tcp_cubic: do not set epoch_start in the future

Tracking idle time in bictcp_cwnd_event() is imprecise, as epoch_start
is normally set at ACK processing time, not at send time.

Doing a proper fix would need to add an additional state variable,
and does not seem worth the trouble, given CUBIC bug has been there
forever before Jana noticed it.

Let's simply not set epoch_start in the future, otherwise
bictcp_update() could overflow and CUBIC would again
grow cwnd too fast.

This was detected thanks to a packetdrill test Neal wrote that was flaky
before applying this fix.

Fixes: 3092752 ("tcp_cubic: better follow cubic curve after idle period")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Jana Iyengar <jri@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/ipv4/tcp_cubic.c

@@ -154,14 +154,20 @@ static void bictcp_init(struct sock *sk)
 static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 {
 	if (event == CA_EVENT_TX_START) {
-		s32 delta = tcp_time_stamp - tcp_sk(sk)->lsndtime;
 		struct bictcp *ca = inet_csk_ca(sk);
+		u32 now = tcp_time_stamp;
+		s32 delta;
+
+		delta = now - tcp_sk(sk)->lsndtime;
 
 		/* We were application limited (idle) for a while.
 		 * Shift epoch_start to keep cwnd growth to cubic curve.
 		 */
-		if (ca->epoch_start && delta > 0)
+		if (ca->epoch_start && delta > 0) {
 			ca->epoch_start += delta;
+			if (after(ca->epoch_start, now))
+				ca->epoch_start = now;
+		}
 		return;
 	}
 }