| 1 | /* |
| 2 | * simple memory allocator, backed by mmap() so that it hands out memory |
| 3 | * that can be shared across processes and threads |
| 4 | */ |
| 5 | #include <sys/mman.h> |
| 6 | #include <stdio.h> |
| 7 | #include <stdlib.h> |
| 8 | #include <assert.h> |
| 9 | #include <string.h> |
| 10 | #include <unistd.h> |
| 11 | #include <inttypes.h> |
| 12 | #include <sys/types.h> |
| 13 | #include <limits.h> |
| 14 | #include <fcntl.h> |
| 15 | |
| 16 | #include "mutex.h" |
| 17 | #include "arch/arch.h" |
| 18 | #include "os/os.h" |
| 19 | #include "smalloc.h" |
| 20 | #include "log.h" |
| 21 | |
| 22 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
| 23 | |
| 24 | #define SMALLOC_BPB 32 /* block size, bytes-per-bit in bitmap */ |
| 25 | #define SMALLOC_BPI (sizeof(unsigned int) * 8) |
| 26 | #define SMALLOC_BPL (SMALLOC_BPB * SMALLOC_BPI) |
| 27 | |
| 28 | #define INITIAL_SIZE 16*1024*1024 /* new pool size */ |
| 29 | #define INITIAL_POOLS 8 /* maximum number of pools to setup */ |
| 30 | |
| 31 | #define MAX_POOLS 16 |
| 32 | |
| 33 | #define SMALLOC_PRE_RED 0xdeadbeefU |
| 34 | #define SMALLOC_POST_RED 0x5aa55aa5U |
| 35 | |
| 36 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
| 37 | #ifdef SMALLOC_REDZONE |
| 38 | static const int int_mask = sizeof(int) - 1; |
| 39 | #endif |
| 40 | |
| 41 | struct pool { |
| 42 | struct fio_mutex *lock; /* protects this pool */ |
| 43 | void *map; /* map of blocks */ |
| 44 | unsigned int *bitmap; /* blocks free/busy map */ |
| 45 | size_t free_blocks; /* free blocks */ |
| 46 | size_t nr_blocks; /* total blocks */ |
| 47 | size_t next_non_full; |
| 48 | size_t mmap_size; |
| 49 | }; |
| 50 | |
| 51 | struct block_hdr { |
| 52 | size_t size; |
| 53 | #ifdef SMALLOC_REDZONE |
| 54 | unsigned int prered; |
| 55 | #endif |
| 56 | }; |
| 57 | |
| 58 | static struct pool mp[MAX_POOLS]; |
| 59 | static unsigned int nr_pools; |
| 60 | static unsigned int last_pool; |
| 61 | |
| 62 | static inline int ptr_valid(struct pool *pool, void *ptr) |
| 63 | { |
| 64 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
| 65 | |
| 66 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); |
| 67 | } |
| 68 | |
| 69 | static inline size_t size_to_blocks(size_t size) |
| 70 | { |
| 71 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; |
| 72 | } |
| 73 | |
| 74 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
| 75 | unsigned int idx, size_t nr_blocks, |
| 76 | int (*func)(unsigned int *map, unsigned int mask)) |
| 77 | { |
| 78 | |
| 79 | while (nr_blocks) { |
| 80 | unsigned int this_blocks, mask; |
| 81 | unsigned int *map; |
| 82 | |
| 83 | if (pool_idx >= pool->nr_blocks) |
| 84 | return 0; |
| 85 | |
| 86 | map = &pool->bitmap[pool_idx]; |
| 87 | |
| 88 | this_blocks = nr_blocks; |
| 89 | if (this_blocks + idx > SMALLOC_BPI) { |
| 90 | this_blocks = SMALLOC_BPI - idx; |
| 91 | idx = SMALLOC_BPI - this_blocks; |
| 92 | } |
| 93 | |
| 94 | if (this_blocks == SMALLOC_BPI) |
| 95 | mask = -1U; |
| 96 | else |
| 97 | mask = ((1U << this_blocks) - 1) << idx; |
| 98 | |
| 99 | if (!func(map, mask)) |
| 100 | return 0; |
| 101 | |
| 102 | nr_blocks -= this_blocks; |
| 103 | idx = 0; |
| 104 | pool_idx++; |
| 105 | } |
| 106 | |
| 107 | return 1; |
| 108 | } |
| 109 | |
| 110 | static int mask_cmp(unsigned int *map, unsigned int mask) |
