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                        !is_zero_pfn(pfn) &&
 190                        !kvm_is_zone_device_pfn(pfn);
 191
 192         return true;
 193 }
 194
 195 bool kvm_is_transparent_hugepage(kvm_pfn_t pfn)
 196 {
 197         struct page *page = pfn_to_page(pfn);
 198
 199         if (!PageTransCompoundMap(page))
 200                 return false;
 201
 202         return is_transparent_hugepage(compound_head(page));
 203 }
 204
 205 /*
 206  * Switches to specified vcpu, until a matching vcpu_put()
 207  */
 208 void vcpu_load(struct kvm_vcpu *vcpu)
 209 {
 210         int cpu = get_cpu();
 211
 212         __this_cpu_write(kvm_running_vcpu, vcpu);
 213         preempt_notifier_register(&vcpu->preempt_notifier);
 214         kvm_arch_vcpu_load(vcpu, cpu);
 215         put_cpu();
 216 }
 217 EXPORT_SYMBOL_GPL(vcpu_load);
 218
 219 void vcpu_put(struct kvm_vcpu *vcpu)
 220 {
 221         preempt_disable();
 222         kvm_arch_vcpu_put(vcpu);
 223         preempt_notifier_unregister(&vcpu->preempt_notifier);
 224         __this_cpu_write(kvm_running_vcpu, NULL);
 225         preempt_enable();
 226 }
 227 EXPORT_SYMBOL_GPL(vcpu_put);
 228
 229 /* TODO: merge with kvm_arch_vcpu_should_kick */
 230 static bool kvm_request_needs_ipi(struct kvm_vcpu *vcpu, unsigned req)
 231 {
 232         int mode = kvm_vcpu_exiting_guest_mode(vcpu);
 233
 234         /*
 235          * We need to wait for the VCPU to reenable interrupts and get out of
 236          * READING_SHADOW_PAGE_TABLES mode.
 237          */
 238         if (req & KVM_REQUEST_WAIT)
 239                 return mode != OUTSIDE_GUEST_MODE;
 240
 241         /*
 242          * Need to kick a running VCPU, but otherwise there is nothing to do.
 243          */
 244         return mode == IN_GUEST_MODE;
 245 }
 246
 247 static void ack_flush(void *_completed)
 248 {
 249 }
 250
 251 static inline bool kvm_kick_many_cpus(const struct cpumask *cpus, bool wait)
 252 {
 253         if (unlikely(!cpus))
 254                 cpus = cpu_online_mask;
 255
 256         if (cpumask_empty(cpus))
 257                 return false;
 258
 259         smp_call_function_many(cpus, ack_flush, NULL, wait);
 260         return true;
 261 }
 262
 263 bool kvm_make_vcpus_request_mask(struct kvm *kvm, unsigned int req,
 264                                  unsigned long *vcpu_bitmap, cpumask_var_t tmp)
 265 {
 266         int i, cpu, me;
 267         struct kvm_vcpu *vcpu;
 268         bool called;
 269
 270         me = get_cpu();
 271
 272         kvm_for_each_vcpu(i, vcpu, kvm) {
 273                 if (vcpu_bitmap && !test_bit(i, vcpu_bitmap))
 274                         continue;
 275
 276                 kvm_make_request(req, vcpu);
 277                 cpu = vcpu->cpu;
 278
 279                 if (!(req & KVM_REQUEST_NO_WAKEUP) && kvm_vcpu_wake_up(vcpu))
 280                         continue;
 281
 282                 if (tmp != NULL && cpu != -1 && cpu != me &&
 283                     kvm_request_needs_ipi(vcpu, req))
 284                         __cpumask_set_cpu(cpu, tmp);
 285         }
 286
 287         called = kvm_kick_many_cpus(tmp, !!(req & KVM_REQUEST_WAIT));
 288         put_cpu();
 289
 290         return called;
 291 }
 292
 293 bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req)
 294 {
 295         cpumask_var_t cpus;
 296         bool called;
 297
 298         zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 299
 300         called = kvm_make_vcpus_request_mask(kvm, req, NULL, cpus);
 301
 302         free_cpumask_var(cpus);
 303         return called;
 304 }
 305
 306 #ifndef CONFIG_HAVE_KVM_ARCH_TLB_FLUSH_ALL
 307 void kvm_flush_remote_tlbs(struct kvm *kvm)
 308 {
 309         /*
 310          * Read tlbs_dirty before setting KVM_REQ_TLB_FLUSH in
 311          * kvm_make_all_cpus_request.
 312          */
 313         long dirty_count = smp_load_acquire(&kvm->tlbs_dirty);
 314
 315         /*
 316          * We want to publish modifications to the page tables before reading
 317          * mode. Pairs with a memory barrier in arch-specific code.
 318          * - x86: smp_mb__after_srcu_read_unlock in vcpu_enter_guest
 319          * and smp_mb in walk_shadow_page_lockless_begin/end.
 320          * - powerpc: smp_mb in kvmppc_prepare_to_enter.
 321          *
 322          * There is already an smp_mb__after_atomic() before
 323          * kvm_make_all_cpus_request() reads vcpu->mode. We reuse that
 324          * barrier here.
 325          */
 326         if (!kvm_arch_flush_remote_tlb(kvm)
 327             || kvm_make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
 328                 ++kvm->stat.remote_tlb_flush;
 329         cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 330 }
 331 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs);
 332 #endif
 333
 334 void kvm_reload_remote_mmus(struct kvm *kvm)
 335 {
 336         kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
 337 }
 338
 339 static void kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
 340 {
 341         mutex_init(&vcpu->mutex);
 342         vcpu->cpu = -1;
 343         vcpu->kvm = kvm;
 344         vcpu->vcpu_id = id;
 345         vcpu->pid = NULL;
 346         init_swait_queue_head(&vcpu->wq);
 347         kvm_async_pf_vcpu_init(vcpu);
 348
 349         vcpu->pre_pcpu = -1;
 350         INIT_LIST_HEAD(&vcpu->blocked_vcpu_list);
 351
 352         kvm_vcpu_set_in_spin_loop(vcpu, false);
 353         kvm_vcpu_set_dy_eligible(vcpu, false);
 354         vcpu->preempted = false;
 355         vcpu->ready = false;
 356         preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
 357 }
 358
 359 void kvm_vcpu_destroy(struct kvm_vcpu *vcpu)
 360 {
 361         kvm_arch_vcpu_destroy(vcpu);
 362
 363         /*
 364          * No need for rcu_read_lock as VCPU_RUN is the only place that changes
 365          * the vcpu->pid pointer, and at destruction time all file descriptors
 366          * are already gone.
 367          */
 368         put_pid(rcu_dereference_protected(vcpu->pid, 1));
 369
 370         free_page((unsigned long)vcpu->run);
 371         kmem_cache_free(kvm_vcpu_cache, vcpu);
 372 }
 373 EXPORT_SYMBOL_GPL(kvm_vcpu_destroy);
 374
 375 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 376 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
 377 {
 378         return container_of(mn, struct kvm, mmu_notifier);
 379 }
 380
 381 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
 382                                         struct mm_struct *mm,
 383                                         unsigned long address,
 384                                         pte_t pte)
 385 {
 386         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 387         int idx;
 388
 389         idx = srcu_read_lock(&kvm->srcu);
 390         spin_lock(&kvm->mmu_lock);
 391         kvm->mmu_notifier_seq++;
 392
 393         if (kvm_set_spte_hva(kvm, address, pte))
 394                 kvm_flush_remote_tlbs(kvm);
 395
 396         spin_unlock(&kvm->mmu_lock);
 397         srcu_read_unlock(&kvm->srcu, idx);
 398 }
 399
 400 static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
 401                                         const struct mmu_notifier_range *range)
 402 {
 403         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 404         int need_tlb_flush = 0, idx;
 405         int ret;
 406
 407         idx = srcu_read_lock(&kvm->srcu);
 408         spin_lock(&kvm->mmu_lock);
 409         /*
 410          * The count increase must become visible at unlock time as no
 411          * spte can be established without taking the mmu_lock and
 412          * count is also read inside the mmu_lock critical section.
 413          */
 414         kvm->mmu_notifier_count++;
 415         need_tlb_flush = kvm_unmap_hva_range(kvm, range->start, range->end);
 416         need_tlb_flush |= kvm->tlbs_dirty;
 417         /* we've to flush the tlb before the pages can be freed */
 418         if (need_tlb_flush)
 419                 kvm_flush_remote_tlbs(kvm);
 420
 421         spin_unlock(&kvm->mmu_lock);
 422
 423         ret = kvm_arch_mmu_notifier_invalidate_range(kvm, range->start,
 424                                         range->end,
 425                                         mmu_notifier_range_blockable(range));
 426
 427         srcu_read_unlock(&kvm->srcu, idx);
 428
 429         return ret;
 430 }
 431
 432 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
 433                                         const struct mmu_notifier_range *range)
 434 {
 435         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 436
 437         spin_lock(&kvm->mmu_lock);
 438         /*
 439          * This sequence increase will notify the kvm page fault that
 440          * the page that is going to be mapped in the spte could have
 441          * been freed.
 442          */
 443         kvm->mmu_notifier_seq++;
 444         smp_wmb();
 445         /*
 446          * The above sequence increase must be visible before the
 447          * below count decrease, which is ensured by the smp_wmb above
 448          * in conjunction with the smp_rmb in mmu_notifier_retry().
 449          */
 450         kvm->mmu_notifier_count--;
 451         spin_unlock(&kvm->mmu_lock);
 452
 453         BUG_ON(kvm->mmu_notifier_count < 0);
 454 }
 455
 456 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
 457                                               struct mm_struct *mm,
 458                                               unsigned long start,
 459                                               unsigned long end)
 460 {
 461         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 462         int young, idx;
 463
 464         idx = srcu_read_lock(&kvm->srcu);
 465         spin_lock(&kvm->mmu_lock);
 466
 467         young = kvm_age_hva(kvm, start, end);
 468         if (young)
 469                 kvm_flush_remote_tlbs(kvm);
 470
 471         spin_unlock(&kvm->mmu_lock);
 472         srcu_read_unlock(&kvm->srcu, idx);
 473
 474         return young;
 475 }
 476
 477 static int kvm_mmu_notifier_clear_young(struct mmu_notifier *mn,
 478                                         struct mm_struct *mm,
 479                                         unsigned long start,
 480                                         unsigned long end)
 481 {
 482         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 483         int young, idx;
 484
 485         idx = srcu_read_lock(&kvm->srcu);
 486         spin_lock(&kvm->mmu_lock);
 487         /*
 488          * Even though we do not flush TLB, this will still adversely
 489          * affect performance on pre-Haswell Intel EPT, where there is
 490          * no EPT Access Bit to clear so that we have to tear down EPT
 491          * tables instead. If we find this unacceptable, we can always
 492          * add a parameter to kvm_age_hva so that it effectively doesn't
 493          * do anything on clear_young.
 494          *
 495          * Also note that currently we never issue secondary TLB flushes
 496          * from clear_young, leaving this job up to the regular system
 497          * cadence. If we find this inaccurate, we might come up with a
 498          * more sophisticated heuristic later.
 499          */
 500         young = kvm_age_hva(kvm, start, end);
 501         spin_unlock(&kvm->mmu_lock);
 502         srcu_read_unlock(&kvm->srcu, idx);
 503
 504         return young;
 505 }
 506
 507 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn,
 508                                        struct mm_struct *mm,
 509                                        unsigned long address)
 510 {
 511         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 512         int young, idx;
 513
 514         idx = srcu_read_lock(&kvm->srcu);
 515         spin_lock(&kvm->mmu_lock);
 516         young = kvm_test_age_hva(kvm, address);
 517         spin_unlock(&kvm->mmu_lock);
 518         srcu_read_unlock(&kvm->srcu, idx);
 519
 520         return young;
 521 }
 522
 523 static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
 524                                      struct mm_struct *mm)
 525 {
 526         struct kvm *kvm = mmu_notifier_to_kvm(mn);
 527         int idx;
 528
 529         idx = srcu_read_lock(&kvm->srcu);
 530         kvm_arch_flush_shadow_all(kvm);
 531         srcu_read_unlock(&kvm->srcu, idx);
 532 }
 533
 534 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
 535         .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
 536         .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
 537         .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
 538         .clear_young            = kvm_mmu_notifier_clear_young,
 539         .test_young             = kvm_mmu_notifier_test_young,
 540         .change_pte             = kvm_mmu_notifier_change_pte,
 541         .release                = kvm_mmu_notifier_release,
 542 };
 543
 544 static int kvm_init_mmu_notifier(struct kvm *kvm)
 545 {
 546         kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
 547         return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
 548 }
 549
 550 #else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
 551
 552 static int kvm_init_mmu_notifier(struct kvm *kvm)
 553 {
 554         return 0;
 555 }
 556
 557 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 558
 559 static struct kvm_memslots *kvm_alloc_memslots(void)
 560 {
 561         int i;
 562         struct kvm_memslots *slots;
 563
 564         slots = kvzalloc(sizeof(struct kvm_memslots), GFP_KERNEL_ACCOUNT);
 565         if (!slots)
 566                 return NULL;
 567
 568         for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 569                 slots->id_to_index[i] = slots->memslots[i].id = i;
 570
 571         return slots;
 572 }
 573
 574 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
 575 {
 576         if (!memslot->dirty_bitmap)
 577                 return;
 578
 579         kvfree(memslot->dirty_bitmap);
 580         memslot->dirty_bitmap = NULL;
 581 }
 582
 583 /*
 584  * Free any memory in @free but not in @dont.
 585  */
 586 static void kvm_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
 587                               struct kvm_memory_slot *dont)
 588 {
 589         if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
 590                 kvm_destroy_dirty_bitmap(free);
 591
 592         kvm_arch_free_memslot(kvm, free, dont);
 593
 594         free->npages = 0;
 595 }
 596
 597 static void kvm_free_memslots(struct kvm *kvm, struct kvm_memslots *slots)
 598 {
 599         struct kvm_memory_slot *memslot;
 600
 601         if (!slots)
 602                 return;
 603
 604         kvm_for_each_memslot(memslot, slots)
 605                 kvm_free_memslot(kvm, memslot, NULL);
 606
 607         kvfree(slots);
 608 }
 609
 610 static void kvm_destroy_vm_debugfs(struct kvm *kvm)
 611 {
 612         int i;
 613
 614         if (!kvm->debugfs_dentry)
 615                 return;
 616
 617         debugfs_remove_recursive(kvm->debugfs_dentry);
 618
 619         if (kvm->debugfs_stat_data) {
 620                 for (i = 0; i < kvm_debugfs_num_entries; i++)
 621                         kfree(kvm->debugfs_stat_data[i]);
 622                 kfree(kvm->debugfs_stat_data);
 623         }
 624 }
 625
 626 static int kvm_create_vm_debugfs(struct kvm *kvm, int fd)
 627 {
 628         char dir_name[ITOA_MAX_LEN * 2];
 629         struct kvm_stat_data *stat_data;
 630         struct kvm_stats_debugfs_item *p;
 631
 632         if (!debugfs_initialized())
 633                 return 0;
 634
 635         snprintf(dir_name, sizeof(dir_name), "%d-%d", task_pid_nr(current), fd);
 636         kvm->debugfs_dentry = debugfs_create_dir(dir_name, kvm_debugfs_dir);
 637
 638         kvm->debugfs_stat_data = kcalloc(kvm_debugfs_num_entries,
 639                                          sizeof(*kvm->debugfs_stat_data),
 640                                          GFP_KERNEL_ACCOUNT);
 641         if (!kvm->debugfs_stat_data)
 642                 return -ENOMEM;
 643
 644         for (p = debugfs_entries; p->name; p++) {
 645                 stat_data = kzalloc(sizeof(*stat_data), GFP_KERNEL_ACCOUNT);
 646                 if (!stat_data)
 647                         return -ENOMEM;
 648
 649                 stat_data->kvm = kvm;
 650                 stat_data->dbgfs_item = p;
 651                 kvm->debugfs_stat_data[p - debugfs_entries] = stat_data;
 652                 debugfs_create_file(p->name, KVM_DBGFS_GET_MODE(p),
 653                                     kvm->debugfs_dentry, stat_data,
 654                                     &stat_fops_per_vm);
 655         }
 656         return 0;
 657 }
 658
 659 /*
 660  * Called after the VM is otherwise initialized, but just before adding it to
 661  * the vm_list.
 662  */
 663 int __weak kvm_arch_post_init_vm(struct kvm *kvm)
 664 {
 665         return 0;
 666 }
 667
 668 /*
 669  * Called just after removing the VM from the vm_list, but before doing any
 670  * other destruction.
 671  */
 672 void __weak kvm_arch_pre_destroy_vm(struct kvm *kvm)
 673 {
 674 }
 675
 676 static struct kvm *kvm_create_vm(unsigned long type)
 677 {
 678         struct kvm *kvm = kvm_arch_alloc_vm();
 679         int r = -ENOMEM;
 680         int i;
 681
 682         if (!kvm)
 683                 return ERR_PTR(-ENOMEM);
 684
 685         spin_lock_init(&kvm->mmu_lock);
 686         mmgrab(current->mm);
 687         kvm->mm = current->mm;
 688         kvm_eventfd_init(kvm);
 689         mutex_init(&kvm->lock);
 690         mutex_init(&kvm->irq_lock);
 691         mutex_init(&kvm->slots_lock);
 692         INIT_LIST_HEAD(&kvm->devices);
 693
 694         BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX);
 695
 696         if (init_srcu_struct(&kvm->srcu))
 697                 goto out_err_no_srcu;
 698         if (init_srcu_struct(&kvm->irq_srcu))
 699                 goto out_err_no_irq_srcu;
 700
 701         refcount_set(&kvm->users_count, 1);
 702         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) {
 703                 struct kvm_memslots *slots = kvm_alloc_memslots();
 704
 705                 if (!slots)
 706                         goto out_err_no_arch_destroy_vm;
 707                 /* Generations must be different for each address space. */
 708                 slots->generation = i;
 709                 rcu_assign_pointer(kvm->memslots[i], slots);
 710         }
 711
 712         for (i = 0; i < KVM_NR_BUSES; i++) {
 713                 rcu_assign_pointer(kvm->buses[i],
 714                         kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL_ACCOUNT));
 715                 if (!kvm->buses[i])
 716                         goto out_err_no_arch_destroy_vm;
 717         }
 718
 719         r = kvm_arch_init_vm(kvm, type);
 720         if (r)
 721                 goto out_err_no_arch_destroy_vm;
 722
 723         r = hardware_enable_all();
 724         if (r)
 725                 goto out_err_no_disable;
 726
 727 #ifdef CONFIG_HAVE_KVM_IRQFD
 728         INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 729 #endif
 730
 731         r = kvm_init_mmu_notifier(kvm);
 732         if (r)
 733                 goto out_err_no_mmu_notifier;
 734
 735         r = kvm_arch_post_init_vm(kvm);
 736         if (r)
 737                 goto out_err;
 738
 739         mutex_lock(&kvm_lock);
 740         list_add(&kvm->vm_list, &vm_list);
 741         mutex_unlock(&kvm_lock);
 742
 743         preempt_notifier_inc();
 744
 745         return kvm;
 746
 747 out_err:
 748 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 749         if (kvm->mmu_notifier.ops)
 750                 mmu_notifier_unregister(&kvm->mmu_notifier, current->mm);
 751 #endif
 752 out_err_no_mmu_notifier:
 753         hardware_disable_all();
 754 out_err_no_disable:
 755         kvm_arch_destroy_vm(kvm);
 756 out_err_no_arch_destroy_vm:
 757         WARN_ON_ONCE(!refcount_dec_and_test(&kvm->users_count));
 758         for (i = 0; i < KVM_NR_BUSES; i++)
 759                 kfree(kvm_get_bus(kvm, i));
 760         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 761                 kvm_free_memslots(kvm, __kvm_memslots(kvm, i));
 762         cleanup_srcu_struct(&kvm->irq_srcu);
 763 out_err_no_irq_srcu:
 764         cleanup_srcu_struct(&kvm->srcu);
 765 out_err_no_srcu:
 766         kvm_arch_free_vm(kvm);
 767         mmdrop(current->mm);
 768         return ERR_PTR(r);
 769 }
 770
 771 static void kvm_destroy_devices(struct kvm *kvm)
 772 {
 773         struct kvm_device *dev, *tmp;
 774
 775         /*
 776          * We do not need to take the kvm->lock here, because nobody else
 777          * has a reference to the struct kvm at this point and therefore
 778          * cannot access the devices list anyhow.
 779          */
 780         list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) {
 781                 list_del(&dev->vm_node);
 782                 dev->ops->destroy(dev);
 783         }
 784 }
 785
 786 static void kvm_destroy_vm(struct kvm *kvm)
 787 {
 788         int i;
 789         struct mm_struct *mm = kvm->mm;
 790
 791         kvm_uevent_notify_change(KVM_EVENT_DESTROY_VM, kvm);
 792         kvm_destroy_vm_debugfs(kvm);
 793         kvm_arch_sync_events(kvm);
 794         mutex_lock(&kvm_lock);
 795         list_del(&kvm->vm_list);
 796         mutex_unlock(&kvm_lock);
 797         kvm_arch_pre_destroy_vm(kvm);
 798
 799         kvm_free_irq_routing(kvm);
 800         for (i = 0; i < KVM_NR_BUSES; i++) {
 801                 struct kvm_io_bus *bus = kvm_get_bus(kvm, i);
 802
 803                 if (bus)
 804                         kvm_io_bus_destroy(bus);
 805                 kvm->buses[i] = NULL;
 806         }
 807         kvm_coalesced_mmio_free(kvm);
 808 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 809         mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 810 #else
 811         kvm_arch_flush_shadow_all(kvm);
 812 #endif
 813         kvm_arch_destroy_vm(kvm);
 814         kvm_destroy_devices(kvm);
 815         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
 816                 kvm_free_memslots(kvm, __kvm_memslots(kvm, i));
 817         cleanup_srcu_struct(&kvm->irq_srcu);
 818         cleanup_srcu_struct(&kvm->srcu);
 819         kvm_arch_free_vm(kvm);
 820         preempt_notifier_dec();
 821         hardware_disable_all();
 822         mmdrop(mm);
 823 }
 824
 825 void kvm_get_kvm(struct kvm *kvm)
 826 {
 827         refcount_inc(&kvm->users_count);
 828 }
 829 EXPORT_SYMBOL_GPL(kvm_get_kvm);
 830
 831 void kvm_put_kvm(struct kvm *kvm)
 832 {
 833         if (refcount_dec_and_test(&kvm->users_count))
 834                 kvm_destroy_vm(kvm);
 835 }
 836 EXPORT_SYMBOL_GPL(kvm_put_kvm);
 837
 838 /*
 839  * Used to put a reference that was taken on behalf of an object associated
 840  * with a user-visible file descriptor, e.g. a vcpu or device, if installation
 841  * of the new file descriptor fails and the reference cannot be transferred to
 842  * its final owner.  In such cases, the caller is still actively using @kvm and
 843  * will fail miserably if the refcount unexpectedly hits zero.
 844  */
 845 void kvm_put_kvm_no_destroy(struct kvm *kvm)
 846 {
 847         WARN_ON(refcount_dec_and_test(&kvm->users_count));
 848 }
 849 EXPORT_SYMBOL_GPL(kvm_put_kvm_no_destroy);
 850
 851 static int kvm_vm_release(struct inode *inode, struct file *filp)
 852 {
 853         struct kvm *kvm = filp->private_data;
 854
 855         kvm_irqfd_release(kvm);
 856
 857         kvm_put_kvm(kvm);
 858         return 0;
 859 }
 860
 861 /*
 862  * Allocation size is twice as large as the actual dirty bitmap size.
 863  * See x86's kvm_vm_ioctl_get_dirty_log() why this is needed.
 864  */
 865 static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
 866 {
 867         unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
 868
 869         memslot->dirty_bitmap = kvzalloc(dirty_bytes, GFP_KERNEL_ACCOUNT);
 870         if (!memslot->dirty_bitmap)
 871                 return -ENOMEM;
 872
 873         return 0;
 874 }
 875
 876 /*
 877  * Insert memslot and re-sort memslots based on their GFN,
 878  * so binary search could be used to lookup GFN.
 879  * Sorting algorithm takes advantage of having initially
 880  * sorted array and known changed memslot position.
 881  */
 882 static void update_memslots(struct kvm_memslots *slots,
 883                             struct kvm_memory_slot *new,
 884                             enum kvm_mr_change change)
 885 {
 886         int id = new->id;
 887         int i = slots->id_to_index[id];
 888         struct kvm_memory_slot *mslots = slots->memslots;
 889
 890         WARN_ON(mslots[i].id != id);
 891         switch (change) {
 892         case KVM_MR_CREATE:
 893                 slots->used_slots++;
 894                 WARN_ON(mslots[i].npages || !new->npages);
 895                 break;
 896         case KVM_MR_DELETE:
 897                 slots->used_slots--;
 898                 WARN_ON(new->npages || !mslots[i].npages);
 899                 break;
 900         default:
 901                 break;
 902         }
 903
 904         while (i < KVM_MEM_SLOTS_NUM - 1 &&
 905                new->base_gfn <= mslots[i + 1].base_gfn) {
 906                 if (!mslots[i + 1].npages)
 907                         break;
 908                 mslots[i] = mslots[i + 1];
 909                 slots->id_to_index[mslots[i].id] = i;
 910                 i++;
 911         }
 912
 913         /*
 914          * The ">=" is needed when creating a slot with base_gfn == 0,
 915          * so that it moves before all those with base_gfn == npages == 0.
 916          *
 917          * On the other hand, if new->npages is zero, the above loop has
 918          * already left i pointing to the beginning of the empty part of
 919          * mslots, and the ">=" would move the hole backwards in this
 920          * case---which is wrong.  So skip the loop when deleting a slot.
 921          */
 922         if (new->npages) {
 923                 while (i > 0 &&
 924                        new->base_gfn >= mslots[i - 1].base_gfn) {
 925                         mslots[i] = mslots[i - 1];
 926                         slots->id_to_index[mslots[i].id] = i;
 927                         i--;
 928                 }
 929         } else
 930                 WARN_ON_ONCE(i != slots->used_slots);
 931
 932         mslots[i] = *new;
 933         slots->id_to_index[mslots[i].id] = i;
 934 }
 935
 936 static int check_memory_region_flags(const struct kvm_userspace_memory_region *mem)
 937 {
 938         u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES;
 939
 940 #ifdef __KVM_HAVE_READONLY_MEM
 941         valid_flags |= KVM_MEM_READONLY;
 942 #endif
 943
 944         if (mem->flags & ~valid_flags)
 945                 return -EINVAL;
 946
 947         return 0;
 948 }
 949
 950 static struct kvm_memslots *install_new_memslots(struct kvm *kvm,
 951                 int as_id, struct kvm_memslots *slots)
 952 {
 953         struct kvm_memslots *old_memslots = __kvm_memslots(kvm, as_id);
 954         u64 gen = old_memslots->generation;
 955
 956         WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS);
 957         slots->generation = gen | KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS;
 958
 959         rcu_assign_pointer(kvm->memslots[as_id], slots);
 960         synchronize_srcu_expedited(&kvm->srcu);
 961
 962         /*
 963          * Increment the new memslot generation a second time, dropping the
 964          * update in-progress flag and incrementing the generation based on
 965          * the number of address spaces.  This provides a unique and easily
 966          * identifiable generation number while the memslots are in flux.
 967          */
 968         gen = slots->generation & ~KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS;
 969
 970         /*
 971          * Generations must be unique even across address spaces.  We do not need
 972          * a global counter for that, instead the generation space is evenly split
 973          * across address spaces.  For example, with two address spaces, address
 974          * space 0 will use generations 0, 2, 4, ... while address space 1 will
 975          * use generations 1, 3, 5, ...
 976          */
 977         gen += KVM_ADDRESS_SPACE_NUM;
 978
 979         kvm_arch_memslots_updated(kvm, gen);
 980
 981         slots->generation = gen;
 982
 983         return old_memslots;
 984 }
 985
 986 /*
 987  * Allocate some memory and give it an address in the guest physical address
 988  * space.
 989  *
 990  * Discontiguous memory is allowed, mostly for framebuffers.
 991  *
 992  * Must be called holding kvm->slots_lock for write.
 993  */
 994 int __kvm_set_memory_region(struct kvm *kvm,
 995                             const struct kvm_userspace_memory_region *mem)
 996 {
 997         int r;
 998         gfn_t base_gfn;
 999         unsigned long npages;
1000         struct kvm_memory_slot *slot;
1001         struct kvm_memory_slot old, new;
1002         struct kvm_memslots *slots = NULL, *old_memslots;
1003         int as_id, id;
1004         enum kvm_mr_change change;
1005
1006         r = check_memory_region_flags(mem);
1007         if (r)
1008                 goto out;
1009
1010         r = -EINVAL;
1011         as_id = mem->slot >> 16;
1012         id = (u16)mem->slot;
1013
1014         /* General sanity checks */
1015         if (mem->memory_size & (PAGE_SIZE - 1))
1016                 goto out;
1017         if (mem->guest_phys_addr & (PAGE_SIZE - 1))
1018                 goto out;
1019         /* We can read the guest memory with __xxx_user() later on. */
1020         if ((id < KVM_USER_MEM_SLOTS) &&
1021             ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
1022              !access_ok((void __user *)(unsigned long)mem->userspace_addr,
1023                         mem->memory_size)))
1024                 goto out;
1025         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_MEM_SLOTS_NUM)
1026                 goto out;
1027         if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
1028                 goto out;
1029
1030         slot = id_to_memslot(__kvm_memslots(kvm, as_id), id);
1031         base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
1032         npages = mem->memory_size >> PAGE_SHIFT;
1033
1034         if (npages > KVM_MEM_MAX_NR_PAGES)
1035                 goto out;
1036
1037         new = old = *slot;
1038
1039         new.id = id;
1040         new.base_gfn = base_gfn;
1041         new.npages = npages;
1042         new.flags = mem->flags;
1043
1044         if (npages) {
1045                 if (!old.npages)
1046                         change = KVM_MR_CREATE;
1047                 else { /* Modify an existing slot. */
1048                         if ((mem->userspace_addr != old.userspace_addr) ||
1049                             (npages != old.npages) ||
1050                             ((new.flags ^ old.flags) & KVM_MEM_READONLY))
1051                                 goto out;
1052
1053                         if (base_gfn != old.base_gfn)
1054                                 change = KVM_MR_MOVE;
1055                         else if (new.flags != old.flags)
1056                                 change = KVM_MR_FLAGS_ONLY;
1057                         else { /* Nothing to change. */
1058                                 r = 0;
1059                                 goto out;
1060                         }
1061                 }
1062         } else {
1063                 if (!old.npages)
1064                         goto out;
1065
1066                 change = KVM_MR_DELETE;
1067                 new.base_gfn = 0;
1068                 new.flags = 0;
1069         }
1070
1071         if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) {
1072                 /* Check for overlaps */
1073                 r = -EEXIST;
1074                 kvm_for_each_memslot(slot, __kvm_memslots(kvm, as_id)) {
1075                         if (slot->id == id)
1076                                 continue;
1077                         if (!((base_gfn + npages <= slot->base_gfn) ||
1078                               (base_gfn >= slot->base_gfn + slot->npages)))
1079                                 goto out;
1080                 }
1081         }
1082
1083         /* Free page dirty bitmap if unneeded */
1084         if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
1085                 new.dirty_bitmap = NULL;
1086
1087         r = -ENOMEM;
1088         if (change == KVM_MR_CREATE) {
1089                 new.userspace_addr = mem->userspace_addr;
1090
1091                 if (kvm_arch_create_memslot(kvm, &new, npages))
1092                         goto out_free;
1093         }
1094
1095         /* Allocate page dirty bitmap if needed */
1096         if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
1097                 if (kvm_create_dirty_bitmap(&new) < 0)
1098                         goto out_free;
1099         }
1100
1101         slots = kvzalloc(sizeof(struct kvm_memslots), GFP_KERNEL_ACCOUNT);
1102         if (!slots)
1103                 goto out_free;
1104         memcpy(slots, __kvm_memslots(kvm, as_id), sizeof(struct kvm_memslots));
1105
1106         if ((change == KVM_MR_DELETE) || (change == KVM_MR_MOVE)) {
1107                 slot = id_to_memslot(slots, id);
1108                 slot->flags |= KVM_MEMSLOT_INVALID;
1109
1110                 old_memslots = install_new_memslots(kvm, as_id, slots);
1111
1112                 /* From this point no new shadow pages pointing to a deleted,
1113                  * or moved, memslot will be created.
1114                  *
1115                  * validation of sp->gfn happens in:
1116                  *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
1117                  *      - kvm_is_visible_gfn (mmu_check_root)
1118                  */
1119                 kvm_arch_flush_shadow_memslot(kvm, slot);
1120
1121                 /*
1122                  * We can re-use the old_memslots from above, the only difference
1123                  * from the currently installed memslots is the invalid flag.  This
1124                  * will get overwritten by update_memslots anyway.
1125                  */
1126                 slots = old_memslots;
1127         }
1128
1129         r = kvm_arch_prepare_memory_region(kvm, &new, mem, change);
1130         if (r)
1131                 goto out_slots;
1132
1133         /* actual memory is freed via old in kvm_free_memslot below */
1134         if (change == KVM_MR_DELETE) {
1135                 new.dirty_bitmap = NULL;
1136                 memset(&new.arch, 0, sizeof(new.arch));
1137         }
1138
1139         update_memslots(slots, &new, change);
1140         old_memslots = install_new_memslots(kvm, as_id, slots);
1141
1142         kvm_arch_commit_memory_region(kvm, mem, &old, &new, change);
1143
1144         kvm_free_memslot(kvm, &old, &new);
1145         kvfree(old_memslots);
1146         return 0;
1147
1148 out_slots:
1149         kvfree(slots);
1150 out_free:
1151         kvm_free_memslot(kvm, &new, &old);
1152 out:
1153         return r;
1154 }
1155 EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
1156
1157 int kvm_set_memory_region(struct kvm *kvm,
1158                           const struct kvm_userspace_memory_region *mem)
1159 {
1160         int r;
1161
1162         mutex_lock(&kvm->slots_lock);
1163         r = __kvm_set_memory_region(kvm, mem);
1164         mutex_unlock(&kvm->slots_lock);
1165         return r;
1166 }
1167 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
1168
1169 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
1170                                           struct kvm_userspace_memory_region *mem)
1171 {
1172         if ((u16)mem->slot >= KVM_USER_MEM_SLOTS)
1173                 return -EINVAL;
1174
1175         return kvm_set_memory_region(kvm, mem);
1176 }
1177
1178 int kvm_get_dirty_log(struct kvm *kvm,
1179                         struct kvm_dirty_log *log, int *is_dirty)
1180 {
1181         struct kvm_memslots *slots;
1182         struct kvm_memory_slot *memslot;
1183         int i, as_id, id;
1184         unsigned long n;
1185         unsigned long any = 0;
1186
1187         as_id = log->slot >> 16;
1188         id = (u16)log->slot;
1189         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1190                 return -EINVAL;
1191
1192         slots = __kvm_memslots(kvm, as_id);
1193         memslot = id_to_memslot(slots, id);
1194         if (!memslot->dirty_bitmap)
1195                 return -ENOENT;
1196
1197         n = kvm_dirty_bitmap_bytes(memslot);
1198
1199         for (i = 0; !any && i < n/sizeof(long); ++i)
1200                 any = memslot->dirty_bitmap[i];
1201
1202         if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
1203                 return -EFAULT;
1204
1205         if (any)
1206                 *is_dirty = 1;
1207         return 0;
1208 }
1209 EXPORT_SYMBOL_GPL(kvm_get_dirty_log);
1210
1211 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
1212 /**
1213  * kvm_get_dirty_log_protect - get a snapshot of dirty pages
1214  *      and reenable dirty page tracking for the corresponding pages.
1215  * @kvm:        pointer to kvm instance
1216  * @log:        slot id and address to which we copy the log
1217  * @flush:      true if TLB flush is needed by caller
1218  *
1219  * We need to keep it in mind that VCPU threads can write to the bitmap
1220  * concurrently. So, to avoid losing track of dirty pages we keep the
1221  * following order:
1222  *
1223  *    1. Take a snapshot of the bit and clear it if needed.
1224  *    2. Write protect the corresponding page.
1225  *    3. Copy the snapshot to the userspace.
1226  *    4. Upon return caller flushes TLB's if needed.
1227  *
1228  * Between 2 and 4, the guest may write to the page using the remaining TLB
1229  * entry.  This is not a problem because the page is reported dirty using
1230  * the snapshot taken before and step 4 ensures that writes done after
1231  * exiting to userspace will be logged for the next call.
1232  *
1233  */
1234 int kvm_get_dirty_log_protect(struct kvm *kvm,
1235                         struct kvm_dirty_log *log, bool *flush)
1236 {
1237         struct kvm_memslots *slots;
1238         struct kvm_memory_slot *memslot;
1239         int i, as_id, id;
1240         unsigned long n;
1241         unsigned long *dirty_bitmap;
1242         unsigned long *dirty_bitmap_buffer;
1243
1244         as_id = log->slot >> 16;
1245         id = (u16)log->slot;
1246         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1247                 return -EINVAL;
1248
1249         slots = __kvm_memslots(kvm, as_id);
1250         memslot = id_to_memslot(slots, id);
1251
1252         dirty_bitmap = memslot->dirty_bitmap;
1253         if (!dirty_bitmap)
1254                 return -ENOENT;
1255
1256         n = kvm_dirty_bitmap_bytes(memslot);
1257         *flush = false;
1258         if (kvm->manual_dirty_log_protect) {
1259                 /*
1260                  * Unlike kvm_get_dirty_log, we always return false in *flush,
1261                  * because no flush is needed until KVM_CLEAR_DIRTY_LOG.  There
1262                  * is some code duplication between this function and
1263                  * kvm_get_dirty_log, but hopefully all architecture
1264                  * transition to kvm_get_dirty_log_protect and kvm_get_dirty_log
1265                  * can be eliminated.
1266                  */
1267                 dirty_bitmap_buffer = dirty_bitmap;
1268         } else {
1269                 dirty_bitmap_buffer = kvm_second_dirty_bitmap(memslot);
1270                 memset(dirty_bitmap_buffer, 0, n);
1271
1272                 spin_lock(&kvm->mmu_lock);
1273                 for (i = 0; i < n / sizeof(long); i++) {
1274                         unsigned long mask;
1275                         gfn_t offset;
1276
1277                         if (!dirty_bitmap[i])
1278                                 continue;
1279
1280                         *flush = true;
1281                         mask = xchg(&dirty_bitmap[i], 0);
1282                         dirty_bitmap_buffer[i] = mask;
1283
1284                         offset = i * BITS_PER_LONG;
1285                         kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot,
1286                                                                 offset, mask);
1287                 }
1288                 spin_unlock(&kvm->mmu_lock);
1289         }
1290
1291         if (copy_to_user(log->dirty_bitmap, dirty_bitmap_buffer, n))
1292                 return -EFAULT;
1293         return 0;
1294 }
1295 EXPORT_SYMBOL_GPL(kvm_get_dirty_log_protect);
1296
1297 /**
1298  * kvm_clear_dirty_log_protect - clear dirty bits in the bitmap
1299  *      and reenable dirty page tracking for the corresponding pages.
1300  * @kvm:        pointer to kvm instance
1301  * @log:        slot id and address from which to fetch the bitmap of dirty pages
1302  * @flush:      true if TLB flush is needed by caller
1303  */
1304 int kvm_clear_dirty_log_protect(struct kvm *kvm,
1305                                 struct kvm_clear_dirty_log *log, bool *flush)
1306 {
1307         struct kvm_memslots *slots;
1308         struct kvm_memory_slot *memslot;
1309         int as_id, id;
1310         gfn_t offset;
1311         unsigned long i, n;
1312         unsigned long *dirty_bitmap;
1313         unsigned long *dirty_bitmap_buffer;
1314
1315         as_id = log->slot >> 16;
1316         id = (u16)log->slot;
1317         if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS)
1318                 return -EINVAL;
1319
1320         if (log->first_page & 63)
1321                 return -EINVAL;
1322
1323         slots = __kvm_memslots(kvm, as_id);
1324         memslot = id_to_memslot(slots, id);
1325
1326         dirty_bitmap = memslot->dirty_bitmap;
1327         if (!dirty_bitmap)
1328                 return -ENOENT;
1329
1330         n = ALIGN(log->num_pages, BITS_PER_LONG) / 8;
1331
1332         if (log->first_page > memslot->npages ||
1333             log->num_pages > memslot->npages - log->first_page ||
1334             (log->num_pages < memslot->npages - log->first_page && (log->num_pages & 63)))
1335             return -EINVAL;
1336
1337         *flush = false;
1338         dirty_bitmap_buffer = kvm_second_dirty_bitmap(memslot);
1339         if (copy_from_user(dirty_bitmap_buffer, log->dirty_bitmap, n))
1340                 return -EFAULT;
1341
1342         spin_lock(&kvm->mmu_lock);
1343         for (offset = log->first_page, i = offset / BITS_PER_LONG,
1344                  n = DIV_ROUND_UP(log->num_pages, BITS_PER_LONG); n--;
1345              i++, offset += BITS_PER_LONG) {
1346                 unsigned long mask = *dirty_bitmap_buffer++;
1347                 atomic_long_t *p = (atomic_long_t *) &dirty_bitmap[i];
1348                 if (!mask)
1349                         continue;
1350
1351                 mask &= atomic_long_fetch_andnot(mask, p);
1352
1353                 /*
1354                  * mask contains the bits that really have been cleared.  This
1355                  * never includes any bits beyond the length of the memslot (if
1356                  * the length is not aligned to 64 pages), therefore it is not
1357                  * a problem if userspace sets them in log->dirty_bitmap.
1358                 */
1359                 if (mask) {
1360                         *flush = true;
1361                         kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot,
1362                                                                 offset, mask);
1363                 }
1364         }
1365         spin_unlock(&kvm->mmu_lock);
1366
1367         return 0;
1368 }
1369 EXPORT_SYMBOL_GPL(kvm_clear_dirty_log_protect);
1370 #endif
1371
1372 bool kvm_largepages_enabled(void)
1373 {
1374         return largepages_enabled;
1375 }
1376
1377 void kvm_disable_largepages(void)
1378 {
1379         largepages_enabled = false;
1380 }
1381 EXPORT_SYMBOL_GPL(kvm_disable_largepages);
1382
1383 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
1384 {
1385         return __gfn_to_memslot(kvm_memslots(kvm), gfn);
1386 }
1387 EXPORT_SYMBOL_GPL(gfn_to_memslot);
1388
1389 struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn)
1390 {
1391         return __gfn_to_memslot(kvm_vcpu_memslots(vcpu), gfn);
1392 }
1393
1394 bool kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
1395 {
1396         struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
1397
1398         if (!memslot || memslot->id >= KVM_USER_MEM_SLOTS ||
1399               memslot->flags & KVM_MEMSLOT_INVALID)
1400                 return false;
1401
1402         return true;
1403 }
1404 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
1405
1406 unsigned long kvm_host_page_size(struct kvm_vcpu *vcpu, gfn_t gfn)
1407 {
1408         struct vm_area_struct *vma;
1409         unsigned long addr, size;
1410
1411         size = PAGE_SIZE;
1412
1413         addr = kvm_vcpu_gfn_to_hva_prot(vcpu, gfn, NULL);
1414         if (kvm_is_error_hva(addr))
1415                 return PAGE_SIZE;
1416
1417         down_read(&current->mm->mmap_sem);
1418         vma = find_vma(current->mm, addr);
1419         if (!vma)
1420                 goto out;
1421
1422         size = vma_kernel_pagesize(vma);
1423
1424 out:
1425         up_read(&current->mm->mmap_sem);
1426
1427         return size;
1428 }
1429
1430 static bool memslot_is_readonly(struct kvm_memory_slot *slot)
1431 {
1432         return slot->flags & KVM_MEM_READONLY;
1433 }
1434
1435 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1436                                        gfn_t *nr_pages, bool write)
1437 {
1438         if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
1439                 return KVM_HVA_ERR_BAD;
1440
1441         if (memslot_is_readonly(slot) && write)
1442                 return KVM_HVA_ERR_RO_BAD;
1443
1444         if (nr_pages)
1445                 *nr_pages = slot->npages - (gfn - slot->base_gfn);
1446
1447         return __gfn_to_hva_memslot(slot, gfn);
1448 }
1449
1450 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
1451                                      gfn_t *nr_pages)
1452 {
1453         return __gfn_to_hva_many(slot, gfn, nr_pages, true);
1454 }
1455
1456 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot,
1457                                         gfn_t gfn)
1458 {
1459         return gfn_to_hva_many(slot, gfn, NULL);
1460 }
1461 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot);
1462
1463 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
1464 {
1465         return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
1466 }
1467 EXPORT_SYMBOL_GPL(gfn_to_hva);
1468
1469 unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn)
1470 {
1471         return gfn_to_hva_many(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, NULL);
1472 }
1473 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_hva);
1474
1475 /*
1476  * Return the hva of a @gfn and the R/W attribute if possible.
1477  *
1478  * @slot: the kvm_memory_slot which contains @gfn
1479  * @gfn: the gfn to be translated
1480  * @writable: used to return the read/write attribute of the @slot if the hva
1481  * is valid and @writable is not NULL
1482  */
1483 unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot,
1484                                       gfn_t gfn, bool *writable)
1485 {
1486         unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false);
1487
1488         if (!kvm_is_error_hva(hva) && writable)
1489                 *writable = !memslot_is_readonly(slot);
1490
1491         return hva;
1492 }
1493
1494 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable)
1495 {
1496         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
1497
1498         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1499 }
1500
1501 unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable)
1502 {
1503         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
1504
1505         return gfn_to_hva_memslot_prot(slot, gfn, writable);
1506 }
1507
1508 static inline int check_user_page_hwpoison(unsigned long addr)
1509 {
1510         int rc, flags = FOLL_HWPOISON | FOLL_WRITE;
1511
1512         rc = get_user_pages(addr, 1, flags, NULL, NULL);
1513         return rc == -EHWPOISON;
1514 }
1515
1516 /*
1517  * The fast path to get the writable pfn which will be stored in @pfn,
1518  * true indicates success, otherwise false is returned.  It's also the
1519  * only part that runs if we can in atomic context.
1520  */
1521 static bool hva_to_pfn_fast(unsigned long addr, bool write_fault,
1522                             bool *writable, kvm_pfn_t *pfn)
1523 {
1524         struct page *page[1];
1525         int npages;
1526
1527         /*
1528          * Fast pin a writable pfn only if it is a write fault request
1529          * or the caller allows to map a writable pfn for a read fault
1530          * request.
1531          */
1532         if (!(write_fault || writable))
1533                 return false;
1534
1535         npages = __get_user_pages_fast(addr, 1, 1, page);
1536         if (npages == 1) {
1537                 *pfn = page_to_pfn(page[0]);
1538
1539                 if (writable)
1540                         *writable = true;
1541                 return true;
1542         }
1543
1544         return false;
1545 }
1546
1547 /*
1548  * The slow path to get the pfn of the specified host virtual address,
1549  * 1 indicates success, -errno is returned if error is detected.
1550  */
1551 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault,
1552                            bool *writable, kvm_pfn_t *pfn)
1553 {
1554         unsigned int flags = FOLL_HWPOISON;
1555         struct page *page;
1556         int npages = 0;
1557
1558         might_sleep();
1559
1560         if (writable)
1561                 *writable = write_fault;
1562
1563         if (write_fault)
1564                 flags |= FOLL_WRITE;
1565         if (async)
1566                 flags |= FOLL_NOWAIT;
1567
1568         npages = get_user_pages_unlocked(addr, 1, &page, flags);
1569         if (npages != 1)
1570                 return npages;
1571
1572         /* map read fault as writable if possible */
1573         if (unlikely(!write_fault) && writable) {
1574                 struct page *wpage;
1575
1576                 if (__get_user_pages_fast(addr, 1, 1, &wpage) == 1) {
1577                         *writable = true;
1578                         put_page(page);
1579                         page = wpage;
1580                 }
1581         }
1582         *pfn = page_to_pfn(page);
1583         return npages;
1584 }
1585
1586 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault)
1587 {
1588         if (unlikely(!(vma->vm_flags & VM_READ)))
1589                 return false;
1590
1591         if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE))))
1592                 return false;
1593
1594         return true;
1595 }
1596
1597 static int hva_to_pfn_remapped(struct vm_area_struct *vma,
1598                                unsigned long addr, bool *async,
1599                                bool write_fault, bool *writable,
1600                                kvm_pfn_t *p_pfn)
1601 {
1602         unsigned long pfn;
1603         int r;
1604
1605         r = follow_pfn(vma, addr, &pfn);
1606         if (r) {
1607                 /*
1608                  * get_user_pages fails for VM_IO and VM_PFNMAP vmas and does
1609                  * not call the fault handler, so do it here.
1610                  */
1611                 bool unlocked = false;
1612                 r = fixup_user_fault(current, current->mm, addr,
1613                                      (write_fault ? FAULT_FLAG_WRITE : 0),
1614                                      &unlocked);
1615                 if (unlocked)
1616                         return -EAGAIN;
1617                 if (r)
1618                         return r;
1619
1620                 r = follow_pfn(vma, addr, &pfn);
1621                 if (r)
1622                         return r;
1623
1624         }
1625
1626         if (writable)
1627                 *writable = true;
1628
1629         /*
1630          * Get a reference here because callers of *hva_to_pfn* and
1631          * *gfn_to_pfn* ultimately call kvm_release_pfn_clean on the
1632          * returned pfn.  This is only needed if the VMA has VM_MIXEDMAP
1633          * set, but the kvm_get_pfn/kvm_release_pfn_clean pair will
1634          * simply do nothing for reserved pfns.
1635          *
1636          * Whoever called remap_pfn_range is also going to call e.g.
1637          * unmap_mapping_range before the underlying pages are freed,
1638          * causing a call to our MMU notifier.
1639          */
1640         kvm_get_pfn(pfn);
1641
1642         *p_pfn = pfn;
1643         return 0;
1644 }
1645
1646 /*
1647  * Pin guest page in memory and return its pfn.
1648  * @addr: host virtual address which maps memory to the guest
1649  * @atomic: whether this function can sleep
1650  * @async: whether this function need to wait IO complete if the
1651  *         host page is not in the memory
1652  * @write_fault: whether we should get a writable host page
1653  * @writable: whether it allows to map a writable host page for !@write_fault
1654  *
1655  * The function will map a writable host page for these two cases:
1656  * 1): @write_fault = true
1657  * 2): @write_fault = false && @writable, @writable will tell the caller
1658  *     whether the mapping is writable.
1659  */
1660 static kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async,
1661                         bool write_fault, bool *writable)
1662 {
1663         struct vm_area_struct *vma;
1664         kvm_pfn_t pfn = 0;
1665         int npages, r;
1666
1667         /* we can do it either atomically or asynchronously, not both */
1668         BUG_ON(atomic && async);
1669
1670         if (hva_to_pfn_fast(addr, write_fault, writable, &pfn))
1671                 return pfn;
1672
1673         if (atomic)
1674                 return KVM_PFN_ERR_FAULT;
1675
1676         npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn);
1677         if (npages == 1)
1678                 return pfn;
1679
1680         down_read(&current->mm->mmap_sem);
1681         if (npages == -EHWPOISON ||
1682               (!async && check_user_page_hwpoison(addr))) {
1683                 pfn = KVM_PFN_ERR_HWPOISON;
1684                 goto exit;
1685         }
1686
1687 retry:
1688         vma = find_vma_intersection(current->mm, addr, addr + 1);
1689
1690         if (vma == NULL)
1691                 pfn = KVM_PFN_ERR_FAULT;
1692         else if (vma->vm_flags & (VM_IO | VM_PFNMAP)) {
1693                 r = hva_to_pfn_remapped(vma, addr, async, write_fault, writable, &pfn);
1694                 if (r == -EAGAIN)
1695                         goto retry;
1696                 if (r < 0)
1697                         pfn = KVM_PFN_ERR_FAULT;
1698         } else {
1699                 if (async && vma_is_valid(vma, write_fault))
1700                         *async = true;
1701                 pfn = KVM_PFN_ERR_FAULT;
1702         }
1703 exit:
1704         up_read(&current->mm->mmap_sem);
1705         return pfn;
1706 }
1707
1708 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn,
1709                                bool atomic, bool *async, bool write_fault,
1710                                bool *writable)
1711 {
1712         unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault);
1713
1714         if (addr == KVM_HVA_ERR_RO_BAD) {
1715                 if (writable)
1716                         *writable = false;
1717                 return KVM_PFN_ERR_RO_FAULT;
1718         }
1719
1720         if (kvm_is_error_hva(addr)) {
1721                 if (writable)
1722                         *writable = false;
1723                 return KVM_PFN_NOSLOT;
1724         }
1725
1726         /* Do not map writable pfn in the readonly memslot. */
1727         if (writable && memslot_is_readonly(slot)) {
1728                 *writable = false;
1729                 writable = NULL;
1730         }
1731
1732         return hva_to_pfn(addr, atomic, async, write_fault,
1733                           writable);
1734 }
1735 EXPORT_SYMBOL_GPL(__gfn_to_pfn_memslot);
1736
1737 kvm_pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
1738                       bool *writable)
1739 {
1740         return __gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn, false, NULL,
1741                                     write_fault, writable);
1742 }
1743 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
1744
1745 kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn)
1746 {
1747         return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL);
1748 }
1749 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot);
1750
1751 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn)
1752 {
1753         return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL);
1754 }
1755 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic);
1756
1757 kvm_pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
1758 {
1759         return gfn_to_pfn_memslot_atomic(gfn_to_memslot(kvm, gfn), gfn);
1760 }
1761 EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
1762
1763 kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn)
1764 {
1765         return gfn_to_pfn_memslot_atomic(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1766 }
1767 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn_atomic);
1768
1769 kvm_pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
1770 {
1771         return gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn);
1772 }
1773 EXPORT_SYMBOL_GPL(gfn_to_pfn);
1774
1775 kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn)
1776 {
1777         return gfn_to_pfn_memslot(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn);
1778 }
1779 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn);
1780
1781 int gfn_to_page_many_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
1782                             struct page **pages, int nr_pages)
1783 {
1784         unsigned long addr;
1785         gfn_t entry = 0;
1786
1787         addr = gfn_to_hva_many(slot, gfn, &entry);
1788         if (kvm_is_error_hva(addr))
1789                 return -1;
1790
1791         if (entry < nr_pages)
1792                 return 0;
1793
1794         return __get_user_pages_fast(addr, nr_pages, 1, pages);
1795 }
1796 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
1797
1798 static struct page *kvm_pfn_to_page(kvm_pfn_t pfn)
1799 {
1800         if (is_error_noslot_pfn(pfn))
1801                 return KVM_ERR_PTR_BAD_PAGE;
1802
1803         if (kvm_is_reserved_pfn(pfn)) {
1804                 WARN_ON(1);
1805                 return KVM_ERR_PTR_BAD_PAGE;
1806         }
1807
1808         return pfn_to_page(pfn);
1809 }
1810
1811 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
1812 {
1813         kvm_pfn_t pfn;
1814
1815         pfn = gfn_to_pfn(kvm, gfn);
1816
1817         return kvm_pfn_to_page(pfn);
1818 }
1819 EXPORT_SYMBOL_GPL(gfn_to_page);
1820
1821 void kvm_release_pfn(kvm_pfn_t pfn, bool dirty, struct gfn_to_pfn_cache *cache)
1822 {
1823         if (pfn == 0)
1824                 return;
1825
1826         if (cache)
1827                 cache->pfn = cache->gfn = 0;
1828
1829         if (dirty)
1830                 kvm_release_pfn_dirty(pfn);
1831         else
1832                 kvm_release_pfn_clean(pfn);
1833 }
1834
1835 static void kvm_cache_gfn_to_pfn(struct kvm_memory_slot *slot, gfn_t gfn,
1836                                  struct gfn_to_pfn_cache *cache, u64 gen)
1837 {
1838         kvm_release_pfn(cache->pfn, cache->dirty, cache);
1839
1840         cache->pfn = gfn_to_pfn_memslot(slot, gfn);
1841         cache->gfn = gfn;
1842         cache->dirty = false;
1843         cache->generation = gen;
1844 }
1845
1846 static int __kvm_map_gfn(struct kvm_memslots *slots, gfn_t gfn,
1847                          struct kvm_host_map *map,
1848                          struct gfn_to_pfn_cache *cache,
1849                          bool atomic)
1850 {
1851         kvm_pfn_t pfn;
1852         void *hva = NULL;
1853         struct page *page = KVM_UNMAPPED_PAGE;
1854         struct kvm_memory_slot *slot = __gfn_to_memslot(slots, gfn);
1855         u64 gen = slots->generation;
1856
1857         if (!map)
1858                 return -EINVAL;
1859
1860         if (cache) {
1861                 if (!cache->pfn || cache->gfn != gfn ||
1862                         cache->generation != gen) {
1863                         if (atomic)
1864                                 return -EAGAIN;
1865                         kvm_cache_gfn_to_pfn(slot, gfn, cache, gen);
1866                 }
1867                 pfn = cache->pfn;
1868         } else {
1869                 if (atomic)
1870                         return -EAGAIN;
1871                 pfn = gfn_to_pfn_memslot(slot, gfn);
1872         }
1873         if (is_error_noslot_pfn(pfn))
1874                 return -EINVAL;
1875
1876         if (pfn_valid(pfn)) {
1877                 page = pfn_to_page(pfn);
1878                 if (atomic)
1879                         hva = kmap_atomic(page);
1880                 else
1881                         hva = kmap(page);
1882 #ifdef CONFIG_HAS_IOMEM
1883         } else if (!atomic) {
1884                 hva = memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB);
1885         } else {
1886                 return -EINVAL;
1887 #endif
1888         }
1889
1890         if (!hva)
1891                 return -EFAULT;
1892
1893         map->page = page;
1894         map->hva = hva;
1895         map->pfn = pfn;
1896         map->gfn = gfn;
1897
1898         return 0;
1899 }
1900
1901 int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map,
1902                 struct gfn_to_pfn_cache *cache, bool atomic)
1903 {
1904         return __kvm_map_gfn(kvm_memslots(vcpu->kvm), gfn, map,
1905                         cache, atomic);
1906 }
1907 EXPORT_SYMBOL_GPL(kvm_map_gfn);
1908
1909 int kvm_vcpu_map(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map)
1910 {
1911         return __kvm_map_gfn(kvm_vcpu_memslots(vcpu), gfn, map,
1912                 NULL, false);
1913 }
1914 EXPORT_SYMBOL_GPL(kvm_vcpu_map);
1915
1916 static void __kvm_unmap_gfn(struct kvm_memory_slot *memslot,
1917                         struct kvm_host_map *map,
1918                         struct gfn_to_pfn_cache *cache,
1919                         bool dirty, bool atomic)
1920 {
1921         if (!map)
1922                 return;
1923
1924         if (!map->hva)
1925                 return;
1926
1927         if (map->page != KVM_UNMAPPED_PAGE) {
1928                 if (atomic)
1929                         kunmap_atomic(map->hva);
1930                 else
1931                         kunmap(map->page);
1932         }
1933 #ifdef CONFIG_HAS_IOMEM
1934         else if (!atomic)
1935                 memunmap(map->hva);
1936         else
1937                 WARN_ONCE(1, "Unexpected unmapping in atomic context");
1938 #endif
1939
1940         if (dirty)
1941                 mark_page_dirty_in_slot(memslot, map->gfn);
1942
1943         if (cache)
1944                 cache->dirty |= dirty;
1945         else
1946                 kvm_release_pfn(map->pfn, dirty, NULL);
1947
1948         map->hva = NULL;
1949         map->page = NULL;
1950 }
1951
1952 int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map,
1953                   struct gfn_to_pfn_cache *cache, bool dirty, bool atomic)
1954 {
1955         __kvm_unmap_gfn(gfn_to_memslot(vcpu->kvm, map->gfn), map,
1956                         cache, dirty, atomic);
1957         return 0;
1958 }
1959 EXPORT_SYMBOL_GPL(kvm_unmap_gfn);
1960
1961 void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty)
1962 {
1963         __kvm_unmap_gfn(kvm_vcpu_gfn_to_memslot(vcpu, map->gfn), map, NULL,
1964                         dirty, false);
1965 }
1966 EXPORT_SYMBOL_GPL(kvm_vcpu_unmap);
1967
1968 struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn)
1969 {
1970         kvm_pfn_t pfn;
1971
1972         pfn = kvm_vcpu_gfn_to_pfn(vcpu, gfn);
1973
1974         return kvm_pfn_to_page(pfn);
1975 }
1976 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_page);
1977
1978 void kvm_release_page_clean(struct page *page)
1979 {
1980         WARN_ON(is_error_page(page));
1981
1982         kvm_release_pfn_clean(page_to_pfn(page));
1983 }
1984 EXPORT_SYMBOL_GPL(kvm_release_page_clean);
1985
1986 void kvm_release_pfn_clean(kvm_pfn_t pfn)
1987 {
1988         if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn))
1989                 put_page(pfn_to_page(pfn));
1990 }
1991 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
1992
1993 void kvm_release_page_dirty(struct page *page)
1994 {
1995         WARN_ON(is_error_page(page));
1996
1997         kvm_release_pfn_dirty(page_to_pfn(page));
1998 }
1999 EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
2000
2001 void kvm_release_pfn_dirty(kvm_pfn_t pfn)
2002 {
2003         kvm_set_pfn_dirty(pfn);
2004         kvm_release_pfn_clean(pfn);
2005 }
2006 EXPORT_SYMBOL_GPL(kvm_release_pfn_dirty);
2007
2008 void kvm_set_pfn_dirty(kvm_pfn_t pfn)
2009 {
2010         if (!kvm_is_reserved_pfn(pfn) && !kvm_is_zone_device_pfn(pfn))
2011                 SetPageDirty(pfn_to_page(pfn));
2012 }
2013 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
2014
2015 void kvm_set_pfn_accessed(kvm_pfn_t pfn)
2016 {
2017         if (!kvm_is_reserved_pfn(pfn) && !kvm_is_zone_device_pfn(pfn))
2018                 mark_page_accessed(pfn_to_page(pfn));
2019 }
2020 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
2021
2022 void kvm_get_pfn(kvm_pfn_t pfn)
2023 {
2024         if (!kvm_is_reserved_pfn(pfn))
2025                 get_page(pfn_to_page(pfn));
2026 }
2027 EXPORT_SYMBOL_GPL(kvm_get_pfn);
2028
2029 static int next_segment(unsigned long len, int offset)
2030 {
2031         if (len > PAGE_SIZE - offset)
2032                 return PAGE_SIZE - offset;
2033         else
2034                 return len;
2035 }
2036
2037 static int __kvm_read_guest_page(struct kvm_memory_slot *slot, gfn_t gfn,
2038                                  void *data, int offset, int len)
2039 {
2040         int r;
2041         unsigned long addr;
2042
2043         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
2044         if (kvm_is_error_hva(addr))
2045                 return -EFAULT;
2046         r = __copy_from_user(data, (void __user *)addr + offset, len);
2047         if (r)
2048                 return -EFAULT;
2049         return 0;
2050 }
2051
2052 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
2053                         int len)
2054 {
2055         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
2056
2057         return __kvm_read_guest_page(slot, gfn, data, offset, len);
2058 }
2059 EXPORT_SYMBOL_GPL(kvm_read_guest_page);
2060
2061 int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data,
2062                              int offset, int len)
2063 {
2064         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2065
2066         return __kvm_read_guest_page(slot, gfn, data, offset, len);
2067 }
2068 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_page);
2069
2070 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
2071 {
2072         gfn_t gfn = gpa >> PAGE_SHIFT;
2073         int seg;
2074         int offset = offset_in_page(gpa);
2075         int ret;
2076
2077         while ((seg = next_segment(len, offset)) != 0) {
2078                 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
2079                 if (ret < 0)
2080                         return ret;
2081                 offset = 0;
2082                 len -= seg;
2083                 data += seg;
2084                 ++gfn;
2085         }
2086         return 0;
2087 }
2088 EXPORT_SYMBOL_GPL(kvm_read_guest);
2089
2090 int kvm_vcpu_read_guest(struct kvm_vcpu *vcpu, gpa_t gpa, void *data, unsigned long len)
2091 {
2092         gfn_t gfn = gpa >> PAGE_SHIFT;
2093         int seg;
2094         int offset = offset_in_page(gpa);
2095         int ret;
2096
2097         while ((seg = next_segment(len, offset)) != 0) {
2098                 ret = kvm_vcpu_read_guest_page(vcpu, gfn, data, offset, seg);
2099                 if (ret < 0)
2100                         return ret;
2101                 offset = 0;
2102                 len -= seg;
2103                 data += seg;
2104                 ++gfn;
2105         }
2106         return 0;
2107 }
2108 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest);
2109
2110 static int __kvm_read_guest_atomic(struct kvm_memory_slot *slot, gfn_t gfn,
2111                                    void *data, int offset, unsigned long len)
2112 {
2113         int r;
2114         unsigned long addr;
2115
2116         addr = gfn_to_hva_memslot_prot(slot, gfn, NULL);
2117         if (kvm_is_error_hva(addr))
2118                 return -EFAULT;
2119         pagefault_disable();
2120         r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len);
2121         pagefault_enable();
2122         if (r)
2123                 return -EFAULT;
2124         return 0;
2125 }
2126
2127 int kvm_vcpu_read_guest_atomic(struct kvm_vcpu *vcpu, gpa_t gpa,
2128                                void *data, unsigned long len)
2129 {
2130         gfn_t gfn = gpa >> PAGE_SHIFT;
2131         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2132         int offset = offset_in_page(gpa);
2133
2134         return __kvm_read_guest_atomic(slot, gfn, data, offset, len);
2135 }
2136 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_atomic);
2137
2138 static int __kvm_write_guest_page(struct kvm_memory_slot *memslot, gfn_t gfn,
2139                                   const void *data, int offset, int len)
2140 {
2141         int r;
2142         unsigned long addr;
2143
2144         addr = gfn_to_hva_memslot(memslot, gfn);
2145         if (kvm_is_error_hva(addr))
2146                 return -EFAULT;
2147         r = __copy_to_user((void __user *)addr + offset, data, len);
2148         if (r)
2149                 return -EFAULT;
2150         mark_page_dirty_in_slot(memslot, gfn);
2151         return 0;
2152 }
2153
2154 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn,
2155                          const void *data, int offset, int len)
2156 {
2157         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
2158
2159         return __kvm_write_guest_page(slot, gfn, data, offset, len);
2160 }
2161 EXPORT_SYMBOL_GPL(kvm_write_guest_page);
2162
2163 int kvm_vcpu_write_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn,
2164                               const void *data, int offset, int len)
2165 {
2166         struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2167
2168         return __kvm_write_guest_page(slot, gfn, data, offset, len);
2169 }
2170 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest_page);
2171
2172 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
2173                     unsigned long len)
2174 {
2175         gfn_t gfn = gpa >> PAGE_SHIFT;
2176         int seg;
2177         int offset = offset_in_page(gpa);
2178         int ret;
2179
2180         while ((seg = next_segment(len, offset)) != 0) {
2181                 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
2182                 if (ret < 0)
2183                         return ret;
2184                 offset = 0;
2185                 len -= seg;
2186                 data += seg;
2187                 ++gfn;
2188         }
2189         return 0;
2190 }
2191 EXPORT_SYMBOL_GPL(kvm_write_guest);
2192
2193 int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data,
2194                          unsigned long len)
2195 {
2196         gfn_t gfn = gpa >> PAGE_SHIFT;
2197         int seg;
2198         int offset = offset_in_page(gpa);
2199         int ret;
2200
2201         while ((seg = next_segment(len, offset)) != 0) {
2202                 ret = kvm_vcpu_write_guest_page(vcpu, gfn, data, offset, seg);
2203                 if (ret < 0)
2204                         return ret;
2205                 offset = 0;
2206                 len -= seg;
2207                 data += seg;
2208                 ++gfn;
2209         }
2210         return 0;
2211 }
2212 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest);
2213
2214 static int __kvm_gfn_to_hva_cache_init(struct kvm_memslots *slots,
2215                                        struct gfn_to_hva_cache *ghc,
2216                                        gpa_t gpa, unsigned long len)
2217 {
2218         int offset = offset_in_page(gpa);
2219         gfn_t start_gfn = gpa >> PAGE_SHIFT;
2220         gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT;
2221         gfn_t nr_pages_needed = end_gfn - start_gfn + 1;
2222         gfn_t nr_pages_avail;
2223
2224         /* Update ghc->generation before performing any error checks. */
2225         ghc->generation = slots->generation;
2226
2227         if (start_gfn > end_gfn) {
2228                 ghc->hva = KVM_HVA_ERR_BAD;
2229                 return -EINVAL;
2230         }
2231
2232         /*
2233          * If the requested region crosses two memslots, we still
2234          * verify that the entire region is valid here.
2235          */
2236         for ( ; start_gfn <= end_gfn; start_gfn += nr_pages_avail) {
2237                 ghc->memslot = __gfn_to_memslot(slots, start_gfn);
2238                 ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn,
2239                                            &nr_pages_avail);
2240                 if (kvm_is_error_hva(ghc->hva))
2241                         return -EFAULT;
2242         }
2243
2244         /* Use the slow path for cross page reads and writes. */
2245         if (nr_pages_needed == 1)
2246                 ghc->hva += offset;
2247         else
2248                 ghc->memslot = NULL;
2249
2250         ghc->gpa = gpa;
2251         ghc->len = len;
2252         return 0;
2253 }
2254
2255 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2256                               gpa_t gpa, unsigned long len)
2257 {
2258         struct kvm_memslots *slots = kvm_memslots(kvm);
2259         return __kvm_gfn_to_hva_cache_init(slots, ghc, gpa, len);
2260 }
2261 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
2262
2263 int kvm_write_guest_offset_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2264                                   void *data, unsigned int offset,
2265                                   unsigned long len)
2266 {
2267         struct kvm_memslots *slots = kvm_memslots(kvm);
2268         int r;
2269         gpa_t gpa = ghc->gpa + offset;
2270
2271         BUG_ON(len + offset > ghc->len);
2272
2273         if (slots->generation != ghc->generation) {
2274                 if (__kvm_gfn_to_hva_cache_init(slots, ghc, ghc->gpa, ghc->len))
2275                         return -EFAULT;
2276         }
2277
2278         if (kvm_is_error_hva(ghc->hva))
2279                 return -EFAULT;
2280
2281         if (unlikely(!ghc->memslot))
2282                 return kvm_write_guest(kvm, gpa, data, len);
2283
2284         r = __copy_to_user((void __user *)ghc->hva + offset, data, len);
2285         if (r)
2286                 return -EFAULT;
2287         mark_page_dirty_in_slot(ghc->memslot, gpa >> PAGE_SHIFT);
2288
2289         return 0;
2290 }
2291 EXPORT_SYMBOL_GPL(kvm_write_guest_offset_cached);
2292
2293 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2294                            void *data, unsigned long len)
2295 {
2296         return kvm_write_guest_offset_cached(kvm, ghc, data, 0, len);
2297 }
2298 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
2299
2300 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
2301                            void *data, unsigned long len)
2302 {
2303         struct kvm_memslots *slots = kvm_memslots(kvm);
2304         int r;
2305
2306         BUG_ON(len > ghc->len);
2307
2308         if (slots->generation != ghc->generation) {
2309                 if (__kvm_gfn_to_hva_cache_init(slots, ghc, ghc->gpa, ghc->len))
2310                         return -EFAULT;
2311         }
2312
2313         if (kvm_is_error_hva(ghc->hva))
2314                 return -EFAULT;
2315
2316         if (unlikely(!ghc->memslot))
2317                 return kvm_read_guest(kvm, ghc->gpa, data, len);
2318
2319         r = __copy_from_user(data, (void __user *)ghc->hva, len);
2320         if (r)
2321                 return -EFAULT;
2322
2323         return 0;
2324 }
2325 EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
2326
2327 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
2328 {
2329         const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0)));
2330
2331         return kvm_write_guest_page(kvm, gfn, zero_page, offset, len);
2332 }
2333 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
2334
2335 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
2336 {
2337         gfn_t gfn = gpa >> PAGE_SHIFT;
2338         int seg;
2339         int offset = offset_in_page(gpa);
2340         int ret;
2341
2342         while ((seg = next_segment(len, offset)) != 0) {
2343                 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
2344                 if (ret < 0)
2345                         return ret;
2346                 offset = 0;
2347                 len -= seg;
2348                 ++gfn;
2349         }
2350         return 0;
2351 }
2352 EXPORT_SYMBOL_GPL(kvm_clear_guest);
2353
2354 static void mark_page_dirty_in_slot(struct kvm_memory_slot *memslot,
2355                                     gfn_t gfn)
2356 {
2357         if (memslot && memslot->dirty_bitmap) {
2358                 unsigned long rel_gfn = gfn - memslot->base_gfn;
2359
2360                 set_bit_le(rel_gfn, memslot->dirty_bitmap);
2361         }
2362 }
2363
2364 void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
2365 {
2366         struct kvm_memory_slot *memslot;
2367
2368         memslot = gfn_to_memslot(kvm, gfn);
2369         mark_page_dirty_in_slot(memslot, gfn);
2370 }
2371 EXPORT_SYMBOL_GPL(mark_page_dirty);
2372
2373 void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn)
2374 {
2375         struct kvm_memory_slot *memslot;
2376
2377         memslot = kvm_vcpu_gfn_to_memslot(vcpu, gfn);
2378         mark_page_dirty_in_slot(memslot, gfn);
2379 }
2380 EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty);
2381
2382 void kvm_sigset_activate(struct kvm_vcpu *vcpu)
2383 {
2384         if (!vcpu->sigset_active)
2385                 return;
2386
2387         /*
2388          * This does a lockless modification of ->real_blocked, which is fine
2389          * because, only current can change ->real_blocked and all readers of
2390          * ->real_blocked don't care as long ->real_blocked is always a subset
2391          * of ->blocked.
2392          */
2393         sigprocmask(SIG_SETMASK, &vcpu->sigset, &current->real_blocked);
2394 }
2395
2396 void kvm_sigset_deactivate(struct kvm_vcpu *vcpu)
2397 {
2398         if (!vcpu->sigset_active)
2399                 return;
2400
2401         sigprocmask(SIG_SETMASK, &current->real_blocked, NULL);
2402         sigemptyset(&current->real_blocked);
2403 }
2404
2405 static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
2406 {
2407         unsigned int old, val, grow, grow_start;
2408
2409         old = val = vcpu->halt_poll_ns;
2410         grow_start = READ_ONCE(halt_poll_ns_grow_start);
2411         grow = READ_ONCE(halt_poll_ns_grow);
2412         if (!grow)
2413                 goto out;
2414
2415         val *= grow;
2416         if (val < grow_start)
2417                 val = grow_start;
2418
2419         if (val > halt_poll_ns)
2420                 val = halt_poll_ns;
2421
2422         vcpu->halt_poll_ns = val;
2423 out:
2424         trace_kvm_halt_poll_ns_grow(vcpu->vcpu_id, val, old);
2425 }
2426
2427 static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu)
2428 {
2429         unsigned int old, val, shrink;
2430
2431         old = val = vcpu->halt_poll_ns;
2432         shrink = READ_ONCE(halt_poll_ns_shrink);
2433         if (shrink == 0)
2434                 val = 0;
2435         else
2436                 val /= shrink;
2437
2438         vcpu->halt_poll_ns = val;
2439         trace_kvm_halt_poll_ns_shrink(vcpu->vcpu_id, val, old);
2440 }
2441
2442 static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu)
2443 {
2444         int ret = -EINTR;
2445         int idx = srcu_read_lock(&vcpu->kvm->srcu);
2446
2447         if (kvm_arch_vcpu_runnable(vcpu)) {
2448                 kvm_make_request(KVM_REQ_UNHALT, vcpu);
2449                 goto out;
2450         }
2451         if (kvm_cpu_has_pending_timer(vcpu))
2452                 goto out;
2453         if (signal_pending(current))
2454                 goto out;
2455
2456         ret = 0;
2457 out:
2458         srcu_read_unlock(&vcpu->kvm->srcu, idx);
2459         return ret;
2460 }
2461
2462 /*
2463  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
2464  */
2465 void kvm_vcpu_block(struct kvm_vcpu *vcpu)
2466 {
2467         ktime_t start, cur;
2468         DECLARE_SWAITQUEUE(wait);
2469         bool waited = false;
2470         u64 block_ns;
2471
2472         kvm_arch_vcpu_blocking(vcpu);
2473
2474         start = cur = ktime_get();
2475         if (vcpu->halt_poll_ns && !kvm_arch_no_poll(vcpu)) {
2476                 ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
2477
2478                 ++vcpu->stat.halt_attempted_poll;
2479                 do {
2480                         /*
2481                          * This sets KVM_REQ_UNHALT if an interrupt
2482                          * arrives.
2483                          */
2484                         if (kvm_vcpu_check_block(vcpu) < 0) {
2485                                 ++vcpu->stat.halt_successful_poll;
2486                                 if (!vcpu_valid_wakeup(vcpu))
2487                                         ++vcpu->stat.halt_poll_invalid;
2488                                 goto out;
2489                         }
2490                         cur = ktime_get();
2491                 } while (single_task_running() && ktime_before(cur, stop));
2492         }
2493
2494         for (;;) {
2495                 prepare_to_swait_exclusive(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
2496
2497                 if (kvm_vcpu_check_block(vcpu) < 0)
2498                         break;
2499
2500                 waited = true;
2501                 schedule();
2502         }
2503
2504         finish_swait(&vcpu->wq, &wait);
2505         cur = ktime_get();
2506 out:
2507         kvm_arch_vcpu_unblocking(vcpu);
2508         block_ns = ktime_to_ns(cur) - ktime_to_ns(start);
2509
2510         if (!kvm_arch_no_poll(vcpu)) {
2511                 if (!vcpu_valid_wakeup(vcpu)) {
2512                         shrink_halt_poll_ns(vcpu);
2513                 } else if (halt_poll_ns) {
2514                         if (block_ns <= vcpu->halt_poll_ns)
2515                                 ;
2516                         /* we had a long block, shrink polling */
2517                         else if (vcpu->halt_poll_ns && block_ns > halt_poll_ns)
2518                                 shrink_halt_poll_ns(vcpu);
2519                         /* we had a short halt and our poll time is too small */
2520                         else if (vcpu->halt_poll_ns < halt_poll_ns &&
2521                                 block_ns < halt_poll_ns)
2522                                 grow_halt_poll_ns(vcpu);
2523                 } else {
2524                         vcpu->halt_poll_ns = 0;
2525                 }
2526         }
2527
2528         trace_kvm_vcpu_wakeup(block_ns, waited, vcpu_valid_wakeup(vcpu));
2529         kvm_arch_vcpu_block_finish(vcpu);
2530 }
2531 EXPORT_SYMBOL_GPL(kvm_vcpu_block);
2532
2533 bool kvm_vcpu_wake_up(struct kvm_vcpu *vcpu)
2534 {
2535         struct swait_queue_head *wqp;
2536
2537         wqp = kvm_arch_vcpu_wq(vcpu);
2538         if (swq_has_sleeper(wqp)) {
2539                 swake_up_one(wqp);
2540                 WRITE_ONCE(vcpu->ready, true);
2541                 ++vcpu->stat.halt_wakeup;
2542                 return true;
2543         }
2544
2545         return false;
2546 }
2547 EXPORT_SYMBOL_GPL(kvm_vcpu_wake_up);
2548
2549 #ifndef CONFIG_S390
2550 /*
2551  * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode.
2552  */
2553 void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
2554 {
2555         int me;
2556         int cpu = vcpu->cpu;
2557
2558         if (kvm_vcpu_wake_up(vcpu))
2559                 return;
2560
2561         me = get_cpu();
2562         if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
2563                 if (kvm_arch_vcpu_should_kick(vcpu))
2564                         smp_send_reschedule(cpu);
2565         put_cpu();
2566 }
2567 EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
2568 #endif /* !CONFIG_S390 */
2569
2570 int kvm_vcpu_yield_to(struct kvm_vcpu *target)
2571 {
2572         struct pid *pid;
2573         struct task_struct *task = NULL;
2574         int ret = 0;
2575
2576         rcu_read_lock();
2577         pid = rcu_dereference(target->pid);
2578         if (pid)
2579                 task = get_pid_task(pid, PIDTYPE_PID);
2580         rcu_read_unlock();
2581         if (!task)
2582                 return ret;
2583         ret = yield_to(task, 1);
2584         put_task_struct(task);
2585
2586         return ret;
2587 }
2588 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to);
2589
2590 /*
2591  * Helper that checks whether a VCPU is eligible for directed yield.
2592  * Most eligible candidate to yield is decided by following heuristics:
2593  *
2594  *  (a) VCPU which has not done pl-exit or cpu relax intercepted recently
2595  *  (preempted lock holder), indicated by @in_spin_loop.
2596  *  Set at the beiginning and cleared at the end of interception/PLE handler.
2597  *
2598  *  (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get
2599  *  chance last time (mostly it has become eligible now since we have probably
2600  *  yielded to lockholder in last iteration. This is done by toggling
2601  *  @dy_eligible each time a VCPU checked for eligibility.)
2602  *
2603  *  Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding
2604  *  to preempted lock-holder could result in wrong VCPU selection and CPU
2605  *  burning. Giving priority for a potential lock-holder increases lock
2606  *  progress.
2607  *
2608  *  Since algorithm is based on heuristics, accessing another VCPU data without
2609  *  locking does not harm. It may result in trying to yield to  same VCPU, fail
2610  *  and continue with next VCPU and so on.
2611  */
2612 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu)
2613 {
2614 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT
2615         bool eligible;
2616
2617         eligible = !vcpu->spin_loop.in_spin_loop ||
2618                     vcpu->spin_loop.dy_eligible;
2619
2620         if (vcpu->spin_loop.in_spin_loop)
2621                 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible);
2622
2623         return eligible;
2624 #else
2625         return true;
2626 #endif
2627 }
2628
2629 /*
2630  * Unlike kvm_arch_vcpu_runnable, this function is called outside
2631  * a vcpu_load/vcpu_put pair.  However, for most architectures
2632  * kvm_arch_vcpu_runnable does not require vcpu_load.
2633  */
2634 bool __weak kvm_arch_dy_runnable(struct kvm_vcpu *vcpu)
2635 {
2636         return kvm_arch_vcpu_runnable(vcpu);
2637 }
2638
2639 static bool vcpu_dy_runnable(struct kvm_vcpu *vcpu)
2640 {
2641         if (kvm_arch_dy_runnable(vcpu))
2642                 return true;
2643
2644 #ifdef CONFIG_KVM_ASYNC_PF
2645         if (!list_empty_careful(&vcpu->async_pf.done))
2646                 return true;
2647 #endif
2648
2649         return false;
2650 }
2651
2652 void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode)
2653 {
2654         struct kvm *kvm = me->kvm;
2655         struct kvm_vcpu *vcpu;
2656         int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
2657         int yielded = 0;
2658         int try = 3;
2659         int pass;
2660         int i;
2661
2662         kvm_vcpu_set_in_spin_loop(me, true);
2663         /*
2664          * We boost the priority of a VCPU that is runnable but not
2665          * currently running, because it got preempted by something
2666          * else and called schedule in __vcpu_run.  Hopefully that
2667          * VCPU is holding the lock that we need and will release it.
2668          * We approximate round-robin by starting at the last boosted VCPU.
2669          */
2670         for (pass = 0; pass < 2 && !yielded && try; pass++) {
2671                 kvm_for_each_vcpu(i, vcpu, kvm) {
2672                         if (!pass && i <= last_boosted_vcpu) {
2673                                 i = last_boosted_vcpu;
2674                                 continue;
2675                         } else if (pass && i > last_boosted_vcpu)
2676                                 break;
2677                         if (!READ_ONCE(vcpu->ready))
2678                                 continue;
2679                         if (vcpu == me)
2680                                 continue;
2681                         if (swait_active(&vcpu->wq) && !vcpu_dy_runnable(vcpu))
2682                                 continue;
2683                         if (READ_ONCE(vcpu->preempted) && yield_to_kernel_mode &&
2684                                 !kvm_arch_vcpu_in_kernel(vcpu))
2685                                 continue;
2686                         if (!kvm_vcpu_eligible_for_directed_yield(vcpu))
2687                                 continue;
2688
2689                         yielded = kvm_vcpu_yield_to(vcpu);
2690                         if (yielded > 0) {
2691                                 kvm->last_boosted_vcpu = i;
2692                                 break;
2693                         } else if (yielded < 0) {
2694                                 try--;
2695                                 if (!try)
2696                                         break;
2697                         }
2698                 }
2699         }
2700         kvm_vcpu_set_in_spin_loop(me, false);
2701
2702         /* Ensure vcpu is not eligible during next spinloop */
2703         kvm_vcpu_set_dy_eligible(me, false);
2704 }
2705 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
2706
2707 static vm_fault_t kvm_vcpu_fault(struct vm_fault *vmf)
2708 {
2709         struct kvm_vcpu *vcpu = vmf->vma->vm_file->private_data;
2710         struct page *page;
2711
2712         if (vmf->pgoff == 0)
2713                 page = virt_to_page(vcpu->run);
2714 #ifdef CONFIG_X86
2715         else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
2716                 page = virt_to_page(vcpu->arch.pio_data);
2717 #endif
2718 #ifdef CONFIG_KVM_MMIO
2719         else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
2720                 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
2721 #endif
2722         else
2723                 return kvm_arch_vcpu_fault(vcpu, vmf);
2724         get_page(page);
2725         vmf->page = page;
2726         return 0;
2727 }
2728
2729 static const struct vm_operations_struct kvm_vcpu_vm_ops = {
2730         .fault = kvm_vcpu_fault,
2731 };
2732
2733 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
2734 {
2735         vma->vm_ops = &kvm_vcpu_vm_ops;
2736         return 0;
2737 }
2738
2739 static int kvm_vcpu_release(struct inode *inode, struct file *filp)
2740 {
2741         struct kvm_vcpu *vcpu = filp->private_data;
2742
2743         debugfs_remove_recursive(vcpu->debugfs_dentry);
2744         kvm_put_kvm(vcpu->kvm);
2745         return 0;
2746 }
2747
2748 static struct file_operations kvm_vcpu_fops = {
2749         .release        = kvm_vcpu_release,
2750         .unlocked_ioctl = kvm_vcpu_ioctl,
2751         .mmap           = kvm_vcpu_mmap,
2752         .llseek         = noop_llseek,
2753         KVM_COMPAT(kvm_vcpu_compat_ioctl),
2754 };
2755
2756 /*
2757  * Allocates an inode for the vcpu.
2758  */
2759 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
2760 {
2761         char name[8 + 1 + ITOA_MAX_LEN + 1];
2762
2763         snprintf(name, sizeof(name), "kvm-vcpu:%d", vcpu->vcpu_id);
2764         return anon_inode_getfd(name, &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC);
2765 }
2766
2767 static void kvm_create_vcpu_debugfs(struct kvm_vcpu *vcpu)
2768 {
2769 #ifdef __KVM_HAVE_ARCH_VCPU_DEBUGFS
2770         char dir_name[ITOA_MAX_LEN * 2];
2771
2772         if (!debugfs_initialized())
2773                 return;
2774
2775         snprintf(dir_name, sizeof(dir_name), "vcpu%d", vcpu->vcpu_id);
2776         vcpu->debugfs_dentry = debugfs_create_dir(dir_name,
2777                                                   vcpu->kvm->debugfs_dentry);
2778
2779         kvm_arch_create_vcpu_debugfs(vcpu);
2780 #endif
2781 }
2782
2783 /*
2784  * Creates some virtual cpus.  Good luck creating more than one.
2785  */
2786 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
2787 {
2788         int r;
2789         struct kvm_vcpu *vcpu;
2790         struct page *page;
2791
2792         if (id >= KVM_MAX_VCPU_ID)
2793                 return -EINVAL;
2794
2795         mutex_lock(&kvm->lock);
2796         if (kvm->created_vcpus == KVM_MAX_VCPUS) {
2797                 mutex_unlock(&kvm->lock);
2798                 return -EINVAL;
2799         }
2800
2801         kvm->created_vcpus++;
2802         mutex_unlock(&kvm->lock);
2803
2804         r = kvm_arch_vcpu_precreate(kvm, id);
2805         if (r)
2806                 goto vcpu_decrement;
2807
2808         vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
2809         if (!vcpu) {
2810                 r = -ENOMEM;
2811                 goto vcpu_decrement;
2812         }
2813
2814         BUILD_BUG_ON(sizeof(struct kvm_run) > PAGE_SIZE);
2815         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2816         if (!page) {
2817                 r = -ENOMEM;
2818                 goto vcpu_free;
2819         }
2820         vcpu->run = page_address(page);
2821
2822         kvm_vcpu_init(vcpu, kvm, id);
2823
2824         r = kvm_arch_vcpu_create(vcpu);
2825         if (r)
2826                 goto vcpu_free_run_page;
2827
2828         kvm_create_vcpu_debugfs(vcpu);
2829
2830         mutex_lock(&kvm->lock);
2831         if (kvm_get_vcpu_by_id(kvm, id)) {
2832                 r = -EEXIST;
2833                 goto unlock_vcpu_destroy;
2834         }
2835
2836         vcpu->vcpu_idx = atomic_read(&kvm->online_vcpus);
2837         BUG_ON(kvm->vcpus[vcpu->vcpu_idx]);
2838
2839         /* Now it's all set up, let userspace reach it */
2840         kvm_get_kvm(kvm);
2841         r = create_vcpu_fd(vcpu);
2842         if (r < 0) {
2843                 kvm_put_kvm_no_destroy(kvm);
2844                 goto unlock_vcpu_destroy;
2845         }
2846
2847         kvm->vcpus[vcpu->vcpu_idx] = vcpu;
2848
2849         /*
2850          * Pairs with smp_rmb() in kvm_get_vcpu.  Write kvm->vcpus
2851          * before kvm->online_vcpu's incremented value.
2852          */
2853         smp_wmb();
2854         atomic_inc(&kvm->online_vcpus);
2855
2856         mutex_unlock(&kvm->lock);
2857         kvm_arch_vcpu_postcreate(vcpu);
2858         return r;
2859
2860 unlock_vcpu_destroy:
2861         mutex_unlock(&kvm->lock);
2862         debugfs_remove_recursive(vcpu->debugfs_dentry);
2863         kvm_arch_vcpu_destroy(vcpu);
2864 vcpu_free_run_page:
2865         free_page((unsigned long)vcpu->run);
2866 vcpu_free:
2867         kmem_cache_free(kvm_vcpu_cache, vcpu);
2868 vcpu_decrement:
2869         mutex_lock(&kvm->lock);
2870         kvm->created_vcpus--;
2871         mutex_unlock(&kvm->lock);
2872         return r;
2873 }
2874
2875 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
2876 {
2877         if (sigset) {
2878                 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
2879                 vcpu->sigset_active = 1;
2880                 vcpu->sigset = *sigset;
2881         } else
2882                 vcpu->sigset_active = 0;
2883         return 0;
2884 }
2885
2886 static long kvm_vcpu_ioctl(struct file *filp,
2887                            unsigned int ioctl, unsigned long arg)
2888 {
2889         struct kvm_vcpu *vcpu = filp->private_data;
2890         void __user *argp = (void __user *)arg;
2891         int r;
2892         struct kvm_fpu *fpu = NULL;
2893         struct kvm_sregs *kvm_sregs = NULL;
2894
2895         if (vcpu->kvm->mm != current->mm)
2896                 return -EIO;
2897
2898         if (unlikely(_IOC_TYPE(ioctl) != KVMIO))
2899                 return -EINVAL;
2900
2901         /*
2902          * Some architectures have vcpu ioctls that are asynchronous to vcpu
2903          * execution; mutex_lock() would break them.
