Re: [RFC][PATCH 2/2] memcg: new lock for mutual execution ofaccount_move and file stats

[KAMEZAWA Hiroyuki]

On Sun, 17 Oct 2010 14:35:47 +0900
Minchan Kim <minchan.kim@xxxxxxxxx> wrote:

> On Sun, Oct 17, 2010 at 2:33 PM, Minchan Kim <minchan.kim@xxxxxxxxx> wrote:
> > On Fri, Oct 15, 2010 at 5:12 PM, KAMEZAWA Hiroyuki
> > <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> >> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> >>
> >> When we try to enhance page's status update to support other flags,
> >> one of problem is updating status from IRQ context.
> >>
> >> Now, mem_cgroup_update_file_stat() takes lock_page_cgroup() to avoid
> >> race with _account move_. IOW, there are no races with charge/uncharge
> >> in nature. Considering an update from IRQ context, it seems better
> >> to disable IRQ at lock_page_cgroup() to avoid deadlock.
> >>
> >> But lock_page_cgroup() is used too widerly and adding IRQ disable
> >> there makes the performance bad. To avoid the big hammer, this patch
> >> adds a new lock for update_stat().
> >>
> >> This lock is for mutual execustion of updating stat and accout moving.
> >> This adds a new lock to move_account..so, this makes move_account slow.
> >> But considering trade-off, I think it's acceptable.
> >>
> >> A score of moving 8GB anon pages, 8cpu Xeon(3.1GHz) is here.
> >>
> >> [before patch] (mmotm + optimization patch (#1 in this series)
> >> [root@bluextal kamezawa]# time echo 2257 > /cgroup/B/tasks
> >>
> >> real  Â0m0.694s
> >> user  Â0m0.000s
> >> sys   0m0.683s
> >>
> >> [After patch]
> >> [root@bluextal kamezawa]# time echo 2238 > /cgroup/B/tasks
> >>
> >> real  Â0m0.741s
> >> user  Â0m0.000s
> >> sys   0m0.730s
> >>
> >> This moves 8Gbytes == 2048k pages. But no bad effects to codes
> >> other than "move".
> >>
> >> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> >
> > It looks good than old approach.
> > Just a below nitpick.
> >
> >> ---
> >> Âinclude/linux/page_cgroup.h | Â 29 +++++++++++++++++++++++++++++
> >> Âmm/memcontrol.c       |  11 +++++++++--
> >> Â2 files changed, 38 insertions(+), 2 deletions(-)
> >>
> >> Index: mmotm-1013/include/linux/page_cgroup.h
> >> ===================================================================
> >> --- mmotm-1013.orig/include/linux/page_cgroup.h
> >> +++ mmotm-1013/include/linux/page_cgroup.h
> >> @@ -36,6 +36,7 @@ struct page_cgroup *lookup_page_cgroup(s
> >> Âenum {
> >> Â Â Â Â/* flags for mem_cgroup */
> >> Â Â Â ÂPCG_LOCK, Â/* page cgroup is locked */
> >> + Â Â Â PCG_LOCK_STATS, /* page cgroup's stat accounting flags are locked */
> >
> > Hmm, I think naming isn't a good. Aren't both for stat?

PCG_LOCK is for page_cgroup->mem_cgroup, not for stat.


But hmm...how about
{
  PCG_LOCK  /* For CACEH, USED and pc->mem_cgroup */
  PCG_CACHE
  PCG_USED
  PCG_ACCT_LRU /* no lock is used */
  PCG_MOVE_FLAGS_LOCK  /* For MAPPED and I/O flags v.s account_move races*/
  PCG_FILE_MAPPED,
  ..
  PCG_MIGRATION, /* For remembering Page Migration */
}

Anyway, documentation should be updated.
...


> > As I understand, Both are used for stat.
> > One is just used by charge/uncharge and the other is used by
> > pdate_file_stat/move_account.
> > If you guys who are expert in mcg feel it with easy, I am not against.
> > But at least, mcg-not-familiar people like me don't feel it comfortable.
> >
> 
> And I think this patch would be better to be part of Greg Thelen's series.
> 

Hmm. Greg, can you merge my new version (I'll post today) into your series ?

Thanks,
-Kame



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/