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/kasan-enabled.h> 17 #include <linux/kcsan-checks.h> 18 #include <linux/kfence.h> 19 #include <linux/kmemleak.h> 20 #include <linux/list.h> 21 #include <linux/lockdep.h> 22 #include <linux/log2.h> 23 #include <linux/memblock.h> 24 #include <linux/moduleparam.h> 25 #include <linux/nodemask.h> 26 #include <linux/notifier.h> 27 #include <linux/panic_notifier.h> 28 #include <linux/random.h> 29 #include <linux/rcupdate.h> 30 #include <linux/reboot.h> 31 #include <linux/sched/clock.h> 32 #include <linux/seq_file.h> 33 #include <linux/slab.h> 34 #include <linux/spinlock.h> 35 #include <linux/string.h> 36 37 #include <asm/kfence.h> 38 39 #include "kfence.h" 40 41 /* Disables KFENCE on the first warning assuming an irrecoverable error. */ 42 #define KFENCE_WARN_ON(cond) \ 43 ({ \ 44 const bool __cond = WARN_ON(cond); \ 45 if (unlikely(__cond)) { \ 46 WRITE_ONCE(kfence_enabled, false); \ 47 disabled_by_warn = true; \ 48 } \ 49 __cond; \ 50 }) 51 52 /* === Data ================================================================= */ 53 54 bool kfence_enabled __read_mostly; 55 static bool disabled_by_warn __read_mostly; 56 57 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL; 58 EXPORT_SYMBOL_GPL(kfence_sample_interval); /* Export for test modules. */ 59 60 #ifdef MODULE_PARAM_PREFIX 61 #undef MODULE_PARAM_PREFIX 62 #endif 63 #define MODULE_PARAM_PREFIX "kfence." 64 65 static int kfence_enable_late(void); 66 static int param_set_sample_interval(const char *val, const struct kernel_param *kp) 67 { 68 unsigned long num; 69 int ret = kstrtoul(val, 0, &num); 70 71 if (ret < 0) 72 return ret; 73 74 /* Using 0 to indicate KFENCE is disabled. */ 75 if (!num && READ_ONCE(kfence_enabled)) { 76 pr_info("disabled\n"); 77 WRITE_ONCE(kfence_enabled, false); 78 } 79 80 if (num && kasan_hw_tags_enabled()) { 81 pr_info("disabled as KASAN HW tags are enabled\n"); 82 return -EINVAL; 83 } 84 85 *((unsigned long *)kp->arg) = num; 86 87 if (num && !READ_ONCE(kfence_enabled) && system_state != SYSTEM_BOOTING) 88 return disabled_by_warn ? -EINVAL : kfence_enable_late(); 89 return 0; 90 } 91 92 static int param_get_sample_interval(char *buffer, const struct kernel_param *kp) 93 { 94 if (!READ_ONCE(kfence_enabled)) 95 return sprintf(buffer, "0\n"); 96 97 return param_get_ulong(buffer, kp); 98 } 99 100 static const struct kernel_param_ops sample_interval_param_ops = { 101 .set = param_set_sample_interval, 102 .get = param_get_sample_interval, 103 }; 104 module_param_cb(sample_interval, &sample_interval_param_ops, &kfence_sample_interval, 0600); 105 106 /* Pool usage% threshold when currently covered allocations are skipped. */ 107 static unsigned long kfence_skip_covered_thresh __read_mostly = 75; 108 module_param_named(skip_covered_thresh, kfence_skip_covered_thresh, ulong, 0644); 109 110 /* Allocation burst count: number of excess KFENCE allocations per sample. */ 111 static unsigned int kfence_burst __read_mostly; 112 module_param_named(burst, kfence_burst, uint, 0644); 113 114 /* If true, use a deferrable timer. */ 115 static bool kfence_deferrable __read_mostly = IS_ENABLED(CONFIG_KFENCE_DEFERRABLE); 116 module_param_named(deferrable, kfence_deferrable, bool, 0444); 117 118 /* If true, check all canary bytes on panic. */ 119 static bool kfence_check_on_panic __read_mostly; 120 module_param_named(check_on_panic, kfence_check_on_panic, bool, 0444); 121 122 /* The pool of pages used for guard pages and objects. */ 123 char *__kfence_pool __read_mostly; 124 EXPORT_SYMBOL(__kfence_pool); /* Export for test modules. */ 125 126 /* 127 * Per-object metadata, with one-to-one mapping of object metadata to 128 * backing pages (in __kfence_pool). 129 */ 130 static_assert(CONFIG_KFENCE_NUM_OBJECTS > 0); 131 struct kfence_metadata *kfence_metadata __read_mostly; 132 133 /* 134 * If kfence_metadata is not NULL, it may be accessed by kfence_shutdown_cache(). 135 * So introduce kfence_metadata_init to initialize metadata, and then make 136 * kfence_metadata visible after initialization is successful. This prevents 137 * potential UAF or access to uninitialized metadata. 138 */ 139 static struct kfence_metadata *kfence_metadata_init __read_mostly; 140 141 /* Freelist with available objects. */ 142 DEFINE_RAW_SPINLOCK(kfence_freelist_lock); /* Lock protecting freelist. */ 143 static struct list_head kfence_freelist __guarded_by(&kfence_freelist_lock) = LIST_HEAD_INIT(kfence_freelist); 144 145 /* 146 * The static key to set up a KFENCE allocation; or if static keys are not used 147 * to gate allocations, to avoid a load and compare if KFENCE is disabled. 148 */ 149 DEFINE_STATIC_KEY_FALSE(kfence_allocation_key); 150 151 /* Gates the allocation, ensuring only one succeeds in a given period. */ 152 atomic_t kfence_allocation_gate = ATOMIC_INIT(1); 153 154 /* 155 * A Counting Bloom filter of allocation coverage: limits currently covered 156 * allocations of the same source filling up the pool. 157 * 158 * Assuming a range of 15%-85% unique allocations in the pool at any point in 159 * time, the below parameters provide a probablity of 0.02-0.33 for false 160 * positive hits respectively: 161 * 162 * P(alloc_traces) = (1 - e^(-HNUM * (alloc_traces / SIZE)) ^ HNUM 163 */ 164 #define ALLOC_COVERED_HNUM 2 165 #define ALLOC_COVERED_ORDER (const_ilog2(CONFIG_KFENCE_NUM_OBJECTS) + 2) 166 #define ALLOC_COVERED_SIZE (1 << ALLOC_COVERED_ORDER) 167 #define ALLOC_COVERED_HNEXT(h) hash_32(h, ALLOC_COVERED_ORDER) 168 #define ALLOC_COVERED_MASK (ALLOC_COVERED_SIZE - 1) 169 static atomic_t alloc_covered[ALLOC_COVERED_SIZE]; 170 171 /* Stack depth used to determine uniqueness of an allocation. */ 172 #define UNIQUE_ALLOC_STACK_DEPTH ((size_t)8) 173 174 /* 175 * Randomness for stack hashes, making the same collisions across reboots and 176 * different machines less likely. 