Re: [PATCH 4/5] AF_UNIX: find peers on multicast Unix stream sockets

[Alban Crequy]

Le Fri, 24 Sep 2010 20:00:37 +0200,
Eric Dumazet <eric.dumazet@xxxxxxxxx> a Ãcrit :

> Le vendredi 24 septembre 2010 Ã 18:25 +0100, Alban Crequy a Ãcrit :
> 
> > @@ -1612,7 +1671,12 @@ static int unix_stream_sendmsg(struct kiocb
> > *kiocb, struct socket *sock, } else {
> >  		sunaddr = NULL;
> >  		err = -ENOTCONN;
> > -		other = NULL; /* FIXME: get the list of other
> > connection */
> > +		max_others = atomic_read(&unix_nr_multicast_socks);
> > +		others = kzalloc((max_others + 1) * sizeof(void
> > *), GFP_KERNEL);
> > +		unix_find_other(sock_net(sk), u->addr->name,
> > +		    u->addr->len, 0, u->addr->hash, 1, others,
> > max_others, &err);
> > +		other = others[0];
> > +		kfree(others);
> >  		if (!other)
> >  			goto out_err;
> >  	}
> 
> Seriously, this block sizing against unix_nr_multicast_socks is not
> scalable. What happens if we have 1000 sockets ?
> kzalloc() to clear 8000 bytes ?
> Its also unsafe.
> 
> (say you kzalloc() a buffer for 2 sockets, and another cpu inserts a
> new socket. unix_find_socket_byname() can overflow the buffer)
> 
> 
> You should use a list, and allocates elements in
> unix_find_socket_byname()
> 
> struct item {
> 	struct item *next;
> 	struct sock *s;
> };

Thanks for your review.

I cannot allocate elements directly in unix_find_socket_byname()
iteration after iteration because the spinlock "unix_table_lock" is
held. If I release the spinlock to allocate, the number of sockets in
the table may change.

I changed the code to count the sockets with the lock held and then
allocate. In the unfortunate case where the allocation is not big
enough (if another process inserts a new socket), it just tries again.

The code is available here. Please pull from:

git://git.collabora.co.uk/git/user/alban/linux-2.6.35.y/.git unix-multicast2

It is still a work in progress. Missing pieces:

- The flow control does not work correctly: poll/select does not match
  the reality

- Atomic delivery: if a process is killed or interrupted in the middle
  of a delivery, only a subset of the recipients will get the message 

- Some locking to provide the same delivery order to all the recipients
  when several senders run concurrently.

Feedback welcome,

Alban Crequy

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/