Re: [PATCH v4 3/3] KVM: nVMX: Emulate EPTP switching for the L1 hypervisor

[Radim KrÄmÃÅ]

2017-07-11 15:50-0400, Bandan Das:
> Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx> writes:
> > 2017-07-11 14:24-0400, Bandan Das:
> >> Bandan Das <bsd@xxxxxxxxxx> writes:
> >> > If there's a triple fault, I think it's a good idea to inject it
> >> > back. Basically, there's no need to take care of damage control
> >> > that L1 is intentionally doing.
> >> >
> >> >>> +			goto fail;
> >> >>> +		kvm_mmu_unload(vcpu);
> >> >>> +		vmcs12->ept_pointer = address;
> >> >>> +		kvm_mmu_reload(vcpu);
> >> >>
> >> >> I was thinking about something like this:
> >> >>
> >> >> kvm_mmu_unload(vcpu);
> >> >> old = vmcs12->ept_pointer;
> >> >> vmcs12->ept_pointer = address;
> >> >> if (kvm_mmu_reload(vcpu)) {
> >> >> 	/* pointer invalid, restore previous state */
> >> >> 	kvm_clear_request(KVM_REQ_TRIPLE_FAULT, vcpu);
> >> >> 	vmcs12->ept_pointer = old;
> >> >> 	kvm_mmu_reload(vcpu);
> >> >> 	goto fail;
> >> >> }
> >> >>
> >> >> The you can inherit the checks from mmu_check_root().
> >> 
> >> Actually, thinking about this a bit more, I agree with you. Any fault
> >> with a vmfunc operation should end with a vmfunc vmexit, so this
> >> is a good thing to have. Thank you for this idea! :)
> >
> > SDM says
> >
> >   IF tent_EPTP is not a valid EPTP value (would cause VM entry to fail
> >   if in EPTP) THEN VMexit;
> 
> This section here:
> As noted in Section 25.5.5.2, an execution of the
> EPTP-switching VM function that causes a VM exit (as specified
> above), uses the basic exit reason 59, indicating âVMFUNCâ.
> The length of the VMFUNC instruction is saved into the
> VM-exit instruction-length field. No additional VM-exit
> information is provided.
> 
> Although, it adds (as specified above), from testing, any vmexit that
> happens as a result of the execution of the vmfunc instruction always
> has exit reason 59.
> 
> IMO, the case David pointed out comes under "as a result of the
> execution of the vmfunc instruction", so I would prefer exiting
> with reason 59.

Right, the exit reason is 59 for reasons that trigger a VM exit
(i.e. invalid EPTP value, the four below), but kvm_mmu_reload() checks
unrelated stuff.

If the EPTP value is correct, then the switch should succeed.
If the EPTP is correct, but bogus, then the guest should get
EPT_MISCONFIG VM exit on its first access (when reading the
instruction).  Source: I added

  vmcs_write64(EPT_POINTER, vmcs_read64(EPT_POINTER) | (1ULL << 40));

shortly before a VMLAUNCH on L0. :)

I think that we might be emulating this case incorrectly and throwing
triple faults when it should be VM exits in vcpu_run().

> > and no other mentions of a VM exit, so I think that the VM exit happens
> > only under these conditions:
> >
> >   â The EPT memory type (bits 2:0) must be a value supported by the
> >     processor as indicated in the IA32_VMX_EPT_VPID_CAP MSR (see
> >     Appendix A.10).
> >   â Bits 5:3 (1 less than the EPT page-walk length) must be 3, indicating
> >     an EPT page-walk length of 4; see Section 28.2.2.
> >   â Bit 6 (enable bit for accessed and dirty flags for EPT) must be 0 if
> >     bit 21 of the IA32_VMX_EPT_VPID_CAP MSR (see Appendix A.10) is read
> >     as 0, indicating that the processor does not support accessed and
> >     dirty flags for EPT.
> >   â Reserved bits 11:7 and 63:N (where N is the processorâs
> >     physical-address width) must all be 0.
> >
> > And it looks like we need parts of nested_ept_init_mmu_context() to
> > properly handle VMX_EPT_AD_ENABLE_BIT.
> 
> I completely ignored AD and the #VE sections. I will add a TODO item
> in the comment section.

AFAIK, we don't support #VE, but AD would be nice to handle from the
beginning.  (I think that caling nested_ept_init_mmu_context() as-is
isn't that bad.)