Re: [PATCH] x86/tsc: Add kernel options to disable CPUID and MSR calibrations

[Prarit Bhargava]

On 2/26/20 6:27 PM, Thomas Gleixner wrote:
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>> On Wed, Feb 26, 2020 at 11:43:08AM -0500, Prarit Bhargava wrote:
>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
>>> index dbc22d684627..0316aadfff08 100644
>>> --- a/Documentation/admin-guide/kernel-parameters.txt
>>> +++ b/Documentation/admin-guide/kernel-parameters.txt
>>> @@ -4942,7 +4942,7 @@
>>>  			See Documentation/admin-guide/mm/transhuge.rst
>>>  			for more details.
>>>  
>>> -	tsc=		Disable clocksource stability checks for TSC.
>>> +	tsc=option[,option...]	Various TSC options.
>>>  			Format: <string>
>>>  			[x86] reliable: mark tsc clocksource as reliable, this
>>>  			disables clocksource verification at runtime, as well
>>> @@ -4960,6 +4960,12 @@
>>>  			in situations with strict latency requirements (where
>>>  			interruptions from clocksource watchdog are not
>>>  			acceptable).
>>> +			[x86] no_cpuid_calibration: Disable the CPUID TSC
>>> +			calibration.  Used in situations where the CPUID
>>> +			TSC khz does not match the actual CPU TSC khz
>>> +			[x86] no_msr_calibration: Disable the MSR TSC
>>> +			calibration.  Used in situations where the MSR
>>> +			TSC khz does not match the actual CPU TSC khz.
>>
>> Do we want to mention that these situations are mostly broken firmware?
>> Also do mention that if you disable these you might not boot due to not
>> having a PIT/HPET at all?
> 
> Right. Same discussion as before.
> 
> Also why do we want no_cpuid_calibration and no_msr_calibration? How


> should Joe User figure out which one to use? This does not make
> sense. The point is that the BIOS/Firmware supplied value in system
> registers is bogus. So something like "skip_firmware_calibration" might
> be better suitable.


no_cpuid_calibration was required for Patrick's case where the CPU was
overclocked and therefore the CPUID khz value was invalid, but the MSR value is
good.  I had to skip both to get to the PIT calibration because I had broken FW.
 I don't see how a single skip_firmware_calibration covers these cases.

> 
> Aside of that this really wants to be combined with the ability to
> supply the actual frequency on the command line as I suggested in the
> other thread to cope with machines which do not expose PIT/HPET or have
> broken variants of them.

tglx, can you give a lore link to the thread?

Thanks,

P.
> 
> Thanks,
> 
>         tglx
> 
>