tcp: Reject invalid ack_seq to Fast Open sockets
authorJerry Chu <hkchu@google.com>
Mon, 22 Oct 2012 11:26:36 +0000 (11:26 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 23 Oct 2012 06:42:56 +0000 (02:42 -0400)
A packet with an invalid ack_seq may cause a TCP Fast Open socket to switch
to the unexpected TCP_CLOSING state, triggering a BUG_ON kernel panic.

When a FIN packet with an invalid ack_seq# arrives at a socket in
the TCP_FIN_WAIT1 state, rather than discarding the packet, the current
code will accept the FIN, causing state transition to TCP_CLOSING.

This may be a small deviation from RFC793, which seems to say that the
packet should be dropped. Unfortunately I did not expect this case for
Fast Open hence it will trigger a BUG_ON panic.

It turns out there is really nothing bad about a TFO socket going into
TCP_CLOSING state so I could just remove the BUG_ON statements. But after
some thought I think it's better to treat this case like TCP_SYN_RECV
and return a RST to the confused peer who caused the unacceptable ack_seq
to be generated in the first place.

Signed-off-by: H.K. Jerry Chu <hkchu@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_input.c
net/ipv4/tcp_timer.c

index 036f857381412c4ca98b8cb7eb9ed34457ba520b..1db6639835877bfef79df2d88034bb0e1fdfa082 100644 (file)
@@ -5964,7 +5964,7 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
 
        req = tp->fastopen_rsk;
        if (req != NULL) {
-               BUG_ON(sk->sk_state != TCP_SYN_RECV &&
+               WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
                    sk->sk_state != TCP_FIN_WAIT1);
 
                if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
@@ -6053,7 +6053,15 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
                         * ACK we have received, this would have acknowledged
                         * our SYNACK so stop the SYNACK timer.
                         */
-                       if (acceptable && req != NULL) {
+                       if (req != NULL) {
+                               /* Return RST if ack_seq is invalid.
+                                * Note that RFC793 only says to generate a
+                                * DUPACK for it but for TCP Fast Open it seems
+                                * better to treat this case like TCP_SYN_RECV
+                                * above.
+                                */
+                               if (!acceptable)
+                                       return 1;
                                /* We no longer need the request sock. */
                                reqsk_fastopen_remove(sk, req, false);
                                tcp_rearm_rto(sk);
index fc04711e80c89dd0dc92ff1027efaa7324b218c3..d47c1b4421a339a3378ef23646c8f99b41d7490b 100644 (file)
@@ -347,8 +347,8 @@ void tcp_retransmit_timer(struct sock *sk)
                return;
        }
        if (tp->fastopen_rsk) {
-               BUG_ON(sk->sk_state != TCP_SYN_RECV &&
-                   sk->sk_state != TCP_FIN_WAIT1);
+               WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
+                            sk->sk_state != TCP_FIN_WAIT1);
                tcp_fastopen_synack_timer(sk);
                /* Before we receive ACK to our SYN-ACK don't retransmit
                 * anything else (e.g., data or FIN segments).