lkml.org 
[lkml]   [2011]   [Aug]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 2/8] mm: memcg-aware global reclaim
On Mon, Aug 29, 2011 at 12:22:02AM -0700, Ying Han wrote:
> On Mon, Aug 29, 2011 at 12:15 AM, Ying Han <yinghan@google.com> wrote:
> > On Thu, Aug 11, 2011 at 2:09 PM, Johannes Weiner <hannes@cmpxchg.org> wrote:
> >>
> >> On Thu, Aug 11, 2011 at 01:39:45PM -0700, Ying Han wrote:
> >> > Please consider including the following patch for the next post. It causes
> >> > crash on some of the tests where sc->mem_cgroup is NULL (global kswapd).
> >> >
> >> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> >> > index b72a844..12ab25d 100644
> >> > --- a/mm/vmscan.c
> >> > +++ b/mm/vmscan.c
> >> > @@ -2768,7 +2768,8 @@ loop_again:
> >> >                          * Do some background aging of the anon list, to
> >> > give
> >> >                          * pages a chance to be referenced before
> >> > reclaiming.
> >> >                          */
> >> > -                       if (inactive_anon_is_low(zone, &sc))
> >> > +                       if (scanning_global_lru(&sc) &&
> >> > +                                       inactive_anon_is_low(zone, &sc))
> >> >                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
> >> >                                                         &sc, priority, 0);
> >>
> >> Thanks!  I completely overlooked this one and only noticed it after
> >> changing the arguments to shrink_active_list().
> >>
> >> On memcg configurations, scanning_global_lru() will essentially never
> >> be true again, so I moved the anon pre-aging to a separate function
> >> that also does a hierarchy loop to preage the per-memcg anon lists.
> >>
> >> I hope to send out the next revision soon.
> >
> > Also, please consider to fold in the following patch as well. It fixes
> > the root cgroup lru accounting and we could easily trigger OOM while
> > doing some swapoff test w/o it.
> >
> > mm:fix the lru accounting for root cgroup.
> >
> > This patch is applied on top of:
> > "
> > mm: memcg-aware global reclaim
> > Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
> > "
> >
> > This patch fixes the lru accounting for root cgroup.
> >
> > After the "memcg-aware global reclaim" patch, one of the changes is to have
> > lru pages linked back to root. Under the global memory pressure, we start from
> > the root cgroup lru and walk through the memcg hierarchy of the system. For
> > each memcg, we reclaim pages based on the its lru size.
> >
> > However for root cgroup, we used not having a seperate lru and only counting
> > the pages charged to root as part of root lru size. Without this patch, all
> > the pages which are linked to root lru but not charged to root like swapcache
> > readahead are not visible to page reclaim code and we are easily to get OOM.
> >
> > After this patch, all the pages linked under root lru are counted in the lru
> > size, including Used and !Used.
> >
> > Signed-off-by: Hugh Dickins <hughd@google.com>
> > Signed-off-by: Ying Han <yinghan@google.com>
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 5518f54..f6c5f29 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -888,19 +888,21 @@ void mem_cgroup_del_lru_list(struct page *page,
> > enum lru_list lru)
> >  {
> >  >------struct page_cgroup *pc;
> >  >------struct mem_cgroup_per_zone *mz;
> > +>------struct mem_cgroup *mem;
> > ·
> >  >------if (mem_cgroup_disabled())
> >  >------>-------return;
> >  >------pc = lookup_page_cgroup(page);
> > ->------/* can happen while we handle swapcache. */
> > ->------if (!TestClearPageCgroupAcctLRU(pc))
> > ->------>-------return;
> > ->------VM_BUG_ON(!pc->mem_cgroup);
> > ->------/*
> > ->------ * We don't check PCG_USED bit. It's cleared when the "page" is finally
> > ->------ * removed from global LRU.
> > ->------ */
> > ->------mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
> > +
> > +>------if (TestClearPageCgroupAcctLRU(pc) || PageCgroupUsed(pc)) {

This PageCgroupUsed part confuses me. A page that is being isolated
shortly after being charged while on the LRU may reach here, and then
it is unaccounted from pc->mem_cgroup, which it never was accounted
to.

Could you explain why you added it?

I now made it so that PageCgroupAcctLRU on the LRU means accounted to
pc->mem_cgroup, and !PageCgroupAcctLRU on the LRU means accounted to
and babysitted by root_mem_cgroup. Always. Which also means that
before_commit now ensures an LRU page is moved to root_mem_cgroup for
babysitting during the charge, so that concurrent isolations/putbacks
are always accounted correctly. Is this what you had in mind? Did I
miss something?
--
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/

\
 
 \ /
  Last update: 2011-08-29 21:07    [from the cache]
©2003-2011 Jasper Spaans