Re: [PATCH V4 4/4] thermal/drivers/cpu_cooling: Rename to cpufreq_cooling

From: Martin Kepplinger
Date: Tue Dec 10 2019 - 03:57:44 EST




On 09.12.19 20:29, Daniel Lezcano wrote:
> On 09/12/2019 13:03, Daniel Lezcano wrote:
>> On 09/12/2019 10:54, Martin Kepplinger wrote:
>>>
>>>
>>> On 06.12.19 15:15, Daniel Lezcano wrote:
>>>> On 06/12/2019 12:33, Martin Kepplinger wrote:
>>>>> I tested this on the librem5-devkit and see the
>>>>> cooling devices in sysfs. I configure ARM_PSCI_CPUIDLE, not ARM_CPUIDLE and
>>>>> add the patch below in register the cooling device there. "psci_idle"
>>>>> is listed as the cpuidle_driver.
>>>>>
>>>>> That's what I'm running, in case you want to see it all:
>>>>> https://source.puri.sm/martin.kepplinger/linux-next/commits/next-20191205/librem5_cpuidle_mainline_atf
>>>>>
>>>>> so I add a trip temperature description like this:
>>>>> https://source.puri.sm/martin.kepplinger/linux-next/commit/361f49f93ae2c477fd012790831cabd0ed976660
>>>>>
>>>>> When I let the SoC heat up, cpuidle cooling won't kick it. In sysfs:
>>>>>
>>>>> catting the relevant files in /sys/class/thermal after heating up,
>>>>> if that makes sense:
>>>>>
>>>>> 87000
>>>>> 85000
>>>>> 85000
>>>>> thermal-cpufreq-0
>>>>> 1
>>>>> thermal-idle-0
>>>>> 0
>>>>> thermal-idle-1
>>>>> 0
>>>>> thermal-idle-2
>>>>> 0
>>>>> thermal-idle-3
>>>>> 0
>>>>>
>>>>> with ARM_CPUIDLE instead of ARM_PSCI_CPUIDLE (and registering the cooling dev
>>>>> during cpuidle-arm.c init) I won't have a cpuidle driver and thus no cpu-sleep
>>>>> state at all.
>>>>>
>>>>> Can you see where the problem here lies?
>>>>
>>>> Yes, I removed the registration via the DT.
>>>>
>>>> Can you try the following:
>>>>
>>>> diff --git a/drivers/cpuidle/dt_idle_states.c
>>>> b/drivers/cpuidle/dt_idle_states.c
>>>> index d06d21a9525d..01367ddec49a 100644
>>>> --- a/drivers/cpuidle/dt_idle_states.c
>>>> +++ b/drivers/cpuidle/dt_idle_states.c
>>>> @@ -13,6 +13,7 @@
>>>> #include <linux/errno.h>
>>>> #include <linux/kernel.h>
>>>> #include <linux/module.h>
>>>> +#include <linux/cpu_cooling.h>
>>>> #include <linux/of.h>
>>>> #include <linux/of_device.h>
>>>>
>>>> @@ -205,6 +206,9 @@ int dt_init_idle_driver(struct cpuidle_driver *drv,
>>>> err = -EINVAL;
>>>> break;
>>>> }
>>>> +
>>>> + cpuidle_of_cooling_register(state_node, drv);
>>>> +
>>>> of_node_put(state_node);
>>>> }
>>>>
>>>> That's a hack for the moment.
>>>>
>>>
>>> thanks. I could test that successfully. The only question would be: Is
>>> is intentional how "non-aggressive" the cooling driver cools? I would
>>> have expected it to basically inject more idle cycles earlier. I'd set
>>> 75 degrees as trip point and at 85 degress is would only inject about 30
>>> (of 100).
>
> By the way, how many CPUs are injecting idle cycle when the mitigation
> happens ?

all 4 are injecting the same.

>
>>> You describe the "config values" in question in the documentation, but
>>> I'm not sure what's the correct way to change them.
>>
>> That is difficult to say without knowing the board behavior. Are you
>> able to profile the temperature with the load? How fast the temperature
>> increases? The aggressive behavior of the cooling device will depend on
>> the governor which depends on the slope of the temperature increase and
>> the sampling.
>>
>> Can you give the pointer to the git tree with the DT definition of your
>> board?

https://source.puri.sm/martin.kepplinger/linux-next/blob/next-20191205/librem5_cpuidle_mainline_atf/arch/arm64/boot/dts/freescale/imx8mq-librem5-devkit.dts

you can browse in that branch.

>>
>> You can try by changing the idle duration to 10ms instead of the default
>> 4ms.

where is that set?

>>
>> You can also change the cooling states in the DT <&state 20 70>, so it
>> will begin to mitigate at state 20. But I wouldn't recommend that.

where would we assign that? I'm not sure who reads that -.-
it's still something to consider, but a longer idle duration makes more
sense, yes.

>>
>> Do you have the energy power model, so we can try with the IPA governor?
>>
>>

thanks for the reminder. I'd look at that later.

martin