2904          */
2905         r = kvm_arch_vcpu_async_ioctl(filp, ioctl, arg);
2906         if (r != -ENOIOCTLCMD)
2907                 return r;
2908
2909         if (mutex_lock_killable(&vcpu->mutex))
2910                 return -EINTR;
2911         switch (ioctl) {
2912         case KVM_RUN: {
2913                 struct pid *oldpid;
2914                 r = -EINVAL;
2915                 if (arg)
2916                         goto out;
2917                 oldpid = rcu_access_pointer(vcpu->pid);
2918                 if (unlikely(oldpid != task_pid(current))) {
2919                         /* The thread running this VCPU changed. */
2920                         struct pid *newpid;
2921
2922                         r = kvm_arch_vcpu_run_pid_change(vcpu);
2923                         if (r)
2924                                 break;
2925
2926                         newpid = get_task_pid(current, PIDTYPE_PID);
2927                         rcu_assign_pointer(vcpu->pid, newpid);
2928                         if (oldpid)
2929                                 synchronize_rcu();
2930                         put_pid(oldpid);
2931                 }
2932                 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
2933                 trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
2934                 break;
2935         }
2936         case KVM_GET_REGS: {
2937                 struct kvm_regs *kvm_regs;
2938
2939                 r = -ENOMEM;
2940                 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL_ACCOUNT);
2941                 if (!kvm_regs)
2942                         goto out;
2943                 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
2944                 if (r)
2945                         goto out_free1;
2946                 r = -EFAULT;
2947                 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
2948                         goto out_free1;
2949                 r = 0;
2950 out_free1:
2951                 kfree(kvm_regs);
2952                 break;
2953         }
2954         case KVM_SET_REGS: {
2955                 struct kvm_regs *kvm_regs;
2956
2957                 r = -ENOMEM;
2958                 kvm_regs = memdup_user(argp, sizeof(*kvm_regs));
2959                 if (IS_ERR(kvm_regs)) {
2960                         r = PTR_ERR(kvm_regs);
2961                         goto out;
2962                 }
2963                 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
2964                 kfree(kvm_regs);
2965                 break;
2966         }
2967         case KVM_GET_SREGS: {
2968                 kvm_sregs = kzalloc(sizeof(struct kvm_sregs),
2969                                     GFP_KERNEL_ACCOUNT);
2970                 r = -ENOMEM;
2971                 if (!kvm_sregs)
2972                         goto out;
2973                 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
2974                 if (r)
2975                         goto out;
2976                 r = -EFAULT;
2977                 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
2978                         goto out;
2979                 r = 0;
2980                 break;
2981         }
2982         case KVM_SET_SREGS: {
2983                 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs));
2984                 if (IS_ERR(kvm_sregs)) {
2985                         r = PTR_ERR(kvm_sregs);
2986                         kvm_sregs = NULL;
2987                         goto out;
2988                 }
2989                 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
2990                 break;
2991         }
2992         case KVM_GET_MP_STATE: {
2993                 struct kvm_mp_state mp_state;
2994
2995                 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
2996                 if (r)
2997                         goto out;
2998                 r = -EFAULT;
2999                 if (copy_to_user(argp, &mp_state, sizeof(mp_state)))
3000                         goto out;
3001                 r = 0;
3002                 break;
3003         }
3004         case KVM_SET_MP_STATE: {
3005                 struct kvm_mp_state mp_state;
3006
3007                 r = -EFAULT;
3008                 if (copy_from_user(&mp_state, argp, sizeof(mp_state)))
3009                         goto out;
3010                 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
3011                 break;
3012         }
3013         case KVM_TRANSLATE: {
3014                 struct kvm_translation tr;
3015
3016                 r = -EFAULT;
3017                 if (copy_from_user(&tr, argp, sizeof(tr)))
3018                         goto out;
3019                 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
3020                 if (r)
3021                         goto out;
3022                 r = -EFAULT;
3023                 if (copy_to_user(argp, &tr, sizeof(tr)))
3024                         goto out;
3025                 r = 0;
3026                 break;
3027         }
3028         case KVM_SET_GUEST_DEBUG: {
3029                 struct kvm_guest_debug dbg;
3030
3031                 r = -EFAULT;
3032                 if (copy_from_user(&dbg, argp, sizeof(dbg)))
3033                         goto out;
3034                 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
3035                 break;
3036         }
3037         case KVM_SET_SIGNAL_MASK: {
3038                 struct kvm_signal_mask __user *sigmask_arg = argp;
3039                 struct kvm_signal_mask kvm_sigmask;
3040                 sigset_t sigset, *p;
3041
3042                 p = NULL;
3043                 if (argp) {
3044                         r = -EFAULT;
3045                         if (copy_from_user(&kvm_sigmask, argp,
3046                                            sizeof(kvm_sigmask)))
3047                                 goto out;
3048                         r = -EINVAL;
3049                         if (kvm_sigmask.len != sizeof(sigset))
3050                                 goto out;
3051                         r = -EFAULT;
3052                         if (copy_from_user(&sigset, sigmask_arg->sigset,
3053                                            sizeof(sigset)))
3054                                 goto out;
3055                         p = &sigset;
3056                 }
3057                 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
3058                 break;
3059         }
3060         case KVM_GET_FPU: {
3061                 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL_ACCOUNT);
3062                 r = -ENOMEM;
3063                 if (!fpu)
3064                         goto out;
3065                 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
3066                 if (r)
3067                         goto out;
3068                 r = -EFAULT;
3069                 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
3070                         goto out;
3071                 r = 0;
3072                 break;
3073         }
3074         case KVM_SET_FPU: {
3075                 fpu = memdup_user(argp, sizeof(*fpu));
3076                 if (IS_ERR(fpu)) {
3077                         r = PTR_ERR(fpu);
3078                         fpu = NULL;
3079                         goto out;
3080                 }
3081                 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
3082                 break;
3083         }
3084         default:
3085                 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
3086         }
3087 out:
3088         mutex_unlock(&vcpu->mutex);
3089         kfree(fpu);
3090         kfree(kvm_sregs);
3091         return r;
3092 }
3093
3094 #ifdef CONFIG_KVM_COMPAT
3095 static long kvm_vcpu_compat_ioctl(struct file *filp,
3096                                   unsigned int ioctl, unsigned long arg)
3097 {
3098         struct kvm_vcpu *vcpu = filp->private_data;
3099         void __user *argp = compat_ptr(arg);
3100         int r;
3101
3102         if (vcpu->kvm->mm != current->mm)
3103                 return -EIO;
3104
3105         switch (ioctl) {
3106         case KVM_SET_SIGNAL_MASK: {
3107                 struct kvm_signal_mask __user *sigmask_arg = argp;
3108                 struct kvm_signal_mask kvm_sigmask;
3109                 sigset_t sigset;
3110
3111                 if (argp) {
3112                         r = -EFAULT;
3113                         if (copy_from_user(&kvm_sigmask, argp,
3114                                            sizeof(kvm_sigmask)))
3115                                 goto out;
3116                         r = -EINVAL;
3117                         if (kvm_sigmask.len != sizeof(compat_sigset_t))
3118                                 goto out;
3119                         r = -EFAULT;
3120                         if (get_compat_sigset(&sigset, (void *)sigmask_arg->sigset))
3121                                 goto out;
3122                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
3123                 } else
3124                         r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL);
3125                 break;
3126         }
3127         default:
3128                 r = kvm_vcpu_ioctl(filp, ioctl, arg);
3129         }
3130
3131 out:
3132         return r;
3133 }
3134 #endif
3135
3136 static int kvm_device_mmap(struct file *filp, struct vm_area_struct *vma)
3137 {
3138         struct kvm_device *dev = filp->private_data;
3139
3140         if (dev->ops->mmap)
3141                 return dev->ops->mmap(dev, vma);
3142
3143         return -ENODEV;
3144 }
3145
3146 static int kvm_device_ioctl_attr(struct kvm_device *dev,
3147                                  int (*accessor)(struct kvm_device *dev,
3148                                                  struct kvm_device_attr *attr),
3149                                  unsigned long arg)
3150 {
3151         struct kvm_device_attr attr;
3152
3153         if (!accessor)
3154                 return -EPERM;
3155
3156         if (copy_from_user(&attr, (void __user *)arg, sizeof(attr)))
3157                 return -EFAULT;
3158
3159         return accessor(dev, &attr);
3160 }
3161
3162 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl,
3163                              unsigned long arg)
3164 {
3165         struct kvm_device *dev = filp->private_data;
3166
3167         if (dev->kvm->mm != current->mm)
3168                 return -EIO;
3169
3170         switch (ioctl) {
3171         case KVM_SET_DEVICE_ATTR:
3172                 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg);
3173         case KVM_GET_DEVICE_ATTR:
3174                 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg);
3175         case KVM_HAS_DEVICE_ATTR:
3176                 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg);
3177         default:
3178                 if (dev->ops->ioctl)
3179                         return dev->ops->ioctl(dev, ioctl, arg);
3180
3181                 return -ENOTTY;
3182         }
3183 }
3184
3185 static int kvm_device_release(struct inode *inode, struct file *filp)
3186 {
3187         struct kvm_device *dev = filp->private_data;
3188         struct kvm *kvm = dev->kvm;
3189
3190         if (dev->ops->release) {
3191                 mutex_lock(&kvm->lock);
3192                 list_del(&dev->vm_node);
3193                 dev->ops->release(dev);
3194                 mutex_unlock(&kvm->lock);
3195         }
3196
3197         kvm_put_kvm(kvm);
3198         return 0;
3199 }
3200
3201 static const struct file_operations kvm_device_fops = {
3202         .unlocked_ioctl = kvm_device_ioctl,
3203         .release = kvm_device_release,
3204         KVM_COMPAT(kvm_device_ioctl),
3205         .mmap = kvm_device_mmap,
3206 };
3207
3208 struct kvm_device *kvm_device_from_filp(struct file *filp)
3209 {
3210         if (filp->f_op != &kvm_device_fops)
3211                 return NULL;
3212
3213         return filp->private_data;
3214 }
3215
3216 static const struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = {
3217 #ifdef CONFIG_KVM_MPIC
3218         [KVM_DEV_TYPE_FSL_MPIC_20]      = &kvm_mpic_ops,
3219         [KVM_DEV_TYPE_FSL_MPIC_42]      = &kvm_mpic_ops,
3220 #endif
3221 };
3222
3223 int kvm_register_device_ops(const struct kvm_device_ops *ops, u32 type)
3224 {
3225         if (type >= ARRAY_SIZE(kvm_device_ops_table))
3226                 return -ENOSPC;
3227
3228         if (kvm_device_ops_table[type] != NULL)
3229                 return -EEXIST;
3230
3231         kvm_device_ops_table[type] = ops;
3232         return 0;
3233 }
3234
3235 void kvm_unregister_device_ops(u32 type)
3236 {
3237         if (kvm_device_ops_table[type] != NULL)
3238                 kvm_device_ops_table[type] = NULL;
3239 }
3240
3241 static int kvm_ioctl_create_device(struct kvm *kvm,
3242                                    struct kvm_create_device *cd)
3243 {
3244         const struct kvm_device_ops *ops = NULL;
3245         struct kvm_device *dev;
3246         bool test = cd->flags & KVM_CREATE_DEVICE_TEST;
3247         int type;
3248         int ret;
3249
3250         if (cd->type >= ARRAY_SIZE(kvm_device_ops_table))
3251                 return -ENODEV;
3252
3253         type = array_index_nospec(cd->type, ARRAY_SIZE(kvm_device_ops_table));
3254         ops = kvm_device_ops_table[type];
3255         if (ops == NULL)
3256                 return -ENODEV;
3257
3258         if (test)
3259                 return 0;
3260
3261         dev = kzalloc(sizeof(*dev), GFP_KERNEL_ACCOUNT);
3262         if (!dev)
3263                 return -ENOMEM;
3264
3265         dev->ops = ops;
3266         dev->kvm = kvm;
3267
3268         mutex_lock(&kvm->lock);
3269         ret = ops->create(dev, type);
3270         if (ret < 0) {
3271                 mutex_unlock(&kvm->lock);
3272                 kfree(dev);
3273                 return ret;
3274         }
3275         list_add(&dev->vm_node, &kvm->devices);
3276         mutex_unlock(&kvm->lock);
3277
3278         if (ops->init)
3279                 ops->init(dev);
3280
3281         kvm_get_kvm(kvm);
3282         ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC);
3283         if (ret < 0) {
3284                 kvm_put_kvm_no_destroy(kvm);
3285                 mutex_lock(&kvm->lock);
3286                 list_del(&dev->vm_node);
3287                 mutex_unlock(&kvm->lock);
3288                 ops->destroy(dev);
3289                 return ret;
3290         }
3291
3292         cd->fd = ret;
3293         return 0;
3294 }
3295
3296 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
3297 {
3298         switch (arg) {
3299         case KVM_CAP_USER_MEMORY:
3300         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
3301         case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
3302         case KVM_CAP_INTERNAL_ERROR_DATA:
3303 #ifdef CONFIG_HAVE_KVM_MSI
3304         case KVM_CAP_SIGNAL_MSI:
3305 #endif
3306 #ifdef CONFIG_HAVE_KVM_IRQFD
3307         case KVM_CAP_IRQFD:
3308         case KVM_CAP_IRQFD_RESAMPLE:
3309 #endif
3310         case KVM_CAP_IOEVENTFD_ANY_LENGTH:
3311         case KVM_CAP_CHECK_EXTENSION_VM:
3312         case KVM_CAP_ENABLE_CAP_VM:
3313 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
3314         case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
3315 #endif
3316                 return 1;
3317 #ifdef CONFIG_KVM_MMIO
3318         case KVM_CAP_COALESCED_MMIO:
3319                 return KVM_COALESCED_MMIO_PAGE_OFFSET;
3320         case KVM_CAP_COALESCED_PIO:
3321                 return 1;
3322 #endif
3323 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
3324         case KVM_CAP_IRQ_ROUTING:
3325                 return KVM_MAX_IRQ_ROUTES;
3326 #endif
3327 #if KVM_ADDRESS_SPACE_NUM > 1
3328         case KVM_CAP_MULTI_ADDRESS_SPACE:
3329                 return KVM_ADDRESS_SPACE_NUM;
3330 #endif
3331         case KVM_CAP_NR_MEMSLOTS:
3332                 return KVM_USER_MEM_SLOTS;
3333         default:
3334                 break;
3335         }
3336         return kvm_vm_ioctl_check_extension(kvm, arg);
3337 }
3338
3339 int __attribute__((weak)) kvm_vm_ioctl_enable_cap(struct kvm *kvm,
3340                                                   struct kvm_enable_cap *cap)
3341 {
3342         return -EINVAL;
3343 }
3344
3345 static int kvm_vm_ioctl_enable_cap_generic(struct kvm *kvm,
3346                                            struct kvm_enable_cap *cap)
3347 {
3348         switch (cap->cap) {
3349 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
3350         case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
3351                 if (cap->flags || (cap->args[0] & ~1))
3352                         return -EINVAL;
3353                 kvm->manual_dirty_log_protect = cap->args[0];
3354                 return 0;
3355 #endif
3356         default:
3357                 return kvm_vm_ioctl_enable_cap(kvm, cap);
3358         }
3359 }
3360
3361 static long kvm_vm_ioctl(struct file *filp,
3362                            unsigned int ioctl, unsigned long arg)
3363 {
3364         struct kvm *kvm = filp->private_data;
3365         void __user *argp = (void __user *)arg;
3366         int r;
3367
3368         if (kvm->mm != current->mm)
3369                 return -EIO;
3370         switch (ioctl) {
3371         case KVM_CREATE_VCPU:
3372                 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
3373                 break;
3374         case KVM_ENABLE_CAP: {
3375                 struct kvm_enable_cap cap;
3376
3377                 r = -EFAULT;
3378                 if (copy_from_user(&cap, argp, sizeof(cap)))
3379                         goto out;
3380                 r = kvm_vm_ioctl_enable_cap_generic(kvm, &cap);
3381                 break;
3382         }
3383         case KVM_SET_USER_MEMORY_REGION: {
3384                 struct kvm_userspace_memory_region kvm_userspace_mem;
3385
3386                 r = -EFAULT;
3387                 if (copy_from_user(&kvm_userspace_mem, argp,
3388                                                 sizeof(kvm_userspace_mem)))
3389                         goto out;
3390
3391                 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem);
3392                 break;
3393         }
3394         case KVM_GET_DIRTY_LOG: {
3395                 struct kvm_dirty_log log;
3396
3397                 r = -EFAULT;
3398                 if (copy_from_user(&log, argp, sizeof(log)))
3399                         goto out;
3400                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
3401                 break;
3402         }
3403 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
3404         case KVM_CLEAR_DIRTY_LOG: {
3405                 struct kvm_clear_dirty_log log;
3406
3407                 r = -EFAULT;
3408                 if (copy_from_user(&log, argp, sizeof(log)))
3409                         goto out;
3410                 r = kvm_vm_ioctl_clear_dirty_log(kvm, &log);
3411                 break;
3412         }
3413 #endif
3414 #ifdef CONFIG_KVM_MMIO
3415         case KVM_REGISTER_COALESCED_MMIO: {
3416                 struct kvm_coalesced_mmio_zone zone;
3417
3418                 r = -EFAULT;
3419                 if (copy_from_user(&zone, argp, sizeof(zone)))
3420                         goto out;
3421                 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
3422                 break;
3423         }
3424         case KVM_UNREGISTER_COALESCED_MMIO: {
3425                 struct kvm_coalesced_mmio_zone zone;
3426
3427                 r = -EFAULT;
3428                 if (copy_from_user(&zone, argp, sizeof(zone)))
3429                         goto out;
3430                 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
3431                 break;
3432         }
3433 #endif
3434         case KVM_IRQFD: {
3435                 struct kvm_irqfd data;
3436
3437                 r = -EFAULT;
3438                 if (copy_from_user(&data, argp, sizeof(data)))
3439                         goto out;
3440                 r = kvm_irqfd(kvm, &data);
3441                 break;
3442         }
3443         case KVM_IOEVENTFD: {
3444                 struct kvm_ioeventfd data;
3445
3446                 r = -EFAULT;
3447                 if (copy_from_user(&data, argp, sizeof(data)))
3448                         goto out;
3449                 r = kvm_ioeventfd(kvm, &data);
3450                 break;
3451         }
3452 #ifdef CONFIG_HAVE_KVM_MSI
3453         case KVM_SIGNAL_MSI: {
3454                 struct kvm_msi msi;
3455
3456                 r = -EFAULT;
3457                 if (copy_from_user(&msi, argp, sizeof(msi)))
3458                         goto out;
3459                 r = kvm_send_userspace_msi(kvm, &msi);
3460                 break;
3461         }
3462 #endif
3463 #ifdef __KVM_HAVE_IRQ_LINE
3464         case KVM_IRQ_LINE_STATUS:
3465         case KVM_IRQ_LINE: {
3466                 struct kvm_irq_level irq_event;
3467
3468                 r = -EFAULT;
3469                 if (copy_from_user(&irq_event, argp, sizeof(irq_event)))
3470                         goto out;
3471
3472                 r = kvm_vm_ioctl_irq_line(kvm, &irq_event,
3473                                         ioctl == KVM_IRQ_LINE_STATUS);
3474                 if (r)
3475                         goto out;
3476
3477                 r = -EFAULT;
3478                 if (ioctl == KVM_IRQ_LINE_STATUS) {
3479                         if (copy_to_user(argp, &irq_event, sizeof(irq_event)))
3480                                 goto out;
3481                 }
3482
3483                 r = 0;
3484                 break;
3485         }
3486 #endif
3487 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
3488         case KVM_SET_GSI_ROUTING: {
3489                 struct kvm_irq_routing routing;
3490                 struct kvm_irq_routing __user *urouting;
3491                 struct kvm_irq_routing_entry *entries = NULL;
3492
3493                 r = -EFAULT;
3494                 if (copy_from_user(&routing, argp, sizeof(routing)))
3495                         goto out;
3496                 r = -EINVAL;
3497                 if (!kvm_arch_can_set_irq_routing(kvm))
3498                         goto out;
3499                 if (routing.nr > KVM_MAX_IRQ_ROUTES)
3500                         goto out;
3501                 if (routing.flags)
3502                         goto out;
3503                 if (routing.nr) {
3504                         r = -ENOMEM;
3505                         entries = vmalloc(array_size(sizeof(*entries),
3506                                                      routing.nr));
3507                         if (!entries)
3508                                 goto out;
3509                         r = -EFAULT;
3510                         urouting = argp;
3511                         if (copy_from_user(entries, urouting->entries,
3512                                            routing.nr * sizeof(*entries)))
3513                                 goto out_free_irq_routing;
3514                 }
3515                 r = kvm_set_irq_routing(kvm, entries, routing.nr,
3516                                         routing.flags);
3517 out_free_irq_routing:
3518                 vfree(entries);
3519                 break;
3520         }
3521 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */
3522         case KVM_CREATE_DEVICE: {
3523                 struct kvm_create_device cd;
3524
3525                 r = -EFAULT;
3526                 if (copy_from_user(&cd, argp, sizeof(cd)))
3527                         goto out;
3528
3529                 r = kvm_ioctl_create_device(kvm, &cd);
3530                 if (r)
3531                         goto out;
3532
3533                 r = -EFAULT;
3534                 if (copy_to_user(argp, &cd, sizeof(cd)))
3535                         goto out;
3536
3537                 r = 0;
3538                 break;
3539         }
3540         case KVM_CHECK_EXTENSION:
3541                 r = kvm_vm_ioctl_check_extension_generic(kvm, arg);
3542                 break;
3543         default:
3544                 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
3545         }
3546 out:
3547         return r;
3548 }
3549
3550 #ifdef CONFIG_KVM_COMPAT
3551 struct compat_kvm_dirty_log {
3552         __u32 slot;
3553         __u32 padding1;
3554         union {
3555                 compat_uptr_t dirty_bitmap; /* one bit per page */
3556                 __u64 padding2;
3557         };
3558 };
3559
3560 static long kvm_vm_compat_ioctl(struct file *filp,
3561                            unsigned int ioctl, unsigned long arg)
3562 {
3563         struct kvm *kvm = filp->private_data;
3564         int r;
3565
3566         if (kvm->mm != current->mm)
3567                 return -EIO;
3568         switch (ioctl) {
3569         case KVM_GET_DIRTY_LOG: {
3570                 struct compat_kvm_dirty_log compat_log;
3571                 struct kvm_dirty_log log;
3572
3573                 if (copy_from_user(&compat_log, (void __user *)arg,
3574                                    sizeof(compat_log)))
3575                         return -EFAULT;
3576                 log.slot         = compat_log.slot;
3577                 log.padding1     = compat_log.padding1;
3578                 log.padding2     = compat_log.padding2;
3579                 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
3580
3581                 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
3582                 break;
3583         }
3584         default:
3585                 r = kvm_vm_ioctl(filp, ioctl, arg);
3586         }
3587         return r;
3588 }
3589 #endif
3590
3591 static struct file_operations kvm_vm_fops = {
3592         .release        = kvm_vm_release,
3593         .unlocked_ioctl = kvm_vm_ioctl,
3594         .llseek         = noop_llseek,
3595         KVM_COMPAT(kvm_vm_compat_ioctl),
3596 };
3597
3598 static int kvm_dev_ioctl_create_vm(unsigned long type)
3599 {
3600         int r;
3601         struct kvm *kvm;
3602         struct file *file;
3603
3604         kvm = kvm_create_vm(type);
3605         if (IS_ERR(kvm))
3606                 return PTR_ERR(kvm);
3607 #ifdef CONFIG_KVM_MMIO
3608         r = kvm_coalesced_mmio_init(kvm);
3609         if (r < 0)
3610                 goto put_kvm;
3611 #endif
3612         r = get_unused_fd_flags(O_CLOEXEC);
3613         if (r < 0)
3614                 goto put_kvm;
3615
3616         file = anon_inode_getfile("kvm-vm", &kvm_vm_fops, kvm, O_RDWR);
3617         if (IS_ERR(file)) {
3618                 put_unused_fd(r);
3619                 r = PTR_ERR(file);
3620                 goto put_kvm;
3621         }
3622
3623         /*
3624          * Don't call kvm_put_kvm anymore at this point; file->f_op is
3625          * already set, with ->release() being kvm_vm_release().  In error
3626          * cases it will be called by the final fput(file) and will take
3627          * care of doing kvm_put_kvm(kvm).
