Re: [PATCH] sched/fair: Prevent cpu_busy_time from exceeding actual_cpu_capacity

[Xuewen Yan]

On Thu, Jun 20, 2024 at 2:03 AM Qais Yousef <qyousef@xxxxxxxxxxx> wrote:
>
> On 06/19/24 10:46, Xuewen Yan wrote:
> > On Tue, Jun 18, 2024 at 10:58 PM Qais Yousef <qyousef@xxxxxxxxxxx> wrote:
> > >
> > > On 06/17/24 12:03, Qais Yousef wrote:
> > >
> > > > > Sorry, I miss the "fits_capacity() use capacity_of()", and without
> > > > > uclamp_max, the rd is over-utilized,
> > > > > and would not use feec().
> > > > > But I notice the uclamp_max, if the rq's uclamp_max is smaller than
> > > > > SCHED_CAPACITY_SCALE,
> > > > > and is bigger than actual_cpu_capacity, the util_fits_cpu() would
> > > > > return true, and the rd is not over-utilized.
> > > > > Is this setting intentional?
> > > >
> > > > Hmm. To a great extent yes. We didn't want to take all types of rq pressure
> > > > into account for uclamp_max. But this corner case could be debatable.
> > > >
> > > > Is this the source of your problem? If you change util_fits_cpu() to return
> > > > false here, would this fix the problem you're seeing?
> > >
> > > FWIW, if this happens due to uclamp_max, then this patch to do the capping is
> > > still needed.
> > >
> > > I think it's good to understand first how we end up in feec() when a CPU is
> > > supposed to be overutlized. uclamp_max is the only way to override this
> > > decision AFAICT..
> >
> > Sorry for the late reply...
> > In our own tree, we removed the check for rd overutil in feec(), so
> > the above case often occurs.
>
> Yeah, the current definition of overutilized is not good enough. So I can see
> the need to do that..
>
> > And now it seems that on the mainline, uclamp_max is the only way to
> > override this.
>
> Thanks for checking. Your patch is correct and if we combine this with another
> patch to use get_actual_cpu_capacity() in util_fits_cpu(), they are good
> improvements.
>
> Are you still happy to send an updated patches?

Okay, I'll send the new patches soon:)

Thanks!

--
xuewen