Re: [PATCH v2 9/9] sched/fair: Consider capacity inversion in util_fits_cpu()

[Qais Yousef]

On 11/07/22 18:58, Valentin Schneider wrote:
> On 05/11/22 20:41, Qais Yousef wrote:
> > On 11/04/22 17:35, Valentin Schneider wrote:
> >> On 04/08/22 15:36, Qais Yousef wrote:
> >> > We do consider thermal pressure in util_fits_cpu() for uclamp_min only.
> >> > With the exception of the biggest cores which by definition are the max
> >> > performance point of the system and all tasks by definition should fit.
> >> >
> >> > Even under thermal pressure, the capacity of the biggest CPU is the
> >> > highest in the system and should still fit every task. Except when it
> >> > reaches capacity inversion point, then this is no longer true.
> >> >
> >> > We can handle this by using the inverted capacity as capacity_orig in
> >> > util_fits_cpu(). Which not only addresses the problem above, but also
> >> > ensure uclamp_max now considers the inverted capacity. Force fitting
> >> > a task when a CPU is in this adverse state will contribute to making the
> >> > thermal throttling last longer.
> >> >
> >> > Signed-off-by: Qais Yousef <qais.yousef@xxxxxxx>
> >> > ---
> >> >  kernel/sched/fair.c | 14 +++++++++-----
> >> >  1 file changed, 9 insertions(+), 5 deletions(-)
> >> >
> >> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >> > index cb32dc9a057f..77ae343e32a3 100644
> >> > --- a/kernel/sched/fair.c
> >> > +++ b/kernel/sched/fair.c
> >> > @@ -4293,12 +4293,16 @@ static inline int util_fits_cpu(unsigned long util,
> >> >        * For uclamp_max, we can tolerate a drop in performance level as the
> >> >        * goal is to cap the task. So it's okay if it's getting less.
> >> >        *
> >> > -	 * In case of capacity inversion, which is not handled yet, we should
> >> > -	 * honour the inverted capacity for both uclamp_min and uclamp_max all
> >> > -	 * the time.
> >> > +	 * In case of capacity inversion we should honour the inverted capacity
> >> > +	 * for both uclamp_min and uclamp_max all the time.
> >> >        */
> >> > -	capacity_orig = capacity_orig_of(cpu);
> >> > -	capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
> >> > +	capacity_orig = cpu_in_capacity_inversion(cpu);
> >> > +	if (capacity_orig) {
> >> > +		capacity_orig_thermal = capacity_orig;
> >> > +	} else {
> >> > +		capacity_orig = capacity_orig_of(cpu);
> >> > +		capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
> >> > +	}
> >> >
> >>
> >> IIUC the rq->cpu_capacity_inverted computation in update_cpu_capacity() can be
> >> summarised as:
> >>
> >> - If there is a PD with equal cap_orig, but higher effective (orig - thermal)
> >>   capacity
> >>   OR
> >>   there is a PD with pd_cap_orig > cpu_effective_cap:
> >>   rq->cpu_capacity_inverted = capacity_orig - thermal_load_avg(rq)
> >>
> >> - Else:
> >>   rq->cpu_capacity_inverted = 0
> >>
> >> Then, the code above uses either rq->cpu_capacity_inverted if it is
> >> non-zero, otherwise:
> >>
> >>   capacity_orig - arch_scale_thermal_pressure(cpu);
> >>
> >> Why use average thermal pressure in one case, and use instantaneous
> >> thermal pressure in the other?
> >
> > There was a big debate on [1] about using avg vs instantaneous.
> >
> 
> Interesting thread, thanks for the link!
> 
> > I used avg for detecting inversion to be consistent with using average in in
> > scale_rt_capacity(). I didn't want the inversion state to be flipping too
> > quickly too.
> >
> > I used the instantaneous in the other check based on that discussion. It seemed
> > using the average is hurtful when for example the medium drops an OPP and by
> > not reacting quickly at wake up we lose the chance to place it on a big; which
> > if my memory didn't fail me is what Xuewen was seeing.
> >
> 
> OK So IIUC by using the inst. pressure you start excluding CPUs sooner, and
> with the avg pressure you keep those CPUs out (if the pressure remained
> long enough).

Yes. I hope the discussion on avg vs instantaneous will be continued and we can
unify the usages.

> 
> > [1] https://lore.kernel.org/lkml/24631a27-42d9-229f-d9b0-040ac993b749@xxxxxxx/
> >
> 
> >>
> >> Can't we get rid of rq->cpu_capacity_inverted and replace this whole thing
> >> with an unconditional
> >>
> >>   capacity_orig_thermal = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
> >>
> >> ?
> >
> > I can't see how we end up with equivalent behavior then. Or address the
> > concerns raised by Xuewen and Lukasz on the RT thread in regards to avg vs
> > instantaneous.
> >
> > Specifically, if we don't use the new rq->cpu_capacity_inverted we can't handle
> > the case where the task is requesting to run at maximum performance but a small
> > drop in thermal pressure means it won't fit anywhere. That PD is the best fit
> > until it hits an inversion.
> >
> > Originally I wanted to defer handling thermal pressure into a different series.
> > But Vincent thought it's better to handle it now. We want more data points from
> > more systems tbh. But I think what we have now is still a good improvement over
> > what we had before.
> >
> > Lukasz had a patch [2] which could allow making thermal_load_avg() more
> > acceptable for systems that care about faster response times.
> >
> > [2] https://lore.kernel.org/lkml/20220429091245.12423-1-lukasz.luba@xxxxxxx/
> >
> >
> > Thanks
> >
> > --
> > Qais Yousef
>