Re: [PATCH 1/2] perf tool AMD: Use non-precise cycles as default event on certain Zen2 processors

[Namhyung Kim]

On Mon, Dec 11, 2023 at 5:54 AM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
>
> Hi Namhyung,
>
> >>>> By default, Perf uses precise cycles event when no explicit event is
> >>>> specified by user. Precise cycles event is forwarded to ibs_op// pmu
> >>>> on AMD. However, IBS has hw issue on certain Zen2 processors where
> >>>> it might raise NMI without sample_valid bit set, which causes Unknown
> >>>> NMI warnings. (Erratum #1215: IBS (Instruction Based Sampling) Counter
> >>>> Valid Value May be Incorrect After Exit From Core C6 (CC6) State.) So,
> >>>> use non-precise cycles as default event on affected processors.
> >>>
> >>> It seems like a kernel issue, do we have a kernel patch not to forward
> >>> precise cycles or instructions events to IBS on the affected CPUs?
> >>
> >> I'm not sure how it's a kernel issue. User can use ibs_op// pmu directly
> >> and might hit the hw bug.
> >
> > Sorry for the late reply.  I know it's the user's fault when using ibs_op//
> > directly but I think the kernel should not forward cycles:pp to IBS.
>
> For all AMD zen uarchs, cycles:p is same as ibs_op// (man perf-list has
> more detail). It's confusing if ibs_op// works but cycles:p fails on the
> same machine.

Make sense.  Probably perf tools can provide a more detailed message
about the situation.

>
> No strong opinion though. I can prepare a kernel patch, if you want.

I have a report that these unknown NMI messages caused some
performance issues.  So maybe we need to be safe by not using
IBS automatically.  I've also posted a patch to rate-limit the message
itself.

Thanks,
Namhyung