Re: sysbus_create_simple Vs qdev_create

[Markus Armbruster <armbru@redhat.com>]

Paolo Bonzini <pbonzini@redhat.com> writes:

> On 30/07/20 12:03, Markus Armbruster wrote:
>> qdev C layer:
>> 
>>     frob->prop = 42;
>> 
>> Least cognitive load.
>> 
>> QOM has no C layer.
>
> Not really, a QOM object is totally free to do frob->prop = 42.  And
> just like we didn't do that outside device implementation in qdev as our
> tithe to the Church of Information Hiding; the same applies to QOM.

I screwed up the part of my argument that actually has a hope to be
valid, let me try again.

With qdev, you can always do frob->prop = 42, because properties are
always backed by a struct member.

With QOM, properties are built around visitor-based getters and setters.
This means you can always do (but never actually would do) something
like

    fortytwo = qnum_from_int(42);
    v = qobject_input_visitor_new(fortytwo);
    set_prop(OBJECT(frob), v, "prop", cookie, &err);
    visit_free(v);
    qobject_unref(fortytwo);

where set_prop() is the setter you passed to object_property_add(), and
@cookie is the opaque value you passed along with it.

*Maybe* set_prop() wraps around a simpler setter you can call directly,
or even a struct member you can set directy.  QOM does not care.

And that's my point: QOM does not care for the C layer.

>> qdev property layer works even when @frob has incomplete type:
>> 
>>     qdev_prop_set_int32(DEVICE(frob), "prop", 42);
>> 
>> This used to map property name to struct offset & copy the value.
>> Simple, stupid.
>> 
>> Nowadays, it is the same as
>> 
>>     object_property_set_int(OBJECT(frob), "frob", 42, &error_abort);
>> 
>> which first converts the int to a QObject, then uses a QObject input
>> visitor with a virtual walk to convert it back to int and store it in
>> @frob.  It's quite a sight in the debugger.
>
> Yes, but thatt's just because we never bothered to create single-type
> visitors.  For a good reason though: I don't think the extra QAPI code
> is worth (not even that much) nicer backtraces when we already have a
> QObject as a battle-tested variant type.
>
>> qdev "text" layer is really a QemuOpts layer (because that's what we had
>> back then).  If we have prop=42 in a QemuOpts, it calls
>> 
>>     set_property("prop", "42", frob, &err);
>> 
>> Nowadays, this is a thin wrapper around object_property_parse(),
>> basically
>> 
>>     object_property_parse(frob, "prop", 42, &err);
>> 
>> Fine print: except set_property() does nothing for @prop "driver" and
>> "bus", which look just like properties in -device / device-add, but
>> aren't.
>
> Ugly indeed.  They should be special cased up in the caller, probably,
> or use the long-discussed "remainder" feature of the QAPI schema.

qdev_device_add() is still stuck in the QemuOpts age.

>> object_property_parse() uses the string input visitor, which I loathe.
>
> Apart from the list syntax, the string input visitor is decent I think.

It's a death trap:

/*
 * The string input visitor does not implement support for visiting
 * QAPI structs, alternates, null, or arbitrary QTypes. Only flat lists
 * of integers (except type "size") are supported.
 */

"Does not implement support for visiting" is polite language for
"crashes when you dare to visit".

>>>> I've long had the nagging feeling that if we had special-cased
>>>> containers, children and links, we could have made a QOM that was easier
>>>> to reason about, and much easier to integrate with a QAPI schema.
>>>
>>> That's at least plausible.  But I have a nagging feeling that it would
>>> only cover 99% of what we're doing with QOM. :)
>> 
>> The question is whether that 1% really should be done the way it is done
>> :)
>
> And that's a very fair question, but it implies non-trivial design work,
> so the smiley changes to a frown. :(

True!