| 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/debugfs.h> |
| 18 | #include <linux/gfp.h> |
| 19 | #include <linux/jhash.h> |
| 20 | #include <linux/kernel.h> |
| 21 | #include <linux/kmsan.h> |
| 22 | #include <linux/list.h> |
| 23 | #include <linux/mm.h> |
| 24 | #include <linux/mutex.h> |
| 25 | #include <linux/poison.h> |
| 26 | #include <linux/printk.h> |
| 27 | #include <linux/rculist.h> |
| 28 | #include <linux/rcupdate.h> |
| 29 | #include <linux/refcount.h> |
| 30 | #include <linux/slab.h> |
| 31 | #include <linux/spinlock.h> |
| 32 | #include <linux/stacktrace.h> |
| 33 | #include <linux/stackdepot.h> |
| 34 | #include <linux/string.h> |
| 35 | #include <linux/types.h> |
| 36 | #include <linux/memblock.h> |
| 37 | #include <linux/kasan-enabled.h> |
| 38 | |
| 39 | /* |
| 40 | * The pool_index is offset by 1 so the first record does not have a 0 handle. |
| 41 | */ |
| 42 | static unsigned int stack_max_pools __read_mostly = |
| 43 | MIN((1LL << DEPOT_POOL_INDEX_BITS) - 1, 8192); |
| 44 | |
| 45 | static bool stack_depot_disabled; |
| 46 | static bool __stack_depot_early_init_requested __initdata = IS_ENABLED(CONFIG_STACKDEPOT_ALWAYS_INIT); |
| 47 | static bool __stack_depot_early_init_passed __initdata; |
| 48 | |
| 49 | /* Use one hash table bucket per 16 KB of memory. */ |
| 50 | #define STACK_HASH_TABLE_SCALE 14 |
| 51 | /* Limit the number of buckets between 4K and 1M. */ |
| 52 | #define STACK_BUCKET_NUMBER_ORDER_MIN 12 |
| 53 | #define STACK_BUCKET_NUMBER_ORDER_MAX 20 |
| 54 | /* Initial seed for jhash2. */ |
| 55 | #define STACK_HASH_SEED 0x9747b28c |
| 56 | |
| 57 | /* Hash table of stored stack records. */ |
| 58 | static struct list_head *stack_table; |
| 59 | /* Fixed order of the number of table buckets. Used when KASAN is enabled. */ |
| 60 | static unsigned int stack_bucket_number_order; |
| 61 | /* Hash mask for indexing the table. */ |
| 62 | static unsigned int stack_hash_mask; |
| 63 | |
| 64 | /* Array of memory regions that store stack records. */ |
| 65 | static void **stack_pools; |
| 66 | /* Newly allocated pool that is not yet added to stack_pools. */ |
| 67 | static void *new_pool; |
| 68 | /* Number of pools in stack_pools. */ |
| 69 | static int pools_num; |
| 70 | /* Offset to the unused space in the currently used pool. */ |
| 71 | static size_t pool_offset = DEPOT_POOL_SIZE; |
| 72 | /* Freelist of stack records within stack_pools. */ |
| 73 | static LIST_HEAD(free_stacks); |
| 74 | /* The lock must be held when performing pool or freelist modifications. */ |
| 75 | static DEFINE_RAW_SPINLOCK(pool_lock); |
| 76 | |
| 77 | /* Statistics counters for debugfs. */ |
| 78 | enum depot_counter_id { |
| 79 | DEPOT_COUNTER_REFD_ALLOCS, |
| 80 | DEPOT_COUNTER_REFD_FREES, |
| 81 | DEPOT_COUNTER_REFD_INUSE, |
| 82 | DEPOT_COUNTER_FREELIST_SIZE, |
| 83 | DEPOT_COUNTER_PERSIST_COUNT, |
| 84 | DEPOT_COUNTER_PERSIST_BYTES, |
| 85 | DEPOT_COUNTER_COUNT, |
| 86 | }; |
| 87 | static long counters[DEPOT_COUNTER_COUNT]; |
| 88 | static const char *const counter_names[] = { |
| 89 | [DEPOT_COUNTER_REFD_ALLOCS] = "refcounted_allocations" , |
| 90 | [DEPOT_COUNTER_REFD_FREES] = "refcounted_frees" , |
| 91 | [DEPOT_COUNTER_REFD_INUSE] = "refcounted_in_use" , |
| 92 | [DEPOT_COUNTER_FREELIST_SIZE] = "freelist_size" , |
| 93 | [DEPOT_COUNTER_PERSIST_COUNT] = "persistent_count" , |
| 94 | [DEPOT_COUNTER_PERSIST_BYTES] = "persistent_bytes" , |
| 95 | }; |
| 96 | static_assert(ARRAY_SIZE(counter_names) == DEPOT_COUNTER_COUNT); |
| 97 | |
| 98 | static int __init disable_stack_depot(char *str) |
| 99 | { |
| 100 | return kstrtobool(s: str, res: &stack_depot_disabled); |
| 101 | } |
| 102 | early_param("stack_depot_disable" , disable_stack_depot); |
| 103 | |
| 104 | static int __init parse_max_pools(char *str) |
| 105 | { |
| 106 | const long long limit = (1LL << (DEPOT_POOL_INDEX_BITS)) - 1; |
| 107 | unsigned int max_pools; |
| 108 | int rv; |
