Re: [Qemu-devel] [PATCH RFC 0/3] qom: Generalize qdev init to "realize"

[Anthony Liguori]

On 03/27/2012 11:46 AM, Paolo Bonzini wrote:
> Il 27/03/2012 18:33, Peter Maydell ha scritto:
> > > > Yes, I haven't reviewed this series yet, but my expectation would be that
> > > > realize propagates and that the default implementation of
> > > > DeviceClass::realize() would explicitly *not* propagate and instead just
> > > > call ::init.
> > So who calls realize for non-qdev QOM objects which are children of
> > qdev QOM objects?
> >
> > I really don't like having the object lifecycle methods be different
> > for DeviceClass than for base objects if we can avoid it.
>
> The way I read it was that the "realize" property propagates to the
> children and calls either the "realize" or "reset" method (in the
> appropriate order).

Correct.

We would probably want to have two methods in Object:

/* realize yourself */
void realize(Error **errp);

/* realize your children */
void realize_children(Error **errp);

The actual realize property would call this->realize(errp); then 
this->realize_children(errp) modulo some cleverness to support transversal ordering.

That means usually, you can just implement realize() and rely on the standard 
child realize support.  But you could overload realize_children() if you needed to.

realize_children() would propagate to child<> properties.  All child<> 
properties trace their parentage back to the object_get_root() device.  So you 
can realize() the entire device/object model by setting the top level realize 
property.

> DeviceClass::realize() would also do other things done currently by
> qdev_init, such as register vmstate,

Correct.  I think DeviceClass::realize() should call DeviceClass::init() to 
start out with.  Then we can refactor incrementally to move stuff from ::init() 
to ::realize().  Some parts of :init() will get moved to the constructor.

> so there would still be a
> separation between DeviceClass::realize and DeviceClass::init (we do not
> want yet another pass through the whole tree).
>
> qdev_init would be a thin wrapper around object_realize that takes care
> of freeing the object when init fails.

Absolutely.

Regards,

Anthony Liguori

> Is this correct?
>
> Paolo