Re: [Qemu-devel] [PATCH v9 05/10] move out net queue structs define

[Yang Hongyang <yanghy@cn.fujitsu.com>]

On 09/07/2015 07:43 PM, Daniel P. Berrange wrote:
> On Mon, Sep 07, 2015 at 07:41:26PM +0800, Yang Hongyang wrote:
>>
>>
>> On 09/07/2015 07:00 PM, Daniel P. Berrange wrote:
>>> On Mon, Sep 07, 2015 at 06:53:49PM +0800, Yang Hongyang wrote:
>>>> On 09/07/2015 05:11 PM, Stefan Hajnoczi wrote:
>>>> [...]
>>>>>> Thanks very much for the suggestion, I've already implemented it and tested,
>>>>>> the code looks cleaner now.
>>>>>>
>>>>>> The last issue is the QOM thing, do Markus and Andreas have more input
>>>>>> about that?
>>>>>
>>>>> If you would like to see examples of QOM usage, take a look at
>>>>> iothread.c and/or backends/hostmem.c.
>>>>>
>>>>> The key things are:
>>>>>
>>>>> 1. They use include/qom/object.h to register a type based on
>>>>>     TYPE_OBJECT and their properties are registered using
>>>>>     object_property_add_*().
>>>>>
>>>>> 2. They implement the TYPE_USER_CREATABLE interface so the -object
>>>>>     command-line option can be used to instantiate them.  See
>>>>>     object_interfaces.h.
>>>>>
>>>>> As a result, a lot of code becomes unnecessary and iothread.c, in
>>>>> particular, is quite short.
>>>>
>>>> After looking into this, I have some questions on the implement, could you
>>>> please help me on this because I don't know much about the object mechanism:
>>>>
>>>> The netfilter need to be initialized after the net_init_clients, because we
>>>> need to attach the filter to the net client. But currently, net client is not
>>>> using QOM, and seems that it is initialized after objects been created. So here
>>>> comes the problem: how can I initialize a certain object later, is it possiable?
>>>
>>> It is currently a bit hacky - most objects are initialized very early,
>>> but we have a similar problem with rng-egd which must be created /after/
>>> chardevs. To deal with this in vl.c main() we have two helper methods
>>> object_create_initial and object_create_delayed. The delayed method though
>>> still happens before the net clients are created. We could probably just
>>> move the objec_create_delayed method invokation to later on, after net
>>> clients are created, but if that doesn't work just add an extra helper
>>> object_create_very_delayed :-)
>>
>> If it's ok to move creation of "rng-egd" later, then "move the
>> objec_create_delayed method invokation after net clients are created" should
>> solve the problem.
>
> I think it should be ok, because IIRC, the only requirement from rng-egd
> was that it be created /after/ -device is processed, so moving it even
> later shouldn't hurt it

Got it, thanks!

>
> Regards,
> Daniel
>

-- 
Thanks,
Yang.