Re: [PATCH v20 0/6] Hierarchical Percpu Counters for RSS

[Lorenzo Stoakes (ARM)]

On Wed, Aug 05, 2026 at 03:03:39PM -0400, Mathieu Desnoyers wrote:
> On 2026-08-05 14:57, Andrew Morton wrote:
> > On Wed, 5 Aug 2026 14:37:08 -0400 Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
> >
> > > On 2026-07-23 13:53, Mathieu Desnoyers wrote:
> > > > On 2026-07-07 09:15, Mathieu Desnoyers wrote:
> > > > > Hi Andrew,
> > > > >
> > > > > Here is the hierarchical percpu counters series rebased on top of
> > > > > v7.2-rc2. It includes small bootup fixes which were needed to fix
> > > > > bootup sequence on specific architectures, and a rename of the
> > > > > test config option to include "KUNIT_".
> > > > >
> > > > > This aims at replacing the prior version of the series you had
> > > > > in mm.
> > > > >
> > > > > As a reminder, the goal here is to provide more precise RSS counters
> > > > > through /proc.
> > > >
> > > > Hello,
> > > >
> > > > Just a gentle ping for feedback whenever it's convenient.
> > >
> > > Hi Andrew,
> > >
> > > Can you pick this up after the upcoming merge window ?
> >
> > Resending after -rc1 would be appropriate.
>
> Will do, thanks!
>
> >
> > > That's of course assuming this is solving an issue which is
> > > still relevant.
> > >
> > > If there is no interest in solving this issue anymore, kindly
> > > let me know and I'll drop this series.
> >
> > This was 1000-2000 patches ago so I've forgotten what the issue was!
> > Please ensure that changelogging describes the issue in glorious
> > detail and I'm sure it'll all come back.
>
> Certainly. For the immediate records, the issue addressed by this series
> is RSS stats inaccuracy when reading /proc files on large SMP boxes. It
> gets increasingly worse for long-lifetime processes which jump around
> various cores while allocating/freeing memory over their lifetime.
>
> And once we get this in, it would be relevant to use this infrastructure
> to address the tail latency of memcg oom killer, which AFAIU is now a
> scan in the order of num_possible_cpus * nb processes. Although not so
> bad when the OOM killer fires due to a machine memory use ballooning out
> of proportions, it's rather inconvenient to cause large latencies when
> it's used to apply the mem cgroup limits, which is a rather more common
> routine scenario nowadays.

Perhaps Shakeel might have an interest from the memcg side here? :)

Also similarly Johannes and Michal/Roman too perhaps?

>
> Thanks,
>
> Mathieu
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com

--
Cheers, Lorenzo