]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/commitdiff
KVM: nVMX: fix msr bitmaps to prevent L2 from accessing L0 x2APIC
authorRadim Krčmář <rkrcmar@redhat.com>
Mon, 8 Aug 2016 18:16:22 +0000 (20:16 +0200)
committerPaolo Bonzini <pbonzini@redhat.com>
Thu, 18 Aug 2016 10:19:07 +0000 (12:19 +0200)
msr bitmap can be used to avoid a VM exit (interception) on guest MSR
accesses.  In some configurations of VMX controls, the guest can even
directly access host's x2APIC MSRs.  See SDM 29.5 VIRTUALIZING MSR-BASED
APIC ACCESSES.

L2 could read all L0's x2APIC MSRs and write TPR, EOI, and SELF_IPI.
To do so, L1 would first trick KVM to disable all possible interceptions
by enabling APICv features and then would turn those features off;
nested_vmx_merge_msr_bitmap() only disabled interceptions, so VMX would
not intercept previously enabled MSRs even though they were not safe
with the new configuration.

Correctly re-enabling interceptions is not enough as a second bug would
still allow L1+L2 to access host's MSRs: msr bitmap was shared for all
VMCSs, so L1 could trigger a race to get the desired combination of msr
bitmap and VMX controls.

This fix allocates a msr bitmap for every L1 VCPU, allows only safe
x2APIC MSRs from L1's msr bitmap, and disables msr bitmaps if they would
have to intercept everything anyway.

Fixes: 3af18d9c5fe9 ("KVM: nVMX: Prepare for using hardware MSR bitmap")
Reported-by: Jim Mattson <jmattson@google.com>
Suggested-by: Wincy Van <fanwenyi0529@gmail.com>
Reviewed-by: Wanpeng Li <wanpeng.li@hotmail.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/kvm/vmx.c

index a45d8580f91e7e8459666d7c1f0076ab1803b821..c66ac2c70d228bba570811285d6fdb40281af292 100644 (file)
@@ -435,6 +435,8 @@ struct nested_vmx {
        bool pi_pending;
        u16 posted_intr_nv;
 
+       unsigned long *msr_bitmap;
+
        struct hrtimer preemption_timer;
        bool preemption_timer_expired;
 
@@ -924,7 +926,6 @@ static unsigned long *vmx_msr_bitmap_legacy;
 static unsigned long *vmx_msr_bitmap_longmode;
 static unsigned long *vmx_msr_bitmap_legacy_x2apic;
 static unsigned long *vmx_msr_bitmap_longmode_x2apic;
-static unsigned long *vmx_msr_bitmap_nested;
 static unsigned long *vmx_vmread_bitmap;
 static unsigned long *vmx_vmwrite_bitmap;
 
@@ -2508,7 +2509,7 @@ static void vmx_set_msr_bitmap(struct kvm_vcpu *vcpu)
        unsigned long *msr_bitmap;
 
        if (is_guest_mode(vcpu))
-               msr_bitmap = vmx_msr_bitmap_nested;
+               msr_bitmap = to_vmx(vcpu)->nested.msr_bitmap;
        else if (cpu_has_secondary_exec_ctrls() &&
                 (vmcs_read32(SECONDARY_VM_EXEC_CONTROL) &
                  SECONDARY_EXEC_VIRTUALIZE_X2APIC_MODE)) {
@@ -6363,13 +6364,6 @@ static __init int hardware_setup(void)
        if (!vmx_msr_bitmap_longmode_x2apic)
                goto out4;
 
