On 05/04/2016 09:45 AM, Markus Armbruster wrote:
>>>>> +void json_output_visitor_reset(JsonOutputVisitor *v);
>>>>
>>>> Hmm.  Why is "reset" not a Visitor method?
>>>>
>>>> I think this would let us put the things enforced by your "qmp: Tighten
>>>> output visitor rules" in the Visitor contract.
>>>
>>> I thought about that, and now that you've mentioned it, I'll probably
>>> give it a try (that is, make visit_reset() a top-level construct that
>>> ALL visitors must support, rather than just qmp-output and json-output).
>>
>> Yes, please.
> 
> Same question for "cleanup".  Stupid name for a destructor, by the way.

Interface question - all of the FOO_visitor_new() functions return a
subtype pointer Foo*, rather than Visitor*; along with a Visitor
*FOO_get_visitor(FOO*) for up-casting, so that FOO* can be used in the
per-type cleanup function; the FOO* pointers are also useful for two
additional output functions in the two output visitors.  We're proposing
hiding the per-type cleanup function behind a simpler visit_free(Visitor
*v).  So all that's left are the two output functions.  Can we get rid
of those, and make Visitor* the only public interface, rather than
making every caller have to do upcasts?

It looks like outside of the testsuite, all uses of these visitors are
local to a single function; and improving the testsuite is not the end
of the world.  Particularly if only the testsuite is using reset, it may
be easier to just patch the testsuite to use a new visitor in the places
where it currently does a reset.  How ugly would it be to require that
the caller pass in a pointer to the result as part of creating the
visitor, with the promise that the result is populated at the end of a
successful visit, and left NULL if the visit is reset early?  Or maybe a
visit_complete() function that is a no-op on input visitors, but
populates the parameter passed at creation for output visitors.  If we
do that, we could rewrite things like the existing:

QObject *object_property_get_qobject(Object *obj, const char *name,
                                     Error **errp)
{
    QObject *ret = NULL;
    Error *local_err = NULL;
    QmpOutputVisitor *qov;

    qov = qmp_output_visitor_new();
    object_property_get(obj, qmp_output_get_visitor(qov), name, &local_err);
    if (!local_err) {
        ret = qmp_output_get_qobject(qov);
    }
    error_propagate(errp, local_err);
    qmp_output_visitor_cleanup(qov);
    return ret;
}

to instead be:

QObject *object_property_get_qobject(Object *obj, const char *name,
                                     Error **errp)
{
    QObject *ret = NULL;
    Error *local_err = NULL;
    Visitor *v;

    v = qmp_output_visitor_new(&ret);
    object_property_get(obj, v, name, &local_err);
    if (!local_err) {
        visit_complete(v); /* populates ret */
    }
    error_propagate(errp, local_err);
    visit_free(v);
    return ret;
}

Slightly shorter, but populating 'ret' at a distance feels a bit weird.
 Maybe we need to keep the FOO_new() functions returning FOO* rather
than Visitor*, along with the FOO_get_visitor() functions, after all.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org