Hi Andy,

On Thu, Mar 19, 2020 at 08:35:59AM -0700, Andy Lutomirski wrote:

On Thu, Mar 19, 2020 at 2:14 AM Joerg Roedel [off-list ref] wrote:

quoted

From: Joerg Roedel <redacted>

Keep NMI state in SEV-ES code so the kernel can re-enable NMIs for the
vCPU when it reaches IRET.

IIRC I suggested just re-enabling NMI in C from do_nmi().  What was
wrong with that approach?

If I understand the code correctly a nested NMI will just reset the
interrupted NMI handler to start executing again at 'restart_nmi'.
The interrupted NMI handler could be in the #VC handler, and it is not
safe to just jump back to the start of the NMI handler from somewhere
within the #VC handler.

So I decided to not allow NMI nesting for SEV-ES and only re-enable the
NMI window when the first NMI returns. This is not implemented in this
patch, but I will do that once Thomas' entry-code rewrite is upstream.

This causes us to pop the NMI frame off the stack.  Assuming the NMI
restart logic is invoked (which is maybe impossible?), we get #DB,
which presumably is actually delivered.  And we end up on the #DB
stack, which might already have been in use, so we have a potential
increase in nesting.  Also, #DB may be called from an unexpected
context.

An SEV-ES hypervisor is required to intercept #DB, which means that the
#DB exception actually ends up being a #VC exception. So it will not end
up on the #DB stack.

Now somehow #DB is supposed to invoke #VC, which is supposed to do the
magic hypercall, and all of this is supposed to be safe?  Or is #DB
unconditionally redirected to #VC?  What happens if we had no stack
(e.g. we interrupted SYSCALL) or we were already in #VC to begin with?

Yeah, as I said above, the #DB is redirected to #VC, as the hypervisor
has to intercept #DB.

The stack-problem is the one that prevents the Single-step-over-iret
approach right now, because the NMI can hit while in kernel mode and on
entry stack, which the generic entry code (besided NMI) does not handle.
Getting a #VC exception there (like after an IRET to that state) breaks
things.

Last, in this version of the patch-set the #VC handler became
nesting-safe. It detects whether the per-cpu GHCB is in use and
safes/restores its contents in this case.

I think there are two credible ways to approach this:

1. Just put the NMI unmask in do_nmi().  The kernel *already* knows
how to handle running do_nmi() with NMIs unmasked.  This is much, much
simpler than your code.

Right, and I thought about that, but the implication is that the
complexity is moved somewhere else, namely into the #VC handler, which
then has to be restartable.

2. Have an entirely separate NMI path for the
SEV-ES-on-misdesigned-CPU case.  And have very clear documentation for
what prevents this code from being executed on future CPUs (Zen3?)
that have this issue fixed for real?

That sounds like a good alternative, I will investigate this approach.
The NMI handler should be much simpler as it doesn't need to allow NMI
nesting. The question is, does the C code down the NMI path depend on
the NMI handlers stack frame layout (e.g. the in-nmi flag)?

Regards,

	Joerg