Re: [PATCH 08/10] Move IO APIC to its own lock.

[Gleb Natapov <gleb@redhat.com>]

On Sun, Aug 09, 2009 at 06:10:07PM +0300, Avi Kivity wrote:
> On 08/09/2009 05:57 PM, Gleb Natapov wrote:
>> On Sun, Aug 09, 2009 at 05:54:30PM +0300, Avi Kivity wrote:
>>    
>>> On 08/09/2009 03:41 PM, Gleb Natapov wrote:
>>>      
>>>> Introduce new function kvm_notifier_set_irq() that should be used
>>>> to change irq line level from irq notifiers. When irq notifier
>>>> change irq line level it calls into irq chip code recursively. The
>>>> function avoids taking a lock recursively.
>>>>
>>>>        
>>> This looks really horrible.  I don't have an alternative yet, but I'll
>>> think of one.
>>>      
>> I agree this is not nice. This is needed only for device assignment
>> case. That explains why I don't like device assignment :)
>
> Well, implementation problems can be fixed.  Other issues with device  
> assignment cannot.
>
>> The problem
>> is that the only communication channel from guest to assigned device that
>> goes through the host is interrupt injection/acknowledgement, so we try to
>> do things (lowering IRQ) on this path that usually are done somewhere else.
>>    
>
> You can queue the injection somehow (work struct? special purpose queue  
> examined after unlock?) and avoid the recursive locking.
>
We can't. Line status should be update here and now. Otherwise interrupt
will be immediately reinjected. 

--
			Gleb.