Re: [RFC 2/3] perf tool: stat: say more about why event is not supported by the kernel

From: Kim Phillips
Date: Tue Oct 24 2017 - 20:47:13 EST


On Tue, 24 Oct 2017 14:35:30 +0100
Will Deacon <will.deacon@xxxxxxx> wrote:

> On Tue, Oct 24, 2017 at 03:04:08AM -0500, Kim Phillips wrote:
> > +++ b/tools/perf/builtin-stat.c
> > @@ -625,9 +625,13 @@ static int __run_perf_stat(int argc, const char **argv)
> > if (errno == EINVAL || errno == ENOSYS ||
> > errno == ENOENT || errno == EOPNOTSUPP ||
> > errno == ENXIO) {
> > - if (verbose > 0)
> > + if (verbose > 0) {
> > ui__warning("%s event is not supported by the kernel.\n",
> > perf_evsel__name(counter));
> > + perf_evsel__open_strerror(counter, &target,
> > + errno, msg, sizeof(msg));
> > + ui__error("%s\n", msg);
> > + }
>
> Hmm, perf_evsel__open_strerror can already get called a few lines later, so

That's in the case when perf decides to fully abort, whereas this
change allows users to know why perf stat output lists '<not
supported>' instead of having a count figure next to the event requested:

BEFORE THIS SERIES:

# ./oldperf stat -v -e ccn/cycles/ku sleep 1
Warning:
ccn/cycles/ku event is not supported by the kernel.
failed to read counter ccn/cycles/ku

Performance counter stats for 'system wide':

<not supported> ccn/cycles/ku

1.002756731 seconds time elapsed

AFTER THIS SERIES (amended to show only strerror_arch output):

The tool itself tells the user the event can't exclude execution
levels, rather than the user having to refer to dmesg:

# ./newperf stat -v -e ccn/cycles/ku sleep 1
Warning:
ccn/cycles/ku: Can't exclude execution levels!
failed to read counter ccn/cycles/ku

Performance counter stats for 'system wide':

<not supported> ccn/cycles/ku

1.002935620 seconds time elapsed

> I don't think this change is right. It looks like the intention here is that
> we can push ahead skipping unsupported events unless they are the leader of
> a populated group. If this isn't working as intended, then it looks like we
> need a helper to identify unsupported events instead.

I could be wrong, but I think it's working as intended: this patch
just adds more user-visible resolution into the event opening errors.

Kim