Re: [RFT][PATCH v4 4/7] cpuidle: Return nohz hint from cpuidle_select()

[Peter Zijlstra]

On Mon, Mar 12, 2018 at 10:54:18AM +0100, Rafael J. Wysocki wrote:
> @@ -378,6 +384,26 @@ static int menu_select(struct cpuidle_dr
>  	if (idx == -1)
>  		idx = 0; /* No states enabled. Must use 0. */
>  
> +	if (drv->states[idx].flags & CPUIDLE_FLAG_POLLING) {
> +		*nohz_ret = false;
> +	} else if (drv->states[idx].target_residency < TICK_USEC_HZ) {
> +		first_idx = idx;
> +		for (i = idx + 1; i < drv->state_count; i++) {
> +			if (!drv->states[i].disabled &&
> +			    !dev->states_usage[i].disable) {
> +				first_idx = i;
> +				break;
> +			}
		}
> +
> +		/*
> +		 * Do not stop the tick if there is at least one more state
> +		 * within the tick period range that could be used if longer
> +		 * idle duration was predicted.
> +		 */
> +		*nohz_ret = !(first_idx > idx &&
> +			      drv->states[first_idx].target_residency < TICK_USEC_HZ);


Why though? That comment only states what it does, but gives no clue as
to why we're doing this. What was wrong with disabling NOHZ if the
selected state (@idx) has shorter target residency.


> +	}
> +
>  	data->last_state_idx = idx;
>  
>  	return data->last_state_idx;
>