route: set the deleted fnhe fnhe_daddr to 0 in ip_del_fnhe to fix a race

The race occurs in __mkroute_output() when 2 threads lookup a dst:

  CPU A                 CPU B
  find_exception()
                        find_exception() [fnhe expires]
                        ip_del_fnhe() [fnhe is deleted]
  rt_bind_exception()

In rt_bind_exception() it will bind a deleted fnhe with the new dst, and
this dst will get no chance to be freed. It causes a dev defcnt leak and
consecutive dmesg warnings:

  unregister_netdevice: waiting for ethX to become free. Usage count = 1

Especially thanks Jon to identify the issue.

This patch fixes it by setting fnhe_daddr to 0 in ip_del_fnhe() to stop
binding the deleted fnhe with a new dst when checking fnhe's fnhe_daddr
and daddr in rt_bind_exception().

It works as both ip_del_fnhe() and rt_bind_exception() are protected by
fnhe_lock and the fhne is freed by kfree_rcu().

Fixes: deed49d ("route: check and remove route cache when we get route")
Signed-off-by: Jon Maxwell <[email protected]>
Signed-off-by: Xin Long <[email protected]>
Reviewed-by: David Ahern <[email protected]>
Signed-off-by: David S. Miller <[email protected]>

net/ipv4/route.c

@@ -1303,6 +1303,10 @@ static void ip_del_fnhe(struct fib_nh *nh, __be32 daddr)
 		if (fnhe->fnhe_daddr == daddr) {
 			rcu_assign_pointer(*fnhe_p, rcu_dereference_protected(
 				fnhe->fnhe_next, lockdep_is_held(&fnhe_lock)));
+			/* set fnhe_daddr to 0 to ensure it won't bind with
+			 * new dsts in rt_bind_exception().
+			 */
+			fnhe->fnhe_daddr = 0;
 			fnhe_flush_routes(fnhe);
 			kfree_rcu(fnhe, rcu);
 			break;