| 111 | { |
| 112 | return !(*map & mask); |
| 113 | } |
| 114 | |
| 115 | static int mask_clear(unsigned int *map, unsigned int mask) |
| 116 | { |
| 117 | assert((*map & mask) == mask); |
| 118 | *map &= ~mask; |
| 119 | return 1; |
| 120 | } |
| 121 | |
| 122 | static int mask_set(unsigned int *map, unsigned int mask) |
| 123 | { |
| 124 | assert(!(*map & mask)); |
| 125 | *map |= mask; |
| 126 | return 1; |
| 127 | } |
| 128 | |
| 129 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
| 130 | unsigned int idx, size_t nr_blocks) |
| 131 | { |
| 132 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
| 133 | } |
| 134 | |
| 135 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
| 136 | unsigned int idx, size_t nr_blocks) |
| 137 | { |
| 138 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
| 139 | } |
| 140 | |
| 141 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
| 142 | unsigned int idx, size_t nr_blocks) |
| 143 | { |
| 144 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
| 145 | } |
| 146 | |
| 147 | static int find_next_zero(int word, int start) |
| 148 | { |
| 149 | assert(word != -1U); |
| 150 | word >>= start; |
| 151 | return ffz(word) + start; |
| 152 | } |
| 153 | |
| 154 | static bool add_pool(struct pool *pool, unsigned int alloc_size) |
| 155 | { |
| 156 | int bitmap_blocks; |
| 157 | int mmap_flags; |
| 158 | void *ptr; |
| 159 | |
| 160 | if (nr_pools == MAX_POOLS) |
| 161 | return false; |
| 162 | |
| 163 | #ifdef SMALLOC_REDZONE |
| 164 | alloc_size += sizeof(unsigned int); |
| 165 | #endif |
| 166 | alloc_size += sizeof(struct block_hdr); |
| 167 | if (alloc_size < INITIAL_SIZE) |
| 168 | alloc_size = INITIAL_SIZE; |
| 169 | |
| 170 | /* round up to nearest full number of blocks */ |
| 171 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); |
| 172 | bitmap_blocks = alloc_size / SMALLOC_BPL; |
| 173 | alloc_size += bitmap_blocks * sizeof(unsigned int); |
| 174 | pool->mmap_size = alloc_size; |
| 175 | |
| 176 | pool->nr_blocks = bitmap_blocks; |
| 177 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; |
| 178 | |
| 179 | mmap_flags = OS_MAP_ANON; |
| 180 | #ifdef CONFIG_ESX |
| 181 | mmap_flags |= MAP_PRIVATE; |
| 182 | #else |
| 183 | mmap_flags |= MAP_SHARED; |
| 184 | #endif |
| 185 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, mmap_flags, -1, 0); |
| 186 | |
| 187 | if (ptr == MAP_FAILED) |
| 188 | goto out_fail; |
| 189 | |
| 190 | pool->map = ptr; |
| 191 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); |
| 192 | memset(pool->bitmap, 0, bitmap_blocks * sizeof(unsigned int)); |
| 193 | |
| 194 | pool->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED); |
| 195 | if (!pool->lock) |
| 196 | goto out_fail; |
| 197 | |
| 198 | nr_pools++; |
| 199 | return true; |
| 200 | out_fail: |
| 201 | log_err("smalloc: failed adding pool\n"); |
| 202 | if (pool->map) |
| 203 | munmap(pool->map, pool->mmap_size); |
| 204 | return false; |
| 205 | } |
| 206 | |
| 207 | void sinit(void) |
| 208 | { |
| 209 | bool ret; |
| 210 | int i; |
| 211 | |
| 212 | for (i = 0; i < INITIAL_POOLS; i++) { |
| 213 | ret = add_pool(&mp[nr_pools], smalloc_pool_size); |
| 214 | if (!ret) |
| 215 | break; |
| 216 | } |
| 217 | |
| 218 | /* |
| 219 | * If we added at least one pool, we should be OK for most |
| 220 | * cases. |
| 221 | */ |
