On 2011-12-20 22:38, Anthony Liguori wrote:
> On 12/20/2011 03:23 PM, Jan Kiszka wrote:
>> On 2011-12-20 20:14, Anthony Liguori wrote:
>>> On 12/20/2011 11:02 AM, Jan Kiszka wrote:
>>>> On 2011-12-20 15:07, Anthony Liguori wrote:
>>>>> On 12/20/2011 07:57 AM, Paolo Bonzini wrote:
>>>>>> On 12/20/2011 02:54 PM, Anthony Liguori wrote:
>>>>>>>> In QOM parlance Jan implemented this:
>>>>>>>>
>>>>>>>> abstract class Object
>>>>>>>> abstract class Device
>>>>>>>> class APIC: { backend: link<APICBackend> }
>>>>>>>> abstract class APICBackend
>>>>>>>> class QEMU_APICBackend
>>>>>>>> class KVM_APICBackend
>>>>>>>
>>>>>>> I don't fundamentally object to modeling it like this provided that
>>>>>>> it's
>>>>>>> modeled (and visible) through qdev and not done through a one-off
>>>>>>> infrastructure.
>>>>>>
>>>>>> There is no superclass of DeviceState, hence doing it through qdev
>>>>>> would mean
>>>>>> introducing a new bus type and so on. This would be a superb example
>>>>>> of a
>>>>>> useless bus that can disappear with QOM, but I don't see why we
>>>>>> should
>>>>>> take the
>>>>>> pain to add it in the first place. :)
>>>>>
>>>>> Right, so let's modeled it for now as inheritance which qdev can cope
>>>>> with.
>>>>
>>>> Do we have a clear plan now how to sort out the addressing issues in
>>>> this model? I mean when registering two devices under different names
>>>> that are supposed to be addressable under the same alias once
>>>> instantiated. I didn't follow recent qtree naming changes in details
>>>> unfortunately, if they already enable this.
>>>
>>> I think everyone is in agreement. We'll start with an APICBase type
>>> that's modeled in qdev as a base class.
>>>
>>> There will be an APICBaseInfo that will replace APICBackend.
>>>
>>> There will be two classes that implement APICBaseInfo, KvmAPIC and
>>> APIC. They will be separate devices.
>>>
>>> APICBase will register the vmsd and will use the name "apic" to register
>>> it. You can just set the qdev.vmsd field in the apic_qdev_register()
>>> function to ensure that both use the same implementation.
>>
>> I'm not talking about migration here, I'm talking about qtree
>> addressability. That is orthogonal, at least right now.
>
> qtree is not an ABI. The output of info qtree can (and will) change
> over time.
That's not the point. The point is that at least some branch of the
qtree should be identically named for both the KVM and the user space
incarnations of a particular device (given a certain qemu version).
The request was that /qtree/path/to/apic should not change if you enable
KVM in-kernel acceleration in the very same qemu release. There can also
be some /qtree/path/to/kvm-apic then, but as alias (or as primary name
and the other becomes an alias). I think this makes sense if the user is
still able to clearly differentiate between both versions when listing
devices.
Jan
Attachment:
signature.asc
Description: OpenPGP digital signature
