Re: [PATCH] tracing: Add task activate/deactivate tracepoints

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, May 31, 2010 at 04:43:33PM +0200, Peter Zijlstra wrote:
> On Mon, 2010-05-31 at 16:36 +0200, Frederic Weisbecker wrote:
> > On Mon, May 31, 2010 at 10:54:59AM +0200, Peter Zijlstra wrote:
> > > On Mon, 2010-05-31 at 10:12 +0200, Peter Zijlstra wrote:
> > > > On Mon, 2010-05-31 at 10:00 +0200, Ingo Molnar wrote:
> > > > > > 
> > > > > > NAK, aside from a few corner cases wakeup and sleep are the important 
> > > > > > points.
> > > > > > 
> > > > > > The activate and deactivate functions are implementation details.
> > > > > 
> > > > > Frederic, can you show us a concrete example of where we dont know what is 
> > > > > going on due to inadequate instrumentation? Can we fix that be extending the 
> > > > > existing tracepoints? 
> > > > 
> > > > Right, so a few of those corner cases I mentioned above are things like
> > > > re-nice, PI-boosts etc.. Those use deactivate, modify task-state,
> > > > activate cycles. so if you want to see those, we can add an explicit
> > > > tracepoint for those actions.
> > > > 
> > > > An explicit nice/PI-boost tracepoint is much clearer than trying to
> > > > figure out wth the deactivate/activate cycle was for.
> > > 
> > > Another advantage of explicit tracepoints is that you'd see them even
> > > for non-running tasks, because we only do the deactivate/activate thingy
> > > for runnable tasks.
> > 
> > 
> > Yeah. So I agree with you that activate/deactivate are too much
> > implementation related, they even don't give much sense as we
> > don't know the cause of the event, could be a simple renice, or
> > could be a sleep.
> > 
> > So agreed, this sucks.
> > 
> > For the corner cases like re-nice and PI-boost or so, we can indeed plug
> > some higher level tracepoints there.
> > 
> > But there is one more important problem these tracepoints were solving and
> > that still need something:
> > 
> > We don't know when a task goes to sleep. We have two wait tracepoints,
> > sched_wait_task() to wait for a task to unschedule, and sched_process_wait()
> > that is a hooks for waitid and wait4 syscalls. So we are missing all
> > the event waiting from inside the kernel. But even with that, wait and sleep
> > doesn't mean the same thing. Sleeping don't always involve using the waiting
> > API.
> > 
> >   I think we need such tracepoint:
> > 
> > diff --git a/kernel/sched.c b/kernel/sched.c
> > index 8c0b90d..5f67c04 100644
> > --- a/kernel/sched.c
> > +++ b/kernel/sched.c
> > @@ -3628,8 +3628,10 @@ need_resched_nonpreemptible:
> >  	if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
> >  		if (unlikely(signal_pending_state(prev->state, prev)))
> >  			prev->state = TASK_RUNNING;
> > -		else
> > +		else {
> > +			trace_sched_task_sleep(prev);
> >  			deactivate_task(rq, prev, DEQUEUE_SLEEP);
> > +		}
> >  		switch_count = &prev->nvcsw;
> >  	}
> 
> > And concerning the task waking up, if it is not migrated, it means it stays
> > on its orig cpu. This is something that can be dealt from the post-processing.
> 
> Hurm,.. I was thinking trace_sched_switch(.prev_state != TASK_RUNNING)
> would be enough, but its not for preemptible kernels.
> 
> Should we maybe cure this and rely on sched_switch() to detect sleeps?
> It seems natural since only the current task can go to sleep, its just
> that the whole preempt state gets a bit iffy.


Sounds good, we have the preempt depth in the common tracepoint headers, I'll
try to rebuild a reliable cpu runqueue from post-processing and see if all that
is enough.

Thanks.