Re: [Qemu-devel] [PATCH] hw/arm/bcm283x: Fix crash with device_add bcm2837 on unsupported machines

[Eduardo Habkost <ehabkost@redhat.com>]

On Mon, Jul 16, 2018 at 08:43:24AM +0200, Markus Armbruster wrote:
> Eduardo Habkost <ehabkost@redhat.com> writes:
> 
> > On Thu, Jul 12, 2018 at 10:05:46AM +0200, Paolo Bonzini wrote:
> >> On 11/07/2018 22:23, Eduardo Habkost wrote:
> >> > On Wed, Jul 11, 2018 at 10:16:42PM +0200, Paolo Bonzini wrote:
> >> >> On 11/07/2018 20:30, Eduardo Habkost wrote:
> >> >>>> The theoretical behavior should be:
> >> >>> It's not clear below where you expect
> >> >>>   qdev_set_parent_bus(..., sysbus_get_default())
> >> >>> to be called (if it should be called at all).
> >> >>>
> >> >>> I don't know where it should be called, but I'm absolutely sure
> >> >>> instance_init is not the right place.
> >> >>
> >> >> I think instance_init is fine to call qdev_set_parent_bus on contained
> >> >> devices.  Why do you say it's not?
> >> > 
> >> > Because object_unref(object_new(...)) is not supposed to affect
> >> > QEMU global state at all.
> >> 
> >> It should not affect it.  Any changes to the global state done by
> >> instance_init are immediately undone when object_unref destroys the
> >> child properties of the object.
> >
> > I would prefer if it didn't, but not a big deal as long as all
> > QOM code is protected by the BQL (it is, right?).
> >
> > If we get rid of object_new() in qmp_device_list_properties(),
> > then most of the restrictions on instance_init can go away,
> > anyway.
> 
> How could we get rid of object_new()?  As long as we create properties
> in code, we need to run the code to find the properties.

By stopping registering properties at instance_init, and making
them introspectable at the class object.

-- 
Eduardo