Re: [PATCH 0/3] net: time stamping fixes

[Johannes Berg <johannes@sipsolutions.net>]

On Wed, 2011-10-19 at 15:25 +0200, Eric Dumazet wrote:

> > Given the complexity of all this, I'm not sure we shouldn't do something
> > like this, but I have no idea what the cost would be:
> > 
> > --- wireless-testing.orig/include/net/sock.h	2011-10-18 22:28:41.000000000 +0200
> > +++ wireless-testing/include/net/sock.h	2011-10-19 15:08:45.000000000 +0200
> > @@ -434,7 +434,10 @@ static __inline__ int __sk_del_node_init
> >  
> >  static inline void sock_hold(struct sock *sk)
> >  {
> > -	atomic_inc(&sk->sk_refcnt);
> > +	if (atomic_inc_return(&sk->sk_refcnt) == 1) {
> > +		/* was zero -- we must've gotten an sk_wmem_alloc reference */
> > +		atomic_inc(&sk->sk_wmem_alloc);
> > +	}
> >  }
> >  
> 
> Hmm, it will be difficult to handle two atomics without adding races,

Where do you see a race? If you do sock_hold() while you have a 'proper
sk_refcnt reference', this does nothing but increase sk_refcnt. If you
do sock_hold() while you have an 'sk_wmem_alloc reference', this will
actually increase sk_wmem_alloc by 1; then when the original
'sk_wmem_alloc reference' you had when calling sock_hold() is released,
sk_wmem_alloc will still have 1 matching a non-zero sk_refcnt.


> and add quite expensive atomic_inc_return() on some arches.

Yes I was afraid of that.

> I would just change the skb tx cloning to take a normal reference on
> sk_wmem_alloc
> 
> 	atomic_add(skb->truesize, &sk->sk_wmem_alloc);
> instead of
> 	sock_hold(sk);

Even with that fixed I'm not really convinced of it all -- need to
really really really make sure that no skb->sk that was owned by a TX
skb is ever passed to sock_hold(). Can we really guarantee that?

johannes