Re: [PATCH] perf print-events: make is_event_supported() more robust

[Ian Rogers]

On Sat, Jan 20, 2024 at 10:27 AM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>
> On Tue, Jan 16, 2024 at 9:04 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> >
> > Currently the perf tool doesn't deteect support for extneded event types
>
> nit: s/deteect/detect/
> nit: s/extneded/extended/
>
> > on Apple M1/M2 systems, and will not auto-expand plain PERF_EVENT_TYPE
> > hardware events into per-PMU events. This is due to the detection of
> > extended event types not handling mandatory filters required by the
> > M1/M2 PMU driver.
> >
> > PMU drivers and the core perf_events code can require that
> > perf_event_attr::exclude_* filters are configured in a specific way and
> > may reject certain configurations of filters, for example:
> >
> > (a) Many PMUs lack support for any event filtering, and require all
> >     perf_event_attr::exclude_* bits to be clear. This includes Alpha's
> >     CPU PMU, and ARM CPU PMUs prior to the introduction of PMUv2 in
> >     ARMv7,
> >
> > (b) When /proc/sys/kernel/perf_event_paranoid >= 2, the perf core
> >     requires that perf_event_attr::exclude_kernel is set.
> >
> > (c) The Apple M1/M2 PMU requires that perf_event_attr::exclude_guest is
> >     set as the hardware PMU does not count while a guest is running (but
> >     might be extended in future to do so).
> >
> > In is_event_supported(), we try to account for cases (a) and (b), first
> > attempting to open an event without any filters, and if this fails,
> > retrying with perf_event_attr::exclude_kernel set. We do not account for
> > case (c), or any other filters that drivers could theoretically require
> > to be set.
> >
> > Thus is_event_supported() will fail to detect support for any events
> > targetting an Apple M1/M2 PMU, even where events would be supported with
>
> nit: s/targetting/targeting/
>
> > perf_event_attr:::exclude_guest set.
> >
> > Since commit:
> >
> >   82fe2e45cdb00de4 ("perf pmus: Check if we can encode the PMU number in perf_event_attr.type")
> >
> > ... we use is_event_supported() to detect support for extended types,
> > with the PMU ID encoded into the perf_event_attr::type. As above, on an
> > Apple M1/M2 system this will always fail to detect that the event is
> > supported, and consequently we fail to detect support for extended types
> > even when these are supported, as they have been since commit:
> >
> >   5c816728651ae425 ("arm_pmu: Add PERF_PMU_CAP_EXTENDED_HW_TYPE capability")
> >
> > Due to this, the perf tool will not automatically expand plain
> > PERF_TYPE_HARDWARE events into per-PMU events, even when all the
> > necessary kernel support is present.
> >
> > This patch updates is_event_supported() to additionally try opening
> > events with perf_event_attr::exclude_guest set, allowing support for
> > events to be detected on Apple M1/M2 systems. I beleive that this is
>
> nit: s/beleive/believe/
>
> > sufficient for all contemporary CPU PMU drivers, though in future it may
> > be necessary to check for other combinations of filter bits.
> >
> > I've deliberately changed the check to not expect a specific error code
> > for missing filters, as today ;the kernel may return a number of
> > different error codes for missing filters (e.g. -EACCESS, -EINVAL, or
> > -EOPNOTSUPP) depending on why and where the filter configuration is
> > rejected, and retrying for any error is more robust.
> >
> > Note that this does not remove the need for commit:
> >
> >   a24d9d9dc096fc0d ("perf parse-events: Make legacy events lower priority than sysfs/JSON")
> >
> > ... which is still necessary so that named-pmu/event/ events work on
> > kernels without extended type support, even if the event name happens to
> > be the same as a PERF_EVENT_TYPE_HARDWARE event (e.g. as is the case for
> > the M1/M2 PMU's 'cycles' and 'instructions' events).
> >
> > Fixes: 82fe2e45cdb00de4 ("perf pmus: Check if we can encode the PMU number in perf_event_attr.type")
> > Signed-off-by: Mark Rutland <mark.rutland@xxxxxxx>
> > Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> > Cc: Hector Martin <marcan@xxxxxxxxx>
> > Cc: Ian Rogers <irogers@xxxxxxxxxx>
> > Cc: James Clark <james.clark@xxxxxxx>
> > Cc: John Garry <john.g.garry@xxxxxxxxxx>
> > Cc: Leo Yan <leo.yan@xxxxxxxxxx>
> > Cc: Marc Zyngier <maz@xxxxxxxxxx>
> > Cc: Mike Leach <mike.leach@xxxxxxxxxx>
> > Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
> > Cc: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > Cc: Thomas Richter <tmricht@xxxxxxxxxxxxx>
> > Cc: Will Deacon <will@xxxxxxxxxx>

