net: rds: don't hold sock lock when cancelling work from rds_tcp_reset_callbacks()

syzbot is reporting lockdep warning at rds_tcp_reset_callbacks() [1], for
commit ac3615e ("RDS: TCP: Reduce code duplication in
rds_tcp_reset_callbacks()") added cancel_delayed_work_sync() into a section
protected by lock_sock() without realizing that rds_send_xmit() might call
lock_sock().

We don't need to protect cancel_delayed_work_sync() using lock_sock(), for
even if rds_{send,recv}_worker() re-queued this work while __flush_work()
 from cancel_delayed_work_sync() was waiting for this work to complete,
retried rds_{send,recv}_worker() is no-op due to the absence of RDS_CONN_UP
bit.

Link: https://syzkaller.appspot.com/bug?extid=78c55c7bc6f66e53dce2 [1]
Reported-by: syzbot <[email protected]>
Co-developed-by: Hillf Danton <[email protected]>
Signed-off-by: Hillf Danton <[email protected]>
Signed-off-by: Tetsuo Handa <[email protected]>
Tested-by: syzbot <[email protected]>
Fixes: ac3615e ("RDS: TCP: Reduce code duplication in rds_tcp_reset_callbacks()")
Signed-off-by: David S. Miller <[email protected]>

Author: Tetsuo Handa

net/rds/tcp.c

@@ -166,10 +166,10 @@ void rds_tcp_reset_callbacks(struct socket *sock,
 	 */
 	atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
 	wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
-	lock_sock(osock->sk);
 	/* reset receive side state for rds_tcp_data_recv() for osock  */
 	cancel_delayed_work_sync(&cp->cp_send_w);
 	cancel_delayed_work_sync(&cp->cp_recv_w);
+	lock_sock(osock->sk);
 	if (tc->t_tinc) {
 		rds_inc_put(&tc->t_tinc->ti_inc);
 		tc->t_tinc = NULL;