1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Stack depot - a stack trace storage that avoids duplication. 4 * 5 * Internally, stack depot maintains a hash table of unique stacktraces. The 6 * stack traces themselves are stored contiguously one after another in a set 7 * of separate page allocations. 8 * 9 * Author: Alexander Potapenko <glider@google.com> 10 * Copyright (C) 2016 Google, Inc. 11 * 12 * Based on the code by Dmitry Chernenkov. 13 */ 14 15 #define pr_fmt(fmt) "stackdepot: " fmt 16 17 #include <linux/gfp.h> 18 #include <linux/jhash.h> 19 #include <linux/kernel.h> 20 #include <linux/kmsan.h> 21 #include <linux/mm.h> 22 #include <linux/mutex.h> 23 #include <linux/percpu.h> 24 #include <linux/printk.h> 25 #include <linux/slab.h> 26 #include <linux/stacktrace.h> 27 #include <linux/stackdepot.h> 28 #include <linux/string.h> 29 #include <linux/types.h> 30 #include <linux/memblock.h> 31 #include <linux/kasan-enabled.h> 32 33 #define DEPOT_HANDLE_BITS (sizeof(depot_stack_handle_t) * 8) 34 35 #define DEPOT_VALID_BITS 1 36 #define DEPOT_POOL_ORDER 2 /* Pool size order, 4 pages */ 37 #define DEPOT_POOL_SIZE (1LL << (PAGE_SHIFT + DEPOT_POOL_ORDER)) 38 #define DEPOT_STACK_ALIGN 4 39 #define DEPOT_OFFSET_BITS (DEPOT_POOL_ORDER + PAGE_SHIFT - DEPOT_STACK_ALIGN) 40 #define DEPOT_POOL_INDEX_BITS (DEPOT_HANDLE_BITS - DEPOT_VALID_BITS - \ 41 DEPOT_OFFSET_BITS - STACK_DEPOT_EXTRA_BITS) 42 #define DEPOT_POOLS_CAP 8192 43 #define DEPOT_MAX_POOLS \ 44 (((1LL << (DEPOT_POOL_INDEX_BITS)) < DEPOT_POOLS_CAP) ? \ 45 (1LL << (DEPOT_POOL_INDEX_BITS)) : DEPOT_POOLS_CAP) 46 47 /* Compact structure that stores a reference to a stack. */ 48 union handle_parts { 49 depot_stack_handle_t handle; 50 struct { 51 u32 pool_index : DEPOT_POOL_INDEX_BITS; 52 u32 offset : DEPOT_OFFSET_BITS; 53 u32 valid : DEPOT_VALID_BITS; 54 u32 extra : STACK_DEPOT_EXTRA_BITS; 55 }; 56 }; 57 58 struct stack_record { 59 struct stack_record *next; /* Link in the hash table */ 60 u32 hash; /* Hash in the hash table */ 61 u32 size; /* Number of stored frames */ 62 union handle_parts handle; 63 unsigned long entries[]; /* Variable-sized array of frames */ 64 }; 65 66 static bool stack_depot_disabled; 67 static bool __stack_depot_early_init_requested __initdata = IS_ENABLED(CONFIG_STACKDEPOT_ALWAYS_INIT); 68 static bool __stack_depot_early_init_passed __initdata; 69 70 /* Use one hash table bucket per 16 KB of memory. */ 71 #define STACK_HASH_TABLE_SCALE 14 72 /* Limit the number of buckets between 4K and 1M. */ 73 #define STACK_BUCKET_NUMBER_ORDER_MIN 12 74 #define STACK_BUCKET_NUMBER_ORDER_MAX 20 75 /* Initial seed for jhash2. */ 76 #define STACK_HASH_SEED 0x9747b28c 77 78 /* Hash table of pointers to stored stack traces. */ 79 static struct stack_record **stack_table; 80 /* Fixed order of the number of table buckets. Used when KASAN is enabled. */ 81 static unsigned int stack_bucket_number_order; 82 /* Hash mask for indexing the table. */ 83 static unsigned int stack_hash_mask; 84 85 /* Array of memory regions that store stack traces. */ 86 static void *stack_pools[DEPOT_MAX_POOLS]; 87 /* Currently used pool in stack_pools. */ 88 static int pool_index; 89 /* Offset to the unused space in the currently used pool. */ 90 static size_t pool_offset; 91 /* Lock that protects the variables above. */ 92 static DEFINE_RAW_SPINLOCK(pool_lock); 93 /* 94 * Stack depot tries to keep an extra pool allocated even before it runs out 95 * of space in the currently used pool. 96 * This flag marks that this next extra pool needs to be allocated and 97 * initialized. It has the value 0 when either the next pool is not yet 98 * initialized or the limit on the number of pools is reached. 99 */ 100 static int next_pool_required = 1; 101 102 static int __init disable_stack_depot(char *str) 103 { 104 int ret; 105 106 ret = kstrtobool(str, &stack_depot_disabled); 107 if (!ret && stack_depot_disabled) { 108 pr_info("disabled\n"); 109 stack_table = NULL; 110 } 111 return 0; 112 } 113 early_param("stack_depot_disable", disable_stack_depot); 114 115 void __init stack_depot_request_early_init(void) 116 { 117 /* Too late to request early init now. */ 118 WARN_ON(__stack_depot_early_init_passed); 119 120 __stack_depot_early_init_requested = true; 121 } 122 123 /* Allocates a hash table via memblock. Can only be used during early boot. */ 124 int __init stack_depot_early_init(void) 125 { 126 unsigned long entries = 0; 127 128 /* This function must be called only once, from mm_init(). */ 129 if (WARN_ON(__stack_depot_early_init_passed)) 130 return 0; 131 __stack_depot_early_init_passed = true; 132 133 /* 134 * If KASAN is enabled, use the maximum order: KASAN is frequently used 135 * in fuzzing scenarios, which leads to a large number of different 136 * stack traces being stored in stack depot. 137 */ 138 if (kasan_enabled() && !stack_bucket_number_order) 139 stack_bucket_number_order = STACK_BUCKET_NUMBER_ORDER_MAX; 140 141 if (!__stack_depot_early_init_requested || stack_depot_disabled) 142 return 0; 143 144 /* 145 * If stack_bucket_number_order is not set, leave entries as 0 to rely 146 * on the automatic calculations performed by alloc_large_system_hash. 147 */ 148 if (stack_bucket_number_order) 149 entries = 1UL << stack_bucket_number_order; 150 pr_info("allocating hash table via alloc_large_system_hash\n"); 151 stack_table = alloc_large_system_hash("stackdepot", 152 sizeof(struct stack_record *), 153 entries, 154 STACK_HASH_TABLE_SCALE, 155 HASH_EARLY | HASH_ZERO, 156 NULL, 157 &stack_hash_mask, 158 1UL << STACK_BUCKET_NUMBER_ORDER_MIN, 159 1UL << STACK_BUCKET_NUMBER_ORDER_MAX); 160 if (!stack_table) { 161 pr_err("hash table allocation failed, disabling\n"); 162 stack_depot_disabled = true; 163 return -ENOMEM; 164 } 165 166 return 0; 167 } 168 169 /* Allocates a hash table via kvcalloc. Can be used after boot. */ 170 int stack_depot_init(void) 171 { 172 static DEFINE_MUTEX(stack_depot_init_mutex); 173 unsigned long entries; 174 int ret = 0; 175 176 mutex_lock(&stack_depot_init_mutex); 177 178 if (stack_depot_disabled || stack_table) 179 goto out_unlock; 180 181 /* 182 * Similarly to stack_depot_early_init, use stack_bucket_number_order 183 * if assigned, and rely on automatic scaling otherwise. 184 */ 185 if (stack_bucket_number_order) { 186 entries = 1UL << stack_bucket_number_order; 187 } else { 188 int scale = STACK_HASH_TABLE_SCALE; 189 190 entries = nr_free_buffer_pages(); 191 entries = roundup_pow_of_two(entries); 192 193 if (scale > PAGE_SHIFT) 194 entries >>= (scale - PAGE_SHIFT); 195 else 196 entries <<= (PAGE_SHIFT - scale); 197 } 198 199 if (entries < 1UL << STACK_BUCKET_NUMBER_ORDER_MIN) 200 entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MIN; 201 if (entries > 1UL << STACK_BUCKET_NUMBER_ORDER_MAX) 202 entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MAX; 203 204 pr_info("allocating hash table of %lu entries via