lkml.org 
[lkml]   [2009]   [Jun]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/2] memcg: cgroup fix rmdir hang
On Thu, 25 Jun 2009 14:28:09 -0700
Andrew Morton <akpm@linux-foundation.org> wrote:

> On Tue, 23 Jun 2009 16:07:20 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
>
> > previous discussion was this => http://marc.info/?t=124478543600001&r=1&w=2
> >
> > This patch tries to fix problem as
> > - rmdir can sleep very very long if swap entry is shared between multiple
> > cgroups
> >
> > Now, cgroup's rmdir path does following
> >
> > ==
> > again:
> > check there are no tasks and children group.
> > call pre_destroy()
> > check css's refcnt
> > if (refcnt > 0) {
> > sleep until css's refcnt goes down to 0.
> > goto again
> > }
> > ==
> >
> > Unfortunately, memory cgroup does following at charge.
> >
> > css_get(&memcg->css)
> > ....
> > charge(memcg) (increase USAGE)
> > ...
> > And this "memcg" is not necessary to include the caller, task.
> >
> > pre_destroy() tries to reduce memory usage until USAGE goes down to 0.
> > Then, there is a race that
> > - css's refcnt > 0 (and memcg's usage > 0)
> > - rmdir() caller sleeps until css->refcnt goes down 0.
> > - But to make css->refcnt be 0, pre_destroy() should be called again.
> >
> > This patch tries to fix this in asyhcnrounos way (i.e. without big lock.)
> > Any comments are welcome.
> >
>
> Do you believe that these fixes should be backported into 2.6.30.x?

Yes, I think so. (If it's easy)

To be honest:

To cause the problem,
- swap cgroup should be shared between cgroup.
- rmdir should be called in critical chance.

Considering usual usage of cgroup is "container", there will be no share of swap
in typical users. But, 2.6.30 can be a base kernel of a major distro. So,
I hope this in 2.6.30 if we have no difficulties.

Thanks,
-Kame












\
 
 \ /
  Last update: 2009-06-26 01:35    [W:0.146 / U:0.216 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site