| 109 | |
| 110 | rv = kstrtouint(s: str, base: 0, res: &max_pools); |
| 111 | if (rv) |
| 112 | return rv; |
| 113 | |
| 114 | if (max_pools < 1024) { |
| 115 | pr_err("stack_depot_max_pools below 1024, using default of %u\n" , |
| 116 | stack_max_pools); |
| 117 | goto out; |
| 118 | } |
| 119 | |
| 120 | if (max_pools > limit) { |
| 121 | pr_err("stack_depot_max_pools exceeds %lld, using default of %u\n" , |
| 122 | limit, stack_max_pools); |
| 123 | goto out; |
| 124 | } |
| 125 | |
| 126 | stack_max_pools = max_pools; |
| 127 | out: |
| 128 | return 0; |
| 129 | } |
| 130 | early_param("stack_depot_max_pools" , parse_max_pools); |
| 131 | |
| 132 | void __init stack_depot_request_early_init(void) |
| 133 | { |
| 134 | /* Too late to request early init now. */ |
| 135 | WARN_ON(__stack_depot_early_init_passed); |
| 136 | |
| 137 | __stack_depot_early_init_requested = true; |
| 138 | } |
| 139 | |
| 140 | /* Initialize list_head's within the hash table. */ |
| 141 | static void init_stack_table(unsigned long entries) |
| 142 | { |
| 143 | unsigned long i; |
| 144 | |
| 145 | for (i = 0; i < entries; i++) |
| 146 | INIT_LIST_HEAD(list: &stack_table[i]); |
| 147 | } |
| 148 | |
| 149 | /* Allocates a hash table via memblock. Can only be used during early boot. */ |
| 150 | int __init stack_depot_early_init(void) |
| 151 | { |
| 152 | unsigned long entries = 0; |
| 153 | |
| 154 | /* This function must be called only once, from mm_init(). */ |
| 155 | if (WARN_ON(__stack_depot_early_init_passed)) |
| 156 | return 0; |
| 157 | __stack_depot_early_init_passed = true; |
| 158 | |
| 159 | /* |
| 160 | * Print disabled message even if early init has not been requested: |
| 161 | * stack_depot_init() will not print one. |
| 162 | */ |
| 163 | if (stack_depot_disabled) { |
| 164 | pr_info("disabled\n" ); |
| 165 | return 0; |
| 166 | } |
| 167 | |
| 168 | /* |
| 169 | * If KASAN is enabled, use the maximum order: KASAN is frequently used |
| 170 | * in fuzzing scenarios, which leads to a large number of different |
| 171 | * stack traces being stored in stack depot. |
| 172 | */ |
| 173 | if (kasan_enabled() && !stack_bucket_number_order) |
| 174 | stack_bucket_number_order = STACK_BUCKET_NUMBER_ORDER_MAX; |
| 175 | |
| 176 | /* |
| 177 | * Check if early init has been requested after setting |
| 178 | * stack_bucket_number_order: stack_depot_init() uses its value. |
| 179 | */ |
| 180 | if (!__stack_depot_early_init_requested) |
| 181 | return 0; |
| 182 | |
| 183 | /* |
| 184 | * If stack_bucket_number_order is not set, leave entries as 0 to rely |
| 185 | * on the automatic calculations performed by alloc_large_system_hash(). |
| 186 | */ |
| 187 | if (stack_bucket_number_order) |
| 188 | entries = 1UL << stack_bucket_number_order; |
| 189 | pr_info("allocating hash table via alloc_large_system_hash\n" ); |
| 190 | stack_table = alloc_large_system_hash(tablename: "stackdepot" , |
| 191 | bucketsize: sizeof(struct list_head), |
| 192 | numentries: entries, |
| 193 | STACK_HASH_TABLE_SCALE, |
| 194 | HASH_EARLY, |
| 195 | NULL, |
| 196 | hash_mask: &stack_hash_mask, |
| 197 | low_limit: 1UL << STACK_BUCKET_NUMBER_ORDER_MIN, |
| 198 | high_limit: 1UL << STACK_BUCKET_NUMBER_ORDER_MAX); |
| 199 | if (!stack_table) { |
| 200 | pr_err("hash table allocation failed, disabling\n" ); |
| 201 | stack_depot_disabled = true; |
| 202 | return -ENOMEM; |
| 203 | } |
| 204 | if (!entries) { |
| 205 | /* |
| 206 | * Obtain the number of entries that was calculated by |
| 207 | * alloc_large_system_hash(). |
| 208 | */ |
| 209 | entries = stack_hash_mask + 1; |
| 210 | } |
| 211 | init_stack_table(entries); |
| 212 | |
| 213 | pr_info("allocating space for %u stack pools via memblock\n" , |
| 214 | stack_max_pools); |
| 215 | stack_pools = |
| 216 | memblock_alloc(size: stack_max_pools * sizeof(void *), PAGE_SIZE); |
| 217 | if (!stack_pools) { |
| 218 | pr_err("stack pools allocation failed, disabling\n" ); |
