[PATCH] mm, slab: release slab_mutex earlier in kmem_cache_destroy()(was Re: Lockdep complains about commit 1331e7a1bb ("rcu: Remove _rcu_barrier()dependency on __stop_machine()"))

From: Jiri Kosina
Date: Tue Oct 02 2012 - 20:45:50 EST


On Tue, 2 Oct 2012, Paul E. McKenney wrote:

> On Wed, Oct 03, 2012 at 01:48:21AM +0200, Jiri Kosina wrote:
> > On Tue, 2 Oct 2012, Paul E. McKenney wrote:
> >
> > > Indeed. Slab seems to be doing an rcu_barrier() in a CPU hotplug
> > > notifier, which doesn't sit so well with rcu_barrier() trying to exclude
> > > CPU hotplug events. I could go back to the old approach, but it is
> > > significantly more complex. I cannot say that I am all that happy about
> > > anyone calling rcu_barrier() from a CPU hotplug notifier because it
> > > doesn't help CPU hotplug latency, but that is a separate issue.
> > >
> > > But the thing is that rcu_barrier()'s assumptions work just fine if either
> > > (1) it excludes hotplug operations or (2) if it is called from a hotplug
> > > notifier. You see, either way, the CPU cannot go away while rcu_barrier()
> > > is executing. So the right way to resolve this seems to be to do the
> > > get_online_cpus() only if rcu_barrier() is -not- executing in the context
> > > of a hotplug notifier. Should be fixable without too much hassle...
> >
> > Sorry, I don't think I understand what you are proposing just yet.
> >
> > If I understand it correctly, you are proposing to introduce some magic
> > into _rcu_barrier() such as (pseudocode of course):
> >
> > if (!being_called_from_hotplug_notifier_callback)
> > get_online_cpus()
> >
> > How does that protect from the scenario I've outlined before though?
> >
> > CPU 0 CPU 1
> > kmem_cache_destroy()
> > mutex_lock(slab_mutex)
> > _cpu_up()
> > cpu_hotplug_begin()
> > mutex_lock(cpu_hotplug.lock)
> > rcu_barrier()
> > _rcu_barrier()
> > get_online_cpus()
> > mutex_lock(cpu_hotplug.lock)
> > (blocks, CPU 1 has the mutex)
> > __cpu_notify()
> > mutex_lock(slab_mutex)
> >
> > CPU 0 grabs both locks anyway (it's not running from notifier callback).
> > CPU 1 grabs both locks as well, as there is no _rcu_barrier() being called
> > from notifier callback either.
> >
> > What did I miss?
>
> You didn't miss anything, I was suffering a failure to read carefully.
>
> So my next stupid question is "Why can't kmem_cache_destroy drop
> slab_mutex early?" like the following:
>
> void kmem_cache_destroy(struct kmem_cache *cachep)
> {
> BUG_ON(!cachep || in_interrupt());
>
> /* Find the cache in the chain of caches. */
> get_online_cpus();
> mutex_lock(&slab_mutex);
> /*
> * the chain is never empty, cache_cache is never destroyed
> */
> list_del(&cachep->list);
> if (__cache_shrink(cachep)) {
> slab_error(cachep, "Can't free all objects");
> list_add(&cachep->list, &slab_caches);
> mutex_unlock(&slab_mutex);
> put_online_cpus();
> return;
> }
> mutex_unlock(&slab_mutex);
>
> if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
> rcu_barrier();
>
> __kmem_cache_destroy(cachep);
> put_online_cpus();
> }
>
> Or did I miss some reason why __kmem_cache_destroy() needs that lock?
> Looks to me like it is just freeing now-disconnected memory.

Good question. I believe it should be safe to drop slab_mutex earlier, as
cachep has already been unlinked. I am adding slab people and linux-mm to
CC (the whole thread on LKML can be found at
https://lkml.org/lkml/2012/10/2/296 for reference).

How about the patch below? Pekka, Christoph, please?

It makes the lockdep happy again, and obviously removes the deadlock (I
tested it).



From: Jiri Kosina <jkosina@xxxxxxx>
Subject: mm, slab: release slab_mutex earlier in kmem_cache_destroy()

Commit 1331e7a1bbe1 ("rcu: Remove _rcu_barrier() dependency on
__stop_machine()") introduced slab_mutex -> cpu_hotplug.lock
dependency through kmem_cache_destroy() -> rcu_barrier() ->
_rcu_barrier() -> get_online_cpus().

This opens a possibilty for deadlock:

CPU 0 CPU 1
kmem_cache_destroy()
mutex_lock(slab_mutex)
_cpu_up()
cpu_hotplug_begin()
mutex_lock(cpu_hotplug.lock)
rcu_barrier()
_rcu_barrier()
get_online_cpus()
mutex_lock(cpu_hotplug.lock)
(blocks, CPU 1 has the mutex)
__cpu_notify()
mutex_lock(slab_mutex)

It turns out that slab's kmem_cache_destroy() might release slab_mutex
earlier before calling out to rcu_barrier(), as cachep has already been
unlinked.

This patch removes the AB-BA dependency by calling rcu_barrier() with
slab_mutex already unlocked.

Signed-off-by: Jiri Kosina <jkosina@xxxxxxx>
---
mm/slab.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/mm/slab.c b/mm/slab.c
index 1133911..693c7cb 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -2801,12 +2801,12 @@ void kmem_cache_destroy(struct kmem_cache *cachep)
put_online_cpus();
return;
}
+ mutex_unlock(&slab_mutex);

if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
rcu_barrier();

__kmem_cache_destroy(cachep);
- mutex_unlock(&slab_mutex);
put_online_cpus();
}
EXPORT_SYMBOL(kmem_cache_destroy);

--
Jiri Kosina
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/