KVM: nVMX: initialize PML fields in vmcs02
authorLadi Prosek <lprosek@redhat.com>
Tue, 4 Apr 2017 12:18:53 +0000 (14:18 +0200)
committerRadim Krčmář <rkrcmar@redhat.com>
Tue, 4 Apr 2017 14:24:43 +0000 (16:24 +0200)
L2 was running with uninitialized PML fields which led to incomplete
dirty bitmap logging. This manifested as all kinds of subtle erratic
behavior of the nested guest.

Fixes: 843e4330573c ("KVM: VMX: Add PML support in VMX")
Signed-off-by: Ladi Prosek <lprosek@redhat.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/kvm/vmx.c

index 6051832910698f6aa403b93b21f4ac82c8409aaf..259e9b28ccf8e7ff2434d6911ffcfbfc4ef89f42 100644 (file)
@@ -10270,6 +10270,18 @@ static int prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
 
        }
 
+       if (enable_pml) {
+               /*
+                * Conceptually we want to copy the PML address and index from
+                * vmcs01 here, and then back to vmcs01 on nested vmexit. But,
+                * since we always flush the log on each vmexit, this happens
+                * to be equivalent to simply resetting the fields in vmcs02.
+                */
+               ASSERT(vmx->pml_pg);
+               vmcs_write64(PML_ADDRESS, page_to_phys(vmx->pml_pg));
+               vmcs_write16(GUEST_PML_INDEX, PML_ENTITY_NUM - 1);
+       }
+
        if (nested_cpu_has_ept(vmcs12)) {
                kvm_mmu_unload(vcpu);
                nested_ept_init_mmu_context(vcpu);