Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753235Ab1FJIq4 (ORCPT ); Fri, 10 Jun 2011 04:46:56 -0400 Received: from fgwmail5.fujitsu.co.jp ([192.51.44.35]:47700 "EHLO fgwmail5.fujitsu.co.jp" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752815Ab1FJIqy (ORCPT ); Fri, 10 Jun 2011 04:46:54 -0400 X-SecurityPolicyCheck-FJ: OK by FujitsuOutboundMailChecker v1.3.1 Date: Fri, 10 Jun 2011 17:39:58 +0900 From: KAMEZAWA Hiroyuki To: Michal Hocko Cc: "linux-mm@kvack.org" , "linux-kernel@vger.kernel.org" , "akpm@linux-foundation.org" , "nishimura@mxp.nes.nec.co.jp" , "bsingharora@gmail.com" , Ying Han Subject: Re: [BUGFIX][PATCH v3] memcg: fix behavior of per cpu charge cache draining. Message-Id: <20110610173958.d9ab901c.kamezawa.hiroyu@jp.fujitsu.com> In-Reply-To: <20110610081218.GC4832@tiehlicka.suse.cz> References: <20110609093045.1f969d30.kamezawa.hiroyu@jp.fujitsu.com> <20110610081218.GC4832@tiehlicka.suse.cz> Organization: FUJITSU Co. LTD. X-Mailer: Sylpheed 3.1.0 (GTK+ 2.10.14; i686-pc-mingw32) Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4051 Lines: 114 On Fri, 10 Jun 2011 10:12:19 +0200 Michal Hocko wrote: > On Thu 09-06-11 09:30:45, KAMEZAWA Hiroyuki wrote: > > From 0ebd8a90a91d50c512e7c63e5529a22e44e84c42 Mon Sep 17 00:00:00 2001 > > From: KAMEZAWA Hiroyuki > > Date: Wed, 8 Jun 2011 13:51:11 +0900 > > Subject: [PATCH] Fix behavior of per-cpu charge cache draining in memcg. > > > > For performance, memory cgroup caches some "charge" from res_counter > > into per cpu cache. This works well but because it's cache, > > it needs to be flushed in some cases. Typical cases are > > 1. when someone hit limit. > > 2. when rmdir() is called and need to charges to be 0. > > > > But "1" has problem. > > > > Recently, with large SMP machines, we many kworker runs because > > of flushing memcg's cache. Bad things in implementation are > > > > a) it's called before calling try_to_free_mem_cgroup_pages() > > so, it's called immidiately when a task hit limit. > > (I though it was better to avoid to run into memory reclaim. > > But it was wrong decision.) > > > > b) Even if a cpu contains a cache for memcg not related to > > a memcg which hits limit, drain code is called. > > > > This patch fixes a) and b) by > > > > A) delay calling of flushing until one run of try_to_free... > > Then, the number of calling is decreased. > > B) check percpu cache contains a useful data or not. > > plus > > C) check asynchronous percpu draining doesn't run. > > > > BTW, why this patch relpaces atomic_t counter with mutex is > > to guarantee a memcg which is pointed by stock->cacne is > > not destroyed while we check css_id. > > > > Reported-by: Ying Han > > Reviewed-by: Michal Hocko > > Signed-off-by: KAMEZAWA Hiroyuki > > > > Changelog: > > - fixed typo. > > - fixed rcu_read_lock() and add strict mutal execution between > > asynchronous and synchronous flushing. It's requred for validness > > of cached pointer. > > - add root_mem->use_hierarchy check. > > --- > > mm/memcontrol.c | 54 +++++++++++++++++++++++++++++++++++------------------- > > 1 files changed, 35 insertions(+), 19 deletions(-) > > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c > > index bd9052a..3baddcb 100644 > > --- a/mm/memcontrol.c > > +++ b/mm/memcontrol.c > [...] > > static struct mem_cgroup_per_zone * > > mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) > > @@ -1670,8 +1670,6 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, > > victim = mem_cgroup_select_victim(root_mem); > > if (victim == root_mem) { > > loop++; > > - if (loop >= 1) > > - drain_all_stock_async(); > > if (loop >= 2) { > > /* > > * If we have not been able to reclaim > > @@ -1723,6 +1721,7 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, > > return total; > > } else if (mem_cgroup_margin(root_mem)) > > return total; > > + drain_all_stock_async(root_mem); > > } > > return total; > > } > > I still think that we pointlessly reclaim even though we could have a > lot of pages pre-charged in the cache (the more CPUs we have the more > significant this might be). The more CPUs, the more scan cost for each per-cpu memory, which makes cache-miss. I know placement of drain_all_stock_async() is not big problem on my host, which has 2socket/8core cpus. But, assuming 1000+ cpu host, "when you hit limit, you'll see 1000*128bytes cache miss and need to call test_and_set for 1000+ cpus in bad case." doesn't seem much win. If we implement "call-drain-only-nearby-cpus", I think we can call it before calling try_to_free_mem_cgroup_pages(). I'll add it to my TO-DO-LIST. How do you think ? Thanks, -Kame -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/