On 09/06/2015 10:42, Markus Armbruster wrote: > Part 2: When to use it? > > We use 'gen': false when we can't (be bothered to) specify the exact > type of an argument or result. > > Bad example: netdev_add > > We have arguments 'type': 'str' and '*props': '**'. > > We should have a union tagged by network backend type. For each > type, the union holds the type's properties (if any).
The problem with this is that netdev_add was not type safe, because it uses qemu_opts_from_qdict and QemuOpts is exclusively string-based. So you could write 'port': '123' or 'port': 123, and both would work, the conversion to integer is done by the QemuOptsVisitor. Note that device_add would have the same problem. Paolo > Better example: device_add (but that's not even QAPIfied, yet) > > If QAPIfied, we'd have arguments like 'driver': 'str' and '*props': > '**'. > > Looks just like netdev_add. The difference is that network backends > and their properties are defined in one place, but device models and > their properties aren't. They get collected at run time. As long > as the schema is fixed at compile-time, it can't express the > resulting tagged union. > > Another good example: qom-get > > We have a return value '**'. > > The run time type is the type of the property identified by the > arguments. Therefore, the compile time type can only be the union > of all property types, which effectively boils down to "anything". > > The only way to say "anything" right now is '**'. Requires 'gen': > false. I figure we could extend the generators to support '**' in a > few places, which may let us avoid 'gen': false here. > > Drawback of '**': introspection knows nothing. > > Introspection knowing nothing about netdev_add's and device_add's > acceptable properties is a big, painful gap. > > Please don't invent new reasons for 'gen': false without a very > compelling use case. If you think you have one, we need to talk to make > sure use of 'gen': false really beats the alternatives. Alternatives > may include extending the generators. > >