| 219 | memblock_free(ptr: stack_table, size: entries * sizeof(struct list_head)); |
| 220 | stack_depot_disabled = true; |
| 221 | return -ENOMEM; |
| 222 | } |
| 223 | |
| 224 | return 0; |
| 225 | } |
| 226 | |
| 227 | /* Allocates a hash table via kvcalloc. Can be used after boot. */ |
| 228 | int stack_depot_init(void) |
| 229 | { |
| 230 | static DEFINE_MUTEX(stack_depot_init_mutex); |
| 231 | unsigned long entries; |
| 232 | int ret = 0; |
| 233 | |
| 234 | mutex_lock(lock: &stack_depot_init_mutex); |
| 235 | |
| 236 | if (stack_depot_disabled || stack_table) |
| 237 | goto out_unlock; |
| 238 | |
| 239 | /* |
| 240 | * Similarly to stack_depot_early_init, use stack_bucket_number_order |
| 241 | * if assigned, and rely on automatic scaling otherwise. |
| 242 | */ |
| 243 | if (stack_bucket_number_order) { |
| 244 | entries = 1UL << stack_bucket_number_order; |
| 245 | } else { |
| 246 | int scale = STACK_HASH_TABLE_SCALE; |
| 247 | |
| 248 | entries = nr_free_buffer_pages(); |
| 249 | entries = roundup_pow_of_two(entries); |
| 250 | |
| 251 | if (scale > PAGE_SHIFT) |
| 252 | entries >>= (scale - PAGE_SHIFT); |
| 253 | else |
| 254 | entries <<= (PAGE_SHIFT - scale); |
| 255 | } |
| 256 | |
| 257 | if (entries < 1UL << STACK_BUCKET_NUMBER_ORDER_MIN) |
| 258 | entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MIN; |
| 259 | if (entries > 1UL << STACK_BUCKET_NUMBER_ORDER_MAX) |
| 260 | entries = 1UL << STACK_BUCKET_NUMBER_ORDER_MAX; |
| 261 | |
| 262 | pr_info("allocating hash table of %lu entries via kvcalloc\n" , entries); |
| 263 | stack_table = kvcalloc(entries, sizeof(struct list_head), GFP_KERNEL); |
| 264 | if (!stack_table) { |
| 265 | pr_err("hash table allocation failed, disabling\n" ); |
| 266 | stack_depot_disabled = true; |
| 267 | ret = -ENOMEM; |
| 268 | goto out_unlock; |
| 269 | } |
| 270 | stack_hash_mask = entries - 1; |
| 271 | init_stack_table(entries); |
| 272 | |
| 273 | pr_info("allocating space for %u stack pools via kvcalloc\n" , |
| 274 | stack_max_pools); |
| 275 | stack_pools = kvcalloc(stack_max_pools, sizeof(void *), GFP_KERNEL); |
| 276 | if (!stack_pools) { |
| 277 | pr_err("stack pools allocation failed, disabling\n" ); |
| 278 | kvfree(addr: stack_table); |
| 279 | stack_depot_disabled = true; |
| 280 | ret = -ENOMEM; |
| 281 | } |
| 282 | |
| 283 | out_unlock: |
| 284 | mutex_unlock(lock: &stack_depot_init_mutex); |
| 285 | |
| 286 | return ret; |
| 287 | } |
| 288 | EXPORT_SYMBOL_GPL(stack_depot_init); |
| 289 | |
| 290 | /* |
| 291 | * Initializes new stack pool, and updates the list of pools. |
| 292 | */ |
| 293 | static bool depot_init_pool(void **prealloc) |
| 294 | { |
| 295 | lockdep_assert_held(&pool_lock); |
| 296 | |
| 297 | if (unlikely(pools_num >= stack_max_pools)) { |
| 298 | /* Bail out if we reached the pool limit. */ |
| 299 | WARN_ON_ONCE(pools_num > stack_max_pools); /* should never happen */ |
| 300 | WARN_ON_ONCE(!new_pool); /* to avoid unnecessary pre-allocation */ |
| 301 | WARN_ONCE(1, "Stack depot reached limit capacity" ); |
| 302 | return false; |
| 303 | } |
| 304 | |
| 305 | if (!new_pool && *prealloc) { |
| 306 | /* We have preallocated memory, use it. */ |
| 307 | WRITE_ONCE(new_pool, *prealloc); |
| 308 | *prealloc = NULL; |
| 309 | } |
| 310 | |
| 311 | if (!new_pool) |
| 312 | return false; /* new_pool and *prealloc are NULL */ |
| 313 | |
| 314 | /* Save reference to the pool to be used by depot_fetch_stack(). */ |
| 315 | stack_pools[pools_num] = new_pool; |
| 316 | |
| 317 | /* |
| 318 | * Stack depot tries to keep an extra pool allocated even before it runs |
| 319 | * out of space in the currently used pool. |
| 320 | * |
| 321 | * To indicate that a new preallocation is needed new_pool is reset to |
| 322 | * NULL; do not reset to NULL if we have reached the maximum number of |
| 323 | * pools. |
| 324 | */ |
| 325 | if (pools_num < stack_max_pools) |
