Re: [PATCH] x86, smpboot: Disable frequency invariance when it's unsupported
From: Giovanni Gherdovich
Date: Thu Apr 16 2020 - 14:09:40 EST
On Thu, 2020-04-16 at 20:54 +0800, Like Xu wrote:
> On 2020/4/16 16:40, Giovanni Gherdovich wrote:
> > [...]
> > 1) Can you share the output of "turbostat --interval 1 sleep 0"? I'm
> > interested in the headers of the output, where all the various pm-related
> > MSRs are decoded.
> >
>
> I couldn't disclose more information about this.
No worries, I understand.
>
> > 2) Despite not being in the Intel SDM, I was under the assumption that all
> > Intel CPUs declare the "all-cores turbo" frequency, but it's not the case
> > for this one. Eg: if you have 24 cores, somewhere in your MSRs I'd expect
> > to find "24C turbo" (or even "30C turbo", anything greater or equal than 24).
> > My understanding from
> > https://ark.intel.com/content/www/us/en/ark/products/202682/intel-atom-processor-p5962b-27m-cache-2-20-ghz.html
> > is that this CPU doesn't support turbo boost at all; in other CPUs without
> > turbo I've seen MSRs saying the all-cores turbo freq is equal to the base
> > freq (for compatibility I suppose). Here MSR_TURBO_RATIO_LIMIT says that 1C
> > turbo is the same as base frequency (2.2GHz), but turbo for larger sets of
> > cores is declared as zero, which I find a little odd.
>
> That's odd and we could only rely on the Intel specification
> about the assumption "Intel CPUs declare the all-cores turbo frequency"
> and I may report this issue if something does mismatch.
Ok.
>
> >
> > 3) The parsing of MSRs in the frequency invariance code is modeled after
> > turbostat, and classifies CPUs in 5 groups: Atom up to Goldmont, Atom from
> > Goldmont onwards, Xeon Phi, Xeon Scalable Processors onwards and "generic
> > Core". As you've already found out from where your panic happens, your Atom
> > falls into the "generic Core" category (function core_set_max_freq_ratio()),
> > but given that it's an Atom and it's been released this very quarter I'd
> > have guessed it to behave like a Goldmont. Something for me to keep in mind.
>
> It's INTEL_FAM6_ATOM_TREMONT or INTEL_FAM6_ATOM_TREMONT_D.
>
Thanks! The model name from intel-family.h is useful!
Giovanni