Last WQE Reached event treatment

To avoid theoretical leakage for QPs assocoated with SRQ, according to IB spec (section 10.3.1):

"Note, for QPs that are associated with an SRQ, the Consumer should take the QP through the Error State before invoking a Destroy QP or a Modify QP to the Reset State. The Consumer may invoke the Destroy QP without first performing a Modify QP to the Error State and waiting for the

Asynchronous Last WQE Reached Event. However, if the Consumer does not wait for the Affiliated Asynchronous Last WQE Reached Event, then WQE and Data Segment leakage may occur. Therefore, it is good programming practice to tear down a QP that is associated with an SRQ by using the following process:
- Put the QP in the Error State;
- wait for the Affiliated Asynchronous Last WQE Reached Event;
- either:
- drain the CQ by invoking the Poll CQ verb and either wait for CQ to be empty or the number of Poll CQ operations has exceeded CQ capacity size; or
- post another WR that completes on the same CQ and wait for this

     WR to return as a WC;

- and then invoke a Destroy QP or Reset QP."

index 7868727..7604450 100644

drivers/infiniband/core/verbs.c +19 -1

@@ -886,8 +886,10 @@ struct ib_qp *ib_create_qp(struct ib_pd *pd,
 		if (qp_init_attr->recv_cq)
 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
 		qp->srq = qp_init_attr->srq;
-		if (qp->srq)
+		if (qp->srq) {
 			atomic_inc(&qp_init_attr->srq->usecnt);
+			init_completion(&qp->srq_completion);
+		}
 	}
 
 	qp->pd	    = pd;

@@ -1405,6 +1407,22 @@ int ib_get_eth_speed(struct ib_device *dev, u8
 port_num, u8 *speed, u8 *width)
 }
 EXPORT_SYMBOL(ib_get_eth_speed);
 
+int ib_notify_qp(struct ib_qp *qp, enum ib_event_type event)
+{
+	int ret = 0;
+
+	switch (event) {
+	case IB_EVENT_QP_LAST_WQE_REACHED:
+		complete(&qp->srq_completion);
+		break;
+	default:
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL(ib_notify_qp);
+
 int ib_modify_qp(struct ib_qp *qp,
 		 struct ib_qp_attr *qp_attr,

drivers/infiniband/core/verbs.c +32 -0

 		 int qp_attr_mask)

@@ -2213,6 +2231,53 @@ static void __ib_drain_rq(struct ib_qp *qp)
 		wait_for_completion(&rdrain.done);
 }
 
+/*
+ * __ib_drain_srq() - Block until all Last WQE Reached event arrives, or
+ *                    timeout expires (best effort).
+ * @qp:               queue pair associated with SRQ to drain
+ *
+ * In order to avoid WQE and data segment leakage, one should destroy
+ * QP associated after performing the following:
+ *  - moving QP to err state
+ *  - wait for the Affiliated Asynchronous Last WQE Reached Event
+ *  - drain the CQ
+ */
+static void __ib_drain_srq(struct ib_qp *qp)
+{
+	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
+	struct ib_cq *cq;
+	int ret;
+
+	if (!qp->srq) {
+		WARN_ONCE(1, "QP 0x%p is not associated with SRQ\n", qp);
+		return;
+	}
+
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+	if (ret) {
+		WARN_ONCE(ret, "failed to drain shared recv queue: %d\n",
 ret);
+		return;
+	}
+
+	if (ib_srq_has_cq(qp->srq->srq_type)) {
+		cq = qp->srq->ext.cq;
+	} else if (qp->recv_cq) {
+		cq = qp->recv_cq;

drivers/infiniband/core/verbs.c +0 -0

 /**

* ib_drain_sq() - Block until all SQ CQEs have been consumed by the
* application.

