tcp: fix TFO SYNACK undo to avoid double-timestamp-undo
authorNeal Cardwell <ncardwell@google.com>
Sat, 22 Feb 2020 16:21:15 +0000 (11:21 -0500)
committerDavid S. Miller <davem@davemloft.net>
Mon, 24 Feb 2020 01:23:35 +0000 (17:23 -0800)
commitdad8cea7add96a353fa1898b5ccefbb72da66f29
tree71c9b40690b3ef06f0dbeeb48fbf79f82624d667
parentf6f13c125e05603f68f5bf31f045b95e6d493598
tcp: fix TFO SYNACK undo to avoid double-timestamp-undo

In a rare corner case the new logic for undo of SYNACK RTO could
result in triggering the warning in tcp_fastretrans_alert() that says:
        WARN_ON(tp->retrans_out != 0);

The warning looked like:

WARNING: CPU: 1 PID: 1 at net/ipv4/tcp_input.c:2818 tcp_ack+0x13e0/0x3270

The sequence that tickles this bug is:
 - Fast Open server receives TFO SYN with data, sends SYNACK
 - (client receives SYNACK and sends ACK, but ACK is lost)
 - server app sends some data packets
 - (N of the first data packets are lost)
 - server receives client ACK that has a TS ECR matching first SYNACK,
   and also SACKs suggesting the first N data packets were lost
    - server performs TS undo of SYNACK RTO, then immediately
      enters recovery
    - buggy behavior then performed a *second* undo that caused
      the connection to be in CA_Open with retrans_out != 0

Basically, the incoming ACK packet with SACK blocks causes us to first
undo the cwnd reduction from the SYNACK RTO, but then immediately
enters fast recovery, which then makes us eligible for undo again. And
then tcp_rcv_synrecv_state_fastopen() accidentally performs an undo
using a "mash-up" of state from two different loss recovery phases: it
uses the timestamp info from the ACK of the original SYNACK, and the
undo_marker from the fast recovery.

This fix refines the logic to only invoke the tcp_try_undo_loss()
inside tcp_rcv_synrecv_state_fastopen() if the connection is still in
CA_Loss.  If peer SACKs triggered fast recovery, then
tcp_rcv_synrecv_state_fastopen() can't safely undo.

Fixes: 794200d66273 ("tcp: undo cwnd on Fast Open spurious SYNACK retransmit")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_input.c