Re: [Qemu-devel] [PATCH v5] qom, qmp, hmp, qapi: create qom-type-prop-list for class properties

[Eduardo Habkost]

On Tue, Jan 26, 2016 at 10:19:13PM +0000, Daniel P. Berrange wrote:
> On Tue, Jan 26, 2016 at 03:26:35PM -0200, Eduardo Habkost wrote:
> > On Tue, Jan 26, 2016 at 03:51:21PM +0000, Daniel P. Berrange wrote:
> > > On Tue, Jan 26, 2016 at 01:35:38PM -0200, Eduardo Habkost wrote:
> > > > On Mon, Jan 25, 2016 at 11:24:47AM +0300, Valentin Rakush wrote:
> > > > > This patch adds support for qom-type-prop-list command to list object
> > > > > class properties. A later patch will use this functionality to
> > > > > implement x86_64-cpu properties.
> > > > > 
> > > > > Signed-off-by: Valentin Rakush <address@hidden>
> > > > > Cc: Luiz Capitulino <address@hidden>
> > > > > Cc: Eric Blake <address@hidden>
> > > > > Cc: Markus Armbruster <address@hidden>
> > > > > Cc: Andreas Färber <address@hidden>
> > > > > Cc: Daniel P. Berrange <address@hidden>
> > > > > Cc: Eduardo Habkost <address@hidden>
> > > > > ---
> > > > [...]
> > > > > diff --git a/qmp.c b/qmp.c
> > > > > index 53affe2..baf25c0 100644
> > > > > --- a/qmp.c
> > > > > +++ b/qmp.c
> > > > > @@ -460,6 +460,37 @@ ObjectTypeInfoList *qmp_qom_list_types(bool 
> > > > > has_implements,
> > > > >      return ret;
> > > > >  }
> > > > >  
> > > > > +ObjectPropertyInfoList *qmp_qom_type_prop_list(const char *typename, 
> > > > > Error **errp)
> > > > > +{
> > > > > +    ObjectClass *klass;
> > > > > +    ObjectPropertyInfoList *props = NULL;
> > > > > +    ObjectProperty *prop;
> > > > > +    ObjectPropertyIterator iter;
> > > > > +
> > > > > +    klass = object_class_by_name(typename);
> > > > > +    if (!klass) {
> > > > > +        error_set(errp, ERROR_CLASS_DEVICE_NOT_FOUND,
> > > > > +                  "Object class '%s' not found", typename);
> > > > > +        return NULL;
> > > > > +    }
> > > > > +
> > > > > +    object_class_property_iter_init(&iter, klass);
> > > > > +    while ((prop = object_property_iter_next(&iter))) {
> > > > > +        ObjectPropertyInfoList *entry = 
> > > > > g_new0(ObjectPropertyInfoList, 1);
> > > > > +
> > > > > +        if (entry) {
> > > > > +            entry->value = g_new0(ObjectPropertyInfo, 1);
> > > > > +            entry->next = props;
> > > > > +            props = entry;
> > > > > +
> > > > > +            entry->value->name = g_strdup(prop->name);
> > > > > +            entry->value->type = g_strdup(prop->type);
> > > > > +        }
> > > > > +    }
> > > > > +
> > > > > +    return props;
> > > > > +}
> > > > > +
> > > > 
> > > > We already have "-device <type>,help", and it uses a completely
> > > > different mechanism for listing properties. There's no reason for
> > > > having two arbitrarily different APIs for listing properties
> > > > returning different results.
> > > > 
> > > > If qmp_device_list_properties() is not enough for you, please
> > > > clarify why, so we can consider improving it.
> > > 
> > > qmp_device_list_properties() has to actually instantiate an instance
> > > of objects it is reporting properties against, since it is reporting
> > > properties registered against object instances. In fact it only
> > > reports properties against things which are TYPE_DEVICE - it'll refuse
> > > to report other object types. Having to instantiate objects is inherantly
> > > limiting to the command because there are some objects that cannot be
> > > instantiated for this purpose. eg abstract objects and objects marked
> > > "cannot_destroy_with_object_finalize_yet". Finally there is also a
> > > performance and memory overhead in having to instantiate objects which
> > > is best avoided.
> > > 
> > > This new API is reporting properties that are statically registered
> > > against the *class* rather than than object instance. It is guaranteed
> > > that you can always report these properties for any class without any
> > > restrictions, nor any chance of side effects during instantiation.
> > 
> > The existing implementation has its limitations, but we can
> > address those limitations without exporting a new API that return
> > arbitrarily different results (that aren't even a superset of the
> > existing API).
> > 
> > About the existing qmp_device_list_properties() limitations:
> > 
> > cannot_destroy_with_object_finalize_yet is supposed to eventually
> > go away. If there are use cases that depend on listing properties
> > for cannot_destroy_with_object_finalize_yet classes, we can fix
> > that.
> > 
> > The TYPE_DEVICE requirement can be removed, as long as the
> > non-device QOM classes are object_new()-safe like the existing
> > cannot_destroy_with_object_finalize_yet=false device classes
> > (they are supposed to be).
> > 
> > About having to instantiate objects: if optimizing that is so
> > important, we can gradually convert the existing classes to use
> > class-properties. While we convert them, we can even have a
> > doesnt_need_to_instantiate_object_to_query_properties flag to
> > indicate classes that were already converted. No need to export a
> > new API.
> > 
> > Abstract classes are harder, but if they are important we can
> > make them a special case inside the existing implementation
> > instead of having two APIs.
> >
> >                              * * *
> > 
> > So, now we have enumerated the current API limitations. Can we
> > enumerate the real world use cases that are affected by them, so
> > we know which ones we need to address first?
> 
> Being able to list properties of arbitrary non-device objects is
> really the critical thing that's missing right now, with abstract
> types a close second.

About abstract types: I thought we didn't export any class
hierarchy information. Should we do it? I guess we wouldn't want
clients to make assumptions about the class hierarchy.

-- 
Eduardo