Re: [PATCH V7 3/6] perf, x86: handle multiple records in PEBS buffer

[Peter Zijlstra <peterz@infradead.org>]

On Tue, May 05, 2015 at 04:30:25PM +0000, Liang, Kan wrote:
> > > +	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
> > >  		struct pebs_record_nhm *p = at;
> > >
> > >  		for_each_set_bit(bit, (unsigned long *)&p->status,
> > >  				 x86_pmu.max_pebs_events) {
> > >  			event = cpuc->events[bit];
> > >  			WARN_ON_ONCE(!event);
> > >
> > > +			if (event->attr.precise_ip)
> > > +				break;
> > > +		}
> > 
> > Would it make sense to delay looking for the event until you've found
> > there is a single bit set -- and already know which bit that is?
> >
> 
> Yes, I think we can test cpuc->pebs_enabled here.
> It should be better than attr.precise_ip checking.
> 
> -       for (; at < top; at += x86_pmu.pebs_record_size) {
> +       for (at = base; at < top; at += x86_pmu.pebs_record_size) {
>                 struct pebs_record_nhm *p = at;
> 
>                 for_each_set_bit(bit, (unsigned long *)&p->status,
>                                  x86_pmu.max_pebs_events) {
> -                       event = cpuc->events[bit];
> -                       if (!test_bit(bit, cpuc->active_mask))
> -                               continue;
> -
> -                       WARN_ON_ONCE(!event);
> 
> -                       if (!event->attr.precise_ip)
> -                               continue;
> +                       if (test_bit(bit, cpuc->pebs_enabled))
> +                               break;
> +               }
> 

Can't we take that entire for_each_set_bit() loop out?

It appears to me we effectively do that single test_bit() test you left
in there already with the & cpuc->pebs_enabled later on.


>  
> +       for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
> +               if (counts[bit] == 0)
>                         continue;
> -
> -               __intel_pmu_pebs_event(event, iregs, at);
> +               event = cpuc->events[bit];
> +               WARN_ON_ONCE(!event);
> +               WARN_ON_ONCE(!event->attr.precise_ip);
> +               __intel_pmu_pebs_event(event, iregs, base,
> +                                      top, bit, counts[bit]);
>         }

Right bit of paranoia :-)