Re: [Qemu-devel] Re: [RFC] qapi: events in QMP

[Kevin Wolf <kwolf@redhat.com>]

Am 16.02.2011 01:59, schrieb Anthony Liguori:
> On 02/15/2011 07:38 AM, Luiz Capitulino wrote:
>> On Tue, 15 Feb 2011 10:20:01 +0100
>> Kevin Wolf<kwolf@redhat.com>  wrote:
>>
>>    
>>> Am 14.02.2011 20:34, schrieb Anthony Liguori:
>>>      
>>>> On 02/14/2011 12:34 PM, Luiz Capitulino wrote:
>>>>        
>>>>> On Mon, 14 Feb 2011 08:39:11 -0600
>>>>> Anthony Liguori<anthony@codemonkey.ws>   wrote:
>>>>>
>>>>>
>>>>>          
>>>>>> On 02/14/2011 06:45 AM, Luiz Capitulino wrote:
>>>>>>
>>>>>>            
>>>>>>> So the question is: how does the schema based design support extending
>>>>>>> commands or events? Does it require adding new commands/events?
>>>>>>>
>>>>>>>
>>>>>>>              
>>>>>> Well, let me ask you, how do we do that today?
>>>>>>
>>>>>> Let's say that I want to add a new parameter to the `change' function so
>>>>>> that I can include a salt parameter as part of the password.
>>>>>>
>>>>>> The way we'd do this today is by checking for the 'salt' parameter in
>>>>>> qdict, and if it's not present, use a random salt or something like that.
>>>>>>
>>>>>>            
>>>>> You likely want to do what you did before. Of course that you have to
>>>>> consider if what you're doing is extending an existing command or badly
>>>>> overloading it (like change is today), in this case you'll want to add
>>>>> a new command instead.
>>>>>
>>>>> But yes, the use-case here is extending an existing command.
>>>>>
>>>>>
>>>>>          
>>>>>> However, if I'm a QMP client, how can I tell whether you're going to
>>>>>> ignore my salt parameter or actually use it?  Nothing in QMP tells me
>>>>>> this today.  If I set the salt parameter in the `change' command, I'll
>>>>>> just get a success message.
>>>>>>
>>>>>>            
>>>>> I'm sorry?
>>>>>
>>>>> { "execute": "change", "arguments": { "device": "vnc", "target": "password", "arg": "1234", "salt": "r1" } }
>>>>> {"error": {"class": "InvalidParameter", "desc": "Invalid parameter 'salt'", "data": {"name": "salt"}}}
>>>>>
>>>>>          
>>>> So I'm supposed to execute the command, and if execution fails, drop the
>>>> new parameter?  If we add a few optional parameters, does that mean I
>>>> have to try every possible combination of parameters?
>>>>        
>>> How is that different from trying out multiple commands? In the end, you
>>> always need some meta information like a schema in order to avoid trying
>>> out which parameters the server supports.
>>>
>>> Anyway, I think there's a second interesting point: Adding parameters
>>> does cause these problems, but it's different for data sent from qemu to
>>> the client (return values and events). If we add more information there,
>>> an older client can just ignore it, without even looking at a schema.
>>>
>>> So I think we should consider this for return values and definitely do
>>> it for events. Sending out five different messages for a single event
>>> that are completely redundant and only differ in the number of fields is
>>> just insane (okay, they wouldn't actually get on the wire because a
>>> client registers only for one of them, but the code for generating them
>>> must exist).
>>>      
>> That's my point when I asked about events in the other thread.
>>    
> 
> Okay, I had confused myself about this.  It's not quite as bad as I had 
> been saying.
> 
> One of the reasons to have generated allocation function is so that we 
> can make sure to always pad structures.  Since all optional fields has a 
> bool to indicate the fields presence, by setting the allocated structure 
> to zero, we can support forwards compatibility for structures.

I think in most cases we would even get away with a default value
instead of the bool. For example for strings, NULL would be a very clear
indication that the field wasn't there in the JSON message. I think it
works also for most (if not all) integer fields, but maybe having a bool
has_* field is safer there.

Or maybe we can add an indication to the schema that specifies if a
default value should be generated or (for the rare cases where it's
needed) a has_* field is added.

> So we can add new members to structures with no libqmp compatibility 
> problem.  It's only if we add new arguments to signals or commands that 
> we will break the ABI.

Yeah, that makes sense.

> If we expect to add fields later, we just have to make sure we use a 
> structure to encapsulate things.

As stated before, I think we should use structures for all events. I
still don't understand why we should have an exception for events. Any
other command returns structures, too, and you don't automagically pull
their fields up one level anywhere except for events.

Kevin