| 326 | WRITE_ONCE(new_pool, NULL); |
| 327 | else |
| 328 | WRITE_ONCE(new_pool, STACK_DEPOT_POISON); |
| 329 | |
| 330 | /* Pairs with concurrent READ_ONCE() in depot_fetch_stack(). */ |
| 331 | WRITE_ONCE(pools_num, pools_num + 1); |
| 332 | ASSERT_EXCLUSIVE_WRITER(pools_num); |
| 333 | |
| 334 | pool_offset = 0; |
| 335 | |
| 336 | return true; |
| 337 | } |
| 338 | |
| 339 | /* Keeps the preallocated memory to be used for a new stack depot pool. */ |
| 340 | static void depot_keep_new_pool(void **prealloc) |
| 341 | { |
| 342 | lockdep_assert_held(&pool_lock); |
| 343 | |
| 344 | /* |
| 345 | * If a new pool is already saved or the maximum number of |
| 346 | * pools is reached, do not use the preallocated memory. |
| 347 | */ |
| 348 | if (new_pool) |
| 349 | return; |
| 350 | |
| 351 | WRITE_ONCE(new_pool, *prealloc); |
| 352 | *prealloc = NULL; |
| 353 | } |
| 354 | |
| 355 | /* |
| 356 | * Try to initialize a new stack record from the current pool, a cached pool, or |
| 357 | * the current pre-allocation. |
| 358 | */ |
| 359 | static struct stack_record *depot_pop_free_pool(void **prealloc, size_t size) |
| 360 | { |
| 361 | struct stack_record *stack; |
| 362 | void *current_pool; |
| 363 | u32 pool_index; |
| 364 | |
| 365 | lockdep_assert_held(&pool_lock); |
| 366 | |
| 367 | if (pool_offset + size > DEPOT_POOL_SIZE) { |
| 368 | if (!depot_init_pool(prealloc)) |
| 369 | return NULL; |
| 370 | } |
| 371 | |
| 372 | if (WARN_ON_ONCE(pools_num < 1)) |
| 373 | return NULL; |
| 374 | pool_index = pools_num - 1; |
| 375 | current_pool = stack_pools[pool_index]; |
| 376 | if (WARN_ON_ONCE(!current_pool)) |
| 377 | return NULL; |
| 378 | |
| 379 | stack = current_pool + pool_offset; |
| 380 | |
| 381 | /* Pre-initialize handle once. */ |
| 382 | stack->handle.pool_index_plus_1 = pool_index + 1; |
| 383 | stack->handle.offset = pool_offset >> DEPOT_STACK_ALIGN; |
| 384 | stack->handle.extra = 0; |
| 385 | INIT_LIST_HEAD(list: &stack->hash_list); |
| 386 | |
| 387 | pool_offset += size; |
| 388 | |
| 389 | return stack; |
| 390 | } |
| 391 | |
| 392 | /* Try to find next free usable entry from the freelist. */ |
| 393 | static struct stack_record *depot_pop_free(void) |
| 394 | { |
| 395 | struct stack_record *stack; |
| 396 | |
| 397 | lockdep_assert_held(&pool_lock); |
| 398 | |
| 399 | if (list_empty(head: &free_stacks)) |
| 400 | return NULL; |
| 401 | |
| 402 | /* |
| 403 | * We maintain the invariant that the elements in front are least |
| 404 | * recently used, and are therefore more likely to be associated with an |
| 405 | * RCU grace period in the past. Consequently it is sufficient to only |
| 406 | * check the first entry. |
| 407 | */ |
| 408 | stack = list_first_entry(&free_stacks, struct stack_record, free_list); |
| 409 | if (!poll_state_synchronize_rcu(oldstate: stack->rcu_state)) |
| 410 | return NULL; |
| 411 | |
| 412 | list_del(entry: &stack->free_list); |
| 413 | counters[DEPOT_COUNTER_FREELIST_SIZE]--; |
| 414 | |
| 415 | return stack; |
| 416 | } |
| 417 | |
| 418 | static inline size_t depot_stack_record_size(struct stack_record *s, unsigned int nr_entries) |
| 419 | { |
| 420 | const size_t used = flex_array_size(s, entries, nr_entries); |
| 421 | const size_t unused = sizeof(s->entries) - used; |
| 422 | |
| 423 | WARN_ON_ONCE(sizeof(s->entries) < used); |
| 424 | |
| 425 | return ALIGN(sizeof(struct stack_record) - unused, 1 << DEPOT_STACK_ALIGN); |
| 426 | } |
| 427 | |
| 428 | /* Allocates a new stack in a stack depot pool. */ |
| 429 | static struct stack_record * |
| 430 | depot_alloc_stack(unsigned long *entries, unsigned int nr_entries, u32 hash, depot_flags_t flags, void **prealloc) |
| 431 | { |
| 432 | struct stack_record *stack = NULL; |
| 433 | size_t record_size; |
| 434 | |
| 435 | lockdep_assert_held(&pool_lock); |
| 436 | |
| 437 | /* This should already be checked by public API entry points. */ |
| 438 | if (WARN_ON_ONCE(!nr_entries)) |
| 439 | return NULL; |
| 440 | |
| 441 | /* Limit number of saved frames to CONFIG_STACKDEPOT_MAX_FRAMES. */ |
| 442 | if (nr_entries > CONFIG_STACKDEPOT_MAX_FRAMES) |
| 443 | nr_entries = CONFIG_STACKDEPOT_MAX_FRAMES; |
| 444 | |
| 445 | if (flags & STACK_DEPOT_FLAG_GET) { |
| 446 | /* |
| 447 | * Evictable entries have to allocate the max. size so they may |
| 448 | * safely be re-used by differently sized allocations. |
| 449 | */ |
| 450 | record_size = depot_stack_record_size(s: stack, CONFIG_STACKDEPOT_MAX_FRAMES); |
| 451 | stack = depot_pop_free(); |
| 452 | } else { |
| 453 | record_size = depot_stack_record_size(s: stack, nr_entries); |
| 454 | } |
| 455 | |
| 456 | if (!stack) { |
| 457 | stack = depot_pop_free_pool(prealloc, size: record_size); |
| 458 | if (!stack) |
| 459 | return NULL; |
| 460 | } |
| 461 | |
| 462 | /* Save the stack trace. */ |
| 463 | stack->hash = hash; |
| 464 | stack->size = nr_entries; |
| 465 | /* stack->handle is already filled in by depot_pop_free_pool(). */ |
| 466 | memcpy(to: stack->entries, from: entries, flex_array_size(stack, entries, nr_entries)); |
| 467 | |
| 468 | if (flags & STACK_DEPOT_FLAG_GET) { |
| 469 | refcount_set(r: &stack->count, n: 1); |
| 470 | counters[DEPOT_COUNTER_REFD_ALLOCS]++; |
| 471 | counters[DEPOT_COUNTER_REFD_INUSE]++; |
| 472 | } else { |
| 473 | /* Warn on attempts to switch to refcounting this entry. */ |
| 474 | refcount_set(r: &stack->count, REFCOUNT_SATURATED); |
| 475 | counters[DEPOT_COUNTER_PERSIST_COUNT]++; |
| 476 | counters[DEPOT_COUNTER_PERSIST_BYTES] += record_size; |
| 477 | } |
| 478 | |
| 479 | /* |
| 480 | * Let KMSAN know the stored stack record is initialized. This shall |
| 481 | * prevent false positive reports if instrumented code accesses it. |
| 482 | */ |
| 483 | kmsan_unpoison_memory(address: stack, size: record_size); |
| 484 | |
| 485 | return stack; |
| 486 | } |
| 487 | |
| 488 | static struct stack_record *depot_fetch_stack(depot_stack_handle_t handle) |
| 489 | { |
| 490 | const int pools_num_cached = READ_ONCE(pools_num); |
| 491 | union handle_parts parts = { .handle = handle }; |
| 492 | void *pool; |
| 493 | u32 pool_index = parts.pool_index_plus_1 - 1; |
| 494 | size_t offset = parts.offset << DEPOT_STACK_ALIGN; |
| 495 | struct stack_record *stack; |
| 496 | |
| 497 | lockdep_assert_not_held(&pool_lock); |
| 498 | |
| 499 | if (pool_index >= pools_num_cached) { |
| 500 | WARN(1, "pool index %d out of bounds (%d) for stack id %08x\n" , |
| 501 | pool_index, pools_num_cached, handle); |
| 502 | return NULL; |
| 503 | } |
| 504 | |
| 505 | pool = stack_pools[pool_index]; |
| 506 | if (WARN_ON(!pool)) |
| 507 | return NULL; |
| 508 | |
| 509 | stack = pool + offset; |
| 510 | if (WARN_ON(!refcount_read(&stack->count))) |
| 511 | return NULL; |
| 512 | |
| 513 | return stack; |
| 514 | } |
| 515 | |
| 516 | /* Links stack into the freelist. */ |
| 517 | static void depot_free_stack(struct stack_record *stack) |
| 518 | { |
| 519 | unsigned long flags; |
| 520 | |
| 521 | lockdep_assert_not_held(&pool_lock); |
| 522 | |
| 523 | raw_spin_lock_irqsave(&pool_lock, flags); |
| 524 | printk_deferred_enter(); |
| 525 | |
| 526 | /* |
| 527 | * Remove the entry from the hash list. Concurrent list traversal may |
| 528 | * still observe the entry, but since the refcount is zero, this entry |
| 529 | * will no longer be considered as valid. |
| 530 | */ |
| 531 | list_del_rcu(entry: &stack->hash_list); |
| 532 | |
| 533 | /* |
| 534 | * Due to being used from constrained contexts such as the allocators, |
| 535 | * NMI, or even RCU itself, stack depot cannot rely on primitives that |
| 536 | * would sleep (such as synchronize_rcu()) or recursively call into |
| 537 | * stack depot again (such as call_rcu()). |
| 538 | * |
| 539 | * Instead, get an RCU cookie, so that we can ensure this entry isn't |
| 540 | * moved onto another list until the next grace period, and concurrent |
