On Wed, 3 Apr 2019, Thomas Gleixner wrote:
> For newer CPUs we might assume that:
>
> 1) The TSC and APIC timer are actually usable
>
> 2) The frequencies can be retrieved from CPUID or MSRs
>
> If #1 and #2 are reliable we can avoid the whole calibration and interrupt
> delivery mess.
>
> That means we need the following decision logic:
>
> 1) If HPET is available in ACPI, boot normal.
>
> 2) If HPET is not available, verify that the PIT actually counts. If it
> does, boot normal.
>
> If it does not either:
>
> 2A) Verify that this is a PCH 300/C240 and fiddle with that ISST bit.
>
> But that means that we need to chase PCH ids forever...
>
> 2B) Shrug and just avoid the whole PIT/HPET magic all over the place:
>
> - Avoid the interrupt delivery check in the IOAPIC code as it's
> uninteresting in that case. Trivial to do.
>
> - Prevent the TSC calibration code from touching PIT/HPET. It
> should do that already when the TSC frequency can be retrieved
> via CPUID or MSR. Should work, emphasis on should ...
>
> See the mess in: native_calibrate_tsc() and the magic tables in
> tsc_msr.c how well that stuff works.
>
> The cpu_khz_from_cpuid() case at seems to not have these
> issues. Knock on wood!
>
> - Prevent the APIC calibration code from touching PIT/HPET. That's
> only happening right now when the TSC frequency comes from
> the MSRs. No idea why the CPUID method does not provide that.
>
> CPUID leaf 0x16 provides the bus frequency, so we can deduce the
> APIC timer frequency from there and spare the whole APIC timer
> calibration mess:
>
> ECX Bits 15 - 00: Bus (Reference) Frequency (in MHz).
>
> It's usually not required on these newer CPUs because they
> support TSC deadline timer, but you can disable that on the
> kernel command line and some implementations of that were
> broken. With that we are back to square one.
>
> So we need to make sure that these things work under all
> circumstances.
>
> Rafael?
And we have to think hard about how we are going to provide that for
backporting in a digestable form. People are supposed to run recent stable
kernels (I'm not talking about dead kernels ...).
Thanks,
tglx
