Re: [RFC PATCH 06/15] virt: tdx-guest: Support TDI bind and unbind operations

From: Edgecombe, Rick P

Date: Tue Nov 17 2026 - 17:00:36 EST


On Tue, 2026-11-17 at 18:36 +0800, Xu Yilun wrote:
> On Mon, Nov 16, 2026 at 11:57:23PM +0000, Edgecombe, Rick P wrote:
> > On Thu, 2026-11-12 at 20:56 +0800, Xu Yilun wrote:
> > > > > >
> > > > > > For errors that indicate uncooperative or buggy host, we elsewhere
> > > > > > warn
> > > > > > in
> > > > > the
> > > > > > TDX wrappers. But in coco, not handling an error in the caller can
> > > > > > be an
> > > > > avenue
> > > > > > to attack the guest. So I guess we would commonly need to handle
> > > > > > these
> > >
> > > The secure firmware should guarantee the attack can't impact TSM
> > > security.
> >
> > We had some issues in core TDX guest where if the VMM lied to the guest it
> > could
> > cause shared pages to be returned page allocator. Due to an error not
> > getting
> > handled properly in many callers. This led to a debate about whether we
> > should
> > even handle the error, or just panic the guest if we see weird responses
> > from
> > the VMM. Ultimately we settled on a warn, due to the face that panic_on_warn
> > should be set. For x86 code at least, we should be consistent.
>
> I see. IIUC the problem is not the VMM lies or not, it is the guest
> itself not be consistent about the page attribute: The caller ignores
> the error and see the page private, but the guest PTE maps it share.
>
> Actually the guest can't tell if the VMM lies by GHCI return code,
> success or failure.

Exactly.

> The guest could alternatively ignore the return
> code, update its PTE and return success, later the HW can guard the
> inconsistency between the guest's view and the physical state when
> accessing the page.
>
> Of course the GHCI error detecting is useful to keep system stable when
> working with benigh VMM. So I agree that we propagate errors and handle
> in the caller for TDI LOCK/ATTESTATION/ACCEPT phase. But I also feel OK
> to treat UNLOCK as always success - the kernel rarely fails on exit.
> As long as the guest doesn't mess its own TDI state tracking, we are safe.
>
> >
> > So it was more about exposure to bugs then higher level issues.
> >
> > >   For DDOS attack, the guest could always move the device to
> > > untrusted mode.
> >
> > DDOS attack?
>
> Sorry for the typo, DoS (Denial of Service).

That is what I thought you meant. But DOS of guest by the host is not in the
coco threat model. What is DOSing what here?