Re: kernel 3.7+ cpufreq regression on AMD system running as dom0

[Matt Wilson <msw@amazon.com>]

On Mon, Jan 14, 2013 at 04:58:54PM +0100, Stefan Bader wrote:
> Starting with kernel v3.7 the following commit added a quirk
> to obtain the real frequencies of certain AMD systems:
> 
> commit f594065faf4f9067c2283a34619fc0714e79a98d
> Author: Matthew Garrett <mjg@redhat.com>
> Date:   Tue Sep 4 08:28:06 2012 +0000
> 
>     ACPI: Add fixups for AMD P-state figures
> 
> When running bare-metal, on my Opteron 6128 test box results
> in the frequencies remaining effectively unchanged:
> [    5.475735] P0: MSR(hi,lo): 8000015c-50004004
> [    5.479049] P0: fid=0x4, did=0x0, freq: 2000 -> 2000
> [    5.484001] P1: MSR(hi,lo): 8000014c-50004a4e
> [    5.487314] P1: fid=0xe, did=0x1, freq: 1500 -> 1500
> [    5.492272] P2: MSR(hi,lo): 80000141-50005048
> [    5.495584] P2: fid=0x8, did=0x1, freq: 1200 -> 1200
> [    5.500540] P3: MSR(hi,lo): 80000138-50005844
> [    5.503853] P3: fid=0x4, did=0x1, freq: 1000 -> 1000
> [    5.508812] P4: MSR(hi,lo): 80000131-50005c40
> [    5.512125] P4: fid=0x0, did=0x1, freq: 800 -> 800
> 
> However running as dom0 under Xen 4.2, reading this MSR returns
> null:
> [   11.613068] P0: MSR(hi,lo): 00000000-00000000
> [   11.613074] P0: fid=0x0, did=0x0, freq: 2000 -> 1600
> [   11.613078] P1: MSR(hi,lo): 00000000-00000000
> [   11.613081] P1: fid=0x0, did=0x0, freq: 1500 -> 1600
> [   11.613085] P2: MSR(hi,lo): 00000000-00000000
> [   11.613088] P2: fid=0x0, did=0x0, freq: 1200 -> 1600
> [   11.613091] P3: MSR(hi,lo): 00000000-00000000
> [   11.613094] P3: fid=0x0, did=0x0, freq: 1000 -> 1600
> [   11.613098] P4: MSR(hi,lo): 00000000-00000000
> [   11.613101] P4: fid=0x0, did=0x0, freq: 800 -> 1600
> 
> And this results in Xen failing to change the governor:
>   "(XEN) Fail change to ondemand governor"
> 
> I suppose this ultimately requires some support in the hypervisor
> to pass through the real values. But since this is at least on my
> combination of Xen 4.2 + kernel v3.7+ and AMD family 0x10 CPU a
> regression compared to older kernels, I wonder whether the following
> change might be something that should go into mainline:

The correct values should be returned already via rdmsr if
"cpureq=dom0-kernel" is specified on the Xen command line. Looking at
the LP report, it doesn't seem that this option was used.

Likely you will also need to use "dom0_vcpu_pin=1" if you want dom0 to
be the cpufreq controller on AMD.

Matt

> --- a/drivers/acpi/processor_perflib.c
> +++ b/drivers/acpi/processor_perflib.c
> @@ -340,6 +340,9 @@ static void amd_fixup_frequency(struct acpi_processor_px *px
>         if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
>             || boot_cpu_data.x86 == 0x11) {
>                 rdmsr(MSR_AMD_PSTATE_DEF_BASE + index, lo, hi);
> +               /* Bit 63 indicates whether contents are valid */
> +               if (!(hi & 0x8000000))
> +                       return;
>                 fid = lo & 0x3f;
>                 did = (lo >> 6) & 7;
>                 if (boot_cpu_data.x86 == 0x10)
> 
> I tested something similar (so hopefully I have not failed on slapping
> together a cleaned up version), which did resolve the problem.
> 
> -Stefan
> 
> Reference: https://bugs.launchpad.net/bugs/1078619