virt/kvm/kvm_main.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * Copyright (C) 2006 Qumranet, Inc.
   8  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   9  *
  10  * Authors:
  11  *   Avi Kivity   <avi@qumranet.com>
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *
  14  * This work is licensed under the terms of the GNU GPL, version 2.  See
  15  * the COPYING file in the top-level directory.
  16  *
  17  */
  18
  19 #include "iodev.h"
  20
  21 #include <linux/kvm_host.h>
  22 #include <linux/kvm.h>
  23 #include <linux/module.h>
  24 #include <linux/errno.h>
  25 #include <linux/percpu.h>
  26 #include <linux/mm.h>
  27 #include <linux/miscdevice.h>
  28 #include <linux/vmalloc.h>
  29 #include <linux/reboot.h>
  30 #include <linux/debugfs.h>
  31 #include <linux/highmem.h>
  32 #include <linux/file.h>
  33 #include <linux/syscore_ops.h>
  34 #include <linux/cpu.h>
  35 #include <linux/sched.h>
  36 #include <linux/cpumask.h>
  37 #include <linux/smp.h>
  38 #include <linux/anon_inodes.h>
  39 #include <linux/profile.h>
  40 #include <linux/kvm_para.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/mman.h>
  43 #include <linux/swap.h>
  44 #include <linux/bitops.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/compat.h>
  47 #include <linux/srcu.h>
  48 #include <linux/hugetlb.h>
  49 #include <linux/slab.h>
  50 #include <linux/sort.h>
  51 #include <linux/bsearch.h>
  52
  53 #include <asm/processor.h>
  54 #include <asm/io.h>
  55 #include <asm/uaccess.h>
  56 #include <asm/pgtable.h>
  57
  58 #include "coalesced_mmio.h"
  59 #include "async_pf.h"
  60
  61 #define CREATE_TRACE_POINTS
  62 #include <trace/events/kvm.h>
  63
  64 MODULE_AUTHOR("Qumranet");
  65 MODULE_LICENSE("GPL");
  66
  67 /*
  68  * Ordering of locks:
  69  *
  70  *              kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  71  */
  72
  73 DEFINE_SPINLOCK(kvm_lock);
  74 static DEFINE_RAW_SPINLOCK(kvm_count_lock);
  75 LIST_HEAD(vm_list);
  76
  77 static cpumask_var_t cpus_hardware_enabled;
  78 static int kvm_usage_count = 0;
  79 static atomic_t hardware_enable_failed;
  80
  81 struct kmem_cache *kvm_vcpu_cache;
  82 EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
  83
  84 static __read_mostly struct preempt_ops kvm_preempt_ops;
  85
  86 struct dentry *kvm_debugfs_dir;
  87
  88 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
  89                            unsigned long arg);
  90 #ifdef CONFIG_COMPAT
  91 static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl,
  92                                   unsigned long arg);
  93 #endif
  94 static int hardware_enable_all(void);
  95 static void hardware_disable_all(void);
  96
  97 static void kvm_io_bus_destroy(struct kvm_io_bus *bus);
  98
  99 static void kvm_release_pfn_dirty(pfn_t pfn);
 100 static void mark_page_dirty_in_slot(struct kvm *kvm,
 101                                     struct kvm_memory_slot *memslot, gfn_t gfn);
 102
 103 __visible bool kvm_rebooting;
 104 EXPORT_SYMBOL_GPL(kvm_rebooting);
 105
 106 static bool largepages_enabled = true;
 107
 108 bool kvm_is_mmio_pfn(pfn_t pfn)
 109 {
 110         if (pfn_valid(pfn))
 111                 return PageReserved(pfn_to_page(pfn));
 112
 113         return true;
 114 }
 115
 116 /*
 117  * Switches to specified vcpu, until a matching vcpu_put()
 118  */
 119 int vcpu_load(struct kvm_vcpu *vcpu)
 120 {
 121         int cpu;
 122
 123         if (mutex_lock_killable(&vcpu->mutex))
 124                 return -EINTR;
 125         if (unlikely(vcpu->pid != current->pids[PIDTYPE_PID].pid)) {
 126                 /* The thread running this VCPU changed. */
 127                 struct pid *oldpid = vcpu->pid;
 128                 struct pid *newpid = get_task_pid(current, PIDTYPE_PID);
 129                 rcu_assign_pointer(vcpu->pid, newpid);
 130                 if (oldpid)
 131                         synchronize_rcu();
 132                 put_pid(oldpid);
 133         }
 134         cpu = get_cpu();
 135         preempt_notifier_register(&vcpu->preempt_notifier);
 136         kvm_arch_vcpu_load(vcpu, cpu);
 137         put_cpu();
 138         return 0;
 139 }
 140
 141 void vcpu_put(struct kvm_vcpu *vcpu)
 142 {
 143         preempt_disable();
 144         kvm_arch_vcpu_put(vcpu);
 145         preempt_notifier_unregister(&vcpu->preempt_notifier);
 146         preempt_enable();
 147         mutex_unlock(&vcpu->mutex);
 148 }
 149
 150 static void ack_flush(void *_completed)
 151 {
 152 }
 153
 154 static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
 155 {
 156         int i, cpu, me;
 157         cpumask_var_t cpus;
 158         bool called = true;
 159         struct kvm_vcpu *vcpu;
 160
 161         zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 162
 163         me = get_cpu();
 164         kvm_for_each_vcpu(i, vcpu, kvm) {
 165                 kvm_make_request(req, vcpu);
 166                 cpu = vcpu->cpu;
 167
 168                 /* Set ->requests bit before we read ->mode */
 169                 smp_mb();
 170
 171                 if (cpus != NULL && cpu != -1 && cpu != me &&
 172                       kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE)
 173                         cpumask_set_cpu(cpu, cpus);
 174         }
 175         if (unlikely(cpus == NULL))
 176                 smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1);
 177         else if (!cpumask_empty(cpus))
 178                 smp_call_function_many(cpus, ack_flush, NULL, 1);
 179         else
 180                 called = false;
 181         put_cpu();
 182         free_cpumask_var(cpus);
 183         return called;
 184 }
 185
 186 void kvm_flush_remote_tlbs(struct kvm *kvm)
 187 {
 188         long dirty_count = kvm->tlbs_dirty;
 189
 190         smp_mb();
 191         if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 192                 ++kvm->stat.remote_tlb_flush;
 193         cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 194 }
 195 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs);
 196
 197 void kvm_reload_remote_mmus(struct kvm *kvm)
 198 {
 199         make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
 200 }
 201
 202 void kvm_make_mclock_inprogress_request(struct kvm *kvm)
 203 {
 204         make_all_cpus_request(kvm, KVM_REQ_MCLOCK_INPROGRESS);
 205 }
 206
 207 void kvm_make_scan_ioapic_request(struct kvm *kvm)
 208 {
 209         make_all_cpus_request(kvm, KVM_REQ_SCAN_IOAPIC);
 210 }
 211
 212 int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
 213 {
 214         struct page *page;
 215         int r;
 216
 217         mutex_init(&vcpu->mutex);
 218         vcpu->cpu = -1;
 219         vcpu->kvm = kvm;
 220         vcpu->vcpu_id = id;
 221         vcpu->pid = NULL;
 222         init_waitqueue_head(&vcpu->wq);
 223         kvm_async_pf_vcpu_init(vcpu);
 224
 225         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 226         if (!page) {
 227                 r = -ENOMEM;
 228                 goto fail;
 229         }
 230         vcpu->run = page_address(page);
 231
 232         kvm_vcpu_set_in_spin_loop(vcpu, false);
 233         kvm_vcpu_set_dy_eligible(vcpu, false);
 234         vcpu->preempted = false;
 235
 236         r = kvm_arch_vcpu_init(vcpu);
 237         if (r < 0)
 238                 goto fail_free_run;
 239         return 0;
 240
 241 fail_free_run:
 242         free_page((unsigned long)vcpu->run);
 243 fail:
 244         return r;
 245 }
 246 EXPORT_SYMBOL_GPL(kvm_vcpu_init);
 247
 248 void kvm_vcpu_uninit(struct kvm_vcpu *vcpu)
 249 {
 250         put_pid(vcpu->pid);
 251         kvm_arch_vcpu_uninit(vcpu);
 252         free_page((unsigned long)vcpu->run);
 253 }
 254 EXPORT_SYMBOL_GPL(kvm_vcpu_uninit);
 255
 256 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 257 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
 258 {
 259         return container_of(mn, struct kvm, mmu_notifier);
 260 }
 261
 262 static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 263                                              struct mm_struct *mm,
 264                                              unsigned long address)
 265 {
 266         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 267         int need_tlb_flush, idx;
 268
 269         /*
 270          * When ->invalidate_page runs, the linux pte has been zapped
 271          * already but the page is still allocated until
 272          * ->invalidate_page returns. So if we increase the sequence
 273          * here the kvm page fault will notice if the spte can't be
 274          * established because the page is going to be freed. If
 275          * instead the kvm page fault establishes the spte before
 276          * ->invalidate_page runs, kvm_unmap_hva will release it
 277          * before returning.
 278          *
 279          * The sequence increase only need to be seen at spin_unlock
 280          * time, and not at spin_lock time.
 281          *
 282          * Increasing the sequence after the spin_unlock would be
 283          * unsafe because the kvm page fault could then establish the
 284          * pte after kvm_unmap_hva returned, without noticing the page
 285          * is going to be freed.
 286          */
 287         idx = srcu_read_lock(&kvm->srcu);
 288         spin_lock(&kvm->mmu_lock);
 289
 290         kvm->mmu_notifier_seq++;
 291         need_tlb_flush = kvm_unmap_hva(kvm, address) | kvm->tlbs_dirty;
 292         /* we've to flush the tlb before the pages can be freed */
 293         if (need_tlb_flush)
 294                 kvm_flush_remote_tlbs(kvm);
 295
 296         spin_unlock(&kvm->mmu_lock);
 297         srcu_read_unlock(&kvm->srcu, idx);
 298 }
 299
 300 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
 301                                         struct mm_struct *mm,
 302                                         unsigned long address,
 303                                         pte_t pte)
 304 {
 305         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 306         int idx;
 307
 308         idx = srcu_read_lock(&kvm->srcu);
 309         spin_lock(&kvm->mmu_lock);
 310         kvm->mmu_notifier_seq++;
 311         kvm_set_spte_hva(kvm, address, pte);
 312         spin_unlock(&kvm->mmu_lock);
 313         srcu_read_unlock(&kvm->srcu, idx);
 314 }
 315
 316 static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 317                                                     struct mm_struct *mm,
 318                                                     unsigned long start,
 319                                                     unsigned long end)
 320 {
 321         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 322         int need_tlb_flush = 0, idx;
 323
 324         idx = srcu_read_lock(&kvm->srcu);
 325         spin_lock(&kvm->mmu_lock);
 326         /*
 327          * The count increase must become visible at unlock time as no
 328          * spte can be established without taking the mmu_lock and
 329          * count is also read inside the mmu_lock critical section.
 330          */
 331         kvm->mmu_notifier_count++;
 332         need_tlb_flush = kvm_unmap_hva_range(kvm, start, end);
 333         need_tlb_flush |= kvm->tlbs_dirty;
 334         /* we've to flush the tlb before the pages can be freed */
 335         if (need_tlb_flush)
 336                 kvm_flush_remote_tlbs(kvm);
 337
 338         spin_unlock(&kvm->mmu_lock);
 339         srcu_read_unlock(&kvm->srcu, idx);
 340 }
 341
 342 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
 343                                                   struct mm_struct *mm,
 344                                                   unsigned long start,
 345                                                   unsigned long end)
 346 {
 347         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 348
 349         spin_lock(&kvm->mmu_lock);
 350         /*
 351          * This sequence increase will notify the kvm page fault that
 352          * the page that is going to be mapped in the spte could have
 353          * been freed.
 354          */
 355         kvm->mmu_notifier_seq++;
 356         smp_wmb();
 357         /*
 358          * The above sequence increase must be visible before the
 359          * below count decrease, which is ensured by the smp_wmb above
 360          * in conjunction with the smp_rmb in mmu_notifier_retry().
