Re: [Qemu-devel] [PATCH 07/27] add memdev backend infrastructure

[Igor Mammedov <imammedo@redhat.com>]

On Mon, 25 Nov 2013 17:09:37 +0100
Paolo Bonzini <pbonzini@redhat.com> wrote:

> Il 25/11/2013 17:01, Igor Mammedov ha scritto:
> >> > So this is why you need a new command-line option.
> >> > 
> >> > I think we need a generic mechanism for post-initialization of whatever
> >> > is given on the command line.  Perhaps you can do that with an
> >> > interface, and get rid of -memdev and memdev_add altogether?
> >> > 
> >> > MemoryBackend's implementation of the interface's sole method would call
> >> > get_memory, of course.
> > 
> > What I would use instead of memdev_add in CLI/HMP/QMP?
> 
> We could add a new object_add command.
> 
> > Could you explain it a bit more, please?
> 
> The interface would look like
> 
> struct QOMCommandLineIface {
>     void complete(Object *object, Error **errp);
>     Object *get_base_path(void);
> }
> 
> MemoryBackend could implement it like this:
> 
> void memory_backend_complete(Object *object, Error **errp)
> {
>     MemoryBackend *backend = MEMORY_BACKEND(object);
>     MemoryBackendClass *bc = MEMORY_BACKEND_GET_CLASS(obj);
>     if (bc->get_memory) {
>         bc->get_memory(backend, errp);
>     }
> }
> 
> Object *memory_backend_get_base_path(void);
> {
>     return container_get(qdev_get_backend(), "/memdev"),
> }
> 
> A default implementation can be added to RNGBackend and TPMBackend.
> vl.c can use the interface like this in object_create:
> 
>     obj = object_new(type);
>     QOMCommandLineIface *cmdline_iface;
> 
>     if (IS_QOM_COMMAND_LINE(obj)) {
>         object_unref(obj);
>         error...
>         return -1;
>     }
> 
>     if (qemu_opt_foreach(opts, object_set_property, obj, 1) < 0) {
>         object_unref(obj);
>         return -1;
>     }
> 
>     cmdline_iface = QOM_COMMAND_LINE_GET_IFACE(obj);
>     cmdline_iface->complete(obj, &local_err);
>     if (local_err)) {
>         error_propagate(...)
>         object_unref(obj);
>         return -1;
>     }
> 
>     object_property_add_child(cmdline_iface->get_base_path(),
>                               id, obj, NULL);
> 
> Then you can just use -object instead of -memdev.
It looks like "realize" for -object / object-add implemented via
an interface.

Maybe it should be renamed from QOMCommandLineIface to QOMRealizeIface
and s/complete/realize/ so anyone who knows about Device.realize would
get meaning without digging in complete() implementations.

Alternative would be to behave just like Rng/Tpm do, i.e. use -object
to do late initialization in a backend user (DimmDevice.realize).
Draw back of it would be user won't get error during the first command
"object-add" and only will get error when creating DimmDevice calling
"device_add".