177 */ 178 static u32 stack_hash_seed __ro_after_init; 179 180 /* Statistics counters for debugfs. */ 181 enum kfence_counter_id { 182 KFENCE_COUNTER_ALLOCATED, 183 KFENCE_COUNTER_ALLOCS, 184 KFENCE_COUNTER_FREES, 185 KFENCE_COUNTER_ZOMBIES, 186 KFENCE_COUNTER_BUGS, 187 KFENCE_COUNTER_SKIP_INCOMPAT, 188 KFENCE_COUNTER_SKIP_CAPACITY, 189 KFENCE_COUNTER_SKIP_COVERED, 190 KFENCE_COUNTER_COUNT, 191 }; 192 static atomic_long_t counters[KFENCE_COUNTER_COUNT]; 193 static const char *const counter_names[] = { 194 [KFENCE_COUNTER_ALLOCATED] = "currently allocated", 195 [KFENCE_COUNTER_ALLOCS] = "total allocations", 196 [KFENCE_COUNTER_FREES] = "total frees", 197 [KFENCE_COUNTER_ZOMBIES] = "zombie allocations", 198 [KFENCE_COUNTER_BUGS] = "total bugs", 199 [KFENCE_COUNTER_SKIP_INCOMPAT] = "skipped allocations (incompatible)", 200 [KFENCE_COUNTER_SKIP_CAPACITY] = "skipped allocations (capacity)", 201 [KFENCE_COUNTER_SKIP_COVERED] = "skipped allocations (covered)", 202 }; 203 static_assert(ARRAY_SIZE(counter_names) == KFENCE_COUNTER_COUNT); 204 205 /* === Internals ============================================================ */ 206 207 static inline bool should_skip_covered(void) 208 { 209 unsigned long thresh = (CONFIG_KFENCE_NUM_OBJECTS * kfence_skip_covered_thresh) / 100; 210 211 return atomic_long_read(&counters[KFENCE_COUNTER_ALLOCATED]) > thresh; 212 } 213 214 static u32 get_alloc_stack_hash(unsigned long *stack_entries, size_t num_entries) 215 { 216 num_entries = min(num_entries, UNIQUE_ALLOC_STACK_DEPTH); 217 num_entries = filter_irq_stacks(stack_entries, num_entries); 218 return jhash(stack_entries, num_entries * sizeof(stack_entries[0]), stack_hash_seed); 219 } 220 221 /* 222 * Adds (or subtracts) count @val for allocation stack trace hash 223 * @alloc_stack_hash from Counting Bloom filter. 224 */ 225 static void alloc_covered_add(u32 alloc_stack_hash, int val) 226 { 227 int i; 228 229 for (i = 0; i < ALLOC_COVERED_HNUM; i++) { 230 atomic_add(val, &alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK]); 231 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash); 232 } 233 } 234 235 /* 236 * Returns true if the allocation stack trace hash @alloc_stack_hash is 237 * currently contained (non-zero count) in Counting Bloom filter. 238 */ 239 static bool alloc_covered_contains(u32 alloc_stack_hash) 240 { 241 int i; 242 243 for (i = 0; i < ALLOC_COVERED_HNUM; i++) { 244 if (!atomic_read(&alloc_covered[alloc_stack_hash & ALLOC_COVERED_MASK])) 245 return false; 246 alloc_stack_hash = ALLOC_COVERED_HNEXT(alloc_stack_hash); 247 } 248 249 return true; 250 } 251 252 static bool kfence_protect(unsigned long addr) 253 { 254 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), true)); 255 } 256 257 static bool kfence_unprotect(unsigned long addr) 258 { 259 return !KFENCE_WARN_ON(!kfence_protect_page(ALIGN_DOWN(addr, PAGE_SIZE), false)); 260 } 261 262 static inline unsigned long metadata_to_pageaddr(const struct kfence_metadata *meta) 263 __must_hold(&meta->lock) 264 { 265 unsigned long offset = (meta - kfence_metadata + 1) * PAGE_SIZE * 2; 266 unsigned long pageaddr = (unsigned long)&__kfence_pool[offset]; 267 268 /* The checks do not affect performance; only called from slow-paths. */ 269 270 /* Only call with a pointer into kfence_metadata. */ 271 if (KFENCE_WARN_ON(meta < kfence_metadata || 272 meta >= kfence_metadata + CONFIG_KFENCE_NUM_OBJECTS)) 273 return 0; 274 275 /* 276 * This metadata object only ever maps to 1 page; verify that the stored 277 * address is in the expected range. 278 */ 279 if (KFENCE_WARN_ON(ALIGN_DOWN(meta->addr, PAGE_SIZE) != pageaddr)) 280 return 0; 281 282 return pageaddr; 283 } 284 285 static inline bool kfence_obj_allocated(const struct kfence_metadata *meta) 286 { 287 enum kfence_object_state state = READ_ONCE(meta->state); 288 289 return state == KFENCE_OBJECT_ALLOCATED || state == KFENCE_OBJECT_RCU_FREEING; 290 } 291 292 /* 293 * Update the object's metadata state, including updating the alloc/free stacks 294 * depending on the state transition. 295 */ 296 static noinline void 297 metadata_update_state(struct kfence_metadata *meta, enum kfence_object_state next, 298 unsigned long *stack_entries, size_t num_stack_entries) 299 __must_hold(&meta->lock) 300 { 301 struct kfence_track *track = 302 next == KFENCE_OBJECT_ALLOCATED ? &meta->alloc_track : &meta->free_track; 303 304 lockdep_assert_held(&meta->lock); 305 306 /* Stack has been saved when calling rcu, skip. */ 307 if (READ_ONCE(meta->state) == KFENCE_OBJECT_RCU_FREEING) 308 goto out; 309 310 if (stack_entries) { 311 memcpy(track->stack_entries, stack_entries, 312 num_stack_entries * sizeof(stack_entries[0])); 313 } else { 314 /* 315 * Skip over 1 (this) functions; noinline ensures we do not 316 * accidentally skip over the caller by never inlining. 317 */ 318 num_stack_entries = stack_trace_save(track->stack_entries, KFENCE_STACK_DEPTH, 1); 319 } 320 track->num_stack_entries = num_stack_entries; 321 track->pid = task_pid_nr(current); 322 track->cpu = raw_smp_processor_id(); 323 track->ts_nsec = local_clock(); /* Same source as printk timestamps. */ 324 325 out: 326 /* 327 * Pairs with READ_ONCE() in 328 * kfence_shutdown_cache(), 329 * kfence_handle_page_fault(). 