Re: [Qemu-devel] [PATCH 6/6] qdev: Generate IDs for anonymous devices

[Jan Kiszka]

On 2011-09-07 12:34, Gleb Natapov wrote:
> On Wed, Sep 07, 2011 at 12:27:23PM +0200, Jan Kiszka wrote:
>> On 2011-09-07 11:50, Gleb Natapov wrote:
>>> On Wed, Aug 31, 2011 at 08:31:26PM +0200, Jan Kiszka wrote:
>>>> On 2011-08-29 23:19, Anthony Liguori wrote:
>>>>> On 08/29/2011 03:56 PM, Jan Kiszka wrote:
>>>>>> On 2011-08-29 21:23, Anthony Liguori wrote:
>>>>>>> On 08/26/2011 09:48 AM, Jan Kiszka wrote:
>>>>>>>> In order to address devices for that the user forgot or is even unable
>>>>>>>> (no_user) to provide an ID, assign an automatically generated one. Such
>>>>>>>> IDs have the format #<number>, thus are outside the name space availing
>>>>>>>> to users. Don't use them for bus naming to avoid any other user-visible
>>>>>>>> change.
>>>>>>>
>>>>>>> I don't think this is a very nice approach.  Why not eliminate anonymous
>>>>>>> devices entirely and use a parent derived name for devices that are not
>>>>>>> created by the user?
>>>>>>
>>>>>> This eliminates anonymous devices completely. So I guess you are asking
>>>>>> for a different naming scheme, something like<parent-id>.child#<no>
>>>>>> e.g.? Well, we would end up with fairly long names when a complete
>>>>>> hierarchy is anonymous. What would be the benefit?
>>>>>
>>>>> No, I'm saying that whenever a device is created, it should be given a
>>>>> non-random name.  IOW, the names of these devices should be stable.
>>>>>
>>>>>> I'm really just looking for some simple, temporary workaround without
>>>>>> touching the existing fragile naming scheme. What we really need is full
>>>>>> path addressing, but that without preserving all the legacy.
>>>>>
>>>>> Yeah, I understand, and I hesitated making any grander suggestions here,
>>>>> but I'm not sure how much work it would be to just remove any caller
>>>>> that passes NULL for ID and replace it with something more meaningful. I
>>>>> think that's a helpful clean up long term no matter what.
>>>>
>>>> That won't solve the problem of finding a unique device name. If we want
>>>> to derive it from stable device properties (bus addresses etc.), we
>>>> first of all have to define them for all types of devices. And that's
>>>> basically were the discussion exploded last year IIRC.
>>>>
>>> Why not use the OpenFirmware naming that we already have for some
>>> devices instead of inventing something new?
>>
>> Because I do not want to establish any path names before QOM conversion
>> (including potential device reorganization) has been started.
> In theory device paths are dictated by HW topology, not today's flavor of
> QEMU object model.

There will be changes in the object composition, but predicting them
today and modeling this on top of current qdev is nothing I want to try.

> 
>> Specifically as I do not need naming for "some" devices, but for all.
>>
> It can be extended. We already have three types of device naming. One is
> used in qdev, another is used for migration and yet another one for
> passing device names to firmware. We should converge to a single one :)

Yes, but that's beyond what this patch set can achieve or what will
happen in foreseeable time.

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux