Re: [PATCH 4/5] workqueue: don't expose workqueue_attrs to users

[Lai Jiangshan <laijs@cn.fujitsu.com>]

On 05/12/2015 09:22 PM, Tejun Heo wrote:
> Hello, Lai.
> 
> On Tue, May 12, 2015 at 10:15:28AM +0800, Lai Jiangshan wrote:
>>> I'm not sure about this.  Yeah, sure, it's a bit more lines of code
>>> but at the same time this'd allow us to make the public interface
>>> atomic too.  What we prolly should do is changing the interface so
>>> that we do
>>>
>>> 	attrs = prepare_workqueue_attrs(gfp_mask);	/* allocate, lock & copy */
>>> 	/* modify attrs as desired */
>>> 	commit_workqueue_attrs(attrs);			/* apply, unlock and free */
>>
>> I think the workqueue.c has too much complicated and rarely used APIs
>> and exposes too much in this way.  No one can set the nice value
>> and the cpuallowed of a task atomically.
> 
> What do you mean no one can?

normal/general task. not kworker.

no one can set the nice value and the cpumallowed of a normal task atomically.

The kernel doesn't have such APIs:

lock_and_get_task_cpus_allowed(task);
/* modify cpumask */
set_cpus_allowed_ptr_and_unlock();


> 
>> If the user want atomic-able, Her/he can just disable WQ_SYSFS
>> on its workqueue and maintain a copy of the cpumask, nice, numa values
>> under its own lock.
> 
> So, we're now requiring workqueue users to take care of
> synchronization, disabling and reinstating WQ_SYSFS (what if userland
> hits those knobs at the same time?) 

I think there is no userland knobs when !WQ_SYSFS.

> and poking into workqueue struct to determine the current values of the

I think the copy version of cpumask, nice, numa values are same as
the workqueue struct have. No poking is required.
(Its own lock-protect-region is the ONLY entry to call apply_workqueue_attrs()).

> attributes that the user is not
> intereted in changing?  This is a horrible interface.