lkml.org 
[lkml]   [2008]   [May]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: 2.6.26-rc4: RIP __call_for_each_cic+0x20/0x50
On Wed, May 28, 2008 at 12:07:21PM +0200, Jens Axboe wrote:
> On Tue, May 27 2008, Paul E. McKenney wrote:
> > On Tue, May 27, 2008 at 03:35:10PM +0200, Jens Axboe wrote:
> > > On Tue, May 27 2008, Alexey Dobriyan wrote:
> > > > On Sat, May 10, 2008 at 02:37:19PM +0400, Alexey Dobriyan wrote:
> > > > > > > > > @@ -41,8 +41,8 @@ int put_io_context(struct io_context *ioc)
> > > > > > > > > rcu_read_lock();
> > > > > > > > > if (ioc->aic && ioc->aic->dtor)
> > > > > > > > > ioc->aic->dtor(ioc->aic);
> > > > > > > > > - rcu_read_unlock();
> > > > > > > > > cfq_dtor(ioc);
> > > > > > > > > + rcu_read_unlock();
> > > > > > > > >
> > > > > > > > > kmem_cache_free(iocontext_cachep, ioc);
> > > > > > > > > return 1;
> > > > > > > >
> > > > > > > > This helps in sense that 3 times bulk cross-compiles finish to the end.
> > > > > > > > You'll hear me if another such oops will resurface.
> > > > > > >
> > > > > > > Still looking good?
> > > > > >
> > > > > > Yup!
> > > > >
> > > > > And this with patch in mainline, again with PREEMPT_RCU.
> > > >
> > > > Ping, this happened again with 2.6.26-rc4 and PREEMPT_RCU.
> > >
> > > Worrisome... Paul, would you mind taking a quick look at cfq
> > > and see if you can detect why breaks with preempt rcu? It's
> > > clearly a use-after-free symptom, but I don't see how it can
> > > happen.
> >
> > Some quick and probably off-the-mark questions...
>
> Thanks!

Glad it actually was of help! ;-)

> > o What is the purpose of __call_for_each_cic()? When called
> > from call_for_each_cic(), it is under rcu_read_lock(), as
> > required, but it is also called from cfq_free_io_context(),
> > which is assigned to the ->dtor and ->exit members of the
> > cfq_io_context struct. What protects calls through these
> > members?
> >
> > (This is for the ->cic_list field of the cfq_io_context structure.
> > One possibility is that the io_context's ->lock member is held,
> > but I don't see this. Not that I looked all that hard...)
> >
> > My suggestion would be to simply change all invocations of
> > __call_for_each_cic() to instead invoke call_for_each_cic().
> > The rcu_read_lock()/rcu_read_unlock() pair is pretty
> > lightweight, even in CONFIG_PREEMPT_RCU.
>
> __call_for_each_cic() is always called under rcu_read_lock(), it merely
> exists to avoid a double rcu_read_lock(). Even if it is cheap. The
> convention follows the usual __lock_is_already_held() double under
> score, but I guess it could do with a comment! There are only two
> callers of the function, call_for_each_cic() which does the
> rcu_read_lock(), and cfq_free_io_context() which is called from ->dtor
> (and holds the rcu_read_lock() and ->trim which actually does not. That
> looks like it could be problematic, but it's only called when an io
> scheduler module is removed so not really critical. I'll add it, though!
> Actually, the task_lock() should be enough there. So no bug, but (again)
> it could do with a comment.

Sounds good!

> > o When calling cfq_slab_kill(), for example from cfq_exit(),
> > what ensures that all previous RCU callbacks have completed?
> >
> > I suspect that you need an rcu_barrier() at the beginning
> > of cfq_slab_kill(), but I could be missing something.
>
> So we have two callers of that, one is from the error path at init time
> and is obviously ok. The other does need rcu_barrier()! I'll add that.

OK, that does make my brain hurt less. ;-)

> > o Updates to the ->ioc_data field of the cfq_io_context
> > seem to be protected by the request_queue ->queue_lock
> > field. This seems very strange to me. It is OK if every
> > cfq_io_context is associated with only one request_queue
> > structure -- is this the case?
>
> ->ioc_data is part of the io_context, not cfq_io_context. And it can be
> shared now, so the correct locking for that would be ioc->lock and not
> the queue lock. __cfq_exit_single_io_context() is serialized in the
> sense that only one process gets to call the exit path.

