Re: [PATCH 10/10] Introduce MSI message sending interface that bypass IRQ routing.

[Gleb Natapov <gleb@redhat.com>]

On Sun, Aug 09, 2009 at 06:01:15PM +0300, Avi Kivity wrote:
> On 08/09/2009 05:52 PM, Gleb Natapov wrote:
>> On Sun, Aug 09, 2009 at 05:56:30PM +0300, Avi Kivity wrote:
>>    
>>> On 08/09/2009 03:41 PM, Gleb Natapov wrote:
>>>      
>>>> Sending of MSI using IRQ routing is an artificial concept and potentially
>>>> big number of MSIs (2048 per device) make it also inefficient. This
>>>> patch adds an interface to inject MSI messages from userspace to lapic
>>>> logic directly. The patch also reduces the maximum number of IRQ routing
>>>> entries to 128 since MSIs will no longer go there and 128 entries cover
>>>> 5 ioapics and this ought to be enough for anybody.
>>>>        
>>> In the future many MSIs will be triggered via irqfds, and those
>>> require irq routing.
>>>
>>>      
>> Why? My plan is to change irqfd to use the MSI functions.
>>
>>    
>
> It's still an "install handle, call handle" interface.  Maybe it would  
> have been better to start off with your new interface, but having both  
> is too much for too little gain.
>
Is it not too late to change interface? There was no released kernel with
irqfd yet. And this just adds another level of indirection and one more
point of false cache sharing.

--
			Gleb.