Re: [PATCH 1/2] perf record: Fix "--per-thread" option for hybrid machines

[Xing Zhengjun]

On 7/4/2022 9:22 PM, Liang, Kan wrote:
> On 7/1/2022 10:35 PM, zhengjun.xing@xxxxxxxxxxxxxxx wrote:
> > From: Zhengjun Xing <zhengjun.xing@xxxxxxxxxxxxxxx>
> >
> > Commit b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems to
> > collect metadata records") adds a dummy event on hybrid systems to fix 
> > the
> > symbol "unknown" issue when the workload is created in a P-core but runs
> > on an E-core. When "--per-thread" is enabled, the nr_cpus is reduced 
> > to 1,
> >   adding a dummy event is useless for this issue, and it will also cause
>
> A dummy event is required since a mmap event may be loaded at runtime on 
> any CPU. Thanks Adrian to point it out.

There will be a kernel patch to fix the "mmap fail" issue, I will drop 
this one and continue to fix other "per-thread" issues based on the 
kernel patch.


> Thanks,
> Kan
>
> > "failed to mmap with 22 (Invalid argument)". This patch stops adding 
> > dummy
> > events when the option "--per-thread" is enabled, then the option can 
> > work
> > on hybrid machines.
> >
> > Before:
> >
> >   # ./perf record -e cycles:u --per-thread  sleep 1
> >   failed to mmap with 22 (Invalid argument)
> >
> > After:
> >
> >   # ./perf record -e cycles:u --per-thread  sleep 1
> > [ perf record: Woken up 1 times to write data ]
> > [ perf record: Captured and wrote 0.002 MB perf.data (6 samples) ]
> >
> > Fixes: b91e5492f9d7 ("perf record: Add a dummy event on hybrid systems 
> > to collect metadata records")
> > Signed-off-by: Zhengjun Xing <zhengjun.xing@xxxxxxxxxxxxxxx>
> > Reviewed-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > ---
> >   tools/perf/builtin-record.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> > index e1edd4e98358..44ea2dd424fe 100644
> > --- a/tools/perf/builtin-record.c
> > +++ b/tools/perf/builtin-record.c
> > @@ -1223,7 +1223,7 @@ static int record__open(struct record *rec)
> >        * of waiting or event synthesis.
> >        */
> >       if (opts->initial_delay || target__has_cpu(&opts->target) ||
> > -        perf_pmu__has_hybrid()) {
> > +        (perf_pmu__has_hybrid() && !opts->target.per_thread)) {
> >           pos = evlist__get_tracking_event(evlist);
> >           if (!evsel__is_dummy_event(pos)) {
> >               /* Set up dummy event. */

--
Zhengjun Xing