Re: [PATCH] perf evsel: Get group fd from CPU0 for system wide event

[Jin, Yao]

Hi Jiri,

On 5/15/2020 4:33 PM, Jiri Olsa wrote:
> On Fri, May 15, 2020 at 02:04:57PM +0800, Jin, Yao wrote:
>
> SNIP
>
> > I think I get the root cause. That should be a serious bug in get_group_fd, access violation!
> >
> > For a group mixed with system-wide event and per-core event and the group
> > leader is system-wide event, access violation will happen.
> >
> > perf_evsel__alloc_fd allocates one FD member for system-wide event (only FD(evsel, 0, 0) is valid).
> >
> > But for per core event, perf_evsel__alloc_fd allocates N FD members (N =
> > ncpus). For example, for ncpus is 8, FD(evsel, 0, 0) to FD(evsel, 7, 0) are
> > valid.
> >
> > get_group_fd(struct evsel *evsel, int cpu, int thread)
> > {
> >      struct evsel *leader = evsel->leader;
> >
> >      fd = FD(leader, cpu, thread);    /* access violation may happen here */
> > }
> >
> > If leader is system-wide event, only the FD(leader, 0, 0) is valid.
> >
> > When get_group_fd accesses FD(leader, 1, 0), access violation happens.
> >
> > My fix is:
> >
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index 28683b0eb738..db05b8a1e1a8 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -1440,6 +1440,9 @@ static int get_group_fd(struct evsel *evsel, int cpu, int thread)
> >          if (evsel__is_group_leader(evsel))
> >                  return -1;
> >
> > +       if (leader->core.system_wide && !evsel->core.system_wide)
> > +               return -2;
>
> so this effectively stops grouping system_wide events with others,
> and I think it's correct, how about events that differ in cpumask?

My understanding for the events that differ in cpumaks is, if the leader's cpumask is not fully 
matched with the evsel's cpumask then we stop the grouping. Is this understanding correct?

I have done some tests and get some conclusions:

1. If the group is mixed with core and uncore events, the system_wide checking can distinguish them.

2. If the group is mixed with core and uncore events and "-a" is specified, the system_wide for core 
event is also false. So system_wide checking can distinguish them too

3. In my test, the issue only occurs when we collect the metric which is mixed with uncore event and 
core event, so maybe checking the system_wide is OK.

> should we perhaps ensure this before we call open? go throught all
> groups and check they are on the same cpus?

The issue doesn't happen at most of the time (only for the metric consisting of uncore event and 
core event), so fallback to stop grouping if call open is failed looks reasonable.

Thanks
Jin Yao

> thanks,
> jirka
>
>
> > +
> >          /*
> >           * Leader must be already processed/open,
> >           * if not it's a bug.
> > @@ -1665,6 +1668,11 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
> >                                  pid = perf_thread_map__pid(threads, thread);
> >
> >                          group_fd = get_group_fd(evsel, cpu, thread);
> > +                       if (group_fd == -2) {
> > +                               errno = EINVAL;
> > +                               err = -EINVAL;
> > +                               goto out_close;
> > +                       }
> >   retry_open:
> >                          test_attr__ready();
> >
> > It enables the perf_evlist__reset_weak_group. And in the second_pass (in
> > __run_perf_stat), the events will be opened successfully.
> >
> > I have tested OK for this fix on cascadelakex.
> >
> > Thanks
> > Jin Yao