Re: [PATCH 1/2] sched/fair: Filter out locally-unsolvable misfit imbalances

[Dietmar Eggemann]

On 15/04/2021 19:58, Valentin Schneider wrote:

[...]

> Aggregate a misfit task's load into sgs->group_misfit_task_load only if
> env->dst_cpu would grant it a capacity uplift.
> 
> Note that the aforementioned capacity vs sgc->max_capacity comparison was
> meant to prevent misfit task downmigration: candidate groups classified as
> group_misfit_task but with a higher (max) CPU capacity than the destination CPU
> would be discarded. This change makes it so said group_misfit_task
> classification can't happen anymore, which may cause some undesired
> downmigrations.
> 
> Further tweak find_busiest_queue() to ensure this doesn't happen. 

Maybe you can describe shortly here what's the new mechanism is you
replace the old 'prevent misfit task downmigration' with.

Also note
> find_busiest_queue() can now iterate over CPUs with a higher capacity than
> the local CPU's, so add a capacity check there.

[...]

> +static inline void update_sg_lb_misfit_stats(struct lb_env *env,
> +					     struct sched_group *group,

Seems to be not used.

> +					     struct sg_lb_stats *sgs,
> +					     int *sg_status,
> +					     int cpu)
> +{
> +	struct rq *rq = cpu_rq(cpu);
> +
> +	if (!(env->sd->flags & SD_ASYM_CPUCAPACITY) ||
> +	    !rq->misfit_task_load)
> +		return;
> +
> +	*sg_status |= SG_OVERLOAD;
> +
> +	/*
> +	 * Don't attempt to maximize load for misfit tasks that can't be
> +	 * granted a CPU capacity uplift.
> +	 */
> +	if (cpu_capacity_greater(env->dst_cpu, cpu)) {
> +		sgs->group_misfit_task_load = max(
> +			sgs->group_misfit_task_load,
> +			rq->misfit_task_load);
> +	}
> +
> +}
> +

[...]

> @@ -9288,6 +9310,8 @@ static struct sched_group *find_busiest_group(struct lb_env *env)
>  	if (!sds.busiest)
>  		goto out_balanced;
>  
> +	env->src_grp_type = busiest->group_type;

Maybe a short comment why you set it here in fbg(). It's only used later
in fbq() for asym. CPU capacity sd but is set unconditionally.

[...]