Testing determined that the recent commit 9e046bb111 ("tcp: clear
tp->retrans_stamp in tcp_rcv_fastopen_synack()") has a race, and does
not always ensure retrans_stamp is 0 after a TFO payload retransmit.
If transmit completion for the SYN+data skb happens after the client
TCP stack receives the SYNACK (which sometimes happens), then
retrans_stamp can erroneously remain non-zero for the lifetime of the
connection, causing a premature ETIMEDOUT later.
Testing and tracing showed that the buggy scenario is the following
somewhat tricky sequence:
+ Client attempts a TFO handshake. tcp_send_syn_data() sends SYN + TFO
cookie + data in a single packet in the syn_data skb. It hands the
syn_data skb to tcp_transmit_skb(), which makes a clone. Crucially,
it then reuses the same original (non-clone) syn_data skb,
transforming it by advancing the seq by one byte and removing the
FIN bit, and enques the resulting payload-only skb in the
sk->tcp_rtx_queue.
+ Client sets retrans_stamp to the start time of the three-way
handshake.
+ Cookie mismatches or server has TFO disabled, and server only ACKs
SYN.
+ tcp_ack() sees SYN is acked, tcp_clean_rtx_queue() clears
retrans_stamp.
+ Since the client SYN was acked but not the payload, the TFO failure
code path in tcp_rcv_fastopen_synack() tries to retransmit the
payload skb. However, in some cases the transmit completion for the
clone of the syn_data (which had SYN + TFO cookie + data) hasn't
happened. In those cases, skb_still_in_host_queue() returns true
for the retransmitted TFO payload, because the clone of the syn_data
skb has not had its tx completetion.
+ Because skb_still_in_host_queue() finds skb_fclone_busy() is true,
it sets the TSQ_THROTTLED bit and the retransmit does not happen in
the tcp_rcv_fastopen_synack() call chain.
+ The tcp_rcv_fastopen_synack() code next implicitly assumes the
retransmit process is finished, and sets retrans_stamp to 0 to clear
it, but this is later overwritten (see below).
+ Later, upon tx completion, tcp_tsq_write() calls
tcp_xmit_retransmit_queue(), which puts the retransmit in flight and
sets retrans_stamp to a non-zero value.
+ The client receives an ACK for the retransmitted TFO payload data.
+ Since we're in CA_Open and there are no dupacks/SACKs/DSACKs/ECN to
make tcp_ack_is_dubious() true and make us call
tcp_fastretrans_alert() and reach a code path that clears
retrans_stamp, retrans_stamp stays nonzero.
+ Later, if there is a TLP, RTO, RTO sequence, then the connection
will suffer an early ETIMEDOUT due to the erroneously ancient
retrans_stamp.
The fix: this commit refactors the code to have
tcp_rcv_fastopen_synack() retransmit by reusing the relevant parts of
tcp_simple_retransmit() that enter CA_Loss (without changing cwnd) and
call tcp_xmit_retransmit_queue(). We have tcp_simple_retransmit() and
tcp_rcv_fastopen_synack() share code in this way because in both cases
we get a packet indicating non-congestion loss (MTU reduction or TFO
failure) and thus in both cases we want to retransmit as many packets
as cwnd allows, without reducing cwnd. And given that retransmits will
set retrans_stamp to a non-zero value (and may do so in a later
calling context due to TSQ), we also want to enter CA_Loss so that we
track when all retransmitted packets are ACked and clear retrans_stamp
when that happens (to ensure later recurring RTOs are using the
correct retrans_stamp and don't declare ETIMEDOUT prematurely).
Fixes: 9e046bb111 ("tcp: clear tp->retrans_stamp in tcp_rcv_fastopen_synack()")
Fixes: a7abf3cd76 ("tcp: consider using standard rtx logic in tcp_rcv_fastopen_synack()")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Link: https://patch.msgid.link/20240624144323.2371403-1-ncardwell.sw@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
5dfe9d2739