SLUB: kmem_cache_destroy doesn't

SLUB: kmem_cache_destroy doesn't

[Neil Brown]

I tried out SLOB in latest mm.
raid5 uses kmem_cache_create to create a cache, a deletes it with
kmem_cache_destroy.

After I change the name for the slab to not include a '/' the create
worked fine, but the destroy seems incomplete.
If I stop an array and restart it, the second kmem_cache_create fails
because it thinks the name is already in use.

If I look in sysfs, I can see it there.
And 

  # cat /sys/slab/raid5-md1/aliases
  -1

is a bit suspicious.

And

void kmem_cache_destroy(struct kmem_cache *s)
{
	down_write(&slub_lock);
	if (s->refcount)
		s->refcount--;
	else {
		list_del(&s->list);
		WARN_ON(kmem_cache_close(s));
		sysfs_slab_remove(s);
		kfree(s);
	}
	up_write(&slub_lock);
}

looks a bit odd.
I would have thought the test should be:

        if (s->refcount > 1)

??

(2.6.21-rc6-mm1 to be precise)

NeilBrown

Re: SLUB: kmem_cache_destroy doesn't

[Christoph Lameter]

On Thu, 12 Apr 2007, Neil Brown wrote:

> And 
> 
>   # cat /sys/slab/raid5-md1/aliases
>   -1
> 
> is a bit suspicious.

Yuck!

> void kmem_cache_destroy(struct kmem_cache *s)
> {
> 	down_write(&slub_lock);
> 	if (s->refcount)
> 		s->refcount--;
> 	else {
> 		list_del(&s->list);
> 		WARN_ON(kmem_cache_close(s));
> 		sysfs_slab_remove(s);
> 		kfree(s);
> 	}
> 	up_write(&slub_lock);
> }
> 
> looks a bit odd.
> I would have thought the test should be:
> 
>         if (s->refcount > 1)

Correct. Does the change fix it?