virt/kvm/kvm_main.c

   1 /*
   2  * Kernel-based Virtual Machine driver for Linux
   3  *
   4  * This module enables machines with Intel VT-x extensions to run virtual
   5  * machines without emulation or binary translation.
   6  *
   7  * Copyright (C) 2006 Qumranet, Inc.
   8  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   9  *
  10  * Authors:
  11  *   Avi Kivity   <avi@qumranet.com>
  12  *   Yaniv Kamay  <yaniv@qumranet.com>
  13  *
  14  * This work is licensed under the terms of the GNU GPL, version 2.  See
  15  * the COPYING file in the top-level directory.
  16  *
  17  */
  18
  19 #include <kvm/iodev.h>
  20
  21 #include <linux/kvm_host.h>
  22 #include <linux/kvm.h>
  23 #include <linux/module.h>
  24 #include <linux/errno.h>
  25 #include <linux/percpu.h>
  26 #include <linux/mm.h>
  27 #include <linux/miscdevice.h>
  28 #include <linux/vmalloc.h>
  29 #include <linux/reboot.h>
  30 #include <linux/debugfs.h>
  31 #include <linux/highmem.h>
  32 #include <linux/file.h>
  33 #include <linux/syscore_ops.h>
  34 #include <linux/cpu.h>
  35 #include <linux/sched.h>
  36 #include <linux/cpumask.h>
  37 #include <linux/smp.h>
  38 #include <linux/anon_inodes.h>
  39 #include <linux/profile.h>
  40 #include <linux/kvm_para.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/mman.h>
  43 #include <linux/swap.h>
  44 #include <linux/bitops.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/compat.h>
  47 #include <linux/srcu.h>
  48 #include <linux/hugetlb.h>
  49 #include <linux/slab.h>
  50 #include <linux/sort.h>
  51 #include <linux/bsearch.h>
  52
  53 #include <asm/processor.h>
  54 #include <asm/io.h>
  55 #include <asm/ioctl.h>
  56 #include <asm/uaccess.h>
  57 #include <asm/pgtable.h>
  58
  59 #include "coalesced_mmio.h"
  60 #include "async_pf.h"
  61 #include "vfio.h"
  62
  63 #define CREATE_TRACE_POINTS
  64 #include <trace/events/kvm.h>
  65
  66 MODULE_AUTHOR("Qumranet");
  67 MODULE_LICENSE("GPL");
  68
  69 /* Architectures should define their poll value according to the halt latency */
  70 static unsigned int halt_poll_ns = KVM_HALT_POLL_NS_DEFAULT;
  71 module_param(halt_poll_ns, uint, S_IRUGO | S_IWUSR);
  72
  73 /* Default doubles per-vcpu halt_poll_ns. */
  74 static unsigned int halt_poll_ns_grow = 2;
  75 module_param(halt_poll_ns_grow, uint, S_IRUGO | S_IWUSR);
  76
  77 /* Default resets per-vcpu halt_poll_ns . */
  78 static unsigned int halt_poll_ns_shrink;
  79 module_param(halt_poll_ns_shrink, uint, S_IRUGO | S_IWUSR);
  80
  81 /*
  82  * Ordering of locks:
  83  *
  84  *      kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  85  */
  86
  87 DEFINE_SPINLOCK(kvm_lock);
  88 static DEFINE_RAW_SPINLOCK(kvm_count_lock);
  89 LIST_HEAD(vm_list);
  90
  91 static cpumask_var_t cpus_hardware_enabled;
  92 static int kvm_usage_count;
  93 static atomic_t hardware_enable_failed;
  94
  95 struct kmem_cache *kvm_vcpu_cache;
  96 EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
  97
  98 static __read_mostly struct preempt_ops kvm_preempt_ops;
  99
 100 struct dentry *kvm_debugfs_dir;
 101 EXPORT_SYMBOL_GPL(kvm_debugfs_dir);
 102
 103 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
 104                            unsigned long arg);
 105 #ifdef CONFIG_KVM_COMPAT
 106 static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl,
 107                                   unsigned long arg);
 108 #endif
 109 static int hardware_enable_all(void);
 110 static void hardware_disable_all(void);
 111
 112 static void kvm_io_bus_destroy(struct kvm_io_bus *bus);
 113
 114 static void kvm_release_pfn_dirty(kvm_pfn_t pfn);
 115 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot, gfn_t gfn);
 116
 117 __visible bool kvm_rebooting;
 118 EXPORT_SYMBOL_GPL(kvm_rebooting);
 119
 120 static bool largepages_enabled = true;
 121
 122 bool kvm_is_reserved_pfn(kvm_pfn_t pfn)
 123 {
 124         if (pfn_valid(pfn))
 125                 return PageReserved(pfn_to_page(pfn));
 126
 127         return true;
 128 }
 129
 130 /*
 131  * Switches to specified vcpu, until a matching vcpu_put()
 132  */
 133 int vcpu_load(struct kvm_vcpu *vcpu)
 134 {
 135         int cpu;
 136
 137         if (mutex_lock_killable(&vcpu->mutex))
 138                 return -EINTR;
 139         cpu = get_cpu();
 140         preempt_notifier_register(&vcpu->preempt_notifier);
 141         kvm_arch_vcpu_load(vcpu, cpu);
 142         put_cpu();
 143         return 0;
 144 }
 145
 146 void vcpu_put(struct kvm_vcpu *vcpu)
 147 {
 148         preempt_disable();
 149         kvm_arch_vcpu_put(vcpu);
 150         preempt_notifier_unregister(&vcpu->preempt_notifier);
 151         preempt_enable();
 152         mutex_unlock(&vcpu->mutex);
 153 }
 154
 155 static void ack_flush(void *_completed)
 156 {
 157 }
 158
 159 bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req)
 160 {
 161         int i, cpu, me;
 162         cpumask_var_t cpus;
 163         bool called = true;
 164         struct kvm_vcpu *vcpu;
 165
 166         zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 167
 168         me = get_cpu();
 169         kvm_for_each_vcpu(i, vcpu, kvm) {
 170                 kvm_make_request(req, vcpu);
 171                 cpu = vcpu->cpu;
 172
 173                 /* Set ->requests bit before we read ->mode */
 174                 smp_mb();
 175
 176                 if (cpus != NULL && cpu != -1 && cpu != me &&
 177                       kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE)
 178                         cpumask_set_cpu(cpu, cpus);
 179         }
 180         if (unlikely(cpus == NULL))
 181                 smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1);
 182         else if (!cpumask_empty(cpus))
 183                 smp_call_function_many(cpus, ack_flush, NULL, 1);
 184         else
 185                 called = false;
 186         put_cpu();
 187         free_cpumask_var(cpus);
 188         return called;
 189 }
 190
 191 #ifndef CONFIG_HAVE_KVM_ARCH_TLB_FLUSH_ALL
 192 void kvm_flush_remote_tlbs(struct kvm *kvm)
 193 {
 194         long dirty_count = kvm->tlbs_dirty;
 195
 196         smp_mb();
 197         if (kvm_make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 198                 ++kvm->stat.remote_tlb_flush;
 199         cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 200 }
 201 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs);
 202 #endif
 203
 204 void kvm_reload_remote_mmus(struct kvm *kvm)
 205 {
 206         kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
 207 }
 208
 209 int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
 210 {
 211         struct page *page;
 212         int r;
 213
 214         mutex_init(&vcpu->mutex);
 215         vcpu->cpu = -1;
 216         vcpu->kvm = kvm;
 217         vcpu->vcpu_id = id;
 218         vcpu->pid = NULL;
 219         init_swait_queue_head(&vcpu->wq);
 220         kvm_async_pf_vcpu_init(vcpu);
 221
 222         vcpu->pre_pcpu = -1;
 223         INIT_LIST_HEAD(&vcpu->blocked_vcpu_list);
 224
 225         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 226         if (!page) {
 227                 r = -ENOMEM;
 228                 goto fail;
 229         }
 230         vcpu->run = page_address(page);
 231
 232         kvm_vcpu_set_in_spin_loop(vcpu, false);
 233         kvm_vcpu_set_dy_eligible(vcpu, false);
 234         vcpu->preempted = false;
 235
 236         r = kvm_arch_vcpu_init(vcpu);
 237         if (r < 0)
 238                 goto fail_free_run;
 239         return 0;
 240
 241 fail_free_run:
 242         free_page((unsigned long)vcpu->run);
 243 fail:
 244         return r;
 245 }
 246 EXPORT_SYMBOL_GPL(kvm_vcpu_init);
 247
 248 void kvm_vcpu_uninit(struct kvm_vcpu *vcpu)
 249 {
 250         put_pid(vcpu->pid);
 251         kvm_arch_vcpu_uninit(vcpu);
 252         free_page((unsigned long)vcpu->run);
 253 }
 254 EXPORT_SYMBOL_GPL(kvm_vcpu_uninit);
 255
 256 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 257 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
 258 {
 259         return container_of(mn, struct kvm, mmu_notifier);
 260 }
 261
 262 static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 263                                              struct mm_struct *mm,
 264                                              unsigned long address)
 265 {
 266         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 267         int need_tlb_flush, idx;
 268
 269         /*
 270          * When ->invalidate_page runs, the linux pte has been zapped
 271          * already but the page is still allocated until
 272          * ->invalidate_page returns. So if we increase the sequence
 273          * here the kvm page fault will notice if the spte can't be
 274          * established because the page is going to be freed. If
 275          * instead the kvm page fault establishes the spte before
 276          * ->invalidate_page runs, kvm_unmap_hva will release it
 277          * before returning.
 278          *
 279          * The sequence increase only need to be seen at spin_unlock
 280          * time, and not at spin_lock time.
 281          *
 282          * Increasing the sequence after the spin_unlock would be
 283          * unsafe because the kvm page fault could then establish the
 284          * pte after kvm_unmap_hva returned, without noticing the page
 285          * is going to be freed.
 286          */
 287         idx = srcu_read_lock(&kvm->srcu);
 288         spin_lock(&kvm->mmu_lock);
 289
 290         kvm->mmu_notifier_seq++;
 291         need_tlb_flush = kvm_unmap_hva(kvm, address) | kvm->tlbs_dirty;
 292         /* we've to flush the tlb before the pages can be freed */
 293         if (need_tlb_flush)
 294                 kvm_flush_remote_tlbs(kvm);
 295
 296         spin_unlock(&kvm->mmu_lock);
 297
 298         kvm_arch_mmu_notifier_invalidate_page(kvm, address);
 299
 300         srcu_read_unlock(&kvm->srcu, idx);
 301 }
 302
 303 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
 304                                         struct mm_struct *mm,
 305                                         unsigned long address,
 306                                         pte_t pte)
 307 {
 308         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 309         int idx;
 310
 311         idx = srcu_read_lock(&kvm->srcu);
 312         spin_lock(&kvm->mmu_lock);
 313         kvm->mmu_notifier_seq++;
 314         kvm_set_spte_hva(kvm, address, pte);
 315         spin_unlock(&kvm->mmu_lock);
 316         srcu_read_unlock(&kvm->srcu, idx);
 317 }
 318
 319 static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 320                                                     struct mm_struct *mm,
 321                                                     unsigned long start,
 322                                                     unsigned long end)
 323 {
 324         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 325         int need_tlb_flush = 0, idx;
 326
 327         idx = srcu_read_lock(&kvm->srcu);
 328         spin_lock(&kvm->mmu_lock);
 329         /*
 330          * The count increase must become visible at unlock time as no
 331          * spte can be established without taking the mmu_lock and
 332          * count is also read inside the mmu_lock critical section.
 333          */
 334         kvm->mmu_notifier_count++;
 335         need_tlb_flush = kvm_unmap_hva_range(kvm, start, end);
 336         need_tlb_flush |= kvm->tlbs_dirty;
 337         /* we've to flush the tlb before the pages can be freed */
 338         if (need_tlb_flush)
 339                 kvm_flush_remote_tlbs(kvm);
 340
 341         spin_unlock(&kvm->mmu_lock);
 342         srcu_read_unlock(&kvm->srcu, idx);
 343 }
 344
 345 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
 346                                                   struct mm_struct *mm,
 347                                                   unsigned long start,
 348                                                   unsigned long end)
 349 {
 350         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 351
 352         spin_lock(&kvm->mmu_lock);
 353         /*
 354          * This sequence increase will notify the kvm page fault that
 355          * the page that is going to be mapped in the spte could have
 356          * been freed.
 357          */
 358         kvm->mmu_notifier_seq++;
 359         smp_wmb();
 360         /*
 361          * The above sequence increase must be visible before the
 362          * below count decrease, which is ensured by the smp_wmb above
 363          * in conjunction with the smp_rmb in mmu_notifier_retry().
 364          */
 365         kvm->mmu_notifier_count--;
 366         spin_unlock(&kvm->mmu_lock);
 367
 368         BUG_ON(kvm->mmu_notifier_count < 0);
 369 }
 370
 371 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
 372                                               struct mm_struct *mm,
 373                                               unsigned long start,
 374                                               unsigned long end)
 375 {
 376         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 377         int young, idx;
 378
 379         idx = srcu_read_lock(&kvm->srcu);
 380         spin_lock(&kvm->mmu_lock);
 381
 382         young = kvm_age_hva(kvm, start, end);
 383         if (young)
 384                 kvm_flush_remote_tlbs(kvm);
 385
 386         spin_unlock(&kvm->mmu_lock);
 387         srcu_read_unlock(&kvm->srcu, idx);
 388
 389         return young;
 390 }
 391
 392 static int kvm_mmu_notifier_clear_young(struct mmu_notifier *mn,
 393                                         struct mm_struct *mm,
 394                                         unsigned long start,
 395                                         unsigned long end)
 396 {
 397         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 398         int young, idx;
 399
 400         idx = srcu_read_lock(&kvm->srcu);
 401         spin_lock(&kvm->mmu_lock);
 402         /*
 403          * Even though we do not flush TLB, this will still adversely
 404          * affect performance on pre-Haswell Intel EPT, where there is
 405          * no EPT Access Bit to clear so that we have to tear down EPT
 406          * tables instead. If we find this unacceptable, we can always
 407          * add a parameter to kvm_age_hva so that it effectively doesn't
 408          * do anything on clear_young.
 409          *
 410          * Also note that currently we never issue secondary TLB flushes
 411          * from clear_young, leaving this job up to the regular system
 412          * cadence. If we find this inaccurate, we might come up with a
 413          * more sophisticated heuristic later.
