[PATCH v2] KVM: selftests: Add guest-side test for AMD HWCR.McStatusWrEn
From: Costas Argyris
Date: Fri May 08 2026 - 12:15:38 EST
Add a guest-side test to msrs_test that verifies HWCR.McStatusWrEn
(bit 18) correctly gates non-zero writes to MCi_STATUS MSRs. With
the bit set, the write must succeed and read back unchanged. With
the bit clear, the write must raise #GP and leave the register
unmodified.
This exercises can_set_mci_status(), which is only reachable via the
guest WRMSR path (host_initiated=false).
Signed-off-by: Costas Argyris <costas.argyris@xxxxxxx>
---
v1 -> v2:
- Fold test into msrs_test.c
- Reuse existing helpers for the guest code
- Drop unnecessary while loop
I tried fitting into the existing __msrs[] table but it doesn't seem to
map well — the proposed test has one MSR (HWCR) gating writes to another
(MCi_STATUS), which doesn't seem to fit the per-MSR CPUID-feature model.
Tried a self-contained function alongside test_msrs() instead.
tools/testing/selftests/kvm/x86/msrs_test.c | 59 +++++++++++++++++++++
1 file changed, 59 insertions(+)
diff --git a/tools/testing/selftests/kvm/x86/msrs_test.c b/tools/testing/selftests/kvm/x86/msrs_test.c
index f7e39bf887ad..348811559831 100644
--- a/tools/testing/selftests/kvm/x86/msrs_test.c
+++ b/tools/testing/selftests/kvm/x86/msrs_test.c
@@ -476,6 +476,62 @@ static void test_msrs(void)
kvm_vm_free(vm);
}
+/*
+ * AMD-specific: test that HWCR.McStatusWrEn (bit 18) gates guest writes to
+ * MCi_STATUS MSRs. With the bit set, a non-zero write to MC0_STATUS must
+ * succeed and read back unchanged. With the bit clear, the write must take
+ * a #GP.
+ *
+ * This exercises can_set_mci_status(), which is only reachable via the guest
+ * WRMSR path (host_initiated=false).
+ */
+static void guest_test_mcstatus_wren(void)
+{
+ const u64 val = 1;
+ u8 vec;
+
+ /* McStatusWrEn=1: non-zero write to MCi_STATUS must succeed. */
+ wrmsr(MSR_K7_HWCR, BIT_ULL(18));
+ __wrmsr(MSR_IA32_MC0_STATUS, val);
+
+ /* Clear before disabling the write-enable bit. */
+ __wrmsr(MSR_IA32_MC0_STATUS, 0);
+
+ /* McStatusWrEn=0: non-zero write to MCi_STATUS must #GP. */
+ wrmsr(MSR_K7_HWCR, 0);
+ vec = wrmsr_safe(MSR_IA32_MC0_STATUS, val);
+ __GUEST_ASSERT(vec == GP_VECTOR,
+ "Wanted #GP on WRMSR(0x%x, 0x%lx), got %s",
+ MSR_IA32_MC0_STATUS, val, ex_str(vec));
+
+ /* Confirm the failed write left the register at zero. */
+ __rdmsr(MSR_IA32_MC0_STATUS, 0);
+
+ GUEST_DONE();
+}
+
+static void test_mcstatus_wren(void)
+{
+ struct kvm_vcpu *vcpu;
+ struct kvm_vm *vm;
+ struct ucall uc;
+
+ vm = vm_create_with_one_vcpu(&vcpu, guest_test_mcstatus_wren);
+ vcpu_run(vcpu);
+ TEST_ASSERT_KVM_EXIT_REASON(vcpu, KVM_EXIT_IO);
+
+ switch (get_ucall(vcpu, &uc)) {
+ case UCALL_DONE:
+ break;
+ case UCALL_ABORT:
+ REPORT_GUEST_ASSERT(uc);
+ default:
+ TEST_FAIL("Unexpected ucall %lu", uc.cmd);
+ }
+
+ kvm_vm_free(vm);
+}
+
int main(void)
{
has_one_reg = kvm_has_cap(KVM_CAP_ONE_REG);
@@ -486,4 +542,7 @@ int main(void)
use_one_reg = true;
test_msrs();
}
+
+ if (host_cpu_is_amd_compatible)
+ test_mcstatus_wren();
}
--
2.43.0