virt/kvm/kvm_main.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * Copyright (C) 2006 Qumranet, Inc.
   8  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   9  *
  10  * Authors:
  11  *   Avi Kivity   <avi@qumranet.com>
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *
  14  * This work is licensed under the terms of the GNU GPL, version 2.  See
  15  * the COPYING file in the top-level directory.
  16  *
  17  */
  18
  19 #include "iodev.h"
  20
  21 #include <linux/kvm_host.h>
  22 #include <linux/kvm.h>
  23 #include <linux/module.h>
  24 #include <linux/errno.h>
  25 #include <linux/percpu.h>
  26 #include <linux/mm.h>
  27 #include <linux/miscdevice.h>
  28 #include <linux/vmalloc.h>
  29 #include <linux/reboot.h>
  30 #include <linux/debugfs.h>
  31 #include <linux/highmem.h>
  32 #include <linux/file.h>
  33 #include <linux/syscore_ops.h>
  34 #include <linux/cpu.h>
  35 #include <linux/sched.h>
  36 #include <linux/cpumask.h>
  37 #include <linux/smp.h>
  38 #include <linux/anon_inodes.h>
  39 #include <linux/profile.h>
  40 #include <linux/kvm_para.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/mman.h>
  43 #include <linux/swap.h>
  44 #include <linux/bitops.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/compat.h>
  47 #include <linux/srcu.h>
  48 #include <linux/hugetlb.h>
  49 #include <linux/slab.h>
  50 #include <linux/sort.h>
  51 #include <linux/bsearch.h>
  52
  53 #include <asm/processor.h>
  54 #include <asm/io.h>
  55 #include <asm/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 MODULE_AUTHOR("Qumranet");
  67 MODULE_LICENSE("GPL");
  68
  69 /*
  70  * Ordering of locks:
  71  *
  72  *              kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  73  */
  74
  75 DEFINE_SPINLOCK(kvm_lock);
  76 static DEFINE_RAW_SPINLOCK(kvm_count_lock);
  77 LIST_HEAD(vm_list);
  78
  79 static cpumask_var_t cpus_hardware_enabled;
  80 static int kvm_usage_count = 0;
  81 static atomic_t hardware_enable_failed;
  82
  83 struct kmem_cache *kvm_vcpu_cache;
  84 EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
  85
  86 static __read_mostly struct preempt_ops kvm_preempt_ops;
  87
  88 struct dentry *kvm_debugfs_dir;
  89
  90 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
  91                            unsigned long arg);
  92 #ifdef CONFIG_COMPAT
  93 static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl,
  94                                   unsigned long arg);
  95 #endif
  96 static int hardware_enable_all(void);
  97 static void hardware_disable_all(void);
  98
  99 static void kvm_io_bus_destroy(struct kvm_io_bus *bus);
 100
 101 static void kvm_release_pfn_dirty(pfn_t pfn);
 102 static void mark_page_dirty_in_slot(struct kvm *kvm,
 103                                     struct kvm_memory_slot *memslot, gfn_t gfn);
 104
 105 __visible bool kvm_rebooting;
 106 EXPORT_SYMBOL_GPL(kvm_rebooting);
 107
 108 static bool largepages_enabled = true;
 109
 110 bool kvm_is_mmio_pfn(pfn_t pfn)
 111 {
 112         if (pfn_valid(pfn))
 113                 return PageReserved(pfn_to_page(pfn));
 114
 115         return true;
 116 }
 117
 118 /*
 119  * Switches to specified vcpu, until a matching vcpu_put()
 120  */
 121 int vcpu_load(struct kvm_vcpu *vcpu)
 122 {
 123         int cpu;
 124
 125         if (mutex_lock_killable(&vcpu->mutex))
 126                 return -EINTR;
 127         if (unlikely(vcpu->pid != current->pids[PIDTYPE_PID].pid)) {
 128                 /* The thread running this VCPU changed. */
 129                 struct pid *oldpid = vcpu->pid;
 130                 struct pid *newpid = get_task_pid(current, PIDTYPE_PID);
 131                 rcu_assign_pointer(vcpu->pid, newpid);
 132                 if (oldpid)
 133                         synchronize_rcu();
 134                 put_pid(oldpid);
 135         }
 136         cpu = get_cpu();
 137         preempt_notifier_register(&vcpu->preempt_notifier);
 138         kvm_arch_vcpu_load(vcpu, cpu);
 139         put_cpu();
 140         return 0;
 141 }
 142
 143 void vcpu_put(struct kvm_vcpu *vcpu)
 144 {
 145         preempt_disable();
 146         kvm_arch_vcpu_put(vcpu);
 147         preempt_notifier_unregister(&vcpu->preempt_notifier);
 148         preempt_enable();
 149         mutex_unlock(&vcpu->mutex);
 150 }
 151
 152 static void ack_flush(void *_completed)
 153 {
 154 }
 155
 156 static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
 157 {
 158         int i, cpu, me;
 159         cpumask_var_t cpus;
 160         bool called = true;
 161         struct kvm_vcpu *vcpu;
 162
 163         zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 164
 165         me = get_cpu();
 166         kvm_for_each_vcpu(i, vcpu, kvm) {
 167                 kvm_make_request(req, vcpu);
 168                 cpu = vcpu->cpu;
 169
 170                 /* Set ->requests bit before we read ->mode */
 171                 smp_mb();
 172
 173                 if (cpus != NULL && cpu != -1 && cpu != me &&
 174                       kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE)
 175                         cpumask_set_cpu(cpu, cpus);
 176         }
 177         if (unlikely(cpus == NULL))
 178                 smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1);
 179         else if (!cpumask_empty(cpus))
 180                 smp_call_function_many(cpus, ack_flush, NULL, 1);
 181         else
 182                 called = false;
 183         put_cpu();
 184         free_cpumask_var(cpus);
 185         return called;
 186 }
 187
 188 void kvm_flush_remote_tlbs(struct kvm *kvm)
 189 {
 190         long dirty_count = kvm->tlbs_dirty;
 191
 192         smp_mb();
 193         if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 194                 ++kvm->stat.remote_tlb_flush;
 195         cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 196 }
 197 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs);
 198
 199 void kvm_reload_remote_mmus(struct kvm *kvm)
 200 {
 201         make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
 202 }
 203
 204 void kvm_make_mclock_inprogress_request(struct kvm *kvm)
 205 {
 206         make_all_cpus_request(kvm, KVM_REQ_MCLOCK_INPROGRESS);
 207 }
 208
 209 void kvm_make_scan_ioapic_request(struct kvm *kvm)
 210 {
 211         make_all_cpus_request(kvm, KVM_REQ_SCAN_IOAPIC);
 212 }
 213
 214 int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
 215 {
 216         struct page *page;
 217         int r;
 218
 219         mutex_init(&vcpu->mutex);
 220         vcpu->cpu = -1;
 221         vcpu->kvm = kvm;
 222         vcpu->vcpu_id = id;
 223         vcpu->pid = NULL;
 224         init_waitqueue_head(&vcpu->wq);
 225         kvm_async_pf_vcpu_init(vcpu);
 226
 227         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 228         if (!page) {
 229                 r = -ENOMEM;
 230                 goto fail;
 231         }
 232         vcpu->run = page_address(page);
 233
 234         kvm_vcpu_set_in_spin_loop(vcpu, false);
 235         kvm_vcpu_set_dy_eligible(vcpu, false);
 236         vcpu->preempted = false;
 237
 238         r = kvm_arch_vcpu_init(vcpu);
 239         if (r < 0)
 240                 goto fail_free_run;
 241         return 0;
 242
 243 fail_free_run:
 244         free_page((unsigned long)vcpu->run);
 245 fail:
 246         return r;
 247 }
 248 EXPORT_SYMBOL_GPL(kvm_vcpu_init);
 249
 250 void kvm_vcpu_uninit(struct kvm_vcpu *vcpu)
 251 {
 252         put_pid(vcpu->pid);
 253         kvm_arch_vcpu_uninit(vcpu);
 254         free_page((unsigned long)vcpu->run);
 255 }
 256 EXPORT_SYMBOL_GPL(kvm_vcpu_uninit);
 257
 258 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 259 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
 260 {
 261         return container_of(mn, struct kvm, mmu_notifier);
 262 }
 263
 264 static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 265                                              struct mm_struct *mm,
 266                                              unsigned long address)
 267 {
 268         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 269         int need_tlb_flush, idx;
 270
 271         /*
 272          * When ->invalidate_page runs, the linux pte has been zapped
 273          * already but the page is still allocated until
 274          * ->invalidate_page returns. So if we increase the sequence
 275          * here the kvm page fault will notice if the spte can't be
 276          * established because the page is going to be freed. If
 277          * instead the kvm page fault establishes the spte before
 278          * ->invalidate_page runs, kvm_unmap_hva will release it
 279          * before returning.
 280          *
 281          * The sequence increase only need to be seen at spin_unlock
 282          * time, and not at spin_lock time.
 283          *
 284          * Increasing the sequence after the spin_unlock would be
 285          * unsafe because the kvm page fault could then establish the
 286          * pte after kvm_unmap_hva returned, without noticing the page
 287          * is going to be freed.
 288          */
 289         idx = srcu_read_lock(&kvm->srcu);
 290         spin_lock(&kvm->mmu_lock);
 291
 292         kvm->mmu_notifier_seq++;
 293         need_tlb_flush = kvm_unmap_hva(kvm, address) | kvm->tlbs_dirty;
 294         /* we've to flush the tlb before the pages can be freed */
 295         if (need_tlb_flush)
 296                 kvm_flush_remote_tlbs(kvm);
 297
 298         spin_unlock(&kvm->mmu_lock);
 299         srcu_read_unlock(&kvm->srcu, idx);
 300 }
 301
 302 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
 303                                         struct mm_struct *mm,
 304                                         unsigned long address,
 305                                         pte_t pte)
 306 {
 307         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 308         int idx;
 309
 310         idx = srcu_read_lock(&kvm->srcu);
 311         spin_lock(&kvm->mmu_lock);
 312         kvm->mmu_notifier_seq++;
 313         kvm_set_spte_hva(kvm, address, pte);
 314         spin_unlock(&kvm->mmu_lock);
 315         srcu_read_unlock(&kvm->srcu, idx);
 316 }
 317
 318 static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 319                                                     struct mm_struct *mm,
 320                                                     unsigned long start,
 321                                                     unsigned long end)
 322 {
 323         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 324         int need_tlb_flush = 0, idx;
 325
 326         idx = srcu_read_lock(&kvm->srcu);
 327         spin_lock(&kvm->mmu_lock);
 328         /*
 329          * The count increase must become visible at unlock time as no
 330          * spte can be established without taking the mmu_lock and
 331          * count is also read inside the mmu_lock critical section.
 332          */
 333         kvm->mmu_notifier_count++;
 334         need_tlb_flush = kvm_unmap_hva_range(kvm, start, end);
 335         need_tlb_flush |= kvm->tlbs_dirty;
 336         /* we've to flush the tlb before the pages can be freed */
 337         if (need_tlb_flush)
 338                 kvm_flush_remote_tlbs(kvm);
 339
 340         spin_unlock(&kvm->mmu_lock);
 341         srcu_read_unlock(&kvm->srcu, idx);
 342 }
 343
 344 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
 345                                                   struct mm_struct *mm,
 346                                                   unsigned long start,
 347                                                   unsigned long end)
 348 {
 349         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 350
 351         spin_lock(&kvm->mmu_lock);
 352         /*
 353          * This sequence increase will notify the kvm page fault that
 354          * the page that is going to be mapped in the spte could have
 355          * been freed.
 356          */
 357         kvm->mmu_notifier_seq++;
 358         smp_wmb();
 359         /*
 360          * The above sequence increase must be visible before the
 361          * below count decrease, which is ensured by the smp_wmb above
 362          * in conjunction with the smp_rmb in mmu_notifier_retry().
