Re: [PATCH v4 17/19] sched: Add migrate_disable() tracepoints

[Peter Zijlstra]

On Thu, Oct 29, 2020 at 04:27:26PM +0000, Valentin Schneider wrote:
> 
> On 23/10/20 11:12, Peter Zijlstra wrote:
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -1732,6 +1732,8 @@ void migrate_disable(void)
> >               return;
> >       }
> >
> > +	trace_sched_migrate_disable_tp(p);
> > +
> >       preempt_disable();
> >       this_rq()->nr_pinned++;
> >       p->migration_disabled = 1;
> > @@ -1764,6 +1766,8 @@ void migrate_enable(void)
> >       p->migration_disabled = 0;
> >       this_rq()->nr_pinned--;
> >       preempt_enable();
> > +
> > +	trace_sched_migrate_enable_tp(p);
> 
> Don't you want those directly after the ->migration_disabled write?
> esp. for migrate_enable(), if that preempt_enable() leads to a context
> switch then the disable->enable deltas won't reflect the kernel view.
> 
> That delta may indeed include the time it took to run the stopper and
> fix the task's affinity on migrate_enable(), but it could include all
> sorts of other higher-priority tasks.

I can put them in the preempt_disable() section I suppose, but these
tracers should be looking at task_sched_runtime(), not walltime, and
then the preemption doesn't matter.

Also, a distinct lack of actual users atm.. :/