mptcp: fix race in release_cb

If we receive a MPTCP_PUSH_PENDING even from a subflow when
mptcp_release_cb() is serving the previous one, the latter
will be delayed up to the next release_sock(msk).

Address the issue implementing a test/serve loop for such
event.

Additionally rename the push helper to __mptcp_push_pending()
to be more consistent with the existing code.

Fixes: 6e628cd ("mptcp: use mptcp release_cb for delayed tasks")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/mptcp/protocol.c

@@ -1445,7 +1445,7 @@ static void mptcp_push_release(struct sock *sk, struct sock *ssk,
 	release_sock(ssk);
 }
 
-static void mptcp_push_pending(struct sock *sk, unsigned int flags)
+static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
 {
 	struct sock *prev_ssk = NULL, *ssk = NULL;
 	struct mptcp_sock *msk = mptcp_sk(sk);
@@ -1697,14 +1697,14 @@ static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 
 wait_for_memory:
 		mptcp_set_nospace(sk);
-		mptcp_push_pending(sk, msg->msg_flags);
+		__mptcp_push_pending(sk, msg->msg_flags);
 		ret = sk_stream_wait_memory(sk, &timeo);
 		if (ret)
 			goto out;
 	}
 
 	if (copied)
-		mptcp_push_pending(sk, msg->msg_flags);
+		__mptcp_push_pending(sk, msg->msg_flags);
 
 out:
 	release_sock(sk);
@@ -2959,13 +2959,14 @@ static void mptcp_release_cb(struct sock *sk)
 {
 	unsigned long flags, nflags;
 
-	/* push_pending may touch wmem_reserved, do it before the later
-	 * cleanup
-	 */
-	if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
-		__mptcp_clean_una(sk);
-	if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags)) {
-		/* mptcp_push_pending() acquires the subflow socket lock
+	for (;;) {
+		flags = 0;
+		if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
+			flags |= MPTCP_PUSH_PENDING;
+		if (!flags)
+			break;
+
+		/* the following actions acquire the subflow socket lock
 		 *
 		 * 1) can't be invoked in atomic scope
 		 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
@@ -2974,13 +2975,21 @@ static void mptcp_release_cb(struct sock *sk)
 		 */
 
 		spin_unlock_bh(&sk->sk_lock.slock);
-		mptcp_push_pending(sk, 0);
+		if (flags & MPTCP_PUSH_PENDING)
+			__mptcp_push_pending(sk, 0);
+
+		cond_resched();
 		spin_lock_bh(&sk->sk_lock.slock);
 	}
+
+	if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
+		__mptcp_clean_una(sk);
 	if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
 		__mptcp_error_report(sk);
 
-	/* clear any wmem reservation and errors */
+	/* push_pending may touch wmem_reserved, ensure we do the cleanup
+	 * later
+	 */
 	__mptcp_update_wmem(sk);
 	__mptcp_update_rmem(sk);