 363          */
 364         kvm->mmu_notifier_count--;
 365         spin_unlock(&kvm->mmu_lock);
 366
 367         BUG_ON(kvm->mmu_notifier_count < 0);
 368 }
 369
 370 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
 371                                               struct mm_struct *mm,
 372                                               unsigned long address)
 373 {
 374         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 375         int young, idx;
 376
 377         idx = srcu_read_lock(&kvm->srcu);
 378         spin_lock(&kvm->mmu_lock);
 379
 380         young = kvm_age_hva(kvm, address);
 381         if (young)
 382                 kvm_flush_remote_tlbs(kvm);
 383
 384         spin_unlock(&kvm->mmu_lock);
 385         srcu_read_unlock(&kvm->srcu, idx);
 386
 387         return young;
 388 }
 389
 390 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn,
 391                                        struct mm_struct *mm,
 392                                        unsigned long address)
 393 {
 394         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 395         int young, idx;
 396
 397         idx = srcu_read_lock(&kvm->srcu);
 398         spin_lock(&kvm->mmu_lock);
 399         young = kvm_test_age_hva(kvm, address);
 400         spin_unlock(&kvm->mmu_lock);
 401         srcu_read_unlock(&kvm->srcu, idx);
 402
 403         return young;
 404 }
 405
 406 static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
 407                                      struct mm_struct *mm)
 408 {
 409         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 410         int idx;
 411
 412         idx = srcu_read_lock(&kvm->srcu);
 413         kvm_arch_flush_shadow_all(kvm);
 414         srcu_read_unlock(&kvm->srcu, idx);
 415 }
 416
 417 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
 418         .invalidate_page        = kvm_mmu_notifier_invalidate_page,
 419         .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
 420         .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
 421         .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
 422         .test_young             = kvm_mmu_notifier_test_young,
 423         .change_pte             = kvm_mmu_notifier_change_pte,
 424         .release                = kvm_mmu_notifier_release,
 425 };
 426
 427 static int kvm_init_mmu_notifier(struct kvm *kvm)
 428 {
 429         kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
 430         return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
 431 }
 432
 433 #else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
 434
 435 static int kvm_init_mmu_notifier(struct kvm *kvm)
 436 {
 437         return 0;
 438 }
 439
 440 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 441
 442 static void kvm_init_memslots_id(struct kvm *kvm)
 443 {
 444         int i;
 445         struct kvm_memslots *slots = kvm->memslots;
 446
 447         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 448                 slots->id_to_index[i] = slots->memslots[i].id = i;
 449 }
 450
 451 static struct kvm *kvm_create_vm(unsigned long type)
 452 {
 453         int r, i;
 454         struct kvm *kvm = kvm_arch_alloc_vm();
 455
 456         if (!kvm)
 457                 return ERR_PTR(-ENOMEM);
 458
 459         r = kvm_arch_init_vm(kvm, type);
 460         if (r)
 461                 goto out_err_no_disable;
 462
 463         r = hardware_enable_all();
 464         if (r)
 465                 goto out_err_no_disable;
 466
 467 #ifdef CONFIG_HAVE_KVM_IRQCHIP
 468         INIT_HLIST_HEAD(&kvm->mask_notifier_list);
 469 #endif
 470 #ifdef CONFIG_HAVE_KVM_IRQFD
 471         INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 472 #endif
 473
 474         BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX);
 475
 476         r = -ENOMEM;
 477         kvm->memslots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
 478         if (!kvm->memslots)
 479                 goto out_err_no_srcu;
 480
 481         /*
 482          * Init kvm generation close to the maximum to easily test the
 483          * code of handling generation number wrap-around.
 484          */
 485         kvm->memslots->generation = -150;
 486
 487         kvm_init_memslots_id(kvm);
 488         if (init_srcu_struct(&kvm->srcu))
 489                 goto out_err_no_srcu;
 490         if (init_srcu_struct(&kvm->irq_srcu))
 491                 goto out_err_no_irq_srcu;
 492         for (i = 0; i < KVM_NR_BUSES; i++) {
 493                 kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
 494                                         GFP_KERNEL);
 495                 if (!kvm->buses[i])
 496                         goto out_err;
 497         }
 498
 499         spin_lock_init(&kvm->mmu_lock);
 500         kvm->mm = current->mm;
 501         atomic_inc(&kvm->mm->mm_count);
 502         kvm_eventfd_init(kvm);
 503         mutex_init(&kvm->lock);
 504         mutex_init(&kvm->irq_lock);
 505         mutex_init(&kvm->slots_lock);
 506         atomic_set(&kvm->users_count, 1);
 507         INIT_LIST_HEAD(&kvm->devices);
 508
 509         r = kvm_init_mmu_notifier(kvm);
 510         if (r)
 511                 goto out_err;
 512
 513         spin_lock(&kvm_lock);
 514         list_add(&kvm->vm_list, &vm_list);
 515         spin_unlock(&kvm_lock);
 516
 517         return kvm;
 518
 519 out_err:
 520         cleanup_srcu_struct(&kvm->irq_srcu);
 521 out_err_no_irq_srcu:
 522         cleanup_srcu_struct(&kvm->srcu);
 523 out_err_no_srcu:
 524         hardware_disable_all();
 525 out_err_no_disable:
 526         for (i = 0; i < KVM_NR_BUSES; i++)
 527                 kfree(kvm->buses[i]);
 528         kfree(kvm->memslots);
 529         kvm_arch_free_vm(kvm);
 530         return ERR_PTR(r);
 531 }
 532
 533 /*
 534  * Avoid using vmalloc for a small buffer.
 535  * Should not be used when the size is statically known.
 536  */
 537 void *kvm_kvzalloc(unsigned long size)
 538 {
 539         if (size > PAGE_SIZE)
 540                 return vzalloc(size);
 541         else
 542                 return kzalloc(size, GFP_KERNEL);
 543 }
 544
 545 void kvm_kvfree(const void *addr)
 546 {
 547         if (is_vmalloc_addr(addr))
 548                 vfree(addr);
 549         else
 550                 kfree(addr);
 551 }
 552
 553 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
 554 {
 555         if (!memslot->dirty_bitmap)
 556                 return;
 557
 558         kvm_kvfree(memslot->dirty_bitmap);
 559         memslot->dirty_bitmap = NULL;
 560 }
 561
 562 /*
 563  * Free any memory in @free but not in @dont.
 564  */
 565 static void kvm_free_physmem_slot(struct kvm *kvm, struct kvm_memory_slot *free,
 566                                   struct kvm_memory_slot *dont)
 567 {
 568         if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
 569                 kvm_destroy_dirty_bitmap(free);
 570
 571         kvm_arch_free_memslot(kvm, free, dont);
 572
 573         free->npages = 0;
 574 }
 575
 576 static void kvm_free_physmem(struct kvm *kvm)
 577 {
 578         struct kvm_memslots *slots = kvm->memslots;
 579         struct kvm_memory_slot *memslot;
 580
 581         kvm_for_each_memslot(memslot, slots)
 582                 kvm_free_physmem_slot(kvm, memslot, NULL);
 583
 584         kfree(kvm->memslots);
 585 }
 586
 587 static void kvm_destroy_devices(struct kvm *kvm)
 588 {
 589         struct list_head *node, *tmp;
 590
 591         list_for_each_safe(node, tmp, &kvm->devices) {
 592                 struct kvm_device *dev =
 593                         list_entry(node, struct kvm_device, vm_node);
 594
 595                 list_del(node);
 596                 dev->ops->destroy(dev);
 597         }
 598 }
 599
 600 static void kvm_destroy_vm(struct kvm *kvm)
 601 {
 602         int i;
 603         struct mm_struct *mm = kvm->mm;
 604
 605         kvm_arch_sync_events(kvm);
 606         spin_lock(&kvm_lock);
 607         list_del(&kvm->vm_list);
 608         spin_unlock(&kvm_lock);
 609         kvm_free_irq_routing(kvm);
 610         for (i = 0; i < KVM_NR_BUSES; i++)
 611                 kvm_io_bus_destroy(kvm->buses[i]);
 612         kvm_coalesced_mmio_free(kvm);
 613 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 614         mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 615 #else
 616         kvm_arch_flush_shadow_all(kvm);
 617 #endif
 618         kvm_arch_destroy_vm(kvm);
 619         kvm_destroy_devices(kvm);
 620         kvm_free_physmem(kvm);
 621         cleanup_srcu_struct(&kvm->irq_srcu);
 622         cleanup_srcu_struct(&kvm->srcu);
 623         kvm_arch_free_vm(kvm);
 624         hardware_disable_all();
 625         mmdrop(mm);
 626 }
 627
 628 void kvm_get_kvm(struct kvm *kvm)
 629 {
 630         atomic_inc(&kvm->users_count);
 631 }
 632 EXPORT_SYMBOL_GPL(kvm_get_kvm);
 633
 634 void kvm_put_kvm(struct kvm *kvm)
 635 {
 636         if (atomic_dec_and_test(&kvm->users_count))
 637                 kvm_destroy_vm(kvm);
 638 }
 639 EXPORT_SYMBOL_GPL(kvm_put_kvm);
 640
 641
 642 static int kvm_vm_release(struct inode *inode, struct file *filp)
 643 {
 644         struct kvm *kvm = filp->private_data;
 645
 646         kvm_irqfd_release(kvm);
 647
 648         kvm_put_kvm(kvm);
 649         return 0;
 650 }
 651
 652 /*
 653  * Allocation size is twice as large as the actual dirty bitmap size.
 654  * See x86's kvm_vm_ioctl_get_dirty_log() why this is needed.
 655  */
 656 static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
 657 {
 658         unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
 659
 660         memslot->dirty_bitmap = kvm_kvzalloc(dirty_bytes);
 661         if (!memslot->dirty_bitmap)
 662                 return -ENOMEM;
 663
 664         return 0;
 665 }
 666
 667 static int cmp_memslot(const void *slot1, const void *slot2)
 668 {
 669         struct kvm_memory_slot *s1, *s2;
 670
 671         s1 = (struct kvm_memory_slot *)slot1;
 672         s2 = (struct kvm_memory_slot *)slot2;
 673
 674         if (s1->npages < s2->npages)
 675                 return 1;
 676         if (s1->npages > s2->npages)
 677                 return -1;
 678
 679         return 0;
 680 }
 681
 682 /*
 683  * Sort the memslots base on its size, so the larger slots
 684  * will get better fit.
 685  */
 686 static void sort_memslots(struct kvm_memslots *slots)
 687 {
 688         int i;
 689
 690         sort(slots->memslots, KVM_MEM_SLOTS_NUM,
 691               sizeof(struct kvm_memory_slot), cmp_memslot, NULL);
 692
 693         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 694                 slots->id_to_index[slots->memslots[i].id] = i;
 695 }
 696
 697 static void update_memslots(struct kvm_memslots *slots,
 698                             struct kvm_memory_slot *new)
 699 {
 700         if (new) {
 701                 int id = new->id;
 702                 struct kvm_memory_slot *old = id_to_memslot(slots, id);
 703                 unsigned long npages = old->npages;
 704
 705                 *old = *new;
 706                 if (new->npages != npages)
 707                         sort_memslots(slots);
 708         }
 709 }
 710
 711 static int check_memory_region_flags(struct kvm_userspace_memory_region *mem)
 712 {
 713         u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES;
 714
 715 #ifdef __KVM_HAVE_READONLY_MEM
 716         valid_flags |= KVM_MEM_READONLY;
 717 #endif
 718
 719         if (mem->flags & ~valid_flags)
 720                 return -EINVAL;
 721
 722         return 0;
 723 }
 724
 725 static struct kvm_memslots *install_new_memslots(struct kvm *kvm,
 726                 struct kvm_memslots *slots, struct kvm_memory_slot *new)
 727 {
 728         struct kvm_memslots *old_memslots = kvm->memslots;
 729
 730         /*
 731          * Set the low bit in the generation, which disables SPTE caching
 732          * until the end of synchronize_srcu_expedited.
 733          */
 734         WARN_ON(old_memslots->generation & 1);
 735         slots->generation = old_memslots->generation + 1;
 736
 737         update_memslots(slots, new);
 738         rcu_assign_pointer(kvm->memslots, slots);
 739         synchronize_srcu_expedited(&kvm->srcu);
 740
 741         /*
 742          * Increment the new memslot generation a second time. This prevents
 743          * vm exits that race with memslot updates from caching a memslot
 744          * generation that will (potentially) be valid forever.
 745          */
 746         slots->generation++;
 747
 748         kvm_arch_memslots_updated(kvm);
 749
 750         return old_memslots;
 751 }
 752
 753 /*
 754  * Allocate some memory and give it an address in the guest physical address
 755  * space.
 756  *
 757  * Discontiguous memory is allowed, mostly for framebuffers.
 758  *
 759  * Must be called holding mmap_sem for write.
 760  */
 761 int __kvm_set_memory_region(struct kvm *kvm,
 762                             struct kvm_userspace_memory_region *mem)
 763 {
 764         int r;
 765         gfn_t base_gfn;
 766         unsigned long npages;
 767         struct kvm_memory_slot *slot;
 768         struct kvm_memory_slot old, new;
 769         struct kvm_memslots *slots = NULL, *old_memslots;
 770         enum kvm_mr_change change;
 771
 772         r = check_memory_region_flags(mem);
 773         if (r)
 774                 goto out;
 775
 776         r = -EINVAL;
 777         /* General sanity checks */
 778         if (mem->memory_size & (PAGE_SIZE - 1))
 779                 goto out;
 780         if (mem->guest_phys_addr & (PAGE_SIZE - 1))
 781                 goto out;
 782         /* We can read the guest memory with __xxx_user() later on. */
 783         if ((mem->slot < KVM_USER_MEM_SLOTS) &&
 784             ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
 785              !access_ok(VERIFY_WRITE,
 786                         (void __user *)(unsigned long)mem->userspace_addr,
 787                         mem->memory_size)))
 788                 goto out;
 789         if (mem->slot >= KVM_MEM_SLOTS_NUM)
 790                 goto out;
 791         if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
 792                 goto out;
 793
 794         slot = id_to_memslot(kvm->memslots, mem->slot);
 795         base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
 796         npages = mem->memory_size >> PAGE_SHIFT;
 797
 798         if (npages > KVM_MEM_MAX_NR_PAGES)
 799                 goto out;
 800
 801         if (!npages)
 802                 mem->flags &= ~KVM_MEM_LOG_DIRTY_PAGES;
 803
 804         new = old = *slot;
 805
 806         new.id = mem->slot;
 807         new.base_gfn = base_gfn;
 808         new.npages = npages;
 809         new.flags = mem->flags;
 810
 811         if (npages) {
 812                 if (!old.npages)
 813                         change = KVM_MR_CREATE;
 814                 else { /* Modify an existing slot. */
 815                         if ((mem->userspace_addr != old.userspace_addr) ||
 816                             (npages != old.npages) ||
 817                             ((new.flags ^ old.flags) & KVM_MEM_READONLY))
 818                                 goto out;
 819
 820                         if (base_gfn != old.base_gfn)
 821                                 change = KVM_MR_MOVE;
 822                         else if (new.flags != old.flags)
 823                                 change = KVM_MR_FLAGS_ONLY;
 824                         else { /* Nothing to change. */
 825                                 r = 0;
 826                                 goto out;
 827                         }
 828                 }
 829         } else if (old.npages) {
 830                 change = KVM_MR_DELETE;
 831         } else /* Modify a non-existent slot: disallowed. */
 832                 goto out;
 833
 834         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 835                 /* Check for overlaps */
 836                 r = -EEXIST;
 837                 kvm_for_each_memslot(slot, kvm->memslots) {
 838                         if ((slot->id >= KVM_USER_MEM_SLOTS) ||
 839                             (slot->id == mem->slot))
 840                                 continue;
 841                         if (!((base_gfn + npages <= slot->base_gfn) ||
 842                               (base_gfn >= slot->base_gfn + slot->npages)))
 843                                 goto out;
 844                 }
 845         }
 846
 847         /* Free page dirty bitmap if unneeded */
 848         if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
 849                 new.dirty_bitmap = NULL;
 850
 851         r = -ENOMEM;
 852         if (change == KVM_MR_CREATE) {
 853                 new.userspace_addr = mem->userspace_addr;
 854
 855                 if (kvm_arch_create_memslot(kvm, &new, npages))
 856                         goto out_free;
 857         }
 858
 859         /* Allocate page dirty bitmap if needed */
 860         if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
 861                 if (kvm_create_dirty_bitmap(&new) < 0)
 862                         goto out_free;
 863         }
 864
 865         if ((change == KVM_MR_DELETE) || (change == KVM_MR_MOVE)) {
 866                 slots = kmemdup(kvm->memslots, sizeof(struct kvm_memslots),
 867                                 GFP_KERNEL);
 868                 if (!slots)
 869                         goto out_free;
 870                 slot = id_to_memslot(slots, mem->slot);
 871                 slot->flags |= KVM_MEMSLOT_INVALID;
 872
 873                 old_memslots = install_new_memslots(kvm, slots, NULL);
 874
 875                 /* slot was deleted or moved, clear iommu mapping */
 876                 kvm_iommu_unmap_pages(kvm, &old);
 877                 /* From this point no new shadow pages pointing to a deleted,
 878                  * or moved, memslot will be created.
 879                  *
 880                  * validation of sp->gfn happens in:
 881                  *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
 882                  *      - kvm_is_visible_gfn (mmu_check_roots)
 883                  */
 884                 kvm_arch_flush_shadow_memslot(kvm, slot);
 885                 slots = old_memslots;
 886         }
 887
 888         r = kvm_arch_prepare_memory_region(kvm, &new, mem, change);
 889         if (r)
 890                 goto out_slots;
 891
 892         r = -ENOMEM;
 893         /*
 894          * We can re-use the old_memslots from above, the only difference
 895          * from the currently installed memslots is the invalid flag.  This
 896          * will get overwritten by update_memslots anyway.
 897          */
 898         if (!slots) {
 899                 slots = kmemdup(kvm->memslots, sizeof(struct kvm_memslots),
 900                                 GFP_KERNEL);
 901                 if (!slots)
 902                         goto out_free;
 903         }
 904
 905         /* actual memory is freed via old in kvm_free_physmem_slot below */
 906         if (change == KVM_MR_DELETE) {
 907                 new.dirty_bitmap = NULL;
 908                 memset(&new.arch, 0, sizeof(new.arch));
 909         }
 910
 911         old_memslots = install_new_memslots(kvm, slots, &new);
 912
 913         kvm_arch_commit_memory_region(kvm, mem, &old, change);
 914
 915         kvm_free_physmem_slot(kvm, &old, &new);
 916         kfree(old_memslots);
 917
 918         /*
 919          * IOMMU mapping:  New slots need to be mapped.  Old slots need to be
 920          * un-mapped and re-mapped if their base changes.  Since base change
 921          * unmapping is handled above with slot deletion, mapping alone is
 922          * needed here.  Anything else the iommu might care about for existing
 923          * slots (size changes, userspace addr changes and read-only flag
 924          * changes) is disallowed above, so any other attribute changes getting
 925          * here can be skipped.
 926          */
 927         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 928                 r = kvm_iommu_map_pages(kvm, &new);
 929                 return r;
 930         }
 931
 932         return 0;
 933
 934 out_slots:
 935         kfree(slots);
 936 out_free:
 937         kvm_free_physmem_slot(kvm, &new, &old);
 938 out:
 939         return r;
 940 }
 941 EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
 942
 943 int kvm_set_memory_region(struct kvm *kvm,
 944                           struct kvm_userspace_memory_region *mem)
 945 {
 946         int r;
 947
 948         mutex_lock(&kvm->slots_lock);
 949         r = __kvm_set_memory_region(kvm, mem);
 950         mutex_unlock(&kvm->slots_lock);
 951         return r;
 952 }
 953 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
 954
 955 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
 956                                           struct kvm_userspace_memory_region *mem)
 957 {
 958         if (mem->slot >= KVM_USER_MEM_SLOTS)
 959                 return -EINVAL;
 960         return kvm_set_memory_region(kvm, mem);
 961 }
 962
 963 int kvm_get_dirty_log(struct kvm *kvm,
 964                         struct kvm_dirty_log *log, int *is_dirty)
 965 {
 966         struct kvm_memory_slot *memslot;
 967         int r, i;
 968         unsigned long n;
 969         unsigned long any = 0;
 970
 971         r = -EINVAL;
 972         if (log->slot >= KVM_USER_MEM_SLOTS)
 973                 goto out;
 974
 975         memslot = id_to_memslot(kvm->memslots, log->slot);
 976         r = -ENOENT;
 977         if (!memslot->dirty_bitmap)
 978                 goto out;
 979
 980         n = kvm_dirty_bitmap_bytes(memslot);
 981
 982         for (i = 0; !any && i < n/sizeof(long); ++i)
 983                 any = memslot->dirty_bitmap[i];
 984
 985         r = -EFAULT;
 986         if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
 987                 goto out;
 988
 989         if (any)
 990                 *is_dirty = 1;
 991
 992         r = 0;
 993 out:
 994         return r;
 995 }
 996 EXPORT_SYMBOL_GPL(kvm_get_dirty_log);
 997
 998 bool kvm_largepages_enabled(void)
 999 {
1000         return largepages_enabled;
1001 }
1002
1003 void kvm_disable_largepages(void)
1004 {
1005         largepages_enabled = false;
1006 }
1007 EXPORT_SYMBOL_GPL(kvm_disable_largepages);
1008
1009 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
1010 {
1011         return __gfn_to_memslot(kvm_memslots(kvm), gfn);
1012 }
1013 EXPORT_SYMBOL_GPL(gfn_to_memslot);
1014
1015 int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
1016 {
1017         struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
1018
1019         if (!memslot || memslot->id >= KVM_USER_MEM_SLOTS ||
1020               memslot->flags & KVM_MEMSLOT_INVALID)
1021                 return 0;
1022
1023         return 1;
1024 }
1025 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
1026
1027 unsigned long kvm_host_page_size(struct kvm *kvm, gfn_t gfn)
1028 {
1029         struct vm_area_struct *vma;
1030         unsigned long addr, size;
1031
1032         size = PAGE_SIZE;
1033
1034         addr = gfn_to_hva(kvm, gfn);
1035         if (kvm_is_error_hva(addr))
1036                 return PAGE_SIZE;
1037
1038         down_read(&current->mm->mmap_sem);
1039         vma = find_vma(current->mm, addr);
1040         if (!vma)
1041                 goto out;
1042
1043         size = vma_kernel_pagesize(vma);
1044
1045 out:
1046         up_read(&current->mm->mmap_sem);
1047
1048         return size;
1049 }
1050
1051 static bool memslot_is_readonly(struct kvm_memory_slot *slot)
1052 {
1053         return slot->flags & KVM_MEM_READONLY;
1054 }
1055
1056 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1057                                        gfn_t *nr_pages, bool write)
1058 {
1059         if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1060                 return KVM_HVA_ERR_BAD;
1061
1062         if (memslot_is_readonly(slot) && write)
1063                 return KVM_HVA_ERR_RO_BAD;
1064
1065         if (nr_pages)
1066                 *nr_pages = slot->npages - (gfn - slot->base_gfn);
1067
1068         return __gfn_to_hva_memslot(slot, gfn);
1069 }
1070
1071 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1072                                      gfn_t *nr_pages)
1073 {
1074         return __gfn_to_hva_many(slot, gfn, nr_pages, true);
1075 }
1076
1077 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot,
1078                                         gfn_t gfn)
1079 {
1080         return gfn_to_hva_many(slot, gfn, NULL);
1081 }
1082 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot);
1083
1084 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
1085 {
1086         return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
1087 }
1088 EXPORT_SYMBOL_GPL(gfn_to_hva);
1089
1090 /*
1091  * If writable is set to false, the hva returned by this function is only
1092  * allowed to be read.
1093  */
1094 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable)
1095 {
1096         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1097         unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false);
1098
1099         if (!kvm_is_error_hva(hva) && writable)
1100                 *writable = !memslot_is_readonly(slot);
1101
1102         return hva;
1103 }
1104
1105 static int kvm_read_hva(void *data, void __user *hva, int len)
1106 {
1107         return __copy_from_user(data, hva, len);
1108 }
1109
1110 static int kvm_read_hva_atomic(void *data, void __user *hva, int len)
1111 {
1112         return __copy_from_user_inatomic(data, hva, len);
1113 }
1114
1115 static int get_user_page_nowait(struct task_struct *tsk, struct mm_struct *mm,
1116         unsigned long start, int write, struct page **page)
1117 {
1118         int flags = FOLL_TOUCH | FOLL_NOWAIT | FOLL_HWPOISON | FOLL_GET;
1119
1120         if (write)
1121                 flags |= FOLL_WRITE;
1122
1123         return __get_user_pages(tsk, mm, start, 1, flags, page, NULL, NULL);
1124 }
1125
1126 int kvm_get_user_page_io(struct task_struct *tsk, struct mm_struct *mm,
1127                          unsigned long addr, bool write_fault,
1128                          struct page **pagep)
1129 {
1130         int npages;
1131         int locked = 1;
1132         int flags = FOLL_TOUCH | FOLL_HWPOISON |
1133                     (pagep ? FOLL_GET : 0) |
1134                     (write_fault ? FOLL_WRITE : 0);
1135
1136         /*
1137          * If retrying the fault, we get here *not* having allowed the filemap
1138          * to wait on the page lock. We should now allow waiting on the IO with
1139          * the mmap semaphore released.
1140          */
1141         down_read(&mm->mmap_sem);
1142         npages = __get_user_pages(tsk, mm, addr, 1, flags, pagep, NULL,
1143                                   &locked);
1144         if (!locked) {
1145                 VM_BUG_ON(npages != -EBUSY);
1146
1147                 if (!pagep)
1148                         return 0;
1149
1150                 /*
1151                  * The previous call has now waited on the IO. Now we can
1152                  * retry and complete. Pass TRIED to ensure we do not re
1153                  * schedule async IO (see e.g. filemap_fault).
1154                  */
1155                 down_read(&mm->mmap_sem);
1156                 npages = __get_user_pages(tsk, mm, addr, 1, flags | FOLL_TRIED,
1157                                           pagep, NULL, NULL);
1158         }
1159         up_read(&mm->mmap_sem);
1160         return npages;
1161 }
1162
1163 static inline int check_user_page_hwpoison(unsigned long addr)
1164 {
1165         int rc, flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_WRITE;
1166
1167         rc = __get_user_pages(current, current->mm, addr, 1,
1168                               flags, NULL, NULL, NULL);
1169         return rc == -EHWPOISON;
1170 }
1171
1172 /*
1173  * The atomic path to get the writable pfn which will be stored in @pfn,
1174  * true indicates success, otherwise false is returned.
1175  */
1176 static bool hva_to_pfn_fast(unsigned long addr, bool atomic, bool *async,
1177                             bool write_fault, bool *writable, pfn_t *pfn)
1178 {
1179         struct page *page[1];
1180         int npages;
1181
1182         if (!(async || atomic))
1183                 return false;
1184
1185         /*
1186          * Fast pin a writable pfn only if it is a write fault request
1187          * or the caller allows to map a writable pfn for a read fault
1188          * request.
1189          */
1190         if (!(write_fault || writable))
1191                 return false;
1192
1193         npages = __get_user_pages_fast(addr, 1, 1, page);
1194         if (npages == 1) {
1195                 *pfn = page_to_pfn(page[0]);
1196
1197                 if (writable)
1198                         *writable = true;
1199                 return true;
1200         }
1201
1202         return false;
1203 }
1204
1205 /*
1206  * The slow path to get the pfn of the specified host virtual address,
1207  * 1 indicates success, -errno is returned if error is detected.
1208  */
1209 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault,
1210                            bool *writable, pfn_t *pfn)
1211 {
1212         struct page *page[1];
1213         int npages = 0;
1214
1215         might_sleep();
1216
1217         if (writable)
1218                 *writable = write_fault;
1219
1220         if (async) {
1221                 down_read(&current->mm->mmap_sem);
1222                 npages = get_user_page_nowait(current, current->mm,
1223                                               addr, write_fault, page);
1224                 up_read(&current->mm->mmap_sem);
1225         } else {
1226                 /*
1227                  * By now we have tried gup_fast, and possibly async_pf, and we
1228                  * are certainly not atomic. Time to retry the gup, allowing
1229                  * mmap semaphore to be relinquished in the case of IO.
1230                  */
1231                 npages = kvm_get_user_page_io(current, current->mm, addr,
1232                                               write_fault, page);
1233         }
1234         if (npages != 1)
1235                 return npages;
1236
1237         /* map read fault as writable if possible */
1238         if (unlikely(!write_fault) && writable) {
1239                 struct page *wpage[1];
1240
1241                 npages = __get_user_pages_fast(addr, 1, 1, wpage);
1242                 if (npages == 1) {
1243                         *writable = true;
1244                         put_page(page[0]);
1245                         page[0] = wpage[0];
1246                 }
1247
1248                 npages = 1;
1249         }
1250         *pfn = page_to_pfn(page[0]);
1251         return npages;
1252 }
1253
1254 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault)
1255 {
1256         if (unlikely(!(vma->vm_flags & VM_READ)))
1257                 return false;
1258
1259         if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE))))
1260                 return false;
1261
1262         return true;
1263 }
1264
1265 /*
1266  * Pin guest page in memory and return its pfn.
1267  * @addr: host virtual address which maps memory to the guest
1268  * @atomic: whether this function can sleep
1269  * @async: whether this function need to wait IO complete if the
1270  *         host page is not in the memory
1271  * @write_fault: whether we should get a writable host page
1272  * @writable: whether it allows to map a writable host page for !@write_fault
1273  *
1274  * The function will map a writable host page for these two cases:
1275  * 1): @write_fault = true
1276  * 2): @write_fault = false && @writable, @writable will tell the caller
1277  *     whether the mapping is writable.
1278  */
1279 static pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
1280                         bool write_fault, bool *writable)
1281 {
1282         struct vm_area_struct *vma;
1283         pfn_t pfn = 0;
1284         int npages;
1285
1286         /* we can do it either atomically or asynchronously, not both */
1287         BUG_ON(atomic && async);
1288
1289         if (hva_to_pfn_fast(addr, atomic, async, write_fault, writable, &pfn))
1290                 return pfn;
1291
1292         if (atomic)
1293                 return KVM_PFN_ERR_FAULT;
1294
1295         npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn);
1296         if (npages == 1)
1297                 return pfn;
1298
1299         down_read(&current->mm->mmap_sem);
1300         if (npages == -EHWPOISON ||
1301               (!async && check_user_page_hwpoison(addr))) {
1302                 pfn = KVM_PFN_ERR_HWPOISON;
1303                 goto exit;
1304         }
1305
1306         vma = find_vma_intersection(current->mm, addr, addr + 1);
1307
1308         if (vma == NULL)
1309                 pfn = KVM_PFN_ERR_FAULT;
1310         else if ((vma->vm_flags & VM_PFNMAP)) {
1311                 pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) +
1312                         vma->vm_pgoff;
1313                 BUG_ON(!kvm_is_mmio_pfn(pfn));
1314         } else {
1315                 if (async && vma_is_valid(vma, write_fault))
1316                         *async = true;
1317                 pfn = KVM_PFN_ERR_FAULT;
1318         }
1319 exit:
1320         up_read(&current->mm->mmap_sem);
1321         return pfn;
1322 }
1323
1324 static pfn_t
1325 __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn, bool atomic,
1326                      bool *async, bool write_fault, bool *writable)
1327 {
1328         unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
1329
1330         if (addr == KVM_HVA_ERR_RO_BAD)
1331                 return KVM_PFN_ERR_RO_FAULT;
1332
1333         if (kvm_is_error_hva(addr))
1334                 return KVM_PFN_NOSLOT;
1335
1336         /* Do not map writable pfn in the readonly memslot. */
1337         if (writable && memslot_is_readonly(slot)) {
1338                 *writable = false;
1339                 writable = NULL;
1340         }
1341
1342         return hva_to_pfn(addr, atomic, async, write_fault,
1343                           writable);
1344 }
1345
1346 static pfn_t __gfn_to_pfn(struct kvm *kvm, gfn_t gfn, bool atomic, bool *async,
1347                           bool write_fault, bool *writable)
1348 {
1349         struct kvm_memory_slot *slot;
1350
1351         if (async)
1352                 *async = false;
1353
1354         slot = gfn_to_memslot(kvm, gfn);
1355
1356         return __gfn_to_pfn_memslot(slot, gfn, atomic, async, write_fault,
1357                                     writable);
1358 }
1359
1360 pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
1361 {
1362         return __gfn_to_pfn(kvm, gfn, true, NULL, true, NULL);
1363 }
1364 EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
1365
1366 pfn_t gfn_to_pfn_async(struct kvm *kvm, gfn_t gfn, bool *async,
1367                        bool write_fault, bool *writable)
1368 {
1369         return __gfn_to_pfn(kvm, gfn, false, async, write_fault, writable);
1370 }
1371 EXPORT_SYMBOL_GPL(gfn_to_pfn_async);
1372
1373 pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
1374 {
1375         return __gfn_to_pfn(kvm, gfn, false, NULL, true, NULL);
1376 }
1377 EXPORT_SYMBOL_GPL(gfn_to_pfn);
1378
1379 pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
1380                       bool *writable)
1381 {
1382         return __gfn_to_pfn(kvm, gfn, false, NULL, write_fault, writable);
1383 }
1384 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
1385
1386 pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
1387 {
1388         return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
1389 }
1390
1391 pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
1392 {
1393         return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
1394 }
1395 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
1396
1397 int gfn_to_page_many_atomic(struct kvm *kvm, gfn_t gfn, struct page **pages,
1398                                                                   int nr_pages)
1399 {
1400         unsigned long addr;
1401         gfn_t entry;
1402
1403         addr = gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, &entry);
1404         if (kvm_is_error_hva(addr))
1405                 return -1;
1406
1407         if (entry < nr_pages)
1408                 return 0;
1409
1410         return __get_user_pages_fast(addr, nr_pages, 1, pages);
1411 }
1412 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
1413
1414 static struct page *kvm_pfn_to_page(pfn_t pfn)
1415 {
1416         if (is_error_noslot_pfn(pfn))
1417                 return KVM_ERR_PTR_BAD_PAGE;
1418
1419         if (kvm_is_mmio_pfn(pfn)) {
1420                 WARN_ON(1);
1421                 return KVM_ERR_PTR_BAD_PAGE;
1422         }
1423
1424         return pfn_to_page(pfn);
1425 }
1426
1427 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
1428 {
1429         pfn_t pfn;
1430
1431         pfn = gfn_to_pfn(kvm, gfn);
1432
1433         return kvm_pfn_to_page(pfn);
1434 }
1435
1436 EXPORT_SYMBOL_GPL(gfn_to_page);
1437
1438 void kvm_release_page_clean(struct page *page)
1439 {
1440         WARN_ON(is_error_page(page));
1441
1442         kvm_release_pfn_clean(page_to_pfn(page));
1443 }
1444 EXPORT_SYMBOL_GPL(kvm_release_page_clean);
1445
1446 void kvm_release_pfn_clean(pfn_t pfn)
1447 {
1448         if (!is_error_noslot_pfn(pfn) && !kvm_is_mmio_pfn(pfn))
1449                 put_page(pfn_to_page(pfn));
1450 }
1451 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
1452
1453 void kvm_release_page_dirty(struct page *page)
1454 {
1455         WARN_ON(is_error_page(page));
1456
1457         kvm_release_pfn_dirty(page_to_pfn(page));
1458 }
1459 EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
1460
1461 static void kvm_release_pfn_dirty(pfn_t pfn)
1462 {
1463         kvm_set_pfn_dirty(pfn);
1464         kvm_release_pfn_clean(pfn);
1465 }
1466
1467 void kvm_set_pfn_dirty(pfn_t pfn)
1468 {
1469         if (!kvm_is_mmio_pfn(pfn)) {
1470                 struct page *page = pfn_to_page(pfn);
1471                 if (!PageReserved(page))
1472                         SetPageDirty(page);
1473         }
1474 }
1475 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
1476
1477 void kvm_set_pfn_accessed(pfn_t pfn)
1478 {
1479         if (!kvm_is_mmio_pfn(pfn))
1480                 mark_page_accessed(pfn_to_page(pfn));
1481 }
1482 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
1483
1484 void kvm_get_pfn(pfn_t pfn)
1485 {
1486         if (!kvm_is_mmio_pfn(pfn))
1487                 get_page(pfn_to_page(pfn));
1488 }
1489 EXPORT_SYMBOL_GPL(kvm_get_pfn);
1490
1491 static int next_segment(unsigned long len, int offset)
1492 {
1493         if (len > PAGE_SIZE - offset)
1494                 return PAGE_SIZE - offset;
1495         else
1496                 return len;
1497 }
1498
1499 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
1500                         int len)
1501 {
1502         int r;
1503         unsigned long addr;
1504
1505         addr = gfn_to_hva_prot(kvm, gfn, NULL);
1506         if (kvm_is_error_hva(addr))
1507                 return -EFAULT;
1508         r = kvm_read_hva(data, (void __user *)addr + offset, len);
1509         if (r)
1510                 return -EFAULT;
1511         return 0;
1512 }
1513 EXPORT_SYMBOL_GPL(kvm_read_guest_page);
1514
1515 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
1516 {
1517         gfn_t gfn = gpa >> PAGE_SHIFT;
1518         int seg;
1519         int offset = offset_in_page(gpa);
1520         int ret;
1521
1522         while ((seg = next_segment(len, offset)) != 0) {
1523                 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
1524                 if (ret < 0)
1525                         return ret;
1526                 offset = 0;
1527                 len -= seg;
1528                 data += seg;
1529                 ++gfn;
1530         }
1531         return 0;
1532 }
1533 EXPORT_SYMBOL_GPL(kvm_read_guest);
1534
1535 int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data,
1536                           unsigned long len)
1537 {
1538         int r;
1539         unsigned long addr;
1540         gfn_t gfn = gpa >> PAGE_SHIFT;
1541         int offset = offset_in_page(gpa);
1542
1543         addr = gfn_to_hva_prot(kvm, gfn, NULL);
1544         if (kvm_is_error_hva(addr))
1545                 return -EFAULT;
1546         pagefault_disable();
1547         r = kvm_read_hva_atomic(data, (void __user *)addr + offset, len);
1548         pagefault_enable();
1549         if (r)
1550                 return -EFAULT;
1551         return 0;
1552 }
1553 EXPORT_SYMBOL(kvm_read_guest_atomic);
1554
1555 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn, const void *data,
1556                          int offset, int len)
1557 {
1558         int r;
1559         unsigned long addr;
1560
1561         addr = gfn_to_hva(kvm, gfn);
1562         if (kvm_is_error_hva(addr))
1563                 return -EFAULT;
1564         r = __copy_to_user((void __user *)addr + offset, data, len);
1565         if (r)
1566                 return -EFAULT;
1567         mark_page_dirty(kvm, gfn);
1568         return 0;
1569 }
1570 EXPORT_SYMBOL_GPL(kvm_write_guest_page);
1571
1572 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
1573                     unsigned long len)
1574 {
1575         gfn_t gfn = gpa >> PAGE_SHIFT;
1576         int seg;
1577         int offset = offset_in_page(gpa);
1578         int ret;
1579
1580         while ((seg = next_segment(len, offset)) != 0) {
1581                 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
1582                 if (ret < 0)
1583                         return ret;
1584                 offset = 0;
1585                 len -= seg;
1586                 data += seg;
1587                 ++gfn;
1588         }
1589         return 0;
1590 }
1591
1592 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1593                               gpa_t gpa, unsigned long len)
1594 {
1595         struct kvm_memslots *slots = kvm_memslots(kvm);
1596         int offset = offset_in_page(gpa);
1597         gfn_t start_gfn = gpa >> PAGE_SHIFT;
1598         gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT;
1599         gfn_t nr_pages_needed = end_gfn - start_gfn + 1;
1600         gfn_t nr_pages_avail;
1601
1602         ghc->gpa = gpa;
1603         ghc->generation = slots->generation;
1604         ghc->len = len;
1605         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1606         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn, &nr_pages_avail);
1607         if (!kvm_is_error_hva(ghc->hva) && nr_pages_avail >= nr_pages_needed) {
1608                 ghc->hva += offset;
1609         } else {
1610                 /*
1611                  * If the requested region crosses two memslots, we still
1612                  * verify that the entire region is valid here.
