Re: [PATCH 8/8] thermal/drivers/cpu_cooling: Add the combo cpu cooling device

[Daniel Lezcano]

On 05/02/2018 05:17, Viresh Kumar wrote:
> On 02-02-18, 15:30, Daniel Lezcano wrote:
>> On 02/02/2018 11:42, Viresh Kumar wrote:
>>> Here is how I see the whole thing now:
>>>
>>> - Yes we need individual support for both cpufreq and cpuidle cooling devices,
>>>   and no one disagrees on that I believe.
>>>
>>> - There is nothing in the thermal framework that disallows both cpufreq and
>>>   cpuidle cooling devices to co-exist. Both would be part of the same thermal
>>>   zone and so will get throttled with the same thermal sensor event. And so we
>>>   will end up trying to cool down the SoC using both cpufreq and cpuidle
>>>   technique.
>>
>> No. It does not work because we will need different state for each
>> cooling device and we need some logic behind.
> 
> Right, but I thought the cooling-maps can help us specify different cooling
> states for different cooling devices for the same trip point. Maybe my
> understanding of that is incorrect.
> 
>>> - Now I am just wondering if we really need the "combo" functionality or not.
>>>   Can we fine tune the DT cpu-cooling properties (existing ones) for a platform,
>>>   so that it automatically acts as a combo cooling device? I am not 100% sure
>>>   its gonna fly, but just wanted to make sure its not possible to work around
>>>   with and then only try the combo device thing.
>>>
>>> For example, suppose that with just cpufreq-cooling device we need to take the
>>> CPU down to 1 GHz from 2 GHz if we cross temperature 'X'. What if we can change
>>> this policy from DT and say the cpufreq-cooling device goes to 1.5 GHz and
>>> cpuidle-cooling device takes us to idle for 'y' us, and the effect of
>>> combination of these two is >= the effect of the 1 GHz for just the
>>> cpufreq-cooling device.
>>>
>>> Is there any possibility of this to work ?
>>
>> It does not make sense. The combo does that automatically by computing
>> the power equivalence more precisely.
> 
> Sure, but that works by creating a virtual combo-cooling device instead of two
> separate cooling devices and then there are several limitation (at least right
> now) where it doesn't sense the real situation automagically. For example I
> would expect the combo to just work with cpuidle if cpufreq isn't present and as
> soon as cpufreq comes in, covert itself to cpufreq+cpuidle. I was just trying to
> present another view at solving the problem at hand, not that one is better
> than the other.
At the first glance, it sounds interesting but I'm afraid that raises
more corner-cases than it solves because we have to take into account
all the combinations: cpuidle=0 && cpufreq=1, cpuidle=1 && cpufreq=0,
cpuidle=1 && cpufreq=1 with dynamic code changes when the cpufreq driver
is loaded/unloaded.

I'm not against this approach as well as merging all the cpu cooling
devices into a single one but that won't be trivial and will need
several iterations before reaching this level of features.

IMO, we should keep the current approach (but handle the cpufreq
loading/unloading) and then iteratively merge all the cooling device
into a single one with policy change at runtime which will automatically
handle the cpufreq load/unload.

However I'm open to suggestion.



-- 
 <http://www.linaro.org/> Linaro.org â Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog