mptcp: more conservative check for zero probes
authorPaolo Abeni <pabeni@redhat.com>
Wed, 18 Oct 2023 18:23:54 +0000 (11:23 -0700)
committerJakub Kicinski <kuba@kernel.org>
Thu, 19 Oct 2023 16:10:00 +0000 (09:10 -0700)
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 <cpaasch@apple.com>
Closes: https://github.com/multipath-tcp/mptcp_net-next/issues/444
Fixes: f70cad1085d1 ("mptcp: stop relying on tcp_tx_skb_cache")
Cc: stable@vger.kernel.org
Reviewed-by: Mat Martineau <martineau@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <martineau@kernel.org>
Link: https://lore.kernel.org/r/20231018-send-net-20231018-v1-3-17ecb002e41d@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net/mptcp/protocol.c

index d190237..4e30e5b 100644 (file)
@@ -1298,7 +1298,7 @@ alloc_skb:
        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;
                }
@@ -1306,11 +1306,6 @@ alloc_skb:
                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);
@@ -1339,7 +1334,6 @@ alloc_skb:
        if (reuse_skb) {
                TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
                mpext->data_len += copy;
-               WARN_ON_ONCE(zero_window_probe);
                goto out;
        }