RDS:TCP: Synchronize rds_tcp_accept_one with rds_send_xmit when resetting t_sock
authorSowmini Varadhan <sowmini.varadhan@oracle.com>
Mon, 2 May 2016 18:24:51 +0000 (11:24 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 3 May 2016 20:03:44 +0000 (16:03 -0400)
There is a race condition between rds_send_xmit -> rds_tcp_xmit
and the code that deals with resolution of duelling syns added
by commit 241b271952eb ("RDS-TCP: Reset tcp callbacks if re-using an
outgoing socket in rds_tcp_accept_one()").

Specifically, we may end up derefencing a null pointer in rds_send_xmit
if we have the interleaving sequence:
           rds_tcp_accept_one                  rds_send_xmit

                                             conn is RDS_CONN_UP, so
      invoke rds_tcp_xmit

                                             tc = conn->c_transport_data
        rds_tcp_restore_callbacks
            /* reset t_sock */
      null ptr deref from tc->t_sock

The race condition can be avoided without adding the overhead of
additional locking in the xmit path: have rds_tcp_accept_one wait
for rds_tcp_xmit threads to complete before resetting callbacks.
The synchronization can be done in the same manner as rds_conn_shutdown().
First set the rds_conn_state to something other than RDS_CONN_UP
(so that new threads cannot get into rds_tcp_xmit()), then wait for
RDS_IN_XMIT to be cleared in the conn->c_flags indicating that any
threads in rds_tcp_xmit are done.

Fixes: 241b271952eb ("RDS-TCP: Reset tcp callbacks if re-using an
outgoing socket in rds_tcp_accept_one()")
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/rds/tcp.c
net/rds/tcp_listen.c

index 61ed2a8764ba4c7a66788d7843cef18abb9e904a..9134544941c22fc000ca7e6702c7d9ec0aab9e5a 100644 (file)
@@ -127,7 +127,7 @@ void rds_tcp_restore_callbacks(struct socket *sock,
 
 /*
  * This is the only path that sets tc->t_sock.  Send and receive trust that
- * it is set.  The RDS_CONN_CONNECTED bit protects those paths from being
+ * it is set.  The RDS_CONN_UP bit protects those paths from being
  * called while it isn't set.
  */
 void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
index 0936a4a32b476fdde5c7208fc465ec3324bbcf09..0896187243d68c0ad1d3a324e7b396be367fa24e 100644 (file)
@@ -115,24 +115,32 @@ int rds_tcp_accept_one(struct socket *sock)
         * rds_tcp_state_change() will do that cleanup
         */
        rs_tcp = (struct rds_tcp_connection *)conn->c_transport_data;
-       if (rs_tcp->t_sock &&
-           ntohl(inet->inet_saddr) < ntohl(inet->inet_daddr)) {
-               struct sock *nsk = new_sock->sk;
-
-               nsk->sk_user_data = NULL;
-               nsk->sk_prot->disconnect(nsk, 0);
-               tcp_done(nsk);
-               new_sock = NULL;
-               ret = 0;
-               goto out;
-       } else if (rs_tcp->t_sock) {
-               rds_tcp_restore_callbacks(rs_tcp->t_sock, rs_tcp);
-               conn->c_outgoing = 0;
-       }
-
        rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
+       if (rs_tcp->t_sock) {
+               /* Need to resolve a duelling SYN between peers.
+                * We have an outstanding SYN to this peer, which may
+                * potentially have transitioned to the RDS_CONN_UP state,
+                * so we must quiesce any send threads before resetting
+                * c_transport_data.
+                */
+               wait_event(conn->c_waitq,
+                          !test_bit(RDS_IN_XMIT, &conn->c_flags));
+               if (ntohl(inet->inet_saddr) < ntohl(inet->inet_daddr)) {
+                       struct sock *nsk = new_sock->sk;
+
+                       nsk->sk_user_data = NULL;
+                       nsk->sk_prot->disconnect(nsk, 0);
+                       tcp_done(nsk);
+                       new_sock = NULL;
+                       ret = 0;
+                       goto out;
+               } else if (rs_tcp->t_sock) {
+                       rds_tcp_restore_callbacks(rs_tcp->t_sock, rs_tcp);
+                       conn->c_outgoing = 0;
+               }
+       }
        rds_tcp_set_callbacks(new_sock, conn);
-       rds_connect_complete(conn);
+       rds_connect_complete(conn); /* marks RDS_CONN_UP */
        new_sock = NULL;
        ret = 0;