Since the blamed commit, closing the first subflow resets the first
subflow socket state to SS_UNCONNECTED.
The current mptcp listen implementation relies only on such
state to prevent touching not-fully-disconnected sockets.
Incoming mptcp fastclose (or paired endpoint removal) unconditionally
closes the first subflow.
All the above allows an incoming fastclose followed by a listen() call
to successfully race with a blocking recvmsg(), potentially causing the
latter to hit a divide by zero bug in cleanup_rbuf/__tcp_select_window().
Address the issue explicitly checking the msk socket state in
mptcp_listen(). An alternative solution would be moving the first
subflow socket state update into mptcp_disconnect(), but in the long
term the first subflow socket should be removed: better avoid relaying
on it for internal consistency check.
Fixes: b29fcfb54cd7 ("mptcp: full disconnect implementation")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
This is basically a new version of "mptcp: fix bogus socket state update",
but since it uses a quite different logic, no revision increase ;)
Should close issues/414
---
net/mptcp/protocol.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/net/mptcp/protocol.c b/net/mptcp/protocol.c
index a50eaa01ba8a..d3ff2a2c60de 100644
--- a/net/mptcp/protocol.c
+++ b/net/mptcp/protocol.c
@@ -3751,6 +3751,11 @@ static int mptcp_listen(struct socket *sock, int backlog)
pr_debug("msk=%p", msk);
lock_sock(sk);
+
+ err = -EINVAL;
+ if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
+ goto unlock;
+
ssock = __mptcp_nmpc_socket(msk);
if (IS_ERR(ssock)) {
err = PTR_ERR(ssock);
--
2.40.1