Also:
Tested-by: Ian Rogers <irogers@xxxxxxxxxx>

No regressions on Alderlake except a pre-existing problem I noted here:
https://lore.kernel.org/lkml/CAP-5=fWVQ-7ijjK3-w1q+k2WYVNHbAcejb-xY0ptbjRw476VKA@xxxxxxxxxxxxxx/

Thanks,
Ian

> > ---
> >  tools/perf/util/print-events.c | 27 +++++++++++++++++++--------
> >  1 file changed, 19 insertions(+), 8 deletions(-)
> >
> > Hector, Marc, I'd appreciate if either of you could give this a spin on
> > your M1/M2 machines. I've given it local testing with the arm_pmuv3
> > driver modified to behave the same as the apple_m1_pmu driver (requiring
> > exclude_guest, having a 'cycles' event in sysfs), but that might not
> > perfectly replicate your setup.
> >
> > The patch is based on the 'perf-tools-for-v6.8-1-2024-01-09' tag in the
> > perf-tools tree:
> >
> >   https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools.git/
> >
> > ... and I've pushed it out to the 'perf-tools/event-supported-filters'
> > branch in my tree:
> >
> >   https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/
> >
> > This patch *should* make it possible to do:
> >
> >         perf stat -e cycles ./workload
> >         perf stat -e instructions ./workload
> >
> > ... with those 'cycles' and 'instructions' events being automatically
> > expanded and reported as separate events per-PMU, which is a nice
> > quality-of-life improvement.
> >
> > Comparing before and after this patch:
> >
> > | # ./perf-before stat -e cycles true
> > |
> > |  Performance counter stats for 'true':
> > |
> > |      <not counted>      cycles                                                                  (0.00%)
> > |
> > |        0.000990250 seconds time elapsed
> > |
> > |        0.000934000 seconds user
> > |        0.000000000 seconds sys
> > |
> > | # ./perf-after stat -e cycles true
> > |
> > |  Performance counter stats for 'true':
> > |
> > |             965175      armv8_pmuv3_0/cycles/
> > |      <not counted>      armv8_pmuv3_1/cycles/                                                   (0.00%)
> > |      <not counted>      armv8_pmuv3_2/cycles/                                                   (0.00%)
> > |      <not counted>      armv8_pmuv3_3/cycles/                                                   (0.00%)
> > |
> > |        0.000836555 seconds time elapsed
> > |
> > |        0.000884000 seconds user
> > |        0.000000000 seconds sys
>
> Just to check, this is the expected expansion of cycles? I'm unclear
> why 4 core PMUs.
>
> >
> > This *shouldn't* change the interpetation of named-pmu events, e.g.
> >
> >         perf stat -e apple_whichever_pmu/cycles/ ./workload
> >
> > ... should behave the same as without this patch
> >
> > Comparing before and after this patch:
> >
> > | # ./perf-before stat -e armv8_pmuv3_0/cycles/ -e armv8_pmuv3_1/cycles/ -e armv8_pmuv3_2/cycles/ -e armv8_pmuv3_3/cycles/ true
> > |
> > |  Performance counter stats for 'true':
> > |
> > |      <not counted>      armv8_pmuv3_0/cycles/                                                   (0.00%)
> > |      <not counted>      armv8_pmuv3_1/cycles/                                                   (0.00%)
> > |      <not counted>      armv8_pmuv3_2/cycles/                                                   (0.00%)
> > |             901415      armv8_pmuv3_3/cycles/
> > |
> > |        0.000756590 seconds time elapsed
> > |
> > |        0.000811000 seconds user
> > |        0.000000000 seconds sys
> > |
> > | # ./perf-after stat -e armv8_pmuv3_0/cycles/ -e armv8_pmuv3_1/cycles/ -e armv8_pmuv3_2/cycles/ -e armv8_pmuv3_3/cycles/ true
> > |
> > |  Performance counter stats for 'true':
> > |
> > |             923314      armv8_pmuv3_0/cycles/
> > |      <not counted>      armv8_pmuv3_1/cycles/                                                   (0.00%)
> > |      <not counted>      armv8_pmuv3_2/cycles/                                                   (0.00%)
> > |      <not counted>      armv8_pmuv3_3/cycles/                                                   (0.00%)
> > |
> > |        0.000782420 seconds time elapsed
> > |
> > |        0.000836000 seconds user
> > |        0.000000000 seconds sys
> >
> > One thing I'm still looing into is that this doesn't seem to do anything
> > for a default perf stat session, e.g.
> >
> >         perf stat ./workload
> >
> > ... doesn't automatically expand the implicitly-created events into per-pmu
> > events.
>
> Ugh, weak symbols. x86 has overridden the default adding of attributes
> to do it for hybrid:
> https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/arch/x86/util/evlist.c?h=perf-tools-next#n36
> I think we should lose the adding events via attributes and just go to
> using parse events for everything. I'll see if I can do some cleanup
> and that should resolve this - I also want to cleanup the default
> events/metrics and the detailed ones as we can drop the unsupported
> events, etc.
>
> > Comparing before and after this patch:
