Re: [PATCH] memcg: allow exiting tasks to write back data to swap

[Rik van Riel <riel@surriel.com>]

On Wed, 2024-12-11 at 09:00 -0800, Yosry Ahmed wrote:
> On Wed, Dec 11, 2024 at 8:34 AM Rik van Riel <riel@surriel.com>
> wrote:
> > 
> > On Wed, 2024-12-11 at 08:26 -0800, Yosry Ahmed wrote:
> > > On Wed, Dec 11, 2024 at 7:54 AM Rik van Riel <riel@surriel.com>
> > > wrote:
> > > > 
> > > > +++ b/mm/memcontrol.c
> > > > @@ -5371,6 +5371,15 @@ bool
> > > > mem_cgroup_zswap_writeback_enabled(struct mem_cgroup *memcg)
> > > >         if (!zswap_is_enabled())
> > > >                 return true;
> > > > 
> > > > +       /*
> > > > +        * Always allow exiting tasks to push data to swap. A
> > > > process in
> > > > +        * the middle of exit cannot get OOM killed, but may
> > > > need
> > > > to push
> > > > +        * uncompressible data to swap in order to get the
> > > > cgroup
> > > > memory
> > > > +        * use below the limit, and make progress with the
> > > > exit.
> > > > +        */
> > > > +       if ((current->flags & PF_EXITING) && memcg ==
> > > > mem_cgroup_from_task(current))
> > > > +               return true;
> > > > +
> > > 
> > > I have a few questions:
> > > (a) If the task is being OOM killed it should be able to charge
> > > memory
> > > beyond memory.max, so why do we need to get the usage down below
> > > the
> > > limit?
> > > 
> > If it is a kernel directed memcg OOM kill, that is
> > true.
> > 
> > However, if the exit comes from somewhere else,
> > like a userspace oomd kill, we might not hit that
> > code path.
> 
> Why do we treat dying tasks differently based on the source of the
> kill?
> 
Are you saying we should fail allocations for
every dying task, and add a check for PF_EXITING
in here?


        if (unlikely(task_in_memcg_oom(current)))
                goto nomem;


> > However, we don't know until the attempted zswap write
> > whether the memory is compressible, and whether doing
> > a bunch of zswap writes will help us bring our memcg
> > down below its memory.max limit.
> 
> If we are at memory.max (or memory.zswap.max), we can't compress
> pages
> into zswap anyway, regardless of their compressibility.
> 
Wait, this is news to me.

This seems like something we should fix, rather
than live with, since compressing the data to
a smaller size could bring us below memory.max.

Is this "cannot compress when at memory.max"
behavior intentional, or just a side effect of
how things happen to be?

Won't the allocations made from zswap_store
ignore the memory.max limit because PF_MEMALLOC
is set?

> > > 
> > > (b) Should we use mem_cgroup_is_descendant() or mm_match_memcg()
> > > in
> > > case we are reclaiming from an ancestor and we hit the limit of
> > > that
> > > ancestor?
> > > 
> > I don't know if we need or want to reclaim from any
> > other memcgs than those of the exiting process itself.
> > 
> > A small blast radius seems like it could be desirable,
> > but I'm open to other ideas :)
> 
> The exiting process is part of all the ancestor cgroups by the
> hierarchy.
> 
> If we have the following hierarchy:
> root
>    |
>   A
>    |
>   B
> 
> Then a process in cgroup B could be getting OOM killed due to hitting
> the limit of A, not B. In which case, reclaiming from A helps us get
> below the limit. We can check if the cgroup is an ancestor and it hit
> its limit, but maybe that's an overkill.

Since we're dealing with a corner case anyway, I
suppose there's no harm using mm_match_cgroup,
which also happens to be cleaner than the code
I have right now.
-- 
All Rights Reversed.