Re: [RFC PATCH 00/20] TD-Preserving updates

[Chao Gao <chao.gao@intel.com>]

On Fri, Jul 11, 2025 at 07:06:21AM -0700, Sean Christopherson wrote:
>On Fri, Jul 11, 2025, Chao Gao wrote:
>> >2. P-SEAMLDR seamcalls (specificially SEAMRET from P-SEAMLDR) clear current
>> >   VMCS pointers, which may disrupt KVM. To prevent VMX instructions in IRQ
>> >   context from encountering NULL current-VMCS pointers, P-SEAMLDR
>> >   seamcalls are called with IRQ disabled. I'm uncertain if NMIs could
>> >   cause a problem, but I believe they won't. See more information in patch 3.
>
>NMIs shouldn't be a problem.  KVM does access the current VMCS in NMI context
>(to do VMREAD(GUEST_RIP) in response to a perf NMI), but only when KVM knows the

Yes. I also think the guest NMI handler is the only case where VMREAD/VMWRITE is
done in NMI context.

>NMI occurred in KVM's run loop.  So in effect, only in KVM_RUN context, which I
>gotta image is mutually exclusive with tdx_fw_write().

Just a heads-up: P-SEAMLDR may gain other functions and be called from other
code paths, but they won't interfere with the guest NMI handler or KVM_RUN
context.

>
>It'd be nice if we could make the P-SEAMLDR calls completely NMI safe, but
>practically speaking, if KVM (or any other hypervisor) is playing with the VMCS
>in arbitrary NMI handlers, then we've probably got bigger issues.

Agreed.

It's a little late to change the CPU behavior about SEAMRET, as several CPU
generations have already been shipped. Implementing new behavior would require
a new feature bit, which could complicate the host kernel code because the
kernel would need to perform save/restore VMCS conditionally based on this new
feature. So, let's pursue a hardware change unless it becomes a practical issue
for hypervisors.