On Mon, Feb 08, 2021, Peter Zijlstra wrote:
> On Mon, Feb 08, 2021 at 08:23:01AM -0800, Andi Kleen wrote:
> > > > +#ifdef CONFIG_INTEL_TDX_GUEST
> > > > +DEFINE_IDTENTRY(exc_virtualization_exception)
> > > > +{
> > > > + struct ve_info ve;
> > > > + int ret;
> > > > +
> > > > + RCU_LOCKDEP_WARN(!rcu_is_watching(), "entry code didn't wake
> > > > RCU");
> > > > +
> > > > + /* Consume #VE info before re-enabling interrupts */
> > >
> > > So what happens if NMI happens here, and triggers a nested #VE ?
> >
> > Yes that's a gap. We should probably bail out and reexecute the original
> > instruction. The VE handler would need to set a flag for that.
No, NMI cannot happen here. The TDX-Module "blocks" NMIs until the #VE info is
consumed by the guest.
> > Or alternatively the NMI always gets the VE information and puts
> > it on some internal stack, but that would seem clunkier.
>
> The same is possible with MCE and #DB I imagine.
The MCE "architecture" for a TDX guest is rather stupid. The guest is required
to keep CR4.MCE=1, but at least for TDX 1.0 the VMM is not allowed to inject
#MC.
So, for better or worse, #MC is a non-issue.
#VE->#DB->#VE would be an issue, presumably this needs to be noinstr (or
whatever
it is that prevents #DBs on functions).
