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 <kvm/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/ioctl.h>
  56 #include <asm/uaccess.h>
  57 #include <asm/pgtable.h>
  58
  59 #include "coalesced_mmio.h"
  60 #include "async_pf.h"
  61 #include "vfio.h"
  62
  63 #define CREATE_TRACE_POINTS
  64 #include <trace/events/kvm.h>
  65
  66 /* Worst case buffer size needed for holding an integer. */
  67 #define ITOA_MAX_LEN 12
  68
  69 MODULE_AUTHOR("Qumranet");
  70 MODULE_LICENSE("GPL");
  71
  72 /* Architectures should define their poll value according to the halt latency */
  73 static unsigned int halt_poll_ns = KVM_HALT_POLL_NS_DEFAULT;
  74 module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
  75
  76 /* Default doubles per-vcpu halt_poll_ns. */
  77 static unsigned int halt_poll_ns_grow = 2;
  78 module_param(halt_poll_ns_grow, uint, S_IRUGO | S_IWUSR);
  79
  80 /* Default resets per-vcpu halt_poll_ns . */
  81 static unsigned int halt_poll_ns_shrink;
  82 module_param(halt_poll_ns_shrink, uint, S_IRUGO | S_IWUSR);
  83
  84 /*
  85  * Ordering of locks:
  86  *
  87  *      kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  88  */
  89
  90 DEFINE_SPINLOCK(kvm_lock);
  91 static DEFINE_RAW_SPINLOCK(kvm_count_lock);
  92 LIST_HEAD(vm_list);
  93
  94 static cpumask_var_t cpus_hardware_enabled;
  95 static int kvm_usage_count;
  96 static atomic_t hardware_enable_failed;
  97
  98 struct kmem_cache *kvm_vcpu_cache;
  99 EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
 100
 101 static __read_mostly struct preempt_ops kvm_preempt_ops;
 102
 103 struct dentry *kvm_debugfs_dir;
 104 EXPORT_SYMBOL_GPL(kvm_debugfs_dir);
 105
 106 static int kvm_debugfs_num_entries;
 107 static const struct file_operations *stat_fops_per_vm[];
 108
 109 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
 110                            unsigned long arg);
 111 #ifdef CONFIG_KVM_COMPAT
 112 static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl,
 113                                   unsigned long arg);
 114 #endif
 115 static int hardware_enable_all(void);
 116 static void hardware_disable_all(void);
 117
 118 static void kvm_io_bus_destroy(struct kvm_io_bus *bus);
 119
 120 static void kvm_release_pfn_dirty(kvm_pfn_t pfn);
 121 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot, gfn_t gfn);
 122
 123 __visible bool kvm_rebooting;
 124 EXPORT_SYMBOL_GPL(kvm_rebooting);
 125
 126 static bool largepages_enabled = true;
 127
 128 bool kvm_is_reserved_pfn(kvm_pfn_t pfn)
 129 {
 130         if (pfn_valid(pfn))
 131                 return PageReserved(pfn_to_page(pfn));
 132
 133         return true;
 134 }
 135
 136 /*
 137  * Switches to specified vcpu, until a matching vcpu_put()
 138  */
 139 int vcpu_load(struct kvm_vcpu *vcpu)
 140 {
 141         int cpu;
 142
 143         if (mutex_lock_killable(&vcpu->mutex))
 144                 return -EINTR;
 145         cpu = get_cpu();
 146         preempt_notifier_register(&vcpu->preempt_notifier);
 147         kvm_arch_vcpu_load(vcpu, cpu);
 148         put_cpu();
 149         return 0;
 150 }
 151 EXPORT_SYMBOL_GPL(vcpu_load);
 152
 153 void vcpu_put(struct kvm_vcpu *vcpu)
 154 {
 155         preempt_disable();
 156         kvm_arch_vcpu_put(vcpu);
 157         preempt_notifier_unregister(&vcpu->preempt_notifier);
 158         preempt_enable();
 159         mutex_unlock(&vcpu->mutex);
 160 }
 161 EXPORT_SYMBOL_GPL(vcpu_put);
 162
 163 static void ack_flush(void *_completed)
 164 {
 165 }
 166
 167 bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req)
 168 {
 169         int i, cpu, me;
 170         cpumask_var_t cpus;
 171         bool called = true;
 172         struct kvm_vcpu *vcpu;
 173
 174         zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 175
 176         me = get_cpu();
 177         kvm_for_each_vcpu(i, vcpu, kvm) {
 178                 kvm_make_request(req, vcpu);
 179                 cpu = vcpu->cpu;
 180
 181                 /* Set ->requests bit before we read ->mode. */
 182                 smp_mb__after_atomic();
 183
 184                 if (cpus != NULL && cpu != -1 && cpu != me &&
 185                       kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE)
 186                         cpumask_set_cpu(cpu, cpus);
 187         }
 188         if (unlikely(cpus == NULL))
 189                 smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1);
 190         else if (!cpumask_empty(cpus))
 191                 smp_call_function_many(cpus, ack_flush, NULL, 1);
 192         else
 193                 called = false;
 194         put_cpu();
 195         free_cpumask_var(cpus);
 196         return called;
 197 }
 198
 199 #ifndef CONFIG_HAVE_KVM_ARCH_TLB_FLUSH_ALL
 200 void kvm_flush_remote_tlbs(struct kvm *kvm)
 201 {
 202         /*
 203          * Read tlbs_dirty before setting KVM_REQ_TLB_FLUSH in
 204          * kvm_make_all_cpus_request.
 205          */
 206         long dirty_count = smp_load_acquire(&kvm->tlbs_dirty);
 207
 208         /*
 209          * We want to publish modifications to the page tables before reading
 210          * mode. Pairs with a memory barrier in arch-specific code.
 211          * - x86: smp_mb__after_srcu_read_unlock in vcpu_enter_guest
 212          * and smp_mb in walk_shadow_page_lockless_begin/end.
 213          * - powerpc: smp_mb in kvmppc_prepare_to_enter.
 214          *
 215          * There is already an smp_mb__after_atomic() before
 216          * kvm_make_all_cpus_request() reads vcpu->mode. We reuse that
 217          * barrier here.
 218          */
 219         if (kvm_make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 220                 ++kvm->stat.remote_tlb_flush;
 221         cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 222 }
 223 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs);
 224 #endif
 225
 226 void kvm_reload_remote_mmus(struct kvm *kvm)
 227 {
 228         kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
 229 }
 230
 231 int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
 232 {
 233         struct page *page;
 234         int r;
 235
 236         mutex_init(&vcpu->mutex);
 237         vcpu->cpu = -1;
 238         vcpu->kvm = kvm;
 239         vcpu->vcpu_id = id;
 240         vcpu->pid = NULL;
 241         init_swait_queue_head(&vcpu->wq);
 242         kvm_async_pf_vcpu_init(vcpu);
 243
 244         vcpu->pre_pcpu = -1;
 245         INIT_LIST_HEAD(&vcpu->blocked_vcpu_list);
 246
 247         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 248         if (!page) {
 249                 r = -ENOMEM;
 250                 goto fail;
 251         }
 252         vcpu->run = page_address(page);
 253
 254         kvm_vcpu_set_in_spin_loop(vcpu, false);
 255         kvm_vcpu_set_dy_eligible(vcpu, false);
 256         vcpu->preempted = false;
 257
 258         r = kvm_arch_vcpu_init(vcpu);
 259         if (r < 0)
 260                 goto fail_free_run;
 261         return 0;
 262
 263 fail_free_run:
 264         free_page((unsigned long)vcpu->run);
 265 fail:
 266         return r;
 267 }
 268 EXPORT_SYMBOL_GPL(kvm_vcpu_init);
 269
 270 void kvm_vcpu_uninit(struct kvm_vcpu *vcpu)
 271 {
 272         put_pid(vcpu->pid);
 273         kvm_arch_vcpu_uninit(vcpu);
 274         free_page((unsigned long)vcpu->run);
 275 }
 276 EXPORT_SYMBOL_GPL(kvm_vcpu_uninit);
 277
 278 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 279 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
 280 {
 281         return container_of(mn, struct kvm, mmu_notifier);
 282 }
 283
 284 static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 285                                              struct mm_struct *mm,
 286                                              unsigned long address)
 287 {
 288         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 289         int need_tlb_flush, idx;
 290
 291         /*
 292          * When ->invalidate_page runs, the linux pte has been zapped
 293          * already but the page is still allocated until
 294          * ->invalidate_page returns. So if we increase the sequence
 295          * here the kvm page fault will notice if the spte can't be
 296          * established because the page is going to be freed. If
 297          * instead the kvm page fault establishes the spte before
 298          * ->invalidate_page runs, kvm_unmap_hva will release it
 299          * before returning.
 300          *
 301          * The sequence increase only need to be seen at spin_unlock
 302          * time, and not at spin_lock time.
 303          *
 304          * Increasing the sequence after the spin_unlock would be
 305          * unsafe because the kvm page fault could then establish the
 306          * pte after kvm_unmap_hva returned, without noticing the page
 307          * is going to be freed.
 308          */
 309         idx = srcu_read_lock(&kvm->srcu);
 310         spin_lock(&kvm->mmu_lock);
 311
 312         kvm->mmu_notifier_seq++;
 313         need_tlb_flush = kvm_unmap_hva(kvm, address) | kvm->tlbs_dirty;
 314         /* we've to flush the tlb before the pages can be freed */
 315         if (need_tlb_flush)
 316                 kvm_flush_remote_tlbs(kvm);
 317
 318         spin_unlock(&kvm->mmu_lock);
 319
 320         kvm_arch_mmu_notifier_invalidate_page(kvm, address);
 321
 322         srcu_read_unlock(&kvm->srcu, idx);
 323 }
 324
 325 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
 326                                         struct mm_struct *mm,
 327                                         unsigned long address,
 328                                         pte_t pte)
 329 {
 330         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 331         int idx;
 332
 333         idx = srcu_read_lock(&kvm->srcu);
 334         spin_lock(&kvm->mmu_lock);
 335         kvm->mmu_notifier_seq++;
 336         kvm_set_spte_hva(kvm, address, pte);
 337         spin_unlock(&kvm->mmu_lock);
 338         srcu_read_unlock(&kvm->srcu, idx);
 339 }
 340
 341 static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 342                                                     struct mm_struct *mm,
 343                                                     unsigned long start,
 344                                                     unsigned long end)
 345 {
 346         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 347         int need_tlb_flush = 0, idx;
 348
 349         idx = srcu_read_lock(&kvm->srcu);
 350         spin_lock(&kvm->mmu_lock);
 351         /*
 352          * The count increase must become visible at unlock time as no
 353          * spte can be established without taking the mmu_lock and
 354          * count is also read inside the mmu_lock critical section.
 355          */
 356         kvm->mmu_notifier_count++;
 357         need_tlb_flush = kvm_unmap_hva_range(kvm, start, end);
 358         need_tlb_flush |= kvm->tlbs_dirty;
 359         /* we've to flush the tlb before the pages can be freed */
 360         if (need_tlb_flush)
 361                 kvm_flush_remote_tlbs(kvm);
 362
 363         spin_unlock(&kvm->mmu_lock);
 364         srcu_read_unlock(&kvm->srcu, idx);
 365 }
 366
 367 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
 368                                                   struct mm_struct *mm,
 369                                                   unsigned long start,
 370                                                   unsigned long end)
 371 {
 372         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 373
 374         spin_lock(&kvm->mmu_lock);
 375         /*
 376          * This sequence increase will notify the kvm page fault that
 377          * the page that is going to be mapped in the spte could have
 378          * been freed.
 379          */
 380         kvm->mmu_notifier_seq++;
 381         smp_wmb();
 382         /*
 383          * The above sequence increase must be visible before the
 384          * below count decrease, which is ensured by the smp_wmb above
 385          * in conjunction with the smp_rmb in mmu_notifier_retry().
 386          */
 387         kvm->mmu_notifier_count--;
 388         spin_unlock(&kvm->mmu_lock);
 389
 390         BUG_ON(kvm->mmu_notifier_count < 0);
 391 }
 392
 393 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
 394                                               struct mm_struct *mm,
 395                                               unsigned long start,
 396                                               unsigned long end)
 397 {
 398         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 399         int young, idx;
 400
 401         idx = srcu_read_lock(&kvm->srcu);
 402         spin_lock(&kvm->mmu_lock);
 403
 404         young = kvm_age_hva(kvm, start, end);
 405         if (young)
 406                 kvm_flush_remote_tlbs(kvm);
 407
 408         spin_unlock(&kvm->mmu_lock);
 409         srcu_read_unlock(&kvm->srcu, idx);
 410
 411         return young;
 412 }
 413
 414 static int kvm_mmu_notifier_clear_young(struct mmu_notifier *mn,
 415                                         struct mm_struct *mm,
 416                                         unsigned long start,
 417                                         unsigned long end)
 418 {
 419         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 420         int young, idx;
 421
 422         idx = srcu_read_lock(&kvm->srcu);
 423         spin_lock(&kvm->mmu_lock);
 424         /*
 425          * Even though we do not flush TLB, this will still adversely
 426          * affect performance on pre-Haswell Intel EPT, where there is
 427          * no EPT Access Bit to clear so that we have to tear down EPT
 428          * tables instead. If we find this unacceptable, we can always
 429          * add a parameter to kvm_age_hva so that it effectively doesn't
 430          * do anything on clear_young.
 431          *
 432          * Also note that currently we never issue secondary TLB flushes
 433          * from clear_young, leaving this job up to the regular system
 434          * cadence. If we find this inaccurate, we might come up with a
 435          * more sophisticated heuristic later.
 436          */
 437         young = kvm_age_hva(kvm, start, end);
 438         spin_unlock(&kvm->mmu_lock);
 439         srcu_read_unlock(&kvm->srcu, idx);
 440
 441         return young;
 442 }
 443
 444 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn,
 445                                        struct mm_struct *mm,
 446                                        unsigned long address)
 447 {
 448         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 449         int young, idx;
 450
 451         idx = srcu_read_lock(&kvm->srcu);
 452         spin_lock(&kvm->mmu_lock);
 453         young = kvm_test_age_hva(kvm, address);
 454         spin_unlock(&kvm->mmu_lock);
 455         srcu_read_unlock(&kvm->srcu, idx);
 456
 457         return young;
 458 }
 459
 460 static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
 461                                      struct mm_struct *mm)
 462 {
 463         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 464         int idx;
 465
 466         idx = srcu_read_lock(&kvm->srcu);
 467         kvm_arch_flush_shadow_all(kvm);
 468         srcu_read_unlock(&kvm->srcu, idx);
 469 }
 470
 471 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
 472         .invalidate_page        = kvm_mmu_notifier_invalidate_page,
 473         .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
 474         .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
 475         .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
 476         .clear_young            = kvm_mmu_notifier_clear_young,
 477         .test_young             = kvm_mmu_notifier_test_young,
 478         .change_pte             = kvm_mmu_notifier_change_pte,
 479         .release                = kvm_mmu_notifier_release,
 480 };
 481
 482 static int kvm_init_mmu_notifier(struct kvm *kvm)
 483 {
 484         kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
 485         return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
 486 }
 487
 488 #else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
 489
 490 static int kvm_init_mmu_notifier(struct kvm *kvm)
 491 {
 492         return 0;
 493 }
 494
 495 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 496
 497 static struct kvm_memslots *kvm_alloc_memslots(void)
 498 {
 499         int i;
 500         struct kvm_memslots *slots;
 501
 502         slots = kvm_kvzalloc(sizeof(struct kvm_memslots));
 503         if (!slots)
 504                 return NULL;
 505
 506         /*
 507          * Init kvm generation close to the maximum to easily test the
 508          * code of handling generation number wrap-around.
 509          */
 510         slots->generation = -150;
 511         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 512                 slots->id_to_index[i] = slots->memslots[i].id = i;
 513
 514         return slots;
 515 }
 516
 517 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
 518 {
 519         if (!memslot->dirty_bitmap)
 520                 return;
 521
 522         kvfree(memslot->dirty_bitmap);
 523         memslot->dirty_bitmap = NULL;
 524 }
 525
 526 /*
 527  * Free any memory in @free but not in @dont.
 528  */
 529 static void kvm_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
 530                               struct kvm_memory_slot *dont)
 531 {
 532         if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
 533                 kvm_destroy_dirty_bitmap(free);
 534
 535         kvm_arch_free_memslot(kvm, free, dont);
 536
 537         free->npages = 0;
 538 }
 539
 540 static void kvm_free_memslots(struct kvm *kvm, struct kvm_memslots *slots)
 541 {
 542         struct kvm_memory_slot *memslot;
 543
 544         if (!slots)
 545                 return;
 546
 547         kvm_for_each_memslot(memslot, slots)
 548                 kvm_free_memslot(kvm, memslot, NULL);
 549
 550         kvfree(slots);
 551 }
 552
 553 static void kvm_destroy_vm_debugfs(struct kvm *kvm)
 554 {
 555         int i;
 556
 557         if (!kvm->debugfs_dentry)
 558                 return;
 559
 560         debugfs_remove_recursive(kvm->debugfs_dentry);
 561
 562         for (i = 0; i < kvm_debugfs_num_entries; i++)
 563                 kfree(kvm->debugfs_stat_data[i]);
 564         kfree(kvm->debugfs_stat_data);
 565 }
 566
 567 static int kvm_create_vm_debugfs(struct kvm *kvm, int fd)
 568 {
 569         char dir_name[ITOA_MAX_LEN * 2];
 570         struct kvm_stat_data *stat_data;
 571         struct kvm_stats_debugfs_item *p;
 572
 573         if (!debugfs_initialized())
 574                 return 0;
 575
 576         snprintf(dir_name, sizeof(dir_name), "%d-%d", task_pid_nr(current), fd);
 577         kvm->debugfs_dentry = debugfs_create_dir(dir_name,
 578                                                  kvm_debugfs_dir);
 579         if (!kvm->debugfs_dentry)
 580                 return -ENOMEM;
 581
 582         kvm->debugfs_stat_data = kcalloc(kvm_debugfs_num_entries,
 583                                          sizeof(*kvm->debugfs_stat_data),
 584                                          GFP_KERNEL);
 585         if (!kvm->debugfs_stat_data)
 586                 return -ENOMEM;
 587
 588         for (p = debugfs_entries; p->name; p++) {
 589                 stat_data = kzalloc(sizeof(*stat_data), GFP_KERNEL);
 590                 if (!stat_data)
 591                         return -ENOMEM;
 592
 593                 stat_data->kvm = kvm;
 594                 stat_data->offset = p->offset;
 595                 kvm->debugfs_stat_data[p - debugfs_entries] = stat_data;
 596                 if (!debugfs_create_file(p->name, 0444,
 597                                          kvm->debugfs_dentry,
 598                                          stat_data,
 599                                          stat_fops_per_vm[p->kind]))
 600                         return -ENOMEM;
 601         }
 602         return 0;
 603 }
 604
 605 static struct kvm *kvm_create_vm(unsigned long type)
 606 {
 607         int r, i;
 608         struct kvm *kvm = kvm_arch_alloc_vm();
 609
 610         if (!kvm)
 611                 return ERR_PTR(-ENOMEM);
 612
 613         spin_lock_init(&kvm->mmu_lock);
 614         atomic_inc(&current->mm->mm_count);
 615         kvm->mm = current->mm;
 616         kvm_eventfd_init(kvm);
 617         mutex_init(&kvm->lock);
 618         mutex_init(&kvm->irq_lock);
 619         mutex_init(&kvm->slots_lock);
 620         atomic_set(&kvm->users_count, 1);
 621         INIT_LIST_HEAD(&kvm->devices);
 622
 623         r = kvm_arch_init_vm(kvm, type);
 624         if (r)
 625                 goto out_err_no_disable;
 626
 627         r = hardware_enable_all();
 628         if (r)
 629                 goto out_err_no_disable;
 630
 631 #ifdef CONFIG_HAVE_KVM_IRQFD
 632         INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 633 #endif
 634
 635         BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX);
 636
 637         r = -ENOMEM;
 638         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
 639                 kvm->memslots[i] = kvm_alloc_memslots();
 640                 if (!kvm->memslots[i])
 641                         goto out_err_no_srcu;
 642         }
 643
 644         if (init_srcu_struct(&kvm->srcu))
 645                 goto out_err_no_srcu;
 646         if (init_srcu_struct(&kvm->irq_srcu))
 647                 goto out_err_no_irq_srcu;
 648         for (i = 0; i < KVM_NR_BUSES; i++) {
 649                 kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
 650                                         GFP_KERNEL);
 651                 if (!kvm->buses[i])
 652                         goto out_err;
 653         }
 654
 655         r = kvm_init_mmu_notifier(kvm);
 656         if (r)
 657                 goto out_err;
 658
 659         spin_lock(&kvm_lock);
 660         list_add(&kvm->vm_list, &vm_list);
 661         spin_unlock(&kvm_lock);
 662
 663         preempt_notifier_inc();
 664
 665         return kvm;
 666
 667 out_err:
 668         cleanup_srcu_struct(&kvm->irq_srcu);
 669 out_err_no_irq_srcu:
 670         cleanup_srcu_struct(&kvm->srcu);
 671 out_err_no_srcu:
 672         hardware_disable_all();
 673 out_err_no_disable:
 674         for (i = 0; i < KVM_NR_BUSES; i++)
 675                 kfree(kvm->buses[i]);
 676         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 677                 kvm_free_memslots(kvm, kvm->memslots[i]);
 678         kvm_arch_free_vm(kvm);
 679         mmdrop(current->mm);
 680         return ERR_PTR(r);
 681 }
 682
 683 /*
 684  * Avoid using vmalloc for a small buffer.
 685  * Should not be used when the size is statically known.
 686  */
 687 void *kvm_kvzalloc(unsigned long size)
 688 {
 689         if (size > PAGE_SIZE)
 690                 return vzalloc(size);
 691         else
 692                 return kzalloc(size, GFP_KERNEL);
 693 }
 694
 695 static void kvm_destroy_devices(struct kvm *kvm)
 696 {
 697         struct kvm_device *dev, *tmp;
 698
 699         list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) {
 700                 list_del(&dev->vm_node);
 701                 dev->ops->destroy(dev);
 702         }
 703 }
 704
 705 static void kvm_destroy_vm(struct kvm *kvm)
 706 {
 707         int i;
 708         struct mm_struct *mm = kvm->mm;
 709
 710         kvm_destroy_vm_debugfs(kvm);
 711         kvm_arch_sync_events(kvm);
 712         spin_lock(&kvm_lock);
 713         list_del(&kvm->vm_list);
 714         spin_unlock(&kvm_lock);
 715         kvm_free_irq_routing(kvm);
 716         for (i = 0; i < KVM_NR_BUSES; i++)
 717                 kvm_io_bus_destroy(kvm->buses[i]);
 718         kvm_coalesced_mmio_free(kvm);
 719 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 720         mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 721 #else
 722         kvm_arch_flush_shadow_all(kvm);
 723 #endif
 724         kvm_arch_destroy_vm(kvm);
 725         kvm_destroy_devices(kvm);
 726         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 727                 kvm_free_memslots(kvm, kvm->memslots[i]);
 728         cleanup_srcu_struct(&kvm->irq_srcu);
 729         cleanup_srcu_struct(&kvm->srcu);
 730         kvm_arch_free_vm(kvm);
 731         preempt_notifier_dec();
 732         hardware_disable_all();
 733         mmdrop(mm);
 734 }
 735
 736 void kvm_get_kvm(struct kvm *kvm)
 737 {
 738         atomic_inc(&kvm->users_count);
 739 }
 740 EXPORT_SYMBOL_GPL(kvm_get_kvm);
 741
 742 void kvm_put_kvm(struct kvm *kvm)
 743 {
 744         if (atomic_dec_and_test(&kvm->users_count))
 745                 kvm_destroy_vm(kvm);
 746 }
 747 EXPORT_SYMBOL_GPL(kvm_put_kvm);
 748
 749
 750 static int kvm_vm_release(struct inode *inode, struct file *filp)
 751 {
 752         struct kvm *kvm = filp->private_data;
 753
 754         kvm_irqfd_release(kvm);
 755
 756         kvm_put_kvm(kvm);
 757         return 0;
 758 }
 759
 760 /*
 761  * Allocation size is twice as large as the actual dirty bitmap size.
 762  * See x86's kvm_vm_ioctl_get_dirty_log() why this is needed.
 763  */
 764 static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
 765 {
 766         unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
 767
 768         memslot->dirty_bitmap = kvm_kvzalloc(dirty_bytes);
 769         if (!memslot->dirty_bitmap)
 770                 return -ENOMEM;
 771
 772         return 0;
 773 }
 774
 775 /*
 776  * Insert memslot and re-sort memslots based on their GFN,
 777  * so binary search could be used to lookup GFN.
 778  * Sorting algorithm takes advantage of having initially
 779  * sorted array and known changed memslot position.
 780  */
 781 static void update_memslots(struct kvm_memslots *slots,
 782                             struct kvm_memory_slot *new)
 783 {
 784         int id = new->id;
 785         int i = slots->id_to_index[id];
 786         struct kvm_memory_slot *mslots = slots->memslots;
 787
 788         WARN_ON(mslots[i].id != id);
 789         if (!new->npages) {
 790                 WARN_ON(!mslots[i].npages);
 791                 if (mslots[i].npages)
 792                         slots->used_slots--;
 793         } else {
 794                 if (!mslots[i].npages)
 795                         slots->used_slots++;
 796         }
 797
 798         while (i < KVM_MEM_SLOTS_NUM - 1 &&
 799                new->base_gfn <= mslots[i + 1].base_gfn) {
 800                 if (!mslots[i + 1].npages)
 801                         break;
 802                 mslots[i] = mslots[i + 1];
 803                 slots->id_to_index[mslots[i].id] = i;
 804                 i++;
 805         }
 806
 807         /*
 808          * The ">=" is needed when creating a slot with base_gfn == 0,
 809          * so that it moves before all those with base_gfn == npages == 0.
 810          *
 811          * On the other hand, if new->npages is zero, the above loop has
 812          * already left i pointing to the beginning of the empty part of
 813          * mslots, and the ">=" would move the hole backwards in this
 814          * case---which is wrong.  So skip the loop when deleting a slot.
 815          */
 816         if (new->npages) {
 817                 while (i > 0 &&
 818                        new->base_gfn >= mslots[i - 1].base_gfn) {
 819                         mslots[i] = mslots[i - 1];
 820                         slots->id_to_index[mslots[i].id] = i;
 821                         i--;
 822                 }
 823         } else
 824                 WARN_ON_ONCE(i != slots->used_slots);
 825
 826         mslots[i] = *new;
 827         slots->id_to_index[mslots[i].id] = i;
 828 }
 829
 830 static int check_memory_region_flags(const struct kvm_userspace_memory_region *mem)
 831 {
 832         u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES;
 833
 834 #ifdef __KVM_HAVE_READONLY_MEM
 835         valid_flags |= KVM_MEM_READONLY;
 836 #endif
 837
 838         if (mem->flags & ~valid_flags)
 839                 return -EINVAL;
 840
 841         return 0;
 842 }
 843
 844 static struct kvm_memslots *install_new_memslots(struct kvm *kvm,
 845                 int as_id, struct kvm_memslots *slots)
 846 {
 847         struct kvm_memslots *old_memslots = __kvm_memslots(kvm, as_id);
 848
 849         /*
 850          * Set the low bit in the generation, which disables SPTE caching
 851          * until the end of synchronize_srcu_expedited.
 852          */
 853         WARN_ON(old_memslots->generation & 1);
 854         slots->generation = old_memslots->generation + 1;
 855
 856         rcu_assign_pointer(kvm->memslots[as_id], slots);
 857         synchronize_srcu_expedited(&kvm->srcu);
 858
 859         /*
 860          * Increment the new memslot generation a second time. This prevents
 861          * vm exits that race with memslot updates from caching a memslot
 862          * generation that will (potentially) be valid forever.
 863          */
 864         slots->generation++;
 865
 866         kvm_arch_memslots_updated(kvm, slots);
 867
 868         return old_memslots;
 869 }
 870
 871 /*
 872  * Allocate some memory and give it an address in the guest physical address
 873  * space.
 874  *
 875  * Discontiguous memory is allowed, mostly for framebuffers.
 876  *
 877  * Must be called holding kvm->slots_lock for write.
