Re: [PATCH v4 1/2] dt-bindings: cpufreq: add virtual cpufreq device

[Marc Zyngier <maz@kernel.org>]

On Fri, 12 Jan 2024 22:02:39 +0000,
Saravana Kannan <saravanak@google.com> wrote:
> 
> Sorry for the delay in response. Was very busy for a while and then
> holidays started.
> 
> On Fri, Dec 8, 2023 at 12:52 AM Marc Zyngier <maz@kernel.org> wrote:
> >
> > On Thu, 07 Dec 2023 22:44:36 +0000,
> > Saravana Kannan <saravanak@google.com> wrote:
> > >
> > > On Wed, Nov 15, 2023 at 12:49 AM Marc Zyngier <maz@kernel.org> wrote:
> > > >
> > > > On Sat, 11 Nov 2023 01:49:29 +0000,
> > > > David Dai <davidai@google.com> wrote:
> > > > >
> > > > > Adding bindings to represent a virtual cpufreq device.
> > > > >
> > > > > Virtual machines may expose MMIO regions for a virtual cpufreq device
> > > > > for guests to read frequency information or to request frequency
> > > > > selection. The virtual cpufreq device has an individual controller for
> > > > > each frequency domain.
> > > >
> > > > I would really refrain form having absolute frequencies here. A
> > > > virtual machine can be migrated, and there are *zero* guarantees that
> > > > the target system has the same clock range as the source.
> > > >
> > > > This really should be a relative number, much like the capacity. That,
> > > > at least, can be migrated across systems.
> > >
> > > There's nothing in this patch that mandates absolute frequency.
> > > In true KVM philosophy, we leave it to the VMM to decide.
> >
> > This has nothing to do with KVM. It would apply to any execution
> > environment, including QEMU in TCG mode.
> >
> > To quote the original patch:
> >
> > +    description:
> > +      Address and size of region containing frequency controls for each of the
> > +      frequency domains. Regions for each frequency domain is placed
> > +      contiugously and contain registers for controlling DVFS(Dynamic Frequency
> > +      and Voltage) characteristics. The size of the region is proportional to
> > +      total number of frequency domains.
> >
> > What part of that indicates that *relative* frequencies are
> > acceptable? The example explicitly uses the opp-v2 binding, which
> > clearly is about absolute frequency.
> 
> We can update the doc to make that clearer and update the example too.
> 
> > To reiterate: absolute frequencies are not the right tool for the job,
> > and they should explicitly be described as relative in the spec. Not
> > left as a "whatev'" option for the execution environment to interpret.
> 
> I think it depends on the use case. If there's no plan to migrate the
> VM across different devices, there's no need to do the unnecessary
> normalization back and forth.

VM migration is a given, specially when QEMU is involved. Designing
something that doesn't support it is a bug, plain and simple.

> And if we can translate between pCPU frequency and a normalized
> frequency, we can do the same for whatever made up frequencies too. In
> fact, we plan to do exactly that in our internal use cases for this.
> There's nothing here that prevents the VMM from doing that.
>
> Also, if there are hardware virtualized performance counters (AMU,
> CPPC, etc) that are used for frequency normalization, then we have to
> use the real frequencies in those devices otherwise the "current
> frequency" can be 2 GHz while the max normalized frequency is 1024
> KHz. That'll mess up load tracking.

And that's exactly why this shouldn't be a *frequency*, but a
performance scale or some other unit-less coefficient. Just like the
big-little capacity.

	M.

-- 
Without deviation from the norm, progress is not possible.