Re: [PATCH v2 1/2] hugetlb: memcg: account hugetlb-backed memory in memory controller

[Michal Hocko <mhocko@suse.com>]

On Fri 29-09-23 13:42:21, Johannes Weiner wrote:
> On Fri, Sep 29, 2023 at 08:11:54AM -0700, Yosry Ahmed wrote:
> > On Fri, Sep 29, 2023 at 8:08 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
> > >
> > > On Thu, Sep 28, 2023 at 06:18:19PM -0700, Yosry Ahmed wrote:
> > > > My concern is the scenario where the memory controller is mounted in
> > > > cgroup v1, and cgroup v2 is mounted with memory_hugetlb_accounting.
> > > >
> > > > In this case it seems like the current code will only check whether
> > > > memory_hugetlb_accounting was set on cgroup v2 or not, disregarding
> > > > the fact that cgroup v1 did not enable hugetlb accounting.
> > > >
> > > > I obviously prefer that any features are also added to cgroup v1,
> > > > because we still didn't make it to cgroup v2, especially when the
> > > > infrastructure is shared. On the other hand, I am pretty sure the
> > > > maintainers will not like what I am saying :)
> > >
> > > I have a weak preference.
> > >
> > > It's definitely a little weird that the v1 controller's behavior
> > > changes based on the v2 mount flag. And that if you want it as an
> > > otherwise exclusive v1 user, you'd have to mount a dummy v2.
> > >
> > > But I also don't see a scenario where it would hurt, or where there
> > > would be an unresolvable conflict between v1 and v2 in expressing
> > > desired behavior, since the memory controller is exclusive to one.
> > >
> > > While we could eliminate this quirk with a simple
> > > !cgroup_subsys_on_dfl(memory_cgrp_subsys) inside the charge function,
> > > it would seem almost punitive to add extra code just to take something
> > > away that isn't really a problem and could be useful to some people.
> > >
> > > If Tejun doesn't object, I'd say let's just keep implied v1 behavior.
> > 
> > I agree that adding extra code to take a feature away from v1 is
> > probably too much, but I also think relying on a v2 mount option is
> > weird. Would it be too much to just have a v1-specific flag as well
> > and use cgroup_subsys_on_dfl(memory_cgrp_subsys) to decide which flag
> > to read?
> 
> Yeah, let's not preemptively add explicit new features to cgroup1.
> 
> Since we agree the incidental support is weird, let's filter hugetlb
> charging on cgroup_subsys_on_dfl(memory_cgrp_subsys) after all.

Agreed. It would be a bad idea to have an implicit behavior change based
on v2 mounting options. And I really do not think we want to add this
feature to v1. I am not supper thrilled about enabling this for v2 to be
completely honest but I do see a demand so I will not object to that.
-- 
Michal Hocko
SUSE Labs