| 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 <assert.h> |
| 7 | #include <string.h> |
| 8 | |
| 9 | #include "fio.h" |
| 10 | #include "fio_sem.h" |
| 11 | #include "os/os.h" |
| 12 | #include "smalloc.h" |
| 13 | #include "log.h" |
| 14 | |
| 15 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
| 16 | |
| 17 | #define SMALLOC_BPB 32 /* block size, bytes-per-bit in bitmap */ |
| 18 | #define SMALLOC_BPI (sizeof(unsigned int) * 8) |
| 19 | #define SMALLOC_BPL (SMALLOC_BPB * SMALLOC_BPI) |
| 20 | |
| 21 | #define INITIAL_SIZE 16*1024*1024 /* new pool size */ |
| 22 | #define INITIAL_POOLS 8 /* maximum number of pools to setup */ |
| 23 | |
| 24 | #define MAX_POOLS 16 |
| 25 | |
| 26 | #define SMALLOC_PRE_RED 0xdeadbeefU |
| 27 | #define SMALLOC_POST_RED 0x5aa55aa5U |
| 28 | |
| 29 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
| 30 | #ifdef SMALLOC_REDZONE |
| 31 | static const int int_mask = sizeof(int) - 1; |
| 32 | #endif |
| 33 | |
| 34 | struct pool { |
| 35 | struct fio_sem *lock; /* protects this pool */ |
| 36 | void *map; /* map of blocks */ |
| 37 | unsigned int *bitmap; /* blocks free/busy map */ |
| 38 | size_t free_blocks; /* free blocks */ |
| 39 | size_t nr_blocks; /* total blocks */ |
| 40 | size_t next_non_full; |
| 41 | size_t mmap_size; |
| 42 | }; |
| 43 | |
| 44 | struct block_hdr { |
| 45 | size_t size; |
| 46 | #ifdef SMALLOC_REDZONE |
| 47 | unsigned int prered; |
| 48 | #endif |
| 49 | }; |
| 50 | |
| 51 | /* |
| 52 | * This suppresses the voluminous potential bitmap printout when |
| 53 | * smalloc encounters an OOM error |
| 54 | */ |
| 55 | static const bool enable_smalloc_debug = false; |
| 56 | |
| 57 | static struct pool *mp; |
| 58 | static unsigned int nr_pools; |
| 59 | static unsigned int last_pool; |
| 60 | |
| 61 | static inline int ptr_valid(struct pool *pool, void *ptr) |
| 62 | { |
| 63 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
| 64 | |
| 65 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); |
| 66 | } |
| 67 | |
| 68 | static inline size_t size_to_blocks(size_t size) |
| 69 | { |
| 70 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; |
| 71 | } |
| 72 | |
| 73 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
| 74 | unsigned int idx, size_t nr_blocks, |
| 75 | int (*func)(unsigned int *map, unsigned int mask)) |
| 76 | { |
| 77 | |
| 78 | while (nr_blocks) { |
| 79 | unsigned int this_blocks, mask; |
| 80 | unsigned int *map; |
| 81 | |
| 82 | if (pool_idx >= pool->nr_blocks) |
| 83 | return 0; |
| 84 | |
| 85 | map = &pool->bitmap[pool_idx]; |
| 86 | |
| 87 | this_blocks = nr_blocks; |
| 88 | if (this_blocks + idx > SMALLOC_BPI) { |
| 89 | this_blocks = SMALLOC_BPI - idx; |
| 90 | idx = SMALLOC_BPI - this_blocks; |
| 91 | } |
| 92 | |
| 93 | if (this_blocks == SMALLOC_BPI) |
| 94 | mask = -1U; |
| 95 | else |
| 96 | mask = ((1U << this_blocks) - 1) << idx; |
| 97 | |
| 98 | if (!func(map, mask)) |
| 99 | return 0; |
| 100 | |
| 101 | nr_blocks -= this_blocks; |
| 102 | idx = 0; |
| 103 | pool_idx++; |
| 104 | } |
| 105 | |
| 106 | return 1; |
| 107 | } |
| 108 | |
| 109 | static int mask_cmp(unsigned int *map, unsigned int mask) |
| 110 | { |
| 111 | return !(*map & mask); |
| 112 | } |
| 113 | |
| 114 | static int mask_clear(unsigned int *map, unsigned int mask) |
| 115 | { |
| 116 | assert((*map & mask) == mask); |
| 117 | *map &= ~mask; |
| 118 | return 1; |
| 119 | } |
| 120 | |
| 121 | static int mask_set(unsigned int *map, unsigned int mask) |
| 122 | { |
| 123 | assert(!(*map & mask)); |
| 124 | *map |= mask; |
| 125 | return 1; |
| 126 | } |
| 127 | |
| 128 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
| 129 | unsigned int idx, size_t nr_blocks) |
| 130 | { |
| 131 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
| 132 | } |
| 133 | |
| 134 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
| 135 | unsigned int idx, size_t nr_blocks) |
| 136 | { |
| 137 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
| 138 | } |
| 139 | |
| 140 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
| 141 | unsigned int idx, size_t nr_blocks) |
| 142 | { |
| 143 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
| 144 | } |
| 145 | |
| 146 | static int find_next_zero(int word, int start) |
| 147 | { |
| 148 | assert(word != -1U); |
| 149 | word >>= start; |
| 150 | return ffz(word) + start; |
| 151 | } |
| 152 | |
| 153 | static bool add_pool(struct pool *pool, unsigned int alloc_size) |
| 154 | { |
| 155 | int bitmap_blocks; |
| 156 | int mmap_flags; |
| 157 | void *ptr; |
| 158 | |
| 159 | if (nr_pools == MAX_POOLS) |
| 160 | return false; |
| 161 | |
| 162 | #ifdef SMALLOC_REDZONE |
| 163 | alloc_size += sizeof(unsigned int); |
| 164 | #endif |
| 165 | alloc_size += sizeof(struct block_hdr); |
| 166 | if (alloc_size < INITIAL_SIZE) |
| 167 | alloc_size = INITIAL_SIZE; |
| 168 | |
| 169 | /* round up to nearest full number of blocks */ |
| 170 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); |
| 171 | bitmap_blocks = alloc_size / SMALLOC_BPL; |
| 172 | alloc_size += bitmap_blocks * sizeof(unsigned int); |
| 173 | pool->mmap_size = alloc_size; |
| 174 | |
| 175 | pool->nr_blocks = bitmap_blocks; |
| 176 | pool->free_blocks = bitmap_blocks * SMALLOC_BPI; |
| 177 | |
| 178 | mmap_flags = OS_MAP_ANON; |
| 179 | #ifdef CONFIG_ESX |
| 180 | mmap_flags |= MAP_PRIVATE; |
| 181 | #else |
| 182 | mmap_flags |= MAP_SHARED; |
| 183 | #endif |
| 184 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, mmap_flags, -1, 0); |
| 185 | |
| 186 | if (ptr == MAP_FAILED) |
| 187 | goto out_fail; |
| 188 | |
| 189 | pool->map = ptr; |
| 190 | pool->bitmap = (unsigned int *)((char *) ptr + (pool->nr_blocks * SMALLOC_BPL)); |
| 191 | memset(pool->bitmap, 0, bitmap_blocks * sizeof(unsigned int)); |
| 192 | |
| 193 | pool->lock = fio_sem_init(FIO_SEM_UNLOCKED); |
| 194 | if (!pool->lock) |
| 195 | goto out_fail; |
| 196 | |
| 197 | nr_pools++; |
| 198 | return true; |
| 199 | out_fail: |
| 200 | log_err("smalloc: failed adding pool\n"); |
| 201 | if (pool->map) |
| 202 | munmap(pool->map, pool->mmap_size); |
| 203 | return false; |
| 204 | } |
| 205 | |
| 206 | void sinit(void) |
| 207 | { |
| 208 | bool ret; |
| 209 | int i; |
| 210 | |
| 211 | /* |
| 212 | * sinit() can be called more than once if alloc-size is |
| 213 | * set. But we want to allocate space for the struct pool |
| 214 | * instances only once. |
| 215 | */ |
| 216 | if (!mp) { |
| 217 | mp = (struct pool *) mmap(NULL, |
| 218 | MAX_POOLS * sizeof(struct pool), |
| 219 | PROT_READ | PROT_WRITE, |
| 220 | OS_MAP_ANON | MAP_SHARED, -1, 0); |
| 221 | |
