On Thu, Sep 11, 2025 at 6:00 AM James Clark <james.clark@xxxxxxxxxx> wrote:
>
>
>
> On 10/09/2025 4:00 pm, Ian Rogers wrote:
> > On Wed, Sep 10, 2025 at 4:14 AM James Clark <james.clark@xxxxxxxxxx> wrote:
> >>
> >> On 28/08/2025 9:59 pm, Ian Rogers wrote:
> >>> Mirroring similar work for software events in commit 6e9fa4131abb
> >>> ("perf parse-events: Remove non-json software events"). These changes
> >>> migrate the legacy hardware and cache events to json. With no hard
> >>> coded legacy hardware or cache events the wild card, case
> >>> insensitivity, etc. is consistent for events. This does, however, mean
> >>> events like cycles will wild card against all PMUs. A change doing the
> >>> same was originally posted and merged from:
> >>> https://lore.kernel.org/r/20240416061533.921723-10-irogers@xxxxxxxxxx
> >>> and reverted by Linus in commit 4f1b067359ac ("Revert "perf
> >>> parse-events: Prefer sysfs/JSON hardware events over legacy"") due to
> >>> his dislike for the cycles behavior on ARM with perf record. Earlier
> >>> patches in this series make perf record event opening failures
> >>> non-fatal and hide the cycles event's failure to open on ARM in perf
> >>> record, so it is expected the behavior will now be transparent in perf
> >>> record on ARM. perf stat with a cycles event will wildcard open the
> >>> event on all PMUs.
> >>
> >> Hi Ian,
> >>
> >> Briefly testing perf record and perf stat seem to work now. i.e "perf
> >> record -e cycles" doesn't fail and just skips the uncore cycles event.
> >> And "perf stat" now includes the uncore cycles event which I think is
> >> harmless.
> >
> > Thanks for confirming this.
> >
> >> But there are a few perf test failures. For example "test event parsing":
> >>
> >> evlist after sorting/fixing: 'arm_cmn_0/cycles/,{cycles,cache-
> >> misses,branch-misses}'
> >> FAILED tests/parse-events.c:1589 wrong number of entries
> >> Event test failure: test 57 '{cycles,cache-misses,branch-
> >> misses}:e'running test 58 'cycles/name=name/'
> >
> > I suspect the easiest fix for this is to change "cycles" to the
> > "cpu-cycles" legacy hardware event for this test. The test has always
> > had issues on ARM due to hardcoded expectations of the core PMU being
> > "cpu".
> >
> >> The tests "Perf time to TSC" and "Use a dummy software event to keep
> >> tracking" are using libperf to open the cycles event as a sampling event
> >> which now fails. It seems like we've fixed Perf record to ignore this
> >> failure, but we didn't think about libperf until now.
> >
> > I'm not clear on the connection here. libperf doesn't do event parsing
> > and so there are no changes in tools/lib/perf. If a test has an
> > expectation that "cycles" is a core event, again we can change it to
> > "cpu-cycles" as a workaround for ARM. As "cycles" will wildcard now,
> > we don't want that behavior in say API probing as we'll end up never
> > lazily processing the PMUs. That code has been altered in these
> > changes to specify the core PMU. For tests it is less of an issue and
> > so the changes are more limited.
> >
> > Thanks,
> > Ian
>
> Sure makes sense if there's an easy fix for the tests, we can do that. I
> suppose the main reason I mentioned it was that the tests might be
> highlighting that other genuine non-Perf and non-test users would see
> the same breakage though.
For a non-perf user to see a perf change they must transitively depend
on perf to care. I think the complaint is that we've gone from 1 event
(ignoring BIG.little/hybrid) to possibly many, particularly on ARM.
What I'm thinking is we should have something like:
#if defined(__aarch64__) || defined(__arm__)
#define HW_CYCLES_STR "cpu-cycles"
#else
#define HW_CYCLES_STR "cycles"
#endif
and remove all use of just raw "cycles" in the code to use this
#define. This should avoid the >1 event issue on ARM in things like
tests. It does cause a new problem if the evsel->name is assumed to be
cycles, which is something that can happen a lot in shell scripts.
Perhaps all those use-cases should switch to specifying a PMU, which
would be a good thing performance wise to avoid scanning lots of PMUs.
I'll add somethings to v4 to do a mix of this.
Thanks,
Ian