 361          */
 362         kvm->mmu_notifier_count--;
 363         spin_unlock(&kvm->mmu_lock);
 364
 365         BUG_ON(kvm->mmu_notifier_count < 0);
 366 }
 367
 368 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
 369                                               struct mm_struct *mm,
 370                                               unsigned long address)
 371 {
 372         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 373         int young, idx;
 374
 375         idx = srcu_read_lock(&kvm->srcu);
 376         spin_lock(&kvm->mmu_lock);
 377
 378         young = kvm_age_hva(kvm, address);
 379         if (young)
 380                 kvm_flush_remote_tlbs(kvm);
 381
 382         spin_unlock(&kvm->mmu_lock);
 383         srcu_read_unlock(&kvm->srcu, idx);
 384
 385         return young;
 386 }
 387
 388 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn,
 389                                        struct mm_struct *mm,
 390                                        unsigned long address)
 391 {
 392         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 393         int young, idx;
 394
 395         idx = srcu_read_lock(&kvm->srcu);
 396         spin_lock(&kvm->mmu_lock);
 397         young = kvm_test_age_hva(kvm, address);
 398         spin_unlock(&kvm->mmu_lock);
 399         srcu_read_unlock(&kvm->srcu, idx);
 400
 401         return young;
 402 }
 403
 404 static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
 405                                      struct mm_struct *mm)
 406 {
 407         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 408         int idx;
 409
 410         idx = srcu_read_lock(&kvm->srcu);
 411         kvm_arch_flush_shadow_all(kvm);
 412         srcu_read_unlock(&kvm->srcu, idx);
 413 }
 414
 415 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
 416         .invalidate_page        = kvm_mmu_notifier_invalidate_page,
 417         .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
 418         .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
 419         .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
 420         .test_young             = kvm_mmu_notifier_test_young,
 421         .change_pte             = kvm_mmu_notifier_change_pte,
 422         .release                = kvm_mmu_notifier_release,
 423 };
 424
 425 static int kvm_init_mmu_notifier(struct kvm *kvm)
 426 {
 427         kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
 428         return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
 429 }
 430
 431 #else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
 432
 433 static int kvm_init_mmu_notifier(struct kvm *kvm)
 434 {
 435         return 0;
 436 }
 437
 438 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 439
 440 static void kvm_init_memslots_id(struct kvm *kvm)
 441 {
 442         int i;
 443         struct kvm_memslots *slots = kvm->memslots;
 444
 445         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 446                 slots->id_to_index[i] = slots->memslots[i].id = i;
 447 }
 448
 449 static struct kvm *kvm_create_vm(unsigned long type)
 450 {
 451         int r, i;
 452         struct kvm *kvm = kvm_arch_alloc_vm();
 453
 454         if (!kvm)
 455                 return ERR_PTR(-ENOMEM);
 456
 457         r = kvm_arch_init_vm(kvm, type);
 458         if (r)
 459                 goto out_err_no_disable;
 460
 461         r = hardware_enable_all();
 462         if (r)
 463                 goto out_err_no_disable;
 464
 465 #ifdef CONFIG_HAVE_KVM_IRQCHIP
 466         INIT_HLIST_HEAD(&kvm->mask_notifier_list);
 467 #endif
 468 #ifdef CONFIG_HAVE_KVM_IRQFD
 469         INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 470 #endif
 471
 472         BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX);
 473
 474         r = -ENOMEM;
 475         kvm->memslots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
 476         if (!kvm->memslots)
 477                 goto out_err_no_srcu;
 478
 479         /*
 480          * Init kvm generation close to the maximum to easily test the
 481          * code of handling generation number wrap-around.
 482          */
 483         kvm->memslots->generation = -150;
 484
 485         kvm_init_memslots_id(kvm);
 486         if (init_srcu_struct(&kvm->srcu))
 487                 goto out_err_no_srcu;
 488         if (init_srcu_struct(&kvm->irq_srcu))
 489                 goto out_err_no_irq_srcu;
 490         for (i = 0; i < KVM_NR_BUSES; i++) {
 491                 kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
 492                                         GFP_KERNEL);
 493                 if (!kvm->buses[i])
 494                         goto out_err;
 495         }
 496
 497         spin_lock_init(&kvm->mmu_lock);
 498         kvm->mm = current->mm;
 499         atomic_inc(&kvm->mm->mm_count);
 500         kvm_eventfd_init(kvm);
 501         mutex_init(&kvm->lock);
 502         mutex_init(&kvm->irq_lock);
 503         mutex_init(&kvm->slots_lock);
 504         atomic_set(&kvm->users_count, 1);
 505         INIT_LIST_HEAD(&kvm->devices);
 506
 507         r = kvm_init_mmu_notifier(kvm);
 508         if (r)
 509                 goto out_err;
 510
 511         spin_lock(&kvm_lock);
 512         list_add(&kvm->vm_list, &vm_list);
 513         spin_unlock(&kvm_lock);
 514
 515         return kvm;
 516
 517 out_err:
 518         cleanup_srcu_struct(&kvm->irq_srcu);
 519 out_err_no_irq_srcu:
 520         cleanup_srcu_struct(&kvm->srcu);
 521 out_err_no_srcu:
 522         hardware_disable_all();
 523 out_err_no_disable:
 524         for (i = 0; i < KVM_NR_BUSES; i++)
 525                 kfree(kvm->buses[i]);
 526         kfree(kvm->memslots);
 527         kvm_arch_free_vm(kvm);
 528         return ERR_PTR(r);
 529 }
 530
 531 /*
 532  * Avoid using vmalloc for a small buffer.
 533  * Should not be used when the size is statically known.
 534  */
 535 void *kvm_kvzalloc(unsigned long size)
 536 {
 537         if (size > PAGE_SIZE)
 538                 return vzalloc(size);
 539         else
 540                 return kzalloc(size, GFP_KERNEL);
 541 }
 542
 543 void kvm_kvfree(const void *addr)
 544 {
 545         if (is_vmalloc_addr(addr))
 546                 vfree(addr);
 547         else
 548                 kfree(addr);
 549 }
 550
 551 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
 552 {
 553         if (!memslot->dirty_bitmap)
 554                 return;
 555
 556         kvm_kvfree(memslot->dirty_bitmap);
 557         memslot->dirty_bitmap = NULL;
 558 }
 559
 560 /*
 561  * Free any memory in @free but not in @dont.
 562  */
 563 static void kvm_free_physmem_slot(struct kvm *kvm, struct kvm_memory_slot *free,
 564                                   struct kvm_memory_slot *dont)
 565 {
 566         if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
 567                 kvm_destroy_dirty_bitmap(free);
 568
 569         kvm_arch_free_memslot(kvm, free, dont);
 570
 571         free->npages = 0;
 572 }
 573
 574 static void kvm_free_physmem(struct kvm *kvm)
 575 {
 576         struct kvm_memslots *slots = kvm->memslots;
 577         struct kvm_memory_slot *memslot;
 578
 579         kvm_for_each_memslot(memslot, slots)
 580                 kvm_free_physmem_slot(kvm, memslot, NULL);
 581
 582         kfree(kvm->memslots);
 583 }
 584
 585 static void kvm_destroy_devices(struct kvm *kvm)
 586 {
 587         struct list_head *node, *tmp;
 588
 589         list_for_each_safe(node, tmp, &kvm->devices) {
 590                 struct kvm_device *dev =
 591                         list_entry(node, struct kvm_device, vm_node);
 592
 593                 list_del(node);
 594                 dev->ops->destroy(dev);
 595         }
 596 }
 597
 598 static void kvm_destroy_vm(struct kvm *kvm)
 599 {
 600         int i;
 601         struct mm_struct *mm = kvm->mm;
 602
 603         kvm_arch_sync_events(kvm);
 604         spin_lock(&kvm_lock);
 605         list_del(&kvm->vm_list);
 606         spin_unlock(&kvm_lock);
 607         kvm_free_irq_routing(kvm);
 608         for (i = 0; i < KVM_NR_BUSES; i++)
 609                 kvm_io_bus_destroy(kvm->buses[i]);
 610         kvm_coalesced_mmio_free(kvm);
 611 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 612         mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 613 #else
 614         kvm_arch_flush_shadow_all(kvm);
 615 #endif
 616         kvm_arch_destroy_vm(kvm);
 617         kvm_destroy_devices(kvm);
 618         kvm_free_physmem(kvm);
 619         cleanup_srcu_struct(&kvm->irq_srcu);
 620         cleanup_srcu_struct(&kvm->srcu);
 621         kvm_arch_free_vm(kvm);
 622         hardware_disable_all();
 623         mmdrop(mm);
 624 }
 625
 626 void kvm_get_kvm(struct kvm *kvm)
 627 {
 628         atomic_inc(&kvm->users_count);
 629 }
 630 EXPORT_SYMBOL_GPL(kvm_get_kvm);
 631
 632 void kvm_put_kvm(struct kvm *kvm)
 633 {
 634         if (atomic_dec_and_test(&kvm->users_count))
 635                 kvm_destroy_vm(kvm);
 636 }
 637 EXPORT_SYMBOL_GPL(kvm_put_kvm);
 638
 639
 640 static int kvm_vm_release(struct inode *inode, struct file *filp)
 641 {
 642         struct kvm *kvm = filp->private_data;
 643
 644         kvm_irqfd_release(kvm);
 645
 646         kvm_put_kvm(kvm);
 647         return 0;
 648 }
 649
 650 /*
 651  * Allocation size is twice as large as the actual dirty bitmap size.
 652  * See x86's kvm_vm_ioctl_get_dirty_log() why this is needed.
 653  */
 654 static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
 655 {
 656         unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
 657
 658         memslot->dirty_bitmap = kvm_kvzalloc(dirty_bytes);
 659         if (!memslot->dirty_bitmap)
 660                 return -ENOMEM;
 661
 662         return 0;
 663 }
 664
 665 static int cmp_memslot(const void *slot1, const void *slot2)
 666 {
 667         struct kvm_memory_slot *s1, *s2;
 668
 669         s1 = (struct kvm_memory_slot *)slot1;
 670         s2 = (struct kvm_memory_slot *)slot2;
 671
 672         if (s1->npages < s2->npages)
 673                 return 1;
 674         if (s1->npages > s2->npages)
 675                 return -1;
 676
 677         return 0;
 678 }
 679
 680 /*
 681  * Sort the memslots base on its size, so the larger slots
 682  * will get better fit.
 683  */
 684 static void sort_memslots(struct kvm_memslots *slots)
 685 {
 686         int i;
 687
 688         sort(slots->memslots, KVM_MEM_SLOTS_NUM,
 689               sizeof(struct kvm_memory_slot), cmp_memslot, NULL);
 690
 691         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 692                 slots->id_to_index[slots->memslots[i].id] = i;
 693 }
 694
 695 static void update_memslots(struct kvm_memslots *slots,
 696                             struct kvm_memory_slot *new)
 697 {
 698         if (new) {
 699                 int id = new->id;
 700                 struct kvm_memory_slot *old = id_to_memslot(slots, id);
 701                 unsigned long npages = old->npages;
 702
 703                 *old = *new;
 704                 if (new->npages != npages)
 705                         sort_memslots(slots);
 706         }
 707 }
 708
 709 static int check_memory_region_flags(struct kvm_userspace_memory_region *mem)
 710 {
 711         u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES;
 712
 713 #ifdef __KVM_HAVE_READONLY_MEM
 714         valid_flags |= KVM_MEM_READONLY;
 715 #endif
 716
 717         if (mem->flags & ~valid_flags)
 718                 return -EINVAL;
 719
 720         return 0;
 721 }
 722
 723 static struct kvm_memslots *install_new_memslots(struct kvm *kvm,
 724                 struct kvm_memslots *slots, struct kvm_memory_slot *new)
 725 {
 726         struct kvm_memslots *old_memslots = kvm->memslots;
 727
 728         /*
 729          * Set the low bit in the generation, which disables SPTE caching
 730          * until the end of synchronize_srcu_expedited.
 731          */
 732         WARN_ON(old_memslots->generation & 1);
 733         slots->generation = old_memslots->generation + 1;
 734
 735         update_memslots(slots, new);
 736         rcu_assign_pointer(kvm->memslots, slots);
 737         synchronize_srcu_expedited(&kvm->srcu);
 738
 739         /*
 740          * Increment the new memslot generation a second time. This prevents
 741          * vm exits that race with memslot updates from caching a memslot
 742          * generation that will (potentially) be valid forever.
 743          */
 744         slots->generation++;
 745
 746         kvm_arch_memslots_updated(kvm);
 747
 748         return old_memslots;
 749 }
 750
 751 /*
 752  * Allocate some memory and give it an address in the guest physical address
 753  * space.
 754  *
 755  * Discontiguous memory is allowed, mostly for framebuffers.
 756  *
 757  * Must be called holding mmap_sem for write.
