Re: [RFC] KVM: mm: fd-based approach for supporting KVM guest private memory

On Tue, Aug 31, 2021 at 01:39:31PM -0700, Andi Kleen wrote:

On 8/31/2021 1:15 PM, David Hildenbrand wrote:

quoted

On 31.08.21 22:01, Andi Kleen wrote:

quoted

quoted

quoted

Thanks a lot for this summary. A question about the requirement: do
we or
do we not have plan to support assigned device to the protected VM?

Good question, I assume that is stuff for the far far future.

It is in principle possible with the current TDX, but not secure. But
someone might decide to do it. So it would be good to have basic support
at least.

Can you elaborate the "not secure" part? Do you mean, making the device
only access "shared" memory, not secure/encrypted/whatsoever?

Yes that's right. It can only access shared areas.

Thanks, Andy & David.

Actually, enabling of device assinment needs quite some effort, e.g.,
to guarantee only shared pages are mapped in IOMMU page table (using
shared GFNs). And the buffer copying inside TD is still unavoidable,
thus not much performance benefit.

Maybe we should just *disable* VFIO device in TDX first. 

As to the fd-based private memory, enventually we will have to tolerate
its impact on any place where GUP is needed in virtualization. :)

B.R.
Yu