]> git.baikalelectronics.ru Git - kernel.git/commitdiff
RDS: TCP: fix race windows in send-path quiescence by rds_tcp_accept_one()
authorSowmini Varadhan <sowmini.varadhan@oracle.com>
Sat, 4 Jun 2016 21:00:00 +0000 (14:00 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 7 Jun 2016 22:10:15 +0000 (15:10 -0700)
The send path needs to be quiesced before resetting callbacks from
rds_tcp_accept_one(), and commit eb192840266f ("RDS:TCP: Synchronize
rds_tcp_accept_one with rds_send_xmit when resetting t_sock") achieves
this using the c_state and RDS_IN_XMIT bit following the pattern
used by rds_conn_shutdown(). However this leaves the possibility
of a race window as shown in the sequence below
    take t_conn_lock in rds_tcp_conn_connect
    send outgoing syn to peer
    drop t_conn_lock in rds_tcp_conn_connect
    incoming from peer triggers rds_tcp_accept_one, conn is
marked CONNECTING
    wait for RDS_IN_XMIT to quiesce any rds_send_xmit threads
    call rds_tcp_reset_callbacks
    [.. race-window where incoming syn-ack can cause the conn
to be marked UP from rds_tcp_state_change ..]
    lock_sock called from rds_tcp_reset_callbacks, and we set
t_sock to null
As soon as the conn is marked UP in the race-window above, rds_send_xmit()
threads will proceed to rds_tcp_xmit and may encounter a null-pointer
deref on the t_sock.

Given that rds_tcp_state_change() is invoked in softirq context, whereas
rds_tcp_reset_callbacks() is in workq context, and testing for RDS_IN_XMIT
after lock_sock could result in a deadlock with tcp_sendmsg, this
commit fixes the race by using a new c_state, RDS_TCP_RESETTING, which
will prevent a transition to RDS_CONN_UP from rds_tcp_state_change().

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/rds.h
net/rds/tcp.c
net/rds/tcp_connect.c
net/rds/tcp_listen.c
net/rds/threads.c

index 80256b08eac0388d702f5981f00703d2118f5cce..387df5f32e49b9084363dc3da3ff62aa252e5507 100644 (file)
@@ -74,6 +74,7 @@ enum {
        RDS_CONN_CONNECTING,
        RDS_CONN_DISCONNECTING,
        RDS_CONN_UP,
+       RDS_CONN_RESETTING,
        RDS_CONN_ERROR,
 };
 
@@ -813,6 +814,7 @@ void rds_connect_worker(struct work_struct *);
 void rds_shutdown_worker(struct work_struct *);
 void rds_send_worker(struct work_struct *);
 void rds_recv_worker(struct work_struct *);
+void rds_connect_path_complete(struct rds_connection *conn, int curr);
 void rds_connect_complete(struct rds_connection *conn);
 
 /* transport.c */
index 7ab1b41ffc88aa5586b2fa6321477a6be4f44904..74ee126a6fe6c00ce8bce45053ec2b8805fc222e 100644 (file)
@@ -148,11 +148,23 @@ void rds_tcp_reset_callbacks(struct socket *sock,
         * potentially have transitioned to the RDS_CONN_UP state,
         * so we must quiesce any send threads before resetting
         * c_transport_data. We quiesce these threads by setting
-        * cp_state to something other than RDS_CONN_UP, and then
+        * c_state to something other than RDS_CONN_UP, and then
         * waiting for any existing threads in rds_send_xmit to
         * complete release_in_xmit(). (Subsequent threads entering
         * rds_send_xmit() will bail on !rds_conn_up().
+        *
+        * However an incoming syn-ack at this point would end up
+        * marking the conn as RDS_CONN_UP, and would again permit
+        * rds_send_xmi() threads through, so ideally we would
+        * synchronize on RDS_CONN_UP after lock_sock(), but cannot
+        * do that: waiting on !RDS_IN_XMIT after lock_sock() may
+        * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
+        * would not get set. As a result, we set c_state to
+        * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
+        * cannot mark rds_conn_path_up() in the window before lock_sock()
         */
+       atomic_set(&conn->c_state, RDS_CONN_RESETTING);
+       wait_event(conn->c_waitq, !test_bit(RDS_IN_XMIT, &conn->c_flags));
        lock_sock(osock->sk);
        /* reset receive side state for rds_tcp_data_recv() for osock  */
        if (tc->t_tinc) {
index fb82e0a0bf89361e005a23c371f0e95dfa4970bc..fba13d0305fb25f12c476e6014a2a9d4c5559075 100644 (file)
@@ -60,7 +60,7 @@ void rds_tcp_state_change(struct sock *sk)
                case TCP_SYN_RECV:
                        break;
                case TCP_ESTABLISHED:
-                       rds_connect_complete(conn);
+                       rds_connect_path_complete(conn, RDS_CONN_CONNECTING);
                        break;
                case TCP_CLOSE_WAIT:
                case TCP_CLOSE:
index d9fe53675d9584b1a51ab44064a5c7f51556f2e2..686b1d03a55858aeb59d9863339c2c8e99e03d90 100644 (file)
@@ -135,16 +135,15 @@ int rds_tcp_accept_one(struct socket *sock)
                    !conn->c_outgoing) {
                        goto rst_nsk;
                } else {
-                       atomic_set(&conn->c_state, RDS_CONN_CONNECTING);
-                       wait_event(conn->c_waitq,
-                                  !test_bit(RDS_IN_XMIT, &conn->c_flags));
                        rds_tcp_reset_callbacks(new_sock, conn);
                        conn->c_outgoing = 0;
+                       /* rds_connect_path_complete() marks RDS_CONN_UP */
+                       rds_connect_path_complete(conn, RDS_CONN_DISCONNECTING);
                }
        } else {
                rds_tcp_set_callbacks(new_sock, conn);
+               rds_connect_path_complete(conn, RDS_CONN_CONNECTING);
        }
-       rds_connect_complete(conn); /* marks RDS_CONN_UP */
        new_sock = NULL;
        ret = 0;
        goto out;
index 454aa6d23327b2a302bcfc72b6b3d857656e33ed..4a323045719b5c0aa0407916a772fe65aa900b0b 100644 (file)
@@ -71,9 +71,9 @@
 struct workqueue_struct *rds_wq;
 EXPORT_SYMBOL_GPL(rds_wq);
 
-void rds_connect_complete(struct rds_connection *conn)
+void rds_connect_path_complete(struct rds_connection *conn, int curr)
 {
-       if (!rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_UP)) {
+       if (!rds_conn_transition(conn, curr, RDS_CONN_UP)) {
                printk(KERN_WARNING "%s: Cannot transition to state UP, "
                                "current state is %d\n",
                                __func__,
@@ -90,6 +90,12 @@ void rds_connect_complete(struct rds_connection *conn)
        queue_delayed_work(rds_wq, &conn->c_send_w, 0);
        queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
 }
+EXPORT_SYMBOL_GPL(rds_connect_path_complete);
+
+void rds_connect_complete(struct rds_connection *conn)
+{
+       rds_connect_path_complete(conn, RDS_CONN_CONNECTING);
+}
 EXPORT_SYMBOL_GPL(rds_connect_complete);
 
 /*