KVM: nVMX: Eliminate vmcs02 pool
authorJim Mattson <jmattson@google.com>
Mon, 27 Nov 2017 23:22:25 +0000 (17:22 -0600)
committerPaolo Bonzini <pbonzini@redhat.com>
Sat, 27 Jan 2018 08:43:03 +0000 (09:43 +0100)
The potential performance advantages of a vmcs02 pool have never been
realized. To simplify the code, eliminate the pool. Instead, a single
vmcs02 is allocated per VCPU when the VCPU enters VMX operation.

Cc: stable@vger.kernel.org # prereq for Spectre mitigation
Signed-off-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Mark Kanda <mark.kanda@oracle.com>
Reviewed-by: Ameya More <ameya.more@oracle.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/kvm/vmx.c

index c829d89..ad6a883 100644 (file)
@@ -185,7 +185,6 @@ module_param(ple_window_max, int, S_IRUGO);
 extern const ulong vmx_return;
 
 #define NR_AUTOLOAD_MSRS 8
-#define VMCS02_POOL_SIZE 1
 
 struct vmcs {
        u32 revision_id;
@@ -226,7 +225,7 @@ struct shared_msr_entry {
  * stored in guest memory specified by VMPTRLD, but is opaque to the guest,
  * which must access it using VMREAD/VMWRITE/VMCLEAR instructions.
  * More than one of these structures may exist, if L1 runs multiple L2 guests.
- * nested_vmx_run() will use the data here to build a vmcs02: a VMCS for the
+ * nested_vmx_run() will use the data here to build the vmcs02: a VMCS for the
  * underlying hardware which will be used to run L2.
  * This structure is packed to ensure that its layout is identical across
  * machines (necessary for live migration).
@@ -409,13 +408,6 @@ struct __packed vmcs12 {
  */
 #define VMCS12_SIZE 0x1000
 
-/* Used to remember the last vmcs02 used for some recently used vmcs12s */
-struct vmcs02_list {
-       struct list_head list;
-       gpa_t vmptr;
-       struct loaded_vmcs vmcs02;
-};
-
 /*
  * The nested_vmx structure is part of vcpu_vmx, and holds information we need
  * for correct emulation of VMX (i.e., nested VMX) on this vcpu.
@@ -440,15 +432,15 @@ struct nested_vmx {
         */
        bool sync_shadow_vmcs;
 
-       /* vmcs02_list cache of VMCSs recently used to run L2 guests */
-       struct list_head vmcs02_pool;
-       int vmcs02_num;
        bool change_vmcs01_virtual_x2apic_mode;
        /* L2 must run next, and mustn't decide to exit to L1. */
        bool nested_run_pending;
+
+       struct loaded_vmcs vmcs02;
+
        /*
-        * Guest pages referred to in vmcs02 with host-physical pointers, so
-        * we must keep them pinned while L2 runs.
+        * Guest pages referred to in the vmcs02 with host-physical
+        * pointers, so we must keep them pinned while L2 runs.
         */
        struct page *apic_access_page;
        struct page *virtual_apic_page;
@@ -6974,94 +6966,6 @@ static int handle_monitor(struct kvm_vcpu *vcpu)
 }
 
 /*
- * To run an L2 guest, we need a vmcs02 based on the L1-specified vmcs12.
- * We could reuse a single VMCS for all the L2 guests, but we also want the
- * option to allocate a separate vmcs02 for each separate loaded vmcs12 - this
- * allows keeping them loaded on the processor, and in the future will allow
- * optimizations where prepare_vmcs02 doesn't need to set all the fields on
- * every entry if they never change.
- * So we keep, in vmx->nested.vmcs02_pool, a cache of size VMCS02_POOL_SIZE
- * (>=0) with a vmcs02 for each recently loaded vmcs12s, most recent first.
- *
- * The following functions allocate and free a vmcs02 in this pool.
- */
-
-/* Get a VMCS from the pool to use as vmcs02 for the current vmcs12. */
-static struct loaded_vmcs *nested_get_current_vmcs02(struct vcpu_vmx *vmx)
-{
-       struct vmcs02_list *item;
-       list_for_each_entry(item, &vmx->nested.vmcs02_pool, list)
-               if (item->vmptr == vmx->nested.current_vmptr) {
-                       list_move(&item->list, &vmx->nested.vmcs02_pool);
-                       return &item->vmcs02;
-               }
-
-       if (vmx->nested.vmcs02_num >= max(VMCS02_POOL_SIZE, 1)) {
-               /* Recycle the least recently used VMCS. */
-               item = list_last_entry(&vmx->nested.vmcs02_pool,
-                                      struct vmcs02_list, list);
-               item->vmptr = vmx->nested.current_vmptr;
-               list_move(&item->list, &vmx->nested.vmcs02_pool);
-               return &item->vmcs02;
-       }
-
-       /* Create a new VMCS */
-       item = kzalloc(sizeof(struct vmcs02_list), GFP_KERNEL);
-       if (!item)
-               return NULL;
-       item->vmcs02.vmcs = alloc_vmcs();
-       item->vmcs02.shadow_vmcs = NULL;
-       if (!item->vmcs02.vmcs) {
-               kfree(item);
-               return NULL;
-       }
-       loaded_vmcs_init(&item->vmcs02);
-       item->vmptr = vmx->nested.current_vmptr;
-       list_add(&(item->list), &(vmx->nested.vmcs02_pool));
-       vmx->nested.vmcs02_num++;
-       return &item->vmcs02;
-}
-
-/* Free and remove from pool a vmcs02 saved for a vmcs12 (if there is one) */
-static void nested_free_vmcs02(struct vcpu_vmx *vmx, gpa_t vmptr)
-{
-       struct vmcs02_list *item;
-       list_for_each_entry(item, &vmx->nested.vmcs02_pool, list)
-               if (item->vmptr == vmptr) {
-                       free_loaded_vmcs(&item->vmcs02);
-                       list_del(&item->list);
-                       kfree(item);
-                       vmx->nested.vmcs02_num--;
-                       return;
-               }
-}
-
-/*
- * Free all VMCSs saved for this vcpu, except the one pointed by
- * vmx->loaded_vmcs. We must be running L1, so vmx->loaded_vmcs
- * must be &vmx->vmcs01.
- */
-static void nested_free_all_saved_vmcss(struct vcpu_vmx *vmx)
-{
-       struct vmcs02_list *item, *n;
-
-       WARN_ON(vmx->loaded_vmcs != &vmx->vmcs01);
-       list_for_each_entry_safe(item, n, &vmx->nested.vmcs02_pool, list) {
-               /*
-                * Something will leak if the above WARN triggers.  Better than
-                * a use-after-free.
-                */
-               if (vmx->loaded_vmcs == &item->vmcs02)
-                       continue;
-
-               free_loaded_vmcs(&item->vmcs02);
-               list_del(&item->list);
-               kfree(item);
-               vmx->nested.vmcs02_num--;
-       }
-}
-
-/*
  * The following 3 functions, nested_vmx_succeed()/failValid()/failInvalid(),
  * set the success or error code of an emulated VMX instruction, as specified
  * by Vol 2B, VMX Instruction Reference, "Conventions".
@@ -7242,6 +7146,12 @@ static int enter_vmx_operation(struct kvm_vcpu *vcpu)
        struct vcpu_vmx *vmx = to_vmx(vcpu);
        struct vmcs *shadow_vmcs;
 
+       vmx->nested.vmcs02.vmcs = alloc_vmcs();
+       vmx->nested.vmcs02.shadow_vmcs = NULL;
+       if (!vmx->nested.vmcs02.vmcs)
+               goto out_vmcs02;
+       loaded_vmcs_init(&vmx->nested.vmcs02);
+
        if (cpu_has_vmx_msr_bitmap()) {
                vmx->nested.msr_bitmap =
                                (unsigned long *)__get_free_page(GFP_KERNEL);
@@ -7264,9 +7174,6 @@ static int enter_vmx_operation(struct kvm_vcpu *vcpu)
                vmx->vmcs01.shadow_vmcs = shadow_vmcs;
        }
 
-       INIT_LIST_HEAD(&(vmx->nested.vmcs02_pool));
-       vmx->nested.vmcs02_num = 0;
-
        hrtimer_init(&vmx->nested.preemption_timer, CLOCK_MONOTONIC,
                     HRTIMER_MODE_REL_PINNED);
        vmx->nested.preemption_timer.function = vmx_preemption_timer_fn;
@@ -7281,6 +7188,9 @@ out_cached_vmcs12:
        free_page((unsigned long)vmx->nested.msr_bitmap);
 
 out_msr_bitmap:
+       free_loaded_vmcs(&vmx->nested.vmcs02);
+
+out_vmcs02:
        return -ENOMEM;
 }
 
@@ -7434,7 +7344,7 @@ static void free_nested(struct vcpu_vmx *vmx)
                vmx->vmcs01.shadow_vmcs = NULL;
        }
        kfree(vmx->nested.cached_vmcs12);
-       /* Unpin physical memory we referred to in current vmcs02 */
+       /* Unpin physical memory we referred to in the vmcs02 */
        if (vmx->nested.apic_access_page) {
                kvm_release_page_dirty(vmx->nested.apic_access_page);
                vmx->nested.apic_access_page = NULL;
@@ -7450,7 +7360,7 @@ static void free_nested(struct vcpu_vmx *vmx)
                vmx->nested.pi_desc = NULL;
        }
 
-       nested_free_all_saved_vmcss(vmx);
+       free_loaded_vmcs(&vmx->nested.vmcs02);
 }
 
 /* Emulate the VMXOFF instruction */
@@ -7493,8 +7403,6 @@ static int handle_vmclear(struct kvm_vcpu *vcpu)
                        vmptr + offsetof(struct vmcs12, launch_state),
                        &zero, sizeof(zero));
 
-       nested_free_vmcs02(vmx, vmptr);
-
        nested_vmx_succeed(vcpu);
        return kvm_skip_emulated_instruction(vcpu);
 }
@@ -8406,10 +8314,11 @@ static bool nested_vmx_exit_reflected(struct kvm_vcpu *vcpu, u32 exit_reason)
 
        /*
         * The host physical addresses of some pages of guest memory
-        * are loaded into VMCS02 (e.g. L1's Virtual APIC Page). The CPU
-        * may write to these pages via their host physical address while
-        * L2 is running, bypassing any address-translation-based dirty
-        * tracking (e.g. EPT write protection).
+        * are loaded into the vmcs02 (e.g. vmcs12's Virtual APIC
+        * Page). The CPU may write to these pages via their host
+        * physical address while L2 is running, bypassing any
+        * address-translation-based dirty tracking (e.g. EPT write
+        * protection).
         *
         * Mark them dirty on every exit from L2 to prevent them from
         * getting out of sync with dirty tracking.
@@ -10903,20 +10812,15 @@ static int enter_vmx_non_root_mode(struct kvm_vcpu *vcpu, bool from_vmentry)
 {
        struct vcpu_vmx *vmx = to_vmx(vcpu);
        struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
-       struct loaded_vmcs *vmcs02;
        u32 msr_entry_idx;
        u32 exit_qual;
 
-       vmcs02 = nested_get_current_vmcs02(vmx);
-       if (!vmcs02)
-               return -ENOMEM;
-
        enter_guest_mode(vcpu);
 
        if (!(vmcs12->vm_entry_controls & VM_ENTRY_LOAD_DEBUG_CONTROLS))
                vmx->nested.vmcs01_debugctl = vmcs_read64(GUEST_IA32_DEBUGCTL);
 
-       vmx_switch_vmcs(vcpu, vmcs02);
+       vmx_switch_vmcs(vcpu, &vmx->nested.vmcs02);
        vmx_segment_cache_clear(vmx);
 
        if (prepare_vmcs02(vcpu, vmcs12, from_vmentry, &exit_qual)) {
@@ -11534,10 +11438,6 @@ static void nested_vmx_vmexit(struct kvm_vcpu *vcpu, u32 exit_reason,
        vm_exit_controls_reset_shadow(vmx);
        vmx_segment_cache_clear(vmx);
 
-       /* if no vmcs02 cache requested, remove the one we used */
-       if (VMCS02_POOL_SIZE == 0)
-               nested_free_vmcs02(vmx, vmx->nested.current_vmptr);
-
        /* Update any VMCS fields that might have changed while L2 ran */
        vmcs_write32(VM_EXIT_MSR_LOAD_COUNT, vmx->msr_autoload.nr);
        vmcs_write32(VM_ENTRY_MSR_LOAD_COUNT, vmx->msr_autoload.nr);