330 */ 331 WRITE_ONCE(meta->state, next); 332 } 333 334 #ifdef CONFIG_KMSAN 335 #define check_canary_attributes noinline __no_kmsan_checks 336 #else 337 #define check_canary_attributes inline 338 #endif 339 340 /* Check canary byte at @addr. */ 341 static check_canary_attributes bool check_canary_byte(u8 *addr) 342 { 343 struct kfence_metadata *meta; 344 enum kfence_fault fault; 345 unsigned long flags; 346 347 if (likely(*addr == KFENCE_CANARY_PATTERN_U8(addr))) 348 return true; 349 350 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]); 351 352 meta = addr_to_metadata((unsigned long)addr); 353 raw_spin_lock_irqsave(&meta->lock, flags); 354 fault = kfence_report_error((unsigned long)addr, false, NULL, meta, KFENCE_ERROR_CORRUPTION); 355 raw_spin_unlock_irqrestore(&meta->lock, flags); 356 kfence_handle_fault(fault); 357 358 return false; 359 } 360 361 static inline void set_canary(const struct kfence_metadata *meta) 362 { 363 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE); 364 unsigned long addr = pageaddr; 365 366 /* 367 * The canary may be written to part of the object memory, but it does 368 * not affect it. The user should initialize the object before using it. 369 */ 370 for (; addr < meta->addr; addr += sizeof(u64)) 371 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64; 372 373 addr = ALIGN_DOWN(meta->addr + meta->size, sizeof(u64)); 374 for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) 375 *((u64 *)addr) = KFENCE_CANARY_PATTERN_U64; 376 } 377 378 static check_canary_attributes void 379 check_canary(const struct kfence_metadata *meta) 380 { 381 const unsigned long pageaddr = ALIGN_DOWN(meta->addr, PAGE_SIZE); 382 unsigned long addr = pageaddr; 383 384 /* 385 * We'll iterate over each canary byte per-side until a corrupted byte 386 * is found. However, we'll still iterate over the canary bytes to the 387 * right of the object even if there was an error in the canary bytes to 388 * the left of the object. Specifically, if check_canary_byte() 389 * generates an error, showing both sides might give more clues as to 390 * what the error is about when displaying which bytes were corrupted. 391 */ 392 393 /* Apply to left of object. */ 394 for (; meta->addr - addr >= sizeof(u64); addr += sizeof(u64)) { 395 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) 396 break; 397 } 398 399 /* 400 * If the canary is corrupted in a certain 64 bytes, or the canary 401 * memory cannot be completely covered by multiple consecutive 64 bytes, 402 * it needs to be checked one by one. 403 */ 404 for (; addr < meta->addr; addr++) { 405 if (unlikely(!check_canary_byte((u8 *)addr))) 406 break; 407 } 408 409 /* Apply to right of object. */ 410 for (addr = meta->addr + meta->size; addr % sizeof(u64) != 0; addr++) { 411 if (unlikely(!check_canary_byte((u8 *)addr))) 412 return; 413 } 414 for (; addr - pageaddr < PAGE_SIZE; addr += sizeof(u64)) { 415 if (unlikely(*((u64 *)addr) != KFENCE_CANARY_PATTERN_U64)) { 416 417 for (; addr - pageaddr < PAGE_SIZE; addr++) { 418 if (!check_canary_byte((u8 *)addr)) 419 return; 420 } 421 } 422 } 423 } 424 425 static void *kfence_guarded_alloc(struct kmem_cache *cache, size_t size, gfp_t gfp, 426 unsigned long *stack_entries, size_t num_stack_entries, 427 u32 alloc_stack_hash) 428 { 429 struct kfence_metadata *meta = NULL; 430 unsigned long flags; 431 struct slab *slab; 432 void *addr; 433 const bool random_right_allocate = get_random_u32_below(2); 434 const bool random_fault = CONFIG_KFENCE_STRESS_TEST_FAULTS && 435 !get_random_u32_below(CONFIG_KFENCE_STRESS_TEST_FAULTS); 436 437 /* Try to obtain a free object. */ 438 raw_spin_lock_irqsave(&kfence_freelist_lock, flags); 439 if (!list_empty(&kfence_freelist)) { 440 meta = list_entry(kfence_freelist.next, struct kfence_metadata, list); 441 list_del_init(&meta->list); 442 } 443 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags); 444 if (!meta) { 445 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_CAPACITY]); 446 return NULL; 447 } 448 449 if (unlikely(!raw_spin_trylock_irqsave(&meta->lock, flags))) { 450 /* 451 * This is extremely unlikely -- we are reporting on a 452 * use-after-free, which locked meta->lock, and the reporting 453 * code via printk calls kmalloc() which ends up in 454 * kfence_alloc() and tries to grab the same object that we're 455 * reporting on. While it has never been observed, lockdep does 456 * report that there is a possibility of deadlock. Fix it by 457 * using trylock and bailing out gracefully. 458 */ 459 raw_spin_lock_irqsave(&kfence_freelist_lock, flags); 460 /* Put the object back on the freelist. */ 461 list_add_tail(&meta->list, &kfence_freelist); 462 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags); 463 464 return NULL; 465 } 466 467 meta->addr = metadata_to_pageaddr(meta); 468 /* Unprotect if we're reusing this page. */ 469 if (meta->state == KFENCE_OBJECT_FREED) 470 kfence_unprotect(meta->addr); 471 472 /* 473 * Note: for allocations made before RNG initialization, will always 474 * return zero. We still benefit from enabling KFENCE as early as 475 * possible, even when the RNG is not yet available, as this will allow 476 * KFENCE to detect bugs due to earlier allocations. The only downside 477 * is that the out-of-bounds accesses detected are deterministic for 478 * such allocations. 479 */ 480 if (random_right_allocate) { 481 /* Allocate on the "right" side, re-calculate address. */ 482 meta->addr += PAGE_SIZE - size; 483 meta->addr = ALIGN_DOWN(meta->addr, cache->align); 484 } 485 486 addr = (void *)meta->addr; 487 488 /* Update remaining metadata. */ 489 metadata_update_state(meta, KFENCE_OBJECT_ALLOCATED, stack_entries, num_stack_entries); 490 /* Pairs with READ_ONCE() in kfence_shutdown_cache(). */ 491 WRITE_ONCE(meta->cache, cache); 492 meta->size = size; 493 meta->alloc_stack_hash = alloc_stack_hash; 494 raw_spin_unlock_irqrestore(&meta->lock, flags); 495 496 alloc_covered_add(alloc_stack_hash, 1); 497 498 /* Set required slab fields. */ 499 slab = virt_to_slab(addr); 500 slab->slab_cache = cache; 501 slab->objects = 1; 502 503 /* Memory initialization. */ 504 set_canary(meta); 505 506 /* 507 * We check slab_want_init_on_alloc() ourselves, rather than letting 508 * slab do the initialization, as otherwise it might overwrite KFENCE's 509 * redzone. 510 */ 511 if (unlikely(slab_want_init_on_alloc(gfp, cache))) 512 memzero_explicit(addr, size); 513 if (cache->ctor) 514 cache->ctor(addr); 515 516 if (random_fault) 517 kfence_protect(meta->addr); /* Random "faults" by protecting the object. */ 518 519 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCATED]); 520 atomic_long_inc(&counters[KFENCE_COUNTER_ALLOCS]); 521 522 return addr; 523 } 524 525 static void kfence_guarded_free(void *addr, struct kfence_metadata *meta, bool zombie) 526 { 527 struct kcsan_scoped_access assert_page_exclusive; 528 u32 alloc_stack_hash; 529 unsigned long flags; 530 bool init; 531 532 raw_spin_lock_irqsave(&meta->lock, flags); 533 534 if (!kfence_obj_allocated(meta) || meta->addr != (unsigned long)addr) { 535 enum kfence_fault fault; 536 537 /* Invalid or double-free, bail out. */ 538 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]); 539 fault = kfence_report_error((unsigned long)addr, false, NULL, meta, 540 KFENCE_ERROR_INVALID_FREE); 541 raw_spin_unlock_irqrestore(&meta->lock, flags); 542 kfence_handle_fault(fault); 543 return; 544 } 545 546 /* Detect racy use-after-free, or incorrect reallocation of this page by KFENCE. */ 547 kcsan_begin_scoped_access((void *)ALIGN_DOWN((unsigned long)addr, PAGE_SIZE), PAGE_SIZE, 548 KCSAN_ACCESS_SCOPED | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT, 549 &assert_page_exclusive); 550 551 if (CONFIG_KFENCE_STRESS_TEST_FAULTS) 552 kfence_unprotect((unsigned long)addr); /* To check canary bytes. */ 553 554 /* Restore page protection if there was an OOB access. */ 555 if (meta->unprotected_page) { 556 memzero_explicit((void *)ALIGN_DOWN(meta->unprotected_page, PAGE_SIZE), PAGE_SIZE); 557 kfence_protect(meta->unprotected_page); 558 meta->unprotected_page = 0; 559 } 560 561 /* Mark the object as freed. */ 562 metadata_update_state(meta, KFENCE_OBJECT_FREED, NULL, 0); 563 init = slab_want_init_on_free(meta->cache); 564 alloc_stack_hash = meta->alloc_stack_hash; 565 raw_spin_unlock_irqrestore(&meta->lock, flags); 566 567 alloc_covered_add(alloc_stack_hash, -1); 568 569 /* Check canary bytes for memory corruption. */ 570 check_canary(meta); 571 572 /* 573 * Clear memory if init-on-free is set. While we protect the page, the 574 * data is still there, and after a use-after-free is detected, we 575 * unprotect the page, so the data is still accessible. 576 */ 577 if (!zombie && unlikely(init)) 578 memzero_explicit(addr, meta->size); 579 580 /* Protect to detect use-after-frees. */ 581 kfence_protect((unsigned long)addr); 582 583 kcsan_end_scoped_access(&assert_page_exclusive); 584 if (!zombie) { 585 /* Add it to the tail of the freelist for reuse. */ 586 raw_spin_lock_irqsave(&kfence_freelist_lock, flags); 587 KFENCE_WARN_ON(!list_empty(&meta->list)); 588 list_add_tail(&meta->list, &kfence_freelist); 589 raw_spin_unlock_irqrestore(&kfence_freelist_lock, flags); 590 591 atomic_long_dec(&counters[KFENCE_COUNTER_ALLOCATED]); 592 atomic_long_inc(&counters[KFENCE_COUNTER_FREES]); 593 } else { 594 /* See kfence_shutdown_cache(). */ 595 atomic_long_inc(&counters[KFENCE_COUNTER_ZOMBIES]); 596 } 597 } 598 599 static void rcu_guarded_free(struct rcu_head *h) 600 { 601 struct kfence_metadata *meta = container_of(h, struct kfence_metadata, rcu_head); 602 603 kfence_guarded_free((void *)meta->addr, meta, false); 604 } 605 606 /* 607 * Initialization of the KFENCE pool after its allocation. 608 * Returns 0 on success; otherwise returns the address up to 609 * which partial initialization succeeded. 610 */ 611 static unsigned long kfence_init_pool(void) 612 __context_unsafe(/* constructor */) 613 { 614 unsigned long addr, start_pfn; 615 int i, rand; 616 617 if (!arch_kfence_init_pool()) 618 return (unsigned long)__kfence_pool; 619 620 addr = (unsigned long)__kfence_pool; 621 start_pfn = PHYS_PFN(virt_to_phys(__kfence_pool)); 622 623 /* 624 * Set up object pages: they must have PGTY_slab set to avoid freeing 625 * them as real pages. 626 * 627 * We also want to avoid inserting kfence_free() in the kfree() 628 * fast-path in SLUB, and therefore need to ensure kfree() correctly 629 * enters __slab_free() slow-path. 630 */ 631 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) { 632 struct page *page; 633 634 if (!i || (i % 2)) 635 continue; 636 637 page = pfn_to_page(start_pfn + i); 638 __SetPageSlab(page); 639 #ifdef CONFIG_MEMCG 640 struct slab *slab = page_slab(page); 641 slab->obj_exts = (unsigned long)&kfence_metadata_init[i / 2 - 1].obj_exts | 642 MEMCG_DATA_OBJEXTS; 643 #endif 644 } 645 646 /* 647 * Protect the first 2 pages. The first page is mostly unnecessary, and 648 * merely serves as an extended guard page. However, adding one 649 * additional page in the beginning gives us an even number of pages, 650 * which simplifies the mapping of address to metadata index. 651 */ 652 for (i = 0; i < 2; i++) { 653 if (unlikely(!kfence_protect(addr))) 654 return addr; 655 656 addr += PAGE_SIZE; 657 } 658 659 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 660 struct kfence_metadata *meta = &kfence_metadata_init[i]; 661 662 /* Initialize metadata. */ 663 INIT_LIST_HEAD(&meta->list); 664 raw_spin_lock_init(&meta->lock); 665 meta->state = KFENCE_OBJECT_UNUSED; 666 /* Use addr to randomize the freelist. */ 667 meta->addr = i; 668 669 /* Protect the right redzone. */ 670 if (unlikely(!kfence_protect(addr + 2 * i * PAGE_SIZE + PAGE_SIZE))) 671 goto reset_slab; 672 } 673 674 for (i = CONFIG_KFENCE_NUM_OBJECTS; i > 0; i--) { 675 rand = get_random_u32_below(i); 676 swap(kfence_metadata_init[i - 1].addr, kfence_metadata_init[rand].addr); 677 } 678 679 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 680 struct kfence_metadata *meta_1 = &kfence_metadata_init[i]; 681 struct kfence_metadata *meta_2 = &kfence_metadata_init[meta_1->addr]; 682 683 list_add_tail(&meta_2->list, &kfence_freelist); 684 } 685 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 686 kfence_metadata_init[i].addr = addr; 687 addr += 2 * PAGE_SIZE; 688 } 689 690 /* 691 * Make kfence_metadata visible only when initialization is successful. 692 * Otherwise, if the initialization fails and kfence_metadata is freed, 693 * it may cause UAF in kfence_shutdown_cache(). 694 */ 695 smp_store_release(&kfence_metadata, kfence_metadata_init); 696 return 0; 697 698 reset_slab: 699 addr += 2 * i * PAGE_SIZE; 700 for (i = 0; i < KFENCE_POOL_SIZE / PAGE_SIZE; i++) { 701 struct page *page; 702 703 if (!i || (i % 2)) 704 continue; 705 706 page = pfn_to_page(start_pfn + i); 707 #ifdef CONFIG_MEMCG 708 struct slab *slab = page_slab(page); 709 slab->obj_exts = 0; 710 #endif 711 __ClearPageSlab(page); 712 } 713 714 return addr; 715 } 716 717 static bool __init kfence_init_pool_early(void) 718 { 719 unsigned long addr; 720 721 if (!__kfence_pool) 722 return false; 723 724 addr = kfence_init_pool(); 725 726 if (!addr) { 727 /* 728 * The pool is live and will never be deallocated from this point on. 729 * Ignore the pool object from the kmemleak phys object tree, as it would 730 * otherwise overlap with allocations returned by kfence_alloc(), which 731 * are registered with kmemleak through the slab post-alloc hook. 732 */ 733 kmemleak_ignore_phys(__pa(__kfence_pool)); 734 return true; 735 } 736 737 /* 738 * Only release unprotected pages, and do not try to go back and change 739 * page attributes due to risk of failing to do so as well. If changing 740 * page attributes for some pages fails, it is very likely that it also 741 * fails for the first page, and therefore expect addr==__kfence_pool in 742 * most failure cases. 743 */ 744 memblock_free((void *)addr, KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool)); 745 __kfence_pool = NULL; 746 747 memblock_free(kfence_metadata_init, KFENCE_METADATA_SIZE); 748 kfence_metadata_init = NULL; 749 750 return false; 751 } 752 753 /* === DebugFS Interface ==================================================== */ 754 755 static int stats_show(struct seq_file *seq, void *v) 756 { 757 int i; 758 759 seq_printf(seq, "enabled: %i\n", READ_ONCE(kfence_enabled)); 760 for (i = 0; i < KFENCE_COUNTER_COUNT; i++) 761 seq_printf(seq, "%s: %ld\n", counter_names[i], atomic_long_read(&counters[i])); 762 763 return 0; 764 } 765 DEFINE_SHOW_ATTRIBUTE(stats); 766 767 /* 768 * debugfs seq_file operations for /sys/kernel/debug/kfence/objects. 769 * start_object() and next_object() return the object index + 1, because NULL is used 770 * to stop iteration. 771 */ 772 static void *start_object(struct seq_file *seq, loff_t *pos) 773 { 774 if (*pos < CONFIG_KFENCE_NUM_OBJECTS) 775 return (void *)((long)*pos + 1); 776 return NULL; 777 } 778 779 static void stop_object(struct seq_file *seq, void *v) 780 { 781 } 782 783 static void *next_object(struct seq_file *seq, void *v, loff_t *pos) 784 { 785 ++*pos; 786 if (*pos < CONFIG_KFENCE_NUM_OBJECTS) 787 return (void *)((long)*pos + 1); 788 return NULL; 789 } 790 791 static int show_object(struct seq_file *seq, void *v) 792 { 793 struct kfence_metadata *meta = &kfence_metadata[(long)v - 1]; 794 unsigned long flags; 795 796 raw_spin_lock_irqsave(&meta->lock, flags); 797 kfence_print_object(seq, meta); 798 raw_spin_unlock_irqrestore(&meta->lock, flags); 799 seq_puts(seq, "---------------------------------\n"); 800 801 return 0; 802 } 803 804 static const struct seq_operations objects_sops = { 805 .start = start_object, 806 .next = next_object, 807 .stop = stop_object, 808 .show = show_object, 809 }; 810 DEFINE_SEQ_ATTRIBUTE(objects); 811 812 static int kfence_debugfs_init(void) 813 { 814 struct dentry *kfence_dir; 815 816 if (!READ_ONCE(kfence_enabled)) 817 return 0; 818 819 kfence_dir = debugfs_create_dir("kfence", NULL); 820 debugfs_create_file("stats", 0444, kfence_dir, NULL, &stats_fops); 821 debugfs_create_file("objects", 0400, kfence_dir, NULL, &objects_fops); 822 return 0; 823 } 824 825 late_initcall(kfence_debugfs_init); 826 827 /* === Panic Notifier ====================================================== */ 828 829 static void kfence_check_all_canary(void) 830 { 831 int i; 832 833 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 834 struct kfence_metadata *meta = &kfence_metadata[i]; 835 836 if (kfence_obj_allocated(meta)) 837 check_canary(meta); 838 } 839 } 840 841 static int kfence_check_canary_callback(struct notifier_block *nb, 842 unsigned long reason, void *arg) 843 { 844 if (READ_ONCE(kfence_enabled)) 845 kfence_check_all_canary(); 846 return NOTIFY_OK; 847 } 848 849 static struct notifier_block kfence_check_canary_notifier = { 850 .notifier_call = kfence_check_canary_callback, 851 }; 852 853 /* === Allocation Gate Timer ================================================ */ 854 855 static struct delayed_work kfence_timer; 856 857 #ifdef CONFIG_KFENCE_STATIC_KEYS 858 /* Wait queue to wake up allocation-gate timer task. */ 859 static DECLARE_WAIT_QUEUE_HEAD(allocation_wait); 860 861 static int kfence_reboot_callback(struct notifier_block *nb, 862 unsigned long action, void *data) 863 { 864 /* 865 * Disable kfence to avoid static keys IPI synchronization during 866 * late shutdown/kexec 867 */ 868 WRITE_ONCE(kfence_enabled, false); 869 /* Cancel any pending timer work */ 870 cancel_delayed_work(&kfence_timer); 871 /* 872 * Wake up any blocked toggle_allocation_gate() so it can complete 873 * early while the system is still able to handle IPIs. 874 */ 875 wake_up(&allocation_wait); 876 877 return NOTIFY_OK; 878 } 879 880 static struct notifier_block kfence_reboot_notifier = { 881 .notifier_call = kfence_reboot_callback, 882 .priority = INT_MAX, /* Run early to stop timers ASAP */ 883 }; 884 885 static void wake_up_kfence_timer(struct irq_work *work) 886 { 887 wake_up(&allocation_wait); 888 } 889 static DEFINE_IRQ_WORK(wake_up_kfence_timer_work, wake_up_kfence_timer); 890 #endif 891 892 /* 893 * Set up delayed work, which will enable and disable the static key. We need to 894 * use a work queue (rather than a simple timer), since enabling and disabling a 895 * static key cannot be done from an interrupt. 896 * 897 * Note: Toggling a static branch currently causes IPIs, and here we'll end up 898 * with a total of 2 IPIs to all CPUs. If this ends up a problem in future (with 899 * more aggressive sampling intervals), we could get away with a variant that 900 * avoids IPIs, at the cost of not immediately capturing allocations if the 901 * instructions remain cached. 902 */ 903 static void toggle_allocation_gate(struct work_struct *work) 904 { 905 if (!READ_ONCE(kfence_enabled)) 906 return; 907 908 atomic_set(&kfence_allocation_gate, -kfence_burst); 909 #ifdef CONFIG_KFENCE_STATIC_KEYS 910 /* Enable static key, and await allocation to happen. */ 911 static_branch_enable(&kfence_allocation_key); 912 913 wait_event_idle(allocation_wait, 914 atomic_read(&kfence_allocation_gate) > 0 || 915 !READ_ONCE(kfence_enabled)); 916 917 /* Disable static key and reset timer. */ 918 static_branch_disable(&kfence_allocation_key); 919 #endif 920 queue_delayed_work(system_dfl_wq, &kfence_timer, 921 msecs_to_jiffies(kfence_sample_interval)); 922 } 923 924 /* === Public interface ===================================================== */ 925 926 void __init kfence_alloc_pool_and_metadata(void) 927 { 928 if (!kfence_sample_interval) 929 return; 930 931 /* 932 * If KASAN hardware tags are enabled, disable KFENCE, because it 933 * does not support MTE yet. 934 */ 935 if (kasan_hw_tags_enabled()) { 936 pr_info("disabled as KASAN HW tags are enabled\n"); 937 if (__kfence_pool) { 938 memblock_free(__kfence_pool, KFENCE_POOL_SIZE); 939 __kfence_pool = NULL; 940 } 941 kfence_sample_interval = 0; 942 return; 943 } 944 945 /* 946 * If the pool has already been initialized by arch, there is no need to 947 * re-allocate the memory pool. 948 */ 949 if (!__kfence_pool) 950 __kfence_pool = memblock_alloc(KFENCE_POOL_SIZE, PAGE_SIZE); 951 952 if (!__kfence_pool) { 953 pr_err("failed to allocate pool\n"); 954 return; 955 } 956 957 /* The memory allocated by memblock has been zeroed out. */ 958 kfence_metadata_init = memblock_alloc(KFENCE_METADATA_SIZE, PAGE_SIZE); 959 if (!kfence_metadata_init) { 960 pr_err("failed to allocate metadata\n"); 961 memblock_free(__kfence_pool, KFENCE_POOL_SIZE); 962 __kfence_pool = NULL; 963 } 964 } 965 966 static void kfence_init_enable(void) 967 { 968 if (!IS_ENABLED(CONFIG_KFENCE_STATIC_KEYS)) 969 static_branch_enable(&kfence_allocation_key); 970 971 if (kfence_deferrable) 972 INIT_DEFERRABLE_WORK(&kfence_timer, toggle_allocation_gate); 973 else 974 INIT_DELAYED_WORK(&kfence_timer, toggle_allocation_gate); 975 976 if (kfence_check_on_panic) 977 atomic_notifier_chain_register(&panic_notifier_list, &kfence_check_canary_notifier); 978 979 #ifdef CONFIG_KFENCE_STATIC_KEYS 980 register_reboot_notifier(&kfence_reboot_notifier); 981 #endif 982 983 WRITE_ONCE(kfence_enabled, true); 984 queue_delayed_work(system_dfl_wq, &kfence_timer, 0); 985 986 pr_info("initialized - using %lu bytes for %d objects at 0x%p-0x%p\n", KFENCE_POOL_SIZE, 987 CONFIG_KFENCE_NUM_OBJECTS, (void *)__kfence_pool, 988 (void *)(__kfence_pool + KFENCE_POOL_SIZE)); 989 } 990 991 void __init kfence_init(void) 992 { 993 stack_hash_seed = get_random_u32(); 994 995 /* Setting kfence_sample_interval to 0 on boot disables KFENCE. */ 996 if (!kfence_sample_interval) 997 return; 998 999 if (!kfence_init_pool_early()) { 1000 pr_err("%s failed\n", __func__); 1001 return; 1002 } 1003 1004 kfence_init_enable(); 1005 } 1006 1007 static int kfence_init_late(void) 1008 { 1009 const unsigned long nr_pages_pool = KFENCE_POOL_SIZE / PAGE_SIZE; 1010 const unsigned long nr_pages_meta = KFENCE_METADATA_SIZE / PAGE_SIZE; 1011 unsigned long addr = (unsigned long)__kfence_pool; 1012 unsigned long free_size = KFENCE_POOL_SIZE; 1013 int err = -ENOMEM; 1014 1015 #ifdef CONFIG_CONTIG_ALLOC 1016 struct page *pages; 1017 1018 pages = alloc_contig_pages(nr_pages_pool, GFP_KERNEL | __GFP_SKIP_KASAN, 1019 first_online_node, NULL); 1020 if (!pages) 1021 return -ENOMEM; 1022 1023 __kfence_pool = page_to_virt(pages); 1024 pages = alloc_contig_pages(nr_pages_meta, GFP_KERNEL | __GFP_SKIP_KASAN, 1025 first_online_node, NULL); 1026 if (pages) 1027 kfence_metadata_init = page_to_virt(pages); 1028 #else 1029 if (nr_pages_pool > MAX_ORDER_NR_PAGES || 1030 nr_pages_meta > MAX_ORDER_NR_PAGES) { 1031 pr_warn("KFENCE_NUM_OBJECTS too large for buddy allocator\n"); 1032 return -EINVAL; 1033 } 1034 1035 __kfence_pool = alloc_pages_exact(KFENCE_POOL_SIZE, 1036 GFP_KERNEL | __GFP_SKIP_KASAN); 1037 if (!