On Thu, 2022-04-07 at 18:11 +0000, Sean Christopherson wrote:
> On Thu, Apr 07, 2022, Paolo Bonzini wrote:
> > On 4/7/22 16:39, Sean Christopherson wrote:
> > > On Thu, Apr 07, 2022, Paolo Bonzini wrote:
> > > > On 4/7/22 06:15, Kai Huang wrote:
> > > > > > +static int handle_tdvmcall(struct kvm_vcpu *vcpu)
> > > > > > +{
> > > > > > + struct vcpu_tdx *tdx = to_tdx(vcpu);
> > > > > > +
> > > > > > + if (unlikely(tdx->tdvmcall.xmm_mask))
> > > > > > + goto unsupported;
> > > > > Put a comment explaining this logic?
> > > > >
> > > >
> > > > This only seems to be necessary for Hyper-V hypercalls, which however are
> > > > not supported by this series in TDX guests (because the kvm_hv_hypercall
> > > > still calls kvm_*_read, likewise for Xen).
> > > >
> > > > So for now this conditional can be dropped.
> > >
> > > I'd prefer to keep the sanity check, it's a cheap and easy way to detect a clear
> > > cut guest bug.
> >
> > I don't think it's necessarily a guest bug, just silly but valid behavior.
>
> It's a bug from a security perspective given that letting the host unnecessarily
> manipulate register state is an exploit waiting to happen.
Security perspective from guest's or host's respective? It's guest's
responsibility to make sure itself doesn't expose unnecessarily security holes.
In this particular case, if guest does, then it's guest's business, but I don't
see how it can compromise host's security, or other VM's security?
As Paolo said it's a valid operation from guest, perhaps an alternative is KVM
can unconditionally clear XMM registers, instead of rejecting this VMCALL.
>
> Though for KVM to reject the TDVMCALLs, the GHCI should really be updated to state
> that exposing more state than is required _may_ be considered invalid ("may" so
> that KVM isn't required to check the mask on every exit, which IMO is beyond tedious).
Independent from this issue, I guess it's good for GHCI to have this anyway.