-       if (nested) {
-               vmx_msr_bitmap_nested =
-                       (unsigned long *)__get_free_page(GFP_KERNEL);
-               if (!vmx_msr_bitmap_nested)
-                       goto out5;
-       }
-
        vmx_vmread_bitmap = (unsigned long *)__get_free_page(GFP_KERNEL);
        if (!vmx_vmread_bitmap)
                goto out6;
@@ -6392,8 +6386,6 @@ static __init int hardware_setup(void)
 
        memset(vmx_msr_bitmap_legacy, 0xff, PAGE_SIZE);
        memset(vmx_msr_bitmap_longmode, 0xff, PAGE_SIZE);
-       if (nested)
-               memset(vmx_msr_bitmap_nested, 0xff, PAGE_SIZE);
 
        if (setup_vmcs_config(&vmcs_config) < 0) {
                r = -EIO;
@@ -6529,9 +6521,6 @@ out8:
 out7:
        free_page((unsigned long)vmx_vmread_bitmap);
 out6:
-       if (nested)
-               free_page((unsigned long)vmx_msr_bitmap_nested);
-out5:
        free_page((unsigned long)vmx_msr_bitmap_longmode_x2apic);
 out4:
        free_page((unsigned long)vmx_msr_bitmap_longmode);
@@ -6557,8 +6546,6 @@ static __exit void hardware_unsetup(void)
        free_page((unsigned long)vmx_io_bitmap_a);
        free_page((unsigned long)vmx_vmwrite_bitmap);
        free_page((unsigned long)vmx_vmread_bitmap);
-       if (nested)
-               free_page((unsigned long)vmx_msr_bitmap_nested);
 
        free_kvm_area();
 }
@@ -6995,16 +6982,21 @@ static int handle_vmon(struct kvm_vcpu *vcpu)
                return 1;
        }
 
+       if (cpu_has_vmx_msr_bitmap()) {
+               vmx->nested.msr_bitmap =
+                               (unsigned long *)__get_free_page(GFP_KERNEL);
+               if (!vmx->nested.msr_bitmap)
+                       goto out_msr_bitmap;
+       }
+
        vmx->nested.cached_vmcs12 = kmalloc(VMCS12_SIZE, GFP_KERNEL);
        if (!vmx->nested.cached_vmcs12)
-               return -ENOMEM;
+               goto out_cached_vmcs12;
 
        if (enable_shadow_vmcs) {
                shadow_vmcs = alloc_vmcs();
-               if (!shadow_vmcs) {
-                       kfree(vmx->nested.cached_vmcs12);
-                       return -ENOMEM;
-               }
+               if (!shadow_vmcs)
+                       goto out_shadow_vmcs;
                /* mark vmcs as shadow */
                shadow_vmcs->revision_id |= (1u << 31);
                /* init shadow vmcs */
@@ -7024,6 +7016,15 @@ static int handle_vmon(struct kvm_vcpu *vcpu)
        skip_emulated_instruction(vcpu);
        nested_vmx_succeed(vcpu);
        return 1;
+
+out_shadow_vmcs:
+       kfree(vmx->nested.cached_vmcs12);
+
+out_cached_vmcs12:
+       free_page((unsigned long)vmx->nested.msr_bitmap);
+
+out_msr_bitmap:
+       return -ENOMEM;
 }
 
 /*
@@ -7098,6 +7099,10 @@ static void free_nested(struct vcpu_vmx *vmx)
        vmx->nested.vmxon = false;
        free_vpid(vmx->nested.vpid02);
        nested_release_vmcs12(vmx);
+       if (vmx->nested.msr_bitmap) {
+               free_page((unsigned long)vmx->nested.msr_bitmap);
+               vmx->nested.msr_bitmap = NULL;
+       }
        if (enable_shadow_vmcs)
                free_vmcs(vmx->nested.current_shadow_vmcs);
        kfree(vmx->nested.cached_vmcs12);
@@ -9472,8 +9477,10 @@ static inline bool nested_vmx_merge_msr_bitmap(struct kvm_vcpu *vcpu,
 {
        int msr;
        struct page *page;
-       unsigned long *msr_bitmap;
+       unsigned long *msr_bitmap_l1;
+       unsigned long *msr_bitmap_l0 = to_vmx(vcpu)->nested.msr_bitmap;
 
+       /* This shortcut is ok because we support only x2APIC MSRs so far. */
        if (!nested_cpu_has_virt_x2apic_mode(vmcs12))
                return false;
 
