tcp: fastopen: fix on syn-data transmit failure
authorEric Dumazet <edumazet@google.com>
Tue, 19 Sep 2017 17:05:57 +0000 (10:05 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 12 Oct 2017 09:51:21 +0000 (11:51 +0200)
commitb463521db854a0b73a14b34d782c51d6f7c87a77
treea463181bf4f8477cf681e2d939327b514fea3659
parentb13bc543b1e64edd5786dda8be614303f55b68f1
tcp: fastopen: fix on syn-data transmit failure

[ Upstream commit b5b7db8d680464b1d631fd016f5e093419f0bfd9 ]

Our recent change exposed a bug in TCP Fastopen Client that syzkaller
found right away [1]

When we prepare skb with SYN+DATA, we attempt to transmit it,
and we update socket state as if the transmit was a success.

In socket RTX queue we have two skbs, one with the SYN alone,
and a second one containing the DATA.

When (malicious) ACK comes in, we now complain that second one had no
skb_mstamp.

The proper fix is to make sure that if the transmit failed, we do not
pretend we sent the DATA skb, and make it our send_head.

When 3WHS completes, we can now send the DATA right away, without having
to wait for a timeout.

[1]
WARNING: CPU: 0 PID: 100189 at net/ipv4/tcp_input.c:3117 tcp_clean_rtx_queue+0x2057/0x2ab0 net/ipv4/tcp_input.c:3117()

 WARN_ON_ONCE(last_ackt == 0);

Modules linked in:
CPU: 0 PID: 100189 Comm: syz-executor1 Not tainted
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
 0000000000000000 ffff8800b35cb1d8 ffffffff81cad00d 0000000000000000
 ffffffff828a4347 ffff88009f86c080 ffffffff8316eb20 0000000000000d7f
 ffff8800b35cb220 ffffffff812c33c2 ffff8800baad2440 00000009d46575c0
Call Trace:
 [<ffffffff81cad00d>] __dump_stack
 [<ffffffff81cad00d>] dump_stack+0xc1/0x124
 [<ffffffff812c33c2>] warn_slowpath_common+0xe2/0x150
 [<ffffffff812c361e>] warn_slowpath_null+0x2e/0x40
 [<ffffffff828a4347>] tcp_clean_rtx_queue+0x2057/0x2ab0 n
 [<ffffffff828ae6fd>] tcp_ack+0x151d/0x3930
 [<ffffffff828baa09>] tcp_rcv_state_process+0x1c69/0x4fd0
 [<ffffffff828efb7f>] tcp_v4_do_rcv+0x54f/0x7c0
 [<ffffffff8258aacb>] sk_backlog_rcv
 [<ffffffff8258aacb>] __release_sock+0x12b/0x3a0
 [<ffffffff8258ad9e>] release_sock+0x5e/0x1c0
 [<ffffffff8294a785>] inet_wait_for_connect
 [<ffffffff8294a785>] __inet_stream_connect+0x545/0xc50
 [<ffffffff82886f08>] tcp_sendmsg_fastopen
 [<ffffffff82886f08>] tcp_sendmsg+0x2298/0x35a0
 [<ffffffff82952515>] inet_sendmsg+0xe5/0x520
 [<ffffffff8257152f>] sock_sendmsg_nosec
 [<ffffffff8257152f>] sock_sendmsg+0xcf/0x110

Fixes: 8c72c65b426b ("tcp: update skb->skb_mstamp more carefully")
Fixes: 783237e8daf1 ("net-tcp: Fast Open client - sending SYN-data")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
net/ipv4/tcp_output.c