kvcalloc\n", entries); 205 stack_table = kvcalloc(entries, sizeof(struct stack_record *), GFP_KERNEL); 206 if (!stack_table) { 207 pr_err("hash table allocation failed, disabling\n"); 208 stack_depot_disabled = true; 209 ret = -ENOMEM; 210 goto out_unlock; 211 } 212 stack_hash_mask = entries - 1; 213 214 out_unlock: 215 mutex_unlock(&stack_depot_init_mutex); 216 217 return ret; 218 } 219 EXPORT_SYMBOL_GPL(stack_depot_init); 220 221 /* Uses preallocated memory to initialize a new stack depot pool. */ 222 static void depot_init_pool(void **prealloc) 223 { 224 /* 225 * If the next pool is already initialized or the maximum number of 226 * pools is reached, do not use the preallocated memory. 227 * smp_load_acquire() here pairs with smp_store_release() below and 228 * in depot_alloc_stack(). 229 */ 230 if (!smp_load_acquire(&next_pool_required)) 231 return; 232 233 /* Check if the current pool is not yet allocated. */ 234 if (stack_pools[pool_index] == NULL) { 235 /* Use the preallocated memory for the current pool. */ 236 stack_pools[pool_index] = *prealloc; 237 *prealloc = NULL; 238 } else { 239 /* 240 * Otherwise, use the preallocated memory for the next pool 241 * as long as we do not exceed the maximum number of pools. 242 */ 243 if (pool_index + 1 < DEPOT_MAX_POOLS) { 244 stack_pools[pool_index + 1] = *prealloc; 245 *prealloc = NULL; 246 } 247 /* 248 * At this point, either the next pool is initialized or the 249 * maximum number of pools is reached. In either case, take 250 * note that initializing another pool is not required. 251 * This smp_store_release pairs with smp_load_acquire() above 252 * and in stack_depot_save(). 253 */ 254 smp_store_release(&next_pool_required, 0); 255 } 256 } 257 258 /* Allocates a new stack in a stack depot pool. */ 259 static struct stack_record * 260 depot_alloc_stack(unsigned long *entries, int size, u32 hash, void **prealloc) 261 { 262 struct stack_record *stack; 263 size_t required_size = struct_size(stack, entries, size); 264 265 required_size = ALIGN(required_size, 1 << DEPOT_STACK_ALIGN); 266 267 /* Check if there is not enough space in the current pool. */ 268 if (unlikely(pool_offset + required_size > DEPOT_POOL_SIZE)) { 269 /* Bail out if we reached the pool limit. */ 270 if (unlikely(pool_index + 1 >= DEPOT_MAX_POOLS)) { 271 WARN_ONCE(1, "Stack depot reached limit capacity"); 272 return NULL; 273 } 274 275 /* 276 * Move on to the next pool. 277 * WRITE_ONCE pairs with potential concurrent read in 278 * stack_depot_fetch(). 279 */ 280 WRITE_ONCE(pool_index, pool_index + 1); 281 pool_offset = 0; 282 /* 283 * If the maximum number of pools is not reached, take note 284 * that the next pool needs to initialized. 285 * smp_store_release() here pairs with smp_load_acquire() in 286 * stack_depot_save() and depot_init_pool(). 287 */ 288 if (pool_index + 1 < DEPOT_MAX_POOLS) 289 smp_store_release(&next_pool_required, 1); 290 } 291 292 /* Assign the preallocated memory to a pool if required. */ 293 if (*prealloc) 294 depot_init_pool(prealloc); 295 296 /* Check if we have a pool to save the stack trace. */ 297 if (stack_pools[pool_index] == NULL) 298 return NULL; 299 300 /* Save the stack trace. */ 301 stack = stack_pools[pool_index] + pool_offset; 302 stack->hash = hash; 303 stack->size = size; 304 stack->handle.pool_index = pool_index; 305 stack->handle.offset = pool_offset >> DEPOT_STACK_ALIGN; 306 stack->handle.valid = 1; 307 stack->handle.extra = 0; 