KVM: nVMX: Fix loss of L2's NMI blocking state
authorWanpeng Li <kernellwp@gmail.com>
Tue, 25 Jul 2017 10:40:46 +0000 (03:40 -0700)
committerPaolo Bonzini <pbonzini@redhat.com>
Wed, 26 Jul 2017 16:57:46 +0000 (18:57 +0200)
Run kvm-unit-tests/eventinj.flat in L1 w/ ept=0 on both L0 and L1:

Before NMI IRET test
Sending NMI to self
NMI isr running stack 0x461000
Sending nested NMI to self
After nested NMI to self
Nested NMI isr running rip=40038e
After iret
After NMI to self

Commit 4c4a6f790ee862 (KVM: nVMX: track NMI blocking state separately
for each VMCS) tracks NMI blocking state separately for vmcs01 and
vmcs02. However it is not enough:

 - The L2 (kvm-unit-tests/eventinj.flat) generates NMI that will fault
   on IRET, so the L2 can generate #PF which can be intercepted by L0.
 - L0 walks L1's guest page table and sees the mapping is invalid, it
   resumes the L1 guest and injects the #PF into L1.  At this point the
   vmcs02 has nmi_known_unmasked=true.
 - L1 sets set bit 3 (blocking by NMI) in the interruptibility-state field
   of vmcs12 (and fixes the shadow page table) before resuming L2 guest.
 - L1 executes VMRESUME to resume L2, causing a vmexit to L0
 - during VMRESUME emulation, prepare_vmcs02 sets bit 3 in the
   interruptibility-state field of vmcs02, but nmi_known_unmasked is
   still true.
 - L2 immediately exits to L0 with another page fault, because L0 still has
   not updated the NGVA->HPA page tables.  However, nmi_known_unmasked is
   true so vmx_recover_nmi_blocking does not do anything.

The fix is to update nmi_known_unmasked when preparing vmcs02 from vmcs12.

Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

index 497cf64794fd96ba64ef400a2e2cf21974a097a9..39a6222bf9687fcc47dea3ee564eaa92704026e9 100644 (file)
@@ -10042,6 +10042,8 @@ static int prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
+               vmx->loaded_vmcs->nmi_known_unmasked =
+                       !(vmcs12->guest_interruptibility_info & GUEST_INTR_STATE_NMI);
        } else {
                vmcs_write32(VM_ENTRY_INTR_INFO_FIELD, 0);