| 222 | assert(i); |
| 223 | } |
| 224 | |
| 225 | static void cleanup_pool(struct pool *pool) |
| 226 | { |
| 227 | /* |
| 228 | * This will also remove the temporary file we used as a backing |
| 229 | * store, it was already unlinked |
| 230 | */ |
| 231 | munmap(pool->map, pool->mmap_size); |
| 232 | |
| 233 | if (pool->lock) |
| 234 | fio_mutex_remove(pool->lock); |
| 235 | } |
| 236 | |
| 237 | void scleanup(void) |
| 238 | { |
| 239 | unsigned int i; |
| 240 | |
| 241 | for (i = 0; i < nr_pools; i++) |
| 242 | cleanup_pool(&mp[i]); |
| 243 | } |
| 244 | |
| 245 | #ifdef SMALLOC_REDZONE |
| 246 | static void *postred_ptr(struct block_hdr *hdr) |
| 247 | { |
| 248 | uintptr_t ptr; |
| 249 | |
| 250 | ptr = (uintptr_t) hdr + hdr->size - sizeof(unsigned int); |
| 251 | ptr = (ptr + int_mask) & ~int_mask; |
| 252 | |
| 253 | return (void *) ptr; |
| 254 | } |
| 255 | |
| 256 | static void fill_redzone(struct block_hdr *hdr) |
| 257 | { |
| 258 | unsigned int *postred = postred_ptr(hdr); |
| 259 | |
| 260 | hdr->prered = SMALLOC_PRE_RED; |
| 261 | *postred = SMALLOC_POST_RED; |
| 262 | } |
| 263 | |
| 264 | static void sfree_check_redzone(struct block_hdr *hdr) |
| 265 | { |
| 266 | unsigned int *postred = postred_ptr(hdr); |
| 267 | |
| 268 | if (hdr->prered != SMALLOC_PRE_RED) { |
| 269 | log_err("smalloc pre redzone destroyed!\n" |
| 270 | " ptr=%p, prered=%x, expected %x\n", |
| 271 | hdr, hdr->prered, SMALLOC_PRE_RED); |
| 272 | assert(0); |
| 273 | } |
| 274 | if (*postred != SMALLOC_POST_RED) { |
| 275 | log_err("smalloc post redzone destroyed!\n" |
| 276 | " ptr=%p, postred=%x, expected %x\n", |
| 277 | hdr, *postred, SMALLOC_POST_RED); |
| 278 | assert(0); |
| 279 | } |
| 280 | } |
| 281 | #else |
| 282 | static void fill_redzone(struct block_hdr *hdr) |
| 283 | { |
| 284 | } |
| 285 | |
| 286 | static void sfree_check_redzone(struct block_hdr *hdr) |
| 287 | { |
| 288 | } |
| 289 | #endif |
| 290 | |
| 291 | static void sfree_pool(struct pool *pool, void *ptr) |
| 292 | { |
| 293 | struct block_hdr *hdr; |
| 294 | unsigned int i, idx; |
| 295 | unsigned long offset; |
| 296 | |
| 297 | if (!ptr) |
| 298 | return; |
| 299 | |
| 300 | ptr -= sizeof(*hdr); |
| 301 | hdr = ptr; |
| 302 | |
| 303 | assert(ptr_valid(pool, ptr)); |
| 304 | |
| 305 | sfree_check_redzone(hdr); |
| 306 | |
| 307 | offset = ptr - pool->map; |
| 308 | i = offset / SMALLOC_BPL; |
| 309 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; |
| 310 | |
| 311 | fio_mutex_down(pool->lock); |
| 312 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
| 313 | if (i < pool->next_non_full) |
| 314 | pool->next_non_full = i; |
| 315 | pool->free_blocks += size_to_blocks(hdr->size); |
| 316 | fio_mutex_up(pool->lock); |
| 317 | } |
| 318 | |
| 319 | void sfree(void *ptr) |
| 320 | { |
| 321 | struct pool *pool = NULL; |
| 322 | unsigned int i; |
| 323 | |
| 324 | if (!ptr) |
| 325 | return; |
| 326 | |
| 327 | for (i = 0; i < nr_pools; i++) { |
| 328 | if (ptr_valid(&mp[i], ptr)) { |
| 329 | pool = &mp[i]; |
| 330 | break; |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | if (pool) { |
| 335 | sfree_pool(pool, ptr); |
| 336 | return; |
| 337 | } |
| 338 | |
| 339 | log_err("smalloc: ptr %p not from smalloc pool\n", ptr); |
| 340 | } |
| 341 | |
| 342 | static void *__smalloc_pool(struct pool *pool, size_t size) |
