Re: [PATCH v4 RESEND 0/5] mm: restore per-memcg reclaim for NONSLAB shrinkers under nokmem
From: Andrew Morton
Date: Fri Nov 20 2026 - 22:11:54 EST
On Fri, 20 Nov 2026 15:04:37 +0800 Qinyun Tan <qinyuntan@xxxxxxxxxxxxxxxxx> wrote:
>
> >> objects of all cgroups share one list per node: the per-memcg
> >> shrinker bit is only set for whichever memcg happens to repopulate
> >> the empty list, so pressure in other cgroups may not even trigger
> >> the scan, and when it does, the scan walks everyone's objects.
> >
> > Thanks. Sashiko still has significant-sounding things to say:
> >
> > https://sashiko.dev/#/patchset/20261105092233.3558950-1-qinyuntan@xxxxxxxxxxxxxxxxx
> >
> > It also says "failed" for reasons I don't understand and it doesn't
> > seem to have emailed anyone. Please check it out?
>
>
> Hi Andrew,
>
> Sorry it's taken me this long to get back to you. I've had a very busy
> couple of weeks.
>
> Thanks for pointing me to the report. I looked into the two issues
> Sashiko flagged in patch 2/5.
>
> Sashiko pointed out that mem_cgroup_id() identifies the cgroup
> directory. We reproduced the ID collision on v4: after disabling and
> re-enabling the memory controller on the same cgroup, the old and new
> memcgs had the same mem_cgroup_id(). Once the new memcg has its own
> LRU list, a lookup for the old memcg can therefore find the new list.
>
> I think memcg->css.id would be a better fit for the list_lru index.
> Each per-memcg list belongs to a particular memory CSS, and its ID can
> only be reused after the last reference to that CSS is dropped and an
> RCU grace period has elapsed. That keeps the indices distinct for as
> long as the old CSS can still be accessed.
Yes, it sounds like ->css.id better matches the lifetime needs. And
the root memcg list_lru retains index=0?
> The second finding is the same race that Shakeel's patch [1] addresses.
> It predates this series: a folio can move to the root memcg between
> the two lookups in folio_memcg_list_lru_alloc(). Shakeel's patch
> rechecks whether allocation is needed after taking the memcg reference.
> That also prevents the unnecessary allocation at index 0 that Sashiko
> flagged in v4.
>
> Does using the CSS ID for the list_lru index sound reasonable? I'd be
> happy to incorporate any suggestions and send a v5.
Shakeel has a couple of related fixes in mm-usable ("mm: list_lru: fix
NULL deref when a folio moves to the root memcg" and "mm: list_lru:
don't use a new memcg's lists for an offline memcg") so please base on
that.