Re: [PATCH V2] sched/fair: Fix that tasks are not constrained by cfs_b->quota on hotplug core, when hotplug core is offline and then online.

[bsegall@google.com]

Jeehong Kim <jhez.kim@samsung.com> writes:

>>Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>>
>>> You forgot to Cc Ben, who gave you feedback on v1, which is rather poor
>>> style. Also, I don't see how kernel-janitors is relevant to this patch.
>>> This is very much not a janitorial thing.
>>>
>>> (also, why send it twice?)
>>>
>>> On Tue, Aug 30, 2016 at 10:12:40PM +0900, Jeehong Kim wrote:
>>>> In case that CONFIG_HOTPLUG_CPU and CONFIG_CFS_BANDWIDTH is turned on
>>>> and tasks in bandwidth controlled task group run on hotplug core,
>>>> the tasks are not controlled by cfs_b->quota when hotplug core is offline
>>>> and then online. The remaining tasks in task group consume all of
>>>> cfs_b->quota on other cores.
>>>>
>>>> The cause of this problem is described as below:
>>>>
>>>> 1. When hotplug core is offline while tasks in task group run
>>>> on hotplug core, unregister_fair_sched_group() deletes
>>>> leaf_cfs_rq_list of tg->cfs_rq[cpu] from &rq_of(cfs_rq)->leaf_cfs_rq_list.
>>>>
>>>> 2. Then, when hotplug core is online, update_runtime_enabled()
>>>Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>>
>>> You forgot to Cc Ben, who gave you feedback on v1, which is rather poor
>>> style. Also, I don't see how kernel-janitors is relevant to this patch.
>>> This is very much not a janitorial thing.
>>>
>>> (also, why send it twice?)
>>>
>>> On Tue, Aug 30, 2016 at 10:12:40PM +0900, Jeehong Kim wrote:
>>>> In case that CONFIG_HOTPLUG_CPU and CONFIG_CFS_BANDWIDTH is turned on
>>>> and tasks in bandwidth controlled task group run on hotplug core,
>>>> the tasks are not controlled by cfs_b->quota when hotplug core is offline
>>>> and then online. The remaining tasks in task group consume all of
>>>> cfs_b->quota on other cores.
>>>>
>>>> The cause of this problem is described as below:
>>>>
>>>> 1. When hotplug core is offline while tasks in task group run
>>>> on hotplug core, unregister_fair_sched_group() deletes
>>>> leaf_cfs_rq_list of tg->cfs_rq[cpu] from &rq_of(cfs_rq)->leaf_cfs_rq_list.
>>>>
>>>> 2. Then, when hotplug core is online, update_runtime_enabled()
>>>> registers cfs_b->quota on cfs_rq->runtime_enabled of all leaf cfs_rq
>>>> on runqueue. However, because this is before enqueue_entity() adds
>>>> &cfs_rq->leaf_cfs_rq_list on &rq_of(cfs_rq)->leaf_cfs_rq_list,
>>>> cfs->quota is not register on cfs_rq->runtime_enabled.
>>>>
>>>> To resolve this problem, this patch makes update_runtime_enabled()
>>>> registers cfs_b->quota by using walk_tg_tree_from().
>>>
>>>
>>>> +static int __maybe_unused __update_runtime_enabled(struct task_group *tg, void *data)
>>>>  {
>>>> +    struct rq *rq = data;
>>>> +    struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
>>>> +    struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
>>>>
>>>> +    raw_spin_lock(&cfs_b->lock);
>>>> +    raw_spin_unlock(&cfs_b->lock);
>>>>
>>>> +    return 0;
>>>> +}
>>>> +
>>>> +static void __maybe_unused update_runtime_enabled(struct rq *rq)
>>>> +{
>>>> +    struct cfs_rq *cfs_rq = &rq->cfs;
>>>> +
>>>> +    /* register cfs_b->quota on the whole tg tree */
>>>> +    rcu_read_lock();
>>>> +    walk_tg_tree_from(cfs_rq->tg, __update_runtime_enabled, tg_nop, (void *)rq);
>>>> +    rcu_read_unlock();
>>>>  }
>>>
>>> Looks ok, performance on hotplug doesn't really matter. Ben, you happy
>>> with this?
>>
>> I'm not 100% sure about the exact timings and mechanics of hotplug, but
>> cfs-bandwidth wise this is ok. We may still have runtime_remaining = 1,
>> or we may have < 0 and yet be unthrottled, but either case is ok, even
>> if hotplug allows tasks to have migrated here already (I'm not sure,
>> looking at the code).
>>
>> Now that I check again you can just loop over the list of tgs rather
>> than the hierarchical walk_tg_tree_from, but there's certainly no harm
>> in it.
>
> Ben,
>
> Is there additional revision which I have to do?
> If so, could you let me know about that?
>
> Regards,
> Jeehong Kim

Oh, no, this is fine by me.