Re: [Qemu-devel] Re: [PATCH] virtio: Use ioeventfd for virtqueue notify

[Anthony Liguori <aliguori@linux.vnet.ibm.com>]

On 10/04/2010 11:12 AM, Michael S. Tsirkin wrote:
> On Mon, Oct 04, 2010 at 09:01:14AM -0500, Anthony Liguori wrote:
>    
>> On 10/04/2010 03:04 AM, Avi Kivity wrote:
>>      
>>> On 10/04/2010 03:18 AM, Anthony Liguori wrote:
>>>        
>>>> On 10/03/2010 09:28 AM, Michael S. Tsirkin wrote:
>>>>          
>>>>>            
>>>>>> This is using eventfd as well.
>>>>>>              
>>>>> Sorry, I meant irqfd.
>>>>>            
>>>> I've tried using irqfd in userspace.  It hurts performance quite
>>>> a bit compared to doing an ioctl so I would suspect this too.
>>>>
>>>> A last_used_idx or similar mechanism should help performance
>>>> quite a bit on top of ioeventfd too.
>>>>
>>>>          
>>> Any idea why?  While irqfd does quite a bit of extra locking, it
>>> shouldn't be that bad.
>>>        
>> Not really.  It was somewhat counter intuitive.
>>
>> A worthwhile experiment might be to do some layering violations and
>> have vhost do an irq injection via an ioctl and see what the
>> performance delta is.
>>      
> I think you don't even need to try that hard.
> Just comment this line:
> //   proxy->pci_dev.msix_mask_notifier = virtio_pci_mask_notifier;
> this is what switches to irqfd when msi vector is unmasked.
>    

That drops to userspace though for all irqs, no?

Or did you mean that commenting that line out improves performance 
demonstrating the overhead of irqfd?

Regards,

Anthony Liguori

>    
>>   I suspect it could give vhost a nice boost.
>>
>> Regards,
>>
>> Anthony Liguori
>>