Re: [PATCH 2/2] sched/debug: fix deadlock when enabling sched events

[Ingo Molnar <mingo@kernel.org>]

* Josh Poimboeuf <jpoimboe@redhat.com> wrote:

> On Wed, Jun 08, 2016 at 09:56:12AM +0200, Ingo Molnar wrote:
> > 
> > * Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> > 
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -789,6 +789,13 @@ static void update_curr_fair(struct rq *rq)
> > >  	update_curr(cfs_rq_of(&rq->curr->se));
> > >  }
> > >  
> > > +void trace_sched_stat_register(void)
> > > +{
> > > +#ifdef CONFIG_SCHEDSTATS
> > > +	force_schedstat_enabled();
> > > +#endif
> > > +}
> > 
> > I think it would be cleaner to provide an empty force_schedstat_enabled() 
> > definition in sched.h, on !CONFIG_SCHEDSTATS.
> 
> Yes, agreed.
> 
> > But it might make sense to further decouple schedstats from tracing?
> 
> Looking at how we could do that:
> 
> - The sched_stat_wait tracepoint is dependent on the wait_start schedstat.

Can we maintain wait_start unconditionally? Having config dependent tracepoints 
sucks.

> - The sched_stat_sleep tracepoint is dependent on the sleep_start
>   schedstat.

Ditto.

> - The sched_stat_iowait and sched_stat_blocked tracepoints are dependent
>   on the block_start schedstat.

Ditto.

> We could move those three schedstats values out of sched_statistics and
> into sched_entity, and then always update them regardless of
> CONFIG_SCHEDSTATS.
> 
> That would ensure these tracepoints always work.  But then again it
> would add a little bit of runtime overhead.
> 
> I don't have a strong opinion either way.  Thoughts?

Please do a before/after 'size vmlinux' dump of a defconfig build with tracing 
disabled. I.e. how much code overhead does the updating of those variables add?

Thanks,

	Ingo