mptcp: Only update meta-RTO when we have a valid RTT measurement

Commit fc29b3a ("mptcp: Don't update meta-RTO from subflows that are
retransmitting") for the most part already addressed the issue where a
subflow with an inflated RTO due to retransmissions might push the
meta-RTO to huge values.

However, there is one possibility where icsk_retransmits is 0, while the
RTO is still inflated. When an ACK comes in that ends up not updating
the RTT (see tcp_ack_update_rtt) but is still acknowledging data, we end
up in a situation where icsk_retransmits is 0 but the subflow's RTO is
still huge. The one additional input to whether we have a valid RTO is
icsk_backoff (see comment in tcp_ack_update_rtt()).

So, we need to take this into account as well.

Fixes: fc29b3a ("mptcp: Don't update meta-RTO from subflows that are retransmitting")
Signed-off-by: Christoph Paasch <[email protected]>
Signed-off-by: Matthieu Baerts <[email protected]>
(cherry picked from commit 3944acd)
Signed-off-by: Matthieu Baerts <[email protected]>

include/net/mptcp.h

@@ -1196,6 +1196,7 @@ static inline void mptcp_set_rto(struct sock *sk)
 
 		if ((mptcp_sk_can_send(sk_it) || sk_it->sk_state == TCP_SYN_RECV) &&
 		    inet_csk(sk_it)->icsk_retransmits == 0 &&
+		    inet_csk(sk_it)->icsk_backoff == 0 &&
 		    inet_csk(sk_it)->icsk_rto > max_rto)
 			max_rto = inet_csk(sk_it)->icsk_rto;
 	}