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

From: Yu Zhang
Date: Thu Sep 02 2021 - 04:19:40 EST


On Wed, Sep 01, 2021 at 09:07:59AM -0700, Andy Lutomirski wrote:
> On 9/1/21 3:24 AM, Yu Zhang wrote:
> > On Tue, Aug 31, 2021 at 09:53:27PM -0700, Andy Lutomirski wrote:
> >>
> >>
> >> On Thu, Aug 26, 2021, at 7:31 PM, Yu Zhang wrote:
> >>> On Thu, Aug 26, 2021 at 12:15:48PM +0200, David Hildenbrand wrote:
> >>
> >>> 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?
> >>>
> >>> If yes. The fd based solution may need change the VFIO interface as well(
> >>> though the fake swap entry solution need mess with VFIO too). Because:
> >>>
> >>> 1> KVM uses VFIO when assigning devices into a VM.
> >>>
> >>> 2> Not knowing which GPA ranges may be used by the VM as DMA buffer, all
> >>> guest pages will have to be mapped in host IOMMU page table to host pages,
> >>> which are pinned during the whole life cycle fo the VM.
> >>>
> >>> 3> IOMMU mapping is done during VM creation time by VFIO and IOMMU driver,
> >>> in vfio_dma_do_map().
> >>>
> >>> 4> However, vfio_dma_do_map() needs the HVA to perform a GUP to get the HPA
> >>> and pin the page.
> >>>
> >>> But if we are using fd based solution, not every GPA can have a HVA, thus
> >>> the current VFIO interface to map and pin the GPA(IOVA) wont work. And I
> >>> doubt if VFIO can be modified to support this easily.
> >>>
> >>>
> >>
> >> Do you mean assigning a normal device to a protected VM or a hypothetical protected-MMIO device?
> >>
> >> If the former, it should work more or less like with a non-protected VM. mmap the VFIO device, set up a memslot, and use it. I'm not sure whether anyone will actually do this, but it should be possible, at least in principle. Maybe someone will want to assign a NIC to a TDX guest. An NVMe device with the understanding that the guest can't trust it wouldn't be entirely crazy ether.
> >>
> >> If the latter, AFAIK there is no spec for how it would work even in principle. Presumably it wouldn't work quite like VFIO -- instead, the kernel could have a protection-virtual-io-fd mechanism, and that fd could be bound to a memslot in whatever way we settle on for binding secure memory to a memslot.
> >>
> >
> > Thanks Andy. I was asking the first scenario.
> >
> > Well, I agree it is doable if someone really want some assigned
> > device in TD guest. As Kevin mentioned in his reply, HPA can be
> > generated, by extending VFIO with a new mapping protocol which
> > uses fd+offset, instead of HVA.
>
> I'm confused. I don't see why any new code is needed for this at all.
> Every proposal I've seen for handling TDX memory continues to handle TDX
> *shared* memory exactly like regular guest memory today. The only
> differences are that more hole punching will be needed, which will
> require lightweight memslots (to have many of them), memslots with
> holes, or mappings backing memslots with holes (which can be done with
> munmap() on current kernels).

Thanks for pointing this out. And yes, for DMAs not capable of encryption(
which is the case in current TDX). GUP shall work as it is in VFIO. :)

>
> So you can literally just mmap a VFIO device and expect it to work,
> exactly like it does right now. Whether the guest will be willing to
> use the device will depend on the guest security policy (all kinds of
> patches about that are flying around), but if the guest tries to use it,
> it really should just work.
>

But I think there's still problem. For now,

1> Qemu mmap()s all GPAs into its HVA space, when the VM is created.
2> With no idea which part of guest memory shall be shared, VFIO will just
set up the IOPT, by mapping whole GPA ranges in IOPT.
3> And those GPAs are actually private ones, with no shared-bit set.

Later when guest tries to perform a DMA(using a shared GPA), IO page fault
shall happen.

> >
> > Another issue is current TDX does not support DMA encryption, and
> > only shared GPA memory shall be mapped in the VT-d. So to support
> > this, KVM may need to work with VFIO to dynamically program host
> > IOPT(IOMMU Page Table) when TD guest notifies a shared GFN range(e.g.,
> > with a MAP_GPA TDVMCALL), instead of prepopulating the IOPT at VM
> > creation time, by mapping entire GFN ranges of a guest.
>
> Given that there is no encrypted DMA support, shouldn't the only IOMMU
> mappings (real host-side IOMMU) that point at guest memory be for
> non-encrypted DMA? I don't see how this interacts at all. If the guest
> tries to MapGPA to turn a shared MMIO page into private, the host should
> fail the hypercall because the operation makes no sense.
>
> It is indeed the case that, with a TDX guest, MapGPA shared->private to
> a page that was previously used for unencrypted DMA will need to avoid
> having IOPT entries to the new private page, but even that doesn't seem
> particularly bad. The fd+special memslot proposal for private memory
> means that shared *backing store* pages never actually transition
> between shared and private without being completely freed.
>
> As far as I can tell, the actual problem you're referring to is:
>
> >>> 2> Not knowing which GPA ranges may be used by the VM as DMA buffer, all
> >>> guest pages will have to be mapped in host IOMMU page table to host
> pages,
> >>> which are pinned during the whole life cycle fo the VM.

Yes. That's the primary concern. :)

>
> In principle, you could actually initialize a TDX guest with all of its
> memory shared and all of it mapped in the host IOMMU. When a guest
> turns some pages private, user code could punch a hole in the memslot,
> allocate private memory at that address, but leave the shared backing
> store in place and still mapped in the host IOMMU. The result would be
> that guest-initiated DMA to the previously shared address would actually
> work but would hit pages that are invisible to the guest. And a whole
> bunch of memory would be waste, but the whole system should stll work.

Do you mean to let VFIO & IOMMU to treat all guest memory as shared first,
and then just allocate the private pages in another backing store? I guess
that could work, but with the cost of allocating roughly 2x physical pages
of the guest RAM size. After all, the shared pages shall be only a small
part of guest memory.

If device assignment is desired in current TDX. My understanding of the
enabling work would be like this:
1> Change qemu to not trigger VFIO_IOMMU_MAP_DMA for the TD, thus no IOPT
prepopulated, and no physical page allocated.
2> KVM forwards MapGPA(private -> shared) request to Qemu.
3> Qemu asks VFIO to pin and map the shared GPAs.

For private -> shared transitions, the memslot punching, IOPT unmapping,
and iotlb flushing are necessary. Possibly new interface between VFIO and
KVM is needed.

But actually I am not sure if people really want assigned device in current
TDX. Bottleneck of the performance should be the copying to/from swiotlb
buffers.

B.R.
Yu