Re: [PATCH v5 07/17] KVM: x86: add fields to struct kvm_arch for CoCo features

[Sean Christopherson <seanjc@google.com>]

On Wed, May 08, 2024, Rick P Edgecombe wrote:
> On Tue, 2024-05-07 at 17:21 -0700, Sean Christopherson wrote:
> > > Can you elaborate on the reason for a per-memslot flag? We are discussing
> > > this
> > > design point internally, and also the intersection with the previous
> > > attempts to
> > > do something similar with a per-vm flag[0].
> > > 
> > > I'm wondering if the intention is to try to make a memslot flag, so it can
> > > be
> > > expanded for the normal VM usage.
> > 
> > Sure, I'll go with that answer.  Like I said, off-the-cuff.
> > 
> > There's no concrete motiviation, it's more that _if_ we're going to expose
> > a knob to userspace, then I'd prefer to make it as precise as possible to
> > minimize the changes of KVM ending up back in ABI hell again.
> > 
> > > Because the discussion on the original attempts, it seems safer to keep this
> > > behavior more limited (TDX only) for now.  And for TDX's usage a struct kvm
> > > bool fits best because all memslots need to be set to zap_leafs_only = true,
> > > anyway.
> > 
> > No they don't.  They might be set that way in practice for QEMU, but it's
> > not strictly required.  E.g. nothing would prevent a VMM from exposing a
> > shared- only memslot to a guest.  The memslots that burned KVM the first
> > time around were related to VFIO devices, and I wouldn't put it past
> > someone to be crazy enough
> > to expose an passhtrough an untrusted device to a TDX guest.
> 
> Ok, thanks for clarification. So it's more of a strategic thing to move more
> zapping logic into userspace so the logic can change without introducing kernel
> regressions.

You're _really_ reading too much into my suggestion.  As above, my suggestion
was very spur of the momemnt.  I haven't put much thought into the tradeoffs and
side effects.