Re: [External] Re: [PATCH] mm: memcontrol: fix missing wakeup oom task

[Michal Hocko <mhocko@suse.com>]

On Fri 05-02-21 23:30:36, Muchun Song wrote:
> On Fri, Feb 5, 2021 at 8:20 PM Michal Hocko <mhocko@suse.com> wrote:
> >
> > On Fri 05-02-21 19:04:19, Muchun Song wrote:
> > > On Fri, Feb 5, 2021 at 6:21 PM Michal Hocko <mhocko@suse.com> wrote:
> > > >
> > > > On Fri 05-02-21 17:55:10, Muchun Song wrote:
> > > > > On Fri, Feb 5, 2021 at 4:24 PM Michal Hocko <mhocko@suse.com> wrote:
> > > > > >
> > > > > > On Fri 05-02-21 14:23:10, Muchun Song wrote:
> > > > > > > We call memcg_oom_recover() in the uncharge_batch() to wakeup OOM task
> > > > > > > when page uncharged, but for the slab pages, we do not do this when page
> > > > > > > uncharged.
> > > > > >
> > > > > > How does the patch deal with this?
> > > > >
> > > > > When we uncharge a slab page via __memcg_kmem_uncharge,
> > > > > actually, this path forgets to do this for us compared to
> > > > > uncharge_batch(). Right?
> > > >
> > > > Yes this was more more or less clear (still would have been nicer to be
> > > > explicit). But you still haven't replied to my question I believe. I
> > > > assume you rely on refill_stock doing draining but how does this address
> > > > the problem? Is it sufficient to do wakeups in the batched way?
> > >
> > > Sorry, the subject title may not be suitable. IIUC, memcg_oom_recover
> > > aims to wake up the OOM task when we uncharge the page.
> >
> > Yes, your understanding is correct. This is a way to pro-actively wake
> > up oom victims when the memcg oom handling is outsourced to the
> > userspace. Please note that I haven't objected to the problem statement.
> >
> > I was questioning the fix for the problem.
> >
> > > I see uncharge_batch always do this. I am confused why
> > > __memcg_kmem_uncharge does not.
> >
> > Very likely an omission. I haven't checked closely but I suspect this
> > has been introduced by the recent kmem accounting changes.
> >
> > Why didn't you simply do the same thing and call memcg_oom_recover
> > unconditionally and instead depend on the draining? I suspect this was
> > because you wanted to recover also when draining which is not necessary
> > as pointed out in other email.
> 
> Thanks for your explanations. You are right. It is my fault to depend
> on the draining. I should call memcg_oom_recover directly in the
> __memcg_kmem_uncharge. Right?

Yes.

> > [...]
> > > > > > Does this lead to any code generation improvements? I would expect
> > > > > > compiler to be clever enough to inline static functions if that pays
> > > > > > off. If yes make this a patch on its own.
> > > > >
> > > > > I have disassembled the code, I see memcg_oom_recover is not
> > > > > inline. Maybe because memcg_oom_recover has a lot of callers.
> > > > > Just guess.
> > > > >
> > > > > (gdb) disassemble uncharge_batch
> > > > >  [...]
> > > > >  0xffffffff81341c73 <+227>: callq  0xffffffff8133c420 <page_counter_uncharge>
> > > > >  0xffffffff81341c78 <+232>: jmpq   0xffffffff81341bc0 <uncharge_batch+48>
> > > > >  0xffffffff81341c7d <+237>: callq  0xffffffff8133e2c0 <memcg_oom_recover>
> > > >
> > > > So does it really help to do the inlining?
> > >
> > > I just think memcg_oom_recover is very small, inline maybe
> > > a good choice. Maybe I am wrong.
> >
> > In general I am not overly keen on changes without a proper
> > justification. In this particular case I would understand that a
> > function call that will almost never do anything but the test (because
> > oom_disabled is a rarely used) is just waste of cycles in some hot
> > paths (e.g. kmem uncharge). Maybe this even has some visible performance
> > benefit. If this is really the case then would it make sense to guard
> > this test by the existing cgroup_subsys_on_dfl(memory_cgrp_subsys)?
> 
> Agree. I think it can improve performance when this
> function is inline. Guarding the test should be also
> an improvement on cgroup v2.

I would be surprised if this was measurable but you can give it a try. A
static key would be a reasonable argument for inlining on its own.
-- 
Michal Hocko
SUSE Labs