On Wed, Jul 11, 2012 at 04:55:08PM +0200, Peter Zijlstra wrote: > On Wed, 2012-07-11 at 16:48 +0200, Frederic Weisbecker wrote: > > On Wed, Jul 11, 2012 at 04:38:19PM +0200, Peter Zijlstra wrote: > > > On Wed, 2012-07-11 at 16:36 +0200, Frederic Weisbecker wrote: > > > > > > > > In this case he can just record sched wakeup as well. With sched_switch > > > > + sched_wakeup, he'll unlikely lose events. > > > > > > > > With sched_stat_sleep he will lose events, unless we fix this period > > > > demux thing. > > > > > > But without this patch, the sched_wakeup will belong to another task, so > > > if you trace task A, and B wakes you, you'll never see the wakeup. > > > > Ah so the goal is to minimize the amount of events by only tracing task A? > > Right, or just not having sufficient privs to trace the world. And a > wakeup of A is very much also part of A, not only the task doing the > wakeup. > > Hence the proposed mechanism.
Yeah that's fair. > > > Ok then. Still we need to fix these events that use __perf_count() because > > wide tracing of sched_switch/wake_up still generate less events than > > sched stat sleep. > > > > I believe: > > > > perf record -e sched:sched_stat_sleep sleep 1 > > > > produces 1 billion events because we sleep 1 billion nanosecs. Or > > something like that. > > Right.. back when I did that the plan was to make PERF_SAMPLE_PERIOD fix > that, of course that never seemed to have happened. > > With PERF_SAMPLE_PERIOD you can simply write the 1b into the period of 1 > event and be done with it. I believe the perf tools handle pretty well variable periods of an event on top of PERF_SAMPLE_PERIOD. We just need to tweak the maths in perf_swevent_overflow() I think... -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
