Re: [PATCH bpf-next v1 5/5] bpf: add a selftest for cgroup hierarchical stats collection

[Yosry Ahmed]

On Mon, Jun 6, 2022 at 5:32 AM Michal Koutný <mkoutny@xxxxxxxx> wrote:
>
> On Fri, Jun 03, 2022 at 12:52:27PM -0700, Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> > Good catch. I get confused between cgrp->subsys and
> > task->cgroups->subsys sometimes because of different fallback
> > behavior. IIUC cgrp->subsys should have NULL if the memory controller
> > is not enabled (no nearest ancestor fallback), and hence I can use
> > memory_subsys_enabled() that I defined just above task_memcg() to test
> > for this (I have no idea why I am not already using it here). Is my
> > understanding correct?
>
> You're correct, css_set (task->cgroups) has a css (memcg) always defined
> (be it root only (or even a css from v1 hierarchy but that should not
> relevant here)). A particular cgroup can have the css set to NULL.
>
> When I think about your stats collecting example now, task_memcg() looks
> more suitable to achieve proper hierarchical counting in the end (IOW
> you'd lose info from tasks who don't reside in memcg-enabled leaf).

I guess it depends on how userspace reasons about this, and whether or
not you want to collect stats from leaves that don't reside in a
memcg-enabled leaf. I will go through all the memcg-enabled checks and
make sure they make sense and are consistent, maybe add some comments
to make the userspace policy here clear.

>
> (It's just that task_memcg won't return NULL. Unless the kernel is
> compiled without memcg support completely, which makes me think how do
> the config-dependent values propagate to BPF programs?)

I don't know if there is a standard way to handle this, but I think
you should know the configs of your kernel when you are loading a bpf
program? In this particular case, if CONFIG_CGROUPS=0 then the bpf
programs will not even load due to lack of hook points or kfuncs won't
exist. If the CONFIG_CGROUPS=1 but CONFIG_MEMCG=0 I think everything
will work normally except that task_memcg() will always return NULL so
no stats will be collected, which makes sense. There will be some
overhead to running bpf programs that will always do nothing, but I
would argue that it's the userspace's fault here for loading bpf
programs on a non-compatible kernel.

>
> Thanks,
> Michal