Re: [PATCH v5 28/34] perf pmus: Split pmus list into core and other

[Ian Rogers]

On Fri, Jun 9, 2023 at 12:59 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
>
> On Thu, Jun 08, 2023 at 10:35:02PM -0700, Ian Rogers wrote:
> > On Thu, Jun 8, 2023 at 10:30 PM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
> > >
> > > On 09-Jun-23 10:10 AM, Ian Rogers wrote:
> > > > On Thu, Jun 8, 2023 at 9:01 PM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
> > > >>
> > > >> Hi Ian,
> > > >
> > > > Hi Ravi,
> > > >
> > > >> On 27-May-23 12:52 PM, Ian Rogers wrote:
> > > >>> Split the pmus list into core and other. This will later allow for
> > > >>> the core and other pmus to be populated separately.
> > > >>>
> > > >>> Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> > > >>> Reviewed-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > > >>> ---
> > > >>>  tools/perf/util/pmus.c | 52 ++++++++++++++++++++++++++++++------------
> > > >>>  1 file changed, 38 insertions(+), 14 deletions(-)
> > > >>>
> > > >>> diff --git a/tools/perf/util/pmus.c b/tools/perf/util/pmus.c
> > > >>> index 58ff7937e9b7..4ef4fecd335f 100644
> > > >>> --- a/tools/perf/util/pmus.c
> > > >>> +++ b/tools/perf/util/pmus.c
> > > >>> @@ -12,13 +12,19 @@
> > > >>>  #include "pmu.h"
> > > >>>  #include "print-events.h"
> > > >>>
> > > >>> -static LIST_HEAD(pmus);
> > > >>> +static LIST_HEAD(core_pmus);
> > > >>> +static LIST_HEAD(other_pmus);
> > > >>
> > > >> AMD ibs_fetch// and ibs_op// PMUs are per SMT-thread and are independent of
> > > >> core hw pmu. I wonder where does IBS fit. Currently it's part of other_pmus.
> > > >> So, is it safe to assume that other_pmus are not just uncore pmus? In that
> > > >> case shall we add a comment here?
> > > >
> > > > I'm a fan of comments. The code has landed in perf-tools-next:
> > > > https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/pmus.c?h=perf-tools-next
> > > > Do you have any suggestions on wording? I've had limited success
> > > > adding glossary terms, for example, offcore vs uncore:
> > > > https://perf.wiki.kernel.org/index.php/Glossary#Offcore
> > > > I think offcore is a more interconnect related term, but I'd prefer
> > > > not to be inventing the definitions. I'd like it if we could be less
> > > > ambiguous in the code and provide useful information on the wiki, so
> > > > help appreciated :-)
> > >
> > > Does this look good?
> > >
> > > /*
> > >  * core_pmus:  A PMU belongs to core_pmus if it's name is "cpu" or it's sysfs
> > >  *             directory contains "cpus" file. All PMUs belonging to core_pmus
> > >  *             must have pmu->is_core=1. If there are more than one PMUs in
> > >  *             this list, perf interprets it as a heterogeneous platform.
> >
> >
> > Looks good but a nit here. It is heterogeneous from point-of-view of
> > PMUs, there are ARM systems where they are heterogenous with big and
> > little cores but they have a single homogeneous PMU driver. The perf
> > tool will treat them as homogeneous.
>
> For the sake of the comment: there's a little more nuance here.
>
> The intent is that each distinct micro-architecture has its own PMU instance,
> but some people write their device trees incorrectly with a single pmu node
> rather than separate pmu nodes per micro-architecture.
>
> That should be viewed as a FW bug, even if we have to deal with it here.

Thanks for the clarification Mark. For heterogeneous ARM I was
primarily looking at a Pixel 4 phone, which has a homogeneous PMU. The
normal way to make sure Android configurations are sensible is to have
a CTS test. Would that be appropriate here?

Given Intel contributed the original heterogeneous PMU support to the
perf tool, and hard coded the PMU names to 'cpu_core' and 'cpu_atom',
are there any correctness tests that exist for ARM heterogeneous PMUs?
Could we make this part of the 'perf test' command? Tests acting both
as a correctness feature and documentation.

Thanks,
Ian

> Thanks,
> Mark.