Re: sched: trace: update trace_sched_wakeup()

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Peter Zijlstra (peterz@infradead.org) wrote:
> On Tue, 2008-12-16 at 10:47 -0500, Mathieu Desnoyers wrote:
> > * Peter Zijlstra (peterz@infradead.org) wrote:
> > > On Tue, 2008-12-16 at 08:07 +0100, Peter Zijlstra wrote:
> > > > Add the information needed to distinguish 'real' wakeups from 'false'
> > > > wakeups.
> > > 
> > > I don't seem to be doing so well this morning...
> > > 
> > > Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > > 
> > > Thanks Dhaval!
> > > 
> > 
> > Applied a modified version into the 2.6.27.9 LTTng tree. Thanks!
> > (remembering that I should really move to 2.6.28-rc!)
> 
> Ingo pointed out that this generates build crap in ftrace which made me
> look at the thing again, and I just realized we don't need it.
> 
> trace_sched_wakeup() gets called before we set p->state = TASK_RUNNING,
> so by checking p->state in the callback we can see if its a real of
> false wakeup.
> 
> The scheduler tracer in ftrace already exploits this.
> 

We do something similar in LTTng : we save the state along with the
event in the trace stream and let the userspace analyzer see what was
the state of the process prior to the wakeup.

Thanks,

Mathieu

> > > > --- a/kernel/sched.c
> > > > +++ b/kernel/sched.c
> > > > @@ -2329,7 +2329,7 @@ out_activate:
> > > >  	success = 1;
> > > >  
> > > >  out_running:
> > > > -	trace_sched_wakeup(rq, p);
> > > > +	trace_sched_wakeup(rq, p, success);
> > > >  	check_preempt_curr(rq, p, sync);
> > > >  
> > > >  	p->state = TASK_RUNNING;
> 
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68