tcp: fix incorrect undo caused by DSACK of TLP retransmit

Loss recovery undo_retrans bookkeeping had a long-standing bug where a
DSACK from a spurious TLP retransmit packet could cause an erroneous
undo of a fast recovery or RTO recovery that repaired a single
really-lost packet (in a sequence range outside that of the TLP
retransmit). Basically, because the loss recovery state machine didn't
account for the fact that it sent a TLP retransmit, the DSACK for the
TLP retransmit could erroneously be implicitly be interpreted as
corresponding to the normal fast recovery or RTO recovery retransmit
that plugged a real hole, thus resulting in an improper undo.

For example, consider the following buggy scenario where there is a
real packet loss but the congestion control response is improperly
undone because of this bug:

+ send packets P1, P2, P3, P4
+ P1 is really lost
+ send TLP retransmit of P4
+ receive SACK for original P2, P3, P4
+ enter fast recovery, fast-retransmit P1, increment undo_retrans to 1
+ receive DSACK for TLP P4, decrement undo_retrans to 0, undo (bug!)
+ receive cumulative ACK for P1-P4 (fast retransmit plugged real hole)

The fix: when we initialize undo machinery in tcp_init_undo(), if
there is a TLP retransmit in flight, then increment tp->undo_retrans
so that we make sure that we receive a DSACK corresponding to the TLP
retransmit, as well as DSACKs for all later normal retransmits, before
triggering a loss recovery undo. Note that we also have to move the
line that clears tp->tlp_high_seq for RTO recovery, so that upon RTO
we remember the tp->tlp_high_seq value until tcp_init_undo() and clear
it only afterward.

Also note that the bug dates back to the original 2013 TLP
implementation, commit 6ba8a3b1 ("tcp: Tail loss probe (TLP)").

However, this patch will only compile and work correctly with kernels
that have tp->tlp_retrans, which was added only in v5.8 in 2020 in
commit 76be93fc ("tcp: allow at most one TLP probe per flight").
So we associate this fix with that later commit.

Fixes: 76be93fc ("tcp: allow at most one TLP probe per flight")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Kevin Yang <yyd@google.com>
Link: https://patch.msgid.link/20240703171246.1739561-1-ncardwell.sw@gmail.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

net/ipv4/tcp_input.c

@@ -2129,8 +2129,16 @@ void tcp_clear_retrans(struct tcp_sock *tp)
 static inline void tcp_init_undo(struct tcp_sock *tp)
 {
 	tp->undo_marker = tp->snd_una;
+
 	/* Retransmission still in flight may cause DSACKs later. */
-	tp->undo_retrans = tp->retrans_out ? : -1;
+	/* First, account for regular retransmits in flight: */
+	tp->undo_retrans = tp->retrans_out;
+	/* Next, account for TLP retransmits in flight: */
+	if (tp->tlp_high_seq && tp->tlp_retrans)
+		tp->undo_retrans++;
+	/* Finally, avoid 0, because undo_retrans==0 means "can undo now": */
+	if (!tp->undo_retrans)
+		tp->undo_retrans = -1;
 }
 
 static bool tcp_is_rack(const struct sock *sk)
@@ -2209,6 +2217,7 @@ void tcp_enter_loss(struct sock *sk)
 
 	tcp_set_ca_state(sk, TCP_CA_Loss);
 	tp->high_seq = tp->snd_nxt;
+	tp->tlp_high_seq = 0;
 	tcp_ecn_queue_cwr(tp);
 
 	/* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous

net/ipv4/tcp_timer.c

@@ -536,8 +536,6 @@ void tcp_retransmit_timer(struct sock *sk)
 	if (WARN_ON_ONCE(!skb))
 		return;
 
-	tp->tlp_high_seq = 0;
-
 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
 		/* Receiver dastardly shrinks window. Our retransmits