Re: [Qemu-devel] [PATCH 15/18] qom: optimize qdev_get_canonical_path using a parent link

[Anthony Liguori]

On 12/01/2011 05:21 AM, Stefan Hajnoczi wrote:
> On Wed, Nov 30, 2011 at 03:03:45PM -0600, Anthony Liguori wrote:
> > @@ -1210,6 +1210,9 @@ void qdev_property_add_child(DeviceState *dev, const char 
> > *name,
> >       qdev_property_add(dev, name, type, qdev_get_child_property,
> >                         NULL, NULL, child, errp);
> >
> > +    g_assert(child->parent == NULL);
> > +    child->parent = dev;
>
> The implications are:
>
> 1. A DeviceState must be a child or the root.  It is not okay to create
>     a DeviceState and inquire its canonical path before making it a child in
>     the graph.

Correct.

> 2. A DeviceState can only be the child of one parent.  Since
>     user-created devices are added to /peripheral or /peripheral-anon this
>     means that the /i440fx only has links to them, never a parent-child
>     relationship.

Correct.

/peripheral[-anon] are where user created devices live.  Machine created devices 
live directly off of /

Each "directory" ends up being a unique namespace.  This ends up being a nice 
way to support compatibly since we have so many namespace today.

I imagine, for instance, that we will end up with a /block directory too and 
that is where blockdev objects will live.

> Is this right?
>
> > +    /* Do not, under any circumstance, use this parent link below anywhere
> > +     * outside of qdev.c.  You have been warned. */
> > +    DeviceState *parent;
>
> It would be nice to explain why parent is private to qdev.c.

Composition is a unidirectional relationship.  The parent pointer is there 
simply as an optimization.  I'll make the comment a big scarier :-)

Regards,

Anthony Liguori