Re: [PATCH 2/3] workqueues: implement flush_work()

[Oleg Nesterov <oleg@tv-sign.ru>]

On 07/01, Jarek Poplawski wrote:
>
> On Tue, Jul 01, 2008 at 04:50:18PM +0400, Oleg Nesterov wrote:
> ...
> > Yes, cwq can be "stale", but this doesn't matter and we can't have
> > the false positive here.
> > 
> > cwq->current_work is always changed under cwq->lock, and we hold this
> > lock. If we see "cwq->current_work == work" we can safely insert the
> > barrier and wait. Even if this work was already re-queued on another
> > CPU or another workqueue_struct.
> > 
> > Note also that rmb() can't really help here.
> 
> Right! The question is how "stale" this cwq could be when read without
> any lock or barrier. Of course, there can't be the false positive, but
> I wonder if we really do enough, to check if a work isn't current on
> some other cwq, even without any immediate re-queuing.

Not sure I understand...

Of course, the work can be current on _all_ CPUs. So no, we don't do
enough. Please look at the changelog, in particular the note about
flush_work_sync().

But without re-queuing cwq can't be wrong? Once again, flush_work()
flushes the result of the last visible queue_work(). If not requeued,
the work is either current, or it is pending and list_empty() == F.

Oleg.