Re: [PATCH v2] memcg: keep folio's objcg same as its node

On Thu, Aug 06, 2026 at 03:22:50PM -0400, Johannes Weiner wrote:
> Hm, it's still not quite right.
> 
> When there is a root mismatch, we cannot fall back to committing the
> old objcg. That would reintroduce Karl's issue. Not as broadly as
> before, but can still happen if the memcg tree died up to the root.
> 
> So we have to commit to the new objcg, always.
> 
> What obj_cgroup_is_root() then comes down to is whether uncharge will
> balance the page counters or not.
> 
> mem_cgroup_replace_folio() gets a new charge for the new page. We can
> just conditionalize that right away on whether the new objcg will
> actually uncharge.
> 
> mem_cgroup_migrate() currently trades the charge, but that won't work
> if the new objcg is root and won't uncharge. So we have to settle it
> right then and there.
> 
> This?

Thanks a lot Johannes, this looks much better. I will incorporate this and run
through the reproducer before sending out the v3.