1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * KFENCE guarded object allocator and fault handling. 4 * 5 * Copyright (C) 2020, Google LLC. 6 */ 7 8 #define pr_fmt(fmt) "kfence: " fmt 9 10 #include <linux/atomic.h> 11 #include <linux/bug.h> 12 #include <linux/debugfs.h> 13 #include <linux/hash.h> 14 #include <linux/irq_work.h> 15 #include <linux/jhash.h> 16 #include <linux/kcsan-checks.h> 17 #include <linux/kfence.h> 18 #include <linux/kmemleak.h> 19 #include <linux/list.h> 20 #include <linux/lockdep.h> 21 #include <linux/log2.h> 22 #include <linux/memblock.h> 23 #include <linux/moduleparam.h> 24 #include <linux/nodemask.h> 25 #include <linux/notifier.h> 26 #include <linux/panic_notifier.h> 27 #include <linux/random.h> 28 #include <linux/rcupdate.h> 29 #include <linux/sched/clock.h> 30 #include <linux/seq_file.h> 31 #include <linux/slab.h> 32 #include <linux/spinlock.h> 33 #include <linux/string.h> 34 35 #include <asm/kfence.h> 36 37 #include "kfence.h" 38 39 /* Disables KFENCE on the first warning assuming an irrecoverable error. */ 40 #define KFENCE_WARN_ON(cond) \ 41 ({ \ 42 const bool __cond = WARN_ON(cond); \ 43 if (unlikely(__cond)) { \ 44 WRITE_ONCE(kfence_enabled, false); \ 45 disabled_by_warn = true; \ 46 } \ 47 __cond; \ 48 }) 49 50 /* === Data ================================================================= */ 51 52 static bool kfence_enabled __read_mostly; 53 static bool disabled_by_warn __read_mostly; 54 55 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL; 56 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */ 57 58 #ifdef MODULE_PARAM_PREFIX 59 #undef MODULE_PARAM_PREFIX 60 #endif 61 #define MODULE_PARAM_PREFIX "kfence." 62 63 static int kfence_enable_late(void); 64 static int param_set_sample_interval(const char *val, const struct kernel_param *kp) 65 { 66 unsigned long num; 67 int ret = kstrtoul(val, 0, &num); 68 69 if (ret < 0) 70 return ret; 71 72 /* Using 0 to indicate KFENCE is disabled. */ 73 if (!num && READ_ONCE(kfence_enabled)) { 74 pr_info("disabled\n"); 75 WRITE_ONCE(kfence_enabled, false); 76 } 77 78 *((unsigned long *)kp->arg) = num; 79 80 if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING) 81 return disabled_by_warn ? -EINVAL : kfence_enable_late(); 82 return 0; 83 } 84 85 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp) 86 { 87 if (!READ_ONCE(kfence_enabled)) 88 return sprintf(buffer, "0\n"); 89 90 return param_get_ulong(buffer, kp); 91 } 92 93 static const struct kernel_param_ops sample_interval_param_ops = { 94 .set = param_set_sample_interval, 95 .get = param_get_sample_interval, 96 }; 97 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600); 98 99 /* Pool usage% threshold when currently covered allocations are skipped. */ 100 static unsigned long kfence_skip_covered_thresh __read_mostly = 75; 101 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644); 102 103 /* Allocation burst count: number of excess KFENCE allocations per sample. */ 104 static unsigned int kfence_burst __read_mostly; 105 module_param_named(burst, kfence_burst, uint, 0644); 106 107 /* If true, use a deferrable timer. */ 108 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE); 109 module_param_named(deferrable, kfence_deferrable, bool, 0444); 110 111 /* If true, check all canary bytes on panic. */ 112 static bool kfence_check_on_panic __read_mostly; 113 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444); 114 115 /* The pool of pages used for guard pages and objects. */ 116 char *__kfence_pool __read_mostly; 117 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */ 118 119 /* 120 * Per-object metadata, with one-to-one mapping of object metadata to 121 * backing pages (in __kfence_pool). 122 */ 123 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0); 124 struct kfence_metadata *kfence_metadata __read_mostly; 125 126 /* 127 * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache(). 128 * So introduce kfence_metadata_init to initialize metadata, and then make 129 * kfence_metadata visible after initialization is successful. This prevents 130 * potential UAF or access to uninitialized metadata. 131 */ 132 static struct kfence_metadata *kfence_metadata_init __read_mostly; 133 134 /* Freelist with available objects. */ 135 static struct list_head kfence_freelist = LIST_HEAD_INIT(kfence_freelist); 136 static DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */ 137 138 /* 139 * The static key to set up a KFENCE allocation; or if static keys are not used 140 * to gate allocations, to avoid a load and compare if KFENCE is disabled. 