IB/srp: Fix a race condition triggered by destroying a queue pair
authorBart Van Assche <bvanassche@acm.org>
Tue, 21 Oct 2014 16:00:35 +0000 (18:00 +0200)
committerChristoph Hellwig <hch@lst.de>
Wed, 12 Nov 2014 11:05:24 +0000 (12:05 +0100)
At least LID reassignment can trigger a race condition in the SRP
initiator driver, namely the receive completion handler trying to
post a request on a QP during or after QP destruction and before
the CQ's have been destroyed. Avoid this race by modifying a QP
into the error state and by waiting until all receive completions
have been processed before destroying a QP.

Reported-by: Max Gurtuvoy <maxg@mellanox.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
drivers/infiniband/ulp/srp/ib_srp.c
drivers/infiniband/ulp/srp/ib_srp.h

index aac844a6eef630a0996d7070fca10927a3920a60..57b5ff1bbcb6597452d8bf99b8ee7bf247cc1d4e 100644 (file)
@@ -453,6 +453,41 @@ static struct srp_fr_pool *srp_alloc_fr_pool(struct srp_target_port *target)
                                  dev->max_pages_per_mr);
 }
 
+/**
+ * srp_destroy_qp() - destroy an RDMA queue pair
+ * @ch: SRP RDMA channel.
+ *
+ * Change a queue pair into the error state and wait until all receive
+ * completions have been processed before destroying it. This avoids that
+ * the receive completion handler can access the queue pair while it is
+ * being destroyed.
+ */
+static void srp_destroy_qp(struct srp_rdma_ch *ch)
+{
+       struct srp_target_port *target = ch->target;
+       static struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
+       static struct ib_recv_wr wr = { .wr_id = SRP_LAST_WR_ID };
+       struct ib_recv_wr *bad_wr;
+       int ret;
+
+       /* Destroying a QP and reusing ch->done is only safe if not connected */
+       WARN_ON_ONCE(target->connected);
+
+       ret = ib_modify_qp(ch->qp, &attr, IB_QP_STATE);
+       WARN_ONCE(ret, "ib_cm_init_qp_attr() returned %d\n", ret);
+       if (ret)
+               goto out;
+
+       init_completion(&ch->done);
+       ret = ib_post_recv(ch->qp, &wr, &bad_wr);
+       WARN_ONCE(ret, "ib_post_recv() returned %d\n", ret);
+       if (ret == 0)
+               wait_for_completion(&ch->done);
+
+out:
+       ib_destroy_qp(ch->qp);
+}
+
 static int srp_create_ch_ib(struct srp_rdma_ch *ch)
 {
        struct srp_target_port *target = ch->target;
@@ -469,8 +504,9 @@ static int srp_create_ch_ib(struct srp_rdma_ch *ch)
        if (!init_attr)
                return -ENOMEM;
 
+       /* + 1 for SRP_LAST_WR_ID */
        recv_cq = ib_create_cq(dev->dev, srp_recv_completion, NULL, ch,
-                              target->queue_size, ch->comp_vector);
+                              target->queue_size + 1, ch->comp_vector);
        if (IS_ERR(recv_cq)) {
                ret = PTR_ERR(recv_cq);
                goto err;
@@ -487,7 +523,7 @@ static int srp_create_ch_ib(struct srp_rdma_ch *ch)
 
        init_attr->event_handler       = srp_qp_event;
        init_attr->cap.max_send_wr     = m * target->queue_size;
-       init_attr->cap.max_recv_wr     = target->queue_size;
+       init_attr->cap.max_recv_wr     = target->queue_size + 1;
        init_attr->cap.max_recv_sge    = 1;
        init_attr->cap.max_send_sge    = 1;
        init_attr->sq_sig_type         = IB_SIGNAL_REQ_WR;
@@ -530,7 +566,7 @@ static int srp_create_ch_ib(struct srp_rdma_ch *ch)
        }
 
        if (ch->qp)
-               ib_destroy_qp(ch->qp);
+               srp_destroy_qp(ch);
        if (ch->recv_cq)
                ib_destroy_cq(ch->recv_cq);
        if (ch->send_cq)
@@ -586,7 +622,7 @@ static void srp_free_ch_ib(struct srp_target_port *target,
                if (ch->fmr_pool)
                        ib_destroy_fmr_pool(ch->fmr_pool);
        }
-       ib_destroy_qp(ch->qp);
+       srp_destroy_qp(ch);
        ib_destroy_cq(ch->send_cq);
        ib_destroy_cq(ch->recv_cq);
 
@@ -1883,8 +1919,15 @@ static void srp_tl_err_work(struct work_struct *work)
 }
 
 static void srp_handle_qp_err(u64 wr_id, enum ib_wc_status wc_status,
-                             bool send_err, struct srp_target_port *target)
+                             bool send_err, struct srp_rdma_ch *ch)
 {
+       struct srp_target_port *target = ch->target;
+
+       if (wr_id == SRP_LAST_WR_ID) {
+               complete(&ch->done);
+               return;
+       }
+
        if (target->connected && !target->qp_in_error) {
                if (wr_id & LOCAL_INV_WR_ID_MASK) {
                        shost_printk(KERN_ERR, target->scsi_host, PFX
@@ -1915,8 +1958,7 @@ static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr)
                if (likely(wc.status == IB_WC_SUCCESS)) {
                        srp_handle_recv(ch, &wc);
                } else {
-                       srp_handle_qp_err(wc.wr_id, wc.status, false,
-                                         ch->target);
+                       srp_handle_qp_err(wc.wr_id, wc.status, false, ch);
                }
        }
 }
@@ -1932,8 +1974,7 @@ static void srp_send_completion(struct ib_cq *cq, void *ch_ptr)
                        iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
                        list_add(&iu->list, &ch->free_tx);
                } else {
-                       srp_handle_qp_err(wc.wr_id, wc.status, true,
-                                         ch->target);
+                       srp_handle_qp_err(wc.wr_id, wc.status, true, ch);
                }
        }
 }
index ca7c6f0654340823c565d1ea22dabe54da6b789c..a611556406ac05bb1a4befbc870f510880fce2fe 100644 (file)
@@ -70,6 +70,8 @@ enum {
 
        LOCAL_INV_WR_ID_MASK    = 1,
        FAST_REG_WR_ID_MASK     = 2,
+
+       SRP_LAST_WR_ID          = 0xfffffffcU,
 };
 
 enum srp_target_state {