Re: [RFC,PATCH] workqueues: turn queue_work() into the "barrier" for work->func()

[Oleg Nesterov <oleg@redhat.com>]

On 11/11, David Howells wrote:
>
> Oleg Nesterov <oleg@redhat.com> wrote:
>
> > I think the caller of queue_work() has all rights to expect that
> > the next invocation of work_func() must see "VAR == 1", but this
> > is not true if the work is already pending.
>
> As you said, queue_work() does test_and_set_bit() which implies smp_mb()
> either side of the function, so you're half way there, and run_workqueue()
> calls spin_unlock_irq() just before calling work_clear_pending()...  So might
> it make sense to move the work_clear_pending() into locked section?  Or would
> that require an smp_mb__before_clear_bit()?

This can't really help, afaics. We still need mb() between clear_bit(_PENDING)
and LOAD(VAR). Because unlock() is the "one way" barrier, LOAD(VAR) can leak
into the critical section, and it can be re-ordered with clear_bit() inside
the critical section.

Oleg.