Re: [PATCH] cpufreq, intel_pstate, set max_sysfs_pct and min_sysfs_pct on governor switch

[Prarit Bhargava <prarit@redhat.com>]

On 10/07/2015 06:26 PM, Doug Smythies wrote:
> On 2015.10.07 15:06 Rafael J. Wysocki wrote:
>> On Wednesday, October 07, 2015 05:31:25 PM Prarit Bhargava wrote:
>>> On 10/07/2015 02:52 PM, Doug Smythies wrote:
>>>> On 2015.10.07 08:46 Prarit Bhargava wrote:
>>>>> On 10/07/2015 11:40 AM, Doug Smythies wrote:
>>>>>>
>>>>>> Do we agree or disagree that the root issue seems to be (from your test)?:
>>>>>>
>>>>>> \#  echo 100 > /sys/devices/system/cpu/intel_pstate/min_perf_pct
>>>>>>
>>>>>> [   21.483436] store_min_perf_pct[453] min_sysfs_pct = 100
>>>>>> [   21.489373] store_min_perf_pct[456] min_perf_pct = 100
>>>>>> [   21.495203] store_min_perf_pct[459] min_perf_pct = 100
>>>>>> [   21.501050] store_min_perf_pct[462] min_perf_pct = 100
>>>>>
>>>>> Yep, and it appears to be done by default in Fedora & RHEL :/ ... the issue is
>>>>> still the same IMO that min_sysfs_pct & max_sysfs_pct are not cleared on a
>>>>> governor switch.
>>>>
>>>> Clearing them will break some other things. For example, and as
>>>> shown in my original reply, resume from suspend.
>>>>
>>>> Why? Because, at least on my computer, the governor is changed to
>>>> "performance" during suspend, and the "powersave" governor is
>>>> restored sometime during resume. The users wants the settings they had
>>>> before the suspend.
>>>>
>>> Looking at this in more detail after having tested on a Intel(R) Core(TM)
>>> i7-2600 CPU @ 3.40GHz in Fedora and RHEL.
>>>
>>> I have a feeling that the switch you're seeing (poweersave->performance, suspend
>>> ... resume, performance->powersave) is occurring in userspace, and not as a
>>> result of the kernel.
> Agreed. It is pm-suspend doing it.
> 
>>>  IMO if userspace changes the governor, all bets are off
>>> on maintaining max_sysfs_pct and min_sysfs_pct.
>>>
>>> Here's something I cannot figure out (because I do not have an Ubuntu install).
>>>  *Why* is Ubuntu making the governor switch during suspend/resume?  Is it
>>> because of archaic brokeness they were trying to paper over?
> 
>>> That's not limited to Ubuntu, pm-utils has been doing that forever.
> 
> Agreed. This in pm-utils, and not limited to Ubuntu.
> We can ignore this issue if everyone wants, but I can envision bug reports.
> 

Yeah -- I'd rather not just ignore it.  Sorry to repeat myself but I just sent
out a reply that said I'm going to submit a patch to pm-utils that blacklists
the intel-pstate driver.  Suspend/Resume is known to work with intel-pstate and
we shouldn't be doing anything fancy there.  I will put it on my TODO list to
also later on take a look at acpi-cpufreq.

>> I have no idea why has it been doing that, though.  I guess the reason
>> was to "speed up" PM transitions (in case it started when you were in a
>> low-frequency P-state and then there was no time to bump it up before
>> things got too far).
> 
> I have no idea either, but the stated theory seems sound.

Looking at various comments here and there online, it sounds like they wanted to
put the system into a "sane state" before attempting to suspend/resume.  The
code has been untouched since 2008.  We've come a long way and I'm willing to
argue that we don't need it anymore.

P.
> 
>