1613                  */
1614                 while (start_gfn <= end_gfn) {
1615                         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1616                         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn,
1617                                                    &nr_pages_avail);
1618                         if (kvm_is_error_hva(ghc->hva))
1619                                 return -EFAULT;
1620                         start_gfn += nr_pages_avail;
1621                 }
1622                 /* Use the slow path for cross page reads and writes. */
1623                 ghc->memslot = NULL;
1624         }
1625         return 0;
1626 }
1627 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
1628
1629 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1630                            void *data, unsigned long len)
1631 {
1632         struct kvm_memslots *slots = kvm_memslots(kvm);
1633         int r;
1634
1635         BUG_ON(len > ghc->len);
1636
1637         if (slots->generation != ghc->generation)
1638                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1639
1640         if (unlikely(!ghc->memslot))
1641                 return kvm_write_guest(kvm, ghc->gpa, data, len);
1642
1643         if (kvm_is_error_hva(ghc->hva))
1644                 return -EFAULT;
1645
1646         r = __copy_to_user((void __user *)ghc->hva, data, len);
1647         if (r)
1648                 return -EFAULT;
1649         mark_page_dirty_in_slot(kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT);
1650
1651         return 0;
1652 }
1653 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
1654
1655 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1656                            void *data, unsigned long len)
1657 {
1658         struct kvm_memslots *slots = kvm_memslots(kvm);
1659         int r;
1660
1661         BUG_ON(len > ghc->len);
1662
1663         if (slots->generation != ghc->generation)
1664                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1665
1666         if (unlikely(!ghc->memslot))
1667                 return kvm_read_guest(kvm, ghc->gpa, data, len);
1668
1669         if (kvm_is_error_hva(ghc->hva))
1670                 return -EFAULT;
1671
1672         r = __copy_from_user(data, (void __user *)ghc->hva, len);
1673         if (r)
1674                 return -EFAULT;
1675
1676         return 0;
1677 }
1678 EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
1679
1680 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
1681 {
1682         const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0)));
1683
1684         return kvm_write_guest_page(kvm, gfn, zero_page, offset, len);
1685 }
1686 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
1687
1688 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
1689 {
1690         gfn_t gfn = gpa >> PAGE_SHIFT;
1691         int seg;
1692         int offset = offset_in_page(gpa);
1693         int ret;
1694
1695         while ((seg = next_segment(len, offset)) != 0) {
1696                 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
1697                 if (ret < 0)
1698                         return ret;
1699                 offset = 0;
1700                 len -= seg;
1701                 ++gfn;
1702         }
1703         return 0;
1704 }
1705 EXPORT_SYMBOL_GPL(kvm_clear_guest);
1706
1707 static void mark_page_dirty_in_slot(struct kvm *kvm,
1708                                     struct kvm_memory_slot *memslot,
1709                                     gfn_t gfn)
1710 {
1711         if (memslot && memslot->dirty_bitmap) {
1712                 unsigned long rel_gfn = gfn - memslot->base_gfn;
1713
1714                 set_bit_le(rel_gfn, memslot->dirty_bitmap);
1715         }
1716 }
1717
1718 void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
1719 {
1720         struct kvm_memory_slot *memslot;
1721
1722         memslot = gfn_to_memslot(kvm, gfn);
1723         mark_page_dirty_in_slot(kvm, memslot, gfn);
1724 }
1725 EXPORT_SYMBOL_GPL(mark_page_dirty);
1726
1727 /*
1728  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
1729  */
1730 void kvm_vcpu_block(struct kvm_vcpu *vcpu)
1731 {
1732         DEFINE_WAIT(wait);
1733
1734         for (;;) {
1735                 prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
1736
1737                 if (kvm_arch_vcpu_runnable(vcpu)) {
1738                         kvm_make_request(KVM_REQ_UNHALT, vcpu);
1739                         break;
1740                 }
1741                 if (kvm_cpu_has_pending_timer(vcpu))
1742                         break;
1743                 if (signal_pending(current))
1744                         break;
1745
1746                 schedule();
1747         }
1748
1749         finish_wait(&vcpu->wq, &wait);
1750 }
1751 EXPORT_SYMBOL_GPL(kvm_vcpu_block);
1752
1753 #ifndef CONFIG_S390
1754 /*
1755  * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode.
1756  */
1757 void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
1758 {
1759         int me;
1760         int cpu = vcpu->cpu;
1761         wait_queue_head_t *wqp;
1762
1763         wqp = kvm_arch_vcpu_wq(vcpu);
1764         if (waitqueue_active(wqp)) {
1765                 wake_up_interruptible(wqp);
1766                 ++vcpu->stat.halt_wakeup;
1767         }
1768
1769         me = get_cpu();
1770         if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
1771                 if (kvm_arch_vcpu_should_kick(vcpu))
1772                         smp_send_reschedule(cpu);
1773         put_cpu();
1774 }
1775 EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
1776 #endif /* !CONFIG_S390 */
1777
1778 int kvm_vcpu_yield_to(struct kvm_vcpu *target)
1779 {
1780         struct pid *pid;
1781         struct task_struct *task = NULL;
1782         int ret = 0;
1783
1784         rcu_read_lock();
1785         pid = rcu_dereference(target->pid);
1786         if (pid)
1787                 task = get_pid_task(target->pid, PIDTYPE_PID);
1788         rcu_read_unlock();
1789         if (!task)
1790                 return ret;
1791         if (task->flags & PF_VCPU) {
1792                 put_task_struct(task);
1793                 return ret;
1794         }
1795         ret = yield_to(task, 1);
1796         put_task_struct(task);
1797
1798         return ret;
1799 }
1800 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to);
1801
1802 /*
1803  * Helper that checks whether a VCPU is eligible for directed yield.
1804  * Most eligible candidate to yield is decided by following heuristics:
1805  *
1806  *  (a) VCPU which has not done pl-exit or cpu relax intercepted recently
1807  *  (preempted lock holder), indicated by @in_spin_loop.
1808  *  Set at the beiginning and cleared at the end of interception/PLE handler.
1809  *
1810  *  (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get
1811  *  chance last time (mostly it has become eligible now since we have probably
1812  *  yielded to lockholder in last iteration. This is done by toggling
1813  *  @dy_eligible each time a VCPU checked for eligibility.)
1814  *
1815  *  Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding
1816  *  to preempted lock-holder could result in wrong VCPU selection and CPU
1817  *  burning. Giving priority for a potential lock-holder increases lock
1818  *  progress.
1819  *
1820  *  Since algorithm is based on heuristics, accessing another VCPU data without
1821  *  locking does not harm. It may result in trying to yield to  same VCPU, fail
1822  *  and continue with next VCPU and so on.
1823  */
1824 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu)
1825 {
1826 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT
1827         bool eligible;
1828
1829         eligible = !vcpu->spin_loop.in_spin_loop ||
1830                     vcpu->spin_loop.dy_eligible;
1831
1832         if (vcpu->spin_loop.in_spin_loop)
1833                 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible);
1834
1835         return eligible;
1836 #else
1837         return true;
1838 #endif
1839 }
1840
1841 void kvm_vcpu_on_spin(struct kvm_vcpu *me)
1842 {
1843         struct kvm *kvm = me->kvm;
1844         struct kvm_vcpu *vcpu;
1845         int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
1846         int yielded = 0;
1847         int try = 3;
1848         int pass;
1849         int i;
1850
1851         kvm_vcpu_set_in_spin_loop(me, true);
1852         /*
1853          * We boost the priority of a VCPU that is runnable but not
1854          * currently running, because it got preempted by something
1855          * else and called schedule in __vcpu_run.  Hopefully that
1856          * VCPU is holding the lock that we need and will release it.
1857          * We approximate round-robin by starting at the last boosted VCPU.
1858          */
1859         for (pass = 0; pass < 2 && !yielded && try; pass++) {
1860                 kvm_for_each_vcpu(i, vcpu, kvm) {
1861                         if (!pass && i <= last_boosted_vcpu) {
1862                                 i = last_boosted_vcpu;
1863                                 continue;
1864                         } else if (pass && i > last_boosted_vcpu)
1865                                 break;
1866                         if (!ACCESS_ONCE(vcpu->preempted))
1867                                 continue;
1868                         if (vcpu == me)
1869                                 continue;
1870                         if (waitqueue_active(&vcpu->wq) && !kvm_arch_vcpu_runnable(vcpu))
1871                                 continue;
1872                         if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
1873                                 continue;
1874
1875                         yielded = kvm_vcpu_yield_to(vcpu);
1876                         if (yielded > 0) {
1877                                 kvm->last_boosted_vcpu = i;
1878                                 break;
1879                         } else if (yielded < 0) {
1880                                 try--;
1881                                 if (!try)
1882                                         break;
1883                         }
1884                 }
1885         }
1886         kvm_vcpu_set_in_spin_loop(me, false);
1887
1888         /* Ensure vcpu is not eligible during next spinloop */
1889         kvm_vcpu_set_dy_eligible(me, false);
1890 }
1891 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
1892
1893 static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1894 {
1895         struct kvm_vcpu *vcpu = vma->vm_file->private_data;
1896         struct page *page;
1897
1898         if (vmf->pgoff == 0)
1899                 page = virt_to_page(vcpu->run);
1900 #ifdef CONFIG_X86
1901         else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
1902                 page = virt_to_page(vcpu->arch.pio_data);
1903 #endif
1904 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
1905         else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
1906                 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
1907 #endif
1908         else
1909                 return kvm_arch_vcpu_fault(vcpu, vmf);
1910         get_page(page);
1911         vmf->page = page;
1912         return 0;
1913 }
1914
1915 static const struct vm_operations_struct kvm_vcpu_vm_ops = {
1916         .fault = kvm_vcpu_fault,
1917 };
1918
1919 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
1920 {
1921         vma->vm_ops = &kvm_vcpu_vm_ops;
1922         return 0;
1923 }
1924
1925 static int kvm_vcpu_release(struct inode *inode, struct file *filp)
1926 {
1927         struct kvm_vcpu *vcpu = filp->private_data;
1928
1929         kvm_put_kvm(vcpu->kvm);
1930         return 0;
1931 }
1932
1933 static struct file_operations kvm_vcpu_fops = {
1934         .release        = kvm_vcpu_release,
1935         .unlocked_ioctl = kvm_vcpu_ioctl,
1936 #ifdef CONFIG_COMPAT
1937         .compat_ioctl   = kvm_vcpu_compat_ioctl,
1938 #endif
1939         .mmap           = kvm_vcpu_mmap,
1940         .llseek         = noop_llseek,
1941 };
1942
1943 /*
1944  * Allocates an inode for the vcpu.
1945  */
1946 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
1947 {
1948         return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC);
1949 }
1950
1951 /*
1952  * Creates some virtual cpus.  Good luck creating more than one.
1953  */
1954 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
1955 {
1956         int r;
1957         struct kvm_vcpu *vcpu, *v;
1958
1959         if (id >= KVM_MAX_VCPUS)
1960                 return -EINVAL;
1961
1962         vcpu = kvm_arch_vcpu_create(kvm, id);
1963         if (IS_ERR(vcpu))
1964                 return PTR_ERR(vcpu);
1965
1966         preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
1967
1968         r = kvm_arch_vcpu_setup(vcpu);
1969         if (r)
1970                 goto vcpu_destroy;
1971
1972         mutex_lock(&kvm->lock);
1973         if (!kvm_vcpu_compatible(vcpu)) {
1974                 r = -EINVAL;
1975                 goto unlock_vcpu_destroy;
1976         }
1977         if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
1978                 r = -EINVAL;
1979                 goto unlock_vcpu_destroy;
1980         }
1981
1982         kvm_for_each_vcpu(r, v, kvm)
1983                 if (v->vcpu_id == id) {
1984                         r = -EEXIST;
1985                         goto unlock_vcpu_destroy;
1986                 }
1987
1988         BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]);
1989
1990         /* Now it's all set up, let userspace reach it */
1991         kvm_get_kvm(kvm);
1992         r = create_vcpu_fd(vcpu);
1993         if (r < 0) {
1994                 kvm_put_kvm(kvm);
1995                 goto unlock_vcpu_destroy;
1996         }
1997
1998         kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu;
1999         smp_wmb();
2000         atomic_inc(&kvm->online_vcpus);
2001
2002         mutex_unlock(&kvm->lock);
2003         kvm_arch_vcpu_postcreate(vcpu);
2004         return r;
2005
2006 unlock_vcpu_destroy:
2007         mutex_unlock(&kvm->lock);
2008 vcpu_destroy:
2009         kvm_arch_vcpu_destroy(vcpu);
2010         return r;
2011 }
2012
2013 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
2014 {
2015         if (sigset) {
2016                 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
2017                 vcpu->sigset_active = 1;
2018                 vcpu->sigset = *sigset;
2019         } else
2020                 vcpu->sigset_active = 0;
2021         return 0;
2022 }
2023
2024 static long kvm_vcpu_ioctl(struct file *filp,
2025                            unsigned int ioctl, unsigned long arg)
2026 {
2027         struct kvm_vcpu *vcpu = filp->private_data;
2028         void __user *argp = (void __user *)arg;
2029         int r;
2030         struct kvm_fpu *fpu = NULL;
2031         struct kvm_sregs *kvm_sregs = NULL;
2032
2033         if (vcpu->kvm->mm != current->mm)
2034                 return -EIO;
2035
2036         if (unlikely(_IOC_TYPE(ioctl) != KVMIO))
2037                 return -EINVAL;
2038
2039 #if defined(CONFIG_S390) || defined(CONFIG_PPC) || defined(CONFIG_MIPS)
2040         /*
2041          * Special cases: vcpu ioctls that are asynchronous to vcpu execution,
2042          * so vcpu_load() would break it.
