]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
tcp: avoid fastopen API to be used on AF_UNSPEC
authorWei Wang <weiwan@google.com>
Wed, 24 May 2017 16:59:31 +0000 (09:59 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 25 May 2017 17:30:34 +0000 (13:30 -0400)
Fastopen API should be used to perform fastopen operations on the TCP
socket. It does not make sense to use fastopen API to perform disconnect
by calling it with AF_UNSPEC. The fastopen data path is also prone to
race conditions and bugs when using with AF_UNSPEC.

One issue reported and analyzed by Vegard Nossum is as follows:
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Thread A:                            Thread B:
------------------------------------------------------------------------
sendto()
 - tcp_sendmsg()
     - sk_stream_memory_free() = 0
         - goto wait_for_sndbuf
     - sk_stream_wait_memory()
        - sk_wait_event() // sleep
          |                          sendto(flags=MSG_FASTOPEN, dest_addr=AF_UNSPEC)
  |                           - tcp_sendmsg()
  |                              - tcp_sendmsg_fastopen()
  |                                 - __inet_stream_connect()
  |                                    - tcp_disconnect() //because of AF_UNSPEC
  |                                       - tcp_transmit_skb()// send RST
  |                                    - return 0; // no reconnect!
  |                           - sk_stream_wait_connect()
  |                                 - sock_error()
  |                                    - xchg(&sk->sk_err, 0)
  |                                    - return -ECONNRESET
- ... // wake up, see sk->sk_err == 0
    - skb_entail() on TCP_CLOSE socket

If the connection is reopened then we will send a brand new SYN packet
after thread A has already queued a buffer. At this point I think the
socket internal state (sequence numbers etc.) becomes messed up.

When the new connection is closed, the FIN-ACK is rejected because the
sequence number is outside the window. The other side tries to
retransmit,
but __tcp_retransmit_skb() calls tcp_trim_head() on an empty skb which
corrupts the skb data length and hits a BUG() in copy_and_csum_bits().
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Hence, this patch adds a check for AF_UNSPEC in the fastopen data path
and return EOPNOTSUPP to user if such case happens.

Fixes: cf60af03ca4e7 ("tcp: Fast Open client - sendmsg(MSG_FASTOPEN)")
Reported-by: Vegard Nossum <vegard.nossum@oracle.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp.c

index 842b575f8fdddc41a41aa6f03fb9086cec7ee451..59792d283ff8c19048904cb790dbeebef14da73d 100644 (file)
@@ -1084,9 +1084,12 @@ static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
 {
        struct tcp_sock *tp = tcp_sk(sk);
        struct inet_sock *inet = inet_sk(sk);
+       struct sockaddr *uaddr = msg->msg_name;
        int err, flags;
 
-       if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
+       if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
+           (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
+            uaddr->sa_family == AF_UNSPEC))
                return -EOPNOTSUPP;
        if (tp->fastopen_req)
                return -EALREADY; /* Another Fast Open is in progress */
@@ -1108,7 +1111,7 @@ static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
                }
        }
        flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
-       err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
+       err = __inet_stream_connect(sk->sk_socket, uaddr,
                                    msg->msg_namelen, flags, 1);
        /* fastopen_req could already be freed in __inet_stream_connect
         * if the connection times out or gets rst