Re: [Qemu-devel] [RFC][PATCH 14/16] kvm: x86: Add user space part for in-kernel i8259

[Jan Kiszka]

On 2011-12-05 14:36, Avi Kivity wrote:
> On 12/05/2011 03:29 PM, Jan Kiszka wrote:
>> On 2011-12-05 14:14, Avi Kivity wrote:
>>> On 12/05/2011 02:47 PM, Jan Kiszka wrote:
>>>>>
>>>>> (the memory API added unstable names, hopefully the QOM can take over
>>>>> the stable ones and we'll have a good way to denote the unstable ones).
>>>>>
>>>>
>>>> OK, maybe - or likely - we should make those device models have the same
>>>> names in QOM once instantiated. But I'm still convinced they should
>>>> remain separated models in contrast to a single model with a property.
>>>
>>> What do you mean by separate models?  You share all the code you can,
>>> and don't share the code you can't.  To me, single model == single name.
>>
>> But different configuration.
> 
> Right, just like IDE with different backends.

Except that there is a comparably large infrastructure to manage those
backends.

> 
>>>
>>>> The kvm ioapic, e.g., requires an additional property (gsi_base) that is
>>>> meaningless for user space devices. And its interrupts have to be
>>>> wired&configured differently at board model level. So, from the QEMU
>>>> POV, it is a very different device. Just the guest does not notice.
>>>
>>> It's like qcow2 and raw/native IO are wire differently, or virtio-net
>>> and vhost-net.  But it's the same IDE device or virtio NIC.
>>
>> That would mean introducing a backend/frontend concept for irqchips.
> 
> We could do it, have one ioapic model with ioapic_ops->eoi_broadcast(). 
> Most of the interfaces already dispatch dynamically (qdev gpio/irq) so
> there wouldn't be much more there.

The problem is configuration. Just by setting ioapic.backend=xxx, we
cannot pass down parameters that are backend-specific. We could ignore
this issue and make all specific parameters visible via the frontend.
Would be slightly ugly.

> 
> To me, how it's actually implemented is not important.  What is
> important is that save/restore, the monitor, and the guest don't notice
> any changes.

I widely agree, except that differentiation (or backend awareness) has
to be preserved in the monitor.

Jan

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