Re: [PATCH v3 1/2] perf/core: Share an event with multiple cgroups

[Namhyung Kim]

On Fri, Apr 16, 2021 at 8:59 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Fri, Apr 16, 2021 at 08:22:38PM +0900, Namhyung Kim wrote:
> > On Fri, Apr 16, 2021 at 7:28 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Apr 16, 2021 at 11:29:30AM +0200, Peter Zijlstra wrote:
> > >
> > > > > So I think we've had proposals for being able to close fds in the past;
> > > > > while preserving groups etc. We've always pushed back on that because of
> > > > > the resource limit issue. By having each counter be a filedesc we get a
> > > > > natural limit on the amount of resources you can consume. And in that
> > > > > respect, having to use 400k fds is things working as designed.
> > > > >
> > > > > Anyway, there might be a way around this..
> > >
> > > So how about we flip the whole thing sideways, instead of doing one
> > > event for multiple cgroups, do an event for multiple-cpus.
> > >
> > > Basically, allow:
> > >
> > >         perf_event_open(.pid=fd, cpu=-1, .flag=PID_CGROUP);
> > >
> > > Which would have the kernel create nr_cpus events [the corrolary is that
> > > we'd probably also allow: (.pid=-1, cpu=-1) ].
> >
> > Do you mean it'd have separate perf_events per cpu internally?
> > From a cpu's perspective, there's nothing changed, right?
> > Then it will have the same performance problem as of now.
>
> Yes, but we'll not end up in ioctl() hell. The interface is sooo much
> better. The performance thing just means we need to think harder.
>
> I thought cgroup scheduling got a lot better with the work Ian did a
> while back? What's the actual bottleneck now?

Yep, that's true but it still comes with a high cost of multiplexing in
context (cgroup) switch.  It's inefficient that it programs the PMU
with exactly the same config just for a different cgroup.  You know
accessing the MSRs is no cheap operation.

>
> > > Output could be done by adding FORMAT_PERCPU, which takes the current
> > > read() format and writes a copy for each CPU event. (p)read(v)() could
> > > be used to explode or partial read that.
> >
> > Yeah, I think it's good for read.  But what about mmap?
> > I don't think we can use file offset since it's taken for auxtrace.
> > Maybe we can simply disallow that..
>
> Are you actually using mmap() to read? I had a proposal for FORMAT_GROUP
> like thing for mmap(), but I never implemented that (didn't get the
> enthousiatic response I thought it would). But yeah, there's nowhere
> near enough space in there for PERCPU.

Recently there's a patch to do it with rdpmc which needs to mmap first.

https://lore.kernel.org/lkml/20210414155412.3697605-1-robh@xxxxxxxxxx/

>
> Not sure how to do that, these counters must not be sampling counters
> because we can't be sharing a buffer from multiple CPUs, so data/aux
> just isn't a concern. But it's weird to have them magically behave
> differently.

Yeah it's weird, and we should limit the sampling use case.

Thanks,
Namhyung