Re: [RFC PATCH] mm, memcg: fix (Re: OOM: Better, but still there on)

[Mel Gorman <mgorman@suse.de>]

On Mon, Dec 26, 2016 at 01:48:40PM +0100, Michal Hocko wrote:
> On Fri 23-12-16 23:26:00, Nils Holland wrote:
> > On Fri, Dec 23, 2016 at 03:47:39PM +0100, Michal Hocko wrote:
> > > 
> > > Nils, even though this is still highly experimental, could you give it a
> > > try please?
> > 
> > Yes, no problem! So I kept the very first patch you sent but had to
> > revert the latest version of the debugging patch (the one in
> > which you added the "mm_vmscan_inactive_list_is_low" event) because
> > otherwise the patch you just sent wouldn't apply. Then I rebooted with
> > memory cgroups enabled again, and the first thing that strikes the eye
> > is that I get this during boot:
> > 
> > [    1.568174] ------------[ cut here ]------------
> > [    1.568327] WARNING: CPU: 0 PID: 1 at mm/memcontrol.c:1032 mem_cgroup_update_lru_size+0x118/0x130
> > [    1.568543] mem_cgroup_update_lru_size(f4406400, 2, 1): lru_size 0 but not empty
> 
> Ohh, I can see what is wrong! a) there is a bug in the accounting in
> my patch (I double account) and b) the detection for the empty list
> cannot work after my change because per node zone will not match per
> zone statistics. The updated patch is below. So I hope my brain already
> works after it's been mostly off last few days...
> ---
> From 397adf46917b2d9493180354a7b0182aee280a8b Mon Sep 17 00:00:00 2001
> From: Michal Hocko <mhocko@suse.com>
> Date: Fri, 23 Dec 2016 15:11:54 +0100
> Subject: [PATCH] mm, memcg: fix the active list aging for lowmem requests when
>  memcg is enabled
> 
> Nils Holland has reported unexpected OOM killer invocations with 32b
> kernel starting with 4.8 kernels
> 

I think it's unfortunate that per-zone stats are reintroduced to the
memcg structure. I can't help but think that it would have also worked
to always rotate a small number of pages if !inactive_list_is_low and
reclaiming for memcg even if it distorted page aging. However, given
that such an approach would be less robust and this has been heavily
tested;

Acked-by: Mel Gorman <mgorman@suse.de>

-- 
Mel Gorman
SUSE Labs