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

[Mel Gorman <mgorman@suse.de>]

On Fri, Dec 30, 2016 at 12:05:45PM +0100, Michal Hocko wrote:
> On Fri 30-12-16 10:19:26, Mel Gorman wrote:
> > 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.
> 
> the original patch I had didn't add per zone stats but rather did a
> nr_highmem counter to mem_cgroup_per_node (inside ifdeff CONFIG_HIGMEM).
> This would help for this particular case but it wouldn't work for other
> lowmem requests (e.g. GFP_DMA32) and with the kmem accounting this might
> be a problem in future.

That did occur to me.

> So I've decided to go with a more generic
> approach which requires per-zone tracking. I cannot say I would be
> overly happy about this at all.
> 
> > 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.
> 
> I am not really sure how that would work. Do you mean something like the
> following?
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index fa30010a5277..563ada3c02ac 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -2044,6 +2044,9 @@ static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
>  	inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
>  	active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
>  
> +	if (!mem_cgroup_disabled())
> +		goto out;
> +
>  	/*
>  	 * For zone-constrained allocations, it is necessary to check if
>  	 * deactivations are required for lowmem to be reclaimed. This
> @@ -2063,6 +2066,7 @@ static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
>  		active -= min(active, active_zone);
>  	}
>  
> +out:
>  	gb = (inactive + active) >> (30 - PAGE_SHIFT);
>  	if (gb)
>  		inactive_ratio = int_sqrt(10 * gb);
> 
> The problem I see with such an approach is that chances are that this
> would reintroduce what f8d1a31163fc ("mm: consider whether to decivate
> based on eligible zones inactive ratio") tried to fix. But maybe I have
> missed your point.
> 

No, you didn't miss the point. It was something like that I had in mind
but as I thought about it, I could see some cases where it might not work
and still cause a premature OOM. The per-zone accounting is unfortunate
but it's robust hence the Ack.

-- 
Mel Gorman
SUSE Labs