Re: [PATCH v4 3/5] x86/umwait: Add sysfs interface to control umwait C0.2 state

[Andy Lutomirski]

On Tue, Jun 11, 2019 at 10:27 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
>
> (can you, perchance, look at a MUA that isn't 'broken' ?)
>
> On Tue, Jun 11, 2019 at 09:04:30AM -0700, Andy Lutomirski wrote:
> >
> >
> > > On Jun 11, 2019, at 1:54 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > >> On Fri, Jun 07, 2019 at 03:00:35PM -0700, Fenghua Yu wrote:
> > >> C0.2 state in umwait and tpause instructions can be enabled or disabled
> > >> on a processor through IA32_UMWAIT_CONTROL MSR register.
> > >>
> > >> By default, C0.2 is enabled and the user wait instructions result in
> > >> lower power consumption with slower wakeup time.
> > >>
> > >> But in real time systems which require faster wakeup time although power
> > >> savings could be smaller, the administrator needs to disable C0.2 and all
> > >> C0.2 requests from user applications revert to C0.1.
> > >>
> > >> A sysfs interface "/sys/devices/system/cpu/umwait_control/enable_c02" is
> > >> created to allow the administrator to control C0.2 state during run time.
> > >
> > > We already have an interface for applications to convey their latency
> > > requirements (pm-qos). We do not need another magic sys variable.
> >
> > Iâm not sure I agree.  This isnât an overall latency request, and
> > setting an absurdly low pm_qos will badly hurt idle power and turbo
> > performance.  Also, pm_qos isnât exactly beautiful.
> >
> > (I speak from some experience. I may be literally the only person to
> > write a driver that listens to dev_pm_qos latency requests. And, in my
> > production box, I directly disable c states instead of messing with
> > pm_qos.)
> >
> > I do wonder whether anyone will ever use this particular control, though.
>
> I agree that pm-qos is pretty terrible; but that doesn't mean we should
> just add random control files all over the place.

I don't think pm-qos is expressive enough.  It seems entirely
reasonable to want to do a C0.1 wait for lower latency *while waiting*
but still want full power-saving idle when not waiting.

Do we even know what the C0.2 and C0.1 latencies are?  And why is this
thing an MSR instead of a flag passed to UMWAIT?

--Andy