Re: [PATCH 4/7 V6] workqueue: fix idle worker depletion

[Lai Jiangshan <eag0628@gmail.com>]

On Sun, Sep 9, 2012 at 3:02 AM, Tejun Heo <tj@kernel.org> wrote:
> Hello, Lai.
>
> On Sun, Sep 09, 2012 at 02:34:02AM +0800, Lai Jiangshan wrote:
>> in 3.6 busy_worker_rebind() handle WORKER_REBIND bit,
>> not WORKER_UNBOUND bit.
>>
>> busy_worker_rebind() takes struct work_struct *work argument, we have to
>> add new patch to add a helper and restruct it at first.
>
> What's wrong with just treating manager as busy.  Factor out,
> rebind_work scheduling from rebind_workers() and call it for busy
> workers and the manager if it exists.  manage_workers() only need to
> call process_scheduled_works().  Wouldn't that work?
>
>> worker_maybe_bind_and_lock() 's mean is very clear
>> here. busy_worker_rebind() seems for busy workers, manager is not
>> busy workers.
>
> I don't know.  It just seems unnecessarily wordy.  If you don't like
> reusing the busy worker path, how about just calling
> maybe_bind_and_lock() unconditionally after locking manager_mutex?  I
> mean, can't it just do the following?
>
>         spin_unlock_irq(&gcwq->lock);
>
>         /*
>          * Explain what's going on.
>          */
>         mutex_lock(&pool->manager_mutex);
>         if (worker_maybe_bind_and_lock(worker))
>                 worker_clr_flags(worker, WORKER_UNBOUND);
>         ret = true;
>


This code is correct. worker_maybe_bind_and_lock() can be called any time.
but I like to call it only when it is really needed.

Thanks.
Lai