KVM: nVMX: Fix mmu context after VMLAUNCH/VMRESUME failure
authorWanpeng Li <wanpeng.li@hotmail.com>
Mon, 6 Nov 2017 00:54:49 +0000 (16:54 -0800)
committerPaolo Bonzini <pbonzini@redhat.com>
Fri, 17 Nov 2017 12:20:14 +0000 (13:20 +0100)
Commit 4f350c6dbcb (kvm: nVMX: Handle deferred early VMLAUNCH/VMRESUME failure
properly) can result in L1(run kvm-unit-tests/run_tests.sh vmx_controls in L1)
null pointer deference and also L0 calltrace when EPT=0 on both L0 and L1.

In L1:

BUG: unable to handle kernel paging request at ffffffffc015bf8f
 IP: vmx_vcpu_run+0x202/0x510 [kvm_intel]
 PGD 146e13067 P4D 146e13067 PUD 146e15067 PMD 3d2686067 PTE 3d4af9161
 Oops: 0003 [#1] PREEMPT SMP
 CPU: 2 PID: 1798 Comm: qemu-system-x86 Not tainted 4.14.0-rc4+ #6
 RIP: 0010:vmx_vcpu_run+0x202/0x510 [kvm_intel]
 Call Trace:
 WARNING: kernel stack frame pointer at ffffb86f4988bc18 in qemu-system-x86:1798 has bad value 0000000000000002

In L0:

-----------[ cut here ]------------
 WARNING: CPU: 6 PID: 4460 at /home/kernel/linux/arch/x86/kvm//vmx.c:9845 vmx_inject_page_fault_nested+0x130/0x140 [kvm_intel]
 CPU: 6 PID: 4460 Comm: qemu-system-x86 Tainted: G           OE   4.14.0-rc7+ #25
 RIP: 0010:vmx_inject_page_fault_nested+0x130/0x140 [kvm_intel]
 Call Trace:
  paging64_page_fault+0x500/0xde0 [kvm]
  ? paging32_gva_to_gpa_nested+0x120/0x120 [kvm]
  ? nonpaging_page_fault+0x3b0/0x3b0 [kvm]
  ? __asan_storeN+0x12/0x20
  ? paging64_gva_to_gpa+0xb0/0x120 [kvm]
  ? paging64_walk_addr_generic+0x11a0/0x11a0 [kvm]
  ? lock_acquire+0x2c0/0x2c0
  ? vmx_read_guest_seg_ar+0x97/0x100 [kvm_intel]
  ? vmx_get_segment+0x2a6/0x310 [kvm_intel]
  ? sched_clock+0x1f/0x30
  ? check_chain_key+0x137/0x1e0
  ? __lock_acquire+0x83c/0x2420
  ? kvm_multiple_exception+0xf2/0x220 [kvm]
  ? debug_check_no_locks_freed+0x240/0x240
  ? debug_smp_processor_id+0x17/0x20
  ? __lock_is_held+0x9e/0x100
  kvm_mmu_page_fault+0x90/0x180 [kvm]
  kvm_handle_page_fault+0x15c/0x310 [kvm]
  ? __lock_is_held+0x9e/0x100
  handle_exception+0x3c7/0x4d0 [kvm_intel]
  vmx_handle_exit+0x103/0x1010 [kvm_intel]
  ? kvm_arch_vcpu_ioctl_run+0x1628/0x2e20 [kvm]

The commit avoids to load host state of vmcs12 as vmcs01's guest state
since vmcs12 is not modified (except for the VM-instruction error field)
if the checking of vmcs control area fails. However, the mmu context is
switched to nested mmu in prepare_vmcs02() and it will not be reloaded
since load_vmcs12_host_state() is skipped when nested VMLAUNCH/VMRESUME
fails. This patch fixes it by reloading mmu context when nested
VMLAUNCH/VMRESUME fails.

Reviewed-by: Jim Mattson <jmattson@google.com>
Reviewed-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Jim Mattson <jmattson@google.com>
Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/kvm/vmx.c

index 707aaa954b3ddfab03982fb99bd79f9dfe97e9d9..10474d26a00020e4bc2518106d4e87205ef39134 100644 (file)
@@ -11330,6 +11330,24 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
        kvm_clear_interrupt_queue(vcpu);
 }
 
+static void load_vmcs12_mmu_host_state(struct kvm_vcpu *vcpu,
+                       struct vmcs12 *vmcs12)
+{
+       u32 entry_failure_code;
+
+       nested_ept_uninit_mmu_context(vcpu);
+
+       /*
+        * Only PDPTE load can fail as the value of cr3 was checked on entry and
+        * couldn't have changed.
+        */
+       if (nested_vmx_load_cr3(vcpu, vmcs12->host_cr3, false, &entry_failure_code))
+               nested_vmx_abort(vcpu, VMX_ABORT_LOAD_HOST_PDPTE_FAIL);
+
+       if (!enable_ept)
+               vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
+}
+
 /*
  * A part of what we need to when the nested L2 guest exits and we want to
  * run its L1 parent, is to reset L1's guest state to the host state specified
@@ -11343,7 +11361,6 @@ static void load_vmcs12_host_state(struct kvm_vcpu *vcpu,
                                   struct vmcs12 *vmcs12)
 {
        struct kvm_segment seg;
-       u32 entry_failure_code;
 
        if (vmcs12->vm_exit_controls & VM_EXIT_LOAD_IA32_EFER)
                vcpu->arch.efer = vmcs12->host_ia32_efer;
@@ -11370,17 +11387,7 @@ static void load_vmcs12_host_state(struct kvm_vcpu *vcpu,
        vcpu->arch.cr4_guest_owned_bits = ~vmcs_readl(CR4_GUEST_HOST_MASK);
        vmx_set_cr4(vcpu, vmcs12->host_cr4);
 
-       nested_ept_uninit_mmu_context(vcpu);
-
-       /*
-        * Only PDPTE load can fail as the value of cr3 was checked on entry and
-        * couldn't have changed.
-        */
-       if (nested_vmx_load_cr3(vcpu, vmcs12->host_cr3, false, &entry_failure_code))
-               nested_vmx_abort(vcpu, VMX_ABORT_LOAD_HOST_PDPTE_FAIL);
-
-       if (!enable_ept)
-               vcpu->arch.walk_mmu->inject_page_fault = kvm_inject_page_fault;
+       load_vmcs12_mmu_host_state(vcpu, vmcs12);
 
        if (enable_vpid) {
                /*
@@ -11610,6 +11617,9 @@ static void nested_vmx_vmexit(struct kvm_vcpu *vcpu, u32 exit_reason,
         * accordingly.
         */
        nested_vmx_failValid(vcpu, VMXERR_ENTRY_INVALID_CONTROL_FIELD);
+
+       load_vmcs12_mmu_host_state(vcpu, vmcs12);
+
        /*
         * The emulated instruction was already skipped in
         * nested_vmx_run, but the updated RIP was never