Hi,

On Wed, Aug 31, 2022 at 02:01:54PM +0200, Markus Armbruster wrote:
> Victor Toso <victortoso@redhat.com> writes:
>
> > The goal of this generator is to validate QAPI examples and transform
> > them into a format that can be used for 3rd party applications to
> > validate their QAPI/QMP introspection.
> >
> > For each Example section, we parse server and client messages into a
> > python dictionary. This step alone has found several ill formatted
> > JSON messages in the examples.
> >
> > The generator outputs another JSON file with all the examples in the
> > QAPI module that they came from. This can be used to validate the
> > introspection between QAPI/QMP to language bindings.
> >
> > When used with the POC qapi-go branch, we have found bad QMP messages
> > with wrong member names, mandatory members that were missing and
> > optional members that were being set with null (not needed).
> >
> > A simple example of the output format is:
> >
> >  { "examples": [
> >    {
> >      "id": "ksuxwzfayw",
> >      "client": [
> >      {
> >        "sequence-order": 1
> >        "message-type": "command",
> >        "message":
> >        { "arguments":
> >          { "device": "scratch", "size": 1073741824 },
> >          "execute": "block_resize"
> >        },
> >     } ],
> >     "server": [
> >     {
> >       "sequence-order": 2
> >       "message-type": "return",
> >       "message": { "return": {} },
> >     } ]
> >     }
> >   ] }
> >
> > If this idea seems reasonable, we can add python-qemu-qmp to validate
> > each message at generation time already.
> >
> > Signed-off-by: Victor Toso <victortoso@redhat.com>
>
> If I understand you correctly, there are two benefits:
>
> 1. Mechanical syntax check for examples
>
>    Love it.

Not just JSON syntax but can be extend to the introspection
layer. Errors like wrong member names would fail while parsing
the examples (issues such as fixed by patches 11 and 13/16 should
not happen anymore).

> 2. Can extract examples for use as test cases
>
>    Sounds good to me.  Possible redundancy with existing tests.
>    Probably nothing to worry about.
>
>    Can you explain in a bit more detail how the extracted data
>    is (to be) used?

Sure.

The Golang test that consumes this is 152 lines of code [0]. The
idea is that we can use the examples to feed Golang unmarshalling
code and then marshall it back to JSON and compare input JSON
with output JSON and see that their content matches.

[0] https://gitlab.com/victortoso/qapi-go/-/blob/wip-v3/test/examples_test.go

I have generated the examples with this patch series and stored
the output here [1]

[1] https://gitlab.com/victortoso/qapi-go/-/tree/wip-v3/test/data/examples

The examples are QMP messages that are either sent by Client "->"
or sent by Server "<-". The order matters so I take the order set
in the examples and store it as "sequence-order".

In the Go test code, I follow the sequence-order. One example of
this being useful is that we know which Return type to expect
after a Command is issued.

I've also included metadata about the type of message, which is
one of three options: command, event or return. (Errors are
return too).

This is important because it makes the tests very easy to write.
Different Unmarshal/Marshal code can be set in the code block of
the specific message type.

--

The things that makes me quite excited with this idea are:

 1. We have valid functional examples documented. If the examples
    break, we would have the software in place to know it (plug
    to ci or some other ninja check seems reasonable to me)

 2. Developers should get more interested in documenting examples
    as that alone is is a valid test case, even if only useful
    for language binding's syntax.

Cheers,
Victor