quoted text > On Wed, Apr 02 2008, Peter Zijlstra wrote:
> > On Wed, 2008-04-02 at 13:42 +0200, Jens Axboe wrote:
> > > On Wed, Apr 02 2008, Peter Zijlstra wrote:
> > > > On Wed, 2008-04-02 at 13:32 +0200, Jens Axboe wrote:
> > > > > On Wed, Apr 02 2008, Peter Zijlstra wrote:
> > > > > > On Wed, 2008-04-02 at 13:14 +0200, Jens Axboe wrote:
> > > > > > > On Wed, Apr 02 2008, Peter Zijlstra wrote:
> > > > > > > > On Wed, 2008-04-02 at 13:07 +0200, Jens Axboe wrote:
> > > > > > > > > On Wed, Apr 02 2008, Pekka Enberg wrote:
> > > > > > > > > > Hi Paul,
> > > > > > > > > >
> > > > > > > > > > On Wed, Apr 2, 2008 at 1:55 PM, Paul E. McKenney
> > > > > > > > > > <paulmck@linux.vnet.ibm.com> wrote:
> > > > > > > > > > > I will check this when I get back to some bandwidth -- but in the meantime,
> > > > > > > > > > > does kmemcheck special-case SLAB_DESTROY_BY_RCU? It is legal to access
> > > > > > > > > > > newly-freed items in that case, as long as you did rcu_read_lock()
> > > > > > > > > > > before gaining a reference to them and don't hold the reference past
> > > > > > > > > > > the matching rcu_read_unlock().
> > > > > > > > > >
> > > > > > > > > > No, kmemcheck is work in progress and does not know about
> > > > > > > > > > SLAB_DESTROY_BY_RCU yet. The reason I asked Vegard to post the warning
> > > > > > > > > > was because Peter, Vegard, and myself identified this particular
> > > > > > > > > > warning as a real problem. But yeah, kmemcheck can cause false
> > > > > > > > > > positives for RCU for now.
> > > > > > > > >
> > > > > > > > > Makes sense, and to me Pauls analysis of the code looks totally correct
> > > > > > > > > - there's no bug there, at least related to hlist traversal and
> > > > > > > > > kmem_cache_free(), since we are under rcu_read_lock() and thus hold off
> > > > > > > > > the grace for freeing.
> > > > > > > >
> > > > > > > > but what holds off the slab allocator re-issueing that same object and
> > > > > > > > someone else writing other stuff into it?
> > > > > > >
> > > > > > > Nothing, that's how rcu destry works here. But for the validation to be
> > > > > > > WRONG radix_tree_lookup(..., old_key) must return cic for new_key, not
> > > > > > > NULL.
> > > > > > >
> > > > > >
> > > > > >
> > > > > > A B C
> > > > > >
> > > > > > cfq_cic_lookup(cfqd_1, ioc)
> > > > > >
> > > > > > rcu_read_lock()
> > > > > > cic = radix_tree_lookup(, cfqd_q);
> > > > > >
> > > > > > cfq_cic_free()
> > > > > >
> > > > > > cfq_cic_link(cfqd_2, ioc,)
> > > > > >
> > > > > > rcu_read_unlock()
> > > > > >
> > > > > >
> > > > > > and now we have that:
> > > > > >
> > > > > > cic->key == cfqd_2
> > > > > >
> > > > > >
> > > > > > I'm not seeing anything stopping this from happening.
> > > > >
> > > > > I don't follow your A-B-C here, what do they refer to?
> > > >
> > > > A does a radix_tree_lookup() of cfqd_1 (darn typos)
> > > > B does a kfree of the same cic found by A
> > > > C does an alloc and gets the same cic as freed by B and inserts it
> > > > in a different location.
> > > >
> > > > So that when we return to A, cic->key == cfqd_2 even though we did a
> > > > lookup for cfqd_1.
> > >
> > > That I follow, my question was if A, B, and C refer to different
> > > processes but with a shared io context? I'm assuming that is correct...
> >
> > Ah, yeah, whatever is needed to make this race happen :-)
>
> The only place where you'll have multiple processes involved with this
> at all is if they share io contexts. That is also why the bug isn't that
> critical, since it's not possible right now (CLONE_IO flag must be
> used).
There are 3 races here:
1) A continues with another object than intended
(requires CLONE_IO)
2) A does hlist_for_each_rcu() and races with B,C so that
we continue the iteration on a possibly unrelated list.
3) cic is freed after the !cic->key check.
I'm not familiar enough with the code yet to see if 3 really is an
possibility. But from what I can see there is nothing guarding its
existence.
--