drivers/infiniband/core/verbs.c +2 -0

@@ -2289,5 +2354,7 @@ void ib_drain_qp(struct ib_qp *qp)
 	ib_drain_sq(qp);
 	if (!qp->srq)
 		ib_drain_rq(qp);
+	else
+		__ib_drain_srq(qp);
 }
 EXPORT_SYMBOL(ib_drain_qp);

include/rdma/ib_verbs.h +8 -1

index fd84cda..c5febae 100644
@@ -1728,6 +1728,7 @@ struct ib_qp {
 	struct list_head	rdma_mrs;
 	struct list_head	sig_mrs;
 	struct ib_srq	       *srq;
+	struct completion	srq_completion;
 	struct ib_xrcd	       *xrcd; /* XRC TGT QPs only */
 	struct list_head	xrcd_list;
 

@@ -3060,6 +3061,13 @@ int ib_modify_qp(struct ib_qp *qp,
 		 int qp_attr_mask);
 
 /**
+ * ib_notify_qp - Notifies the QP for event arrival
+ * @qp: The QP to notify.
+ * @event: Specifies the event to notify.
+ */
+int ib_notify_qp(struct ib_qp *qp, enum ib_event_type event);
+
+/**
  * ib_query_qp - Returns the attribute list and current values for the
  *   specified QP.
  * @qp: The QP to query.
--

1.8.3.1

drivers/infiniband/core/verbs.c static void __ib_drain_rq(struct ib_qp *qp) · line 2279

Quoted 3 lines

drivers/infiniband/core/verbs.c +3 -0

+	if (wait_for_completion_timeout(&qp->srq_completion, 10 * HZ) > 0)
+		ib_process_cq_direct(cq, -1);
+}

Perhaps a WARN_ONCE is warranted?

drivers/infiniband/core/verbs.c +36 -0

+/*
+ * __ib_drain_srq() - Block until all Last WQE Reached event arrives, or
+ *                    timeout expires (best effort).
+ * @qp:               queue pair associated with SRQ to drain
+ *
+ * In order to avoid WQE and data segment leakage, one should destroy
+ * QP associated after performing the following:
+ *  - moving QP to err state
+ *  - wait for the Affiliated Asynchronous Last WQE Reached Event
+ *  - drain the CQ
+ */
+static void __ib_drain_srq(struct ib_qp *qp)
+{
+	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
+	struct ib_cq *cq;
+	int ret;
+
+	if (!qp->srq) {
+		WARN_ONCE(1, "QP 0x%p is not associated with SRQ\n", qp);
+		return;
+	}
+
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+	if (ret) {
+		WARN_ONCE(ret, "failed to drain shared recv queue: %d\n", ret);
+		return;
+	}
+
+	if (ib_srq_has_cq(qp->srq->srq_type)) {
+		cq = qp->srq->ext.cq;
+	} else if (qp->recv_cq) {
+		cq = qp->recv_cq;
+	} else {
+		WARN_ONCE(1, "QP 0x%p has no CQ associated with SRQ\n", qp);
+		return;
+	}

Quoted 40 lines · Max Gurtovoy

drivers/infiniband/core/verbs.c +40 -0

+/*
+ * __ib_drain_srq() - Block until all Last WQE Reached event arrives, or
+ *                    timeout expires (best effort).
+ * @qp:               queue pair associated with SRQ to drain
+ *
+ * In order to avoid WQE and data segment leakage, one should destroy
+ * QP associated after performing the following:
+ *  - moving QP to err state
+ *  - wait for the Affiliated Asynchronous Last WQE Reached Event
+ *  - drain the CQ
+ */
+static void __ib_drain_srq(struct ib_qp *qp)
+{
+	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
+	struct ib_cq *cq;
+	int ret;
+
+	if (!qp->srq) {
+		WARN_ONCE(1, "QP 0x%p is not associated with SRQ\n", qp);
+		return;
+	}
+
+	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+	if (ret) {
+		WARN_ONCE(ret, "failed to drain shared recv queue: %d\n", ret);
+		return;
+	}
+
+	if (ib_srq_has_cq(qp->srq->srq_type)) {
+		cq = qp->srq->ext.cq;
+	} else if (qp->recv_cq) {
+		cq = qp->recv_cq;
+	} else {
+		WARN_ONCE(1, "QP 0x%p has no CQ associated with SRQ\n", qp);
+		return;
+	}
+
+	/*
+         * ULP should invoke ib_notify_qp on IB_EVENT_QP_LAST_WQE_REACHED
+         * arrival, otherwise timeout will expire and leakage may occur.

