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