Christoph reported a possible deadlock while the TCP stack
destroys an unaccepted subflow due to an incoming reset: the
MPTCP socket error path tries to acquire the msk-level socket
lock while TCP still owns the listener socket accept queue
spinlock, and the reverse dependency already exists in the
TCP stack.
Note that the above is actually a lockdep false positive, as
the chain involves two separate sockets. A different per-socket
lockdep key will address the issue, but such a change will be
quite invasive.
Instead, we can simply stop earlier the socket error handling
for orphaned or unaccepted subflows, breaking the critical
lockdep chain. Error handling in such a scenario is a no-op.
Reported-and-tested-by: Christoph Paasch <cpaasch@apple.com>
Fixes: 15cc10453398 ("mptcp: deliver ssk errors to msk")
Closes: https://github.com/multipath-tcp/mptcp_net-next/issues/355
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
Note that the fixes tag above is different from the bisected, much
more recent, commit pointed by syzkaller, but I believe (code
inspection shows) that the DL scenario is present since it.
---
net/mptcp/subflow.c | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
index 8b46311b8d5e..a14c3a0c7c00 100644
--- a/net/mptcp/subflow.c
+++ b/net/mptcp/subflow.c
@@ -1437,6 +1437,13 @@ static void subflow_error_report(struct sock *ssk)
{
struct sock *sk = mptcp_subflow_ctx(ssk)->conn;
+ /* bail early if this is a no-op, so that we avoid introducing a
+ * problematic lockdep dependency between TCP accept queue lock
+ * and msk socket spinlock
+ */
+ if (!sk->sk_socket)
+ return;
+
mptcp_data_lock(sk);
if (!sock_owned_by_user(sk))
__mptcp_error_report(sk);
--
2.39.2