308 memcpy(stack->entries, entries, flex_array_size(stack, entries, size)); 309 pool_offset += required_size; 310 /* 311 * Let KMSAN know the stored stack record is initialized. This shall 312 * prevent false positive reports if instrumented code accesses it. 313 */ 314 kmsan_unpoison_memory(stack, required_size); 315 316 return stack; 317 } 318 319 /* Calculates the hash for a stack. */ 320 static inline u32 hash_stack(unsigned long *entries, unsigned int size) 321 { 322 return jhash2((u32 *)entries, 323 array_size(size, sizeof(*entries)) / sizeof(u32), 324 STACK_HASH_SEED); 325 } 326 327 /* 328 * Non-instrumented version of memcmp(). 329 * Does not check the lexicographical order, only the equality. 330 */ 331 static inline 332 int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2, 333 unsigned int n) 334 { 335 for ( ; n-- ; u1++, u2++) { 336 if (*u1 != *u2) 337 return 1; 338 } 339 return 0; 340 } 341 342 /* Finds a stack in a bucket of the hash table. */ 343 static inline struct stack_record *find_stack(struct stack_record *bucket, 344 unsigned long *entries, int size, 345 u32 hash) 346 { 347 struct stack_record *found; 348 349 for (found = bucket; found; found = found->next) { 350 if (found->hash == hash && 351 found->size == size && 352 !stackdepot_memcmp(entries, found->entries, size)) 353 return found; 354 } 355 return NULL; 356 } 357 358 depot_stack_handle_t __stack_depot_save(unsigned long *entries, 359 unsigned int nr_entries, 360 gfp_t alloc_flags, bool can_alloc) 361 { 362 struct stack_record *found = NULL, **bucket; 363 union handle_parts retval = { .handle = 0 }; 364 struct page *page = NULL; 365 void *prealloc = NULL; 366 unsigned long flags; 367 u32 hash; 368 369 /* 370 * If this stack trace is from an interrupt, including anything before 371 * interrupt entry usually leads to unbounded stack depot growth. 372 * 373 * Since use of filter_irq_stacks() is a requirement to ensure stack 374 * depot can efficiently deduplicate interrupt stacks, always 375 * filter_irq_stacks() to simplify all callers' use of stack depot. 376 */ 377 nr_entries = filter_irq_stacks(entries, nr_entries); 378 379 if (unlikely(nr_entries == 0) || stack_depot_disabled) 380 goto fast_exit; 381 382 hash = hash_stack(entries, nr_entries); 383 bucket = &stack_table[hash & stack_hash_mask]; 384 385 /* 386 * Fast path: look the stack trace up without locking. 387 * The smp_load_acquire() here pairs with smp_store_release() to 388 * |bucket| below. 389 */ 390 found = find_stack(smp_load_acquire(bucket), entries, nr_entries, hash); 391 if (found) 392 goto exit; 393 394 /* 395 * Check if another stack pool needs to be initialized. If so, allocate 396 * the memory now - we won't be able to do that under the lock. 397 * 398 * The smp_load_acquire() here pairs with smp_store_release() to 399 * |next_pool_inited| in depot_alloc_stack() and depot_init_pool(). 400 */ 401 if (unlikely(can_alloc && smp_load_acquire(&next_pool_required))) { 402 /* 403 * Zero out zone modifiers, as we don't have specific zone 404 * requirements. Keep the flags related to allocation in atomic 405 * contexts and I/O. 406 */ 407 alloc_flags &= ~GFP_ZONEMASK; 408 alloc_flags &= (GFP_ATOMIC | GFP_KERNEL); 409 alloc_flags |= __GFP_NOWARN; 410 page = alloc_pages(alloc_flags, DEPOT_POOL_ORDER); 411 if (page) 412 prealloc = page_address(page); 413 } 414 415 raw_spin_lock_irqsave(&pool_lock, flags); 416 417 found = find_stack(*bucket, entries, nr_entries, hash); 418 if (!found) { 419 struct stack_record *new = 420 depot_alloc_stack(entries, nr_entries, hash, &prealloc); 421 422 if (new) { 423 new->next = *bucket; 424 /* 425 * This smp_store_release() pairs with 426 * smp_load_acquire() from |bucket| above. 