Re: Found the commit that causes the OOMs

[Minchan Kim <minchan.kim@gmail.com>]

On Sun, Jun 28, 2009 at 12:36 AM, Johannes Weiner<hannes@cmpxchg.org> wrote:
> On Sat, Jun 27, 2009 at 10:50:25PM +0900, Minchan Kim wrote:
>> Hi, Hannes.
>>
>> On Sat, Jun 27, 2009 at 9:54 PM, Johannes Weiner<hannes@cmpxchg.org> wrote:
>> > On Sat, Jun 27, 2009 at 08:12:49AM +0100, David Howells wrote:
>> >>
>> >> I've managed to bisect things to find the commit that causes the OOMs.  It's:
>> >>
>> >>       commit 69c854817566db82c362797b4a6521d0b00fe1d8
>> >>       Author: MinChan Kim <minchan.kim@gmail.com>
>> >>       Date:   Tue Jun 16 15:32:44 2009 -0700
>> >>
>> >>           vmscan: prevent shrinking of active anon lru list in case of no swap space V3
>> >>
>> >>           shrink_zone() can deactivate active anon pages even if we don't have a
>> >>           swap device.  Many embedded products don't have a swap device.  So the
>> >>           deactivation of anon pages is unnecessary.
>> >>
>> >>           This patch prevents unnecessary deactivation of anon lru pages.  But, it
>> >>           don't prevent aging of anon pages to swap out.
>> >>
>> >>           Signed-off-by: Minchan Kim <minchan.kim@gmail.com>
>> >>           Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
>> >>           Cc: Johannes Weiner <hannes@cmpxchg.org>
>> >>           Acked-by: Rik van Riel <riel@redhat.com>
>> >>           Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
>> >>           Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
>> >>
>> >> This exhibits the problem.  The previous commit:
>> >>
>> >>       commit 35282a2de4e5e4e173ab61aa9d7015886021a821
>> >>       Author: Brice Goglin <Brice.Goglin@ens-lyon.org>
>> >>       Date:   Tue Jun 16 15:32:43 2009 -0700
>> >>
>> >>           migration: only migrate_prep() once per move_pages()
>> >>
>> >> survives 16 iterations of the LTP syscall testsuite without exhibiting the
>> >> problem.
>> >
>> > Here is the patch in question:
>> >
>> > diff --git a/mm/vmscan.c b/mm/vmscan.c
>> > index 7592d8e..879d034 100644
>> > --- a/mm/vmscan.c
>> > +++ b/mm/vmscan.c
>> > @@ -1570,7 +1570,7 @@ static void shrink_zone(int priority, struct zone *zone,
>> >         * Even if we did not try to evict anon pages at all, we want to
>> >         * rebalance the anon lru active/inactive ratio.
>> >         */
>> > -       if (inactive_anon_is_low(zone, sc))
>> > +       if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
>> >                shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
>> >
>> >        throttle_vm_writeout(sc->gfp_mask);
>> >
>> > When this was discussed, I think we missed that nr_swap_pages can
>> > actually get zero on swap systems as well and this should have been
>> > total_swap_pages - otherwise we also stop balancing the two anon lists
>> > when swap is _full_ which was not the intention of this change at all.
>>
>> At that time we considered it so that we didn't prevent anon list
>> aging for background reclaim.
>> Do you think it is not enough ?
>
> With a heavy multiprocess anon load, direct reclaimers will likely
> reuse the reclaimed pages for anon mappings, so you have a handful of
> processes shuffling pages on the active list and only one thread that
> tries to balance.  I can imagine that it can not keep up for long.

I agree. :)
total_swap_pages is better than nr_swap_pages although it isn't
related this problem.


>



-- 
Kinds regards,
Minchan Kim