 878  */
 879 int __kvm_set_memory_region(struct kvm *kvm,
 880                             const struct kvm_userspace_memory_region *mem)
 881 {
 882         int r;
 883         gfn_t base_gfn;
 884         unsigned long npages;
 885         struct kvm_memory_slot *slot;
 886         struct kvm_memory_slot old, new;
 887         struct kvm_memslots *slots = NULL, *old_memslots;
 888         int as_id, id;
 889         enum kvm_mr_change change;
 890
 891         r = check_memory_region_flags(mem);
 892         if (r)
 893                 goto out;
 894
 895         r = -EINVAL;
 896         as_id = mem->slot >> 16;
 897         id = (u16)mem->slot;
 898
 899         /* General sanity checks */
 900         if (mem->memory_size & (PAGE_SIZE - 1))
 901                 goto out;
 902         if (mem->guest_phys_addr & (PAGE_SIZE - 1))
 903                 goto out;
 904         /* We can read the guest memory with __xxx_user() later on. */
 905         if ((id < KVM_USER_MEM_SLOTS) &&
 906             ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
 907              !access_ok(VERIFY_WRITE,
 908                         (void __user *)(unsigned long)mem->userspace_addr,
 909                         mem->memory_size)))
 910                 goto out;
 911         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_MEM_SLOTS_NUM)
 912                 goto out;
 913         if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
 914                 goto out;
 915
 916         slot = id_to_memslot(__kvm_memslots(kvm, as_id), id);
 917         base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
 918         npages = mem->memory_size >> PAGE_SHIFT;
 919
 920         if (npages > KVM_MEM_MAX_NR_PAGES)
 921                 goto out;
 922
 923         new = old = *slot;
 924
 925         new.id = id;
 926         new.base_gfn = base_gfn;
 927         new.npages = npages;
 928         new.flags = mem->flags;
 929
 930         if (npages) {
 931                 if (!old.npages)
 932                         change = KVM_MR_CREATE;
 933                 else { /* Modify an existing slot. */
 934                         if ((mem->userspace_addr != old.userspace_addr) ||
 935                             (npages != old.npages) ||
 936                             ((new.flags ^ old.flags) & KVM_MEM_READONLY))
 937                                 goto out;
 938
 939                         if (base_gfn != old.base_gfn)
 940                                 change = KVM_MR_MOVE;
 941                         else if (new.flags != old.flags)
 942                                 change = KVM_MR_FLAGS_ONLY;
 943                         else { /* Nothing to change. */
 944                                 r = 0;
 945                                 goto out;
 946                         }
 947                 }
 948         } else {
 949                 if (!old.npages)
 950                         goto out;
 951
 952                 change = KVM_MR_DELETE;
 953                 new.base_gfn = 0;
 954                 new.flags = 0;
 955         }
 956
 957         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 958                 /* Check for overlaps */
 959                 r = -EEXIST;
 960                 kvm_for_each_memslot(slot, __kvm_memslots(kvm, as_id)) {
 961                         if ((slot->id >= KVM_USER_MEM_SLOTS) ||
 962                             (slot->id == id))
 963                                 continue;
 964                         if (!((base_gfn + npages <= slot->base_gfn) ||
 965                               (base_gfn >= slot->base_gfn + slot->npages)))
 966                                 goto out;
 967                 }
 968         }
 969
 970         /* Free page dirty bitmap if unneeded */
 971         if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
 972                 new.dirty_bitmap = NULL;
 973
 974         r = -ENOMEM;
 975         if (change == KVM_MR_CREATE) {
 976                 new.userspace_addr = mem->userspace_addr;
 977
 978                 if (kvm_arch_create_memslot(kvm, &new, npages))
 979                         goto out_free;
 980         }
 981
 982         /* Allocate page dirty bitmap if needed */
 983         if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
 984                 if (kvm_create_dirty_bitmap(&new) < 0)
 985                         goto out_free;
 986         }
 987
 988         slots = kvm_kvzalloc(sizeof(struct kvm_memslots));
 989         if (!slots)
 990                 goto out_free;
 991         memcpy(slots, __kvm_memslots(kvm, as_id), sizeof(struct kvm_memslots));
 992
 993         if ((change == KVM_MR_DELETE) || (change == KVM_MR_MOVE)) {
 994                 slot = id_to_memslot(slots, id);
 995                 slot->flags |= KVM_MEMSLOT_INVALID;
 996
 997                 old_memslots = install_new_memslots(kvm, as_id, slots);
 998
 999                 /* slot was deleted or moved, clear iommu mapping */
1000                 kvm_iommu_unmap_pages(kvm, &old);
1001                 /* From this point no new shadow pages pointing to a deleted,
1002                  * or moved, memslot will be created.
1003                  *
1004                  * validation of sp->gfn happens in:
1005                  *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
1006                  *      - kvm_is_visible_gfn (mmu_check_roots)
1007                  */
1008                 kvm_arch_flush_shadow_memslot(kvm, slot);
1009
1010                 /*
1011                  * We can re-use the old_memslots from above, the only difference
1012                  * from the currently installed memslots is the invalid flag.  This
1013                  * will get overwritten by update_memslots anyway.
1014                  */
1015                 slots = old_memslots;
1016         }
1017
1018         r = kvm_arch_prepare_memory_region(kvm, &new, mem, change);
1019         if (r)
1020                 goto out_slots;
1021
1022         /* actual memory is freed via old in kvm_free_memslot below */
1023         if (change == KVM_MR_DELETE) {
1024                 new.dirty_bitmap = NULL;
1025                 memset(&new.arch, 0, sizeof(new.arch));
1026         }
1027
1028         update_memslots(slots, &new);
1029         old_memslots = install_new_memslots(kvm, as_id, slots);
1030
1031         kvm_arch_commit_memory_region(kvm, mem, &old, &new, change);
1032
1033         kvm_free_memslot(kvm, &old, &new);
1034         kvfree(old_memslots);
1035
1036         /*
1037          * IOMMU mapping:  New slots need to be mapped.  Old slots need to be
1038          * un-mapped and re-mapped if their base changes.  Since base change
1039          * unmapping is handled above with slot deletion, mapping alone is
1040          * needed here.  Anything else the iommu might care about for existing
1041          * slots (size changes, userspace addr changes and read-only flag
1042          * changes) is disallowed above, so any other attribute changes getting
1043          * here can be skipped.
1044          */
1045         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
1046                 r = kvm_iommu_map_pages(kvm, &new);
1047                 return r;
1048         }
1049
1050         return 0;
1051
1052 out_slots:
1053         kvfree(slots);
1054 out_free:
1055         kvm_free_memslot(kvm, &new, &old);
1056 out:
1057         return r;
1058 }
1059 EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
1060
1061 int kvm_set_memory_region(struct kvm *kvm,
1062                           const struct kvm_userspace_memory_region *mem)
1063 {
1064         int r;
1065
1066         mutex_lock(&kvm->slots_lock);
1067         r = __kvm_set_memory_region(kvm, mem);
1068         mutex_unlock(&kvm->slots_lock);
1069         return r;
1070 }
1071 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
1072
1073 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
1074                                           struct kvm_userspace_memory_region *mem)
1075 {
1076         if ((u16)mem->slot >= KVM_USER_MEM_SLOTS)
1077                 return -EINVAL;
1078
1079         return kvm_set_memory_region(kvm, mem);
1080 }
1081
1082 int kvm_get_dirty_log(struct kvm *kvm,
1083                         struct kvm_dirty_log *log, int *is_dirty)
1084 {
1085         struct kvm_memslots *slots;
1086         struct kvm_memory_slot *memslot;
1087         int r, i, as_id, id;
1088         unsigned long n;
1089         unsigned long any = 0;
1090
1091         r = -EINVAL;
1092         as_id = log->slot >> 16;
1093         id = (u16)log->slot;
1094         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1095                 goto out;
1096
1097         slots = __kvm_memslots(kvm, as_id);
1098         memslot = id_to_memslot(slots, id);
1099         r = -ENOENT;
1100         if (!memslot->dirty_bitmap)
1101                 goto out;
1102
1103         n = kvm_dirty_bitmap_bytes(memslot);
1104
1105         for (i = 0; !any && i < n/sizeof(long); ++i)
1106                 any = memslot->dirty_bitmap[i];
1107
1108         r = -EFAULT;
1109         if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
1110                 goto out;
1111
1112         if (any)
1113                 *is_dirty = 1;
1114
1115         r = 0;
1116 out:
1117         return r;
1118 }
1119 EXPORT_SYMBOL_GPL(kvm_get_dirty_log);
1120
1121 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
1122 /**
1123  * kvm_get_dirty_log_protect - get a snapshot of dirty pages, and if any pages
1124  *      are dirty write protect them for next write.
1125  * @kvm:        pointer to kvm instance
1126  * @log:        slot id and address to which we copy the log
1127  * @is_dirty:   flag set if any page is dirty
1128  *
1129  * We need to keep it in mind that VCPU threads can write to the bitmap
1130  * concurrently. So, to avoid losing track of dirty pages we keep the
1131  * following order:
1132  *
1133  *    1. Take a snapshot of the bit and clear it if needed.
1134  *    2. Write protect the corresponding page.
1135  *    3. Copy the snapshot to the userspace.
1136  *    4. Upon return caller flushes TLB's if needed.
1137  *
1138  * Between 2 and 4, the guest may write to the page using the remaining TLB
1139  * entry.  This is not a problem because the page is reported dirty using
1140  * the snapshot taken before and step 4 ensures that writes done after
1141  * exiting to userspace will be logged for the next call.
1142  *
1143  */
1144 int kvm_get_dirty_log_protect(struct kvm *kvm,
1145                         struct kvm_dirty_log *log, bool *is_dirty)
1146 {
1147         struct kvm_memslots *slots;
1148         struct kvm_memory_slot *memslot;
1149         int r, i, as_id, id;
1150         unsigned long n;
1151         unsigned long *dirty_bitmap;
1152         unsigned long *dirty_bitmap_buffer;
1153
1154         r = -EINVAL;
1155         as_id = log->slot >> 16;
1156         id = (u16)log->slot;
1157         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1158                 goto out;
1159
1160         slots = __kvm_memslots(kvm, as_id);
1161         memslot = id_to_memslot(slots, id);
1162
1163         dirty_bitmap = memslot->dirty_bitmap;
1164         r = -ENOENT;
1165         if (!dirty_bitmap)
1166                 goto out;
1167
1168         n = kvm_dirty_bitmap_bytes(memslot);
1169
1170         dirty_bitmap_buffer = dirty_bitmap + n / sizeof(long);
1171         memset(dirty_bitmap_buffer, 0, n);
1172
1173         spin_lock(&kvm->mmu_lock);
1174         *is_dirty = false;
1175         for (i = 0; i < n / sizeof(long); i++) {
1176                 unsigned long mask;
1177                 gfn_t offset;
1178
1179                 if (!dirty_bitmap[i])
1180                         continue;
1181
1182                 *is_dirty = true;
1183
1184                 mask = xchg(&dirty_bitmap[i], 0);
1185                 dirty_bitmap_buffer[i] = mask;
1186
1187                 if (mask) {
1188                         offset = i * BITS_PER_LONG;
1189                         kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot,
1190                                                                 offset, mask);
1191                 }
1192         }
1193
1194         spin_unlock(&kvm->mmu_lock);
1195
1196         r = -EFAULT;
1197         if (copy_to_user(log->dirty_bitmap, dirty_bitmap_buffer, n))
1198                 goto out;
1199
1200         r = 0;
1201 out:
1202         return r;
1203 }
1204 EXPORT_SYMBOL_GPL(kvm_get_dirty_log_protect);
1205 #endif
1206
1207 bool kvm_largepages_enabled(void)
1208 {
1209         return largepages_enabled;
1210 }
1211
1212 void kvm_disable_largepages(void)
1213 {
1214         largepages_enabled = false;
1215 }
1216 EXPORT_SYMBOL_GPL(kvm_disable_largepages);
1217
1218 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
1219 {
1220         return __gfn_to_memslot(kvm_memslots(kvm), gfn);
1221 }
1222 EXPORT_SYMBOL_GPL(gfn_to_memslot);
1223
1224 struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn)
1225 {
1226         return __gfn_to_memslot(kvm_vcpu_memslots(vcpu), gfn);
1227 }
1228
1229 bool kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
1230 {
1231         struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
1232
1233         if (!memslot || memslot->id >= KVM_USER_MEM_SLOTS ||
1234               memslot->flags & KVM_MEMSLOT_INVALID)
1235                 return false;
1236
1237         return true;
1238 }
1239 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
1240
1241 unsigned long kvm_host_page_size(struct kvm *kvm, gfn_t gfn)
1242 {
1243         struct vm_area_struct *vma;
1244         unsigned long addr, size;
1245
1246         size = PAGE_SIZE;
1247
1248         addr = gfn_to_hva(kvm, gfn);
1249         if (kvm_is_error_hva(addr))
1250                 return PAGE_SIZE;
1251
1252         down_read(&current->mm->mmap_sem);
1253         vma = find_vma(current->mm, addr);
1254         if (!vma)
1255                 goto out;
1256
1257         size = vma_kernel_pagesize(vma);
1258
1259 out:
1260         up_read(&current->mm->mmap_sem);
1261
1262         return size;
1263 }
1264
1265 static bool memslot_is_readonly(struct kvm_memory_slot *slot)
1266 {
1267         return slot->flags & KVM_MEM_READONLY;
1268 }
1269
1270 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1271                                        gfn_t *nr_pages, bool write)
1272 {
1273         if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1274                 return KVM_HVA_ERR_BAD;
1275
1276         if (memslot_is_readonly(slot) && write)
1277                 return KVM_HVA_ERR_RO_BAD;
1278
1279         if (nr_pages)
1280                 *nr_pages = slot->npages - (gfn - slot->base_gfn);
1281
1282         return __gfn_to_hva_memslot(slot, gfn);
1283 }
1284
1285 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1286                                      gfn_t *nr_pages)
1287 {
1288         return __gfn_to_hva_many(slot, gfn, nr_pages, true);
1289 }
1290
1291 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot,
1292                                         gfn_t gfn)
1293 {
1294         return gfn_to_hva_many(slot, gfn, NULL);
1295 }
1296 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot);
1297
1298 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
1299 {
1300         return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
1301 }
1302 EXPORT_SYMBOL_GPL(gfn_to_hva);
1303
1304 unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn)
1305 {
1306         return gfn_to_hva_many(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, NULL);
1307 }
1308 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_hva);
1309
1310 /*
1311  * If writable is set to false, the hva returned by this function is only
1312  * allowed to be read.
1313  */
1314 unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot,
1315                                       gfn_t gfn, bool *writable)
1316 {
1317         unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false);
1318
1319         if (!kvm_is_error_hva(hva) && writable)
1320                 *writable = !memslot_is_readonly(slot);
1321
1322         return hva;
1323 }
1324
1325 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable)
1326 {
1327         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1328
1329         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1330 }
1331
1332 unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable)
1333 {
1334         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1335
1336         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1337 }
1338
1339 static int get_user_page_nowait(unsigned long start, int write,
1340                 struct page **page)
1341 {
1342         int flags = FOLL_TOUCH | FOLL_NOWAIT | FOLL_HWPOISON | FOLL_GET;
1343
1344         if (write)
1345                 flags |= FOLL_WRITE;
1346
1347         return __get_user_pages(current, current->mm, start, 1, flags, page,
1348                         NULL, NULL);
1349 }
1350
1351 static inline int check_user_page_hwpoison(unsigned long addr)
1352 {
1353         int rc, flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_WRITE;
1354
1355         rc = __get_user_pages(current, current->mm, addr, 1,
1356                               flags, NULL, NULL, NULL);
1357         return rc == -EHWPOISON;
1358 }
1359
1360 /*
1361  * The atomic path to get the writable pfn which will be stored in @pfn,
1362  * true indicates success, otherwise false is returned.
1363  */
1364 static bool hva_to_pfn_fast(unsigned long addr, bool atomic, bool *async,
1365                             bool write_fault, bool *writable, kvm_pfn_t *pfn)
1366 {
1367         struct page *page[1];
1368         int npages;
1369
1370         if (!(async || atomic))
1371                 return false;
1372
1373         /*
1374          * Fast pin a writable pfn only if it is a write fault request
1375          * or the caller allows to map a writable pfn for a read fault
1376          * request.
1377          */
1378         if (!(write_fault || writable))
1379                 return false;
1380
1381         npages = __get_user_pages_fast(addr, 1, 1, page);
1382         if (npages == 1) {
1383                 *pfn = page_to_pfn(page[0]);
1384
1385                 if (writable)
1386                         *writable = true;
1387                 return true;
1388         }
1389
1390         return false;
1391 }
1392
1393 /*
1394  * The slow path to get the pfn of the specified host virtual address,
1395  * 1 indicates success, -errno is returned if error is detected.
1396  */
1397 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault,
1398                            bool *writable, kvm_pfn_t *pfn)
1399 {
1400         struct page *page[1];
1401         int npages = 0;
1402
1403         might_sleep();
1404
1405         if (writable)
1406                 *writable = write_fault;
1407
1408         if (async) {
1409                 down_read(&current->mm->mmap_sem);
1410                 npages = get_user_page_nowait(addr, write_fault, page);
1411                 up_read(&current->mm->mmap_sem);
1412         } else
1413                 npages = __get_user_pages_unlocked(current, current->mm, addr, 1,
1414                                                    write_fault, 0, page,
1415                                                    FOLL_TOUCH|FOLL_HWPOISON);
1416         if (npages != 1)
1417                 return npages;
1418
1419         /* map read fault as writable if possible */
1420         if (unlikely(!write_fault) && writable) {
1421                 struct page *wpage[1];
1422
1423                 npages = __get_user_pages_fast(addr, 1, 1, wpage);
1424                 if (npages == 1) {
1425                         *writable = true;
1426                         put_page(page[0]);
1427                         page[0] = wpage[0];
1428                 }
1429
1430                 npages = 1;
1431         }
1432         *pfn = page_to_pfn(page[0]);
1433         return npages;
1434 }
1435
1436 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault)
1437 {
1438         if (unlikely(!(vma->vm_flags & VM_READ)))
1439                 return false;
1440
1441         if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE))))
1442                 return false;
1443
1444         return true;
1445 }
1446
1447 /*
1448  * Pin guest page in memory and return its pfn.
