Re: [PATCH 1/3] libperf: Add processing to scale the counters obtained during the read() system call when multiplexing

[Ian Rogers]

On Tue, Aug 31, 2021 at 5:26 AM Rob Herring <robh@xxxxxxxxxx> wrote:
>
> On Tue, Aug 24, 2021 at 5:12 AM nakamura.shun@xxxxxxxxxxx
> <nakamura.shun@xxxxxxxxxxx> wrote:
> >
> > Hi, Rob
> >
> > > On Fri, Aug 20, 2021 at 06:39:06PM +0900, Shunsuke Nakamura wrote:
> > > > perf_evsel__read() scales counters obtained by RDPMC during multiplexing, but
> > > > does not scale counters obtained by read() system call.
> > > >
> > > > Add processing to perf_evsel__read() to scale the counters obtained during the
> > > > read() system call when multiplexing.
> > >
> > > Which one is right though? Changing what read() returns could break
> > > users, right? Or are you implying that the RDPMC path is correct and
> > > read() was not. More likely the former case since I wrote the latter.
> >
> > perf_evsel__read() returns both the count obtained by RDPMC and the count obtained
> > by the read() system call when multiplexed with RDPMC enabled.
> >
> > That is, there is a mix of scaled and unscaled values.
> >
> > As Rob says, when this patch is applied, rescaling the count obtained from
> > perf_evsel__read() during multiplexing will break the count.
> >
> > I think the easiest solution is to change the value you get from RDPMC to not scale
> > and let the user scale it, but I thought it would be a little inconvenient.
>
> Agreed, unless someone else has an opinion. It would be good to do the
> scaling in libperf with the optimized math op, but I assume there's
> some reason the user may need unscaled values?

Hi, something I've mentioned on other threads [1] is that running may
be zero due to multiplexing but enabled be greater. This can lead to a
divide by zero when scaling. Giving the ratio to the caller gives more
information - I may be misunderstanding this thread, apologies if so.

Thanks,
Ian

[1] https://lore.kernel.org/lkml/CAL_JsqKc_qFA59L9e-xXOhE4yBTdVg-Ea9EddimWwqj3XXxhGw@xxxxxxxxxxxxxx/

> Rob