RDMA/cxgb3: Serialize calls to CQ's comp_handler
authorKumar Sanghvi <kumaras@chelsio.com>
Mon, 24 Oct 2011 15:50:22 +0000 (21:20 +0530)
committerRoland Dreier <roland@purestorage.com>
Mon, 31 Oct 2011 18:33:17 +0000 (11:33 -0700)
iw_cxgb3 has a potential problem where a CQ's comp_handler can get
called simultaneously from different places in iw_cxgb3 driver.  This
does not comply with Documentation/infiniband/core_locking.txt, which
states that at a given point of time, there should be only one
callback per CQ should be active.

Such problem was reported by Parav Pandit <Parav.Pandit@Emulex.Com>
for iw_cxgb4 driver.  Based on discussion between Parav Pandit and
Steve Wise, this patch fixes the above problem by serializing the
calls to a CQ's comp_handler using a spin_lock.

Signed-off-by: Kumar Sanghvi <kumaras@chelsio.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
drivers/infiniband/hw/cxgb3/iwch_ev.c
drivers/infiniband/hw/cxgb3/iwch_provider.c
drivers/infiniband/hw/cxgb3/iwch_provider.h
drivers/infiniband/hw/cxgb3/iwch_qp.c

index 71e0d845da3d976588542aebd2803e9d4a6ef2e9..abcc9e76962bc11bf442118c546cf49c399271b3 100644 (file)
@@ -46,6 +46,7 @@ static void post_qp_event(struct iwch_dev *rnicp, struct iwch_cq *chp,
        struct ib_event event;
        struct iwch_qp_attributes attrs;
        struct iwch_qp *qhp;
+       unsigned long flag;
 
        spin_lock(&rnicp->lock);
        qhp = get_qhp(rnicp, CQE_QPID(rsp_msg->cqe));
@@ -94,7 +95,9 @@ static void post_qp_event(struct iwch_dev *rnicp, struct iwch_cq *chp,
        if (qhp->ibqp.event_handler)
                (*qhp->ibqp.event_handler)(&event, qhp->ibqp.qp_context);
 
+       spin_lock_irqsave(&chp->comp_handler_lock, flag);
        (*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);
+       spin_unlock_irqrestore(&chp->comp_handler_lock, flag);
 
        if (atomic_dec_and_test(&qhp->refcnt))
                wake_up(&qhp->wait);
@@ -107,6 +110,7 @@ void iwch_ev_dispatch(struct cxio_rdev *rdev_p, struct sk_buff *skb)
        struct iwch_cq *chp;
        struct iwch_qp *qhp;
        u32 cqid = RSPQ_CQID(rsp_msg);
+       unsigned long flag;
 
        rnicp = (struct iwch_dev *) rdev_p->ulp;
        spin_lock(&rnicp->lock);
@@ -170,7 +174,9 @@ void iwch_ev_dispatch(struct cxio_rdev *rdev_p, struct sk_buff *skb)
                 */
                if (qhp->ep && SQ_TYPE(rsp_msg->cqe))
                        dst_confirm(qhp->ep->dst);
+               spin_lock_irqsave(&chp->comp_handler_lock, flag);
                (*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);
+               spin_unlock_irqrestore(&chp->comp_handler_lock, flag);
                break;
 
        case TPT_ERR_STAG:
index c7d9411f295427e9fe1ad946ebce3bfba212bb5f..37c224fc3ad9b8f31e0886507fc43f543ee547b0 100644 (file)
@@ -190,6 +190,7 @@ static struct ib_cq *iwch_create_cq(struct ib_device *ibdev, int entries, int ve
        chp->rhp = rhp;
        chp->ibcq.cqe = 1 << chp->cq.size_log2;
        spin_lock_init(&chp->lock);
+       spin_lock_init(&chp->comp_handler_lock);
        atomic_set(&chp->refcnt, 1);
        init_waitqueue_head(&chp->wait);
        if (insert_handle(rhp, &rhp->cqidr, chp, chp->cq.cqid)) {
index 9a342c9b220d240afebce55e5abfac6f67ee4953..87c14b0c5ac0c9514d9df59ddd39da08cfef8342 100644 (file)
@@ -103,6 +103,7 @@ struct iwch_cq {
        struct iwch_dev *rhp;
        struct t3_cq cq;
        spinlock_t lock;
+       spinlock_t comp_handler_lock;
        atomic_t refcnt;
        wait_queue_head_t wait;
        u32 __user *user_rptr_addr;
index ecd313f359a45ac6d90c0184ae1ccea3eeda7a6c..bea5839d89ee6b5f57b40aea7dc6b30f18e2fe68 100644 (file)
@@ -822,8 +822,11 @@ static void __flush_qp(struct iwch_qp *qhp, struct iwch_cq *rchp,
        flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count);
        spin_unlock(&qhp->lock);
        spin_unlock_irqrestore(&rchp->lock, *flag);
-       if (flushed)
+       if (flushed) {
+               spin_lock_irqsave(&rchp->comp_handler_lock, *flag);
                (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
+               spin_unlock_irqrestore(&rchp->comp_handler_lock, *flag);
+       }
 
        /* locking hierarchy: cq lock first, then qp lock. */
        spin_lock_irqsave(&schp->lock, *flag);
@@ -833,8 +836,11 @@ static void __flush_qp(struct iwch_qp *qhp, struct iwch_cq *rchp,
        flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count);
        spin_unlock(&qhp->lock);
        spin_unlock_irqrestore(&schp->lock, *flag);
-       if (flushed)
+       if (flushed) {
+               spin_lock_irqsave(&schp->comp_handler_lock, *flag);
                (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
+               spin_unlock_irqrestore(&schp->comp_handler_lock, *flag);
+       }
 
        /* deref */
        if (atomic_dec_and_test(&qhp->refcnt))
@@ -853,11 +859,15 @@ static void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
        if (qhp->ibqp.uobject) {
                cxio_set_wq_in_error(&qhp->wq);
                cxio_set_cq_in_error(&rchp->cq);
+               spin_lock_irqsave(&rchp->comp_handler_lock, *flag);
                (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
+               spin_unlock_irqrestore(&rchp->comp_handler_lock, *flag);
                if (schp != rchp) {
                        cxio_set_cq_in_error(&schp->cq);
+                       spin_lock_irqsave(&schp->comp_handler_lock, *flag);
                        (*schp->ibcq.comp_handler)(&schp->ibcq,
                                                   schp->ibcq.cq_context);
+                       spin_unlock_irqrestore(&schp->comp_handler_lock, *flag);
                }
                return;
        }