Re: [RFC PATCH] ata: libata-scsi: Move long delayed work on system_dfl_long_wq

[Niklas Cassel]

On Mon, May 11, 2026 at 02:54:26PM +0200, Marco Crivellari wrote:
> On Mon, May 11, 2026 at 2:48 PM Niklas Cassel <cassel@xxxxxxxxxx> wrote:
> > [...]
> > Looks good to me.
> >
> > Any particular reason that you sent this as an RFC?
> >
> > I can see similar patches queued up in linux-next already.
> 
> I just wanted to be sure I didn't miss any other reason for being
> per-cpu, and in case
> receive comments on it.

Hmm... I can see that:
drivers/ata/libata-eh.c:ata_scsi_port_error_handler()
does:

schedule_delayed_work(&ap->hotplug_task, 0);

schedule_delayed_work() does:
queue_delayed_work(system_percpu_wq, dwork, delay);

So this will schedule the work on a per-cpu workqueue.


It seems that we are already queueing the same work (&ap->hotplug_task)
on different workqueues, so I guess that is fine.

Right now, both workqueues are per-cpu. Is it fine to change one of them
to be not be bound to a specific CPU?

>From looking at the work, ata_scsi_hotplug(), I can't think of a reason
why this would have to run on the same CPU as the CPU that queued the
work.



>From looking at workqueue.h:

 * system_dfl_wq is unbound workqueue.  Workers are not bound to
 * any specific CPU, not concurrency managed, and all queued works are
 * executed immediately as long as max_active limit is not reached and
 * resources are available.

 [...]

 * system_dfl_long_wq is similar to system_dfl_wq but it may host long running
 * works.

"not concurrency managed"

That sounds like a big change, since the per-cpu workqueues do seem to be
concurrency managed (unlike the _dfl_ ones).

However, considering that the work (&ap->hotplug_task / ata_scsi_hotplug())
does:
mutex_lock(&ap->scsi_scan_mutex);

I also don't see a problem with the workqueue not being concurrency managed,
since the work is taking a mutex anyway.



If anyone sees a problem, please say something, otherwise intend to queue
this up in a few days.


Kind regards,
Niklas