2043          */
2044         if (ioctl == KVM_S390_INTERRUPT || ioctl == KVM_INTERRUPT)
2045                 return kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2046 #endif
2047
2048
2049         r = vcpu_load(vcpu);
2050         if (r)
2051                 return r;
2052         switch (ioctl) {
2053         case KVM_RUN:
2054                 r = -EINVAL;
2055                 if (arg)
2056                         goto out;
2057                 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
2058                 trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
2059                 break;
2060         case KVM_GET_REGS: {
2061                 struct kvm_regs *kvm_regs;
2062
2063                 r = -ENOMEM;
2064                 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL);
2065                 if (!kvm_regs)
2066                         goto out;
2067                 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
2068                 if (r)
2069                         goto out_free1;
2070                 r = -EFAULT;
2071                 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
2072                         goto out_free1;
2073                 r = 0;
2074 out_free1:
2075                 kfree(kvm_regs);
2076                 break;
2077         }
2078         case KVM_SET_REGS: {
2079                 struct kvm_regs *kvm_regs;
2080
2081                 r = -ENOMEM;
2082                 kvm_regs = memdup_user(argp, sizeof(*kvm_regs));
2083                 if (IS_ERR(kvm_regs)) {
2084                         r = PTR_ERR(kvm_regs);
2085                         goto out;
2086                 }
2087                 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
2088                 kfree(kvm_regs);
2089                 break;
2090         }
2091         case KVM_GET_SREGS: {
2092                 kvm_sregs = kzalloc(sizeof(struct kvm_sregs), GFP_KERNEL);
2093                 r = -ENOMEM;
2094                 if (!kvm_sregs)
2095                         goto out;
2096                 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
2097                 if (r)
2098                         goto out;
2099                 r = -EFAULT;
2100                 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
2101                         goto out;
2102                 r = 0;
2103                 break;
2104         }
2105         case KVM_SET_SREGS: {
2106                 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs));
2107                 if (IS_ERR(kvm_sregs)) {
2108                         r = PTR_ERR(kvm_sregs);
2109                         kvm_sregs = NULL;
2110                         goto out;
2111                 }
2112                 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
2113                 break;
2114         }
2115         case KVM_GET_MP_STATE: {
2116                 struct kvm_mp_state mp_state;
2117
2118                 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
2119                 if (r)
2120                         goto out;
2121                 r = -EFAULT;
2122                 if (copy_to_user(argp, &mp_state, sizeof mp_state))
2123                         goto out;
2124                 r = 0;
2125                 break;
2126         }
2127         case KVM_SET_MP_STATE: {
2128                 struct kvm_mp_state mp_state;
2129
2130                 r = -EFAULT;
2131                 if (copy_from_user(&mp_state, argp, sizeof mp_state))
2132                         goto out;
2133                 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
2134                 break;
2135         }
2136         case KVM_TRANSLATE: {
2137                 struct kvm_translation tr;
2138
2139                 r = -EFAULT;
2140                 if (copy_from_user(&tr, argp, sizeof tr))
2141                         goto out;
2142                 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
2143                 if (r)
2144                         goto out;
2145                 r = -EFAULT;
2146                 if (copy_to_user(argp, &tr, sizeof tr))
2147                         goto out;
2148                 r = 0;
2149                 break;
2150         }
2151         case KVM_SET_GUEST_DEBUG: {
2152                 struct kvm_guest_debug dbg;
2153
2154                 r = -EFAULT;
2155                 if (copy_from_user(&dbg, argp, sizeof dbg))
2156                         goto out;
2157                 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
2158                 break;
2159         }
2160         case KVM_SET_SIGNAL_MASK: {
2161                 struct kvm_signal_mask __user *sigmask_arg = argp;
2162                 struct kvm_signal_mask kvm_sigmask;
2163                 sigset_t sigset, *p;
2164
2165                 p = NULL;
2166                 if (argp) {
2167                         r = -EFAULT;
2168                         if (copy_from_user(&kvm_sigmask, argp,
2169                                            sizeof kvm_sigmask))
2170                                 goto out;
2171                         r = -EINVAL;
2172                         if (kvm_sigmask.len != sizeof sigset)
2173                                 goto out;
2174                         r = -EFAULT;
2175                         if (copy_from_user(&sigset, sigmask_arg->sigset,
2176                                            sizeof sigset))
2177                                 goto out;
2178                         p = &sigset;
2179                 }
2180                 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
2181                 break;
2182         }
2183         case KVM_GET_FPU: {
2184                 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL);
2185                 r = -ENOMEM;
2186                 if (!fpu)
2187                         goto out;
2188                 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
2189                 if (r)
2190                         goto out;
2191                 r = -EFAULT;
2192                 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
2193                         goto out;
2194                 r = 0;
2195                 break;
2196         }
2197         case KVM_SET_FPU: {
2198                 fpu = memdup_user(argp, sizeof(*fpu));
2199                 if (IS_ERR(fpu)) {
2200                         r = PTR_ERR(fpu);
2201                         fpu = NULL;
2202                         goto out;
2203                 }
2204                 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
2205                 break;
2206         }
2207         default:
2208                 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2209         }
2210 out:
2211         vcpu_put(vcpu);
2212         kfree(fpu);
2213         kfree(kvm_sregs);
2214         return r;
2215 }
2216
2217 #ifdef CONFIG_COMPAT
2218 static long kvm_vcpu_compat_ioctl(struct file *filp,
2219                                   unsigned int ioctl, unsigned long arg)
2220 {
2221         struct kvm_vcpu *vcpu = filp->private_data;
2222         void __user *argp = compat_ptr(arg);
2223         int r;
2224
2225         if (vcpu->kvm->mm != current->mm)
2226                 return -EIO;
2227
2228         switch (ioctl) {
2229         case KVM_SET_SIGNAL_MASK: {
2230                 struct kvm_signal_mask __user *sigmask_arg = argp;
2231                 struct kvm_signal_mask kvm_sigmask;
2232                 compat_sigset_t csigset;
2233                 sigset_t sigset;
2234
2235                 if (argp) {
2236                         r = -EFAULT;
2237                         if (copy_from_user(&kvm_sigmask, argp,
2238                                            sizeof kvm_sigmask))
2239                                 goto out;
2240                         r = -EINVAL;
2241                         if (kvm_sigmask.len != sizeof csigset)
2242                                 goto out;
2243                         r = -EFAULT;
2244                         if (copy_from_user(&csigset, sigmask_arg->sigset,
2245                                            sizeof csigset))
2246                                 goto out;
2247                         sigset_from_compat(&sigset, &csigset);
2248                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
2249                 } else
2250                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL);
2251                 break;
2252         }
2253         default:
2254                 r = kvm_vcpu_ioctl(filp, ioctl, arg);
2255         }
2256
2257 out:
2258         return r;
2259 }
2260 #endif
2261
2262 static int kvm_device_ioctl_attr(struct kvm_device *dev,
2263                                  int (*accessor)(struct kvm_device *dev,
2264                                                  struct kvm_device_attr *attr),
2265                                  unsigned long arg)
2266 {
2267         struct kvm_device_attr attr;
2268
2269         if (!accessor)
2270                 return -EPERM;
2271
2272         if (copy_from_user(&attr, (void __user *)arg, sizeof(attr)))
2273                 return -EFAULT;
2274
2275         return accessor(dev, &attr);
2276 }
2277
2278 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl,
2279                              unsigned long arg)
2280 {
2281         struct kvm_device *dev = filp->private_data;
2282
2283         switch (ioctl) {
2284         case KVM_SET_DEVICE_ATTR:
2285                 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg);
2286         case KVM_GET_DEVICE_ATTR:
2287                 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg);
2288         case KVM_HAS_DEVICE_ATTR:
2289                 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg);
2290         default:
2291                 if (dev->ops->ioctl)
2292                         return dev->ops->ioctl(dev, ioctl, arg);
2293
2294                 return -ENOTTY;
2295         }
2296 }
2297
2298 static int kvm_device_release(struct inode *inode, struct file *filp)
2299 {
2300         struct kvm_device *dev = filp->private_data;
2301         struct kvm *kvm = dev->kvm;
2302
2303         kvm_put_kvm(kvm);
2304         return 0;
2305 }
2306
2307 static const struct file_operations kvm_device_fops = {
2308         .unlocked_ioctl = kvm_device_ioctl,
2309 #ifdef CONFIG_COMPAT
2310         .compat_ioctl = kvm_device_ioctl,
2311 #endif
2312         .release = kvm_device_release,
2313 };
2314
2315 struct kvm_device *kvm_device_from_filp(struct file *filp)
2316 {
2317         if (filp->f_op != &kvm_device_fops)
2318                 return NULL;
2319
2320         return filp->private_data;
2321 }
2322
2323 static struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = {
2324 #ifdef CONFIG_KVM_MPIC
2325         [KVM_DEV_TYPE_FSL_MPIC_20]      = &kvm_mpic_ops,
2326         [KVM_DEV_TYPE_FSL_MPIC_42]      = &kvm_mpic_ops,
2327 #endif
2328
2329 #ifdef CONFIG_KVM_XICS
2330         [KVM_DEV_TYPE_XICS]             = &kvm_xics_ops,
2331 #endif
2332 };
2333
2334 int kvm_register_device_ops(struct kvm_device_ops *ops, u32 type)
2335 {
2336         if (type >= ARRAY_SIZE(kvm_device_ops_table))
2337                 return -ENOSPC;
2338
2339         if (kvm_device_ops_table[type] != NULL)
2340                 return -EEXIST;
2341
2342         kvm_device_ops_table[type] = ops;
2343         return 0;
2344 }
2345
2346 static int kvm_ioctl_create_device(struct kvm *kvm,
2347                                    struct kvm_create_device *cd)
2348 {
2349         struct kvm_device_ops *ops = NULL;
2350         struct kvm_device *dev;
2351         bool test = cd->flags & KVM_CREATE_DEVICE_TEST;
2352         int ret;
2353
2354         if (cd->type >= ARRAY_SIZE(kvm_device_ops_table))
2355                 return -ENODEV;
2356
2357         ops = kvm_device_ops_table[cd->type];
2358         if (ops == NULL)
2359                 return -ENODEV;
2360
2361         if (test)
2362                 return 0;
2363
2364         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2365         if (!dev)
2366                 return -ENOMEM;
2367
2368         dev->ops = ops;
2369         dev->kvm = kvm;
2370
2371         ret = ops->create(dev, cd->type);
2372         if (ret < 0) {
2373                 kfree(dev);
2374                 return ret;
2375         }
2376
2377         ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC);
2378         if (ret < 0) {
2379                 ops->destroy(dev);
2380                 return ret;
2381         }
2382
2383         list_add(&dev->vm_node, &kvm->devices);
2384         kvm_get_kvm(kvm);
2385         cd->fd = ret;
2386         return 0;
2387 }
2388
2389 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
2390 {
2391         switch (arg) {
2392         case KVM_CAP_USER_MEMORY:
2393         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
2394         case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
2395 #ifdef CONFIG_KVM_APIC_ARCHITECTURE
2396         case KVM_CAP_SET_BOOT_CPU_ID:
2397 #endif
2398         case KVM_CAP_INTERNAL_ERROR_DATA:
2399 #ifdef CONFIG_HAVE_KVM_MSI
2400         case KVM_CAP_SIGNAL_MSI:
2401 #endif
2402 #ifdef CONFIG_HAVE_KVM_IRQFD
2403         case KVM_CAP_IRQFD_RESAMPLE:
2404 #endif
2405         case KVM_CAP_CHECK_EXTENSION_VM:
2406                 return 1;
2407 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2408         case KVM_CAP_IRQ_ROUTING:
2409                 return KVM_MAX_IRQ_ROUTES;
2410 #endif
2411         default:
2412                 break;
2413         }
2414         return kvm_vm_ioctl_check_extension(kvm, arg);
2415 }
2416
2417 static long kvm_vm_ioctl(struct file *filp,
2418                            unsigned int ioctl, unsigned long arg)
2419 {
2420         struct kvm *kvm = filp->private_data;
2421         void __user *argp = (void __user *)arg;
2422         int r;
2423
2424         if (kvm->mm != current->mm)
2425                 return -EIO;
2426         switch (ioctl) {
2427         case KVM_CREATE_VCPU:
