Re: [PATCH 2/9] KVM: nVMX: Initialize #VE info page for vmcs02 when proving #VE support

[Huang, Kai]

On 21/05/2024 12:21 pm, Sean Christopherson wrote:
> On Tue, May 21, 2024, Kai Huang wrote:
> > On 21/05/2024 11:22 am, Sean Christopherson wrote:
> > > On Tue, May 21, 2024, Kai Huang wrote:
> > > > On 18/05/2024 12:04 pm, Sean Christopherson wrote:
> > > > > Point vmcs02.VE_INFORMATION_ADDRESS at the vCPU's #VE info page when
> > > > > initializing vmcs02, otherwise KVM will run L2 with EPT Violation #VE
> > > > > enabled and a VE info address pointing at pfn 0.
> > > >
> > > > How about we just clear EPT_VIOLATION_VE bit in 2nd_exec_control
> > > > unconditionally for vmcs02?
> > >
> > > Because then KVM wouldn't get any EPT Violation #VE coverage for L2, and as
> > > evidence by the KVM-Unit-Test failure, running L2 with EPT Violation #VEs enabled
> > > provides unique coverage.  Doing so definitely provides coverage beyond what is
> > > strictly needed for TDX, but it's just as easy to set the VE info page in vmcs02
> > > as it is so clear EPT_VIOLATION_VE, so why not.
> > >
> > > > Your next patch says:
> > > >
> > > > "
> > > > Always handle #VEs, e.g. due to prove EPT Violation #VE failures, in L0,
> > > > as KVM does not expose any #VE capabilities to L1, i.e. any and all #VEs
> > > > are KVM's responsibility.
> > > > "
> > >
> > > I don't see how that's relevant to whether or not KVM enables EPT Violation #VEs
> > > while L2 is running.  That patch simply routes all #VEs to L0, it doesn't affect
> > > whether or not it's safe to enable EPT Violation #VEs for L2.
> >
> > My logic is, if #VE exit cannot possibly happen for L2, then we don't need
> > to deal whether to route #VE exits to L1. :-)
> >
> > Well, actually I think conceptually, it kinda makes sense to route #VE exits
> > to L1:
> >
> > L1 should never enable #VE related bits so L1 is certainly not expecting to
>
> Not "should never", "can never".  If L1 attempts to enable EPT_VIOLATION_VE, then
> VM-Enter will VM-Fail.
>
> > see #VE from L2.  But how to act should be depending on L1's logic? E.g., it
> > can choose to ignore, or just kill the L2 etc?
>
> No.  Architecturally, from L1's perspective, a #VE VM-Exit _cannot_ occur in L2.
> L1 can inject a #VE into L2, but a #VE cannot be generated by the CPU and thus
> cannot cause a VM-Exit.

OK.  The point is not to argue about L1 how to handle, but whether we 
should inject to L1 -- L1 can do whatever it believes legal/sane.

But I understand the purpose is to test/validate, so it's fine for L0 to 
handle, and by handle it eventually means we want to just dump that #VE 
exit.

But now L0 always handles #VE exits from L2, and AFAICT L0 will just 
kill the L1, until the patch:

	KVM: VMX: Don't kill the VM on an unexpected #VE

lands.

So looks that patch at least should be done first.  Otherwise it doesn't 
make a lot sense to kill L1 for #VE exits from L2.

> > Unconditionally disable #VE in vmcs02 can avoid such issue because it's just
> > not possible for L2 to have the #VE exit.
>
> Sure, but by that argument we could just avoid all nested VMX issues by never
> enabling anything for L2.
>
> If there's an argument to be made for disabling EPT_VIOLATION_VE in vmcs02, it's
> that the potential maintenance cost of keeping nEPT, nVMX, and the shadow MMU
> healthy outweighs the benefits.  I.e. we don't have a use case for enabling
> EPT_VIOLATION_VE while L2 is running, so why validate it?

Yeah.  I am not sure the purpose of validating #VE exits from L2.

> If whatever bug the KUT EPT found ends up being a KVM bug that specifically only
> affects nVMX, then it'd be worth revisiting whether or not it's worth enabling
> EPT_VIOLATION_VE in vmcs02.  But that's a rather big "if" at this point.

OK.