__kfence_pool) 1038 return -ENOMEM; 1039 1040 kfence_metadata_init = alloc_pages_exact(KFENCE_METADATA_SIZE, 1041 GFP_KERNEL | __GFP_SKIP_KASAN); 1042 #endif 1043 1044 if (!kfence_metadata_init) 1045 goto free_pool; 1046 1047 memzero_explicit(kfence_metadata_init, KFENCE_METADATA_SIZE); 1048 addr = kfence_init_pool(); 1049 if (!addr) { 1050 kfence_init_enable(); 1051 kfence_debugfs_init(); 1052 return 0; 1053 } 1054 1055 pr_err("%s failed\n", __func__); 1056 free_size = KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool); 1057 err = -EBUSY; 1058 1059 #ifdef CONFIG_CONTIG_ALLOC 1060 free_contig_range(page_to_pfn(virt_to_page((void *)kfence_metadata_init)), 1061 nr_pages_meta); 1062 free_pool: 1063 free_contig_range(page_to_pfn(virt_to_page((void *)addr)), 1064 free_size / PAGE_SIZE); 1065 #else 1066 free_pages_exact((void *)kfence_metadata_init, KFENCE_METADATA_SIZE); 1067 free_pool: 1068 free_pages_exact((void *)addr, free_size); 1069 #endif 1070 1071 kfence_metadata_init = NULL; 1072 __kfence_pool = NULL; 1073 return err; 1074 } 1075 1076 static int kfence_enable_late(void) 1077 { 1078 if (!__kfence_pool) 1079 return kfence_init_late(); 1080 1081 WRITE_ONCE(kfence_enabled, true); 1082 queue_delayed_work(system_dfl_wq, &kfence_timer, 0); 1083 pr_info("re-enabled\n"); 1084 return 0; 1085 } 1086 1087 void kfence_shutdown_cache(struct kmem_cache *s) 1088 { 1089 unsigned long flags; 1090 struct kfence_metadata *meta; 1091 int i; 1092 1093 /* Pairs with release in kfence_init_pool(). */ 1094 if (!smp_load_acquire(&kfence_metadata)) 1095 return; 1096 1097 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 1098 bool in_use; 1099 1100 meta = &kfence_metadata[i]; 1101 1102 /* 1103 * If we observe some inconsistent cache and state pair where we 1104 * should have returned false here, cache destruction is racing 1105 * with either kmem_cache_alloc() or kmem_cache_free(). Taking 1106 * the lock will not help, as different critical section 1107 * serialization will have the same outcome. 1108 */ 1109 if (READ_ONCE(meta->cache) != s || !kfence_obj_allocated(meta)) 1110 continue; 1111 1112 raw_spin_lock_irqsave(&meta->lock, flags); 1113 in_use = meta->cache == s && kfence_obj_allocated(meta); 1114 raw_spin_unlock_irqrestore(&meta->lock, flags); 1115 1116 if (in_use) { 1117 /* 1118 * This cache still has allocations, and we should not 1119 * release them back into the freelist so they can still 1120 * safely be used and retain the kernel's default 1121 * behaviour of keeping the allocations alive (leak the 1122 * cache); however, they effectively become "zombie 1123 * allocations" as the KFENCE objects are the only ones 1124 * still in use and the owning cache is being destroyed. 1125 * 1126 * We mark them freed, so that any subsequent use shows 1127 * more useful error messages that will include stack 1128 * traces of the user of the object, the original 1129 * allocation, and caller to shutdown_cache(). 1130 */ 1131 kfence_guarded_free((void *)meta->addr, meta, /*zombie=*/true); 1132 } 1133 } 1134 1135 for (i = 0; i < CONFIG_KFENCE_NUM_OBJECTS; i++) { 1136 meta = &kfence_metadata[i]; 1137 1138 /* See above. */ 1139 if (READ_ONCE(meta->cache) != s || READ_ONCE(meta->state) != KFENCE_OBJECT_FREED) 1140 continue; 1141 1142 raw_spin_lock_irqsave(&meta->lock, flags); 1143 if (meta->cache == s && meta->state == KFENCE_OBJECT_FREED) 1144 meta->cache = NULL; 1145 raw_spin_unlock_irqrestore(&meta->lock, flags); 1146 } 1147 } 1148 1149 void *__kfence_alloc(struct kmem_cache *s, size_t size, gfp_t flags) 1150 { 1151 unsigned long stack_entries[KFENCE_STACK_DEPTH]; 1152 size_t num_stack_entries; 1153 u32 alloc_stack_hash; 1154 int allocation_gate; 1155 1156 /* 1157 * Perform size check before switching kfence_allocation_gate, so that 1158 * we don't disable KFENCE without making an allocation. 1159 */ 1160 if (size > PAGE_SIZE) { 1161 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]); 1162 return NULL; 1163 } 1164 1165 /* 1166 * Skip allocations from non-default zones, including DMA. We cannot 1167 * guarantee that pages in the KFENCE pool will have the requested 1168 * properties (e.g. reside in DMAable memory). 1169 */ 1170 if ((flags & GFP_ZONEMASK) || 1171 ((flags & __GFP_THISNODE) && num_online_nodes() > 1) || 1172 (s->flags & (SLAB_CACHE_DMA | SLAB_CACHE_DMA32))) { 1173 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_INCOMPAT]); 1174 return NULL; 1175 } 1176 1177 /* 1178 * Skip allocations for this slab, if KFENCE has been disabled for 1179 * this slab. 1180 */ 1181 if (s->flags & SLAB_SKIP_KFENCE) 1182 return NULL; 1183 1184 allocation_gate = atomic_inc_return(&kfence_allocation_gate); 1185 if (allocation_gate > 1) 1186 return NULL; 1187 #ifdef CONFIG_KFENCE_STATIC_KEYS 1188 /* 1189 * waitqueue_active() is fully ordered after the update of 1190 * kfence_allocation_gate per atomic_inc_return(). 1191 */ 1192 if (allocation_gate == 1 && waitqueue_active(&allocation_wait)) { 1193 /* 1194 * Calling wake_up() here may deadlock when allocations happen 1195 * from within timer code. Use an irq_work to defer it. 1196 */ 1197 irq_work_queue(&wake_up_kfence_timer_work); 1198 } 1199 #endif 1200 1201 if (!READ_ONCE(kfence_enabled)) 1202 return NULL; 1203 1204 num_stack_entries = stack_trace_save(stack_entries, KFENCE_STACK_DEPTH, 0); 1205 1206 /* 1207 * Do expensive check for coverage of allocation in slow-path after 1208 * allocation_gate has already become non-zero, even though it might 1209 * mean not making any allocation within a given sample interval. 1210 * 1211 * This ensures reasonable allocation coverage when the pool is almost 1212 * full, including avoiding long-lived allocations of the same source 1213 * filling up the pool (e.g. pagecache allocations). 1214 */ 1215 alloc_stack_hash = get_alloc_stack_hash(stack_entries, num_stack_entries); 1216 if (should_skip_covered() && alloc_covered_contains(alloc_stack_hash)) { 1217 atomic_long_inc(&counters[KFENCE_COUNTER_SKIP_COVERED]); 1218 return NULL; 1219 } 1220 1221 return kfence_guarded_alloc(s, size, flags, stack_entries, num_stack_entries, 1222 alloc_stack_hash); 1223 } 1224 1225 size_t kfence_ksize(const void *addr) 1226 { 1227 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr); 1228 1229 /* 1230 * Read locklessly -- if there is a race with __kfence_alloc(), this is 1231 * either a use-after-free or invalid access. 1232 */ 1233 return meta ? meta->size : 0; 1234 } 1235 1236 void *kfence_object_start(const void *addr) 1237 { 1238 const struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr); 1239 1240 /* 1241 * Read locklessly -- if there is a race with __kfence_alloc(), this is 1242 * either a use-after-free or invalid access. 1243 */ 1244 return meta ? (void *)meta->addr : NULL; 1245 } 1246 1247 void __kfence_free(void *addr) 1248 { 1249 struct kfence_metadata *meta = addr_to_metadata((unsigned long)addr); 1250 1251 #ifdef CONFIG_MEMCG 1252 KFENCE_WARN_ON(meta->obj_exts.objcg); 1253 #endif 1254 /* 1255 * If the objects of the cache are SLAB_TYPESAFE_BY_RCU, defer freeing 1256 * the object, as the object page may be recycled for other-typed 1257 * objects once it has been freed. meta->cache may be NULL if the cache 1258 * was destroyed. 1259 * Save the stack trace here so that reports show where the user freed 1260 * the object. 1261 */ 1262 if (unlikely(meta->cache && (meta->cache->flags & SLAB_TYPESAFE_BY_RCU))) { 1263 unsigned long flags; 1264 1265 raw_spin_lock_irqsave(&meta->lock, flags); 1266 metadata_update_state(meta, KFENCE_OBJECT_RCU_FREEING, NULL, 0); 1267 raw_spin_unlock_irqrestore(&meta->lock, flags); 1268 call_rcu(&meta->rcu_head, rcu_guarded_free); 1269 } else { 1270 kfence_guarded_free(addr, meta, false); 1271 } 1272 } 1273 1274 bool kfence_handle_page_fault(unsigned long addr, bool is_write, struct pt_regs *regs) 1275 { 1276 const int page_index = (addr - (unsigned long)__kfence_pool) / PAGE_SIZE; 1277 struct kfence_metadata *to_report = NULL; 1278 unsigned long unprotected_page = 0; 1279 enum kfence_error_type error_type; 1280 enum kfence_fault fault; 1281 unsigned long flags; 1282 1283 if (!is_kfence_address((void *)addr)) 1284 return false; 1285 1286 if (!READ_ONCE(kfence_enabled)) /* If disabled at runtime ... */ 1287 return kfence_unprotect(addr); /* ... unprotect and proceed. */ 1288 1289 atomic_long_inc(&counters[KFENCE_COUNTER_BUGS]); 1290 1291 if (page_index % 2) { 1292 /* This is a redzone, report a buffer overflow. */ 1293 struct kfence_metadata *meta; 1294 int distance = 0; 1295 1296 meta = addr_to_metadata(addr - PAGE_SIZE); 1297 if (meta && kfence_obj_allocated(meta)) { 1298 to_report = meta; 1299 /* Data race ok; distance calculation approximate. */ 1300 distance = addr - data_race(meta->addr + meta->size); 1301 } 1302 1303 meta = addr_to_metadata(addr + PAGE_SIZE); 1304 if (meta && kfence_obj_allocated(meta)) { 1305 /* Data race ok; distance calculation approximate. */ 1306 if (!to_report || distance > data_race(meta->addr) - addr) 1307 to_report = meta; 1308 } 1309 1310 if (!to_report) 1311 goto out; 1312 1313 error_type = KFENCE_ERROR_OOB; 1314 unprotected_page = addr; 1315 1316 /* 1317 * If the object was freed before we took the look we can still 1318 * report this as an OOB -- the report will simply show the 1319 * stacktrace of the free as well. 1320 */ 1321 } else { 1322 to_report = addr_to_metadata(addr); 1323 if (!to_report) 1324 goto out; 1325 1326 error_type = KFENCE_ERROR_UAF; 1327 /* 1328 * We may race with __kfence_alloc(), and it is possible that a 1329 * freed object may be reallocated. We simply report this as a 1330 * use-after-free, with the stack trace showing the place where 1331 * the object was re-allocated. 1332 */ 1333 } 1334 1335 out: 1336 if (to_report) { 1337 raw_spin_lock_irqsave(&to_report->lock, flags); 1338 to_report->unprotected_page = unprotected_page; 1339 fault = kfence_report_error(addr, is_write, regs, to_report, error_type); 1340 raw_spin_unlock_irqrestore(&to_report->lock, flags); 1341 } else { 1342 /* This may be a UAF or OOB access, but we can't be sure. */ 1343 fault = kfence_report_error(addr, is_write, regs, NULL, KFENCE_ERROR_INVALID); 1344 } 1345 1346 kfence_handle_fault(fault); 1347 1348 return kfence_unprotect(addr); /* Unprotect and let access proceed. */ 1349 } 1350