RDMA/iwcm: Don't call provider reject func with irqs disabled

In commit cb58160e ("RDMA/iwcm: Reject the connection when the cm_id
is destroyed") a call to the provider's reject handler was added to
destroy_cm_id() to fix a provider endpoint leak.  This call needs to
be done with interrupts enabled.  So unlock and relock around this
call.  This is safe because:

1) the provider will do nothing with this endpoint until the iwcm either
   accepts or rejects.
2) the lock is only released after the iwcm state is changed, so an
   errant iwcm app that is destroying -and- rejecting the connection
   concurrently will get a failure on one of the calls.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

1 files changed, 2 insertions, 0 deletions

diff --git a/drivers/infiniband/core/iwcm.c b/drivers/infiniband/core/iwcm.c
index 55d093a36ae4..625fec5a741c 100644
--- a/drivers/infiniband/core/iwcm.c
+++ b/drivers/infiniband/core/iwcm.c
@@ -362,7 +362,9 @@ static void destroy_cm_id(struct iw_cm_id *cm_id)
                  * In either case, must tell the provider to reject.
                  */
                 cm_id_priv->state = IW_CM_STATE_DESTROYING;
+                spin_unlock_irqrestore(&cm_id_priv->lock, flags);
                 cm_id->device->iwcm->reject(cm_id, NULL, 0);
+                spin_lock_irqsave(&cm_id_priv->lock, flags);
                 break;
         case IW_CM_STATE_CONN_SENT:
         case IW_CM_STATE_DESTROYING: