kvm: vmx: ensure VMCS is current while enabling PML
authorPeter Feiner <pfeiner@google.com>
Thu, 7 Jul 2016 21:49:58 +0000 (14:49 -0700)
committerPaolo Bonzini <pbonzini@redhat.com>
Thu, 14 Jul 2016 17:11:19 +0000 (19:11 +0200)
Between loading the new VMCS and enabling PML, the CPU was unpinned.
If the vCPU thread were migrated to another CPU in the interim (e.g.,
due to preemption or sleeping alloc_page), then the VMWRITEs to enable
PML would target the wrong VMCS -- or no VMCS at all:

  [ 2087.266950] vmwrite error: reg 200e value 3fe1d52000 (err -506126336)
  [ 2087.267062] vmwrite error: reg 812 value 1ff (err 511)
  [ 2087.267125] vmwrite error: reg 401e value 12229c00 (err 304258048)

This patch ensures that the VMCS remains current while enabling PML by
doing the VMWRITEs while the CPU is pinned. Allocation of the PML buffer
is hoisted out of the critical section.

Signed-off-by: Peter Feiner <pfeiner@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/vmx.c

index 64a79f2..e34965b 100644 (file)
@@ -4979,6 +4979,12 @@ static int vmx_vcpu_setup(struct vcpu_vmx *vmx)
        if (vmx_xsaves_supported())
                vmcs_write64(XSS_EXIT_BITMAP, VMX_XSS_EXIT_BITMAP);
 
+       if (enable_pml) {
+               ASSERT(vmx->pml_pg);
+               vmcs_write64(PML_ADDRESS, page_to_phys(vmx->pml_pg));
+               vmcs_write16(GUEST_PML_INDEX, PML_ENTITY_NUM - 1);
+       }
+
        return 0;
 }
 
@@ -7937,22 +7943,6 @@ static void vmx_get_exit_info(struct kvm_vcpu *vcpu, u64 *info1, u64 *info2)
        *info2 = vmcs_read32(VM_EXIT_INTR_INFO);
 }
 
-static int vmx_create_pml_buffer(struct vcpu_vmx *vmx)
-{
-       struct page *pml_pg;
-
-       pml_pg = alloc_page(GFP_KERNEL | __GFP_ZERO);
-       if (!pml_pg)
-               return -ENOMEM;
-
-       vmx->pml_pg = pml_pg;
-
-       vmcs_write64(PML_ADDRESS, page_to_phys(vmx->pml_pg));
-       vmcs_write16(GUEST_PML_INDEX, PML_ENTITY_NUM - 1);
-
-       return 0;
-}
-
 static void vmx_destroy_pml_buffer(struct vcpu_vmx *vmx)
 {
        if (vmx->pml_pg) {
@@ -8885,14 +8875,26 @@ static struct kvm_vcpu *vmx_create_vcpu(struct kvm *kvm, unsigned int id)
        if (err)
                goto free_vcpu;
 
+       err = -ENOMEM;
+
+       /*
+        * If PML is turned on, failure on enabling PML just results in failure
+        * of creating the vcpu, therefore we can simplify PML logic (by
+        * avoiding dealing with cases, such as enabling PML partially on vcpus
+        * for the guest, etc.
+        */
+       if (enable_pml) {
+               vmx->pml_pg = alloc_page(GFP_KERNEL | __GFP_ZERO);
+               if (!vmx->pml_pg)
+                       goto uninit_vcpu;
+       }
+
        vmx->guest_msrs = kmalloc(PAGE_SIZE, GFP_KERNEL);
        BUILD_BUG_ON(ARRAY_SIZE(vmx_msr_index) * sizeof(vmx->guest_msrs[0])
                     > PAGE_SIZE);
 
-       err = -ENOMEM;
-       if (!vmx->guest_msrs) {
-               goto uninit_vcpu;
-       }
+       if (!vmx->guest_msrs)
+               goto free_pml;
 
        vmx->loaded_vmcs = &vmx->vmcs01;
        vmx->loaded_vmcs->vmcs = alloc_vmcs();
@@ -8936,18 +8938,6 @@ static struct kvm_vcpu *vmx_create_vcpu(struct kvm *kvm, unsigned int id)
        vmx->nested.current_vmptr = -1ull;
        vmx->nested.current_vmcs12 = NULL;
 
-       /*
-        * If PML is turned on, failure on enabling PML just results in failure
-        * of creating the vcpu, therefore we can simplify PML logic (by
-        * avoiding dealing with cases, such as enabling PML partially on vcpus
-        * for the guest, etc.
-        */
-       if (enable_pml) {
-               err = vmx_create_pml_buffer(vmx);
-               if (err)
-                       goto free_vmcs;
-       }
-
        return &vmx->vcpu;
 
 free_vmcs:
@@ -8955,6 +8945,8 @@ free_vmcs:
        free_loaded_vmcs(vmx->loaded_vmcs);
 free_msrs:
        kfree(vmx->guest_msrs);
+free_pml:
+       vmx_destroy_pml_buffer(vmx);
 uninit_vcpu:
        kvm_vcpu_uninit(&vmx->vcpu);
 free_vcpu: