xprtrdma: Ensure ia->ri_id->qp is not NULL when reconnecting
authorChuck Lever <chuck.lever@oracle.com>
Wed, 28 May 2014 14:34:07 +0000 (10:34 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Wed, 4 Jun 2014 12:56:47 +0000 (08:56 -0400)
Devesh Sharma <Devesh.Sharma@Emulex.Com> reports that after a
disconnect, his HCA is failing to create a fresh QP, leaving
ia_ri->ri_id->qp set to NULL. But xprtrdma still allows RPCs to
wake up and post LOCAL_INV as they exit, causing an oops.

rpcrdma_ep_connect() is allowing the wake-up by leaking the QP
creation error code (-EPERM in this case) to the RPC client's
generic layer. xprt_connect_status() does not recognize -EPERM, so
it kills pending RPC tasks immediately rather than retrying the
connect.

Re-arrange the QP creation logic so that when it fails on reconnect,
it leaves ->qp with the old QP rather than NULL.  If pending RPC
tasks wake and exit, LOCAL_INV work requests will flush rather than
oops.

On initial connect, leaving ->qp == NULL is OK, since there are no
pending RPCs that might use ->qp. But be sure not to try to destroy
a NULL QP when rpcrdma_ep_connect() is retried.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
net/sunrpc/xprtrdma/verbs.c

index c80995af82de819dee79f463ed7254b75f08191e..54edf2ac48a1d685447a7ae3ee512a42cc73e263 100644 (file)
@@ -867,6 +867,7 @@ rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
        if (ep->rep_connected != 0) {
                struct rpcrdma_xprt *xprt;
 retry:
+               dprintk("RPC:       %s: reconnecting...\n", __func__);
                rc = rpcrdma_ep_disconnect(ep, ia);
                if (rc && rc != -ENOTCONN)
                        dprintk("RPC:       %s: rpcrdma_ep_disconnect"
@@ -879,7 +880,7 @@ retry:
                id = rpcrdma_create_id(xprt, ia,
                                (struct sockaddr *)&xprt->rx_data.addr);
                if (IS_ERR(id)) {
-                       rc = PTR_ERR(id);
+                       rc = -EHOSTUNREACH;
                        goto out;
                }
                /* TEMP TEMP TEMP - fail if new device:
@@ -893,20 +894,30 @@ retry:
                        printk("RPC:       %s: can't reconnect on "
                                "different device!\n", __func__);
                        rdma_destroy_id(id);
-                       rc = -ENETDOWN;
+                       rc = -ENETUNREACH;
                        goto out;
                }
                /* END TEMP */
+               rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
+               if (rc) {
+                       dprintk("RPC:       %s: rdma_create_qp failed %i\n",
+                               __func__, rc);
+                       rdma_destroy_id(id);
+                       rc = -ENETUNREACH;
+                       goto out;
+               }
                rdma_destroy_qp(ia->ri_id);
                rdma_destroy_id(ia->ri_id);
                ia->ri_id = id;
-       }
-
-       rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
-       if (rc) {
-               dprintk("RPC:       %s: rdma_create_qp failed %i\n",
-                       __func__, rc);
-               goto out;
+       } else {
+               dprintk("RPC:       %s: connecting...\n", __func__);
+               rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
+               if (rc) {
+                       dprintk("RPC:       %s: rdma_create_qp failed %i\n",
+                               __func__, rc);
+                       /* do not update ep->rep_connected */
+                       return -ENETUNREACH;
+               }
        }
 
 /* XXX Tavor device performs badly with 2K MTU! */