Re: [PATCH v3 1/7] sched/fair: Ignore percpu threads for imbalance pulls

[Valentin Schneider <valentin.schneider@arm.com>]

On 16/03/21 16:49, Dietmar Eggemann wrote:
> On 11/03/2021 13:05, Valentin Schneider wrote:
>> From: Lingutla Chandrasekhar <clingutla@codeaurora.org>
>>
>> In load balancing, when balancing group is unable to pull task
>> due to ->cpus_ptr constraints from busy group, then it sets
>> LBF_SOME_PINNED to lb env flags, as a consequence, sgc->imbalance
>> is set for its parent domain level. which makes the group
>> classified as imbalance to get help from another balancing cpu.
>>
>> Consider a 4-CPU big.LITTLE system with CPUs 0-1 as LITTLEs and
>
> Does it have to be a big.LITTLE system? I assume this issue also happens
> on an SMP system.
>

Aye, though the consequences are "worse" on asym CPU capacity systems.

>> CPUs 2-3 as Bigs with below scenario:
>> - CPU0 doing newly_idle balancing
>> - CPU1 running percpu kworker and RT task (small tasks)
>
> What's the role of the small RT task here in the story?
>

I don't think it matters much here.

>> - CPU2 running 2 big tasks
>> - CPU3 running 1 medium task
>>
>> While CPU0 is doing newly_idle load balance at MC level, it fails to
>> pull percpu kworker from CPU1 and sets LBF_SOME_PINNED to lb env flag
>> and set sgc->imbalance at DIE level domain. As LBF_ALL_PINNED not cleared,
>> it tries to redo the balancing by clearing CPU1 in env cpus, but it don't
>> find other busiest_group, so CPU0 stops balacing at MC level without
>> clearing 'sgc->imbalance' and restart the load balacing at DIE level.
>>
>> And CPU0 (balancing cpu) finds LITTLE's group as busiest_group with group
>> type as imbalance, and Bigs that classified the level below imbalance type
>> would be ignored to pick as busiest, and the balancing would be aborted
>> without pulling any tasks (by the time, CPU1 might not have running tasks).
>>
>> It is suboptimal decision to classify the group as imbalance due to
>> percpu threads. So don't use LBF_SOME_PINNED for per cpu threads.
>
> This sentence mentioned per-cpu threads (and so does the patch name) but
> the implementation (only) deals with per-cpu kernel threads. IMHO, it
> would be good to align this.
>

Tell you what, I'll go for:
1) how can pcpu kthreads cause LBF_SOME_PINNED
2) why we may not want this, but still ignore !kthread pcpu tasks
3) why this is even more important for big.LITTLE