Re: [PATCH v2 1/2] libperf: Add processing to scale the counters obtained during the read() system call when multiplexing

From: Jiri Olsa
Date: Sun Nov 14 2021 - 11:16:57 EST


On Mon, Nov 08, 2021 at 12:49:24AM +0000, nakamura.shun@xxxxxxxxxxx wrote:
> Hi Jirka
>
> > > > > On Wed, Sep 22, 2021 at 07:16:26PM +0900, Shunsuke Nakamura wrote:
> > > > > > From: nakamura shunsuke <nakamura.shun@xxxxxxxxxxx>
> > > > > >
> > > > > > 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.
> > > > > >
> > > > > >
> > > > > > Signed-off-by: Shunsuke Nakamura <nakamura.shun@xxxxxxxxxxx>
> > > > > > ---
> > > > > >  tools/lib/perf/evsel.c | 6 ++++++
> > > > > >  1 file changed, 6 insertions(+)
> > > > > >
> > > > > > diff --git a/tools/lib/perf/evsel.c b/tools/lib/perf/evsel.c
> > > > > > index 8441e3e1aaac..0ebd1d34436f 100644
> > > > > > --- a/tools/lib/perf/evsel.c
> > > > > +++ b/tools/lib/perf/evsel.c
> > > > > > @@ -18,6 +18,7 @@
> > > > > >  #include <sys/ioctl.h>
> > > > > >  #include <sys/mman.h>
> > > > > >  #include <asm/bug.h>
> > > > > > +#include <linux/math64.h>
> > > > > > 
> > > > > >  void perf_evsel__init(struct perf_evsel *evsel, struct perf_event_attr *attr,
> > > > > >                      int idx)
> > > > > > @@ -321,6 +322,11 @@ int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
> > > > > >        if (readn(*fd, count->values, size) <= 0)
> > > > > >                return -errno;
> > > > > > 
> > > > > > +     if (count->ena != count->run) {
> > > > > > +             if (count->run != 0)
> > > > > > +                     count->val = mul_u64_u64_div64(count->val, count->ena, count->run);
> > > > > > +     }
> > > > >
> > > > > so I think perf stat expect raw values in there and does the
> > > > > scaling by itself, please check following code:
> > > > >
> > > > > read_counters
> > > > >   read_affinity_counters
> > > > >     read_counter_cpu
> > > > >       read_single_counter
> > > > >         evsel__read_counter
> > > > >
> > > > >   perf_stat_process_counter
> > > > >     process_counter_maps
> > > > >       process_counter_values
> > > > >         perf_counts_values__scale
> > > > >
> > > > >
> > > > > perhaps we could export perf_counts_values__scale if it'd be any help
> > > >
> > > > Thank you for your comment.
> > > >
> > > > The purpose of this patch is to unify the counters obtained with
> > > > perf_evsel__read() to scaled or unscaled values.
> > > >
> > > > perf_evsel__read() gets counter by perf_mmap__read_self() if RDPMC is
> > > > available, else gets by readn(). In current implementation, caller
> > > > gets scaled counter if goes through RDPMC path, otherwise gets unscaled
> > > > counter via readn() path.
> > > >
> > > > However caller cannnot know which path were taken.
> > > >
> > > > If caller expects a raw value, I think the RDPMC path should also
> > > > return an unscaled counter.
> > > >
> > > > diff --git a/tools/lib/perf/mmap.c b/tools/lib/perf/mmap.c
> > > > index c89dfa5..aaa4579 100644
> > > > --- a/tools/lib/perf/mmap.c
> > > > +++ b/tools/lib/perf/mmap.c
> > > > @@ -353,8 +353,6 @@ int perf_mmap__read_self(struct perf_mmap *map, struct perf_counts_values *count
> > > >                 count->ena += delta;
> > > >                 if (idx)
> > > >                         count->run += delta;
> > > > -
> > > > -               cnt = mul_u64_u64_div64(cnt, count->ena, count->run);
> > >
> > > perf stat does not mmap counters so this should not be invoked
> > > within perf stat.. but we should be consistent and scale after
> > > calling perf_evsel__read.. and give user the possibility to get
> > > un-scaled counts
> > >
> > > that perhaps brings new feature.. mmap perf stat counters to invoke
> > > the fast reading path for counters.. IIRC it should be matter just
> > > to mmap the first 'user' page
> >
> > Thank you for your comment.
> > I think it will be good that perf stat supports rdpmc.
> >
> > I will modify the patch.
>
> I think rdpmc cannot measure the command/program specified in perf stat
> because it measures the calling thread of perf_event_open.
> If my understanding is wrong, please point it out to me.

right, I guess we could use that just for system wide monitoring,
where we open counter for each cpu

jirka