mptcp: more conservative check for zero probes

commit 72377ab upstream.

Christoph reported that the MPTCP protocol can find the subflow-level
write queue unexpectedly not empty while crafting a zero-window probe,
hitting a warning:

------------[ cut here ]------------
WARNING: CPU: 0 PID: 188 at net/mptcp/protocol.c:1312 mptcp_sendmsg_frag+0xc06/0xe70
Modules linked in:
CPU: 0 PID: 188 Comm: kworker/0:2 Not tainted 6.6.0-rc2-g1176aa719d7a #47
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-2.el7 04/01/2014
Workqueue: events mptcp_worker
RIP: 0010:mptcp_sendmsg_frag+0xc06/0xe70 net/mptcp/protocol.c:1312
RAX: 47d0530de347ff6a RBX: 47d0530de347ff6b RCX: ffff8881015d3c00
RDX: ffff8881015d3c00 RSI: 47d0530de347ff6b RDI: 47d0530de347ff6b
RBP: 47d0530de347ff6b R08: ffffffff8243c6a8 R09: ffffffff82042d9c
R10: 0000000000000002 R11: ffffffff82056850 R12: ffff88812a13d580
R13: 0000000000000001 R14: ffff88812b375e50 R15: ffff88812bbf3200
FS:  0000000000000000(0000) GS:ffff88813bc00000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000695118 CR3: 0000000115dfc001 CR4: 0000000000170ef0
Call Trace:
 <TASK>
 __subflow_push_pending+0xa4/0x420 net/mptcp/protocol.c:1545
 __mptcp_push_pending+0x128/0x3b0 net/mptcp/protocol.c:1614
 mptcp_release_cb+0x218/0x5b0 net/mptcp/protocol.c:3391
 release_sock+0xf6/0x100 net/core/sock.c:3521
 mptcp_worker+0x6e8/0x8f0 net/mptcp/protocol.c:2746
 process_scheduled_works+0x341/0x690 kernel/workqueue.c:2630
 worker_thread+0x3a7/0x610 kernel/workqueue.c:2784
 kthread+0x143/0x180 kernel/kthread.c:388
 ret_from_fork+0x4d/0x60 arch/x86/kernel/process.c:147
 ret_from_fork_asm+0x1b/0x30 arch/x86/entry/entry_64.S:304
 </TASK>

The root cause of the issue is that expectations are wrong: e.g. due
to MPTCP-level re-injection we can hit the critical condition.

Explicitly avoid the zero-window probe when the subflow write queue
is not empty and drop the related warnings.

Reported-by: Christoph Paasch <[email protected]>
Closes: multipath-tcp/mptcp_net-next#444
Fixes: f70cad1 ("mptcp: stop relying on tcp_tx_skb_cache")
Cc: [email protected]
Reviewed-by: Mat Martineau <[email protected]>
Signed-off-by: Paolo Abeni <[email protected]>
Signed-off-by: Mat Martineau <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>

net/mptcp/protocol.c

@@ -1342,19 +1342,14 @@ static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
 	if (copy == 0) {
 		u64 snd_una = READ_ONCE(msk->snd_una);
 
-		if (snd_una != msk->snd_nxt) {
+		if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
 			tcp_remove_empty_skb(ssk);
 			return 0;
 		}
 
 		zero_window_probe = true;
 		data_seq = snd_una - 1;
 		copy = 1;
-
-		/* all mptcp-level data is acked, no skbs should be present into the
-		 * ssk write queue
-		 */
-		WARN_ON_ONCE(reuse_skb);
 	}
 
 	copy = min_t(size_t, copy, info->limit - info->sent);
@@ -1383,7 +1378,6 @@ static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
 	if (reuse_skb) {
 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
 		mpext->data_len += copy;
-		WARN_ON_ONCE(zero_window_probe);
 		goto out;
 	}