Date:	Thu, 2 Jun 2011 19:29:05 +0200
From:	Johannes Weiner <hannes@...xchg.org>
To:	Hiroyuki Kamezawa <kamezawa.hiroyuki@...il.com>
Cc:	KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>,
	Daisuke Nishimura <nishimura@....nes.nec.co.jp>,
	Balbir Singh <balbir@...ux.vnet.ibm.com>,
	Ying Han <yinghan@...gle.com>, Michal Hocko <mhocko@...e.cz>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Rik van Riel <riel@...hat.com>,
	Minchan Kim <minchan.kim@...il.com>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>,
	Mel Gorman <mgorman@...e.de>, Greg Thelen <gthelen@...gle.com>,
	Michel Lespinasse <walken@...gle.com>, linux-mm@...ck.org,
	linux-kernel@...r.kernel.org
Subject: Re: [patch 2/8] mm: memcg-aware global reclaim

On Fri, Jun 03, 2011 at 01:14:12AM +0900, Hiroyuki Kamezawa wrote:
> 2011/6/3 Johannes Weiner <hannes@...xchg.org>:
> > On Thu, Jun 02, 2011 at 10:59:01PM +0900, Hiroyuki Kamezawa wrote:
> >> 2011/6/1 Johannes Weiner <hannes@...xchg.org>:
> 
> >> > @@ -1927,8 +1980,7 @@ static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
> >> >        if (!(gfp_mask & __GFP_WAIT))
> >> >                return CHARGE_WOULDBLOCK;
> >> >
> >> > -       ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
> >> > -                                             gfp_mask, flags);
> >> > +       ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
> >> >        if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
> >> >                return CHARGE_RETRY;
> >> >        /*
> >>
> >> It seems this clean-up around hierarchy and softlimit can be in an
> >> independent patch, no ?
> >
> > Hm, why do you think it's a cleanup?  The hierarchical target reclaim
> > code is moved to vmscan.c and as a result the entry points for hard
> > limit and soft limit reclaim differ.  This is why the original
> > function, mem_cgroup_hierarchical_reclaim() has to be split into two
> > parts.
> >
> If functionality is unchanged, I think it's clean up.
> I agree to move hierarchy walk to vmscan.c. but it can be done as
> a clean up patch for current code.
> (Make current try_to_free_mem_cgroup_pages() to use this code.)
>  and then, you can write a patch which only includes a core
> logic/purpose of this patch
> "use root cgroup's LRU for global and make global reclaim as full-scan
> of memcgroup."
> 
> In short, I felt this patch is long....and maybe watchers of -mm are
> not interested in rewritie of hierarchy walk but are intetested in the
> chages in shrink_zone() itself very much.

But the split up is, unfortunately, a change in functionality.  The
current code selects one memcg and reclaims all zones on all priority
levels on behalf of that memcg.  My code changes that such that it
reclaims a bunch of memcgs from the hierarchy for each zone and
priority level instead.  From memcgs -> priorities -> zones to
priorities -> zones -> memcgs.

I don't want to pass that off as a cleanup.

But it is long, I agree with you.  I'll split up the 'move
hierarchical target reclaim to generic code' from 'make global reclaim
hierarchical' and see if this makes the changes more straight-forward.

Because I suspect the perceived unwieldiness does not stem from the
amount of lines changed, but from the number of different logical
changes.

> >> > +       for (;;) {
> >> > +               unsigned long nr_reclaimed;
> >> > +
> >> > +               sc->mem_cgroup = mem;
> >> > +               do_shrink_zone(priority, zone, sc);
> >> > +
> >> > +               nr_reclaimed = sc->nr_reclaimed - nr_reclaimed_before;
> >> > +               if (nr_reclaimed >= sc->nr_to_reclaim)
> >> > +                       break;
> >>
> >> what this calculation means ?  Shouldn't we do this quit based on the
> >> number of "scan"
> >> rather than "reclaimed" ?
> >
> > It aborts the loop once sc->nr_to_reclaim pages have been reclaimed
> > from that zone during that hierarchy walk, to prevent overreclaim.
> >
> > If you have unbalanced sizes of memcgs in the system, it is not
> > desirable to have every reclaimer scan all memcgs, but let those quit
> > early that have made some progress on the bigger memcgs.
> >
> Hmm, why not if (sc->nr_reclaimed >= sc->nr_to_reclaim) ?
> 
> I'm sorry if I miss something..

It's a bit awkward and undocumented, I'm afraid.  The loop is like
this:

	for each zone:
	  for each memcg:
	    shrink
	    if sc->nr_reclaimed >= sc->nr_to_reclaim:
	      break

sc->nr_reclaimed is never reset, so once you reclaimed enough pages
from one zone, you will only try the first memcg in all the other
zones, which might well be empty, so no pressure at all on subsequent
zones.

That's why I use the per-zone delta like this:

       for each zone:
         before = sc->nr_reclaimed
	 for each memcg:
	   shrink
	   if sc->nr_reclaimed - before >= sc->nr_to_reclaim

which still ensures on one hand that we don't keep hammering a zone if
we reclaimed the overall reclaim target already, but on the other hand
that we apply some pressure to the other zones as well.

It's the same concept as in do_shrink_zone().  It breaks the loop when

	nr_reclaimed >= sc->nr_to_reclaim

where nr_reclaimed refers to the number of pages reclaimed from the
current zone, not the accumulated total of the whole reclaim cycle.

> >> > +               mem = mem_cgroup_hierarchy_walk(root, mem);
> >> > +               if (mem == first)
> >> > +                       break;
> >>
> >> Why we quit loop  ?
> >
> > get_scan_count() for traditional global reclaim returns the scan
> > target for the zone.
> >
> > With this per-memcg reclaimer, get_scan_count() will return scan
> > targets for the respective per-memcg zone subsizes.
> >
> > So once we have gone through all memcgs, we should have scanned the
> > amount of pages that global reclaim would have deemed sensible for
> > that zone at that priority level.
> >
> > As such, this is the exit condition based on scan count you referred
> > to above.
> >
> That's what I want as a comment in codes.

Will do, for both exit conditions ;-)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists