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 #include <linux/suspend.h> 55 56 #include <asm/processor.h> 57 #include <asm/ioctl.h> 58 #include <linux/uaccess.h> 59 60 #include "coalesced_mmio.h" 61 #include "async_pf.h" 62 #include "mmu_lock.h" 63 #include "vfio.h" 64 65 #define CREATE_TRACE_POINTS 66 #include <trace/events/kvm.h> 67 68 #include <linux/kvm_dirty_ring.h> 69 70 /* Worst case buffer size needed for holding an integer. */ 71 #define ITOA_MAX_LEN 12 72 73 MODULE_AUTHOR("Qumranet"); 74 MODULE_LICENSE("GPL"); 75 76 /* Architectures should define their poll value according to the halt latency */ 77 unsigned int halt_poll_ns = KVM_HALT_POLL_NS_DEFAULT; 78 module_param(halt_poll_ns, uint, 0644); 79 EXPORT_SYMBOL_GPL(halt_poll_ns); 80 81 /* Default doubles per-vcpu halt_poll_ns. */ 82 unsigned int halt_poll_ns_grow = 2; 83 module_param(halt_poll_ns_grow, uint, 0644); 84 EXPORT_SYMBOL_GPL(halt_poll_ns_grow); 85 86 /* The start value to grow halt_poll_ns from */ 87 unsigned int halt_poll_ns_grow_start = 10000; /* 10us */ 88 module_param(halt_poll_ns_grow_start, uint, 0644); 89 EXPORT_SYMBOL_GPL(halt_poll_ns_grow_start); 90 91 /* Default resets per-vcpu halt_poll_ns . */ 92 unsigned int halt_poll_ns_shrink; 93 module_param(halt_poll_ns_shrink, uint, 0644); 94 EXPORT_SYMBOL_GPL(halt_poll_ns_shrink); 95 96 /* 97 * Ordering of locks: 98 * 99 * kvm->lock --> kvm->slots_lock --> kvm->irq_lock 100 */ 101 102 DEFINE_MUTEX(kvm_lock); 103 static DEFINE_RAW_SPINLOCK(kvm_count_lock); 104 LIST_HEAD(vm_list); 105 106 static cpumask_var_t cpus_hardware_enabled; 107 static int kvm_usage_count; 108 static atomic_t hardware_enable_failed; 109 110 static struct kmem_cache *kvm_vcpu_cache; 111 112 static __read_mostly struct preempt_ops kvm_preempt_ops; 113 static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_running_vcpu); 114 115 struct dentry *kvm_debugfs_dir; 116 EXPORT_SYMBOL_GPL(kvm_debugfs_dir); 117 118 static const struct file_operations stat_fops_per_vm; 119 120 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl, 121 unsigned long arg); 122 #ifdef CONFIG_KVM_COMPAT 123 static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl, 124 unsigned long arg); 125 #define KVM_COMPAT(c) .compat_ioctl = (c) 126 #else 127 /* 128 * For architectures that don't implement a compat infrastructure, 129 * adopt a double line of defense: 130 * - Prevent a compat task from opening /dev/kvm 131 * - If the open has been done by a 64bit task, and the KVM fd 132 * passed to a compat task, let the ioctls fail. 133 */ 134 static long kvm_no_compat_ioctl(struct file *file, unsigned int ioctl, 135 unsigned long arg) { return -EINVAL; } 136 137 static int kvm_no_compat_open(struct inode *inode, struct file *file) 138 { 139 return is_compat_task() ? -ENODEV : 0; 140 } 141 #define KVM_COMPAT(c) .compat_ioctl = kvm_no_compat_ioctl, \ 142 .open = kvm_no_compat_open 143 #endif 144 static int hardware_enable_all(void); 145 static void hardware_disable_all(void); 146 147 static void kvm_io_bus_destroy(struct kvm_io_bus *bus); 148 149 __visible bool kvm_rebooting; 150 EXPORT_SYMBOL_GPL(kvm_rebooting); 151 152 #define KVM_EVENT_CREATE_VM 0 153 #define KVM_EVENT_DESTROY_VM 1 154 static void kvm_uevent_notify_change(unsigned int type, struct kvm *kvm); 155 static unsigned long long kvm_createvm_count; 156 static unsigned long long kvm_active_vms; 157 158 __weak void kvm_arch_mmu_notifier_invalidate_range(struct kvm *kvm, 159 unsigned long start, unsigned long end) 160 { 161 } 162 163 bool kvm_is_zone_device_pfn(kvm_pfn_t pfn) 164 { 165 /* 166 * The metadata used by is_zone_device_page() to determine whether or 167 * not a page is ZONE_DEVICE is guaranteed to be valid if and only if 168 * the device has been pinned, e.g. by get_user_pages(). WARN if the 169 * page_count() is zero to help detect bad usage of this helper. 170 */ 171 if (!pfn_valid(pfn) || WARN_ON_ONCE(!page_count(pfn_to_page(pfn)))) 172 return false; 173 174 return is_zone_device_page(pfn_to_page(pfn)); 175 } 176 177 bool kvm_is_reserved_pfn(kvm_pfn_t pfn) 178 { 179 /* 180 * ZONE_DEVICE pages currently set PG_reserved, but from a refcounting 181 * perspective they are "normal" pages, albeit with slightly different 182 * usage rules. 183 */ 184 if (pfn_valid(pfn)) 185 return PageReserved(pfn_to_page(pfn)) && 186 !is_zero_pfn(pfn) && 187 !kvm_is_zone_device_pfn(pfn); 188 189 return true; 190 } 191 192 /* 193 * Switches to specified vcpu, until a matching vcpu_put() 194 */ 195 void vcpu_load(struct kvm_vcpu *vcpu) 196 { 197 int cpu = get_cpu(); 198 199 __this_cpu_write(kvm_running_vcpu, vcpu); 200 preempt_notifier_register(&vcpu->preempt_notifier); 201 kvm_arch_vcpu_load(vcpu, cpu); 202 put_cpu(); 203 } 204 EXPORT_SYMBOL_GPL(vcpu_load); 205 206 void vcpu_put(struct kvm_vcpu *vcpu) 207 { 208 preempt_disable(); 209 kvm_arch_vcpu_put(vcpu); 210 preempt_notifier_unregister(&vcpu->preempt_notifier); 211 __this_cpu_write(kvm_running_vcpu, NULL); 212 preempt_enable(); 213 } 214 EXPORT_SYMBOL_GPL(vcpu_put); 215 216 /* TODO: merge with kvm_arch_vcpu_should_kick */ 217 static bool kvm_request_needs_ipi(struct kvm_vcpu *vcpu, unsigned req) 218 { 219 int mode = kvm_vcpu_exiting_guest_mode(vcpu); 220 221 /* 222 * We need to wait for the VCPU to reenable interrupts and get out of 223 * READING_SHADOW_PAGE_TABLES mode. 224 */ 225 if (req & KVM_REQUEST_WAIT) 226 return mode != OUTSIDE_GUEST_MODE; 227 228 /* 229 * Need to kick a running VCPU, but otherwise there is nothing to do. 230 */ 231 return mode == IN_GUEST_MODE; 232 } 233 234 static void ack_flush(void *_completed) 235 { 236 } 237 238 static inline bool kvm_kick_many_cpus(cpumask_var_t tmp, bool wait) 239 { 240 const struct cpumask *cpus; 241 242 if (likely(cpumask_available(tmp))) 243 cpus = tmp; 244 else 245 cpus = cpu_online_mask; 246 247 if (cpumask_empty(cpus)) 248 return false; 249 250 smp_call_function_many(cpus, ack_flush, NULL, wait); 251 return true; 252 } 253 254 bool kvm_make_vcpus_request_mask(struct kvm *kvm, unsigned int req, 255 struct kvm_vcpu *except, 256 unsigned long *vcpu_bitmap, cpumask_var_t tmp) 257 { 258 int i, cpu, me; 259 struct kvm_vcpu *vcpu; 260 bool called; 261 262 me = get_cpu(); 263 264 kvm_for_each_vcpu(i, vcpu, kvm) { 265 if ((vcpu_bitmap && !test_bit(i, vcpu_bitmap)) || 266 vcpu == except) 267 continue; 268 269 kvm_make_request(req, vcpu); 270 271 if (!(req & KVM_REQUEST_NO_WAKEUP) && kvm_vcpu_wake_up(vcpu)) 272 continue; 273 274 /* 275 * tmp can be "unavailable" if cpumasks are allocated off stack 276 * as allocation of the mask is deliberately not fatal and is 277 * handled by falling back to kicking all online CPUs. 278 */ 279 if (!cpumask_available(tmp)) 280 continue; 281 282 /* 283 * Note, the vCPU could get migrated to a different pCPU at any 284 * point after kvm_request_needs_ipi(), which could result in 285 * sending an IPI to the previous pCPU. But, that's ok because 286 * the purpose of the IPI is to ensure the vCPU returns to 287 * OUTSIDE_GUEST_MODE, which is satisfied if the vCPU migrates. 288 * Entering READING_SHADOW_PAGE_TABLES after this point is also 289 * ok, as the requirement is only that KVM wait for vCPUs that 290 * were reading SPTEs _before_ any changes were finalized. See 291 * kvm_vcpu_kick() for more details on handling requests. 292 */ 293 if (kvm_request_needs_ipi(vcpu, req)) { 294 cpu = READ_ONCE(vcpu->cpu); 295 if (cpu != -1 && cpu != me) 296 __cpumask_set_cpu(cpu, tmp); 297 } 298 } 299 300 called = kvm_kick_many_cpus(tmp, !!(req & KVM_REQUEST_WAIT)); 301 put_cpu(); 302 303 return called; 304 } 305 306 bool kvm_make_all_cpus_request_except(struct kvm *kvm, unsigned int req, 307 struct kvm_vcpu *except) 308 { 309 cpumask_var_t cpus; 310 bool called; 311 312 zalloc_cpumask_var(&cpus, GFP_ATOMIC); 313 314 called = kvm_make_vcpus_request_mask(kvm, req, except, NULL, cpus); 315 316 free_cpumask_var(cpus); 317 return called; 318 } 319 320 bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req) 321 { 322 return kvm_make_all_cpus_request_except(kvm, req, NULL); 323 } 324 EXPORT_SYMBOL_GPL(kvm_make_all_cpus_request); 325 326 #ifndef CONFIG_HAVE_KVM_ARCH_TLB_FLUSH_ALL 327 void kvm_flush_remote_tlbs(struct kvm *kvm) 328 { 329 ++kvm->stat.generic.remote_tlb_flush_requests; 330 331 /* 332 * We want to publish modifications to the page tables before reading 333 * mode. Pairs with a memory barrier in arch-specific code. 334 * - x86: smp_mb__after_srcu_read_unlock in vcpu_enter_guest 335 * and smp_mb in walk_shadow_page_lockless_begin/end. 336 * - powerpc: smp_mb in kvmppc_prepare_to_enter. 337 * 338 * There is already an smp_mb__after_atomic() before 339 * kvm_make_all_cpus_request() reads vcpu->mode. We reuse that 340 * barrier here. 341 */ 342 if (!kvm_arch_flush_remote_tlb(kvm) 343 || kvm_make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH)) 344 ++kvm->stat.generic.remote_tlb_flush; 345 } 346 EXPORT_SYMBOL_GPL(kvm_flush_remote_tlbs); 347 #endif 348 349 void kvm_reload_remote_mmus(struct kvm *kvm) 350 { 351 kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD); 352 } 353 354 #ifdef KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE 355 static inline void *mmu_memory_cache_alloc_obj(struct kvm_mmu_memory_cache *mc, 356 gfp_t gfp_flags) 357 { 358 gfp_flags |= mc->gfp_zero; 359 360 if (mc->kmem_cache) 361 return kmem_cache_alloc(mc->kmem_cache, gfp_flags); 362 else 363 return (void *)__get_free_page(gfp_flags); 364 } 365 366 int kvm_mmu_topup_memory_cache(struct kvm_mmu_memory_cache *mc, int min) 367 { 368 void *obj; 369 370 if (mc->nobjs >= min) 371 return 0; 372 while (mc->nobjs < ARRAY_SIZE(mc->objects)) { 373 obj = mmu_memory_cache_alloc_obj(mc, GFP_KERNEL_ACCOUNT); 374 if (!obj) 375 return mc->nobjs >= min ? 0 : -ENOMEM; 376 mc->objects[mc->nobjs++] = obj; 377 } 378 return 0; 379 } 380 381 int kvm_mmu_memory_cache_nr_free_objects(struct kvm_mmu_memory_cache *mc) 382 { 383 return mc->nobjs; 384 } 385 386 void kvm_mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc) 387 { 388 while (mc->nobjs) { 389 if (mc->kmem_cache) 390 kmem_cache_free(mc->kmem_cache, mc->objects[--mc->nobjs]); 391 else 392 free_page((unsigned long)mc->objects[--mc->nobjs]); 393 } 394 } 395 396 void *kvm_mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc) 397 { 398 void *p; 399 400 if (WARN_ON(!mc->nobjs)) 401 p = mmu_memory_cache_alloc_obj(mc, GFP_ATOMIC | __GFP_ACCOUNT); 402 else 403 p = mc->objects[--mc->nobjs]; 404 BUG_ON(!p); 405 return p; 406 } 407 #endif 408 409 static void kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id) 410 { 411 mutex_init(&vcpu->mutex); 412 vcpu->cpu = -1; 413 vcpu->kvm = kvm; 414 vcpu->vcpu_id = id; 415 vcpu->pid = NULL; 416 rcuwait_init(&vcpu->wait); 417 kvm_async_pf_vcpu_init(vcpu); 418 419 vcpu->pre_pcpu = -1; 420 INIT_LIST_HEAD(&vcpu->blocked_vcpu_list); 421 422 kvm_vcpu_set_in_spin_loop(vcpu, false); 423 kvm_vcpu_set_dy_eligible(vcpu, false); 424 vcpu->preempted = false; 425 vcpu->ready = false; 426 preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops); 427 vcpu->last_used_slot = 0; 428 } 429 430 void kvm_vcpu_destroy(struct kvm_vcpu *vcpu) 431 { 432 kvm_dirty_ring_free(&vcpu->dirty_ring); 433 kvm_arch_vcpu_destroy(vcpu); 434 435 /* 436 * No need for rcu_read_lock as VCPU_RUN is the only place that changes 437 * the vcpu->pid pointer, and at destruction time all file descriptors 438 * are already gone. 439 */ 440 put_pid(rcu_dereference_protected(vcpu->pid, 1)); 441 442 free_page((unsigned long)vcpu->run); 443 kmem_cache_free(kvm_vcpu_cache, vcpu); 444 } 445 EXPORT_SYMBOL_GPL(kvm_vcpu_destroy); 446 447 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) 448 static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn) 449 { 450 return container_of(mn, struct kvm, mmu_notifier); 451 } 452 453 static void kvm_mmu_notifier_invalidate_range(struct mmu_notifier *mn, 454 struct mm_struct *mm, 455 unsigned long start, unsigned long end) 456 { 457 struct kvm *kvm = mmu_notifier_to_kvm(mn); 458 int idx; 459 460 idx = srcu_read_lock(&kvm->srcu); 461 kvm_arch_mmu_notifier_invalidate_range(kvm, start, end); 462 srcu_read_unlock(&kvm->srcu, idx); 463 } 464 465 typedef bool (*hva_handler_t)(struct kvm *kvm, struct kvm_gfn_range *range); 466 467 typedef void (*on_lock_fn_t)(struct kvm *kvm, unsigned long start, 468 unsigned long end); 469 470 struct kvm_hva_range { 471 unsigned long start; 472 unsigned long end; 473 pte_t pte; 474 hva_handler_t handler; 475 on_lock_fn_t on_lock; 476 bool flush_on_ret; 477 bool may_block; 478 }; 479 480 /* 481 * Use a dedicated stub instead of NULL to indicate that there is no callback 482 * function/handler. The compiler technically can't guarantee that a real 483 * function will have a non-zero address, and so it will generate code to 484 * check for !NULL, whereas comparing against a stub will be elided at compile 485 * time (unless the compiler is getting long in the tooth, e.g. gcc 4.9). 486 */ 487 static void kvm_null_fn(void) 488 { 489 490 } 491 #define IS_KVM_NULL_FN(fn) ((fn) == (void *)kvm_null_fn) 492 493 static __always_inline int __kvm_handle_hva_range(struct kvm *kvm, 494 const struct kvm_hva_range *range) 495 { 496 bool ret = false, locked = false; 497 struct kvm_gfn_range gfn_range; 498 struct kvm_memory_slot *slot; 499 struct kvm_memslots *slots; 500 int i, idx; 501 502 /* A null handler is allowed if and only if on_lock() is provided. */ 503 if (WARN_ON_ONCE(IS_KVM_NULL_FN(range->on_lock) && 504 IS_KVM_NULL_FN(range->handler))) 505 return 0; 506 507 idx = srcu_read_lock(&kvm->srcu); 508 509 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) { 510 slots = __kvm_memslots(kvm, i); 511 kvm_for_each_memslot(slot, slots) { 512 unsigned long hva_start, hva_end; 513 514 hva_start = max(range->start, slot->userspace_addr); 515 hva_end = min(range->end, slot->userspace_addr + 516 (slot->npages << PAGE_SHIFT)); 517 if (hva_start >= hva_end) 518 continue; 519 520 /* 521 * To optimize for the likely case where the address 522 * range is covered by zero or one memslots, don't 523 * bother making these conditional (to avoid writes on 524 * the second or later invocation of the handler). 525 */ 526 gfn_range.pte = range->pte; 527 gfn_range.may_block = range->may_block; 528 529 /* 530 * {gfn(page) | page intersects with [hva_start, hva_end)} = 531 * {gfn_start, gfn_start+1, ..., gfn_end-1}. 532 */ 533 gfn_range.start = hva_to_gfn_memslot(hva_start, slot); 534 gfn_range.end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, slot); 535 gfn_range.slot = slot; 536 537 if (!locked) { 538 locked = true; 539 KVM_MMU_LOCK(kvm); 540 if (!IS_KVM_NULL_FN(range->on_lock)) 541 range->on_lock(kvm, range->start, range->end); 542 if (IS_KVM_NULL_FN(range->handler)) 543 break; 544 } 545 ret |= range->handler(kvm, &gfn_range); 546 } 547 } 548 549 if (range->flush_on_ret && ret) 550 kvm_flush_remote_tlbs(kvm); 551 552 if (locked) 553 KVM_MMU_UNLOCK(kvm); 554 555 srcu_read_unlock(&kvm->srcu, idx); 556 557 /* The notifiers are averse to booleans. :-( */ 558 return (int)ret; 559 } 560 561 static __always_inline int kvm_handle_hva_range(struct mmu_notifier *mn, 562 unsigned long start, 563 unsigned long end, 564 pte_t pte, 565 hva_handler_t handler) 566 { 567 struct kvm *kvm = mmu_notifier_to_kvm(mn); 568 const struct kvm_hva_range range = { 569 .start = start, 570 .end = end, 571 .pte = pte, 572 .handler = handler, 573 .on_lock = (void *)kvm_null_fn, 574 .flush_on_ret = true, 575 .may_block = false, 576 }; 577 578 return __kvm_handle_hva_range(kvm, &range); 579 } 580 581 static __always_inline int kvm_handle_hva_range_no_flush(struct mmu_notifier *mn, 582 unsigned long start, 583 unsigned long end, 584 hva_handler_t handler) 585 { 586 struct kvm *kvm = mmu_notifier_to_kvm(mn); 587 const struct kvm_hva_range range = { 588 .start = start, 589 .end = end, 590 .pte = __pte(0), 591 .handler = handler, 592 .on_lock = (void *)kvm_null_fn, 593 .flush_on_ret = false, 594 .may_block = false, 595 }; 596 597 return __kvm_handle_hva_range(kvm, &range); 598 } 599 static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn, 600 struct mm_struct *mm, 601 unsigned long address, 602 pte_t pte) 603 { 604 struct kvm *kvm = mmu_notifier_to_kvm(mn); 605 606 trace_kvm_set_spte_hva(address); 607 608 /* 609 * .change_pte() must be surrounded by .invalidate_range_{start,end}(). 610 * If mmu_notifier_count is zero, then no in-progress invalidations, 611 * including this one, found a relevant memslot at start(); rechecking 612 * memslots here is unnecessary. Note, a false positive (count elevated 613 * by a different invalidation) is sub-optimal but functionally ok. 614 */ 615 WARN_ON_ONCE(!READ_ONCE(kvm->mn_active_invalidate_count)); 616 if (!READ_ONCE(kvm->mmu_notifier_count)) 617 return; 618 619 kvm_handle_hva_range(mn, address, address + 1, pte, kvm_set_spte_gfn); 620 } 621 622 void kvm_inc_notifier_count(struct kvm *kvm, unsigned long start, 623 unsigned long end) 624 { 625 /* 626 * The count increase must become visible at unlock time as no 627 * spte can be established without taking the mmu_lock and 628 * count is also read inside the mmu_lock critical section. 629 */ 630 kvm->mmu_notifier_count++; 631 if (likely(kvm->mmu_notifier_count == 1)) { 632 kvm->mmu_notifier_range_start = start; 633 kvm->mmu_notifier_range_end = end; 634 } else { 635 /* 636 * Fully tracking multiple concurrent ranges has dimishing 637 * returns. Keep things simple and just find the minimal range 638 * which includes the current and new ranges. As there won't be 639 * enough information to subtract a range after its invalidate 640 * completes, any ranges invalidated concurrently will 641 * accumulate and persist until all outstanding invalidates 642 * complete. 643 */ 644 kvm->mmu_notifier_range_start = 645 min(kvm->mmu_notifier_range_start, start); 646 kvm->mmu_notifier_range_end = 647 max(kvm->mmu_notifier_range_end, end); 648 } 649 } 650 651 static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, 652 const struct mmu_notifier_range *range) 653 { 654 struct kvm *kvm = mmu_notifier_to_kvm(mn); 655 const struct kvm_hva_range hva_range = { 656 .start = range->start, 657 .end = range->end, 658 .pte = __pte(0), 659 .handler = kvm_unmap_gfn_range, 660 .on_lock = kvm_inc_notifier_count, 661 .flush_on_ret = true, 662 .may_block = mmu_notifier_range_blockable(range), 663 }; 664 665 trace_kvm_unmap_hva_range(range->start, range->end); 666 667 /* 668 * Prevent memslot modification between range_start() and range_end() 669 * so that conditionally locking provides the same result in both 670 * functions. Without that guarantee, the mmu_notifier_count 671 * adjustments will be imbalanced. 672 * 673 * Pairs with the decrement in range_end(). 674 */ 675 spin_lock(&kvm->mn_invalidate_lock); 676 kvm->mn_active_invalidate_count++; 677 spin_unlock(&kvm->mn_invalidate_lock); 678 679 __kvm_handle_hva_range(kvm, &hva_range); 680 681 return 0; 682 } 683 684 void kvm_dec_notifier_count(struct kvm *kvm, unsigned long start, 685 unsigned long end) 686 { 687 /* 688 * This sequence increase will notify the kvm page fault that 689 * the page that is going to be mapped in the spte could have 690 * been freed. 691 */ 692 kvm->mmu_notifier_seq++; 693 smp_wmb(); 694 /* 695 * The above sequence increase must be visible before the 696 * below count decrease, which is ensured by the smp_wmb above 697 * in conjunction with the smp_rmb in mmu_notifier_retry(). 698 */ 699 kvm->mmu_notifier_count--; 700 } 701 702 static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, 703 const struct mmu_notifier_range *range) 704 { 705 struct kvm *kvm = mmu_notifier_to_kvm(mn); 706 const struct kvm_hva_range hva_range = { 707 .start = range->start, 708 .end = range->end, 709 .pte = __pte(0), 710 .handler = (void *)kvm_null_fn, 711 .on_lock = kvm_dec_notifier_count, 712 .flush_on_ret = false, 713 .may_block = mmu_notifier_range_blockable(range), 714 }; 715 bool wake; 716 717 __kvm_handle_hva_range(kvm, &hva_range); 718 719 /* Pairs with the increment in range_start(). */ 720 spin_lock(&kvm->mn_invalidate_lock); 721 wake = (--kvm->mn_active_invalidate_count == 0); 722 spin_unlock(&kvm->mn_invalidate_lock); 723 724 /* 725 * There can only be one waiter, since the wait happens under 726 * slots_lock. 727 */ 728 if (wake) 729 rcuwait_wake_up(&kvm->mn_memslots_update_rcuwait); 730 731 BUG_ON(kvm->mmu_notifier_count < 0); 732 } 733 734 static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn, 735 struct mm_struct *mm, 736 unsigned long start, 737 unsigned long end) 738 { 739 trace_kvm_age_hva(start, end); 740 741 return kvm_handle_hva_range(mn, start, end, __pte(0), kvm_age_gfn); 742 } 743 744 static int kvm_mmu_notifier_clear_young(struct mmu_notifier *mn, 745 struct mm_struct *mm, 746 unsigned long start, 747 unsigned long end) 748 { 749 trace_kvm_age_hva(start, end); 750 751 /* 752 * Even though we do not flush TLB, this will still adversely 753 * affect performance on pre-Haswell Intel EPT, where there is 754 * no EPT Access Bit to clear so that we have to tear down EPT 755 * tables instead. If we find this unacceptable, we can always 756 * add a parameter to kvm_age_hva so that it effectively doesn't 757 * do anything on clear_young. 758 * 759 * Also note that currently we never issue secondary TLB flushes 760 * from clear_young, leaving this job up to the regular system 761 * cadence. If we find this inaccurate, we might come up with a 762 * more sophisticated heuristic later. 763 */ 764 return kvm_handle_hva_range_no_flush(mn, start, end, kvm_age_gfn); 765 } 766 767 static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn, 768 struct mm_struct *mm, 769 unsigned long address) 770 { 771 trace_kvm_test_age_hva(address); 772 773 return kvm_handle_hva_range_no_flush(mn, address, address + 1, 774 kvm_test_age_gfn); 775 } 776 777 static void kvm_mmu_notifier_release(struct mmu_notifier *mn, 778 struct mm_struct *mm) 779 { 780 struct kvm *kvm = mmu_notifier_to_kvm(mn); 781 int idx; 782 783 idx = srcu_read_lock(&kvm->srcu); 784 kvm_arch_flush_shadow_all(kvm); 785 srcu_read_unlock(&kvm->srcu, idx); 786 } 787 788 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = { 789 .invalidate_range = kvm_mmu_notifier_invalidate_range, 790 .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start, 791 .invalidate_range_end = kvm_mmu_notifier_invalidate_range_end, 792 .clear_flush_young = kvm_mmu_notifier_clear_flush_young, 793 .clear_young = kvm_mmu_notifier_clear_young, 794 .test_young = kvm_mmu_notifier_test_young, 795 .change_pte = kvm_mmu_notifier_change_pte, 796 .release = kvm_mmu_notifier_release, 797 }; 798 799 static int kvm_init_mmu_notifier(struct kvm *kvm) 800 { 801 kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops; 802 return mmu_notifier_register(&kvm->mmu_notifier, current->mm); 803 } 804 805 #else /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */ 806 807 static int kvm_init_mmu_notifier(struct kvm *kvm) 808 { 809 return 0; 810 } 811 812 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */ 813 814 #ifdef CONFIG_HAVE_KVM_PM_NOTIFIER 815 static int kvm_pm_notifier_call(struct notifier_block *bl, 816 unsigned long state, 817 void *unused) 818 { 819 struct kvm *kvm = container_of(bl, struct kvm, pm_notifier); 820 821 return kvm_arch_pm_notifier(kvm, state); 822 } 823 824 static void kvm_init_pm_notifier(struct kvm *kvm) 825 { 826 kvm->pm_notifier.notifier_call = kvm_pm_notifier_call; 827 /* Suspend KVM before we suspend ftrace, RCU, etc. */ 828 kvm->pm_notifier.priority = INT_MAX; 829 register_pm_notifier(&kvm->pm_notifier); 830 } 831 832 static void kvm_destroy_pm_notifier(struct kvm *kvm) 833 { 834 unregister_pm_notifier(&kvm->pm_notifier); 835 } 836 #else /* !CONFIG_HAVE_KVM_PM_NOTIFIER */ 837 static void kvm_init_pm_notifier(struct kvm *kvm) 838 { 839 } 840 841 static void kvm_destroy_pm_notifier(struct kvm *kvm) 842 { 843 } 844 #endif /* CONFIG_HAVE_KVM_PM_NOTIFIER */ 845 846 static struct kvm_memslots *kvm_alloc_memslots(void) 847 { 848 int i; 849 struct kvm_memslots *slots; 850 851 slots = kvzalloc(sizeof(struct kvm_memslots), GFP_KERNEL_ACCOUNT); 852 if (!slots) 853 return NULL; 854 855 for (i = 0; i < KVM_MEM_SLOTS_NUM; i++) 856 slots->id_to_index[i] = -1; 857 858 return slots; 859 } 860 861 static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot) 862 { 863 if (!memslot->dirty_bitmap) 864 return; 865 866 kvfree(memslot->dirty_bitmap); 867 memslot->dirty_bitmap = NULL; 868 } 869 870 static void kvm_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot) 871 { 872 kvm_destroy_dirty_bitmap(slot); 873 874 kvm_arch_free_memslot(kvm, slot); 875 876 slot->flags = 0; 877 slot->npages = 0; 878 } 879 880 static void kvm_free_memslots(struct kvm *kvm, struct kvm_memslots *slots) 881 { 882 struct kvm_memory_slot *memslot; 883 884 if (!slots) 885 return; 886 887 kvm_for_each_memslot(memslot, slots) 888 kvm_free_memslot(kvm, memslot); 889 890 kvfree(slots); 891 } 892 893 static umode_t kvm_stats_debugfs_mode(const struct _kvm_stats_desc *pdesc) 894 { 895 switch (pdesc->desc.flags & KVM_STATS_TYPE_MASK) { 896 case KVM_STATS_TYPE_INSTANT: 897 return 0444; 898 case KVM_STATS_TYPE_CUMULATIVE: 899 case KVM_STATS_TYPE_PEAK: 900 default: 901 return 0644; 902 } 903 } 904 905 906 static void kvm_destroy_vm_debugfs(struct kvm *kvm) 907 { 908 int i; 909 int kvm_debugfs_num_entries = kvm_vm_stats_header.num_desc + 910 kvm_vcpu_stats_header.num_desc; 911 912 if (!kvm->debugfs_dentry) 913 return; 914 915 debugfs_remove_recursive(kvm->debugfs_dentry); 916 917 if (kvm->debugfs_stat_data) { 918 for (i = 0; i < kvm_debugfs_num_entries; i++) 919 kfree(kvm->debugfs_stat_data[i]); 920 kfree(kvm->debugfs_stat_data); 921 } 922 } 923 924 static int kvm_create_vm_debugfs(struct kvm *kvm, int fd) 925 { 926 static DEFINE_MUTEX(kvm_debugfs_lock); 927 struct dentry *dent; 928 char dir_name[ITOA_MAX_LEN * 2]; 929 struct kvm_stat_data *stat_data; 930 const struct _kvm_stats_desc *pdesc; 931 int i, ret; 932 int kvm_debugfs_num_entries = kvm_vm_stats_header.num_desc + 933 kvm_vcpu_stats_header.num_desc; 934 935 if (!debugfs_initialized()) 936 return 0; 937 938 snprintf(dir_name, sizeof(dir_name), "%d-%d", task_pid_nr(current), fd); 939 mutex_lock(&kvm_debugfs_lock); 940 dent = debugfs_lookup(dir_name, kvm_debugfs_dir); 941 if (dent) { 942 pr_warn_ratelimited("KVM: debugfs: duplicate directory %s\n", dir_name); 943 dput(dent); 944 mutex_unlock(&kvm_debugfs_lock); 945 return 0; 946 } 947 dent = debugfs_create_dir(dir_name, kvm_debugfs_dir); 948 mutex_unlock(&kvm_debugfs_lock); 949 if (IS_ERR(dent)) 950 return 0; 951 952 kvm->debugfs_dentry = dent; 953 kvm->debugfs_stat_data = kcalloc(kvm_debugfs_num_entries, 954 sizeof(*kvm->debugfs_stat_data), 955 GFP_KERNEL_ACCOUNT); 956 if (!kvm->debugfs_stat_data) 957 return -ENOMEM; 958 959 for (i = 0; i < kvm_vm_stats_header.num_desc; ++i) { 960 pdesc = &kvm_vm_stats_desc[i]; 961 stat_data = kzalloc(sizeof(*stat_data), GFP_KERNEL_ACCOUNT); 962 if (!stat_data) 963 return -ENOMEM; 964 965 stat_data->kvm = kvm; 966 stat_data->desc = pdesc; 967 stat_data->kind = KVM_STAT_VM; 968 kvm->debugfs_stat_data[i] = stat_data; 969 debugfs_create_file(pdesc->name, kvm_stats_debugfs_mode(pdesc), 970 kvm->debugfs_dentry, stat_data, 971 &stat_fops_per_vm); 972 } 973 974 for (i = 0; i < kvm_vcpu_stats_header.num_desc; ++i) { 975 pdesc = &kvm_vcpu_stats_desc[i]; 976 stat_data = kzalloc(sizeof(*stat_data), GFP_KERNEL_ACCOUNT); 977 if (!stat_data) 978 return -ENOMEM; 979 980 stat_data->kvm = kvm; 981 stat_data->desc = pdesc; 982 stat_data->kind = KVM_STAT_VCPU; 983 kvm->debugfs_stat_data[i + kvm_vm_stats_header.num_desc] = stat_data; 984 debugfs_create_file(pdesc->name, kvm_stats_debugfs_mode(pdesc), 985 kvm->debugfs_dentry, stat_data, 986 &stat_fops_per_vm); 987 } 988 989 ret = kvm_arch_create_vm_debugfs(kvm); 990 if (ret) { 991 kvm_destroy_vm_debugfs(kvm); 992 return i; 993 } 994 995 return 0; 996 } 997 998 /* 999 * Called after the VM is otherwise initialized, but just before adding it to 1000 * the vm_list. 1001 */ 1002 int __weak kvm_arch_post_init_vm(struct kvm *kvm) 1003 { 1004 return 0; 1005 } 1006 1007 /* 1008 * Called just after removing the VM from the vm_list, but before doing any 1009 * other destruction. 1010 */ 1011 void __weak kvm_arch_pre_destroy_vm(struct kvm *kvm) 1012 { 1013 } 1014 1015 /* 1016 * Called after per-vm debugfs created. When called kvm->debugfs_dentry should 1017 * be setup already, so we can create arch-specific debugfs entries under it. 1018 * Cleanup should be automatic done in kvm_destroy_vm_debugfs() recursively, so 1019 * a per-arch destroy interface is not needed. 1020 */ 1021 int __weak kvm_arch_create_vm_debugfs(struct kvm *kvm) 1022 { 1023 return 0; 1024 } 1025 1026 static struct kvm *kvm_create_vm(unsigned long type) 1027 { 1028 struct kvm *kvm = kvm_arch_alloc_vm(); 1029 int r = -ENOMEM; 1030 int i; 1031 1032 if (!kvm) 1033 return ERR_PTR(-ENOMEM); 1034 1035 KVM_MMU_LOCK_INIT(kvm); 1036 mmgrab(current->mm); 1037 kvm->mm = current->mm; 1038 kvm_eventfd_init(kvm); 1039 mutex_init(&kvm->lock); 1040 mutex_init(&kvm->irq_lock); 1041 mutex_init(&kvm->slots_lock); 1042 mutex_init(&kvm->slots_arch_lock); 1043 spin_lock_init(&kvm->mn_invalidate_lock); 1044 rcuwait_init(&kvm->mn_memslots_update_rcuwait); 1045 1046 INIT_LIST_HEAD(&kvm->devices); 1047 1048 BUILD_BUG_ON(KVM_MEM_SLOTS_NUM > SHRT_MAX); 1049 1050 if (init_srcu_struct(&kvm->srcu)) 1051 goto out_err_no_srcu; 1052 if (init_srcu_struct(&kvm->irq_srcu)) 1053 goto out_err_no_irq_srcu; 1054 1055 refcount_set(&kvm->users_count, 1); 1056 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) { 1057 struct kvm_memslots *slots = kvm_alloc_memslots(); 1058 1059 if (!slots) 1060 goto out_err_no_arch_destroy_vm; 1061 /* Generations must be different for each address space. */ 1062 slots->generation = i; 1063 rcu_assign_pointer(kvm->memslots[i], slots); 1064 } 1065 1066 for (i = 0; i < KVM_NR_BUSES; i++) { 1067 rcu_assign_pointer(kvm->buses[i], 1068 kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL_ACCOUNT)); 1069 if (!kvm->buses[i]) 1070 goto out_err_no_arch_destroy_vm; 1071 } 1072 1073 kvm->max_halt_poll_ns = halt_poll_ns; 1074 1075 r = kvm_arch_init_vm(kvm, type); 1076 if (r) 1077 goto out_err_no_arch_destroy_vm; 1078 1079 r = hardware_enable_all(); 1080 if (r) 1081 goto out_err_no_disable; 1082 1083 #ifdef CONFIG_HAVE_KVM_IRQFD 1084 INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list); 1085 #endif 1086 1087 r = kvm_init_mmu_notifier(kvm); 1088 if (r) 1089 goto out_err_no_mmu_notifier; 1090 1091 r = kvm_arch_post_init_vm(kvm); 1092 if (r) 1093 goto out_err; 1094 1095 mutex_lock(&kvm_lock); 1096 list_add(&kvm->vm_list, &vm_list); 1097 mutex_unlock(&kvm_lock); 1098 1099 preempt_notifier_inc(); 1100 kvm_init_pm_notifier(kvm); 1101 1102 return kvm; 1103 1104 out_err: 1105 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) 1106 if (kvm->mmu_notifier.ops) 1107 mmu_notifier_unregister(&kvm->mmu_notifier, current->mm); 1108 #endif 1109 out_err_no_mmu_notifier: 1110 hardware_disable_all(); 1111 out_err_no_disable: 1112 kvm_arch_destroy_vm(kvm); 1113 out_err_no_arch_destroy_vm: 1114 WARN_ON_ONCE(!refcount_dec_and_test(&kvm->users_count)); 1115 for (i = 0; i < KVM_NR_BUSES; i++) 1116 kfree(kvm_get_bus(kvm, i)); 1117 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) 1118 kvm_free_memslots(kvm, __kvm_memslots(kvm, i)); 1119 cleanup_srcu_struct(&kvm->irq_srcu); 1120 out_err_no_irq_srcu: 1121 cleanup_srcu_struct(&kvm->srcu); 1122 out_err_no_srcu: 1123 kvm_arch_free_vm(kvm); 1124 mmdrop(current->mm); 1125 return ERR_PTR(r); 1126 } 1127 1128 static void kvm_destroy_devices(struct kvm *kvm) 1129 { 1130 struct kvm_device *dev, *tmp; 1131 1132 /* 1133 * We do not need to take the kvm->lock here, because nobody else 1134 * has a reference to the struct kvm at this point and therefore 1135 * cannot access the devices list anyhow. 1136 */ 1137 list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) { 1138 list_del(&dev->vm_node); 1139 dev->ops->destroy(dev); 1140 } 1141 } 1142 1143 static void kvm_destroy_vm(struct kvm *kvm) 1144 { 1145 int i; 1146 struct mm_struct *mm = kvm->mm; 1147 1148 kvm_destroy_pm_notifier(kvm); 1149 kvm_uevent_notify_change(KVM_EVENT_DESTROY_VM, kvm); 1150 kvm_destroy_vm_debugfs(kvm); 1151 kvm_arch_sync_events(kvm); 1152 mutex_lock(&kvm_lock); 1153 list_del(&kvm->vm_list); 1154 mutex_unlock(&kvm_lock); 1155 kvm_arch_pre_destroy_vm(kvm); 1156 1157 kvm_free_irq_routing(kvm); 1158 for (i = 0; i < KVM_NR_BUSES; i++) { 1159 struct kvm_io_bus *bus = kvm_get_bus(kvm, i); 1160 1161 if (bus) 1162 kvm_io_bus_destroy(bus); 1163 kvm->buses[i] = NULL; 1164 } 1165 kvm_coalesced_mmio_free(kvm); 1166 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) 1167 mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm); 1168 /* 1169 * At this point, pending calls to invalidate_range_start() 1170 * have completed but no more MMU notifiers will run, so 1171 * mn_active_invalidate_count may remain unbalanced. 1172 * No threads can be waiting in install_new_memslots as the 1173 * last reference on KVM has been dropped, but freeing 1174 * memslots would deadlock without this manual intervention. 1175 */ 1176 WARN_ON(rcuwait_active(&kvm->mn_memslots_update_rcuwait)); 1177 kvm->mn_active_invalidate_count = 0; 1178 #else 1179 kvm_arch_flush_shadow_all(kvm); 1180 #endif 1181 kvm_arch_destroy_vm(kvm); 1182 kvm_destroy_devices(kvm); 1183 for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) 1184 kvm_free_memslots(kvm, __kvm_memslots(kvm, i)); 1185 cleanup_srcu_struct(&kvm->irq_srcu); 1186 cleanup_srcu_struct(&kvm->srcu); 1187 kvm_arch_free_vm(kvm); 1188 preempt_notifier_dec(); 1189 hardware_disable_all(); 1190 mmdrop(mm); 1191 } 1192 1193 void kvm_get_kvm(struct kvm *kvm) 1194 { 1195 refcount_inc(&kvm->users_count); 1196 } 1197 EXPORT_SYMBOL_GPL(kvm_get_kvm); 1198 1199 /* 1200 * Make sure the vm is not during destruction, which is a safe version of 1201 * kvm_get_kvm(). Return true if kvm referenced successfully, false otherwise. 1202 */ 1203 bool kvm_get_kvm_safe(struct kvm *kvm) 1204 { 1205 return refcount_inc_not_zero(&kvm->users_count); 1206 } 1207 EXPORT_SYMBOL_GPL(kvm_get_kvm_safe); 1208 1209 void kvm_put_kvm(struct kvm *kvm) 1210 { 1211 if (refcount_dec_and_test(&kvm->users_count)) 1212 kvm_destroy_vm(kvm); 1213 } 1214 EXPORT_SYMBOL_GPL(kvm_put_kvm); 1215 1216 /* 1217 * Used to put a reference that was taken on behalf of an object associated 1218 * with a user-visible file descriptor, e.g. a vcpu or device, if installation 1219 * of the new file descriptor fails and the reference cannot be transferred to 1220 * its final owner. In such cases, the caller is still actively using @kvm and 1221 * will fail miserably if the refcount unexpectedly hits zero. 1222 */ 1223 void kvm_put_kvm_no_destroy(struct kvm *kvm) 1224 { 1225 WARN_ON(refcount_dec_and_test(&kvm->users_count)); 1226 } 1227 EXPORT_SYMBOL_GPL(kvm_put_kvm_no_destroy); 1228 1229 static int kvm_vm_release(struct inode *inode, struct file *filp) 1230 { 1231 struct kvm *kvm = filp->private_data; 1232 1233 kvm_irqfd_release(kvm); 1234 1235 kvm_put_kvm(kvm); 1236 return 0; 1237 } 1238 1239 /* 1240 * Allocation size is twice as large as the actual dirty bitmap size. 1241 * See kvm_vm_ioctl_get_dirty_log() why this is needed. 1242 */ 1243 static int kvm_alloc_dirty_bitmap(struct kvm_memory_slot *memslot) 1244 { 1245 unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot); 1246 1247 memslot->dirty_bitmap = kvzalloc(dirty_bytes, GFP_KERNEL_ACCOUNT); 1248 if (!memslot->dirty_bitmap) 1249 return -ENOMEM; 1250 1251 return 0; 1252 } 1253 1254 /* 1255 * Delete a memslot by decrementing the number of used slots and shifting all 1256 * other entries in the array forward one spot. 1257 */ 1258 static inline void kvm_memslot_delete(struct kvm_memslots *slots, 1259 struct kvm_memory_slot *memslot) 1260 { 1261 struct kvm_memory_slot *mslots = slots->memslots; 1262 int i; 1263 1264 if (WARN_ON(slots->id_to_index[memslot->id] == -1)) 1265 return; 1266 1267 slots->used_slots--; 1268 1269 if (atomic_read(&slots->last_used_slot) >= slots->used_slots) 1270 atomic_set(&slots->last_used_slot, 0); 1271 1272 for (i = slots->id_to_index[memslot->id]; i < slots->used_slots; i++) { 1273 mslots[i] = mslots[i + 1]; 1274 slots->id_to_index[mslots[i].id] = i; 1275 } 1276 mslots[i] = *memslot; 1277 slots->id_to_index[memslot->id] = -1; 1278 } 1279 1280 /* 1281 * "Insert" a new memslot by incrementing the number of used slots. Returns 1282 * the new slot's initial index into the memslots array. 1283 */ 1284 static inline int kvm_memslot_insert_back(struct kvm_memslots *slots) 1285 { 1286 return slots->used_slots++; 1287 } 1288 1289 /* 1290 * Move a changed memslot backwards in the array by shifting existing slots 1291 * with a higher GFN toward the front of the array. Note, the changed memslot 1292 * itself is not preserved in the array, i.e. not swapped at this time, only 1293 * its new index into the array is tracked. Returns the changed memslot's 1294 * current index into the memslots array. 1295 */ 1296 static inline int kvm_memslot_move_backward(struct kvm_memslots *slots, 1297 struct kvm_memory_slot *memslot) 1298 { 1299 struct kvm_memory_slot *mslots = slots->memslots; 1300 int i; 1301 1302 if (WARN_ON_ONCE(slots->id_to_index[memslot->id] == -1) || 1303 WARN_ON_ONCE(!slots->used_slots)) 1304 return -1; 1305 1306 /* 1307 * Move the target memslot backward in the array by shifting existing 1308 * memslots with a higher GFN (than the target memslot) towards the 1309 * front of the array. 1310 */ 1311 for (i = slots->id_to_index[memslot->id]; i < slots->used_slots - 1; i++) { 1312 if (memslot->base_gfn > mslots[i + 1].base_gfn) 1313 break; 1314 1315 WARN_ON_ONCE(memslot->base_gfn == mslots[i + 1].base_gfn); 1316 1317 /* Shift the next memslot forward one and update its index. */ 1318 mslots[i] = mslots[i + 1]; 1319 slots->id_to_index[mslots[i].id] = i; 1320 } 1321 return i; 1322 } 1323 1324 /* 1325 * Move a changed memslot forwards in the array by shifting existing slots with 1326 * a lower GFN toward the back of the array. Note, the changed memslot itself 1327 * is not preserved in the array, i.e. not swapped at this time, only its new 1328 * index into the array is tracked. Returns the changed memslot's final index 1329 * into the memslots array. 1330 */ 1331 static inline int kvm_memslot_move_forward(struct kvm_memslots *slots, 1332 struct kvm_memory_slot *memslot, 1333 int start) 1334 { 1335 struct kvm_memory_slot *mslots = slots->memslots; 1336 int i; 1337 1338 for (i = start; i > 0; i--) { 1339 if (memslot->base_gfn < mslots[i - 1].base_gfn) 1340 break; 1341 1342 WARN_ON_ONCE(memslot->base_gfn == mslots[i - 1].base_gfn); 1343 1344 /* Shift the next memslot back one and update its index. */ 1345 mslots[i] = mslots[i - 1]; 1346 slots->id_to_index[mslots[i].id] = i; 1347 } 1348 return i; 1349 } 1350 1351 /* 1352 * Re-sort memslots based on their GFN to account for an added, deleted, or 1353 * moved memslot. Sorting memslots by GFN allows using a binary search during 1354 * memslot lookup. 1355 * 1356 * IMPORTANT: Slots are sorted from highest GFN to lowest GFN! I.e. the entry 1357 * at memslots[0] has the highest GFN. 1358 * 1359 * The sorting algorithm takes advantage of having initially sorted memslots 1360 * and knowing the position of the changed memslot. Sorting is also optimized 1361 * by not swapping the updated memslot and instead only shifting other memslots 1362 * and tracking the new index for the update memslot. Only once its final 1363 * index is known is the updated memslot copied into its position in the array. 1364 * 1365 * - When deleting a memslot, the deleted memslot simply needs to be moved to 1366 * the end of the array. 1367 * 1368 * - When creating a memslot, the algorithm "inserts" the new memslot at the 1369 * end of the array and then it forward to its correct location. 1370 * 1371 * - When moving a memslot, the algorithm first moves the updated memslot 1372 * backward to handle the scenario where the memslot's GFN was changed to a 1373 * lower value. update_memslots() then falls through and runs the same flow 1374 * as creating a memslot to move the memslot forward to handle the scenario 1375 * where its GFN was changed to a higher value. 1376 * 1377 * Note, slots are sorted from highest->lowest instead of lowest->highest for 1378 * historical reasons. Originally, invalid memslots where denoted by having 1379 * GFN=0, thus sorting from highest->lowest naturally sorted invalid memslots 1380 * to the end of the array. The current algorithm uses dedicated logic to 1381 * delete a memslot and thus does not rely on invalid memslots having GFN=0. 1382 * 1383 * The other historical motiviation for highest->lowest was to improve the 1384 * performance of memslot lookup. KVM originally used a linear search starting 1385 * at memslots[0]. On x86, the largest memslot usually has one of the highest, 1386 * if not *the* highest, GFN, as the bulk of the guest's RAM is located in a 1387 * single memslot above the 4gb boundary. As the largest memslot is also the 1388 * most likely to be referenced, sorting it to the front of the array was 1389 * advantageous. The current binary search starts from the middle of the array 1390 * and uses an LRU pointer to improve performance for all memslots and GFNs. 1391 */ 1392 static void update_memslots(struct kvm_memslots *slots, 1393 struct kvm_memory_slot *memslot, 1394 enum kvm_mr_change change) 1395 { 1396 int i; 1397 1398 if (change == KVM_MR_DELETE) { 1399 kvm_memslot_delete(slots, memslot); 1400 } else { 1401 if (change == KVM_MR_CREATE) 1402 i = kvm_memslot_insert_back(slots); 1403 else 1404 i = kvm_memslot_move_backward(slots, memslot); 1405 i = kvm_memslot_move_forward(slots, memslot, i); 1406 1407 /* 1408 * Copy the memslot to its new position in memslots and update 1409 * its index accordingly. 1410 */ 1411 slots->memslots[i] = *memslot; 1412 slots->id_to_index[memslot->id] = i; 1413 } 1414 } 1415 1416 static int check_memory_region_flags(const struct kvm_userspace_memory_region *mem) 1417 { 1418 u32 valid_flags = KVM_MEM_LOG_DIRTY_PAGES; 1419 1420 #ifdef __KVM_HAVE_READONLY_MEM 1421 valid_flags |= KVM_MEM_READONLY; 1422 #endif 1423 1424 if (mem->flags & ~valid_flags) 1425 return -EINVAL; 1426 1427 return 0; 1428 } 1429 1430 static struct kvm_memslots *install_new_memslots(struct kvm *kvm, 1431 int as_id, struct kvm_memslots *slots) 1432 { 1433 struct kvm_memslots *old_memslots = __kvm_memslots(kvm, as_id); 1434 u64 gen = old_memslots->generation; 1435 1436 WARN_ON(gen & KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS); 1437 slots->generation = gen | KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS; 1438 1439 /* 1440 * Do not store the new memslots while there are invalidations in 1441 * progress, otherwise the locking in invalidate_range_start and 1442 * invalidate_range_end will be unbalanced. 1443 */ 1444 spin_lock(&kvm->mn_invalidate_lock); 1445 prepare_to_rcuwait(&kvm->mn_memslots_update_rcuwait); 1446 while (kvm->mn_active_invalidate_count) { 1447 set_current_state(TASK_UNINTERRUPTIBLE); 1448 spin_unlock(&kvm->mn_invalidate_lock); 1449 schedule(); 1450 spin_lock(&kvm->mn_invalidate_lock); 1451 } 1452 finish_rcuwait(&kvm->mn_memslots_update_rcuwait); 1453 rcu_assign_pointer(kvm->memslots[as_id], slots); 1454 spin_unlock(&kvm->mn_invalidate_lock); 1455 1456 /* 1457 * Acquired in kvm_set_memslot. Must be released before synchronize 1458 * SRCU below in order to avoid deadlock with another thread 1459 * acquiring the slots_arch_lock in an srcu critical section. 1460 */ 1461 mutex_unlock(&kvm->slots_arch_lock); 1462 1463 synchronize_srcu_expedited(&kvm->srcu); 1464 1465 /* 1466 * Increment the new memslot generation a second time, dropping the 1467 * update in-progress flag and incrementing the generation based on 1468 * the number of address spaces. This provides a unique and easily 1469 * identifiable generation number while the memslots are in flux. 1470 */ 1471 gen = slots->generation & ~KVM_MEMSLOT_GEN_UPDATE_IN_PROGRESS; 1472 1473 /* 1474 * Generations must be unique even across address spaces. We do not need 1475 * a global counter for that, instead the generation space is evenly split 1476 * across address spaces. For example, with two address spaces, address 1477 * space 0 will use generations 0, 2, 4, ... while address space 1 will 1478 * use generations 1, 3, 5, ... 1479 */ 1480 gen += KVM_ADDRESS_SPACE_NUM; 1481 1482 kvm_arch_memslots_updated(kvm, gen); 1483 1484 slots->generation = gen; 1485 1486 return old_memslots; 1487 } 1488 1489 static size_t kvm_memslots_size(int slots) 1490 { 1491 return sizeof(struct kvm_memslots) + 1492 (sizeof(struct kvm_memory_slot) * slots); 1493 } 1494 1495 static void kvm_copy_memslots(struct kvm_memslots *to, 1496 struct kvm_memslots *from) 1497 { 1498 memcpy(to, from, kvm_memslots_size(from->used_slots)); 1499 } 1500 1501 /* 1502 * Note, at a minimum, the current number of used slots must be allocated, even 1503 * when deleting a memslot, as we need a complete duplicate of the memslots for 1504 * use when invalidating a memslot prior to deleting/moving the memslot. 1505 */ 1506 static struct kvm_memslots *kvm_dup_memslots(struct kvm_memslots *old, 1507 enum kvm_mr_change change) 1508 { 1509 struct kvm_memslots *slots; 1510 size_t new_size; 1511 1512 if (change == KVM_MR_CREATE) 1513 new_size = kvm_memslots_size(old->used_slots + 1); 1514 else 1515 new_size = kvm_memslots_size(old->used_slots); 1516 1517 slots = kvzalloc(new_size, GFP_KERNEL_ACCOUNT); 1518 if (likely(slots)) 1519 kvm_copy_memslots(slots, old); 1520 1521 return slots; 1522 } 1523 1524 static int kvm_set_memslot(struct kvm *kvm, 1525 const struct kvm_userspace_memory_region *mem, 1526 struct kvm_memory_slot *old, 1527 struct kvm_memory_slot *new, int as_id, 1528 enum kvm_mr_change change) 1529 { 1530 struct kvm_memory_slot *slot; 1531 struct kvm_memslots *slots; 1532 int r; 1533 1534 /* 1535 * Released in install_new_memslots. 1536 * 1537 * Must be held from before the current memslots are copied until 1538 * after the new memslots are installed with rcu_assign_pointer, 1539 * then released before the synchronize srcu in install_new_memslots. 1540 * 1541 * When modifying memslots outside of the slots_lock, must be held 1542 * before reading the pointer to the current memslots until after all 1543 * changes to those memslots are complete. 1544 * 1545 * These rules ensure that installing new memslots does not lose 1546 * changes made to the previous memslots. 1547 */ 1548 mutex_lock(&kvm->slots_arch_lock); 1549 1550 slots = kvm_dup_memslots(__kvm_memslots(kvm, as_id), change); 1551 if (!slots) { 1552 mutex_unlock(&kvm->slots_arch_lock); 1553 return -ENOMEM; 1554 } 1555 1556 if (change == KVM_MR_DELETE || change == KVM_MR_MOVE) { 1557 /* 1558 * Note, the INVALID flag needs to be in the appropriate entry 1559 * in the freshly allocated memslots, not in @old or @new. 1560 */ 1561 slot = id_to_memslot(slots, old->id); 1562 slot->flags |= KVM_MEMSLOT_INVALID; 1563 1564 /* 1565 * We can re-use the memory from the old memslots. 1566 * It will be overwritten with a copy of the new memslots 1567 * after reacquiring the slots_arch_lock below. 1568 */ 1569 slots = install_new_memslots(kvm, as_id, slots); 1570 1571 /* From this point no new shadow pages pointing to a deleted, 1572 * or moved, memslot will be created. 1573 * 1574 * validation of sp->gfn happens in: 1575 * - gfn_to_hva (kvm_read_guest, gfn_to_pfn) 1576 * - kvm_is_visible_gfn (mmu_check_root) 1577 */ 1578 kvm_arch_flush_shadow_memslot(kvm, slot); 1579 1580 /* Released in install_new_memslots. */ 1581 mutex_lock(&kvm->slots_arch_lock); 1582 1583 /* 1584 * The arch-specific fields of the memslots could have changed 1585 * between releasing the slots_arch_lock in 1586 * install_new_memslots and here, so get a fresh copy of the 1587 * slots. 1588 */ 1589 kvm_copy_memslots(slots, __kvm_memslots(kvm, as_id)); 1590 } 1591 1592 r = kvm_arch_prepare_memory_region(kvm, new, mem, change); 1593 if (r) 1594 goto out_slots; 1595 1596 update_memslots(slots, new, change); 1597 slots = install_new_memslots(kvm, as_id, slots); 1598 1599 kvm_arch_commit_memory_region(kvm, mem, old, new, change); 1600 1601 kvfree(slots); 1602 return 0; 1603 1604 out_slots: 1605 if (change == KVM_MR_DELETE || change == KVM_MR_MOVE) { 1606 slot = id_to_memslot(slots, old->id); 1607 slot->flags &= ~KVM_MEMSLOT_INVALID; 1608 slots = install_new_memslots(kvm, as_id, slots); 1609 } else { 1610 mutex_unlock(&kvm->slots_arch_lock); 1611 } 1612 kvfree(slots); 1613 return r; 1614 } 1615 1616 static int kvm_delete_memslot(struct kvm *kvm, 1617 const struct kvm_userspace_memory_region *mem, 1618 struct kvm_memory_slot *old, int as_id) 1619 { 1620 struct kvm_memory_slot new; 1621 int r; 1622 1623 if (!old->npages) 1624 return -EINVAL; 1625 1626 memset(&new, 0, sizeof(new)); 1627 new.id = old->id; 1628 /* 1629 * This is only for debugging purpose; it should never be referenced 1630 * for a removed memslot. 1631 */ 1632 new.as_id = as_id; 1633 1634 r = kvm_set_memslot(kvm, mem, old, &new, as_id, KVM_MR_DELETE); 1635 if (r) 1636 return r; 1637 1638 kvm_free_memslot(kvm, old); 1639 return 0; 1640 } 1641 1642 /* 1643 * Allocate some memory and give it an address in the guest physical address 1644 * space. 1645 * 1646 * Discontiguous memory is allowed, mostly for framebuffers. 1647 * 1648 * Must be called holding kvm->slots_lock for write. 1649 */ 1650 int __kvm_set_memory_region(struct kvm *kvm, 1651 const struct kvm_userspace_memory_region *mem) 1652 { 1653 struct kvm_memory_slot old, new; 1654 struct kvm_memory_slot *tmp; 1655 enum kvm_mr_change change; 1656 int as_id, id; 1657 int r; 1658 1659 r = check_memory_region_flags(mem); 1660 if (r) 1661 return r; 1662 1663 as_id = mem->slot >> 16; 1664 id = (u16)mem->slot; 1665 1666 /* General sanity checks */ 1667 if (mem->memory_size & (PAGE_SIZE - 1)) 1668 return -EINVAL; 1669 if (mem->guest_phys_addr & (PAGE_SIZE - 1)) 1670 return -EINVAL; 1671 /* We can read the guest memory with __xxx_user() later on. */ 1672 if ((mem->userspace_addr & (PAGE_SIZE - 1)) || 1673 (mem->userspace_addr != untagged_addr(mem->userspace_addr)) || 1674 !access_ok((void __user *)(unsigned long)mem->userspace_addr, 1675 mem->memory_size)) 1676 return -EINVAL; 1677 if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_MEM_SLOTS_NUM) 1678 return -EINVAL; 1679 if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr) 1680 return -EINVAL; 1681 1682 /* 1683 * Make a full copy of the old memslot, the pointer will become stale 1684 * when the memslots are re-sorted by update_memslots(), and the old 1685 * memslot needs to be referenced after calling update_memslots(), e.g. 1686 * to free its resources and for arch specific behavior. 1687 */ 1688 tmp = id_to_memslot(__kvm_memslots(kvm, as_id), id); 1689 if (tmp) { 1690 old = *tmp; 1691 tmp = NULL; 1692 } else { 1693 memset(&old, 0, sizeof(old)); 1694 old.id = id; 1695 } 1696 1697 if (!mem->memory_size) 1698 return kvm_delete_memslot(kvm, mem, &old, as_id); 1699 1700 new.as_id = as_id; 1701 new.id = id; 1702 new.base_gfn = mem->guest_phys_addr >> PAGE_SHIFT; 1703 new.npages = mem->memory_size >> PAGE_SHIFT; 1704 new.flags = mem->flags; 1705 new.userspace_addr = mem->userspace_addr; 1706 1707 if (new.npages > KVM_MEM_MAX_NR_PAGES) 1708 return -EINVAL; 1709 1710 if (!old.npages) { 1711 change = KVM_MR_CREATE; 1712 new.dirty_bitmap = NULL; 1713 memset(&new.arch, 0, sizeof(new.arch)); 1714 } else { /* Modify an existing slot. */ 1715 if ((new.userspace_addr != old.userspace_addr) || 1716 (new.npages != old.npages) || 1717 ((new.flags ^ old.flags) & KVM_MEM_READONLY)) 1718 return -EINVAL; 1719 1720 if (new.base_gfn != old.base_gfn) 1721 change = KVM_MR_MOVE; 1722 else if (new.flags != old.flags) 1723 change = KVM_MR_FLAGS_ONLY; 1724 else /* Nothing to change. */ 1725 return 0; 1726 1727 /* Copy dirty_bitmap and arch from the current memslot. */ 1728 new.dirty_bitmap = old.dirty_bitmap; 1729 memcpy(&new.arch, &old.arch, sizeof(new.arch)); 1730 } 1731 1732 if ((change == KVM_MR_CREATE) || (change == KVM_MR_MOVE)) { 1733 /* Check for overlaps */ 1734 kvm_for_each_memslot(tmp, __kvm_memslots(kvm, as_id)) { 1735 if (tmp->id == id) 1736 continue; 1737 if (!((new.base_gfn + new.npages <= tmp->base_gfn) || 1738 (new.base_gfn >= tmp->base_gfn + tmp->npages))) 1739 return -EEXIST; 1740 } 1741 } 1742 1743 /* Allocate/free page dirty bitmap as needed */ 1744 if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES)) 1745 new.dirty_bitmap = NULL; 1746 else if (!new.dirty_bitmap && !kvm->dirty_ring_size) { 1747 r = kvm_alloc_dirty_bitmap(&new); 1748 if (r) 1749 return r; 1750 1751 if (kvm_dirty_log_manual_protect_and_init_set(kvm)) 1752 bitmap_set(new.dirty_bitmap, 0, new.npages); 1753 } 1754 1755 r = kvm_set_memslot(kvm, mem, &old, &new, as_id, change); 1756 if (r) 1757 goto out_bitmap; 1758 1759 if (old.dirty_bitmap && !new.dirty_bitmap) 1760 kvm_destroy_dirty_bitmap(&old); 1761 return 0; 1762 1763 out_bitmap: 1764 if (new.dirty_bitmap && !old.dirty_bitmap) 1765 kvm_destroy_dirty_bitmap(&new); 1766 return r; 1767 } 1768 EXPORT_SYMBOL_GPL(__kvm_set_memory_region); 1769 1770 int kvm_set_memory_region(struct kvm *kvm, 1771 const struct kvm_userspace_memory_region *mem) 1772 { 1773 int r; 1774 1775 mutex_lock(&kvm->slots_lock); 1776 r = __kvm_set_memory_region(kvm, mem); 1777 mutex_unlock(&kvm->slots_lock); 1778 return r; 1779 } 1780 EXPORT_SYMBOL_GPL(kvm_set_memory_region); 1781 1782 static int kvm_vm_ioctl_set_memory_region(struct kvm *kvm, 1783 struct kvm_userspace_memory_region *mem) 1784 { 1785 if ((u16)mem->slot >= KVM_USER_MEM_SLOTS) 1786 return -EINVAL; 1787 1788 return kvm_set_memory_region(kvm, mem); 1789 } 1790 1791 #ifndef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT 1792 /** 1793 * kvm_get_dirty_log - get a snapshot of dirty pages 1794 * @kvm: pointer to kvm instance 1795 * @log: slot id and address to which we copy the log 1796 * @is_dirty: set to '1' if any dirty pages were found 1797 * @memslot: set to the associated memslot, always valid on success 1798 */ 1799 int kvm_get_dirty_log(struct kvm *kvm, struct kvm_dirty_log *log, 1800 int *is_dirty, struct kvm_memory_slot **memslot) 1801 { 1802 struct kvm_memslots *slots; 1803 int i, as_id, id; 1804 unsigned long n; 1805 unsigned long any = 0; 1806 1807 /* Dirty ring tracking is exclusive to dirty log tracking */ 1808 if (kvm->dirty_ring_size) 1809 return -ENXIO; 1810 1811 *memslot = NULL; 1812 *is_dirty = 0; 1813 1814 as_id = log->slot >> 16; 1815 id = (u16)log->slot; 1816 if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS) 1817 return -EINVAL; 1818 1819 slots = __kvm_memslots(kvm, as_id); 1820 *memslot = id_to_memslot(slots, id); 1821 if (!(*memslot) || !(*memslot)->dirty_bitmap) 1822 return -ENOENT; 1823 1824 kvm_arch_sync_dirty_log(kvm, *memslot); 1825 1826 n = kvm_dirty_bitmap_bytes(*memslot); 1827 1828 for (i = 0; !any && i < n/sizeof(long); ++i) 1829 any = (*memslot)->dirty_bitmap[i]; 1830 1831 if (copy_to_user(log->dirty_bitmap, (*memslot)->dirty_bitmap, n)) 1832 return -EFAULT; 1833 1834 if (any) 1835 *is_dirty = 1; 1836 return 0; 1837 } 1838 EXPORT_SYMBOL_GPL(kvm_get_dirty_log); 1839 1840 #else /* CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT */ 1841 /** 1842 * kvm_get_dirty_log_protect - get a snapshot of dirty pages 1843 * and reenable dirty page tracking for the corresponding pages. 1844 * @kvm: pointer to kvm instance 1845 * @log: slot id and address to which we copy the log 1846 * 1847 * We need to keep it in mind that VCPU threads can write to the bitmap 1848 * concurrently. So, to avoid losing track of dirty pages we keep the 1849 * following order: 1850 * 1851 * 1. Take a snapshot of the bit and clear it if needed. 1852 * 2. Write protect the corresponding page. 1853 * 3. Copy the snapshot to the userspace. 1854 * 4. Upon return caller flushes TLB's if needed. 1855 * 1856 * Between 2 and 4, the guest may write to the page using the remaining TLB 1857 * entry. This is not a problem because the page is reported dirty using 1858 * the snapshot taken before and step 4 ensures that writes done after 1859 * exiting to userspace will be logged for the next call. 1860 * 1861 */ 1862 static int kvm_get_dirty_log_protect(struct kvm *kvm, struct kvm_dirty_log *log) 1863 { 1864 struct kvm_memslots *slots; 1865 struct kvm_memory_slot *memslot; 1866 int i, as_id, id; 1867 unsigned long n; 1868 unsigned long *dirty_bitmap; 1869 unsigned long *dirty_bitmap_buffer; 1870 bool flush; 1871 1872 /* Dirty ring tracking is exclusive to dirty log tracking */ 1873 if (kvm->dirty_ring_size) 1874 return -ENXIO; 1875 1876 as_id = log->slot >> 16; 1877 id = (u16)log->slot; 1878 if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS) 1879 return -EINVAL; 1880 1881 slots = __kvm_memslots(kvm, as_id); 1882 memslot = id_to_memslot(slots, id); 1883 if (!memslot || !memslot->dirty_bitmap) 1884 return -ENOENT; 1885 1886 dirty_bitmap = memslot->dirty_bitmap; 1887 1888 kvm_arch_sync_dirty_log(kvm, memslot); 1889 1890 n = kvm_dirty_bitmap_bytes(memslot); 1891 flush = false; 1892 if (kvm->manual_dirty_log_protect) { 1893 /* 1894 * Unlike kvm_get_dirty_log, we always return false in *flush, 1895 * because no flush is needed until KVM_CLEAR_DIRTY_LOG. There 1896 * is some code duplication between this function and 1897 * kvm_get_dirty_log, but hopefully all architecture 1898 * transition to kvm_get_dirty_log_protect and kvm_get_dirty_log 1899 * can be eliminated. 1900 */ 1901 dirty_bitmap_buffer = dirty_bitmap; 1902 } else { 1903 dirty_bitmap_buffer = kvm_second_dirty_bitmap(memslot); 1904 memset(dirty_bitmap_buffer, 0, n); 1905 1906 KVM_MMU_LOCK(kvm); 1907 for (i = 0; i < n / sizeof(long); i++) { 1908 unsigned long mask; 1909 gfn_t offset; 1910 1911 if (!dirty_bitmap[i]) 1912 continue; 1913 1914 flush = true; 1915 mask = xchg(&dirty_bitmap[i], 0); 1916 dirty_bitmap_buffer[i] = mask; 1917 1918 offset = i * BITS_PER_LONG; 1919 kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot, 1920 offset, mask); 1921 } 1922 KVM_MMU_UNLOCK(kvm); 1923 } 1924 1925 if (flush) 1926 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot); 1927 1928 if (copy_to_user(log->dirty_bitmap, dirty_bitmap_buffer, n)) 1929 return -EFAULT; 1930 return 0; 1931 } 1932 1933 1934 /** 1935 * kvm_vm_ioctl_get_dirty_log - get and clear the log of dirty pages in a slot 1936 * @kvm: kvm instance 1937 * @log: slot id and address to which we copy the log 1938 * 1939 * Steps 1-4 below provide general overview of dirty page logging. See 1940 * kvm_get_dirty_log_protect() function description for additional details. 1941 * 1942 * We call kvm_get_dirty_log_protect() to handle steps 1-3, upon return we 1943 * always flush the TLB (step 4) even if previous step failed and the dirty 1944 * bitmap may be corrupt. Regardless of previous outcome the KVM logging API 1945 * does not preclude user space subsequent dirty log read. Flushing TLB ensures 1946 * writes will be marked dirty for next log read. 1947 * 1948 * 1. Take a snapshot of the bit and clear it if needed. 1949 * 2. Write protect the corresponding page. 1950 * 3. Copy the snapshot to the userspace. 1951 * 4. Flush TLB's if needed. 1952 */ 1953 static int kvm_vm_ioctl_get_dirty_log(struct kvm *kvm, 1954 struct kvm_dirty_log *log) 1955 { 1956 int r; 1957 1958 mutex_lock(&kvm->slots_lock); 1959 1960 r = kvm_get_dirty_log_protect(kvm, log); 1961 1962 mutex_unlock(&kvm->slots_lock); 1963 return r; 1964 } 1965 1966 /** 1967 * kvm_clear_dirty_log_protect - clear dirty bits in the bitmap 1968 * and reenable dirty page tracking for the corresponding pages. 1969 * @kvm: pointer to kvm instance 1970 * @log: slot id and address from which to fetch the bitmap of dirty pages 1971 */ 1972 static int kvm_clear_dirty_log_protect(struct kvm *kvm, 1973 struct kvm_clear_dirty_log *log) 1974 { 1975 struct kvm_memslots *slots; 1976 struct kvm_memory_slot *memslot; 1977 int as_id, id; 1978 gfn_t offset; 1979 unsigned long i, n; 1980 unsigned long *dirty_bitmap; 1981 unsigned long *dirty_bitmap_buffer; 1982 bool flush; 1983 1984 /* Dirty ring tracking is exclusive to dirty log tracking */ 1985 if (kvm->dirty_ring_size) 1986 return -ENXIO; 1987 1988 as_id = log->slot >> 16; 1989 id = (u16)log->slot; 1990 if (as_id >= KVM_ADDRESS_SPACE_NUM || id >= KVM_USER_MEM_SLOTS) 1991 return -EINVAL; 1992 1993 if (log->first_page & 63) 1994 return -EINVAL; 1995 1996 slots = __kvm_memslots(kvm, as_id); 1997 memslot = id_to_memslot(slots, id); 1998 if (!memslot || !memslot->dirty_bitmap) 1999 return -ENOENT; 2000 2001 dirty_bitmap = memslot->dirty_bitmap; 2002 2003 n = ALIGN(log->num_pages, BITS_PER_LONG) / 8; 2004 2005 if (log->first_page > memslot->npages || 2006 log->num_pages > memslot->npages - log->first_page || 2007 (log->num_pages < memslot->npages - log->first_page && (log->num_pages & 63))) 2008 return -EINVAL; 2009 2010 kvm_arch_sync_dirty_log(kvm, memslot); 2011 2012 flush = false; 2013 dirty_bitmap_buffer = kvm_second_dirty_bitmap(memslot); 2014 if (copy_from_user(dirty_bitmap_buffer, log->dirty_bitmap, n)) 2015 return -EFAULT; 2016 2017 KVM_MMU_LOCK(kvm); 2018 for (offset = log->first_page, i = offset / BITS_PER_LONG, 2019 n = DIV_ROUND_UP(log->num_pages, BITS_PER_LONG); n--; 2020 i++, offset += BITS_PER_LONG) { 2021 unsigned long mask = *dirty_bitmap_buffer++; 2022 atomic_long_t *p = (atomic_long_t *) &dirty_bitmap[i]; 2023 if (!mask) 2024 continue; 2025 2026 mask &= atomic_long_fetch_andnot(mask, p); 2027 2028 /* 2029 * mask contains the bits that really have been cleared. This 2030 * never includes any bits beyond the length of the memslot (if 2031 * the length is not aligned to 64 pages), therefore it is not 2032 * a problem if userspace sets them in log->dirty_bitmap. 2033 */ 2034 if (mask) { 2035 flush = true; 2036 kvm_arch_mmu_enable_log_dirty_pt_masked(kvm, memslot, 2037 offset, mask); 2038 } 2039 } 2040 KVM_MMU_UNLOCK(kvm); 2041 2042 if (flush) 2043 kvm_arch_flush_remote_tlbs_memslot(kvm, memslot); 2044 2045 return 0; 2046 } 2047 2048 static int kvm_vm_ioctl_clear_dirty_log(struct kvm *kvm, 2049 struct kvm_clear_dirty_log *log) 2050 { 2051 int r; 2052 2053 mutex_lock(&kvm->slots_lock); 2054 2055 r = kvm_clear_dirty_log_protect(kvm, log); 2056 2057 mutex_unlock(&kvm->slots_lock); 2058 return r; 2059 } 2060 #endif /* CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT */ 2061 2062 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn) 2063 { 2064 return __gfn_to_memslot(kvm_memslots(kvm), gfn); 2065 } 2066 EXPORT_SYMBOL_GPL(gfn_to_memslot); 2067 2068 struct kvm_memory_slot *kvm_vcpu_gfn_to_memslot(struct kvm_vcpu *vcpu, gfn_t gfn) 2069 { 2070 struct kvm_memslots *slots = kvm_vcpu_memslots(vcpu); 2071 struct kvm_memory_slot *slot; 2072 int slot_index; 2073 2074 slot = try_get_memslot(slots, vcpu->last_used_slot, gfn); 2075 if (slot) 2076 return slot; 2077 2078 /* 2079 * Fall back to searching all memslots. We purposely use 2080 * search_memslots() instead of __gfn_to_memslot() to avoid 2081 * thrashing the VM-wide last_used_index in kvm_memslots. 2082 */ 2083 slot = search_memslots(slots, gfn, &slot_index); 2084 if (slot) { 2085 vcpu->last_used_slot = slot_index; 2086 return slot; 2087 } 2088 2089 return NULL; 2090 } 2091 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_memslot); 2092 2093 bool kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn) 2094 { 2095 struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn); 2096 2097 return kvm_is_visible_memslot(memslot); 2098 } 2099 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn); 2100 2101 bool kvm_vcpu_is_visible_gfn(struct kvm_vcpu *vcpu, gfn_t gfn) 2102 { 2103 struct kvm_memory_slot *memslot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 2104 2105 return kvm_is_visible_memslot(memslot); 2106 } 2107 EXPORT_SYMBOL_GPL(kvm_vcpu_is_visible_gfn); 2108 2109 unsigned long kvm_host_page_size(struct kvm_vcpu *vcpu, gfn_t gfn) 2110 { 2111 struct vm_area_struct *vma; 2112 unsigned long addr, size; 2113 2114 size = PAGE_SIZE; 2115 2116 addr = kvm_vcpu_gfn_to_hva_prot(vcpu, gfn, NULL); 2117 if (kvm_is_error_hva(addr)) 2118 return PAGE_SIZE; 2119 2120 mmap_read_lock(current->mm); 2121 vma = find_vma(current->mm, addr); 2122 if (!vma) 2123 goto out; 2124 2125 size = vma_kernel_pagesize(vma); 2126 2127 out: 2128 mmap_read_unlock(current->mm); 2129 2130 return size; 2131 } 2132 2133 static bool memslot_is_readonly(struct kvm_memory_slot *slot) 2134 { 2135 return slot->flags & KVM_MEM_READONLY; 2136 } 2137 2138 static unsigned long __gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn, 2139 gfn_t *nr_pages, bool write) 2140 { 2141 if (!slot || slot->flags & KVM_MEMSLOT_INVALID) 2142 return KVM_HVA_ERR_BAD; 2143 2144 if (memslot_is_readonly(slot) && write) 2145 return KVM_HVA_ERR_RO_BAD; 2146 2147 if (nr_pages) 2148 *nr_pages = slot->npages - (gfn - slot->base_gfn); 2149 2150 return __gfn_to_hva_memslot(slot, gfn); 2151 } 2152 2153 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn, 2154 gfn_t *nr_pages) 2155 { 2156 return __gfn_to_hva_many(slot, gfn, nr_pages, true); 2157 } 2158 2159 unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot, 2160 gfn_t gfn) 2161 { 2162 return gfn_to_hva_many(slot, gfn, NULL); 2163 } 2164 EXPORT_SYMBOL_GPL(gfn_to_hva_memslot); 2165 2166 unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn) 2167 { 2168 return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL); 2169 } 2170 EXPORT_SYMBOL_GPL(gfn_to_hva); 2171 2172 unsigned long kvm_vcpu_gfn_to_hva(struct kvm_vcpu *vcpu, gfn_t gfn) 2173 { 2174 return gfn_to_hva_many(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn, NULL); 2175 } 2176 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_hva); 2177 2178 /* 2179 * Return the hva of a @gfn and the R/W attribute if possible. 2180 * 2181 * @slot: the kvm_memory_slot which contains @gfn 2182 * @gfn: the gfn to be translated 2183 * @writable: used to return the read/write attribute of the @slot if the hva 2184 * is valid and @writable is not NULL 2185 */ 2186 unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot, 2187 gfn_t gfn, bool *writable) 2188 { 2189 unsigned long hva = __gfn_to_hva_many(slot, gfn, NULL, false); 2190 2191 if (!kvm_is_error_hva(hva) && writable) 2192 *writable = !memslot_is_readonly(slot); 2193 2194 return hva; 2195 } 2196 2197 unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable) 2198 { 2199 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn); 2200 2201 return gfn_to_hva_memslot_prot(slot, gfn, writable); 2202 } 2203 2204 unsigned long kvm_vcpu_gfn_to_hva_prot(struct kvm_vcpu *vcpu, gfn_t gfn, bool *writable) 2205 { 2206 struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 2207 2208 return gfn_to_hva_memslot_prot(slot, gfn, writable); 2209 } 2210 2211 static inline int check_user_page_hwpoison(unsigned long addr) 2212 { 2213 int rc, flags = FOLL_HWPOISON | FOLL_WRITE; 2214 2215 rc = get_user_pages(addr, 1, flags, NULL, NULL); 2216 return rc == -EHWPOISON; 2217 } 2218 2219 /* 2220 * The fast path to get the writable pfn which will be stored in @pfn, 2221 * true indicates success, otherwise false is returned. It's also the 2222 * only part that runs if we can in atomic context. 2223 */ 2224 static bool hva_to_pfn_fast(unsigned long addr, bool write_fault, 2225 bool *writable, kvm_pfn_t *pfn) 2226 { 2227 struct page *page[1]; 2228 2229 /* 2230 * Fast pin a writable pfn only if it is a write fault request 2231 * or the caller allows to map a writable pfn for a read fault 2232 * request. 2233 */ 2234 if (!(write_fault || writable)) 2235 return false; 2236 2237 if (get_user_page_fast_only(addr, FOLL_WRITE, page)) { 2238 *pfn = page_to_pfn(page[0]); 2239 2240 if (writable) 2241 *writable = true; 2242 return true; 2243 } 2244 2245 return false; 2246 } 2247 2248 /* 2249 * The slow path to get the pfn of the specified host virtual address, 2250 * 1 indicates success, -errno is returned if error is detected. 2251 */ 2252 static int hva_to_pfn_slow(unsigned long addr, bool *async, bool write_fault, 2253 bool *writable, kvm_pfn_t *pfn) 2254 { 2255 unsigned int flags = FOLL_HWPOISON; 2256 struct page *page; 2257 int npages = 0; 2258 2259 might_sleep(); 2260 2261 if (writable) 2262 *writable = write_fault; 2263 2264 if (write_fault) 2265 flags |= FOLL_WRITE; 2266 if (async) 2267 flags |= FOLL_NOWAIT; 2268 2269 npages = get_user_pages_unlocked(addr, 1, &page, flags); 2270 if (npages != 1) 2271 return npages; 2272 2273 /* map read fault as writable if possible */ 2274 if (unlikely(!write_fault) && writable) { 2275 struct page *wpage; 2276 2277 if (get_user_page_fast_only(addr, FOLL_WRITE, &wpage)) { 2278 *writable = true; 2279 put_page(page); 2280 page = wpage; 2281 } 2282 } 2283 *pfn = page_to_pfn(page); 2284 return npages; 2285 } 2286 2287 static bool vma_is_valid(struct vm_area_struct *vma, bool write_fault) 2288 { 2289 if (unlikely(!(vma->vm_flags & VM_READ))) 2290 return false; 2291 2292 if (write_fault && (unlikely(!(vma->vm_flags & VM_WRITE)))) 2293 return false; 2294 2295 return true; 2296 } 2297 2298 static int kvm_try_get_pfn(kvm_pfn_t pfn) 2299 { 2300 if (kvm_is_reserved_pfn(pfn)) 2301 return 1; 2302 return get_page_unless_zero(pfn_to_page(pfn)); 2303 } 2304 2305 static int hva_to_pfn_remapped(struct vm_area_struct *vma, 2306 unsigned long addr, bool *async, 2307 bool write_fault, bool *writable, 2308 kvm_pfn_t *p_pfn) 2309 { 2310 kvm_pfn_t pfn; 2311 pte_t *ptep; 2312 spinlock_t *ptl; 2313 int r; 2314 2315 r = follow_pte(vma->vm_mm, addr, &ptep, &ptl); 2316 if (r) { 2317 /* 2318 * get_user_pages fails for VM_IO and VM_PFNMAP vmas and does 2319 * not call the fault handler, so do it here. 2320 */ 2321 bool unlocked = false; 2322 r = fixup_user_fault(current->mm, addr, 2323 (write_fault ? FAULT_FLAG_WRITE : 0), 2324 &unlocked); 2325 if (unlocked) 2326 return -EAGAIN; 2327 if (r) 2328 return r; 2329 2330 r = follow_pte(vma->vm_mm, addr, &ptep, &ptl); 2331 if (r) 2332 return r; 2333 } 2334 2335 if (write_fault && !pte_write(*ptep)) { 2336 pfn = KVM_PFN_ERR_RO_FAULT; 2337 goto out; 2338 } 2339 2340 if (writable) 2341 *writable = pte_write(*ptep); 2342 pfn = pte_pfn(*ptep); 2343 2344 /* 2345 * Get a reference here because callers of *hva_to_pfn* and 2346 * *gfn_to_pfn* ultimately call kvm_release_pfn_clean on the 2347 * returned pfn. This is only needed if the VMA has VM_MIXEDMAP 2348 * set, but the kvm_try_get_pfn/kvm_release_pfn_clean pair will 2349 * simply do nothing for reserved pfns. 2350 * 2351 * Whoever called remap_pfn_range is also going to call e.g. 2352 * unmap_mapping_range before the underlying pages are freed, 2353 * causing a call to our MMU notifier. 2354 * 2355 * Certain IO or PFNMAP mappings can be backed with valid 2356 * struct pages, but be allocated without refcounting e.g., 2357 * tail pages of non-compound higher order allocations, which 2358 * would then underflow the refcount when the caller does the 2359 * required put_page. Don't allow those pages here. 2360 */ 2361 if (!kvm_try_get_pfn(pfn)) 2362 r = -EFAULT; 2363 2364 out: 2365 pte_unmap_unlock(ptep, ptl); 2366 *p_pfn = pfn; 2367 2368 return r; 2369 } 2370 2371 /* 2372 * Pin guest page in memory and return its pfn. 2373 * @addr: host virtual address which maps memory to the guest 2374 * @atomic: whether this function can sleep 2375 * @async: whether this function need to wait IO complete if the 2376 * host page is not in the memory 2377 * @write_fault: whether we should get a writable host page 2378 * @writable: whether it allows to map a writable host page for !@write_fault 2379 * 2380 * The function will map a writable host page for these two cases: 2381 * 1): @write_fault = true 2382 * 2): @write_fault = false && @writable, @writable will tell the caller 2383 * whether the mapping is writable. 2384 */ 2385 static kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool *async, 2386 bool write_fault, bool *writable) 2387 { 2388 struct vm_area_struct *vma; 2389 kvm_pfn_t pfn = 0; 2390 int npages, r; 2391 2392 /* we can do it either atomically or asynchronously, not both */ 2393 BUG_ON(atomic && async); 2394 2395 if (hva_to_pfn_fast(addr, write_fault, writable, &pfn)) 2396 return pfn; 2397 2398 if (atomic) 2399 return KVM_PFN_ERR_FAULT; 2400 2401 npages = hva_to_pfn_slow(addr, async, write_fault, writable, &pfn); 2402 if (npages == 1) 2403 return pfn; 2404 2405 mmap_read_lock(current->mm); 2406 if (npages == -EHWPOISON || 2407 (!async && check_user_page_hwpoison(addr))) { 2408 pfn = KVM_PFN_ERR_HWPOISON; 2409 goto exit; 2410 } 2411 2412 retry: 2413 vma = vma_lookup(current->mm, addr); 2414 2415 if (vma == NULL) 2416 pfn = KVM_PFN_ERR_FAULT; 2417 else if (vma->vm_flags & (VM_IO | VM_PFNMAP)) { 2418 r = hva_to_pfn_remapped(vma, addr, async, write_fault, writable, &pfn); 2419 if (r == -EAGAIN) 2420 goto retry; 2421 if (r < 0) 2422 pfn = KVM_PFN_ERR_FAULT; 2423 } else { 2424 if (async && vma_is_valid(vma, write_fault)) 2425 *async = true; 2426 pfn = KVM_PFN_ERR_FAULT; 2427 } 2428 exit: 2429 mmap_read_unlock(current->mm); 2430 return pfn; 2431 } 2432 2433 kvm_pfn_t __gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn, 2434 bool atomic, bool *async, bool write_fault, 2435 bool *writable, hva_t *hva) 2436 { 2437 unsigned long addr = __gfn_to_hva_many(slot, gfn, NULL, write_fault); 2438 2439 if (hva) 2440 *hva = addr; 2441 2442 if (addr == KVM_HVA_ERR_RO_BAD) { 2443 if (writable) 2444 *writable = false; 2445 return KVM_PFN_ERR_RO_FAULT; 2446 } 2447 2448 if (kvm_is_error_hva(addr)) { 2449 if (writable) 2450 *writable = false; 2451 return KVM_PFN_NOSLOT; 2452 } 2453 2454 /* Do not map writable pfn in the readonly memslot. */ 2455 if (writable && memslot_is_readonly(slot)) { 2456 *writable = false; 2457 writable = NULL; 2458 } 2459 2460 return hva_to_pfn(addr, atomic, async, write_fault, 2461 writable); 2462 } 2463 EXPORT_SYMBOL_GPL(__gfn_to_pfn_memslot); 2464 2465 kvm_pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault, 2466 bool *writable) 2467 { 2468 return __gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn, false, NULL, 2469 write_fault, writable, NULL); 2470 } 2471 EXPORT_SYMBOL_GPL(gfn_to_pfn_prot); 2472 2473 kvm_pfn_t gfn_to_pfn_memslot(struct kvm_memory_slot *slot, gfn_t gfn) 2474 { 2475 return __gfn_to_pfn_memslot(slot, gfn, false, NULL, true, NULL, NULL); 2476 } 2477 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot); 2478 2479 kvm_pfn_t gfn_to_pfn_memslot_atomic(struct kvm_memory_slot *slot, gfn_t gfn) 2480 { 2481 return __gfn_to_pfn_memslot(slot, gfn, true, NULL, true, NULL, NULL); 2482 } 2483 EXPORT_SYMBOL_GPL(gfn_to_pfn_memslot_atomic); 2484 2485 kvm_pfn_t kvm_vcpu_gfn_to_pfn_atomic(struct kvm_vcpu *vcpu, gfn_t gfn) 2486 { 2487 return gfn_to_pfn_memslot_atomic(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn); 2488 } 2489 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn_atomic); 2490 2491 kvm_pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn) 2492 { 2493 return gfn_to_pfn_memslot(gfn_to_memslot(kvm, gfn), gfn); 2494 } 2495 EXPORT_SYMBOL_GPL(gfn_to_pfn); 2496 2497 kvm_pfn_t kvm_vcpu_gfn_to_pfn(struct kvm_vcpu *vcpu, gfn_t gfn) 2498 { 2499 return gfn_to_pfn_memslot(kvm_vcpu_gfn_to_memslot(vcpu, gfn), gfn); 2500 } 2501 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_pfn); 2502 2503 int gfn_to_page_many_atomic(struct kvm_memory_slot *slot, gfn_t gfn, 2504 struct page **pages, int nr_pages) 2505 { 2506 unsigned long addr; 2507 gfn_t entry = 0; 2508 2509 addr = gfn_to_hva_many(slot, gfn, &entry); 2510 if (kvm_is_error_hva(addr)) 2511 return -1; 2512 2513 if (entry < nr_pages) 2514 return 0; 2515 2516 return get_user_pages_fast_only(addr, nr_pages, FOLL_WRITE, pages); 2517 } 2518 EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic); 2519 2520 static struct page *kvm_pfn_to_page(kvm_pfn_t pfn) 2521 { 2522 if (is_error_noslot_pfn(pfn)) 2523 return KVM_ERR_PTR_BAD_PAGE; 2524 2525 if (kvm_is_reserved_pfn(pfn)) { 2526 WARN_ON(1); 2527 return KVM_ERR_PTR_BAD_PAGE; 2528 } 2529 2530 return pfn_to_page(pfn); 2531 } 2532 2533 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn) 2534 { 2535 kvm_pfn_t pfn; 2536 2537 pfn = gfn_to_pfn(kvm, gfn); 2538 2539 return kvm_pfn_to_page(pfn); 2540 } 2541 EXPORT_SYMBOL_GPL(gfn_to_page); 2542 2543 void kvm_release_pfn(kvm_pfn_t pfn, bool dirty, struct gfn_to_pfn_cache *cache) 2544 { 2545 if (pfn == 0) 2546 return; 2547 2548 if (cache) 2549 cache->pfn = cache->gfn = 0; 2550 2551 if (dirty) 2552 kvm_release_pfn_dirty(pfn); 2553 else 2554 kvm_release_pfn_clean(pfn); 2555 } 2556 2557 static void kvm_cache_gfn_to_pfn(struct kvm_memory_slot *slot, gfn_t gfn, 2558 struct gfn_to_pfn_cache *cache, u64 gen) 2559 { 2560 kvm_release_pfn(cache->pfn, cache->dirty, cache); 2561 2562 cache->pfn = gfn_to_pfn_memslot(slot, gfn); 2563 cache->gfn = gfn; 2564 cache->dirty = false; 2565 cache->generation = gen; 2566 } 2567 2568 static int __kvm_map_gfn(struct kvm_memslots *slots, gfn_t gfn, 2569 struct kvm_host_map *map, 2570 struct gfn_to_pfn_cache *cache, 2571 bool atomic) 2572 { 2573 kvm_pfn_t pfn; 2574 void *hva = NULL; 2575 struct page *page = KVM_UNMAPPED_PAGE; 2576 struct kvm_memory_slot *slot = __gfn_to_memslot(slots, gfn); 2577 u64 gen = slots->generation; 2578 2579 if (!map) 2580 return -EINVAL; 2581 2582 if (cache) { 2583 if (!cache->pfn || cache->gfn != gfn || 2584 cache->generation != gen) { 2585 if (atomic) 2586 return -EAGAIN; 2587 kvm_cache_gfn_to_pfn(slot, gfn, cache, gen); 2588 } 2589 pfn = cache->pfn; 2590 } else { 2591 if (atomic) 2592 return -EAGAIN; 2593 pfn = gfn_to_pfn_memslot(slot, gfn); 2594 } 2595 if (is_error_noslot_pfn(pfn)) 2596 return -EINVAL; 2597 2598 if (pfn_valid(pfn)) { 2599 page = pfn_to_page(pfn); 2600 if (atomic) 2601 hva = kmap_atomic(page); 2602 else 2603 hva = kmap(page); 2604 #ifdef CONFIG_HAS_IOMEM 2605 } else if (!atomic) { 2606 hva = memremap(pfn_to_hpa(pfn), PAGE_SIZE, MEMREMAP_WB); 2607 } else { 2608 return -EINVAL; 2609 #endif 2610 } 2611 2612 if (!hva) 2613 return -EFAULT; 2614 2615 map->page = page; 2616 map->hva = hva; 2617 map->pfn = pfn; 2618 map->gfn = gfn; 2619 2620 return 0; 2621 } 2622 2623 int kvm_map_gfn(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map, 2624 struct gfn_to_pfn_cache *cache, bool atomic) 2625 { 2626 return __kvm_map_gfn(kvm_memslots(vcpu->kvm), gfn, map, 2627 cache, atomic); 2628 } 2629 EXPORT_SYMBOL_GPL(kvm_map_gfn); 2630 2631 int kvm_vcpu_map(struct kvm_vcpu *vcpu, gfn_t gfn, struct kvm_host_map *map) 2632 { 2633 return __kvm_map_gfn(kvm_vcpu_memslots(vcpu), gfn, map, 2634 NULL, false); 2635 } 2636 EXPORT_SYMBOL_GPL(kvm_vcpu_map); 2637 2638 static void __kvm_unmap_gfn(struct kvm *kvm, 2639 struct kvm_memory_slot *memslot, 2640 struct kvm_host_map *map, 2641 struct gfn_to_pfn_cache *cache, 2642 bool dirty, bool atomic) 2643 { 2644 if (!map) 2645 return; 2646 2647 if (!map->hva) 2648 return; 2649 2650 if (map->page != KVM_UNMAPPED_PAGE) { 2651 if (atomic) 2652 kunmap_atomic(map->hva); 2653 else 2654 kunmap(map->page); 2655 } 2656 #ifdef CONFIG_HAS_IOMEM 2657 else if (!atomic) 2658 memunmap(map->hva); 2659 else 2660 WARN_ONCE(1, "Unexpected unmapping in atomic context"); 2661 #endif 2662 2663 if (dirty) 2664 mark_page_dirty_in_slot(kvm, memslot, map->gfn); 2665 2666 if (cache) 2667 cache->dirty |= dirty; 2668 else 2669 kvm_release_pfn(map->pfn, dirty, NULL); 2670 2671 map->hva = NULL; 2672 map->page = NULL; 2673 } 2674 2675 int kvm_unmap_gfn(struct kvm_vcpu *vcpu, struct kvm_host_map *map, 2676 struct gfn_to_pfn_cache *cache, bool dirty, bool atomic) 2677 { 2678 __kvm_unmap_gfn(vcpu->kvm, gfn_to_memslot(vcpu->kvm, map->gfn), map, 2679 cache, dirty, atomic); 2680 return 0; 2681 } 2682 EXPORT_SYMBOL_GPL(kvm_unmap_gfn); 2683 2684 void kvm_vcpu_unmap(struct kvm_vcpu *vcpu, struct kvm_host_map *map, bool dirty) 2685 { 2686 __kvm_unmap_gfn(vcpu->kvm, kvm_vcpu_gfn_to_memslot(vcpu, map->gfn), 2687 map, NULL, dirty, false); 2688 } 2689 EXPORT_SYMBOL_GPL(kvm_vcpu_unmap); 2690 2691 struct page *kvm_vcpu_gfn_to_page(struct kvm_vcpu *vcpu, gfn_t gfn) 2692 { 2693 kvm_pfn_t pfn; 2694 2695 pfn = kvm_vcpu_gfn_to_pfn(vcpu, gfn); 2696 2697 return kvm_pfn_to_page(pfn); 2698 } 2699 EXPORT_SYMBOL_GPL(kvm_vcpu_gfn_to_page); 2700 2701 void kvm_release_page_clean(struct page *page) 2702 { 2703 WARN_ON(is_error_page(page)); 2704 2705 kvm_release_pfn_clean(page_to_pfn(page)); 2706 } 2707 EXPORT_SYMBOL_GPL(kvm_release_page_clean); 2708 2709 void kvm_release_pfn_clean(kvm_pfn_t pfn) 2710 { 2711 if (!is_error_noslot_pfn(pfn) && !kvm_is_reserved_pfn(pfn)) 2712 put_page(pfn_to_page(pfn)); 2713 } 2714 EXPORT_SYMBOL_GPL(kvm_release_pfn_clean); 2715 2716 void kvm_release_page_dirty(struct page *page) 2717 { 2718 WARN_ON(is_error_page(page)); 2719 2720 kvm_release_pfn_dirty(page_to_pfn(page)); 2721 } 2722 EXPORT_SYMBOL_GPL(kvm_release_page_dirty); 2723 2724 void kvm_release_pfn_dirty(kvm_pfn_t pfn) 2725 { 2726 kvm_set_pfn_dirty(pfn); 2727 kvm_release_pfn_clean(pfn); 2728 } 2729 EXPORT_SYMBOL_GPL(kvm_release_pfn_dirty); 2730 2731 void kvm_set_pfn_dirty(kvm_pfn_t pfn) 2732 { 2733 if (!kvm_is_reserved_pfn(pfn) && !kvm_is_zone_device_pfn(pfn)) 2734 SetPageDirty(pfn_to_page(pfn)); 2735 } 2736 EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty); 2737 2738 void kvm_set_pfn_accessed(kvm_pfn_t pfn) 2739 { 2740 if (!kvm_is_reserved_pfn(pfn) && !kvm_is_zone_device_pfn(pfn)) 2741 mark_page_accessed(pfn_to_page(pfn)); 2742 } 2743 EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed); 2744 2745 static int next_segment(unsigned long len, int offset) 2746 { 2747 if (len > PAGE_SIZE - offset) 2748 return PAGE_SIZE - offset; 2749 else 2750 return len; 2751 } 2752 2753 static int __kvm_read_guest_page(struct kvm_memory_slot *slot, gfn_t gfn, 2754 void *data, int offset, int len) 2755 { 2756 int r; 2757 unsigned long addr; 2758 2759 addr = gfn_to_hva_memslot_prot(slot, gfn, NULL); 2760 if (kvm_is_error_hva(addr)) 2761 return -EFAULT; 2762 r = __copy_from_user(data, (void __user *)addr + offset, len); 2763 if (r) 2764 return -EFAULT; 2765 return 0; 2766 } 2767 2768 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset, 2769 int len) 2770 { 2771 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn); 2772 2773 return __kvm_read_guest_page(slot, gfn, data, offset, len); 2774 } 2775 EXPORT_SYMBOL_GPL(kvm_read_guest_page); 2776 2777 int kvm_vcpu_read_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, void *data, 2778 int offset, int len) 2779 { 2780 struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 2781 2782 return __kvm_read_guest_page(slot, gfn, data, offset, len); 2783 } 2784 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_page); 2785 2786 int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len) 2787 { 2788 gfn_t gfn = gpa >> PAGE_SHIFT; 2789 int seg; 2790 int offset = offset_in_page(gpa); 2791 int ret; 2792 2793 while ((seg = next_segment(len, offset)) != 0) { 2794 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg); 2795 if (ret < 0) 2796 return ret; 2797 offset = 0; 2798 len -= seg; 2799 data += seg; 2800 ++gfn; 2801 } 2802 return 0; 2803 } 2804 EXPORT_SYMBOL_GPL(kvm_read_guest); 2805 2806 int kvm_vcpu_read_guest(struct kvm_vcpu *vcpu, gpa_t gpa, void *data, unsigned long len) 2807 { 2808 gfn_t gfn = gpa >> PAGE_SHIFT; 2809 int seg; 2810 int offset = offset_in_page(gpa); 2811 int ret; 2812 2813 while ((seg = next_segment(len, offset)) != 0) { 2814 ret = kvm_vcpu_read_guest_page(vcpu, gfn, data, offset, seg); 2815 if (ret < 0) 2816 return ret; 2817 offset = 0; 2818 len -= seg; 2819 data += seg; 2820 ++gfn; 2821 } 2822 return 0; 2823 } 2824 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest); 2825 2826 static int __kvm_read_guest_atomic(struct kvm_memory_slot *slot, gfn_t gfn, 2827 void *data, int offset, unsigned long len) 2828 { 2829 int r; 2830 unsigned long addr; 2831 2832 addr = gfn_to_hva_memslot_prot(slot, gfn, NULL); 2833 if (kvm_is_error_hva(addr)) 2834 return -EFAULT; 2835 pagefault_disable(); 2836 r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len); 2837 pagefault_enable(); 2838 if (r) 2839 return -EFAULT; 2840 return 0; 2841 } 2842 2843 int kvm_vcpu_read_guest_atomic(struct kvm_vcpu *vcpu, gpa_t gpa, 2844 void *data, unsigned long len) 2845 { 2846 gfn_t gfn = gpa >> PAGE_SHIFT; 2847 struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 2848 int offset = offset_in_page(gpa); 2849 2850 return __kvm_read_guest_atomic(slot, gfn, data, offset, len); 2851 } 2852 EXPORT_SYMBOL_GPL(kvm_vcpu_read_guest_atomic); 2853 2854 static int __kvm_write_guest_page(struct kvm *kvm, 2855 struct kvm_memory_slot *memslot, gfn_t gfn, 2856 const void *data, int offset, int len) 2857 { 2858 int r; 2859 unsigned long addr; 2860 2861 addr = gfn_to_hva_memslot(memslot, gfn); 2862 if (kvm_is_error_hva(addr)) 2863 return -EFAULT; 2864 r = __copy_to_user((void __user *)addr + offset, data, len); 2865 if (r) 2866 return -EFAULT; 2867 mark_page_dirty_in_slot(kvm, memslot, gfn); 2868 return 0; 2869 } 2870 2871 int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn, 2872 const void *data, int offset, int len) 2873 { 2874 struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn); 2875 2876 return __kvm_write_guest_page(kvm, slot, gfn, data, offset, len); 2877 } 2878 EXPORT_SYMBOL_GPL(kvm_write_guest_page); 2879 2880 int kvm_vcpu_write_guest_page(struct kvm_vcpu *vcpu, gfn_t gfn, 2881 const void *data, int offset, int len) 2882 { 2883 struct kvm_memory_slot *slot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 2884 2885 return __kvm_write_guest_page(vcpu->kvm, slot, gfn, data, offset, len); 2886 } 2887 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest_page); 2888 2889 int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data, 2890 unsigned long len) 2891 { 2892 gfn_t gfn = gpa >> PAGE_SHIFT; 2893 int seg; 2894 int offset = offset_in_page(gpa); 2895 int ret; 2896 2897 while ((seg = next_segment(len, offset)) != 0) { 2898 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg); 2899 if (ret < 0) 2900 return ret; 2901 offset = 0; 2902 len -= seg; 2903 data += seg; 2904 ++gfn; 2905 } 2906 return 0; 2907 } 2908 EXPORT_SYMBOL_GPL(kvm_write_guest); 2909 2910 int kvm_vcpu_write_guest(struct kvm_vcpu *vcpu, gpa_t gpa, const void *data, 2911 unsigned long len) 2912 { 2913 gfn_t gfn = gpa >> PAGE_SHIFT; 2914 int seg; 2915 int offset = offset_in_page(gpa); 2916 int ret; 2917 2918 while ((seg = next_segment(len, offset)) != 0) { 2919 ret = kvm_vcpu_write_guest_page(vcpu, gfn, data, offset, seg); 2920 if (ret < 0) 2921 return ret; 2922 offset = 0; 2923 len -= seg; 2924 data += seg; 2925 ++gfn; 2926 } 2927 return 0; 2928 } 2929 EXPORT_SYMBOL_GPL(kvm_vcpu_write_guest); 2930 2931 static int __kvm_gfn_to_hva_cache_init(struct kvm_memslots *slots, 2932 struct gfn_to_hva_cache *ghc, 2933 gpa_t gpa, unsigned long len) 2934 { 2935 int offset = offset_in_page(gpa); 2936 gfn_t start_gfn = gpa >> PAGE_SHIFT; 2937 gfn_t end_gfn = (gpa + len - 1) >> PAGE_SHIFT; 2938 gfn_t nr_pages_needed = end_gfn - start_gfn + 1; 2939 gfn_t nr_pages_avail; 2940 2941 /* Update ghc->generation before performing any error checks. */ 2942 ghc->generation = slots->generation; 2943 2944 if (start_gfn > end_gfn) { 2945 ghc->hva = KVM_HVA_ERR_BAD; 2946 return -EINVAL; 2947 } 2948 2949 /* 2950 * If the requested region crosses two memslots, we still 2951 * verify that the entire region is valid here. 2952 */ 2953 for ( ; start_gfn <= end_gfn; start_gfn += nr_pages_avail) { 2954 ghc->memslot = __gfn_to_memslot(slots, start_gfn); 2955 ghc->hva = gfn_to_hva_many(ghc->memslot, start_gfn, 2956 &nr_pages_avail); 2957 if (kvm_is_error_hva(ghc->hva)) 2958 return -EFAULT; 2959 } 2960 2961 /* Use the slow path for cross page reads and writes. */ 2962 if (nr_pages_needed == 1) 2963 ghc->hva += offset; 2964 else 2965 ghc->memslot = NULL; 2966 2967 ghc->gpa = gpa; 2968 ghc->len = len; 2969 return 0; 2970 } 2971 2972 int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc, 2973 gpa_t gpa, unsigned long len) 2974 { 2975 struct kvm_memslots *slots = kvm_memslots(kvm); 2976 return __kvm_gfn_to_hva_cache_init(slots, ghc, gpa, len); 2977 } 2978 EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init); 2979 2980 int kvm_write_guest_offset_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc, 2981 void *data, unsigned int offset, 2982 unsigned long len) 2983 { 2984 struct kvm_memslots *slots = kvm_memslots(kvm); 2985 int r; 2986 gpa_t gpa = ghc->gpa + offset; 2987 2988 BUG_ON(len + offset > ghc->len); 2989 2990 if (slots->generation != ghc->generation) { 2991 if (__kvm_gfn_to_hva_cache_init(slots, ghc, ghc->gpa, ghc->len)) 2992 return -EFAULT; 2993 } 2994 2995 if (kvm_is_error_hva(ghc->hva)) 2996 return -EFAULT; 2997 2998 if (unlikely(!ghc->memslot)) 2999 return kvm_write_guest(kvm, gpa, data, len); 3000 3001 r = __copy_to_user((void __user *)ghc->hva + offset, data, len); 3002 if (r) 3003 return -EFAULT; 3004 mark_page_dirty_in_slot(kvm, ghc->memslot, gpa >> PAGE_SHIFT); 3005 3006 return 0; 3007 } 3008 EXPORT_SYMBOL_GPL(kvm_write_guest_offset_cached); 3009 3010 int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc, 3011 void *data, unsigned long len) 3012 { 3013 return kvm_write_guest_offset_cached(kvm, ghc, data, 0, len); 3014 } 3015 EXPORT_SYMBOL_GPL(kvm_write_guest_cached); 3016 3017 int kvm_read_guest_offset_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc, 3018 void *data, unsigned int offset, 3019 unsigned long len) 3020 { 3021 struct kvm_memslots *slots = kvm_memslots(kvm); 3022 int r; 3023 gpa_t gpa = ghc->gpa + offset; 3024 3025 BUG_ON(len + offset > ghc->len); 3026 3027 if (slots->generation != ghc->generation) { 3028 if (__kvm_gfn_to_hva_cache_init(slots, ghc, ghc->gpa, ghc->len)) 3029 return -EFAULT; 3030 } 3031 3032 if (kvm_is_error_hva(ghc->hva)) 3033 return -EFAULT; 3034 3035 if (unlikely(!ghc->memslot)) 3036 return kvm_read_guest(kvm, gpa, data, len); 3037 3038 r = __copy_from_user(data, (void __user *)ghc->hva + offset, len); 3039 if (r) 3040 return -EFAULT; 3041 3042 return 0; 3043 } 3044 EXPORT_SYMBOL_GPL(kvm_read_guest_offset_cached); 3045 3046 int kvm_read_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc, 3047 void *data, unsigned long len) 3048 { 3049 return kvm_read_guest_offset_cached(kvm, ghc, data, 0, len); 3050 } 3051 EXPORT_SYMBOL_GPL(kvm_read_guest_cached); 3052 3053 int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len) 3054 { 3055 const void *zero_page = (const void *) __va(page_to_phys(ZERO_PAGE(0))); 3056 gfn_t gfn = gpa >> PAGE_SHIFT; 3057 int seg; 3058 int offset = offset_in_page(gpa); 3059 int ret; 3060 3061 while ((seg = next_segment(len, offset)) != 0) { 3062 ret = kvm_write_guest_page(kvm, gfn, zero_page, offset, len); 3063 if (ret < 0) 3064 return ret; 3065 offset = 0; 3066 len -= seg; 3067 ++gfn; 3068 } 3069 return 0; 3070 } 3071 EXPORT_SYMBOL_GPL(kvm_clear_guest); 3072 3073 void mark_page_dirty_in_slot(struct kvm *kvm, 3074 struct kvm_memory_slot *memslot, 3075 gfn_t gfn) 3076 { 3077 if (memslot && kvm_slot_dirty_track_enabled(memslot)) { 3078 unsigned long rel_gfn = gfn - memslot->base_gfn; 3079 u32 slot = (memslot->as_id << 16) | memslot->id; 3080 3081 if (kvm->dirty_ring_size) 3082 kvm_dirty_ring_push(kvm_dirty_ring_get(kvm), 3083 slot, rel_gfn); 3084 else 3085 set_bit_le(rel_gfn, memslot->dirty_bitmap); 3086 } 3087 } 3088 EXPORT_SYMBOL_GPL(mark_page_dirty_in_slot); 3089 3090 void mark_page_dirty(struct kvm *kvm, gfn_t gfn) 3091 { 3092 struct kvm_memory_slot *memslot; 3093 3094 memslot = gfn_to_memslot(kvm, gfn); 3095 mark_page_dirty_in_slot(kvm, memslot, gfn); 3096 } 3097 EXPORT_SYMBOL_GPL(mark_page_dirty); 3098 3099 void kvm_vcpu_mark_page_dirty(struct kvm_vcpu *vcpu, gfn_t gfn) 3100 { 3101 struct kvm_memory_slot *memslot; 3102 3103 memslot = kvm_vcpu_gfn_to_memslot(vcpu, gfn); 3104 mark_page_dirty_in_slot(vcpu->kvm, memslot, gfn); 3105 } 3106 EXPORT_SYMBOL_GPL(kvm_vcpu_mark_page_dirty); 3107 3108 void kvm_sigset_activate(struct kvm_vcpu *vcpu) 3109 { 3110 if (!vcpu->sigset_active) 3111 return; 3112 3113 /* 3114 * This does a lockless modification of ->real_blocked, which is fine 3115 * because, only current can change ->real_blocked and all readers of 3116 * ->real_blocked don't care as long ->real_blocked is always a subset 3117 * of ->blocked. 3118 */ 3119 sigprocmask(SIG_SETMASK, &vcpu->sigset, ¤t->real_blocked); 3120 } 3121 3122 void kvm_sigset_deactivate(struct kvm_vcpu *vcpu) 3123 { 3124 if (!vcpu->sigset_active) 3125 return; 3126 3127 sigprocmask(SIG_SETMASK, ¤t->real_blocked, NULL); 3128 sigemptyset(¤t->real_blocked); 3129 } 3130 3131 static void grow_halt_poll_ns(struct kvm_vcpu *vcpu) 3132 { 3133 unsigned int old, val, grow, grow_start; 3134 3135 old = val = vcpu->halt_poll_ns; 3136 grow_start = READ_ONCE(halt_poll_ns_grow_start); 3137 grow = READ_ONCE(halt_poll_ns_grow); 3138 if (!grow) 3139 goto out; 3140 3141 val *= grow; 3142 if (val < grow_start) 3143 val = grow_start; 3144 3145 if (val > vcpu->kvm->max_halt_poll_ns) 3146 val = vcpu->kvm->max_halt_poll_ns; 3147 3148 vcpu->halt_poll_ns = val; 3149 out: 3150 trace_kvm_halt_poll_ns_grow(vcpu->vcpu_id, val, old); 3151 } 3152 3153 static void shrink_halt_poll_ns(struct kvm_vcpu *vcpu) 3154 { 3155 unsigned int old, val, shrink, grow_start; 3156 3157 old = val = vcpu->halt_poll_ns; 3158 shrink = READ_ONCE(halt_poll_ns_shrink); 3159 grow_start = READ_ONCE(halt_poll_ns_grow_start); 3160 if (shrink == 0) 3161 val = 0; 3162 else 3163 val /= shrink; 3164 3165 if (val < grow_start) 3166 val = 0; 3167 3168 vcpu->halt_poll_ns = val; 3169 trace_kvm_halt_poll_ns_shrink(vcpu->vcpu_id, val, old); 3170 } 3171 3172 static int kvm_vcpu_check_block(struct kvm_vcpu *vcpu) 3173 { 3174 int ret = -EINTR; 3175 int idx = srcu_read_lock(&vcpu->kvm->srcu); 3176 3177 if (kvm_arch_vcpu_runnable(vcpu)) { 3178 kvm_make_request(KVM_REQ_UNHALT, vcpu); 3179 goto out; 3180 } 3181 if (kvm_cpu_has_pending_timer(vcpu)) 3182 goto out; 3183 if (signal_pending(current)) 3184 goto out; 3185 if (kvm_check_request(KVM_REQ_UNBLOCK, vcpu)) 3186 goto out; 3187 3188 ret = 0; 3189 out: 3190 srcu_read_unlock(&vcpu->kvm->srcu, idx); 3191 return ret; 3192 } 3193 3194 static inline void 3195 update_halt_poll_stats(struct kvm_vcpu *vcpu, u64 poll_ns, bool waited) 3196 { 3197 if (waited) 3198 vcpu->stat.generic.halt_poll_fail_ns += poll_ns; 3199 else 3200 vcpu->stat.generic.halt_poll_success_ns += poll_ns; 3201 } 3202 3203 /* 3204 * The vCPU has executed a HLT instruction with in-kernel mode enabled. 3205 */ 3206 void kvm_vcpu_block(struct kvm_vcpu *vcpu) 3207 { 3208 ktime_t start, cur, poll_end; 3209 bool waited = false; 3210 u64 block_ns; 3211 3212 kvm_arch_vcpu_blocking(vcpu); 3213 3214 start = cur = poll_end = ktime_get(); 3215 if (vcpu->halt_poll_ns && !kvm_arch_no_poll(vcpu)) { 3216 ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns); 3217 3218 ++vcpu->stat.generic.halt_attempted_poll; 3219 do { 3220 /* 3221 * This sets KVM_REQ_UNHALT if an interrupt 3222 * arrives. 3223 */ 3224 if (kvm_vcpu_check_block(vcpu) < 0) { 3225 ++vcpu->stat.generic.halt_successful_poll; 3226 if (!vcpu_valid_wakeup(vcpu)) 3227 ++vcpu->stat.generic.halt_poll_invalid; 3228 3229 KVM_STATS_LOG_HIST_UPDATE( 3230 vcpu->stat.generic.halt_poll_success_hist, 3231 ktime_to_ns(ktime_get()) - 3232 ktime_to_ns(start)); 3233 goto out; 3234 } 3235 cpu_relax(); 3236 poll_end = cur = ktime_get(); 3237 } while (kvm_vcpu_can_poll(cur, stop)); 3238 3239 KVM_STATS_LOG_HIST_UPDATE( 3240 vcpu->stat.generic.halt_poll_fail_hist, 3241 ktime_to_ns(ktime_get()) - ktime_to_ns(start)); 3242 } 3243 3244 3245 prepare_to_rcuwait(&vcpu->wait); 3246 for (;;) { 3247 set_current_state(TASK_INTERRUPTIBLE); 3248 3249 if (kvm_vcpu_check_block(vcpu) < 0) 3250 break; 3251 3252 waited = true; 3253 schedule(); 3254 } 3255 finish_rcuwait(&vcpu->wait); 3256 cur = ktime_get(); 3257 if (waited) { 3258 vcpu->stat.generic.halt_wait_ns += 3259 ktime_to_ns(cur) - ktime_to_ns(poll_end); 3260 KVM_STATS_LOG_HIST_UPDATE(vcpu->stat.generic.halt_wait_hist, 3261 ktime_to_ns(cur) - ktime_to_ns(poll_end)); 3262 } 3263 out: 3264 kvm_arch_vcpu_unblocking(vcpu); 3265 block_ns = ktime_to_ns(cur) - ktime_to_ns(start); 3266 3267 update_halt_poll_stats( 3268 vcpu, ktime_to_ns(ktime_sub(poll_end, start)), waited); 3269 3270 if (!kvm_arch_no_poll(vcpu)) { 3271 if (!vcpu_valid_wakeup(vcpu)) { 3272 shrink_halt_poll_ns(vcpu); 3273 } else if (vcpu->kvm->max_halt_poll_ns) { 3274 if (block_ns <= vcpu->halt_poll_ns) 3275 ; 3276 /* we had a long block, shrink polling */ 3277 else if (vcpu->halt_poll_ns && 3278 block_ns > vcpu->kvm->max_halt_poll_ns) 3279 shrink_halt_poll_ns(vcpu); 3280 /* we had a short halt and our poll time is too small */ 3281 else if (vcpu->halt_poll_ns < vcpu->kvm->max_halt_poll_ns && 3282 block_ns < vcpu->kvm->max_halt_poll_ns) 3283 grow_halt_poll_ns(vcpu); 3284 } else { 3285 vcpu->halt_poll_ns = 0; 3286 } 3287 } 3288 3289 trace_kvm_vcpu_wakeup(block_ns, waited, vcpu_valid_wakeup(vcpu)); 3290 kvm_arch_vcpu_block_finish(vcpu); 3291 } 3292 EXPORT_SYMBOL_GPL(kvm_vcpu_block); 3293 3294 bool kvm_vcpu_wake_up(struct kvm_vcpu *vcpu) 3295 { 3296 struct rcuwait *waitp; 3297 3298 waitp = kvm_arch_vcpu_get_wait(vcpu); 3299 if (rcuwait_wake_up(waitp)) { 3300 WRITE_ONCE(vcpu->ready, true); 3301 ++vcpu->stat.generic.halt_wakeup; 3302 return true; 3303 } 3304 3305 return false; 3306 } 3307 EXPORT_SYMBOL_GPL(kvm_vcpu_wake_up); 3308 3309 #ifndef CONFIG_S390 3310 /* 3311 * Kick a sleeping VCPU, or a guest VCPU in guest mode, into host kernel mode. 3312 */ 3313 void kvm_vcpu_kick(struct kvm_vcpu *vcpu) 3314 { 3315 int me, cpu; 3316 3317 if (kvm_vcpu_wake_up(vcpu)) 3318 return; 3319 3320 /* 3321 * Note, the vCPU could get migrated to a different pCPU at any point 3322 * after kvm_arch_vcpu_should_kick(), which could result in sending an 3323 * IPI to the previous pCPU. But, that's ok because the purpose of the 3324 * IPI is to force the vCPU to leave IN_GUEST_MODE, and migrating the 3325 * vCPU also requires it to leave IN_GUEST_MODE. 3326 */ 3327 me = get_cpu(); 3328 if (kvm_arch_vcpu_should_kick(vcpu)) { 3329 cpu = READ_ONCE(vcpu->cpu); 3330 if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) 3331 smp_send_reschedule(cpu); 3332 } 3333 put_cpu(); 3334 } 3335 EXPORT_SYMBOL_GPL(kvm_vcpu_kick); 3336 #endif /* !CONFIG_S390 */ 3337 3338 int kvm_vcpu_yield_to(struct kvm_vcpu *target) 3339 { 3340 struct pid *pid; 3341 struct task_struct *task = NULL; 3342 int ret = 0; 3343 3344 rcu_read_lock(); 3345 pid = rcu_dereference(target->pid); 3346 if (pid) 3347 task = get_pid_task(pid, PIDTYPE_PID); 3348 rcu_read_unlock(); 3349 if (!task) 3350 return ret; 3351 ret = yield_to(task, 1); 3352 put_task_struct(task); 3353 3354 return ret; 3355 } 3356 EXPORT_SYMBOL_GPL(kvm_vcpu_yield_to); 3357 3358 /* 3359 * Helper that checks whether a VCPU is eligible for directed yield. 3360 * Most eligible candidate to yield is decided by following heuristics: 3361 * 3362 * (a) VCPU which has not done pl-exit or cpu relax intercepted recently 3363 * (preempted lock holder), indicated by @in_spin_loop. 3364 * Set at the beginning and cleared at the end of interception/PLE handler. 3365 * 3366 * (b) VCPU which has done pl-exit/ cpu relax intercepted but did not get 3367 * chance last time (mostly it has become eligible now since we have probably 3368 * yielded to lockholder in last iteration. This is done by toggling 3369 * @dy_eligible each time a VCPU checked for eligibility.) 3370 * 3371 * Yielding to a recently pl-exited/cpu relax intercepted VCPU before yielding 3372 * to preempted lock-holder could result in wrong VCPU selection and CPU 3373 * burning. Giving priority for a potential lock-holder increases lock 3374 * progress. 3375 * 3376 * Since algorithm is based on heuristics, accessing another VCPU data without 3377 * locking does not harm. It may result in trying to yield to same VCPU, fail 3378 * and continue with next VCPU and so on. 3379 */ 3380 static bool kvm_vcpu_eligible_for_directed_yield(struct kvm_vcpu *vcpu) 3381 { 3382 #ifdef CONFIG_HAVE_KVM_CPU_RELAX_INTERCEPT 3383 bool eligible; 3384 3385 eligible = !vcpu->spin_loop.in_spin_loop || 3386 vcpu->spin_loop.dy_eligible; 3387 3388 if (vcpu->spin_loop.in_spin_loop) 3389 kvm_vcpu_set_dy_eligible(vcpu, !vcpu->spin_loop.dy_eligible); 3390 3391 return eligible; 3392 #else 3393 return true; 3394 #endif 3395 } 3396 3397 /* 3398 * Unlike kvm_arch_vcpu_runnable, this function is called outside 3399 * a vcpu_load/vcpu_put pair. However, for most architectures 3400 * kvm_arch_vcpu_runnable does not require vcpu_load. 3401 */ 3402 bool __weak kvm_arch_dy_runnable(struct kvm_vcpu *vcpu) 3403 { 3404 return kvm_arch_vcpu_runnable(vcpu); 3405 } 3406 3407 static bool vcpu_dy_runnable(struct kvm_vcpu *vcpu) 3408 { 3409 if (kvm_arch_dy_runnable(vcpu)) 3410 return true; 3411 3412 #ifdef CONFIG_KVM_ASYNC_PF 3413 if (!list_empty_careful(&vcpu->async_pf.done)) 3414 return true; 3415 #endif 3416 3417 return false; 3418 } 3419 3420 bool __weak kvm_arch_dy_has_pending_interrupt(struct kvm_vcpu *vcpu) 3421 { 3422 return false; 3423 } 3424 3425 void kvm_vcpu_on_spin(struct kvm_vcpu *me, bool yield_to_kernel_mode) 3426 { 3427 struct kvm *kvm = me->kvm; 3428 struct kvm_vcpu *vcpu; 3429 int last_boosted_vcpu = me->kvm->last_boosted_vcpu; 3430 int yielded = 0; 3431 int try = 3; 3432 int pass; 3433 int i; 3434 3435 kvm_vcpu_set_in_spin_loop(me, true); 3436 /* 3437 * We boost the priority of a VCPU that is runnable but not 3438 * currently running, because it got preempted by something 3439 * else and called schedule in __vcpu_run. Hopefully that 3440 * VCPU is holding the lock that we need and will release it. 3441 * We approximate round-robin by starting at the last boosted VCPU. 3442 */ 3443 for (pass = 0; pass < 2 && !yielded && try; pass++) { 3444 kvm_for_each_vcpu(i, vcpu, kvm) { 3445 if (!pass && i <= last_boosted_vcpu) { 3446 i = last_boosted_vcpu; 3447 continue; 3448 } else if (pass && i > last_boosted_vcpu) 3449 break; 3450 if (!READ_ONCE(vcpu->ready)) 3451 continue; 3452 if (vcpu == me) 3453 continue; 3454 if (rcuwait_active(&vcpu->wait) && 3455 !vcpu_dy_runnable(vcpu)) 3456 continue; 3457 if (READ_ONCE(vcpu->preempted) && yield_to_kernel_mode && 3458 !kvm_arch_dy_has_pending_interrupt(vcpu) && 3459 !kvm_arch_vcpu_in_kernel(vcpu)) 3460 continue; 3461 if (!kvm_vcpu_eligible_for_directed_yield(vcpu)) 3462 continue; 3463 3464 yielded = kvm_vcpu_yield_to(vcpu); 3465 if (yielded > 0) { 3466 kvm->last_boosted_vcpu = i; 3467 break; 3468 } else if (yielded < 0) { 3469 try--; 3470 if (!try) 3471 break; 3472 } 3473 } 3474 } 3475 kvm_vcpu_set_in_spin_loop(me, false); 3476 3477 /* Ensure vcpu is not eligible during next spinloop */ 3478 kvm_vcpu_set_dy_eligible(me, false); 3479 } 3480 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin); 3481 3482 static bool kvm_page_in_dirty_ring(struct kvm *kvm, unsigned long pgoff) 3483 { 3484 #if KVM_DIRTY_LOG_PAGE_OFFSET > 0 3485 return (pgoff >= KVM_DIRTY_LOG_PAGE_OFFSET) && 3486 (pgoff < KVM_DIRTY_LOG_PAGE_OFFSET + 3487 kvm->dirty_ring_size / PAGE_SIZE); 3488 #else 3489 return false; 3490 #endif 3491 } 3492 3493 static vm_fault_t kvm_vcpu_fault(struct vm_fault *vmf) 3494 { 3495 struct kvm_vcpu *vcpu = vmf->vma->vm_file->private_data; 3496 struct page *page; 3497 3498 if (vmf->pgoff == 0) 3499 page = virt_to_page(vcpu->run); 3500 #ifdef CONFIG_X86 3501 else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET) 3502 page = virt_to_page(vcpu->arch.pio_data); 3503 #endif 3504 #ifdef CONFIG_KVM_MMIO 3505 else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET) 3506 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring); 3507 #endif 3508 else if (kvm_page_in_dirty_ring(vcpu->kvm, vmf->pgoff)) 3509 page = kvm_dirty_ring_get_page( 3510 &vcpu->dirty_ring, 3511 vmf->pgoff - KVM_DIRTY_LOG_PAGE_OFFSET); 3512 else 3513 return kvm_arch_vcpu_fault(vcpu, vmf); 3514 get_page(page); 3515 vmf->page = page; 3516 return 0; 3517 } 3518 3519 static const struct vm_operations_struct kvm_vcpu_vm_ops = { 3520 .fault = kvm_vcpu_fault, 3521 }; 3522 3523 static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma) 3524 { 3525 struct kvm_vcpu *vcpu = file->private_data; 3526 unsigned long pages = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 3527 3528 if ((kvm_page_in_dirty_ring(vcpu->kvm, vma->vm_pgoff) || 3529 kvm_page_in_dirty_ring(vcpu->kvm, vma->vm_pgoff + pages - 1)) && 3530 ((vma->vm_flags & VM_EXEC) || !(vma->vm_flags & VM_SHARED))) 3531 return -EINVAL; 3532 3533 vma->vm_ops = &kvm_vcpu_vm_ops; 3534 return 0; 3535 } 3536 3537 static int kvm_vcpu_release(struct inode *inode, struct file *filp) 3538 { 3539 struct kvm_vcpu *vcpu = filp->private_data; 3540 3541 kvm_put_kvm(vcpu->kvm); 3542 return 0; 3543 } 3544 3545 static struct file_operations kvm_vcpu_fops = { 3546 .release = kvm_vcpu_release, 3547 .unlocked_ioctl = kvm_vcpu_ioctl, 3548 .mmap = kvm_vcpu_mmap, 3549 .llseek = noop_llseek, 3550 KVM_COMPAT(kvm_vcpu_compat_ioctl), 3551 }; 3552 3553 /* 3554 * Allocates an inode for the vcpu. 3555 */ 3556 static int create_vcpu_fd(struct kvm_vcpu *vcpu) 3557 { 3558 char name[8 + 1 + ITOA_MAX_LEN + 1]; 3559 3560 snprintf(name, sizeof(name), "kvm-vcpu:%d", vcpu->vcpu_id); 3561 return anon_inode_getfd(name, &kvm_vcpu_fops, vcpu, O_RDWR | O_CLOEXEC); 3562 } 3563 3564 static void kvm_create_vcpu_debugfs(struct kvm_vcpu *vcpu) 3565 { 3566 #ifdef __KVM_HAVE_ARCH_VCPU_DEBUGFS 3567 struct dentry *debugfs_dentry; 3568 char dir_name[ITOA_MAX_LEN * 2]; 3569 3570 if (!debugfs_initialized()) 3571 return; 3572 3573 snprintf(dir_name, sizeof(dir_name), "vcpu%d", vcpu->vcpu_id); 3574 debugfs_dentry = debugfs_create_dir(dir_name, 3575 vcpu->kvm->debugfs_dentry); 3576 3577 kvm_arch_create_vcpu_debugfs(vcpu, debugfs_dentry); 3578 #endif 3579 } 3580 3581 /* 3582 * Creates some virtual cpus. Good luck creating more than one. 3583 */ 3584 static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id) 3585 { 3586 int r; 3587 struct kvm_vcpu *vcpu; 3588 struct page *page; 3589 3590 if (id >= KVM_MAX_VCPU_ID) 3591 return -EINVAL; 3592 3593 mutex_lock(&kvm->lock); 3594 if (kvm->created_vcpus == KVM_MAX_VCPUS) { 3595 mutex_unlock(&kvm->lock); 3596 return -EINVAL; 3597 } 3598 3599 kvm->created_vcpus++; 3600 mutex_unlock(&kvm->lock); 3601 3602 r = kvm_arch_vcpu_precreate(kvm, id); 3603 if (r) 3604 goto vcpu_decrement; 3605 3606 vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL_ACCOUNT); 3607 if (!vcpu) { 3608 r = -ENOMEM; 3609 goto vcpu_decrement; 3610 } 3611 3612 BUILD_BUG_ON(sizeof(struct kvm_run) > PAGE_SIZE); 3613 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO); 3614 if (!page) { 3615 r = -ENOMEM; 3616 goto vcpu_free; 3617 } 3618 vcpu->run = page_address(page); 3619 3620 kvm_vcpu_init(vcpu, kvm, id); 3621 3622 r = kvm_arch_vcpu_create(vcpu); 3623 if (r) 3624 goto vcpu_free_run_page; 3625 3626 if (kvm->dirty_ring_size) { 3627 r = kvm_dirty_ring_alloc(&vcpu->dirty_ring, 3628 id, kvm->dirty_ring_size); 3629 if (r) 3630 goto arch_vcpu_destroy; 3631 } 3632 3633 mutex_lock(&kvm->lock); 3634 if (kvm_get_vcpu_by_id(kvm, id)) { 3635 r = -EEXIST; 3636 goto unlock_vcpu_destroy; 3637 } 3638 3639 vcpu->vcpu_idx = atomic_read(&kvm->online_vcpus); 3640 BUG_ON(kvm->vcpus[vcpu->vcpu_idx]); 3641 3642 /* Fill the stats id string for the vcpu */ 3643 snprintf(vcpu->stats_id, sizeof(vcpu->stats_id), "kvm-%d/vcpu-%d", 3644 task_pid_nr(current), id); 3645 3646 /* Now it's all set up, let userspace reach it */ 3647 kvm_get_kvm(kvm); 3648 r = create_vcpu_fd(vcpu); 3649 if (r < 0) { 3650 kvm_put_kvm_no_destroy(kvm); 3651 goto unlock_vcpu_destroy; 3652 } 3653 3654 kvm->vcpus[vcpu->vcpu_idx] = vcpu; 3655 3656 /* 3657 * Pairs with smp_rmb() in kvm_get_vcpu. Write kvm->vcpus 3658 * before kvm->online_vcpu's incremented value. 3659 */ 3660 smp_wmb(); 3661 atomic_inc(&kvm->online_vcpus); 3662 3663 mutex_unlock(&kvm->lock); 3664 kvm_arch_vcpu_postcreate(vcpu); 3665 kvm_create_vcpu_debugfs(vcpu); 3666 return r; 3667 3668 unlock_vcpu_destroy: 3669 mutex_unlock(&kvm->lock); 3670 kvm_dirty_ring_free(&vcpu->dirty_ring); 3671 arch_vcpu_destroy: 3672 kvm_arch_vcpu_destroy(vcpu); 3673 vcpu_free_run_page: 3674 free_page((unsigned long)vcpu->run); 3675 vcpu_free: 3676 kmem_cache_free(kvm_vcpu_cache, vcpu); 3677 vcpu_decrement: 3678 mutex_lock(&kvm->lock); 3679 kvm->created_vcpus--; 3680 mutex_unlock(&kvm->lock); 3681 return r; 3682 } 3683 3684 static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset) 3685 { 3686 if (sigset) { 3687 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP)); 3688 vcpu->sigset_active = 1; 3689 vcpu->sigset = *sigset; 3690 } else 3691 vcpu->sigset_active = 0; 3692 return 0; 3693 } 3694 3695 static ssize_t kvm_vcpu_stats_read(struct file *file, char __user *user_buffer, 3696 size_t size, loff_t *offset) 3697 { 3698 struct kvm_vcpu *vcpu = file->private_data; 3699 3700 return kvm_stats_read(vcpu->stats_id, &kvm_vcpu_stats_header, 3701 &kvm_vcpu_stats_desc[0], &vcpu->stat, 3702 sizeof(vcpu->stat), user_buffer, size, offset); 3703 } 3704 3705 static const struct file_operations kvm_vcpu_stats_fops = { 3706 .read = kvm_vcpu_stats_read, 3707 .llseek = noop_llseek, 3708 }; 3709 3710 static int kvm_vcpu_ioctl_get_stats_fd(struct kvm_vcpu *vcpu) 3711 { 3712 int fd; 3713 struct file *file; 3714 char name[15 + ITOA_MAX_LEN + 1]; 3715 3716 snprintf(name, sizeof(name), "kvm-vcpu-stats:%d", vcpu->vcpu_id); 3717 3718 fd = get_unused_fd_flags(O_CLOEXEC); 3719 if (fd < 0) 3720 return fd; 3721 3722 file = anon_inode_getfile(name, &kvm_vcpu_stats_fops, vcpu, O_RDONLY); 3723 if (IS_ERR(file)) { 3724 put_unused_fd(fd); 3725 return PTR_ERR(file); 3726 } 3727 file->f_mode |= FMODE_PREAD; 3728 fd_install(fd, file); 3729 3730 return fd; 3731 } 3732 3733 static long kvm_vcpu_ioctl(struct file *filp, 3734 unsigned int ioctl, unsigned long arg) 3735 { 3736 struct kvm_vcpu *vcpu = filp->private_data; 3737 void __user *argp = (void __user *)arg; 3738 int r; 3739 struct kvm_fpu *fpu = NULL; 3740 struct kvm_sregs *kvm_sregs = NULL; 3741 3742 if (vcpu->kvm->mm != current->mm || vcpu->kvm->vm_bugged) 3743 return -EIO; 3744 3745 if (unlikely(_IOC_TYPE(ioctl) != KVMIO)) 3746 return -EINVAL; 3747 3748 /* 3749 * Some architectures have vcpu ioctls that are asynchronous to vcpu 3750 * execution; mutex_lock() would break them. 3751 */ 3752 r = kvm_arch_vcpu_async_ioctl(filp, ioctl, arg); 3753 if (r != -ENOIOCTLCMD) 3754 return r; 3755 3756 if (mutex_lock_killable(&vcpu->mutex)) 3757 return -EINTR; 3758 switch (ioctl) { 3759 case KVM_RUN: { 3760 struct pid *oldpid; 3761 r = -EINVAL; 3762 if (arg) 3763 goto out; 3764 oldpid = rcu_access_pointer(vcpu->pid); 3765 if (unlikely(oldpid != task_pid(current))) { 3766 /* The thread running this VCPU changed. */ 3767 struct pid *newpid; 3768 3769 r = kvm_arch_vcpu_run_pid_change(vcpu); 3770 if (r) 3771 break; 3772 3773 newpid = get_task_pid(current, PIDTYPE_PID); 3774 rcu_assign_pointer(vcpu->pid, newpid); 3775 if (oldpid) 3776 synchronize_rcu(); 3777 put_pid(oldpid); 3778 } 3779 r = kvm_arch_vcpu_ioctl_run(vcpu); 3780 trace_kvm_userspace_exit(vcpu->run->exit_reason, r); 3781 break; 3782 } 3783 case KVM_GET_REGS: { 3784 struct kvm_regs *kvm_regs; 3785 3786 r = -ENOMEM; 3787 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL_ACCOUNT); 3788 if (!kvm_regs) 3789 goto out; 3790 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs); 3791 if (r) 3792 goto out_free1; 3793 r = -EFAULT; 3794 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs))) 3795 goto out_free1; 3796 r = 0; 3797 out_free1: 3798 kfree(kvm_regs); 3799 break; 3800 } 3801 case KVM_SET_REGS: { 3802 struct kvm_regs *kvm_regs; 3803 3804 kvm_regs = memdup_user(argp, sizeof(*kvm_regs)); 3805 if (IS_ERR(kvm_regs)) { 3806 r = PTR_ERR(kvm_regs); 3807 goto out; 3808 } 3809 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs); 3810 kfree(kvm_regs); 3811 break; 3812 } 3813 case KVM_GET_SREGS: { 3814 kvm_sregs = kzalloc(sizeof(struct kvm_sregs), 3815 GFP_KERNEL_ACCOUNT); 3816 r = -ENOMEM; 3817 if (!kvm_sregs) 3818 goto out; 3819 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs); 3820 if (r) 3821 goto out; 3822 r = -EFAULT; 3823 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs))) 3824 goto out; 3825 r = 0; 3826 break; 3827 } 3828 case KVM_SET_SREGS: { 3829 kvm_sregs = memdup_user(argp, sizeof(*kvm_sregs)); 3830 if (IS_ERR(kvm_sregs)) { 3831 r = PTR_ERR(kvm_sregs); 3832 kvm_sregs = NULL; 3833 goto out; 3834 } 3835 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs); 3836 break; 3837 } 3838 case KVM_GET_MP_STATE: { 3839 struct kvm_mp_state mp_state; 3840 3841 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state); 3842 if (r) 3843 goto out; 3844 r = -EFAULT; 3845 if (copy_to_user(argp, &mp_state, sizeof(mp_state))) 3846 goto out; 3847 r = 0; 3848 break; 3849 } 3850 case KVM_SET_MP_STATE: { 3851 struct kvm_mp_state mp_state; 3852 3853 r = -EFAULT; 3854 if (copy_from_user(&mp_state, argp, sizeof(mp_state))) 3855 goto out; 3856 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state); 3857 break; 3858 } 3859 case KVM_TRANSLATE: { 3860 struct kvm_translation tr; 3861 3862 r = -EFAULT; 3863 if (copy_from_user(&tr, argp, sizeof(tr))) 3864 goto out; 3865 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr); 3866 if (r) 3867 goto out; 3868 r = -EFAULT; 3869 if (copy_to_user(argp, &tr, sizeof(tr))) 3870 goto out; 3871 r = 0; 3872 break; 3873 } 3874 case KVM_SET_GUEST_DEBUG: { 3875 struct kvm_guest_debug dbg; 3876 3877 r = -EFAULT; 3878 if (copy_from_user(&dbg, argp, sizeof(dbg))) 3879 goto out; 3880 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg); 3881 break; 3882 } 3883 case KVM_SET_SIGNAL_MASK: { 3884 struct kvm_signal_mask __user *sigmask_arg = argp; 3885 struct kvm_signal_mask kvm_sigmask; 3886 sigset_t sigset, *p; 3887 3888 p = NULL; 3889 if (argp) { 3890 r = -EFAULT; 3891 if (copy_from_user(&kvm_sigmask, argp, 3892 sizeof(kvm_sigmask))) 3893 goto out; 3894 r = -EINVAL; 3895 if (kvm_sigmask.len != sizeof(sigset)) 3896 goto out; 3897 r = -EFAULT; 3898 if (copy_from_user(&sigset, sigmask_arg->sigset, 3899 sizeof(sigset))) 3900 goto out; 3901 p = &sigset; 3902 } 3903 r = kvm_vcpu_ioctl_set_sigmask(vcpu, p); 3904 break; 3905 } 3906 case KVM_GET_FPU: { 3907 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL_ACCOUNT); 3908 r = -ENOMEM; 3909 if (!fpu) 3910 goto out; 3911 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu); 3912 if (r) 3913 goto out; 3914 r = -EFAULT; 3915 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu))) 3916 goto out; 3917 r = 0; 3918 break; 3919 } 3920 case KVM_SET_FPU: { 3921 fpu = memdup_user(argp, sizeof(*fpu)); 3922 if (IS_ERR(fpu)) { 3923 r = PTR_ERR(fpu); 3924 fpu = NULL; 3925 goto out; 3926 } 3927 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu); 3928 break; 3929 } 3930 case KVM_GET_STATS_FD: { 3931 r = kvm_vcpu_ioctl_get_stats_fd(vcpu); 3932 break; 3933 } 3934 default: 3935 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg); 3936 } 3937 out: 3938 mutex_unlock(&vcpu->mutex); 3939 kfree(fpu); 3940 kfree(kvm_sregs); 3941 return r; 3942 } 3943 3944 #ifdef CONFIG_KVM_COMPAT 3945 static long kvm_vcpu_compat_ioctl(struct file *filp, 3946 unsigned int ioctl, unsigned long arg) 3947 { 3948 struct kvm_vcpu *vcpu = filp->private_data; 3949 void __user *argp = compat_ptr(arg); 3950 int r; 3951 3952 if (vcpu->kvm->mm != current->mm || vcpu->kvm->vm_bugged) 3953 return -EIO; 3954 3955 switch (ioctl) { 3956 case KVM_SET_SIGNAL_MASK: { 3957 struct kvm_signal_mask __user *sigmask_arg = argp; 3958 struct kvm_signal_mask kvm_sigmask; 3959 sigset_t sigset; 3960 3961 if (argp) { 3962 r = -EFAULT; 3963 if (copy_from_user(&kvm_sigmask, argp, 3964 sizeof(kvm_sigmask))) 3965 goto out; 3966 r = -EINVAL; 3967 if (kvm_sigmask.len != sizeof(compat_sigset_t)) 3968 goto out; 3969 r = -EFAULT; 3970 if (get_compat_sigset(&sigset, 3971 (compat_sigset_t __user *)sigmask_arg->sigset)) 3972 goto out; 3973 r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset); 3974 } else 3975 r = kvm_vcpu_ioctl_set_sigmask(vcpu, NULL); 3976 break; 3977 } 3978 default: 3979 r = kvm_vcpu_ioctl(filp, ioctl, arg); 3980 } 3981 3982 out: 3983 return r; 3984 } 3985 #endif 3986 3987 static int kvm_device_mmap(struct file *filp, struct vm_area_struct *vma) 3988 { 3989 struct kvm_device *dev = filp->private_data; 3990 3991 if (dev->ops->mmap) 3992 return dev->ops->mmap(dev, vma); 3993 3994 return -ENODEV; 3995 } 3996 3997 static int kvm_device_ioctl_attr(struct kvm_device *dev, 3998 int (*accessor)(struct kvm_device *dev, 3999 struct kvm_device_attr *attr), 4000 unsigned long arg) 4001 { 4002 struct kvm_device_attr attr; 4003 4004 if (!accessor) 4005 return -EPERM; 4006 4007 if (copy_from_user(&attr, (void __user *)arg, sizeof(attr))) 4008 return -EFAULT; 4009 4010 return accessor(dev, &attr); 4011 } 4012 4013 static long kvm_device_ioctl(struct file *filp, unsigned int ioctl, 4014 unsigned long arg) 4015 { 4016 struct kvm_device *dev = filp->private_data; 4017 4018 if (dev->kvm->mm != current->mm || dev->kvm->vm_bugged) 4019 return -EIO; 4020 4021 switch (ioctl) { 4022 case KVM_SET_DEVICE_ATTR: 4023 return kvm_device_ioctl_attr(dev, dev->ops->set_attr, arg); 4024 case KVM_GET_DEVICE_ATTR: 4025 return kvm_device_ioctl_attr(dev, dev->ops->get_attr, arg); 4026 case KVM_HAS_DEVICE_ATTR: 4027 return kvm_device_ioctl_attr(dev, dev->ops->has_attr, arg); 4028 default: 4029 if (dev->ops->ioctl) 4030 return dev->ops->ioctl(dev, ioctl, arg); 4031 4032 return -ENOTTY; 4033 } 4034 } 4035 4036 static int kvm_device_release(struct inode *inode, struct file *filp) 4037 { 4038 struct kvm_device *dev = filp->private_data; 4039 struct kvm *kvm = dev->kvm; 4040 4041 if (dev->ops->release) { 4042 mutex_lock(&kvm->lock); 4043 list_del(&dev->vm_node); 4044 dev->ops->release(dev); 4045 mutex_unlock(&kvm->lock); 4046 } 4047 4048 kvm_put_kvm(kvm); 4049 return 0; 4050 } 4051 4052 static const struct file_operations kvm_device_fops = { 4053 .unlocked_ioctl = kvm_device_ioctl, 4054 .release = kvm_device_release, 4055 KVM_COMPAT(kvm_device_ioctl), 4056 .mmap = kvm_device_mmap, 4057 }; 4058 4059 struct kvm_device *kvm_device_from_filp(struct file *filp) 4060 { 4061 if (filp->f_op != &kvm_device_fops) 4062 return NULL; 4063 4064 return filp->private_data; 4065 } 4066 4067 static const struct kvm_device_ops *kvm_device_ops_table[KVM_DEV_TYPE_MAX] = { 4068 #ifdef CONFIG_KVM_MPIC 4069 [KVM_DEV_TYPE_FSL_MPIC_20] = &kvm_mpic_ops, 4070 [KVM_DEV_TYPE_FSL_MPIC_42] = &kvm_mpic_ops, 4071 #endif 4072 }; 4073 4074 int kvm_register_device_ops(const struct kvm_device_ops *ops, u32 type) 4075 { 4076 if (type >= ARRAY_SIZE(kvm_device_ops_table)) 4077 return -ENOSPC; 4078 4079 if (kvm_device_ops_table[type] != NULL) 4080 return -EEXIST; 4081 4082 kvm_device_ops_table[type] = ops; 4083 return 0; 4084 } 4085 4086 void kvm_unregister_device_ops(u32 type) 4087 { 4088 if (kvm_device_ops_table[type] != NULL) 4089 kvm_device_ops_table[type] = NULL; 4090 } 4091 4092 static int kvm_ioctl_create_device(struct kvm *kvm, 4093 struct kvm_create_device *cd) 4094 { 4095 const struct kvm_device_ops *ops = NULL; 4096 struct kvm_device *dev; 4097 bool test = cd->flags & KVM_CREATE_DEVICE_TEST; 4098 int type; 4099 int ret; 4100 4101 if (cd->type >= ARRAY_SIZE(kvm_device_ops_table)) 4102 return -ENODEV; 4103 4104 type = array_index_nospec(cd->type, ARRAY_SIZE(kvm_device_ops_table)); 4105 ops = kvm_device_ops_table[type]; 4106 if (ops == NULL) 4107 return -ENODEV; 4108 4109 if (test) 4110 return 0; 4111 4112 dev = kzalloc(sizeof(*dev), GFP_KERNEL_ACCOUNT); 4113 if (!dev) 4114 return -ENOMEM; 4115 4116 dev->ops = ops; 4117 dev->kvm = kvm; 4118 4119 mutex_lock(&kvm->lock); 4120 ret = ops->create(dev, type); 4121 if (ret < 0) { 4122 mutex_unlock(&kvm->lock); 4123 kfree(dev); 4124 return ret; 4125 } 4126 list_add(&dev->vm_node, &kvm->devices); 4127 mutex_unlock(&kvm->lock); 4128 4129 if (ops->init) 4130 ops->init(dev); 4131 4132 kvm_get_kvm(kvm); 4133 ret = anon_inode_getfd(ops->name, &kvm_device_fops, dev, O_RDWR | O_CLOEXEC); 4134 if (ret < 0) { 4135 kvm_put_kvm_no_destroy(kvm); 4136 mutex_lock(&kvm->lock); 4137 list_del(&dev->vm_node); 4138 mutex_unlock(&kvm->lock); 4139 ops->destroy(dev); 4140 return ret; 4141 } 4142 4143 cd->fd = ret; 4144 return 0; 4145 } 4146 4147 static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg) 4148 { 4149 switch (arg) { 4150 case KVM_CAP_USER_MEMORY: 4151 case KVM_CAP_DESTROY_MEMORY_REGION_WORKS: 4152 case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS: 4153 case KVM_CAP_INTERNAL_ERROR_DATA: 4154 #ifdef CONFIG_HAVE_KVM_MSI 4155 case KVM_CAP_SIGNAL_MSI: 4156 #endif 4157 #ifdef CONFIG_HAVE_KVM_IRQFD 4158 case KVM_CAP_IRQFD: 4159 case KVM_CAP_IRQFD_RESAMPLE: 4160 #endif 4161 case KVM_CAP_IOEVENTFD_ANY_LENGTH: 4162 case KVM_CAP_CHECK_EXTENSION_VM: 4163 case KVM_CAP_ENABLE_CAP_VM: 4164 case KVM_CAP_HALT_POLL: 4165 return 1; 4166 #ifdef CONFIG_KVM_MMIO 4167 case KVM_CAP_COALESCED_MMIO: 4168 return KVM_COALESCED_MMIO_PAGE_OFFSET; 4169 case KVM_CAP_COALESCED_PIO: 4170 return 1; 4171 #endif 4172 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT 4173 case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2: 4174 return KVM_DIRTY_LOG_MANUAL_CAPS; 4175 #endif 4176 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING 4177 case KVM_CAP_IRQ_ROUTING: 4178 return KVM_MAX_IRQ_ROUTES; 4179 #endif 4180 #if KVM_ADDRESS_SPACE_NUM > 1 4181 case KVM_CAP_MULTI_ADDRESS_SPACE: 4182 return KVM_ADDRESS_SPACE_NUM; 4183 #endif 4184 case KVM_CAP_NR_MEMSLOTS: 4185 return KVM_USER_MEM_SLOTS; 4186 case KVM_CAP_DIRTY_LOG_RING: 4187 #if KVM_DIRTY_LOG_PAGE_OFFSET > 0 4188 return KVM_DIRTY_RING_MAX_ENTRIES * sizeof(struct kvm_dirty_gfn); 4189 #else 4190 return 0; 4191 #endif 4192 case KVM_CAP_BINARY_STATS_FD: 4193 return 1; 4194 default: 4195 break; 4196 } 4197 return kvm_vm_ioctl_check_extension(kvm, arg); 4198 } 4199 4200 static int kvm_vm_ioctl_enable_dirty_log_ring(struct kvm *kvm, u32 size) 4201 { 4202 int r; 4203 4204 if (!KVM_DIRTY_LOG_PAGE_OFFSET) 4205 return -EINVAL; 4206 4207 /* the size should be power of 2 */ 4208 if (!size || (size & (size - 1))) 4209 return -EINVAL; 4210 4211 /* Should be bigger to keep the reserved entries, or a page */ 4212 if (size < kvm_dirty_ring_get_rsvd_entries() * 4213 sizeof(struct kvm_dirty_gfn) || size < PAGE_SIZE) 4214 return -EINVAL; 4215 4216 if (size > KVM_DIRTY_RING_MAX_ENTRIES * 4217 sizeof(struct kvm_dirty_gfn)) 4218 return -E2BIG; 4219 4220 /* We only allow it to set once */ 4221 if (kvm->dirty_ring_size) 4222 return -EINVAL; 4223 4224 mutex_lock(&kvm->lock); 4225 4226 if (kvm->created_vcpus) { 4227 /* We don't allow to change this value after vcpu created */ 4228 r = -EINVAL; 4229 } else { 4230 kvm->dirty_ring_size = size; 4231 r = 0; 4232 } 4233 4234 mutex_unlock(&kvm->lock); 4235 return r; 4236 } 4237 4238 static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm) 4239 { 4240 int i; 4241 struct kvm_vcpu *vcpu; 4242 int cleared = 0; 4243 4244 if (!kvm->dirty_ring_size) 4245 return -EINVAL; 4246 4247 mutex_lock(&kvm->slots_lock); 4248 4249 kvm_for_each_vcpu(i, vcpu, kvm) 4250 cleared += kvm_dirty_ring_reset(vcpu->kvm, &vcpu->dirty_ring); 4251 4252 mutex_unlock(&kvm->slots_lock); 4253 4254 if (cleared) 4255 kvm_flush_remote_tlbs(kvm); 4256 4257 return cleared; 4258 } 4259 4260 int __attribute__((weak)) kvm_vm_ioctl_enable_cap(struct kvm *kvm, 4261 struct kvm_enable_cap *cap) 4262 { 4263 return -EINVAL; 4264 } 4265 4266 static int kvm_vm_ioctl_enable_cap_generic(struct kvm *kvm, 4267 struct kvm_enable_cap *cap) 4268 { 4269 switch (cap->cap) { 4270 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT 4271 case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2: { 4272 u64 allowed_options = KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE; 4273 4274 if (cap->args[0] & KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE) 4275 allowed_options = KVM_DIRTY_LOG_MANUAL_CAPS; 4276 4277 if (cap->flags || (cap->args[0] & ~allowed_options)) 4278 return -EINVAL; 4279 kvm->manual_dirty_log_protect = cap->args[0]; 4280 return 0; 4281 } 4282 #endif 4283 case KVM_CAP_HALT_POLL: { 4284 if (cap->flags || cap->args[0] != (unsigned int)cap->args[0]) 4285 return -EINVAL; 4286 4287 kvm->max_halt_poll_ns = cap->args[0]; 4288 return 0; 4289 } 4290 case KVM_CAP_DIRTY_LOG_RING: 4291 return kvm_vm_ioctl_enable_dirty_log_ring(kvm, cap->args[0]); 4292 default: 4293 return kvm_vm_ioctl_enable_cap(kvm, cap); 4294 } 4295 } 4296 4297 static ssize_t kvm_vm_stats_read(struct file *file, char __user *user_buffer, 4298 size_t size, loff_t *offset) 4299 { 4300 struct kvm *kvm = file->private_data; 4301 4302 return kvm_stats_read(kvm->stats_id, &kvm_vm_stats_header, 4303 &kvm_vm_stats_desc[0], &kvm->stat, 4304 sizeof(kvm->stat), user_buffer, size, offset); 4305 } 4306 4307 static const struct file_operations kvm_vm_stats_fops = { 4308 .read = kvm_vm_stats_read, 4309 .llseek = noop_llseek, 4310 }; 4311 4312 static int kvm_vm_ioctl_get_stats_fd(struct kvm *kvm) 4313 { 4314 int fd; 4315 struct file *file; 4316 4317 fd = get_unused_fd_flags(O_CLOEXEC); 4318 if (fd < 0) 4319 return fd; 4320 4321 file = anon_inode_getfile("kvm-vm-stats", 4322 &kvm_vm_stats_fops, kvm, O_RDONLY); 4323 if (IS_ERR(file)) { 4324 put_unused_fd(fd); 4325 return PTR_ERR(file); 4326 } 4327 file->f_mode |= FMODE_PREAD; 4328 fd_install(fd, file); 4329 4330 return fd; 4331 } 4332 4333 static long kvm_vm_ioctl(struct file *filp, 4334 unsigned int ioctl, unsigned long arg) 4335 { 4336 struct kvm *kvm = filp->private_data; 4337 void __user *argp = (void __user *)arg; 4338 int r; 4339 4340 if (kvm->mm != current->mm || kvm->vm_bugged) 4341 return -EIO; 4342 switch (ioctl) { 4343 case KVM_CREATE_VCPU: 4344 r = kvm_vm_ioctl_create_vcpu(kvm, arg); 4345 break; 4346 case KVM_ENABLE_CAP: { 4347 struct kvm_enable_cap cap; 4348 4349 r = -EFAULT; 4350 if (copy_from_user(&cap, argp, sizeof(cap))) 4351 goto out; 4352 r = kvm_vm_ioctl_enable_cap_generic(kvm, &cap); 4353 break; 4354 } 4355 case KVM_SET_USER_MEMORY_REGION: { 4356 struct kvm_userspace_memory_region kvm_userspace_mem; 4357 4358 r = -EFAULT; 4359 if (copy_from_user(&kvm_userspace_mem, argp, 4360 sizeof(kvm_userspace_mem))) 4361 goto out; 4362 4363 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem); 4364 break; 4365 } 4366 case KVM_GET_DIRTY_LOG: { 4367 struct kvm_dirty_log log; 4368 4369 r = -EFAULT; 4370 if (copy_from_user(&log, argp, sizeof(log))) 4371 goto out; 4372 r = kvm_vm_ioctl_get_dirty_log(kvm, &log); 4373 break; 4374 } 4375 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT 4376 case KVM_CLEAR_DIRTY_LOG: { 4377 struct kvm_clear_dirty_log log; 4378 4379 r = -EFAULT; 4380 if (copy_from_user(&log, argp, sizeof(log))) 4381 goto out; 4382 r = kvm_vm_ioctl_clear_dirty_log(kvm, &log); 4383 break; 4384 } 4385 #endif 4386 #ifdef CONFIG_KVM_MMIO 4387 case KVM_REGISTER_COALESCED_MMIO: { 4388 struct kvm_coalesced_mmio_zone zone; 4389 4390 r = -EFAULT; 4391 if (copy_from_user(&zone, argp, sizeof(zone))) 4392 goto out; 4393 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone); 4394 break; 4395 } 4396 case KVM_UNREGISTER_COALESCED_MMIO: { 4397 struct kvm_coalesced_mmio_zone zone; 4398 4399 r = -EFAULT; 4400 if (copy_from_user(&zone, argp, sizeof(zone))) 4401 goto out; 4402 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone); 4403 break; 4404 } 4405 #endif 4406 case KVM_IRQFD: { 4407 struct kvm_irqfd data; 4408 4409 r = -EFAULT; 4410 if (copy_from_user(&data, argp, sizeof(data))) 4411 goto out; 4412 r = kvm_irqfd(kvm, &data); 4413 break; 4414 } 4415 case KVM_IOEVENTFD: { 4416 struct kvm_ioeventfd data; 4417 4418 r = -EFAULT; 4419 if (copy_from_user(&data, argp, sizeof(data))) 4420 goto out; 4421 r = kvm_ioeventfd(kvm, &data); 4422 break; 4423 } 4424 #ifdef CONFIG_HAVE_KVM_MSI 4425 case KVM_SIGNAL_MSI: { 4426 struct kvm_msi msi; 4427 4428 r = -EFAULT; 4429 if (copy_from_user(&msi, argp, sizeof(msi))) 4430 goto out; 4431 r = kvm_send_userspace_msi(kvm, &msi); 4432 break; 4433 } 4434 #endif 4435 #ifdef __KVM_HAVE_IRQ_LINE 4436 case KVM_IRQ_LINE_STATUS: 4437 case KVM_IRQ_LINE: { 4438 struct kvm_irq_level irq_event; 4439 4440 r = -EFAULT; 4441 if (copy_from_user(&irq_event, argp, sizeof(irq_event))) 4442 goto out; 4443 4444 r = kvm_vm_ioctl_irq_line(kvm, &irq_event, 4445 ioctl == KVM_IRQ_LINE_STATUS); 4446 if (r) 4447 goto out; 4448 4449 r = -EFAULT; 4450 if (ioctl == KVM_IRQ_LINE_STATUS) { 4451 if (copy_to_user(argp, &irq_event, sizeof(irq_event))) 4452 goto out; 4453 } 4454 4455 r = 0; 4456 break; 4457 } 4458 #endif 4459 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING 4460 case KVM_SET_GSI_ROUTING: { 4461 struct kvm_irq_routing routing; 4462 struct kvm_irq_routing __user *urouting; 4463 struct kvm_irq_routing_entry *entries = NULL; 4464 4465 r = -EFAULT; 4466 if (copy_from_user(&routing, argp, sizeof(routing))) 4467 goto out; 4468 r = -EINVAL; 4469 if (!kvm_arch_can_set_irq_routing(kvm)) 4470 goto out; 4471 if (routing.nr > KVM_MAX_IRQ_ROUTES) 4472 goto out; 4473 if (routing.flags) 4474 goto out; 4475 if (routing.nr) { 4476 urouting = argp; 4477 entries = vmemdup_user(urouting->entries, 4478 array_size(sizeof(*entries), 4479 routing.nr)); 4480 if (IS_ERR(entries)) { 4481 r = PTR_ERR(entries); 4482 goto out; 4483 } 4484 } 4485 r = kvm_set_irq_routing(kvm, entries, routing.nr, 4486 routing.flags); 4487 kvfree(entries); 4488 break; 4489 } 4490 #endif /* CONFIG_HAVE_KVM_IRQ_ROUTING */ 4491 case KVM_CREATE_DEVICE: { 4492 struct kvm_create_device cd; 4493 4494 r = -EFAULT; 4495 if (copy_from_user(&cd, argp, sizeof(cd))) 4496 goto out; 4497 4498 r = kvm_ioctl_create_device(kvm, &cd); 4499 if (r) 4500 goto out; 4501 4502 r = -EFAULT; 4503 if (copy_to_user(argp, &cd, sizeof(cd))) 4504 goto out; 4505 4506 r = 0; 4507 break; 4508 } 4509 case KVM_CHECK_EXTENSION: 4510 r = kvm_vm_ioctl_check_extension_generic(kvm, arg); 4511 break; 4512 case KVM_RESET_DIRTY_RINGS: 4513 r = kvm_vm_ioctl_reset_dirty_pages(kvm); 4514 break; 4515 case KVM_GET_STATS_FD: 4516 r = kvm_vm_ioctl_get_stats_fd(kvm); 4517 break; 4518 default: 4519 r = kvm_arch_vm_ioctl(filp, ioctl, arg); 4520 } 4521 out: 4522 return r; 4523 } 4524 4525 #ifdef CONFIG_KVM_COMPAT 4526 struct compat_kvm_dirty_log { 4527 __u32 slot; 4528 __u32 padding1; 4529 union { 4530 compat_uptr_t dirty_bitmap; /* one bit per page */ 4531 __u64 padding2; 4532 }; 4533 }; 4534 4535 struct compat_kvm_clear_dirty_log { 4536 __u32 slot; 4537 __u32 num_pages; 4538 __u64 first_page; 4539 union { 4540 compat_uptr_t dirty_bitmap; /* one bit per page */ 4541 __u64 padding2; 4542 }; 4543 }; 4544 4545 static long kvm_vm_compat_ioctl(struct file *filp, 4546 unsigned int ioctl, unsigned long arg) 4547 { 4548 struct kvm *kvm = filp->private_data; 4549 int r; 4550 4551 if (kvm->mm != current->mm || kvm->vm_bugged) 4552 return -EIO; 4553 switch (ioctl) { 4554 #ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT 4555 case KVM_CLEAR_DIRTY_LOG: { 4556 struct compat_kvm_clear_dirty_log compat_log; 4557 struct kvm_clear_dirty_log log; 4558 4559 if (copy_from_user(&compat_log, (void __user *)arg, 4560 sizeof(compat_log))) 4561 return -EFAULT; 4562 log.slot = compat_log.slot; 4563 log.num_pages = compat_log.num_pages; 4564 log.first_page = compat_log.first_page; 4565 log.padding2 = compat_log.padding2; 4566 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap); 4567 4568 r = kvm_vm_ioctl_clear_dirty_log(kvm, &log); 4569 break; 4570 } 4571 #endif 4572 case KVM_GET_DIRTY_LOG: { 4573 struct compat_kvm_dirty_log compat_log; 4574 struct kvm_dirty_log log; 4575 4576 if (copy_from_user(&compat_log, (void __user *)arg, 4577 sizeof(compat_log))) 4578 return -EFAULT; 4579 log.slot = compat_log.slot; 4580 log.padding1 = compat_log.padding1; 4581 log.padding2 = compat_log.padding2; 4582 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap); 4583 4584 r = kvm_vm_ioctl_get_dirty_log(kvm, &log); 4585 break; 4586 } 4587 default: 4588 r = kvm_vm_ioctl(filp, ioctl, arg); 4589 } 4590 return r; 4591 } 4592 #endif 4593 4594 static struct file_operations kvm_vm_fops = { 4595 .release = kvm_vm_release, 4596 .unlocked_ioctl = kvm_vm_ioctl, 4597 .llseek = noop_llseek, 4598 KVM_COMPAT(kvm_vm_compat_ioctl), 4599 }; 4600 4601 bool file_is_kvm(struct file *file) 4602 { 4603 return file && file->f_op == &kvm_vm_fops; 4604 } 4605 EXPORT_SYMBOL_GPL(file_is_kvm); 4606 4607 static int kvm_dev_ioctl_create_vm(unsigned long type) 4608 { 4609 int r; 4610 struct kvm *kvm; 4611 struct file *file; 4612 4613 kvm = kvm_create_vm(type); 4614 if (IS_ERR(kvm)) 4615 return PTR_ERR(kvm); 4616 #ifdef CONFIG_KVM_MMIO 4617 r = kvm_coalesced_mmio_init(kvm); 4618 if (r < 0) 4619 goto put_kvm; 4620 #endif 4621 r = get_unused_fd_flags(O_CLOEXEC); 4622 if (r < 0) 4623 goto put_kvm; 4624 4625 snprintf(kvm->stats_id, sizeof(kvm->stats_id), 4626 "kvm-%d", task_pid_nr(current)); 4627 4628 file = anon_inode_getfile("kvm-vm", &kvm_vm_fops, kvm, O_RDWR); 4629 if (IS_ERR(file)) { 4630 put_unused_fd(r); 4631 r = PTR_ERR(file); 4632 goto put_kvm; 4633 } 4634 4635 /* 4636 * Don't call kvm_put_kvm anymore at this point; file->f_op is 4637 * already set, with ->release() being kvm_vm_release(). In error 4638 * cases it will be called by the final fput(file) and will take 4639 * care of doing kvm_put_kvm(kvm). 4640 */ 4641 if (kvm_create_vm_debugfs(kvm, r) < 0) { 4642 put_unused_fd(r); 4643 fput(file); 4644 return -ENOMEM; 4645 } 4646 kvm_uevent_notify_change(KVM_EVENT_CREATE_VM, kvm); 4647 4648 fd_install(r, file); 4649 return r; 4650 4651 put_kvm: 4652 kvm_put_kvm(kvm); 4653 return r; 4654 } 4655 4656 static long kvm_dev_ioctl(struct file *filp, 4657 unsigned int ioctl, unsigned long arg) 4658 { 4659 long r = -EINVAL; 4660 4661 switch (ioctl) { 4662 case KVM_GET_API_VERSION: 4663 if (arg) 4664 goto out; 4665 r = KVM_API_VERSION; 4666 break; 4667 case KVM_CREATE_VM: 4668 r = kvm_dev_ioctl_create_vm(arg); 4669 break; 4670 case KVM_CHECK_EXTENSION: 4671 r = kvm_vm_ioctl_check_extension_generic(NULL, arg); 4672 break; 4673 case KVM_GET_VCPU_MMAP_SIZE: 4674 if (arg) 4675 goto out; 4676 r = PAGE_SIZE; /* struct kvm_run */ 4677 #ifdef CONFIG_X86 4678 r += PAGE_SIZE; /* pio data page */ 4679 #endif 4680 #ifdef CONFIG_KVM_MMIO 4681 r += PAGE_SIZE; /* coalesced mmio ring page */ 4682 #endif 4683 break; 4684 case KVM_TRACE_ENABLE: 4685 case KVM_TRACE_PAUSE: 4686 case KVM_TRACE_DISABLE: 4687 r = -EOPNOTSUPP; 4688 break; 4689 default: 4690 return kvm_arch_dev_ioctl(filp, ioctl, arg); 4691 } 4692 out: 4693 return r; 4694 } 4695 4696 static struct file_operations kvm_chardev_ops = { 4697 .unlocked_ioctl = kvm_dev_ioctl, 4698 .llseek = noop_llseek, 4699 KVM_COMPAT(kvm_dev_ioctl), 4700 }; 4701 4702 static struct miscdevice kvm_dev = { 4703 KVM_MINOR, 4704 "kvm", 4705 &kvm_chardev_ops, 4706 }; 4707 4708 static void hardware_enable_nolock(void *junk) 4709 { 4710 int cpu = raw_smp_processor_id(); 4711 int r; 4712 4713 if (cpumask_test_cpu(cpu, cpus_hardware_enabled)) 4714 return; 4715 4716 cpumask_set_cpu(cpu, cpus_hardware_enabled); 4717 4718 r = kvm_arch_hardware_enable(); 4719 4720 if (r) { 4721 cpumask_clear_cpu(cpu, cpus_hardware_enabled); 4722 atomic_inc(&hardware_enable_failed); 4723 pr_info("kvm: enabling virtualization on CPU%d failed\n", cpu); 4724 } 4725 } 4726 4727 static int kvm_starting_cpu(unsigned int cpu) 4728 { 4729 raw_spin_lock(&kvm_count_lock); 4730 if (kvm_usage_count) 4731 hardware_enable_nolock(NULL); 4732 raw_spin_unlock(&kvm_count_lock); 4733 return 0; 4734 } 4735 4736 static void hardware_disable_nolock(void *junk) 4737 { 4738 int cpu = raw_smp_processor_id(); 4739 4740 if (!cpumask_test_cpu(cpu, cpus_hardware_enabled)) 4741 return; 4742 cpumask_clear_cpu(cpu, cpus_hardware_enabled); 4743 kvm_arch_hardware_disable(); 4744 } 4745 4746 static int kvm_dying_cpu(unsigned int cpu) 4747 { 4748 raw_spin_lock(&kvm_count_lock); 4749 if (kvm_usage_count) 4750 hardware_disable_nolock(NULL); 4751 raw_spin_unlock(&kvm_count_lock); 4752 return 0; 4753 } 4754 4755 static void hardware_disable_all_nolock(void) 4756 { 4757 BUG_ON(!kvm_usage_count); 4758 4759 kvm_usage_count--; 4760 if (!kvm_usage_count) 4761 on_each_cpu(hardware_disable_nolock, NULL, 1); 4762 } 4763 4764 static void hardware_disable_all(void) 4765 { 4766 raw_spin_lock(&kvm_count_lock); 4767 hardware_disable_all_nolock(); 4768 raw_spin_unlock(&kvm_count_lock); 4769 } 4770 4771 static int hardware_enable_all(void) 4772 { 4773 int r = 0; 4774 4775 raw_spin_lock(&kvm_count_lock); 4776 4777 kvm_usage_count++; 4778 if (kvm_usage_count == 1) { 4779 atomic_set(&hardware_enable_failed, 0); 4780 on_each_cpu(hardware_enable_nolock, NULL, 1); 4781 4782 if (atomic_read(&hardware_enable_failed)) { 4783 hardware_disable_all_nolock(); 4784 r = -EBUSY; 4785 } 4786 } 4787 4788 raw_spin_unlock(&kvm_count_lock); 4789 4790 return r; 4791 } 4792 4793 static int kvm_reboot(struct notifier_block *notifier, unsigned long val, 4794 void *v) 4795 { 4796 /* 4797 * Some (well, at least mine) BIOSes hang on reboot if 4798 * in vmx root mode. 4799 * 4800 * And Intel TXT required VMX off for all cpu when system shutdown. 4801 */ 4802 pr_info("kvm: exiting hardware virtualization\n"); 4803 kvm_rebooting = true; 4804 on_each_cpu(hardware_disable_nolock, NULL, 1); 4805 return NOTIFY_OK; 4806 } 4807 4808 static struct notifier_block kvm_reboot_notifier = { 4809 .notifier_call = kvm_reboot, 4810 .priority = 0, 4811 }; 4812 4813 static void kvm_io_bus_destroy(struct kvm_io_bus *bus) 4814 { 4815 int i; 4816 4817 for (i = 0; i < bus->dev_count; i++) { 4818 struct kvm_io_device *pos = bus->range[i].dev; 4819 4820 kvm_iodevice_destructor(pos); 4821 } 4822 kfree(bus); 4823 } 4824 4825 static inline int kvm_io_bus_cmp(const struct kvm_io_range *r1, 4826 const struct kvm_io_range *r2) 4827 { 4828 gpa_t addr1 = r1->addr; 4829 gpa_t addr2 = r2->addr; 4830 4831 if (addr1 < addr2) 4832 return -1; 4833 4834 /* If r2->len == 0, match the exact address. If r2->len != 0, 4835 * accept any overlapping write. Any order is acceptable for 4836 * overlapping ranges, because kvm_io_bus_get_first_dev ensures 4837 * we process all of them. 4838 */ 4839 if (r2->len) { 4840 addr1 += r1->len; 4841 addr2 += r2->len; 4842 } 4843 4844 if (addr1 > addr2) 4845 return 1; 4846 4847 return 0; 4848 } 4849 4850 static int kvm_io_bus_sort_cmp(const void *p1, const void *p2) 4851 { 4852 return kvm_io_bus_cmp(p1, p2); 4853 } 4854 4855 static int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus, 4856 gpa_t addr, int len) 4857 { 4858 struct kvm_io_range *range, key; 4859 int off; 4860 4861 key = (struct kvm_io_range) { 4862 .addr = addr, 4863 .len = len, 4864 }; 4865 4866 range = bsearch(&key, bus->range, bus->dev_count, 4867 sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp); 4868 if (range == NULL) 4869 return -ENOENT; 4870 4871 off = range - bus->range; 4872 4873 while (off > 0 && kvm_io_bus_cmp(&key, &bus->range[off-1]) == 0) 4874 off--; 4875 4876 return off; 4877 } 4878 4879 static int __kvm_io_bus_write(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus, 4880 struct kvm_io_range *range, const void *val) 4881 { 4882 int idx; 4883 4884 idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len); 4885 if (idx < 0) 4886 return -EOPNOTSUPP; 4887 4888 while (idx < bus->dev_count && 4889 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) { 4890 if (!kvm_iodevice_write(vcpu, bus->range[idx].dev, range->addr, 4891 range->len, val)) 4892 return idx; 4893 idx++; 4894 } 4895 4896 return -EOPNOTSUPP; 4897 } 4898 4899 /* kvm_io_bus_write - called under kvm->slots_lock */ 4900 int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr, 4901 int len, const void *val) 4902 { 4903 struct kvm_io_bus *bus; 4904 struct kvm_io_range range; 4905 int r; 4906 4907 range = (struct kvm_io_range) { 4908 .addr = addr, 4909 .len = len, 4910 }; 4911 4912 bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu); 4913 if (!bus) 4914 return -ENOMEM; 4915 r = __kvm_io_bus_write(vcpu, bus, &range, val); 4916 return r < 0 ? r : 0; 4917 } 4918 EXPORT_SYMBOL_GPL(kvm_io_bus_write); 4919 4920 /* kvm_io_bus_write_cookie - called under kvm->slots_lock */ 4921 int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, 4922 gpa_t addr, int len, const void *val, long cookie) 4923 { 4924 struct kvm_io_bus *bus; 4925 struct kvm_io_range range; 4926 4927 range = (struct kvm_io_range) { 4928 .addr = addr, 4929 .len = len, 4930 }; 4931 4932 bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu); 4933 if (!bus) 4934 return -ENOMEM; 4935 4936 /* First try the device referenced by cookie. */ 4937 if ((cookie >= 0) && (cookie < bus->dev_count) && 4938 (kvm_io_bus_cmp(&range, &bus->range[cookie]) == 0)) 4939 if (!kvm_iodevice_write(vcpu, bus->range[cookie].dev, addr, len, 4940 val)) 4941 return cookie; 4942 4943 /* 4944 * cookie contained garbage; fall back to search and return the 4945 * correct cookie value. 4946 */ 4947 return __kvm_io_bus_write(vcpu, bus, &range, val); 4948 } 4949 4950 static int __kvm_io_bus_read(struct kvm_vcpu *vcpu, struct kvm_io_bus *bus, 4951 struct kvm_io_range *range, void *val) 4952 { 4953 int idx; 4954 4955 idx = kvm_io_bus_get_first_dev(bus, range->addr, range->len); 4956 if (idx < 0) 4957 return -EOPNOTSUPP; 4958 4959 while (idx < bus->dev_count && 4960 kvm_io_bus_cmp(range, &bus->range[idx]) == 0) { 4961 if (!kvm_iodevice_read(vcpu, bus->range[idx].dev, range->addr, 4962 range->len, val)) 4963 return idx; 4964 idx++; 4965 } 4966 4967 return -EOPNOTSUPP; 4968 } 4969 4970 /* kvm_io_bus_read - called under kvm->slots_lock */ 4971 int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr, 4972 int len, void *val) 4973 { 4974 struct kvm_io_bus *bus; 4975 struct kvm_io_range range; 4976 int r; 4977 4978 range = (struct kvm_io_range) { 4979 .addr = addr, 4980 .len = len, 4981 }; 4982 4983 bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu); 4984 if (!bus) 4985 return -ENOMEM; 4986 r = __kvm_io_bus_read(vcpu, bus, &range, val); 4987 return r < 0 ? r : 0; 4988 } 4989 4990 /* Caller must hold slots_lock. */ 4991 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr, 4992 int len, struct kvm_io_device *dev) 4993 { 4994 int i; 4995 struct kvm_io_bus *new_bus, *bus; 4996 struct kvm_io_range range; 4997 4998 bus = kvm_get_bus(kvm, bus_idx); 4999 if (!bus) 5000 return -ENOMEM; 5001 5002 /* exclude ioeventfd which is limited by maximum fd */ 5003 if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1) 5004 return -ENOSPC; 5005 5006 new_bus = kmalloc(struct_size(bus, range, bus->dev_count + 1), 5007 GFP_KERNEL_ACCOUNT); 5008 if (!new_bus) 5009 return -ENOMEM; 5010 5011 range = (struct kvm_io_range) { 5012 .addr = addr, 5013 .len = len, 5014 .dev = dev, 5015 }; 5016 5017 for (i = 0; i < bus->dev_count; i++) 5018 if (kvm_io_bus_cmp(&bus->range[i], &range) > 0) 5019 break; 5020 5021 memcpy(new_bus, bus, sizeof(*bus) + i * sizeof(struct kvm_io_range)); 5022 new_bus->dev_count++; 5023 new_bus->range[i] = range; 5024 memcpy(new_bus->range + i + 1, bus->range + i, 5025 (bus->dev_count - i) * sizeof(struct kvm_io_range)); 5026 rcu_assign_pointer(kvm->buses[bus_idx], new_bus); 5027 synchronize_srcu_expedited(&kvm->srcu); 5028 kfree(bus); 5029 5030 return 0; 5031 } 5032 5033 int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx, 5034 struct kvm_io_device *dev) 5035 { 5036 int i, j; 5037 struct kvm_io_bus *new_bus, *bus; 5038 5039 lockdep_assert_held(&kvm->slots_lock); 5040 5041 bus = kvm_get_bus(kvm, bus_idx); 5042 if (!bus) 5043 return 0; 5044 5045 for (i = 0; i < bus->dev_count; i++) { 5046 if (bus->range[i].dev == dev) { 5047 break; 5048 } 5049 } 5050 5051 if (i == bus->dev_count) 5052 return 0; 5053 5054 new_bus = kmalloc(struct_size(bus, range, bus->dev_count - 1), 5055 GFP_KERNEL_ACCOUNT); 5056 if (new_bus) { 5057 memcpy(new_bus, bus, struct_size(bus, range, i)); 5058 new_bus->dev_count--; 5059 memcpy(new_bus->range + i, bus->range + i + 1, 5060 flex_array_size(new_bus, range, new_bus->dev_count - i)); 5061 } 5062 5063 rcu_assign_pointer(kvm->buses[bus_idx], new_bus); 5064 synchronize_srcu_expedited(&kvm->srcu); 5065 5066 /* Destroy the old bus _after_ installing the (null) bus. */ 5067 if (!new_bus) { 5068 pr_err("kvm: failed to shrink bus, removing it completely\n"); 5069 for (j = 0; j < bus->dev_count; j++) { 5070 if (j == i) 5071 continue; 5072 kvm_iodevice_destructor(bus->range[j].dev); 5073 } 5074 } 5075 5076 kfree(bus); 5077 return new_bus ? 0 : -ENOMEM; 5078 } 5079 5080 struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx, 5081 gpa_t addr) 5082 { 5083 struct kvm_io_bus *bus; 5084 int dev_idx, srcu_idx; 5085 struct kvm_io_device *iodev = NULL; 5086 5087 srcu_idx = srcu_read_lock(&kvm->srcu); 5088 5089 bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu); 5090 if (!bus) 5091 goto out_unlock; 5092 5093 dev_idx = kvm_io_bus_get_first_dev(bus, addr, 1); 5094 if (dev_idx < 0) 5095 goto out_unlock; 5096 5097 iodev = bus->range[dev_idx].dev; 5098 5099 out_unlock: 5100 srcu_read_unlock(&kvm->srcu, srcu_idx); 5101 5102 return iodev; 5103 } 5104 EXPORT_SYMBOL_GPL(kvm_io_bus_get_dev); 5105 5106 static int kvm_debugfs_open(struct inode *inode, struct file *file, 5107 int (*get)(void *, u64 *), int (*set)(void *, u64), 5108 const char *fmt) 5109 { 5110 struct kvm_stat_data *stat_data = (struct kvm_stat_data *) 5111 inode->i_private; 5112 5113 /* 5114 * The debugfs files are a reference to the kvm struct which 5115 * is still valid when kvm_destroy_vm is called. kvm_get_kvm_safe 5116 * avoids the race between open and the removal of the debugfs directory. 5117 */ 5118 if (!kvm_get_kvm_safe(stat_data->kvm)) 5119 return -ENOENT; 5120 5121 if (simple_attr_open(inode, file, get, 5122 kvm_stats_debugfs_mode(stat_data->desc) & 0222 5123 ? set : NULL, 5124 fmt)) { 5125 kvm_put_kvm(stat_data->kvm); 5126 return -ENOMEM; 5127 } 5128 5129 return 0; 5130 } 5131 5132 static int kvm_debugfs_release(struct inode *inode, struct file *file) 5133 { 5134 struct kvm_stat_data *stat_data = (struct kvm_stat_data *) 5135 inode->i_private; 5136 5137 simple_attr_release(inode, file); 5138 kvm_put_kvm(stat_data->kvm); 5139 5140 return 0; 5141 } 5142 5143 static int kvm_get_stat_per_vm(struct kvm *kvm, size_t offset, u64 *val) 5144 { 5145 *val = *(u64 *)((void *)(&kvm->stat) + offset); 5146 5147 return 0; 5148 } 5149 5150 static int kvm_clear_stat_per_vm(struct kvm *kvm, size_t offset) 5151 { 5152 *(u64 *)((void *)(&kvm->stat) + offset) = 0; 5153 5154 return 0; 5155 } 5156 5157 static int kvm_get_stat_per_vcpu(struct kvm *kvm, size_t offset, u64 *val) 5158 { 5159 int i; 5160 struct kvm_vcpu *vcpu; 5161 5162 *val = 0; 5163 5164 kvm_for_each_vcpu(i, vcpu, kvm) 5165 *val += *(u64 *)((void *)(&vcpu->stat) + offset); 5166 5167 return 0; 5168 } 5169 5170 static int kvm_clear_stat_per_vcpu(struct kvm *kvm, size_t offset) 5171 { 5172 int i; 5173 struct kvm_vcpu *vcpu; 5174 5175 kvm_for_each_vcpu(i, vcpu, kvm) 5176 *(u64 *)((void *)(&vcpu->stat) + offset) = 0; 5177 5178 return 0; 5179 } 5180 5181 static int kvm_stat_data_get(void *data, u64 *val) 5182 { 5183 int r = -EFAULT; 5184 struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data; 5185 5186 switch (stat_data->kind) { 5187 case KVM_STAT_VM: 5188 r = kvm_get_stat_per_vm(stat_data->kvm, 5189 stat_data->desc->desc.offset, val); 5190 break; 5191 case KVM_STAT_VCPU: 5192 r = kvm_get_stat_per_vcpu(stat_data->kvm, 5193 stat_data->desc->desc.offset, val); 5194 break; 5195 } 5196 5197 return r; 5198 } 5199 5200 static int kvm_stat_data_clear(void *data, u64 val) 5201 { 5202 int r = -EFAULT; 5203 struct kvm_stat_data *stat_data = (struct kvm_stat_data *)data; 5204 5205 if (val) 5206 return -EINVAL; 5207 5208 switch (stat_data->kind) { 5209 case KVM_STAT_VM: 5210 r = kvm_clear_stat_per_vm(stat_data->kvm, 5211 stat_data->desc->desc.offset); 5212 break; 5213 case KVM_STAT_VCPU: 5214 r = kvm_clear_stat_per_vcpu(stat_data->kvm, 5215 stat_data->desc->desc.offset); 5216 break; 5217 } 5218 5219 return r; 5220 } 5221 5222 static int kvm_stat_data_open(struct inode *inode, struct file *file) 5223 { 5224 __simple_attr_check_format("%llu\n", 0ull); 5225 return kvm_debugfs_open(inode, file, kvm_stat_data_get, 5226 kvm_stat_data_clear, "%llu\n"); 5227 } 5228 5229 static const struct file_operations stat_fops_per_vm = { 5230 .owner = THIS_MODULE, 5231 .open = kvm_stat_data_open, 5232 .release = kvm_debugfs_release, 5233 .read = simple_attr_read, 5234 .write = simple_attr_write, 5235 .llseek = no_llseek, 5236 }; 5237 5238 static int vm_stat_get(void *_offset, u64 *val) 5239 { 5240 unsigned offset = (long)_offset; 5241 struct kvm *kvm; 5242 u64 tmp_val; 5243 5244 *val = 0; 5245 mutex_lock(&kvm_lock); 5246 list_for_each_entry(kvm, &vm_list, vm_list) { 5247 kvm_get_stat_per_vm(kvm, offset, &tmp_val); 5248 *val += tmp_val; 5249 } 5250 mutex_unlock(&kvm_lock); 5251 return 0; 5252 } 5253 5254 static int vm_stat_clear(void *_offset, u64 val) 5255 { 5256 unsigned offset = (long)_offset; 5257 struct kvm *kvm; 5258 5259 if (val) 5260 return -EINVAL; 5261 5262 mutex_lock(&kvm_lock); 5263 list_for_each_entry(kvm, &vm_list, vm_list) { 5264 kvm_clear_stat_per_vm(kvm, offset); 5265 } 5266 mutex_unlock(&kvm_lock); 5267 5268 return 0; 5269 } 5270 5271 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, vm_stat_clear, "%llu\n"); 5272 DEFINE_SIMPLE_ATTRIBUTE(vm_stat_readonly_fops, vm_stat_get, NULL, "%llu\n"); 5273 5274 static int vcpu_stat_get(void *_offset, u64 *val) 5275 { 5276 unsigned offset = (long)_offset; 5277 struct kvm *kvm; 5278 u64 tmp_val; 5279 5280 *val = 0; 5281 mutex_lock(&kvm_lock); 5282 list_for_each_entry(kvm, &vm_list, vm_list) { 5283 kvm_get_stat_per_vcpu(kvm, offset, &tmp_val); 5284 *val += tmp_val; 5285 } 5286 mutex_unlock(&kvm_lock); 5287 return 0; 5288 } 5289 5290 static int vcpu_stat_clear(void *_offset, u64 val) 5291 { 5292 unsigned offset = (long)_offset; 5293 struct kvm *kvm; 5294 5295 if (val) 5296 return -EINVAL; 5297 5298 mutex_lock(&kvm_lock); 5299 list_for_each_entry(kvm, &vm_list, vm_list) { 5300 kvm_clear_stat_per_vcpu(kvm, offset); 5301 } 5302 mutex_unlock(&kvm_lock); 5303 5304 return 0; 5305 } 5306 5307 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, vcpu_stat_clear, 5308 "%llu\n"); 5309 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_readonly_fops, vcpu_stat_get, NULL, "%llu\n"); 5310 5311 static void kvm_uevent_notify_change(unsigned int type, struct kvm *kvm) 5312 { 5313 struct kobj_uevent_env *env; 5314 unsigned long long created, active; 5315 5316 if (!kvm_dev.this_device || !kvm) 5317 return; 5318 5319 mutex_lock(&kvm_lock); 5320 if (type == KVM_EVENT_CREATE_VM) { 5321 kvm_createvm_count++; 5322 kvm_active_vms++; 5323 } else if (type == KVM_EVENT_DESTROY_VM) { 5324 kvm_active_vms--; 5325 } 5326 created = kvm_createvm_count; 5327 active = kvm_active_vms; 5328 mutex_unlock(&kvm_lock); 5329 5330 env = kzalloc(sizeof(*env), GFP_KERNEL_ACCOUNT); 5331 if (!env) 5332 return; 5333 5334 add_uevent_var(env, "CREATED=%llu", created); 5335 add_uevent_var(env, "COUNT=%llu", active); 5336 5337 if (type == KVM_EVENT_CREATE_VM) { 5338 add_uevent_var(env, "EVENT=create"); 5339 kvm->userspace_pid = task_pid_nr(current); 5340 } else if (type == KVM_EVENT_DESTROY_VM) { 5341 add_uevent_var(env, "EVENT=destroy"); 5342 } 5343 add_uevent_var(env, "PID=%d", kvm->userspace_pid); 5344 5345 if (kvm->debugfs_dentry) { 5346 char *tmp, *p = kmalloc(PATH_MAX, GFP_KERNEL_ACCOUNT); 5347 5348 if (p) { 5349 tmp = dentry_path_raw(kvm->debugfs_dentry, p, PATH_MAX); 5350 if (!IS_ERR(tmp)) 5351 add_uevent_var(env, "STATS_PATH=%s", tmp); 5352 kfree(p); 5353 } 5354 } 5355 /* no need for checks, since we are adding at most only 5 keys */ 5356 env->envp[env->envp_idx++] = NULL; 5357 kobject_uevent_env(&kvm_dev.this_device->kobj, KOBJ_CHANGE, env->envp); 5358 kfree(env); 5359 } 5360 5361 static void kvm_init_debug(void) 5362 { 5363 const struct file_operations *fops; 5364 const struct _kvm_stats_desc *pdesc; 5365 int i; 5366 5367 kvm_debugfs_dir = debugfs_create_dir("kvm", NULL); 5368 5369 for (i = 0; i < kvm_vm_stats_header.num_desc; ++i) { 5370 pdesc = &kvm_vm_stats_desc[i]; 5371 if (kvm_stats_debugfs_mode(pdesc) & 0222) 5372 fops = &vm_stat_fops; 5373 else 5374 fops = &vm_stat_readonly_fops; 5375 debugfs_create_file(pdesc->name, kvm_stats_debugfs_mode(pdesc), 5376 kvm_debugfs_dir, 5377 (void *)(long)pdesc->desc.offset, fops); 5378 } 5379 5380 for (i = 0; i < kvm_vcpu_stats_header.num_desc; ++i) { 5381 pdesc = &kvm_vcpu_stats_desc[i]; 5382 if (kvm_stats_debugfs_mode(pdesc) & 0222) 5383 fops = &vcpu_stat_fops; 5384 else 5385 fops = &vcpu_stat_readonly_fops; 5386 debugfs_create_file(pdesc->name, kvm_stats_debugfs_mode(pdesc), 5387 kvm_debugfs_dir, 5388 (void *)(long)pdesc->desc.offset, fops); 5389 } 5390 } 5391 5392 static int kvm_suspend(void) 5393 { 5394 if (kvm_usage_count) 5395 hardware_disable_nolock(NULL); 5396 return 0; 5397 } 5398 5399 static void kvm_resume(void) 5400 { 5401 if (kvm_usage_count) { 5402 #ifdef CONFIG_LOCKDEP 5403 WARN_ON(lockdep_is_held(&kvm_count_lock)); 5404 #endif 5405 hardware_enable_nolock(NULL); 5406 } 5407 } 5408 5409 static struct syscore_ops kvm_syscore_ops = { 5410 .suspend = kvm_suspend, 5411 .resume = kvm_resume, 5412 }; 5413 5414 static inline 5415 struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn) 5416 { 5417 return container_of(pn, struct kvm_vcpu, preempt_notifier); 5418 } 5419 5420 static void kvm_sched_in(struct preempt_notifier *pn, int cpu) 5421 { 5422 struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn); 5423 5424 WRITE_ONCE(vcpu->preempted, false); 5425 WRITE_ONCE(vcpu->ready, false); 5426 5427 __this_cpu_write(kvm_running_vcpu, vcpu); 5428 kvm_arch_sched_in(vcpu, cpu); 5429 kvm_arch_vcpu_load(vcpu, cpu); 5430 } 5431 5432 static void kvm_sched_out(struct preempt_notifier *pn, 5433 struct task_struct *next) 5434 { 5435 struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn); 5436 5437 if (current->on_rq) { 5438 WRITE_ONCE(vcpu->preempted, true); 5439 WRITE_ONCE(vcpu->ready, true); 5440 } 5441 kvm_arch_vcpu_put(vcpu); 5442 __this_cpu_write(kvm_running_vcpu, NULL); 5443 } 5444 5445 /** 5446 * kvm_get_running_vcpu - get the vcpu running on the current CPU. 5447 * 5448 * We can disable preemption locally around accessing the per-CPU variable, 5449 * and use the resolved vcpu pointer after enabling preemption again, 5450 * because even if the current thread is migrated to another CPU, reading 5451 * the per-CPU value later will give us the same value as we update the 5452 * per-CPU variable in the preempt notifier handlers. 5453 */ 5454 struct kvm_vcpu *kvm_get_running_vcpu(void) 5455 { 5456 struct kvm_vcpu *vcpu; 5457 5458 preempt_disable(); 5459 vcpu = __this_cpu_read(kvm_running_vcpu); 5460 preempt_enable(); 5461 5462 return vcpu; 5463 } 5464 EXPORT_SYMBOL_GPL(kvm_get_running_vcpu); 5465 5466 /** 5467 * kvm_get_running_vcpus - get the per-CPU array of currently running vcpus. 5468 */ 5469 struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void) 5470 { 5471 return &kvm_running_vcpu; 5472 } 5473 5474 struct kvm_cpu_compat_check { 5475 void *opaque; 5476 int *ret; 5477 }; 5478 5479 static void check_processor_compat(void *data) 5480 { 5481 struct kvm_cpu_compat_check *c = data; 5482 5483 *c->ret = kvm_arch_check_processor_compat(c->opaque); 5484 } 5485 5486 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align, 5487 struct module *module) 5488 { 5489 struct kvm_cpu_compat_check c; 5490 int r; 5491 int cpu; 5492 5493 r = kvm_arch_init(opaque); 5494 if (r) 5495 goto out_fail; 5496 5497 /* 5498 * kvm_arch_init makes sure there's at most one caller 5499 * for architectures that support multiple implementations, 5500 * like intel and amd on x86. 5501 * kvm_arch_init must be called before kvm_irqfd_init to avoid creating 5502 * conflicts in case kvm is already setup for another implementation. 5503 */ 5504 r = kvm_irqfd_init(); 5505 if (r) 5506 goto out_irqfd; 5507 5508 if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) { 5509 r = -ENOMEM; 5510 goto out_free_0; 5511 } 5512 5513 r = kvm_arch_hardware_setup(opaque); 5514 if (r < 0) 5515 goto out_free_1; 5516 5517 c.ret = &r; 5518 c.opaque = opaque; 5519 for_each_online_cpu(cpu) { 5520 smp_call_function_single(cpu, check_processor_compat, &c, 1); 5521 if (r < 0) 5522 goto out_free_2; 5523 } 5524 5525 r = cpuhp_setup_state_nocalls(CPUHP_AP_KVM_STARTING, "kvm/cpu:starting", 5526 kvm_starting_cpu, kvm_dying_cpu); 5527 if (r) 5528 goto out_free_2; 5529 register_reboot_notifier(&kvm_reboot_notifier); 5530 5531 /* A kmem cache lets us meet the alignment requirements of fx_save. */ 5532 if (!vcpu_align) 5533 vcpu_align = __alignof__(struct kvm_vcpu); 5534 kvm_vcpu_cache = 5535 kmem_cache_create_usercopy("kvm_vcpu", vcpu_size, vcpu_align, 5536 SLAB_ACCOUNT, 5537 offsetof(struct kvm_vcpu, arch), 5538 offsetofend(struct kvm_vcpu, stats_id) 5539 - offsetof(struct kvm_vcpu, arch), 5540 NULL); 5541 if (!kvm_vcpu_cache) { 5542 r = -ENOMEM; 5543 goto out_free_3; 5544 } 5545 5546 r = kvm_async_pf_init(); 5547 if (r) 5548 goto out_free; 5549 5550 kvm_chardev_ops.owner = module; 5551 kvm_vm_fops.owner = module; 5552 kvm_vcpu_fops.owner = module; 5553 5554 r = misc_register(&kvm_dev); 5555 if (r) { 5556 pr_err("kvm: misc device register failed\n"); 5557 goto out_unreg; 5558 } 5559 5560 register_syscore_ops(&kvm_syscore_ops); 5561 5562 kvm_preempt_ops.sched_in = kvm_sched_in; 5563 kvm_preempt_ops.sched_out = kvm_sched_out; 5564 5565 kvm_init_debug(); 5566 5567 r = kvm_vfio_ops_init(); 5568 WARN_ON(r); 5569 5570 return 0; 5571 5572 out_unreg: 5573 kvm_async_pf_deinit(); 5574 out_free: 5575 kmem_cache_destroy(kvm_vcpu_cache); 5576 out_free_3: 5577 unregister_reboot_notifier(&kvm_reboot_notifier); 5578 cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING); 5579 out_free_2: 5580 kvm_arch_hardware_unsetup(); 5581 out_free_1: 5582 free_cpumask_var(cpus_hardware_enabled); 5583 out_free_0: 5584 kvm_irqfd_exit(); 5585 out_irqfd: 5586 kvm_arch_exit(); 5587 out_fail: 5588 return r; 5589 } 5590 EXPORT_SYMBOL_GPL(kvm_init); 5591 5592 void kvm_exit(void) 5593 { 5594 debugfs_remove_recursive(kvm_debugfs_dir); 5595 misc_deregister(&kvm_dev); 5596 kmem_cache_destroy(kvm_vcpu_cache); 5597 kvm_async_pf_deinit(); 5598 unregister_syscore_ops(&kvm_syscore_ops); 5599 unregister_reboot_notifier(&kvm_reboot_notifier); 5600 cpuhp_remove_state_nocalls(CPUHP_AP_KVM_STARTING); 5601 on_each_cpu(hardware_disable_nolock, NULL, 1); 5602 kvm_arch_hardware_unsetup(); 5603 kvm_arch_exit(); 5604 kvm_irqfd_exit(); 5605 free_cpumask_var(cpus_hardware_enabled); 5606 kvm_vfio_ops_exit(); 5607 } 5608 EXPORT_SYMBOL_GPL(kvm_exit); 5609 5610 struct kvm_vm_worker_thread_context { 5611 struct kvm *kvm; 5612 struct task_struct *parent; 5613 struct completion init_done; 5614 kvm_vm_thread_fn_t thread_fn; 5615 uintptr_t data; 5616 int err; 5617 }; 5618 5619 static int kvm_vm_worker_thread(void *context) 5620 { 5621 /* 5622 * The init_context is allocated on the stack of the parent thread, so 5623 * we have to locally copy anything that is needed beyond initialization 5624 */ 5625 struct kvm_vm_worker_thread_context *init_context = context; 5626 struct kvm *kvm = init_context->kvm; 5627 kvm_vm_thread_fn_t thread_fn = init_context->thread_fn; 5628 uintptr_t data = init_context->data; 5629 int err; 5630 5631 err = kthread_park(current); 5632 /* kthread_park(current) is never supposed to return an error */ 5633 WARN_ON(err != 0); 5634 if (err) 5635 goto init_complete; 5636 5637 err = cgroup_attach_task_all(init_context->parent, current); 5638 if (err) { 5639 kvm_err("%s: cgroup_attach_task_all failed with err %d\n", 5640 __func__, err); 5641 goto init_complete; 5642 } 5643 5644 set_user_nice(current, task_nice(init_context->parent)); 5645 5646 init_complete: 5647 init_context->err = err; 5648 complete(&init_context->init_done); 5649 init_context = NULL; 5650 5651 if (err) 5652 return err; 5653 5654 /* Wait to be woken up by the spawner before proceeding. */ 5655 kthread_parkme(); 5656 5657 if (!kthread_should_stop()) 5658 err = thread_fn(kvm, data); 5659 5660 return err; 5661 } 5662 5663 int kvm_vm_create_worker_thread(struct kvm *kvm, kvm_vm_thread_fn_t thread_fn, 5664 uintptr_t data, const char *name, 5665 struct task_struct **thread_ptr) 5666 { 5667 struct kvm_vm_worker_thread_context init_context = {}; 5668 struct task_struct *thread; 5669 5670 *thread_ptr = NULL; 5671 init_context.kvm = kvm; 5672 init_context.parent = current; 5673 init_context.thread_fn = thread_fn; 5674 init_context.data = data; 5675 init_completion(&init_context.init_done); 5676 5677 thread = kthread_run(kvm_vm_worker_thread, &init_context, 5678 "%s-%d", name, task_pid_nr(current)); 5679 if (IS_ERR(thread)) 5680 return PTR_ERR(thread); 5681 5682 /* kthread_run is never supposed to return NULL */ 5683 WARN_ON(thread == NULL); 5684 5685 wait_for_completion(&init_context.init_done); 5686 5687 if (!init_context.err) 5688 *thread_ptr = thread; 5689 5690 return init_context.err; 5691 } 5692