lkml.org 
[lkml]   [2012]   [Mar]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v3 2/2] memcg: avoid THP split in task migration
On Fri, 9 Mar 2012 12:24:48 +0900
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:

> On Thu, 8 Mar 2012 18:33:14 -0800 (PST)
> Hugh Dickins <hughd@google.com> wrote:
>
> > On Fri, 9 Mar 2012, KAMEZAWA Hiroyuki wrote:
> > > > +
> > > > + page = pmd_page(pmd);
> > > > + VM_BUG_ON(!page || !PageHead(page));
> > > > + if (!move_anon() || page_mapcount(page) != 1)
> > > > + return 0;
> > >
> > > Could you add this ?
> > > ==
> > > static bool move_check_shared_map(struct page *page)
> > > {
> > > /*
> > > * Handling of shared pages between processes is a big trouble in memcg.
> > > * Now, we never move shared-mapped pages between memcg at 'task' moving because
> > > * we have no hint which task the page is really belongs to. For example,
> > > * When a task does "fork()-> move to the child other group -> exec()", the charges
> > > * should be stay in the original cgroup.
> > > * So, check mapcount to determine we can move or not.
> > > */
> > > return page_mapcount(page) != 1;
> > > }
> >
> > That's a helpful elucidation, thank you. However...
> >
> > That is not how it has actually been behaving for the last 18 months
> > (because of the "> 2" bug), so in practice you are asking for a change
> > in behaviour there.
> >
> Yes.
>
>
> > And it's not how it has been and continues to behave with file pages.
> >
> It's ok to add somethink like..
>
> if (PageAnon(page) && !move_anon())
> return false;
> ...
>
> > Isn't getting that behaviour in fork-move-exec just a good reason not
> > to set move_charge_at_immigrate?
> >
> Hmm. Maybe.
>
> > I think there are other scenarios where you do want all the pages to
> > move if move_charge_at_immigrate: and that's certainly easier to
> > describe and to understand and to code.
> >
> > But if you do insist on not moving the shared, then it needs to involve
> > something like mem_cgroup_count_swap_user() on PageSwapCache pages,
> > rather than just the bare page_mapcount().
> >
>
> This 'moving swap account' was a requirement from a user (NEC?).
> But no user doesn't say 'I want to move shared pages between cgroups at task
> move !' and I don't like to move shared objects.
>
> > I'd rather delete than add code here!
> >
>
> As a user, for Fujitsu, I believe it's insane to move task between cgroups.
> So, I have no benefit from this code, at all.
> Ok, maybe I'm not a stakeholder,here.
>

Considering again, in my personal opinion,
at fork() -> move() -> exec(), parent's RSS charge should not be moved.

Thanks,
-Kame



\
 
 \ /
  Last update: 2012-03-09 05:05    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site