Re: [External] Re: [RFC PATCH 00/15] Use obj_cgroup APIs to charge the LRU pages

[Johannes Weiner]

On Thu, Apr 01, 2021 at 10:15:45AM -0700, Shakeel Butt wrote:
> On Thu, Apr 1, 2021 at 9:08 AM Muchun Song <songmuchun@xxxxxxxxxxxxx> wrote:
> >
> [...]
> > > The zombie issue is a pretty urgent concern that has caused several
> > > production emergencies now. It needs a fix sooner rather than later.
> >
> > Thank you very much for clarifying the problem for me. I do agree
> > with you. This issue should be fixed ASAP. Using objcg is a good
> > choice. Dying objcg should not be a problem. Because the size of
> > objcg is so small compared to memcg.
> >
> 
> Just wanted to say out loud that yes this patchset will reduce the
> memcg zombie issue but this is not the final destination. We should
> continue the discussions on sharing/reusing scenarios.

Absolutely. I think it's an important discussion to have.

My only concern is that Muchun's patches fix a regression, which
admittedly has built over a few years, but is a regression nonetheless
that can leave machines in undesirable states and may require reboots.

The sharing and reuse semantics on the other hand have been the same
since the beginning of cgroups. Users have had some time to structure
their requirements around these semantics :-)

If there were a concrete proposal or an RFC on the table for how
sharing and reusing could be implemented, and this proposal would be
in direct conflict with the reparenting patches, I would say let's try
to figure out a way whether the bug could be fixed in a way that is
compatible with such another imminent change.

But we shouldn't hold up a bug fix to start planning a new feature.