virt/kvm/kvm_main.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * This module enables machines with Intel VT-x extensions to run virtual
   6  * machines without emulation or binary translation.
   7  *
   8  * Copyright (C) 2006 Qumranet, Inc.
   9  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
  10  *
  11  * Authors:
  12  *   Avi Kivity   <avi@qumranet.com>
  13  *   Yaniv Kamay  <yaniv@qumranet.com>
  14  */
  15
  16 #include <kvm/iodev.h>
  17
  18 #include <linux/kvm_host.h>
  19 #include <linux/kvm.h>
  20 #include <linux/module.h>
  21 #include <linux/errno.h>
  22 #include <linux/percpu.h>
  23 #include <linux/mm.h>
  24 #include <linux/miscdevice.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/reboot.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/highmem.h>
  29 #include <linux/file.h>
  30 #include <linux/syscore_ops.h>
  31 #include <linux/cpu.h>
  32 #include <linux/sched/signal.h>
  33 #include <linux/sched/mm.h>
  34 #include <linux/sched/stat.h>
  35 #include <linux/cpumask.h>
  36 #include <linux/smp.h>
  37 #include <linux/anon_inodes.h>
  38 #include <linux/profile.h>
  39 #include <linux/kvm_para.h>
  40 #include <linux/pagemap.h>
  41 #include <linux/mman.h>
  42 #include <linux/swap.h>
  43 #include <linux/bitops.h>
  44 #include <linux/spinlock.h>
  45 #include <linux/compat.h>
  46 #include <linux/srcu.h>
  47 #include <linux/hugetlb.h>
  48 #include <linux/slab.h>
  49 #include <linux/sort.h>
  50 #include <linux/bsearch.h>
  51 #include <linux/io.h>
  52 #include <linux/lockdep.h>
  53 #include <linux/kthread.h>
  54
  55 #include <asm/processor.h>
  56 #include <asm/ioctl.h>
  57 #include <linux/uaccess.h>
  58 #include <asm/pgtable.h>
  59
  60 #include "coalesced_mmio.h"
  61 #include "async_pf.h"
  62 #include "vfio.h"
  63
  64 #define CREATE_TRACE_POINTS
  65 #include <trace/events/kvm.h>
  66
  67 /* Worst case buffer size needed for holding an integer. */
  68 #define ITOA_MAX_LEN 12
  69
  70 MODULE_AUTHOR("Qumranet");
  71 MODULE_LICENSE("GPL");
  72
  73 /* Architectures should define their poll value according to the halt latency */
  74 unsigned int halt_poll_ns = KVM_HALT_POLL_NS_DEFAULT;
  75 module_param(halt_poll_ns, uint, 0644);
  76 EXPORT_SYMBOL_GPL(halt_poll_ns);
  77
  78 /* Default doubles per-vcpu halt_poll_ns. */
  79 unsigned int halt_poll_ns_grow = 2;
  80 module_param(halt_poll_ns_grow, uint, 0644);
  81 EXPORT_SYMBOL_GPL(halt_poll_ns_grow);
  82
  83 /* The start value to grow halt_poll_ns from */
  84 unsigned int halt_poll_ns_grow_start = 10000; /* 10us */
  85 module_param(halt_poll_ns_grow_start, uint, 0644);
  86 EXPORT_SYMBOL_GPL(halt_poll_ns_grow_start);
  87
  88 /* Default resets per-vcpu halt_poll_ns . */
  89 unsigned int halt_poll_ns_shrink;
  90 module_param(halt_poll_ns_shrink, uint, 0644);
  91 EXPORT_SYMBOL_GPL(halt_poll_ns_shrink);
  92
  93 /*
  94  * Ordering of locks:
  95  *
  96  *      kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  97  */
  98
  99 DEFINE_MUTEX(kvm_lock);
 100 static DEFINE_RAW_SPINLOCK(kvm_count_lock);
 101 LIST_HEAD(vm_list);
 102
 103 static cpumask_var_t cpus_hardware_enabled;
 104 static int kvm_usage_count;
 105 static atomic_t hardware_enable_failed;
 106
 107 static struct kmem_cache *kvm_vcpu_cache;
 108
 109 static __read_mostly struct preempt_ops kvm_preempt_ops;
 110 static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_running_vcpu);
 111
 112 struct dentry *kvm_debugfs_dir;
 113 EXPORT_SYMBOL_GPL(kvm_debugfs_dir);
 114
 115 static int kvm_debugfs_num_entries;
 116 static const struct file_operations stat_fops_per_vm;
 117
 118 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
 119                            unsigned long arg);
 120 #ifdef CONFIG_KVM_COMPAT
 121 static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl,
 122                                   unsigned long arg);
 123 #define KVM_COMPAT(c)   .compat_ioctl   = (c)
 124 #else
 125 /*
 126  * For architectures that don't implement a compat infrastructure,
 127  * adopt a double line of defense:
 128  * - Prevent a compat task from opening /dev/kvm
 129  * - If the open has been done by a 64bit task, and the KVM fd
 130  *   passed to a compat task, let the ioctls fail.
 131  */
 132 static long kvm_no_compat_ioctl(struct file *file, unsigned int ioctl,
 133                                 unsigned long arg) { return -EINVAL; }
 134
 135 static int kvm_no_compat_open(struct inode *inode, struct file *file)
 136 {
 137         return is_compat_task() ? -ENODEV : 0;
 138 }
 139 #define KVM_COMPAT(c)   .compat_ioctl   = kvm_no_compat_ioctl,  \
 140                         .open           = kvm_no_compat_open
 141 #endif
 142 static int hardware_enable_all(void);
 143 static void hardware_disable_all(void);
 144
 145 static void kvm_io_bus_destroy(struct kvm_io_bus *bus);
 146
 147 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot, gfn_t gfn);
 148
 149 __visible bool kvm_rebooting;
 150 EXPORT_SYMBOL_GPL(kvm_rebooting);
 151
 152 static bool largepages_enabled = true;
 153
 154 #define KVM_EVENT_CREATE_VM 0
 155 #define KVM_EVENT_DESTROY_VM 1
 156 static void kvm_uevent_notify_change(unsigned int type, struct kvm *kvm);
 157 static unsigned long long kvm_createvm_count;
 158 static unsigned long long kvm_active_vms;
 159
 160 __weak int kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm,
 161                 unsigned long start, unsigned long end, bool blockable)
 162 {
 163         return 0;
 164 }
 165
 166 bool kvm_is_zone_device_pfn(kvm_pfn_t pfn)
 167 {
 168         /*
 169          * The metadata used by is_zone_device_page() to determine whether or
 170          * not a page is ZONE_DEVICE is guaranteed to be valid if and only if
 171          * the device has been pinned, e.g. by get_user_pages().  WARN if the
 172          * page_count() is zero to help detect bad usage of this helper.
 173          */
 174         if (!pfn_valid(pfn) || WARN_ON_ONCE(!page_count(pfn_to_page(pfn))))
 175                 return false;
 176
 177         return is_zone_device_page(pfn_to_page(pfn));
 178 }
 179
 180 bool kvm_is_reserved_pfn(kvm_pfn_t pfn)
 181 {
 182         /*
 183          * ZONE_DEVICE pages currently set PG_reserved, but from a refcounting
 184          * perspective they are "normal" pages, albeit with slightly different
 185          * usage rules.
 186          */
 187         if (pfn_valid(pfn))
 188                 return PageReserved(pfn_to_page(pfn)) &&
 189                        !kvm_is_zone_device_pfn(pfn);
 190
 191         return true;
 192 }
 193
 194 bool kvm_is_transparent_hugepage(kvm_pfn_t pfn)
 195 {
 196         struct page *page = pfn_to_page(pfn);
 197
 198         if (!PageTransCompoundMap(page))
 199                 return false;
 200
 201         return is_transparent_hugepage(compound_head(page));
 202 }
 203
 204 /*
 205  * Switches to specified vcpu, until a matching vcpu_put()
 206  */
 207 void vcpu_load(struct kvm_vcpu *vcpu)
 208 {
 209         int cpu = get_cpu();
 210
 211         __this_cpu_write(kvm_running_vcpu, vcpu);
 212         preempt_notifier_register(&vcpu->preempt_notifier);
 213         kvm_arch_vcpu_load(vcpu, cpu);
 214         put_cpu();
 215 }
 216 EXPORT_SYMBOL_GPL(vcpu_load);
 217
 218 void vcpu_put(struct kvm_vcpu *vcpu)
 219 {
 220         preempt_disable();
 221         kvm_arch_vcpu_put(vcpu);
 222         preempt_notifier_unregister(&vcpu->preempt_notifier);
 223         __this_cpu_write(kvm_running_vcpu, NULL);
 224         preempt_enable();
 225 }
 226 EXPORT_SYMBOL_GPL(vcpu_put);
 227
 228 /* TODO: merge with kvm_arch_vcpu_should_kick */
 229 static bool kvm_request_needs_ipi(struct kvm_vcpu *vcpu, unsigned req)
 230 {
 231         int mode = kvm_vcpu_exiting_guest_mode(vcpu);
 232
 233         /*
 234          * We need to wait for the VCPU to reenable interrupts and get out of
 235          * READING_SHADOW_PAGE_TABLES mode.
 236          */
 237         if (req & KVM_REQUEST_WAIT)
 238                 return mode != OUTSIDE_GUEST_MODE;
 239
 240         /*
 241          * Need to kick a running VCPU, but otherwise there is nothing to do.
 242          */
 243         return mode == IN_GUEST_MODE;
 244 }
 245
 246 static void ack_flush(void *_completed)
 247 {
 248 }
 249
 250 static inline bool kvm_kick_many_cpus(const struct cpumask *cpus, bool wait)
 251 {
 252         if (unlikely(!cpus))
 253                 cpus = cpu_online_mask;
 254
 255         if (cpumask_empty(cpus))
 256                 return false;
 257
 258         smp_call_function_many(cpus, ack_flush, NULL, wait);
 259         return true;
 260 }
 261
 262 bool kvm_make_vcpus_request_mask(struct kvm *kvm, unsigned int req,
 263                                  unsigned long *vcpu_bitmap, cpumask_var_t tmp)
 264 {
 265         int i, cpu, me;
 266         struct kvm_vcpu *vcpu;
 267         bool called;
 268
 269         me = get_cpu();
 270
 271         kvm_for_each_vcpu(i, vcpu, kvm) {
 272                 if (vcpu_bitmap && !test_bit(i, vcpu_bitmap))
 273                         continue;
 274
 275                 kvm_make_request(req, vcpu);
 276                 cpu = vcpu->cpu;
 277
 278                 if (!(req & KVM_REQUEST_NO_WAKEUP) && kvm_vcpu_wake_up(vcpu))
 279                         continue;
 280
 281                 if (tmp != NULL && cpu != -1 && cpu != me &&
 282                     kvm_request_needs_ipi(vcpu, req))
 283                         __cpumask_set_cpu(cpu, tmp);
 284         }
 285
 286         called = kvm_kick_many_cpus(tmp, !!(req & KVM_REQUEST_WAIT));
 287         put_cpu();
 288
 289         return called;
 290 }
 291
 292 bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req)
 293 {
 294         cpumask_var_t cpus;
 295         bool called;
 296
 297         zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 298
 299         called = kvm_make_vcpus_request_mask(kvm, req, NULL, cpus);
 300
 301         free_cpumask_var(cpus);
 302         return called;
 303 }
 304
 305 #ifndef CONFIG_HAVE_KVM_ARCH_TLB_FLUSH_ALL
 306 void kvm_flush_remote_tlbs(struct kvm *kvm)
 307 {
 308         /*
 309          * Read tlbs_dirty before setting KVM_REQ_TLB_FLUSH in
 310          * kvm_make_all_cpus_request.
 311          */
 312         long dirty_count = smp_load_acquire(&kvm->tlbs_dirty);
 313
 314         /*
 315          * We want to publish modifications to the page tables before reading
 316          * mode. Pairs with a memory barrier in arch-specific code.
 317          * - x86: smp_mb__after_srcu_read_unlock in vcpu_enter_guest
 318          * and smp_mb in walk_shadow_page_lockless_begin/end.
 319          * - powerpc: smp_mb in kvmppc_prepare_to_enter.
 320          *
 321          * There is already an smp_mb__after_atomic() before
 322          * kvm_make_all_cpus_request() reads vcpu->mode. We reuse that
 323          * barrier here.
 324          */
 325         if (!kvm_arch_flush_remote_tlb(kvm)
 326             || kvm_make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 327                 ++kvm->stat.remote_tlb_flush;
 328         cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 329 }
 330 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs);
 331 #endif
 332
 333 void kvm_reload_remote_mmus(struct kvm *kvm)
 334 {
 335         kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
 336 }
 337
 338 static void kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
 339 {
 340         mutex_init(&vcpu->mutex);
 341         vcpu->cpu = -1;
 342         vcpu->kvm = kvm;
 343         vcpu->vcpu_id = id;
 344         vcpu->pid = NULL;
 345         init_swait_queue_head(&vcpu->wq);
 346         kvm_async_pf_vcpu_init(vcpu);
 347
 348         vcpu->pre_pcpu = -1;
 349         INIT_LIST_HEAD(&vcpu->blocked_vcpu_list);
 350
 351         kvm_vcpu_set_in_spin_loop(vcpu, false);
 352         kvm_vcpu_set_dy_eligible(vcpu, false);
 353         vcpu->preempted = false;
 354         vcpu->ready = false;
 355         preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
 356 }
 357
 358 void kvm_vcpu_destroy(struct kvm_vcpu *vcpu)
 359 {
 360         kvm_arch_vcpu_destroy(vcpu);
 361
 362         /*
 363          * No need for rcu_read_lock as VCPU_RUN is the only place that changes
 364          * the vcpu->pid pointer, and at destruction time all file descriptors
 365          * are already gone.
 366          */
 367         put_pid(rcu_dereference_protected(vcpu->pid, 1));
 368
 369         free_page((unsigned long)vcpu->run);
 370         kmem_cache_free(kvm_vcpu_cache, vcpu);
 371 }
 372 EXPORT_SYMBOL_GPL(kvm_vcpu_destroy);
 373
 374 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 375 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
 376 {
 377         return container_of(mn, struct kvm, mmu_notifier);
 378 }
 379
 380 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
 381                                         struct mm_struct *mm,
 382                                         unsigned long address,
 383                                         pte_t pte)
 384 {
 385         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 386         int idx;
 387
 388         idx = srcu_read_lock(&kvm->srcu);
 389         spin_lock(&kvm->mmu_lock);
 390         kvm->mmu_notifier_seq++;
 391
 392         if (kvm_set_spte_hva(kvm, address, pte))
 393                 kvm_flush_remote_tlbs(kvm);
 394
 395         spin_unlock(&kvm->mmu_lock);
 396         srcu_read_unlock(&kvm->srcu, idx);
 397 }
 398
 399 static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 400                                         const struct mmu_notifier_range *range)
 401 {
 402         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 403         int need_tlb_flush = 0, idx;
 404         int ret;
 405
 406         idx = srcu_read_lock(&kvm->srcu);
 407         spin_lock(&kvm->mmu_lock);
 408         /*
 409          * The count increase must become visible at unlock time as no
 410          * spte can be established without taking the mmu_lock and
 411          * count is also read inside the mmu_lock critical section.
 412          */
 413         kvm->mmu_notifier_count++;
 414         need_tlb_flush = kvm_unmap_hva_range(kvm, range->start, range->end);
 415         need_tlb_flush |= kvm->tlbs_dirty;
 416         /* we've to flush the tlb before the pages can be freed */
 417         if (need_tlb_flush)
 418                 kvm_flush_remote_tlbs(kvm);
 419
 420         spin_unlock(&kvm->mmu_lock);
 421
 422         ret = kvm_arch_mmu_notifier_invalidate_range(kvm, range->start,
 423                                         range->end,
 424                                         mmu_notifier_range_blockable(range));
 425
 426         srcu_read_unlock(&kvm->srcu, idx);
 427
 428         return ret;
 429 }
 430
 431 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
 432                                         const struct mmu_notifier_range *range)
 433 {
 434         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 435
 436         spin_lock(&kvm->mmu_lock);
 437         /*
 438          * This sequence increase will notify the kvm page fault that
 439          * the page that is going to be mapped in the spte could have
 440          * been freed.
 441          */
 442         kvm->mmu_notifier_seq++;
 443         smp_wmb();
 444         /*
 445          * The above sequence increase must be visible before the
 446          * below count decrease, which is ensured by the smp_wmb above
 447          * in conjunction with the smp_rmb in mmu_notifier_retry().
 448          */
 449         kvm->mmu_notifier_count--;
 450         spin_unlock(&kvm->mmu_lock);
 451
 452         BUG_ON(kvm->mmu_notifier_count < 0);
 453 }
 454
 455 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
 456                                               struct mm_struct *mm,
 457                                               unsigned long start,
 458                                               unsigned long end)
 459 {
 460         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 461         int young, idx;
 462
 463         idx = srcu_read_lock(&kvm->srcu);
 464         spin_lock(&kvm->mmu_lock);
 465
 466         young = kvm_age_hva(kvm, start, end);
 467         if (young)
 468                 kvm_flush_remote_tlbs(kvm);
 469
 470         spin_unlock(&kvm->mmu_lock);
 471         srcu_read_unlock(&kvm->srcu, idx);
 472
 473         return young;
 474 }
 475
 476 static int kvm_mmu_notifier_clear_young(struct mmu_notifier *mn,
 477                                         struct mm_struct *mm,
 478                                         unsigned long start,
 479                                         unsigned long end)
 480 {
 481         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 482         int young, idx;
 483
 484         idx = srcu_read_lock(&kvm->srcu);
 485         spin_lock(&kvm->mmu_lock);
 486         /*
 487          * Even though we do not flush TLB, this will still adversely
 488          * affect performance on pre-Haswell Intel EPT, where there is
 489          * no EPT Access Bit to clear so that we have to tear down EPT
 490          * tables instead. If we find this unacceptable, we can always
 491          * add a parameter to kvm_age_hva so that it effectively doesn't
 492          * do anything on clear_young.
 493          *
 494          * Also note that currently we never issue secondary TLB flushes
 495          * from clear_young, leaving this job up to the regular system
 496          * cadence. If we find this inaccurate, we might come up with a
 497          * more sophisticated heuristic later.
 498          */
 499         young = kvm_age_hva(kvm, start, end);
 500         spin_unlock(&kvm->mmu_lock);
 501         srcu_read_unlock(&kvm->srcu, idx);
 502
 503         return young;
 504 }
 505
 506 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn,
 507                                        struct mm_struct *mm,
 508                                        unsigned long address)
 509 {
 510         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 511         int young, idx;
 512
 513         idx = srcu_read_lock(&kvm->srcu);
 514         spin_lock(&kvm->mmu_lock);
 515         young = kvm_test_age_hva(kvm, address);
 516         spin_unlock(&kvm->mmu_lock);
 517         srcu_read_unlock(&kvm->srcu, idx);
 518
 519         return young;
 520 }
 521
 522 static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
 523                                      struct mm_struct *mm)
 524 {
 525         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 526         int idx;
 527
 528         idx = srcu_read_lock(&kvm->srcu);
 529         kvm_arch_flush_shadow_all(kvm);
 530         srcu_read_unlock(&kvm->srcu, idx);
 531 }
 532
 533 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
 534         .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
 535         .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
 536         .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
 537         .clear_young            = kvm_mmu_notifier_clear_young,
 538         .test_young             = kvm_mmu_notifier_test_young,
 539         .change_pte             = kvm_mmu_notifier_change_pte,
 540         .release                = kvm_mmu_notifier_release,
 541 };
 542
 543 static int kvm_init_mmu_notifier(struct kvm *kvm)
 544 {
 545         kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
 546         return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
 547 }
 548
 549 #else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
 550
 551 static int kvm_init_mmu_notifier(struct kvm *kvm)
 552 {
 553         return 0;
 554 }
 555
 556 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 557
 558 static struct kvm_memslots *kvm_alloc_memslots(void)
 559 {
 560         int i;
 561         struct kvm_memslots *slots;
 562
 563         slots = kvzalloc(sizeof(struct kvm_memslots), GFP_KERNEL_ACCOUNT);
 564         if (!slots)
 565                 return NULL;
 566
 567         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 568                 slots->id_to_index[i] = slots->memslots[i].id = i;
 569
 570         return slots;
 571 }
 572
 573 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
 574 {
 575         if (!memslot->dirty_bitmap)
 576                 return;
 577
 578         kvfree(memslot->dirty_bitmap);
 579         memslot->dirty_bitmap = NULL;
 580 }
 581
 582 /*
 583  * Free any memory in @free but not in @dont.
 584  */
 585 static void kvm_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
 586                               struct kvm_memory_slot *dont)
 587 {
 588         if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
 589                 kvm_destroy_dirty_bitmap(free);
 590
 591         kvm_arch_free_memslot(kvm, free, dont);
 592
 593         free->npages = 0;
 594 }
 595
 596 static void kvm_free_memslots(struct kvm *kvm, struct kvm_memslots *slots)
 597 {
 598         struct kvm_memory_slot *memslot;
 599
 600         if (!slots)
 601                 return;
 602
 603         kvm_for_each_memslot(memslot, slots)
 604                 kvm_free_memslot(kvm, memslot, NULL);
 605
 606         kvfree(slots);
 607 }
 608
 609 static void kvm_destroy_vm_debugfs(struct kvm *kvm)
 610 {
 611         int i;
 612
 613         if (!kvm->debugfs_dentry)
 614                 return;
 615
 616         debugfs_remove_recursive(kvm->debugfs_dentry);
 617
 618         if (kvm->debugfs_stat_data) {
 619                 for (i = 0; i < kvm_debugfs_num_entries; i++)
 620                         kfree(kvm->debugfs_stat_data[i]);
 621                 kfree(kvm->debugfs_stat_data);
 622         }
 623 }
 624
 625 static int kvm_create_vm_debugfs(struct kvm *kvm, int fd)
 626 {
 627         char dir_name[ITOA_MAX_LEN * 2];
 628         struct kvm_stat_data *stat_data;
 629         struct kvm_stats_debugfs_item *p;
 630
 631         if (!debugfs_initialized())
 632                 return 0;
 633
 634         snprintf(dir_name, sizeof(dir_name), "%d-%d", task_pid_nr(current), fd);
 635         kvm->debugfs_dentry = debugfs_create_dir(dir_name, kvm_debugfs_dir);
 636
 637         kvm->debugfs_stat_data = kcalloc(kvm_debugfs_num_entries,
 638                                          sizeof(*kvm->debugfs_stat_data),
 639                                          GFP_KERNEL_ACCOUNT);
 640         if (!kvm->debugfs_stat_data)
 641                 return -ENOMEM;
 642
 643         for (p = debugfs_entries; p->name; p++) {
 644                 stat_data = kzalloc(sizeof(*stat_data), GFP_KERNEL_ACCOUNT);
 645                 if (!stat_data)
 646                         return -ENOMEM;
 647
 648                 stat_data->kvm = kvm;
 649                 stat_data->dbgfs_item = p;
 650                 kvm->debugfs_stat_data[p - debugfs_entries] = stat_data;
 651                 debugfs_create_file(p->name, KVM_DBGFS_GET_MODE(p),
 652                                     kvm->debugfs_dentry, stat_data,
 653                                     &stat_fops_per_vm);
 654         }
 655         return 0;
 656 }
 657
 658 /*
 659  * Called after the VM is otherwise initialized, but just before adding it to
 660  * the vm_list.
 661  */
 662 int __weak kvm_arch_post_init_vm(struct kvm *kvm)
 663 {
 664         return 0;
 665 }
 666
 667 /*
 668  * Called just after removing the VM from the vm_list, but before doing any
 669  * other destruction.
 670  */
 671 void __weak kvm_arch_pre_destroy_vm(struct kvm *kvm)
 672 {
 673 }
 674
 675 static struct kvm *kvm_create_vm(unsigned long type)
 676 {
 677         struct kvm *kvm = kvm_arch_alloc_vm();
 678         int r = -ENOMEM;
 679         int i;
 680
 681         if (!kvm)
 682                 return ERR_PTR(-ENOMEM);
 683
 684         spin_lock_init(&kvm->mmu_lock);
 685         mmgrab(current->mm);
 686         kvm->mm = current->mm;
 687         kvm_eventfd_init(kvm);
 688         mutex_init(&kvm->lock);
 689         mutex_init(&kvm->irq_lock);
 690         mutex_init(&kvm->slots_lock);
 691         INIT_LIST_HEAD(&kvm->devices);
 692
 693         BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX);
 694
 695         if (init_srcu_struct(&kvm->srcu))
 696                 goto out_err_no_srcu;
 697         if (init_srcu_struct(&kvm->irq_srcu))
 698                 goto out_err_no_irq_srcu;
 699
 700         refcount_set(&kvm->users_count, 1);
 701         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
 702                 struct kvm_memslots *slots = kvm_alloc_memslots();
 703
 704                 if (!slots)
 705                         goto out_err_no_arch_destroy_vm;
 706                 /* Generations must be different for each address space. */
 707                 slots->generation = i;
 708                 rcu_assign_pointer(kvm->memslots[i], slots);
 709         }
 710
 711         for (i = 0; i < KVM_NR_BUSES; i++) {
 712                 rcu_assign_pointer(kvm->buses[i],
 713                         kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL_ACCOUNT));
 714                 if (!kvm->buses[i])
 715                         goto out_err_no_arch_destroy_vm;
 716         }
 717
 718         r = kvm_arch_init_vm(kvm, type);
 719         if (r)
 720                 goto out_err_no_arch_destroy_vm;
 721
 722         r = hardware_enable_all();
 723         if (r)
 724                 goto out_err_no_disable;
 725
 726 #ifdef CONFIG_HAVE_KVM_IRQFD
 727         INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 728 #endif
 729
 730         r = kvm_init_mmu_notifier(kvm);
 731         if (r)
 732                 goto out_err_no_mmu_notifier;
 733
 734         r = kvm_arch_post_init_vm(kvm);
 735         if (r)
 736                 goto out_err;
 737
 738         mutex_lock(&kvm_lock);
 739         list_add(&kvm->vm_list, &vm_list);
 740         mutex_unlock(&kvm_lock);
 741
 742         preempt_notifier_inc();
 743
 744         return kvm;
 745
 746 out_err:
 747 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 748         if (kvm->mmu_notifier.ops)
 749                 mmu_notifier_unregister(&kvm->mmu_notifier, current->mm);
 750 #endif
 751 out_err_no_mmu_notifier:
 752         hardware_disable_all();
 753 out_err_no_disable:
 754         kvm_arch_destroy_vm(kvm);
 755 out_err_no_arch_destroy_vm:
 756         WARN_ON_ONCE(!refcount_dec_and_test(&kvm->users_count));
 757         for (i = 0; i < KVM_NR_BUSES; i++)
 758                 kfree(kvm_get_bus(kvm, i));
 759         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 760                 kvm_free_memslots(kvm, __kvm_memslots(kvm, i));
 761         cleanup_srcu_struct(&kvm->irq_srcu);
 762 out_err_no_irq_srcu:
 763         cleanup_srcu_struct(&kvm->srcu);
 764 out_err_no_srcu:
 765         kvm_arch_free_vm(kvm);
 766         mmdrop(current->mm);
 767         return ERR_PTR(r);
 768 }
 769
 770 static void kvm_destroy_devices(struct kvm *kvm)
 771 {
 772         struct kvm_device *dev, *tmp;
 773
 774         /*
 775          * We do not need to take the kvm->lock here, because nobody else
 776          * has a reference to the struct kvm at this point and therefore
 777          * cannot access the devices list anyhow.
 778          */
 779         list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) {
 780                 list_del(&dev->vm_node);
 781                 dev->ops->destroy(dev);
 782         }
 783 }
 784
 785 static void kvm_destroy_vm(struct kvm *kvm)
 786 {
 787         int i;
 788         struct mm_struct *mm = kvm->mm;
 789
 790         kvm_uevent_notify_change(KVM_EVENT_DESTROY_VM, kvm);
 791         kvm_destroy_vm_debugfs(kvm);
 792         kvm_arch_sync_events(kvm);
 793         mutex_lock(&kvm_lock);
 794         list_del(&kvm->vm_list);
 795         mutex_unlock(&kvm_lock);
 796         kvm_arch_pre_destroy_vm(kvm);
 797
 798         kvm_free_irq_routing(kvm);
 799         for (i = 0; i < KVM_NR_BUSES; i++) {
 800                 struct kvm_io_bus *bus = kvm_get_bus(kvm, i);
 801
 802                 if (bus)
 803                         kvm_io_bus_destroy(bus);
 804                 kvm->buses[i] = NULL;
 805         }
 806         kvm_coalesced_mmio_free(kvm);
 807 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 808         mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 809 #else
 810         kvm_arch_flush_shadow_all(kvm);
 811 #endif
 812         kvm_arch_destroy_vm(kvm);
 813         kvm_destroy_devices(kvm);
 814         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 815                 kvm_free_memslots(kvm, __kvm_memslots(kvm, i));
 816         cleanup_srcu_struct(&kvm->irq_srcu);
 817         cleanup_srcu_struct(&kvm->srcu);
 818         kvm_arch_free_vm(kvm);
 819         preempt_notifier_dec();
 820         hardware_disable_all();
 821         mmdrop(mm);
 822 }
 823
 824 void kvm_get_kvm(struct kvm *kvm)
 825 {
 826         refcount_inc(&kvm->users_count);
 827 }
 828 EXPORT_SYMBOL_GPL(kvm_get_kvm);
 829
 830 void kvm_put_kvm(struct kvm *kvm)
 831 {
 832         if (refcount_dec_and_test(&kvm->users_count))
 833                 kvm_destroy_vm(kvm);
 834 }
 835 EXPORT_SYMBOL_GPL(kvm_put_kvm);
 836
 837 /*
 838  * Used to put a reference that was taken on behalf of an object associated
 839  * with a user-visible file descriptor, e.g. a vcpu or device, if installation
 840  * of the new file descriptor fails and the reference cannot be transferred to
 841  * its final owner.  In such cases, the caller is still actively using @kvm and
 842  * will fail miserably if the refcount unexpectedly hits zero.
 843  */
 844 void kvm_put_kvm_no_destroy(struct kvm *kvm)
 845 {
 846         WARN_ON(refcount_dec_and_test(&kvm->users_count));
 847 }
 848 EXPORT_SYMBOL_GPL(kvm_put_kvm_no_destroy);
 849
 850 static int kvm_vm_release(struct inode *inode, struct file *filp)
 851 {
 852         struct kvm *kvm = filp->private_data;
 853
 854         kvm_irqfd_release(kvm);
 855
 856         kvm_put_kvm(kvm);
 857         return 0;
 858 }
 859
 860 /*
 861  * Allocation size is twice as large as the actual dirty bitmap size.
 862  * See x86's kvm_vm_ioctl_get_dirty_log() why this is needed.
 863  */
 864 static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
 865 {
 866         unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
 867
 868         memslot->dirty_bitmap = kvzalloc(dirty_bytes, GFP_KERNEL_ACCOUNT);
 869         if (!memslot->dirty_bitmap)
 870                 return -ENOMEM;
 871
 872         return 0;
 873 }
 874
 875 /*
 876  * Insert memslot and re-sort memslots based on their GFN,
 877  * so binary search could be used to lookup GFN.
 878  * Sorting algorithm takes advantage of having initially
 879  * sorted array and known changed memslot position.
 880  */
 881 static void update_memslots(struct kvm_memslots *slots,
 882                             struct kvm_memory_slot *new,
 883                             enum kvm_mr_change change)
 884 {
 885         int id = new->id;
 886         int i = slots->id_to_index[id];
 887         struct kvm_memory_slot *mslots = slots->memslots;
 888
 889         WARN_ON(mslots[i].id != id);
 890         switch (change) {
 891         case KVM_MR_CREATE:
 892                 slots->used_slots++;
 893                 WARN_ON(mslots[i].npages || !new->npages);
 894                 break;
 895         case KVM_MR_DELETE:
 896                 slots->used_slots--;
 897                 WARN_ON(new->npages || !mslots[i].npages);
 898                 break;
 899         default:
 900                 break;
 901         }
 902
 903         while (i < KVM_MEM_SLOTS_NUM - 1 &&
 904                new->base_gfn <= mslots[i + 1].base_gfn) {
 905                 if (!mslots[i + 1].npages)
 906                         break;
 907                 mslots[i] = mslots[i + 1];
 908                 slots->id_to_index[mslots[i].id] = i;
 909                 i++;
 910         }
 911
 912         /*
 913          * The ">=" is needed when creating a slot with base_gfn == 0,
 914          * so that it moves before all those with base_gfn == npages == 0.
 915          *
 916          * On the other hand, if new->npages is zero, the above loop has
 917          * already left i pointing to the beginning of the empty part of
 918          * mslots, and the ">=" would move the hole backwards in this
 919          * case---which is wrong.  So skip the loop when deleting a slot.
 920          */
 921         if (new->npages) {
 922                 while (i > 0 &&
 923                        new->base_gfn >= mslots[i - 1].base_gfn) {
 924                         mslots[i] = mslots[i - 1];
 925                         slots->id_to_index[mslots[i].id] = i;
 926                         i--;
 927                 }
 928         } else
 929                 WARN_ON_ONCE(i != slots->used_slots);
 930
 931         mslots[i] = *new;
 932         slots->id_to_index[mslots[i].id] = i;
 933 }
 934
 935 static int check_memory_region_flags(const struct kvm_userspace_memory_region *mem)
 936 {
 937         u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES;
 938
 939 #ifdef __KVM_HAVE_READONLY_MEM
 940         valid_flags |= KVM_MEM_READONLY;
 941 #endif
 942
 943         if (mem->flags & ~valid_flags)
 944                 return -EINVAL;
 945
 946         return 0;
 947 }
 948
 949 static struct kvm_memslots *install_new_memslots(struct kvm *kvm,
 950                 int as_id, struct kvm_memslots *slots)
 951 {
 952         struct kvm_memslots *old_memslots = __kvm_memslots(kvm, as_id);
 953         u64 gen = old_memslots->generation;
 954
 955         WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
 956         slots->generation = gen | KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS;
 957
 958         rcu_assign_pointer(kvm->memslots[as_id], slots);
 959         synchronize_srcu_expedited(&kvm->srcu);
 960
 961         /*
 962          * Increment the new memslot generation a second time, dropping the
 963          * update in-progress flag and incrementing the generation based on
 964          * the number of address spaces.  This provides a unique and easily
 965          * identifiable generation number while the memslots are in flux.
 966          */
 967         gen = slots->generation & ~KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS;
 968
 969         /*
 970          * Generations must be unique even across address spaces.  We do not need
 971          * a global counter for that, instead the generation space is evenly split
 972          * across address spaces.  For example, with two address spaces, address
 973          * space 0 will use generations 0, 2, 4, ... while address space 1 will
 974          * use generations 1, 3, 5, ...
 975          */
 976         gen += KVM_ADDRESS_SPACE_NUM;
 977
 978         kvm_arch_memslots_updated(kvm, gen);
 979
 980         slots->generation = gen;
 981
 982         return old_memslots;
 983 }
 984
 985 /*
 986  * Allocate some memory and give it an address in the guest physical address
 987  * space.
 988  *
 989  * Discontiguous memory is allowed, mostly for framebuffers.
 990  *
 991  * Must be called holding kvm->slots_lock for write.
 992  */
 993 int __kvm_set_memory_region(struct kvm *kvm,
 994                             const struct kvm_userspace_memory_region *mem)
 995 {
 996         int r;
 997         gfn_t base_gfn;
 998         unsigned long npages;
 999         struct kvm_memory_slot *slot;
1000         struct kvm_memory_slot old, new;
1001         struct kvm_memslots *slots = NULL, *old_memslots;
1002         int as_id, id;
1003         enum kvm_mr_change change;
1004
1005         r = check_memory_region_flags(mem);
1006         if (r)
1007                 goto out;
1008
1009         r = -EINVAL;
1010         as_id = mem->slot >> 16;
1011         id = (u16)mem->slot;
1012
1013         /* General sanity checks */
1014         if (mem->memory_size & (PAGE_SIZE - 1))
1015                 goto out;
1016         if (mem->guest_phys_addr & (PAGE_SIZE - 1))
1017                 goto out;
1018         /* We can read the guest memory with __xxx_user() later on. */
1019         if ((id < KVM_USER_MEM_SLOTS) &&
1020             ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
1021              !access_ok((void __user *)(unsigned long)mem->userspace_addr,
1022                         mem->memory_size)))
1023                 goto out;
1024         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_MEM_SLOTS_NUM)
1025                 goto out;
1026         if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
1027                 goto out;
1028
1029         slot = id_to_memslot(__kvm_memslots(kvm, as_id), id);
1030         base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
1031         npages = mem->memory_size >> PAGE_SHIFT;
1032
1033         if (npages > KVM_MEM_MAX_NR_PAGES)
1034                 goto out;
1035
1036         new = old = *slot;
1037
1038         new.id = id;
1039         new.base_gfn = base_gfn;
1040         new.npages = npages;
1041         new.flags = mem->flags;
1042
1043         if (npages) {
1044                 if (!old.npages)
1045                         change = KVM_MR_CREATE;
1046                 else { /* Modify an existing slot. */
1047                         if ((mem->userspace_addr != old.userspace_addr) ||
1048                             (npages != old.npages) ||
1049                             ((new.flags ^ old.flags) & KVM_MEM_READONLY))
1050                                 goto out;
1051
1052                         if (base_gfn != old.base_gfn)
1053                                 change = KVM_MR_MOVE;
1054                         else if (new.flags != old.flags)
1055                                 change = KVM_MR_FLAGS_ONLY;
1056                         else { /* Nothing to change. */
1057                                 r = 0;
1058                                 goto out;
1059                         }
1060                 }
1061         } else {
1062                 if (!old.npages)
1063                         goto out;
1064
1065                 change = KVM_MR_DELETE;
1066                 new.base_gfn = 0;
1067                 new.flags = 0;
1068         }
1069
1070         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
1071                 /* Check for overlaps */
1072                 r = -EEXIST;
1073                 kvm_for_each_memslot(slot, __kvm_memslots(kvm, as_id)) {
1074                         if (slot->id == id)
1075                                 continue;
1076                         if (!((base_gfn + npages <= slot->base_gfn) ||
1077                               (base_gfn >= slot->base_gfn + slot->npages)))
1078                                 goto out;
1079                 }
1080         }
1081
1082         /* Free page dirty bitmap if unneeded */
1083         if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
1084                 new.dirty_bitmap = NULL;
1085
1086         r = -ENOMEM;
1087         if (change == KVM_MR_CREATE) {
1088                 new.userspace_addr = mem->userspace_addr;
1089
1090                 if (kvm_arch_create_memslot(kvm, &new, npages))
1091                         goto out_free;
1092         }
1093
1094         /* Allocate page dirty bitmap if needed */
1095         if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
1096                 if (kvm_create_dirty_bitmap(&new) < 0)
1097                         goto out_free;
1098         }
1099
1100         slots = kvzalloc(sizeof(struct kvm_memslots), GFP_KERNEL_ACCOUNT);
1101         if (!slots)
1102                 goto out_free;
1103         memcpy(slots, __kvm_memslots(kvm, as_id), sizeof(struct kvm_memslots));
1104
1105         if ((change == KVM_MR_DELETE) || (change == KVM_MR_MOVE)) {
1106                 slot = id_to_memslot(slots, id);
1107                 slot->flags |= KVM_MEMSLOT_INVALID;
1108
1109                 old_memslots = install_new_memslots(kvm, as_id, slots);
1110
1111                 /* From this point no new shadow pages pointing to a deleted,
1112                  * or moved, memslot will be created.
1113                  *
1114                  * validation of sp->gfn happens in:
1115                  *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
1116                  *      - kvm_is_visible_gfn (mmu_check_root)
1117                  */
1118                 kvm_arch_flush_shadow_memslot(kvm, slot);
1119
1120                 /*
1121                  * We can re-use the old_memslots from above, the only difference
1122                  * from the currently installed memslots is the invalid flag.  This
1123                  * will get overwritten by update_memslots anyway.
1124                  */
1125                 slots = old_memslots;
1126         }
1127
1128         r = kvm_arch_prepare_memory_region(kvm, &new, mem, change);
1129         if (r)
1130                 goto out_slots;
1131
1132         /* actual memory is freed via old in kvm_free_memslot below */
1133         if (change == KVM_MR_DELETE) {
1134                 new.dirty_bitmap = NULL;
1135                 memset(&new.arch, 0, sizeof(new.arch));
1136         }
1137
1138         update_memslots(slots, &new, change);
1139         old_memslots = install_new_memslots(kvm, as_id, slots);
1140
1141         kvm_arch_commit_memory_region(kvm, mem, &old, &new, change);
1142
1143         kvm_free_memslot(kvm, &old, &new);
1144         kvfree(old_memslots);
1145         return 0;
1146
1147 out_slots:
1148         kvfree(slots);
1149 out_free:
1150         kvm_free_memslot(kvm, &new, &old);
1151 out:
1152         return r;
1153 }
1154 EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
1155
1156 int kvm_set_memory_region(struct kvm *kvm,
1157                           const struct kvm_userspace_memory_region *mem)
1158 {
1159         int r;
1160
1161         mutex_lock(&kvm->slots_lock);
1162         r = __kvm_set_memory_region(kvm, mem);
1163         mutex_unlock(&kvm->slots_lock);
1164         return r;
1165 }
1166 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
1167
1168 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
1169                                           struct kvm_userspace_memory_region *mem)
1170 {
1171         if ((u16)mem->slot >= KVM_USER_MEM_SLOTS)
1172                 return -EINVAL;
1173
1174         return kvm_set_memory_region(kvm, mem);
1175 }
1176
1177 int kvm_get_dirty_log(struct kvm *kvm,
1178                         struct kvm_dirty_log *log, int *is_dirty)
1179 {
1180         struct kvm_memslots *slots;
1181         struct kvm_memory_slot *memslot;
1182         int i, as_id, id;
1183         unsigned long n;
1184         unsigned long any = 0;
1185
1186         as_id = log->slot >> 16;
1187         id = (u16)log->slot;
1188         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1189                 return -EINVAL;
1190
1191         slots = __kvm_memslots(kvm, as_id);
1192         memslot = id_to_memslot(slots, id);
1193         if (!memslot->dirty_bitmap)
1194                 return -ENOENT;
1195
1196         n = kvm_dirty_bitmap_bytes(memslot);
1197
1198         for (i = 0; !any && i < n/sizeof(long); ++i)
1199                 any = memslot->dirty_bitmap[i];
1200
1201         if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
1202                 return -EFAULT;
1203
1204         if (any)
1205                 *is_dirty = 1;
1206         return 0;
1207 }
1208 EXPORT_SYMBOL_GPL(kvm_get_dirty_log);
1209
1210 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
1211 /**
1212  * kvm_get_dirty_log_protect - get a snapshot of dirty pages
1213  *      and reenable dirty page tracking for the corresponding pages.
1214  * @kvm:        pointer to kvm instance
1215  * @log:        slot id and address to which we copy the log
1216  * @flush:      true if TLB flush is needed by caller
1217  *
1218  * We need to keep it in mind that VCPU threads can write to the bitmap
1219  * concurrently. So, to avoid losing track of dirty pages we keep the
1220  * following order:
1221  *
1222  *    1. Take a snapshot of the bit and clear it if needed.
1223  *    2. Write protect the corresponding page.
1224  *    3. Copy the snapshot to the userspace.
1225  *    4. Upon return caller flushes TLB's if needed.
1226  *
1227  * Between 2 and 4, the guest may write to the page using the remaining TLB
1228  * entry.  This is not a problem because the page is reported dirty using
1229  * the snapshot taken before and step 4 ensures that writes done after
1230  * exiting to userspace will be logged for the next call.
1231  *
1232  */
1233 int kvm_get_dirty_log_protect(struct kvm *kvm,
1234                         struct kvm_dirty_log *log, bool *flush)
1235 {
1236         struct kvm_memslots *slots;
1237         struct kvm_memory_slot *memslot;
1238         int i, as_id, id;
1239         unsigned long n;
1240         unsigned long *dirty_bitmap;
1241         unsigned long *dirty_bitmap_buffer;
1242
1243         as_id = log->slot >> 16;
1244         id = (u16)log->slot;
1245         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1246                 return -EINVAL;
1247
1248         slots = __kvm_memslots(kvm, as_id);
1249         memslot = id_to_memslot(slots, id);
1250
1251         dirty_bitmap = memslot->dirty_bitmap;
1252         if (!dirty_bitmap)
1253                 return -ENOENT;
1254
1255         n = kvm_dirty_bitmap_bytes(memslot);
1256         *flush = false;
1257         if (kvm->manual_dirty_log_protect) {
1258                 /*
1259                  * Unlike kvm_get_dirty_log, we always return false in *flush,
1260                  * because no flush is needed until KVM_CLEAR_DIRTY_LOG.  There
1261                  * is some code duplication between this function and
1262                  * kvm_get_dirty_log, but hopefully all architecture
1263                  * transition to kvm_get_dirty_log_protect and kvm_get_dirty_log
1264                  * can be eliminated.
1265                  */
1266                 dirty_bitmap_buffer = dirty_bitmap;
1267         } else {
1268                 dirty_bitmap_buffer = kvm_second_dirty_bitmap(memslot);
1269                 memset(dirty_bitmap_buffer, 0, n);
1270
1271                 spin_lock(&kvm->mmu_lock);
1272                 for (i = 0; i < n / sizeof(long); i++) {
1273                         unsigned long mask;
1274                         gfn_t offset;
1275
1276                         if (!dirty_bitmap[i])
1277                                 continue;
1278
1279                         *flush = true;
1280                         mask = xchg(&dirty_bitmap[i], 0);
1281                         dirty_bitmap_buffer[i] = mask;
1282
1283                         offset = i * BITS_PER_LONG;
1284                         kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot,
1285                                                                 offset, mask);
1286                 }
1287                 spin_unlock(&kvm->mmu_lock);
1288         }
1289
1290         if (copy_to_user(log->dirty_bitmap, dirty_bitmap_buffer, n))
1291                 return -EFAULT;
1292         return 0;
1293 }
1294 EXPORT_SYMBOL_GPL(kvm_get_dirty_log_protect);
1295
1296 /**
1297  * kvm_clear_dirty_log_protect - clear dirty bits in the bitmap
1298  *      and reenable dirty page tracking for the corresponding pages.
1299  * @kvm:        pointer to kvm instance
1300  * @log:        slot id and address from which to fetch the bitmap of dirty pages
1301  * @flush:      true if TLB flush is needed by caller
1302  */
1303 int kvm_clear_dirty_log_protect(struct kvm *kvm,
1304                                 struct kvm_clear_dirty_log *log, bool *flush)
1305 {
1306         struct kvm_memslots *slots;
1307         struct kvm_memory_slot *memslot;
1308         int as_id, id;
1309         gfn_t offset;
1310         unsigned long i, n;
1311         unsigned long *dirty_bitmap;
1312         unsigned long *dirty_bitmap_buffer;
1313
1314         as_id = log->slot >> 16;
1315         id = (u16)log->slot;
1316         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1317                 return -EINVAL;
1318
1319         if (log->first_page & 63)
1320                 return -EINVAL;
1321
1322         slots = __kvm_memslots(kvm, as_id);
1323         memslot = id_to_memslot(slots, id);
1324
1325         dirty_bitmap = memslot->dirty_bitmap;
1326         if (!dirty_bitmap)
1327                 return -ENOENT;
1328
1329         n = ALIGN(log->num_pages, BITS_PER_LONG) / 8;
1330
1331         if (log->first_page > memslot->npages ||
1332             log->num_pages > memslot->npages - log->first_page ||
1333             (log->num_pages < memslot->npages - log->first_page && (log->num_pages & 63)))
1334             return -EINVAL;
1335
1336         *flush = false;
1337         dirty_bitmap_buffer = kvm_second_dirty_bitmap(memslot);
1338         if (copy_from_user(dirty_bitmap_buffer, log->dirty_bitmap, n))
1339                 return -EFAULT;
1340
1341         spin_lock(&kvm->mmu_lock);
1342         for (offset = log->first_page, i = offset / BITS_PER_LONG,
1343                  n = DIV_ROUND_UP(log->num_pages, BITS_PER_LONG); n--;
1344              i++, offset += BITS_PER_LONG) {
1345                 unsigned long mask = *dirty_bitmap_buffer++;
1346                 atomic_long_t *p = (atomic_long_t *) &dirty_bitmap[i];
1347                 if (!mask)
1348                         continue;
1349
1350                 mask &= atomic_long_fetch_andnot(mask, p);
1351
1352                 /*
1353                  * mask contains the bits that really have been cleared.  This
1354                  * never includes any bits beyond the length of the memslot (if
1355                  * the length is not aligned to 64 pages), therefore it is not
1356                  * a problem if userspace sets them in log->dirty_bitmap.
1357                 */
1358                 if (mask) {
1359                         *flush = true;
1360                         kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot,
1361                                                                 offset, mask);
1362                 }
1363         }
1364         spin_unlock(&kvm->mmu_lock);
1365
1366         return 0;
1367 }
1368 EXPORT_SYMBOL_GPL(kvm_clear_dirty_log_protect);
1369 #endif
1370
1371 bool kvm_largepages_enabled(void)
1372 {
1373         return largepages_enabled;
1374 }
1375
1376 void kvm_disable_largepages(void)
1377 {
1378         largepages_enabled = false;
1379 }
1380 EXPORT_SYMBOL_GPL(kvm_disable_largepages);
1381
1382 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
1383 {
1384         return __gfn_to_memslot(kvm_memslots(kvm), gfn);
1385 }
1386 EXPORT_SYMBOL_GPL(gfn_to_memslot);
1387
1388 struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn)
1389 {
1390         return __gfn_to_memslot(kvm_vcpu_memslots(vcpu), gfn);
1391 }
1392
1393 bool kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
1394 {
1395         struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
1396
1397         if (!memslot || memslot->id >= KVM_USER_MEM_SLOTS ||
1398               memslot->flags & KVM_MEMSLOT_INVALID)
1399                 return false;
1400
1401         return true;
1402 }
1403 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
1404
1405 unsigned long kvm_host_page_size(struct kvm_vcpu *vcpu, gfn_t gfn)
1406 {
1407         struct vm_area_struct *vma;
1408         unsigned long addr, size;
1409
1410         size = PAGE_SIZE;
1411
1412         addr = kvm_vcpu_gfn_to_hva_prot(vcpu, gfn, NULL);
1413         if (kvm_is_error_hva(addr))
1414                 return PAGE_SIZE;
1415
1416         down_read(&current->mm->mmap_sem);
1417         vma = find_vma(current->mm, addr);
1418         if (!vma)
1419                 goto out;
1420
1421         size = vma_kernel_pagesize(vma);
1422
1423 out:
1424         up_read(&current->mm->mmap_sem);
1425
1426         return size;
1427 }
1428
1429 static bool memslot_is_readonly(struct kvm_memory_slot *slot)
1430 {
1431         return slot->flags & KVM_MEM_READONLY;
1432 }
1433
1434 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1435                                        gfn_t *nr_pages, bool write)
1436 {
1437         if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1438                 return KVM_HVA_ERR_BAD;
1439
1440         if (memslot_is_readonly(slot) && write)
1441                 return KVM_HVA_ERR_RO_BAD;
1442
1443         if (nr_pages)
1444                 *nr_pages = slot->npages - (gfn - slot->base_gfn);
1445
1446         return __gfn_to_hva_memslot(slot, gfn);
1447 }
1448
1449 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1450                                      gfn_t *nr_pages)
1451 {
1452         return __gfn_to_hva_many(slot, gfn, nr_pages, true);
1453 }
1454
1455 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot,
1456                                         gfn_t gfn)
1457 {
1458         return gfn_to_hva_many(slot, gfn, NULL);
1459 }
1460 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot);
1461
1462 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
1463 {
1464         return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
1465 }
1466 EXPORT_SYMBOL_GPL(gfn_to_hva);
1467
1468 unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn)
1469 {
1470         return gfn_to_hva_many(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, NULL);
1471 }
1472 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_hva);
1473
1474 /*
1475  * Return the hva of a @gfn and the R/W attribute if possible.
1476  *
1477  * @slot: the kvm_memory_slot which contains @gfn
1478  * @gfn: the gfn to be translated
1479  * @writable: used to return the read/write attribute of the @slot if the hva
1480  * is valid and @writable is not NULL
1481  */
1482 unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot,
1483                                       gfn_t gfn, bool *writable)
1484 {
1485         unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false);
1486
1487         if (!kvm_is_error_hva(hva) && writable)
1488                 *writable = !memslot_is_readonly(slot);
1489
1490         return hva;
1491 }
1492
1493 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable)
1494 {
1495         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1496
1497         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1498 }
1499
1500 unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable)
1501 {
1502         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1503
1504         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1505 }
1506
1507 static inline int check_user_page_hwpoison(unsigned long addr)
1508 {
1509         int rc, flags = FOLL_HWPOISON | FOLL_WRITE;
1510
1511         rc = get_user_pages(addr, 1, flags, NULL, NULL);
1512         return rc == -EHWPOISON;
1513 }
1514
1515 /*
1516  * The fast path to get the writable pfn which will be stored in @pfn,
1517  * true indicates success, otherwise false is returned.  It's also the
1518  * only part that runs if we can in atomic context.
1519  */
1520 static bool hva_to_pfn_fast(unsigned long addr, bool write_fault,
1521                             bool *writable, kvm_pfn_t *pfn)
1522 {
1523         struct page *page[1];
1524         int npages;
1525
1526         /*
1527          * Fast pin a writable pfn only if it is a write fault request
1528          * or the caller allows to map a writable pfn for a read fault
1529          * request.
1530          */
1531         if (!(write_fault || writable))
1532                 return false;
1533
1534         npages = __get_user_pages_fast(addr, 1, 1, page);
1535         if (npages == 1) {
1536                 *pfn = page_to_pfn(page[0]);
1537
1538                 if (writable)
1539                         *writable = true;
1540                 return true;
1541         }
1542
1543         return false;
1544 }
1545
1546 /*
1547  * The slow path to get the pfn of the specified host virtual address,
1548  * 1 indicates success, -errno is returned if error is detected.
1549  */
1550 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault,
1551                            bool *writable, kvm_pfn_t *pfn)
1552 {
1553         unsigned int flags = FOLL_HWPOISON;
1554         struct page *page;
1555         int npages = 0;
1556
1557         might_sleep();
1558
1559         if (writable)
1560                 *writable = write_fault;
1561
1562         if (write_fault)
1563                 flags |= FOLL_WRITE;
1564         if (async)
1565                 flags |= FOLL_NOWAIT;
1566
1567         npages = get_user_pages_unlocked(addr, 1, &page, flags);
1568         if (npages != 1)
1569                 return npages;
1570
1571         /* map read fault as writable if possible */
1572         if (unlikely(!write_fault) && writable) {
1573                 struct page *wpage;
1574
1575                 if (__get_user_pages_fast(addr, 1, 1, &wpage) == 1) {
1576                         *writable = true;
1577                         put_page(page);
1578                         page = wpage;
1579                 }
1580         }
1581         *pfn = page_to_pfn(page);
1582         return npages;
1583 }
1584
1585 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault)
1586 {
1587         if (unlikely(!(vma->vm_flags & VM_READ)))
1588                 return false;
1589
1590         if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE))))
1591                 return false;
1592
1593         return true;
1594 }
1595
1596 static int hva_to_pfn_remapped(struct vm_area_struct *vma,
1597                                unsigned long addr, bool *async,
1598                                bool write_fault, bool *writable,
1599                                kvm_pfn_t *p_pfn)
1600 {
1601         unsigned long pfn;
1602         int r;
1603
1604         r = follow_pfn(vma, addr, &pfn);
1605         if (r) {
1606                 /*
1607                  * get_user_pages fails for VM_IO and VM_PFNMAP vmas and does
1608                  * not call the fault handler, so do it here.
1609                  */
1610                 bool unlocked = false;
1611                 r = fixup_user_fault(current, current->mm, addr,
1612                                      (write_fault ? FAULT_FLAG_WRITE : 0),
1613                                      &unlocked);
1614                 if (unlocked)
1615                         return -EAGAIN;
1616                 if (r)
1617                         return r;
1618
1619                 r = follow_pfn(vma, addr, &pfn);
1620                 if (r)
1621                         return r;
1622
1623         }
1624
1625         if (writable)
1626                 *writable = true;
1627
1628         /*
1629          * Get a reference here because callers of *hva_to_pfn* and
1630          * *gfn_to_pfn* ultimately call kvm_release_pfn_clean on the
1631          * returned pfn.  This is only needed if the VMA has VM_MIXEDMAP
1632          * set, but the kvm_get_pfn/kvm_release_pfn_clean pair will
1633          * simply do nothing for reserved pfns.
1634          *
1635          * Whoever called remap_pfn_range is also going to call e.g.
1636          * unmap_mapping_range before the underlying pages are freed,
1637          * causing a call to our MMU notifier.
1638          */
1639         kvm_get_pfn(pfn);
1640
1641         *p_pfn = pfn;
1642         return 0;
1643 }
1644
1645 /*
1646  * Pin guest page in memory and return its pfn.
1647  * @addr: host virtual address which maps memory to the guest
1648  * @atomic: whether this function can sleep
1649  * @async: whether this function need to wait IO complete if the
1650  *         host page is not in the memory
1651  * @write_fault: whether we should get a writable host page
1652  * @writable: whether it allows to map a writable host page for !@write_fault
1653  *
1654  * The function will map a writable host page for these two cases:
1655  * 1): @write_fault = true
1656  * 2): @write_fault = false && @writable, @writable will tell the caller
1657  *     whether the mapping is writable.
1658  */
1659 static kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
1660                         bool write_fault, bool *writable)
1661 {
1662         struct vm_area_struct *vma;
1663         kvm_pfn_t pfn = 0;
1664         int npages, r;
1665
1666         /* we can do it either atomically or asynchronously, not both */
1667         BUG_ON(atomic && async);
1668
1669         if (hva_to_pfn_fast(addr, write_fault, writable, &pfn))
1670                 return pfn;
1671
1672         if (atomic)
1673                 return KVM_PFN_ERR_FAULT;
1674
1675         npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn);
1676         if (npages == 1)
1677                 return pfn;
1678
1679         down_read(&current->mm->mmap_sem);
1680         if (npages == -EHWPOISON ||
1681               (!async && check_user_page_hwpoison(addr))) {
1682                 pfn = KVM_PFN_ERR_HWPOISON;
1683                 goto exit;
1684         }
1685
1686 retry:
1687         vma = find_vma_intersection(current->mm, addr, addr + 1);
1688
1689         if (vma == NULL)
1690                 pfn = KVM_PFN_ERR_FAULT;
1691         else if (vma->vm_flags & (VM_IO | VM_PFNMAP)) {
1692                 r = hva_to_pfn_remapped(vma, addr, async, write_fault, writable, &pfn);
1693                 if (r == -EAGAIN)
1694                         goto retry;
1695                 if (r < 0)
1696                         pfn = KVM_PFN_ERR_FAULT;
1697         } else {
1698                 if (async && vma_is_valid(vma, write_fault))
1699                         *async = true;
1700                 pfn = KVM_PFN_ERR_FAULT;
1701         }
1702 exit:
1703         up_read(&current->mm->mmap_sem);
1704         return pfn;
1705 }
1706
1707 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn,
1708                                bool atomic, bool *async, bool write_fault,
1709                                bool *writable)
1710 {
1711         unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
1712
1713         if (addr == KVM_HVA_ERR_RO_BAD) {
1714                 if (writable)
1715                         *writable = false;
1716                 return KVM_PFN_ERR_RO_FAULT;
1717         }
1718
1719         if (kvm_is_error_hva(addr)) {
1720                 if (writable)
1721                         *writable = false;
1722                 return KVM_PFN_NOSLOT;
1723         }
1724
1725         /* Do not map writable pfn in the readonly memslot. */
1726         if (writable && memslot_is_readonly(slot)) {
1727                 *writable = false;
1728                 writable = NULL;
1729         }
1730
1731         return hva_to_pfn(addr, atomic, async, write_fault,
1732                           writable);
1733 }
1734 EXPORT_SYMBOL_GPL(__gfn_to_pfn_memslot);
1735
1736 kvm_pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
1737                       bool *writable)
1738 {
1739         return __gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn, false, NULL,
1740                                     write_fault, writable);
1741 }
1742 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
1743
1744 kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
1745 {
1746         return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
1747 }
1748 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot);
1749
1750 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
1751 {
1752         return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
1753 }
1754 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
1755
1756 kvm_pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
1757 {
1758         return gfn_to_pfn_memslot_atomic(gfn_to_memslot(kvm, gfn), gfn);
1759 }
1760 EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
1761
1762 kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn)
1763 {
1764         return gfn_to_pfn_memslot_atomic(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1765 }
1766 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn_atomic);
1767
1768 kvm_pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
1769 {
1770         return gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn);
1771 }
1772 EXPORT_SYMBOL_GPL(gfn_to_pfn);
1773
1774 kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn)
1775 {
1776         return gfn_to_pfn_memslot(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1777 }
1778 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn);
1779
1780 int gfn_to_page_many_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
1781                             struct page **pages, int nr_pages)
1782 {
1783         unsigned long addr;
1784         gfn_t entry = 0;
1785
1786         addr = gfn_to_hva_many(slot, gfn, &entry);
1787         if (kvm_is_error_hva(addr))
1788                 return -1;
1789
1790         if (entry < nr_pages)
1791                 return 0;
1792
1793         return __get_user_pages_fast(addr, nr_pages, 1, pages);
1794 }
1795 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
1796
1797 static struct page *kvm_pfn_to_page(kvm_pfn_t pfn)
1798 {
1799         if (is_error_noslot_pfn(pfn))
1800                 return KVM_ERR_PTR_BAD_PAGE;
1801
1802         if (kvm_is_reserved_pfn(pfn)) {
1803                 WARN_ON(1);
1804                 return KVM_ERR_PTR_BAD_PAGE;
1805         }
1806
1807         return pfn_to_page(pfn);
1808 }
1809
1810 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
1811 {
1812         kvm_pfn_t pfn;
1813
1814         pfn = gfn_to_pfn(kvm, gfn);
1815
1816         return kvm_pfn_to_page(pfn);
1817 }
1818 EXPORT_SYMBOL_GPL(gfn_to_page);
1819
1820 void kvm_release_pfn(kvm_pfn_t pfn, bool dirty, struct gfn_to_pfn_cache *cache)
1821 {
1822         if (pfn == 0)
1823                 return;
1824
1825         if (cache)
1826                 cache->pfn = cache->gfn = 0;
1827
1828         if (dirty)
1829                 kvm_release_pfn_dirty(pfn);
1830         else
1831                 kvm_release_pfn_clean(pfn);
1832 }
1833
1834 static void kvm_cache_gfn_to_pfn(struct kvm_memory_slot *slot, gfn_t gfn,
1835                                  struct gfn_to_pfn_cache *cache, u64 gen)
1836 {
1837         kvm_release_pfn(cache->pfn, cache->dirty, cache);
1838
1839         cache->pfn = gfn_to_pfn_memslot(slot, gfn);
1840         cache->gfn = gfn;
1841         cache->dirty = false;
1842         cache->generation = gen;
1843 }
1844
1845 static int __kvm_map_gfn(struct kvm_memslots *slots, gfn_t gfn,
1846                          struct kvm_host_map *map,
1847                          struct gfn_to_pfn_cache *cache,
1848                          bool atomic)
1849 {
1850         kvm_pfn_t pfn;
1851         void *hva = NULL;
1852         struct page *page = KVM_UNMAPPED_PAGE;
1853         struct kvm_memory_slot *slot = __gfn_to_memslot(slots, gfn);
1854         u64 gen = slots->generation;
1855
1856         if (!map)
1857                 return -EINVAL;
1858
1859         if (cache) {
1860                 if (!cache->pfn || cache->gfn != gfn ||
1861                         cache->generation != gen) {
1862                         if (atomic)
1863                                 return -EAGAIN;
1864                         kvm_cache_gfn_to_pfn(slot, gfn, cache, gen);
1865                 }
1866                 pfn = cache->pfn;
1867         } else {
1868                 if (atomic)
1869                         return -EAGAIN;
1870                 pfn = gfn_to_pfn_memslot(slot, gfn);
1871         }
1872         if (is_error_noslot_pfn(pfn))
1873                 return -EINVAL;
1874
1875         if (pfn_valid(pfn)) {
1876                 page = pfn_to_page(pfn);
1877                 if (atomic)
1878                         hva = kmap_atomic(page);
1879                 else
1880                         hva = kmap(page);
1881 #ifdef CONFIG_HAS_IOMEM
1882         } else if (!atomic) {
1883                 hva = memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB);
1884         } else {
1885                 return -EINVAL;
1886 #endif
1887         }
1888
1889         if (!hva)
1890                 return -EFAULT;
1891
1892         map->page = page;
1893         map->hva = hva;
1894         map->pfn = pfn;
1895         map->gfn = gfn;
1896
1897         return 0;
1898 }
1899
1900 int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map,
1901                 struct gfn_to_pfn_cache *cache, bool atomic)
1902 {
1903         return __kvm_map_gfn(kvm_memslots(vcpu->kvm), gfn, map,
1904                         cache, atomic);
1905 }
1906 EXPORT_SYMBOL_GPL(kvm_map_gfn);
1907
1908 int kvm_vcpu_map(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
1909 {
1910         return __kvm_map_gfn(kvm_vcpu_memslots(vcpu), gfn, map,
1911                 NULL, false);
1912 }
1913 EXPORT_SYMBOL_GPL(kvm_vcpu_map);
1914
1915 static void __kvm_unmap_gfn(struct kvm_memory_slot *memslot,
1916                         struct kvm_host_map *map,
1917                         struct gfn_to_pfn_cache *cache,
1918                         bool dirty, bool atomic)
1919 {
1920         if (!map)
1921                 return;
1922
1923         if (!map->hva)
1924                 return;
1925
1926         if (map->page != KVM_UNMAPPED_PAGE) {
1927                 if (atomic)
1928                         kunmap_atomic(map->hva);
1929                 else
1930                         kunmap(map->page);
1931         }
1932 #ifdef CONFIG_HAS_IOMEM
1933         else if (!atomic)
1934                 memunmap(map->hva);
1935         else
1936                 WARN_ONCE(1, "Unexpected unmapping in atomic context");
1937 #endif
1938
1939         if (dirty)
1940                 mark_page_dirty_in_slot(memslot, map->gfn);
1941
1942         if (cache)
1943                 cache->dirty |= dirty;
1944         else
1945                 kvm_release_pfn(map->pfn, dirty, NULL);
1946
1947         map->hva = NULL;
1948         map->page = NULL;
1949 }
1950
1951 int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
1952                   struct gfn_to_pfn_cache *cache, bool dirty, bool atomic)
1953 {
1954         __kvm_unmap_gfn(gfn_to_memslot(vcpu->kvm, map->gfn), map,
1955                         cache, dirty, atomic);
1956         return 0;
1957 }
1958 EXPORT_SYMBOL_GPL(kvm_unmap_gfn);
1959
1960 void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty)
1961 {
1962         __kvm_unmap_gfn(kvm_vcpu_gfn_to_memslot(vcpu, map->gfn), map, NULL,
1963                         dirty, false);
1964 }
1965 EXPORT_SYMBOL_GPL(kvm_vcpu_unmap);
1966
1967 struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1968 {
1969         kvm_pfn_t pfn;
1970
1971         pfn = kvm_vcpu_gfn_to_pfn(vcpu, gfn);
1972
1973         return kvm_pfn_to_page(pfn);
1974 }
1975 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_page);
1976
1977 void kvm_release_page_clean(struct page *page)
1978 {
1979         WARN_ON(is_error_page(page));
1980
1981         kvm_release_pfn_clean(page_to_pfn(page));
1982 }
1983 EXPORT_SYMBOL_GPL(kvm_release_page_clean);
1984
1985 void kvm_release_pfn_clean(kvm_pfn_t pfn)
1986 {
1987         if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn))
1988                 put_page(pfn_to_page(pfn));
1989 }
1990 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
1991
1992 void kvm_release_page_dirty(struct page *page)
1993 {
1994         WARN_ON(is_error_page(page));
1995
1996         kvm_release_pfn_dirty(page_to_pfn(page));
1997 }
1998 EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
1999
2000 void kvm_release_pfn_dirty(kvm_pfn_t pfn)
2001 {
2002         kvm_set_pfn_dirty(pfn);
2003         kvm_release_pfn_clean(pfn);
2004 }
2005 EXPORT_SYMBOL_GPL(kvm_release_pfn_dirty);
2006
2007 void kvm_set_pfn_dirty(kvm_pfn_t pfn)
2008 {
2009         if (!kvm_is_reserved_pfn(pfn) && !kvm_is_zone_device_pfn(pfn))
2010                 SetPageDirty(pfn_to_page(pfn));
2011 }
2012 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
2013
2014 void kvm_set_pfn_accessed(kvm_pfn_t pfn)
2015 {
2016         if (!kvm_is_reserved_pfn(pfn) && !kvm_is_zone_device_pfn(pfn))
2017                 mark_page_accessed(pfn_to_page(pfn));
2018 }
2019 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
2020
2021 void kvm_get_pfn(kvm_pfn_t pfn)
2022 {
2023         if (!kvm_is_reserved_pfn(pfn))
2024                 get_page(pfn_to_page(pfn));
2025 }
2026 EXPORT_SYMBOL_GPL(kvm_get_pfn);
2027
2028 static int next_segment(unsigned long len, int offset)
2029 {
2030         if (len > PAGE_SIZE - offset)
2031                 return PAGE_SIZE - offset;
2032         else
2033                 return len;
2034 }
2035
2036 static int __kvm_read_guest_page(struct kvm_memory_slot *slot, gfn_t gfn,
2037                                  void *data, int offset, int len)
2038 {
2039         int r;
2040         unsigned long addr;
2041
2042         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
2043         if (kvm_is_error_hva(addr))
2044                 return -EFAULT;
2045         r = __copy_from_user(data, (void __user *)addr + offset, len);
2046         if (r)
2047                 return -EFAULT;
2048         return 0;
2049 }
2050
2051 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
2052                         int len)
2053 {
2054         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
2055
2056         return __kvm_read_guest_page(slot, gfn, data, offset, len);
2057 }
2058 EXPORT_SYMBOL_GPL(kvm_read_guest_page);
2059
2060 int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data,
2061                              int offset, int len)
2062 {
2063         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2064
2065         return __kvm_read_guest_page(slot, gfn, data, offset, len);
2066 }
2067 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_page);
2068
2069 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
2070 {
2071         gfn_t gfn = gpa >> PAGE_SHIFT;
2072         int seg;
2073         int offset = offset_in_page(gpa);
2074         int ret;
2075
2076         while ((seg = next_segment(len, offset)) != 0) {
2077                 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
2078                 if (ret < 0)
2079                         return ret;
2080                 offset = 0;
2081                 len -= seg;
2082                 data += seg;
2083                 ++gfn;
2084         }
2085         return 0;
2086 }
2087 EXPORT_SYMBOL_GPL(kvm_read_guest);
2088
2089 int kvm_vcpu_read_guest(struct kvm_vcpu *vcpu, gpa_t gpa, void *data, unsigned long len)
2090 {
2091         gfn_t gfn = gpa >> PAGE_SHIFT;
2092         int seg;
2093         int offset = offset_in_page(gpa);
2094         int ret;
2095
2096         while ((seg = next_segment(len, offset)) != 0) {
2097                 ret = kvm_vcpu_read_guest_page(vcpu, gfn, data, offset, seg);
2098                 if (ret < 0)
2099                         return ret;
2100                 offset = 0;
2101                 len -= seg;
2102                 data += seg;
2103                 ++gfn;
2104         }
2105         return 0;
2106 }
2107 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest);
2108
2109 static int __kvm_read_guest_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
2110                                    void *data, int offset, unsigned long len)
2111 {
2112         int r;
2113         unsigned long addr;
2114
2115         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
2116         if (kvm_is_error_hva(addr))
2117                 return -EFAULT;
2118         pagefault_disable();
2119         r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len);
2120         pagefault_enable();
2121         if (r)
2122                 return -EFAULT;
2123         return 0;
2124 }
2125
2126 int kvm_vcpu_read_guest_atomic(struct kvm_vcpu *vcpu, gpa_t gpa,
2127                                void *data, unsigned long len)
2128 {
2129         gfn_t gfn = gpa >> PAGE_SHIFT;
2130         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2131         int offset = offset_in_page(gpa);
2132
2133         return __kvm_read_guest_atomic(slot, gfn, data, offset, len);
2134 }
2135 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_atomic);
2136
2137 static int __kvm_write_guest_page(struct kvm_memory_slot *memslot, gfn_t gfn,
2138                                   const void *data, int offset, int len)
2139 {
2140         int r;
2141         unsigned long addr;
2142
2143         addr = gfn_to_hva_memslot(memslot, gfn);
2144         if (kvm_is_error_hva(addr))
2145                 return -EFAULT;
2146         r = __copy_to_user((void __user *)addr + offset, data, len);
2147         if (r)
2148                 return -EFAULT;
2149         mark_page_dirty_in_slot(memslot, gfn);
2150         return 0;
2151 }
2152
2153 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn,
2154                          const void *data, int offset, int len)
2155 {
2156         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
2157
2158         return __kvm_write_guest_page(slot, gfn, data, offset, len);
2159 }
2160 EXPORT_SYMBOL_GPL(kvm_write_guest_page);
2161
2162 int kvm_vcpu_write_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn,
2163                               const void *data, int offset, int len)
2164 {
2165         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2166
2167         return __kvm_write_guest_page(slot, gfn, data, offset, len);
2168 }
2169 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest_page);
2170
2171 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
2172                     unsigned long len)
2173 {
2174         gfn_t gfn = gpa >> PAGE_SHIFT;
2175         int seg;
2176         int offset = offset_in_page(gpa);
2177         int ret;
2178
2179         while ((seg = next_segment(len, offset)) != 0) {
2180                 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
2181                 if (ret < 0)
2182                         return ret;
2183                 offset = 0;
2184                 len -= seg;
2185                 data += seg;
2186                 ++gfn;
2187         }
2188         return 0;
2189 }
2190 EXPORT_SYMBOL_GPL(kvm_write_guest);
2191
2192 int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data,
2193                          unsigned long len)
2194 {
2195         gfn_t gfn = gpa >> PAGE_SHIFT;
2196         int seg;
2197         int offset = offset_in_page(gpa);
2198         int ret;
2199
2200         while ((seg = next_segment(len, offset)) != 0) {
2201                 ret = kvm_vcpu_write_guest_page(vcpu, gfn, data, offset, seg);
2202                 if (ret < 0)
2203                         return ret;
2204                 offset = 0;
2205                 len -= seg;
2206                 data += seg;
2207                 ++gfn;
2208         }
2209         return 0;
2210 }
2211 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest);
2212
2213 static int __kvm_gfn_to_hva_cache_init(struct kvm_memslots *slots,
2214                                        struct gfn_to_hva_cache *ghc,
2215                                        gpa_t gpa, unsigned long len)
2216 {
2217         int offset = offset_in_page(gpa);
2218         gfn_t start_gfn = gpa >> PAGE_SHIFT;
2219         gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT;
2220         gfn_t nr_pages_needed = end_gfn - start_gfn + 1;
2221         gfn_t nr_pages_avail;
2222
2223         /* Update ghc->generation before performing any error checks. */
2224         ghc->generation = slots->generation;
2225
2226         if (start_gfn > end_gfn) {
2227                 ghc->hva = KVM_HVA_ERR_BAD;
2228                 return -EINVAL;
2229         }
2230
2231         /*
2232          * If the requested region crosses two memslots, we still
2233          * verify that the entire region is valid here.
