tipc: fix link overflow issue at socket shutdown
authorTuong Lien <tuong.t.lien@dektech.com.au>
Wed, 8 Jan 2020 02:18:15 +0000 (09:18 +0700)
committerDavid S. Miller <davem@davemloft.net>
Wed, 8 Jan 2020 23:57:07 +0000 (15:57 -0800)
When a socket is suddenly shutdown or released, it will reject all the
unreceived messages in its receive queue. This applies to a connected
socket too, whereas there is only one 'FIN' message required to be sent
back to its peer in this case.

In case there are many messages in the queue and/or some connections
with such messages are shutdown at the same time, the link layer will
easily get overflowed at the 'TIPC_SYSTEM_IMPORTANCE' backlog level
because of the message rejections. As a result, the link will be taken
down. Moreover, immediately when the link is re-established, the socket
layer can continue to reject the messages and the same issue happens...

The commit refactors the '__tipc_shutdown()' function to only send one
'FIN' in the situation mentioned above. For the connectionless case, it
is unavoidable but usually there is no rejections for such socket
messages because they are 'dest-droppable' by default.

In addition, the new code makes the other socket states clear
(e.g.'TIPC_LISTEN') and treats as a separate case to avoid misbehaving.

Acked-by: Ying Xue <ying.xue@windriver.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/tipc/socket.c

index 6552f98..6ebd809 100644 (file)
@@ -287,12 +287,12 @@ static void tipc_sk_respond(struct sock *sk, struct sk_buff *skb, int err)
  *
  * Caller must hold socket lock
  */
-static void tsk_rej_rx_queue(struct sock *sk)
+static void tsk_rej_rx_queue(struct sock *sk, int error)
 {
        struct sk_buff *skb;
 
        while ((skb = __skb_dequeue(&sk->sk_receive_queue)))
-               tipc_sk_respond(sk, skb, TIPC_ERR_NO_PORT);
+               tipc_sk_respond(sk, skb, error);
 }
 
 static bool tipc_sk_connected(struct sock *sk)
@@ -545,34 +545,45 @@ static void __tipc_shutdown(struct socket *sock, int error)
        /* Remove pending SYN */
        __skb_queue_purge(&sk->sk_write_queue);
 
-       /* Reject all unreceived messages, except on an active connection
-        * (which disconnects locally & sends a 'FIN+' to peer).
-        */
-       while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
-               if (TIPC_SKB_CB(skb)->bytes_read) {
-                       kfree_skb(skb);
-                       continue;
-               }
-               if (!tipc_sk_type_connectionless(sk) &&
-                   sk->sk_state != TIPC_DISCONNECTING) {
-                       tipc_set_sk_state(sk, TIPC_DISCONNECTING);
-                       tipc_node_remove_conn(net, dnode, tsk->portid);
-               }
-               tipc_sk_respond(sk, skb, error);
+       /* Remove partially received buffer if any */
+       skb = skb_peek(&sk->sk_receive_queue);
+       if (skb && TIPC_SKB_CB(skb)->bytes_read) {
+               __skb_unlink(skb, &sk->sk_receive_queue);
+               kfree_skb(skb);
        }
 
-       if (tipc_sk_type_connectionless(sk))
+       /* Reject all unreceived messages if connectionless */
+       if (tipc_sk_type_connectionless(sk)) {
+               tsk_rej_rx_queue(sk, error);
                return;
+       }
 
-       if (sk->sk_state != TIPC_DISCONNECTING) {
+       switch (sk->sk_state) {
+       case TIPC_CONNECTING:
+       case TIPC_ESTABLISHED:
+               tipc_set_sk_state(sk, TIPC_DISCONNECTING);
+               tipc_node_remove_conn(net, dnode, tsk->portid);
+               /* Send a FIN+/- to its peer */
+               skb = __skb_dequeue(&sk->sk_receive_queue);
+               if (skb) {
+                       __skb_queue_purge(&sk->sk_receive_queue);
+                       tipc_sk_respond(sk, skb, error);
+                       break;
+               }
                skb = tipc_msg_create(TIPC_CRITICAL_IMPORTANCE,
                                      TIPC_CONN_MSG, SHORT_H_SIZE, 0, dnode,
                                      tsk_own_node(tsk), tsk_peer_port(tsk),
                                      tsk->portid, error);
                if (skb)
                        tipc_node_xmit_skb(net, skb, dnode, tsk->portid);
-               tipc_node_remove_conn(net, dnode, tsk->portid);
-               tipc_set_sk_state(sk, TIPC_DISCONNECTING);
+               break;
+       case TIPC_LISTEN:
+               /* Reject all SYN messages */
+               tsk_rej_rx_queue(sk, error);
+               break;
+       default:
+               __skb_queue_purge(&sk->sk_receive_queue);
+               break;
        }
 }
 
@@ -2643,7 +2654,7 @@ static int tipc_accept(struct socket *sock, struct socket *new_sock, int flags,
         * Reject any stray messages received by new socket
         * before the socket lock was taken (very, very unlikely)
         */
-       tsk_rej_rx_queue(new_sk);
+       tsk_rej_rx_queue(new_sk, TIPC_ERR_NO_PORT);
 
        /* Connect new socket to it's peer */
        tipc_sk_finish_conn(new_tsock, msg_origport(msg), msg_orignode(msg));