| 222 | assert(mp != MAP_FAILED); |
| 223 | } |
| 224 | |
| 225 | for (i = 0; i < INITIAL_POOLS; i++) { |
| 226 | ret = add_pool(&mp[nr_pools], smalloc_pool_size); |
| 227 | if (!ret) |
| 228 | break; |
| 229 | } |
| 230 | |
| 231 | /* |
| 232 | * If we added at least one pool, we should be OK for most |
| 233 | * cases. |
| 234 | */ |
| 235 | assert(i); |
| 236 | } |
| 237 | |
| 238 | static void cleanup_pool(struct pool *pool) |
| 239 | { |
| 240 | /* |
| 241 | * This will also remove the temporary file we used as a backing |
| 242 | * store, it was already unlinked |
| 243 | */ |
| 244 | munmap(pool->map, pool->mmap_size); |
| 245 | |
| 246 | if (pool->lock) |
| 247 | fio_sem_remove(pool->lock); |
| 248 | } |
| 249 | |
| 250 | void scleanup(void) |
| 251 | { |
| 252 | unsigned int i; |
| 253 | |
| 254 | for (i = 0; i < nr_pools; i++) |
| 255 | cleanup_pool(&mp[i]); |
| 256 | |
| 257 | munmap(mp, MAX_POOLS * sizeof(struct pool)); |
| 258 | } |
| 259 | |
| 260 | #ifdef SMALLOC_REDZONE |
| 261 | static void *postred_ptr(struct block_hdr *hdr) |
| 262 | { |
| 263 | uintptr_t ptr; |
| 264 | |
| 265 | ptr = (uintptr_t) hdr + hdr->size - sizeof(unsigned int); |
| 266 | ptr = (uintptr_t) PTR_ALIGN(ptr, int_mask); |
| 267 | |
| 268 | return (void *) ptr; |
| 269 | } |
| 270 | |
| 271 | static void fill_redzone(struct block_hdr *hdr) |
| 272 | { |
| 273 | unsigned int *postred = postred_ptr(hdr); |
| 274 | |
| 275 | hdr->prered = SMALLOC_PRE_RED; |
| 276 | *postred = SMALLOC_POST_RED; |
| 277 | } |
| 278 | |
| 279 | static void sfree_check_redzone(struct block_hdr *hdr) |
| 280 | { |
| 281 | unsigned int *postred = postred_ptr(hdr); |
| 282 | |
| 283 | if (hdr->prered != SMALLOC_PRE_RED) { |
| 284 | log_err("smalloc pre redzone destroyed!\n" |
| 285 | " ptr=%p, prered=%x, expected %x\n", |
| 286 | hdr, hdr->prered, SMALLOC_PRE_RED); |
| 287 | assert(0); |
| 288 | } |
| 289 | if (*postred != SMALLOC_POST_RED) { |
| 290 | log_err("smalloc post redzone destroyed!\n" |
| 291 | " ptr=%p, postred=%x, expected %x\n", |
| 292 | hdr, *postred, SMALLOC_POST_RED); |
| 293 | assert(0); |
| 294 | } |
| 295 | } |
| 296 | #else |
| 297 | static void fill_redzone(struct block_hdr *hdr) |
| 298 | { |
| 299 | } |
| 300 | |
| 301 | static void sfree_check_redzone(struct block_hdr *hdr) |
| 302 | { |
| 303 | } |
| 304 | #endif |
| 305 | |
| 306 | static void sfree_pool(struct pool *pool, void *ptr) |
| 307 | { |
| 308 | struct block_hdr *hdr; |
| 309 | unsigned int i, idx; |
| 310 | unsigned long offset; |
| 311 | |
| 312 | if (!ptr) |
| 313 | return; |
| 314 | |
| 315 | ptr -= sizeof(*hdr); |
| 316 | hdr = ptr; |
| 317 | |
| 318 | assert(ptr_valid(pool, ptr)); |
| 319 | |
| 320 | sfree_check_redzone(hdr); |
| 321 | |
| 322 | offset = ptr - pool->map; |
| 323 | i = offset / SMALLOC_BPL; |
| 324 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; |
| 325 | |
| 326 | fio_sem_down(pool->lock); |
| 327 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
| 328 | if (i < pool->next_non_full) |
| 329 | pool->next_non_full = i; |
| 330 | pool->free_blocks += size_to_blocks(hdr->size); |
| 331 | fio_sem_up(pool->lock); |
| 332 | } |
| 333 | |
| 334 | void sfree(void *ptr) |
| 335 | { |
| 336 | struct pool *pool = NULL; |
| 337 | unsigned int i; |
| 338 | |
| 339 | if (!ptr) |
| 340 | return; |
| 341 | |
