Re: [Qemu-devel] [PATCH 1/4] Introduce a new hotplug state: Force eject.

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, May 16, 2012 at 11:02:30AM -0500, Anthony Liguori wrote:
> On 05/16/2012 06:23 AM, Paolo Bonzini wrote:
> >Il 16/05/2012 13:15, Anthony PERARD ha scritto:
> >>>>         qdev_unplug(&(d->qdev), NULL);
> >>>>+        qdev_free(&(d->qdev));
> >>>>     }
> >>>>  }
> >>>>
> >>>>
> >>>>Anthony, can you confirm that this solves the problem for you?
> >>This work until I try to hotplug a new device to the guest at wish
> >>point I have this:
> >>ERROR:/local/home/anthony/work/qemu/qom/object.c:389:object_delete:
> >>assertion failed: (obj->ref == 0)
> >>
> >>This is because there is still a pending request of the hotunplug in
> >>the acpi piix4.
> >>If I call qdev_free without qdev_unplug, I hit the same assert, but
> >>rigth away. This is way something new.
> >
> >Because it's missing the object_unparent done by qdev_unplug.  Does
> >object_unparent+qdev_free work?  (I believe object_unparent should be
> >done by qdev_free rather than qdev_unplug, but that's something for 1.2).
> 
> qdev_free() is trivially object_delete today.
> 
> What we should do is make an object_destroy() which emits a destroy
> event and then decrements the reference count.  When ref == 0, we
> should emit a delete event.
> 
> We could then register a slot to object_unparent in the destroy
> event handler, and then object_new() could register a free handler
> in the delete event.
> 
> Then object_delete()/qdev_free() just become trivial invocations of object_unref().
> 
> But for 1.1, we definitely should just do an explicit object_unparent().
> 
> Regards,
> 
> Anthony Liguori
> 
> >
> >Paolo

Okay. Can you fix the bug Amos reported this way too
to avoid confusion?  Or prefer Amos to do it?

-- 
MST