Re: [PATCH v9 09/23] coco/tdx-host: Don't expose P-SEAMLDR information on CPUs with erratum

[Edgecombe, Rick P]

On Mon, 2026-05-18 at 08:29 -0700, Dave Hansen wrote:
> On 5/18/26 05:44, Chao Gao wrote:
> > On Fri, May 15, 2026 at 10:26:19AM -0700, Dave Hansen wrote:
> > > On 5/13/26 08:09, Chao Gao wrote:
> > > > Some TDX-capable CPUs have an erratum, as documented in Intel® Trust
> > > > Domain CPU Architectural Extensions (May 2021 edition) Chapter 2.3:
> > > 2021, eh?
> > The TDX ISA document has not been updated since then; the May 2021
> > edition is still the latest revision. See:
> > 
> > https://www.intel.com/content/www/us/en/developer/tools/trust-domain-
> > extensions/documentation.html
> 
> I think you are saying that the CPUs have an erratum.
> 
> That erratum diverges their implementation from the spec: "Intel® Trust
> Domain CPU Architectural Extensions (May 2021 edition) Chapter 2.3".

It actually is documented in that May 2021 spec as the architectural behavior.
But it looks like not earlier, because the doc said it is new verbiage on that
one.

> 
> But when you combine those two things in one sentence, it's incredibly
> confusing.
> 
> The erratum you are talking about is brand new. I just asked for it to
> be created in the last month or two. Thus, my confusion when you say
> there: "an erratum, as documented in ... May 2021".
> 
> Thus, I'm questioning the 2021 date. You probably also want to mention
> that the erratum is, as of today, not publicly documented.
> 
> Can you rephrase this all and make it clearer, please?

So I guess we want to explain:
1. The problematic VMCS clearing behavior
2. That the problematic behavior is only documented in later docs (right?)
3. That it will be documented as an erratum later, and checked via the bit

Maybe something like?

Some TDX-capable CPUs have an erratum where SEAMRET clears the current VMCS
pointer. The behavior relies on the VMM to reload the current VMCS pointer.
However, that is a problem for KVM because clearing the current VMCS pointer
behind KVM's back will break KVM. While the VMCS clearing is documented as the
actual architecture in later versions of the "Intel® Trust Domain CPU
Architectural Extensions"[0] documents, it is not present in the earlier ones. 

Future docs will describe this SEAMRET VMCS clearing behavior as being present
when IA32_VMX_BASIC[60] is set...