Re: [PATCH] KVM: x86: Disallow EFER.LMSLE when EferLmsleUnsupported is set in guest CPUID

[Sean Christopherson <seanjc@google.com>]

On Tue, Sep 22, 2026, Jim Mattson wrote:
> On Tue, Sep 22, 2026 at 10:53 AM Sean Christopherson <seanjc@google.com> wrote:
> >
> > On Tue, Sep 22, 2026, Jim Mattson wrote:
> > > On Tue, Sep 22, 2026 at 7:26 AM Sean Christopherson <seanjc@google.com> wrote:
> > > >
> > > > On Tue, Sep 22, 2026, Jim Mattson wrote:
> > > > > On Tue, Sep 22, 2026 at 6:57 AM Sean Christopherson <seanjc@google.com> wrote:
> > > > > > Gah, I hate the negative "feature" bits, I've gotten completely turned around
> > > > > > multiple times trying to sort through this.
> > > > >
> > > > > Me too!
> > > > >
> > > > > > IIUC, you're trying to "fix" setups where host CPUID.EFER_LMSLE_MBZ=0, but userspace
> > > > > > configured guest CPUID.EFER_LMSLE_MBZ=1?  If so, then Sashiko is right, it needs to
> > > > > > be declared with EMULATED_F().  I would even argue that technically this isn't a KVM
> > > > > > bug, because KVM won't advertise EFER_LMSLE_MBZ in KVM_GET_SUPPORTED_CPUID on such
> > > > > > hosts.
> > >
> > > > > I'm trying to make it possible to run a Rome vCPU on Milan or later.
> > > > >
> > > > > SInce this is a negative feature bit, it must be retained if the host
> > > > > reports it. That blocks running a Rome vCPU on Milan or later.
> > > > > However, a defeatured Rome (which reports LMSLE_MBZ as 1) can be run
> > > >
> > > > So, a partially defeatured Rome?  Because it sounds like the CPU reports LMSLE_MBZ=1,
> > > > but then still allows setting EFER.LMSLE=1?
> > >
> > > No; it should not allow setting EFER.LMSLE=1. That will fail on later
> > > platforms and would block live migration from Rome to later platforms.
> >
> > Oh, drat, I misread the "Rome vCPU" vs. "defeatured Rome" paragraph.  You want to
> > be able to migrate a vCPU with guest.CPUID.LMSLE_MBZ=1 between a non-defeatured,
> > vanilla Rome and Milan+ (or a defeatured Rome).  Correct?
> >
> > If so, then Sashiko is right, the flag needs to be EMULATED_F() so that KVM will
> > advertise EFER_LMSLE_MBZ to userspace on the vanilla Rome host, and also do the
> > right thing when guest.CPUID.EFER_LMSLE_MBZ=0.
> 
> No; that's wrong.

Heh, it's not "wrong" per se, just unwise :-)

> KVM_GET_SUPPORTED_CPUID should continue to report EferLmsleUnsupported=0 on
> Rome. Otherwise, we break backwards compatibility for userspace that simply
> transfers KVM_GET_SUPPORTED_CPUID to KVM_SET_CPUID[2] or userspace that just
> doesn't know anything about EferLmsleUnsupported. The "default" for Rome and
> older platforms should still be to claim support for EFER.LMSLE, even though
> that has always been a lie.

But you can't have it both ways, KVM can't enforce a bit it doesn't support.

> > If not correct, then I'm even more confused, because ignoring clear_cpuid,
> > guest_cpu_cap_has() will only ever return true if boot_cpu_has() also returns
> > true, in which case KVM *will* reject the value because EFER_LMSLE will not be
> > listed as a supported EFER bit.
> >
> >                 if (!boot_cpu_has(X86_FEATURE_EFER_LMSLE_MBZ))
> >                         kvm_caps.supported_efer_bits |= EFER_LMSLE;
> >
> > > I want KVM to abide by the guest CPUID bit and disallow setting
> > > EFER.LMSLE even on hardware that allows it.
> >
> > Right, and that requires declaring EFER_LMSLE_MBZ with EMULATED_F().
> 
> I think you and Sashiko are both wrong. PASSTHROUGH_F() is correct for this bit.
> 
> I think I need to report EferLmsleUnsupported=1 in
> KVM_GET_EMULATED_CPUID, so that userspace knows it can opt-in if it
> wants to.

Hmm, I think we should have this be a "partially" emulated feature.  KVM definitely
doesn't fully emulate LMSLE, and presumably anyone that cares about migrating
Rome vCPUs to Milan+ hosts is already advertising EFER_LMSLE_MBZ to the guest,
i.e. doesn't need identify which KVM version have the bug and which don't.

E.g. something like this?

diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
index 38450e9fca9a..67807cd95958 100644
--- a/arch/x86/kvm/cpuid.c
+++ b/arch/x86/kvm/cpuid.c
@@ -1416,6 +1416,22 @@ static int cpuid_func_emulated(struct kvm_cpuid_entry2 *entry, u32 func, u32 ind
 		if (kvm_cpu_cap_has(X86_FEATURE_RDTSCP))
 			entry->ecx = feature_bit(RDPID);
 		return 1;
+	case 0x80000008:
+		/*
+		 * Honor the guest's EFER_LMSLE_MBZ even if the underlying CPU
+		 * allows setting EFER.LMSLE, e.g. to allowing migrating a vCPU
+		 * between hosts with and without EFER.LMSLE support.  To avoid
+		 * breaking existing setups that reflect KVM's supported CPUID
+		 * into the guest, KVM doesn't advertise EFER_LMSLE_MBZ unless
+		 * it's supported by hardware, i.e. unless KVM *can't* support
+		 * EFER.LMSLE=1.
+		 */
+		if (include_partially_emulated &&
+		    !kvm_cpu_cap_has(X86_FEATURE_EFER_LMSLE_MBZ)) {
+			entry->ebx |= feature_bit(EFER_LMSLE_MBZ);
+			return 1;
+		}
+		return 0;
 	default:
 		return 0;
 	}