Re: [RFC v2-fix-v5 1/1] x86: Skip WBINVD instruction for VM guest

[Dan Williams]

On Wed, Jun 9, 2021 at 2:03 PM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> This changelog lacks both clear problem statements and a clear solution
> implemented within the patch.
>
> Here's a proposed changelog.  It clearly spells out the two problems
> caused by WBINVD within a guest, and the proposed solution which fixes
> those two problems.

Looks good to me modulo the comment below...

>
> Is this missing anything?
>
> --
>
> VM guests that support ACPI use standard ACPI mechanisms to signal sleep
> state entry to the host.  To ACPI, reboot is simply another sleep state.
>
> ACPI specifies that the platform preserve memory contents over (some)
> sleep states.  It does not specify any requirements for data
> preservation in CPU caches.  The ACPI specification mandates the use of
> WBINVD to flush the contents of the CPU caches to memory before entering
> specific sleep states, thus ensuring data in caches can survive sleep
> state transitions.e
>
> Unlike when entering sleep states bare metal, no actions within a guest
> can cause data in processor caches to be lost.  That makes these WBINVD
> invocations harmless but superfluous within a guest. (<--- problem #1)
>
> In TDX guests, these WBINVD operations cause #VE exceptions.  For debug,
> it would be ideal for the #VE handler to be able to WARN() when an
> unexpected WBINVD occurs. (<--- problem #2)

...but it doesn't WARN() it triggers unhandled #VE, unless I missed
another patch that precedes this that turns it into a WARN()? If a
code path expects WBINVD for correct operation and the guest can't
execute that sounds fatal, not a WARN to me.

> Avoid WBINVD for all ACPI cache-flushing operations which occur while
> running under a hypervisor, which includes TDX guests.  This both avoids
> TDX warnings and optimizes away superfluous WBINVD invocations. (<----
> solution)
>