Re: [PATCH 1/2] perf/x86/rapl: Add support for Intel Meteor Lake

[Zhang, Rui]

On Fri, 2023-01-06 at 12:33 +0100, Ingo Molnar wrote:
> * Borislav Petkov <bp@xxxxxxxxx> wrote:
> 
> > On Fri, Jan 06, 2023 at 11:56:18AM +0100, Ingo Molnar wrote:
> > > It's a trade-off in any case: there's a point where quirk flags
> > > or even 
> > > feature flags become harder to read and harder to maintain than
> > > cleanly 
> > > separated per model driver functions.
> > 
> > Yeah, no, singular: a synthetic feature *flag*: X86_FEATURE_RAPL.
> > 
> > cpu/intel.c can set it and driver can test it.
> > 
> > Everything else inside the driver.
> > 
> > Until Intel can get their act together and actually do a CPUID bit
> > like AMD. :-P
> > 
> > But when you think about it, whether the model matching happens in
> > the driver or
> > in cpu/intel.c doesn't matter a whole lot.
> > 
> > All that matters is, they should finally give it a CPUID bit.
> 
> The other thing that matters here are the RAPL *incompatibilities*
> between 
> model variants, which are significant AFAICS.
> 
> With a CPUID we get a kind of semi-compatible hardware interface with
> well 
> defined semantics & expansion.

Agreed.
> 
> With 'non-architectural', per-model RAPL features we get very little
> of 
> that...

Exactly.

The main purpose of the model list in RAPL PMU code and the intel_rapl
driver is to differentiate the model-specific behavior, say,
some models use standard energy unit retrieved from MSR
some models use a fixed energy unit for Dram Domain
and
some models use a fixed energy unit for Psys Domain
etc.

> 
> Which is why it's a trade-off that is hard to judge in advance: maybe
> we 
> can simplify the code via a synthethic CPUID[s], maybe it will just
> be 
> another zoo of per-model feature flags...

Agreed.

> Likely won't be able to tell for sure until we see patches.
> 
Yeah, let me cook up a RFC series later and we can continue with that.

thanks,
rui