 758  */
 759 int __kvm_set_memory_region(struct kvm *kvm,
 760                             struct kvm_userspace_memory_region *mem)
 761 {
 762         int r;
 763         gfn_t base_gfn;
 764         unsigned long npages;
 765         struct kvm_memory_slot *slot;
 766         struct kvm_memory_slot old, new;
 767         struct kvm_memslots *slots = NULL, *old_memslots;
 768         enum kvm_mr_change change;
 769
 770         r = check_memory_region_flags(mem);
 771         if (r)
 772                 goto out;
 773
 774         r = -EINVAL;
 775         /* General sanity checks */
 776         if (mem->memory_size & (PAGE_SIZE - 1))
 777                 goto out;
 778         if (mem->guest_phys_addr & (PAGE_SIZE - 1))
 779                 goto out;
 780         /* We can read the guest memory with __xxx_user() later on. */
 781         if ((mem->slot < KVM_USER_MEM_SLOTS) &&
 782             ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
 783              !access_ok(VERIFY_WRITE,
 784                         (void __user *)(unsigned long)mem->userspace_addr,
 785                         mem->memory_size)))
 786                 goto out;
 787         if (mem->slot >= KVM_MEM_SLOTS_NUM)
 788                 goto out;
 789         if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
 790                 goto out;
 791
 792         slot = id_to_memslot(kvm->memslots, mem->slot);
 793         base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
 794         npages = mem->memory_size >> PAGE_SHIFT;
 795
 796         if (npages > KVM_MEM_MAX_NR_PAGES)
 797                 goto out;
 798
 799         if (!npages)
 800                 mem->flags &= ~KVM_MEM_LOG_DIRTY_PAGES;
 801
 802         new = old = *slot;
 803
 804         new.id = mem->slot;
 805         new.base_gfn = base_gfn;
 806         new.npages = npages;
 807         new.flags = mem->flags;
 808
 809         if (npages) {
 810                 if (!old.npages)
 811                         change = KVM_MR_CREATE;
 812                 else { /* Modify an existing slot. */
 813                         if ((mem->userspace_addr != old.userspace_addr) ||
 814                             (npages != old.npages) ||
 815                             ((new.flags ^ old.flags) & KVM_MEM_READONLY))
 816                                 goto out;
 817
 818                         if (base_gfn != old.base_gfn)
 819                                 change = KVM_MR_MOVE;
 820                         else if (new.flags != old.flags)
 821                                 change = KVM_MR_FLAGS_ONLY;
 822                         else { /* Nothing to change. */
 823                                 r = 0;
 824                                 goto out;
 825                         }
 826                 }
 827         } else if (old.npages) {
 828                 change = KVM_MR_DELETE;
 829         } else /* Modify a non-existent slot: disallowed. */
 830                 goto out;
 831
 832         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 833                 /* Check for overlaps */
 834                 r = -EEXIST;
 835                 kvm_for_each_memslot(slot, kvm->memslots) {
 836                         if ((slot->id >= KVM_USER_MEM_SLOTS) ||
 837                             (slot->id == mem->slot))
 838                                 continue;
 839                         if (!((base_gfn + npages <= slot->base_gfn) ||
 840                               (base_gfn >= slot->base_gfn + slot->npages)))
 841                                 goto out;
 842                 }
 843         }
 844
 845         /* Free page dirty bitmap if unneeded */
 846         if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
 847                 new.dirty_bitmap = NULL;
 848
 849         r = -ENOMEM;
 850         if (change == KVM_MR_CREATE) {
 851                 new.userspace_addr = mem->userspace_addr;
 852
 853                 if (kvm_arch_create_memslot(kvm, &new, npages))
 854                         goto out_free;
 855         }
 856
 857         /* Allocate page dirty bitmap if needed */
 858         if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
 859                 if (kvm_create_dirty_bitmap(&new) < 0)
 860                         goto out_free;
 861         }
 862
 863         if ((change == KVM_MR_DELETE) || (change == KVM_MR_MOVE)) {
 864                 slots = kmemdup(kvm->memslots, sizeof(struct kvm_memslots),
 865                                 GFP_KERNEL);
 866                 if (!slots)
 867                         goto out_free;
 868                 slot = id_to_memslot(slots, mem->slot);
 869                 slot->flags |= KVM_MEMSLOT_INVALID;
 870
 871                 old_memslots = install_new_memslots(kvm, slots, NULL);
 872
 873                 /* slot was deleted or moved, clear iommu mapping */
 874                 kvm_iommu_unmap_pages(kvm, &old);
 875                 /* From this point no new shadow pages pointing to a deleted,
 876                  * or moved, memslot will be created.
 877                  *
 878                  * validation of sp->gfn happens in:
 879                  *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
 880                  *      - kvm_is_visible_gfn (mmu_check_roots)
 881                  */
 882                 kvm_arch_flush_shadow_memslot(kvm, slot);
 883                 slots = old_memslots;
 884         }
 885
 886         r = kvm_arch_prepare_memory_region(kvm, &new, mem, change);
 887         if (r)
 888                 goto out_slots;
 889
 890         r = -ENOMEM;
 891         /*
 892          * We can re-use the old_memslots from above, the only difference
 893          * from the currently installed memslots is the invalid flag.  This
 894          * will get overwritten by update_memslots anyway.
 895          */
 896         if (!slots) {
 897                 slots = kmemdup(kvm->memslots, sizeof(struct kvm_memslots),
 898                                 GFP_KERNEL);
 899                 if (!slots)
 900                         goto out_free;
 901         }
 902
 903         /* actual memory is freed via old in kvm_free_physmem_slot below */
 904         if (change == KVM_MR_DELETE) {
 905                 new.dirty_bitmap = NULL;
 906                 memset(&new.arch, 0, sizeof(new.arch));
 907         }
 908
 909         old_memslots = install_new_memslots(kvm, slots, &new);
 910
 911         kvm_arch_commit_memory_region(kvm, mem, &old, change);
 912
 913         kvm_free_physmem_slot(kvm, &old, &new);
 914         kfree(old_memslots);
 915
 916         /*
 917          * IOMMU mapping:  New slots need to be mapped.  Old slots need to be
 918          * un-mapped and re-mapped if their base changes.  Since base change
 919          * unmapping is handled above with slot deletion, mapping alone is
 920          * needed here.  Anything else the iommu might care about for existing
 921          * slots (size changes, userspace addr changes and read-only flag
 922          * changes) is disallowed above, so any other attribute changes getting
 923          * here can be skipped.
 924          */
 925         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 926                 r = kvm_iommu_map_pages(kvm, &new);
 927                 return r;
 928         }
 929
 930         return 0;
 931
 932 out_slots:
 933         kfree(slots);
 934 out_free:
 935         kvm_free_physmem_slot(kvm, &new, &old);
 936 out:
 937         return r;
 938 }
 939 EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
 940
 941 int kvm_set_memory_region(struct kvm *kvm,
 942                           struct kvm_userspace_memory_region *mem)
 943 {
 944         int r;
 945
 946         mutex_lock(&kvm->slots_lock);
 947         r = __kvm_set_memory_region(kvm, mem);
 948         mutex_unlock(&kvm->slots_lock);
 949         return r;
 950 }
 951 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
 952
 953 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
 954                                           struct kvm_userspace_memory_region *mem)
 955 {
 956         if (mem->slot >= KVM_USER_MEM_SLOTS)
 957                 return -EINVAL;
 958         return kvm_set_memory_region(kvm, mem);
 959 }
 960
 961 int kvm_get_dirty_log(struct kvm *kvm,
 962                         struct kvm_dirty_log *log, int *is_dirty)
 963 {
 964         struct kvm_memory_slot *memslot;
 965         int r, i;
 966         unsigned long n;
 967         unsigned long any = 0;
 968
 969         r = -EINVAL;
 970         if (log->slot >= KVM_USER_MEM_SLOTS)
 971                 goto out;
 972
 973         memslot = id_to_memslot(kvm->memslots, log->slot);
 974         r = -ENOENT;
 975         if (!memslot->dirty_bitmap)
 976                 goto out;
 977
 978         n = kvm_dirty_bitmap_bytes(memslot);
 979
 980         for (i = 0; !any && i < n/sizeof(long); ++i)
 981                 any = memslot->dirty_bitmap[i];
 982
 983         r = -EFAULT;
 984         if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
 985                 goto out;
 986
 987         if (any)
 988                 *is_dirty = 1;
 989
 990         r = 0;
 991 out:
 992         return r;
 993 }
 994 EXPORT_SYMBOL_GPL(kvm_get_dirty_log);
 995
 996 bool kvm_largepages_enabled(void)
 997 {
 998         return largepages_enabled;
 999 }
1000
1001 void kvm_disable_largepages(void)
1002 {
1003         largepages_enabled = false;
1004 }
1005 EXPORT_SYMBOL_GPL(kvm_disable_largepages);
1006
1007 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
1008 {
1009         return __gfn_to_memslot(kvm_memslots(kvm), gfn);
1010 }
1011 EXPORT_SYMBOL_GPL(gfn_to_memslot);
1012
1013 int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
1014 {
1015         struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
1016
1017         if (!memslot || memslot->id >= KVM_USER_MEM_SLOTS ||
1018               memslot->flags & KVM_MEMSLOT_INVALID)
1019                 return 0;
1020
1021         return 1;
1022 }
1023 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
1024
1025 unsigned long kvm_host_page_size(struct kvm *kvm, gfn_t gfn)
1026 {
1027         struct vm_area_struct *vma;
1028         unsigned long addr, size;
1029
1030         size = PAGE_SIZE;
1031
1032         addr = gfn_to_hva(kvm, gfn);
1033         if (kvm_is_error_hva(addr))
1034                 return PAGE_SIZE;
1035
1036         down_read(&current->mm->mmap_sem);
1037         vma = find_vma(current->mm, addr);
1038         if (!vma)
1039                 goto out;
1040
1041         size = vma_kernel_pagesize(vma);
1042
1043 out:
1044         up_read(&current->mm->mmap_sem);
1045
1046         return size;
1047 }
1048
1049 static bool memslot_is_readonly(struct kvm_memory_slot *slot)
1050 {
1051         return slot->flags & KVM_MEM_READONLY;
1052 }
1053
1054 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1055                                        gfn_t *nr_pages, bool write)
1056 {
1057         if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1058                 return KVM_HVA_ERR_BAD;
1059
1060         if (memslot_is_readonly(slot) && write)
1061                 return KVM_HVA_ERR_RO_BAD;
1062
1063         if (nr_pages)
1064                 *nr_pages = slot->npages - (gfn - slot->base_gfn);
1065
1066         return __gfn_to_hva_memslot(slot, gfn);
1067 }
1068
1069 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1070                                      gfn_t *nr_pages)
1071 {
1072         return __gfn_to_hva_many(slot, gfn, nr_pages, true);
1073 }
1074
1075 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot,
1076                                         gfn_t gfn)
1077 {
1078         return gfn_to_hva_many(slot, gfn, NULL);
1079 }
1080 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot);
1081
1082 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
1083 {
1084         return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
1085 }
1086 EXPORT_SYMBOL_GPL(gfn_to_hva);
1087
1088 /*
1089  * If writable is set to false, the hva returned by this function is only
1090  * allowed to be read.
1091  */
1092 unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot,
1093                                       gfn_t gfn, bool *writable)
1094 {
1095         unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false);
1096
1097         if (!kvm_is_error_hva(hva) && writable)
1098                 *writable = !memslot_is_readonly(slot);
1099
1100         return hva;
1101 }
1102
1103 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable)
1104 {
1105         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1106
1107         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1108 }
1109
1110 static int kvm_read_hva(void *data, void __user *hva, int len)
1111 {
1112         return __copy_from_user(data, hva, len);
1113 }
1114
1115 static int kvm_read_hva_atomic(void *data, void __user *hva, int len)
1116 {
1117         return __copy_from_user_inatomic(data, hva, len);
1118 }
1119
1120 static int get_user_page_nowait(struct task_struct *tsk, struct mm_struct *mm,
1121         unsigned long start, int write, struct page **page)
1122 {
1123         int flags = FOLL_TOUCH | FOLL_NOWAIT | FOLL_HWPOISON | FOLL_GET;
1124
1125         if (write)
1126                 flags |= FOLL_WRITE;
1127
1128         return __get_user_pages(tsk, mm, start, 1, flags, page, NULL, NULL);
1129 }
1130
1131 static inline int check_user_page_hwpoison(unsigned long addr)
1132 {
1133         int rc, flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_WRITE;
1134
1135         rc = __get_user_pages(current, current->mm, addr, 1,
1136                               flags, NULL, NULL, NULL);
1137         return rc == -EHWPOISON;
1138 }
1139
1140 /*
1141  * The atomic path to get the writable pfn which will be stored in @pfn,
1142  * true indicates success, otherwise false is returned.
1143  */
1144 static bool hva_to_pfn_fast(unsigned long addr, bool atomic, bool *async,
1145                             bool write_fault, bool *writable, pfn_t *pfn)
1146 {
1147         struct page *page[1];
1148         int npages;
1149
1150         if (!(async || atomic))
1151                 return false;
1152
1153         /*
1154          * Fast pin a writable pfn only if it is a write fault request
1155          * or the caller allows to map a writable pfn for a read fault
1156          * request.
1157          */
1158         if (!(write_fault || writable))
1159                 return false;
1160
1161         npages = __get_user_pages_fast(addr, 1, 1, page);
1162         if (npages == 1) {
1163                 *pfn = page_to_pfn(page[0]);
1164
1165                 if (writable)
1166                         *writable = true;
1167                 return true;
1168         }
1169
1170         return false;
1171 }
1172
1173 /*
1174  * The slow path to get the pfn of the specified host virtual address,
1175  * 1 indicates success, -errno is returned if error is detected.
1176  */
1177 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault,
1178                            bool *writable, pfn_t *pfn)
1179 {
1180         struct page *page[1];
1181         int npages = 0;
1182
1183         might_sleep();
1184
1185         if (writable)
1186                 *writable = write_fault;
1187
1188         if (async) {
1189                 down_read(&current->mm->mmap_sem);
1190                 npages = get_user_page_nowait(current, current->mm,
1191                                               addr, write_fault, page);
1192                 up_read(&current->mm->mmap_sem);
1193         } else
1194                 npages = get_user_pages_fast(addr, 1, write_fault,
1195                                              page);
1196         if (npages != 1)
1197                 return npages;
1198
1199         /* map read fault as writable if possible */
1200         if (unlikely(!write_fault) && writable) {
1201                 struct page *wpage[1];
1202
1203                 npages = __get_user_pages_fast(addr, 1, 1, wpage);
1204                 if (npages == 1) {
1205                         *writable = true;
1206                         put_page(page[0]);
1207                         page[0] = wpage[0];
1208                 }
1209
1210                 npages = 1;
1211         }
1212         *pfn = page_to_pfn(page[0]);
1213         return npages;
1214 }
1215
1216 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault)
1217 {
1218         if (unlikely(!(vma->vm_flags & VM_READ)))
1219                 return false;
1220
1221         if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE))))
1222                 return false;
1223
1224         return true;
1225 }
1226
1227 /*
1228  * Pin guest page in memory and return its pfn.
1229  * @addr: host virtual address which maps memory to the guest
1230  * @atomic: whether this function can sleep
1231  * @async: whether this function need to wait IO complete if the
1232  *         host page is not in the memory
1233  * @write_fault: whether we should get a writable host page
1234  * @writable: whether it allows to map a writable host page for !@write_fault
1235  *
1236  * The function will map a writable host page for these two cases:
1237  * 1): @write_fault = true
1238  * 2): @write_fault = false && @writable, @writable will tell the caller
1239  *     whether the mapping is writable.
1240  */
1241 static pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
1242                         bool write_fault, bool *writable)
1243 {
1244         struct vm_area_struct *vma;
1245         pfn_t pfn = 0;
1246         int npages;
1247
1248         /* we can do it either atomically or asynchronously, not both */
1249         BUG_ON(atomic && async);
1250
1251         if (hva_to_pfn_fast(addr, atomic, async, write_fault, writable, &pfn))
1252                 return pfn;
1253
1254         if (atomic)
1255                 return KVM_PFN_ERR_FAULT;
1256
1257         npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn);
1258         if (npages == 1)
1259                 return pfn;
1260
1261         down_read(&current->mm->mmap_sem);
1262         if (npages == -EHWPOISON ||
1263               (!async && check_user_page_hwpoison(addr))) {
1264                 pfn = KVM_PFN_ERR_HWPOISON;
1265                 goto exit;
1266         }
1267
1268         vma = find_vma_intersection(current->mm, addr, addr + 1);
1269
1270         if (vma == NULL)
1271                 pfn = KVM_PFN_ERR_FAULT;
1272         else if ((vma->vm_flags & VM_PFNMAP)) {
1273                 pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) +
1274                         vma->vm_pgoff;
1275                 BUG_ON(!kvm_is_mmio_pfn(pfn));
1276         } else {
1277                 if (async && vma_is_valid(vma, write_fault))
1278                         *async = true;
1279                 pfn = KVM_PFN_ERR_FAULT;
1280         }
1281 exit:
1282         up_read(&current->mm->mmap_sem);
1283         return pfn;
1284 }
1285
1286 static pfn_t
1287 __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn, bool atomic,
1288                      bool *async, bool write_fault, bool *writable)
1289 {
1290         unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
1291
1292         if (addr == KVM_HVA_ERR_RO_BAD)
1293                 return KVM_PFN_ERR_RO_FAULT;
1294
1295         if (kvm_is_error_hva(addr))
1296                 return KVM_PFN_NOSLOT;
1297
1298         /* Do not map writable pfn in the readonly memslot. */
1299         if (writable && memslot_is_readonly(slot)) {
1300                 *writable = false;
1301                 writable = NULL;
1302         }
1303
1304         return hva_to_pfn(addr, atomic, async, write_fault,
1305                           writable);
1306 }
1307
1308 static pfn_t __gfn_to_pfn(struct kvm *kvm, gfn_t gfn, bool atomic, bool *async,
1309                           bool write_fault, bool *writable)
1310 {
1311         struct kvm_memory_slot *slot;
1312
1313         if (async)
1314                 *async = false;
1315
1316         slot = gfn_to_memslot(kvm, gfn);
1317
1318         return __gfn_to_pfn_memslot(slot, gfn, atomic, async, write_fault,
1319                                     writable);
1320 }
1321
1322 pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
1323 {
1324         return __gfn_to_pfn(kvm, gfn, true, NULL, true, NULL);
1325 }
1326 EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
1327
1328 pfn_t gfn_to_pfn_async(struct kvm *kvm, gfn_t gfn, bool *async,
1329                        bool write_fault, bool *writable)
1330 {
1331         return __gfn_to_pfn(kvm, gfn, false, async, write_fault, writable);
1332 }
1333 EXPORT_SYMBOL_GPL(gfn_to_pfn_async);
1334
1335 pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
1336 {
1337         return __gfn_to_pfn(kvm, gfn, false, NULL, true, NULL);
1338 }
1339 EXPORT_SYMBOL_GPL(gfn_to_pfn);
1340
1341 pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
1342                       bool *writable)
1343 {
1344         return __gfn_to_pfn(kvm, gfn, false, NULL, write_fault, writable);
1345 }
1346 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
1347
1348 pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
1349 {
1350         return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
1351 }
1352
1353 pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
1354 {
1355         return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
1356 }
1357 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
1358
1359 int gfn_to_page_many_atomic(struct kvm *kvm, gfn_t gfn, struct page **pages,
1360                                                                   int nr_pages)
1361 {
1362         unsigned long addr;
1363         gfn_t entry;
1364
1365         addr = gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, &entry);
1366         if (kvm_is_error_hva(addr))
1367                 return -1;
1368
1369         if (entry < nr_pages)
1370                 return 0;
1371
1372         return __get_user_pages_fast(addr, nr_pages, 1, pages);
1373 }
1374 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
1375
1376 static struct page *kvm_pfn_to_page(pfn_t pfn)
1377 {
1378         if (is_error_noslot_pfn(pfn))
1379                 return KVM_ERR_PTR_BAD_PAGE;
1380
1381         if (kvm_is_mmio_pfn(pfn)) {
1382                 WARN_ON(1);
1383                 return KVM_ERR_PTR_BAD_PAGE;
1384         }
1385
1386         return pfn_to_page(pfn);
1387 }
1388
1389 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
1390 {
1391         pfn_t pfn;
1392
1393         pfn = gfn_to_pfn(kvm, gfn);
1394
1395         return kvm_pfn_to_page(pfn);
1396 }
1397
1398 EXPORT_SYMBOL_GPL(gfn_to_page);
1399
1400 void kvm_release_page_clean(struct page *page)
1401 {
1402         WARN_ON(is_error_page(page));
1403
1404         kvm_release_pfn_clean(page_to_pfn(page));
1405 }
1406 EXPORT_SYMBOL_GPL(kvm_release_page_clean);
1407
1408 void kvm_release_pfn_clean(pfn_t pfn)
1409 {
1410         if (!is_error_noslot_pfn(pfn) && !kvm_is_mmio_pfn(pfn))
1411                 put_page(pfn_to_page(pfn));
1412 }
1413 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
1414
1415 void kvm_release_page_dirty(struct page *page)
1416 {
1417         WARN_ON(is_error_page(page));
1418
1419         kvm_release_pfn_dirty(page_to_pfn(page));
1420 }
1421 EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
1422
1423 static void kvm_release_pfn_dirty(pfn_t pfn)
1424 {
1425         kvm_set_pfn_dirty(pfn);
1426         kvm_release_pfn_clean(pfn);
1427 }
1428
1429 void kvm_set_pfn_dirty(pfn_t pfn)
1430 {
1431         if (!kvm_is_mmio_pfn(pfn)) {
1432                 struct page *page = pfn_to_page(pfn);
1433                 if (!PageReserved(page))
1434                         SetPageDirty(page);
1435         }
1436 }
1437 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
1438
1439 void kvm_set_pfn_accessed(pfn_t pfn)
1440 {
1441         if (!kvm_is_mmio_pfn(pfn))
1442                 mark_page_accessed(pfn_to_page(pfn));
1443 }
1444 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
1445
1446 void kvm_get_pfn(pfn_t pfn)
1447 {
1448         if (!kvm_is_mmio_pfn(pfn))
1449                 get_page(pfn_to_page(pfn));
1450 }
1451 EXPORT_SYMBOL_GPL(kvm_get_pfn);
1452
1453 static int next_segment(unsigned long len, int offset)
1454 {
1455         if (len > PAGE_SIZE - offset)
1456                 return PAGE_SIZE - offset;
1457         else
1458                 return len;
1459 }
1460
1461 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
1462                         int len)
1463 {
1464         int r;
1465         unsigned long addr;
1466
1467         addr = gfn_to_hva_prot(kvm, gfn, NULL);
1468         if (kvm_is_error_hva(addr))
1469                 return -EFAULT;
1470         r = kvm_read_hva(data, (void __user *)addr + offset, len);
1471         if (r)
1472                 return -EFAULT;
1473         return 0;
1474 }
1475 EXPORT_SYMBOL_GPL(kvm_read_guest_page);
1476
1477 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
1478 {
1479         gfn_t gfn = gpa >> PAGE_SHIFT;
1480         int seg;
1481         int offset = offset_in_page(gpa);
1482         int ret;
1483
1484         while ((seg = next_segment(len, offset)) != 0) {
1485                 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
1486                 if (ret < 0)
1487                         return ret;
1488                 offset = 0;
1489                 len -= seg;
1490                 data += seg;
1491                 ++gfn;
1492         }
1493         return 0;
1494 }
1495 EXPORT_SYMBOL_GPL(kvm_read_guest);
1496
1497 int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data,
1498                           unsigned long len)
1499 {
1500         int r;
1501         unsigned long addr;
1502         gfn_t gfn = gpa >> PAGE_SHIFT;
1503         int offset = offset_in_page(gpa);
1504
1505         addr = gfn_to_hva_prot(kvm, gfn, NULL);
1506         if (kvm_is_error_hva(addr))
1507                 return -EFAULT;
1508         pagefault_disable();
1509         r = kvm_read_hva_atomic(data, (void __user *)addr + offset, len);
1510         pagefault_enable();
1511         if (r)
1512                 return -EFAULT;
1513         return 0;
1514 }
1515 EXPORT_SYMBOL(kvm_read_guest_atomic);
1516
1517 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn, const void *data,
1518                          int offset, int len)
1519 {
1520         int r;
1521         unsigned long addr;
1522
1523         addr = gfn_to_hva(kvm, gfn);
1524         if (kvm_is_error_hva(addr))
1525                 return -EFAULT;
1526         r = __copy_to_user((void __user *)addr + offset, data, len);
1527         if (r)
1528                 return -EFAULT;
1529         mark_page_dirty(kvm, gfn);
1530         return 0;
1531 }
1532 EXPORT_SYMBOL_GPL(kvm_write_guest_page);
1533
1534 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
1535                     unsigned long len)
1536 {
1537         gfn_t gfn = gpa >> PAGE_SHIFT;
1538         int seg;
1539         int offset = offset_in_page(gpa);
1540         int ret;
1541
1542         while ((seg = next_segment(len, offset)) != 0) {
1543                 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
1544                 if (ret < 0)
1545                         return ret;
1546                 offset = 0;
1547                 len -= seg;
1548                 data += seg;
1549                 ++gfn;
1550         }
1551         return 0;
1552 }
1553
1554 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1555                               gpa_t gpa, unsigned long len)
1556 {
1557         struct kvm_memslots *slots = kvm_memslots(kvm);
1558         int offset = offset_in_page(gpa);
1559         gfn_t start_gfn = gpa >> PAGE_SHIFT;
1560         gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT;
1561         gfn_t nr_pages_needed = end_gfn - start_gfn + 1;
1562         gfn_t nr_pages_avail;
1563
1564         ghc->gpa = gpa;
1565         ghc->generation = slots->generation;
1566         ghc->len = len;
1567         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1568         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn, &nr_pages_avail);
1569         if (!kvm_is_error_hva(ghc->hva) && nr_pages_avail >= nr_pages_needed) {
1570                 ghc->hva += offset;
1571         } else {
1572                 /*
1573                  * If the requested region crosses two memslots, we still
1574                  * verify that the entire region is valid here.
1575                  */
1576                 while (start_gfn <= end_gfn) {
1577                         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1578                         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn,
1579                                                    &nr_pages_avail);
1580                         if (kvm_is_error_hva(ghc->hva))
1581                                 return -EFAULT;
1582                         start_gfn += nr_pages_avail;
1583                 }
1584                 /* Use the slow path for cross page reads and writes. */
1585                 ghc->memslot = NULL;
1586         }
1587         return 0;
1588 }
1589 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
1590
1591 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1592                            void *data, unsigned long len)
1593 {
1594         struct kvm_memslots *slots = kvm_memslots(kvm);
1595         int r;
1596
1597         BUG_ON(len > ghc->len);
1598
1599         if (slots->generation != ghc->generation)
1600                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1601
1602         if (unlikely(!ghc->memslot))
1603                 return kvm_write_guest(kvm, ghc->gpa, data, len);
1604
1605         if (kvm_is_error_hva(ghc->hva))
1606                 return -EFAULT;
1607
1608         r = __copy_to_user((void __user *)ghc->hva, data, len);
1609         if (r)
1610                 return -EFAULT;
1611         mark_page_dirty_in_slot(kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT);
1612
1613         return 0;
1614 }
1615 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
1616
1617 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1618                            void *data, unsigned long len)
1619 {
1620         struct kvm_memslots *slots = kvm_memslots(kvm);
1621         int r;
1622
1623         BUG_ON(len > ghc->len);
1624
1625         if (slots->generation != ghc->generation)
1626                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1627
1628         if (unlikely(!ghc->memslot))
1629                 return kvm_read_guest(kvm, ghc->gpa, data, len);
1630
1631         if (kvm_is_error_hva(ghc->hva))
1632                 return -EFAULT;
1633
1634         r = __copy_from_user(data, (void __user *)ghc->hva, len);
1635         if (r)
1636                 return -EFAULT;
1637
1638         return 0;
1639 }
1640 EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
1641
1642 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
1643 {
1644         const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0)));
1645
1646         return kvm_write_guest_page(kvm, gfn, zero_page, offset, len);
1647 }
1648 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
1649
1650 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
1651 {
1652         gfn_t gfn = gpa >> PAGE_SHIFT;
1653         int seg;
1654         int offset = offset_in_page(gpa);
1655         int ret;
1656
1657         while ((seg = next_segment(len, offset)) != 0) {
1658                 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
1659                 if (ret < 0)
1660                         return ret;
1661                 offset = 0;
1662                 len -= seg;
1663                 ++gfn;
1664         }
1665         return 0;
1666 }
1667 EXPORT_SYMBOL_GPL(kvm_clear_guest);
1668
1669 static void mark_page_dirty_in_slot(struct kvm *kvm,
1670                                     struct kvm_memory_slot *memslot,
1671                                     gfn_t gfn)
1672 {
1673         if (memslot && memslot->dirty_bitmap) {
1674                 unsigned long rel_gfn = gfn - memslot->base_gfn;
1675
1676                 set_bit_le(rel_gfn, memslot->dirty_bitmap);
1677         }
1678 }
1679
1680 void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
1681 {
1682         struct kvm_memory_slot *memslot;
1683
1684         memslot = gfn_to_memslot(kvm, gfn);
1685         mark_page_dirty_in_slot(kvm, memslot, gfn);
1686 }
1687 EXPORT_SYMBOL_GPL(mark_page_dirty);
1688
1689 /*
1690  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
1691  */
1692 void kvm_vcpu_block(struct kvm_vcpu *vcpu)
1693 {
1694         DEFINE_WAIT(wait);
1695
1696         for (;;) {
1697                 prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
1698
1699                 if (kvm_arch_vcpu_runnable(vcpu)) {
1700                         kvm_make_request(KVM_REQ_UNHALT, vcpu);
1701                         break;
1702                 }
1703                 if (kvm_cpu_has_pending_timer(vcpu))
1704                         break;
1705                 if (signal_pending(current))
1706                         break;
1707
1708                 schedule();
1709         }
1710
1711         finish_wait(&vcpu->wq, &wait);
1712 }
1713 EXPORT_SYMBOL_GPL(kvm_vcpu_block);
1714
1715 #ifndef CONFIG_S390
1716 /*
1717  * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode.
1718  */
1719 void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
1720 {
1721         int me;
1722         int cpu = vcpu->cpu;
1723         wait_queue_head_t *wqp;
1724
1725         wqp = kvm_arch_vcpu_wq(vcpu);
1726         if (waitqueue_active(wqp)) {
1727                 wake_up_interruptible(wqp);
1728                 ++vcpu->stat.halt_wakeup;
1729         }
1730
1731         me = get_cpu();
1732         if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
1733                 if (kvm_arch_vcpu_should_kick(vcpu))
1734                         smp_send_reschedule(cpu);
1735         put_cpu();
1736 }
1737 EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
1738 #endif /* !CONFIG_S390 */
1739
1740 int kvm_vcpu_yield_to(struct kvm_vcpu *target)
1741 {
1742         struct pid *pid;
1743         struct task_struct *task = NULL;
1744         int ret = 0;
1745
1746         rcu_read_lock();
1747         pid = rcu_dereference(target->pid);
1748         if (pid)
1749                 task = get_pid_task(target->pid, PIDTYPE_PID);
1750         rcu_read_unlock();
1751         if (!task)
1752                 return ret;
1753         if (task->flags & PF_VCPU) {
1754                 put_task_struct(task);
1755                 return ret;
1756         }
1757         ret = yield_to(task, 1);
1758         put_task_struct(task);
1759
1760         return ret;
1761 }
1762 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to);
1763
1764 /*
1765  * Helper that checks whether a VCPU is eligible for directed yield.
1766  * Most eligible candidate to yield is decided by following heuristics:
1767  *
1768  *  (a) VCPU which has not done pl-exit or cpu relax intercepted recently
1769  *  (preempted lock holder), indicated by @in_spin_loop.
1770  *  Set at the beiginning and cleared at the end of interception/PLE handler.
1771  *
1772  *  (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get
1773  *  chance last time (mostly it has become eligible now since we have probably
1774  *  yielded to lockholder in last iteration. This is done by toggling
1775  *  @dy_eligible each time a VCPU checked for eligibility.)
1776  *
1777  *  Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding
1778  *  to preempted lock-holder could result in wrong VCPU selection and CPU
1779  *  burning. Giving priority for a potential lock-holder increases lock
1780  *  progress.
1781  *
1782  *  Since algorithm is based on heuristics, accessing another VCPU data without
1783  *  locking does not harm. It may result in trying to yield to  same VCPU, fail
1784  *  and continue with next VCPU and so on.
1785  */
1786 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu)
1787 {
1788 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT
1789         bool eligible;
1790
1791         eligible = !vcpu->spin_loop.in_spin_loop ||
1792                     vcpu->spin_loop.dy_eligible;
1793
1794         if (vcpu->spin_loop.in_spin_loop)
1795                 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible);
1796
1797         return eligible;
1798 #else
1799         return true;
1800 #endif
1801 }
1802
1803 void kvm_vcpu_on_spin(struct kvm_vcpu *me)
1804 {
1805         struct kvm *kvm = me->kvm;
1806         struct kvm_vcpu *vcpu;
1807         int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
1808         int yielded = 0;
1809         int try = 3;
1810         int pass;
1811         int i;
1812
1813         kvm_vcpu_set_in_spin_loop(me, true);
1814         /*
1815          * We boost the priority of a VCPU that is runnable but not
1816          * currently running, because it got preempted by something
1817          * else and called schedule in __vcpu_run.  Hopefully that
1818          * VCPU is holding the lock that we need and will release it.
1819          * We approximate round-robin by starting at the last boosted VCPU.
1820          */
1821         for (pass = 0; pass < 2 && !yielded && try; pass++) {
1822                 kvm_for_each_vcpu(i, vcpu, kvm) {
1823                         if (!pass && i <= last_boosted_vcpu) {
1824                                 i = last_boosted_vcpu;
1825                                 continue;
1826                         } else if (pass && i > last_boosted_vcpu)
1827                                 break;
1828                         if (!ACCESS_ONCE(vcpu->preempted))
1829                                 continue;
1830                         if (vcpu == me)
1831                                 continue;
1832                         if (waitqueue_active(&vcpu->wq) && !kvm_arch_vcpu_runnable(vcpu))
1833                                 continue;
1834                         if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
1835                                 continue;
1836
1837                         yielded = kvm_vcpu_yield_to(vcpu);
1838                         if (yielded > 0) {
1839                                 kvm->last_boosted_vcpu = i;
1840                                 break;
1841                         } else if (yielded < 0) {
1842                                 try--;
1843                                 if (!try)
1844                                         break;
1845                         }
1846                 }
1847         }
1848         kvm_vcpu_set_in_spin_loop(me, false);
1849
1850         /* Ensure vcpu is not eligible during next spinloop */
1851         kvm_vcpu_set_dy_eligible(me, false);
1852 }
1853 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
1854
1855 static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1856 {
1857         struct kvm_vcpu *vcpu = vma->vm_file->private_data;
1858         struct page *page;
1859
1860         if (vmf->pgoff == 0)
1861                 page = virt_to_page(vcpu->run);
1862 #ifdef CONFIG_X86
1863         else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
1864                 page = virt_to_page(vcpu->arch.pio_data);
1865 #endif
1866 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
1867         else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
1868                 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
1869 #endif
1870         else
1871                 return kvm_arch_vcpu_fault(vcpu, vmf);
1872         get_page(page);
1873         vmf->page = page;
1874         return 0;
1875 }
1876
1877 static const struct vm_operations_struct kvm_vcpu_vm_ops = {
1878         .fault = kvm_vcpu_fault,
1879 };
1880
1881 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
1882 {
1883         vma->vm_ops = &kvm_vcpu_vm_ops;
1884         return 0;
1885 }
1886
1887 static int kvm_vcpu_release(struct inode *inode, struct file *filp)
1888 {
1889         struct kvm_vcpu *vcpu = filp->private_data;
1890
1891         kvm_put_kvm(vcpu->kvm);
1892         return 0;
1893 }
1894
1895 static struct file_operations kvm_vcpu_fops = {
1896         .release        = kvm_vcpu_release,
1897         .unlocked_ioctl = kvm_vcpu_ioctl,
1898 #ifdef CONFIG_COMPAT
1899         .compat_ioctl   = kvm_vcpu_compat_ioctl,
1900 #endif
1901         .mmap           = kvm_vcpu_mmap,
1902         .llseek         = noop_llseek,
1903 };
1904
1905 /*
1906  * Allocates an inode for the vcpu.
1907  */
1908 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
1909 {
1910         return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC);
1911 }
1912
1913 /*
1914  * Creates some virtual cpus.  Good luck creating more than one.
1915  */
1916 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
1917 {
1918         int r;
1919         struct kvm_vcpu *vcpu, *v;
1920
1921         if (id >= KVM_MAX_VCPUS)
1922                 return -EINVAL;
1923
1924         vcpu = kvm_arch_vcpu_create(kvm, id);
1925         if (IS_ERR(vcpu))
1926                 return PTR_ERR(vcpu);
1927
1928         preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
1929
1930         r = kvm_arch_vcpu_setup(vcpu);
1931         if (r)
1932                 goto vcpu_destroy;
1933
1934         mutex_lock(&kvm->lock);
1935         if (!kvm_vcpu_compatible(vcpu)) {
1936                 r = -EINVAL;
1937                 goto unlock_vcpu_destroy;
1938         }
1939         if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
1940                 r = -EINVAL;
1941                 goto unlock_vcpu_destroy;
1942         }
1943
1944         kvm_for_each_vcpu(r, v, kvm)
1945                 if (v->vcpu_id == id) {
1946                         r = -EEXIST;
1947                         goto unlock_vcpu_destroy;
1948                 }
1949
1950         BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]);
1951
1952         /* Now it's all set up, let userspace reach it */
1953         kvm_get_kvm(kvm);
1954         r = create_vcpu_fd(vcpu);
1955         if (r < 0) {
1956                 kvm_put_kvm(kvm);
1957                 goto unlock_vcpu_destroy;
1958         }
1959
1960         kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu;
1961         smp_wmb();
1962         atomic_inc(&kvm->online_vcpus);
1963
1964         mutex_unlock(&kvm->lock);
1965         kvm_arch_vcpu_postcreate(vcpu);
1966         return r;
1967
1968 unlock_vcpu_destroy:
1969         mutex_unlock(&kvm->lock);
1970 vcpu_destroy:
1971         kvm_arch_vcpu_destroy(vcpu);
1972         return r;
1973 }
1974
1975 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
1976 {
1977         if (sigset) {
1978                 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
1979                 vcpu->sigset_active = 1;
1980                 vcpu->sigset = *sigset;
1981         } else
1982                 vcpu->sigset_active = 0;
1983         return 0;
1984 }
1985
1986 static long kvm_vcpu_ioctl(struct file *filp,
1987                            unsigned int ioctl, unsigned long arg)
1988 {
1989         struct kvm_vcpu *vcpu = filp->private_data;
1990         void __user *argp = (void __user *)arg;
1991         int r;
1992         struct kvm_fpu *fpu = NULL;
1993         struct kvm_sregs *kvm_sregs = NULL;
1994
1995         if (vcpu->kvm->mm != current->mm)
1996                 return -EIO;
1997
1998 #if defined(CONFIG_S390) || defined(CONFIG_PPC) || defined(CONFIG_MIPS)
1999         /*
2000          * Special cases: vcpu ioctls that are asynchronous to vcpu execution,
2001          * so vcpu_load() would break it.
2002          */
2003         if (ioctl == KVM_S390_INTERRUPT || ioctl == KVM_INTERRUPT)
2004                 return kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2005 #endif
2006
2007
2008         r = vcpu_load(vcpu);
2009         if (r)
2010                 return r;
2011         switch (ioctl) {
2012         case KVM_RUN:
2013                 r = -EINVAL;
2014                 if (arg)
2015                         goto out;
2016                 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
2017                 trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
2018                 break;
2019         case KVM_GET_REGS: {
2020                 struct kvm_regs *kvm_regs;
2021
2022                 r = -ENOMEM;
2023                 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL);
2024                 if (!kvm_regs)
2025                         goto out;
2026                 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
2027                 if (r)
2028                         goto out_free1;
2029                 r = -EFAULT;
2030                 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
2031                         goto out_free1;
2032                 r = 0;
2033 out_free1:
2034                 kfree(kvm_regs);
2035                 break;
2036         }
2037         case KVM_SET_REGS: {
2038                 struct kvm_regs *kvm_regs;
2039
2040                 r = -ENOMEM;
2041                 kvm_regs = memdup_user(argp, sizeof(*kvm_regs));
2042                 if (IS_ERR(kvm_regs)) {
2043                         r = PTR_ERR(kvm_regs);
2044                         goto out;
2045                 }
2046                 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
2047                 kfree(kvm_regs);
2048                 break;
2049         }
2050         case KVM_GET_SREGS: {
2051                 kvm_sregs = kzalloc(sizeof(struct kvm_sregs), GFP_KERNEL);
2052                 r = -ENOMEM;
2053                 if (!kvm_sregs)
2054                         goto out;
2055                 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
2056                 if (r)
2057                         goto out;
2058                 r = -EFAULT;
2059                 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
2060                         goto out;
2061                 r = 0;
2062                 break;
2063         }
2064         case KVM_SET_SREGS: {
2065                 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs));
2066                 if (IS_ERR(kvm_sregs)) {
2067                         r = PTR_ERR(kvm_sregs);
2068                         kvm_sregs = NULL;
2069                         goto out;
2070                 }
2071                 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
2072                 break;
2073         }
2074         case KVM_GET_MP_STATE: {
2075                 struct kvm_mp_state mp_state;
2076
2077                 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
2078                 if (r)
2079                         goto out;
2080                 r = -EFAULT;
2081                 if (copy_to_user(argp, &mp_state, sizeof mp_state))
2082                         goto out;
2083                 r = 0;
2084                 break;
2085         }
2086         case KVM_SET_MP_STATE: {
2087                 struct kvm_mp_state mp_state;
2088
2089                 r = -EFAULT;
2090                 if (copy_from_user(&mp_state, argp, sizeof mp_state))
2091                         goto out;
2092                 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
2093                 break;
2094         }
2095         case KVM_TRANSLATE: {
2096                 struct kvm_translation tr;
2097
2098                 r = -EFAULT;
2099                 if (copy_from_user(&tr, argp, sizeof tr))
2100                         goto out;
2101                 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
2102                 if (r)
2103                         goto out;
2104                 r = -EFAULT;
2105                 if (copy_to_user(argp, &tr, sizeof tr))
2106                         goto out;
2107                 r = 0;
2108                 break;
2109         }
2110         case KVM_SET_GUEST_DEBUG: {
2111                 struct kvm_guest_debug dbg;
2112
2113                 r = -EFAULT;
2114                 if (copy_from_user(&dbg, argp, sizeof dbg))
2115                         goto out;
2116                 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
2117                 break;
2118         }
2119         case KVM_SET_SIGNAL_MASK: {
2120                 struct kvm_signal_mask __user *sigmask_arg = argp;
2121                 struct kvm_signal_mask kvm_sigmask;
2122                 sigset_t sigset, *p;
2123
2124                 p = NULL;
2125                 if (argp) {
2126                         r = -EFAULT;
2127                         if (copy_from_user(&kvm_sigmask, argp,
2128                                            sizeof kvm_sigmask))
2129                                 goto out;
2130                         r = -EINVAL;
2131                         if (kvm_sigmask.len != sizeof sigset)
2132                                 goto out;
2133                         r = -EFAULT;
2134                         if (copy_from_user(&sigset, sigmask_arg->sigset,
2135                                            sizeof sigset))
2136                                 goto out;
2137                         p = &sigset;
2138                 }
2139                 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
2140                 break;
2141         }
2142         case KVM_GET_FPU: {
2143                 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL);
2144                 r = -ENOMEM;
2145                 if (!fpu)
2146                         goto out;
2147                 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
2148                 if (r)
2149                         goto out;
2150                 r = -EFAULT;
2151                 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
2152                         goto out;
2153                 r = 0;
2154                 break;
2155         }
2156         case KVM_SET_FPU: {
2157                 fpu = memdup_user(argp, sizeof(*fpu));
2158                 if (IS_ERR(fpu)) {
2159                         r = PTR_ERR(fpu);
2160                         fpu = NULL;
2161                         goto out;
2162                 }
2163                 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
2164                 break;
2165         }
2166         default:
2167                 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2168         }
2169 out:
2170         vcpu_put(vcpu);
2171         kfree(fpu);
2172         kfree(kvm_sregs);
2173         return r;
2174 }
2175
2176 #ifdef CONFIG_COMPAT
2177 static long kvm_vcpu_compat_ioctl(struct file *filp,
2178                                   unsigned int ioctl, unsigned long arg)
2179 {
2180         struct kvm_vcpu *vcpu = filp->private_data;
2181         void __user *argp = compat_ptr(arg);
2182         int r;
2183
2184         if (vcpu->kvm->mm != current->mm)
2185                 return -EIO;
2186
2187         switch (ioctl) {
2188         case KVM_SET_SIGNAL_MASK: {
2189                 struct kvm_signal_mask __user *sigmask_arg = argp;
2190                 struct kvm_signal_mask kvm_sigmask;
2191                 compat_sigset_t csigset;
2192                 sigset_t sigset;
2193
2194                 if (argp) {
2195                         r = -EFAULT;
2196                         if (copy_from_user(&kvm_sigmask, argp,
2197                                            sizeof kvm_sigmask))
2198                                 goto out;
2199                         r = -EINVAL;
2200                         if (kvm_sigmask.len != sizeof csigset)
2201                                 goto out;
2202                         r = -EFAULT;
2203                         if (copy_from_user(&csigset, sigmask_arg->sigset,
2204                                            sizeof csigset))
2205                                 goto out;
2206                         sigset_from_compat(&sigset, &csigset);
2207                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
2208                 } else
2209                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL);
2210                 break;
2211         }
2212         default:
2213                 r = kvm_vcpu_ioctl(filp, ioctl, arg);
2214         }
2215
2216 out:
2217         return r;
2218 }
2219 #endif
2220
2221 static int kvm_device_ioctl_attr(struct kvm_device *dev,
2222                                  int (*accessor)(struct kvm_device *dev,
2223                                                  struct kvm_device_attr *attr),
2224                                  unsigned long arg)
2225 {
2226         struct kvm_device_attr attr;
2227
2228         if (!accessor)
2229                 return -EPERM;
2230
2231         if (copy_from_user(&attr, (void __user *)arg, sizeof(attr)))
2232                 return -EFAULT;
2233
2234         return accessor(dev, &attr);
2235 }
2236
2237 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl,
2238                              unsigned long arg)
2239 {
2240         struct kvm_device *dev = filp->private_data;
2241
2242         switch (ioctl) {
2243         case KVM_SET_DEVICE_ATTR:
2244                 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg);
2245         case KVM_GET_DEVICE_ATTR:
2246                 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg);
2247         case KVM_HAS_DEVICE_ATTR:
2248                 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg);
2249         default:
2250                 if (dev->ops->ioctl)
2251                         return dev->ops->ioctl(dev, ioctl, arg);
2252
2253                 return -ENOTTY;
2254         }
2255 }
2256
2257 static int kvm_device_release(struct inode *inode, struct file *filp)
2258 {
2259         struct kvm_device *dev = filp->private_data;
2260         struct kvm *kvm = dev->kvm;
2261
2262         kvm_put_kvm(kvm);
2263         return 0;
2264 }
2265
2266 static const struct file_operations kvm_device_fops = {
2267         .unlocked_ioctl = kvm_device_ioctl,
2268 #ifdef CONFIG_COMPAT
2269         .compat_ioctl = kvm_device_ioctl,
2270 #endif
2271         .release = kvm_device_release,
2272 };
2273
2274 struct kvm_device *kvm_device_from_filp(struct file *filp)
2275 {
2276         if (filp->f_op != &kvm_device_fops)
2277                 return NULL;
2278
2279         return filp->private_data;
2280 }
2281
2282 static struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = {
2283 #ifdef CONFIG_KVM_MPIC
2284         [KVM_DEV_TYPE_FSL_MPIC_20]      = &kvm_mpic_ops,
2285         [KVM_DEV_TYPE_FSL_MPIC_42]      = &kvm_mpic_ops,
2286 #endif
2287
2288 #ifdef CONFIG_KVM_XICS
2289         [KVM_DEV_TYPE_XICS]             = &kvm_xics_ops,
2290 #endif
2291 };
2292
2293 int kvm_register_device_ops(struct kvm_device_ops *ops, u32 type)
2294 {
2295         if (type >= ARRAY_SIZE(kvm_device_ops_table))
2296                 return -ENOSPC;
2297
2298         if (kvm_device_ops_table[type] != NULL)
2299                 return -EEXIST;
2300
2301         kvm_device_ops_table[type] = ops;
2302         return 0;
2303 }
2304
2305 static int kvm_ioctl_create_device(struct kvm *kvm,
2306                                    struct kvm_create_device *cd)
2307 {
2308         struct kvm_device_ops *ops = NULL;
2309         struct kvm_device *dev;
2310         bool test = cd->flags & KVM_CREATE_DEVICE_TEST;
2311         int ret;
2312
2313         if (cd->type >= ARRAY_SIZE(kvm_device_ops_table))
2314                 return -ENODEV;
2315
2316         ops = kvm_device_ops_table[cd->type];
2317         if (ops == NULL)
2318                 return -ENODEV;
2319
2320         if (test)
2321                 return 0;
2322
2323         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2324         if (!dev)
2325                 return -ENOMEM;
2326
2327         dev->ops = ops;
2328         dev->kvm = kvm;
2329
2330         ret = ops->create(dev, cd->type);
2331         if (ret < 0) {
2332                 kfree(dev);
2333                 return ret;
2334         }
2335
2336         ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC);
2337         if (ret < 0) {
2338                 ops->destroy(dev);
2339                 return ret;
2340         }
2341
2342         list_add(&dev->vm_node, &kvm->devices);
2343         kvm_get_kvm(kvm);
2344         cd->fd = ret;
2345         return 0;
2346 }
2347
2348 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
2349 {
2350         switch (arg) {
2351         case KVM_CAP_USER_MEMORY:
2352         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
2353         case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
2354 #ifdef CONFIG_KVM_APIC_ARCHITECTURE
2355         case KVM_CAP_SET_BOOT_CPU_ID:
2356 #endif
2357         case KVM_CAP_INTERNAL_ERROR_DATA:
2358 #ifdef CONFIG_HAVE_KVM_MSI
2359         case KVM_CAP_SIGNAL_MSI:
2360 #endif
2361 #ifdef CONFIG_HAVE_KVM_IRQFD
2362         case KVM_CAP_IRQFD_RESAMPLE:
2363 #endif
2364         case KVM_CAP_CHECK_EXTENSION_VM:
2365                 return 1;
2366 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2367         case KVM_CAP_IRQ_ROUTING:
2368                 return KVM_MAX_IRQ_ROUTES;
2369 #endif
2370         default:
2371                 break;
2372         }
2373         return kvm_vm_ioctl_check_extension(kvm, arg);
2374 }
2375
2376 static long kvm_vm_ioctl(struct file *filp,
2377                            unsigned int ioctl, unsigned long arg)
2378 {
2379         struct kvm *kvm = filp->private_data;
2380         void __user *argp = (void __user *)arg;
2381         int r;
2382
2383         if (kvm->mm != current->mm)
2384                 return -EIO;
2385         switch (ioctl) {
2386         case KVM_CREATE_VCPU:
2387                 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
2388                 break;
2389         case KVM_SET_USER_MEMORY_REGION: {
2390                 struct kvm_userspace_memory_region kvm_userspace_mem;
2391
2392                 r = -EFAULT;
2393                 if (copy_from_user(&kvm_userspace_mem, argp,
2394                                                 sizeof kvm_userspace_mem))
2395                         goto out;
2396
2397                 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem);
2398                 break;
2399         }
2400         case KVM_GET_DIRTY_LOG: {
2401                 struct kvm_dirty_log log;
2402
2403                 r = -EFAULT;
2404                 if (copy_from_user(&log, argp, sizeof log))
2405                         goto out;
2406                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2407                 break;
2408         }
2409 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2410         case KVM_REGISTER_COALESCED_MMIO: {
2411                 struct kvm_coalesced_mmio_zone zone;
2412                 r = -EFAULT;
2413                 if (copy_from_user(&zone, argp, sizeof zone))
2414                         goto out;
2415                 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
2416                 break;
2417         }
2418         case KVM_UNREGISTER_COALESCED_MMIO: {
2419                 struct kvm_coalesced_mmio_zone zone;
2420                 r = -EFAULT;
2421                 if (copy_from_user(&zone, argp, sizeof zone))
2422                         goto out;
2423                 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
2424                 break;
2425         }
2426 #endif
2427         case KVM_IRQFD: {
2428                 struct kvm_irqfd data;
2429
2430                 r = -EFAULT;
2431                 if (copy_from_user(&data, argp, sizeof data))
2432                         goto out;
2433                 r = kvm_irqfd(kvm, &data);
2434                 break;
2435         }
2436         case KVM_IOEVENTFD: {
2437                 struct kvm_ioeventfd data;
2438
2439                 r = -EFAULT;
2440                 if (copy_from_user(&data, argp, sizeof data))
2441                         goto out;
2442                 r = kvm_ioeventfd(kvm, &data);
2443                 break;
2444         }
2445 #ifdef CONFIG_KVM_APIC_ARCHITECTURE
2446         case KVM_SET_BOOT_CPU_ID:
2447                 r = 0;
2448                 mutex_lock(&kvm->lock);
2449                 if (atomic_read(&kvm->online_vcpus) != 0)
2450                         r = -EBUSY;
2451                 else
2452                         kvm->bsp_vcpu_id = arg;
2453                 mutex_unlock(&kvm->lock);
2454                 break;
2455 #endif
2456 #ifdef CONFIG_HAVE_KVM_MSI
2457         case KVM_SIGNAL_MSI: {
2458                 struct kvm_msi msi;
2459
2460                 r = -EFAULT;
2461                 if (copy_from_user(&msi, argp, sizeof msi))
2462                         goto out;
2463                 r = kvm_send_userspace_msi(kvm, &msi);
2464                 break;
2465         }
2466 #endif
2467 #ifdef __KVM_HAVE_IRQ_LINE
2468         case KVM_IRQ_LINE_STATUS:
2469         case KVM_IRQ_LINE: {
2470                 struct kvm_irq_level irq_event;
2471
2472                 r = -EFAULT;
2473                 if (copy_from_user(&irq_event, argp, sizeof irq_event))
2474                         goto out;
2475
2476                 r = kvm_vm_ioctl_irq_line(kvm, &irq_event,
2477                                         ioctl == KVM_IRQ_LINE_STATUS);
2478                 if (r)
2479                         goto out;
2480
2481                 r = -EFAULT;
2482                 if (ioctl == KVM_IRQ_LINE_STATUS) {
2483                         if (copy_to_user(argp, &irq_event, sizeof irq_event))
2484                                 goto out;
2485                 }
2486
2487                 r = 0;
2488                 break;
2489         }
2490 #endif
2491 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2492         case KVM_SET_GSI_ROUTING: {
2493                 struct kvm_irq_routing routing;
2494                 struct kvm_irq_routing __user *urouting;
2495                 struct kvm_irq_routing_entry *entries;
2496
2497                 r = -EFAULT;
2498                 if (copy_from_user(&routing, argp, sizeof(routing)))
2499                         goto out;
2500                 r = -EINVAL;
2501                 if (routing.nr >= KVM_MAX_IRQ_ROUTES)
2502                         goto out;
2503                 if (routing.flags)
2504                         goto out;
2505                 r = -ENOMEM;
2506                 entries = vmalloc(routing.nr * sizeof(*entries));
2507                 if (!entries)
2508                         goto out;
2509                 r = -EFAULT;
2510                 urouting = argp;
2511                 if (copy_from_user(entries, urouting->entries,
2512                                    routing.nr * sizeof(*entries)))
2513                         goto out_free_irq_routing;
2514                 r = kvm_set_irq_routing(kvm, entries, routing.nr,
2515                                         routing.flags);
2516         out_free_irq_routing:
2517                 vfree(entries);
2518                 break;
2519         }
2520 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */
2521         case KVM_CREATE_DEVICE: {
2522                 struct kvm_create_device cd;
2523
2524                 r = -EFAULT;
2525                 if (copy_from_user(&cd, argp, sizeof(cd)))
2526                         goto out;
2527
2528                 r = kvm_ioctl_create_device(kvm, &cd);
2529                 if (r)
2530                         goto out;
2531
2532                 r = -EFAULT;
2533                 if (copy_to_user(argp, &cd, sizeof(cd)))
2534                         goto out;
2535
2536                 r = 0;
2537                 break;
2538         }
2539         case KVM_CHECK_EXTENSION:
2540                 r = kvm_vm_ioctl_check_extension_generic(kvm, arg);
2541                 break;
2542         default:
2543                 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
2544                 if (r == -ENOTTY)
2545                         r = kvm_vm_ioctl_assigned_device(kvm, ioctl, arg);
2546         }
2547 out:
2548         return r;
2549 }
2550
2551 #ifdef CONFIG_COMPAT
2552 struct compat_kvm_dirty_log {
2553         __u32 slot;
2554         __u32 padding1;
2555         union {
2556                 compat_uptr_t dirty_bitmap; /* one bit per page */
2557                 __u64 padding2;
2558         };
2559 };
2560
2561 static long kvm_vm_compat_ioctl(struct file *filp,
2562                            unsigned int ioctl, unsigned long arg)
2563 {
2564         struct kvm *kvm = filp->private_data;
2565         int r;
2566
2567         if (kvm->mm != current->mm)
2568                 return -EIO;
2569         switch (ioctl) {
2570         case KVM_GET_DIRTY_LOG: {
2571                 struct compat_kvm_dirty_log compat_log;
2572                 struct kvm_dirty_log log;
2573
2574                 r = -EFAULT;
2575                 if (copy_from_user(&compat_log, (void __user *)arg,
2576                                    sizeof(compat_log)))
2577                         goto out;
2578                 log.slot         = compat_log.slot;
2579                 log.padding1     = compat_log.padding1;
2580                 log.padding2     = compat_log.padding2;
2581                 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
2582
2583                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2584                 break;
2585         }
2586         default:
2587                 r = kvm_vm_ioctl(filp, ioctl, arg);
2588         }
2589
2590 out:
2591         return r;
2592 }
2593 #endif
2594
2595 static struct file_operations kvm_vm_fops = {
2596         .release        = kvm_vm_release,
2597         .unlocked_ioctl = kvm_vm_ioctl,
2598 #ifdef CONFIG_COMPAT
2599         .compat_ioctl   = kvm_vm_compat_ioctl,
2600 #endif
2601         .llseek         = noop_llseek,
2602 };
2603
2604 static int kvm_dev_ioctl_create_vm(unsigned long type)
2605 {
2606         int r;
2607         struct kvm *kvm;
2608
2609         kvm = kvm_create_vm(type);
2610         if (IS_ERR(kvm))
2611                 return PTR_ERR(kvm);
2612 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2613         r = kvm_coalesced_mmio_init(kvm);
2614         if (r < 0) {
2615                 kvm_put_kvm(kvm);
2616                 return r;
2617         }
2618 #endif
2619         r = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, O_RDWR | O_CLOEXEC);
2620         if (r < 0)
2621                 kvm_put_kvm(kvm);
2622
2623         return r;
2624 }
2625
2626 static long kvm_dev_ioctl(struct file *filp,
2627                           unsigned int ioctl, unsigned long arg)
2628 {
2629         long r = -EINVAL;
2630
2631         switch (ioctl) {
2632         case KVM_GET_API_VERSION:
2633                 if (arg)
2634                         goto out;
2635                 r = KVM_API_VERSION;
2636                 break;
2637         case KVM_CREATE_VM:
2638                 r = kvm_dev_ioctl_create_vm(arg);
2639                 break;
2640         case KVM_CHECK_EXTENSION:
2641                 r = kvm_vm_ioctl_check_extension_generic(NULL, arg);
2642                 break;
2643         case KVM_GET_VCPU_MMAP_SIZE:
2644                 if (arg)
2645                         goto out;
2646                 r = PAGE_SIZE;     /* struct kvm_run */
2647 #ifdef CONFIG_X86
2648                 r += PAGE_SIZE;    /* pio data page */
2649 #endif
2650 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2651                 r += PAGE_SIZE;    /* coalesced mmio ring page */
2652 #endif
2653                 break;
2654         case KVM_TRACE_ENABLE:
2655         case KVM_TRACE_PAUSE:
2656         case KVM_TRACE_DISABLE:
2657                 r = -EOPNOTSUPP;
2658                 break;
2659         default:
2660                 return kvm_arch_dev_ioctl(filp, ioctl, arg);
2661         }
2662 out:
2663         return r;
2664 }
2665
2666 static struct file_operations kvm_chardev_ops = {
2667         .unlocked_ioctl = kvm_dev_ioctl,
2668         .compat_ioctl   = kvm_dev_ioctl,
2669         .llseek         = noop_llseek,
2670 };
2671
2672 static struct miscdevice kvm_dev = {
2673         KVM_MINOR,
2674         "kvm",
2675         &kvm_chardev_ops,
2676 };
2677
2678 static void hardware_enable_nolock(void *junk)
2679 {
2680         int cpu = raw_smp_processor_id();
2681         int r;
2682
2683         if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
2684                 return;
2685
2686         cpumask_set_cpu(cpu, cpus_hardware_enabled);
2687
2688         r = kvm_arch_hardware_enable();
2689
2690         if (r) {
2691                 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
2692                 atomic_inc(&hardware_enable_failed);
2693                 printk(KERN_INFO "kvm: enabling virtualization on "
2694                                  "CPU%d failed\n", cpu);
2695         }
2696 }
2697
2698 static void hardware_enable(void)
2699 {
2700         raw_spin_lock(&kvm_count_lock);
2701         if (kvm_usage_count)
2702                 hardware_enable_nolock(NULL);
2703         raw_spin_unlock(&kvm_count_lock);
2704 }
2705
2706 static void hardware_disable_nolock(void *junk)
2707 {
2708         int cpu = raw_smp_processor_id();
2709
2710         if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
2711                 return;
2712         cpumask_clear_cpu(cpu, cpus_hardware_enabled);
2713         kvm_arch_hardware_disable();
2714 }
2715
2716 static void hardware_disable(void)
2717 {
2718         raw_spin_lock(&kvm_count_lock);
2719         if (kvm_usage_count)
2720                 hardware_disable_nolock(NULL);
2721         raw_spin_unlock(&kvm_count_lock);
2722 }
2723
2724 static void hardware_disable_all_nolock(void)
2725 {
2726         BUG_ON(!kvm_usage_count);
2727
2728         kvm_usage_count--;
2729         if (!kvm_usage_count)
2730                 on_each_cpu(hardware_disable_nolock, NULL, 1);
2731 }
2732
2733 static void hardware_disable_all(void)
2734 {
2735         raw_spin_lock(&kvm_count_lock);
2736         hardware_disable_all_nolock();
2737         raw_spin_unlock(&kvm_count_lock);
2738 }
2739
2740 static int hardware_enable_all(void)
2741 {
2742         int r = 0;
2743
2744         raw_spin_lock(&kvm_count_lock);
2745
2746         kvm_usage_count++;
2747         if (kvm_usage_count == 1) {
2748                 atomic_set(&hardware_enable_failed, 0);
2749                 on_each_cpu(hardware_enable_nolock, NULL, 1);
2750
2751                 if (atomic_read(&hardware_enable_failed)) {
2752                         hardware_disable_all_nolock();
2753                         r = -EBUSY;
2754                 }
2755         }
2756
2757         raw_spin_unlock(&kvm_count_lock);
2758
2759         return r;
2760 }
2761
2762 static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
2763                            void *v)
2764 {
2765         int cpu = (long)v;
2766
2767         val &= ~CPU_TASKS_FROZEN;
2768         switch (val) {
2769         case CPU_DYING:
2770                 printk(KERN_INFO "kvm: disabling virtualization on CPU%d\n",
2771                        cpu);
2772                 hardware_disable();
2773                 break;
2774         case CPU_STARTING:
2775                 printk(KERN_INFO "kvm: enabling virtualization on CPU%d\n",
2776                        cpu);
2777                 hardware_enable();
2778                 break;
2779         }
2780         return NOTIFY_OK;
2781 }
2782
2783 static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
2784                       void *v)
2785 {
2786         /*
2787          * Some (well, at least mine) BIOSes hang on reboot if
2788          * in vmx root mode.
2789          *
2790          * And Intel TXT required VMX off for all cpu when system shutdown.
2791          */
2792         printk(KERN_INFO "kvm: exiting hardware virtualization\n");
2793         kvm_rebooting = true;
2794         on_each_cpu(hardware_disable_nolock, NULL, 1);
2795         return NOTIFY_OK;
2796 }
2797
2798 static struct notifier_block kvm_reboot_notifier = {
2799         .notifier_call = kvm_reboot,
2800         .priority = 0,
2801 };
2802
2803 static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
2804 {
2805         int i;
2806
2807         for (i = 0; i < bus->dev_count; i++) {
2808                 struct kvm_io_device *pos = bus->range[i].dev;
2809
2810                 kvm_iodevice_destructor(pos);
2811         }
2812         kfree(bus);
2813 }
2814
2815 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1,
2816                                  const struct kvm_io_range *r2)
2817 {
2818         if (r1->addr < r2->addr)
2819                 return -1;
2820         if (r1->addr + r1->len > r2->addr + r2->len)
2821                 return 1;
2822         return 0;
2823 }
2824
2825 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
2826 {
2827         return kvm_io_bus_cmp(p1, p2);
2828 }
2829
2830 static int kvm_io_bus_insert_dev(struct kvm_io_bus *bus, struct kvm_io_device *dev,
2831                           gpa_t addr, int len)
2832 {
2833         bus->range[bus->dev_count++] = (struct kvm_io_range) {
2834                 .addr = addr,
2835                 .len = len,
2836                 .dev = dev,
2837         };
2838
2839         sort(bus->range, bus->dev_count, sizeof(struct kvm_io_range),
2840                 kvm_io_bus_sort_cmp, NULL);
2841
2842         return 0;
2843 }
2844
2845 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
2846                              gpa_t addr, int len)
2847 {
2848         struct kvm_io_range *range, key;
2849         int off;
2850
2851         key = (struct kvm_io_range) {
2852                 .addr = addr,
2853                 .len = len,
2854         };
2855
2856         range = bsearch(&key, bus->range, bus->dev_count,
2857                         sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
2858         if (range == NULL)
2859                 return -ENOENT;
2860
2861         off = range - bus->range;
2862
2863         while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0)
2864                 off--;
2865
2866         return off;
2867 }
2868
2869 static int __kvm_io_bus_write(struct kvm_io_bus *bus,
2870                               struct kvm_io_range *range, const void *val)
2871 {
2872         int idx;
2873
2874         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
2875         if (idx < 0)
2876                 return -EOPNOTSUPP;
2877
2878         while (idx < bus->dev_count &&
2879                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
2880                 if (!kvm_iodevice_write(bus->range[idx].dev, range->addr,
2881                                         range->len, val))
2882                         return idx;
2883                 idx++;
2884         }
2885
2886         return -EOPNOTSUPP;
2887 }
2888
2889 /* kvm_io_bus_write - called under kvm->slots_lock */
2890 int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2891                      int len, const void *val)
2892 {
2893         struct kvm_io_bus *bus;
2894         struct kvm_io_range range;
2895         int r;
2896
2897         range = (struct kvm_io_range) {
2898                 .addr = addr,
2899                 .len = len,
2900         };
2901
2902         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
2903         r = __kvm_io_bus_write(bus, &range, val);
2904         return r < 0 ? r : 0;
2905 }
2906
2907 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */
2908 int kvm_io_bus_write_cookie(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2909                             int len, const void *val, long cookie)
2910 {
2911         struct kvm_io_bus *bus;
2912         struct kvm_io_range range;
2913
2914         range = (struct kvm_io_range) {
2915                 .addr = addr,
2916                 .len = len,
2917         };
2918
2919         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
2920
2921         /* First try the device referenced by cookie. */
2922         if ((cookie >= 0) && (cookie < bus->dev_count) &&
2923             (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0))
2924                 if (!kvm_iodevice_write(bus->range[cookie].dev, addr, len,
2925                                         val))
2926                         return cookie;
2927
2928         /*
2929          * cookie contained garbage; fall back to search and return the
2930          * correct cookie value.
