Re: [PATCH] workqueue: remove the unneeded cpu_relax() in __queue_work()

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

On 05/26/2014 12:23 PM, Tejun Heo wrote:
> On Thu, May 22, 2014 at 10:21:25PM +0800, Lai Jiangshan wrote:
>> On Thu, May 22, 2014 at 9:47 PM, Tejun Heo <tj@kernel.org> wrote:
>> This is not busy wait, the retry and numa_pwq_tbl() guarantee that
>> the retry will get a new pwq (even without cpu_relax()) as the comments says,
> 
> Yes, *eventually*.  It's not guaranteed to succeed on the immediate
> next try.  This is a busy wait.



changing pwq:
	install pwq
	lock(pool->lock)
	put_pwq();
	unlock(pool->lock)

__queue_work():
	lock(pool->lock)
	test ref and find it zero;
	see the installation here;
	it is guaranteed to get the installed pwq on the immediate next try.
	unlock()
	retry.




> 
>> and the refcnt of this new pwq is very very likely non-zero and
>> cpu_relax() can't
>> increase the probability of non-zero-refcnt. cpu_relax() is useless here.
>>
>> It is different from spin_lock() or some other spin code.
>>
>> it is similar to the loop of __task_rq_lock() which also guarantees progress.
> 
> No, it's not.  __task_rq_lock() *already* sees the updated value to
> use for the next time.  Here, we see the old one dead and the new one
> is guaranteed to show up pretty soon but we're still busy waiting for
> it.
>