141 */ 142 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key); 143 144 /* Gates the allocation, ensuring only one succeeds in a given period. */ 145 atomic_t kfence_allocation_gate = ATOMIC_INIT(1); 146 147 /* 148 * A Counting Bloom filter of allocation coverage: limits currently covered 149 * allocations of the same source filling up the pool. 150 * 151 * Assuming a range of 15%-85% unique allocations in the pool at any point in 152 * time, the below parameters provide a probablity of 0.02-0.33 for false 153 * positive hits respectively: 154 * 155 * P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM 156 */ 157 #define ALLOC_COVERED_HNUM 2 158 #define ALLOC_COVERED_ORDER (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2) 159 #define ALLOC_COVERED_SIZE (1 << ALLOC_COVERED_ORDER) 160 #define ALLOC_COVERED_HNEXT(h) hash_32(h, ALLOC_COVERED_ORDER) 161 #define ALLOC_COVERED_MASK (ALLOC_COVERED_SIZE - 1) 162 static atomic_t alloc_covered[ALLOC_COVERED_SIZE]; 163 164 /* Stack depth used to determine uniqueness of an allocation. */ 165 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8) 166 167 /* 168 * Randomness for stack hashes, making the same collisions across reboots and 169 * different machines less likely. 170 */ 171 static u32 stack_hash_seed __ro_after_init; 172 173 /* Statistics counters for debugfs. */ 174 enum kfence_counter_id { 175 KFENCE_COUNTER_ALLOCATED, 176 KFENCE_COUNTER_ALLOCS, 177 KFENCE_COUNTER_FREES, 178 KFENCE_COUNTER_ZOMBIES, 179 KFENCE_COUNTER_BUGS, 180 KFENCE_COUNTER_SKIP_INCOMPAT, 181 KFENCE_COUNTER_SKIP_CAPACITY, 182 KFENCE_COUNTER_SKIP_COVERED, 183 KFENCE_COUNTER_COUNT, 184 }; 185 static atomic_long_t counters[KFENCE_COUNTER_COUNT]; 186 static const char *const counter_names[] = { 187 [KFENCE_COUNTER_ALLOCATED] = "currently allocated", 188 [KFENCE_COUNTER_ALLOCS] = "total allocations", 189 [KFENCE_COUNTER_FREES] = "total frees", 190 [KFENCE_COUNTER_ZOMBIES] = "zombie allocations", 191 [KFENCE_COUNTER_BUGS] = "total bugs", 192 [KFENCE_COUNTER_SKIP_INCOMPAT] = "skipped allocations (incompatible)", 193 [KFENCE_COUNTER_SKIP_CAPACITY] = "skipped allocations (capacity)", 194 [KFENCE_COUNTER_SKIP_COVERED] = "skipped allocations (covered)", 195 }; 196 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT); 197 198 /* === Internals ============================================================ */ 199 200 static inline bool should_skip_covered(void) 201 { 202 unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100; 203 204 return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh; 205 } 206 207 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries) 208 { 209 num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH); 210 num_entries = filter_irq_stacks(stack_entries, num_entries); 211 return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed); 212 } 213 214 /* 215 * Adds (or subtracts) count @val for allocation stack trace hash 216 * @alloc_stack_hash from Counting Bloom filter. 217 */ 218 static void alloc_covered_add(u32 alloc_stack_hash, int val) 219 { 220 int i; 221 222 for (i = 0; i < ALLOC_COVERED_HNUM; i++) { 223 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]); 224 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash); 225 } 226 } 227 228 /* 229 * Returns true if the allocation stack trace hash @alloc_stack_hash is 230 * currently contained (non-zero count) in Counting Bloom filter. 231 */ 232 static bool alloc_covered_contains(u32 alloc_stack_hash) 233 { 234 int i; 235 236 for (i = 0; i < ALLOC_COVERED_HNUM; i++) { 237 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK])) 238 return false; 239 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash); 240 } 241 242 return true; 243 } 244 245 static bool kfence_protect(unsigned long addr) 246 { 247 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true)); 248 } 249 250 static bool kfence_unprotect(unsigned long addr) 251 { 252 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false)); 253 } 254 255 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta) 256 { 257 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2; 258 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset]; 259 260 /* The checks do not affect performance; only called from slow-paths. */ 261 262 /* Only call with a pointer into kfence_metadata. */ 263 if (KFENCE_WARN_ON(meta < kfence_metadata || 264 meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS)) 265 return 0; 266 267 /* 268 * This metadata object only ever maps to 1 page; verify that the stored 269 * address is in the expected range. 270 */ 271 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr)) 272 return 0; 273 274 return pageaddr; 275 } 276 277 static inline bool kfence_obj_allocated(const struct kfence_metadata *meta) 278 { 279 enum kfence_object_state state = READ_ONCE(meta->state); 280 281 return state == KFENCE_OBJECT_ALLOCATED || state == KFENCE_OBJECT_RCU_FREEING; 282 } 283 284 /* 285 * Update the object's metadata state, including updating the alloc/free stacks 286 * depending on the state transition. 287 */ 288 static noinline void 289 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next, 290 unsigned long *stack_entries, size_t num_stack_entries) 291 { 292 struct kfence_track *track = 293 next == KFENCE_OBJECT_ALLOCATED ? &meta->alloc_track : &meta->free_track; 294 295 lockdep_assert_held(&meta->lock); 296 297 /* Stack has been saved when calling rcu, skip. */ 298 if (READ_ONCE(meta->state) == KFENCE_OBJECT_RCU_FREEING) 299 goto out; 300 301 if (stack_entries) { 302 memcpy(track->stack_entries, stack_entries, 303 num_stack_entries * sizeof(stack_entries[0])); 304 } else { 305 /* 306 * Skip over 1 (this) functions; noinline ensures we do not 307 * accidentally skip over the caller by never inlining. 308 */ 309 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1); 310 } 311 track->num_stack_entries = num_stack_entries; 312 track->pid = task_pid_nr(current); 313 track->cpu = raw_smp_processor_id(); 314 track->ts_nsec = local_clock(); /* Same source as printk timestamps. */ 315 316 out: 317 /* 318 * Pairs with READ_ONCE() in 319 * kfence_shutdown_cache(), 320 * kfence_handle_page_fault(). 321 */ 322 WRITE_ONCE(meta->state, next); 323 } 324 325 #ifdef CONFIG_KMSAN 326 #define check_canary_attributes noinline __no_kmsan_checks 327 #else 328 #define check_canary_attributes inline 329 #endif 330 331 /* Check canary byte at @addr. */ 332 static check_canary_attributes bool check_canary_byte(u8 *addr) 333 { 334 struct kfence_metadata *meta; 335 unsigned long flags; 336 337 if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr))) 338 return true; 339 340 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]); 341 342 meta = addr_to_metadata((unsigned long)addr); 343 raw_spin_lock_irqsave(&meta->lock, flags); 344 kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION); 345 raw_spin_unlock_irqrestore(&meta->lock, flags); 346 347 return false; 348 } 349 350 static inline void set_canary(const struct kfence_metadata *meta) 351 { 352 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE); 353 unsigned long addr = pageaddr; 354 355 /* 356 * The canary may be written to part of the object memory, but it does 357 * not affect it. The user should initialize the object before using it. 358 */ 359 for (; addr < meta->addr; addr += sizeof(u64)) 360 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64; 361 362 addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64)); 363 for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) 364 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64; 365 } 366 367 static check_canary_attributes void 368 check_canary(const struct kfence_metadata *meta) 369 { 370 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE); 371 unsigned long addr = pageaddr; 372 373 /* 374 * We'll iterate over each canary byte per-side until a corrupted byte 375 * is found. However, we'll still iterate over the canary bytes to the 376 * right of the object even if there was an error in the canary bytes to 377 * the left of the object. Specifically, if check_canary_byte() 378 * generates an error, showing both sides might give more clues as to 379 * what the error is about when displaying which bytes were corrupted. 380 */ 381 382 /* Apply to left of object. */ 383 for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) { 384 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) 385 break; 386 } 387 388 /* 389 * If the canary is corrupted in a certain 64 bytes, or the canary 390 * memory cannot be completely covered by multiple consecutive 64 bytes, 391 * it needs to be checked one by one. 392 */ 393 for (; addr < meta->addr; addr++) { 394 if (unlikely(!check_canary_byte((u8 *)addr))) 395 break; 396 } 397 398 /* Apply to right of object. */ 399 for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) { 400 if (unlikely(!check_canary_byte((u8 *)addr))) 401 return; 402 } 403 for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) { 404 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) { 405 406 for (; addr - pageaddr < PAGE_SIZE; addr++) { 407 if (!check_canary_byte((u8 *)addr)) 408 return; 409 } 410 } 411 } 412 } 413 414 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp, 415 unsigned long *stack_entries, size_t num_stack_entries, 416 u32 alloc_stack_hash) 417 { 418 struct kfence_metadata *meta = NULL; 419 unsigned long flags; 420 struct slab *slab; 421 void *addr; 422 const bool random_right_allocate = get_random_u32_below(2); 423 const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS && 424 !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS); 425 426 /* Try to obtain a free object. */ 427 raw_spin_lock_irqsave(&kfence_freelist_lock, flags); 428 if (!list_empty(&kfence_freelist)) { 429 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list); 430 list_del_init(&meta->list); 431 } 432 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags); 433 if (!meta) { 434 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]); 435 return NULL; 436 } 437 438 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) { 439 /* 440 * This is extremely unlikely -- we are reporting on a 441 * use-after-free, which locked meta->lock, and the reporting 442 * code via printk calls kmalloc() which ends up in 443 * kfence_alloc() and tries to grab the same object that we're 444 * reporting on. While it has never been observed, lockdep does 445 * report that there is a possibility of deadlock. Fix it by 446 * using trylock and bailing out gracefully. 447 */ 448 raw_spin_lock_irqsave(&kfence_freelist_lock, flags); 449 /* Put the object back on the freelist. */ 450 list_add_tail(&meta->list, &kfence_freelist); 451 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags); 452 453 return NULL; 454 } 455 456 meta->addr = metadata_to_pageaddr(meta); 457 /* Unprotect if we're reusing this page. */ 458 if (meta->state == KFENCE_OBJECT_FREED) 459 kfence_unprotect(meta->addr); 460 461 /* 462 * Note: for allocations made before RNG initialization, will always 463 * return zero. We still benefit from enabling KFENCE as early as 464 * possible, even when the RNG is not yet available, as this will allow 465 * KFENCE to detect bugs due to earlier allocations. The only downside 466 * is that the out-of-bounds accesses detected are deterministic for 467 * such allocations. 468 */ 469 if (random_right_allocate) { 470 /* Allocate on the "right" side, re-calculate address. */ 471 meta->addr += PAGE_SIZE - size; 472 meta->addr = ALIGN_DOWN(meta->addr, cache->align); 473 } 474 475 addr = (void *)meta->addr; 476 477 /* Update remaining metadata. */ 478 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries); 479 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */ 480 WRITE_ONCE(meta->cache, cache); 481 meta->size = size; 482 meta->alloc_stack_hash = alloc_stack_hash; 483 raw_spin_unlock_irqrestore(&meta->lock, flags); 484 485 alloc_covered_add(alloc_stack_hash, 1); 486 487 /* Set required slab fields. */ 488 slab = virt_to_slab((void *)meta->addr); 489 slab->slab_cache = cache; 490 slab->objects = 1; 491 492 /* Memory initialization. */ 493 set_canary(meta); 494 495 /* 496 * We check slab_want_init_on_alloc() ourselves, rather than letting 497 * SL*B do the initialization, as otherwise we might overwrite KFENCE's 498 * redzone. 499 */ 500 if (unlikely(slab_want_init_on_alloc(gfp, cache))) 501 memzero_explicit(addr, size); 502 if (cache->ctor) 503 cache->ctor(addr); 504 505 if (random_fault) 506 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */ 507 508 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]); 509 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]); 510 511 return addr; 512 } 513 514 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie) 515 { 516 struct kcsan_scoped_access assert_page_exclusive; 517 unsigned long flags; 518 bool init; 519 520 raw_spin_lock_irqsave(&meta->lock, flags); 521 522 if (!kfence_obj_allocated(meta) || meta->addr != (unsigned long)addr) { 523 /* Invalid or double-free, bail out. */ 524 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]); 525 kfence_report_error((unsigned long)addr, false, NULL, meta, 526 KFENCE_ERROR_INVALID_FREE); 527 raw_spin_unlock_irqrestore(&meta->lock, flags); 528 return; 529 } 530 531 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */ 532 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE, 533 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT, 534 &assert_page_exclusive); 535 536 if (CONFIG_KFENCE_STRESS_TEST_FAULTS) 537 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */ 538 539 /* Restore page protection if there was an OOB access. */ 540 if (meta->unprotected_page) { 541 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE); 542 kfence_protect(meta->unprotected_page); 543 meta->unprotected_page = 0; 544 } 545 546 /* Mark the object as freed. */ 547 metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0); 548 init = slab_want_init_on_free(meta->cache); 549 raw_spin_unlock_irqrestore(&meta->lock, flags); 550 551 alloc_covered_add(meta->alloc_stack_hash, -1); 552 553 /* Check canary bytes for memory corruption. */ 554 check_canary(meta); 555 556 /* 557 * Clear memory if init-on-free is set. While we protect the page, the 558 * data is still there, and after a use-after-free is detected, we 559 * unprotect the page, so the data is still accessible. 560 */ 561 if (!zombie && unlikely(init)) 562 memzero_explicit(addr, meta->size); 563 564 /* Protect to detect use-after-frees. */ 565 kfence_protect((unsigned long)addr); 566 567 kcsan_end_scoped_access(&assert_page_exclusive); 568 if (!zombie) { 569 /* Add it to the tail of the freelist for reuse. */ 570 raw_spin_lock_irqsave(&kfence_freelist_lock, flags); 571 KFENCE_WARN_ON(!list_empty(&meta->list)); 572 list_add_tail(&meta->list, &kfence_freelist); 573 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags); 574 575 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]); 576 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]); 577 } else { 578 /* See kfence_shutdown_cache(). */ 579 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]); 580 } 581 } 582 583 static void rcu_guarded_free(struct rcu_head *h) 584 { 585 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head); 586 587 kfence_guarded_free((void *)meta->addr, meta, false); 588 } 589 590 /* 591 * Initialization of the KFENCE pool after its allocation. 592 * Returns 0 on success; otherwise returns the address up to 593 * which partial initialization succeeded. 594 */ 595 static unsigned long kfence_init_pool(void) 596 { 597 unsigned long addr, start_pfn; 598 int i; 599 600 if (!arch_kfence_init_pool()) 601 return (unsigned long)__kfence_pool; 602 603 addr = (unsigned long)__kfence_pool; 604 start_pfn = PHYS_PFN(virt_to_phys(__kfence_pool)); 605 606 /* 607 * Set up object pages: they must have PGTY_slab set to avoid freeing 608 * them as real pages. 609 * 610 * We also want to avoid inserting kfence_free() in the kfree() 611 * fast-path in SLUB, and therefore need to ensure kfree() correctly 612 * enters __slab_free() slow-path. 613 */ 614 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) { 615 struct page *page; 616 617 if (!i || (i % 2)) 618 continue; 619 620 page = pfn_to_page(start_pfn + i); 621 __SetPageSlab(page); 622 #ifdef CONFIG_MEMCG 623 struct slab *slab = page_slab(page); 624 slab->obj_exts = (unsigned long)&kfence_metadata_init[i / 2 - 1].obj_exts | 625 MEMCG_DATA_OBJEXTS; 626 #endif 627 } 628 629 /* 630 * Protect the first 2 pages. The first page is mostly unnecessary, and 631 * merely serves as an extended guard page. However, adding one 632 * additional page in the beginning gives us an even number of pages, 633 * which simplifies the mapping of address to metadata index. 634 */ 635 for (i = 0; i < 2; i++) { 636 if (unlikely(!kfence_protect(addr))) 637 return addr; 638 639 addr += PAGE_SIZE; 640 } 641 642 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 643 struct kfence_metadata *meta = &kfence_metadata_init[i]; 644 645 /* Initialize metadata. */ 646 INIT_LIST_HEAD(&meta->list); 647 raw_spin_lock_init(&meta->lock); 648 meta->state = KFENCE_OBJECT_UNUSED; 649 meta->addr = addr; /* Initialize for validation in metadata_to_pageaddr(). */ 650 list_add_tail(&meta->list, &kfence_freelist); 651 652 /* Protect the right redzone. */ 653 if (unlikely(!kfence_protect(addr + PAGE_SIZE))) 654 goto reset_slab; 655 656 addr += 2 * PAGE_SIZE; 657 } 658 659 /* 660 * Make kfence_metadata visible only when initialization is successful. 661 * Otherwise, if the initialization fails and kfence_metadata is freed, 662 * it may cause UAF in kfence_shutdown_cache(). 663 */ 664 smp_store_release(&kfence_metadata, kfence_metadata_init); 665 return 0; 666 667 reset_slab: 668 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) { 669 struct page *page; 670 671 if (!i || (i % 2)) 672 continue; 673 674 page = pfn_to_page(start_pfn + i); 675 #ifdef CONFIG_MEMCG 676 struct slab *slab = page_slab(page); 677 slab->obj_exts = 0; 678 #endif 679 __ClearPageSlab(page); 680 } 681 682 return addr; 683 } 684 685 static bool __init kfence_init_pool_early(void) 686 { 687 unsigned long addr; 688 689 if (!__kfence_pool) 690 return false; 691 692 addr = kfence_init_pool(); 693 694 if (!addr) { 695 /* 696 * The pool is live and will never be deallocated from this point on. 697 * Ignore the pool object from the kmemleak phys object tree, as it would 698 * otherwise overlap with allocations returned by kfence_alloc(), which 699 * are registered with kmemleak through the slab post-alloc hook. 700 */ 701 kmemleak_ignore_phys(__pa(__kfence_pool)); 702 return true; 703 } 704 705 /* 706 * Only release unprotected pages, and do not try to go back and change 707 * page attributes due to risk of failing to do so as well. If changing 708 * page attributes for some pages fails, it is very likely that it also 709 * fails for the first page, and therefore expect addr==__kfence_pool in 710 * most failure cases. 711 */ 712 memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool)); 713 __kfence_pool = NULL; 714 715 memblock_free_late(__pa(kfence_metadata_init), KFENCE_METADATA_SIZE); 716 kfence_metadata_init = NULL; 717 718 return false; 719 } 720 721 /* === DebugFS Interface ==================================================== */ 722 723 static int stats_show(struct seq_file *seq, void *v) 724 { 725 int i; 726 727 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled)); 728 for (i = 0; i < KFENCE_COUNTER_COUNT; i++) 729 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i])); 730 731 return 0; 732 } 733 DEFINE_SHOW_ATTRIBUTE(stats); 734 735 /* 736 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects. 737 * start_object() and next_object() return the object index + 1, because NULL is used 738 * to stop iteration. 739 */ 740 static void *start_object(struct seq_file *seq, loff_t *pos) 741 { 742 if (*pos < CONFIG_KFENCE_NUM_OBJECTS) 743 return (void *)((long)*pos + 1); 744 return NULL; 745 } 746 747 static void stop_object(struct seq_file *seq, void *v) 748 { 749 } 750 751 static void *next_object(struct seq_file *seq, void *v, loff_t *pos) 752 { 753 ++*pos; 754 if (*pos < CONFIG_KFENCE_NUM_OBJECTS) 755 return (void *)((long)*pos + 1); 756 return NULL; 757 } 758 759 static int show_object(struct seq_file *seq, void *v) 760 { 761 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1]; 762 unsigned long flags; 763 764 raw_spin_lock_irqsave(&meta->lock, flags); 765 kfence_print_object(seq, meta); 766 raw_spin_unlock_irqrestore(&meta->lock, flags); 767 seq_puts(seq, "---------------------------------\n"); 768 769 return 0; 770 } 771 772 static const struct seq_operations objects_sops = { 773 .start = start_object, 774 .next = next_object, 775 .stop = stop_object, 776 .show = show_object, 777 }; 778 DEFINE_SEQ_ATTRIBUTE(objects); 779 780 static int kfence_debugfs_init(void) 781 { 782 struct dentry *kfence_dir; 783 784 if (!READ_ONCE(kfence_enabled)) 785 return 0; 786 787 kfence_dir = debugfs_create_dir("kfence", NULL); 788 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops); 789 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops); 790 return 0; 791 } 792 793 late_initcall(kfence_debugfs_init); 794 795 /* === Panic Notifier ====================================================== */ 796 797 static void kfence_check_all_canary(void) 798 { 799 int i; 800 801 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 802 struct kfence_metadata *meta = &kfence_metadata[i]; 803 804 if (kfence_obj_allocated(meta)) 805 check_canary(meta); 806 } 807 } 808 809 static int kfence_check_canary_callback(struct notifier_block *nb, 810 unsigned long reason, void *arg) 811 { 812 kfence_check_all_canary(); 813 return NOTIFY_OK; 814 } 815 816 static struct notifier_block kfence_check_canary_notifier = { 817 .notifier_call = kfence_check_canary_callback, 818 }; 819 820 /* === Allocation Gate Timer ================================================ */ 821 822 static struct delayed_work kfence_timer; 823 824 #ifdef CONFIG_KFENCE_STATIC_KEYS 825 /* Wait queue to wake up allocation-gate timer task. */ 826 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait); 827 828 static void wake_up_kfence_timer(struct irq_work *work) 829 { 830 wake_up(&allocation_wait); 831 } 832 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer); 833 #endif 834 835 /* 836 * Set up delayed work, which will enable and disable the static key. We need to 837 * use a work queue (rather than a simple timer), since enabling and disabling a 838 * static key cannot be done from an interrupt. 839 * 840 * Note: Toggling a static branch currently causes IPIs, and here we'll end up 841 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with 842 * more aggressive sampling intervals), we could get away with a variant that 843 * avoids IPIs, at the cost of not immediately capturing allocations if the 844 * instructions remain cached. 845 */ 846 static void toggle_allocation_gate(struct work_struct *work) 847 { 848 if (!READ_ONCE(kfence_enabled)) 849 return; 850 851 atomic_set(&kfence_allocation_gate, -kfence_burst); 852 #ifdef CONFIG_KFENCE_STATIC_KEYS 853 /* Enable static key, and await allocation to happen. */ 854 static_branch_enable(&kfence_allocation_key); 855 856 wait_event_idle(allocation_wait, atomic_read(&kfence_allocation_gate) > 0); 857 858 /* Disable static key and reset timer. */ 859 static_branch_disable(&kfence_allocation_key); 860 #endif 861 queue_delayed_work(system_unbound_wq, &kfence_timer, 862 msecs_to_jiffies(kfence_sample_interval)); 863 } 864 865 /* === Public interface ===================================================== */ 866 867 void __init kfence_alloc_pool_and_metadata(void) 868 { 869 if (!kfence_sample_interval) 870 return; 871 872 /* 873 * If the pool has already been initialized by arch, there is no need to 874 * re-allocate the memory pool. 875 */ 876 if (!__kfence_pool) 877 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 878 879 if (!__kfence_pool) { 880 pr_err("failed to allocate pool\n"); 881 return; 882 } 883 884 /* The memory allocated by memblock has been zeroed out. */ 885 kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE); 886 if (!kfence_metadata_init) { 887 pr_err("failed to allocate metadata\n"); 888 memblock_free(__kfence_pool, KFENCE_POOL_SIZE); 889 __kfence_pool = NULL; 890 } 891 } 892 893 static void kfence_init_enable(void) 894 { 895 if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS)) 896 static_branch_enable(&kfence_allocation_key); 897 898 if (kfence_deferrable) 899 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate); 900 else 901 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate); 902 903 if (kfence_check_on_panic) 904 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier); 905 906 WRITE_ONCE(kfence_enabled, true); 907 queue_delayed_work(system_unbound_wq, &kfence_timer, 0); 908 909 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE, 910 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool, 911 (void *)(__kfence_pool + KFENCE_POOL_SIZE)); 912 } 913 914 void __init kfence_init(void) 915 { 916 stack_hash_seed = get_random_u32(); 917 918 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */ 919 if (!kfence_sample_interval) 920 return; 921 922 if (!kfence_init_pool_early()) { 923 pr_err("%s failed\n", __func__); 924 return; 925 } 926 927 kfence_init_enable(); 928 } 929 930 static int kfence_init_late(void) 931 { 932 const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE; 933 const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE; 934 unsigned long addr = (unsigned long)__kfence_pool; 935 unsigned long free_size = KFENCE_POOL_SIZE; 936 int err = -ENOMEM; 937 938 #ifdef CONFIG_CONTIG_ALLOC 939 struct page *pages; 940 941 pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL, first_online_node, 942 NULL); 943 if (!pages) 944 return -ENOMEM; 945 946 __kfence_pool = page_to_virt(pages); 947 pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL, first_online_node, 948 NULL); 949 if (pages) 950 kfence_metadata_init = page_to_virt(pages); 951 #else 952 if (nr_pages_pool > MAX_ORDER_NR_PAGES || 953 nr_pages_meta > MAX_ORDER_NR_PAGES) { 954 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n"); 955 return -EINVAL; 956 } 957 958 __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, GFP_KERNEL); 959 if (!__kfence_pool) 960 return -ENOMEM; 961 962 kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE, GFP_KERNEL); 963 #endif 964 965 if (!kfence_metadata_init) 966 goto free_pool; 967 968 memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE); 969 addr = kfence_init_pool(); 970 if (!addr) { 971 kfence_init_enable(); 972 kfence_debugfs_init(); 973 return 0; 974 } 975 976 pr_err("%s failed\n", __func__); 977 free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool); 978 err = -EBUSY; 979 980 #ifdef CONFIG_CONTIG_ALLOC 981 free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)), 982 nr_pages_meta); 983 free_pool: 984 free_contig_range(page_to_pfn(virt_to_page((void *)addr)), 985 free_size / PAGE_SIZE); 986 #else 987 free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE); 988 free_pool: 989 free_pages_exact((void *)addr, free_size); 990 #endif 991 992 kfence_metadata_init = NULL; 993 __kfence_pool = NULL; 994 return err; 995 } 996 997 static int kfence_enable_late(void) 998 { 999 if (!__kfence_pool) 1000 return kfence_init_late(); 1001 1002 WRITE_ONCE(kfence_enabled, true); 1003 queue_delayed_work(system_unbound_wq, &kfence_timer, 0); 1004 pr_info("re-enabled\n"); 1005 return 0; 1006 } 1007 1008 void kfence_shutdown_cache(struct kmem_cache *s) 1009 { 1010 unsigned long flags; 1011 struct kfence_metadata *meta; 1012 int i; 1013 1014 /* Pairs with release in kfence_init_pool(). */ 1015 if (!smp_load_acquire(&kfence_metadata)) 1016 return; 1017 1018 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 1019 bool in_use; 1020 1021 meta = &kfence_metadata[i]; 1022 1023 /* 1024 * If we observe some inconsistent cache and state pair where we 1025 * should have returned false here, cache destruction is racing 1026 * with either kmem_cache_alloc() or kmem_cache_free(). Taking 1027 * the lock will not help, as different critical section 1028 * serialization will have the same outcome. 1029 */ 1030 if (READ_ONCE(meta->cache) != s || !kfence_obj_allocated(meta)) 1031 continue; 1032 1033 raw_spin_lock_irqsave(&meta->lock, flags); 1034 in_use = meta->cache == s && kfence_obj_allocated(meta); 1035 raw_spin_unlock_irqrestore(&meta->lock, flags); 1036 1037 if (in_use) { 1038 /* 1039 * This cache still has allocations, and we should not 1040 * release them back into the freelist so they can still 1041 * safely be used and retain the kernel's default 1042 * behaviour of keeping the allocations alive (leak the 1043 * cache); however, they effectively become "zombie 1044 * allocations" as the KFENCE objects are the only ones 1045 * still in use and the owning cache is being destroyed. 1046 * 1047 * We mark them freed, so that any subsequent use shows 1048 * more useful error messages that will include stack 1049 * traces of the user of the object, the original 1050 * allocation, and caller to shutdown_cache(). 1051 */ 1052 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true); 1053 } 1054 } 1055 1056 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 1057 meta = &kfence_metadata[i]; 1058 1059 /* See above. */ 1060 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED) 1061 continue; 1062 1063 raw_spin_lock_irqsave(&meta->lock, flags); 1064 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED) 1065 meta->cache = NULL; 1066 raw_spin_unlock_irqrestore(&meta->lock, flags); 1067 } 1068 } 1069 1070 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags) 1071 { 1072 unsigned long stack_entries[KFENCE_STACK_DEPTH]; 1073 size_t num_stack_entries; 1074 u32 alloc_stack_hash; 1075 int allocation_gate; 1076 1077 /* 1078 * Perform size check before switching kfence_allocation_gate, so that 1079 * we don't disable KFENCE without making an allocation. 1080 */ 1081 if (size > PAGE_SIZE) { 1082 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]); 1083 return NULL; 1084 } 1085 1086 /* 1087 * Skip allocations from non-default zones, including DMA. We cannot 1088 * guarantee that pages in the KFENCE pool will have the requested 1089 * properties (e.g. reside in DMAable memory). 1090 */ 1091 if ((flags & GFP_ZONEMASK) || 1092 ((flags & __GFP_THISNODE) && num_online_nodes() > 1) || 1093 (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) { 1094 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]); 1095 return NULL; 1096 } 1097 1098 /* 1099 * Skip allocations for this slab, if KFENCE has been disabled for 1100 * this slab. 1101 */ 1102 if (s->flags & SLAB_SKIP_KFENCE) 1103 return NULL; 1104 1105 allocation_gate = atomic_inc_return(&kfence_allocation_gate); 1106 if (allocation_gate > 1) 1107 return NULL; 1108 #ifdef CONFIG_KFENCE_STATIC_KEYS 1109 /* 1110 * waitqueue_active() is fully ordered after the update of 1111 * kfence_allocation_gate per atomic_inc_return(). 