2931          */
2932         return __kvm_io_bus_write(bus, &range, val);
2933 }
2934
2935 static int __kvm_io_bus_read(struct kvm_io_bus *bus, struct kvm_io_range *range,
2936                              void *val)
2937 {
2938         int idx;
2939
2940         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
2941         if (idx < 0)
2942                 return -EOPNOTSUPP;
2943
2944         while (idx < bus->dev_count &&
2945                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
2946                 if (!kvm_iodevice_read(bus->range[idx].dev, range->addr,
2947                                        range->len, val))
2948                         return idx;
2949                 idx++;
2950         }
2951
2952         return -EOPNOTSUPP;
2953 }
2954 EXPORT_SYMBOL_GPL(kvm_io_bus_write);
2955
2956 /* kvm_io_bus_read - called under kvm->slots_lock */
2957 int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2958                     int len, void *val)
2959 {
2960         struct kvm_io_bus *bus;
2961         struct kvm_io_range range;
2962         int r;
2963
2964         range = (struct kvm_io_range) {
2965                 .addr = addr,
2966                 .len = len,
2967         };
2968
2969         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
2970         r = __kvm_io_bus_read(bus, &range, val);
2971         return r < 0 ? r : 0;
2972 }
2973
2974
2975 /* Caller must hold slots_lock. */
2976 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2977                             int len, struct kvm_io_device *dev)
2978 {
2979         struct kvm_io_bus *new_bus, *bus;
2980
2981         bus = kvm->buses[bus_idx];
2982         /* exclude ioeventfd which is limited by maximum fd */
2983         if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
2984                 return -ENOSPC;
2985
2986         new_bus = kzalloc(sizeof(*bus) + ((bus->dev_count + 1) *
2987                           sizeof(struct kvm_io_range)), GFP_KERNEL);
2988         if (!new_bus)
2989                 return -ENOMEM;
2990         memcpy(new_bus, bus, sizeof(*bus) + (bus->dev_count *
2991                sizeof(struct kvm_io_range)));
2992         kvm_io_bus_insert_dev(new_bus, dev, addr, len);
2993         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
2994         synchronize_srcu_expedited(&kvm->srcu);
2995         kfree(bus);
2996
2997         return 0;
2998 }
2999
3000 /* Caller must hold slots_lock. */
3001 int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
3002                               struct kvm_io_device *dev)
3003 {
3004         int i, r;
3005         struct kvm_io_bus *new_bus, *bus;
3006
3007         bus = kvm->buses[bus_idx];
3008         r = -ENOENT;
3009         for (i = 0; i < bus->dev_count; i++)
3010                 if (bus->range[i].dev == dev) {
3011                         r = 0;
3012                         break;
3013                 }
3014
3015         if (r)
3016                 return r;
3017
3018         new_bus = kzalloc(sizeof(*bus) + ((bus->dev_count - 1) *
3019                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3020         if (!new_bus)
3021                 return -ENOMEM;
3022
3023         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
3024         new_bus->dev_count--;
3025         memcpy(new_bus->range + i, bus->range + i + 1,
3026                (new_bus->dev_count - i) * sizeof(struct kvm_io_range));
3027
3028         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3029         synchronize_srcu_expedited(&kvm->srcu);
3030         kfree(bus);
3031         return r;
3032 }
3033
3034 static struct notifier_block kvm_cpu_notifier = {
3035         .notifier_call = kvm_cpu_hotplug,
3036 };
3037
3038 static int vm_stat_get(void *_offset, u64 *val)
3039 {
3040         unsigned offset = (long)_offset;
3041         struct kvm *kvm;
3042
3043         *val = 0;
3044         spin_lock(&kvm_lock);
3045         list_for_each_entry(kvm, &vm_list, vm_list)
3046                 *val += *(u32 *)((void *)kvm + offset);
3047         spin_unlock(&kvm_lock);
3048         return 0;
3049 }
3050
3051 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, NULL, "%llu\n");
3052
3053 static int vcpu_stat_get(void *_offset, u64 *val)
3054 {
3055         unsigned offset = (long)_offset;
3056         struct kvm *kvm;
3057         struct kvm_vcpu *vcpu;
3058         int i;
3059
3060         *val = 0;
3061         spin_lock(&kvm_lock);
3062         list_for_each_entry(kvm, &vm_list, vm_list)
3063                 kvm_for_each_vcpu(i, vcpu, kvm)
3064                         *val += *(u32 *)((void *)vcpu + offset);
3065
3066         spin_unlock(&kvm_lock);
3067         return 0;
3068 }
3069
3070 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n");
3071
3072 static const struct file_operations *stat_fops[] = {
3073         [KVM_STAT_VCPU] = &vcpu_stat_fops,
3074         [KVM_STAT_VM]   = &vm_stat_fops,
3075 };
3076
3077 static int kvm_init_debug(void)
3078 {
3079         int r = -EEXIST;
3080         struct kvm_stats_debugfs_item *p;
3081
3082         kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
3083         if (kvm_debugfs_dir == NULL)
3084                 goto out;
3085
3086         for (p = debugfs_entries; p->name; ++p) {
3087                 p->dentry = debugfs_create_file(p->name, 0444, kvm_debugfs_dir,
3088                                                 (void *)(long)p->offset,
3089                                                 stat_fops[p->kind]);
3090                 if (p->dentry == NULL)
3091                         goto out_dir;
3092         }
3093
3094         return 0;
3095
3096 out_dir:
3097         debugfs_remove_recursive(kvm_debugfs_dir);
3098 out:
3099         return r;
3100 }
3101
3102 static void kvm_exit_debug(void)
3103 {
3104         struct kvm_stats_debugfs_item *p;
3105
3106         for (p = debugfs_entries; p->name; ++p)
3107                 debugfs_remove(p->dentry);
3108         debugfs_remove(kvm_debugfs_dir);
3109 }
3110
3111 static int kvm_suspend(void)
3112 {
3113         if (kvm_usage_count)
3114                 hardware_disable_nolock(NULL);
3115         return 0;
3116 }
3117
3118 static void kvm_resume(void)
3119 {
3120         if (kvm_usage_count) {
3121                 WARN_ON(raw_spin_is_locked(&kvm_count_lock));
3122                 hardware_enable_nolock(NULL);
3123         }
3124 }
3125
3126 static struct syscore_ops kvm_syscore_ops = {
3127         .suspend = kvm_suspend,
3128         .resume = kvm_resume,
3129 };
3130
3131 static inline
3132 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
3133 {
3134         return container_of(pn, struct kvm_vcpu, preempt_notifier);
3135 }
3136
3137 static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
3138 {
3139         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3140         if (vcpu->preempted)
3141                 vcpu->preempted = false;
3142
3143         kvm_arch_sched_in(vcpu, cpu);
3144
3145         kvm_arch_vcpu_load(vcpu, cpu);
3146 }
3147
3148 static void kvm_sched_out(struct preempt_notifier *pn,
3149                           struct task_struct *next)
3150 {
3151         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3152
3153         if (current->state == TASK_RUNNING)
3154                 vcpu->preempted = true;
3155         kvm_arch_vcpu_put(vcpu);
3156 }
3157
3158 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
3159                   struct module *module)
3160 {
3161         int r;
3162         int cpu;
3163
3164         r = kvm_arch_init(opaque);
3165         if (r)
3166                 goto out_fail;
3167
3168         /*
3169          * kvm_arch_init makes sure there's at most one caller
3170          * for architectures that support multiple implementations,
3171          * like intel and amd on x86.
3172          * kvm_arch_init must be called before kvm_irqfd_init to avoid creating
3173          * conflicts in case kvm is already setup for another implementation.
3174          */
3175         r = kvm_irqfd_init();
3176         if (r)
3177                 goto out_irqfd;
3178
3179         if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
3180                 r = -ENOMEM;
3181                 goto out_free_0;
3182         }
3183
3184         r = kvm_arch_hardware_setup();
3185         if (r < 0)
3186                 goto out_free_0a;
3187
3188         for_each_online_cpu(cpu) {
3189                 smp_call_function_single(cpu,
3190                                 kvm_arch_check_processor_compat,
3191                                 &r, 1);
3192                 if (r < 0)
3193                         goto out_free_1;
3194         }
3195
3196         r = register_cpu_notifier(&kvm_cpu_notifier);
3197         if (r)
3198                 goto out_free_2;
3199         register_reboot_notifier(&kvm_reboot_notifier);
3200
3201         /* A kmem cache lets us meet the alignment requirements of fx_save. */
3202         if (!vcpu_align)
3203                 vcpu_align = __alignof__(struct kvm_vcpu);
3204         kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size, vcpu_align,
3205                                            0, NULL);
3206         if (!kvm_vcpu_cache) {
3207                 r = -ENOMEM;
3208                 goto out_free_3;
3209         }
3210
3211         r = kvm_async_pf_init();
3212         if (r)
3213                 goto out_free;
3214
3215         kvm_chardev_ops.owner = module;
3216         kvm_vm_fops.owner = module;
3217         kvm_vcpu_fops.owner = module;
3218
3219         r = misc_register(&kvm_dev);
3220         if (r) {
3221                 printk(KERN_ERR "kvm: misc device register failed\n");
3222                 goto out_unreg;
3223         }
3224
3225         register_syscore_ops(&kvm_syscore_ops);
3226
3227         kvm_preempt_ops.sched_in = kvm_sched_in;
3228         kvm_preempt_ops.sched_out = kvm_sched_out;
3229
3230         r = kvm_init_debug();
3231         if (r) {
3232                 printk(KERN_ERR "kvm: create debugfs files failed\n");
3233                 goto out_undebugfs;
3234         }
3235
3236         return 0;
3237
3238 out_undebugfs:
3239         unregister_syscore_ops(&kvm_syscore_ops);
3240         misc_deregister(&kvm_dev);
3241 out_unreg:
3242         kvm_async_pf_deinit();
3243 out_free:
3244         kmem_cache_destroy(kvm_vcpu_cache);
3245 out_free_3:
3246         unregister_reboot_notifier(&kvm_reboot_notifier);
3247         unregister_cpu_notifier(&kvm_cpu_notifier);
3248 out_free_2:
3249 out_free_1:
3250         kvm_arch_hardware_unsetup();
3251 out_free_0a:
3252         free_cpumask_var(cpus_hardware_enabled);
3253 out_free_0:
3254         kvm_irqfd_exit();
3255 out_irqfd:
3256         kvm_arch_exit();
3257 out_fail:
3258         return r;
3259 }
3260 EXPORT_SYMBOL_GPL(kvm_init);
3261
3262 void kvm_exit(void)
3263 {
3264         kvm_exit_debug();
3265         misc_deregister(&kvm_dev);
3266         kmem_cache_destroy(kvm_vcpu_cache);
3267         kvm_async_pf_deinit();
3268         unregister_syscore_ops(&kvm_syscore_ops);
3269         unregister_reboot_notifier(&kvm_reboot_notifier);
3270         unregister_cpu_notifier(&kvm_cpu_notifier);
3271         on_each_cpu(hardware_disable_nolock, NULL, 1);
3272         kvm_arch_hardware_unsetup();
3273         kvm_arch_exit();
3274         kvm_irqfd_exit();
3275         free_cpumask_var(cpus_hardware_enabled);
3276 }
3277 EXPORT_SYMBOL_GPL(kvm_exit);