3628          */
3629         if (kvm_create_vm_debugfs(kvm, r) < 0) {
3630                 put_unused_fd(r);
3631                 fput(file);
3632                 return -ENOMEM;
3633         }
3634         kvm_uevent_notify_change(KVM_EVENT_CREATE_VM, kvm);
3635
3636         fd_install(r, file);
3637         return r;
3638
3639 put_kvm:
3640         kvm_put_kvm(kvm);
3641         return r;
3642 }
3643
3644 static long kvm_dev_ioctl(struct file *filp,
3645                           unsigned int ioctl, unsigned long arg)
3646 {
3647         long r = -EINVAL;
3648
3649         switch (ioctl) {
3650         case KVM_GET_API_VERSION:
3651                 if (arg)
3652                         goto out;
3653                 r = KVM_API_VERSION;
3654                 break;
3655         case KVM_CREATE_VM:
3656                 r = kvm_dev_ioctl_create_vm(arg);
3657                 break;
3658         case KVM_CHECK_EXTENSION:
3659                 r = kvm_vm_ioctl_check_extension_generic(NULL, arg);
3660                 break;
3661         case KVM_GET_VCPU_MMAP_SIZE:
3662                 if (arg)
3663                         goto out;
3664                 r = PAGE_SIZE;     /* struct kvm_run */
3665 #ifdef CONFIG_X86
3666                 r += PAGE_SIZE;    /* pio data page */
3667 #endif
3668 #ifdef CONFIG_KVM_MMIO
3669                 r += PAGE_SIZE;    /* coalesced mmio ring page */
3670 #endif
3671                 break;
3672         case KVM_TRACE_ENABLE:
3673         case KVM_TRACE_PAUSE:
3674         case KVM_TRACE_DISABLE:
3675                 r = -EOPNOTSUPP;
3676                 break;
3677         default:
3678                 return kvm_arch_dev_ioctl(filp, ioctl, arg);
3679         }
3680 out:
3681         return r;
3682 }
3683
3684 static struct file_operations kvm_chardev_ops = {
3685         .unlocked_ioctl = kvm_dev_ioctl,
3686         .llseek         = noop_llseek,
3687         KVM_COMPAT(kvm_dev_ioctl),
3688 };
3689
3690 static struct miscdevice kvm_dev = {
3691         KVM_MINOR,
3692         "kvm",
3693         &kvm_chardev_ops,
3694 };
3695
3696 static void hardware_enable_nolock(void *junk)
3697 {
3698         int cpu = raw_smp_processor_id();
3699         int r;
3700
3701         if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
3702                 return;
3703
3704         cpumask_set_cpu(cpu, cpus_hardware_enabled);
3705
3706         r = kvm_arch_hardware_enable();
3707
3708         if (r) {
3709                 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3710                 atomic_inc(&hardware_enable_failed);
3711                 pr_info("kvm: enabling virtualization on CPU%d failed\n", cpu);
3712         }
3713 }
3714
3715 static int kvm_starting_cpu(unsigned int cpu)
3716 {
3717         raw_spin_lock(&kvm_count_lock);
3718         if (kvm_usage_count)
3719                 hardware_enable_nolock(NULL);
3720         raw_spin_unlock(&kvm_count_lock);
3721         return 0;
3722 }
3723
3724 static void hardware_disable_nolock(void *junk)
3725 {
3726         int cpu = raw_smp_processor_id();
3727
3728         if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
3729                 return;
3730         cpumask_clear_cpu(cpu, cpus_hardware_enabled);
3731         kvm_arch_hardware_disable();
3732 }
3733
3734 static int kvm_dying_cpu(unsigned int cpu)
3735 {
3736         raw_spin_lock(&kvm_count_lock);
3737         if (kvm_usage_count)
3738                 hardware_disable_nolock(NULL);
3739         raw_spin_unlock(&kvm_count_lock);
3740         return 0;
3741 }
3742
3743 static void hardware_disable_all_nolock(void)
3744 {
3745         BUG_ON(!kvm_usage_count);
3746
3747         kvm_usage_count--;
3748         if (!kvm_usage_count)
3749                 on_each_cpu(hardware_disable_nolock, NULL, 1);
3750 }
3751
3752 static void hardware_disable_all(void)
3753 {
3754         raw_spin_lock(&kvm_count_lock);
3755         hardware_disable_all_nolock();
3756         raw_spin_unlock(&kvm_count_lock);
3757 }
3758
3759 static int hardware_enable_all(void)
3760 {
3761         int r = 0;
3762
3763         raw_spin_lock(&kvm_count_lock);
3764
3765         kvm_usage_count++;
3766         if (kvm_usage_count == 1) {
3767                 atomic_set(&hardware_enable_failed, 0);
3768                 on_each_cpu(hardware_enable_nolock, NULL, 1);
3769
3770                 if (atomic_read(&hardware_enable_failed)) {
3771                         hardware_disable_all_nolock();
3772                         r = -EBUSY;
3773                 }
3774         }
3775
3776         raw_spin_unlock(&kvm_count_lock);
3777
3778         return r;
3779 }
3780
3781 static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
3782                       void *v)
3783 {
3784         /*
3785          * Some (well, at least mine) BIOSes hang on reboot if
3786          * in vmx root mode.
3787          *
3788          * And Intel TXT required VMX off for all cpu when system shutdown.
3789          */
3790         pr_info("kvm: exiting hardware virtualization\n");
3791         kvm_rebooting = true;
3792         on_each_cpu(hardware_disable_nolock, NULL, 1);
3793         return NOTIFY_OK;
3794 }
3795
3796 static struct notifier_block kvm_reboot_notifier = {
3797         .notifier_call = kvm_reboot,
3798         .priority = 0,
3799 };
3800
3801 static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
3802 {
3803         int i;
3804
3805         for (i = 0; i < bus->dev_count; i++) {
3806                 struct kvm_io_device *pos = bus->range[i].dev;
3807
3808                 kvm_iodevice_destructor(pos);
3809         }
3810         kfree(bus);
3811 }
3812
3813 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1,
3814                                  const struct kvm_io_range *r2)
3815 {
3816         gpa_t addr1 = r1->addr;
3817         gpa_t addr2 = r2->addr;
3818
3819         if (addr1 < addr2)
3820                 return -1;
3821
3822         /* If r2->len == 0, match the exact address.  If r2->len != 0,
3823          * accept any overlapping write.  Any order is acceptable for
3824          * overlapping ranges, because kvm_io_bus_get_first_dev ensures
3825          * we process all of them.
3826          */
3827         if (r2->len) {
3828                 addr1 += r1->len;
3829                 addr2 += r2->len;
3830         }
3831
3832         if (addr1 > addr2)
3833                 return 1;
3834
3835         return 0;
3836 }
3837
3838 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
3839 {
3840         return kvm_io_bus_cmp(p1, p2);
3841 }
3842
3843 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
3844                              gpa_t addr, int len)
3845 {
3846         struct kvm_io_range *range, key;
3847         int off;
3848
3849         key = (struct kvm_io_range) {
3850                 .addr = addr,
3851                 .len = len,
3852         };
3853
3854         range = bsearch(&key, bus->range, bus->dev_count,
3855                         sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
3856         if (range == NULL)
3857                 return -ENOENT;
3858
3859         off = range - bus->range;
3860
3861         while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0)
3862                 off--;
3863
3864         return off;
3865 }
3866
3867 static int __kvm_io_bus_write(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3868                               struct kvm_io_range *range, const void *val)
3869 {
3870         int idx;
3871
3872         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3873         if (idx < 0)
3874                 return -EOPNOTSUPP;
3875
3876         while (idx < bus->dev_count &&
3877                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3878                 if (!kvm_iodevice_write(vcpu, bus->range[idx].dev, range->addr,
3879                                         range->len, val))
3880                         return idx;
3881                 idx++;
3882         }
3883
3884         return -EOPNOTSUPP;
3885 }
3886
3887 /* kvm_io_bus_write - called under kvm->slots_lock */
3888 int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3889                      int len, const void *val)
3890 {
3891         struct kvm_io_bus *bus;
3892         struct kvm_io_range range;
3893         int r;
3894
3895         range = (struct kvm_io_range) {
3896                 .addr = addr,
3897                 .len = len,
3898         };
3899
3900         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3901         if (!bus)
3902                 return -ENOMEM;
3903         r = __kvm_io_bus_write(vcpu, bus, &range, val);
3904         return r < 0 ? r : 0;
3905 }
3906 EXPORT_SYMBOL_GPL(kvm_io_bus_write);
3907
3908 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */
3909 int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
3910                             gpa_t addr, int len, const void *val, long cookie)
3911 {
3912         struct kvm_io_bus *bus;
3913         struct kvm_io_range range;
3914
3915         range = (struct kvm_io_range) {
3916                 .addr = addr,
3917                 .len = len,
3918         };
3919
3920         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3921         if (!bus)
3922                 return -ENOMEM;
3923
3924         /* First try the device referenced by cookie. */
3925         if ((cookie >= 0) && (cookie < bus->dev_count) &&
3926             (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0))
3927                 if (!kvm_iodevice_write(vcpu, bus->range[cookie].dev, addr, len,
3928                                         val))
3929                         return cookie;
3930
3931         /*
3932          * cookie contained garbage; fall back to search and return the
3933          * correct cookie value.
3934          */
3935         return __kvm_io_bus_write(vcpu, bus, &range, val);
3936 }
3937
3938 static int __kvm_io_bus_read(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus,
3939                              struct kvm_io_range *range, void *val)
3940 {
3941         int idx;
3942
3943         idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len);
3944         if (idx < 0)
3945                 return -EOPNOTSUPP;
3946
3947         while (idx < bus->dev_count &&
3948                 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) {
3949                 if (!kvm_iodevice_read(vcpu, bus->range[idx].dev, range->addr,
3950                                        range->len, val))
3951                         return idx;
3952                 idx++;
3953         }
3954
3955         return -EOPNOTSUPP;
3956 }
3957
3958 /* kvm_io_bus_read - called under kvm->slots_lock */
3959 int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
3960                     int len, void *val)
3961 {
3962         struct kvm_io_bus *bus;
3963         struct kvm_io_range range;
3964         int r;
3965
3966         range = (struct kvm_io_range) {
3967                 .addr = addr,
3968                 .len = len,
3969         };
3970
3971         bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
3972         if (!bus)
3973                 return -ENOMEM;
3974         r = __kvm_io_bus_read(vcpu, bus, &range, val);
3975         return r < 0 ? r : 0;
3976 }
3977
3978 /* Caller must hold slots_lock. */
3979 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
3980                             int len, struct kvm_io_device *dev)
3981 {
3982         int i;
3983         struct kvm_io_bus *new_bus, *bus;
3984         struct kvm_io_range range;
3985
3986         bus = kvm_get_bus(kvm, bus_idx);
3987         if (!bus)
3988                 return -ENOMEM;
3989
3990         /* exclude ioeventfd which is limited by maximum fd */
3991         if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
3992                 return -ENOSPC;
3993
3994         new_bus = kmalloc(struct_size(bus, range, bus->dev_count + 1),
3995                           GFP_KERNEL_ACCOUNT);
3996         if (!new_bus)
3997                 return -ENOMEM;
3998
3999         range = (struct kvm_io_range) {
4000                 .addr = addr,
4001                 .len = len,
4002                 .dev = dev,
4003         };
4004
4005         for (i = 0; i < bus->dev_count; i++)
4006                 if (kvm_io_bus_cmp(&bus->range[i], &range) > 0)
4007                         break;
4008
4009         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
4010         new_bus->dev_count++;
4011         new_bus->range[i] = range;
4012         memcpy(new_bus->range + i + 1, bus->range + i,
4013                 (bus->dev_count - i) * sizeof(struct kvm_io_range));
4014         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
4015         synchronize_srcu_expedited(&kvm->srcu);
4016         kfree(bus);
4017
4018         return 0;
4019 }
4020
4021 /* Caller must hold slots_lock. */
4022 void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
4023                                struct kvm_io_device *dev)
4024 {
4025         int i;
4026         struct kvm_io_bus *new_bus, *bus;
4027
4028         bus = kvm_get_bus(kvm, bus_idx);
4029         if (!bus)
4030                 return;
4031
4032         for (i = 0; i < bus->dev_count; i++)
4033                 if (bus->range[i].dev == dev) {
4034                         break;
4035                 }
4036
4037         if (i == bus->dev_count)
4038                 return;
4039
4040         new_bus = kmalloc(struct_size(bus, range, bus->dev_count - 1),
4041                           GFP_KERNEL_ACCOUNT);
4042         if (!new_bus)  {
4043                 pr_err("kvm: failed to shrink bus, removing it completely\n");
4044                 goto broken;
4045         }
4046
4047         memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range));
4048         new_bus->dev_count--;
4049         memcpy(new_bus->range + i, bus->range + i + 1,
4050                (new_bus->dev_count - i) * sizeof(struct kvm_io_range));
4051
4052 broken:
4053         rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
4054         synchronize_srcu_expedited(&kvm->srcu);
4055         kfree(bus);
4056         return;
4057 }
4058
4059 struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx,
4060                                          gpa_t addr)
4061 {
4062         struct kvm_io_bus *bus;
4063         int dev_idx, srcu_idx;
4064         struct kvm_io_device *iodev = NULL;
4065
4066         srcu_idx = srcu_read_lock(&kvm->srcu);
4067
4068         bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
4069         if (!bus)
4070                 goto out_unlock;
4071
4072         dev_idx = kvm_io_bus_get_first_dev(bus, addr, 1);
4073         if (dev_idx < 0)
4074                 goto out_unlock;
4075
4076         iodev = bus->range[dev_idx].dev;
4077
4078 out_unlock:
4079         srcu_read_unlock(&kvm->srcu, srcu_idx);
4080
4081         return iodev;
4082 }
4083 EXPORT_SYMBOL_GPL(kvm_io_bus_get_dev);
4084
4085 static int kvm_debugfs_open(struct inode *inode, struct file *file,
4086                            int (*get)(void *, u64 *), int (*set)(void *, u64),
4087                            const char *fmt)
4088 {
4089         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)
4090                                           inode->i_private;
4091
4092         /* The debugfs files are a reference to the kvm struct which
4093          * is still valid when kvm_destroy_vm is called.
4094          * To avoid the race between open and the removal of the debugfs
4095          * directory we test against the users count.
