Re: [PATCH bpf-next v1 3/5] bpf: Introduce cgroup iter

[Yonghong Song]

On 5/20/22 1:11 AM, Tejun Heo wrote:
> Hello,
>
> On Fri, May 20, 2022 at 12:58:52AM -0700, Yosry Ahmed wrote:
> > On Fri, May 20, 2022 at 12:41 AM Tejun Heo <tj@xxxxxxxxxx> wrote:
> > > On Fri, May 20, 2022 at 01:21:31AM +0000, Yosry Ahmed wrote:
> > > > From: Hao Luo <haoluo@xxxxxxxxxx>
> > > >
> > > > Introduce a new type of iter prog: cgroup. Unlike other bpf_iter, this
> > > > iter doesn't iterate a set of kernel objects. Instead, it is supposed to
> > > > be parameterized by a cgroup id and prints only that cgroup. So one
> > > > needs to specify a target cgroup id when attaching this iter. The target
> > > > cgroup's state can be read out via a link of this iter.
> > > >
> > > > Signed-off-by: Hao Luo <haoluo@xxxxxxxxxx>
> > > > Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > >
> > > This could be me not understanding why it's structured this way but it keeps
> > > bothering me that this is adding a cgroup iterator which doesn't iterate
> > > cgroups. If all that's needed is extracting information from a specific
> > > cgroup, why does this need to be an iterator? e.g. why can't I use
> > > BPF_PROG_TEST_RUN which looks up the cgroup with the provided ID, flushes
> > > rstat, retrieves whatever information necessary and returns that as the
> > > result?
> >
> > I will let Hao and Yonghong reply here as they have a lot more
> > context, and they had previous discussions about cgroup_iter. I just
> > want to say that exposing the stats in a file is extremely convenient
> > for userspace apps. It becomes very similar to reading stats from
> > cgroupfs. It also makes migrating cgroup stats that we have
> > implemented in the kernel to BPF a lot easier.
>
> So, if it were upto me, I'd rather direct energy towards making retrieving
> information through TEST_RUN_PROG easier rather than clinging to making
> kernel output text. I get that text interface is familiar but it kinda
> sucks in many ways.
>
> > AFAIK there are also discussions about using overlayfs to have links
> > to the bpffs files in cgroupfs, which makes it even better. So I would
> > really prefer keeping the approach we have here of reading stats
> > through a file from userspace. As for how we go about this (and why a
> > cgroup iterator doesn't iterate cgroups) I will leave this for Hao and
> > Yonghong to explain the rationale behind it. Ideally we can keep the
> > same functionality under a more descriptive name/type.
>
> My answer would be the same here. You guys seem dead set on making the
> kernel emulate cgroup1. I'm not gonna explicitly block that but would
> strongly suggest having a longer term view.
>
> If you *must* do the iterator, can you at least make it a proper iterator
> which supports seeking? AFAICS there's nothing fundamentally preventing bpf
> iterators from supporting seeking. Or is it that you need something which is
> pinned to a cgroup so that you can emulate the directory structure?

The current bpf_iter for cgroup is for the google use case
per previous discussion. But I think a generic cgroup bpf iterator
should help as well.

Maybe you can have a bpf program signature like below:

int BPF_PROG(dump_vmscan, struct bpf_iter_meta *meta, struct cgroup 
*cgrp, struct cgroup *parent_cgrp)

parent_cgrp is NULL when cgrp is the root cgroup.

I would like the bpf program should send the following information to
user space:
   <parent cgroup dir name> <current cgroup dir name>
   <various stats interested by the user>

This way, user space can easily construct the cgroup hierarchy stat like
                           cpu   mem   cpu pressure   mem pressure ...
   cgroup1                 ...
      child1               ...
        grandchild1        ...
      child2               ...
   cgroup 2                ...
      child 3              ...
        ...                ...

the bpf iterator can have additional parameter like
cgroup_id = ... to only call bpf program once with that
cgroup_id if specified.

The kernel part of cgroup_iter can call cgroup_rstat_flush()
before calling cgroup_iter bpf program.

WDYT?

> Thanks.