Re: [PATCH 6/8] sched/fair: Filter out locally-unsolvable misfit imbalances

From: Valentin Schneider
Date: Wed Feb 03 2021 - 13:44:17 EST


On 03/02/21 15:16, Qais Yousef wrote:
> On 01/28/21 18:31, Valentin Schneider wrote:
>> Consider the following (hypothetical) asymmetric CPU capacity topology,
>> with some amount of capacity pressure (RT | DL | IRQ | thermal):
>>
>> DIE [ ]
>> MC [ ][ ]
>> 0 1 2 3
>>
>> | CPU | capacity_orig | capacity |
>> |-----+---------------+----------|
>> | 0 | 870 | 860 |
>> | 1 | 870 | 600 |
>> | 2 | 1024 | 850 |
>> | 3 | 1024 | 860 |
>>
>> If CPU1 has a misfit task, then CPU0, CPU2 and CPU3 are valid candidates to
>> grant the task an uplift in CPU capacity. Consider CPU0 and CPU3 as
>> sufficiently busy, i.e. don't have enough spare capacity to accommodate
>> CPU1's misfit task. This would then fall on CPU2 to pull the task.
>
> I think this scenario would be hard in practice, but not impossible. Maybe
> gaming could push the system that hard.
>

Actually I wouldn't be surprised if a moderatly busy Android environment
could hit this - slight thermal pressure on the bigs, RT pressure because
we know folks love (ab)using RT, a pinch of IRQs in the mix...

>> @@ -8450,11 +8457,21 @@ static inline void update_sg_lb_stats(struct lb_env *env,
>> continue;
>>
>> /* Check for a misfit task on the cpu */
>> - if (sd_has_asym_cpucapacity(env->sd) &&
>> - sgs->group_misfit_task_load < rq->misfit_task_load) {
>> - sgs->group_misfit_task_load = rq->misfit_task_load;
>> - *sg_status |= SG_OVERLOAD;
>> - }
>> + if (!sd_has_asym_cpucapacity(env->sd) ||
>> + !rq->misfit_task_load)
>> + continue;
>> +
>> + *sg_status |= SG_OVERLOAD;
>> + sgs->group_has_misfit_task = true;
>> +
>> + /*
>> + * 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, i))
>> + sgs->group_misfit_task_load = max(
>> + sgs->group_misfit_task_load,
>> + rq->misfit_task_load);
>
> nit: missing curly braces around the if.
>

Ack.

>> @@ -8504,7 +8521,7 @@ static bool update_sd_pick_busiest(struct lb_env *env,
>> /* Don't try to pull misfit tasks we can't help */
>> if (static_branch_unlikely(&sched_asym_cpucapacity) &&
>> sgs->group_type == group_misfit_task &&
>> - (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
>> + (!sgs->group_misfit_task_load ||
>> sds->local_stat.group_type != group_has_spare))
>> return false;
>>
>> @@ -9464,15 +9481,18 @@ static struct rq *find_busiest_queue(struct lb_env *env,
>> case migrate_misfit:
>> /*
>> * For ASYM_CPUCAPACITY domains with misfit tasks we
>> - * simply seek the "biggest" misfit task.
>> + * simply seek the "biggest" misfit task we can
>> + * accommodate.
>> */
>> + if (!cpu_capacity_greater(env->dst_cpu, i))
>> + continue;
>
> Both this hunk and the one above mean we will end up searching harder to pull
> the task into the right cpu taking actual capacity into account. Which is
> a good improvement.
>

Note that those extra checks are to make sure we *don't* downmigrate tasks
(as stated somewhere above, this change lets find_busiest_queue() iterate
over CPUs bigger than the local CPU's, which wasn't the case before). A "big"
CPU will still get the chance to pull a "medium" task, even if a "medium"
CPU would have been a "better" choice.