427 */ 428 smp_store_release(bucket, new); 429 found = new; 430 } 431 } else if (prealloc) { 432 /* 433 * Stack depot already contains this stack trace, but let's 434 * keep the preallocated memory for the future. 435 */ 436 depot_init_pool(&prealloc); 437 } 438 439 raw_spin_unlock_irqrestore(&pool_lock, flags); 440 exit: 441 if (prealloc) { 442 /* Stack depot didn't use this memory, free it. */ 443 free_pages((unsigned long)prealloc, DEPOT_POOL_ORDER); 444 } 445 if (found) 446 retval.handle = found->handle.handle; 447 fast_exit: 448 return retval.handle; 449 } 450 EXPORT_SYMBOL_GPL(__stack_depot_save); 451 452 depot_stack_handle_t stack_depot_save(unsigned long *entries, 453 unsigned int nr_entries, 454 gfp_t alloc_flags) 455 { 456 return __stack_depot_save(entries, nr_entries, alloc_flags, true); 457 } 458 EXPORT_SYMBOL_GPL(stack_depot_save); 459 460 unsigned int stack_depot_fetch(depot_stack_handle_t handle, 461 unsigned long **entries) 462 { 463 union handle_parts parts = { .handle = handle }; 464 /* 465 * READ_ONCE pairs with potential concurrent write in 466 * depot_alloc_stack. 467 */ 468 int pool_index_cached = READ_ONCE(pool_index); 469 void *pool; 470 size_t offset = parts.offset << DEPOT_STACK_ALIGN; 471 struct stack_record *stack; 472 473 *entries = NULL; 474 /* 475 * Let KMSAN know *entries is initialized. This shall prevent false 476 * positive reports if instrumented code accesses it. 477 */ 478 kmsan_unpoison_memory(entries, sizeof(*entries)); 479 480 if (!handle) 481 return 0; 482 483 if (parts.pool_index > pool_index_cached) { 484 WARN(1, "pool index %d out of bounds (%d) for stack id %08x\n", 485 parts.pool_index, pool_index_cached, handle); 486 return 0; 487 } 488 pool = stack_pools[parts.pool_index]; 489 if (!pool) 490 return 0; 491 stack = pool + offset; 492 493 *entries = stack->entries; 494 return stack->size; 495 } 496 EXPORT_SYMBOL_GPL(stack_depot_fetch); 497 498 void stack_depot_print(depot_stack_handle_t stack) 499 { 500 unsigned long *entries; 501 unsigned int nr_entries; 502 503 nr_entries = stack_depot_fetch(stack, &entries); 504 if (nr_entries > 0) 505 stack_trace_print(entries, nr_entries, 0); 506 } 507 EXPORT_SYMBOL_GPL(stack_depot_print); 508 509 int stack_depot_snprint(depot_stack_handle_t handle, char *buf, size_t size, 510 int spaces) 511 { 512 unsigned long *entries; 513 unsigned int nr_entries; 514 515 nr_entries = stack_depot_fetch(handle, &entries); 516 return nr_entries ? stack_trace_snprint(buf, size, entries, nr_entries, 517 spaces) : 0; 518 } 519 EXPORT_SYMBOL_GPL(stack_depot_snprint); 520 521 depot_stack_handle_t __must_check stack_depot_set_extra_bits( 522 depot_stack_handle_t handle, unsigned int extra_bits) 523 { 524 union handle_parts parts = { .handle = handle }; 525 526 /* Don't set extra bits on empty handles. */ 527 if (!handle) 528 return 0; 529 530 parts.extra = extra_bits; 531 return parts.handle; 532 } 533 EXPORT_SYMBOL(stack_depot_set_extra_bits); 534 535 unsigned int stack_depot_get_extra_bits(depot_stack_handle_t handle) 536 { 537 union handle_parts parts = { .handle = handle }; 538 539 return parts.extra; 540 } 541 EXPORT_SYMBOL(stack_depot_get_extra_bits); 542