Re: [Qemu-devel] Re: [PATCH] create kvm-shared-all.c and kvm-shared.h

[Jan Kiszka <jan.kiszka@siemens.com>]

Glauber Costa wrote:
> On Tue, Jun 09, 2009 at 06:42:44PM +0200, Jan Kiszka wrote:
>> Glauber Costa wrote:
>>> On Tue, Jun 09, 2009 at 07:26:58PM +0300, Avi Kivity wrote:
>>>> Jan Kiszka wrote:
>>>>> Glauber Costa wrote:
>>>>>   
>>>>>> Following a suggestion given by Jan, the idea here is to
>>>>>> move shared pieces between qemu and qemu-kvm.git into a common
>>>>>> file, so we can do sharing while avoid clashes.
>>>>>>
>>>>>> In the future, this files should disappear.
>>>>>>
>>>>>>     
>>>>> OK for the header - but why do we have to push the ioctl services into a
>>>>> separate module? Will all functions qemu-kvm start to use from upstream
>>>>> have to be pushed around? Or what is special about the ioctls?
>>>>>
>>>>> I rather think qemu-kvm should build kvm-all.c and #ifdef out those
>>>>> parts which collide with its own implementation. Moreover, when we morph
>>>>> qemu-kvm services for upstream, this could already happen where they
>>>>> shall once be located: in kvm-all.c or target-*/kvm.c.
>>>>>   
>>>> Yes, we could simply append libkvm-all.c and qemu-kvm.c to kvm-all.c,  
>>>> and gradually include more of kvm-all.c as we delete parts of libkvm.c  
>>>> and qemu-kvm.c.
>>> I tried it myself, and it generates tons of conflicts. So scary.
>>> I'd prefer to do it in the way I propose, until there is nothing left on kvm-all.c
>>>
>> kvm-all.c:
>>
>> #ifdef DONT_USE_UPSTREAM_YET
>> <upstream code>
>> #endif
>>
>> int kvm_ioctl(KVMState *s, int type, ...)
>> ...
>>
>> #ifdef DONT_USE_UPSTREAM_YET
>> <some more upstream code>
>> #endif
>>
>> <libkvm code>
>>
>>
>> Can't imagine that this is infeasible. And it would all happily live in
>> qemu-kvm, so upstream would never see these hacks.
> yeah, if it is in fact ifdef'd this way, it of course works.
> 
> I can do that, definitely.
> But by doing that, we probably does not even need the header part here too.
> 

I think we need it as long as qemu-kvm-<arch> fiddles with state fields
directly. If that is easily resolvable, I'm all for doing that first and
skipping the kvm-shared.h part in upstream!

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux