[PATCH RFC 4/4] bpf,cgroup,perf: extend bpf-cgroup to support tracepoint attachment

[Kenny Ho]

On Thu, Nov 18, 2021 at 11:33 PM Alexei Starovoitov
<alexei.starovoitov at gmail.com> wrote:
>
> On Thu, Nov 18, 2021 at 03:28:40PM -0500, Kenny Ho wrote:
> > +     for_each_possible_cpu(cpu) {
> > +             /* allocate first, connect the cgroup later */
> > +             events[i] = perf_event_create_kernel_counter(attr, cpu, NULL, NULL, NULL);
>
> This is a very heavy hammer for this task.
> There is really no need for perf_event to be created.
> Did you consider using raw_tp approach instead?

I came across raw_tp but I don't have a good understanding of it yet.
Initially I was hoping perf event/tracepoint is a stepping stone to
raw tp but that doesn't seem to be the case (and unfortunately I
picked the perf event/tracepoint route to dive in first because I saw
cgroup usage.)  Can you confirm if the following statements are true?

- is raw_tp related to writable tracepoint
- are perf_event/tracepoint/kprobe/uprobe and fentry/fexit/raw_tp
considered two separate 'things' (even though both of their purpose is
tracing)?

> It doesn't need this heavy stuff.
> Also I suspect in follow up you'd be adding tracepoints to GPU code?
> Did you consider just leaving few __weak global functions in GPU code
> and let bpf progs attach to them as fentry?
There are already tracepoints in the GPU code.  And I do like fentry
way of doing things more but my head was very much focused on cgroup,
and tracepoint/kprobe path seems to have something for it.  I
suspected this would be a bit too heavy after seeing the scalability
discussion but I wasn't sure so I whip this up quickly to get some
feedback (while learning more about perf/bpf/cgroup.)

> I suspect the true hierarchical nature of bpf-cgroup framework isn't necessary.
> The bpf program itself can filter for given cgroup.
> We have bpf_current_task_under_cgroup() and friends.
Is there a way to access cgroup local storage from a prog that is not
attached to a bpf-cgroup?  Although, I guess I can just store/read
things using a map with the cg id as key.  And with the
bpf_get_current_ancestor_cgroup_id below I can just simulate the
values being propagated if the hierarchy ends up being relevant.  Then
again, is there a way to atomically update multiple elements of a map?
 I am trying to figure out how to support a multi-user multi-app
sharing use case (like user A given quota X and user B given quota Y
with app 1 and 2 each having a quota assigned by A and app 8 and 9
each having quota assigned by B.)  Is there some kind of 'lock'
mechanism for me to keep quota 1,2,X in sync? (Same for 8,9,Y.)

> I suggest to sprinkle __weak empty funcs in GPU and see what
> you can do with it with fentry and bpf_current_task_under_cgroup.
> There is also bpf_get_current_ancestor_cgroup_id().