mptcp: Avoid tcp_done and mptcp_sock_def_error_report warnings

It seems to be possible that we are in tcp_keepalive_timer with
send_mp_fclose set to 1 and during that time a call to tcp_close ends up
closing the subflow.

Once we successfully manage to pass the bh_lock_sock() and sock_owned*()
barrier in tcp_keepalive_timer we then may end up in the condition
triggering the sending of a fast-close. But, the socket is now in
TCP_CLOSE state, thus leading to an error in tcp_done if we end up
calling tcp_write_err().

All timers handle this kind of a "race" by checking sk_state for
TCP_CLOSE right after passing the bh_lock_sock() barrier. Even
tcp_keepalive_timer makes sure that we won't do anything on a socket in
TCP_CLOSE state.

Thus, we need to check for sk_state before entering the send_mp_fclose
condition.

Fixes: a313bd7 ("mptcp: Correctly wait for RST when FASTCLOSE'ing")
Signed-off-by: Christoph Paasch <cpaasch@apple.com>
Signed-off-by: Matthieu Baerts <matthieu.baerts@tessares.net>
(cherry picked from commit 80a651c)
Signed-off-by: Matthieu Baerts <matthieu.baerts@tessares.net>
(cherry picked from commit db07838)
Signed-off-by: Matthieu Baerts <matthieu.baerts@tessares.net>
(cherry picked from commit ef76cf9)
Signed-off-by: Matthieu Baerts <matthieu.baerts@tessares.net>

net/ipv4/tcp.c

@@ -2168,6 +2168,7 @@ static const unsigned char new_state[16] = {
   [TCP_LISTEN]		= TCP_CLOSE,
   [TCP_CLOSING]		= TCP_CLOSING,
   [TCP_NEW_SYN_RECV]	= TCP_CLOSE,	/* should not happen ! */
+  [TCP_RST_WAIT]	= TCP_CLOSE,
 };
 
 int tcp_close_state(struct sock *sk)

net/ipv4/tcp_timer.c

@@ -681,7 +681,7 @@ static void tcp_keepalive_timer (unsigned long data)
 
 	tcp_mstamp_refresh(tp);
 
-	if (tp->send_mp_fclose) {
+	if (tp->send_mp_fclose && sk->sk_state == TCP_RST_WAIT) {
 		if (icsk->icsk_retransmits >= MPTCP_FASTCLOSE_RETRIES) {
 			tcp_write_err(sk);
 			goto out;