Re: [PATCH 0/2] Support hugetlb charge moving at task migration

[Baolin Wang]

On 2021/10/8 15:12, Michal Hocko wrote:
> On Thu 07-10-21 23:39:15, Baolin Wang wrote:
> > Hi Michal,
> >
> > (Sorry for late reply due to my holidays)
> > On 2021/9/30 18:46, Michal Hocko wrote:
> > > On Wed 29-09-21 18:19:26, Baolin Wang wrote:
> > > > Hi,
> > > >
> > > > Now in the hugetlb cgroup, charges associated with a task aren't moved
> > > > to the new hugetlb cgroup at task migration, which is odd for hugetlb
> > > > cgroup usage.
> > >
> > > Could you elaborate some more about the usecase and/or problems you see
> > > with the existing semantic?
> >
> > The problems is that, it did not check if the tasks can move to the new
> > hugetlb cgroup if the new hugetlb cgroup has a limitation, and the hugetlb
> > cgroup usage is incorrect when moving tasks among hugetlb cgroups.
>
> Could you be more specific please? What do you mean by cgroup usage is
> incorrect? Ideally could you describe your usecase?

Sorry for confusing, what I mean is, when tasks from one hugetlb cgroup 
are migrated to a new hugetlb cgroup, the new hugetlb cgroup's hugetlb 
page usage is not increased accordingly. The issue I found is just from 
my testing for the hugetlb cgroup, and I think this is not resonable if 
we've already set a hugetlb limitation for a cgroup, but we always 
ignore it when tasks migration among hugetlb cgroups.

> > > > This patch set adds hugetlb cgroup charge moving when
> > > > migrate tasks among cgroups, which are based on the memcg charge moving.
> > >
> > > Memcg charge moving has shown some problems over time and hence this is
> > > not part of cgroup v2 interface anymore. Even for cgroup v1 this has
> >
> > Sorry, I missed this part, could you elaborate on the issues? I can have a
> > close look about the problems of memcg charge moving.
>
> The operation is quite expensive. Synchronization with charging is not
> trivial. I do not have the full list handy but you can search the
> mm mailing list archives for more information.

Sure, thanks.

> > > been an opt-in. I do not see anything like that in this patch series.
> > > Why should all existing workloads follow a different semantic during
> > > task migration now?
> >
> > But I think it is reasonable for some cases moving the old charging to the
> > new cgroup when task migration. Maybe I can add a new hugetlb cgroup file to
> > control if need this or not?
>
> It would be good to describe those use cases and why they really need
> this. Because as things stand now, the charge migration is not supported
> in cgroup v2 for memory cgroup controller and there are no plans to add
> the support so it would be quite unexpected that hugetlb controller
> would behave differently. And cgroup v1 is considered legacy and new
> features are ususally not added as there is a hope to move users to v2.

OK, understood. Seems you have a strong opinion that it is unnecessary 
to introduce this feature for cgroup v1 now, then I will drop this patch 
set. Thanks for your input.