On Thu, Oct 25, 2018 at 08:05:40AM -0700, Bart Van Assche wrote:
> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
> index fc9129d5909e..0ef275fe526c 100644
> --- a/kernel/workqueue.c
> +++ b/kernel/workqueue.c
> @@ -1383,6 +1383,10 @@ static void __queue_work(int cpu, struct
> workqueue_struct *wq,
> if (unlikely(wq->flags & __WQ_DRAINING) &&
> WARN_ON_ONCE(!is_chained_work(wq)))
> return;
> +
> + if (!(wq->flags & __WQ_HAS_BEEN_USED))
> + wq->flags |= __WQ_HAS_BEEN_USED;
> +
> retry:
> if (req_cpu == WORK_CPU_UNBOUND)
> cpu = wq_select_unbound_cpu(raw_smp_processor_id());
I was looking to fix this problem as well, and I thought about doing
this, but I thought wq->mutex had to be taken in order to modify
wq->flags --- and that would destroy the scalability of the
add-to-work-to-queue functions.
We could switch all of wq->flags to use the {test,set,clear}_bits()
family of operations, but that seemed like a very large change.
Tejun seemed to be ok with creating a destroy_workqueue_no_drain()
function and using that instead of destroy_workqueue() for the losers
of the cmpxchg lockless initialization code in sb_init_dio_done_wq()
in fs/direct_io.c.
- Ted
