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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



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@xxxxxxxxxxx>
> 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.

> Looking at the other thread with Michal, it looks like it's because
> we
> have to go into reclaim first before we get to the point of force
> charging for dying tasks, and we spend too much time in reclaim. Is
> that correct?
> 
> If that's the case, I am wondering if the real problem is that we
> check  mem_cgroup_zswap_writeback_enabled() too late in the process.
> Reclaim ages the LRUs, isolates pages, unmaps them, allocates swap
> entries, only to realize it cannot swap in swap_writepage().
> 
> Should we check for this in can_reclaim_anon_pages()? If zswap
> writeback is disabled and we are already at the memcg limit (or zswap
> limit for that matter), we should avoid scanning anon memory to begin
> with. The problem is that if we race with memory being freed we may
> have some extra OOM kills, but I am not sure how common this case
> would be.

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.

> 
> (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 :)

> (c) mem_cgroup_from_task() should be called in an RCU read section
> (or
> we need something like rcu_access_point() if we are not dereferencing
> the pointer).
> 
I'll add this in v2.

-- 
All Rights Reversed.





[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux