Re: [RFC PATCH 7/7] workqueue: Replace pool lock with preemption disabling in wq_worker_sleeping()

[Tejun Heo]

Hello,

On Tue, Dec 07, 2021 at 03:35:43PM +0800, Lai Jiangshan wrote:
> From: Lai Jiangshan <laijs@xxxxxxxxxxxxxxxxx>
> 
> Once upon a time,  wq_worker_sleeping() was called with rq lock held,
> so wq_worker_sleeping() can not use pool lock.  Instead it used "X:"
> protection: preemption disabled on local cpu and wq_worker_sleeping()
> didn't depend on rq lock to work even with it held.
> 
> Now, wq_worker_sleeping() isn't called with rq lock held and is using
> pool lock.  But the functionality of "X:" protection isn't removed and
> wq_worker_running() is still using it.
> 
> So we can also use "X:" protection in wq_worker_sleeping() and avoid
> locking the pool lock.
> 
> This patch also documents that only idle_list.next is under "X:"
> protection, not the whole idle_list because destroy_worker() in idle
> timer can remove non-first idle workers.  Idle timer can be possible
> strayed in different CPU, or even in the same CPU, it can interrupt
> preemption disabled context.

It's nice to go back to not needing to grab pool lock in the worker sleeping
path but I'm not sure it actually matters. This isn't in a red-hot path and
we're touching a bunch of stuff in the pool anyway, so the overhead of
grabbing a lock which likely isn't too contended shouldn't matter all that
much. So, maybe it'd be better to just keep things simple?

Thanks.

-- 
tejun