| 342 | for (i = 0; i < nr_pools; i++) { |
| 343 | if (ptr_valid(&mp[i], ptr)) { |
| 344 | pool = &mp[i]; |
| 345 | break; |
| 346 | } |
| 347 | } |
| 348 | |
| 349 | if (pool) { |
| 350 | sfree_pool(pool, ptr); |
| 351 | return; |
| 352 | } |
| 353 | |
| 354 | log_err("smalloc: ptr %p not from smalloc pool\n", ptr); |
| 355 | } |
| 356 | |
| 357 | static unsigned int find_best_index(struct pool *pool) |
| 358 | { |
| 359 | unsigned int i; |
| 360 | |
| 361 | assert(pool->free_blocks); |
| 362 | |
| 363 | for (i = pool->next_non_full; pool->bitmap[i] == -1U; i++) { |
| 364 | if (i == pool->nr_blocks - 1) { |
| 365 | unsigned int j; |
| 366 | |
| 367 | for (j = 0; j < pool->nr_blocks; j++) |
| 368 | if (pool->bitmap[j] != -1U) |
| 369 | return j; |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | return i; |
| 374 | } |
| 375 | |
| 376 | static void *__smalloc_pool(struct pool *pool, size_t size) |
| 377 | { |
| 378 | size_t nr_blocks; |
| 379 | unsigned int i; |
| 380 | unsigned int offset; |
| 381 | unsigned int last_idx; |
| 382 | void *ret = NULL; |
| 383 | |
| 384 | fio_sem_down(pool->lock); |
| 385 | |
| 386 | nr_blocks = size_to_blocks(size); |
| 387 | if (nr_blocks > pool->free_blocks) |
| 388 | goto fail; |
| 389 | |
| 390 | pool->next_non_full = find_best_index(pool); |
| 391 | |
| 392 | last_idx = 0; |
| 393 | offset = -1U; |
| 394 | i = pool->next_non_full; |
| 395 | while (i < pool->nr_blocks) { |
| 396 | unsigned int idx; |
| 397 | |
| 398 | if (pool->bitmap[i] == -1U) { |
| 399 | i++; |
| 400 | last_idx = 0; |
| 401 | continue; |
| 402 | } |
| 403 | |
| 404 | idx = find_next_zero(pool->bitmap[i], last_idx); |
| 405 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
| 406 | idx += nr_blocks; |
| 407 | if (idx < SMALLOC_BPI) |
| 408 | last_idx = idx; |
| 409 | else { |
| 410 | last_idx = 0; |
| 411 | while (idx >= SMALLOC_BPI) { |
| 412 | i++; |
| 413 | idx -= SMALLOC_BPI; |
| 414 | } |
| 415 | } |
| 416 | continue; |
| 417 | } |
| 418 | set_blocks(pool, i, idx, nr_blocks); |
| 419 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
| 420 | break; |
| 421 | } |
| 422 | |
| 423 | if (i < pool->nr_blocks) { |
| 424 | pool->free_blocks -= nr_blocks; |
| 425 | ret = pool->map + offset; |
| 426 | } |
| 427 | fail: |
| 428 | fio_sem_up(pool->lock); |
| 429 | return ret; |
| 430 | } |
| 431 | |
| 432 | static size_t size_to_alloc_size(size_t size) |
| 433 | { |
| 434 | size_t alloc_size = size + sizeof(struct block_hdr); |
| 435 | |
| 436 | /* |
| 437 | * Round to int alignment, so that the postred pointer will |
| 438 | * be naturally aligned as well. |
| 439 | */ |
| 440 | #ifdef SMALLOC_REDZONE |
| 441 | alloc_size += sizeof(unsigned int); |
| 442 | alloc_size = (alloc_size + int_mask) & ~int_mask; |
| 443 | #endif |
| 444 | |
| 445 | return alloc_size; |
| 446 | } |
| 447 | |
| 448 | static void *smalloc_pool(struct pool *pool, size_t size) |
| 449 | { |
| 450 | size_t alloc_size = size_to_alloc_size(size); |
| 451 | void *ptr; |
| 452 | |
| 453 | ptr = __smalloc_pool(pool, alloc_size); |
| 454 | if (ptr) { |
| 455 | struct block_hdr *hdr = ptr; |
| 456 | |
| 457 | hdr->size = alloc_size; |
| 458 | fill_redzone(hdr); |
| 459 | |
| 460 | ptr += sizeof(*hdr); |
| 461 | memset(ptr, 0, size); |
| 462 | } |
| 463 | |
| 464 | return ptr; |
| 465 | } |
