Re: [RFC] x86, tsc: Add kcmdline args for skipping tsc calibration sequences

From: patrickg
Date: Fri Jul 20 2018 - 18:27:38 EST


Sorry for the delay. Expect another large delay if you have any questions. I'm pretty heavily context switching.

I wanted to double check to make sure that I wasn't mis-documenting and mis-remembering things.

On 07/13/2018 07:40 PM, Brown, Len wrote:
> We disabled CPUID-based TSC calibration on SKX in December for several reasons.
> If you still have it enabled, you need this patch:
>
> commit b511203093489eb1829cb4de86e8214752205ac6
> x86/tsc: Fix erroneous TSC rate on Skylake Xeon
So, yeah. I tested against mainline RHEL-alike elrepo builds before and I still saw the TSC running faster.

I've also tested against 3.10.0-862.6.3 which has those patches backported.

>
> If you are referring to another platform that has CPUID-TSC calibration...
> it should still work on an over-clocked system. Over-clocked platforms should
> use exactly the same reference crystal as non-overclocked platforms, but should
> modify the crystal/core multiplier. If you are changing the reference
> crystal, then I believe you are using an un-supported hardware configuration,
> and my ability to support you is limited.
FYI for reference this is SKX Server. Specifically the `gold` series procs.

Now; I'm not sure if we happen to be doing something strange in regards to changing the ref crystal. I'll need to poke at them to figure that out.

I'm working on building something up with a lot of verbosity so that I can see if perhaps something is happening or not happening in an expected way.