2234          */
2235         for ( ; start_gfn <= end_gfn; start_gfn += nr_pages_avail) {
2236                 ghc->memslot = __gfn_to_memslot(slots, start_gfn);
2237                 ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn,
2238                                            &nr_pages_avail);
2239                 if (kvm_is_error_hva(ghc->hva))
2240                         return -EFAULT;
2241         }
2242
2243         /* Use the slow path for cross page reads and writes. */
2244         if (nr_pages_needed == 1)
2245                 ghc->hva += offset;
2246         else
2247                 ghc->memslot = NULL;
2248
2249         ghc->gpa = gpa;
2250         ghc->len = len;
2251         return 0;
2252 }
2253
2254 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2255                               gpa_t gpa, unsigned long len)
2256 {
2257         struct kvm_memslots *slots = kvm_memslots(kvm);
2258         return __kvm_gfn_to_hva_cache_init(slots, ghc, gpa, len);
2259 }
2260 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
2261
2262 int kvm_write_guest_offset_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2263                                   void *data, unsigned int offset,
2264                                   unsigned long len)
2265 {
2266         struct kvm_memslots *slots = kvm_memslots(kvm);
2267         int r;
2268         gpa_t gpa = ghc->gpa + offset;
2269
2270         BUG_ON(len + offset > ghc->len);
2271
2272         if (slots->generation != ghc->generation) {
2273                 if (__kvm_gfn_to_hva_cache_init(slots, ghc, ghc->gpa, ghc->len))
2274                         return -EFAULT;
2275         }
2276
2277         if (kvm_is_error_hva(ghc->hva))
2278                 return -EFAULT;
2279
2280         if (unlikely(!ghc->memslot))
2281                 return kvm_write_guest(kvm, gpa, data, len);
2282
2283         r = __copy_to_user((void __user *)ghc->hva + offset, data, len);
2284         if (r)
2285                 return -EFAULT;
2286         mark_page_dirty_in_slot(ghc->memslot, gpa >> PAGE_SHIFT);
2287
2288         return 0;
2289 }
2290 EXPORT_SYMBOL_GPL(kvm_write_guest_offset_cached);
2291
2292 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2293                            void *data, unsigned long len)
2294 {
2295         return kvm_write_guest_offset_cached(kvm, ghc, data, 0, len);
2296 }
2297 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
2298
2299 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2300                            void *data, unsigned long len)
2301 {
2302         struct kvm_memslots *slots = kvm_memslots(kvm);
2303         int r;
2304
2305         BUG_ON(len > ghc->len);
2306
2307         if (slots->generation != ghc->generation) {
2308                 if (__kvm_gfn_to_hva_cache_init(slots, ghc, ghc->gpa, ghc->len))
2309                         return -EFAULT;
2310         }
2311
2312         if (kvm_is_error_hva(ghc->hva))
2313                 return -EFAULT;
2314
2315         if (unlikely(!ghc->memslot))
2316                 return kvm_read_guest(kvm, ghc->gpa, data, len);
2317
2318         r = __copy_from_user(data, (void __user *)ghc->hva, len);
2319         if (r)
2320                 return -EFAULT;
2321
2322         return 0;
2323 }
2324 EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
2325
2326 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
2327 {
2328         const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0)));
2329
2330         return kvm_write_guest_page(kvm, gfn, zero_page, offset, len);
2331 }
2332 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
2333
2334 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
2335 {
2336         gfn_t gfn = gpa >> PAGE_SHIFT;
2337         int seg;
2338         int offset = offset_in_page(gpa);
2339         int ret;
2340
2341         while ((seg = next_segment(len, offset)) != 0) {
2342                 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
2343                 if (ret < 0)
2344                         return ret;
2345                 offset = 0;
2346                 len -= seg;
2347                 ++gfn;
2348         }
2349         return 0;
2350 }
2351 EXPORT_SYMBOL_GPL(kvm_clear_guest);
2352
2353 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot,
2354                                     gfn_t gfn)
2355 {
2356         if (memslot && memslot->dirty_bitmap) {
2357                 unsigned long rel_gfn = gfn - memslot->base_gfn;
2358
2359                 set_bit_le(rel_gfn, memslot->dirty_bitmap);
2360         }
2361 }
2362
2363 void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
2364 {
2365         struct kvm_memory_slot *memslot;
2366
2367         memslot = gfn_to_memslot(kvm, gfn);
2368         mark_page_dirty_in_slot(memslot, gfn);
2369 }
2370 EXPORT_SYMBOL_GPL(mark_page_dirty);
2371
2372 void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
2373 {
2374         struct kvm_memory_slot *memslot;
2375
2376         memslot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2377         mark_page_dirty_in_slot(memslot, gfn);
2378 }
2379 EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
2380
2381 void kvm_sigset_activate(struct kvm_vcpu *vcpu)
2382 {
2383         if (!vcpu->sigset_active)
2384                 return;
2385
2386         /*
2387          * This does a lockless modification of ->real_blocked, which is fine
2388          * because, only current can change ->real_blocked and all readers of
2389          * ->real_blocked don't care as long ->real_blocked is always a subset
2390          * of ->blocked.
2391          */
2392         sigprocmask(SIG_SETMASK, &vcpu->sigset, &current->real_blocked);
2393 }
2394
2395 void kvm_sigset_deactivate(struct kvm_vcpu *vcpu)
2396 {
2397         if (!vcpu->sigset_active)
2398                 return;
2399
2400         sigprocmask(SIG_SETMASK, &current->real_blocked, NULL);
2401         sigemptyset(&current->real_blocked);
2402 }
2403
2404 static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
2405 {
2406         unsigned int old, val, grow, grow_start;
2407
2408         old = val = vcpu->halt_poll_ns;
2409         grow_start = READ_ONCE(halt_poll_ns_grow_start);
2410         grow = READ_ONCE(halt_poll_ns_grow);
2411         if (!grow)
2412                 goto out;
2413
2414         val *= grow;
2415         if (val < grow_start)
2416                 val = grow_start;
2417
2418         if (val > halt_poll_ns)
2419                 val = halt_poll_ns;
2420
2421         vcpu->halt_poll_ns = val;
2422 out:
2423         trace_kvm_halt_poll_ns_grow(vcpu->vcpu_id, val, old);
2424 }
2425
2426 static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
2427 {
2428         unsigned int old, val, shrink;
2429
2430         old = val = vcpu->halt_poll_ns;
2431         shrink = READ_ONCE(halt_poll_ns_shrink);
2432         if (shrink == 0)
2433                 val = 0;
2434         else
2435                 val /= shrink;
2436
2437         vcpu->halt_poll_ns = val;
2438         trace_kvm_halt_poll_ns_shrink(vcpu->vcpu_id, val, old);
2439 }
2440
2441 static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
2442 {
2443         int ret = -EINTR;
2444         int idx = srcu_read_lock(&vcpu->kvm->srcu);
2445
2446         if (kvm_arch_vcpu_runnable(vcpu)) {
2447                 kvm_make_request(KVM_REQ_UNHALT, vcpu);
2448                 goto out;
2449         }
2450         if (kvm_cpu_has_pending_timer(vcpu))
2451                 goto out;
2452         if (signal_pending(current))
2453                 goto out;
2454
2455         ret = 0;
2456 out:
2457         srcu_read_unlock(&vcpu->kvm->srcu, idx);
2458         return ret;
2459 }
2460
2461 /*
2462  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
2463  */
2464 void kvm_vcpu_block(struct kvm_vcpu *vcpu)
2465 {
2466         ktime_t start, cur;
2467         DECLARE_SWAITQUEUE(wait);
2468         bool waited = false;
2469         u64 block_ns;
2470
2471         kvm_arch_vcpu_blocking(vcpu);
2472
2473         start = cur = ktime_get();
2474         if (vcpu->halt_poll_ns && !kvm_arch_no_poll(vcpu)) {
2475                 ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
2476
2477                 ++vcpu->stat.halt_attempted_poll;
2478                 do {
2479                         /*
2480                          * This sets KVM_REQ_UNHALT if an interrupt
2481                          * arrives.
2482                          */
2483                         if (kvm_vcpu_check_block(vcpu) < 0) {
2484                                 ++vcpu->stat.halt_successful_poll;
2485                                 if (!vcpu_valid_wakeup(vcpu))
2486                                         ++vcpu->stat.halt_poll_invalid;
2487                                 goto out;
2488                         }
2489                         cur = ktime_get();
2490                 } while (single_task_running() && ktime_before(cur, stop));
2491         }
2492
2493         for (;;) {
2494                 prepare_to_swait_exclusive(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
2495
2496                 if (kvm_vcpu_check_block(vcpu) < 0)
2497                         break;
2498
2499                 waited = true;
2500                 schedule();
2501         }
2502
2503         finish_swait(&vcpu->wq, &wait);
2504         cur = ktime_get();
2505 out:
2506         kvm_arch_vcpu_unblocking(vcpu);
2507         block_ns = ktime_to_ns(cur) - ktime_to_ns(start);
2508
2509         if (!kvm_arch_no_poll(vcpu)) {
2510                 if (!vcpu_valid_wakeup(vcpu)) {
2511                         shrink_halt_poll_ns(vcpu);
2512                 } else if (halt_poll_ns) {
2513                         if (block_ns <= vcpu->halt_poll_ns)
2514                                 ;
2515                         /* we had a long block, shrink polling */
2516                         else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
2517                                 shrink_halt_poll_ns(vcpu);
2518                         /* we had a short halt and our poll time is too small */
2519                         else if (vcpu->halt_poll_ns < halt_poll_ns &&
2520                                 block_ns < halt_poll_ns)
2521                                 grow_halt_poll_ns(vcpu);
2522                 } else {
2523                         vcpu->halt_poll_ns = 0;
2524                 }
2525         }
2526
2527         trace_kvm_vcpu_wakeup(block_ns, waited, vcpu_valid_wakeup(vcpu));
2528         kvm_arch_vcpu_block_finish(vcpu);
2529 }
2530 EXPORT_SYMBOL_GPL(kvm_vcpu_block);
2531
2532 bool kvm_vcpu_wake_up(struct kvm_vcpu *vcpu)
2533 {
2534         struct swait_queue_head *wqp;
2535
2536         wqp = kvm_arch_vcpu_wq(vcpu);
2537         if (swq_has_sleeper(wqp)) {
2538                 swake_up_one(wqp);
2539                 WRITE_ONCE(vcpu->ready, true);
2540                 ++vcpu->stat.halt_wakeup;
2541                 return true;
2542         }
2543
2544         return false;
2545 }
2546 EXPORT_SYMBOL_GPL(kvm_vcpu_wake_up);
2547
2548 #ifndef CONFIG_S390
2549 /*
2550  * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode.
2551  */
2552 void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
2553 {
2554         int me;
2555         int cpu = vcpu->cpu;
2556
2557         if (kvm_vcpu_wake_up(vcpu))
2558                 return;
2559
2560         me = get_cpu();
2561         if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
2562                 if (kvm_arch_vcpu_should_kick(vcpu))
2563                         smp_send_reschedule(cpu);
2564         put_cpu();
2565 }
2566 EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
2567 #endif /* !CONFIG_S390 */
2568
2569 int kvm_vcpu_yield_to(struct kvm_vcpu *target)
2570 {
2571         struct pid *pid;
2572         struct task_struct *task = NULL;
2573         int ret = 0;
2574
2575         rcu_read_lock();
2576         pid = rcu_dereference(target->pid);
2577         if (pid)
2578                 task = get_pid_task(pid, PIDTYPE_PID);
2579         rcu_read_unlock();
2580         if (!task)
2581                 return ret;
2582         ret = yield_to(task, 1);
2583         put_task_struct(task);
2584
2585         return ret;
2586 }
2587 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to);
2588
2589 /*
2590  * Helper that checks whether a VCPU is eligible for directed yield.
2591  * Most eligible candidate to yield is decided by following heuristics:
2592  *
2593  *  (a) VCPU which has not done pl-exit or cpu relax intercepted recently
2594  *  (preempted lock holder), indicated by @in_spin_loop.
2595  *  Set at the beiginning and cleared at the end of interception/PLE handler.
2596  *
2597  *  (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get
2598  *  chance last time (mostly it has become eligible now since we have probably
2599  *  yielded to lockholder in last iteration. This is done by toggling
2600  *  @dy_eligible each time a VCPU checked for eligibility.)
2601  *
2602  *  Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding
2603  *  to preempted lock-holder could result in wrong VCPU selection and CPU
2604  *  burning. Giving priority for a potential lock-holder increases lock
2605  *  progress.
2606  *
2607  *  Since algorithm is based on heuristics, accessing another VCPU data without
2608  *  locking does not harm. It may result in trying to yield to  same VCPU, fail
2609  *  and continue with next VCPU and so on.
2610  */
2611 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu)
2612 {
2613 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT
2614         bool eligible;
2615
2616         eligible = !vcpu->spin_loop.in_spin_loop ||
2617                     vcpu->spin_loop.dy_eligible;
2618
2619         if (vcpu->spin_loop.in_spin_loop)
2620                 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible);
2621
2622         return eligible;
2623 #else
2624         return true;
2625 #endif
2626 }
2627
2628 /*
2629  * Unlike kvm_arch_vcpu_runnable, this function is called outside
2630  * a vcpu_load/vcpu_put pair.  However, for most architectures
2631  * kvm_arch_vcpu_runnable does not require vcpu_load.
2632  */
2633 bool __weak kvm_arch_dy_runnable(struct kvm_vcpu *vcpu)
2634 {
2635         return kvm_arch_vcpu_runnable(vcpu);
2636 }
2637
2638 static bool vcpu_dy_runnable(struct kvm_vcpu *vcpu)
2639 {
2640         if (kvm_arch_dy_runnable(vcpu))
2641                 return true;
2642
2643 #ifdef CONFIG_KVM_ASYNC_PF
2644         if (!list_empty_careful(&vcpu->async_pf.done))
2645                 return true;
2646 #endif
2647
2648         return false;
2649 }
2650
2651 void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
2652 {
2653         struct kvm *kvm = me->kvm;
2654         struct kvm_vcpu *vcpu;
2655         int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
2656         int yielded = 0;
2657         int try = 3;
2658         int pass;
2659         int i;
2660
2661         kvm_vcpu_set_in_spin_loop(me, true);
2662         /*
2663          * We boost the priority of a VCPU that is runnable but not
2664          * currently running, because it got preempted by something
2665          * else and called schedule in __vcpu_run.  Hopefully that
2666          * VCPU is holding the lock that we need and will release it.
2667          * We approximate round-robin by starting at the last boosted VCPU.
2668          */
2669         for (pass = 0; pass < 2 && !yielded && try; pass++) {
2670                 kvm_for_each_vcpu(i, vcpu, kvm) {
2671                         if (!pass && i <= last_boosted_vcpu) {
2672                                 i = last_boosted_vcpu;
2673                                 continue;
2674                         } else if (pass && i > last_boosted_vcpu)
2675                                 break;
2676                         if (!READ_ONCE(vcpu->ready))
2677                                 continue;
2678                         if (vcpu == me)
2679                                 continue;
2680                         if (swait_active(&vcpu->wq) && !vcpu_dy_runnable(vcpu))
2681                                 continue;
2682                         if (READ_ONCE(vcpu->preempted) && yield_to_kernel_mode &&
2683                                 !kvm_arch_vcpu_in_kernel(vcpu))
2684                                 continue;
2685                         if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
2686                                 continue;
2687
2688                         yielded = kvm_vcpu_yield_to(vcpu);
2689                         if (yielded > 0) {
2690                                 kvm->last_boosted_vcpu = i;
2691                                 break;
2692                         } else if (yielded < 0) {
2693                                 try--;
2694                                 if (!try)
2695                                         break;
2696                         }
2697                 }
2698         }
2699         kvm_vcpu_set_in_spin_loop(me, false);
2700
2701         /* Ensure vcpu is not eligible during next spinloop */
2702         kvm_vcpu_set_dy_eligible(me, false);
2703 }
2704 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
2705
2706 static vm_fault_t kvm_vcpu_fault(struct vm_fault *vmf)
2707 {
2708         struct kvm_vcpu *vcpu = vmf->vma->vm_file->private_data;
2709         struct page *page;
2710
2711         if (vmf->pgoff == 0)
2712                 page = virt_to_page(vcpu->run);
2713 #ifdef CONFIG_X86
2714         else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
2715                 page = virt_to_page(vcpu->arch.pio_data);
2716 #endif
2717 #ifdef CONFIG_KVM_MMIO
2718         else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
2719                 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
2720 #endif
2721         else
2722                 return kvm_arch_vcpu_fault(vcpu, vmf);
2723         get_page(page);
2724         vmf->page = page;
2725         return 0;
2726 }
2727
2728 static const struct vm_operations_struct kvm_vcpu_vm_ops = {
2729         .fault = kvm_vcpu_fault,
2730 };
2731
2732 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
2733 {
2734         vma->vm_ops = &kvm_vcpu_vm_ops;
2735         return 0;
2736 }
2737
2738 static int kvm_vcpu_release(struct inode *inode, struct file *filp)
2739 {
2740         struct kvm_vcpu *vcpu = filp->private_data;
2741
2742         debugfs_remove_recursive(vcpu->debugfs_dentry);
2743         kvm_put_kvm(vcpu->kvm);
2744         return 0;
2745 }
2746
2747 static struct file_operations kvm_vcpu_fops = {
2748         .release        = kvm_vcpu_release,
2749         .unlocked_ioctl = kvm_vcpu_ioctl,
2750         .mmap           = kvm_vcpu_mmap,
2751         .llseek         = noop_llseek,
2752         KVM_COMPAT(kvm_vcpu_compat_ioctl),
2753 };
2754
2755 /*
2756  * Allocates an inode for the vcpu.
2757  */
2758 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
2759 {
2760         char name[8 + 1 + ITOA_MAX_LEN + 1];
2761
2762         snprintf(name, sizeof(name), "kvm-vcpu:%d", vcpu->vcpu_id);
2763         return anon_inode_getfd(name, &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC);
2764 }
2765
2766 static void kvm_create_vcpu_debugfs(struct kvm_vcpu *vcpu)
2767 {
2768 #ifdef __KVM_HAVE_ARCH_VCPU_DEBUGFS
2769         char dir_name[ITOA_MAX_LEN * 2];
2770
2771         if (!debugfs_initialized())
2772                 return;
2773
2774         snprintf(dir_name, sizeof(dir_name), "vcpu%d", vcpu->vcpu_id);
2775         vcpu->debugfs_dentry = debugfs_create_dir(dir_name,
2776                                                   vcpu->kvm->debugfs_dentry);
2777
2778         kvm_arch_create_vcpu_debugfs(vcpu);
2779 #endif
2780 }
2781
2782 /*
2783  * Creates some virtual cpus.  Good luck creating more than one.
2784  */
2785 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
2786 {
2787         int r;
2788         struct kvm_vcpu *vcpu;
2789         struct page *page;
2790
2791         if (id >= KVM_MAX_VCPU_ID)
2792                 return -EINVAL;
2793
2794         mutex_lock(&kvm->lock);
2795         if (kvm->created_vcpus == KVM_MAX_VCPUS) {
2796                 mutex_unlock(&kvm->lock);
2797                 return -EINVAL;
2798         }
2799
2800         kvm->created_vcpus++;
2801         mutex_unlock(&kvm->lock);
2802
2803         r = kvm_arch_vcpu_precreate(kvm, id);
2804         if (r)
2805                 goto vcpu_decrement;
2806
2807         vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
2808         if (!vcpu) {
2809                 r = -ENOMEM;
2810                 goto vcpu_decrement;
2811         }
2812
2813         BUILD_BUG_ON(sizeof(struct kvm_run) > PAGE_SIZE);
2814         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2815         if (!page) {
2816                 r = -ENOMEM;
2817                 goto vcpu_free;
2818         }
2819         vcpu->run = page_address(page);
2820
2821         kvm_vcpu_init(vcpu, kvm, id);
2822
2823         r = kvm_arch_vcpu_create(vcpu);
2824         if (r)
2825                 goto vcpu_free_run_page;
2826
2827         kvm_create_vcpu_debugfs(vcpu);
2828
2829         mutex_lock(&kvm->lock);
2830         if (kvm_get_vcpu_by_id(kvm, id)) {
2831                 r = -EEXIST;
2832                 goto unlock_vcpu_destroy;
2833         }
2834
2835         vcpu->vcpu_idx = atomic_read(&kvm->online_vcpus);
2836         BUG_ON(kvm->vcpus[vcpu->vcpu_idx]);
2837
2838         /* Now it's all set up, let userspace reach it */
2839         kvm_get_kvm(kvm);
2840         r = create_vcpu_fd(vcpu);
2841         if (r < 0) {
2842                 kvm_put_kvm_no_destroy(kvm);
2843                 goto unlock_vcpu_destroy;
2844         }
2845
2846         kvm->vcpus[vcpu->vcpu_idx] = vcpu;
2847
2848         /*
2849          * Pairs with smp_rmb() in kvm_get_vcpu.  Write kvm->vcpus
2850          * before kvm->online_vcpu's incremented value.
2851          */
2852         smp_wmb();
2853         atomic_inc(&kvm->online_vcpus);
2854
2855         mutex_unlock(&kvm->lock);
2856         kvm_arch_vcpu_postcreate(vcpu);
2857         return r;
2858
2859 unlock_vcpu_destroy:
2860         mutex_unlock(&kvm->lock);
2861         debugfs_remove_recursive(vcpu->debugfs_dentry);
2862         kvm_arch_vcpu_destroy(vcpu);
2863 vcpu_free_run_page:
2864         free_page((unsigned long)vcpu->run);
2865 vcpu_free:
2866         kmem_cache_free(kvm_vcpu_cache, vcpu);
2867 vcpu_decrement:
2868         mutex_lock(&kvm->lock);
2869         kvm->created_vcpus--;
2870         mutex_unlock(&kvm->lock);
2871         return r;
2872 }
2873
2874 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
2875 {
2876         if (sigset) {
2877                 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
2878                 vcpu->sigset_active = 1;
2879                 vcpu->sigset = *sigset;
2880         } else
2881                 vcpu->sigset_active = 0;
2882         return 0;
2883 }
2884
2885 static long kvm_vcpu_ioctl(struct file *filp,
2886                            unsigned int ioctl, unsigned long arg)
2887 {
2888         struct kvm_vcpu *vcpu = filp->private_data;
2889         void __user *argp = (void __user *)arg;
2890         int r;
2891         struct kvm_fpu *fpu = NULL;
2892         struct kvm_sregs *kvm_sregs = NULL;
2893
2894         if (vcpu->kvm->mm != current->mm)
2895                 return -EIO;
2896
2897         if (unlikely(_IOC_TYPE(ioctl) != KVMIO))
2898                 return -EINVAL;
2899
2900         /*
2901          * Some architectures have vcpu ioctls that are asynchronous to vcpu
2902          * execution; mutex_lock() would break them.
2903          */
2904         r = kvm_arch_vcpu_async_ioctl(filp, ioctl, arg);
2905         if (r != -ENOIOCTLCMD)
2906                 return r;
2907
2908         if (mutex_lock_killable(&vcpu->mutex))
2909                 return -EINTR;
2910         switch (ioctl) {
2911         case KVM_RUN: {
2912                 struct pid *oldpid;
2913                 r = -EINVAL;
2914                 if (arg)
2915                         goto out;
2916                 oldpid = rcu_access_pointer(vcpu->pid);
2917                 if (unlikely(oldpid != task_pid(current))) {
2918                         /* The thread running this VCPU changed. */
2919                         struct pid *newpid;
2920
2921                         r = kvm_arch_vcpu_run_pid_change(vcpu);
2922                         if (r)
2923                                 break;
2924
2925                         newpid = get_task_pid(current, PIDTYPE_PID);
2926                         rcu_assign_pointer(vcpu->pid, newpid);
2927                         if (oldpid)
2928                                 synchronize_rcu();
2929                         put_pid(oldpid);
2930                 }
2931                 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
2932                 trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
2933                 break;
2934         }
2935         case KVM_GET_REGS: {
2936                 struct kvm_regs *kvm_regs;
2937
2938                 r = -ENOMEM;
2939                 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL_ACCOUNT);
2940                 if (!kvm_regs)
2941                         goto out;
2942                 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
2943                 if (r)
2944                         goto out_free1;
2945                 r = -EFAULT;
2946                 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
2947                         goto out_free1;
2948                 r = 0;
2949 out_free1:
2950                 kfree(kvm_regs);
2951                 break;
2952         }
2953         case KVM_SET_REGS: {
2954                 struct kvm_regs *kvm_regs;
2955
2956                 r = -ENOMEM;
2957                 kvm_regs = memdup_user(argp, sizeof(*kvm_regs));
2958                 if (IS_ERR(kvm_regs)) {
2959                         r = PTR_ERR(kvm_regs);
2960                         goto out;
2961                 }
2962                 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
2963                 kfree(kvm_regs);
2964                 break;
2965         }
2966         case KVM_GET_SREGS: {
2967                 kvm_sregs = kzalloc(sizeof(struct kvm_sregs),
2968                                     GFP_KERNEL_ACCOUNT);
2969                 r = -ENOMEM;
2970                 if (!kvm_sregs)
2971                         goto out;
2972                 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
2973                 if (r)
2974                         goto out;
2975                 r = -EFAULT;
2976                 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
2977                         goto out;
2978                 r = 0;
2979                 break;
2980         }
2981         case KVM_SET_SREGS: {
2982                 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs));
2983                 if (IS_ERR(kvm_sregs)) {
2984                         r = PTR_ERR(kvm_sregs);
2985                         kvm_sregs = NULL;
2986                         goto out;
2987                 }
2988                 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
2989                 break;
2990         }
2991         case KVM_GET_MP_STATE: {
2992                 struct kvm_mp_state mp_state;
2993
2994                 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
2995                 if (r)
2996                         goto out;
2997                 r = -EFAULT;
2998                 if (copy_to_user(argp, &mp_state, sizeof(mp_state)))
2999                         goto out;
3000                 r = 0;
3001                 break;
3002         }
3003         case KVM_SET_MP_STATE: {
3004                 struct kvm_mp_state mp_state;
3005
3006                 r = -EFAULT;
3007                 if (copy_from_user(&mp_state, argp, sizeof(mp_state)))
3008                         goto out;
3009                 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
3010                 break;
3011         }
3012         case KVM_TRANSLATE: {
3013                 struct kvm_translation tr;
3014
3015                 r = -EFAULT;
3016                 if (copy_from_user(&tr, argp, sizeof(tr)))
3017                         goto out;
3018                 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
3019                 if (r)
3020                         goto out;
3021                 r = -EFAULT;
3022                 if (copy_to_user(argp, &tr, sizeof(tr)))
3023                         goto out;
3024                 r = 0;
3025                 break;
3026         }
3027         case KVM_SET_GUEST_DEBUG: {
3028                 struct kvm_guest_debug dbg;
3029
3030                 r = -EFAULT;
3031                 if (copy_from_user(&dbg, argp, sizeof(dbg)))
3032                         goto out;
3033                 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
3034                 break;
3035         }
3036         case KVM_SET_SIGNAL_MASK: {
3037                 struct kvm_signal_mask __user *sigmask_arg = argp;
3038                 struct kvm_signal_mask kvm_sigmask;
3039                 sigset_t sigset, *p;
3040
3041                 p = NULL;
3042                 if (argp) {
3043                         r = -EFAULT;
3044                         if (copy_from_user(&kvm_sigmask, argp,
3045                                            sizeof(kvm_sigmask)))
3046                                 goto out;
3047                         r = -EINVAL;
3048                         if (kvm_sigmask.len != sizeof(sigset))
3049                                 goto out;
3050                         r = -EFAULT;
3051                         if (copy_from_user(&sigset, sigmask_arg->sigset,
3052                                            sizeof(sigset)))
3053                                 goto out;
3054                         p = &sigset;
3055                 }
3056                 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
3057                 break;
3058         }
3059         case KVM_GET_FPU: {
3060                 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL_ACCOUNT);
3061                 r = -ENOMEM;
3062                 if (!fpu)
3063                         goto out;
3064                 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
3065                 if (r)
3066                         goto out;
3067                 r = -EFAULT;
3068                 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
3069                         goto out;
3070                 r = 0;
3071                 break;
3072         }
3073         case KVM_SET_FPU: {
3074                 fpu = memdup_user(argp, sizeof(*fpu));
3075                 if (IS_ERR(fpu)) {
3076                         r = PTR_ERR(fpu);
3077                         fpu = NULL;
3078                         goto out;
3079                 }
3080                 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
3081                 break;
3082         }
3083         default:
3084                 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
3085         }
3086 out:
3087         mutex_unlock(&vcpu->mutex);
3088         kfree(fpu);
3089         kfree(kvm_sregs);
3090         return r;
3091 }
3092
3093 #ifdef CONFIG_KVM_COMPAT
3094 static long kvm_vcpu_compat_ioctl(struct file *filp,
3095                                   unsigned int ioctl, unsigned long arg)
3096 {
3097         struct kvm_vcpu *vcpu = filp->private_data;
3098         void __user *argp = compat_ptr(arg);
3099         int r;
3100
3101         if (vcpu->kvm->mm != current->mm)
3102                 return -EIO;
3103
3104         switch (ioctl) {
3105         case KVM_SET_SIGNAL_MASK: {
3106                 struct kvm_signal_mask __user *sigmask_arg = argp;
3107                 struct kvm_signal_mask kvm_sigmask;
3108                 sigset_t sigset;
3109
3110                 if (argp) {
3111                         r = -EFAULT;
3112                         if (copy_from_user(&kvm_sigmask, argp,
3113                                            sizeof(kvm_sigmask)))
3114                                 goto out;
3115                         r = -EINVAL;
3116                         if (kvm_sigmask.len != sizeof(compat_sigset_t))
3117                                 goto out;
3118                         r = -EFAULT;
3119                         if (get_compat_sigset(&sigset, (void *)sigmask_arg->sigset))
3120                                 goto out;
3121                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
3122                 } else
3123                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL);
3124                 break;
3125         }
3126         default:
3127                 r = kvm_vcpu_ioctl(filp, ioctl, arg);
3128         }
3129
3130 out:
3131         return r;
3132 }
3133 #endif
3134
3135 static int kvm_device_mmap(struct file *filp, struct vm_area_struct *vma)
3136 {
3137         struct kvm_device *dev = filp->private_data;
3138
3139         if (dev->ops->mmap)
3140                 return dev->ops->mmap(dev, vma);
3141
3142         return -ENODEV;
3143 }
3144
3145 static int kvm_device_ioctl_attr(struct kvm_device *dev,
3146                                  int (*accessor)(struct kvm_device *dev,
3147                                                  struct kvm_device_attr *attr),
3148                                  unsigned long arg)
3149 {
3150         struct kvm_device_attr attr;
3151
3152         if (!accessor)
3153                 return -EPERM;
3154
3155         if (copy_from_user(&attr, (void __user *)arg, sizeof(attr)))
3156                 return -EFAULT;
3157
3158         return accessor(dev, &attr);
3159 }
3160
3161 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl,
3162                              unsigned long arg)
3163 {
3164         struct kvm_device *dev = filp->private_data;
3165
3166         if (dev->kvm->mm != current->mm)
3167                 return -EIO;
3168
3169         switch (ioctl) {
3170         case KVM_SET_DEVICE_ATTR:
3171                 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg);
3172         case KVM_GET_DEVICE_ATTR:
3173                 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg);
3174         case KVM_HAS_DEVICE_ATTR:
3175                 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg);
3176         default:
3177                 if (dev->ops->ioctl)
3178                         return dev->ops->ioctl(dev, ioctl, arg);
3179
3180                 return -ENOTTY;
3181         }
3182 }
3183
3184 static int kvm_device_release(struct inode *inode, struct file *filp)
3185 {
3186         struct kvm_device *dev = filp->private_data;
3187         struct kvm *kvm = dev->kvm;
3188
3189         if (dev->ops->release) {
3190                 mutex_lock(&kvm->lock);
3191                 list_del(&dev->vm_node);
3192                 dev->ops->release(dev);
3193                 mutex_unlock(&kvm->lock);
3194         }
3195
3196         kvm_put_kvm(kvm);
3197         return 0;
3198 }
3199
3200 static const struct file_operations kvm_device_fops = {
3201         .unlocked_ioctl = kvm_device_ioctl,
3202         .release = kvm_device_release,
3203         KVM_COMPAT(kvm_device_ioctl),
3204         .mmap = kvm_device_mmap,
3205 };
3206
3207 struct kvm_device *kvm_device_from_filp(struct file *filp)
3208 {
3209         if (filp->f_op != &kvm_device_fops)
3210                 return NULL;
3211
3212         return filp->private_data;
3213 }
3214
3215 static const struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = {
3216 #ifdef CONFIG_KVM_MPIC
3217         [KVM_DEV_TYPE_FSL_MPIC_20]      = &kvm_mpic_ops,
3218         [KVM_DEV_TYPE_FSL_MPIC_42]      = &kvm_mpic_ops,
3219 #endif
3220 };
3221
3222 int kvm_register_device_ops(const struct kvm_device_ops *ops, u32 type)
3223 {
3224         if (type >= ARRAY_SIZE(kvm_device_ops_table))
3225                 return -ENOSPC;
3226
3227         if (kvm_device_ops_table[type] != NULL)
3228                 return -EEXIST;
3229
3230         kvm_device_ops_table[type] = ops;
3231         return 0;
3232 }
3233
3234 void kvm_unregister_device_ops(u32 type)
3235 {
3236         if (kvm_device_ops_table[type] != NULL)
3237                 kvm_device_ops_table[type] = NULL;
3238 }
3239
3240 static int kvm_ioctl_create_device(struct kvm *kvm,
3241                                    struct kvm_create_device *cd)
3242 {
3243         const struct kvm_device_ops *ops = NULL;
3244         struct kvm_device *dev;
3245         bool test = cd->flags & KVM_CREATE_DEVICE_TEST;
3246         int type;
3247         int ret;
3248
3249         if (cd->type >= ARRAY_SIZE(kvm_device_ops_table))
3250                 return -ENODEV;
3251
3252         type = array_index_nospec(cd->type, ARRAY_SIZE(kvm_device_ops_table));
3253         ops = kvm_device_ops_table[type];
3254         if (ops == NULL)
3255                 return -ENODEV;
3256
3257         if (test)
3258                 return 0;
3259
3260         dev = kzalloc(sizeof(*dev), GFP_KERNEL_ACCOUNT);
3261         if (!dev)
3262                 return -ENOMEM;
3263
3264         dev->ops = ops;
3265         dev->kvm = kvm;
3266
3267         mutex_lock(&kvm->lock);
3268         ret = ops->create(dev, type);
3269         if (ret < 0) {
3270                 mutex_unlock(&kvm->lock);
3271                 kfree(dev);
3272                 return ret;
3273         }
3274         list_add(&dev->vm_node, &kvm->devices);
3275         mutex_unlock(&kvm->lock);
3276
3277         if (ops->init)
3278                 ops->init(dev);
3279
3280         kvm_get_kvm(kvm);
3281         ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC);
3282         if (ret < 0) {
3283                 kvm_put_kvm_no_destroy(kvm);
3284                 mutex_lock(&kvm->lock);
3285                 list_del(&dev->vm_node);
3286                 mutex_unlock(&kvm->lock);
3287                 ops->destroy(dev);
3288                 return ret;
3289         }
3290
3291         cd->fd = ret;
3292         return 0;
3293 }
3294
3295 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
3296 {
3297         switch (arg) {
3298         case KVM_CAP_USER_MEMORY:
3299         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
3300         case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
3301         case KVM_CAP_INTERNAL_ERROR_DATA:
3302 #ifdef CONFIG_HAVE_KVM_MSI
3303         case KVM_CAP_SIGNAL_MSI:
3304 #endif
3305 #ifdef CONFIG_HAVE_KVM_IRQFD
3306         case KVM_CAP_IRQFD:
3307         case KVM_CAP_IRQFD_RESAMPLE:
3308 #endif
3309         case KVM_CAP_IOEVENTFD_ANY_LENGTH:
3310         case KVM_CAP_CHECK_EXTENSION_VM:
3311         case KVM_CAP_ENABLE_CAP_VM:
3312 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
3313         case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
3314 #endif
3315                 return 1;
3316 #ifdef CONFIG_KVM_MMIO
3317         case KVM_CAP_COALESCED_MMIO:
3318                 return KVM_COALESCED_MMIO_PAGE_OFFSET;
3319         case KVM_CAP_COALESCED_PIO:
3320                 return 1;
3321 #endif
3322 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
3323         case KVM_CAP_IRQ_ROUTING:
3324                 return KVM_MAX_IRQ_ROUTES;
3325 #endif
3326 #if KVM_ADDRESS_SPACE_NUM > 1
3327         case KVM_CAP_MULTI_ADDRESS_SPACE:
3328                 return KVM_ADDRESS_SPACE_NUM;
3329 #endif
3330         case KVM_CAP_NR_MEMSLOTS:
3331                 return KVM_USER_MEM_SLOTS;
3332         default:
3333                 break;
3334         }
3335         return kvm_vm_ioctl_check_extension(kvm, arg);
3336 }
3337
3338 int __attribute__((weak)) kvm_vm_ioctl_enable_cap(struct kvm *kvm,
3339                                                   struct kvm_enable_cap *cap)
3340 {
3341         return -EINVAL;
3342 }
3343
3344 static int kvm_vm_ioctl_enable_cap_generic(struct kvm *kvm,
3345                                            struct kvm_enable_cap *cap)
3346 {
3347         switch (cap->cap) {
3348 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
3349         case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
3350                 if (cap->flags || (cap->args[0] & ~1))
3351                         return -EINVAL;
3352                 kvm->manual_dirty_log_protect = cap->args[0];
3353                 return 0;
3354 #endif
3355         default:
3356                 return kvm_vm_ioctl_enable_cap(kvm, cap);
3357         }
3358 }
3359
3360 static long kvm_vm_ioctl(struct file *filp,
3361                            unsigned int ioctl, unsigned long arg)
3362 {
3363         struct kvm *kvm = filp->private_data;
3364         void __user *argp = (void __user *)arg;
3365         int r;
3366
3367         if (kvm->mm != current->mm)
3368                 return -EIO;
3369         switch (ioctl) {
3370         case KVM_CREATE_VCPU:
3371                 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
3372                 break;
3373         case KVM_ENABLE_CAP: {
3374                 struct kvm_enable_cap cap;
3375
3376                 r = -EFAULT;
3377                 if (copy_from_user(&cap, argp, sizeof(cap)))
3378                         goto out;
3379                 r = kvm_vm_ioctl_enable_cap_generic(kvm, &cap);
3380                 break;
3381         }
3382         case KVM_SET_USER_MEMORY_REGION: {
3383                 struct kvm_userspace_memory_region kvm_userspace_mem;
3384
3385                 r = -EFAULT;
3386                 if (copy_from_user(&kvm_userspace_mem, argp,
3387                                                 sizeof(kvm_userspace_mem)))
3388                         goto out;
3389
3390                 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem);
3391                 break;
3392         }
3393         case KVM_GET_DIRTY_LOG: {
3394                 struct kvm_dirty_log log;
3395
3396                 r = -EFAULT;
3397                 if (copy_from_user(&log, argp, sizeof(log)))
3398                         goto out;
3399                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
3400                 break;
3401         }
3402 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
3403         case KVM_CLEAR_DIRTY_LOG: {
3404                 struct kvm_clear_dirty_log log;
3405
3406                 r = -EFAULT;
3407                 if (copy_from_user(&log, argp, sizeof(log)))
3408                         goto out;
3409                 r = kvm_vm_ioctl_clear_dirty_log(kvm, &log);
3410                 break;
3411         }
3412 #endif
3413 #ifdef CONFIG_KVM_MMIO
3414         case KVM_REGISTER_COALESCED_MMIO: {
3415                 struct kvm_coalesced_mmio_zone zone;
3416
3417                 r = -EFAULT;
3418                 if (copy_from_user(&zone, argp, sizeof(zone)))
3419                         goto out;
3420                 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
3421                 break;
3422         }
3423         case KVM_UNREGISTER_COALESCED_MMIO: {
3424                 struct kvm_coalesced_mmio_zone zone;
3425
3426                 r = -EFAULT;
3427                 if (copy_from_user(&zone, argp, sizeof(zone)))
3428                         goto out;
3429                 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
3430                 break;
3431         }
3432 #endif
3433         case KVM_IRQFD: {
3434                 struct kvm_irqfd data;
3435
3436                 r = -EFAULT;
3437                 if (copy_from_user(&data, argp, sizeof(data)))
3438                         goto out;
3439                 r = kvm_irqfd(kvm, &data);
3440                 break;
3441         }
3442         case KVM_IOEVENTFD: {
3443                 struct kvm_ioeventfd data;
3444
3445                 r = -EFAULT;
3446                 if (copy_from_user(&data, argp, sizeof(data)))
3447                         goto out;
3448                 r = kvm_ioeventfd(kvm, &data);
3449                 break;
3450         }
3451 #ifdef CONFIG_HAVE_KVM_MSI
3452         case KVM_SIGNAL_MSI: {
3453                 struct kvm_msi msi;
3454
3455                 r = -EFAULT;
3456                 if (copy_from_user(&msi, argp, sizeof(msi)))
3457                         goto out;
3458                 r = kvm_send_userspace_msi(kvm, &msi);
3459                 break;
3460         }
3461 #endif
3462 #ifdef __KVM_HAVE_IRQ_LINE
3463         case KVM_IRQ_LINE_STATUS:
3464         case KVM_IRQ_LINE: {
3465                 struct kvm_irq_level irq_event;
3466
3467                 r = -EFAULT;
3468                 if (copy_from_user(&irq_event, argp, sizeof(irq_event)))
3469                         goto out;
3470
3471                 r = kvm_vm_ioctl_irq_line(kvm, &irq_event,
3472                                         ioctl == KVM_IRQ_LINE_STATUS);
3473                 if (r)
3474                         goto out;
3475
3476                 r = -EFAULT;
3477                 if (ioctl == KVM_IRQ_LINE_STATUS) {
3478                         if (copy_to_user(argp, &irq_event, sizeof(irq_event)))
3479                                 goto out;
3480                 }
3481
3482                 r = 0;
3483                 break;
3484         }
3485 #endif
3486 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
3487         case KVM_SET_GSI_ROUTING: {
3488                 struct kvm_irq_routing routing;
3489                 struct kvm_irq_routing __user *urouting;
3490                 struct kvm_irq_routing_entry *entries = NULL;
3491
3492                 r = -EFAULT;
3493                 if (copy_from_user(&routing, argp, sizeof(routing)))
3494                         goto out;
3495                 r = -EINVAL;
3496                 if (!kvm_arch_can_set_irq_routing(kvm))
3497                         goto out;
3498                 if (routing.nr > KVM_MAX_IRQ_ROUTES)
3499                         goto out;
3500                 if (routing.flags)
3501                         goto out;
3502                 if (routing.nr) {
3503                         r = -ENOMEM;
3504                         entries = vmalloc(array_size(sizeof(*entries),
3505                                                      routing.nr));
3506                         if (!entries)
3507                                 goto out;
3508                         r = -EFAULT;
3509                         urouting = argp;
3510                         if (copy_from_user(entries, urouting->entries,
3511                                            routing.nr * sizeof(*entries)))
3512                                 goto out_free_irq_routing;
3513                 }
3514                 r = kvm_set_irq_routing(kvm, entries, routing.nr,
3515                                         routing.flags);
3516 out_free_irq_routing:
3517                 vfree(entries);
3518                 break;
3519         }
3520 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */
3521         case KVM_CREATE_DEVICE: {
3522                 struct kvm_create_device cd;
3523
3524                 r = -EFAULT;
3525                 if (copy_from_user(&cd, argp, sizeof(cd)))
3526                         goto out;
3527
3528                 r = kvm_ioctl_create_device(kvm, &cd);
3529                 if (r)
3530                         goto out;
3531
3532                 r = -EFAULT;
3533                 if (copy_to_user(argp, &cd, sizeof(cd)))
3534                         goto out;
3535
3536                 r = 0;
3537                 break;
3538         }
3539         case KVM_CHECK_EXTENSION:
3540                 r = kvm_vm_ioctl_check_extension_generic(kvm, arg);
3541                 break;
3542         default:
3543                 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
3544         }
3545 out:
3546         return r;
3547 }
3548
3549 #ifdef CONFIG_KVM_COMPAT
3550 struct compat_kvm_dirty_log {
3551         __u32 slot;
3552         __u32 padding1;
3553         union {
3554                 compat_uptr_t dirty_bitmap; /* one bit per page */
3555                 __u64 padding2;
3556         };
3557 };
3558
3559 static long kvm_vm_compat_ioctl(struct file *filp,
3560                            unsigned int ioctl, unsigned long arg)
3561 {
3562         struct kvm *kvm = filp->private_data;
3563         int r;
3564
3565         if (kvm->mm != current->mm)
3566                 return -EIO;
3567         switch (ioctl) {
3568         case KVM_GET_DIRTY_LOG: {
3569                 struct compat_kvm_dirty_log compat_log;
3570                 struct kvm_dirty_log log;
3571
3572                 if (copy_from_user(&compat_log, (void __user *)arg,
3573                                    sizeof(compat_log)))
3574                         return -EFAULT;
3575                 log.slot         = compat_log.slot;
3576                 log.padding1     = compat_log.padding1;
3577                 log.padding2     = compat_log.padding2;
3578                 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
3579
3580                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
3581                 break;
3582         }
3583         default:
3584                 r = kvm_vm_ioctl(filp, ioctl, arg);
3585         }
3586         return r;
3587 }
3588 #endif
3589
3590 static struct file_operations kvm_vm_fops = {
3591         .release        = kvm_vm_release,
3592         .unlocked_ioctl = kvm_vm_ioctl,
3593         .llseek         = noop_llseek,
3594         KVM_COMPAT(kvm_vm_compat_ioctl),
3595 };
3596
3597 static int kvm_dev_ioctl_create_vm(unsigned long type)
3598 {
3599         int r;
3600         struct kvm *kvm;
3601         struct file *file;
3602
3603         kvm = kvm_create_vm(type);
3604         if (IS_ERR(kvm))
3605                 return PTR_ERR(kvm);
3606 #ifdef CONFIG_KVM_MMIO
3607         r = kvm_coalesced_mmio_init(kvm);
3608         if (r < 0)
3609                 goto put_kvm;
3610 #endif
3611         r = get_unused_fd_flags(O_CLOEXEC);
3612         if (r < 0)
3613                 goto put_kvm;
3614
3615         file = anon_inode_getfile("kvm-vm", &kvm_vm_fops, kvm, O_RDWR);
3616         if (IS_ERR(file)) {
3617                 put_unused_fd(r);
3618                 r = PTR_ERR(file);
3619                 goto put_kvm;
3620         }
3621
3622         /*
3623          * Don't call kvm_put_kvm anymore at this point; file->f_op is
3624          * already set, with ->release() being kvm_vm_release().  In error
3625          * cases it will be called by the final fput(file) and will take
3626          * care of doing kvm_put_kvm(kvm).
3627          */
3628         if (kvm_create_vm_debugfs(kvm, r) < 0) {
3629                 put_unused_fd(r);
3630                 fput(file);
3631                 return -ENOMEM;
3632         }
3633         kvm_uevent_notify_change(KVM_EVENT_CREATE_VM, kvm);
3634
3635         fd_install(r, file);
3636         return r;
3637
3638 put_kvm:
3639         kvm_put_kvm(kvm);
3640         return r;
3641 }
3642
3643 static long kvm_dev_ioctl(struct file *filp,
3644                           unsigned int ioctl, unsigned long arg)
3645 {
3646         long r = -EINVAL;
3647
3648         switch (ioctl) {
3649         case KVM_GET_API_VERSION:
3650                 if (arg)
3651                         goto out;
3652                 r = KVM_API_VERSION;
3653                 break;
3654         case KVM_CREATE_VM:
3655                 r = kvm_dev_ioctl_create_vm(arg);
3656                 break;
3657         case KVM_CHECK_EXTENSION:
3658                 r = kvm_vm_ioctl_check_extension_generic(NULL, arg);
3659                 break;
3660         case KVM_GET_VCPU_MMAP_SIZE:
3661                 if (arg)
3662                         goto out;
3663                 r = PAGE_SIZE;     /* struct kvm_run */
3664 #ifdef CONFIG_X86
3665                 r += PAGE_SIZE;    /* pio data page */
3666 #endif
3667 #ifdef CONFIG_KVM_MMIO
3668                 r += PAGE_SIZE;    /* coalesced mmio ring page */
3669 #endif
3670                 break;
3671         case KVM_TRACE_ENABLE:
3672         case KVM_TRACE_PAUSE:
3673         case KVM_TRACE_DISABLE:
3674                 r = -EOPNOTSUPP;
3675                 break;
3676         default:
3677                 return kvm_arch_dev_ioctl(filp, ioctl, arg);
3678         }
3679 out:
3680         return r;
3681 }
3682
3683 static struct file_operations kvm_chardev_ops = {
3684         .unlocked_ioctl = kvm_dev_ioctl,
3685         .llseek         = noop_llseek,
3686         KVM_COMPAT(kvm_dev_ioctl),
3687 };
3688
3689 static struct miscdevice kvm_dev = {
3690         KVM_MINOR,
3691         "kvm",
3692         &kvm_chardev_ops,
3693 };
3694
3695 static void hardware_enable_nolock(void *junk)
3696 {
3697         int cpu = raw_smp_processor_id();
3698         int r;
3699
3700         if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
3701                 return;
3702
3703         cpumask_set_cpu(cpu, cpus_hardware_enabled);
3704
3705         r = kvm_arch_hardware_enable();
3706
3707         if (r) {
3708                 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3709                 atomic_inc(&hardware_enable_failed);
3710                 pr_info("kvm: enabling virtualization on CPU%d failed\n", cpu);
3711         }
3712 }
3713
3714 static int kvm_starting_cpu(unsigned int cpu)
3715 {
3716         raw_spin_lock(&kvm_count_lock);
3717         if (kvm_usage_count)
3718                 hardware_enable_nolock(NULL);
3719         raw_spin_unlock(&kvm_count_lock);
3720         return 0;
3721 }
3722
3723 static void hardware_disable_nolock(void *junk)
3724 {
3725         int cpu = raw_smp_processor_id();
3726
3727         if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
3728                 return;
3729         cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3730         kvm_arch_hardware_disable();
3731 }
3732
3733 static int kvm_dying_cpu(unsigned int cpu)
3734 {
3735         raw_spin_lock(&kvm_count_lock);
3736         if (kvm_usage_count)
3737                 hardware_disable_nolock(NULL);
3738         raw_spin_unlock(&kvm_count_lock);
3739         return 0;
3740 }
3741
3742 static void hardware_disable_all_nolock(void)
3743 {
3744         BUG_ON(!kvm_usage_count);
3745
3746         kvm_usage_count--;
3747         if (!kvm_usage_count)
3748                 on_each_cpu(hardware_disable_nolock, NULL, 1);
3749 }
3750
3751 static void hardware_disable_all(void)
3752 {
3753         raw_spin_lock(&kvm_count_lock);
3754         hardware_disable_all_nolock();
3755         raw_spin_unlock(&kvm_count_lock);
3756 }
3757
3758 static int hardware_enable_all(void)
3759 {
3760         int r = 0;
3761
3762         raw_spin_lock(&kvm_count_lock);
3763
3764         kvm_usage_count++;
3765         if (kvm_usage_count == 1) {
3766                 atomic_set(&hardware_enable_failed, 0);
3767                 on_each_cpu(hardware_enable_nolock, NULL, 1);
3768
3769                 if (atomic_read(&hardware_enable_failed)) {
3770                         hardware_disable_all_nolock();
3771                         r = -EBUSY;
3772                 }
3773         }
3774
3775         raw_spin_unlock(&kvm_count_lock);
3776
3777         return r;
3778 }
3779
3780 static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
3781                       void *v)
3782 {
3783         /*
3784          * Some (well, at least mine) BIOSes hang on reboot if
3785          * in vmx root mode.
3786          *
3787          * And Intel TXT required VMX off for all cpu when system shutdown.
3788          */
3789         pr_info("kvm: exiting hardware virtualization\n");
3790         kvm_rebooting = true;
3791         on_each_cpu(hardware_disable_nolock, NULL, 1);
3792         return NOTIFY_OK;
3793 }
3794
3795 static struct notifier_block kvm_reboot_notifier = {
3796         .notifier_call = kvm_reboot,
3797         .priority = 0,
3798 };
3799
3800 static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
3801 {
3802         int i;
3803
3804         for (i = 0; i < bus->dev_count; i++) {
3805                 struct kvm_io_device *pos = bus->range[i].dev;
3806
3807                 kvm_iodevice_destructor(pos);
3808         }
3809         kfree(bus);
3810 }
3811
3812 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1,
3813                                  const struct kvm_io_range *r2)
3814 {
3815         gpa_t addr1 = r1->addr;
3816         gpa_t addr2 = r2->addr;
3817
3818         if (addr1 < addr2)
3819                 return -1;
3820
3821         /* If r2->len == 0, match the exact address.  If r2->len != 0,
3822          * accept any overlapping write.  Any order is acceptable for
3823          * overlapping ranges, because kvm_io_bus_get_first_dev ensures
3824          * we process all of them.
3825          */
3826         if (r2->len) {
3827                 addr1 += r1->len;
3828                 addr2 += r2->len;
3829         }
3830
3831         if (addr1 > addr2)
3832                 return 1;
3833
3834         return 0;
3835 }
3836
3837 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
3838 {
3839         return kvm_io_bus_cmp(p1, p2);
3840 }
3841
3842 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
3843                              gpa_t addr, int len)
3844 {
3845         struct kvm_io_range *range, key;
3846         int off;
3847
3848         key = (struct kvm_io_range) {
3849                 .addr = addr,
3850                 .len = len,
3851         };
3852
3853         range = bsearch(&key, bus->range, bus->dev_count,
3854                         sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
3855         if (range == NULL)
3856                 return -ENOENT;
3857
3858         off = range - bus->range;
3859
3860         while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0)
3861                 off--;
3862
3863         return off;
3864 }
3865
3866 static int __kvm_io_bus_write(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3867                               struct kvm_io_range *range, const void *val)
3868 {
3869         int idx;
3870
3871         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3872         if (idx < 0)
3873                 return -EOPNOTSUPP;
3874
3875         while (idx < bus->dev_count &&
3876                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3877                 if (!kvm_iodevice_write(vcpu, bus->range[idx].dev, range->addr,
3878                                         range->len, val))
3879                         return idx;
3880                 idx++;
3881         }
3882
3883         return -EOPNOTSUPP;
3884 }
3885
3886 /* kvm_io_bus_write - called under kvm->slots_lock */
3887 int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3888                      int len, const void *val)
3889 {
3890         struct kvm_io_bus *bus;
3891         struct kvm_io_range range;
3892         int r;
3893
3894         range = (struct kvm_io_range) {
3895                 .addr = addr,
3896                 .len = len,
3897         };
3898
3899         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3900         if (!bus)
3901                 return -ENOMEM;
3902         r = __kvm_io_bus_write(vcpu, bus, &range, val);
3903         return r < 0 ? r : 0;
3904 }
3905 EXPORT_SYMBOL_GPL(kvm_io_bus_write);
3906
3907 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */
3908 int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
3909                             gpa_t addr, int len, const void *val, long cookie)
3910 {
3911         struct kvm_io_bus *bus;
3912         struct kvm_io_range range;
3913
3914         range = (struct kvm_io_range) {
3915                 .addr = addr,
3916                 .len = len,
3917         };
3918
3919         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3920         if (!bus)
3921                 return -ENOMEM;
3922
3923         /* First try the device referenced by cookie. */
3924         if ((cookie >= 0) && (cookie < bus->dev_count) &&
3925             (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0))
3926                 if (!kvm_iodevice_write(vcpu, bus->range[cookie].dev, addr, len,
3927                                         val))
3928                         return cookie;
3929
3930         /*
3931          * cookie contained garbage; fall back to search and return the
3932          * correct cookie value.
3933          */
3934         return __kvm_io_bus_write(vcpu, bus, &range, val);
3935 }
3936
3937 static int __kvm_io_bus_read(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3938                              struct kvm_io_range *range, void *val)
3939 {
3940         int idx;
3941
3942         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3943         if (idx < 0)
3944                 return -EOPNOTSUPP;
3945
3946         while (idx < bus->dev_count &&
3947                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3948                 if (!kvm_iodevice_read(vcpu, bus->range[idx].dev, range->addr,
3949                                        range->len, val))
3950                         return idx;
3951                 idx++;
3952         }
3953
3954         return -EOPNOTSUPP;
3955 }
3956
3957 /* kvm_io_bus_read - called under kvm->slots_lock */
3958 int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3959                     int len, void *val)
3960 {
3961         struct kvm_io_bus *bus;
3962         struct kvm_io_range range;
3963         int r;
3964
3965         range = (struct kvm_io_range) {
3966                 .addr = addr,
3967                 .len = len,
3968         };
3969
3970         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3971         if (!bus)
3972                 return -ENOMEM;
3973         r = __kvm_io_bus_read(vcpu, bus, &range, val);
3974         return r < 0 ? r : 0;
3975 }
3976
3977 /* Caller must hold slots_lock. */
3978 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
3979                             int len, struct kvm_io_device *dev)
3980 {
3981         int i;
3982         struct kvm_io_bus *new_bus, *bus;
3983         struct kvm_io_range range;
3984
3985         bus = kvm_get_bus(kvm, bus_idx);
3986         if (!bus)
3987                 return -ENOMEM;
3988
3989         /* exclude ioeventfd which is limited by maximum fd */
3990         if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
3991                 return -ENOSPC;
3992
3993         new_bus = kmalloc(struct_size(bus, range, bus->dev_count + 1),
3994                           GFP_KERNEL_ACCOUNT);
3995         if (!new_bus)
3996                 return -ENOMEM;
3997
3998         range = (struct kvm_io_range) {
3999                 .addr = addr,
4000                 .len = len,
4001                 .dev = dev,
4002         };
4003
4004         for (i = 0; i < bus->dev_count; i++)
4005                 if (kvm_io_bus_cmp(&bus->range[i], &range) > 0)
4006                         break;
4007
4008         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
4009         new_bus->dev_count++;
4010         new_bus->range[i] = range;
4011         memcpy(new_bus->range + i + 1, bus->range + i,
4012                 (bus->dev_count - i) * sizeof(struct kvm_io_range));
4013         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
4014         synchronize_srcu_expedited(&kvm->srcu);
4015         kfree(bus);
4016
4017         return 0;
4018 }
4019
4020 /* Caller must hold slots_lock. */
4021 void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
4022                                struct kvm_io_device *dev)
4023 {
4024         int i;
4025         struct kvm_io_bus *new_bus, *bus;
4026
4027         bus = kvm_get_bus(kvm, bus_idx);
4028         if (!bus)
4029                 return;
4030
4031         for (i = 0; i < bus->dev_count; i++)
4032                 if (bus->range[i].dev == dev) {
4033                         break;
4034                 }
4035
4036         if (i == bus->dev_count)
4037                 return;
4038
4039         new_bus = kmalloc(struct_size(bus, range, bus->dev_count - 1),
4040                           GFP_KERNEL_ACCOUNT);
4041         if (!new_bus)  {
4042                 pr_err("kvm: failed to shrink bus, removing it completely\n");
4043                 goto broken;
4044         }
4045
4046         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
4047         new_bus->dev_count--;
4048         memcpy(new_bus->range + i, bus->range + i + 1,
4049                (new_bus->dev_count - i) * sizeof(struct kvm_io_range));
4050
4051 broken:
4052         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
4053         synchronize_srcu_expedited(&kvm->srcu);
4054         kfree(bus);
4055         return;
4056 }
4057
4058 struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx,
4059                                          gpa_t addr)
4060 {
4061         struct kvm_io_bus *bus;
4062         int dev_idx, srcu_idx;
4063         struct kvm_io_device *iodev = NULL;
4064
4065         srcu_idx = srcu_read_lock(&kvm->srcu);
4066
4067         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
4068         if (!bus)
4069                 goto out_unlock;
4070
4071         dev_idx = kvm_io_bus_get_first_dev(bus, addr, 1);
4072         if (dev_idx < 0)
4073                 goto out_unlock;
4074
4075         iodev = bus->range[dev_idx].dev;
4076
4077 out_unlock:
4078         srcu_read_unlock(&kvm->srcu, srcu_idx);
4079
4080         return iodev;
4081 }
4082 EXPORT_SYMBOL_GPL(kvm_io_bus_get_dev);
4083
4084 static int kvm_debugfs_open(struct inode *inode, struct file *file,
4085                            int (*get)(void *, u64 *), int (*set)(void *, u64),
4086                            const char *fmt)
4087 {
4088         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)
4089                                           inode->i_private;
4090
4091         /* The debugfs files are a reference to the kvm struct which
4092          * is still valid when kvm_destroy_vm is called.
4093          * To avoid the race between open and the removal of the debugfs
4094          * directory we test against the users count.
4095          */
4096         if (!refcount_inc_not_zero(&stat_data->kvm->users_count))
4097                 return -ENOENT;
4098
4099         if (simple_attr_open(inode, file, get,
4100                     KVM_DBGFS_GET_MODE(stat_data->dbgfs_item) & 0222
4101                     ? set : NULL,
4102                     fmt)) {
4103                 kvm_put_kvm(stat_data->kvm);
4104                 return -ENOMEM;
4105         }
4106
4107         return 0;
4108 }
4109
4110 static int kvm_debugfs_release(struct inode *inode, struct file *file)
4111 {
4112         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)
4113                                           inode->i_private;
4114
4115         simple_attr_release(inode, file);
4116         kvm_put_kvm(stat_data->kvm);
4117
4118         return 0;
4119 }
4120
4121 static int kvm_get_stat_per_vm(struct kvm *kvm, size_t offset, u64 *val)
4122 {
4123         *val = *(ulong *)((void *)kvm + offset);
4124
4125         return 0;
4126 }
4127
4128 static int kvm_clear_stat_per_vm(struct kvm *kvm, size_t offset)
4129 {
4130         *(ulong *)((void *)kvm + offset) = 0;
4131
4132         return 0;
4133 }
4134
4135 static int kvm_get_stat_per_vcpu(struct kvm *kvm, size_t offset, u64 *val)
4136 {
4137         int i;
4138         struct kvm_vcpu *vcpu;
4139
4140         *val = 0;
4141
4142         kvm_for_each_vcpu(i, vcpu, kvm)
4143                 *val += *(u64 *)((void *)vcpu + offset);
4144
4145         return 0;
4146 }
4147
4148 static int kvm_clear_stat_per_vcpu(struct kvm *kvm, size_t offset)
4149 {
4150         int i;
4151         struct kvm_vcpu *vcpu;
4152
4153         kvm_for_each_vcpu(i, vcpu, kvm)
4154                 *(u64 *)((void *)vcpu + offset) = 0;
4155
4156         return 0;
4157 }
4158
4159 static int kvm_stat_data_get(void *data, u64 *val)
4160 {
4161         int r = -EFAULT;
4162         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data;
4163
4164         switch (stat_data->dbgfs_item->kind) {
4165         case KVM_STAT_VM:
4166                 r = kvm_get_stat_per_vm(stat_data->kvm,
4167                                         stat_data->dbgfs_item->offset, val);
4168                 break;
4169         case KVM_STAT_VCPU:
4170                 r = kvm_get_stat_per_vcpu(stat_data->kvm,
4171                                           stat_data->dbgfs_item->offset, val);
4172                 break;
4173         }
4174
4175         return r;
4176 }
4177
4178 static int kvm_stat_data_clear(void *data, u64 val)
4179 {
4180         int r = -EFAULT;
4181         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data;
4182
4183         if (val)
4184                 return -EINVAL;
4185
4186         switch (stat_data->dbgfs_item->kind) {
4187         case KVM_STAT_VM:
4188                 r = kvm_clear_stat_per_vm(stat_data->kvm,
4189                                           stat_data->dbgfs_item->offset);
4190                 break;
4191         case KVM_STAT_VCPU:
4192                 r = kvm_clear_stat_per_vcpu(stat_data->kvm,
4193                                             stat_data->dbgfs_item->offset);
4194                 break;
4195         }
4196
4197         return r;
4198 }
4199
4200 static int kvm_stat_data_open(struct inode *inode, struct file *file)
4201 {
4202         __simple_attr_check_format("%llu\n", 0ull);
4203         return kvm_debugfs_open(inode, file, kvm_stat_data_get,
4204                                 kvm_stat_data_clear, "%llu\n");
4205 }
4206
4207 static const struct file_operations stat_fops_per_vm = {
4208         .owner = THIS_MODULE,
4209         .open = kvm_stat_data_open,
4210         .release = kvm_debugfs_release,
4211         .read = simple_attr_read,
4212         .write = simple_attr_write,
4213         .llseek = no_llseek,
4214 };
4215
4216 static int vm_stat_get(void *_offset, u64 *val)
4217 {
4218         unsigned offset = (long)_offset;
4219         struct kvm *kvm;
4220         u64 tmp_val;
4221
4222         *val = 0;
4223         mutex_lock(&kvm_lock);
4224         list_for_each_entry(kvm, &vm_list, vm_list) {
4225                 kvm_get_stat_per_vm(kvm, offset, &tmp_val);
4226                 *val += tmp_val;
4227         }
4228         mutex_unlock(&kvm_lock);
4229         return 0;
4230 }
4231
4232 static int vm_stat_clear(void *_offset, u64 val)
4233 {
4234         unsigned offset = (long)_offset;
4235         struct kvm *kvm;
4236
4237         if (val)
4238                 return -EINVAL;
4239
4240         mutex_lock(&kvm_lock);
4241         list_for_each_entry(kvm, &vm_list, vm_list) {
4242                 kvm_clear_stat_per_vm(kvm, offset);
4243         }
4244         mutex_unlock(&kvm_lock);
4245
4246         return 0;
4247 }
4248
4249 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, vm_stat_clear, "%llu\n");
4250
4251 static int vcpu_stat_get(void *_offset, u64 *val)
4252 {
4253         unsigned offset = (long)_offset;
4254         struct kvm *kvm;
4255         u64 tmp_val;
4256
4257         *val = 0;
4258         mutex_lock(&kvm_lock);
4259         list_for_each_entry(kvm, &vm_list, vm_list) {
4260                 kvm_get_stat_per_vcpu(kvm, offset, &tmp_val);
4261                 *val += tmp_val;
4262         }
4263         mutex_unlock(&kvm_lock);
4264         return 0;
4265 }
4266
4267 static int vcpu_stat_clear(void *_offset, u64 val)
4268 {
4269         unsigned offset = (long)_offset;
4270         struct kvm *kvm;
4271
4272         if (val)
4273                 return -EINVAL;
4274
4275         mutex_lock(&kvm_lock);
4276         list_for_each_entry(kvm, &vm_list, vm_list) {
4277                 kvm_clear_stat_per_vcpu(kvm, offset);
4278         }
4279         mutex_unlock(&kvm_lock);
4280
4281         return 0;
4282 }
4283
4284 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, vcpu_stat_clear,
4285                         "%llu\n");
4286
4287 static const struct file_operations *stat_fops[] = {
4288         [KVM_STAT_VCPU] = &vcpu_stat_fops,
4289         [KVM_STAT_VM]   = &vm_stat_fops,
4290 };
4291
4292 static void kvm_uevent_notify_change(unsigned int type, struct kvm *kvm)
4293 {
4294         struct kobj_uevent_env *env;
4295         unsigned long long created, active;
4296
4297         if (!kvm_dev.this_device || !kvm)
4298                 return;
4299
4300         mutex_lock(&kvm_lock);
4301         if (type == KVM_EVENT_CREATE_VM) {
4302                 kvm_createvm_count++;
4303                 kvm_active_vms++;
4304         } else if (type == KVM_EVENT_DESTROY_VM) {
4305                 kvm_active_vms--;
4306         }
4307         created = kvm_createvm_count;
4308         active = kvm_active_vms;
4309         mutex_unlock(&kvm_lock);
4310
4311         env = kzalloc(sizeof(*env), GFP_KERNEL_ACCOUNT);
4312         if (!env)
4313                 return;
4314
4315         add_uevent_var(env, "CREATED=%llu", created);
4316         add_uevent_var(env, "COUNT=%llu", active);
4317
4318         if (type == KVM_EVENT_CREATE_VM) {
4319                 add_uevent_var(env, "EVENT=create");
4320                 kvm->userspace_pid = task_pid_nr(current);
4321         } else if (type == KVM_EVENT_DESTROY_VM) {
4322                 add_uevent_var(env, "EVENT=destroy");
4323         }
4324         add_uevent_var(env, "PID=%d", kvm->userspace_pid);
4325
4326         if (!IS_ERR_OR_NULL(kvm->debugfs_dentry)) {
4327                 char *tmp, *p = kmalloc(PATH_MAX, GFP_KERNEL_ACCOUNT);
4328
4329                 if (p) {
4330                         tmp = dentry_path_raw(kvm->debugfs_dentry, p, PATH_MAX);
4331                         if (!IS_ERR(tmp))
4332                                 add_uevent_var(env, "STATS_PATH=%s", tmp);
4333                         kfree(p);
4334                 }
4335         }
4336         /* no need for checks, since we are adding at most only 5 keys */
4337         env->envp[env->envp_idx++] = NULL;
4338         kobject_uevent_env(&kvm_dev.this_device->kobj, KOBJ_CHANGE, env->envp);
4339         kfree(env);
4340 }
4341
4342 static void kvm_init_debug(void)
4343 {
4344         struct kvm_stats_debugfs_item *p;
4345
4346         kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
4347
4348         kvm_debugfs_num_entries = 0;
4349         for (p = debugfs_entries; p->name; ++p, kvm_debugfs_num_entries++) {
4350                 debugfs_create_file(p->name, KVM_DBGFS_GET_MODE(p),
4351                                     kvm_debugfs_dir, (void *)(long)p->offset,
4352                                     stat_fops[p->kind]);
4353         }
4354 }
4355
4356 static int kvm_suspend(void)
4357 {
4358         if (kvm_usage_count)
4359                 hardware_disable_nolock(NULL);
4360         return 0;
4361 }
4362
4363 static void kvm_resume(void)
4364 {
4365         if (kvm_usage_count) {
4366 #ifdef CONFIG_LOCKDEP
4367                 WARN_ON(lockdep_is_held(&kvm_count_lock));
4368 #endif
4369                 hardware_enable_nolock(NULL);
4370         }
4371 }
4372
4373 static struct syscore_ops kvm_syscore_ops = {
4374         .suspend = kvm_suspend,
4375         .resume = kvm_resume,
4376 };
4377
4378 static inline
4379 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
4380 {
4381         return container_of(pn, struct kvm_vcpu, preempt_notifier);
4382 }
4383
4384 static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
4385 {
4386         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
4387
4388         WRITE_ONCE(vcpu->preempted, false);
4389         WRITE_ONCE(vcpu->ready, false);
4390
4391         __this_cpu_write(kvm_running_vcpu, vcpu);
4392         kvm_arch_sched_in(vcpu, cpu);
4393         kvm_arch_vcpu_load(vcpu, cpu);
4394 }
4395
4396 static void kvm_sched_out(struct preempt_notifier *pn,
4397                           struct task_struct *next)
4398 {
4399         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
4400
4401         if (current->state == TASK_RUNNING) {
4402                 WRITE_ONCE(vcpu->preempted, true);
4403                 WRITE_ONCE(vcpu->ready, true);
4404         }
4405         kvm_arch_vcpu_put(vcpu);
4406         __this_cpu_write(kvm_running_vcpu, NULL);
4407 }
4408
4409 /**
4410  * kvm_get_running_vcpu - get the vcpu running on the current CPU.
4411  * Thanks to preempt notifiers, this can also be called from
4412  * preemptible context.
4413  */
4414 struct kvm_vcpu *kvm_get_running_vcpu(void)
4415 {
4416         return __this_cpu_read(kvm_running_vcpu);
4417 }
4418
4419 /**
4420  * kvm_get_running_vcpus - get the per-CPU array of currently running vcpus.
4421  */
4422 struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void)
4423 {
4424         return &kvm_running_vcpu;
4425 }
4426
4427 static void check_processor_compat(void *rtn)
4428 {
4429         *(int *)rtn = kvm_arch_check_processor_compat();
4430 }
4431
4432 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
4433                   struct module *module)
4434 {
4435         int r;
4436         int cpu;
4437
4438         r = kvm_arch_init(opaque);
4439         if (r)
4440                 goto out_fail;
4441
4442         /*
4443          * kvm_arch_init makes sure there's at most one caller
4444          * for architectures that support multiple implementations,
4445          * like intel and amd on x86.
4446          * kvm_arch_init must be called before kvm_irqfd_init to avoid creating
4447          * conflicts in case kvm is already setup for another implementation.
4448          */
4449         r = kvm_irqfd_init();
4450         if (r)
4451                 goto out_irqfd;
4452
4453         if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
4454                 r = -ENOMEM;
4455                 goto out_free_0;
4456         }
4457
4458         r = kvm_arch_hardware_setup();
4459         if (r < 0)
4460                 goto out_free_1;
4461
4462         for_each_online_cpu(cpu) {
4463                 smp_call_function_single(cpu, check_processor_compat, &r, 1);
4464                 if (r < 0)
4465                         goto out_free_2;
4466         }
4467
4468         r = cpuhp_setup_state_nocalls(CPUHP_AP_KVM_STARTING, "kvm/cpu:starting",
4469                                       kvm_starting_cpu, kvm_dying_cpu);
4470         if (r)
4471                 goto out_free_2;
4472         register_reboot_notifier(&kvm_reboot_notifier);
4473
4474         /* A kmem cache lets us meet the alignment requirements of fx_save. */
4475         if (!vcpu_align)
4476                 vcpu_align = __alignof__(struct kvm_vcpu);
4477         kvm_vcpu_cache =
4478                 kmem_cache_create_usercopy("kvm_vcpu", vcpu_size, vcpu_align,
4479                                            SLAB_ACCOUNT,
4480                                            offsetof(struct kvm_vcpu, arch),
4481                                            sizeof_field(struct kvm_vcpu, arch),
4482                                            NULL);
4483         if (!kvm_vcpu_cache) {
4484                 r = -ENOMEM;
4485                 goto out_free_3;
4486         }
4487
4488         r = kvm_async_pf_init();
4489         if (r)
4490                 goto out_free;
4491
4492         kvm_chardev_ops.owner = module;
4493         kvm_vm_fops.owner = module;
4494         kvm_vcpu_fops.owner = module;
4495
4496         r = misc_register(&kvm_dev);
4497         if (r) {
4498                 pr_err("kvm: misc device register failed\n");
4499                 goto out_unreg;
4500         }
4501
4502         register_syscore_ops(&kvm_syscore_ops);
4503
4504         kvm_preempt_ops.sched_in = kvm_sched_in;
4505         kvm_preempt_ops.sched_out = kvm_sched_out;
4506
4507         kvm_init_debug();
4508
4509         r = kvm_vfio_ops_init();
4510         WARN_ON(r);
4511
4512         return 0;
4513
4514 out_unreg:
4515         kvm_async_pf_deinit();
4516 out_free:
4517         kmem_cache_destroy(kvm_vcpu_cache);
4518 out_free_3:
4519         unregister_reboot_notifier(&kvm_reboot_notifier);
4520         cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING);
4521 out_free_2:
4522         kvm_arch_hardware_unsetup();
4523 out_free_1:
4524         free_cpumask_var(cpus_hardware_enabled);
4525 out_free_0:
4526         kvm_irqfd_exit();
4527 out_irqfd:
4528         kvm_arch_exit();
4529 out_fail:
4530         return r;
4531 }
4532 EXPORT_SYMBOL_GPL(kvm_init);
4533
4534 void kvm_exit(void)
4535 {
4536         debugfs_remove_recursive(kvm_debugfs_dir);
4537         misc_deregister(&kvm_dev);
4538         kmem_cache_destroy(kvm_vcpu_cache);
4539         kvm_async_pf_deinit();
4540         unregister_syscore_ops(&kvm_syscore_ops);
4541         unregister_reboot_notifier(&kvm_reboot_notifier);
4542         cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING);
4543         on_each_cpu(hardware_disable_nolock, NULL, 1);
4544         kvm_arch_hardware_unsetup();
4545         kvm_arch_exit();
4546         kvm_irqfd_exit();
4547         free_cpumask_var(cpus_hardware_enabled);
4548         kvm_vfio_ops_exit();
4549 }
4550 EXPORT_SYMBOL_GPL(kvm_exit);
4551
4552 struct kvm_vm_worker_thread_context {
4553         struct kvm *kvm;
4554         struct task_struct *parent;
4555         struct completion init_done;
4556         kvm_vm_thread_fn_t thread_fn;
4557         uintptr_t data;
4558         int err;
4559 };
4560
4561 static int kvm_vm_worker_thread(void *context)
4562 {
4563         /*
4564          * The init_context is allocated on the stack of the parent thread, so
4565          * we have to locally copy anything that is needed beyond initialization
4566          */
4567         struct kvm_vm_worker_thread_context *init_context = context;
4568         struct kvm *kvm = init_context->kvm;
4569         kvm_vm_thread_fn_t thread_fn = init_context->thread_fn;
4570         uintptr_t data = init_context->data;
4571         int err;
4572
4573         err = kthread_park(current);
4574         /* kthread_park(current) is never supposed to return an error */
4575         WARN_ON(err != 0);
4576         if (err)
4577                 goto init_complete;
4578
4579         err = cgroup_attach_task_all(init_context->parent, current);
4580         if (err) {
4581                 kvm_err("%s: cgroup_attach_task_all failed with err %d\n",
4582                         __func__, err);
4583                 goto init_complete;
4584         }
4585
4586         set_user_nice(current, task_nice(init_context->parent));
4587
4588 init_complete:
4589         init_context->err = err;
4590         complete(&init_context->init_done);
4591         init_context = NULL;
4592
4593         if (err)
4594                 return err;
4595
4596         /* Wait to be woken up by the spawner before proceeding. */
4597         kthread_parkme();
4598
4599         if (!kthread_should_stop())
4600                 err = thread_fn(kvm, data);
4601
4602         return err;
4603 }
4604
4605 int kvm_vm_create_worker_thread(struct kvm *kvm, kvm_vm_thread_fn_t thread_fn,
4606                                 uintptr_t data, const char *name,
4607                                 struct task_struct **thread_ptr)
4608 {
4609         struct kvm_vm_worker_thread_context init_context = {};
4610         struct task_struct *thread;
4611
4612         *thread_ptr = NULL;
4613         init_context.kvm = kvm;
4614         init_context.parent = current;
4615         init_context.thread_fn = thread_fn;
4616         init_context.data = data;
4617         init_completion(&init_context.init_done);
4618
4619         thread = kthread_run(kvm_vm_worker_thread, &init_context,
4620                              "%s-%d", name, task_pid_nr(current));
4621         if (IS_ERR(thread))
4622                 return PTR_ERR(thread);
4623
4624         /* kthread_run is never supposed to return NULL */
4625         WARN_ON(thread == NULL);
4626
4627         wait_for_completion(&init_context.init_done);
4628
4629         if (!init_context.err)
4630                 *thread_ptr = thread;
4631
4632         return init_context.err;
4633 }