Re: [v3] x86/tsc: Unset TSC_KNOWN_FREQ and TSC_RELIABLE flags on Intel Bay Trail SoC

[Thomas Gleixner <tglx@linutronix.de>]

Dave Hansen <dave.hansen@intel.com> writes:
> On 1/29/20 12:57 PM, Thomas Gleixner wrote:
>> Just to make it entirely clear. We are wasting days already due to the
>> fact that Intel, who designs, specifies and most importantly sells these
>> CPUs is either unable or unwilling to provide accurate information about
>> the trivial and essential information to support these CPUs:
>> 
>>     1) The crystal frequency
>> 
>>     2) The nominator/denominator pair to calculate the TSC frequency
>>        from #1
>
> Circling back...  The problem here, as I understand it is that we have
> some of these tables:
>
> static const struct freq_desc freq_desc_byt = {
>         1, { 83300, 100000, 133300, 116700, 80000, 0, 0, 0 }
> };
>
> Where "83300" means "83.3 MHz".  the 83.3 came literally from the SDM.
> Talking to some of the folks who work on the silicon, they confirmed
> that when the SDM says "083.3 MHz", it represents an approximation of
> 2000/24.
> Intel can go through and explain the values more precisely in the
> documentation.  The big-core tables already have more significant
> digits, for instance.  To me, it also seems like the SDM should probably
> just explicitly state the actual ratios rather than a decimal
> approximation.

Yes please.

> But, in the end, the CPU is just enumerating frequencies that are
> derived from crystals outside the CPU.  The hardware in question here
> tended to be put on boards which were not using the highest-end
> components and probably don't have the most accurate crystals.
>
> So, while we can add precision to the numbers in the documentation,
> we're not super confident that it will result in a meaningfully more
> accurate frequency across a big fleet of systems.

Even if you have a cheapo 24MHz crystal it's way less off than the
rounding error in these tables.

Thanks,

        tglx