Re: [PATCH V7 07/14] perf/core: Add a new PERF_EV_CAP_COEXIST event capability

[Liang, Kan]

On 7/24/2020 9:54 AM, Peter Zijlstra wrote:
> On Fri, Jul 24, 2020 at 09:43:44AM -0400, Liang, Kan wrote:
> > On 7/24/2020 7:46 AM, peterz@xxxxxxxxxxxxx wrote:
> > > On Fri, Jul 24, 2020 at 12:55:43PM +0200, peterz@xxxxxxxxxxxxx wrote:
> > > > > +	event_sched_out(event, cpuctx, ctx);
> > > > > +	perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
> > > > > +}
> > > >
> > > > Ah, so the problem here is that ERROR is actually recoverable using
> > > > IOC_ENABLE. We don't want that either. Let me try and figure out of EXIT
> > > > would work.
> > >
> > > EXIT is difficult too.. Also, that COEXIST thing hurt my brian, can't we
> > > do a simpler SIBLING thing that simply requires the event to be a group
> > > sibling?
> > >
> > > The consequence is that SLOTS must be the leader, is that really a
> > > problem? You keep providing the {cycles, slots, metric-things} example,
> > > but afaict {slots, metric-thing, cycles} should work just fine too.
> > > PERF_SAMPLE_READ with PERF_FORMAT_GROUP doesn't need to the the leader.
> >
> > I'm not sure I get your point.
> > For the PERF_SAMPLE_READ with PERF_FORMAT_GROUP case, other events can be
> > the leader, e.g., {cycles, slots, metric-things}:S.
> > In this case, the SLOTS event is not a leader. I don't think we can assume
> > that the SLOTS event must be the leader.
>
> You can have a sibling event with SAMPLE_READ and FORMAT_GROUP just
> fine. The sampling event doesn't need to be the leader.

There will be a problem for the current perf tool, which assumes that 
the leader event is the sampling event.

I will check how can we specially handle it in the perf tool.

Thanks,
Kan