Re: [cpufreq-4.7] New schedutil governor in combination with intel_pstate/x86

["Rafael J. Wysocki" <rafael.j.wysocki@intel.com>]

On 5/18/2016 1:16 AM, Sedat Dilek wrote:
> On 5/17/16, Rafael J. Wysocki <rafael.j.wysocki@intel.com> wrote:
>> On 5/17/2016 5:30 PM, Sedat Dilek wrote:
>>> Hi,
>> Hi,
>>
>> You have a broken address of linux-pm (it leads to nowhere).
>>
> Grrr, yeah you are right.
>
>>> I wonder if the Kconfig logic is correct when using
>>> CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL=y in combination with
>>> intel_pstate driver for X86 architecture
>>>
>>> The available scaling-governors for intel_pstate are: performance
>>> powersave.
>>>
>>> Only acpi-cpufreq can use the new schedutil governor.
>> The right way to say that would be that intel_pstate doesn't work with
>> cpufreq governors.
>>
>> scalling_governor is just used by intel_pstate as an interface for
>> choosing the mode it works in and it always uses its own built-in
>> "governor".  And BTW, intel_pstate "powersave" is not the same as the
>> cpufreq's "powersave" even.
>>
>>> Shouldn't it be prevented from selecting schedutil governor as default
>>> when CONFIG_X86_INTEL_PSTATE=y is set?
>> That should apply to "ondemand" too and it doesn't, so no.
>>
>> "schedutil" is just a regular cpufreq governor and it doesn't care about
>> intel_pstate.
>>
>>> It makes no real sense to me.
>>>
>>> Ubuntu has in it's kernels CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE=y.
>>>
>>> What are the alternatives?
>>> Set CONFIG_CPU_FREQ_DEFAULT_GOV_USERSPACE=y.
>> You can choose every cpufreq governor as the default.  There are 6 of
>> then now in the tree.
>>
>>> BTW, I cannot change my scaling-driver...
>>> ...intel_pstate -> acpi-cpufreq...
>>>
>>> root# for p in policy0 policy1 policy2 policy3 ; do LC_ALL=C echo -n
>>> acpi-cpufreq > /sys/devices/system/cpu/cpufreq/${p}/scaling_driver ;
>>> done
>>>
>>> Attached is also a modified /etc/init.d/ondemand script from my
>>> Ubuntu/precise AMD64 to use and fallback to schedutil governor.
>> No, you can't change the scaling driver this way.
>>
>> It actually is not possible to change from intel_pstate to anything else
>> now.  To run without intel_pstate, you need to add
>> "intel_pstate=disable" to the kernel command line.
>>
> OK, I will try intel_pstate=disable in my k-c-l.
>
> Is there no way to switch between acpi-cpufreq and intel_pstate on the fly?

No, there's no way to do that ATM.

> Did not look in the Kconfig sections...
> Can I built both as modules?
> Then unload and load the other?

No, intel_pstate can't be built as a module.

Thanks,
Rafael