Re: [PATCH 2/2 sched_ext/for-6.12] sched_ext: Use ktime_get_ns() instead of rq_clock_task() in touch_core_sched()

[Peter Zijlstra]

On Fri, Aug 30, 2024 at 12:40:14PM -0500, David Vernet wrote:
> On Fri, Aug 30, 2024 at 12:52:48AM -1000, Tejun Heo wrote:
> > Since sched_ext: Unpin and repin rq lock from balance_scx(), sched_ext's
> > balance path terminates rq_pin in the outermost function. This is simpler
> > and in line with what other balance functions are doing but it loses control
> > over rq->clock_update_flags which makes assert_clock_udpated() trigger if
> > other CPUs pins the rq lock.
> > 
> > The only place this matters is touch_core_sched() which uses the timestamp
> > to order tasks from sibling rq's. For now, switch to ktime_get_ns(). Later,
> > it'd be better to use per-core dispatch sequence number.
> > 
> > Signed-off-by: Tejun Heo <tj@xxxxxxxxxx>
> > Fixes: 3cf78c5d01d6 ("sched_ext: Unpin and repin rq lock from balance_scx()")
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > ---
> >  kernel/sched/ext.c |   10 ++++++++--
> >  1 file changed, 8 insertions(+), 2 deletions(-)
> > 
> > --- a/kernel/sched/ext.c
> > +++ b/kernel/sched/ext.c
> > @@ -1453,13 +1453,20 @@ static void schedule_deferred(struct rq
> >   */
> >  static void touch_core_sched(struct rq *rq, struct task_struct *p)
> >  {
> > +	lockdep_assert_rq_held(rq);
> > +
> >  #ifdef CONFIG_SCHED_CORE
> >  	/*
> >  	 * It's okay to update the timestamp spuriously. Use
> >  	 * sched_core_disabled() which is cheaper than enabled().
> > +	 *
> > +	 * TODO: ktime_get_ns() is used because rq_clock_task() can't be used as
> > +	 * SCX balance path doesn't pin the rq. As this is used to determine
> > +	 * ordering between tasks of sibling CPUs, it'd be better to use
> > +	 * per-core dispatch sequence instead.
> >  	 */
> >  	if (!sched_core_disabled())
> > -		p->scx.core_sched_at = rq_clock_task(rq);
> > +		p->scx.core_sched_at = ktime_get_ns();
> 
> Should we just use sched_clock_cpu()? That's what rq->clock is updated
> from, and it's what fair.c does on the balance path when the rq lock is
> unpinned.

Right, so on x86 with wobbly TSC (still possible in this day and age)
ktime *must* use the HPET, while sched_clock_cpu() makes an 'educated'
guess using TSC and tick based HPET stamps and windows.

IOW, on same machines it doesn't matter much, but for the HPET case the
sched_clock() thing is a lot faster.