Re: [PATCH] thermal: power_allocator: using round the division when re-divvying up power

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

On 3/15/21 8:25 AM, gao.yunxiao6@gmail.com wrote:
> From: "jeson.gao" <jeson.gao@unisoc.com>
> 
> The division is used directly in re-divvying up power, the decimal part will
> be discarded, devices will get less than the extra_actor_power - 1.
> if using round the division to make the calculation more accurate.
> 
> For example:
> actor0 received more than it's max_power, it has the extra_power 759
> actor1 received less than it's max_power, it require extra_actor_power 395
> actor2 received less than it's max_power, it require extra_actor_power 365
> actor1 and actor2 require the total capped_extra_power 760
> 
> using division in re-divvying up power
> actor1 would actually get the extra_actor_power 394
> actor2 would actually get the extra_actor_power 364
> 
> if using round the division in re-divvying up power
> actor1 would actually get the extra_actor_power 394
> actor2 would actually get the extra_actor_power 365
> 
> Signed-off-by: Jeson Gao <jeson.gao@unisoc.com>
> ---
>   drivers/thermal/gov_power_allocator.c | 8 +++++---
>   1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/thermal/gov_power_allocator.c b/drivers/thermal/gov_power_allocator.c
> index 92acae53df49..2802a0e13c88 100644
> --- a/drivers/thermal/gov_power_allocator.c
> +++ b/drivers/thermal/gov_power_allocator.c
> @@ -374,9 +374,11 @@ static void divvy_up_power(u32 *req_power, u32 *max_power, int num_actors,
>   	 */
>   	extra_power = min(extra_power, capped_extra_power);
>   	if (capped_extra_power > 0)
> -		for (i = 0; i < num_actors; i++)
> -			granted_power[i] += (extra_actor_power[i] *
> -					extra_power) / capped_extra_power;
> +		for (i = 0; i < num_actors; i++) {
> +			u64 extra_range = (u64)extra_actor_power[i] * extra_power;
> +			granted_power[i] += DIV_ROUND_CLOSEST_ULL(extra_range,
> +							 capped_extra_power);
> +		}
>   }
>   
>   static int allocate_power(struct thermal_zone_device *tz,
> 

Make sense. There is already DIV_ROUND_CLOSEST_ULL() in that function,
so let's align both. There might be a bit overhead on 32bit machines,
but IPA polling isn't so often there.

Reviewed-by: Lukasz Luba <lukasz.luba@arm.com>

Regards,
Lukasz