Re: [PATCH 2/3] sched/fair: Fixes for capacity inversion detection

[Dietmar Eggemann]

On 27/11/2022 15:17, Qais Yousef wrote:
> Traversing the Perf Domains requires rcu_read_lock() to be held and is
> conditional on sched_energy_enabled(). rcu_read_lock() is held while in
> load_balance(), add an assert to ensure this is always the case.
> 
> Also skip capacity inversion detection for our own pd; which was an
> error.
> 
> Fixes: 44c7b80bffc3 ("sched/fair: Detect capacity inversion")
> Reported-by: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
> Signed-off-by: Qais Yousef (Google) <qyousef@xxxxxxxxxxx>
> ---
>  kernel/sched/fair.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 89dadaafc1ec..7c0dd57e562a 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -8856,16 +8856,22 @@ static void update_cpu_capacity(struct sched_domain *sd, int cpu)
>  	 *   * Thermal pressure will impact all cpus in this perf domain
>  	 *     equally.
>  	 */
> -	if (static_branch_unlikely(&sched_asym_cpucapacity)) {
> +	if (sched_energy_enabled()) {
>  		unsigned long inv_cap = capacity_orig - thermal_load_avg(rq);
>  		struct perf_domain *pd = rcu_dereference(rq->rd->pd);
>  
>  		rq->cpu_capacity_inverted = 0;
>  
> +		SCHED_WARN_ON(!rcu_read_lock_held());

This will trigger in CPU hotplug via build_sched_domains() ->
update_group_capacity() -> update_cpu_capacity() on an EAS system.

> +
>  		for (; pd; pd = pd->next) {
>  			struct cpumask *pd_span = perf_domain_span(pd);
>  			unsigned long pd_cap_orig, pd_cap;
>  
> +			/* We can't be inverted against our own pd */
> +			if (cpumask_test_cpu(cpu_of(rq), pd_span))
> +				continue;
> +

This should fix the issue with `cpu` function parameter in its own PD.

>  			cpu = cpumask_any(pd_span);
>  			pd_cap_orig = arch_scale_cpu_capacity(cpu);
>  

I still don't get the benefit of the CPU capacity inversion patches in
tip/sched/core which should be fixed by this patch:

aa69c36f31aa - sched/fair: Consider capacity inversion in
               util_fits_cpu()
44c7b80bffc3 - sched/fair: Detect capacity inversion

I have to ask again. Why should we use thermal_load_avg() instead of
arch_scale_thermal_pressure() for a CPUx in `CPU capacity inversion
state` (i.e. w/ higher `CPU capacity orig` but lower `CPU capacity` than
CPUy?