[libvirt] [PATCH 09/18] Add virtio-related options to interfaces

Jason Wang jasowang at redhat.com
Thu Jun 1 06:39:59 UTC 2017

On 2017年05月30日 21:09, Daniel P. Berrange wrote:
> On Tue, May 30, 2017 at 02:50:32PM +0200, Ján Tomko wrote:
>> <interface type='user'>
>>    <mac address='52:54:56:5a:5c:5e'/>
>>    <model type='virtio'/>
>>    <driver iommu_platform='on' ats='on'/>
>> </interface>
> Feels like we could just use  'iommu=on' as the name.
> Also, I'm unclear whether we actually need the 'ats' attribute.
>  From the description it sounds like if we have device_iotlb=on
> for the IOMMU device, then we should have ats=on too. Is there
> a  reason to use ats=off when device_iotlb=on, or vica-verca.
> If not, then we should just do the right thing and set ats=on
> whenever we see device_iotlb=on.

Maybe for virtio-net only. Vhost can get best performance only when both 
were set, disabling either will be slower.  For the cases of non vhost 
and other virtio devices, enabling ATS maybe a little bit slower since 
we don't implement device IOTLB in the device now (though we may want to 
implement it in the future). So for now, it's better to leave a choice 
for user.

Another problem is, ATS depends on PCIE, there's some requirement e.g 
qemu does not support an integrated PCIE device which means we may need 
a PCIE switch (pcie-root-port) to make PCIE and ATS works.


> Regards,
> Daniel

More information about the libvir-list mailing list