| 541 | * RCU list traversal remains safe. |
| 542 | */ |
| 543 | stack->rcu_state = get_state_synchronize_rcu(); |
| 544 | |
| 545 | /* |
| 546 | * Add the entry to the freelist tail, so that older entries are |
| 547 | * considered first - their RCU cookie is more likely to no longer be |
| 548 | * associated with the current grace period. |
| 549 | */ |
| 550 | list_add_tail(new: &stack->free_list, head: &free_stacks); |
| 551 | |
| 552 | counters[DEPOT_COUNTER_FREELIST_SIZE]++; |
| 553 | counters[DEPOT_COUNTER_REFD_FREES]++; |
| 554 | counters[DEPOT_COUNTER_REFD_INUSE]--; |
| 555 | |
| 556 | printk_deferred_exit(); |
| 557 | raw_spin_unlock_irqrestore(&pool_lock, flags); |
| 558 | } |
| 559 | |
| 560 | /* Calculates the hash for a stack. */ |
| 561 | static inline u32 hash_stack(unsigned long *entries, unsigned int size) |
| 562 | { |
| 563 | return jhash2(k: (u32 *)entries, |
| 564 | array_size(size, sizeof(*entries)) / sizeof(u32), |
| 565 | STACK_HASH_SEED); |
| 566 | } |
| 567 | |
| 568 | /* |
| 569 | * Non-instrumented version of memcmp(). |
| 570 | * Does not check the lexicographical order, only the equality. |
| 571 | */ |
| 572 | static inline |
| 573 | int stackdepot_memcmp(const unsigned long *u1, const unsigned long *u2, |
| 574 | unsigned int n) |
| 575 | { |
| 576 | for ( ; n-- ; u1++, u2++) { |
| 577 | if (*u1 != *u2) |
| 578 | return 1; |
| 579 | } |
| 580 | return 0; |
| 581 | } |
| 582 | |
| 583 | /* Finds a stack in a bucket of the hash table. */ |
| 584 | static inline struct stack_record *find_stack(struct list_head *bucket, |
| 585 | unsigned long *entries, int size, |
| 586 | u32 hash, depot_flags_t flags) |
| 587 | { |
| 588 | struct stack_record *stack, *ret = NULL; |
| 589 | |
| 590 | /* |
| 591 | * Stack depot may be used from instrumentation that instruments RCU or |
| 592 | * tracing itself; use variant that does not call into RCU and cannot be |
| 593 | * traced. |
| 594 | * |
| 595 | * Note: Such use cases must take care when using refcounting to evict |
| 596 | * unused entries, because the stack record free-then-reuse code paths |
| 597 | * do call into RCU. |
| 598 | */ |
| 599 | rcu_read_lock_sched_notrace(); |
| 600 | |
| 601 | list_for_each_entry_rcu(stack, bucket, hash_list) { |
| 602 | if (stack->hash != hash || stack->size != size) |
| 603 | continue; |
| 604 | |
| 605 | /* |
| 606 | * This may race with depot_free_stack() accessing the freelist |
| 607 | * management state unioned with @entries. The refcount is zero |
| 608 | * in that case and the below refcount_inc_not_zero() will fail. |
| 609 | */ |
| 610 | if (data_race(stackdepot_memcmp(entries, stack->entries, size))) |
| 611 | continue; |
| 612 | |
| 613 | /* |
| 614 | * Try to increment refcount. If this succeeds, the stack record |
| 615 | * is valid and has not yet been freed. |
| 616 | * |
| 617 | * If STACK_DEPOT_FLAG_GET is not used, it is undefined behavior |
| 618 | * to then call stack_depot_put() later, and we can assume that |
| 619 | * a stack record is never placed back on the freelist. |
| 620 | */ |
| 621 | if ((flags & STACK_DEPOT_FLAG_GET) && !refcount_inc_not_zero(r: &stack->count)) |
| 622 | continue; |
| 623 | |
| 624 | ret = stack; |
| 625 | break; |
| 626 | } |
| 627 | |
| 628 | rcu_read_unlock_sched_notrace(); |
| 629 | |
| 630 | return ret; |
| 631 | } |
| 632 | |
| 633 | depot_stack_handle_t stack_depot_save_flags(unsigned long *entries, |
| 634 | unsigned int nr_entries, |
| 635 | gfp_t alloc_flags, |
| 636 | depot_flags_t depot_flags) |
| 637 | { |
| 638 | struct list_head *bucket; |
| 639 | struct stack_record *found = NULL; |
| 640 | depot_stack_handle_t handle = 0; |
| 641 | struct page *page = NULL; |
| 642 | void *prealloc = NULL; |
| 643 | bool allow_spin = gfpflags_allow_spinning(gfp_flags: alloc_flags); |
| 644 | bool can_alloc = (depot_flags & STACK_DEPOT_FLAG_CAN_ALLOC) && allow_spin; |
| 645 | unsigned long flags; |
