@@ -249,6 +249,11 @@ xprt_rdma_connect_worker(struct work_struct *work)
xprt->stat.connect_start;
xprt_set_connected(xprt);
rc = -EAGAIN;
+ } else {
+ /* Force a call to xprt_rdma_close to clean up */
+ spin_lock(&xprt->transport_lock);
+ set_bit(XPRT_CLOSE_WAIT, &xprt->state);
+ spin_unlock(&xprt->transport_lock);
}
xprt_wake_pending_tasks(xprt, rc);
}
@@ -290,7 +290,7 @@ rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
sap, rdma_reject_msg(id, event->status));
ep->re_connect_status = -ECONNREFUSED;
if (event->status == IB_CM_REJ_STALE_CONN)
- ep->re_connect_status = -EAGAIN;
+ ep->re_connect_status = -ENOTCONN;
goto disconnected;
case RDMA_CM_EVENT_DISCONNECTED:
ep->re_connect_status = -ECONNABORTED;
@@ -521,8 +521,6 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
struct rpcrdma_ep *ep;
int rc;
-retry:
- rpcrdma_xprt_disconnect(r_xprt);
rc = rpcrdma_ep_create(r_xprt);
if (rc)
return rc;
@@ -550,17 +548,13 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
wait_event_interruptible(ep->re_connect_wait,
ep->re_connect_status != 0);
if (ep->re_connect_status <= 0) {
- if (ep->re_connect_status == -EAGAIN)
- goto retry;
rc = ep->re_connect_status;
goto out;
}
rc = rpcrdma_reqs_setup(r_xprt);
- if (rc) {
- rpcrdma_xprt_disconnect(r_xprt);
+ if (rc)
goto out;
- }
rpcrdma_mrs_create(r_xprt);
out:
Both Dan and I have observed two processes invoking rpcrdma_xprt_disconnect() concurrently. In my case: 1. The connect worker invokes rpcrdma_xprt_disconnect(), which drains the QP and waits for the final completion 2. This causes the newly posted Receive to flush and invoke xprt_force_disconnect() 3. xprt_force_disconnect() sets CLOSE_WAIT and wakes up the RPC task that is holding the transport lock 4. The RPC task invokes xprt_connect(), which calls ->ops->close 5. xprt_rdma_close() invokes rpcrdma_xprt_disconnect(), which tries to destroy the QP. Deadlock. To prevent xprt_force_disconnect() from waking anything, handle the clean up after a failed connection attempt in the xprt's sndtask. The retry loop is removed from rpcrdma_xprt_connect() to ensure that the newly allocated ep and id are properly released before a REJECTED connection attempt can be retried. Reported-by: Dan Aloni <dan@kernelim.com> Fixes: e28ce90083f0 ("xprtrdma: kmalloc rpcrdma_ep separate from rpcrdma_xprt") Signed-off-by: Chuck Lever <chuck.lever@oracle.com> --- net/sunrpc/xprtrdma/transport.c | 5 +++++ net/sunrpc/xprtrdma/verbs.c | 10 ++-------- 2 files changed, 7 insertions(+), 8 deletions(-)