Re: [PATCH 1/2] KVM: x86: Add emulation support for #GP triggered by VM instructions

[Maxim Levitsky]

On Tue, 2021-01-12 at 15:00 -0500, Bandan Das wrote:
> Sean Christopherson <seanjc@xxxxxxxxxx> writes:
> ...
> > > -	if ((emulation_type & EMULTYPE_VMWARE_GP) &&
> > > -	    !is_vmware_backdoor_opcode(ctxt)) {
> > > -		kvm_queue_exception_e(vcpu, GP_VECTOR, 0);
> > > -		return 1;
> > > +	if (emulation_type & EMULTYPE_PARAVIRT_GP) {
> > > +		vminstr = is_vm_instr_opcode(ctxt);
> > > +		if (!vminstr && !is_vmware_backdoor_opcode(ctxt)) {
> > > +			kvm_queue_exception_e(vcpu, GP_VECTOR, 0);
> > > +			return 1;
> > > +		}
> > > +		if (vminstr)
> > > +			return vminstr;
> > 
> > I'm pretty sure this doesn't correctly handle a VM-instr in L2 that hits a bad
> > L0 GPA and that L1 wants to intercept.  The intercept bitmap isn't checked until
> > x86_emulate_insn(), and the vm*_interception() helpers expect nested VM-Exits to
> > be handled further up the stack.

Actually IMHO this exactly what we want. We want L0 to always intercept
these #GPs, and hide them from the guest.

What we do need to do (and I prepared and attached a patch for that, is that if we run
a guest, we want to inject corresponding vmexit (like SVM_EXIT_VMRUN)
instead of emulating the instruction.
The attached patch does this, and it made my kvm unit test pass,
even if the test was run in a VM (with an unpatched kernel).

This together with setting that X86_FEATURE_SVME_ADDR_CHK bit for
the guest will allow us to hide that errata completely from the guest
which is a very good thing.
(for example for guests that we can't modify)


Best regards,
	Maxim Levitsky



> > 
> So, the condition is that L2 executes a vmload and #GPs on a reserved address, jumps to L0 - L0 doesn't
> check if L1 has asked for the instruction to be intercepted and goes on with emulating
> vmload and returning back to L2 ?



> 
> > >  	}
> > >  
> > >  	/*
> > > -- 
> > > 2.27.0
> > > 

commit 28ab89aaa11380306bafbf49265222f2a2da71da
Author: Maxim Levitsky <mlevitsk@xxxxxxxxxx>
Date:   Thu Jan 14 10:53:25 2021 +0200

    kvm: x86: fix that errata for nested guests

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index c31e005252d69..9cfa5946fac69 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -2027,6 +2027,26 @@ static int svm_emulate_vm_instr(struct kvm_vcpu *vcpu, u8 modrm)
 {
 	struct vcpu_svm *svm = to_svm(vcpu);
 
+	if (is_guest_mode(vcpu)) {
+		switch (modrm) {
+		case 0xd8: /* VMRUN */
+			svm->vmcb->control.exit_code = SVM_EXIT_VMRUN;
+			break;
+		case 0xda: /* VMLOAD */
+			svm->vmcb->control.exit_code = SVM_EXIT_VMLOAD;
+			break;
+		case 0xdb: /* VMSAVE */
+			svm->vmcb->control.exit_code = SVM_EXIT_VMLOAD;
+			break;
+		default:
+			goto inject_exception;
+		}
+
+		svm->vmcb->control.exit_info_1 = 0;
+		svm->vmcb->control.exit_info_2 = 0;
+		return nested_svm_vmexit(svm);
+	}
+
 	switch (modrm) {
 	case 0xd8: /* VMRUN */
 		return vmrun_interception(svm);
@@ -2035,6 +2055,7 @@ static int svm_emulate_vm_instr(struct kvm_vcpu *vcpu, u8 modrm)
 	case 0xdb: /* VMSAVE */
 		return vmsave_interception(svm);
 	default:
+inject_exception:
 		/* inject a #GP for all other cases */
 		kvm_queue_exception_e(vcpu, GP_VECTOR, 0);
 		return 1;