Re: [PATCH 7/8] perf tools: Check fallback error and order

[Namhyung Kim <namhyung@kernel.org>]

On Thu, Oct 03, 2024 at 10:32:47AM -0700, Ian Rogers wrote:
> On Thu, Oct 3, 2024 at 10:06 AM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Tue, Oct 01, 2024 at 03:21:50PM -0700, Ian Rogers wrote:
> > > On Tue, Oct 1, 2024 at 2:36 PM Namhyung Kim <namhyung@kernel.org> wrote:
> > > >
> > > > On Tue, Oct 01, 2024 at 11:00:20AM -0700, Ian Rogers wrote:
> > > > > On Mon, Sep 30, 2024 at 5:20 PM Namhyung Kim <namhyung@kernel.org> wrote:
> > > > > >
> > > > > > The perf_event_open might fail due to various reasons, so blindly
> > > > > > reducing precise_ip level might not be the best way to deal with it.
> > > > > >
> > > > > > It seems the kernel return -EOPNOTSUPP when PMU doesn't support the
> > > > > > given precise level.  Let's try again with the correct error code.
> > > > > >
> > > > > > This caused a problem on AMD, as it stops on precise_ip of 2 for IBS but
> > > > > > user events with exclude_kernel=1 cannot make progress.  Let's add the
> > > > > > evsel__handle_error_quirks() to this case specially.  I plan to work on
> > > > > > the kernel side to improve this situation but it'd still need some
> > > > > > special handling for IBS.
> > > > > >
> > > > > > Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> > > > > > ---
> > > > > >  tools/perf/util/evsel.c | 27 +++++++++++++++++++++------
> > > > > >  1 file changed, 21 insertions(+), 6 deletions(-)
> > > > > >
> > > > > > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > > > > > index 32e30c293d0c6198..ef8356260eea54cd 100644
> > > > > > --- a/tools/perf/util/evsel.c
> > > > > > +++ b/tools/perf/util/evsel.c
> > > > > > @@ -2419,6 +2419,20 @@ static bool evsel__detect_missing_features(struct evsel *evsel)
> > > > > >         return false;
> > > > > >  }
> > > > > >
> > > > > > +static bool evsel__handle_error_quirks(struct evsel *evsel, int error)
> > > > > > +{
> > > > > > +       /* AMD IBS doesn't support exclude_kernel, forward it to core PMU */
> > > > >
> > > > > Should the quirk handling be specific to evsels with the IBS PMU given
> > > > > the comment above? ie this is a PMU specific workaround rather than an
> > > > > AMD specific workaround, however, the PMU only exists on AMD :-)
> > > > >
> > > > > > +       if (error == -EINVAL && evsel->precise_max && evsel->core.attr.precise_ip &&
> > > > > > +           evsel->core.attr.exclude_kernel && x86__is_amd_cpu()) {
> > > > >
> > > > > So here rather than x86__is_amd_cpu it would be
> > > > > !strcmp(evsel->pmu->name, "ibs_...") . But it may be cleaner to move
> > > > > the logic into pmu.c.
> > > >
> > > > I guess the problem is that AMD driver does implicit forwarding to IBS
> > > > if the legacy hardware events have precise_ip.  So it might have just
> > > > core pmu (or no pmu) in the evsel.
> > >
> > > I think the no PMU case should probably have a PMU possibly similar to
> > > the tool PMU in:
> > > https://lore.kernel.org/all/20240912190341.919229-1-irogers@google.com/
> > > But that's not in place yet. You can always use
> > > perf_pmus__scan_core(NULL) or
> > > perf_pmus__find_by_type(evsel->core.attr.type or PERF_TYPE_RAW).
> > > evsel->pmu->is_core && x86__is_amd_cpu() would identify an AMD core
> > > PMU whereas the code above could fire for IBS or other PMUs.
> >
> > But IBS is the only PMU on AMD that provides precise_ip IIRC.  So I
> > don't think other events would have it nor have any effect on changing
> > the value.  So maybe we can skip the PMU scanning and just drop to 0?
> 
> cpu supports precise_ip as it forwards it to IBS, IBS is an ambiguous
> term as there are ibs_op and ibs_fetch PMUs. The code is using values
> in the attribute to infer the PMU that is being used, it feels it
> would be more intention revealing to do things like:
> ```
> if (error == ... && perf_pmu__is_ibs_op_or_fetch(evsel->pmu)) ..
> ```

I guess it'd get a core PMU for the default cycles event.  I think the
intention is already confusing with the implicit forwarding.

Thanks,
Namhyung


> perhaps to not burden the code this can be:
> ```
> if (...) {
>   assert(perf_pmu__is_ibs_op_or_fetch(evsel->pmu));
> ```
> 
> Thanks,
> Ian