From mboxrd@z Thu Jan 1 00:00:00 1970 From: Kamezawa Hiroyuki Subject: Re: [PATCH V3 4/8] memcg: add per cgroup dirty pages accounting Date: Wed, 09 Jan 2013 16:24:34 +0900 Message-ID: <50ED1B32.2030007@jp.fujitsu.com> References: <1356455919-14445-1-git-send-email-handai.szj@taobao.com> <1356456367-14660-1-git-send-email-handai.szj@taobao.com> <20130102104421.GC22160@dhcp22.suse.cz> <50EA7E07.4070902@jp.fujitsu.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: Sha Zhengju , Michal Hocko , Johannes Weiner , linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-mm-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org, linux-fsdevel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, akpm-de/tnXTf+JLsfHDXvbKv3WD2FQJk+8+b@public.gmane.org, gthelen-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org, fengguang.wu-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org, glommer-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org, dchinner-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org, Sha Zhengju To: Hugh Dickins Return-path: In-Reply-To: Sender: cgroups-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org List-Id: linux-fsdevel.vger.kernel.org (2013/01/09 14:15), Hugh Dickins wrote: > On Mon, 7 Jan 2013, Kamezawa Hiroyuki wrote: >> (2013/01/07 5:02), Hugh Dickins wrote: >>> >>> Forgive me, I must confess I'm no more than skimming this thread, >>> and don't like dumping unsigned-off patches on people; but thought >>> that on balance it might be more helpful than not if I offer you a >>> patch I worked on around 3.6-rc2 (but have updated to 3.8-rc2 below). >>> >>> I too was getting depressed by the constraints imposed by >>> mem_cgroup_{begin,end}_update_page_stat (good job though Kamezawa-san >>> did to minimize them), and wanted to replace by something freer, more >>> RCU-like. In the end it seemed more effort than it was worth to go >>> as far as I wanted, but I do think that this is some improvement over >>> what we currently have, and should deal with your recursion issue. >>> >> In what case does this improve performance ? > > Perhaps none. I was aiming to not degrade performance at the stats > update end, and make it more flexible, so new stats can be updated which > would be problematic today (for lock ordering and recursion reasons). > > I've not done any performance measurement on it, and don't have enough > cpus for an interesting report; but if someone thinks it might solve a > problem for them, and has plenty of cpus to test with, please go ahead, > we'd be glad to hear the results. > >> Hi, this patch seems interesting but...doesn't this make move_account() very >> slow if the number of cpus increases because of scanning all cpus per a page >> ? >> And this looks like reader-can-block-writer percpu rwlock..it's too heavy to >> writers if there are many readers. > > I was happy to make the relatively rare move_account end considerably > heavier. I'll be disappointed if it turns out to be prohibitively > heavy at that end - if we're going to make move_account impossible, > there are much easier ways to achieve that! - but it is a possibility. > move_account at task-move has been required feature for NEC and Nishimura-san did good job. I'd like to keep that available as much as possible. > Something you might have missed when considering many readers (stats > updaters): the move_account end does not wait for a moment when there > are no readers, that would indeed be a losing strategy; it just waits > for each cpu that's updating page stats to leave that section, so every > cpu is sure to notice and hold off if it then tries to update the page > which is to be moved. (I may not be explaining that very well!) > Hmm, yeah, maybe I miss somehing. BTW, if nesting, mem_cgroup_end_update_page_stat() seems to make counter minus. Thanks, -Kame