Re: virtio-blk performance and MSI

["Michael S. Tsirkin" <mst@redhat.com>]

On Sun, Aug 09, 2009 at 01:01:35PM +0300, Avi Kivity wrote:
> On 08/06/2009 07:35 PM, Christoph Hellwig wrote:
>> Michael suggested to me a while ago to try MSI with virtio-blk and I
>> played with this small patch:
>>
>>
>> Index: qemu-kvm/hw/virtio-blk.c
>> ===================================================================
>> --- qemu-kvm.orig/hw/virtio-blk.c
>> +++ qemu-kvm/hw/virtio-blk.c
>> @@ -416,6 +416,7 @@ VirtIODevice *virtio_blk_init(DeviceStat
>>       s->vdev.get_config = virtio_blk_update_config;
>>       s->vdev.get_features = virtio_blk_get_features;
>>       s->vdev.reset = virtio_blk_reset;
>> +    s->vdev.nvectors = 2;

some whitespace damage btw

>>       s->bs = bs;
>>       s->rq = NULL;
>>       if (strlen(ps = (char *)drive_get_serial(bs)))
>>
>> which gave about 5% speedups on 4k sized reads and writes, see the full
>> iozone output I attached.  Now getting the information about using
>> multiple MSI vectors from the command line to virtio-blk similar to how
>> virtio-net does seems extremly messy right now.  Waiting for Gerd's
>> additional qdev patches to make it easier as a qdev property.
>>
>>    
>
> Looks good.  Anthony, I think this applies upstream?

This applies upstream, but we also need the flag to change # of vectors:
for loading old images, but also for troubleshooting.
Thus the qdev dependency.

-- 
MST