| 646 | u32 hash; |
| 647 | |
| 648 | if (WARN_ON(depot_flags & ~STACK_DEPOT_FLAGS_MASK)) |
| 649 | return 0; |
| 650 | |
| 651 | /* |
| 652 | * If this stack trace is from an interrupt, including anything before |
| 653 | * interrupt entry usually leads to unbounded stack depot growth. |
| 654 | * |
| 655 | * Since use of filter_irq_stacks() is a requirement to ensure stack |
| 656 | * depot can efficiently deduplicate interrupt stacks, always |
| 657 | * filter_irq_stacks() to simplify all callers' use of stack depot. |
| 658 | */ |
| 659 | nr_entries = filter_irq_stacks(entries, nr_entries); |
| 660 | |
| 661 | if (unlikely(nr_entries == 0) || stack_depot_disabled) |
| 662 | return 0; |
| 663 | |
| 664 | hash = hash_stack(entries, size: nr_entries); |
| 665 | bucket = &stack_table[hash & stack_hash_mask]; |
| 666 | |
| 667 | /* Fast path: look the stack trace up without locking. */ |
| 668 | found = find_stack(bucket, entries, size: nr_entries, hash, flags: depot_flags); |
| 669 | if (found) |
| 670 | goto exit; |
| 671 | |
| 672 | /* |
| 673 | * Allocate memory for a new pool if required now: |
| 674 | * we won't be able to do that under the lock. |
| 675 | */ |
| 676 | if (unlikely(can_alloc && !READ_ONCE(new_pool))) { |
| 677 | page = alloc_pages(gfp_nested_mask(alloc_flags), |
| 678 | DEPOT_POOL_ORDER); |
| 679 | if (page) |
| 680 | prealloc = page_address(page); |
| 681 | } |
| 682 | |
| 683 | if (in_nmi() || !allow_spin) { |
| 684 | /* We can never allocate in NMI context. */ |
| 685 | WARN_ON_ONCE(can_alloc); |
| 686 | /* Best effort; bail if we fail to take the lock. */ |
| 687 | if (!raw_spin_trylock_irqsave(&pool_lock, flags)) |
| 688 | goto exit; |
| 689 | } else { |
| 690 | raw_spin_lock_irqsave(&pool_lock, flags); |
| 691 | } |
| 692 | printk_deferred_enter(); |
| 693 | |
| 694 | /* Try to find again, to avoid concurrently inserting duplicates. */ |
| 695 | found = find_stack(bucket, entries, size: nr_entries, hash, flags: depot_flags); |
| 696 | if (!found) { |
| 697 | struct stack_record *new = |
| 698 | depot_alloc_stack(entries, nr_entries, hash, flags: depot_flags, prealloc: &prealloc); |
| 699 | |
| 700 | if (new) { |
| 701 | /* |
| 702 | * This releases the stack record into the bucket and |
| 703 | * makes it visible to readers in find_stack(). |
| 704 | */ |
| 705 | list_add_rcu(new: &new->hash_list, head: bucket); |
| 706 | found = new; |
| 707 | } |
| 708 | } |
| 709 | |
| 710 | if (prealloc) { |
| 711 | /* |
| 712 | * Either stack depot already contains this stack trace, or |
| 713 | * depot_alloc_stack() did not consume the preallocated memory. |
| 714 | * Try to keep the preallocated memory for future. |
| 715 | */ |
| 716 | depot_keep_new_pool(prealloc: &prealloc); |
| 717 | } |
| 718 | |
| 719 | printk_deferred_exit(); |
| 720 | raw_spin_unlock_irqrestore(&pool_lock, flags); |
| 721 | exit: |
| 722 | if (prealloc) { |
| 723 | /* Stack depot didn't use this memory, free it. */ |
| 724 | if (!allow_spin) |
| 725 | free_pages_nolock(virt_to_page(prealloc), DEPOT_POOL_ORDER); |
| 726 | else |
| 727 | free_pages(addr: (unsigned long)prealloc, DEPOT_POOL_ORDER); |
| 728 | } |
| 729 | if (found) |
| 730 | handle = found->handle.handle; |
| 731 | return handle; |
| 732 | } |
| 733 | EXPORT_SYMBOL_GPL(stack_depot_save_flags); |
| 734 | |
| 735 | depot_stack_handle_t stack_depot_save(unsigned long *entries, |
| 736 | unsigned int nr_entries, |
| 737 | gfp_t alloc_flags) |
| 738 | { |
| 739 | return stack_depot_save_flags(entries, nr_entries, alloc_flags, |
| 740 | STACK_DEPOT_FLAG_CAN_ALLOC); |
| 741 | } |
| 742 | EXPORT_SYMBOL_GPL(stack_depot_save); |
| 743 | |
| 744 | struct stack_record *__stack_depot_get_stack_record(depot_stack_handle_t handle) |
| 745 | { |
| 746 | if (!handle) |
| 747 | return NULL; |
| 748 | |
| 749 | return depot_fetch_stack(handle); |
| 750 | } |
| 751 | |
| 752 | unsigned int stack_depot_fetch(depot_stack_handle_t handle, |
