Re: [PATCH v2] Force cppc_cpufreq to report values in KHz to fix user space reporting

From: rwells
Date: Tue Apr 26 2016 - 14:46:47 EST

On 2016-04-19 18:53, Al Stone wrote:
On 04/19/2016 02:12 PM, Ashwin Chaugule wrote:
+ Ryan

Hi Al,

On 18 April 2016 at 20:11, Al Stone <ahs3@xxxxxxxxxx> wrote:
When CPPC is being used by ACPI on arm64, user space tools such as
cpupower report CPU frequency values from sysfs that are incorrect.

What the driver was doing was reporting the values given by ACPI tables
in whatever scale was used to provide them. However, the ACPI spec
defines the CPPC values as unitless abstract numbers. Internal kernel
structures such as struct perf_cap, in contrast, expect these values
to be in KHz. When these struct values get reported via sysfs, the
user space tools also assume they are in KHz, causing them to report
incorrect values (for example, reporting a CPU frequency of 1MHz when
it should be 1.8GHz).

While the investigation for a long term fix proceeds (several options
are being explored, some of which may require spec changes or other
much more invasive fixes), this patch forces the values read by CPPC
to be read in KHz, regardless of what they actually represent.

The downside is that this approach has some assumptions:

(1) It relies on SMBIOS3 being used, *and* that the Max Frequency
value for a processor is set to a non-zero value.

(2) It assumes that all processors run at the same speed. This
patch retrieves the first CPU Max Frequency from a type 4 DMI
record that it can find. This may not be an issue, however, as a
sampling of DMI data on x86 and arm64 indicates there is often only
one such record regardless.

For arm64 servers, this may be sufficient, but it does rely on
firmware values being set correctly. Hence, other approaches are
also being considered.

This has been tested on three arm64 servers, with and without DMI, with
and without CPPC support.

Changes for v2:
-- Corrected thinko: needed to have DEPENDS on DMI in Kconfig.arm,
not SELECT DMI (found by build daemon)

Signed-off-by: Al Stone <ahs3@xxxxxxxxxx>

This looks like a good short term solution. Does it make more sense to
move this to the cppc_cpufreq driver though? Since that ties more
closely into the cpufreq framework which requires the kHz values in
sysfs. That way we can keep the cppc_acpi.c shim compliant with the
ACPI spec. (i.e. values read in cppc structures remain abstract and
unitless).

Perhaps. Doing it that way made the patch a bit messier since
cppc_acpi.c would set values that then had to be replaced in
cppc_cpufreq.c, so initialization looked odd to me; that's how
I ended up here. You do raise a good point, however; I'll look
at that approach again since I could have missed an easier way
to do it.

Rafael, Viresh, others,

Any other ideas how to handle this better in the long term?

- Decouple the cpufreq sysfs from the cppc driver and introduce its
own entries. Is it possibly to do this cleanly while still allowing
usage of cpufreq registration with existing governors?

- Come up with a scaling factor using the PMU cycle counter at boot
before the CPPC drivers are initialized. This would use the current
freq set by some UEFI var. This would possibly require some messy
perfevents plumbing and added bootup time though.

- .. ?


Cheers,
Ashwin.


The other thought that occurs to me is to go back through the
perf_cap and cpufreq structs and make them more general -- perhaps
store the units being used and pointers to functions to convert them
to KHz. This may require separating sysfs data for perf_cap from the
cpufreq sysfs data from the cppc sysfs data. But, if units are then
reported out to sysfs, user space tools can do whatever conversions
they want, or at least know what they're reporting instead of there
being an implicit ABI between the kernel and the tools. This would
be a far more invasive patch set, I think, but it still may be the
right thing to do for the long term.

The issue is a little more fundamental than that even. We are retrofitting a performance management interface (CPPC) into a frequency management framework (cpufreq) and accompanying tools. Regardless of what scheme we come up with for deriving/exposing frequency, we still havenât completely solved the problem as that assumes a linear relationship between freq and performance. This will work for many but not necessarily all CPPC systems. In fact, making that assumption is explicitly forbidden in the ACPI spec: "OSPM must make no assumption about the exact meaning of the performance values presented by the platform, or how they may correlate to specific hardware metrics like processor frequency."

So to be completely consistent with the current spec, we would need to ween the tools off of frequency altogether and move to abstract performance - either specifically when CPPC driver is loaded or more generally. If we think reporting frequency is required that might still be doable, but would need to be separate interface from the CPPC perf scale. But I agree with Al that is a more invasive change.

For the time being, I don't think it is unreasonable to assume performance is linear with frequency and come up with a scaling factor via one of several mechanisms:
1) SMBIOS as Al proposed (caveats above)
2) Measure at boot using PMU or other mechanism as Ashwin floated (more complicated but removes dependency on SMBIOS and assumption that freq scale is same across all CPUs)
3) Just hardcode a CPPC perf to kHz mapping - maybe everyone is using MHz today? (simplest but obviously least flexible)
4) Others?
None of these are full solution - it is a question of how many different scenarios do we need to cover with initial solution. I think the SMBIOS one is probably a good simplicity/flexibility compromise.

-Ryan

Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project