Re: [PATCH v2] Add hardcoded crystal clock for KabyLake

["Christian A. Ehrhardt" <lk@c--e.de>]

Hi,

On Mon, Nov 14, 2022 at 11:58:54PM +0100, Thomas Gleixner wrote:
> On Thu, Oct 20 2022 at 10:18, Dave Hansen wrote:
> > On 10/20/22 10:13, Peter Zijlstra wrote:
> >> And why, pray *WHY* can't Intel simply write the correct information in
> >> CPUID leaf 15h. I mean, they defined the leaf, might as well use it, no?
> >
> > Is the data that's in the leaf just wrong?  Doesn't that mean that the
> > CPUID leaf on these models is violating the architecture contract?  That
> > sounds like something that deserves an erratum.
> >
> > Is there a documented erratum?
> 
> I don't know. The code has this comment:
> 
> 	/*
> 	 * Some Intel SoCs like Skylake and Kabylake don't report the crystal
> 	 * clock, but we can easily calculate it to a high degree of accuracy
> 	 * by considering the crystal ratio and the CPU speed.
> 	 */

Latest (April 2022) version of the SDM clearly states that the
above comment is wrong. CPUID.16h has the following note:
| Data is returned from this interface in accordance with the processor's
| specification and does not reflect actual values. Suitable use of this
| data includes the display of processor information in like manner to the
| processor brand string and for determining the appropriate range to use
| when displaying processor information e.g. frequency history graphs. The
| returned information should not be used for any other purpose as the
| returned information does not accurately correlate to information /
| counters returned by other processor interfaces.

Thus using CPUID.16h to determine the crystal clock frequency is wrong.
This difference is significant. I have one Kaby Lake latop where
the CPUID.16h reported frequency is 1900MHz but the real frequency is
only 1896MHz. This amounts to a time drift of about 8s/hour if the
wrong TSC frequency is used for time keeping.

Basically, I think this commit:
    604dc9170 (x86/tsc: Use CPUID.0x16 to calculate ...)
needs to be reverted.

> so those SoCs fail to expose clock in leaf 15h and then the information
> in leaf 16h is so inaccurate that the calculation is off.
> 
> Sigh. It's 2022 and we are still relying on crystalball mechanisms to
> figure out the damned crystal frequency.
> 
> The specification of leaf 15h is:
> 
>  15H Time Stamp Counter and Nominal Core Crystal Clock Information Leaf
>     NOTES:
>         If EBX[31:0] is 0, the TSC/”core crystal clock” ratio is not enumerated.
>         If ECX is 0, the nominal core crystal clock frequency is not enumerated.
> 
> IOW, this CPUID leaf is defined to be useless and leaves it up to the
> SoC integration to provide this information or not. It needs even two
> fields to chose from to make it useless...

The SDM (now?) has some hints on how to do this. This is hidden here:

    Vol.3 Chapter 19.7.3: Determining the Processor Base Frequency

This chapter contains a table that lists the correct crystal clock
frequencies for CPU models that do not enumerate it via CPUID.15h.

       regards   Christian