Re: [PATCH 3/4] workqueue: remove per-node unbound pool when node goes offline.

[Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

(2014/12/15 11:06), Lai Jiangshan wrote:
> On 12/14/2014 12:35 AM, Kamezawa Hiroyuki wrote:
>> remove node aware unbound pools if node goes offline.
>>
>> scan unbound workqueue and remove numa affine pool when
>> a node goes offline.
>>
>> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
>> ---
>>   kernel/workqueue.c | 29 +++++++++++++++++++++++++++++
>>   1 file changed, 29 insertions(+)
>>
>> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
>> index 35f4f00..07b4eb5 100644
>> --- a/kernel/workqueue.c
>> +++ b/kernel/workqueue.c
>> @@ -4921,11 +4921,40 @@ early_initcall(init_workqueues);
>>    * cached pools per cpu should be freed at node unplug
>>    */
>>
>> +/*
>> + * Replace per-node pwq with dfl_pwq because this node disappers.
>> + * The new pool will be set at CPU_ONLINE by wq_update_unbound_numa.
>> + */
>> +static void wq_release_unbound_numa(struct workqueue_struct *wq, int nid)
>> +{
>> +	struct pool_workqueue *old_pwq = NULL;
>> +
>> +	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
>> +		return;
>> +	mutex_lock(&wq->mutex);
>> +	if (wq->unbound_attrs->no_numa)
>> +		goto out_unlock;
>> +	spin_lock_irq(&wq->dfl_pwq->pool->lock);
>> +	get_pwq(wq->dfl_pwq);
>> +	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
>> +	old_pwq = numa_pwq_tbl_install(wq, nid, wq->dfl_pwq);
>> +out_unlock:
>> +	mutex_unlock(&wq->mutex);
>> +	put_pwq_unlocked(old_pwq);
>> +	return;
>> +}
>
>
> We have already did it in wq_update_unbound_numa().
>

Ah, hm. CPU_OFFLINE event does this, ok.

Thanks,
-Kame


>> +
>>   void workqueue_register_numanode(int nid)
>>   {
>>   }
>>
>>   void workqueue_unregister_numanode(int nid)
>>   {
>> +	struct workqueue_struct *wq;
>> +
>> +	mutex_lock(&wq_pool_mutex);
>> +	list_for_each_entry(wq, &workqueues, list)
>> +		wq_release_unbound_numa(wq, nid);
>> +	mutex_unlock(&wq_pool_mutex);
>>   }
>>   #endif
>