From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S933926AbcBDUeW (ORCPT ); Thu, 4 Feb 2016 15:34:22 -0500 Received: from gum.cmpxchg.org ([85.214.110.215]:56880 "EHLO gum.cmpxchg.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932861AbcBDUeU (ORCPT ); Thu, 4 Feb 2016 15:34:20 -0500 Date: Thu, 4 Feb 2016 15:33:28 -0500 From: Johannes Weiner To: Vladimir Davydov Cc: Andrew Morton , Michal Hocko , linux-mm@kvack.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH] mm: memcontrol: do not bypass slab charge if memcg is offline Message-ID: <20160204203328.GC8208@cmpxchg.org> References: <1454588275-7615-1-git-send-email-vdavydov@virtuozzo.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1454588275-7615-1-git-send-email-vdavydov@virtuozzo.com> User-Agent: Mutt/1.5.24 (2015-08-30) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu, Feb 04, 2016 at 03:17:55PM +0300, Vladimir Davydov wrote: > Slab pages are charged in two steps. First, an appropriate per memcg > cache is selected (see memcg_kmem_get_cache) basing on the current > context, then the new slab page is charged to the memory cgroup which > the selected cache was created for (see memcg_charge_slab -> > __memcg_kmem_charge_memcg). It is OK to bypass kmemcg charge at step 1, > but if step 1 succeeded and we successfully allocated a new slab page, > step 2 must be performed, otherwise we would get a per memcg kmem cache > which contains a slab that does not hold a reference to the memory > cgroup owning the cache. Since per memcg kmem caches are destroyed on > memcg css free, this could result in freeing a cache while there are > still active objects in it. > > However, currently we will bypass slab page charge if the memory cgroup > owning the cache is offline (see __memcg_kmem_charge_memcg). This is > very unlikely to occur in practice, because for this to happen a process > must be migrated to a different cgroup and the old cgroup must be > removed while the process is in kmalloc somewhere between steps 1 and 2 > (e.g. trying to allocate a new page). Nevertheless, it's still better > to eliminate such a possibility. > > Signed-off-by: Vladimir Davydov Acked-by: Johannes Weiner