On Thu, Nov 17, 2016 at 11:29:59AM +0100, Peter Zijlstra wrote:
> On Thu, Nov 17, 2016 at 05:48:51PM +0800, Boqun Feng wrote:
> > > > But as I said, we actually only need the pairing of orderings:
> > > >
> > > > 1) load part of cmpxchg -> free()
> > > > 2) object accesses -> store part of cmpxchg
> > > >
> > > > Ordering #1 can be achieved via control dependency as you pointed out
> > > > that free()s very much includes stores. And ordering #2 can be achieved
> > > > with RELEASE.
> > > >
> > > > So the code is right, I just thought the comment may be misleading. The
> > > > reason we use cmpxchg_release() is just for achieving ordering #2, and
> > > > not to order "prior loads and stores" with "a subsequent free".
> > > >
> > > > Am I missing some subtle orderings here?
> > >
> > > I would want to further quality 1), it must be no earlier than the load
> > > of the last / successful ll/sc round.
> > >
> >
> > Great, that's more accurate!
> >
> > > At that point we're guaranteed a reference count of 1 that _will_ drop
> > > to 0, and thus nobody else (should) reference that memory anymore.
> > >
> > > If we agree on this, I'll update the comment :-) Will, do you too agree?
> >
> > Agreed ;-)
> >
> > Control dependencies and RELEASE are totally enough for the internal
> > correctness of refcount_t along with its interactivity with free().
> > People better not reply order guarantees other than this ;-)
>
> Hurm.. let me ruin my own argument.
>
> Since the free() stores could leak upwards until that ll, and object
> stores can be delayed until the sc, we still have a problem. Just not
> with the thread that free()s or any other thread that knew about the
> object.
>
> The problem comes from any other thread doing an allocation, since its
> possible to observe the memory as freed while there are stores pending
> to it, we can have those delayed stores trample on our freshly allocated
> and initialized object.
>
> The stores must really not be before the SC, so I fear we must either
> add an smp_wmb() after the release, or punt and use the fully ordered
> cmpxchg().
And let me note here that RCU users can use a fully relaxed put, because
call_rcu() guarantees a grace-period between the call_rcu and the
free(), which in turn provides a full memory barrier that orders things.
We could actually expose that to driver writers by doing something like:
kref_put_rcu(struct kref *kref, struct rcu_head *head, rcu_callback_t
func)
{
if (refcount_dec_and_test_relaxed(&kref->refcount))
call_rcu(head, func);
}
Do we want to go there?
