Re: [GIT PULL] perf tools changes for v6.10

[Arnaldo Carvalho de Melo <acme@kernel.org>]

On Fri, May 24, 2024 at 10:55:11PM -0300, Arnaldo Carvalho de Melo wrote:
> On Fri, May 24, 2024 at 06:31:52PM -0700, Linus Torvalds wrote:
> > On Tue, 21 May 2024 at 12:26, Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
> > >
> > > perf tools fixes and improvements for v6.10:
> > 
> > This actually broke 'perf' completely for me on arm64.
> > 
> > With a 6.9 version of 'perf', I can do this:
> > 
> >     perf record -e cycles:pp make -j199
> > 
> > and it all works fine.
> > 
> > With the current -git version, when I do the same, I instead get
> > 
> >   Error:
> >   cycles:pp: PMU Hardware doesn't support
> > sampling/overflow-interrupts. Try 'perf stat'
> > 
> > and after trying desperately to chase down what went wrong on the
> > kernel side, I finally figured out that it wasn't a kernel change at
> > all, it was the tooling that had changed.
> > 
> > I did a 'git bisect', and it says
> > 
> >   617824a7f0f73e4de325cf8add58e55b28c12493 is the first bad commit
> >   commit 617824a7f0f73e4de325cf8add58e55b28c12493
> >   Author: Ian Rogers <irogers@google.com>
> >   Date:   Mon Apr 15 23:15:25 2024 -0700
> > 
> >       perf parse-events: Prefer sysfs/JSON hardware events over legacy
> > 
> > and very clearly this does *NOT* work at all for me.
> > 
> > I didn't notice until now, simply because I had been busy with the
> > merge window, so I hadn't been doing any profiles, but the merge
> > window is calming down and the end is nigh, and I just wasted more
> > time than I care to admit trying to figure out what went wrong in the
> > kernel.
> > 
> > And no, I don't speak JSON, and I have *no* idea what the legacy
> > events are. Plus I'm not very familiar with the arm64 profiling etc
> > anyway, so I'm just a clueless user here.
> > 
> > I *can* confirm that just reverting that commit makes that trivial
> > "perf record" work for me. So the bisect was good, and it reverts
> > cleanly, but I don't know _why_ my arm64 setup hates it so much.
 
> That is a good data point, we probably could go with the revert, but I
> think Ian submitted a few patches fixing this issue that came up close
> to LSFMM/BPF and the merge window, so didn't have time to sit on
> linux-next for a while, I'm looking those up now.

Couldn't find it quickly, its late here, perhaps Ian can chime in and
point those fixes here. I'll try and continue tomorrow.

- Arnaldo
 
> ARM64 eyes on this would also be good. Adding Mark Rutland and Leo Yan
> to the CC list, maybe they can help us here with the best course of
> action.