Commit | Line | Data |
---|---|---|
d24c33a4 JA |
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> | |
e43606c2 | 11 | #include <inttypes.h> |
d24c33a4 JA |
12 | #include <sys/types.h> |
13 | #include <limits.h> | |
3a8600b4 | 14 | #include <fcntl.h> |
d24c33a4 | 15 | |
6548f47f | 16 | #include "mutex.h" |
b3268b92 | 17 | #include "arch/arch.h" |
3a8600b4 | 18 | #include "os/os.h" |
10aa136b | 19 | #include "smalloc.h" |
b0f0326a | 20 | #include "log.h" |
d24c33a4 | 21 | |
55f6491d | 22 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
d24c33a4 | 23 | |
ec996e9c JA |
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 | ||
23bd40f9 | 28 | #define INITIAL_SIZE 16*1024*1024 /* new pool size */ |
85492cb8 | 29 | #define MAX_POOLS 8 /* maximum number of pools to setup */ |
d24c33a4 | 30 | |
55f6491d JA |
31 | #define SMALLOC_PRE_RED 0xdeadbeefU |
32 | #define SMALLOC_POST_RED 0x5aa55aa5U | |
55f6491d | 33 | |
2b386d25 | 34 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
aa1af5fd | 35 | #ifdef SMALLOC_REDZONE |
10aa136b | 36 | static const int int_mask = sizeof(int) - 1; |
aa1af5fd | 37 | #endif |
2b386d25 | 38 | |
d24c33a4 | 39 | struct pool { |
6548f47f | 40 | struct fio_mutex *lock; /* protects this pool */ |
d24c33a4 | 41 | void *map; /* map of blocks */ |
ec996e9c | 42 | unsigned int *bitmap; /* blocks free/busy map */ |
a3ebe7e0 JA |
43 | size_t free_blocks; /* free blocks */ |
44 | size_t nr_blocks; /* total blocks */ | |
45 | size_t next_non_full; | |
46 | size_t mmap_size; | |
ec996e9c JA |
47 | }; |
48 | ||
49 | struct block_hdr { | |
a3ebe7e0 | 50 | size_t size; |
ec996e9c JA |
51 | #ifdef SMALLOC_REDZONE |
52 | unsigned int prered; | |
53 | #endif | |
d24c33a4 JA |
54 | }; |
55 | ||
56 | static struct pool mp[MAX_POOLS]; | |
57 | static unsigned int nr_pools; | |
58 | static unsigned int last_pool; | |
d7df1d13 | 59 | static struct fio_rwlock *lock; |
d24c33a4 | 60 | |
d24c33a4 JA |
61 | static inline void pool_lock(struct pool *pool) |
62 | { | |
2e3e31e3 | 63 | fio_mutex_down(pool->lock); |
d24c33a4 JA |
64 | } |
65 | ||
66 | static inline void pool_unlock(struct pool *pool) | |
67 | { | |
2e3e31e3 | 68 | fio_mutex_up(pool->lock); |
d24c33a4 JA |
69 | } |
70 | ||
65864cf7 | 71 | static inline void global_read_lock(void) |
d24c33a4 | 72 | { |
d7df1d13 | 73 | fio_rwlock_read(lock); |
d24c33a4 JA |
74 | } |
75 | ||
65864cf7 | 76 | static inline void global_read_unlock(void) |
d24c33a4 | 77 | { |
d7df1d13 | 78 | fio_rwlock_unlock(lock); |
65864cf7 JA |
79 | } |
80 | ||
81 | static inline void global_write_lock(void) | |
82 | { | |
d7df1d13 | 83 | fio_rwlock_write(lock); |
65864cf7 JA |
84 | } |
85 | ||
86 | static inline void global_write_unlock(void) | |
87 | { | |
d7df1d13 | 88 | fio_rwlock_unlock(lock); |
d24c33a4 JA |
89 | } |
90 | ||
d24c33a4 JA |
91 | static inline int ptr_valid(struct pool *pool, void *ptr) |
92 | { | |
dcb69098 | 93 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
94 | |
95 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
96 | } |
97 | ||
a3ebe7e0 | 98 | static inline size_t size_to_blocks(size_t size) |
808e9ea8 JA |
99 | { |
100 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
101 | } | |
102 | ||
dcb69098 | 103 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 104 | unsigned int idx, size_t nr_blocks, |
ec996e9c | 105 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 106 | { |
dcb69098 | 107 | |
ec996e9c JA |
108 | while (nr_blocks) { |
109 | unsigned int this_blocks, mask; | |
dcb69098 JA |
110 | unsigned int *map; |
111 | ||
112 | if (pool_idx >= pool->nr_blocks) | |
113 | return 0; | |
114 | ||
115 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
116 | |
117 | this_blocks = nr_blocks; | |
118 | if (this_blocks + idx > SMALLOC_BPI) { | |
119 | this_blocks = SMALLOC_BPI - idx; | |
120 | idx = SMALLOC_BPI - this_blocks; | |
121 | } | |
122 | ||
123 | if (this_blocks == SMALLOC_BPI) | |
124 | mask = -1U; | |
125 | else | |
126 | mask = ((1U << this_blocks) - 1) << idx; | |
127 | ||
128 | if (!func(map, mask)) | |
129 | return 0; | |
130 | ||
131 | nr_blocks -= this_blocks; | |
132 | idx = 0; | |
dcb69098 | 133 | pool_idx++; |
ec996e9c JA |
134 | } |
135 | ||
136 | return 1; | |
d24c33a4 JA |
137 | } |
138 | ||
ec996e9c | 139 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 140 | { |
ec996e9c | 141 | return !(*map & mask); |
d24c33a4 JA |
142 | } |
143 | ||
ec996e9c | 144 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 145 | { |
dcb69098 | 146 | assert((*map & mask) == mask); |
ec996e9c JA |
147 | *map &= ~mask; |
148 | return 1; | |
d24c33a4 JA |
149 | } |
150 | ||
ec996e9c | 151 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 152 | { |
dcb69098 | 153 | assert(!(*map & mask)); |
ec996e9c JA |
154 | *map |= mask; |
155 | return 1; | |
d24c33a4 JA |
156 | } |
157 | ||
dcb69098 | 158 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 159 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 160 | { |
dcb69098 | 161 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
162 | } |
163 | ||
dcb69098 | 164 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 165 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 166 | { |
dcb69098 | 167 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
168 | } |
169 | ||
dcb69098 | 170 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 171 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 172 | { |
dcb69098 | 173 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
174 | } |
175 | ||
ec996e9c JA |
176 | static int find_next_zero(int word, int start) |
177 | { | |
178 | assert(word != -1U); | |
271067a6 JH |
179 | word >>= start; |
180 | return ffz(word) + start; | |
d24c33a4 JA |
181 | } |
182 | ||
adf57099 | 183 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 184 | { |
8d5844e9 | 185 | int bitmap_blocks; |
c8931876 | 186 | int mmap_flags; |
b8a6582e | 187 | void *ptr; |
ec996e9c | 188 | |
55f6491d | 189 | #ifdef SMALLOC_REDZONE |
ec996e9c | 190 | alloc_size += sizeof(unsigned int); |
55f6491d | 191 | #endif |
ec996e9c JA |
192 | alloc_size += sizeof(struct block_hdr); |
193 | if (alloc_size < INITIAL_SIZE) | |
194 | alloc_size = INITIAL_SIZE; | |
195 | ||
196 | /* round up to nearest full number of blocks */ | |
197 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
198 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
199 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
200 | pool->mmap_size = alloc_size; | |
0b9d69ec | 201 | |
ec996e9c JA |
202 | pool->nr_blocks = bitmap_blocks; |
203 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 204 | |
c8931876 JA |
205 | mmap_flags = OS_MAP_ANON; |
206 | #ifdef CONFIG_ESX | |
207 | mmap_flags |= MAP_PRIVATE; | |
208 | #else | |
209 | mmap_flags |= MAP_SHARED; | |
210 | #endif | |
211 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, mmap_flags, -1, 0); | |
212 | ||
d24c33a4 | 213 | if (ptr == MAP_FAILED) |
