There is a race between netlink_dump_start() and netlink_release()
that can lead to the situation when a netlink socket with non-zero
callback is freed.
Here it is:
CPU1:                           CPU2
netlink_release():              netlink_dump_start():
                                sk = netlink_lookup(); /* OK */
netlink_remove();
spin_lock(&nlk->cb_lock);
if (nlk->cb) { /* false */
  ...
}
spin_unlock(&nlk->cb_lock);
                                spin_lock(&nlk->cb_lock);
                                if (nlk->cb) { /* false */
                                         ...
                                }
                                nlk->cb = cb;
                                spin_unlock(&nlk->cb_lock);
                                ...
sock_orphan(sk);
/*
 * proceed with releasing
 * the socket
 */
The proposal it to make sock_orphan before detaching the callback
in netlink_release() and to check for the sock to be SOCK_DEAD in
netlink_dump_start() before setting a new callback.
Signed-off-by: Denis Lunev <den@openvz.org>
Signed-off-by: Kirill Korotaev <dev@openvz.org>
Signed-off-by: Pavel Emelianov <xemul@openvz.org>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
                return 0;
 
        netlink_remove(sk);
+       sock_orphan(sk);
        nlk = nlk_sk(sk);
 
        spin_lock(&nlk->cb_lock);
        /* OK. Socket is unlinked, and, therefore,
           no new packets will arrive */
 
-       sock_orphan(sk);
        sock->sk = NULL;
        wake_up_interruptible_all(&nlk->wait);
 
                return -ECONNREFUSED;
        }
        nlk = nlk_sk(sk);
-       /* A dump is in progress... */
+       /* A dump or destruction is in progress... */
        spin_lock(&nlk->cb_lock);
-       if (nlk->cb) {
+       if (nlk->cb || sock_flag(sk, SOCK_DEAD)) {
                spin_unlock(&nlk->cb_lock);
                netlink_destroy_callback(cb);
                sock_put(sk);