tcp: fix early ETIMEDOUT after spurious non-SACK RTO

[ Upstream commit 686dc2db ]

Fix a bug reported and analyzed by Nagaraj Arankal, where the handling
of a spurious non-SACK RTO could cause a connection to fail to clear
retrans_stamp, causing a later RTO to very prematurely time out the
connection with ETIMEDOUT.

Here is the buggy scenario, expanding upon Nagaraj Arankal's excellent
report:

(*1) Send one data packet on a non-SACK connection

(*2) Because no ACK packet is received, the packet is retransmitted
     and we enter CA_Loss; but this retransmission is spurious.

(*3) The ACK for the original data is received. The transmitted packet
     is acknowledged.  The TCP timestamp is before the retrans_stamp,
     so tcp_may_undo() returns true, and tcp_try_undo_loss() returns
     true without changing state to Open (because tcp_is_sack() is
     false), and tcp_process_loss() returns without calling
     tcp_try_undo_recovery().  Normally after undoing a CA_Loss
     episode, tcp_fastretrans_alert() would see that the connection
     has returned to CA_Open and fall through and call
     tcp_try_to_open(), which would set retrans_stamp to 0.  However,
     for non-SACK connections we hold the connection in CA_Loss, so do
     not fall through to call tcp_try_to_open() and do not set
     retrans_stamp to 0. So retrans_stamp is (erroneously) still
     non-zero.

     At this point the first "retransmission event" has passed and
     been recovered from. Any future retransmission is a completely
     new "event". However, retrans_stamp is erroneously still
     set. (And we are still in CA_Loss, which is correct.)

(*4) After 16 minutes (to correspond with tcp_retries2=15), a new data
     packet is sent. Note: No data is transmitted between (*3) and
     (*4) and we disabled keep alives.

     The socket's timeout SHOULD be calculated from this point in
     time, but instead it's calculated from the prior "event" 16
     minutes ago (step (*2)).

(*5) Because no ACK packet is received, the packet is retransmitted.

(*6) At the time of the 2nd retransmission, the socket returns
     ETIMEDOUT, prematurely, because retrans_stamp is (erroneously)
     too far in the past (set at the time of (*2)).

This commit fixes this bug by ensuring that we reuse in
tcp_try_undo_loss() the same careful logic for non-SACK connections
that we have in tcp_try_undo_recovery(). To avoid duplicating logic,
we factor out that logic into a new
tcp_is_non_sack_preventing_reopen() helper and call that helper from
both undo functions.

Fixes: da34ac76

 ("tcp: only undo on partial ACKs in CA_Loss")
Reported-by: Nagaraj Arankal <nagaraj.p.arankal@hpe.com>
Link: https://lore.kernel.org/all/SJ0PR84MB1847BE6C24D274C46A1B9B0EB27A9@SJ0PR84MB1847.NAMPRD84.PROD.OUTLOOK.COM/
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20220903121023.866900-1-ncardwell.kernel@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>