[RFC PATCH v2 28/31] KVM: SVM: Keep TPR[3:0] across exits without a guest CR8 write

From: David Woodhouse

Date: Fri Nov 13 2026 - 15:09:18 EST


From: David Woodhouse <dwmw@xxxxxxxxxxxx>

V_TPR holds only TPR[7:4]. While KVM does not intercept CR8 writes,
sync_cr8_to_lapic() pushes V_TPR back into the local APIC after every
VMRUN, whether or not the guest touched CR8. A guest that sets the TPR
through the APIC page or the x2APIC MSR therefore loses TPR[3:0] at the
next exit: it writes 0xff, reads back 0xf0, and interrupts in the
priority sub-class it meant to mask are delivered. VMX keeps all eight
TPR bits in the virtual-APIC page and does not have this problem.

Hardware changes V_TPR only on a guest CR8 write. Record the value
sync_lapic_to_cr8() loads on entry and push V_TPR back into the APIC
only when it differs, which only a guest CR8 write can make it do. A
CR8 write whose value matches the loaded TPR[7:4] is indistinguishable
from no write and now preserves TPR[3:0], where architecturally it
would clear it; the previous behaviour cleared TPR[3:0] on every exit
regardless.

Found while implementing Hyper-V VSM on KVM planes.

Fixes: d7bf8221a303 ("KVM: SVM: sync V_TPR with LAPIC.TPR if CR8 write intercept is disabled")
Signed-off-by: David Woodhouse <dwmw@xxxxxxxxxxxx>
---
arch/x86/kvm/svm/svm.c | 11 ++++++++++-
arch/x86/kvm/svm/svm.h | 13 +++++++++++++
2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index cf049367523c..9bfcc7684504 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -4622,7 +4622,15 @@ static inline void sync_cr8_to_lapic(struct kvm_vcpu *vcpu)

if (!svm_is_intercept(svm, INTERCEPT_CR8_WRITE)) {
int cr8 = svm->vmcb->control.int_ctl & V_TPR_MASK;
- kvm_set_cr8(vcpu, cr8);
+
+ /*
+ * Hardware changes V_TPR only on a guest CR8 write. Pushing
+ * an unchanged value into the APIC would zero TPR[3:0] (CR8
+ * maps TPR[7:4] alone), wiping the low nibble of a TASKPRI
+ * value the guest set through the APIC page or MSR.
+ */
+ if (cr8 != svm->v_tpr_entry)
+ kvm_set_cr8(vcpu, cr8);
}
}

@@ -4637,6 +4645,7 @@ static inline void sync_lapic_to_cr8(struct kvm_vcpu *vcpu)
cr8 = kvm_get_cr8(vcpu);
svm->vmcb->control.int_ctl &= ~V_TPR_MASK;
svm->vmcb->control.int_ctl |= cr8 & V_TPR_MASK;
+ svm->v_tpr_entry = cr8 & V_TPR_MASK;
}

static void svm_complete_soft_interrupt(struct kvm_vcpu *vcpu, u8 vector,
diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
index e7fced399e05..3dfa8b20152b 100644
--- a/arch/x86/kvm/svm/svm.h
+++ b/arch/x86/kvm/svm/svm.h
@@ -323,6 +323,19 @@ struct vcpu_svm {

void *msrpm;

+ /*
+ * V_TPR value written by sync_lapic_to_cr8() on the last entry.
+ * sync_cr8_to_lapic() pushes V_TPR back into the APIC only when it
+ * differs: V_TPR holds TPR[7:4] only, so an unconditional back-sync
+ * would zero TPR[3:0] after every exit, losing the low nibble of an
+ * APIC TASKPRI write. A differing V_TPR implies a guest CR8 write,
+ * but not the converse: a CR8 write of the same value is
+ * indistinguishable from no write, so TPR[3:0] survives it where
+ * the architectural write (and the intercepted path through
+ * kvm_lapic_set_tpr()) would clear it.
+ */
+ u8 v_tpr_entry;
+
ulong nmi_iret_rip;

struct svm_nested_state nested;
--
2.50.1