 414          */
 415         young = kvm_age_hva(kvm, start, end);
 416         spin_unlock(&kvm->mmu_lock);
 417         srcu_read_unlock(&kvm->srcu, idx);
 418
 419         return young;
 420 }
 421
 422 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn,
 423                                        struct mm_struct *mm,
 424                                        unsigned long address)
 425 {
 426         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 427         int young, idx;
 428
 429         idx = srcu_read_lock(&kvm->srcu);
 430         spin_lock(&kvm->mmu_lock);
 431         young = kvm_test_age_hva(kvm, address);
 432         spin_unlock(&kvm->mmu_lock);
 433         srcu_read_unlock(&kvm->srcu, idx);
 434
 435         return young;
 436 }
 437
 438 static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
 439                                      struct mm_struct *mm)
 440 {
 441         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 442         int idx;
 443
 444         idx = srcu_read_lock(&kvm->srcu);
 445         kvm_arch_flush_shadow_all(kvm);
 446         srcu_read_unlock(&kvm->srcu, idx);
 447 }
 448
 449 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
 450         .invalidate_page        = kvm_mmu_notifier_invalidate_page,
 451         .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
 452         .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
 453         .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
 454         .clear_young            = kvm_mmu_notifier_clear_young,
 455         .test_young             = kvm_mmu_notifier_test_young,
 456         .change_pte             = kvm_mmu_notifier_change_pte,
 457         .release                = kvm_mmu_notifier_release,
 458 };
 459
 460 static int kvm_init_mmu_notifier(struct kvm *kvm)
 461 {
 462         kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
 463         return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
 464 }
 465
 466 #else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
 467
 468 static int kvm_init_mmu_notifier(struct kvm *kvm)
 469 {
 470         return 0;
 471 }
 472
 473 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 474
 475 static struct kvm_memslots *kvm_alloc_memslots(void)
 476 {
 477         int i;
 478         struct kvm_memslots *slots;
 479
 480         slots = kvm_kvzalloc(sizeof(struct kvm_memslots));
 481         if (!slots)
 482                 return NULL;
 483
 484         /*
 485          * Init kvm generation close to the maximum to easily test the
 486          * code of handling generation number wrap-around.
 487          */
 488         slots->generation = -150;
 489         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 490                 slots->id_to_index[i] = slots->memslots[i].id = i;
 491
 492         return slots;
 493 }
 494
 495 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
 496 {
 497         if (!memslot->dirty_bitmap)
 498                 return;
 499
 500         kvfree(memslot->dirty_bitmap);
 501         memslot->dirty_bitmap = NULL;
 502 }
 503
 504 /*
 505  * Free any memory in @free but not in @dont.
 506  */
 507 static void kvm_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
 508                               struct kvm_memory_slot *dont)
 509 {
 510         if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
 511                 kvm_destroy_dirty_bitmap(free);
 512
 513         kvm_arch_free_memslot(kvm, free, dont);
 514
 515         free->npages = 0;
 516 }
 517
 518 static void kvm_free_memslots(struct kvm *kvm, struct kvm_memslots *slots)
 519 {
 520         struct kvm_memory_slot *memslot;
 521
 522         if (!slots)
 523                 return;
 524
 525         kvm_for_each_memslot(memslot, slots)
 526                 kvm_free_memslot(kvm, memslot, NULL);
 527
 528         kvfree(slots);
 529 }
 530
 531 static struct kvm *kvm_create_vm(unsigned long type)
 532 {
 533         int r, i;
 534         struct kvm *kvm = kvm_arch_alloc_vm();
 535
 536         if (!kvm)
 537                 return ERR_PTR(-ENOMEM);
 538
 539         r = kvm_arch_init_vm(kvm, type);
 540         if (r)
 541                 goto out_err_no_disable;
 542
 543         r = hardware_enable_all();
 544         if (r)
 545                 goto out_err_no_disable;
 546
 547 #ifdef CONFIG_HAVE_KVM_IRQFD
 548         INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 549 #endif
 550
 551         BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX);
 552
 553         r = -ENOMEM;
 554         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
 555                 kvm->memslots[i] = kvm_alloc_memslots();
 556                 if (!kvm->memslots[i])
 557                         goto out_err_no_srcu;
 558         }
 559
 560         if (init_srcu_struct(&kvm->srcu))
 561                 goto out_err_no_srcu;
 562         if (init_srcu_struct(&kvm->irq_srcu))
 563                 goto out_err_no_irq_srcu;
 564         for (i = 0; i < KVM_NR_BUSES; i++) {
 565                 kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
 566                                         GFP_KERNEL);
 567                 if (!kvm->buses[i])
 568                         goto out_err;
 569         }
 570
 571         spin_lock_init(&kvm->mmu_lock);
 572         kvm->mm = current->mm;
 573         atomic_inc(&kvm->mm->mm_count);
 574         kvm_eventfd_init(kvm);
 575         mutex_init(&kvm->lock);
 576         mutex_init(&kvm->irq_lock);
 577         mutex_init(&kvm->slots_lock);
 578         atomic_set(&kvm->users_count, 1);
 579         INIT_LIST_HEAD(&kvm->devices);
 580
 581         r = kvm_init_mmu_notifier(kvm);
 582         if (r)
 583                 goto out_err;
 584
 585         spin_lock(&kvm_lock);
 586         list_add(&kvm->vm_list, &vm_list);
 587         spin_unlock(&kvm_lock);
 588
 589         preempt_notifier_inc();
 590
 591         return kvm;
 592
 593 out_err:
 594         cleanup_srcu_struct(&kvm->irq_srcu);
 595 out_err_no_irq_srcu:
 596         cleanup_srcu_struct(&kvm->srcu);
 597 out_err_no_srcu:
 598         hardware_disable_all();
 599 out_err_no_disable:
 600         for (i = 0; i < KVM_NR_BUSES; i++)
 601                 kfree(kvm->buses[i]);
 602         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 603                 kvm_free_memslots(kvm, kvm->memslots[i]);
 604         kvm_arch_free_vm(kvm);
 605         return ERR_PTR(r);
 606 }
 607
 608 /*
 609  * Avoid using vmalloc for a small buffer.
 610  * Should not be used when the size is statically known.
 611  */
 612 void *kvm_kvzalloc(unsigned long size)
 613 {
 614         if (size > PAGE_SIZE)
 615                 return vzalloc(size);
 616         else
 617                 return kzalloc(size, GFP_KERNEL);
 618 }
 619
 620 static void kvm_destroy_devices(struct kvm *kvm)
 621 {
 622         struct kvm_device *dev, *tmp;
 623
 624         list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) {
 625                 list_del(&dev->vm_node);
 626                 dev->ops->destroy(dev);
 627         }
 628 }
 629
 630 static void kvm_destroy_vm(struct kvm *kvm)
 631 {
 632         int i;
 633         struct mm_struct *mm = kvm->mm;
 634
 635         kvm_arch_sync_events(kvm);
 636         spin_lock(&kvm_lock);
 637         list_del(&kvm->vm_list);
 638         spin_unlock(&kvm_lock);
 639         kvm_free_irq_routing(kvm);
 640         for (i = 0; i < KVM_NR_BUSES; i++)
 641                 kvm_io_bus_destroy(kvm->buses[i]);
 642         kvm_coalesced_mmio_free(kvm);
 643 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 644         mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 645 #else
 646         kvm_arch_flush_shadow_all(kvm);
 647 #endif
 648         kvm_arch_destroy_vm(kvm);
 649         kvm_destroy_devices(kvm);
 650         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 651                 kvm_free_memslots(kvm, kvm->memslots[i]);
 652         cleanup_srcu_struct(&kvm->irq_srcu);
 653         cleanup_srcu_struct(&kvm->srcu);
 654         kvm_arch_free_vm(kvm);
 655         preempt_notifier_dec();
 656         hardware_disable_all();
 657         mmdrop(mm);
 658 }
 659
 660 void kvm_get_kvm(struct kvm *kvm)
 661 {
 662         atomic_inc(&kvm->users_count);
 663 }
 664 EXPORT_SYMBOL_GPL(kvm_get_kvm);
 665
 666 void kvm_put_kvm(struct kvm *kvm)
 667 {
 668         if (atomic_dec_and_test(&kvm->users_count))
 669                 kvm_destroy_vm(kvm);
 670 }
 671 EXPORT_SYMBOL_GPL(kvm_put_kvm);
 672
 673
 674 static int kvm_vm_release(struct inode *inode, struct file *filp)
 675 {
 676         struct kvm *kvm = filp->private_data;
 677
 678         kvm_irqfd_release(kvm);
 679
 680         kvm_put_kvm(kvm);
 681         return 0;
 682 }
 683
 684 /*
 685  * Allocation size is twice as large as the actual dirty bitmap size.
 686  * See x86's kvm_vm_ioctl_get_dirty_log() why this is needed.
 687  */
 688 static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
 689 {
 690         unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
 691
 692         memslot->dirty_bitmap = kvm_kvzalloc(dirty_bytes);
 693         if (!memslot->dirty_bitmap)
 694                 return -ENOMEM;
 695
 696         return 0;
 697 }
 698
 699 /*
 700  * Insert memslot and re-sort memslots based on their GFN,
 701  * so binary search could be used to lookup GFN.
 702  * Sorting algorithm takes advantage of having initially
 703  * sorted array and known changed memslot position.
 704  */
 705 static void update_memslots(struct kvm_memslots *slots,
 706                             struct kvm_memory_slot *new)
 707 {
 708         int id = new->id;
 709         int i = slots->id_to_index[id];
 710         struct kvm_memory_slot *mslots = slots->memslots;
 711
 712         WARN_ON(mslots[i].id != id);
 713         if (!new->npages) {
 714                 WARN_ON(!mslots[i].npages);
 715                 if (mslots[i].npages)
 716                         slots->used_slots--;
 717         } else {
 718                 if (!mslots[i].npages)
 719                         slots->used_slots++;
 720         }
 721
 722         while (i < KVM_MEM_SLOTS_NUM - 1 &&
 723                new->base_gfn <= mslots[i + 1].base_gfn) {
 724                 if (!mslots[i + 1].npages)
 725                         break;
 726                 mslots[i] = mslots[i + 1];
 727                 slots->id_to_index[mslots[i].id] = i;
 728                 i++;
 729         }
 730
 731         /*
 732          * The ">=" is needed when creating a slot with base_gfn == 0,
 733          * so that it moves before all those with base_gfn == npages == 0.
 734          *
 735          * On the other hand, if new->npages is zero, the above loop has
 736          * already left i pointing to the beginning of the empty part of
 737          * mslots, and the ">=" would move the hole backwards in this
 738          * case---which is wrong.  So skip the loop when deleting a slot.
 739          */
 740         if (new->npages) {
 741                 while (i > 0 &&
 742                        new->base_gfn >= mslots[i - 1].base_gfn) {
 743                         mslots[i] = mslots[i - 1];
 744                         slots->id_to_index[mslots[i].id] = i;
 745                         i--;
 746                 }
 747         } else
 748                 WARN_ON_ONCE(i != slots->used_slots);
 749
 750         mslots[i] = *new;
 751         slots->id_to_index[mslots[i].id] = i;
 752 }
 753
 754 static int check_memory_region_flags(const struct kvm_userspace_memory_region *mem)
 755 {
 756         u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES;
 757
 758 #ifdef __KVM_HAVE_READONLY_MEM
 759         valid_flags |= KVM_MEM_READONLY;
 760 #endif
 761
 762         if (mem->flags & ~valid_flags)
 763                 return -EINVAL;
 764
 765         return 0;
 766 }
 767
 768 static struct kvm_memslots *install_new_memslots(struct kvm *kvm,
 769                 int as_id, struct kvm_memslots *slots)
 770 {
 771         struct kvm_memslots *old_memslots = __kvm_memslots(kvm, as_id);
 772
 773         /*
 774          * Set the low bit in the generation, which disables SPTE caching
 775          * until the end of synchronize_srcu_expedited.
 776          */
 777         WARN_ON(old_memslots->generation & 1);
 778         slots->generation = old_memslots->generation + 1;
 779
 780         rcu_assign_pointer(kvm->memslots[as_id], slots);
 781         synchronize_srcu_expedited(&kvm->srcu);
 782
 783         /*
 784          * Increment the new memslot generation a second time. This prevents
 785          * vm exits that race with memslot updates from caching a memslot
 786          * generation that will (potentially) be valid forever.
 787          */
 788         slots->generation++;
 789
 790         kvm_arch_memslots_updated(kvm, slots);
 791
 792         return old_memslots;
 793 }
 794
 795 /*
 796  * Allocate some memory and give it an address in the guest physical address
 797  * space.
 798  *
 799  * Discontiguous memory is allowed, mostly for framebuffers.
 800  *
 801  * Must be called holding kvm->slots_lock for write.
 802  */
 803 int __kvm_set_memory_region(struct kvm *kvm,
 804                             const struct kvm_userspace_memory_region *mem)
 805 {
 806         int r;
 807         gfn_t base_gfn;
 808         unsigned long npages;
 809         struct kvm_memory_slot *slot;
 810         struct kvm_memory_slot old, new;
 811         struct kvm_memslots *slots = NULL, *old_memslots;
 812         int as_id, id;
 813         enum kvm_mr_change change;
 814
 815         r = check_memory_region_flags(mem);
 816         if (r)
 817                 goto out;
 818
 819         r = -EINVAL;
 820         as_id = mem->slot >> 16;
 821         id = (u16)mem->slot;
 822
 823         /* General sanity checks */
 824         if (mem->memory_size & (PAGE_SIZE - 1))
 825                 goto out;
 826         if (mem->guest_phys_addr & (PAGE_SIZE - 1))
 827                 goto out;
 828         /* We can read the guest memory with __xxx_user() later on. */
 829         if ((id < KVM_USER_MEM_SLOTS) &&
 830             ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
 831              !access_ok(VERIFY_WRITE,
 832                         (void __user *)(unsigned long)mem->userspace_addr,
 833                         mem->memory_size)))
 834                 goto out;
 835         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_MEM_SLOTS_NUM)
 836                 goto out;
 837         if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
 838                 goto out;
 839
 840         slot = id_to_memslot(__kvm_memslots(kvm, as_id), id);
 841         base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
 842         npages = mem->memory_size >> PAGE_SHIFT;
 843
 844         if (npages > KVM_MEM_MAX_NR_PAGES)
 845                 goto out;
 846
 847         new = old = *slot;
 848
 849         new.id = id;
 850         new.base_gfn = base_gfn;
 851         new.npages = npages;
 852         new.flags = mem->flags;
 853
 854         if (npages) {
 855                 if (!old.npages)
 856                         change = KVM_MR_CREATE;
 857                 else { /* Modify an existing slot. */
 858                         if ((mem->userspace_addr != old.userspace_addr) ||
 859                             (npages != old.npages) ||
 860                             ((new.flags ^ old.flags) & KVM_MEM_READONLY))
 861                                 goto out;
 862
 863                         if (base_gfn != old.base_gfn)
 864                                 change = KVM_MR_MOVE;
 865                         else if (new.flags != old.flags)
 866                                 change = KVM_MR_FLAGS_ONLY;
 867                         else { /* Nothing to change. */
 868                                 r = 0;
 869                                 goto out;
 870                         }
 871                 }
 872         } else {
 873                 if (!old.npages)
 874                         goto out;
 875
 876                 change = KVM_MR_DELETE;
 877                 new.base_gfn = 0;
 878                 new.flags = 0;
 879         }
 880
 881         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 882                 /* Check for overlaps */
 883                 r = -EEXIST;
 884                 kvm_for_each_memslot(slot, __kvm_memslots(kvm, as_id)) {
 885                         if ((slot->id >= KVM_USER_MEM_SLOTS) ||
 886                             (slot->id == id))
 887                                 continue;
 888                         if (!((base_gfn + npages <= slot->base_gfn) ||
 889                               (base_gfn >= slot->base_gfn + slot->npages)))
 890                                 goto out;
 891                 }
 892         }
 893
 894         /* Free page dirty bitmap if unneeded */
 895         if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
 896                 new.dirty_bitmap = NULL;
 897
 898         r = -ENOMEM;
 899         if (change == KVM_MR_CREATE) {
 900                 new.userspace_addr = mem->userspace_addr;
 901
 902                 if (kvm_arch_create_memslot(kvm, &new, npages))
 903                         goto out_free;
 904         }
 905
 906         /* Allocate page dirty bitmap if needed */
 907         if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
 908                 if (kvm_create_dirty_bitmap(&new) < 0)
 909                         goto out_free;
 910         }
 911
 912         slots = kvm_kvzalloc(sizeof(struct kvm_memslots));
 913         if (!slots)
 914                 goto out_free;
 915         memcpy(slots, __kvm_memslots(kvm, as_id), sizeof(struct kvm_memslots));
 916
 917         if ((change == KVM_MR_DELETE) || (change == KVM_MR_MOVE)) {
 918                 slot = id_to_memslot(slots, id);
 919                 slot->flags |= KVM_MEMSLOT_INVALID;
 920
 921                 old_memslots = install_new_memslots(kvm, as_id, slots);
 922
 923                 /* slot was deleted or moved, clear iommu mapping */
 924                 kvm_iommu_unmap_pages(kvm, &old);
 925                 /* From this point no new shadow pages pointing to a deleted,
 926                  * or moved, memslot will be created.
 927                  *
 928                  * validation of sp->gfn happens in:
 929                  *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
 930                  *      - kvm_is_visible_gfn (mmu_check_roots)
 931                  */
 932                 kvm_arch_flush_shadow_memslot(kvm, slot);
 933
 934                 /*
 935                  * We can re-use the old_memslots from above, the only difference
 936                  * from the currently installed memslots is the invalid flag.  This
 937                  * will get overwritten by update_memslots anyway.
 938                  */
 939                 slots = old_memslots;
 940         }
 941
 942         r = kvm_arch_prepare_memory_region(kvm, &new, mem, change);
 943         if (r)
 944                 goto out_slots;
 945
 946         /* actual memory is freed via old in kvm_free_memslot below */
 947         if (change == KVM_MR_DELETE) {
 948                 new.dirty_bitmap = NULL;
 949                 memset(&new.arch, 0, sizeof(new.arch));
 950         }
 951
 952         update_memslots(slots, &new);
 953         old_memslots = install_new_memslots(kvm, as_id, slots);
 954
 955         kvm_arch_commit_memory_region(kvm, mem, &old, &new, change);
 956
 957         kvm_free_memslot(kvm, &old, &new);
 958         kvfree(old_memslots);
 959
 960         /*
 961          * IOMMU mapping:  New slots need to be mapped.  Old slots need to be
 962          * un-mapped and re-mapped if their base changes.  Since base change
 963          * unmapping is handled above with slot deletion, mapping alone is
 964          * needed here.  Anything else the iommu might care about for existing
 965          * slots (size changes, userspace addr changes and read-only flag
 966          * changes) is disallowed above, so any other attribute changes getting
 967          * here can be skipped.
 968          */
 969         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
 970                 r = kvm_iommu_map_pages(kvm, &new);
 971                 return r;
 972         }
 973
 974         return 0;
 975
 976 out_slots:
 977         kvfree(slots);
 978 out_free:
 979         kvm_free_memslot(kvm, &new, &old);
 980 out:
 981         return r;
 982 }
 983 EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
 984
 985 int kvm_set_memory_region(struct kvm *kvm,
 986                           const struct kvm_userspace_memory_region *mem)
 987 {
 988         int r;
 989
 990         mutex_lock(&kvm->slots_lock);
 991         r = __kvm_set_memory_region(kvm, mem);
 992         mutex_unlock(&kvm->slots_lock);
 993         return r;
 994 }
 995 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
 996
 997 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
 998                                           struct kvm_userspace_memory_region *mem)
 999 {
1000         if ((u16)mem->slot >= KVM_USER_MEM_SLOTS)
1001                 return -EINVAL;
1002
1003         return kvm_set_memory_region(kvm, mem);
1004 }
1005
1006 int kvm_get_dirty_log(struct kvm *kvm,
1007                         struct kvm_dirty_log *log, int *is_dirty)
1008 {
1009         struct kvm_memslots *slots;
1010         struct kvm_memory_slot *memslot;
1011         int r, i, as_id, id;
1012         unsigned long n;
1013         unsigned long any = 0;
1014
1015         r = -EINVAL;
1016         as_id = log->slot >> 16;
1017         id = (u16)log->slot;
1018         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1019                 goto out;
1020
1021         slots = __kvm_memslots(kvm, as_id);
1022         memslot = id_to_memslot(slots, id);
1023         r = -ENOENT;
1024         if (!memslot->dirty_bitmap)
1025                 goto out;
1026
1027         n = kvm_dirty_bitmap_bytes(memslot);
1028
1029         for (i = 0; !any && i < n/sizeof(long); ++i)
1030                 any = memslot->dirty_bitmap[i];
1031
1032         r = -EFAULT;
1033         if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
1034                 goto out;
1035
1036         if (any)
1037                 *is_dirty = 1;
1038
1039         r = 0;
1040 out:
1041         return r;
1042 }
1043 EXPORT_SYMBOL_GPL(kvm_get_dirty_log);
1044
1045 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
1046 /**
1047  * kvm_get_dirty_log_protect - get a snapshot of dirty pages, and if any pages
1048  *      are dirty write protect them for next write.
1049  * @kvm:        pointer to kvm instance
1050  * @log:        slot id and address to which we copy the log
1051  * @is_dirty:   flag set if any page is dirty
1052  *
1053  * We need to keep it in mind that VCPU threads can write to the bitmap
1054  * concurrently. So, to avoid losing track of dirty pages we keep the
1055  * following order:
1056  *
1057  *    1. Take a snapshot of the bit and clear it if needed.
1058  *    2. Write protect the corresponding page.
1059  *    3. Copy the snapshot to the userspace.
1060  *    4. Upon return caller flushes TLB's if needed.
1061  *
1062  * Between 2 and 4, the guest may write to the page using the remaining TLB
1063  * entry.  This is not a problem because the page is reported dirty using
1064  * the snapshot taken before and step 4 ensures that writes done after
1065  * exiting to userspace will be logged for the next call.
1066  *
1067  */
1068 int kvm_get_dirty_log_protect(struct kvm *kvm,
1069                         struct kvm_dirty_log *log, bool *is_dirty)
1070 {
1071         struct kvm_memslots *slots;
1072         struct kvm_memory_slot *memslot;
1073         int r, i, as_id, id;
1074         unsigned long n;
1075         unsigned long *dirty_bitmap;
1076         unsigned long *dirty_bitmap_buffer;
1077
1078         r = -EINVAL;
1079         as_id = log->slot >> 16;
1080         id = (u16)log->slot;
1081         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1082                 goto out;
1083
1084         slots = __kvm_memslots(kvm, as_id);
1085         memslot = id_to_memslot(slots, id);
1086
1087         dirty_bitmap = memslot->dirty_bitmap;
1088         r = -ENOENT;
1089         if (!dirty_bitmap)
1090                 goto out;
1091
1092         n = kvm_dirty_bitmap_bytes(memslot);
1093
1094         dirty_bitmap_buffer = dirty_bitmap + n / sizeof(long);
1095         memset(dirty_bitmap_buffer, 0, n);
1096
1097         spin_lock(&kvm->mmu_lock);
1098         *is_dirty = false;
1099         for (i = 0; i < n / sizeof(long); i++) {
1100                 unsigned long mask;
1101                 gfn_t offset;
1102
1103                 if (!dirty_bitmap[i])
1104                         continue;
1105
1106                 *is_dirty = true;
1107
1108                 mask = xchg(&dirty_bitmap[i], 0);
1109                 dirty_bitmap_buffer[i] = mask;
1110
1111                 if (mask) {
1112                         offset = i * BITS_PER_LONG;
1113                         kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot,
1114                                                                 offset, mask);
1115                 }
1116         }
1117
1118         spin_unlock(&kvm->mmu_lock);
1119
1120         r = -EFAULT;
1121         if (copy_to_user(log->dirty_bitmap, dirty_bitmap_buffer, n))
1122                 goto out;
1123
1124         r = 0;
1125 out:
1126         return r;
1127 }
1128 EXPORT_SYMBOL_GPL(kvm_get_dirty_log_protect);
1129 #endif
1130
1131 bool kvm_largepages_enabled(void)
1132 {
1133         return largepages_enabled;
1134 }
1135
1136 void kvm_disable_largepages(void)
1137 {
1138         largepages_enabled = false;
1139 }
1140 EXPORT_SYMBOL_GPL(kvm_disable_largepages);
1141
1142 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
1143 {
1144         return __gfn_to_memslot(kvm_memslots(kvm), gfn);
1145 }
1146 EXPORT_SYMBOL_GPL(gfn_to_memslot);
1147
1148 struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn)
1149 {
1150         return __gfn_to_memslot(kvm_vcpu_memslots(vcpu), gfn);
1151 }
1152
1153 bool kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
1154 {
1155         struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
1156
1157         if (!memslot || memslot->id >= KVM_USER_MEM_SLOTS ||
1158               memslot->flags & KVM_MEMSLOT_INVALID)
1159                 return false;
1160
1161         return true;
1162 }
1163 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
1164
1165 unsigned long kvm_host_page_size(struct kvm *kvm, gfn_t gfn)
1166 {
1167         struct vm_area_struct *vma;
1168         unsigned long addr, size;
1169
1170         size = PAGE_SIZE;
1171
1172         addr = gfn_to_hva(kvm, gfn);
1173         if (kvm_is_error_hva(addr))
1174                 return PAGE_SIZE;
1175
1176         down_read(&current->mm->mmap_sem);
1177         vma = find_vma(current->mm, addr);
1178         if (!vma)
1179                 goto out;
1180
1181         size = vma_kernel_pagesize(vma);
1182
1183 out:
1184         up_read(&current->mm->mmap_sem);
1185
1186         return size;
1187 }
1188
1189 static bool memslot_is_readonly(struct kvm_memory_slot *slot)
1190 {
1191         return slot->flags & KVM_MEM_READONLY;
1192 }
1193
1194 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1195                                        gfn_t *nr_pages, bool write)
1196 {
1197         if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1198                 return KVM_HVA_ERR_BAD;
1199
1200         if (memslot_is_readonly(slot) && write)
1201                 return KVM_HVA_ERR_RO_BAD;
1202
1203         if (nr_pages)
1204                 *nr_pages = slot->npages - (gfn - slot->base_gfn);
1205
1206         return __gfn_to_hva_memslot(slot, gfn);
1207 }
1208
1209 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1210                                      gfn_t *nr_pages)
1211 {
1212         return __gfn_to_hva_many(slot, gfn, nr_pages, true);
1213 }
1214
1215 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot,
1216                                         gfn_t gfn)
1217 {
1218         return gfn_to_hva_many(slot, gfn, NULL);
1219 }
1220 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot);
1221
1222 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
1223 {
1224         return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
1225 }
1226 EXPORT_SYMBOL_GPL(gfn_to_hva);
1227
1228 unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn)
1229 {
1230         return gfn_to_hva_many(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, NULL);
1231 }
1232 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_hva);
1233
1234 /*
1235  * If writable is set to false, the hva returned by this function is only
1236  * allowed to be read.
1237  */
1238 unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot,
1239                                       gfn_t gfn, bool *writable)
1240 {
1241         unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false);
1242
1243         if (!kvm_is_error_hva(hva) && writable)
1244                 *writable = !memslot_is_readonly(slot);
1245
1246         return hva;
1247 }
1248
1249 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable)
1250 {
1251         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1252
1253         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1254 }
1255
1256 unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable)
1257 {
1258         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1259
1260         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1261 }
1262
1263 static int get_user_page_nowait(unsigned long start, int write,
1264                 struct page **page)
1265 {
1266         int flags = FOLL_TOUCH | FOLL_NOWAIT | FOLL_HWPOISON | FOLL_GET;
1267
1268         if (write)
1269                 flags |= FOLL_WRITE;
1270
1271         return __get_user_pages(current, current->mm, start, 1, flags, page,
1272                         NULL, NULL);
1273 }
1274
1275 static inline int check_user_page_hwpoison(unsigned long addr)
1276 {
1277         int rc, flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_WRITE;
1278
1279         rc = __get_user_pages(current, current->mm, addr, 1,
1280                               flags, NULL, NULL, NULL);
1281         return rc == -EHWPOISON;
1282 }
1283
1284 /*
1285  * The atomic path to get the writable pfn which will be stored in @pfn,
1286  * true indicates success, otherwise false is returned.
1287  */
1288 static bool hva_to_pfn_fast(unsigned long addr, bool atomic, bool *async,
1289                             bool write_fault, bool *writable, kvm_pfn_t *pfn)
1290 {
1291         struct page *page[1];
1292         int npages;
1293
1294         if (!(async || atomic))
1295                 return false;
1296
1297         /*
1298          * Fast pin a writable pfn only if it is a write fault request
1299          * or the caller allows to map a writable pfn for a read fault
1300          * request.
1301          */
1302         if (!(write_fault || writable))
1303                 return false;
1304
1305         npages = __get_user_pages_fast(addr, 1, 1, page);
1306         if (npages == 1) {
1307                 *pfn = page_to_pfn(page[0]);
1308
1309                 if (writable)
1310                         *writable = true;
1311                 return true;
1312         }
1313
1314         return false;
1315 }
1316
1317 /*
1318  * The slow path to get the pfn of the specified host virtual address,
1319  * 1 indicates success, -errno is returned if error is detected.
1320  */
1321 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault,
1322                            bool *writable, kvm_pfn_t *pfn)
1323 {
1324         struct page *page[1];
1325         int npages = 0;
1326
1327         might_sleep();
1328
1329         if (writable)
1330                 *writable = write_fault;
1331
1332         if (async) {
1333                 down_read(&current->mm->mmap_sem);
1334                 npages = get_user_page_nowait(addr, write_fault, page);
1335                 up_read(&current->mm->mmap_sem);
1336         } else
1337                 npages = __get_user_pages_unlocked(current, current->mm, addr, 1,
1338                                                    write_fault, 0, page,
1339                                                    FOLL_TOUCH|FOLL_HWPOISON);
1340         if (npages != 1)
1341                 return npages;
1342
1343         /* map read fault as writable if possible */
1344         if (unlikely(!write_fault) && writable) {
1345                 struct page *wpage[1];
1346
1347                 npages = __get_user_pages_fast(addr, 1, 1, wpage);
1348                 if (npages == 1) {
1349                         *writable = true;
1350                         put_page(page[0]);
1351                         page[0] = wpage[0];
1352                 }
1353
1354                 npages = 1;
1355         }
1356         *pfn = page_to_pfn(page[0]);
1357         return npages;
1358 }
1359
1360 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault)
1361 {
1362         if (unlikely(!(vma->vm_flags & VM_READ)))
1363                 return false;
1364
1365         if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE))))
1366                 return false;
1367
1368         return true;
1369 }
1370
1371 /*
1372  * Pin guest page in memory and return its pfn.
1373  * @addr: host virtual address which maps memory to the guest
1374  * @atomic: whether this function can sleep
1375  * @async: whether this function need to wait IO complete if the
1376  *         host page is not in the memory
1377  * @write_fault: whether we should get a writable host page
1378  * @writable: whether it allows to map a writable host page for !@write_fault
1379  *
1380  * The function will map a writable host page for these two cases:
1381  * 1): @write_fault = true
1382  * 2): @write_fault = false && @writable, @writable will tell the caller
1383  *     whether the mapping is writable.
1384  */
1385 static kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
1386                         bool write_fault, bool *writable)
1387 {
1388         struct vm_area_struct *vma;
1389         kvm_pfn_t pfn = 0;
1390         int npages;
1391
1392         /* we can do it either atomically or asynchronously, not both */
1393         BUG_ON(atomic && async);
1394
1395         if (hva_to_pfn_fast(addr, atomic, async, write_fault, writable, &pfn))
1396                 return pfn;
1397
1398         if (atomic)
1399                 return KVM_PFN_ERR_FAULT;
1400
1401         npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn);
1402         if (npages == 1)
1403                 return pfn;
1404
1405         down_read(&current->mm->mmap_sem);
1406         if (npages == -EHWPOISON ||
1407               (!async && check_user_page_hwpoison(addr))) {
1408                 pfn = KVM_PFN_ERR_HWPOISON;
1409                 goto exit;
1410         }
1411
1412         vma = find_vma_intersection(current->mm, addr, addr + 1);
1413
1414         if (vma == NULL)
1415                 pfn = KVM_PFN_ERR_FAULT;
1416         else if ((vma->vm_flags & VM_PFNMAP)) {
1417                 pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) +
1418                         vma->vm_pgoff;
1419                 BUG_ON(!kvm_is_reserved_pfn(pfn));
1420         } else {
1421                 if (async && vma_is_valid(vma, write_fault))
1422                         *async = true;
1423                 pfn = KVM_PFN_ERR_FAULT;
1424         }
1425 exit:
1426         up_read(&current->mm->mmap_sem);
1427         return pfn;
1428 }
1429
1430 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn,
1431                                bool atomic, bool *async, bool write_fault,
1432                                bool *writable)
1433 {
1434         unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
1435
1436         if (addr == KVM_HVA_ERR_RO_BAD) {
1437                 if (writable)
1438                         *writable = false;
1439                 return KVM_PFN_ERR_RO_FAULT;
1440         }
1441
1442         if (kvm_is_error_hva(addr)) {
1443                 if (writable)
1444                         *writable = false;
1445                 return KVM_PFN_NOSLOT;
1446         }
1447
1448         /* Do not map writable pfn in the readonly memslot. */
1449         if (writable && memslot_is_readonly(slot)) {
1450                 *writable = false;
1451                 writable = NULL;
1452         }
1453
1454         return hva_to_pfn(addr, atomic, async, write_fault,
1455                           writable);
1456 }
1457 EXPORT_SYMBOL_GPL(__gfn_to_pfn_memslot);
1458
1459 kvm_pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
1460                       bool *writable)
1461 {
1462         return __gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn, false, NULL,
1463                                     write_fault, writable);
1464 }
1465 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
1466
1467 kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
1468 {
1469         return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
1470 }
1471 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot);
1472
1473 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
1474 {
1475         return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
1476 }
1477 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
1478
1479 kvm_pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
1480 {
1481         return gfn_to_pfn_memslot_atomic(gfn_to_memslot(kvm, gfn), gfn);
1482 }
1483 EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
1484
1485 kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn)
1486 {
1487         return gfn_to_pfn_memslot_atomic(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1488 }
1489 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn_atomic);
1490
1491 kvm_pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
1492 {
1493         return gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn);
1494 }
1495 EXPORT_SYMBOL_GPL(gfn_to_pfn);
1496
1497 kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn)
1498 {
1499         return gfn_to_pfn_memslot(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1500 }
1501 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn);
1502
1503 int gfn_to_page_many_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
1504                             struct page **pages, int nr_pages)
1505 {
1506         unsigned long addr;
1507         gfn_t entry;
1508
1509         addr = gfn_to_hva_many(slot, gfn, &entry);
1510         if (kvm_is_error_hva(addr))
1511                 return -1;
1512
1513         if (entry < nr_pages)
1514                 return 0;
1515
1516         return __get_user_pages_fast(addr, nr_pages, 1, pages);
1517 }
1518 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
1519
1520 static struct page *kvm_pfn_to_page(kvm_pfn_t pfn)
1521 {
1522         if (is_error_noslot_pfn(pfn))
1523                 return KVM_ERR_PTR_BAD_PAGE;
1524
1525         if (kvm_is_reserved_pfn(pfn)) {
1526                 WARN_ON(1);
1527                 return KVM_ERR_PTR_BAD_PAGE;
1528         }
1529
1530         return pfn_to_page(pfn);
1531 }
1532
1533 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
1534 {
1535         kvm_pfn_t pfn;
1536
1537         pfn = gfn_to_pfn(kvm, gfn);
1538
1539         return kvm_pfn_to_page(pfn);
1540 }
1541 EXPORT_SYMBOL_GPL(gfn_to_page);
1542
1543 struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1544 {
1545         kvm_pfn_t pfn;
1546
1547         pfn = kvm_vcpu_gfn_to_pfn(vcpu, gfn);
1548
1549         return kvm_pfn_to_page(pfn);
1550 }
1551 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_page);
1552
1553 void kvm_release_page_clean(struct page *page)
1554 {
1555         WARN_ON(is_error_page(page));
1556
1557         kvm_release_pfn_clean(page_to_pfn(page));
1558 }
1559 EXPORT_SYMBOL_GPL(kvm_release_page_clean);
1560
1561 void kvm_release_pfn_clean(kvm_pfn_t pfn)
1562 {
1563         if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn))
1564                 put_page(pfn_to_page(pfn));
1565 }
1566 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
1567
1568 void kvm_release_page_dirty(struct page *page)
1569 {
1570         WARN_ON(is_error_page(page));
1571
1572         kvm_release_pfn_dirty(page_to_pfn(page));
1573 }
1574 EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
1575
1576 static void kvm_release_pfn_dirty(kvm_pfn_t pfn)
1577 {
1578         kvm_set_pfn_dirty(pfn);
1579         kvm_release_pfn_clean(pfn);
1580 }
1581
1582 void kvm_set_pfn_dirty(kvm_pfn_t pfn)
1583 {
1584         if (!kvm_is_reserved_pfn(pfn)) {
1585                 struct page *page = pfn_to_page(pfn);
1586
1587                 if (!PageReserved(page))
1588                         SetPageDirty(page);
1589         }
1590 }
1591 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
1592
1593 void kvm_set_pfn_accessed(kvm_pfn_t pfn)
1594 {
1595         if (!kvm_is_reserved_pfn(pfn))
1596                 mark_page_accessed(pfn_to_page(pfn));
1597 }
1598 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
1599
1600 void kvm_get_pfn(kvm_pfn_t pfn)
1601 {
1602         if (!kvm_is_reserved_pfn(pfn))
1603                 get_page(pfn_to_page(pfn));
1604 }
1605 EXPORT_SYMBOL_GPL(kvm_get_pfn);
1606
1607 static int next_segment(unsigned long len, int offset)
1608 {
1609         if (len > PAGE_SIZE - offset)
1610                 return PAGE_SIZE - offset;
1611         else
1612                 return len;
1613 }
1614
1615 static int __kvm_read_guest_page(struct kvm_memory_slot *slot, gfn_t gfn,
1616                                  void *data, int offset, int len)
1617 {
1618         int r;
1619         unsigned long addr;
1620
1621         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
1622         if (kvm_is_error_hva(addr))
1623                 return -EFAULT;
1624         r = __copy_from_user(data, (void __user *)addr + offset, len);
1625         if (r)
1626                 return -EFAULT;
1627         return 0;
1628 }
1629
1630 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
1631                         int len)
1632 {
1633         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1634
1635         return __kvm_read_guest_page(slot, gfn, data, offset, len);
1636 }
1637 EXPORT_SYMBOL_GPL(kvm_read_guest_page);
1638
1639 int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data,
1640                              int offset, int len)
1641 {
1642         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1643
1644         return __kvm_read_guest_page(slot, gfn, data, offset, len);
1645 }
1646 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_page);
1647
1648 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
1649 {
1650         gfn_t gfn = gpa >> PAGE_SHIFT;
1651         int seg;
1652         int offset = offset_in_page(gpa);
1653         int ret;
1654
1655         while ((seg = next_segment(len, offset)) != 0) {
1656                 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
1657                 if (ret < 0)
1658                         return ret;
1659                 offset = 0;
1660                 len -= seg;
1661                 data += seg;
1662                 ++gfn;
1663         }
1664         return 0;
1665 }
1666 EXPORT_SYMBOL_GPL(kvm_read_guest);
1667
1668 int kvm_vcpu_read_guest(struct kvm_vcpu *vcpu, gpa_t gpa, void *data, unsigned long len)
1669 {
1670         gfn_t gfn = gpa >> PAGE_SHIFT;
1671         int seg;
1672         int offset = offset_in_page(gpa);
1673         int ret;
1674
1675         while ((seg = next_segment(len, offset)) != 0) {
1676                 ret = kvm_vcpu_read_guest_page(vcpu, gfn, data, offset, seg);
1677                 if (ret < 0)
1678                         return ret;
1679                 offset = 0;
1680                 len -= seg;
1681                 data += seg;
1682                 ++gfn;
1683         }
1684         return 0;
1685 }
1686 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest);
1687
1688 static int __kvm_read_guest_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
1689                                    void *data, int offset, unsigned long len)
1690 {
1691         int r;
1692         unsigned long addr;
1693
1694         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
1695         if (kvm_is_error_hva(addr))
1696                 return -EFAULT;
1697         pagefault_disable();
1698         r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len);
1699         pagefault_enable();
1700         if (r)
1701                 return -EFAULT;
1702         return 0;
1703 }
1704
1705 int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data,
1706                           unsigned long len)
1707 {
1708         gfn_t gfn = gpa >> PAGE_SHIFT;
1709         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1710         int offset = offset_in_page(gpa);
1711
1712         return __kvm_read_guest_atomic(slot, gfn, data, offset, len);
1713 }
1714 EXPORT_SYMBOL_GPL(kvm_read_guest_atomic);
1715
1716 int kvm_vcpu_read_guest_atomic(struct kvm_vcpu *vcpu, gpa_t gpa,
1717                                void *data, unsigned long len)
1718 {
1719         gfn_t gfn = gpa >> PAGE_SHIFT;
1720         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1721         int offset = offset_in_page(gpa);
1722
1723         return __kvm_read_guest_atomic(slot, gfn, data, offset, len);
1724 }
1725 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_atomic);
1726
1727 static int __kvm_write_guest_page(struct kvm_memory_slot *memslot, gfn_t gfn,
1728                                   const void *data, int offset, int len)
1729 {
1730         int r;
1731         unsigned long addr;
1732
1733         addr = gfn_to_hva_memslot(memslot, gfn);
1734         if (kvm_is_error_hva(addr))
1735                 return -EFAULT;
1736         r = __copy_to_user((void __user *)addr + offset, data, len);
1737         if (r)
1738                 return -EFAULT;
1739         mark_page_dirty_in_slot(memslot, gfn);
1740         return 0;
1741 }
1742
1743 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn,
1744                          const void *data, int offset, int len)
1745 {
1746         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1747
1748         return __kvm_write_guest_page(slot, gfn, data, offset, len);
1749 }
1750 EXPORT_SYMBOL_GPL(kvm_write_guest_page);
1751
1752 int kvm_vcpu_write_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn,
1753                               const void *data, int offset, int len)
1754 {
1755         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1756
1757         return __kvm_write_guest_page(slot, gfn, data, offset, len);
1758 }
1759 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest_page);
1760
1761 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
1762                     unsigned long len)
1763 {
1764         gfn_t gfn = gpa >> PAGE_SHIFT;
1765         int seg;
1766         int offset = offset_in_page(gpa);
1767         int ret;
1768
1769         while ((seg = next_segment(len, offset)) != 0) {
1770                 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
1771                 if (ret < 0)
1772                         return ret;
1773                 offset = 0;
1774                 len -= seg;
1775                 data += seg;
1776                 ++gfn;
1777         }
1778         return 0;
1779 }
1780 EXPORT_SYMBOL_GPL(kvm_write_guest);
1781
1782 int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data,
1783                          unsigned long len)
1784 {
1785         gfn_t gfn = gpa >> PAGE_SHIFT;
1786         int seg;
1787         int offset = offset_in_page(gpa);
1788         int ret;
1789
1790         while ((seg = next_segment(len, offset)) != 0) {
1791                 ret = kvm_vcpu_write_guest_page(vcpu, gfn, data, offset, seg);
1792                 if (ret < 0)
1793                         return ret;
1794                 offset = 0;
1795                 len -= seg;
1796                 data += seg;
1797                 ++gfn;
1798         }
1799         return 0;
1800 }
1801 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest);
1802
1803 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1804                               gpa_t gpa, unsigned long len)
1805 {
1806         struct kvm_memslots *slots = kvm_memslots(kvm);
1807         int offset = offset_in_page(gpa);
1808         gfn_t start_gfn = gpa >> PAGE_SHIFT;
1809         gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT;
1810         gfn_t nr_pages_needed = end_gfn - start_gfn + 1;
1811         gfn_t nr_pages_avail;
1812
1813         ghc->gpa = gpa;
1814         ghc->generation = slots->generation;
1815         ghc->len = len;
1816         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1817         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn, NULL);
1818         if (!kvm_is_error_hva(ghc->hva) && nr_pages_needed <= 1) {
1819                 ghc->hva += offset;
1820         } else {
1821                 /*
1822                  * If the requested region crosses two memslots, we still
1823                  * verify that the entire region is valid here.
1824                  */
1825                 while (start_gfn <= end_gfn) {
1826                         ghc->memslot = gfn_to_memslot(kvm, start_gfn);
1827                         ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn,
1828                                                    &nr_pages_avail);
1829                         if (kvm_is_error_hva(ghc->hva))
1830                                 return -EFAULT;
1831                         start_gfn += nr_pages_avail;
1832                 }
1833                 /* Use the slow path for cross page reads and writes. */
1834                 ghc->memslot = NULL;
1835         }
1836         return 0;
1837 }
1838 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
1839
1840 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1841                            void *data, unsigned long len)
1842 {
1843         struct kvm_memslots *slots = kvm_memslots(kvm);
1844         int r;
1845
1846         BUG_ON(len > ghc->len);
1847
1848         if (slots->generation != ghc->generation)
1849                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1850
1851         if (unlikely(!ghc->memslot))
1852                 return kvm_write_guest(kvm, ghc->gpa, data, len);
1853
1854         if (kvm_is_error_hva(ghc->hva))
1855                 return -EFAULT;
1856
1857         r = __copy_to_user((void __user *)ghc->hva, data, len);
1858         if (r)
1859                 return -EFAULT;
1860         mark_page_dirty_in_slot(ghc->memslot, ghc->gpa >> PAGE_SHIFT);
1861
1862         return 0;
1863 }
1864 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
1865
1866 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
1867                            void *data, unsigned long len)
1868 {
1869         struct kvm_memslots *slots = kvm_memslots(kvm);
1870         int r;
1871
1872         BUG_ON(len > ghc->len);
1873
1874         if (slots->generation != ghc->generation)
1875                 kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa, ghc->len);
1876
1877         if (unlikely(!ghc->memslot))
1878                 return kvm_read_guest(kvm, ghc->gpa, data, len);
1879
1880         if (kvm_is_error_hva(ghc->hva))
1881                 return -EFAULT;
1882
1883         r = __copy_from_user(data, (void __user *)ghc->hva, len);
1884         if (r)
1885                 return -EFAULT;
1886
1887         return 0;
1888 }
1889 EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
1890
1891 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
1892 {
1893         const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0)));
1894
1895         return kvm_write_guest_page(kvm, gfn, zero_page, offset, len);
1896 }
1897 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
1898
1899 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
1900 {
1901         gfn_t gfn = gpa >> PAGE_SHIFT;
1902         int seg;
1903         int offset = offset_in_page(gpa);
1904         int ret;
1905
1906         while ((seg = next_segment(len, offset)) != 0) {
1907                 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
1908                 if (ret < 0)
1909                         return ret;
1910                 offset = 0;
1911                 len -= seg;
1912                 ++gfn;
1913         }
1914         return 0;
1915 }
1916 EXPORT_SYMBOL_GPL(kvm_clear_guest);
1917
1918 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot,
1919                                     gfn_t gfn)
1920 {
1921         if (memslot && memslot->dirty_bitmap) {
1922                 unsigned long rel_gfn = gfn - memslot->base_gfn;
1923
1924                 set_bit_le(rel_gfn, memslot->dirty_bitmap);
1925         }
1926 }
1927
1928 void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
1929 {
1930         struct kvm_memory_slot *memslot;
1931
1932         memslot = gfn_to_memslot(kvm, gfn);
1933         mark_page_dirty_in_slot(memslot, gfn);
1934 }
1935 EXPORT_SYMBOL_GPL(mark_page_dirty);
1936
1937 void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
1938 {
1939         struct kvm_memory_slot *memslot;
1940
1941         memslot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1942         mark_page_dirty_in_slot(memslot, gfn);
1943 }
1944 EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
1945
1946 static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
1947 {
1948         unsigned int old, val, grow;
1949
1950         old = val = vcpu->halt_poll_ns;
1951         grow = READ_ONCE(halt_poll_ns_grow);
1952         /* 10us base */
1953         if (val == 0 && grow)
1954                 val = 10000;
1955         else
1956                 val *= grow;
1957
1958         vcpu->halt_poll_ns = val;
1959         trace_kvm_halt_poll_ns_grow(vcpu->vcpu_id, val, old);
1960 }
1961
1962 static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
1963 {
1964         unsigned int old, val, shrink;
1965
1966         old = val = vcpu->halt_poll_ns;
1967         shrink = READ_ONCE(halt_poll_ns_shrink);
1968         if (shrink == 0)
1969                 val = 0;
1970         else
1971                 val /= shrink;
1972
1973         vcpu->halt_poll_ns = val;
1974         trace_kvm_halt_poll_ns_shrink(vcpu->vcpu_id, val, old);
1975 }
1976
1977 static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
1978 {
1979         if (kvm_arch_vcpu_runnable(vcpu)) {
1980                 kvm_make_request(KVM_REQ_UNHALT, vcpu);
1981                 return -EINTR;
1982         }
1983         if (kvm_cpu_has_pending_timer(vcpu))
1984                 return -EINTR;
1985         if (signal_pending(current))
1986                 return -EINTR;
1987
1988         return 0;
1989 }
1990
1991 /*
1992  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
1993  */
1994 void kvm_vcpu_block(struct kvm_vcpu *vcpu)
1995 {
1996         ktime_t start, cur;
1997         DECLARE_SWAITQUEUE(wait);
1998         bool waited = false;
1999         u64 block_ns;
2000
2001         start = cur = ktime_get();
2002         if (vcpu->halt_poll_ns) {
2003                 ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
2004
2005                 ++vcpu->stat.halt_attempted_poll;
2006                 do {
2007                         /*
2008                          * This sets KVM_REQ_UNHALT if an interrupt
2009                          * arrives.
2010                          */
2011                         if (kvm_vcpu_check_block(vcpu) < 0) {
2012                                 ++vcpu->stat.halt_successful_poll;
2013                                 goto out;
2014                         }
2015                         cur = ktime_get();
2016                 } while (single_task_running() && ktime_before(cur, stop));
2017         }
2018
2019         kvm_arch_vcpu_blocking(vcpu);
2020
2021         for (;;) {
2022                 prepare_to_swait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
2023
2024                 if (kvm_vcpu_check_block(vcpu) < 0)
2025                         break;
2026
2027                 waited = true;
2028                 schedule();
2029         }
2030
2031         finish_swait(&vcpu->wq, &wait);
2032         cur = ktime_get();
2033
2034         kvm_arch_vcpu_unblocking(vcpu);
2035 out:
2036         block_ns = ktime_to_ns(cur) - ktime_to_ns(start);
2037
2038         if (halt_poll_ns) {
2039                 if (block_ns <= vcpu->halt_poll_ns)
2040                         ;
2041                 /* we had a long block, shrink polling */
2042                 else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
2043                         shrink_halt_poll_ns(vcpu);
2044                 /* we had a short halt and our poll time is too small */
2045                 else if (vcpu->halt_poll_ns < halt_poll_ns &&
2046                         block_ns < halt_poll_ns)
2047                         grow_halt_poll_ns(vcpu);
2048         } else
2049                 vcpu->halt_poll_ns = 0;
2050
2051         trace_kvm_vcpu_wakeup(block_ns, waited);
2052 }
2053 EXPORT_SYMBOL_GPL(kvm_vcpu_block);
2054
2055 #ifndef CONFIG_S390
2056 /*
2057  * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode.
2058  */
2059 void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
2060 {
2061         int me;
2062         int cpu = vcpu->cpu;
2063         struct swait_queue_head *wqp;
2064
2065         wqp = kvm_arch_vcpu_wq(vcpu);
2066         if (swait_active(wqp)) {
2067                 swake_up(wqp);
2068                 ++vcpu->stat.halt_wakeup;
2069         }
2070
2071         me = get_cpu();
2072         if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
2073                 if (kvm_arch_vcpu_should_kick(vcpu))
2074                         smp_send_reschedule(cpu);
2075         put_cpu();
2076 }
2077 EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
2078 #endif /* !CONFIG_S390 */
2079
2080 int kvm_vcpu_yield_to(struct kvm_vcpu *target)
2081 {
2082         struct pid *pid;
2083         struct task_struct *task = NULL;
2084         int ret = 0;
2085
2086         rcu_read_lock();
2087         pid = rcu_dereference(target->pid);
2088         if (pid)
2089                 task = get_pid_task(pid, PIDTYPE_PID);
2090         rcu_read_unlock();
2091         if (!task)
2092                 return ret;
2093         ret = yield_to(task, 1);
2094         put_task_struct(task);
2095
2096         return ret;
2097 }
2098 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to);
2099
2100 /*
2101  * Helper that checks whether a VCPU is eligible for directed yield.
2102  * Most eligible candidate to yield is decided by following heuristics:
2103  *
2104  *  (a) VCPU which has not done pl-exit or cpu relax intercepted recently
2105  *  (preempted lock holder), indicated by @in_spin_loop.
2106  *  Set at the beiginning and cleared at the end of interception/PLE handler.
2107  *
2108  *  (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get
2109  *  chance last time (mostly it has become eligible now since we have probably
2110  *  yielded to lockholder in last iteration. This is done by toggling
2111  *  @dy_eligible each time a VCPU checked for eligibility.)
2112  *
2113  *  Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding
2114  *  to preempted lock-holder could result in wrong VCPU selection and CPU
2115  *  burning. Giving priority for a potential lock-holder increases lock
2116  *  progress.
2117  *
2118  *  Since algorithm is based on heuristics, accessing another VCPU data without
2119  *  locking does not harm. It may result in trying to yield to  same VCPU, fail
2120  *  and continue with next VCPU and so on.
2121  */
2122 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu)
2123 {
2124 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT
2125         bool eligible;
2126
2127         eligible = !vcpu->spin_loop.in_spin_loop ||
2128                     vcpu->spin_loop.dy_eligible;
2129
2130         if (vcpu->spin_loop.in_spin_loop)
2131                 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible);
2132
2133         return eligible;
2134 #else
2135         return true;
2136 #endif
2137 }
2138
2139 void kvm_vcpu_on_spin(struct kvm_vcpu *me)
2140 {
2141         struct kvm *kvm = me->kvm;
2142         struct kvm_vcpu *vcpu;
2143         int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
2144         int yielded = 0;
2145         int try = 3;
2146         int pass;
2147         int i;
2148
2149         kvm_vcpu_set_in_spin_loop(me, true);
2150         /*
2151          * We boost the priority of a VCPU that is runnable but not
2152          * currently running, because it got preempted by something
2153          * else and called schedule in __vcpu_run.  Hopefully that
2154          * VCPU is holding the lock that we need and will release it.
2155          * We approximate round-robin by starting at the last boosted VCPU.
2156          */
2157         for (pass = 0; pass < 2 && !yielded && try; pass++) {
2158                 kvm_for_each_vcpu(i, vcpu, kvm) {
2159                         if (!pass && i <= last_boosted_vcpu) {
2160                                 i = last_boosted_vcpu;
2161                                 continue;
2162                         } else if (pass && i > last_boosted_vcpu)
2163                                 break;
2164                         if (!ACCESS_ONCE(vcpu->preempted))
2165                                 continue;
2166                         if (vcpu == me)
2167                                 continue;
2168                         if (swait_active(&vcpu->wq) && !kvm_arch_vcpu_runnable(vcpu))
2169                                 continue;
2170                         if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
2171                                 continue;
2172
2173                         yielded = kvm_vcpu_yield_to(vcpu);
2174                         if (yielded > 0) {
2175                                 kvm->last_boosted_vcpu = i;
2176                                 break;
2177                         } else if (yielded < 0) {
2178                                 try--;
2179                                 if (!try)
2180                                         break;
2181                         }
2182                 }
2183         }
2184         kvm_vcpu_set_in_spin_loop(me, false);
2185
2186         /* Ensure vcpu is not eligible during next spinloop */
2187         kvm_vcpu_set_dy_eligible(me, false);
2188 }
2189 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
2190
2191 static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2192 {
2193         struct kvm_vcpu *vcpu = vma->vm_file->private_data;
2194         struct page *page;
2195
2196         if (vmf->pgoff == 0)
2197                 page = virt_to_page(vcpu->run);
2198 #ifdef CONFIG_X86
2199         else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
2200                 page = virt_to_page(vcpu->arch.pio_data);
2201 #endif
2202 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2203         else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
2204                 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
2205 #endif
2206         else
2207                 return kvm_arch_vcpu_fault(vcpu, vmf);
2208         get_page(page);
2209         vmf->page = page;
2210         return 0;
2211 }
2212
2213 static const struct vm_operations_struct kvm_vcpu_vm_ops = {
2214         .fault = kvm_vcpu_fault,
2215 };
2216
2217 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
2218 {
2219         vma->vm_ops = &kvm_vcpu_vm_ops;
2220         return 0;
2221 }
2222
2223 static int kvm_vcpu_release(struct inode *inode, struct file *filp)
2224 {
2225         struct kvm_vcpu *vcpu = filp->private_data;
2226
2227         kvm_put_kvm(vcpu->kvm);
2228         return 0;
2229 }
2230
2231 static struct file_operations kvm_vcpu_fops = {
2232         .release        = kvm_vcpu_release,
2233         .unlocked_ioctl = kvm_vcpu_ioctl,
2234 #ifdef CONFIG_KVM_COMPAT
2235         .compat_ioctl   = kvm_vcpu_compat_ioctl,
2236 #endif
2237         .mmap           = kvm_vcpu_mmap,
2238         .llseek         = noop_llseek,
2239 };
2240
2241 /*
2242  * Allocates an inode for the vcpu.
2243  */
2244 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
2245 {
2246         return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC);
2247 }
2248
2249 /*
2250  * Creates some virtual cpus.  Good luck creating more than one.