@@ -9482,63 +9489,37 @@ static inline bool nested_vmx_merge_msr_bitmap(struct kvm_vcpu *vcpu,
                WARN_ON(1);
                return false;
        }
-       msr_bitmap = (unsigned long *)kmap(page);
-       if (!msr_bitmap) {
+       msr_bitmap_l1 = (unsigned long *)kmap(page);
+       if (!msr_bitmap_l1) {
                nested_release_page_clean(page);
                WARN_ON(1);
                return false;
        }
 
+       memset(msr_bitmap_l0, 0xff, PAGE_SIZE);
+
        if (nested_cpu_has_virt_x2apic_mode(vmcs12)) {
                if (nested_cpu_has_apic_reg_virt(vmcs12))
                        for (msr = 0x800; msr <= 0x8ff; msr++)
                                nested_vmx_disable_intercept_for_msr(
-                                       msr_bitmap,
-                                       vmx_msr_bitmap_nested,
+                                       msr_bitmap_l1, msr_bitmap_l0,
                                        msr, MSR_TYPE_R);
-               /* TPR is allowed */
-               nested_vmx_disable_intercept_for_msr(msr_bitmap,
-                               vmx_msr_bitmap_nested,
+
+               nested_vmx_disable_intercept_for_msr(
+                               msr_bitmap_l1, msr_bitmap_l0,
                                APIC_BASE_MSR + (APIC_TASKPRI >> 4),
                                MSR_TYPE_R | MSR_TYPE_W);
+
                if (nested_cpu_has_vid(vmcs12)) {
-                       /* EOI and self-IPI are allowed */
                        nested_vmx_disable_intercept_for_msr(
-                               msr_bitmap,
-                               vmx_msr_bitmap_nested,
+                               msr_bitmap_l1, msr_bitmap_l0,
                                APIC_BASE_MSR + (APIC_EOI >> 4),
                                MSR_TYPE_W);
                        nested_vmx_disable_intercept_for_msr(
-                               msr_bitmap,
-                               vmx_msr_bitmap_nested,
+                               msr_bitmap_l1, msr_bitmap_l0,
                                APIC_BASE_MSR + (APIC_SELF_IPI >> 4),
                                MSR_TYPE_W);
                }
-       } else {
-               /*
-                * Enable reading intercept of all the x2apic
-                * MSRs. We should not rely on vmcs12 to do any
-                * optimizations here, it may have been modified
-                * by L1.
-                */
-               for (msr = 0x800; msr <= 0x8ff; msr++)
-                       __vmx_enable_intercept_for_msr(
-                               vmx_msr_bitmap_nested,
-                               msr,
-                               MSR_TYPE_R);
-
-               __vmx_enable_intercept_for_msr(
-                               vmx_msr_bitmap_nested,
-                               APIC_BASE_MSR + (APIC_TASKPRI >> 4),
-                               MSR_TYPE_W);
-               __vmx_enable_intercept_for_msr(
-                               vmx_msr_bitmap_nested,
-                               APIC_BASE_MSR + (APIC_EOI >> 4),
-                               MSR_TYPE_W);
-               __vmx_enable_intercept_for_msr(
-                               vmx_msr_bitmap_nested,
-                               APIC_BASE_MSR + (APIC_SELF_IPI >> 4),
-                               MSR_TYPE_W);
        }
        kunmap(page);
        nested_release_page_clean(page);
@@ -9957,10 +9938,10 @@ static void prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
        }
 
        if (cpu_has_vmx_msr_bitmap() &&
-           exec_control & CPU_BASED_USE_MSR_BITMAPS) {
-               nested_vmx_merge_msr_bitmap(vcpu, vmcs12);
-               /* MSR_BITMAP will be set by following vmx_set_efer. */
-       else
+           exec_control & CPU_BASED_USE_MSR_BITMAPS &&
+           nested_vmx_merge_msr_bitmap(vcpu, vmcs12))
+               /* MSR_BITMAP will be set by following vmx_set_efer. */
+       else
                exec_control &= ~CPU_BASED_USE_MSR_BITMAPS;
 
        /*