Re: [PATCH v3 4/5] mm: memcg: charge memcg percpu memory to the parent cgroup

[Roman Gushchin <guro@fb.com>]

On Tue, Aug 11, 2020 at 04:47:54PM +0200, Michal Koutny wrote:
> On Thu, Aug 06, 2020 at 09:37:17PM -0700, Roman Gushchin <guro@fb.com> wrote:
> > In general, yes. But in this case I think it wouldn't be a good idea:
> > most often cgroups are created by a centralized daemon (systemd),
> > which is usually located in the root cgroup. Even if it's located not in
> > the root cgroup, limiting it's memory will likely affect the whole system,
> > even if only one specific limit was reached.
> The generic scheme would be (assuming the no internal process
> constraint, in the root too)
> 
> ` root or delegated root
>   ` manager-cgroup (systemd, docker, ...)
>   ` [aggregation group(s)]
>     ` job-group-1
>     ` ...
>     ` job-group-n
> 
> > If there is a containerized workload, which creates sub-cgroups,
> > charging it's parent cgroup is perfectly effective.
> No dispute about this in either approaches.
> 
> > And the opposite, if we'll charge the cgroup of a process, who created
> > a cgroup, we'll not cover the most common case: systemd creating
> > cgroups for all services in the system.
> What I mean is that systemd should be charged for the cgroup creation.
> Or more generally, any container/cgroup manager should be charged.
> Consider a leak when it wouldn't remove spent cgroups, IMO the effect
> (throttling, reclaim) should be exercised on such a culprit.

As I said, there are 2 problems with charging systemd (or a similar daemon):
1) It often belongs to the root cgroup.
2) OOMing or failing some random memory allocations is a bad way
   to "communicate" a memory shortage to the daemon.
   What we really want is to prevent creating a huge number of cgroups
   (including dying cgroups) in some specific sub-tree(s).
   OOMing the daemon or returning -ENOMEM to some random syscalls
   will not help us to reach the goal and likely will bring a bad
   experience to a user.

In a generic case I don't see how we can charge the cgroup which
creates cgroups without solving these problems first.

And if there is a very special case where we have to limit it,
we can just add an additional layer:

` root or delegated root
   ` manager-parent-cgroup-with-a-limit
     ` manager-cgroup (systemd, docker, ...)
   ` [aggregation group(s)]
     ` job-group-1
     ` ...
     ` job-group-n

> 
> I don't think the existing workload (job-group-i above) should
> unnecessarily suffer when only manager is acting up. Is that different
> from your idea?
> 
> > Right, but it's quite unusual for tasks from one cgroup to create sub-cgroups
> > in completely different cgroup. In this particular case there are tons of other
> > ways how a task from C00 can hurt C1.
> I agree with that.
> 
> 
> If I haven't overlooked anything, this should be first case when
> cgroup-related structures are accounted (please correct me).
> So this is setting a precendent, if others show useful to be accounted
> in the future too.

Right.

> I'm thinking about cpu_cgroup_css_alloc() that can
> also allocate a lot (with big CPU count). The current approach would lead
> situations where matching cpu and memory csses needn't to exist and that
> would need special handling.

I'd definitely charge the parent cgroup in all similar cases.

> 
> 
> > On Thu, Aug 06, 2020 at 09:16:03PM -0700, Andrew Morton wrote:
> > > These week-old issues appear to be significant.  Roman?  Or someone
> > > else?
> Despite my concerns, I don't think this is fundamental and can't be
> changed later so it doesn't prevent the inclusion in 5.9 RC1.

Thank you!