Re: [RESEND PATCH V2 3/4] perf/x86/intel: drain PEBS buffer in event read

[Liang, Kan]

On 1/18/2018 4:49 AM, Jiri Olsa wrote:
> On Tue, Jan 16, 2018 at 01:49:13PM -0500, Liang, Kan wrote:
> > On 1/11/2018 10:45 AM, Jiri Olsa wrote:
> > > On Thu, Jan 11, 2018 at 10:21:25AM -0500, Liang, Kan wrote:
> > >
> > > SNIP
> > >
> > > > > hum, but the PEBS drain is specific just for
> > > > > PERF_X86_EVENT_AUTO_RELOAD events, right?
> > > >
> > > > Accurately, PEBS drain is specific for PERF_X86_EVENT_FREERUNNING here.
> > > > PERF_X86_EVENT_FREERUNNING event must be _AUTO_RELOAD event.
> > > > But in some cases, _AUTO_RELOAD event cannot be _FREERUNNING event.
> > > >
> > > > Only the event which is both _FREERUNNING and _AUTO_RELOAD need to do PEBS
> > > > drain in _read().
> > > >
> > > > So it does the check in intel_pmu_pebs_read()
> > > > +	if (pebs_needs_sched_cb(cpuc))
> > > > +		return intel_pmu_drain_pebs_buffer();
> > > >
> > > > > wrt readability maybe you could add function like:
> > > >
> > > > The existing function pebs_needs_sched_cb() can do the check.
> > > > We just need to expose it, and also the intel_pmu_drain_pebs_buffer().
> > > >
> > > > But to be honest, I still cannot see a reason for that.
> > > > It could save a call to intel_pmu_pebs_read(), but _read() is not critical
> > > > path. It doesn't save much.
> > >
> > > hum, pmu->read is also called for PERF_SAMPLE_READ for sample,
> > > check perf_output_read
> > >
> > > for non sampling event you shouldn't be able to create PEBS
> > > event, there's check in x86_pmu_hw_config
> > >
> > > I agree it does not save much, it just confused me while
> > > I was reading the code, like why is this needed for all
> > > events with precise_ip
> >
> >
> > Sorry for the late response.
> >
> > How about the patch as below?
> > The patch will be split into two patches in V3. One is to introduce
> > intel_pmu_large_pebs_read, the other is to introduce intel_pmu_read_event.
> >
> > Thanks,
> > Kan
> >
> > diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
> > index 731153a..1610a9d 100644
> > --- a/arch/x86/events/intel/core.c
> > +++ b/arch/x86/events/intel/core.c
> > @@ -2060,6 +2060,14 @@ static void intel_pmu_del_event(struct perf_event
> > *event)
> >   		intel_pmu_pebs_del(event);
> >   }
> >
> > +static void intel_pmu_read_event(struct perf_event *event)
> > +{
> > +	if (intel_pmu_large_pebs_read(event))
> > +		return;
>
> should this be 'if (!intel_pmu_large_pebs_read(event))'

NO. For large pebs, the event->count has been updated in drain_pebs(). 
So it doesn't need to do x86_perf_event_update() again.

Thanks,
Kan


> but looks better for me without the precise_ip check
>
> thanks,
> jirka