr, bad_wr);
+		} else {
+			siw_dbg_qp(qp, "QP out of state %d\n",
+				   qp->attrs.state);
+			*bad_wr = wr;
+			rv = -ENOTCONN;
+		}
 		up_read(&qp->state_lock);
-		*bad_wr = wr;
-		siw_dbg_qp(qp, "QP out of state %d\n", qp->attrs.state);
-		return -ENOTCONN;
-	}
-	if (wr && !qp->kernel_verbs) {
-		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
-		up_read(&qp->state_lock);
-		*bad_wr = wr;
-		return -EINVAL;
+		return rv;
 	}
 	spin_lock_irqsave(&qp->sq_lock, flags);
 
@@ -919,24 +988,55 @@ int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
 		*bad_wr = wr;
 		return -EOPNOTSUPP; /* what else from errno.h? */
 	}
-	/*
-	 * Try to acquire QP state lock. Must be non-blocking
-	 * to accommodate kernel clients needs.
-	 */
-	if (!down_read_trylock(&qp->state_lock)) {
-		*bad_wr = wr;
-		return -ENOTCONN;
-	}
 	if (!qp->kernel_verbs) {
 		siw_dbg_qp(qp, "no kernel post_recv for user mapped sq\n");
 		up_read(&qp->state_lock);
 		*bad_wr = wr;
 		return -EINVAL;
 	}
+
+	/*
+	 * Try to acquire QP state lock. Must be non-blocking
+	 * to accommodate kernel clients needs.
+	 */
+	if (!down_read_trylock(&qp->state_lock)) {
+		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
+			/*
+			 * ERROR state is final, so we can be sure
+			 * this state will not change as long as the QP
+			 * exists.
+			 *
+			 * This handles an ib_drain_rq() call with
+			 * a concurrent request to set the QP state
+			 * to ERROR.
+			 */
+			rv = siw_rq_flush_wr(qp, wr, bad_wr);
+		} else {
+			siw_dbg_qp(qp, "QP locked, state %d\n",
+				   qp->attrs.state);
+			*bad_wr = wr;
+			rv = -ENOTCONN;
+		}
+		return rv;
+	}
 	if (qp->attrs.state > SIW_QP_STATE_RTS) {
+		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
+			/*
+			 * Immediately flush this WR to CQ, if QP
+			 * is in ERROR state. RQ is guaranteed to
+			 * be empty, so WR complets in-order.
+			 *
+			 * Typically triggered by ib_drain_rq().
+			 */
+			rv = siw_rq_flush_wr(qp, wr, bad_wr);
+		} else {
+			siw_dbg_qp(qp, "QP out of state %d\n",
+				   qp->attrs.state);
+			*bad_wr = wr;
+			rv = -ENOTCONN;
+		}
 		up_read(&qp->state_lock);
-		*bad_wr = wr;
-		return -EINVAL;
+		return rv;
 	}
 	/*
 	 * Serialize potentially multiple producers.