| 343 | { |
| 344 | size_t nr_blocks; |
| 345 | unsigned int i; |
| 346 | unsigned int offset; |
| 347 | unsigned int last_idx; |
| 348 | void *ret = NULL; |
| 349 | |
| 350 | fio_mutex_down(pool->lock); |
| 351 | |
| 352 | nr_blocks = size_to_blocks(size); |
| 353 | if (nr_blocks > pool->free_blocks) |
| 354 | goto fail; |
| 355 | |
| 356 | i = pool->next_non_full; |
| 357 | last_idx = 0; |
| 358 | offset = -1U; |
| 359 | while (i < pool->nr_blocks) { |
| 360 | unsigned int idx; |
| 361 | |
| 362 | if (pool->bitmap[i] == -1U) { |
| 363 | i++; |
| 364 | pool->next_non_full = i; |
| 365 | last_idx = 0; |
| 366 | continue; |
| 367 | } |
| 368 | |
| 369 | idx = find_next_zero(pool->bitmap[i], last_idx); |
| 370 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
| 371 | idx += nr_blocks; |
| 372 | if (idx < SMALLOC_BPI) |
| 373 | last_idx = idx; |
| 374 | else { |
| 375 | last_idx = 0; |
| 376 | while (idx >= SMALLOC_BPI) { |
| 377 | i++; |
| 378 | idx -= SMALLOC_BPI; |
| 379 | } |
| 380 | } |
| 381 | continue; |
| 382 | } |
| 383 | set_blocks(pool, i, idx, nr_blocks); |
| 384 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
| 385 | break; |
| 386 | } |
| 387 | |
| 388 | if (i < pool->nr_blocks) { |
| 389 | pool->free_blocks -= nr_blocks; |
| 390 | ret = pool->map + offset; |
| 391 | } |
| 392 | fail: |
| 393 | fio_mutex_up(pool->lock); |
| 394 | return ret; |
| 395 | } |
| 396 | |
| 397 | static void *smalloc_pool(struct pool *pool, size_t size) |
| 398 | { |
| 399 | size_t alloc_size = size + sizeof(struct block_hdr); |
| 400 | void *ptr; |
| 401 | |
| 402 | /* |
| 403 | * Round to int alignment, so that the postred pointer will |
| 404 | * be naturally aligned as well. |
| 405 | */ |
| 406 | #ifdef SMALLOC_REDZONE |
| 407 | alloc_size += sizeof(unsigned int); |
| 408 | alloc_size = (alloc_size + int_mask) & ~int_mask; |
| 409 | #endif |
| 410 | |
| 411 | ptr = __smalloc_pool(pool, alloc_size); |
| 412 | if (ptr) { |
| 413 | struct block_hdr *hdr = ptr; |
| 414 | |
| 415 | hdr->size = alloc_size; |
| 416 | fill_redzone(hdr); |
| 417 | |
| 418 | ptr += sizeof(*hdr); |
| 419 | memset(ptr, 0, size); |
| 420 | } |
| 421 | |
| 422 | return ptr; |
| 423 | } |
| 424 | |
| 425 | void *smalloc(size_t size) |
| 426 | { |
| 427 | unsigned int i, end_pool; |
| 428 | |
| 429 | if (size != (unsigned int) size) |
| 430 | return NULL; |
| 431 | |
| 432 | i = last_pool; |
| 433 | end_pool = nr_pools; |
| 434 | |
| 435 | do { |
| 436 | for (; i < end_pool; i++) { |
| 437 | void *ptr = smalloc_pool(&mp[i], size); |
| 438 | |
| 439 | if (ptr) { |
| 440 | last_pool = i; |
| 441 | return ptr; |
| 442 | } |
| 443 | } |
| 444 | if (last_pool) { |
| 445 | end_pool = last_pool; |
| 446 | last_pool = i = 0; |
| 447 | continue; |
| 448 | } |
| 449 | |
| 450 | break; |
| 451 | } while (1); |
| 452 | |
| 453 | log_err("smalloc: OOM. Consider using --alloc-size to increase the " |
| 454 | "shared memory available.\n"); |
| 455 | return NULL; |
| 456 | } |
| 457 | |
| 458 | void *scalloc(size_t nmemb, size_t size) |
| 459 | { |
| 460 | return smalloc(nmemb * size); |
| 461 | } |
| 462 | |
| 463 | char *smalloc_strdup(const char *str) |
| 464 | { |
| 465 | char *ptr = NULL; |
| 466 | |
| 467 | ptr = smalloc(strlen(str) + 1); |
| 468 | if (ptr) |
| 469 | strcpy(ptr, str); |
| 470 | return ptr; |
| 471 | } |