Re: [PATCH 1/5] workqueue: Allow changing attributions of ordered workqueues

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, May 19, 2014 at 04:15:31PM -0400, Tejun Heo wrote:
> Hello,
> 
> On Sat, May 17, 2014 at 03:41:55PM +0200, Frederic Weisbecker wrote:
> > > > -	last_pool = get_work_pool(work);
> > > > +	last_pool = wq->flags & __WQ_ORDERED ? NULL : get_work_pool(work);
> > > >  	if (last_pool && last_pool != pwq->pool) {
> > > >  		struct worker *worker;
> > > 
> > > I'm not a big fan of the fact that ordered queues need to be handled
> > > differently when queueing, but as the code is currently written, this
> > > is pretty much necessary to maintain execution order, right?
> > >
> > > Otherwise, you end up with requeueing work items targeting the pwq it
> > > was executing on and new ones targeting the newest one screwing up the
> > > ordering.  I think that'd be a lot more important to note in the
> > > comment.  This is a correctness measure.  Back-to-back requeueing
> > > being affected by this is just a side-effect.
> > 
> > In the case of ordered workqueues it actually doesn't matter much in
> > term of ordering. But it's needed when pwqs are replaced (as it happens
> > in apply_workqueue_attrs(). We must make sure works requeueing themselves
> > don't always requeue to the old pwq otherwise it will never be able to
> > switch and be released. Also the next work items will be queued on the next
> 
> But that's the same for other pwqs too.  Back-to-back requeueing will
> hold back pwq switching on any workqueue.

I don't think so, because non ordered pwqs aren't created with 0 max_active,
so they can run before the old pwq is released. It's not holding back the new
one and creating a starvation there.

But maybe I forget other details.

> 
> > pwq but this one will never be able to run due to the old workqueue still
> > being used by the item requeing itself. So we also risk starvation on the
> > new workqueue.
> > 
> > But the ordering itself is actually fine for ordered workqueue. It's actually
> > enforced by the fact that only one pwq can run at a time for a given workqueue.
> 
> Maybe I'm confused but I don't think it'd be.  Let's say there was an
> attribute change with one work item, A, which is performing
> back-to-back requeueing and another one, B, which queues itself
> intermittently.  If B is queued while A is executing, followed by A
> requeueing itself, the expected execution order is A - B - A; however,
> without the above exception for ordered workqueues, it'd end up A - A
> - B because B will end up on the new pwq while A on the older one and
> max_active won't be transferred to the new pwq before it becomes
> empty.

Ah right AAB instead of ABA is possible indeed. I don't know if some workqueue
rely on such guarantee but it's possible.

In which case we have one more reason to make an exception on ordered workqueues
previous pwq reuse.

> 
> > > Just collapse it into the calling function.  This obfuscates more than
> > > helps.
> > 
> > Yeah but the condition is already big. Lets hope the result won't be too ugly.
> 
> I didn't mean that the condition should be encoded in the if
> conditional.  It's fine to break it out using a separate variable or
> whatever.  I just don't think breaking it out to a separate function
> is helping anything.

Alright.

Thanks.