Re: Detecting x86 LAPIC timer frequency from CPUID data

From: Thomas Gleixner
Date: Fri Apr 19 2019 - 15:03:25 EST


On Fri, 19 Apr 2019, Daniel Drake wrote:
> On Fri, Apr 19, 2019 at 6:30 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> > Time Stamp Counter/Core Crystal Clock Information (0x15):
> > TSC/clock ratio = 168/2
> > nominal core crystal clock = 0 Hz
> >
> > Processor Frequency Information (0x16):
> > Core Base Frequency (MHz) = 0x834 (2100)
> > Core Maximum Frequency (MHz) = 0xed8 (3800)
> > Bus (Reference) Frequency (MHz) = 0x64 (100)
> >
> > Assuming that TSC and local APIC timer run from the same frequency on these
> > modern machines.
> >
> > 2100MHz * 2 / 168 = 25MHz
> >
> > and disabling the tsc deadline timer tells me:
> >
> > ..... calibration result: 24999
> >
> > Close enough.
>
> I tested all the Intel SoC generations we have on hand. The assumption that
> the core crystal clock feeds the APIC seems to be consistently true.
>
> (Please note that all the following results are done with CONFIG_HZ=250,
> which is why the "calibration result" is 4x higher than HZ=1000 as used
> in previous mails)
>
> In the easy case, the low cost platforms do not support CPUID.0x16 (so no
> CPU frequency reporting), but they do tell us the core crystal clock, which
> is consistent with the APIC calibration result:

...

> And the 4 higher-end SoCs that we have available for testing all report
> crystal clock 0Hz from CPUID 0x15, but by combining the CPUID.0x16 base
> frequency with the CPUID.0x15 TSC/clock ratio, the crystal frequency can
> be calculated as you describe, and it consistently matches the APIC timer
> calibration result.

...

> Is this data convincing enough or should we additionally wait for some
> comments from Intel?

For me it's pretty convincing, but having some confirmation from Intel
wouldn't be a bad thing.

> I came up with the patch below. However, upon testing, I realised that, at
> least for the platforms I have in hand, only the first hunk is really needed.
> We don't need to use your magic calculation to find the crystal frequency
> because Intel already told us! native_calibrate_tsc() already hardcodes the
> crystal frequency for Kabylake, and Amber/Whiskey/Coffee also report the
> 0x8e/0x9e Kabylake model codes.

I'd rather replace these model checks with math. These tables are horrible
to maintain.

> Plus ApolloLake/GeminiLake do report the crystal frequency in CPUID.0x15
> so that is covered too.

> While looking around this code I also spotted something curious.
> In calibrate_APIC_clock() for the case where lapic_timer_frequency has been
> externally provided, we have:
> lapic_clockevent.max_delta_ns =
> clockevent_delta2ns(0x7FFFFF, &lapic_clockevent);
> lapic_clockevent.max_delta_ticks = 0x7FFFFF;
>
> But in the case where we calibrate, we have:
> lapic_clockevent.max_delta_ns =
> clockevent_delta2ns(0x7FFFFFFF, &lapic_clockevent);
> lapic_clockevent.max_delta_ticks = 0x7FFFFFFF;
>
> 0x7FFFFF vs 0x7FFFFFFF, is that intentional?

I don't think so. Looks like a failed copy and paste. Cc'ed Jacob, he might
know.

Thanks,

tglx