Re: [PATCH RFC] workqueue: don't grab PENDING bit on some conditions

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

On 07/15/2014 11:58 PM, Tejun Heo wrote:
> Hello, Lai.
> 
> On Tue, Jul 15, 2014 at 05:30:10PM +0800, Lai Jiangshan wrote:
>> Thread1 expects that, after flush_delayed_work() returns, the known pending
>> work is guaranteed finished. But if Thread2 is scheduled a little later than
>> Thread1, the known pending work is dequeued and re-queued, it is considered
>> as two different works in the workqueue subsystem and the guarantee expected
> 
> They are two separate queueing instances of the same work item.

I think the mod_delayed_work() is expected to modify a queueing instances
instead of separate from the name.

> 
>> by Thread1 is broken.
> 
> The guarantee expected by thread 1 is that the most recent queueing
> instance of the work item is finished either through completing
> execution or being cancelled.  No guarantee is broken.

I don't think the mod_delayed_work() is considered as a cancelling operation
to the user. You can add comments to state that it contains a cancelling operation
and a requeue operation.

> 
>> The guarantee expected by Thread1/workqueue-user is reasonable for me,
>> the workqueue subsystem should provide this guarantee. In another aspect,
> 
> You're adding a new component to the existing set of guarantees.  You
> can argue for it but it's a new guarantee regardless.

So, it is an RFC.

> 
>> the flush_delayed_work() is still working when mod_delayed_work_on() returns,
>> it is more acceptable that the flush_delayed_work() beats the
>> mod_delayed_work_on().
>>
>> It is achieved by introducing a KEEP_FLUSHED flag for try_to_grab_pending().
>> If the work is being flushed and KEEP_FLUSHED flags is set,
>> we disallow try_to_grab_pending() to grab the pending of the work.
>>
>> And there is another condition that the user want to speed up a delayed work.
>>
>> When the user use "mod_delayed_work_on(..., 0 /* zero delay */);", his
>> attention is to accelerate the work and queue the work immediately.
>>
>> But the work does be slowed down when it is already queued on the worklist
>> due to the work is dequeued and re-queued. So we also disallow
>> try_to_grab_pending() to grab the pending of the work in this condition
>> by introducing KEEP_QUEUED flag.
> 
> Both are extremely marginal.  
> Do we have any actual cases any of these matters?

No such case.

I only found the WB subsystem (backing-dev.c, fs-writeback.c) uses both
mod_delayed_work() and flush_delayed_work(), but it seems that when
flush_delayed_work() is called, mod_delayed_work() will can't be called.

> I can't see what we're gaining with the extra complexity.

Will you add some comments or let it as before?

> 
>> @@ -1212,6 +1220,13 @@ static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
>>  	 */
>>  	pwq = get_work_pwq(work);
>>  	if (pwq && pwq->pool == pool) {
>> +		if ((keep_flags | KEEP_QUEUED) ||
>> +		    ((keep_flags | KEEP_FLUSHED) &&
> 
> This can't be right.
> 
> Thanks.
>