refcounting on dgram Unix sockets for poll(POLLOUT)

refcounting on dgram Unix sockets for poll(POLLOUT)

[Alban Crequy]

Hi,

When a connected datagram Unix socket is polled for POLLOUT events, the
poller is added in the wait_queue_head_t of the "server" socket:

net/unix/af_unix.c::unix_dgram_poll()
  other = unix_peer_get(sk);
  if (other) {
    if (unix_peer(other) != sk) {
      sock_poll_wait(file, &unix_sk(other)->peer_wait,
                     wait);
      ...

I wonder what prevent the "server" socket ("other") to be released
while the poller is still waiting for POLLOUT events.

There is a reference taken on the "server" socket when the client
connects:
net/unix/af_unix.c::unix_dgram_connect()
  other = unix_find_other(net, sunaddr, alen, sock->type,
            hash, &err);

But that reference could be released when the client socket
disconnects from another thread in one of the 3 possible locations:

1. unix_dgram_connect() when connecting to a different socket
     if (other != old_peer)
       unix_dgram_disconnected(sk, old_peer);
     sock_put(old_peer);

2. unix_dgram_sendmsg() when the server socket is SOCK_DEAD:
     unix_dgram_disconnected(sk, other);
     sock_put(other);

3. unix_release_sock() when the client socket is released:
     skpair = unix_peer(sk);
     if (skpair != NULL) {
       sock_put(skpair); /* It may now die */

I tried to release all the references to server_sockfd with
close(server_sockfd) on the server thread and with
connect(client_sockfd) to a different socket while client_sockfd is
being polled for POLLOUT events in a different thread, hoping to crash
the poller with the stack:
free_poll_entry()->remove_wait_queue()->spin_lock_irqsave()
But I didn't manage to crash the kernel.

Am I missing something? Is there another reference taken on
server_sockfd to protect the kernel from this scenario?

And btw, what is the test (unix_peer(other) != sk) in unix_dgram_poll()?

Alban

Re: refcounting on dgram Unix sockets for poll(POLLOUT)

[Eric Dumazet]

Le mercredi 08 décembre 2010 à 21:01 +0000, Alban Crequy a écrit :
> Hi,
> 
> When a connected datagram Unix socket is polled for POLLOUT events, the
> poller is added in the wait_queue_head_t of the "server" socket:
> 
> net/unix/af_unix.c::unix_dgram_poll()
>   other = unix_peer_get(sk);
>   if (other) {
>     if (unix_peer(other) != sk) {
>       sock_poll_wait(file, &unix_sk(other)->peer_wait,
>                      wait);
>       ...
> 
> I wonder what prevent the "server" socket ("other") to be released
> while the poller is still waiting for POLLOUT events.
> 
> There is a reference taken on the "server" socket when the client
> connects:
> net/unix/af_unix.c::unix_dgram_connect()
>   other = unix_find_other(net, sunaddr, alen, sock->type,
>             hash, &err);
> 
> But that reference could be released when the client socket
> disconnects from another thread in one of the 3 possible locations:
> 
> 1. unix_dgram_connect() when connecting to a different socket
>      if (other != old_peer)
>        unix_dgram_disconnected(sk, old_peer);
>      sock_put(old_peer);
> 
> 2. unix_dgram_sendmsg() when the server socket is SOCK_DEAD:
>      unix_dgram_disconnected(sk, other);
>      sock_put(other);
> 
> 3. unix_release_sock() when the client socket is released:
>      skpair = unix_peer(sk);
>      if (skpair != NULL) {
>        sock_put(skpair); /* It may now die */
> 
> I tried to release all the references to server_sockfd with
> close(server_sockfd) on the server thread and with
> connect(client_sockfd) to a different socket while client_sockfd is
> being polled for POLLOUT events in a different thread, hoping to crash
> the poller with the stack:
> free_poll_entry()->remove_wait_queue()->spin_lock_irqsave()
> But I didn't manage to crash the kernel.
> 

Maybe you need to add some options to detect/trigger this ?

CONFIG_DEBUG_LOCK_ALLOC

and

slub_debug=FZP
cf Documentation/vm/slub.txt

to force slub to overwrite data when freeing master socket.

> Am I missing something? Is there another reference taken on
> server_sockfd to protect the kernel from this scenario?
> 
> And btw, what is the test (unix_peer(other) != sk) in unix_dgram_poll()?
> 
> Alban