On Fri, Nov 19, 2021 at 10:21:39PM +0000, Sean Christopherson wrote:

On Fri, Nov 19, 2021, Jason Gunthorpe wrote:

quoted

On Fri, Nov 19, 2021 at 07:18:00PM +0000, Sean Christopherson wrote:

quoted

No ideas for the kernel API, but that's also less concerning since
it's not set in stone.  I'm also not sure that dedicated APIs for
each high-ish level use case would be a bad thing, as the semantics
are unlikely to be different to some extent.  E.g. for the KVM use
case, there can be at most one guest associated with the fd, but
there can be any number of VFIO devices attached to the fd.

Even the kvm thing is not a hard restriction when you take away
confidential compute.

Why can't we have multiple KVMs linked to the same FD if the memory
isn't encrypted? Sure it isn't actually useful but it should work
fine.

Hmm, true, but I want the KVM semantics to be 1:1 even if memory
isn't encrypted.

That is policy and it doesn't belong hardwired into the kernel.

Your explanation makes me think that the F_SEAL_XX isn't defined
properly. It should be a userspace trap door to prevent any new
external accesses, including establishing new kvms, iommu's, rdmas,
mmaps, read/write, etc.

It's not just avoiding the linked list, there's a trust element as
well.  E.g. in the scenario where a device can access a confidential
VM's encrypted private memory, the guest is still the "owner" of the
memory and needs to explicitly grant access to a third party,
e.g. the device or perhaps another VM.

Authorization is some other issue - the internal kAPI should be able
to indicate it is secured memory and the API user should do whatever
dance to gain access to it. Eg for VFIO ask the realm manager to
associate the pci_device with the owner realm.

Jason