Re: [Qemu-devel] [PATCH 28/40] xenner: libxc emu: evtchn

[Alexander Graf]

On 01.11.2010, at 12:14, Anthony Liguori wrote:

> On 11/01/2010 11:07 AM, Alexander Graf wrote:
>> On 01.11.2010, at 12:01, Anthony Liguori wrote:
>> 
>>   
>>> On 11/01/2010 10:49 AM, Alexander Graf wrote:
>>>     
>>>> On 01.11.2010, at 11:45, Anthony Liguori wrote:
>>>> 
>>>> 
>>>>       
>>>>> On 11/01/2010 10:01 AM, Alexander Graf wrote:
>>>>> 
>>>>>         
>>>>>> Xenner emulates parts of libxc, so we can not use the real xen 
>>>>>> infrastructure
>>>>>> when running xen pv guests without xen.
>>>>>> 
>>>>>> This patch adds support for event channel communication.
>>>>>> 
>>>>>> Signed-off-by: Alexander Graf<address@hidden>
>>>>>> 
>>>>>> 
>>>>>>           
>>>>> Has anyone checked with the Xen folks about supporting this type of 
>>>>> functionality in libxc directly?
>>>>> 
>>>>>         
>>>> The issue I have with libxc is that it goes orthogonal to the qemu 
>>>> infrastructure way of doing things. If we base on libxc, we will never be 
>>>> able to do cross-architecture execution of xen pv guests. Do we really 
>>>> want to go that way?
>>>> 
>>>>       
>>> IIUC, this is a mini-libxc that you enable by mucking with LD_LIBRARY_PATH 
>>> such that you can run things like xenstored unmodified.  What I'm really 
>>> asking is whether there has been a discussion about a more pleasant way to 
>>> do this that the Xen guys would feel comfortable with.
>>> 
>>> I'd feel a little weird if someone was replacing a part of QEMU via 
>>> LD_LIBRARY_PATH trickery.  It's better to try to work out a proper solution 
>>> with the upstream community than to do trickery.
>>> 
>>> I'm not entirely opposed to this if the Xen guys say they don't want 
>>> anything to do with Xenner, but we should have the discussion at least.
>>>     
>> I agree about the discussion part, that's why we're all gathering in Boston 
>> this week, right?
>>   
> 
> Fair enough :-)
> 
>> But technically, this code really just bumps all libxc calls to indirect 
>> function calls that go through a struct. If we're using xenner, we use our 
>> own implementation, if we're using xen, we use xen's. The thing is that with 
>> xenner we usually don't have xen infrastructure available and most likely 
>> don't want to start any either.
>>   
> 
> Yeah, I guess I'd just like to see a more "polite" solution.

We can try and see if we can maybe reuse parts of the event channel and 
xenstored stuff, but when it comes to memory mappings or grant tables, we have 
to have our own code since we're the ones owning the ram.

But yeah, let's move that discussion to LPC :). That way the xen folks can 
participate!


Alex