Re: [PATCH v4 3/4] perf/core: Remove pmu linear searching code

[Ian Rogers <irogers@google.com>]

On Sat, May 27, 2023 at 10:00 AM Ian Rogers <irogers@google.com> wrote:
>
> On Sat, May 27, 2023 at 6:32 AM Marc Zyngier <maz@kernel.org> wrote:
> >
> > On Sat, 27 May 2023 00:00:47 +0100,
> > Ian Rogers <irogers@google.com> wrote:
> > >
> > > On Thu, May 25, 2023 at 8:56 AM Oliver Upton <oliver.upton@linux.dev> wrote:
> > > >
> > > > On Thu, May 25, 2023 at 04:20:31PM +0200, Peter Zijlstra wrote:
> > > > > On Thu, May 25, 2023 at 07:11:41AM +0000, Oliver Upton wrote:
> > > > >
> > > > > > The PMUv3 driver does pass a name, but it relies on getting back an
> > > > > > allocated pmu id as @type is -1 in the call to perf_pmu_register().
> > > > > >
> > > > > > What actually broke is how KVM probes for a default core PMU to use for
> > > > > > a guest. kvm_pmu_probe_armpmu() creates a counter w/ PERF_TYPE_RAW and
> > > > > > reads the pmu from the returned perf_event. The linear search had the
> > > > > > effect of eventually stumbling on the correct core PMU and succeeding.
> > > > > >
> > > > > > Perf folks: is this WAI for heterogenous systems?
> > > > >
> > > > > TBH, I'm not sure. hetero and virt don't mix very well AFAIK and I'm not
> > > > > sure what ARM64 does here.
> > > > >
> > > > > IIRC the only way is to hard affine things; that is, force vCPU of
> > > > > 'type' to the pCPU mask of 'type' CPUs.
> > > >
> > > > We provide absolutely no illusion of consistency across implementations.
> > > > Userspace can select the PMU type, and then it is a userspace problem
> > > > affining vCPUs to the right pCPUs.
> > > >
> > > > And if they get that wrong, we just bail and refuse to run the vCPU.
> > > >
> > > > > If you don't do that; or let userspace 'override' that, things go
> > > > > sideways *real* fast.
> > > >
> > > > Oh yeah, and I wish PMUs were the only problem with these hetero
> > > > systems...
> > >
> > > Just to add some context from what I understand. There are inbuilt
> > > type numbers for PMUs:
> > > https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/include/uapi/linux/perf_event.h?h=perf-tools-next#n34
> > > so the PMU generally called /sys/devices/cpu should have type 4 (ARM
> > > give it another name). For heterogeneous ARM there is a single PMU and
> > > the same events are programmed regardless of whether it is a big or a
> > > little core - the cpumask lists all CPUs.
> >
> > I think you misunderstood the way heterogeneous arm64 systems are
> > described . Each CPU type gets its own PMU type, and its own event
> > list. Case in point:
> >
> > $ grep . /sys/devices/*pmu/{type,cpus}
> > /sys/devices/apple_avalanche_pmu/type:9
> > /sys/devices/apple_blizzard_pmu/type:8
> > /sys/devices/apple_avalanche_pmu/cpus:4-9
> > /sys/devices/apple_blizzard_pmu/cpus:0-3
> >
> > Type 4 (aka PERF_EVENT_RAW) is AFAICT just a way to encode the raw
> > event number, nothing else.
>
> Which PMU will a raw event open on? Note, the raw events don't support
> the extended type that is present in PERF_TYPE_HARDWARE and
> PERF_TYPE_HW_CACHE:
> https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/tree/include/uapi/linux/perf_event.h#n41
> as the bits are already in use for being just plain config values. I
> suspect not being type 4 is a bug on apple ARM here.
>
> Thanks,
> Ian

Btw, thanks for correcting me! :-D These assumptions are under
documented/tested and nobody wants to make interfaces/tools that are
broken. One source of information is:
https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/tree/Documentation/admin-guide/perf
but it doesn't capture this.

Thanks,
Ian

> > Thanks,
> >
> >         M.
> >
> > --
> > Without deviation from the norm, progress is not possible.