4096          */
4097         if (!refcount_inc_not_zero(&stat_data->kvm->users_count))
4098                 return -ENOENT;
4099
4100         if (simple_attr_open(inode, file, get,
4101                     KVM_DBGFS_GET_MODE(stat_data->dbgfs_item) & 0222
4102                     ? set : NULL,
4103                     fmt)) {
4104                 kvm_put_kvm(stat_data->kvm);
4105                 return -ENOMEM;
4106         }
4107
4108         return 0;
4109 }
4110
4111 static int kvm_debugfs_release(struct inode *inode, struct file *file)
4112 {
4113         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)
4114                                           inode->i_private;
4115
4116         simple_attr_release(inode, file);
4117         kvm_put_kvm(stat_data->kvm);
4118
4119         return 0;
4120 }
4121
4122 static int kvm_get_stat_per_vm(struct kvm *kvm, size_t offset, u64 *val)
4123 {
4124         *val = *(ulong *)((void *)kvm + offset);
4125
4126         return 0;
4127 }
4128
4129 static int kvm_clear_stat_per_vm(struct kvm *kvm, size_t offset)
4130 {
4131         *(ulong *)((void *)kvm + offset) = 0;
4132
4133         return 0;
4134 }
4135
4136 static int kvm_get_stat_per_vcpu(struct kvm *kvm, size_t offset, u64 *val)
4137 {
4138         int i;
4139         struct kvm_vcpu *vcpu;
4140
4141         *val = 0;
4142
4143         kvm_for_each_vcpu(i, vcpu, kvm)
4144                 *val += *(u64 *)((void *)vcpu + offset);
4145
4146         return 0;
4147 }
4148
4149 static int kvm_clear_stat_per_vcpu(struct kvm *kvm, size_t offset)
4150 {
4151         int i;
4152         struct kvm_vcpu *vcpu;
4153
4154         kvm_for_each_vcpu(i, vcpu, kvm)
4155                 *(u64 *)((void *)vcpu + offset) = 0;
4156
4157         return 0;
4158 }
4159
4160 static int kvm_stat_data_get(void *data, u64 *val)
4161 {
4162         int r = -EFAULT;
4163         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data;
4164
4165         switch (stat_data->dbgfs_item->kind) {
4166         case KVM_STAT_VM:
4167                 r = kvm_get_stat_per_vm(stat_data->kvm,
4168                                         stat_data->dbgfs_item->offset, val);
4169                 break;
4170         case KVM_STAT_VCPU:
4171                 r = kvm_get_stat_per_vcpu(stat_data->kvm,
4172                                           stat_data->dbgfs_item->offset, val);
4173                 break;
4174         }
4175
4176         return r;
4177 }
4178
4179 static int kvm_stat_data_clear(void *data, u64 val)
4180 {
4181         int r = -EFAULT;
4182         struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data;
4183
4184         if (val)
4185                 return -EINVAL;
4186
4187         switch (stat_data->dbgfs_item->kind) {
4188         case KVM_STAT_VM:
4189                 r = kvm_clear_stat_per_vm(stat_data->kvm,
4190                                           stat_data->dbgfs_item->offset);
4191                 break;
4192         case KVM_STAT_VCPU:
4193                 r = kvm_clear_stat_per_vcpu(stat_data->kvm,
4194                                             stat_data->dbgfs_item->offset);
4195                 break;
4196         }
4197
4198         return r;
4199 }
4200
4201 static int kvm_stat_data_open(struct inode *inode, struct file *file)
4202 {
4203         __simple_attr_check_format("%llu\n", 0ull);
4204         return kvm_debugfs_open(inode, file, kvm_stat_data_get,
4205                                 kvm_stat_data_clear, "%llu\n");
4206 }
4207
4208 static const struct file_operations stat_fops_per_vm = {
4209         .owner = THIS_MODULE,
4210         .open = kvm_stat_data_open,
4211         .release = kvm_debugfs_release,
4212         .read = simple_attr_read,
4213         .write = simple_attr_write,
4214         .llseek = no_llseek,
4215 };
4216
4217 static int vm_stat_get(void *_offset, u64 *val)
4218 {
4219         unsigned offset = (long)_offset;
4220         struct kvm *kvm;
4221         u64 tmp_val;
4222
4223         *val = 0;
4224         mutex_lock(&kvm_lock);
4225         list_for_each_entry(kvm, &vm_list, vm_list) {
4226                 kvm_get_stat_per_vm(kvm, offset, &tmp_val);
4227                 *val += tmp_val;
4228         }
4229         mutex_unlock(&kvm_lock);
4230         return 0;
4231 }
4232
4233 static int vm_stat_clear(void *_offset, u64 val)
4234 {
4235         unsigned offset = (long)_offset;
4236         struct kvm *kvm;
4237
4238         if (val)
4239                 return -EINVAL;
4240
4241         mutex_lock(&kvm_lock);
4242         list_for_each_entry(kvm, &vm_list, vm_list) {
4243                 kvm_clear_stat_per_vm(kvm, offset);
4244         }
4245         mutex_unlock(&kvm_lock);
4246
4247         return 0;
4248 }
4249
4250 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, vm_stat_clear, "%llu\n");
4251
4252 static int vcpu_stat_get(void *_offset, u64 *val)
4253 {
4254         unsigned offset = (long)_offset;
4255         struct kvm *kvm;
4256         u64 tmp_val;
4257
4258         *val = 0;
4259         mutex_lock(&kvm_lock);
4260         list_for_each_entry(kvm, &vm_list, vm_list) {
4261                 kvm_get_stat_per_vcpu(kvm, offset, &tmp_val);
4262                 *val += tmp_val;
4263         }
4264         mutex_unlock(&kvm_lock);
4265         return 0;
4266 }
4267
4268 static int vcpu_stat_clear(void *_offset, u64 val)
4269 {
4270         unsigned offset = (long)_offset;
4271         struct kvm *kvm;
4272
4273         if (val)
4274                 return -EINVAL;
4275
4276         mutex_lock(&kvm_lock);
4277         list_for_each_entry(kvm, &vm_list, vm_list) {
4278                 kvm_clear_stat_per_vcpu(kvm, offset);
4279         }
4280         mutex_unlock(&kvm_lock);
4281
4282         return 0;
4283 }
4284
4285 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, vcpu_stat_clear,
4286                         "%llu\n");
4287
4288 static const struct file_operations *stat_fops[] = {
4289         [KVM_STAT_VCPU] = &vcpu_stat_fops,
4290         [KVM_STAT_VM]   = &vm_stat_fops,
4291 };
4292
4293 static void kvm_uevent_notify_change(unsigned int type, struct kvm *kvm)
4294 {
4295         struct kobj_uevent_env *env;
4296         unsigned long long created, active;
4297
4298         if (!kvm_dev.this_device || !kvm)
4299                 return;
4300
4301         mutex_lock(&kvm_lock);
4302         if (type == KVM_EVENT_CREATE_VM) {
4303                 kvm_createvm_count++;
4304                 kvm_active_vms++;
4305         } else if (type == KVM_EVENT_DESTROY_VM) {
4306                 kvm_active_vms--;
4307         }
4308         created = kvm_createvm_count;
4309         active = kvm_active_vms;
4310         mutex_unlock(&kvm_lock);
4311
4312         env = kzalloc(sizeof(*env), GFP_KERNEL_ACCOUNT);
4313         if (!env)
4314                 return;
4315
4316         add_uevent_var(env, "CREATED=%llu", created);
4317         add_uevent_var(env, "COUNT=%llu", active);
4318
4319         if (type == KVM_EVENT_CREATE_VM) {
4320                 add_uevent_var(env, "EVENT=create");
4321                 kvm->userspace_pid = task_pid_nr(current);
4322         } else if (type == KVM_EVENT_DESTROY_VM) {
4323                 add_uevent_var(env, "EVENT=destroy");
4324         }
4325         add_uevent_var(env, "PID=%d", kvm->userspace_pid);
4326
4327         if (!IS_ERR_OR_NULL(kvm->debugfs_dentry)) {
4328                 char *tmp, *p = kmalloc(PATH_MAX, GFP_KERNEL_ACCOUNT);
4329
4330                 if (p) {
4331                         tmp = dentry_path_raw(kvm->debugfs_dentry, p, PATH_MAX);
4332                         if (!IS_ERR(tmp))
4333                                 add_uevent_var(env, "STATS_PATH=%s", tmp);
4334                         kfree(p);
4335                 }
4336         }
4337         /* no need for checks, since we are adding at most only 5 keys */
4338         env->envp[env->envp_idx++] = NULL;
4339         kobject_uevent_env(&kvm_dev.this_device->kobj, KOBJ_CHANGE, env->envp);
4340         kfree(env);
4341 }
4342
4343 static void kvm_init_debug(void)
4344 {
4345         struct kvm_stats_debugfs_item *p;
4346
4347         kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
4348
4349         kvm_debugfs_num_entries = 0;
4350         for (p = debugfs_entries; p->name; ++p, kvm_debugfs_num_entries++) {
4351                 debugfs_create_file(p->name, KVM_DBGFS_GET_MODE(p),
4352                                     kvm_debugfs_dir, (void *)(long)p->offset,
4353                                     stat_fops[p->kind]);
4354         }
4355 }
4356
4357 static int kvm_suspend(void)
4358 {
4359         if (kvm_usage_count)
4360                 hardware_disable_nolock(NULL);
4361         return 0;
4362 }
4363
4364 static void kvm_resume(void)
4365 {
4366         if (kvm_usage_count) {
4367 #ifdef CONFIG_LOCKDEP
4368                 WARN_ON(lockdep_is_held(&kvm_count_lock));
4369 #endif
4370                 hardware_enable_nolock(NULL);
4371         }
4372 }
4373
4374 static struct syscore_ops kvm_syscore_ops = {
4375         .suspend = kvm_suspend,
4376         .resume = kvm_resume,
4377 };
4378
4379 static inline
4380 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
4381 {
4382         return container_of(pn, struct kvm_vcpu, preempt_notifier);
4383 }
4384
4385 static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
4386 {
4387         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
4388
4389         WRITE_ONCE(vcpu->preempted, false);
4390         WRITE_ONCE(vcpu->ready, false);
4391
4392         __this_cpu_write(kvm_running_vcpu, vcpu);
4393         kvm_arch_sched_in(vcpu, cpu);
4394         kvm_arch_vcpu_load(vcpu, cpu);
4395 }
4396
4397 static void kvm_sched_out(struct preempt_notifier *pn,
4398                           struct task_struct *next)
4399 {
4400         struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
4401
4402         if (current->state == TASK_RUNNING) {
4403                 WRITE_ONCE(vcpu->preempted, true);
4404                 WRITE_ONCE(vcpu->ready, true);
4405         }
4406         kvm_arch_vcpu_put(vcpu);
4407         __this_cpu_write(kvm_running_vcpu, NULL);
4408 }
4409
4410 /**
4411  * kvm_get_running_vcpu - get the vcpu running on the current CPU.
4412  *
4413  * We can disable preemption locally around accessing the per-CPU variable,
4414  * and use the resolved vcpu pointer after enabling preemption again,
4415  * because even if the current thread is migrated to another CPU, reading
4416  * the per-CPU value later will give us the same value as we update the
4417  * per-CPU variable in the preempt notifier handlers.
4418  */
4419 struct kvm_vcpu *kvm_get_running_vcpu(void)
4420 {
4421         struct kvm_vcpu *vcpu;
4422
4423         preempt_disable();
4424         vcpu = __this_cpu_read(kvm_running_vcpu);
4425         preempt_enable();
4426
4427         return vcpu;
4428 }
4429
4430 /**
4431  * kvm_get_running_vcpus - get the per-CPU array of currently running vcpus.
4432  */
4433 struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void)
4434 {
4435         return &kvm_running_vcpu;
4436 }
4437
4438 static void check_processor_compat(void *rtn)
4439 {
4440         *(int *)rtn = kvm_arch_check_processor_compat();
4441 }
4442
4443 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
4444                   struct module *module)
4445 {
4446         int r;
4447         int cpu;
4448
4449         r = kvm_arch_init(opaque);
4450         if (r)
4451                 goto out_fail;
4452
4453         /*
4454          * kvm_arch_init makes sure there's at most one caller
4455          * for architectures that support multiple implementations,
4456          * like intel and amd on x86.
4457          * kvm_arch_init must be called before kvm_irqfd_init to avoid creating
4458          * conflicts in case kvm is already setup for another implementation.
4459          */
4460         r = kvm_irqfd_init();
4461         if (r)
4462                 goto out_irqfd;
4463
4464         if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
4465                 r = -ENOMEM;
4466                 goto out_free_0;
4467         }
4468
4469         r = kvm_arch_hardware_setup();
4470         if (r < 0)
4471                 goto out_free_1;
4472
4473         for_each_online_cpu(cpu) {
4474                 smp_call_function_single(cpu, check_processor_compat, &r, 1);
4475                 if (r < 0)
4476                         goto out_free_2;
4477         }
4478
4479         r = cpuhp_setup_state_nocalls(CPUHP_AP_KVM_STARTING, "kvm/cpu:starting",
4480                                       kvm_starting_cpu, kvm_dying_cpu);
4481         if (r)
4482                 goto out_free_2;
4483         register_reboot_notifier(&kvm_reboot_notifier);
4484
4485         /* A kmem cache lets us meet the alignment requirements of fx_save. */
4486         if (!vcpu_align)
4487                 vcpu_align = __alignof__(struct kvm_vcpu);
4488         kvm_vcpu_cache =
4489                 kmem_cache_create_usercopy("kvm_vcpu", vcpu_size, vcpu_align,
4490                                            SLAB_ACCOUNT,
4491                                            offsetof(struct kvm_vcpu, arch),
4492                                            sizeof_field(struct kvm_vcpu, arch),
4493                                            NULL);
4494         if (!kvm_vcpu_cache) {
4495                 r = -ENOMEM;
4496                 goto out_free_3;
4497         }
4498
4499         r = kvm_async_pf_init();
4500         if (r)
4501                 goto out_free;
4502
4503         kvm_chardev_ops.owner = module;
4504         kvm_vm_fops.owner = module;
4505         kvm_vcpu_fops.owner = module;
4506
4507         r = misc_register(&kvm_dev);
4508         if (r) {
4509                 pr_err("kvm: misc device register failed\n");
4510                 goto out_unreg;
4511         }
4512
4513         register_syscore_ops(&kvm_syscore_ops);
4514
4515         kvm_preempt_ops.sched_in = kvm_sched_in;
4516         kvm_preempt_ops.sched_out = kvm_sched_out;
4517
4518         kvm_init_debug();
4519
4520         r = kvm_vfio_ops_init();
4521         WARN_ON(r);
4522
4523         return 0;
4524
4525 out_unreg:
4526         kvm_async_pf_deinit();
4527 out_free:
4528         kmem_cache_destroy(kvm_vcpu_cache);
4529 out_free_3:
4530         unregister_reboot_notifier(&kvm_reboot_notifier);
4531         cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING);
4532 out_free_2:
4533         kvm_arch_hardware_unsetup();
4534 out_free_1:
4535         free_cpumask_var(cpus_hardware_enabled);
4536 out_free_0:
4537         kvm_irqfd_exit();
4538 out_irqfd:
4539         kvm_arch_exit();
4540 out_fail:
4541         return r;
4542 }
4543 EXPORT_SYMBOL_GPL(kvm_init);
4544
4545 void kvm_exit(void)
4546 {
4547         debugfs_remove_recursive(kvm_debugfs_dir);
4548         misc_deregister(&kvm_dev);
4549         kmem_cache_destroy(kvm_vcpu_cache);
4550         kvm_async_pf_deinit();
4551         unregister_syscore_ops(&kvm_syscore_ops);
4552         unregister_reboot_notifier(&kvm_reboot_notifier);
4553         cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING);
4554         on_each_cpu(hardware_disable_nolock, NULL, 1);
4555         kvm_arch_hardware_unsetup();
4556         kvm_arch_exit();
4557         kvm_irqfd_exit();
4558         free_cpumask_var(cpus_hardware_enabled);
4559         kvm_vfio_ops_exit();
4560 }
4561 EXPORT_SYMBOL_GPL(kvm_exit);
4562
4563 struct kvm_vm_worker_thread_context {
4564         struct kvm *kvm;
4565         struct task_struct *parent;
4566         struct completion init_done;
4567         kvm_vm_thread_fn_t thread_fn;
4568         uintptr_t data;
4569         int err;
4570 };
4571
4572 static int kvm_vm_worker_thread(void *context)
4573 {
4574         /*
4575          * The init_context is allocated on the stack of the parent thread, so
4576          * we have to locally copy anything that is needed beyond initialization
4577          */
4578         struct kvm_vm_worker_thread_context *init_context = context;
4579         struct kvm *kvm = init_context->kvm;
4580         kvm_vm_thread_fn_t thread_fn = init_context->thread_fn;
4581         uintptr_t data = init_context->data;
4582         int err;
4583
4584         err = kthread_park(current);
4585         /* kthread_park(current) is never supposed to return an error */
4586         WARN_ON(err != 0);
4587         if (err)
4588                 goto init_complete;
4589
4590         err = cgroup_attach_task_all(init_context->parent, current);
4591         if (err) {
4592                 kvm_err("%s: cgroup_attach_task_all failed with err %d\n",
4593                         __func__, err);
4594                 goto init_complete;
4595         }
4596
4597         set_user_nice(current, task_nice(init_context->parent));
4598
4599 init_complete:
4600         init_context->err = err;
4601         complete(&init_context->init_done);
4602         init_context = NULL;
4603
4604         if (err)
4605                 return err;
4606
4607         /* Wait to be woken up by the spawner before proceeding. */
4608         kthread_parkme();
4609
4610         if (!kthread_should_stop())
4611                 err = thread_fn(kvm, data);
4612
4613         return err;
4614 }
4615
4616 int kvm_vm_create_worker_thread(struct kvm *kvm, kvm_vm_thread_fn_t thread_fn,
4617                                 uintptr_t data, const char *name,
4618                                 struct task_struct **thread_ptr)
4619 {
4620         struct kvm_vm_worker_thread_context init_context = {};
4621         struct task_struct *thread;
4622
4623         *thread_ptr = NULL;
4624         init_context.kvm = kvm;
4625         init_context.parent = current;
4626         init_context.thread_fn = thread_fn;
4627         init_context.data = data;
4628         init_completion(&init_context.init_done);
4629
4630         thread = kthread_run(kvm_vm_worker_thread, &init_context,
4631                              "%s-%d", name, task_pid_nr(current));
4632         if (IS_ERR(thread))
4633                 return PTR_ERR(thread);
4634
4635         /* kthread_run is never supposed to return NULL */
4636         WARN_ON(thread == NULL);
4637
4638         wait_for_completion(&init_context.init_done);
4639
4640         if (!init_context.err)
4641                 *thread_ptr = thread;
4642
4643         return init_context.err;
4644 }