> >
> > | # ./perf-before stat true
> > |
> > |  Performance counter stats for 'true':
> > |
> > |               0.42 msec task-clock                       #    0.569 CPUs utilized
> > |                  0      context-switches                 #    0.000 /sec
> > |                  0      cpu-migrations                   #    0.000 /sec
> > |                 38      page-faults                      #   89.796 K/sec
> > |      <not counted>      cycles                                                                  (0.00%)
> > |      <not counted>      instructions                                                            (0.00%)
> > |      <not counted>      branches                                                                (0.00%)
> > |      <not counted>      branch-misses                                                           (0.00%)
> > |
> > |        0.000744185 seconds time elapsed
> > |
> > |        0.000795000 seconds user
> > |        0.000000000 seconds sys
> > |
> > | # ./perf-after stat true
> > |
> > |  Performance counter stats for 'true':
> > |
> > |               0.43 msec task-clock                       #    0.582 CPUs utilized
> > |                  0      context-switches                 #    0.000 /sec
> > |                  0      cpu-migrations                   #    0.000 /sec
> > |                 38      page-faults                      #   88.960 K/sec
> > |      <not counted>      cycles                                                                  (0.00%)
> > |      <not counted>      instructions                                                            (0.00%)
> > |      <not counted>      branches                                                                (0.00%)
> > |      <not counted>      branch-misses                                                           (0.00%)
> > |
> > |        0.000734120 seconds time elapsed
> > |
> > |        0.000786000 seconds user
> > |        0.000000000 seconds sys
> >
> > Ian, how does that behave on x86? Is that the same, or do the default
> > events get expanded?
>
> The default events are expanded, the not counted is a feature of a
> fast binary (true here). I'm trying to remove custom code paths so
> that things like this don't happen, sorry that you've come across
> another instance but at least I can fix it.
>
> Thanks,
> Ian
>
> > Thanks,
> > Mark.
> >
> > diff --git a/tools/perf/util/print-events.c b/tools/perf/util/print-events.c
> > index b0fc48be623f3..4f67e8f00a4d6 100644
> > --- a/tools/perf/util/print-events.c
> > +++ b/tools/perf/util/print-events.c
> > @@ -232,7 +232,6 @@ void print_sdt_events(const struct print_callbacks *print_cb, void *print_state)
> >  bool is_event_supported(u8 type, u64 config)
> >  {
> >         bool ret = true;
> > -       int open_return;
> >         struct evsel *evsel;
> >         struct perf_event_attr attr = {
> >                 .type = type,
> > @@ -246,20 +245,32 @@ bool is_event_supported(u8 type, u64 config)
> >
> >         evsel = evsel__new(&attr);
> >         if (evsel) {
> > -               open_return = evsel__open(evsel, NULL, tmap);
> > -               ret = open_return >= 0;
> > +               ret = evsel__open(evsel, NULL, tmap) >= 0;
> >
> > -               if (open_return == -EACCES) {
> > +               if (!ret) {
> >                         /*
> > -                        * This happens if the paranoid value
> > +                        * The event may fail to open if the paranoid value
> >                          * /proc/sys/kernel/perf_event_paranoid is set to 2
> > -                        * Re-run with exclude_kernel set; we don't do that
> > -                        * by default as some ARM machines do not support it.
> > -                        *
> > +                        * Re-run with exclude_kernel set; we don't do that by
> > +                        * default as some ARM machines do not support it.
> >                          */
> >                         evsel->core.attr.exclude_kernel = 1;
> >                         ret = evsel__open(evsel, NULL, tmap) >= 0;
> >                 }
> > +
> > +               if (!ret) {
> > +                       /*
> > +                        * The event may fail to open if the PMU requires
> > +                        * exclude_guest to be set (e.g. as the Apple M1 PMU
> > +                        * requires).
> > +                        * Re-run with exclude_guest set; we don't do that by
> > +                        * default as it's equally legitimate for another PMU
> > +                        * driver to require that exclude_guest is clear.
> > +                        */
> > +                       evsel->core.attr.exclude_guest = 1;
> > +                       ret = evsel__open(evsel, NULL, tmap) >= 0;
> > +               }
> > +
> >                 evsel__delete(evsel);
> >         }
> >
> > --
> > 2.30.2
> >