mptcp: Don't update meta-RTO from subflows that are retransmitting

Because, the RTO of these subflows is inflated. Thus, if you have one
bad subflow hanging around, the MPTCP RTO-timer can blow up to the order
of minutes. Better to ignore the RTO of these subflows.

The following packet-drill script illustrates the behavior:

--tolerance_usecs=100000
+0	`sysctl -w net.mptcp.mptcp_checksum=0`

+0	socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0	setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0	setsockopt(3, SOL_TCP, TCP_NODELAY, [1], 4) = 0
+0	fcntl(3, F_SETFL, O_RDWR|O_NONBLOCK) = 0
+0	bind(3, {sa_family = AF_INET, sin_port = htons(13000), sin_addr = inet_addr("192.168.0.1")}, ...) = 0
+0	listen(3, 1) = 0

+0	socket(..., SOCK_STREAM, IPPROTO_TCP) = 5
+0	setsockopt(5, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0	bind(5, {sa_family = AF_INET, sin_port = htons(13001), sin_addr = inet_addr("192.168.0.1")}, ...) = 0
+0	listen(5,1) = 0

+0	< S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7,mp_capable_no_cs key_a> sock(3)
+0	> S. 0:0(0) ack 1 win 28800 <mss 1460,nop,nop,sackOK,nop,wscale 7,mp_capable_no_cs key_b> sock(3)
+0	< . 1:1(0) ack 1 win 257 <mp_capable_no_cs key_a key_b> sock(3)
+0	accept(3, ..., ...) = 4
+0	fcntl(4, F_SETFL, O_RDWR|O_NONBLOCK) = 0

+0	write(4, ..., 1000) = 1000
+0	> P. 1:1001(1000) ack 1 <dss dack4 dsn4 nocs, nop, nop> sock(4)
+0	< . 1:1(0) ack 1001 win 4242 <dss dack4=1001> sock(4)
+0	write(4, ..., 1000) = 1000
+0	> P. 1001:2001(1000) ack 1 <dss dack4 dsn4 nocs, nop, nop> sock(4)
+0	write(4, ..., 1000) = 1000
+0	> P. 2001:3001(1000) ack 1 <dss dack4 dsn4 nocs, nop, nop> sock(4)
+0	write(4, ..., 1000) = 1000
+0	> P. 3001:4001(1000) ack 1 <dss dack4 dsn4 nocs, nop, nop> sock(4)
+0	write(4, ..., 1000) = 1000
+0	> P. 4001:5001(1000) ack 1 <dss dack4 dsn4 nocs, nop, nop> sock(4)

+0	< . 1:1(0) ack 5001 win 4242 <dss dack4=1001> sock(4)

/* ACK, without advencing dack: Now, the only way to recover is through a meta-retransmission timer */

/* Meta-level rexmit should fire in 400ms */

+0	write(4, ..., 1000) = 1000
+0	> P. 5001:6001(1000) ack 1 <dss dack4 dsn4=5001 nocs, nop, nop> sock(4)

+0.2	> P. 5001:6001(1000) ack 1 <dss dack4 dsn4=5001 nocs, nop, nop> sock(4)

/* Pushing out the timer */
+0.1	< . 1:1(0) ack 5001 win 4242 <dss dack4=2001> sock(4)

+0.2	> P. 5001:6001(1000) ack 1 <dss dack4 dsn4=5001 nocs, nop, nop> sock(4)

/* This one reschedules the meta-timer */
+0.01	< . 1:1(0) ack 5001 win 4242 <dss dack4=3001> sock(4)

+0	< S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7,mp_join_syn backup=0 address_id=0 token=sha1_32(key_b)> sock(5)
+0	> S. 0:0(0) ack 1 win 28800 <mss 1460,nop,nop,sackOK,nop,wscale 7, mp_join_syn_ack backup=0 address_id=0 sender_hmac=trunc_l64_hmac(key_b key_a) > sock(5)
+0	< . 1:1(0) ack 1 win 32792 <mp_join_ack sender_hmac=full_160_hmac(key_a key_b)> sock(5)
+0	mp_join_accept(5) = 6

+0	> . 1:1(0) ack 1 <...> sock(6) // reliably mp_join_ack

/* This first one is a reinjection. */
+0	> P. 1:1001(1000) ack 1 <dss dack4 dsn4=5001 nocs, nop, nop> sock(6)
+0	< . 1:1(0) ack 1001 win 4242 <dss dack4=3001> sock(6)

+0.4	> P. 5001:6001(1000) ack 1 <dss dack4 dsn4=5001 nocs, nop, nop> sock(4)

/* Fire, in 400ms after the above-mentioned ACK as the only working subflow has na RTO of 200ms */
+0	> P. 1001:2001(1000) ack 1 <dss dack4 dsn4=3001 nocs, nop, nop> sock(6)

+0	`sysctl -w net.mptcp.mptcp_checksum=1`

Signed-off-by: Christoph Paasch <[email protected]>
Signed-off-by: Matthieu Baerts <[email protected]>

include/net/mptcp.h

@@ -1195,6 +1195,7 @@ static inline void mptcp_set_rto(struct sock *sk)
 		struct sock *sk_it = mptcp_to_sock(mptcp);
 
 		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_rto > max_rto)
 			max_rto = inet_csk(sk_it)->icsk_rto;
 	}