Re: [PATCH V4 1/1] KVM: TDX: Add sub-ioctl KVM_TDX_TERMINATE_VM

[Vishal Annapurve]

On Mon, Jun 23, 2025 at 2:39 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Mon, Jun 23, 2025, Vishal Annapurve wrote:
> > On Fri, Jun 20, 2025 at 9:14 AM Vishal Annapurve <vannapurve@xxxxxxxxxx> wrote:
> > > Adrian's suggestion makes sense and it should be functional but I am
> > > running into some issues which likely need to be resolved on the
> > > userspace side. I will keep this thread updated.
> > >
> > > Currently testing this reboot flow:
> > > 1) Issue KVM_TDX_TERMINATE_VM on the old VM.
> > > 2) Close the VM fd.
> > > 3) Create a new VM fd.
> > > 4) Link the old guest_memfd handles to the new VM fd.
> > > 5) Close the old guest_memfd handles.
> > > 6) Register memslots on the new VM using the linked guest_memfd handles.
> > >
> >
> > Apparently mmap takes a refcount on backing files.
>
> Heh, yep.
>
> > So basically I had to modify the reboot flow as:
> > 1) Issue KVM_TDX_TERMINATE_VM on the old VM.
> > 2) Close the VM fd.
> > 3) Create a new VM fd.
> > 4) Link the old guest_memfd handles to the new VM fd.
> > 5) Unmap the VMAs backed by the guest memfd
> > 6) Close the old guest_memfd handles. -> Results in VM destruction
> > 7) Setup new VMAs backed by linked guest_memfd handles.
> > 8) Register memslots on the new VM using the linked guest_memfd handles.
> >
> > I think the issue simply is that we have tied guest_memfd lifecycle
> > with VM lifecycle and that discussion is out of scope for this patch.
>
> I wouldn't say it's entirely out of scope.  E.g. if there's a blocking problem
> _in the kernel_ that prevents utilizing KVM_TDX_TERMINATE_VM, then we definitely
> want to sort that out before adding support for KVM_TDX_TERMINATE_VM.
>
> But IIUC, the hiccups you've encountered essentially fall into the category of
> "working as intended", albeit with a lot of not-so-obvious behaviors and dependencies.

Yes, that's correct. The "issue" I referred to above is just extra
steps that need to be taken by userspace VMM and is still in the
"working as intended" category.