1112 */ 1113 if (allocation_gate == 1 && waitqueue_active(&allocation_wait)) { 1114 /* 1115 * Calling wake_up() here may deadlock when allocations happen 1116 * from within timer code. Use an irq_work to defer it. 1117 */ 1118 irq_work_queue(&wake_up_kfence_timer_work); 1119 } 1120 #endif 1121 1122 if (!READ_ONCE(kfence_enabled)) 1123 return NULL; 1124 1125 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0); 1126 1127 /* 1128 * Do expensive check for coverage of allocation in slow-path after 1129 * allocation_gate has already become non-zero, even though it might 1130 * mean not making any allocation within a given sample interval. 1131 * 1132 * This ensures reasonable allocation coverage when the pool is almost 1133 * full, including avoiding long-lived allocations of the same source 1134 * filling up the pool (e.g. pagecache allocations). 1135 */ 1136 alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries); 1137 if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) { 1138 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]); 1139 return NULL; 1140 } 1141 1142 return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries, 1143 alloc_stack_hash); 1144 } 1145 1146 size_t kfence_ksize(const void *addr) 1147 { 1148 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr); 1149 1150 /* 1151 * Read locklessly -- if there is a race with __kfence_alloc(), this is 1152 * either a use-after-free or invalid access. 1153 */ 1154 return meta ? meta->size : 0; 1155 } 1156 1157 void *kfence_object_start(const void *addr) 1158 { 1159 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr); 1160 1161 /* 1162 * Read locklessly -- if there is a race with __kfence_alloc(), this is 1163 * either a use-after-free or invalid access. 1164 */ 1165 return meta ? (void *)meta->addr : NULL; 1166 } 1167 1168 void __kfence_free(void *addr) 1169 { 1170 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr); 1171 1172 #ifdef CONFIG_MEMCG 1173 KFENCE_WARN_ON(meta->obj_exts.objcg); 1174 #endif 1175 /* 1176 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing 1177 * the object, as the object page may be recycled for other-typed 1178 * objects once it has been freed. meta->cache may be NULL if the cache 1179 * was destroyed. 1180 * Save the stack trace here so that reports show where the user freed 1181 * the object. 1182 */ 1183 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU))) { 1184 unsigned long flags; 1185 1186 raw_spin_lock_irqsave(&meta->lock, flags); 1187 metadata_update_state(meta, KFENCE_OBJECT_RCU_FREEING, NULL, 0); 1188 raw_spin_unlock_irqrestore(&meta->lock, flags); 1189 call_rcu(&meta->rcu_head, rcu_guarded_free); 1190 } else { 1191 kfence_guarded_free(addr, meta, false); 1192 } 1193 } 1194 1195 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs) 1196 { 1197 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE; 1198 struct kfence_metadata *to_report = NULL; 1199 enum kfence_error_type error_type; 1200 unsigned long flags; 1201 1202 if (!is_kfence_address((void *)addr)) 1203 return false; 1204 1205 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */ 1206 return kfence_unprotect(addr); /* ... unprotect and proceed. */ 1207 1208 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]); 1209 1210 if (page_index % 2) { 1211 /* This is a redzone, report a buffer overflow. */ 1212 struct kfence_metadata *meta; 1213 int distance = 0; 1214 1215 meta = addr_to_metadata(addr - PAGE_SIZE); 1216 if (meta && kfence_obj_allocated(meta)) { 1217 to_report = meta; 1218 /* Data race ok; distance calculation approximate. */ 1219 distance = addr - data_race(meta->addr + meta->size); 1220 } 1221 1222 meta = addr_to_metadata(addr + PAGE_SIZE); 1223 if (meta && kfence_obj_allocated(meta)) { 1224 /* Data race ok; distance calculation approximate. */ 1225 if (!to_report || distance > data_race(meta->addr) - addr) 1226 to_report = meta; 1227 } 1228 1229 if (!to_report) 1230 goto out; 1231 1232 raw_spin_lock_irqsave(&to_report->lock, flags); 1233 to_report->unprotected_page = addr; 1234 error_type = KFENCE_ERROR_OOB; 1235 1236 /* 1237 * If the object was freed before we took the look we can still 1238 * report this as an OOB -- the report will simply show the 1239 * stacktrace of the free as well. 1240 */ 1241 } else { 1242 to_report = addr_to_metadata(addr); 1243 if (!to_report) 1244 goto out; 1245 1246 raw_spin_lock_irqsave(&to_report->lock, flags); 1247 error_type = KFENCE_ERROR_UAF; 1248 /* 1249 * We may race with __kfence_alloc(), and it is possible that a 1250 * freed object may be reallocated. We simply report this as a 1251 * use-after-free, with the stack trace showing the place where 1252 * the object was re-allocated. 1253 */ 1254 } 1255 1256 out: 1257 if (to_report) { 1258 kfence_report_error(addr, is_write, regs, to_report, error_type); 1259 raw_spin_unlock_irqrestore(&to_report->lock, flags); 1260 } else { 1261 /* This may be a UAF or OOB access, but we can't be sure. */ 1262 kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID); 1263 } 1264 1265 return kfence_unprotect(addr); /* Unprotect and let access proceed. */ 1266 } 1267