2251  */
2252 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
2253 {
2254         int r;
2255         struct kvm_vcpu *vcpu;
2256
2257         if (id >= KVM_MAX_VCPUS)
2258                 return -EINVAL;
2259
2260         vcpu = kvm_arch_vcpu_create(kvm, id);
2261         if (IS_ERR(vcpu))
2262                 return PTR_ERR(vcpu);
2263
2264         preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
2265
2266         r = kvm_arch_vcpu_setup(vcpu);
2267         if (r)
2268                 goto vcpu_destroy;
2269
2270         mutex_lock(&kvm->lock);
2271         if (!kvm_vcpu_compatible(vcpu)) {
2272                 r = -EINVAL;
2273                 goto unlock_vcpu_destroy;
2274         }
2275         if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
2276                 r = -EINVAL;
2277                 goto unlock_vcpu_destroy;
2278         }
2279         if (kvm_get_vcpu_by_id(kvm, id)) {
2280                 r = -EEXIST;
2281                 goto unlock_vcpu_destroy;
2282         }
2283
2284         BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]);
2285
2286         /* Now it's all set up, let userspace reach it */
2287         kvm_get_kvm(kvm);
2288         r = create_vcpu_fd(vcpu);
2289         if (r < 0) {
2290                 kvm_put_kvm(kvm);
2291                 goto unlock_vcpu_destroy;
2292         }
2293
2294         kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu;
2295
2296         /*
2297          * Pairs with smp_rmb() in kvm_get_vcpu.  Write kvm->vcpus
2298          * before kvm->online_vcpu's incremented value.
2299          */
2300         smp_wmb();
2301         atomic_inc(&kvm->online_vcpus);
2302
2303         mutex_unlock(&kvm->lock);
2304         kvm_arch_vcpu_postcreate(vcpu);
2305         return r;
2306
2307 unlock_vcpu_destroy:
2308         mutex_unlock(&kvm->lock);
2309 vcpu_destroy:
2310         kvm_arch_vcpu_destroy(vcpu);
2311         return r;
2312 }
2313
2314 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
2315 {
2316         if (sigset) {
2317                 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
2318                 vcpu->sigset_active = 1;
2319                 vcpu->sigset = *sigset;
2320         } else
2321                 vcpu->sigset_active = 0;
2322         return 0;
2323 }
2324
2325 static long kvm_vcpu_ioctl(struct file *filp,
2326                            unsigned int ioctl, unsigned long arg)
2327 {
2328         struct kvm_vcpu *vcpu = filp->private_data;
2329         void __user *argp = (void __user *)arg;
2330         int r;
2331         struct kvm_fpu *fpu = NULL;
2332         struct kvm_sregs *kvm_sregs = NULL;
2333
2334         if (vcpu->kvm->mm != current->mm)
2335                 return -EIO;
2336
2337         if (unlikely(_IOC_TYPE(ioctl) != KVMIO))
2338                 return -EINVAL;
2339
2340 #if defined(CONFIG_S390) || defined(CONFIG_PPC) || defined(CONFIG_MIPS)
2341         /*
2342          * Special cases: vcpu ioctls that are asynchronous to vcpu execution,
2343          * so vcpu_load() would break it.
2344          */
2345         if (ioctl == KVM_S390_INTERRUPT || ioctl == KVM_S390_IRQ || ioctl == KVM_INTERRUPT)
2346                 return kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2347 #endif
2348
2349
2350         r = vcpu_load(vcpu);
2351         if (r)
2352                 return r;
2353         switch (ioctl) {
2354         case KVM_RUN:
2355                 r = -EINVAL;
2356                 if (arg)
2357                         goto out;
2358                 if (unlikely(vcpu->pid != current->pids[PIDTYPE_PID].pid)) {
2359                         /* The thread running this VCPU changed. */
2360                         struct pid *oldpid = vcpu->pid;
2361                         struct pid *newpid = get_task_pid(current, PIDTYPE_PID);
2362
2363                         rcu_assign_pointer(vcpu->pid, newpid);
2364                         if (oldpid)
2365                                 synchronize_rcu();
2366                         put_pid(oldpid);
2367                 }
2368                 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
2369                 trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
2370                 break;
2371         case KVM_GET_REGS: {
2372                 struct kvm_regs *kvm_regs;
2373
2374                 r = -ENOMEM;
2375                 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL);
2376                 if (!kvm_regs)
2377                         goto out;
2378                 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
2379                 if (r)
2380                         goto out_free1;
2381                 r = -EFAULT;
2382                 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
2383                         goto out_free1;
2384                 r = 0;
2385 out_free1:
2386                 kfree(kvm_regs);
2387                 break;
2388         }
2389         case KVM_SET_REGS: {
2390                 struct kvm_regs *kvm_regs;
2391
2392                 r = -ENOMEM;
2393                 kvm_regs = memdup_user(argp, sizeof(*kvm_regs));
2394                 if (IS_ERR(kvm_regs)) {
2395                         r = PTR_ERR(kvm_regs);
2396                         goto out;
2397                 }
2398                 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
2399                 kfree(kvm_regs);
2400                 break;
2401         }
2402         case KVM_GET_SREGS: {
2403                 kvm_sregs = kzalloc(sizeof(struct kvm_sregs), GFP_KERNEL);
2404                 r = -ENOMEM;
2405                 if (!kvm_sregs)
2406                         goto out;
2407                 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
2408                 if (r)
2409                         goto out;
2410                 r = -EFAULT;
2411                 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
2412                         goto out;
2413                 r = 0;
2414                 break;
2415         }
2416         case KVM_SET_SREGS: {
2417                 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs));
2418                 if (IS_ERR(kvm_sregs)) {
2419                         r = PTR_ERR(kvm_sregs);
2420                         kvm_sregs = NULL;
2421                         goto out;
2422                 }
2423                 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
2424                 break;
2425         }
2426         case KVM_GET_MP_STATE: {
2427                 struct kvm_mp_state mp_state;
2428
2429                 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
2430                 if (r)
2431                         goto out;
2432                 r = -EFAULT;
2433                 if (copy_to_user(argp, &mp_state, sizeof(mp_state)))
2434                         goto out;
2435                 r = 0;
2436                 break;
2437         }
2438         case KVM_SET_MP_STATE: {
2439                 struct kvm_mp_state mp_state;
2440
2441                 r = -EFAULT;
2442                 if (copy_from_user(&mp_state, argp, sizeof(mp_state)))
2443                         goto out;
2444                 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
2445                 break;
2446         }
2447         case KVM_TRANSLATE: {
2448                 struct kvm_translation tr;
2449
2450                 r = -EFAULT;
2451                 if (copy_from_user(&tr, argp, sizeof(tr)))
2452                         goto out;
2453                 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
2454                 if (r)
2455                         goto out;
2456                 r = -EFAULT;
2457                 if (copy_to_user(argp, &tr, sizeof(tr)))
2458                         goto out;
2459                 r = 0;
2460                 break;
2461         }
2462         case KVM_SET_GUEST_DEBUG: {
2463                 struct kvm_guest_debug dbg;
2464
2465                 r = -EFAULT;
2466                 if (copy_from_user(&dbg, argp, sizeof(dbg)))
2467                         goto out;
2468                 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
2469                 break;
2470         }
2471         case KVM_SET_SIGNAL_MASK: {
2472                 struct kvm_signal_mask __user *sigmask_arg = argp;
2473                 struct kvm_signal_mask kvm_sigmask;
2474                 sigset_t sigset, *p;
2475
2476                 p = NULL;
2477                 if (argp) {
2478                         r = -EFAULT;
2479                         if (copy_from_user(&kvm_sigmask, argp,
2480                                            sizeof(kvm_sigmask)))
2481                                 goto out;
2482                         r = -EINVAL;
2483                         if (kvm_sigmask.len != sizeof(sigset))
2484                                 goto out;
2485                         r = -EFAULT;
2486                         if (copy_from_user(&sigset, sigmask_arg->sigset,
2487                                            sizeof(sigset)))
2488                                 goto out;
2489                         p = &sigset;
2490                 }
2491                 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
2492                 break;
2493         }
2494         case KVM_GET_FPU: {
2495                 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL);
2496                 r = -ENOMEM;
2497                 if (!fpu)
2498                         goto out;
2499                 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
2500                 if (r)
2501                         goto out;
2502                 r = -EFAULT;
2503                 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
2504                         goto out;
2505                 r = 0;
2506                 break;
2507         }
2508         case KVM_SET_FPU: {
2509                 fpu = memdup_user(argp, sizeof(*fpu));
2510                 if (IS_ERR(fpu)) {
2511                         r = PTR_ERR(fpu);
2512                         fpu = NULL;
2513                         goto out;
2514                 }
2515                 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
2516                 break;
2517         }
2518         default:
2519                 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
2520         }
2521 out:
2522         vcpu_put(vcpu);
2523         kfree(fpu);
2524         kfree(kvm_sregs);
2525         return r;
2526 }
2527
2528 #ifdef CONFIG_KVM_COMPAT
2529 static long kvm_vcpu_compat_ioctl(struct file *filp,
2530                                   unsigned int ioctl, unsigned long arg)
2531 {
2532         struct kvm_vcpu *vcpu = filp->private_data;
2533         void __user *argp = compat_ptr(arg);
2534         int r;
2535
2536         if (vcpu->kvm->mm != current->mm)
2537                 return -EIO;
2538
2539         switch (ioctl) {
2540         case KVM_SET_SIGNAL_MASK: {
2541                 struct kvm_signal_mask __user *sigmask_arg = argp;
2542                 struct kvm_signal_mask kvm_sigmask;
2543                 compat_sigset_t csigset;
2544                 sigset_t sigset;
2545
2546                 if (argp) {
2547                         r = -EFAULT;
2548                         if (copy_from_user(&kvm_sigmask, argp,
2549                                            sizeof(kvm_sigmask)))
2550                                 goto out;
2551                         r = -EINVAL;
2552                         if (kvm_sigmask.len != sizeof(csigset))
2553                                 goto out;
2554                         r = -EFAULT;
2555                         if (copy_from_user(&csigset, sigmask_arg->sigset,
2556                                            sizeof(csigset)))
2557                                 goto out;
2558                         sigset_from_compat(&sigset, &csigset);
2559                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
2560                 } else
2561                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL);
2562                 break;
2563         }
2564         default:
2565                 r = kvm_vcpu_ioctl(filp, ioctl, arg);
2566         }
2567
2568 out:
2569         return r;
2570 }
2571 #endif
2572
2573 static int kvm_device_ioctl_attr(struct kvm_device *dev,
2574                                  int (*accessor)(struct kvm_device *dev,
2575                                                  struct kvm_device_attr *attr),
2576                                  unsigned long arg)
2577 {
2578         struct kvm_device_attr attr;
2579
2580         if (!accessor)
2581                 return -EPERM;
2582
2583         if (copy_from_user(&attr, (void __user *)arg, sizeof(attr)))
2584                 return -EFAULT;
2585
2586         return accessor(dev, &attr);
2587 }
2588
2589 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl,
2590                              unsigned long arg)
2591 {
2592         struct kvm_device *dev = filp->private_data;
2593
2594         switch (ioctl) {
2595         case KVM_SET_DEVICE_ATTR:
2596                 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg);
2597         case KVM_GET_DEVICE_ATTR:
2598                 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg);
2599         case KVM_HAS_DEVICE_ATTR:
2600                 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg);
2601         default:
2602                 if (dev->ops->ioctl)
2603                         return dev->ops->ioctl(dev, ioctl, arg);
2604
2605                 return -ENOTTY;
2606         }
2607 }
2608
2609 static int kvm_device_release(struct inode *inode, struct file *filp)
2610 {
2611         struct kvm_device *dev = filp->private_data;
2612         struct kvm *kvm = dev->kvm;
2613
2614         kvm_put_kvm(kvm);
2615         return 0;
2616 }
2617
2618 static const struct file_operations kvm_device_fops = {
2619         .unlocked_ioctl = kvm_device_ioctl,
2620 #ifdef CONFIG_KVM_COMPAT
2621         .compat_ioctl = kvm_device_ioctl,
2622 #endif
2623         .release = kvm_device_release,
2624 };
2625
2626 struct kvm_device *kvm_device_from_filp(struct file *filp)
2627 {
2628         if (filp->f_op != &kvm_device_fops)
2629                 return NULL;
2630
2631         return filp->private_data;
2632 }
2633
2634 static struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = {
2635 #ifdef CONFIG_KVM_MPIC
2636         [KVM_DEV_TYPE_FSL_MPIC_20]      = &kvm_mpic_ops,
2637         [KVM_DEV_TYPE_FSL_MPIC_42]      = &kvm_mpic_ops,
2638 #endif
2639
2640 #ifdef CONFIG_KVM_XICS
2641         [KVM_DEV_TYPE_XICS]             = &kvm_xics_ops,
2642 #endif
2643 };
2644
2645 int kvm_register_device_ops(struct kvm_device_ops *ops, u32 type)
2646 {
2647         if (type >= ARRAY_SIZE(kvm_device_ops_table))
2648                 return -ENOSPC;
2649
2650         if (kvm_device_ops_table[type] != NULL)
2651                 return -EEXIST;
2652
2653         kvm_device_ops_table[type] = ops;
2654         return 0;
2655 }
2656
2657 void kvm_unregister_device_ops(u32 type)
2658 {
2659         if (kvm_device_ops_table[type] != NULL)
2660                 kvm_device_ops_table[type] = NULL;
2661 }
2662
2663 static int kvm_ioctl_create_device(struct kvm *kvm,
2664                                    struct kvm_create_device *cd)
2665 {
2666         struct kvm_device_ops *ops = NULL;
2667         struct kvm_device *dev;
2668         bool test = cd->flags & KVM_CREATE_DEVICE_TEST;
2669         int ret;
2670
2671         if (cd->type >= ARRAY_SIZE(kvm_device_ops_table))
2672                 return -ENODEV;
2673
2674         ops = kvm_device_ops_table[cd->type];
2675         if (ops == NULL)
2676                 return -ENODEV;
2677
2678         if (test)
2679                 return 0;
2680
2681         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2682         if (!dev)
2683                 return -ENOMEM;
2684
2685         dev->ops = ops;
2686         dev->kvm = kvm;
2687
2688         ret = ops->create(dev, cd->type);
2689         if (ret < 0) {
2690                 kfree(dev);
2691                 return ret;
2692         }
2693
2694         ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC);
2695         if (ret < 0) {
2696                 ops->destroy(dev);
2697                 return ret;
2698         }
2699
2700         list_add(&dev->vm_node, &kvm->devices);
2701         kvm_get_kvm(kvm);
2702         cd->fd = ret;
2703         return 0;
2704 }
2705
2706 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
2707 {
2708         switch (arg) {
2709         case KVM_CAP_USER_MEMORY:
2710         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
2711         case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
2712         case KVM_CAP_INTERNAL_ERROR_DATA:
2713 #ifdef CONFIG_HAVE_KVM_MSI
2714         case KVM_CAP_SIGNAL_MSI:
2715 #endif
2716 #ifdef CONFIG_HAVE_KVM_IRQFD
2717         case KVM_CAP_IRQFD:
2718         case KVM_CAP_IRQFD_RESAMPLE:
2719 #endif
2720         case KVM_CAP_IOEVENTFD_ANY_LENGTH:
2721         case KVM_CAP_CHECK_EXTENSION_VM:
2722                 return 1;
2723 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2724         case KVM_CAP_IRQ_ROUTING:
2725                 return KVM_MAX_IRQ_ROUTES;
2726 #endif
2727 #if KVM_ADDRESS_SPACE_NUM > 1
2728         case KVM_CAP_MULTI_ADDRESS_SPACE:
2729                 return KVM_ADDRESS_SPACE_NUM;
2730 #endif
2731         default:
2732                 break;
2733         }
2734         return kvm_vm_ioctl_check_extension(kvm, arg);
2735 }
2736
2737 static long kvm_vm_ioctl(struct file *filp,
2738                            unsigned int ioctl, unsigned long arg)
2739 {
2740         struct kvm *kvm = filp->private_data;
2741         void __user *argp = (void __user *)arg;
2742         int r;
2743
2744         if (kvm->mm != current->mm)
2745                 return -EIO;
2746         switch (ioctl) {
2747         case KVM_CREATE_VCPU:
2748                 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
2749                 break;
2750         case KVM_SET_USER_MEMORY_REGION: {
2751                 struct kvm_userspace_memory_region kvm_userspace_mem;
2752
2753                 r = -EFAULT;
2754                 if (copy_from_user(&kvm_userspace_mem, argp,
2755                                                 sizeof(kvm_userspace_mem)))
2756                         goto out;
2757
2758                 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem);
2759                 break;
2760         }
2761         case KVM_GET_DIRTY_LOG: {
2762                 struct kvm_dirty_log log;
2763
2764                 r = -EFAULT;
2765                 if (copy_from_user(&log, argp, sizeof(log)))
2766                         goto out;
2767                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2768                 break;
2769         }
2770 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2771         case KVM_REGISTER_COALESCED_MMIO: {
2772                 struct kvm_coalesced_mmio_zone zone;
2773
2774                 r = -EFAULT;
2775                 if (copy_from_user(&zone, argp, sizeof(zone)))
2776                         goto out;
2777                 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
2778                 break;
2779         }
2780         case KVM_UNREGISTER_COALESCED_MMIO: {
2781                 struct kvm_coalesced_mmio_zone zone;
2782
2783                 r = -EFAULT;
2784                 if (copy_from_user(&zone, argp, sizeof(zone)))
2785                         goto out;
2786                 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
2787                 break;
2788         }
2789 #endif
2790         case KVM_IRQFD: {
2791                 struct kvm_irqfd data;
2792
2793                 r = -EFAULT;
2794                 if (copy_from_user(&data, argp, sizeof(data)))
2795                         goto out;
2796                 r = kvm_irqfd(kvm, &data);
2797                 break;
2798         }
2799         case KVM_IOEVENTFD: {
2800                 struct kvm_ioeventfd data;
2801
2802                 r = -EFAULT;
2803                 if (copy_from_user(&data, argp, sizeof(data)))
2804                         goto out;
2805                 r = kvm_ioeventfd(kvm, &data);
2806                 break;
2807         }
2808 #ifdef CONFIG_HAVE_KVM_MSI
2809         case KVM_SIGNAL_MSI: {
2810                 struct kvm_msi msi;
2811
2812                 r = -EFAULT;
2813                 if (copy_from_user(&msi, argp, sizeof(msi)))
2814                         goto out;
2815                 r = kvm_send_userspace_msi(kvm, &msi);
2816                 break;
2817         }
2818 #endif
2819 #ifdef __KVM_HAVE_IRQ_LINE
2820         case KVM_IRQ_LINE_STATUS:
2821         case KVM_IRQ_LINE: {
2822                 struct kvm_irq_level irq_event;
2823
2824                 r = -EFAULT;
2825                 if (copy_from_user(&irq_event, argp, sizeof(irq_event)))
2826                         goto out;
2827
2828                 r = kvm_vm_ioctl_irq_line(kvm, &irq_event,
2829                                         ioctl == KVM_IRQ_LINE_STATUS);
2830                 if (r)
2831                         goto out;
2832
2833                 r = -EFAULT;
2834                 if (ioctl == KVM_IRQ_LINE_STATUS) {
2835                         if (copy_to_user(argp, &irq_event, sizeof(irq_event)))
2836                                 goto out;
2837                 }
2838
2839                 r = 0;
2840                 break;
2841         }
2842 #endif
2843 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
2844         case KVM_SET_GSI_ROUTING: {
2845                 struct kvm_irq_routing routing;
2846                 struct kvm_irq_routing __user *urouting;
2847                 struct kvm_irq_routing_entry *entries;
2848
2849                 r = -EFAULT;
2850                 if (copy_from_user(&routing, argp, sizeof(routing)))
2851                         goto out;
2852                 r = -EINVAL;
2853                 if (routing.nr >= KVM_MAX_IRQ_ROUTES)
2854                         goto out;
2855                 if (routing.flags)
2856                         goto out;
2857                 r = -ENOMEM;
2858                 entries = vmalloc(routing.nr * sizeof(*entries));
2859                 if (!entries)
2860                         goto out;
2861                 r = -EFAULT;
2862                 urouting = argp;
2863                 if (copy_from_user(entries, urouting->entries,
2864                                    routing.nr * sizeof(*entries)))
2865                         goto out_free_irq_routing;
2866                 r = kvm_set_irq_routing(kvm, entries, routing.nr,
2867                                         routing.flags);
2868 out_free_irq_routing:
2869                 vfree(entries);
2870                 break;
2871         }
2872 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */
2873         case KVM_CREATE_DEVICE: {
2874                 struct kvm_create_device cd;
2875
2876                 r = -EFAULT;
2877                 if (copy_from_user(&cd, argp, sizeof(cd)))
2878                         goto out;
2879
2880                 r = kvm_ioctl_create_device(kvm, &cd);
2881                 if (r)
2882                         goto out;
2883
2884                 r = -EFAULT;
2885                 if (copy_to_user(argp, &cd, sizeof(cd)))
2886                         goto out;
2887
2888                 r = 0;
2889                 break;
2890         }
2891         case KVM_CHECK_EXTENSION:
2892                 r = kvm_vm_ioctl_check_extension_generic(kvm, arg);
2893                 break;
2894         default:
2895                 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
2896         }
2897 out:
2898         return r;
2899 }
2900
2901 #ifdef CONFIG_KVM_COMPAT
2902 struct compat_kvm_dirty_log {
2903         __u32 slot;
2904         __u32 padding1;
2905         union {
2906                 compat_uptr_t dirty_bitmap; /* one bit per page */
2907                 __u64 padding2;
2908         };
2909 };
2910
2911 static long kvm_vm_compat_ioctl(struct file *filp,
2912                            unsigned int ioctl, unsigned long arg)
2913 {
2914         struct kvm *kvm = filp->private_data;
2915         int r;
2916
2917         if (kvm->mm != current->mm)
2918                 return -EIO;
2919         switch (ioctl) {
2920         case KVM_GET_DIRTY_LOG: {
2921                 struct compat_kvm_dirty_log compat_log;
2922                 struct kvm_dirty_log log;
2923
2924                 r = -EFAULT;
2925                 if (copy_from_user(&compat_log, (void __user *)arg,
2926                                    sizeof(compat_log)))
2927                         goto out;
2928                 log.slot         = compat_log.slot;
2929                 log.padding1     = compat_log.padding1;
2930                 log.padding2     = compat_log.padding2;
2931                 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
2932
2933                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
2934                 break;
2935         }
2936         default:
2937                 r = kvm_vm_ioctl(filp, ioctl, arg);
2938         }
2939
2940 out:
2941         return r;
2942 }
2943 #endif
2944
2945 static struct file_operations kvm_vm_fops = {
2946         .release        = kvm_vm_release,
2947         .unlocked_ioctl = kvm_vm_ioctl,
2948 #ifdef CONFIG_KVM_COMPAT
2949         .compat_ioctl   = kvm_vm_compat_ioctl,
2950 #endif
2951         .llseek         = noop_llseek,
2952 };
2953
2954 static int kvm_dev_ioctl_create_vm(unsigned long type)
2955 {
2956         int r;
2957         struct kvm *kvm;
2958
2959         kvm = kvm_create_vm(type);
2960         if (IS_ERR(kvm))
2961                 return PTR_ERR(kvm);
2962 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
2963         r = kvm_coalesced_mmio_init(kvm);
2964         if (r < 0) {
2965                 kvm_put_kvm(kvm);
2966                 return r;
2967         }
2968 #endif
2969         r = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, O_RDWR | O_CLOEXEC);
2970         if (r < 0)
2971                 kvm_put_kvm(kvm);
2972
2973         return r;
2974 }
2975
2976 static long kvm_dev_ioctl(struct file *filp,
2977                           unsigned int ioctl, unsigned long arg)
2978 {
2979         long r = -EINVAL;
2980
2981         switch (ioctl) {
2982         case KVM_GET_API_VERSION:
2983                 if (arg)
2984                         goto out;
2985                 r = KVM_API_VERSION;
2986                 break;
2987         case KVM_CREATE_VM:
2988                 r = kvm_dev_ioctl_create_vm(arg);
2989                 break;
2990         case KVM_CHECK_EXTENSION:
2991                 r = kvm_vm_ioctl_check_extension_generic(NULL, arg);
2992                 break;
2993         case KVM_GET_VCPU_MMAP_SIZE:
2994                 if (arg)
2995                         goto out;
2996                 r = PAGE_SIZE;     /* struct kvm_run */
2997 #ifdef CONFIG_X86
2998                 r += PAGE_SIZE;    /* pio data page */
2999 #endif
3000 #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
3001                 r += PAGE_SIZE;    /* coalesced mmio ring page */
3002 #endif
3003                 break;
3004         case KVM_TRACE_ENABLE:
3005         case KVM_TRACE_PAUSE:
3006         case KVM_TRACE_DISABLE:
3007                 r = -EOPNOTSUPP;
3008                 break;
3009         default:
3010                 return kvm_arch_dev_ioctl(filp, ioctl, arg);
3011         }
3012 out:
3013         return r;
3014 }
3015
3016 static struct file_operations kvm_chardev_ops = {
3017         .unlocked_ioctl = kvm_dev_ioctl,
3018         .compat_ioctl   = kvm_dev_ioctl,
3019         .llseek         = noop_llseek,
3020 };
3021
3022 static struct miscdevice kvm_dev = {
3023         KVM_MINOR,
3024         "kvm",
3025         &kvm_chardev_ops,
3026 };
3027
3028 static void hardware_enable_nolock(void *junk)
3029 {
3030         int cpu = raw_smp_processor_id();
3031         int r;
3032
3033         if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
3034                 return;
3035
3036         cpumask_set_cpu(cpu, cpus_hardware_enabled);
3037
3038         r = kvm_arch_hardware_enable();
3039
3040         if (r) {
3041                 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3042                 atomic_inc(&hardware_enable_failed);
3043                 pr_info("kvm: enabling virtualization on CPU%d failed\n", cpu);
3044         }
3045 }
3046
3047 static void hardware_enable(void)
3048 {
3049         raw_spin_lock(&kvm_count_lock);
3050         if (kvm_usage_count)
3051                 hardware_enable_nolock(NULL);
3052         raw_spin_unlock(&kvm_count_lock);
3053 }
3054
3055 static void hardware_disable_nolock(void *junk)
3056 {
3057         int cpu = raw_smp_processor_id();
3058
3059         if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
3060                 return;
3061         cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3062         kvm_arch_hardware_disable();
3063 }
3064
3065 static void hardware_disable(void)
3066 {
3067         raw_spin_lock(&kvm_count_lock);
3068         if (kvm_usage_count)
3069                 hardware_disable_nolock(NULL);
3070         raw_spin_unlock(&kvm_count_lock);
3071 }
3072
3073 static void hardware_disable_all_nolock(void)
3074 {
3075         BUG_ON(!kvm_usage_count);
3076
3077         kvm_usage_count--;
3078         if (!kvm_usage_count)
3079                 on_each_cpu(hardware_disable_nolock, NULL, 1);
3080 }
3081
3082 static void hardware_disable_all(void)
3083 {
3084         raw_spin_lock(&kvm_count_lock);
3085         hardware_disable_all_nolock();
3086         raw_spin_unlock(&kvm_count_lock);
3087 }
3088
3089 static int hardware_enable_all(void)
3090 {
3091         int r = 0;
3092
3093         raw_spin_lock(&kvm_count_lock);
3094
3095         kvm_usage_count++;
3096         if (kvm_usage_count == 1) {
3097                 atomic_set(&hardware_enable_failed, 0);
3098                 on_each_cpu(hardware_enable_nolock, NULL, 1);
3099
3100                 if (atomic_read(&hardware_enable_failed)) {
3101                         hardware_disable_all_nolock();
3102                         r = -EBUSY;
3103                 }
3104         }
3105
3106         raw_spin_unlock(&kvm_count_lock);
3107
3108         return r;
3109 }
3110
3111 static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
3112                            void *v)
3113 {
3114         val &= ~CPU_TASKS_FROZEN;
3115         switch (val) {
3116         case CPU_DYING:
3117                 hardware_disable();
3118                 break;
3119         case CPU_STARTING:
3120                 hardware_enable();
3121                 break;
3122         }
3123         return NOTIFY_OK;
3124 }
3125
3126 static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
3127                       void *v)
3128 {
3129         /*
3130          * Some (well, at least mine) BIOSes hang on reboot if
3131          * in vmx root mode.
3132          *
3133          * And Intel TXT required VMX off for all cpu when system shutdown.
3134          */
3135         pr_info("kvm: exiting hardware virtualization\n");
3136         kvm_rebooting = true;
3137         on_each_cpu(hardware_disable_nolock, NULL, 1);
3138         return NOTIFY_OK;
3139 }
3140
3141 static struct notifier_block kvm_reboot_notifier = {
3142         .notifier_call = kvm_reboot,
3143         .priority = 0,
3144 };
3145
3146 static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
3147 {
3148         int i;
3149
3150         for (i = 0; i < bus->dev_count; i++) {
3151                 struct kvm_io_device *pos = bus->range[i].dev;
3152
3153                 kvm_iodevice_destructor(pos);
3154         }
3155         kfree(bus);
3156 }
3157
3158 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1,
3159                                  const struct kvm_io_range *r2)
3160 {
3161         gpa_t addr1 = r1->addr;
3162         gpa_t addr2 = r2->addr;
3163
3164         if (addr1 < addr2)
3165                 return -1;
3166
3167         /* If r2->len == 0, match the exact address.  If r2->len != 0,
3168          * accept any overlapping write.  Any order is acceptable for
3169          * overlapping ranges, because kvm_io_bus_get_first_dev ensures
3170          * we process all of them.
3171          */
3172         if (r2->len) {
3173                 addr1 += r1->len;
3174                 addr2 += r2->len;
3175         }
3176
3177         if (addr1 > addr2)
3178                 return 1;
3179
3180         return 0;
3181 }
3182
3183 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
3184 {
3185         return kvm_io_bus_cmp(p1, p2);
3186 }
3187
3188 static int kvm_io_bus_insert_dev(struct kvm_io_bus *bus, struct kvm_io_device *dev,
3189                           gpa_t addr, int len)
3190 {
3191         bus->range[bus->dev_count++] = (struct kvm_io_range) {
3192                 .addr = addr,
3193                 .len = len,
3194                 .dev = dev,
3195         };
3196
3197         sort(bus->range, bus->dev_count, sizeof(struct kvm_io_range),
3198                 kvm_io_bus_sort_cmp, NULL);
3199
3200         return 0;
3201 }
3202
3203 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
3204                              gpa_t addr, int len)
3205 {
3206         struct kvm_io_range *range, key;
3207         int off;
3208
3209         key = (struct kvm_io_range) {
3210                 .addr = addr,
3211                 .len = len,
3212         };
3213
3214         range = bsearch(&key, bus->range, bus->dev_count,
3215                         sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
3216         if (range == NULL)
3217                 return -ENOENT;
3218
3219         off = range - bus->range;
3220
3221         while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0)
3222                 off--;
3223
3224         return off;
3225 }
3226
3227 static int __kvm_io_bus_write(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3228                               struct kvm_io_range *range, const void *val)
3229 {
3230         int idx;
3231
3232         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3233         if (idx < 0)
3234                 return -EOPNOTSUPP;
3235
3236         while (idx < bus->dev_count &&
3237                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3238                 if (!kvm_iodevice_write(vcpu, bus->range[idx].dev, range->addr,
3239                                         range->len, val))
3240                         return idx;
3241                 idx++;
3242         }
3243
3244         return -EOPNOTSUPP;
3245 }
3246
3247 /* kvm_io_bus_write - called under kvm->slots_lock */
3248 int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3249                      int len, const void *val)
3250 {
3251         struct kvm_io_bus *bus;
3252         struct kvm_io_range range;
3253         int r;
3254
3255         range = (struct kvm_io_range) {
3256                 .addr = addr,
3257                 .len = len,
3258         };
3259
3260         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3261         r = __kvm_io_bus_write(vcpu, bus, &range, val);
3262         return r < 0 ? r : 0;
3263 }
3264
3265 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */
3266 int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
3267                             gpa_t addr, int len, const void *val, long cookie)
3268 {
3269         struct kvm_io_bus *bus;
3270         struct kvm_io_range range;
3271
3272         range = (struct kvm_io_range) {
3273                 .addr = addr,
3274                 .len = len,
3275         };
3276
3277         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3278
3279         /* First try the device referenced by cookie. */
3280         if ((cookie >= 0) && (cookie < bus->dev_count) &&
3281             (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0))
3282                 if (!kvm_iodevice_write(vcpu, bus->range[cookie].dev, addr, len,
3283                                         val))
3284                         return cookie;
3285
3286         /*
3287          * cookie contained garbage; fall back to search and return the
3288          * correct cookie value.
3289          */
3290         return __kvm_io_bus_write(vcpu, bus, &range, val);
3291 }
3292
3293 static int __kvm_io_bus_read(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3294                              struct kvm_io_range *range, void *val)
3295 {
3296         int idx;
3297
3298         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3299         if (idx < 0)
3300                 return -EOPNOTSUPP;
3301
3302         while (idx < bus->dev_count &&
3303                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3304                 if (!kvm_iodevice_read(vcpu, bus->range[idx].dev, range->addr,
3305                                        range->len, val))
3306                         return idx;
3307                 idx++;
3308         }
3309
3310         return -EOPNOTSUPP;
3311 }
3312 EXPORT_SYMBOL_GPL(kvm_io_bus_write);
3313
3314 /* kvm_io_bus_read - called under kvm->slots_lock */
3315 int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3316                     int len, void *val)
3317 {
3318         struct kvm_io_bus *bus;
3319         struct kvm_io_range range;
3320         int r;
3321
3322         range = (struct kvm_io_range) {
3323                 .addr = addr,
3324                 .len = len,
3325         };
3326
3327         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3328         r = __kvm_io_bus_read(vcpu, bus, &range, val);
3329         return r < 0 ? r : 0;
3330 }
3331
3332
3333 /* Caller must hold slots_lock. */
3334 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
3335                             int len, struct kvm_io_device *dev)
3336 {
3337         struct kvm_io_bus *new_bus, *bus;
3338
3339         bus = kvm->buses[bus_idx];
3340         /* exclude ioeventfd which is limited by maximum fd */
3341         if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
3342                 return -ENOSPC;
3343
3344         new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count + 1) *
3345                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3346         if (!new_bus)
3347                 return -ENOMEM;
3348         memcpy(new_bus, bus, sizeof(*bus) + (bus->dev_count *
3349                sizeof(struct kvm_io_range)));
3350         kvm_io_bus_insert_dev(new_bus, dev, addr, len);
3351         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3352         synchronize_srcu_expedited(&kvm->srcu);
3353         kfree(bus);
3354
3355         return 0;
3356 }
3357
3358 /* Caller must hold slots_lock. */
3359 int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
3360                               struct kvm_io_device *dev)
3361 {
3362         int i, r;
3363         struct kvm_io_bus *new_bus, *bus;
3364
3365         bus = kvm->buses[bus_idx];
3366         r = -ENOENT;
3367         for (i = 0; i < bus->dev_count; i++)
3368                 if (bus->range[i].dev == dev) {
3369                         r = 0;
3370                         break;
3371                 }
3372
3373         if (r)
3374                 return r;
3375
3376         new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
3377                           sizeof(struct kvm_io_range)), GFP_KERNEL);
3378         if (!new_bus)
3379                 return -ENOMEM;
3380
3381         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
3382         new_bus->dev_count--;
3383         memcpy(new_bus->range + i, bus->range + i + 1,
3384                (new_bus->dev_count - i) * sizeof(struct kvm_io_range));
3385
3386         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
3387         synchronize_srcu_expedited(&kvm->srcu);
3388         kfree(bus);
3389         return r;
3390 }
3391
3392 static struct notifier_block kvm_cpu_notifier = {
3393         .notifier_call = kvm_cpu_hotplug,
3394 };
3395
3396 static int vm_stat_get(void *_offset, u64 *val)
3397 {
3398         unsigned offset = (long)_offset;
3399         struct kvm *kvm;
3400
3401         *val = 0;
3402         spin_lock(&kvm_lock);
3403         list_for_each_entry(kvm, &vm_list, vm_list)
3404                 *val += *(u32 *)((void *)kvm + offset);
3405         spin_unlock(&kvm_lock);
3406         return 0;
3407 }
3408
3409 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, NULL, "%llu\n");
3410
3411 static int vcpu_stat_get(void *_offset, u64 *val)
3412 {
3413         unsigned offset = (long)_offset;
3414         struct kvm *kvm;
3415         struct kvm_vcpu *vcpu;
3416         int i;
3417
3418         *val = 0;
3419         spin_lock(&kvm_lock);
3420         list_for_each_entry(kvm, &vm_list, vm_list)
3421                 kvm_for_each_vcpu(i, vcpu, kvm)
3422                         *val += *(u32 *)((void *)vcpu + offset);
3423
3424         spin_unlock(&kvm_lock);
3425         return 0;
3426 }
3427
3428 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n");
3429
3430 static const struct file_operations *stat_fops[] = {
3431         [KVM_STAT_VCPU] = &vcpu_stat_fops,
3432         [KVM_STAT_VM]   = &vm_stat_fops,
3433 };
3434
3435 static int kvm_init_debug(void)
3436 {
3437         int r = -EEXIST;
3438         struct kvm_stats_debugfs_item *p;
3439
3440         kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
3441         if (kvm_debugfs_dir == NULL)
3442                 goto out;
3443
3444         for (p = debugfs_entries; p->name; ++p) {
3445                 if (!debugfs_create_file(p->name, 0444, kvm_debugfs_dir,
3446                                          (void *)(long)p->offset,
3447                                          stat_fops[p->kind]))
3448                         goto out_dir;
3449         }
3450
3451         return 0;
3452
3453 out_dir:
3454         debugfs_remove_recursive(kvm_debugfs_dir);
3455 out:
3456         return r;
3457 }
3458
3459 static int kvm_suspend(void)
3460 {
3461         if (kvm_usage_count)
3462                 hardware_disable_nolock(NULL);
3463         return 0;
3464 }
3465
3466 static void kvm_resume(void)
3467 {
3468         if (kvm_usage_count) {
3469                 WARN_ON(raw_spin_is_locked(&kvm_count_lock));
3470                 hardware_enable_nolock(NULL);
3471         }
3472 }
3473
3474 static struct syscore_ops kvm_syscore_ops = {
3475         .suspend = kvm_suspend,
3476         .resume = kvm_resume,
3477 };
3478
3479 static inline
3480 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
3481 {
3482         return container_of(pn, struct kvm_vcpu, preempt_notifier);
3483 }
3484
3485 static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
3486 {
3487         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3488
3489         if (vcpu->preempted)
3490                 vcpu->preempted = false;
3491
3492         kvm_arch_sched_in(vcpu, cpu);
3493
3494         kvm_arch_vcpu_load(vcpu, cpu);
3495 }
3496
3497 static void kvm_sched_out(struct preempt_notifier *pn,
3498                           struct task_struct *next)
3499 {
3500         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
3501
3502         if (current->state == TASK_RUNNING)
3503                 vcpu->preempted = true;
3504         kvm_arch_vcpu_put(vcpu);
3505 }
3506
3507 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
3508                   struct module *module)
3509 {
3510         int r;
3511         int cpu;
3512
3513         r = kvm_arch_init(opaque);
3514         if (r)
3515                 goto out_fail;
3516
3517         /*
3518          * kvm_arch_init makes sure there's at most one caller
3519          * for architectures that support multiple implementations,
3520          * like intel and amd on x86.
3521          * kvm_arch_init must be called before kvm_irqfd_init to avoid creating
3522          * conflicts in case kvm is already setup for another implementation.
3523          */
3524         r = kvm_irqfd_init();
3525         if (r)
3526                 goto out_irqfd;
3527
3528         if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
3529                 r = -ENOMEM;
3530                 goto out_free_0;
3531         }
3532
3533         r = kvm_arch_hardware_setup();
3534         if (r < 0)
3535                 goto out_free_0a;
3536
3537         for_each_online_cpu(cpu) {
3538                 smp_call_function_single(cpu,
3539                                 kvm_arch_check_processor_compat,
3540                                 &r, 1);
3541                 if (r < 0)
3542                         goto out_free_1;
3543         }
3544
3545         r = register_cpu_notifier(&kvm_cpu_notifier);
3546         if (r)
3547                 goto out_free_2;
3548         register_reboot_notifier(&kvm_reboot_notifier);
3549
3550         /* A kmem cache lets us meet the alignment requirements of fx_save. */
3551         if (!vcpu_align)
3552                 vcpu_align = __alignof__(struct kvm_vcpu);
3553         kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size, vcpu_align,
3554                                            0, NULL);
3555         if (!kvm_vcpu_cache) {
3556                 r = -ENOMEM;
3557                 goto out_free_3;
3558         }
3559
3560         r = kvm_async_pf_init();
3561         if (r)
3562                 goto out_free;
3563
3564         kvm_chardev_ops.owner = module;
3565         kvm_vm_fops.owner = module;
3566         kvm_vcpu_fops.owner = module;
3567
3568         r = misc_register(&kvm_dev);
3569         if (r) {
3570                 pr_err("kvm: misc device register failed\n");
3571                 goto out_unreg;
3572         }
3573
3574         register_syscore_ops(&kvm_syscore_ops);
3575
3576         kvm_preempt_ops.sched_in = kvm_sched_in;
3577         kvm_preempt_ops.sched_out = kvm_sched_out;
3578
3579         r = kvm_init_debug();
3580         if (r) {
3581                 pr_err("kvm: create debugfs files failed\n");
3582                 goto out_undebugfs;
3583         }
3584
3585         r = kvm_vfio_ops_init();
3586         WARN_ON(r);
3587
3588         return 0;
3589
3590 out_undebugfs:
3591         unregister_syscore_ops(&kvm_syscore_ops);
3592         misc_deregister(&kvm_dev);
3593 out_unreg:
3594         kvm_async_pf_deinit();
3595 out_free:
3596         kmem_cache_destroy(kvm_vcpu_cache);
3597 out_free_3:
3598         unregister_reboot_notifier(&kvm_reboot_notifier);
3599         unregister_cpu_notifier(&kvm_cpu_notifier);
3600 out_free_2:
3601 out_free_1:
3602         kvm_arch_hardware_unsetup();
3603 out_free_0a:
3604         free_cpumask_var(cpus_hardware_enabled);
3605 out_free_0:
3606         kvm_irqfd_exit();
3607 out_irqfd:
3608         kvm_arch_exit();
3609 out_fail:
3610         return r;
3611 }
3612 EXPORT_SYMBOL_GPL(kvm_init);
3613
3614 void kvm_exit(void)
3615 {
3616         debugfs_remove_recursive(kvm_debugfs_dir);
3617         misc_deregister(&kvm_dev);
3618         kmem_cache_destroy(kvm_vcpu_cache);
3619         kvm_async_pf_deinit();
3620         unregister_syscore_ops(&kvm_syscore_ops);
3621         unregister_reboot_notifier(&kvm_reboot_notifier);
3622         unregister_cpu_notifier(&kvm_cpu_notifier);
3623         on_each_cpu(hardware_disable_nolock, NULL, 1);
3624         kvm_arch_hardware_unsetup();
3625         kvm_arch_exit();
3626         kvm_irqfd_exit();
3627         free_cpumask_var(cpus_hardware_enabled);
3628         kvm_vfio_ops_exit();
3629 }
3630 EXPORT_SYMBOL_GPL(kvm_exit);