1449  * @addr: host virtual address which maps memory to the guest
1450  * @atomic: whether this function can sleep
1451  * @async: whether this function need to wait IO complete if the
1452  *         host page is not in the memory
1453  * @write_fault: whether we should get a writable host page
1454  * @writable: whether it allows to map a writable host page for !@write_fault
1455  *
1456  * The function will map a writable host page for these two cases:
1457  * 1): @write_fault = true
1458  * 2): @write_fault = false && @writable, @writable will tell the caller
1459  *     whether the mapping is writable.
1460  */
1461 static kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
1462                         bool write_fault, bool *writable)
1463 {
1464         struct vm_area_struct *vma;
1465         kvm_pfn_t pfn = 0;
1466         int npages;
1467
1468         /* we can do it either atomically or asynchronously, not both */
1469         BUG_ON(atomic && async);
1470
1471         if (hva_to_pfn_fast(addr, atomic, async, write_fault, writable, &pfn))
1472                 return pfn;
1473
1474         if (atomic)
1475                 return KVM_PFN_ERR_FAULT;
1476
1477         npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn);
1478         if (npages == 1)
1479                 return pfn;
1480
1481         down_read(&current->mm->mmap_sem);
1482         if (npages == -EHWPOISON ||
1483               (!async && check_user_page_hwpoison(addr))) {
1484                 pfn = KVM_PFN_ERR_HWPOISON;
1485                 goto exit;
1486         }
1487
1488         vma = find_vma_intersection(current->mm, addr, addr + 1);
1489
1490         if (vma == NULL)
1491                 pfn = KVM_PFN_ERR_FAULT;
1492         else if ((vma->vm_flags & VM_PFNMAP)) {
1493                 pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) +
1494                         vma->vm_pgoff;
1495                 BUG_ON(!kvm_is_reserved_pfn(pfn));
1496         } else {
1497                 if (async && vma_is_valid(vma, write_fault))
1498                         *async = true;
1499                 pfn = KVM_PFN_ERR_FAULT;
1500         }
1501 exit:
1502         up_read(&current->mm->mmap_sem);
1503         return pfn;
1504 }
1505
1506 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn,
1507                                bool atomic, bool *async, bool write_fault,
1508                                bool *writable)
1509 {
1510         unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
1511
1512         if (addr == KVM_HVA_ERR_RO_BAD) {
1513                 if (writable)
1514                         *writable = false;
1515                 return KVM_PFN_ERR_RO_FAULT;
1516         }
1517
1518         if (kvm_is_error_hva(addr)) {
1519                 if (writable)
1520                         *writable = false;
1521                 return KVM_PFN_NOSLOT;
1522         }
1523
1524         /* Do not map writable pfn in the readonly memslot. */
1525         if (writable && memslot_is_readonly(slot)) {
1526                 *writable = false;
1527                 writable = NULL;
1528         }
1529
1530         return hva_to_pfn(addr, atomic, async, write_fault,
1531                           writable);
1532 }
1533 EXPORT_SYMBOL_GPL(__gfn_to_pfn_memslot);
1534
1535 kvm_pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
1536                       bool *writable)
1537 {
1538         return __gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn, false, NULL,
1539                                     write_fault, writable);
1540 }
1541 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
1542
1543 kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
1544 {
1545         return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
1546 }
1547 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot);
1548
1549 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
1550 {
1551         return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
1552 }
1553 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
1554
1555 kvm_pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
1556 {
1557         return gfn_to_pfn_memslot_atomic(gfn_to_memslot(kvm, gfn), gfn);
1558 }
1559 EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
1560
1561 kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn)
1562 {
1563         return gfn_to_pfn_memslot_atomic(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1564 }
1565 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn_atomic);
1566
1567 kvm_pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
1568 {
1569         return gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn);
1570 }
1571 EXPORT_SYMBOL_GPL(gfn_to_pfn);
1572
1573 kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn)
1574 {
1575         return gfn_to_pfn_memslot(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1576 }
1577 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn);
1578
1579 int gfn_to_page_many_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
1580                             struct page **pages, int nr_pages)
1581 {
1582         unsigned long addr;
1583         gfn_t entry;
1584
1585         addr = gfn_to_hva_many(slot, gfn, &entry);
1586         if (kvm_is_error_hva(addr))
1587                 return -1;
1588
1589         if (entry < nr_pages)
1590                 return 0;
1591
1592         return __get_user_pages_fast(addr, nr_pages, 1, pages);
1593 }
1594 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
1595
1596 static struct page *kvm_pfn_to_page(kvm_pfn_t pfn)
1597 {
1598         if (is_error_noslot_pfn(pfn))
1599                 return KVM_ERR_PTR_BAD_PAGE;
1600
1601         if (kvm_is_reserved_pfn(pfn)) {
1602                 WARN_ON(1);
1603                 return KVM_ERR_PTR_BAD_PAGE;
1604         }
1605
1606         return pfn_to_page(pfn);
1607 }
1608
1609 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
1610 {
1611         kvm_pfn_t pfn;
1612
1613         pfn = gfn_to_pfn(kvm, gfn);
1614
1615         return kvm_pfn_to_page(pfn);
1616 }
1617 EXPORT_SYMBOL_GPL(gfn_to_page);
1618
1619 struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1620 {
1621         kvm_pfn_t pfn;
1622
1623         pfn = kvm_vcpu_gfn_to_pfn(vcpu, gfn);
1624
1625         return kvm_pfn_to_page(pfn);
1626 }
1627 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_page);
1628
1629 void kvm_release_page_clean(struct page *page)
1630 {
1631         WARN_ON(is_error_page(page));
1632
1633         kvm_release_pfn_clean(page_to_pfn(page));
1634 }
1635 EXPORT_SYMBOL_GPL(kvm_release_page_clean);
1636
1637 void kvm_release_pfn_clean(kvm_pfn_t pfn)
1638 {
1639         if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn))
1640                 put_page(pfn_to_page(pfn));
1641 }
1642 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
1643
1644 void kvm_release_page_dirty(struct page *page)
1645 {
1646         WARN_ON(is_error_page(page));
1647
1648         kvm_release_pfn_dirty(page_to_pfn(page));
1649 }
1650 EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
1651
1652 static void kvm_release_pfn_dirty(kvm_pfn_t pfn)
1653 {
1654         kvm_set_pfn_dirty(pfn);
1655         kvm_release_pfn_clean(pfn);
1656 }
1657
1658 void kvm_set_pfn_dirty(kvm_pfn_t pfn)
1659 {
1660         if (!kvm_is_reserved_pfn(pfn)) {
1661                 struct page *page = pfn_to_page(pfn);
1662
1663                 if (!PageReserved(page))
1664                         SetPageDirty(page);
1665         }
1666 }
1667 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
1668
1669 void kvm_set_pfn_accessed(kvm_pfn_t pfn)
1670 {
1671         if (!kvm_is_reserved_pfn(pfn))
1672                 mark_page_accessed(pfn_to_page(pfn));
1673 }
1674 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
1675
1676 void kvm_get_pfn(kvm_pfn_t pfn)
1677 {
1678         if (!kvm_is_reserved_pfn(pfn))
1679                 get_page(pfn_to_page(pfn));
1680 }
1681 EXPORT_SYMBOL_GPL(kvm_get_pfn);
1682
1683 static int next_segment(unsigned long len, int offset)
1684 {
1685         if (len > PAGE_SIZE - offset)
1686                 return PAGE_SIZE - offset;
1687         else
1688                 return len;
1689 }
1690
1691 static int __kvm_read_guest_page(struct kvm_memory_slot *slot, gfn_t gfn,
1692                                  void *data, int offset, int len)
1693 {
1694         int r;
1695         unsigned long addr;
1696
1697         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
1698         if (kvm_is_error_hva(addr))
1699                 return -EFAULT;
1700         r = __copy_from_user(data, (void __user *)addr + offset, len);
1701         if (r)
1702                 return -EFAULT;
1703         return 0;
1704 }
1705
1706 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
1707                         int len)
1708 {
1709         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1710
1711         return __kvm_read_guest_page(slot, gfn, data, offset, len);
1712 }
1713 EXPORT_SYMBOL_GPL(kvm_read_guest_page);
1714
1715 int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data,
1716                              int offset, int len)
1717 {
1718         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1719
1720         return __kvm_read_guest_page(slot, gfn, data, offset, len);
1721 }
1722 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_page);
1723
1724 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
1725 {
1726         gfn_t gfn = gpa >> PAGE_SHIFT;
1727         int seg;
1728         int offset = offset_in_page(gpa);
1729         int ret;
1730
1731         while ((seg = next_segment(len, offset)) != 0) {
1732                 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
1733                 if (ret < 0)
1734                         return ret;
1735                 offset = 0;
1736                 len -= seg;
1737                 data += seg;
1738                 ++gfn;
1739         }
1740         return 0;
1741 }
1742 EXPORT_SYMBOL_GPL(kvm_read_guest);
1743
1744 int kvm_vcpu_read_guest(struct kvm_vcpu *vcpu, gpa_t gpa, void *data, unsigned long len)
1745 {
1746         gfn_t gfn = gpa >> PAGE_SHIFT;
1747         int seg;
1748         int offset = offset_in_page(gpa);
1749         int ret;
1750
1751         while ((seg = next_segment(len, offset)) != 0) {
1752                 ret = kvm_vcpu_read_guest_page(vcpu, gfn, data, offset, seg);
1753                 if (ret < 0)
1754                         return ret;
1755                 offset = 0;
1756                 len -= seg;
1757                 data += seg;
1758                 ++gfn;
1759         }
1760         return 0;
1761 }
1762 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest);
1763
1764 static int __kvm_read_guest_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
1765                                    void *data, int offset, unsigned long len)
1766 {
1767         int r;
1768         unsigned long addr;
1769
1770         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
1771         if (kvm_is_error_hva(addr))
1772                 return -EFAULT;
1773         pagefault_disable();
1774         r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len);
1775         pagefault_enable();
1776         if (r)
1777                 return -EFAULT;
1778         return 0;
1779 }
1780
1781 int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data,
1782                           unsigned long len)
1783 {
1784         gfn_t gfn = gpa >> PAGE_SHIFT;
1785         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1786         int offset = offset_in_page(gpa);
1787
1788         return __kvm_read_guest_atomic(slot, gfn, data, offset, len);
1789 }
1790 EXPORT_SYMBOL_GPL(kvm_read_guest_atomic);
1791
1792 int kvm_vcpu_read_guest_atomic(struct kvm_vcpu *vcpu, gpa_t gpa,
1793                                void *data, unsigned long len)
1794 {
1795         gfn_t gfn = gpa >> PAGE_SHIFT;
1796         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1797         int offset = offset_in_page(gpa);
1798
1799         return __kvm_read_guest_atomic(slot, gfn, data, offset, len);
1800 }
1801 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_atomic);
1802
1803 static int __kvm_write_guest_page(struct kvm_memory_slot *memslot, gfn_t gfn,
1804                                   const void *data, int offset, int len)
1805 {
1806         int r;
1807         unsigned long addr;
1808
1809         addr = gfn_to_hva_memslot(memslot, gfn);
1810         if (kvm_is_error_hva(addr))
1811                 return -EFAULT;
1812         r = __copy_to_user((void __user *)addr + offset, data, len);
1813         if (r)
1814                 return -EFAULT;
1815         mark_page_dirty_in_slot(memslot, gfn);
1816         return 0;
1817 }
1818
1819 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn,
1820                          const void *data, int offset, int len)
1821 {
1822         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1823
1824         return __kvm_write_guest_page(slot, gfn, data, offset, len);
1825 }
1826 EXPORT_SYMBOL_GPL(kvm_write_guest_page);
1827
1828 int kvm_vcpu_write_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn,
1829                               const void *data, int offset, int len)
1830 {
1831         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1832
1833         return __kvm_write_guest_page(slot, gfn, data, offset, len);
1834 }
1835 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest_page);
1836
1837 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
1838                     unsigned long len)
1839 {
1840         gfn_t gfn = gpa >> PAGE_SHIFT;
1841         int seg;
1842         int offset = offset_in_page(gpa);
1843         int ret;
1844
1845         while ((seg = next_segment(len, offset)) != 0) {
1846                 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
1847                 if (ret < 0)
1848                         return ret;
1849                 offset = 0;
1850                 len -= seg;
1851                 data += seg;
1852                 ++gfn;
1853         }
1854         return 0;
1855 }
1856 EXPORT_SYMBOL_GPL(kvm_write_guest);
1857
1858 int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data,
1859                          unsigned long len)
1860 {
1861         gfn_t gfn = gpa >> PAGE_SHIFT;
1862         int seg;
1863         int offset = offset_in_page(gpa);
1864         int ret;
1865
1866         while ((seg = next_segment(len, offset)) != 0) {
1867                 ret = kvm_vcpu_write_guest_page(vcpu, gfn, data, offset, seg);
1868                 if (ret < 0)
1869                         return ret;
1870                 offset = 0;
1871                 len -= seg;
1872                 data += seg;
1873                 ++gfn;
1874         }
1875         return 0;
1876 }
1877 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest);
1878
1879 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1880                               gpa_t gpa, unsigned long len)
1881 {
1882         struct kvm_memslots *slots = kvm_memslots(kvm);
1883         int offset = offset_in_page(gpa);
1884         gfn_t start_gfn = gpa >> PAGE_SHIFT;
1885         gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT;
1886         gfn_t nr_pages_needed = end_gfn - start_gfn + 1;
1887         gfn_t nr_pages_avail;
1888
1889         ghc->gpa = gpa;
1890         ghc->generation = slots->generation;
1891         ghc->len = len;
1892         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1893         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn, NULL);
1894         if (!kvm_is_error_hva(ghc->hva) && nr_pages_needed <= 1) {
1895                 ghc->hva += offset;
1896         } else {
1897                 /*
1898                  * If the requested region crosses two memslots, we still
1899                  * verify that the entire region is valid here.
1900                  */
1901                 while (start_gfn <= end_gfn) {
1902                         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1903                         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn,
1904                                                    &nr_pages_avail);
1905                         if (kvm_is_error_hva(ghc->hva))
1906                                 return -EFAULT;
1907                         start_gfn += nr_pages_avail;
1908                 }
1909                 /* Use the slow path for cross page reads and writes. */
1910                 ghc->memslot = NULL;
1911         }
1912         return 0;
1913 }
1914 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
1915
1916 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1917                            void *data, unsigned long len)
1918 {
1919         struct kvm_memslots *slots = kvm_memslots(kvm);
1920         int r;
1921
1922         BUG_ON(len > ghc->len);
1923
1924         if (slots->generation != ghc->generation)
1925                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1926
1927         if (unlikely(!ghc->memslot))
1928                 return kvm_write_guest(kvm, ghc->gpa, data, len);
1929
1930         if (kvm_is_error_hva(ghc->hva))
1931                 return -EFAULT;
1932
1933         r = __copy_to_user((void __user *)ghc->hva, data, len);
1934         if (r)
1935                 return -EFAULT;
1936         mark_page_dirty_in_slot(ghc->memslot, ghc->gpa >> PAGE_SHIFT);
1937
1938         return 0;
1939 }
1940 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
1941
1942 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1943                            void *data, unsigned long len)
1944 {
1945         struct kvm_memslots *slots = kvm_memslots(kvm);
1946         int r;
1947
1948         BUG_ON(len > ghc->len);
1949
1950         if (slots->generation != ghc->generation)
1951                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1952
1953         if (unlikely(!ghc->memslot))
1954                 return kvm_read_guest(kvm, ghc->gpa, data, len);
1955
1956         if (kvm_is_error_hva(ghc->hva))
1957                 return -EFAULT;
1958
1959         r = __copy_from_user(data, (void __user *)ghc->hva, len);
1960         if (r)
1961                 return -EFAULT;
1962
1963         return 0;
1964 }
1965 EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
1966
1967 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
1968 {
1969         const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0)));
1970
1971         return kvm_write_guest_page(kvm, gfn, zero_page, offset, len);
1972 }
1973 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
1974
1975 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
1976 {
1977         gfn_t gfn = gpa >> PAGE_SHIFT;
1978         int seg;
1979         int offset = offset_in_page(gpa);
1980         int ret;
1981
1982         while ((seg = next_segment(len, offset)) != 0) {
1983                 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
1984                 if (ret < 0)
1985                         return ret;
1986                 offset = 0;
1987                 len -= seg;
1988                 ++gfn;
1989         }
1990         return 0;
1991 }
1992 EXPORT_SYMBOL_GPL(kvm_clear_guest);
1993
1994 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot,
1995                                     gfn_t gfn)
1996 {
1997         if (memslot && memslot->dirty_bitmap) {
1998                 unsigned long rel_gfn = gfn - memslot->base_gfn;
1999
2000                 set_bit_le(rel_gfn, memslot->dirty_bitmap);
2001         }
2002 }
2003
2004 void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
2005 {
2006         struct kvm_memory_slot *memslot;
2007
2008         memslot = gfn_to_memslot(kvm, gfn);
2009         mark_page_dirty_in_slot(memslot, gfn);
2010 }
2011 EXPORT_SYMBOL_GPL(mark_page_dirty);
2012
2013 void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
2014 {
2015         struct kvm_memory_slot *memslot;
2016
2017         memslot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2018         mark_page_dirty_in_slot(memslot, gfn);
2019 }
2020 EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
2021
2022 static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
2023 {
2024         unsigned int old, val, grow;
2025
2026         old = val = vcpu->halt_poll_ns;
2027         grow = READ_ONCE(halt_poll_ns_grow);
2028         /* 10us base */
2029         if (val == 0 && grow)
2030                 val = 10000;
2031         else
2032                 val *= grow;
2033
2034         if (val > halt_poll_ns)
2035                 val = halt_poll_ns;
2036
2037         vcpu->halt_poll_ns = val;
2038         trace_kvm_halt_poll_ns_grow(vcpu->vcpu_id, val, old);
2039 }
2040
2041 static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
2042 {
2043         unsigned int old, val, shrink;
2044
2045         old = val = vcpu->halt_poll_ns;
2046         shrink = READ_ONCE(halt_poll_ns_shrink);
2047         if (shrink == 0)
2048                 val = 0;
2049         else
2050                 val /= shrink;
2051
2052         vcpu->halt_poll_ns = val;
2053         trace_kvm_halt_poll_ns_shrink(vcpu->vcpu_id, val, old);
2054 }
2055
2056 static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
2057 {
2058         if (kvm_arch_vcpu_runnable(vcpu)) {
2059                 kvm_make_request(KVM_REQ_UNHALT, vcpu);
2060                 return -EINTR;
2061         }
2062         if (kvm_cpu_has_pending_timer(vcpu))
2063                 return -EINTR;
2064         if (signal_pending(current))
2065                 return -EINTR;
2066
2067         return 0;
2068 }
2069
2070 /*
2071  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
2072  */
2073 void kvm_vcpu_block(struct kvm_vcpu *vcpu)
2074 {
2075         ktime_t start, cur;
2076         DECLARE_SWAITQUEUE(wait);
2077         bool waited = false;
2078         u64 block_ns;
2079
2080         start = cur = ktime_get();
2081         if (vcpu->halt_poll_ns) {
2082                 ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
2083
2084                 ++vcpu->stat.halt_attempted_poll;
2085                 do {
2086                         /*
2087                          * This sets KVM_REQ_UNHALT if an interrupt
2088                          * arrives.
2089                          */
2090                         if (kvm_vcpu_check_block(vcpu) < 0) {
2091                                 ++vcpu->stat.halt_successful_poll;
2092                                 if (!vcpu_valid_wakeup(vcpu))
2093                                         ++vcpu->stat.halt_poll_invalid;
2094                                 goto out;
2095                         }
2096                         cur = ktime_get();
2097                 } while (single_task_running() && ktime_before(cur, stop));
2098         }
2099
2100         kvm_arch_vcpu_blocking(vcpu);
2101
2102         for (;;) {
2103                 prepare_to_swait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
2104
2105                 if (kvm_vcpu_check_block(vcpu) < 0)
2106                         break;
2107
2108                 waited = true;
2109                 schedule();
2110         }
2111
2112         finish_swait(&vcpu->wq, &wait);
2113         cur = ktime_get();
2114
2115         kvm_arch_vcpu_unblocking(vcpu);
2116 out:
2117         block_ns = ktime_to_ns(cur) - ktime_to_ns(start);
2118
2119         if (!vcpu_valid_wakeup(vcpu))
2120                 shrink_halt_poll_ns(vcpu);
2121         else if (halt_poll_ns) {
2122                 if (block_ns <= vcpu->halt_poll_ns)
2123                         ;
2124                 /* we had a long block, shrink polling */
2125                 else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
2126                         shrink_halt_poll_ns(vcpu);
2127                 /* we had a short halt and our poll time is too small */
2128                 else if (vcpu->halt_poll_ns < halt_poll_ns &&
2129                         block_ns < halt_poll_ns)
2130                         grow_halt_poll_ns(vcpu);
2131         } else
2132                 vcpu->halt_poll_ns = 0;
2133
2134         trace_kvm_vcpu_wakeup(block_ns, waited, vcpu_valid_wakeup(vcpu));
2135         kvm_arch_vcpu_block_finish(vcpu);
2136 }
2137 EXPORT_SYMBOL_GPL(kvm_vcpu_block);
2138
2139 #ifndef CONFIG_S390
2140 void kvm_vcpu_wake_up(struct kvm_vcpu *vcpu)
2141 {
2142         struct swait_queue_head *wqp;
2143
2144         wqp = kvm_arch_vcpu_wq(vcpu);
2145         if (swait_active(wqp)) {
2146                 swake_up(wqp);
2147                 ++vcpu->stat.halt_wakeup;
2148         }
2149
2150 }
2151 EXPORT_SYMBOL_GPL(kvm_vcpu_wake_up);
2152
2153 /*
2154  * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode.
2155  */
2156 void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
2157 {
2158         int me;
2159         int cpu = vcpu->cpu;
2160
2161         kvm_vcpu_wake_up(vcpu);
2162         me = get_cpu();
2163         if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
2164                 if (kvm_arch_vcpu_should_kick(vcpu))
2165                         smp_send_reschedule(cpu);
2166         put_cpu();
2167 }
2168 EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
2169 #endif /* !CONFIG_S390 */
2170
2171 int kvm_vcpu_yield_to(struct kvm_vcpu *target)
2172 {
2173         struct pid *pid;
2174         struct task_struct *task = NULL;
2175         int ret = 0;
2176
2177         rcu_read_lock();
2178         pid = rcu_dereference(target->pid);
2179         if (pid)
2180                 task = get_pid_task(pid, PIDTYPE_PID);
2181         rcu_read_unlock();
2182         if (!task)
2183                 return ret;
2184         ret = yield_to(task, 1);
2185         put_task_struct(task);
2186
2187         return ret;
2188 }
2189 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to);
2190
2191 /*
2192  * Helper that checks whether a VCPU is eligible for directed yield.
2193  * Most eligible candidate to yield is decided by following heuristics:
2194  *
2195  *  (a) VCPU which has not done pl-exit or cpu relax intercepted recently
2196  *  (preempted lock holder), indicated by @in_spin_loop.
2197  *  Set at the beiginning and cleared at the end of interception/PLE handler.
2198  *
2199  *  (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get
2200  *  chance last time (mostly it has become eligible now since we have probably
2201  *  yielded to lockholder in last iteration. This is done by toggling
2202  *  @dy_eligible each time a VCPU checked for eligibility.)
2203  *
2204  *  Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding
2205  *  to preempted lock-holder could result in wrong VCPU selection and CPU
2206  *  burning. Giving priority for a potential lock-holder increases lock
2207  *  progress.
2208  *
2209  *  Since algorithm is based on heuristics, accessing another VCPU data without
2210  *  locking does not harm. It may result in trying to yield to  same VCPU, fail
2211  *  and continue with next VCPU and so on.
2212  */
2213 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu)
2214 {
2215 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT
2216         bool eligible;
2217
2218         eligible = !vcpu->spin_loop.in_spin_loop ||
2219                     vcpu->spin_loop.dy_eligible;
2220
2221         if (vcpu->spin_loop.in_spin_loop)
2222                 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible);
2223
2224         return eligible;
2225 #else
2226         return true;
2227 #endif
2228 }
2229
2230 void kvm_vcpu_on_spin(struct kvm_vcpu *me)
2231 {
2232         struct kvm *kvm = me->kvm;
2233         struct kvm_vcpu *vcpu;
2234         int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
2235         int yielded = 0;
2236         int try = 3;
2237         int pass;
2238         int i;
2239
2240         kvm_vcpu_set_in_spin_loop(me, true);
2241         /*
2242          * We boost the priority of a VCPU that is runnable but not
2243          * currently running, because it got preempted by something
2244          * else and called schedule in __vcpu_run.  Hopefully that
2245          * VCPU is holding the lock that we need and will release it.
2246          * We approximate round-robin by starting at the last boosted VCPU.
2247          */
2248         for (pass = 0; pass < 2 && !yielded && try; pass++) {
2249                 kvm_for_each_vcpu(i, vcpu, kvm) {
2250                         if (!pass && i <= last_boosted_vcpu) {
2251                                 i = last_boosted_vcpu;
2252                                 continue;
2253                         } else if (pass && i > last_boosted_vcpu)
2254                                 break;
2255                         if (!ACCESS_ONCE(vcpu->preempted))
2256                                 continue;
2257                         if (vcpu == me)
2258                                 continue;
2259                         if (swait_active(&vcpu->wq) && !kvm_arch_vcpu_runnable(vcpu))
2260                                 continue;
2261                         if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
2262                                 continue;
2263
2264                         yielded = kvm_vcpu_yield_to(vcpu);
2265                         if (yielded > 0) {
2266                                 kvm->last_boosted_vcpu = i;
2267                                 break;
2268                         } else if (yielded < 0) {
2269                                 try--;
2270                                 if (!try)
2271                                         break;
2272                         }
2273                 }
2274         }
2275         kvm_vcpu_set_in_spin_loop(me, false);
2276
2277         /* Ensure vcpu is not eligible during next spinloop */
2278         kvm_vcpu_set_dy_eligible(me, false);
2279 }
2280 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
2281
2282 static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2283 {
2284         struct kvm_vcpu *vcpu = vma->vm_file->private_data;
2285         struct page *page;
2286
2287         if (vmf->pgoff == 0)
2288                 page = virt_to_page(vcpu->run);
2289 #ifdef CONFIG_X86
2290         else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
2291                 page = virt_to_page(vcpu->arch.pio_data);
2292 #endif
2293 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2294         else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
2295                 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
2296 #endif
2297         else
2298                 return kvm_arch_vcpu_fault(vcpu, vmf);
2299         get_page(page);
2300         vmf->page = page;
2301         return 0;
2302 }
2303
2304 static const struct vm_operations_struct kvm_vcpu_vm_ops = {
2305         .fault = kvm_vcpu_fault,
2306 };
2307
2308 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
2309 {
2310         vma->vm_ops = &kvm_vcpu_vm_ops;
2311         return 0;
2312 }
2313
2314 static int kvm_vcpu_release(struct inode *inode, struct file *filp)
2315 {
2316         struct kvm_vcpu *vcpu = filp->private_data;
2317
2318         kvm_put_kvm(vcpu->kvm);
2319         return 0;
2320 }
2321
2322 static struct file_operations kvm_vcpu_fops = {
2323         .release        = kvm_vcpu_release,
2324         .unlocked_ioctl = kvm_vcpu_ioctl,
2325 #ifdef CONFIG_KVM_COMPAT
2326         .compat_ioctl   = kvm_vcpu_compat_ioctl,
2327 #endif
2328         .mmap           = kvm_vcpu_mmap,
2329         .llseek         = noop_llseek,
2330 };
2331
2332 /*
2333  * Allocates an inode for the vcpu.
2334  */
2335 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
2336 {
2337         return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC);
2338 }
2339
2340 /*
2341  * Creates some virtual cpus.  Good luck creating more than one.
2342  */
2343 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
2344 {
2345         int r;
2346         struct kvm_vcpu *vcpu;
2347
2348         if (id >= KVM_MAX_VCPU_ID)
2349                 return -EINVAL;
2350
2351         vcpu = kvm_arch_vcpu_create(kvm, id);
2352         if (IS_ERR(vcpu))
2353                 return PTR_ERR(vcpu);
2354
2355         preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
2356
2357         r = kvm_arch_vcpu_setup(vcpu);
2358         if (r)
2359                 goto vcpu_destroy;
2360
2361         mutex_lock(&kvm->lock);
2362         if (!kvm_vcpu_compatible(vcpu)) {
2363                 r = -EINVAL;
2364                 goto unlock_vcpu_destroy;
2365         }
2366         if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
2367                 r = -EINVAL;
2368                 goto unlock_vcpu_destroy;
2369         }
2370         if (kvm_get_vcpu_by_id(kvm, id)) {
2371                 r = -EEXIST;
2372                 goto unlock_vcpu_destroy;
2373         }
2374
2375         BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]);
2376
2377         /* Now it's all set up, let userspace reach it */
2378         kvm_get_kvm(kvm);
2379         r = create_vcpu_fd(vcpu);
2380         if (r < 0) {
2381                 kvm_put_kvm(kvm);
2382                 goto unlock_vcpu_destroy;
2383         }
2384
2385         kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu;
2386
2387         /*
2388          * Pairs with smp_rmb() in kvm_get_vcpu.  Write kvm->vcpus
2389          * before kvm->online_vcpu's incremented value.
2390          */
2391         smp_wmb();
2392         atomic_inc(&kvm->online_vcpus);
2393
2394         mutex_unlock(&kvm->lock);
2395         kvm_arch_vcpu_postcreate(vcpu);
2396         return r;
2397
2398 unlock_vcpu_destroy:
2399         mutex_unlock(&kvm->lock);
2400 vcpu_destroy:
2401         kvm_arch_vcpu_destroy(vcpu);
2402         return r;
2403 }
2404
2405 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
2406 {
2407         if (sigset) {
2408                 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
2409                 vcpu->sigset_active = 1;
2410                 vcpu->sigset = *sigset;
2411         } else
2412                 vcpu->sigset_active = 0;
2413         return 0;
2414 }
2415
2416 static long kvm_vcpu_ioctl(struct file *filp,
2417                            unsigned int ioctl, unsigned long arg)
2418 {
2419         struct kvm_vcpu *vcpu = filp->private_data;
2420         void __user *argp = (void __user *)arg;
2421         int r;
2422         struct kvm_fpu *fpu = NULL;
2423         struct kvm_sregs *kvm_sregs = NULL;
2424
2425         if (vcpu->kvm->mm != current->mm)
2426                 return -EIO;
2427
2428         if (unlikely(_IOC_TYPE(ioctl) != KVMIO))
2429                 return -EINVAL;
2430
2431 #if defined(CONFIG_S390) || defined(CONFIG_PPC) || defined(CONFIG_MIPS)
2432         /*
2433          * Special cases: vcpu ioctls that are asynchronous to vcpu execution,
2434          * so vcpu_load() would break it.
2435          */
2436         if (ioctl == KVM_S390_INTERRUPT || ioctl == KVM_S390_IRQ || ioctl == KVM_INTERRUPT)
2437                 return kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2438 #endif
2439
2440
2441         r = vcpu_load(vcpu);
2442         if (r)
2443                 return r;
2444         switch (ioctl) {
2445         case KVM_RUN:
2446                 r = -EINVAL;
2447                 if (arg)
2448                         goto out;
2449                 if (unlikely(vcpu->pid != current->pids[PIDTYPE_PID].pid)) {
2450                         /* The thread running this VCPU changed. */
2451                         struct pid *oldpid = vcpu->pid;
2452                         struct pid *newpid = get_task_pid(current, PIDTYPE_PID);
2453
2454                         rcu_assign_pointer(vcpu->pid, newpid);
2455                         if (oldpid)
2456                                 synchronize_rcu();
2457                         put_pid(oldpid);
2458                 }
2459                 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
2460                 trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
2461                 break;
2462         case KVM_GET_REGS: {
2463                 struct kvm_regs *kvm_regs;
2464
2465                 r = -ENOMEM;
2466                 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL);
2467                 if (!kvm_regs)
2468                         goto out;
2469                 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
2470                 if (r)
2471                         goto out_free1;
2472                 r = -EFAULT;
2473                 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
2474                         goto out_free1;
2475                 r = 0;
2476 out_free1:
2477                 kfree(kvm_regs);
2478                 break;
2479         }
2480         case KVM_SET_REGS: {
2481                 struct kvm_regs *kvm_regs;
2482
2483                 r = -ENOMEM;
2484                 kvm_regs = memdup_user(argp, sizeof(*kvm_regs));
2485                 if (IS_ERR(kvm_regs)) {
2486                         r = PTR_ERR(kvm_regs);
2487                         goto out;
2488                 }
2489                 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
2490                 kfree(kvm_regs);
2491                 break;
2492         }
2493         case KVM_GET_SREGS: {
2494                 kvm_sregs = kzalloc(sizeof(struct kvm_sregs), GFP_KERNEL);
2495                 r = -ENOMEM;
2496                 if (!kvm_sregs)
2497                         goto out;
2498                 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
2499                 if (r)
2500                         goto out;
2501                 r = -EFAULT;
2502                 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
2503                         goto out;
2504                 r = 0;
2505                 break;
2506         }
2507         case KVM_SET_SREGS: {
2508                 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs));
2509                 if (IS_ERR(kvm_sregs)) {
2510                         r = PTR_ERR(kvm_sregs);
2511                         kvm_sregs = NULL;
2512                         goto out;
2513                 }
2514                 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
2515                 break;
2516         }
2517         case KVM_GET_MP_STATE: {
2518                 struct kvm_mp_state mp_state;
2519
2520                 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
2521                 if (r)
2522                         goto out;
2523                 r = -EFAULT;
2524                 if (copy_to_user(argp, &mp_state, sizeof(mp_state)))
2525                         goto out;
2526                 r = 0;
2527                 break;
2528         }
2529         case KVM_SET_MP_STATE: {
2530                 struct kvm_mp_state mp_state;
2531
2532                 r = -EFAULT;
2533                 if (copy_from_user(&mp_state, argp, sizeof(mp_state)))
2534                         goto out;
2535                 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
2536                 break;
2537         }
2538         case KVM_TRANSLATE: {
2539                 struct kvm_translation tr;
2540
2541                 r = -EFAULT;
2542                 if (copy_from_user(&tr, argp, sizeof(tr)))
2543                         goto out;
2544                 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
2545                 if (r)
2546                         goto out;
2547                 r = -EFAULT;
2548                 if (copy_to_user(argp, &tr, sizeof(tr)))
2549                         goto out;
2550                 r = 0;
2551                 break;
2552         }
2553         case KVM_SET_GUEST_DEBUG: {
2554                 struct kvm_guest_debug dbg;
2555
2556                 r = -EFAULT;
2557                 if (copy_from_user(&dbg, argp, sizeof(dbg)))
2558                         goto out;
2559                 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
2560                 break;
2561         }
2562         case KVM_SET_SIGNAL_MASK: {
2563                 struct kvm_signal_mask __user *sigmask_arg = argp;
2564                 struct kvm_signal_mask kvm_sigmask;
2565                 sigset_t sigset, *p;
2566
2567                 p = NULL;
2568                 if (argp) {
2569                         r = -EFAULT;
2570                         if (copy_from_user(&kvm_sigmask, argp,
2571                                            sizeof(kvm_sigmask)))
2572                                 goto out;
2573                         r = -EINVAL;
2574                         if (kvm_sigmask.len != sizeof(sigset))
2575                                 goto out;
2576                         r = -EFAULT;
2577                         if (copy_from_user(&sigset, sigmask_arg->sigset,
2578                                            sizeof(sigset)))
2579                                 goto out;
2580                         p = &sigset;
2581                 }
2582                 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
2583                 break;
2584         }
2585         case KVM_GET_FPU: {
2586                 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL);
2587                 r = -ENOMEM;
2588                 if (!fpu)
2589                         goto out;
2590                 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
2591                 if (r)
2592                         goto out;
2593                 r = -EFAULT;
2594                 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
2595                         goto out;
2596                 r = 0;
2597                 break;
2598         }
2599         case KVM_SET_FPU: {
2600                 fpu = memdup_user(argp, sizeof(*fpu));
2601                 if (IS_ERR(fpu)) {
2602                         r = PTR_ERR(fpu);
2603                         fpu = NULL;
2604                         goto out;
2605                 }
2606                 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
2607                 break;
2608         }
2609         default:
2610                 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2611         }
2612 out:
2613         vcpu_put(vcpu);
2614         kfree(fpu);
2615         kfree(kvm_sregs);
2616         return r;
2617 }
2618
2619 #ifdef CONFIG_KVM_COMPAT
2620 static long kvm_vcpu_compat_ioctl(struct file *filp,
2621                                   unsigned int ioctl, unsigned long arg)
2622 {
2623         struct kvm_vcpu *vcpu = filp->private_data;
2624         void __user *argp = compat_ptr(arg);
2625         int r;
2626
2627         if (vcpu->kvm->mm != current->mm)
2628                 return -EIO;
2629
2630         switch (ioctl) {
2631         case KVM_SET_SIGNAL_MASK: {
2632                 struct kvm_signal_mask __user *sigmask_arg = argp;
2633                 struct kvm_signal_mask kvm_sigmask;
2634                 compat_sigset_t csigset;
2635                 sigset_t sigset;
2636
2637                 if (argp) {
2638                         r = -EFAULT;
2639                         if (copy_from_user(&kvm_sigmask, argp,
2640                                            sizeof(kvm_sigmask)))
2641                                 goto out;
2642                         r = -EINVAL;
2643                         if (kvm_sigmask.len != sizeof(csigset))
2644                                 goto out;
2645                         r = -EFAULT;
2646                         if (copy_from_user(&csigset, sigmask_arg->sigset,
2647                                            sizeof(csigset)))
2648                                 goto out;
2649                         sigset_from_compat(&sigset, &csigset);
2650                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
2651                 } else
2652                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL);
2653                 break;
2654         }
2655         default:
2656                 r = kvm_vcpu_ioctl(filp, ioctl, arg);
2657         }
2658
2659 out:
2660         return r;
2661 }
2662 #endif
2663
2664 static int kvm_device_ioctl_attr(struct kvm_device *dev,
2665                                  int (*accessor)(struct kvm_device *dev,
2666                                                  struct kvm_device_attr *attr),
2667                                  unsigned long arg)
2668 {
2669         struct kvm_device_attr attr;
2670
2671         if (!accessor)
2672                 return -EPERM;
2673
2674         if (copy_from_user(&attr, (void __user *)arg, sizeof(attr)))
2675                 return -EFAULT;
2676
2677         return accessor(dev, &attr);
2678 }
2679
2680 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl,
2681                              unsigned long arg)
2682 {
2683         struct kvm_device *dev = filp->private_data;
2684
2685         switch (ioctl) {
2686         case KVM_SET_DEVICE_ATTR:
2687                 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg);
2688         case KVM_GET_DEVICE_ATTR:
2689                 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg);
2690         case KVM_HAS_DEVICE_ATTR:
2691                 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg);
2692         default:
2693                 if (dev->ops->ioctl)
2694                         return dev->ops->ioctl(dev, ioctl, arg);
2695
2696                 return -ENOTTY;
2697         }
2698 }
2699
2700 static int kvm_device_release(struct inode *inode, struct file *filp)
2701 {
2702         struct kvm_device *dev = filp->private_data;
2703         struct kvm *kvm = dev->kvm;
2704
2705         kvm_put_kvm(kvm);
2706         return 0;
2707 }
2708
2709 static const struct file_operations kvm_device_fops = {
2710         .unlocked_ioctl = kvm_device_ioctl,
2711 #ifdef CONFIG_KVM_COMPAT
2712         .compat_ioctl = kvm_device_ioctl,
2713 #endif
2714         .release = kvm_device_release,
2715 };
2716
2717 struct kvm_device *kvm_device_from_filp(struct file *filp)
2718 {
2719         if (filp->f_op != &kvm_device_fops)
2720                 return NULL;
2721
2722         return filp->private_data;
2723 }
2724
2725 static struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = {
2726 #ifdef CONFIG_KVM_MPIC
2727         [KVM_DEV_TYPE_FSL_MPIC_20]      = &kvm_mpic_ops,
2728         [KVM_DEV_TYPE_FSL_MPIC_42]      = &kvm_mpic_ops,
2729 #endif
2730
2731 #ifdef CONFIG_KVM_XICS
2732         [KVM_DEV_TYPE_XICS]             = &kvm_xics_ops,
2733 #endif
2734 };
2735
2736 int kvm_register_device_ops(struct kvm_device_ops *ops, u32 type)
2737 {
2738         if (type >= ARRAY_SIZE(kvm_device_ops_table))
2739                 return -ENOSPC;
2740
2741         if (kvm_device_ops_table[type] != NULL)
2742                 return -EEXIST;
2743
2744         kvm_device_ops_table[type] = ops;
2745         return 0;
2746 }
2747
2748 void kvm_unregister_device_ops(u32 type)
2749 {
2750         if (kvm_device_ops_table[type] != NULL)
2751                 kvm_device_ops_table[type] = NULL;
2752 }
2753
2754 static int kvm_ioctl_create_device(struct kvm *kvm,
2755                                    struct kvm_create_device *cd)
2756 {
2757         struct kvm_device_ops *ops = NULL;
2758         struct kvm_device *dev;
2759         bool test = cd->flags & KVM_CREATE_DEVICE_TEST;
2760         int ret;
2761
2762         if (cd->type >= ARRAY_SIZE(kvm_device_ops_table))
2763                 return -ENODEV;
2764
2765         ops = kvm_device_ops_table[cd->type];
2766         if (ops == NULL)
2767                 return -ENODEV;
2768
2769         if (test)
2770                 return 0;
2771
2772         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2773         if (!dev)
2774                 return -ENOMEM;
2775
2776         dev->ops = ops;
2777         dev->kvm = kvm;
2778
2779         ret = ops->create(dev, cd->type);
2780         if (ret < 0) {
2781                 kfree(dev);
2782                 return ret;
2783         }
2784
2785         ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC);
2786         if (ret < 0) {
2787                 ops->destroy(dev);
2788                 return ret;
2789         }
2790
2791         list_add(&dev->vm_node, &kvm->devices);
2792         kvm_get_kvm(kvm);
2793         cd->fd = ret;
2794         return 0;
2795 }
2796
2797 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
2798 {
2799         switch (arg) {
2800         case KVM_CAP_USER_MEMORY:
2801         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
2802         case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
2803         case KVM_CAP_INTERNAL_ERROR_DATA:
2804 #ifdef CONFIG_HAVE_KVM_MSI
2805         case KVM_CAP_SIGNAL_MSI:
2806 #endif
2807 #ifdef CONFIG_HAVE_KVM_IRQFD
2808         case KVM_CAP_IRQFD:
2809         case KVM_CAP_IRQFD_RESAMPLE:
2810 #endif
2811         case KVM_CAP_IOEVENTFD_ANY_LENGTH:
2812         case KVM_CAP_CHECK_EXTENSION_VM:
2813                 return 1;
2814 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2815         case KVM_CAP_IRQ_ROUTING:
2816                 return KVM_MAX_IRQ_ROUTES;
2817 #endif
2818 #if KVM_ADDRESS_SPACE_NUM > 1
2819         case KVM_CAP_MULTI_ADDRESS_SPACE:
2820                 return KVM_ADDRESS_SPACE_NUM;
2821 #endif
2822         case KVM_CAP_MAX_VCPU_ID:
2823                 return KVM_MAX_VCPU_ID;
2824         default:
2825                 break;
2826         }
2827         return kvm_vm_ioctl_check_extension(kvm, arg);
2828 }
2829
2830 static long kvm_vm_ioctl(struct file *filp,
2831                            unsigned int ioctl, unsigned long arg)
2832 {
2833         struct kvm *kvm = filp->private_data;
2834         void __user *argp = (void __user *)arg;
2835         int r;
2836
2837         if (kvm->mm != current->mm)
2838                 return -EIO;
2839         switch (ioctl) {
2840         case KVM_CREATE_VCPU:
2841                 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
2842                 break;
2843         case KVM_SET_USER_MEMORY_REGION: {
2844                 struct kvm_userspace_memory_region kvm_userspace_mem;
2845
2846                 r = -EFAULT;
2847                 if (copy_from_user(&kvm_userspace_mem, argp,
2848                                                 sizeof(kvm_userspace_mem)))
2849                         goto out;
2850
2851                 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem);
2852                 break;
2853         }
2854         case KVM_GET_DIRTY_LOG: {
2855                 struct kvm_dirty_log log;
2856
2857                 r = -EFAULT;
2858                 if (copy_from_user(&log, argp, sizeof(log)))
2859                         goto out;
2860                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2861                 break;
2862         }
2863 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2864         case KVM_REGISTER_COALESCED_MMIO: {
2865                 struct kvm_coalesced_mmio_zone zone;
2866
2867                 r = -EFAULT;
2868                 if (copy_from_user(&zone, argp, sizeof(zone)))
2869                         goto out;
2870                 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
2871                 break;
2872         }
2873         case KVM_UNREGISTER_COALESCED_MMIO: {
2874                 struct kvm_coalesced_mmio_zone zone;
2875
2876                 r = -EFAULT;
2877                 if (copy_from_user(&zone, argp, sizeof(zone)))
2878                         goto out;
2879                 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
2880                 break;
2881         }
2882 #endif
2883         case KVM_IRQFD: {
2884                 struct kvm_irqfd data;
2885
2886                 r = -EFAULT;
2887                 if (copy_from_user(&data, argp, sizeof(data)))
2888                         goto out;
2889                 r = kvm_irqfd(kvm, &data);
2890                 break;
2891         }
2892         case KVM_IOEVENTFD: {
2893                 struct kvm_ioeventfd data;
2894
2895                 r = -EFAULT;
2896                 if (copy_from_user(&data, argp, sizeof(data)))
2897                         goto out;
2898                 r = kvm_ioeventfd(kvm, &data);
2899                 break;
2900         }
2901 #ifdef CONFIG_HAVE_KVM_MSI
2902         case KVM_SIGNAL_MSI: {
2903                 struct kvm_msi msi;
2904
2905                 r = -EFAULT;
2906                 if (copy_from_user(&msi, argp, sizeof(msi)))
2907                         goto out;
2908                 r = kvm_send_userspace_msi(kvm, &msi);
2909                 break;
2910         }
2911 #endif
2912 #ifdef __KVM_HAVE_IRQ_LINE
2913         case KVM_IRQ_LINE_STATUS:
2914         case KVM_IRQ_LINE: {
2915                 struct kvm_irq_level irq_event;
2916
2917                 r = -EFAULT;
2918                 if (copy_from_user(&irq_event, argp, sizeof(irq_event)))
2919                         goto out;
2920
2921                 r = kvm_vm_ioctl_irq_line(kvm, &irq_event,
2922                                         ioctl == KVM_IRQ_LINE_STATUS);
2923                 if (r)
2924                         goto out;
2925
2926                 r = -EFAULT;
2927                 if (ioctl == KVM_IRQ_LINE_STATUS) {
2928                         if (copy_to_user(argp, &irq_event, sizeof(irq_event)))
2929                                 goto out;
2930                 }
2931
2932                 r = 0;
2933                 break;
2934         }
2935 #endif
2936 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2937         case KVM_SET_GSI_ROUTING: {
2938                 struct kvm_irq_routing routing;
2939                 struct kvm_irq_routing __user *urouting;
2940                 struct kvm_irq_routing_entry *entries = NULL;
2941
2942                 r = -EFAULT;
2943                 if (copy_from_user(&routing, argp, sizeof(routing)))
2944                         goto out;
2945                 r = -EINVAL;
2946                 if (routing.nr > KVM_MAX_IRQ_ROUTES)
2947                         goto out;
2948                 if (routing.flags)
2949                         goto out;
2950                 if (routing.nr) {
2951                         r = -ENOMEM;
2952                         entries = vmalloc(routing.nr * sizeof(*entries));
2953                         if (!entries)
2954                                 goto out;
2955                         r = -EFAULT;
2956                         urouting = argp;
2957                         if (copy_from_user(entries, urouting->entries,
2958                                            routing.nr * sizeof(*entries)))
2959                                 goto out_free_irq_routing;
2960                 }
2961                 r = kvm_set_irq_routing(kvm, entries, routing.nr,
2962                                         routing.flags);
2963 out_free_irq_routing:
2964                 vfree(entries);
2965                 break;
2966         }
2967 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */
2968         case KVM_CREATE_DEVICE: {
2969                 struct kvm_create_device cd;
2970
2971                 r = -EFAULT;
2972                 if (copy_from_user(&cd, argp, sizeof(cd)))
2973                         goto out;
2974
2975                 r = kvm_ioctl_create_device(kvm, &cd);
2976                 if (r)
2977                         goto out;
2978
2979                 r = -EFAULT;
2980                 if (copy_to_user(argp, &cd, sizeof(cd)))
2981                         goto out;
2982
2983                 r = 0;
2984                 break;
2985         }
2986         case KVM_CHECK_EXTENSION:
2987                 r = kvm_vm_ioctl_check_extension_generic(kvm, arg);
2988                 break;
2989         default:
2990                 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
2991         }
2992 out:
2993         return r;
2994 }
2995
2996 #ifdef CONFIG_KVM_COMPAT
2997 struct compat_kvm_dirty_log {
2998         __u32 slot;
2999         __u32 padding1;
3000         union {
3001                 compat_uptr_t dirty_bitmap; /* one bit per page */
3002                 __u64 padding2;
3003         };
3004 };
3005
3006 static long kvm_vm_compat_ioctl(struct file *filp,
3007                            unsigned int ioctl, unsigned long arg)
3008 {
3009         struct kvm *kvm = filp->private_data;
3010         int r;
3011
3012         if (kvm->mm != current->mm)
3013                 return -EIO;
3014         switch (ioctl) {
3015         case KVM_GET_DIRTY_LOG: {
3016                 struct compat_kvm_dirty_log compat_log;
3017                 struct kvm_dirty_log log;
3018
3019                 r = -EFAULT;
3020                 if (copy_from_user(&compat_log, (void __user *)arg,
3021                                    sizeof(compat_log)))
3022                         goto out;
3023                 log.slot         = compat_log.slot;
3024                 log.padding1     = compat_log.padding1;
3025                 log.padding2     = compat_log.padding2;
3026                 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
3027
3028                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
3029                 break;
3030         }
3031         default:
3032                 r = kvm_vm_ioctl(filp, ioctl, arg);
3033         }
3034
3035 out:
3036         return r;
3037 }
3038 #endif
3039
3040 static struct file_operations kvm_vm_fops = {
3041         .release        = kvm_vm_release,
3042         .unlocked_ioctl = kvm_vm_ioctl,
3043 #ifdef CONFIG_KVM_COMPAT
3044         .compat_ioctl   = kvm_vm_compat_ioctl,
3045 #endif
3046         .llseek         = noop_llseek,
3047 };
3048
3049 static int kvm_dev_ioctl_create_vm(unsigned long type)
3050 {
3051         int r;
3052         struct kvm *kvm;
3053         struct file *file;
3054
3055         kvm = kvm_create_vm(type);
3056         if (IS_ERR(kvm))
3057                 return PTR_ERR(kvm);
3058 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
3059         r = kvm_coalesced_mmio_init(kvm);
3060         if (r < 0) {
3061                 kvm_put_kvm(kvm);
3062                 return r;
3063         }
3064 #endif
3065         r = get_unused_fd_flags(O_CLOEXEC);
3066         if (r < 0) {
3067                 kvm_put_kvm(kvm);
3068                 return r;
3069         }
3070         file = anon_inode_getfile("kvm-vm", &kvm_vm_fops, kvm, O_RDWR);
3071         if (IS_ERR(file)) {
3072                 put_unused_fd(r);
3073                 kvm_put_kvm(kvm);
3074                 return PTR_ERR(file);
3075         }
3076
3077         if (kvm_create_vm_debugfs(kvm, r) < 0) {
3078                 put_unused_fd(r);
3079                 fput(file);
3080                 return -ENOMEM;
3081         }
3082
3083         fd_install(r, file);
3084         return r;
3085 }
3086
3087 static long kvm_dev_ioctl(struct file *filp,
3088                           unsigned int ioctl, unsigned long arg)
3089 {
3090         long r = -EINVAL;
3091
3092         switch (ioctl) {
3093         case KVM_GET_API_VERSION:
3094                 if (arg)
3095                         goto out;
3096                 r = KVM_API_VERSION;
3097                 break;
3098         case KVM_CREATE_VM:
3099                 r = kvm_dev_ioctl_create_vm(arg);
3100                 break;
3101         case KVM_CHECK_EXTENSION:
3102                 r = kvm_vm_ioctl_check_extension_generic(NULL, arg);
3103                 break;
3104         case KVM_GET_VCPU_MMAP_SIZE:
3105                 if (arg)
3106                         goto out;
3107                 r = PAGE_SIZE;     /* struct kvm_run */
3108 #ifdef CONFIG_X86
3109                 r += PAGE_SIZE;    /* pio data page */
3110 #endif
3111 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
3112                 r += PAGE_SIZE;    /* coalesced mmio ring page */
3113 #endif
3114                 break;
3115         case KVM_TRACE_ENABLE:
3116         case KVM_TRACE_PAUSE:
3117         case KVM_TRACE_DISABLE:
3118                 r = -EOPNOTSUPP;
3119                 break;
3120         default:
3121                 return kvm_arch_dev_ioctl(filp, ioctl, arg);
3122         }
3123 out:
3124         return r;
3125 }
3126
3127 static struct file_operations kvm_chardev_ops = {
3128         .unlocked_ioctl = kvm_dev_ioctl,
3129         .compat_ioctl   = kvm_dev_ioctl,
3130         .llseek         = noop_llseek,
3131 };
3132
3133 static struct miscdevice kvm_dev = {
3134         KVM_MINOR,
3135         "kvm",
3136         &kvm_chardev_ops,
3137 };
3138
3139 static void hardware_enable_nolock(void *junk)
3140 {
3141         int cpu = raw_smp_processor_id();
3142         int r;
3143
3144         if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
3145                 return;
3146
3147         cpumask_set_cpu(cpu, cpus_hardware_enabled);
3148
3149         r = kvm_arch_hardware_enable();
3150
3151         if (r) {
3152                 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3153                 atomic_inc(&hardware_enable_failed);
3154                 pr_info("kvm: enabling virtualization on CPU%d failed\n", cpu);
3155         }
3156 }
3157
3158 static int kvm_starting_cpu(unsigned int cpu)
3159 {
3160         raw_spin_lock(&kvm_count_lock);
3161         if (kvm_usage_count)
3162                 hardware_enable_nolock(NULL);
3163         raw_spin_unlock(&kvm_count_lock);
3164         return 0;
3165 }
3166
3167 static void hardware_disable_nolock(void *junk)
3168 {
3169         int cpu = raw_smp_processor_id();
3170
3171         if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
3172                 return;
3173         cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3174         kvm_arch_hardware_disable();
3175 }
3176
3177 static int kvm_dying_cpu(unsigned int cpu)
3178 {
3179         raw_spin_lock(&kvm_count_lock);
3180         if (kvm_usage_count)
3181                 hardware_disable_nolock(NULL);
3182         raw_spin_unlock(&kvm_count_lock);
3183         return 0;
3184 }
3185
3186 static void hardware_disable_all_nolock(void)
3187 {
3188         BUG_ON(!kvm_usage_count);
3189
3190         kvm_usage_count--;
3191         if (!kvm_usage_count)
3192                 on_each_cpu(hardware_disable_nolock, NULL, 1);
3193 }
3194
3195 static void hardware_disable_all(void)
3196 {
3197         raw_spin_lock(&kvm_count_lock);
3198         hardware_disable_all_nolock();
3199         raw_spin_unlock(&kvm_count_lock);
3200 }
3201
3202 static int hardware_enable_all(void)
3203 {
3204         int r = 0;
3205
3206         raw_spin_lock(&kvm_count_lock);
3207
3208         kvm_usage_count++;
3209         if (kvm_usage_count == 1) {
3210                 atomic_set(&hardware_enable_failed, 0);
3211                 on_each_cpu(hardware_enable_nolock, NULL, 1);
3212
3213                 if (atomic_read(&hardware_enable_failed)) {
3214                         hardware_disable_all_nolock();
3215                         r = -EBUSY;
3216                 }
3217         }
3218
3219         raw_spin_unlock(&kvm_count_lock);
3220
3221         return r;
3222 }
3223
3224 static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
3225                       void *v)
3226 {
3227         /*
3228          * Some (well, at least mine) BIOSes hang on reboot if
3229          * in vmx root mode.
3230          *
3231          * And Intel TXT required VMX off for all cpu when system shutdown.
3232          */
3233         pr_info("kvm: exiting hardware virtualization\n");
3234         kvm_rebooting = true;
3235         on_each_cpu(hardware_disable_nolock, NULL, 1);
3236         return NOTIFY_OK;
3237 }
3238
3239 static struct notifier_block kvm_reboot_notifier = {
3240         .notifier_call = kvm_reboot,
3241         .priority = 0,
3242 };
3243
3244 static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
3245 {
3246         int i;
3247
3248         for (i = 0; i < bus->dev_count; i++) {
3249                 struct kvm_io_device *pos = bus->range[i].dev;
3250
3251                 kvm_iodevice_destructor(pos);
3252         }
3253         kfree(bus);
3254 }
3255
3256 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1,
3257                                  const struct kvm_io_range *r2)
3258 {
3259         gpa_t addr1 = r1->addr;
3260         gpa_t addr2 = r2->addr;
3261
3262         if (addr1 < addr2)
3263                 return -1;
3264
3265         /* If r2->len == 0, match the exact address.  If r2->len != 0,
3266          * accept any overlapping write.  Any order is acceptable for
3267          * overlapping ranges, because kvm_io_bus_get_first_dev ensures
3268          * we process all of them.
3269          */
3270         if (r2->len) {
3271                 addr1 += r1->len;
3272                 addr2 += r2->len;
3273         }
3274
3275         if (addr1 > addr2)
3276                 return 1;
3277
3278         return 0;
3279 }
3280
3281 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
3282 {
3283         return kvm_io_bus_cmp(p1, p2);
3284 }
3285
3286 static int kvm_io_bus_insert_dev(struct kvm_io_bus *bus, struct kvm_io_device *dev,
3287                           gpa_t addr, int len)
3288 {
3289         bus->range[bus->dev_count++] = (struct kvm_io_range) {
3290                 .addr = addr,
3291                 .len = len,
3292                 .dev = dev,
3293         };
3294
3295         sort(bus->range, bus->dev_count, sizeof(struct kvm_io_range),
3296                 kvm_io_bus_sort_cmp, NULL);
3297
3298         return 0;
3299 }
3300
3301 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
3302                              gpa_t addr, int len)
3303 {
3304         struct kvm_io_range *range, key;
3305         int off;
3306
3307         key = (struct kvm_io_range) {
3308                 .addr = addr,
3309                 .len = len,
3310         };
3311
3312         range = bsearch(&key, bus->range, bus->dev_count,
3313                         sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
3314         if (range == NULL)
3315                 return -ENOENT;
3316
3317         off = range - bus->range;
3318
3319         while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0)
3320                 off--;
3321
3322         return off;
3323 }
3324
3325 static int __kvm_io_bus_write(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3326                               struct kvm_io_range *range, const void *val)
3327 {
3328         int idx;
3329
3330         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3331         if (idx < 0)
3332                 return -EOPNOTSUPP;
3333
3334         while (idx < bus->dev_count &&
3335                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3336                 if (!kvm_iodevice_write(vcpu, bus->range[idx].dev, range->addr,
3337                                         range->len, val))
3338                         return idx;
3339                 idx++;
3340         }
3341
3342         return -EOPNOTSUPP;
3343 }
3344
3345 /* kvm_io_bus_write - called under kvm->slots_lock */
3346 int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3347                      int len, const void *val)
3348 {
3349         struct kvm_io_bus *bus;
3350         struct kvm_io_range range;
3351         int r;
3352
3353         range = (struct kvm_io_range) {
3354                 .addr = addr,
3355                 .len = len,
3356         };
3357
3358         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3359         r = __kvm_io_bus_write(vcpu, bus, &range, val);
3360         return r < 0 ? r : 0;
3361 }
3362
3363 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */
3364 int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
3365                             gpa_t addr, int len, const void *val, long cookie)
3366 {
3367         struct kvm_io_bus *bus;
3368         struct kvm_io_range range;
3369
3370         range = (struct kvm_io_range) {
3371                 .addr = addr,
3372                 .len = len,
3373         };
3374
3375         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3376
3377         /* First try the device referenced by cookie. */
3378         if ((cookie >= 0) && (cookie < bus->dev_count) &&
3379             (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0))
3380                 if (!kvm_iodevice_write(vcpu, bus->range[cookie].dev, addr, len,
3381                                         val))
3382                         return cookie;
3383
3384         /*
3385          * cookie contained garbage; fall back to search and return the
3386          * correct cookie value.