| 753 | unsigned long **entries) |
| 754 | { |
| 755 | struct stack_record *stack; |
| 756 | |
| 757 | *entries = NULL; |
| 758 | /* |
| 759 | * Let KMSAN know *entries is initialized. This shall prevent false |
| 760 | * positive reports if instrumented code accesses it. |
| 761 | */ |
| 762 | kmsan_unpoison_memory(address: entries, size: sizeof(*entries)); |
| 763 | |
| 764 | if (!handle || stack_depot_disabled) |
| 765 | return 0; |
| 766 | |
| 767 | stack = depot_fetch_stack(handle); |
| 768 | /* |
| 769 | * Should never be NULL, otherwise this is a use-after-put (or just a |
| 770 | * corrupt handle). |
| 771 | */ |
| 772 | if (WARN(!stack, "corrupt handle or use after stack_depot_put()" )) |
| 773 | return 0; |
| 774 | |
| 775 | *entries = stack->entries; |
| 776 | return stack->size; |
| 777 | } |
| 778 | EXPORT_SYMBOL_GPL(stack_depot_fetch); |
| 779 | |
| 780 | void stack_depot_put(depot_stack_handle_t handle) |
| 781 | { |
| 782 | struct stack_record *stack; |
| 783 | |
| 784 | if (!handle || stack_depot_disabled) |
| 785 | return; |
| 786 | |
| 787 | stack = depot_fetch_stack(handle); |
| 788 | /* |
| 789 | * Should always be able to find the stack record, otherwise this is an |
| 790 | * unbalanced put attempt (or corrupt handle). |
| 791 | */ |
| 792 | if (WARN(!stack, "corrupt handle or unbalanced stack_depot_put()" )) |
| 793 | return; |
| 794 | |
| 795 | if (refcount_dec_and_test(r: &stack->count)) |
| 796 | depot_free_stack(stack); |
| 797 | } |
| 798 | EXPORT_SYMBOL_GPL(stack_depot_put); |
| 799 | |
| 800 | void stack_depot_print(depot_stack_handle_t stack) |
| 801 | { |
| 802 | unsigned long *entries; |
| 803 | unsigned int nr_entries; |
| 804 | |
| 805 | nr_entries = stack_depot_fetch(stack, &entries); |
| 806 | if (nr_entries > 0) |
| 807 | stack_trace_print(trace: entries, nr_entries, spaces: 0); |
| 808 | } |
| 809 | EXPORT_SYMBOL_GPL(stack_depot_print); |
| 810 | |
| 811 | int stack_depot_snprint(depot_stack_handle_t handle, char *buf, size_t size, |
| 812 | int spaces) |
| 813 | { |
| 814 | unsigned long *entries; |
| 815 | unsigned int nr_entries; |
| 816 | |
| 817 | nr_entries = stack_depot_fetch(handle, &entries); |
| 818 | return nr_entries ? stack_trace_snprint(buf, size, entries, nr_entries, |
| 819 | spaces) : 0; |
| 820 | } |
| 821 | EXPORT_SYMBOL_GPL(stack_depot_snprint); |
| 822 | |
| 823 | depot_stack_handle_t __must_check ( |
| 824 | depot_stack_handle_t handle, unsigned int ) |
| 825 | { |
| 826 | union handle_parts parts = { .handle = handle }; |
| 827 | |
| 828 | /* Don't set extra bits on empty handles. */ |
| 829 | if (!handle) |
| 830 | return 0; |
| 831 | |
| 832 | parts.extra = extra_bits; |
| 833 | return parts.handle; |
| 834 | } |
| 835 | EXPORT_SYMBOL(stack_depot_set_extra_bits); |
| 836 | |
| 837 | unsigned int (depot_stack_handle_t handle) |
| 838 | { |
| 839 | union handle_parts parts = { .handle = handle }; |
| 840 | |
| 841 | return parts.extra; |
| 842 | } |
| 843 | EXPORT_SYMBOL(stack_depot_get_extra_bits); |
| 844 | |
| 845 | static int stats_show(struct seq_file *seq, void *v) |
| 846 | { |
| 847 | /* |
| 848 | * data race ok: These are just statistics counters, and approximate |
| 849 | * statistics are ok for debugging. |
| 850 | */ |
| 851 | seq_printf(m: seq, fmt: "pools: %d\n" , data_race(pools_num)); |
| 852 | for (int i = 0; i < DEPOT_COUNTER_COUNT; i++) |
| 853 | seq_printf(m: seq, fmt: "%s: %ld\n" , counter_names[i], data_race(counters[i])); |
| 854 | |
| 855 | return 0; |
| 856 | } |
| 857 | DEFINE_SHOW_ATTRIBUTE(stats); |
| 858 | |
| 859 | static int depot_debugfs_init(void) |
| 860 | { |
| 861 | struct dentry *dir; |
| 862 | |
| 863 | if (stack_depot_disabled) |
| 864 | return 0; |
| 865 | |
| 866 | dir = debugfs_create_dir(name: "stackdepot" , NULL); |
| 867 | debugfs_create_file("stats" , 0444, dir, NULL, &stats_fops); |
| 868 | return 0; |
| 869 | } |
| 870 | late_initcall(depot_debugfs_init); |
| 871 | |