[PATCH v4 0/5] Thermal devfreq cooling improvements with Energy Model

[Daniel Lezcano]

On 11/12/2020 16:11, Lukasz Luba wrote:
> Hi Daniel,
> 
> Do you think it has chance to go to as material for v5.11?

Yes, it is in the thermal/linux-next material ATM.


> On 12/10/20 2:30 PM, Lukasz Luba wrote:
>> Hi all,
>>
>> This patch set is a continuation of my previous work, which aimed
>> to add Energy Model to all devices [1]. This series is a follow up
>> for the patches which got merged to v5.9-rc1. It aims to change
>> the thermal devfreq cooling and use the Energy Model instead of
>> private power table and structures. The power model is now simplified,
>> static power and dynamic power are removed. The new registration
>> interface
>> in the patch 3/5 helps to register devfreq cooling and the EM in one
>> call.
>> There is also small improvement, patch 2/5 is changing the way how
>> thermal gets the device status (now uses a copy) and normalize the
>> values.
>> The last patch is here for consistency and will probably go through
>> drm tree.
>>
>> The patch set should apply on top of thermal/testing. It does not
>> depend on
>> new EM API change which is queued in the pm/linux-next tree as v5.11
>> material.
>> Thus, could go in parallel. That was the main motiviation for this v4.
>>
>> changes:
>> v4:
>> - patch 3/5 - removed dependency on the EM API change
>> -- removed em_dev_register_perf_domain() and just use
>>     dev_pm_opp_of_register_em() which API has not changed
>> -- removed a helper registration function and renamed
>>     devfreq_cooling_em_register_power() to devfreq_cooling_em_register()
>>     (was actually suggested by Ionela during review)
>> -- moved energy_model.h to include in devfreq_cooling.c not .h, since
>>     there is no EM structure in there anymore
>> - adjusted comments and commit messages
>> v3 [4]:
>> - dropped direct check of device status and used just a copy of 'status';
>>    a separate patch set will be proposed to address this issue
>> - modified _normalize_load() and used 1024 scale to handle ms, us, ns
>> - removed 'em_registered' and called em_dev_unregister_perf_domain()
>>    unconditionally, so the drivers will have to make sure the right
>> order of
>>    all unregister calls to frameworks which might use EM; this call
>> must be last
>>    one; a proper comment added
>> - removed 'em' pointer from struct devfreq_cooling_device,
>> 'dev->em_pd' is used
>> - removed of_node_get/put(), since the code can handle it
>> - removed dfc_em_get_requested_power() (as missed to do it in v2)
>> - collected all Reviewed-by tags
>> v2 [3]:
>> - renamed freq_get_state() and related to perf_idx pattern as
>>    suggested by Ionela
>> v1 [2]
>>
>> Regards,
>> Lukasz Luba
>>
>> Lukasz Luba (5):
>>    thermal: devfreq_cooling: change tracing function and arguments
>>    thermal: devfreq_cooling: use a copy of device status
>>    thermal: devfreq_cooling: add new registration functions with Energy
>>      Model
>>    thermal: devfreq_cooling: remove old power model and use EM
>>    drm/panfrost: Register devfreq cooling and attempt to add Energy Model
>>
>>   drivers/gpu/drm/panfrost/panfrost_devfreq.c |   2 +-
>>   drivers/thermal/devfreq_cooling.c           | 391 +++++++++-----------
>>   include/linux/devfreq_cooling.h             |  27 +-
>>   include/trace/events/thermal.h              |  19 +-
>>   4 files changed, 198 insertions(+), 241 deletions(-)
>>


-- 
<http://www.linaro.org/> Linaro.org │ 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