3387          */
3388         return __kvm_io_bus_write(vcpu, bus, &range, val);
3389 }
3390
3391 static int __kvm_io_bus_read(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3392                              struct kvm_io_range *range, void *val)
3393 {
3394         int idx;
3395
3396         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3397         if (idx < 0)
3398                 return -EOPNOTSUPP;
3399
3400         while (idx < bus->dev_count &&
3401                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3402                 if (!kvm_iodevice_read(vcpu, bus->range[idx].dev, range->addr,
3403                                        range->len, val))
3404                         return idx;
3405                 idx++;
3406         }
3407
3408         return -EOPNOTSUPP;
3409 }
3410 EXPORT_SYMBOL_GPL(kvm_io_bus_write);
3411
3412 /* kvm_io_bus_read - called under kvm->slots_lock */
3413 int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3414                     int len, void *val)
3415 {
3416         struct kvm_io_bus *bus;
3417         struct kvm_io_range range;
3418         int r;
3419
3420         range = (struct kvm_io_range) {
3421                 .addr = addr,
3422                 .len = len,
3423         };
3424
3425         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3426         r = __kvm_io_bus_read(vcpu, bus, &range, val);
3427         return r < 0 ? r : 0;
3428 }
3429
3430
3431 /* Caller must hold slots_lock. */
3432 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
3433                             int len, struct kvm_io_device *dev)
3434 {
3435         struct kvm_io_bus *new_bus, *bus;
3436
3437         bus = kvm->buses[bus_idx];
3438         /* exclude ioeventfd which is limited by maximum fd */
3439         if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
3440                 return -ENOSPC;
3441
3442         new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count + 1) *
3443                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3444         if (!new_bus)
3445                 return -ENOMEM;
3446         memcpy(new_bus, bus, sizeof(*bus) + (bus->dev_count *
3447                sizeof(struct kvm_io_range)));
3448         kvm_io_bus_insert_dev(new_bus, dev, addr, len);
3449         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3450         synchronize_srcu_expedited(&kvm->srcu);
3451         kfree(bus);
3452
3453         return 0;
3454 }
3455
3456 /* Caller must hold slots_lock. */
3457 int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
3458                               struct kvm_io_device *dev)
3459 {
3460         int i, r;
3461         struct kvm_io_bus *new_bus, *bus;
3462
3463         bus = kvm->buses[bus_idx];
3464         r = -ENOENT;
3465         for (i = 0; i < bus->dev_count; i++)
3466                 if (bus->range[i].dev == dev) {
3467                         r = 0;
3468                         break;
3469                 }
3470
3471         if (r)
3472                 return r;
3473
3474         new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
3475                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3476         if (!new_bus)
3477                 return -ENOMEM;
3478
3479         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
3480         new_bus->dev_count--;
3481         memcpy(new_bus->range + i, bus->range + i + 1,
3482                (new_bus->dev_count - i) * sizeof(struct kvm_io_range));
3483
3484         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3485         synchronize_srcu_expedited(&kvm->srcu);
3486         kfree(bus);
3487         return r;
3488 }
3489
3490 static int kvm_debugfs_open(struct inode *inode, struct file *file,
3491                            int (*get)(void *, u64 *), int (*set)(void *, u64),
3492                            const char *fmt)
3493 {
3494         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)
3495                                           inode->i_private;
3496
3497         /* The debugfs files are a reference to the kvm struct which
3498          * is still valid when kvm_destroy_vm is called.
3499          * To avoid the race between open and the removal of the debugfs
3500          * directory we test against the users count.
3501          */
3502         if (!atomic_add_unless(&stat_data->kvm->users_count, 1, 0))
3503                 return -ENOENT;
3504
3505         if (simple_attr_open(inode, file, get, set, fmt)) {
3506                 kvm_put_kvm(stat_data->kvm);
3507                 return -ENOMEM;
3508         }
3509
3510         return 0;
3511 }
3512
3513 static int kvm_debugfs_release(struct inode *inode, struct file *file)
3514 {
3515         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)
3516                                           inode->i_private;
3517
3518         simple_attr_release(inode, file);
3519         kvm_put_kvm(stat_data->kvm);
3520
3521         return 0;
3522 }
3523
3524 static int vm_stat_get_per_vm(void *data, u64 *val)
3525 {
3526         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data;
3527
3528         *val = *(u32 *)((void *)stat_data->kvm + stat_data->offset);
3529
3530         return 0;
3531 }
3532
3533 static int vm_stat_get_per_vm_open(struct inode *inode, struct file *file)
3534 {
3535         __simple_attr_check_format("%llu\n", 0ull);
3536         return kvm_debugfs_open(inode, file, vm_stat_get_per_vm,
3537                                 NULL, "%llu\n");
3538 }
3539
3540 static const struct file_operations vm_stat_get_per_vm_fops = {
3541         .owner   = THIS_MODULE,
3542         .open    = vm_stat_get_per_vm_open,
3543         .release = kvm_debugfs_release,
3544         .read    = simple_attr_read,
3545         .write   = simple_attr_write,
3546         .llseek  = generic_file_llseek,
3547 };
3548
3549 static int vcpu_stat_get_per_vm(void *data, u64 *val)
3550 {
3551         int i;
3552         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data;
3553         struct kvm_vcpu *vcpu;
3554
3555         *val = 0;
3556
3557         kvm_for_each_vcpu(i, vcpu, stat_data->kvm)
3558                 *val += *(u32 *)((void *)vcpu + stat_data->offset);
3559
3560         return 0;
3561 }
3562
3563 static int vcpu_stat_get_per_vm_open(struct inode *inode, struct file *file)
3564 {
3565         __simple_attr_check_format("%llu\n", 0ull);
3566         return kvm_debugfs_open(inode, file, vcpu_stat_get_per_vm,
3567                                  NULL, "%llu\n");
3568 }
3569
3570 static const struct file_operations vcpu_stat_get_per_vm_fops = {
3571         .owner   = THIS_MODULE,
3572         .open    = vcpu_stat_get_per_vm_open,
3573         .release = kvm_debugfs_release,
3574         .read    = simple_attr_read,
3575         .write   = simple_attr_write,
3576         .llseek  = generic_file_llseek,
3577 };
3578
3579 static const struct file_operations *stat_fops_per_vm[] = {
3580         [KVM_STAT_VCPU] = &vcpu_stat_get_per_vm_fops,
3581         [KVM_STAT_VM]   = &vm_stat_get_per_vm_fops,
3582 };
3583
3584 static int vm_stat_get(void *_offset, u64 *val)
3585 {
3586         unsigned offset = (long)_offset;
3587         struct kvm *kvm;
3588         struct kvm_stat_data stat_tmp = {.offset = offset};
3589         u64 tmp_val;
3590
3591         *val = 0;
3592         spin_lock(&kvm_lock);
3593         list_for_each_entry(kvm, &vm_list, vm_list) {
3594                 stat_tmp.kvm = kvm;
3595                 vm_stat_get_per_vm((void *)&stat_tmp, &tmp_val);
3596                 *val += tmp_val;
3597         }
3598         spin_unlock(&kvm_lock);
3599         return 0;
3600 }
3601
3602 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, NULL, "%llu\n");
3603
3604 static int vcpu_stat_get(void *_offset, u64 *val)
3605 {
3606         unsigned offset = (long)_offset;
3607         struct kvm *kvm;
3608         struct kvm_stat_data stat_tmp = {.offset = offset};
3609         u64 tmp_val;
3610
3611         *val = 0;
3612         spin_lock(&kvm_lock);
3613         list_for_each_entry(kvm, &vm_list, vm_list) {
3614                 stat_tmp.kvm = kvm;
3615                 vcpu_stat_get_per_vm((void *)&stat_tmp, &tmp_val);
3616                 *val += tmp_val;
3617         }
3618         spin_unlock(&kvm_lock);
3619         return 0;
3620 }
3621
3622 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n");
3623
3624 static const struct file_operations *stat_fops[] = {
3625         [KVM_STAT_VCPU] = &vcpu_stat_fops,
3626         [KVM_STAT_VM]   = &vm_stat_fops,
3627 };
3628
3629 static int kvm_init_debug(void)
3630 {
3631         int r = -EEXIST;
3632         struct kvm_stats_debugfs_item *p;
3633
3634         kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
3635         if (kvm_debugfs_dir == NULL)
3636                 goto out;
3637
3638         kvm_debugfs_num_entries = 0;
3639         for (p = debugfs_entries; p->name; ++p, kvm_debugfs_num_entries++) {
3640                 if (!debugfs_create_file(p->name, 0444, kvm_debugfs_dir,
3641                                          (void *)(long)p->offset,
3642                                          stat_fops[p->kind]))
3643                         goto out_dir;
3644         }
3645
3646         return 0;
3647
3648 out_dir:
3649         debugfs_remove_recursive(kvm_debugfs_dir);
3650 out:
3651         return r;
3652 }
3653
3654 static int kvm_suspend(void)
3655 {
3656         if (kvm_usage_count)
3657                 hardware_disable_nolock(NULL);
3658         return 0;
3659 }
3660
3661 static void kvm_resume(void)
3662 {
3663         if (kvm_usage_count) {
3664                 WARN_ON(raw_spin_is_locked(&kvm_count_lock));
3665                 hardware_enable_nolock(NULL);
3666         }
3667 }
3668
3669 static struct syscore_ops kvm_syscore_ops = {
3670         .suspend = kvm_suspend,
3671         .resume = kvm_resume,
3672 };
3673
3674 static inline
3675 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
3676 {
3677         return container_of(pn, struct kvm_vcpu, preempt_notifier);
3678 }
3679
3680 static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
3681 {
3682         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3683
3684         if (vcpu->preempted)
3685                 vcpu->preempted = false;
3686
3687         kvm_arch_sched_in(vcpu, cpu);
3688
3689         kvm_arch_vcpu_load(vcpu, cpu);
3690 }
3691
3692 static void kvm_sched_out(struct preempt_notifier *pn,
3693                           struct task_struct *next)
3694 {
3695         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3696
3697         if (current->state == TASK_RUNNING)
3698                 vcpu->preempted = true;
3699         kvm_arch_vcpu_put(vcpu);
3700 }
3701
3702 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
3703                   struct module *module)
3704 {
3705         int r;
3706         int cpu;
3707
3708         r = kvm_arch_init(opaque);
3709         if (r)
3710                 goto out_fail;
3711
3712         /*
3713          * kvm_arch_init makes sure there's at most one caller
3714          * for architectures that support multiple implementations,
3715          * like intel and amd on x86.
3716          * kvm_arch_init must be called before kvm_irqfd_init to avoid creating
3717          * conflicts in case kvm is already setup for another implementation.
3718          */
3719         r = kvm_irqfd_init();
3720         if (r)
3721                 goto out_irqfd;
3722
3723         if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
3724                 r = -ENOMEM;
3725                 goto out_free_0;
3726         }
3727
3728         r = kvm_arch_hardware_setup();
3729         if (r < 0)
3730                 goto out_free_0a;
3731
3732         for_each_online_cpu(cpu) {
3733                 smp_call_function_single(cpu,
3734                                 kvm_arch_check_processor_compat,
3735                                 &r, 1);
3736                 if (r < 0)
3737                         goto out_free_1;
3738         }
3739
3740         r = cpuhp_setup_state_nocalls(CPUHP_AP_KVM_STARTING, "AP_KVM_STARTING",
3741                                       kvm_starting_cpu, kvm_dying_cpu);
3742         if (r)
3743                 goto out_free_2;
3744         register_reboot_notifier(&kvm_reboot_notifier);
3745
3746         /* A kmem cache lets us meet the alignment requirements of fx_save. */
3747         if (!vcpu_align)
3748                 vcpu_align = __alignof__(struct kvm_vcpu);
3749         kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size, vcpu_align,
3750                                            0, NULL);
3751         if (!kvm_vcpu_cache) {
3752                 r = -ENOMEM;
3753                 goto out_free_3;
3754         }
3755
3756         r = kvm_async_pf_init();
3757         if (r)
3758                 goto out_free;
3759
3760         kvm_chardev_ops.owner = module;
3761         kvm_vm_fops.owner = module;
3762         kvm_vcpu_fops.owner = module;
3763
3764         r = misc_register(&kvm_dev);
3765         if (r) {
3766                 pr_err("kvm: misc device register failed\n");
3767                 goto out_unreg;
3768         }
3769
3770         register_syscore_ops(&kvm_syscore_ops);
3771
3772         kvm_preempt_ops.sched_in = kvm_sched_in;
3773         kvm_preempt_ops.sched_out = kvm_sched_out;
3774
3775         r = kvm_init_debug();
3776         if (r) {
3777                 pr_err("kvm: create debugfs files failed\n");
3778                 goto out_undebugfs;
3779         }
3780
3781         r = kvm_vfio_ops_init();
3782         WARN_ON(r);
3783
3784         return 0;
3785
3786 out_undebugfs:
3787         unregister_syscore_ops(&kvm_syscore_ops);
3788         misc_deregister(&kvm_dev);
3789 out_unreg:
3790         kvm_async_pf_deinit();
3791 out_free:
3792         kmem_cache_destroy(kvm_vcpu_cache);
3793 out_free_3:
3794         unregister_reboot_notifier(&kvm_reboot_notifier);
3795         cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING);
3796 out_free_2:
3797 out_free_1:
3798         kvm_arch_hardware_unsetup();
3799 out_free_0a:
3800         free_cpumask_var(cpus_hardware_enabled);
3801 out_free_0:
3802         kvm_irqfd_exit();
3803 out_irqfd:
3804         kvm_arch_exit();
3805 out_fail:
3806         return r;
3807 }
3808 EXPORT_SYMBOL_GPL(kvm_init);
3809
3810 void kvm_exit(void)
3811 {
3812         debugfs_remove_recursive(kvm_debugfs_dir);
3813         misc_deregister(&kvm_dev);
3814         kmem_cache_destroy(kvm_vcpu_cache);
3815         kvm_async_pf_deinit();
3816         unregister_syscore_ops(&kvm_syscore_ops);
3817         unregister_reboot_notifier(&kvm_reboot_notifier);
3818         cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING);
3819         on_each_cpu(hardware_disable_nolock, NULL, 1);
3820         kvm_arch_hardware_unsetup();
3821         kvm_arch_exit();
3822         kvm_irqfd_exit();
3823         free_cpumask_var(cpus_hardware_enabled);
3824         kvm_vfio_ops_exit();
3825 }
3826 EXPORT_SYMBOL_GPL(kvm_exit);