Re: [PATCH v3 15/16] memcg/sl[au]b: shrink dead caches

[Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>]

Hello, Glauber.

On Tue, Sep 18, 2012 at 06:12:09PM +0400, Glauber Costa wrote:
> @@ -764,10 +777,21 @@ static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
>  		goto out;
>  	}
>  
> +	/*
> +	 * Because the cache is expected to duplicate the string,
> +	 * we must make sure it has opportunity to copy its full
> +	 * name. Only now we can remove the dead part from it
> +	 */
> +	name = (char *)new_cachep->name;
> +	if (name)
> +		name[strlen(name) - 4] = '\0';

This is kinda nasty.  Do we really need to do this?  How long would a
dead cache stick around?

> diff --git a/mm/slab.c b/mm/slab.c
> index bd9928f..6cb4abf 100644
> --- a/mm/slab.c
> +++ b/mm/slab.c
> @@ -3785,6 +3785,8 @@ static inline void __cache_free(struct kmem_cache *cachep, void *objp,
>  	}
>  
>  	ac_put_obj(cachep, ac, objp);
> +
> +	kmem_cache_verify_dead(cachep);

Reaping dead caches doesn't exactly sound like a high priority thing
and adding a branch to hot path for that might not be the best way to
do it.  Why not schedule an extremely lazy deferrable delayed_work
which polls for emptiness, say, every miniute or whatever?

Thanks.

-- 
tejun