Re: [PATCH 3/3] net: Convert TCP & DCCP hash tables to use RCU / hlist_nulls

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Thu, 2008-11-13 at 14:51 +0100, Eric Dumazet wrote:
> Peter Zijlstra a écrit :
> > On Thu, 2008-11-13 at 14:15 +0100, Eric Dumazet wrote:
> >> +begin:
> >> +       sk_nulls_for_each_rcu(sk, node, &head->chain) {
> >>                 if (INET_MATCH(sk, net, hash, acookie,
> >> +                                       saddr, daddr, ports, dif)) {
> >> +                       if (unlikely(!atomic_inc_not_zero(&sk->sk_refcnt)))
> >> +                               goto begintw;
> >> +                       if (unlikely(!INET_MATCH(sk, net, hash, acookie,
> >> +                               saddr, daddr, ports, dif))) {
> >> +                               sock_put(sk);
> >> +                               goto begin;
> >> +                       }
> > 
> > This is the validation step that verifies the race opened by using
> > SLAB_DESTROY_BY_RCU, right?
> 
> The atomic_inc_not_zero() is not related to SLAB_DESTROY_BY_RCU but
> classic RCU lookup. A writer can delete the item right before we try to use it.

Yeah, its to stabilize the current situation.

> Next step is necessary in case the deleted item was re-allocated and inserted
> in a hash chain (this one or another one, it doesnt matter). In this case,
> previous atomic_inc_not_zero test will succeed. So we must check again the item
> we selected (and refcounted) is the one we were searching.
> 
> So yes, this bit should be documented, since SLAB_DESTROY_BY_RCU is
> not really used in linux kernel at this moment.

We have one user, anon_vma, and one thing that is very nearly identical
the lockless pagecache.

See page_cache_get_speculative() and find_get_page(). The pagecache gets
away with this due to the simple fact that the page frames are never
freed.

Hmm, I once wrote a comment to go with SLAB_DESTROY_BY_RCU, which seems
to have gotten lost... /me goes dig.

Found it:  http://lkml.org/lkml/2008/4/2/143

I guess I'd better re-submit that..