Re: [PATCH] perf lock: clean the options for perf record

[Hitoshi Mitake]

On 2011å03æ04æ 23:21, Frederic Weisbecker wrote:
> On Fri, Mar 04, 2011 at 02:58:19PM +0100, Peter Zijlstra wrote:
> > On Fri, 2011-03-04 at 14:56 +0100, Frederic Weisbecker wrote:
> > > lockstat is a global measurement since the boot.
> >
> > You can reset lockstat at any given time.
>
> Event though, that's still a global profiling. You can't have a per
> process or thread granularity.
>
> Or even more precise context for some future feature in perf that
> I have in mind and would like to implement soon, like counting/sampling
> an event only between two others. More exactly having two lists for
> each event:
>
> * activate
> * deactivate
>
> Those on the first list activate the target event when they overflow. (->start() )
> Those on the second list deactivate .... (->stop() )
>
> With events in activate and deactivate in the same context than the target,
> locking and permissions should be kept simple. Locking especially shouldn't be
> needed in the fast path. Then if that's needed we could think about a cross
> context things later.
>
> Plus an attr->start_state that decides if the first ->add() made is made
> with PERF_EF_STAT or not. (We then need to keep track of some activated/deactivated
> state across schedules).
>
> That's in fact the exclude_irq and exclude_softirq idea extended to any kind
> of existing event.

Do you mean that the event dropping in perf_tp_event_match() should be
extended for filtering hardirq and softirq? If so, it seems good.
I'd like to measure the effectiveness of it.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/