On Thu, Apr 12, 2018 at 11:43:41AM +0800, Jason Wang wrote: > > > On 2018年04月12日 11:35, Michael S. Tsirkin wrote: > > > There are more advantages to using request with PASID: > > > > > > You can use hardware support for nesting, having guest supply 1st level > > > translation and host second level translation. > > > > > > I actually had an idea to do something like this for AMD > > > and ARM which support nesting even for requests with PASID, > > > having intel benefit too would be nice. > > Something else to consider is implementing PRS capability. > > > > > > In theory this could then go like this: > > > > - get page request from device > > - fetch request from VTD page tables > > - use response to issue a page response message > > > > > > This would match the current vhost-user model. > > This requires IOMMU driver can forward this to VFIO and then VFIO can > forward it to userspace. Looks like a lot of changes and it would be even > slower than what is proposed in this patch. > > Thanks
It would work better for a static table as only accessed pages would need to be sent. Slower for the dynamic case but dynamic case needs hardware support to work properly in any case. -- MST
