Re: [PATCH -mm v3 7/8] slub: make dead memcg caches discard free slabs immediately

From: Vladimir Davydov <vdavydov@parallels.com>
To: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: <akpm@linux-foundation.org>, <cl@linux.com>,
	<rientjes@google.com>, <penberg@kernel.org>, <hannes@cmpxchg.org>,
	<mhocko@suse.cz>, <linux-kernel@vger.kernel.org>,
	<linux-mm@kvack.org>
Subject: Re: [PATCH -mm v3 7/8] slub: make dead memcg caches discard free slabs immediately
Date: Tue, 24 Jun 2014 12:25:26 +0400	[thread overview]
Message-ID: <20140624082526.GD18121@esperanza> (raw)
In-Reply-To: <20140624075011.GD4836@js1304-P5Q-DELUXE>

On Tue, Jun 24, 2014 at 04:50:11PM +0900, Joonsoo Kim wrote:
> On Fri, Jun 13, 2014 at 12:38:21AM +0400, Vladimir Davydov wrote:
> > @@ -3409,6 +3417,9 @@ int __kmem_cache_shrink(struct kmem_cache *s)
> >  		kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
> >  	unsigned long flags;
> >  
> > +	if (memcg_cache_dead(s))
> > +		s->min_partial = 0;
> > +
> >  	if (!slabs_by_inuse) {
> >  		/*
> >  		 * Do not fail shrinking empty slabs if allocation of the
> 
> I think that you should move down n->nr_partial test after holding the
> lock in __kmem_cache_shrink(). Access to n->nr_partial without node lock
> is racy and you can see wrong value. It results in skipping to free empty
> slab so your destroying logic could fail.

You're right! Will fix this.

And there seems to be the same problem in SLAB, where we check
node->slabs_free list emptiness w/o holding node->list_lock (see
drain_freelist) while it can be modified concurrently by free_block.
This will be fixed automatically after we make __kmem_cache_shrink unset
node->free_limit (which must be done under the lock) though.

Thank you!