Re: [PATCH workqueue/for-4.14-fixes] workqueue: replace pool->manager_arb mutex with a flag

[Lai Jiangshan]

On Mon, Oct 9, 2017 at 11:08 PM, Tejun Heo <tj@xxxxxxxxxx> wrote:
> Hello,
>
> On Mon, Oct 09, 2017 at 11:02:34PM +0800, Lai Jiangshan wrote:
>> I was also thinking alternative code when reviewing.
>> The first is quite obvious. Testing POOL_MANAGER_ACTIVE
>> can be replaced by testing pool->manager.
>> And POOL_MANAGER_ACTIVE  is not needed. Isn't it?
>
> put_unbound_pool() doesn't have to be called from a kworker context
> and we don't really have a kworker pointer to set pool->manager to.
> We can use a bogus value and then update pool->manager dereferences
> accordingly but I think it's cleaner to simply use a separate flag.
>
>> The second thing is to make manage_workers()
>> and put_unbound_pool() exclusive.
>> Waiting event on POOL_MANAGER_ACTIVE(or pool->manager)
>> is one way. However, the pool's refcnt is not possible to
>> be dropped to zero now since the caller still hold the pool->lock
>
> wait_event_lock_irq() drops the lock if the condition is not met
> before going to sleep (otherwise it wouldn't be able to sleep).

I think just using get_pwq()/put_pwq() in manage_workers()
as the following said is simpler than using wait_event_lock_irq()

thanks
Lai

>
>> and some pwds of the works in the worklist. So the other way
>> to enforce the exclusive could be just doing
>> get_pwq(the first pwd of the worklist) and put_pwq() when
>> the manage_workers() done. And the code about
>> pool->manager_arb in put_unbound_pool() can be
>> simply removed.
>
> Yeah, that part is removed.
>
> Thanks!
>
> --
> tejun