Re: [Qemu-devel] Proposal for extensions of block job commands in QEMU 1.2

[Anthony Liguori <anthony@codemonkey.ws>]

On 05/21/2012 09:47 AM, Paolo Bonzini wrote:
> Il 21/05/2012 16:40, Anthony Liguori ha scritto:
>> On 05/21/2012 09:26 AM, Paolo Bonzini wrote:
>>> Il 21/05/2012 16:19, Anthony Liguori ha scritto:
>>>>>
>>>>
>>>> I'm not against it in principle, just in practice.  Today, checking
>>>> whether a command exists is:
>>>>
>>>> commands = qmp.query_commands()
>>>>
>>>> if 'block-stream' in commands:
>>>>       # has block-stream
>>>>
>>>> I have a hard time envisioning how schema introspection can be
>>>> reasonably implemented in a client.
>>>
>>> schema = qmp.query_command_schema('block-stream')
>>
>> What would schema return?
>>
>> Did you mean:
>>
>> if schema['arguments'].has_key('on_error'):
>
> Yes, something like that.
>
>> What about adding a parameter to a structure?
>
> schema = qmp.query_type('foo')
> if schema['data'].has_key('xyz')

This may end up being workable.  I hadn't thought of having granular QMP schema 
querying functions.

>> BTW, the other problem with adding arguments like this is that it makes
>> a stable C API impossible.
>
> Adding fields to structs is not a problem as long as "libqmp" takes care
> of all allocations on part of its client.

Yes, the previous version I wrote handled this padding structures appropriately.

> Adding parameters to commands
> requires some smartness, but there are ways to do it:
>
> 1) add the first version number to the schema, generate versioned entry
> points
>
> qmp_block_stream_v1_1
> qmp_block_stream_v1_2
>
> etc.  Provide multiple headers libqmp-1.1.h, libqmp-1.2.h etc. that take
> care of #define'ing qmp_block_stream to one of them

This is pretty nasty :-/  It also gets very challenging if some options are 
backported and others aren't.

I guess let me ask the following question:

Is the problem that should have an 'options' parameter to this command?  Do we 
anticipate future addition of arguments?  If we can at least minimize the C API 
breakages, I'd be happy.

Regards,

Anthony Liguori

> 2) Same as (1) but use qmp_block_stream for the oldest version having
> the command.
>
> 3) have fun with the preprocessor (macro with variable arguments,
> sizeof, designated initializers, whatever) and emulate keyword arguments
> for the C API.



> Paolo