Re: [PATCH 2/4] perf: Add exclude_task perf event attribute

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Jun 08, 2010 at 08:59:17PM +0200, Ingo Molnar wrote:
> 
> * Frederic Weisbecker <fweisbec@gmail.com> wrote:
> 
> > On Tue, May 25, 2010 at 08:58:08AM +0200, Peter Zijlstra wrote:
> > > On Tue, 2010-05-25 at 11:43 +1000, Paul Mackerras wrote:
> > > > On Fri, May 21, 2010 at 04:05:13PM +0200, Frederic Weisbecker wrote:
> > > > 
> > > > > Excluding is useful when you want to trace only hard and softirqs.
> > > > > 
> > > > > For this we use a new generic perf_exclude_event() (the previous
> > > > > one beeing turned into perf_exclude_swevent) to which you can pass
> > > > > the preemption offset to which your events trigger.
> > > > > 
> > > > > Computing preempt_count() - offset gives us the preempt_count() of
> > > > > the context that the event has interrupted, on top of which we
> > > > > can filter the non-irq contexts.
> > > > 
> > > > How does this work for hardware events when we are sampling and
> > > > getting an interrupt every N events?  It seems like the hardware is
> > > > still counting all events and interrupting every N events, but we are
> > > > only recording a sample if the interrupt occurred in the context we
> > > > want.  In other words the context of the Nth event is considered to be
> > > > the context for the N-1 events preceding that, which seems a pretty
> > > > poor approximation.
> > > > 
> > > > Also, for hardware events, if we are counting rather than sampling,
> > > > the exclude_task bit will have no effect.  So perhaps in that case the
> > > > perf_event_open should fail rather than appear to succeed but give
> > > > wrong data.
> > > 
> > > Right, so for hardware event we'd need to go with those irq_{enter,exit}
> > > hooks and either fully disable the call, or do as Ingo suggested, read
> > > the count delta and add that to period_left, so that we'll delay the
> > > sample (and subtract from ->count, which is I think the trickiest bit as
> > > it'll generate a non-monotonic ->count).
> > > 
> > > So I prefer the disable/enable from irq_enter/exit, however I also
> > > suspect that that is by far the most expensive option.
> > 
> > 
> > Playing with that, it's easy to contain the counting on the filtered
> > contexts: I can just flush (event->read()) when we enter/exit a context
> > but filter the update of event->count depending on exclude_* things.
> > 
> > There are several problems with that though:
> > 
> > - overflow interrupts continue, we can block them, but still...
> > - periods become randomly async as the interrupts happen. We
> >   could save the period_left on context enter to solve this
> > 
> > 
> > It would be certainly easier and clearer to use stop/start things on context
> > enter/exit.
> > 
> > And the only thing that seem to happen in these paths is a write
> > to the event config register.
> > Is it what is going to be too slow?
> > If you compare that to all the reads on the counter,
> > the interrupts that still need to be serviced and filtered with the
> > other solution, may be the stop/start solution is eventually better
> > in contrast.
> > 
> > How much time approximately does it take to write in this config register?
> 
> it should be fast enough. I think we should first go for a good, high-quality 
> implementation that has a correct model for collecting information - and then, 
> if in practice there's any significant slowdown, we could perhaps add a 
> speedup that cuts corners.
> 
> If we first cut corners we'll never be able to fully trust the info, and we'll 
> never know how it would all have played out via the disable/enable method.
> 
> Thanks,
> 
> 	Ingo



All agreed, I'm taking that direction then.

Thanks.