Re: [Qemu-devel] [PATCH 5/7] hw/core/sysbus: add fdt_add_node method

[Alexander Graf <agraf@suse.de>]

On 23.07.14 17:33, Eric Auger wrote:
> On 07/08/2014 03:52 PM, Alexander Graf wrote:
>> On 07.07.14 09:08, Eric Auger wrote:
>>> This method is meant to be called on sysbus device dynamic
>>> instantiation (-device option). Devices that support this
>>> kind of instantiation must implement this method.
>>>
>>> Signed-off-by: Eric Auger <eric.auger@linaro.org>
>> For the reason I stated earlier, I don't think it's a good idea to put
>> device tree code into our device models.
> Hi Alex,
>
> I would propose we discuss that topic during next KVM call if you are
> available.

I lost track when that would be. Next week would work fine, the week 
after not :).

> Hope Peter will be available to join too. Because I feel
> stuck between not putting things in the machine file (1) - obviously we
> could put them in a helper module (2) - and not putting them in the
> device (3).
>
> Whatever the solution I fear we are going to pollute something: Any time
> a new device wants to support dynamic instantiation, we would need to
> modify the machine file or the helper module with 1 and 2 resp. In case
> we put it in the device we pollute this latter...
>
> My hope was that quite few QEMU platform devices would need to support
> that feature and hence would need to implement this dt node generation
> method. To me dynamic instantiation of platform device was not the
> mainstream solution.

Quite frankly I don't think it'd be that many. I think we'll cover 99.9% 
of all use cases if we just enable it for the virt machines of e500 and arm.

> Then there is the fundamental question of technical feasibility of
> devising a generic PlatformParams that match all the specialization
> needs? Here I miss experience. In case we know the machine type and a
> small set of additional fields couldn't we do the adaptations you talked
> about, related to IRQs?

The problem is that I don't know all the boards and different things 
people come up with either. There's also no reason machine files have to 
stick to the "platform bus" model - they could just take those devices 
and stick them into an existing other virtual bus.

I don't feel comfortable generalizing something where I'm pretty sure 
things will blow up sooner or later.


Alex