Re: [PATCH v2 3/8] thermal: gov_power_allocator: Move memory allocation out of throttle()

[Lukasz Luba <lukasz.luba@arm.com>]

On 12/20/23 14:35, Rafael J. Wysocki wrote:
> On Tue, Dec 12, 2023 at 2:48 PM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>> The new thermal callback allows to react to the change of cooling
>> instances in the thermal zone. Move the memory allocation to that new
>> callback and save CPU cycles in the throttle() code path.
>>
>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
>> ---
>>   drivers/thermal/gov_power_allocator.c | 144 ++++++++++++++++++++------
>>   1 file changed, 113 insertions(+), 31 deletions(-)
>>
>> diff --git a/drivers/thermal/gov_power_allocator.c b/drivers/thermal/gov_power_allocator.c
>> index 38e1e89ba10c..3328c3ec71f2 100644
>> --- a/drivers/thermal/gov_power_allocator.c
>> +++ b/drivers/thermal/gov_power_allocator.c
>> @@ -61,6 +61,13 @@ static inline s64 div_frac(s64 x, s64 y)
>>    *                     @trip_switch_on should be NULL.
>>    * @trip_max:          last passive trip point of the thermal zone. The
>>    *                     temperature we are controlling for.
>> + * @num_actors:                number of cooling devices supporting IPA callbacks
>> + * @buffer_size:       IPA internal buffer size
>> + * @req_power:         IPA buffer for requested power
>> + * @max_power:         IPA buffer for max allocatable power
>> + * @granted_power:     IPA buffer for granted power
>> + * @extra_actor_power: IPA buffer for extra power
>> + * @weighted_req_power:        IPA buffer for weighted requested power
>>    */
>>   struct power_allocator_params {
>>          bool allocated_tzp;
>> @@ -69,6 +76,13 @@ struct power_allocator_params {
>>          u32 sustainable_power;
>>          const struct thermal_trip *trip_switch_on;
>>          const struct thermal_trip *trip_max;
>> +       int num_actors;
>> +       int buffer_size;
> 
> None of the above can be negative, so maybe consider using unsigned int?

True, I'll change them to unsigned.

> 
>> +       u32 *req_power;
>> +       u32 *max_power;
>> +       u32 *granted_power;
>> +       u32 *extra_actor_power;
>> +       u32 *weighted_req_power;
>>   };
>>
>>   /**
>> @@ -387,39 +401,24 @@ static int allocate_power(struct thermal_zone_device *tz, int control_temp)
>>          u32 *weighted_req_power;
>>          u32 power_range, weight;
>>          int total_weight = 0;
>> -       int num_actors = 0;
> 
> You could retain this local var and set it to params->num_actors.  It
> is kind of inconsistent to drop it and still use the local variables
> above.

OK, I'll do that.

[snip]

>> +
>> +       req_power = kcalloc(num_actors * 5, sizeof(u32), GFP_KERNEL);
> 
> I'd use sizeof(*req_power) instead of sizeof(u32) here and below.

OK

> 
>> +       if (!req_power) {
>> +               ret = -ENOMEM;
>> +               goto clean_buffers;
>> +       }
>> +
>> +       params->num_actors = num_actors;
>> +       params->buffer_size = num_actors * 5 * sizeof(u32);
>> +
>> +       _power_buffers_init(params, req_power, &req_power[params->num_actors],
>> +                           &req_power[2 * params->num_actors],
>> +                           &req_power[3 * params->num_actors],
>> +                           &req_power[4 * params->num_actors]);
> 
> Why don't you use the local var in this instead of the struct member?
> It would be easier to read then IMO.
> 
> Also, I would rather use pointer arithmetic, so it would become
> 
> _power_buffers_init(params, req_power, req_power + num_actors,
> req_power + 2 * num_actors, req_power + 3 * num_actors, req_power + 4
> * num_actors);
> 
> And note that you could introduce something like
> 
> struct power_actor {
>          u32 req_power;
>          u32 max_power;
>          u32 granted_power;
>          u32 extra_actor_power;
>          u32 weighted_req_power;
> };
> 
> and use a single array of these instead of 5 different arrays of u32,
> which would result in more straightforward code if I'm not mistaken.

That sounds like a good idea. Let me implement it and see - but it
should be a better way. Thanks!