Makes sense to me!

> > o What protects the first rcu_dereference() in cfq_cic_lookup()?
> > There needs to be either an enclose rcu_read_lock() on the
> > one hand or the ->queue_lock needs to be held.
> >
> > (My guess is the latter, given the later rcu_assign_pointer()
> > in this same function, but I don't see a lock acquisition
> > in the immediate vicinity -- might be further up the function
> > call stack, though.)
>
> There's no locking going into that function when coming from
> cfq_get_io_context(), the other paths (happen) to hold the queue lock
> already though.

So the call from cfq_get_io_context() needs an rcu_read_lock()?
Not seeing this in the patch below, but maybe you have it up a
function-call level or two?

> > o Why is there no grace period associated with the ioc_data?
> > For example, what happens to the old value of ->ioc_data
> > after the rcu_assign_pointer() in cfq_cic_lookup()? Readers
> > might still be referencing the old version, right? If so,
> > how do we avoid messing them up?
> >
> > Or are we somehow leveraging the call_rcu() in cfq_cic_free()?
>
> The data belonging to ->ioc_data (the cic, or per-process per-queue
> context) is only freed through call_rcu().

Ah, OK, got it.

> > Any of this at all helpful?
>
> Very, perhaps with a few more rounds we can find some more bugs :-). I'm
> attaching a patch below, how does that look?

Looks much improved! Very interested to hear how it does with the testing.

Thanx, Paul

> diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
> index 4df3f05..75db529 100644
> --- a/block/cfq-iosched.c
> +++ b/block/cfq-iosched.c
> @@ -1142,6 +1142,9 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
> kmem_cache_free(cfq_pool, cfqq);
> }
>
> +/*
> + * Must always be called with the rcu_read_lock() held
> + */
> static void
> __call_for_each_cic(struct io_context *ioc,
> void (*func)(struct io_context *, struct cfq_io_context *))
> @@ -1197,6 +1200,11 @@ static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
> cfq_cic_free(cic);
> }
>
> +/*
> + * Must be called with rcu_read_lock() held or preemption otherwise disabled.
> + * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
> + * and ->trim() which is called with the task lock held
> + */
> static void cfq_free_io_context(struct io_context *ioc)
> {
> /*
> @@ -1502,20 +1510,24 @@ static struct cfq_io_context *
> cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
> {
> struct cfq_io_context *cic;
> + unsigned long flags;
> void *k;
>
> if (unlikely(!ioc))
> return NULL;
>
> + rcu_read_lock();
> +
> /*
> * we maintain a last-hit cache, to avoid browsing over the tree
> */
> cic = rcu_dereference(ioc->ioc_data);
> - if (cic && cic->key == cfqd)
> + if (cic && cic->key == cfqd) {
> + rcu_read_unlock();
> return cic;
> + }
>
> do {
> - rcu_read_lock();
> cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
> rcu_read_unlock();
> if (!cic)
> @@ -1524,10 +1536,13 @@ cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
> k = cic->key;
> if (unlikely(!k)) {
> cfq_drop_dead_cic(cfqd, ioc, cic);
> + rcu_read_lock();
> continue;
> }
>
> + spin_lock_irqsave(&ioc->lock, flags);
> rcu_assign_pointer(ioc->ioc_data, cic);
> + spin_unlock_irqrestore(&ioc->lock, flags);
> break;
> } while (1);
>
> @@ -2134,6 +2149,11 @@ static void *cfq_init_queue(struct request_queue *q)
>
> static void cfq_slab_kill(void)
> {
> + /*
> + * Make sure that all existing RCU callbacks have been processed
> + */
> + rcu_barrier();
> +
> if (cfq_pool)
> kmem_cache_destroy(cfq_pool);
> if (cfq_ioc_pool)
>
> --
> Jens Axboe
>


\
 
 \ /
  Last update: 2008-05-28 12:33    [from the cache]
©2003-2011 Jasper Spaans