| 466 | |
| 467 | static void smalloc_print_bitmap(struct pool *pool) |
| 468 | { |
| 469 | size_t nr_blocks = pool->nr_blocks; |
| 470 | unsigned int *bitmap = pool->bitmap; |
| 471 | unsigned int i, j; |
| 472 | char *buffer; |
| 473 | |
| 474 | if (!enable_smalloc_debug) |
| 475 | return; |
| 476 | |
| 477 | buffer = malloc(SMALLOC_BPI + 1); |
| 478 | if (!buffer) |
| 479 | return; |
| 480 | buffer[SMALLOC_BPI] = '\0'; |
| 481 | |
| 482 | for (i = 0; i < nr_blocks; i++) { |
| 483 | unsigned int line = bitmap[i]; |
| 484 | |
| 485 | /* skip completely full lines */ |
| 486 | if (line == -1U) |
| 487 | continue; |
| 488 | |
| 489 | for (j = 0; j < SMALLOC_BPI; j++) |
| 490 | if ((1 << j) & line) |
| 491 | buffer[SMALLOC_BPI-1-j] = '1'; |
| 492 | else |
| 493 | buffer[SMALLOC_BPI-1-j] = '0'; |
| 494 | |
| 495 | log_err("smalloc: bitmap %5u, %s\n", i, buffer); |
| 496 | } |
| 497 | |
| 498 | free(buffer); |
| 499 | } |
| 500 | |
| 501 | void smalloc_debug(size_t size) |
| 502 | { |
| 503 | unsigned int i; |
| 504 | size_t alloc_size = size_to_alloc_size(size); |
| 505 | size_t alloc_blocks; |
| 506 | |
| 507 | alloc_blocks = size_to_blocks(alloc_size); |
| 508 | |
| 509 | if (size) |
| 510 | log_err("smalloc: size = %lu, alloc_size = %lu, blocks = %lu\n", |
| 511 | (unsigned long) size, (unsigned long) alloc_size, |
| 512 | (unsigned long) alloc_blocks); |
| 513 | for (i = 0; i < nr_pools; i++) { |
| 514 | log_err("smalloc: pool %u, free/total blocks %u/%u\n", i, |
| 515 | (unsigned int) (mp[i].free_blocks), |
| 516 | (unsigned int) (mp[i].nr_blocks*sizeof(unsigned int)*8)); |
| 517 | if (size && mp[i].free_blocks >= alloc_blocks) { |
| 518 | void *ptr = smalloc_pool(&mp[i], size); |
| 519 | if (ptr) { |
| 520 | sfree(ptr); |
| 521 | last_pool = i; |
| 522 | log_err("smalloc: smalloc_pool %u succeeded\n", i); |
| 523 | } else { |
| 524 | log_err("smalloc: smalloc_pool %u failed\n", i); |
| 525 | log_err("smalloc: next_non_full=%u, nr_blocks=%u\n", |
| 526 | (unsigned int) mp[i].next_non_full, (unsigned int) mp[i].nr_blocks); |
| 527 | smalloc_print_bitmap(&mp[i]); |
| 528 | } |
| 529 | } |
| 530 | } |
| 531 | } |
| 532 | |
| 533 | void *smalloc(size_t size) |
| 534 | { |
| 535 | unsigned int i, end_pool; |
| 536 | |
| 537 | if (size != (unsigned int) size) |
| 538 | return NULL; |
| 539 | |
| 540 | i = last_pool; |
| 541 | end_pool = nr_pools; |
| 542 | |
| 543 | do { |
| 544 | for (; i < end_pool; i++) { |
| 545 | void *ptr = smalloc_pool(&mp[i], size); |
| 546 | |
| 547 | if (ptr) { |
| 548 | last_pool = i; |
| 549 | return ptr; |
| 550 | } |
| 551 | } |
| 552 | if (last_pool) { |
| 553 | end_pool = last_pool; |
| 554 | last_pool = i = 0; |
| 555 | continue; |
| 556 | } |
| 557 | |
| 558 | break; |
| 559 | } while (1); |
| 560 | |
| 561 | log_err("smalloc: OOM. Consider using --alloc-size to increase the " |
| 562 | "shared memory available.\n"); |
| 563 | smalloc_debug(size); |
| 564 | return NULL; |
| 565 | } |
| 566 | |
| 567 | void *scalloc(size_t nmemb, size_t size) |
| 568 | { |
| 569 | return smalloc(nmemb * size); |
| 570 | } |
| 571 | |
| 572 | char *smalloc_strdup(const char *str) |
| 573 | { |
| 574 | char *ptr = NULL; |
| 575 | |
| 576 | ptr = smalloc(strlen(str) + 1); |
| 577 | if (ptr) |
| 578 | strcpy(ptr, str); |
| 579 | return ptr; |
| 580 | } |