Re: [Patch 2/2] ACPI: processor: Add support to configure CPUFREQ reduction pctg

[Sumit Gupta <sumitg@nvidia.com>]

On 19/08/23 00:10, Rafael J. Wysocki wrote:
> External email: Use caution opening links or attachments
> 
> 
> On Thu, Aug 17, 2023 at 11:31 AM Sumit Gupta <sumitg@nvidia.com> wrote:
>>
>> From: Srikar Srimath Tirumala <srikars@nvidia.com>
>>
>> Add support to configure the CPUFREQ reduction percentage and set the
>> maximum number of throttling steps accordingly. Current implementation
>> of processor_thermal performs software throttling in fixed steps of
>> "20%" which can be too coarse for some platforms. Change that by adding
>> new config to provide the reduction percentage.
>>
>> Signed-off-by: Srikar Srimath Tirumala <srikars@nvidia.com>
>> Signed-off-by: Sumit Gupta <sumitg@nvidia.com>
>> ---
>>   drivers/acpi/Kconfig             | 15 +++++++++++++++
>>   drivers/acpi/processor_thermal.c | 19 ++++++++++++++++---
>>   2 files changed, 31 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/acpi/Kconfig b/drivers/acpi/Kconfig
>> index 00dd309b6682..287cf58defbf 100644
>> --- a/drivers/acpi/Kconfig
>> +++ b/drivers/acpi/Kconfig
>> @@ -254,6 +254,21 @@ config ACPI_DOCK
>>   config ACPI_CPU_FREQ_PSS
>>          bool
>>
>> +config ACPI_CPU_FREQ_THERM_HAS_PARAMS
>> +       bool "CPU frequency throttling control"
>> +       depends on ACPI_PROCESSOR
>> +
>> +config ACPI_CPU_FREQ_THERM_MIN_THROT_PCTG
>> +       int "Minimum throttle percentage for processor_thermal cooling device"
>> +       depends on ACPI_CPU_FREQ_THERM_HAS_PARAMS
>> +       default 20
>> +       help
>> +         The processor_thermal driver uses this config to calculate the
>> +         percentage amount by which cpu frequency must be reduced for each
>> +         cooling state. The config is also used to calculate the maximum number
>> +         of throttling steps or cooling states supported by the driver. Value
>> +         must be an unsigned integer in the range [1, 50].
>> +
> 
> I don't think that the new Kconfig symbols are particularly useful.
> At least they don't help the distro vendors that each would need to
> pick up a specific value for their kernel anyway.
> 
> I also wonder how the users building their own kernels are supposed to
> determine the values suitable for the target systems.
> 

We observed some perf gain after reducing the throttle percentage.
Currently, kept the default to '20%' as before.

Based on need, a vendor can overwrite the default value with macro
'CONFIG_ACPI_CPU_FREQ_THERM_MIN_THROT_PCTG'. Otherwise, the behavior
will remain same.

Thank you,
Sumit Gupta

>>   config ACPI_PROCESSOR_CSTATE
>>          def_bool y
>>          depends on ACPI_PROCESSOR
>> diff --git a/drivers/acpi/processor_thermal.c b/drivers/acpi/processor_thermal.c
>> index b7c6287eccca..ee443cc69b73 100644
>> --- a/drivers/acpi/processor_thermal.c
>> +++ b/drivers/acpi/processor_thermal.c
>> @@ -25,8 +25,20 @@
>>    * _any_ cpufreq driver and not only the acpi-cpufreq driver.
>>    */
>>
>> -#define CPUFREQ_THERMAL_MIN_STEP 0
>> -#define CPUFREQ_THERMAL_MAX_STEP 3
>> +#define CPUFREQ_THERMAL_MIN_STEP       0
>> +#ifdef CONFIG_ACPI_CPU_FREQ_THERM_HAS_PARAMS
>> +#define CPUFREQ_THERMAL_PCTG           CONFIG_ACPI_CPU_FREQ_THERM_MIN_THROT_PCTG
>> +
>> +/* Derive the MAX_STEP from minimum throttle percentage so that the reduction
>> + * percentage does end up becoming negative. Also cap the MAX_STEP so that
>> + * the CPU performance doesn't become 0.
>> + */
>> +#define CPUFREQ_THERMAL_MAX_STEP       ((100 / CPUFREQ_THERMAL_PCTG) - 1)
>> +
>> +#else
>> +#define CPUFREQ_THERMAL_MAX_STEP       3
>> +#define CPUFREQ_THERMAL_PCTG           20
>> +#endif
>>
>>   static DEFINE_PER_CPU(unsigned int, cpufreq_thermal_reduction_pctg);
>>
>> @@ -113,7 +125,8 @@ static int cpufreq_set_cur_state(unsigned int cpu, int state)
>>                  if (!policy)
>>                          return -EINVAL;
>>
>> -               max_freq = (policy->cpuinfo.max_freq * (100 - reduction_pctg(i) * 20)) / 100;
>> +               max_freq = (policy->cpuinfo.max_freq *
>> +                           (100 - reduction_pctg(i) * CPUFREQ_THERMAL_PCTG)) / 100;
>>
>>                  cpufreq_cpu_put(policy);
>>
>> --