2428                 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
2429                 break;
2430         case KVM_SET_USER_MEMORY_REGION: {
2431                 struct kvm_userspace_memory_region kvm_userspace_mem;
2432
2433                 r = -EFAULT;
2434                 if (copy_from_user(&kvm_userspace_mem, argp,
2435                                                 sizeof kvm_userspace_mem))
2436                         goto out;
2437
2438                 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem);
2439                 break;
2440         }
2441         case KVM_GET_DIRTY_LOG: {
2442                 struct kvm_dirty_log log;
2443
2444                 r = -EFAULT;
2445                 if (copy_from_user(&log, argp, sizeof log))
2446                         goto out;
2447                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2448                 break;
2449         }
2450 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2451         case KVM_REGISTER_COALESCED_MMIO: {
2452                 struct kvm_coalesced_mmio_zone zone;
2453                 r = -EFAULT;
2454                 if (copy_from_user(&zone, argp, sizeof zone))
2455                         goto out;
2456                 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
2457                 break;
2458         }
2459         case KVM_UNREGISTER_COALESCED_MMIO: {
2460                 struct kvm_coalesced_mmio_zone zone;
2461                 r = -EFAULT;
2462                 if (copy_from_user(&zone, argp, sizeof zone))
2463                         goto out;
2464                 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
2465                 break;
2466         }
2467 #endif
2468         case KVM_IRQFD: {
2469                 struct kvm_irqfd data;
2470
2471                 r = -EFAULT;
2472                 if (copy_from_user(&data, argp, sizeof data))
2473                         goto out;
2474                 r = kvm_irqfd(kvm, &data);
2475                 break;
2476         }
2477         case KVM_IOEVENTFD: {
2478                 struct kvm_ioeventfd data;
2479
2480                 r = -EFAULT;
2481                 if (copy_from_user(&data, argp, sizeof data))
2482                         goto out;
2483                 r = kvm_ioeventfd(kvm, &data);
2484                 break;
2485         }
2486 #ifdef CONFIG_KVM_APIC_ARCHITECTURE
2487         case KVM_SET_BOOT_CPU_ID:
2488                 r = 0;
2489                 mutex_lock(&kvm->lock);
2490                 if (atomic_read(&kvm->online_vcpus) != 0)
2491                         r = -EBUSY;
2492                 else
2493                         kvm->bsp_vcpu_id = arg;
2494                 mutex_unlock(&kvm->lock);
2495                 break;
2496 #endif
2497 #ifdef CONFIG_HAVE_KVM_MSI
2498         case KVM_SIGNAL_MSI: {
2499                 struct kvm_msi msi;
2500
2501                 r = -EFAULT;
2502                 if (copy_from_user(&msi, argp, sizeof msi))
2503                         goto out;
2504                 r = kvm_send_userspace_msi(kvm, &msi);
2505                 break;
2506         }
2507 #endif
2508 #ifdef __KVM_HAVE_IRQ_LINE
2509         case KVM_IRQ_LINE_STATUS:
2510         case KVM_IRQ_LINE: {
2511                 struct kvm_irq_level irq_event;
2512
2513                 r = -EFAULT;
2514                 if (copy_from_user(&irq_event, argp, sizeof irq_event))
2515                         goto out;
2516
2517                 r = kvm_vm_ioctl_irq_line(kvm, &irq_event,
2518                                         ioctl == KVM_IRQ_LINE_STATUS);
2519                 if (r)
2520                         goto out;
2521
2522                 r = -EFAULT;
2523                 if (ioctl == KVM_IRQ_LINE_STATUS) {
2524                         if (copy_to_user(argp, &irq_event, sizeof irq_event))
2525                                 goto out;
2526                 }
2527
2528                 r = 0;
2529                 break;
2530         }
2531 #endif
2532 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2533         case KVM_SET_GSI_ROUTING: {
2534                 struct kvm_irq_routing routing;
2535                 struct kvm_irq_routing __user *urouting;
2536                 struct kvm_irq_routing_entry *entries;
2537
2538                 r = -EFAULT;
2539                 if (copy_from_user(&routing, argp, sizeof(routing)))
2540                         goto out;
2541                 r = -EINVAL;
2542                 if (routing.nr >= KVM_MAX_IRQ_ROUTES)
2543                         goto out;
2544                 if (routing.flags)
2545                         goto out;
2546                 r = -ENOMEM;
2547                 entries = vmalloc(routing.nr * sizeof(*entries));
2548                 if (!entries)
2549                         goto out;
2550                 r = -EFAULT;
2551                 urouting = argp;
2552                 if (copy_from_user(entries, urouting->entries,
2553                                    routing.nr * sizeof(*entries)))
2554                         goto out_free_irq_routing;
2555                 r = kvm_set_irq_routing(kvm, entries, routing.nr,
2556                                         routing.flags);
2557         out_free_irq_routing:
2558                 vfree(entries);
2559                 break;
2560         }
2561 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */
2562         case KVM_CREATE_DEVICE: {
2563                 struct kvm_create_device cd;
2564
2565                 r = -EFAULT;
2566                 if (copy_from_user(&cd, argp, sizeof(cd)))
2567                         goto out;
2568
2569                 r = kvm_ioctl_create_device(kvm, &cd);
2570                 if (r)
2571                         goto out;
2572
2573                 r = -EFAULT;
2574                 if (copy_to_user(argp, &cd, sizeof(cd)))
2575                         goto out;
2576
2577                 r = 0;
2578                 break;
2579         }
2580         case KVM_CHECK_EXTENSION:
2581                 r = kvm_vm_ioctl_check_extension_generic(kvm, arg);
2582                 break;
2583         default:
2584                 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
2585                 if (r == -ENOTTY)
2586                         r = kvm_vm_ioctl_assigned_device(kvm, ioctl, arg);
2587         }
2588 out:
2589         return r;
2590 }
2591
2592 #ifdef CONFIG_COMPAT
2593 struct compat_kvm_dirty_log {
2594         __u32 slot;
2595         __u32 padding1;
2596         union {
2597                 compat_uptr_t dirty_bitmap; /* one bit per page */
2598                 __u64 padding2;
2599         };
2600 };
2601
2602 static long kvm_vm_compat_ioctl(struct file *filp,
2603                            unsigned int ioctl, unsigned long arg)
2604 {
2605         struct kvm *kvm = filp->private_data;
2606         int r;
2607
2608         if (kvm->mm != current->mm)
2609                 return -EIO;
2610         switch (ioctl) {
2611         case KVM_GET_DIRTY_LOG: {
2612                 struct compat_kvm_dirty_log compat_log;
2613                 struct kvm_dirty_log log;
2614
2615                 r = -EFAULT;
2616                 if (copy_from_user(&compat_log, (void __user *)arg,
2617                                    sizeof(compat_log)))
2618                         goto out;
2619                 log.slot         = compat_log.slot;
2620                 log.padding1     = compat_log.padding1;
2621                 log.padding2     = compat_log.padding2;
2622                 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
2623
2624                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2625                 break;
2626         }
2627         default:
2628                 r = kvm_vm_ioctl(filp, ioctl, arg);
2629         }
2630
2631 out:
2632         return r;
2633 }
2634 #endif
2635
2636 static struct file_operations kvm_vm_fops = {
2637         .release        = kvm_vm_release,
2638         .unlocked_ioctl = kvm_vm_ioctl,
2639 #ifdef CONFIG_COMPAT
2640         .compat_ioctl   = kvm_vm_compat_ioctl,
2641 #endif
2642         .llseek         = noop_llseek,
2643 };
2644
2645 static int kvm_dev_ioctl_create_vm(unsigned long type)
2646 {
2647         int r;
2648         struct kvm *kvm;
2649
2650         kvm = kvm_create_vm(type);
2651         if (IS_ERR(kvm))
2652                 return PTR_ERR(kvm);
2653 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2654         r = kvm_coalesced_mmio_init(kvm);
2655         if (r < 0) {
2656                 kvm_put_kvm(kvm);
2657                 return r;
2658         }
2659 #endif
2660         r = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, O_RDWR | O_CLOEXEC);
2661         if (r < 0)
2662                 kvm_put_kvm(kvm);
2663
2664         return r;
2665 }
2666
2667 static long kvm_dev_ioctl(struct file *filp,
2668                           unsigned int ioctl, unsigned long arg)
2669 {
2670         long r = -EINVAL;
2671
2672         switch (ioctl) {
2673         case KVM_GET_API_VERSION:
2674                 if (arg)
2675                         goto out;
2676                 r = KVM_API_VERSION;
2677                 break;
2678         case KVM_CREATE_VM:
2679                 r = kvm_dev_ioctl_create_vm(arg);
2680                 break;
2681         case KVM_CHECK_EXTENSION:
2682                 r = kvm_vm_ioctl_check_extension_generic(NULL, arg);
2683                 break;
2684         case KVM_GET_VCPU_MMAP_SIZE:
2685                 if (arg)
2686                         goto out;
2687                 r = PAGE_SIZE;     /* struct kvm_run */
2688 #ifdef CONFIG_X86
2689                 r += PAGE_SIZE;    /* pio data page */
2690 #endif
2691 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2692                 r += PAGE_SIZE;    /* coalesced mmio ring page */
2693 #endif
2694                 break;
2695         case KVM_TRACE_ENABLE:
2696         case KVM_TRACE_PAUSE:
2697         case KVM_TRACE_DISABLE:
2698                 r = -EOPNOTSUPP;
2699                 break;
2700         default:
2701                 return kvm_arch_dev_ioctl(filp, ioctl, arg);
2702         }
2703 out:
2704         return r;
2705 }
2706
2707 static struct file_operations kvm_chardev_ops = {
2708         .unlocked_ioctl = kvm_dev_ioctl,
2709         .compat_ioctl   = kvm_dev_ioctl,
2710         .llseek         = noop_llseek,
2711 };
2712
2713 static struct miscdevice kvm_dev = {
2714         KVM_MINOR,
2715         "kvm",
2716         &kvm_chardev_ops,
2717 };
2718
2719 static void hardware_enable_nolock(void *junk)
2720 {
2721         int cpu = raw_smp_processor_id();
2722         int r;
2723
2724         if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
2725                 return;
2726
2727         cpumask_set_cpu(cpu, cpus_hardware_enabled);
2728
2729         r = kvm_arch_hardware_enable();
2730
2731         if (r) {
2732                 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
2733                 atomic_inc(&hardware_enable_failed);
2734                 printk(KERN_INFO "kvm: enabling virtualization on "
2735                                  "CPU%d failed\n", cpu);
2736         }
2737 }
2738
2739 static void hardware_enable(void)
2740 {
2741         raw_spin_lock(&kvm_count_lock);
2742         if (kvm_usage_count)
2743                 hardware_enable_nolock(NULL);
2744         raw_spin_unlock(&kvm_count_lock);
2745 }
2746
2747 static void hardware_disable_nolock(void *junk)
2748 {
2749         int cpu = raw_smp_processor_id();
2750
2751         if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
2752                 return;
2753         cpumask_clear_cpu(cpu, cpus_hardware_enabled);
2754         kvm_arch_hardware_disable();
2755 }
2756
2757 static void hardware_disable(void)
2758 {
2759         raw_spin_lock(&kvm_count_lock);
2760         if (kvm_usage_count)
2761                 hardware_disable_nolock(NULL);
2762         raw_spin_unlock(&kvm_count_lock);
2763 }
2764
2765 static void hardware_disable_all_nolock(void)
2766 {
2767         BUG_ON(!kvm_usage_count);
2768
2769         kvm_usage_count--;
2770         if (!kvm_usage_count)
2771                 on_each_cpu(hardware_disable_nolock, NULL, 1);
2772 }
2773
2774 static void hardware_disable_all(void)
2775 {
2776         raw_spin_lock(&kvm_count_lock);
2777         hardware_disable_all_nolock();
2778         raw_spin_unlock(&kvm_count_lock);
2779 }
2780
2781 static int hardware_enable_all(void)
2782 {
2783         int r = 0;
2784
2785         raw_spin_lock(&kvm_count_lock);
2786
2787         kvm_usage_count++;
2788         if (kvm_usage_count == 1) {
2789                 atomic_set(&hardware_enable_failed, 0);
2790                 on_each_cpu(hardware_enable_nolock, NULL, 1);
2791
2792                 if (atomic_read(&hardware_enable_failed)) {
2793                         hardware_disable_all_nolock();
2794                         r = -EBUSY;
2795                 }
2796         }
2797
2798         raw_spin_unlock(&kvm_count_lock);
2799
2800         return r;
2801 }
2802
2803 static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
2804                            void *v)
2805 {
2806         int cpu = (long)v;
2807
2808         val &= ~CPU_TASKS_FROZEN;
2809         switch (val) {
2810         case CPU_DYING:
2811                 printk(KERN_INFO "kvm: disabling virtualization on CPU%d\n",
2812                        cpu);
2813                 hardware_disable();
2814                 break;
2815         case CPU_STARTING:
2816                 printk(KERN_INFO "kvm: enabling virtualization on CPU%d\n",
2817                        cpu);
2818                 hardware_enable();
2819                 break;
2820         }
2821         return NOTIFY_OK;
2822 }
2823
2824 static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
2825                       void *v)
2826 {
2827         /*
2828          * Some (well, at least mine) BIOSes hang on reboot if
2829          * in vmx root mode.
2830          *
2831          * And Intel TXT required VMX off for all cpu when system shutdown.
