Re: [RFC PATCH 3/4] kernel/workqueue: Distinguish between general unbound and WQ_SYSFS cpumask changes

[Waiman Long]

On 1/17/24 12:12, Tejun Heo wrote:
> Hello,
>
> On Wed, Jan 17, 2024 at 02:06:08PM +0100, Juri Lelli wrote:
> > > This looks rather hacky. Can you elaborate how the current code misbehaves
> > > with an example?
> > I was trying to address the fact that ordered unbound workqueues didn't
> > seem to reflect unbound_cpumask changes, e.g.
> >
> > wq_unbound_cpumask=00000003
> >
> > edac-poller              ordered,E  0xffffffff 000000ff      kworker/R-edac-            351 0xffffffff 000000ff
> >
> > vs.
> >
> > edac-poller              ordered,E  00000003                 kworker/R-edac-            349 00000003
> >
> > with the patch applied. But honestly, I'm now also not convinced what
> > I'm proposing is correct, so I'll need to think more about it.
> >
> > Can you please confirm though that ordered unbound workqueues are not
> > "special" for some reason and we would like them to follow
> > unbound_cpumask changes as normal ubound workqueues?
> They aren't special and should follow the normal unbound workqueue cpumask.

My impression is that changing the workqueue cpumask of ordered unbound 
workqueue may break the ordering guarantee momentarily. I was planning 
to look into this further to see if that is true when I have time. If it 
is not a concern, we should certainly apply the global unbound cpumask 
change to those workqueues as well.

Cheers,
Longman