drivers/infiniband/core/verbs.c static void __ib_drain_rq(struct ib_qp *qp) · line 2279

Quoted 6 lines

drivers/infiniband/core/verbs.c +6 -0

+         * Use long timeout, for the buggy ULPs/HCAs that don't notify the
+         * QP nor raising IB_EVENT_QP_LAST_WQE_REACHED event.
+         */
+	if (wait_for_completion_timeout(&qp->srq_completion, 10 * HZ) > 0)
+		ib_process_cq_direct(cq, -1);
+}

Hello Max,

It seems weird to me that __ib_drain_srq() does not follow the same approach as __ib_drain_rq(). Have you considered to post an additional receive work entry on the SRQ and to wait until the completion for that work entry is signaled?

That would avoid that a completion has to be added in the ib_qp data structure and would also avoid that all ULPs that use SRQs have to be modified.

Thanks,

Bart.

drivers/infiniband/core/verbs.c +0 -0

index 7868727..7604450 100644

Hi Max,

Quoted 12 lines

drivers/infiniband/core/verbs.c +1 -1

index 7868727..7604450 100644
@@ -886,8 +886,10 @@ struct ib_qp *ib_create_qp(struct ib_pd *pd,
 ????????? if (qp_init_attr->recv_cq)
 ????????????? atomic_inc(&qp_init_attr->recv_cq->usecnt);
 ????????? qp->srq = qp_init_attr->srq;
-??????? if (qp->srq)
+??????? if (qp->srq) {

drivers/infiniband/core/verbs.c struct ib_qp *ib_create_qp(struct ib_pd *pd, · line 893

Quoted 4 lines

drivers/infiniband/core/verbs.c +2 -0

 ????????????? atomic_inc(&qp_init_attr->srq->usecnt);
+??????????? init_completion(&qp->srq_completion);
+??????? }
 ????? }

How about initializing the completion at ib_drain_srq and complete it always in last wqe reached event in the core instead of relying on ULPs to call ib_notify_qp?

The simplest way I can think of is to have the core register its own event handler for that..

In order to avoid resource leakage for QP associated with a Shared Receive Queue (SRQ), notify it on Last WQE Reached event arrival.

Signed-off-by Max Gurtovoy <maxg at mellanox.com>

---
 drivers/nvme/target/rdma.c | 3 +++
 1 file changed, 3 insertions(+)

drivers/nvme/target/rdma.c +3 -0

index 4991290..99a14a7 100644
@@ -1151,6 +1151,9 @@ static void nvmet_rdma_qp_event(struct ib_event
 *event, void *priv)
 	case IB_EVENT_COMM_EST:
 		rdma_notify(queue->cm_id, event->event);
 		break;
+	case IB_EVENT_QP_LAST_WQE_REACHED:
+		ib_notify_qp(queue->cm_id->qp, event->event);
+		break;
 	default:
 		pr_err("received IB QP event: %s (%d)\n",

ib_event_msg(event->event), event->event);

Quoted 19 lines

In order to avoid resource leakage for QP associated with a Shared Receive Queue (SRQ), notify it on Last WQE Reached event arrival.

Signed-off-by Max Gurtovoy <maxg at mellanox.com>

---
 drivers/nvme/target/rdma.c | 3 +++
 1 file changed, 3 insertions(+)

drivers/nvme/target/rdma.c +0 -0

index 4991290..99a14a7 100644
@@ -1151,6 +1151,9 @@ static void nvmet_rdma_qp_event(struct ib_event
 *event, void *priv)
 	case IB_EVENT_COMM_EST:
 		rdma_notify(queue->cm_id, event->event);
 		break;

Quoted 6 lines

drivers/nvme/target/rdma.c +3 -0

+	case IB_EVENT_QP_LAST_WQE_REACHED:
+		ib_notify_qp(queue->cm_id->qp, event->event);
+		break;
 	default:
 		pr_err("received IB QP event: %s (%d)\n",

ib_event_msg(event->event), event->event);

I wonder if this could be handled in ib_dispatch_event() for all ULPS?