2832          */
2833         printk(KERN_INFO "kvm: exiting hardware virtualization\n");
2834         kvm_rebooting = true;
2835         on_each_cpu(hardware_disable_nolock, NULL, 1);
2836         return NOTIFY_OK;
2837 }
2838
2839 static struct notifier_block kvm_reboot_notifier = {
2840         .notifier_call = kvm_reboot,
2841         .priority = 0,
2842 };
2843
2844 static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
2845 {
2846         int i;
2847
2848         for (i = 0; i < bus->dev_count; i++) {
2849                 struct kvm_io_device *pos = bus->range[i].dev;
2850
2851                 kvm_iodevice_destructor(pos);
2852         }
2853         kfree(bus);
2854 }
2855
2856 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1,
2857                                  const struct kvm_io_range *r2)
2858 {
2859         if (r1->addr < r2->addr)
2860                 return -1;
2861         if (r1->addr + r1->len > r2->addr + r2->len)
2862                 return 1;
2863         return 0;
2864 }
2865
2866 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
2867 {
2868         return kvm_io_bus_cmp(p1, p2);
2869 }
2870
2871 static int kvm_io_bus_insert_dev(struct kvm_io_bus *bus, struct kvm_io_device *dev,
2872                           gpa_t addr, int len)
2873 {
2874         bus->range[bus->dev_count++] = (struct kvm_io_range) {
2875                 .addr = addr,
2876                 .len = len,
2877                 .dev = dev,
2878         };
2879
2880         sort(bus->range, bus->dev_count, sizeof(struct kvm_io_range),
2881                 kvm_io_bus_sort_cmp, NULL);
2882
2883         return 0;
2884 }
2885
2886 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
2887                              gpa_t addr, int len)
2888 {
2889         struct kvm_io_range *range, key;
2890         int off;
2891
2892         key = (struct kvm_io_range) {
2893                 .addr = addr,
2894                 .len = len,
2895         };
2896
2897         range = bsearch(&key, bus->range, bus->dev_count,
2898                         sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
2899         if (range == NULL)
2900                 return -ENOENT;
2901
2902         off = range - bus->range;
2903
2904         while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0)
2905                 off--;
2906
2907         return off;
2908 }
2909
2910 static int __kvm_io_bus_write(struct kvm_io_bus *bus,
2911                               struct kvm_io_range *range, const void *val)
2912 {
2913         int idx;
2914
2915         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
2916         if (idx < 0)
2917                 return -EOPNOTSUPP;
2918
2919         while (idx < bus->dev_count &&
2920                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
2921                 if (!kvm_iodevice_write(bus->range[idx].dev, range->addr,
2922                                         range->len, val))
2923                         return idx;
2924                 idx++;
2925         }
2926
2927         return -EOPNOTSUPP;
2928 }
2929
2930 /* kvm_io_bus_write - called under kvm->slots_lock */
2931 int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2932                      int len, const void *val)
2933 {
2934         struct kvm_io_bus *bus;
2935         struct kvm_io_range range;
2936         int r;
2937
2938         range = (struct kvm_io_range) {
2939                 .addr = addr,
2940                 .len = len,
2941         };
2942
2943         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
2944         r = __kvm_io_bus_write(bus, &range, val);
2945         return r < 0 ? r : 0;
2946 }
2947
2948 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */
2949 int kvm_io_bus_write_cookie(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2950                             int len, const void *val, long cookie)
2951 {
2952         struct kvm_io_bus *bus;
2953         struct kvm_io_range range;
2954
2955         range = (struct kvm_io_range) {
2956                 .addr = addr,
2957                 .len = len,
2958         };
2959
2960         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
2961
2962         /* First try the device referenced by cookie. */
2963         if ((cookie >= 0) && (cookie < bus->dev_count) &&
2964             (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0))
2965                 if (!kvm_iodevice_write(bus->range[cookie].dev, addr, len,
2966                                         val))
2967                         return cookie;
2968
2969         /*
2970          * cookie contained garbage; fall back to search and return the
2971          * correct cookie value.
2972          */
2973         return __kvm_io_bus_write(bus, &range, val);
2974 }
2975
2976 static int __kvm_io_bus_read(struct kvm_io_bus *bus, struct kvm_io_range *range,
2977                              void *val)
2978 {
2979         int idx;
2980
2981         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
2982         if (idx < 0)
2983                 return -EOPNOTSUPP;
2984
2985         while (idx < bus->dev_count &&
2986                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
2987                 if (!kvm_iodevice_read(bus->range[idx].dev, range->addr,
2988                                        range->len, val))
2989                         return idx;
2990                 idx++;
2991         }
2992
2993         return -EOPNOTSUPP;
2994 }
2995 EXPORT_SYMBOL_GPL(kvm_io_bus_write);
2996
2997 /* kvm_io_bus_read - called under kvm->slots_lock */
2998 int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
2999                     int len, void *val)
3000 {
3001         struct kvm_io_bus *bus;
3002         struct kvm_io_range range;
3003         int r;
3004
3005         range = (struct kvm_io_range) {
3006                 .addr = addr,
3007                 .len = len,
3008         };
3009
3010         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
3011         r = __kvm_io_bus_read(bus, &range, val);
3012         return r < 0 ? r : 0;
3013 }
3014
3015
3016 /* Caller must hold slots_lock. */
3017 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
3018                             int len, struct kvm_io_device *dev)
3019 {
3020         struct kvm_io_bus *new_bus, *bus;
3021
3022         bus = kvm->buses[bus_idx];
3023         /* exclude ioeventfd which is limited by maximum fd */
3024         if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
3025                 return -ENOSPC;
3026
3027         new_bus = kzalloc(sizeof(*bus) + ((bus->dev_count + 1) *
3028                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3029         if (!new_bus)
3030                 return -ENOMEM;
3031         memcpy(new_bus, bus, sizeof(*bus) + (bus->dev_count *
3032                sizeof(struct kvm_io_range)));
3033         kvm_io_bus_insert_dev(new_bus, dev, addr, len);
3034         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3035         synchronize_srcu_expedited(&kvm->srcu);
3036         kfree(bus);
3037
3038         return 0;
3039 }
3040
3041 /* Caller must hold slots_lock. */
3042 int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
3043                               struct kvm_io_device *dev)
3044 {
3045         int i, r;
3046         struct kvm_io_bus *new_bus, *bus;
3047
3048         bus = kvm->buses[bus_idx];
3049         r = -ENOENT;
3050         for (i = 0; i < bus->dev_count; i++)
3051                 if (bus->range[i].dev == dev) {
3052                         r = 0;
3053                         break;
3054                 }
3055
3056         if (r)
3057                 return r;
3058
3059         new_bus = kzalloc(sizeof(*bus) + ((bus->dev_count - 1) *
3060                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3061         if (!new_bus)
3062                 return -ENOMEM;
3063
3064         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
3065         new_bus->dev_count--;
3066         memcpy(new_bus->range + i, bus->range + i + 1,
3067                (new_bus->dev_count - i) * sizeof(struct kvm_io_range));
3068
3069         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3070         synchronize_srcu_expedited(&kvm->srcu);
3071         kfree(bus);
3072         return r;
3073 }
3074
3075 static struct notifier_block kvm_cpu_notifier = {
3076         .notifier_call = kvm_cpu_hotplug,
3077 };
3078
3079 static int vm_stat_get(void *_offset, u64 *val)
3080 {
3081         unsigned offset = (long)_offset;
3082         struct kvm *kvm;
3083
3084         *val = 0;
3085         spin_lock(&kvm_lock);
3086         list_for_each_entry(kvm, &vm_list, vm_list)
3087                 *val += *(u32 *)((void *)kvm + offset);
3088         spin_unlock(&kvm_lock);
3089         return 0;
3090 }
3091
3092 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, NULL, "%llu\n");
3093
3094 static int vcpu_stat_get(void *_offset, u64 *val)
3095 {
3096         unsigned offset = (long)_offset;
3097         struct kvm *kvm;
3098         struct kvm_vcpu *vcpu;
3099         int i;
3100
3101         *val = 0;
3102         spin_lock(&kvm_lock);
3103         list_for_each_entry(kvm, &vm_list, vm_list)
3104                 kvm_for_each_vcpu(i, vcpu, kvm)
3105                         *val += *(u32 *)((void *)vcpu + offset);
3106
3107         spin_unlock(&kvm_lock);
3108         return 0;
3109 }
3110
3111 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n");
3112
3113 static const struct file_operations *stat_fops[] = {
3114         [KVM_STAT_VCPU] = &vcpu_stat_fops,
3115         [KVM_STAT_VM]   = &vm_stat_fops,
3116 };
3117
3118 static int kvm_init_debug(void)
3119 {
3120         int r = -EEXIST;
3121         struct kvm_stats_debugfs_item *p;
3122
3123         kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
3124         if (kvm_debugfs_dir == NULL)
3125                 goto out;
3126
3127         for (p = debugfs_entries; p->name; ++p) {
3128                 p->dentry = debugfs_create_file(p->name, 0444, kvm_debugfs_dir,
3129                                                 (void *)(long)p->offset,
3130                                                 stat_fops[p->kind]);
3131                 if (p->dentry == NULL)
3132                         goto out_dir;
3133         }
3134
3135         return 0;
3136
3137 out_dir:
3138         debugfs_remove_recursive(kvm_debugfs_dir);
3139 out:
3140         return r;
3141 }
3142
3143 static void kvm_exit_debug(void)
3144 {
3145         struct kvm_stats_debugfs_item *p;
3146
3147         for (p = debugfs_entries; p->name; ++p)
3148                 debugfs_remove(p->dentry);
3149         debugfs_remove(kvm_debugfs_dir);
3150 }
3151
3152 static int kvm_suspend(void)
3153 {
3154         if (kvm_usage_count)
3155                 hardware_disable_nolock(NULL);
3156         return 0;
3157 }
3158
3159 static void kvm_resume(void)
3160 {
3161         if (kvm_usage_count) {
3162                 WARN_ON(raw_spin_is_locked(&kvm_count_lock));
3163                 hardware_enable_nolock(NULL);
3164         }
3165 }
3166
3167 static struct syscore_ops kvm_syscore_ops = {
3168         .suspend = kvm_suspend,
3169         .resume = kvm_resume,
3170 };
3171
3172 static inline
3173 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
3174 {
3175         return container_of(pn, struct kvm_vcpu, preempt_notifier);
3176 }
3177
3178 static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
3179 {
3180         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3181         if (vcpu->preempted)
3182                 vcpu->preempted = false;
3183
3184         kvm_arch_sched_in(vcpu, cpu);
3185
3186         kvm_arch_vcpu_load(vcpu, cpu);
3187 }
3188
3189 static void kvm_sched_out(struct preempt_notifier *pn,
3190                           struct task_struct *next)
3191 {
3192         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3193
3194         if (current->state == TASK_RUNNING)
3195                 vcpu->preempted = true;
3196         kvm_arch_vcpu_put(vcpu);
3197 }
3198
3199 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
3200                   struct module *module)
3201 {
3202         int r;
3203         int cpu;
3204
3205         r = kvm_arch_init(opaque);
3206         if (r)
3207                 goto out_fail;
3208
3209         /*
3210          * kvm_arch_init makes sure there's at most one caller
3211          * for architectures that support multiple implementations,
3212          * like intel and amd on x86.
3213          * kvm_arch_init must be called before kvm_irqfd_init to avoid creating
3214          * conflicts in case kvm is already setup for another implementation.
3215          */
3216         r = kvm_irqfd_init();
3217         if (r)
3218                 goto out_irqfd;
3219
3220         if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
3221                 r = -ENOMEM;
3222                 goto out_free_0;
3223         }
3224
3225         r = kvm_arch_hardware_setup();
3226         if (r < 0)
3227                 goto out_free_0a;
3228
3229         for_each_online_cpu(cpu) {
3230                 smp_call_function_single(cpu,
3231                                 kvm_arch_check_processor_compat,
3232                                 &r, 1);
3233                 if (r < 0)
3234                         goto out_free_1;
3235         }
3236
3237         r = register_cpu_notifier(&kvm_cpu_notifier);
3238         if (r)
3239                 goto out_free_2;
3240         register_reboot_notifier(&kvm_reboot_notifier);
3241
3242         /* A kmem cache lets us meet the alignment requirements of fx_save. */
3243         if (!vcpu_align)
3244                 vcpu_align = __alignof__(struct kvm_vcpu);
3245         kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size, vcpu_align,
3246                                            0, NULL);
3247         if (!kvm_vcpu_cache) {
3248                 r = -ENOMEM;
3249                 goto out_free_3;
3250         }
3251
3252         r = kvm_async_pf_init();
3253         if (r)
3254                 goto out_free;
3255
3256         kvm_chardev_ops.owner = module;
3257         kvm_vm_fops.owner = module;
3258         kvm_vcpu_fops.owner = module;
3259
3260         r = misc_register(&kvm_dev);
3261         if (r) {
3262                 printk(KERN_ERR "kvm: misc device register failed\n");
3263                 goto out_unreg;
3264         }
3265
3266         register_syscore_ops(&kvm_syscore_ops);
3267
3268         kvm_preempt_ops.sched_in = kvm_sched_in;
3269         kvm_preempt_ops.sched_out = kvm_sched_out;
3270
3271         r = kvm_init_debug();
3272         if (r) {
3273                 printk(KERN_ERR "kvm: create debugfs files failed\n");
3274                 goto out_undebugfs;
3275         }
3276
3277         r = kvm_vfio_ops_init();
3278         WARN_ON(r);
3279
3280         return 0;
3281
3282 out_undebugfs:
3283         unregister_syscore_ops(&kvm_syscore_ops);
3284         misc_deregister(&kvm_dev);
3285 out_unreg:
3286         kvm_async_pf_deinit();
3287 out_free:
3288         kmem_cache_destroy(kvm_vcpu_cache);
3289 out_free_3:
3290         unregister_reboot_notifier(&kvm_reboot_notifier);
3291         unregister_cpu_notifier(&kvm_cpu_notifier);
3292 out_free_2:
3293 out_free_1:
3294         kvm_arch_hardware_unsetup();
3295 out_free_0a:
3296         free_cpumask_var(cpus_hardware_enabled);
3297 out_free_0:
3298         kvm_irqfd_exit();
3299 out_irqfd:
3300         kvm_arch_exit();
3301 out_fail:
3302         return r;
3303 }
3304 EXPORT_SYMBOL_GPL(kvm_init);
3305
3306 void kvm_exit(void)
3307 {
3308         kvm_exit_debug();
3309         misc_deregister(&kvm_dev);
3310         kmem_cache_destroy(kvm_vcpu_cache);
3311         kvm_async_pf_deinit();
3312         unregister_syscore_ops(&kvm_syscore_ops);
3313         unregister_reboot_notifier(&kvm_reboot_notifier);
3314         unregister_cpu_notifier(&kvm_cpu_notifier);
3315         on_each_cpu(hardware_disable_nolock, NULL, 1);
3316         kvm_arch_hardware_unsetup();
3317         kvm_arch_exit();
3318         kvm_irqfd_exit();
3319         free_cpumask_var(cpus_hardware_enabled);
3320 }
3321 EXPORT_SYMBOL_GPL(kvm_exit);