8d5844e9 | 214 | goto out_fail; |
d24c33a4 | 215 | |
ec996e9c JA |
216 | memset(ptr, 0, alloc_size); |
217 | pool->map = ptr; | |
218 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 | 219 | |
521da527 | 220 | pool->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED); |
d24c33a4 | 221 | if (!pool->lock) |
8d5844e9 | 222 | goto out_fail; |
d24c33a4 | 223 | |
d24c33a4 JA |
224 | nr_pools++; |
225 | return 0; | |
8d5844e9 | 226 | out_fail: |
b0f0326a | 227 | log_err("smalloc: failed adding pool\n"); |
d24c33a4 | 228 | if (pool->map) |
ec996e9c | 229 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
230 | return 1; |
231 | } | |
232 | ||
233 | void sinit(void) | |
234 | { | |
85492cb8 | 235 | int i, ret; |
d24c33a4 | 236 | |
d7df1d13 | 237 | lock = fio_rwlock_init(); |
85492cb8 JA |
238 | |
239 | for (i = 0; i < MAX_POOLS; i++) { | |
240 | ret = add_pool(&mp[i], INITIAL_SIZE); | |
241 | if (ret) | |
242 | break; | |
243 | } | |
244 | ||
245 | /* | |
246 | * If we added at least one pool, we should be OK for most | |
247 | * cases. | |
248 | */ | |
249 | assert(i); | |
d24c33a4 JA |
250 | } |
251 | ||
252 | static void cleanup_pool(struct pool *pool) | |
253 | { | |
443bb114 JA |
254 | /* |
255 | * This will also remove the temporary file we used as a backing | |
256 | * store, it was already unlinked | |
257 | */ | |
ec996e9c | 258 | munmap(pool->map, pool->mmap_size); |
6548f47f JA |
259 | |
260 | if (pool->lock) | |
261 | fio_mutex_remove(pool->lock); | |
d24c33a4 JA |
262 | } |
263 | ||
264 | void scleanup(void) | |
265 | { | |
266 | unsigned int i; | |
267 | ||
268 | for (i = 0; i < nr_pools; i++) | |
269 | cleanup_pool(&mp[i]); | |
270 | ||
6548f47f | 271 | if (lock) |
d7df1d13 | 272 | fio_rwlock_remove(lock); |
d24c33a4 JA |
273 | } |
274 | ||
89da54e8 | 275 | #ifdef SMALLOC_REDZONE |
cf98708d JA |
276 | static void *postred_ptr(struct block_hdr *hdr) |
277 | { | |
e43606c2 | 278 | uintptr_t ptr; |
cf98708d | 279 | |
e43606c2 | 280 | ptr = (uintptr_t) hdr + hdr->size - sizeof(unsigned int); |
cf98708d JA |
281 | ptr = (ptr + int_mask) & ~int_mask; |
282 | ||
283 | return (void *) ptr; | |
284 | } | |
285 | ||
ec996e9c | 286 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 287 | { |
cf98708d | 288 | unsigned int *postred = postred_ptr(hdr); |
55f6491d | 289 | |
ec996e9c JA |
290 | hdr->prered = SMALLOC_PRE_RED; |
291 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 292 | } |
55f6491d | 293 | |
ec996e9c JA |
294 | static void sfree_check_redzone(struct block_hdr *hdr) |
295 | { | |
cf98708d | 296 | unsigned int *postred = postred_ptr(hdr); |
ec996e9c JA |
297 | |
298 | if (hdr->prered != SMALLOC_PRE_RED) { | |
b0f0326a JA |
299 | log_err("smalloc pre redzone destroyed!\n" |
300 | " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 301 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
302 | assert(0); |
303 | } | |
304 | if (*postred != SMALLOC_POST_RED) { | |
b0f0326a JA |
305 | log_err("smalloc post redzone destroyed!\n" |
306 | " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 307 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
308 | assert(0); |
309 | } | |
89da54e8 JA |
310 | } |
311 | #else | |
312 | static void fill_redzone(struct block_hdr *hdr) | |
313 | { | |
55f6491d JA |
314 | } |
315 | ||
89da54e8 JA |
316 | static void sfree_check_redzone(struct block_hdr *hdr) |
317 | { | |
318 | } | |
319 | #endif | |
320 | ||
d24c33a4 JA |
321 | static void sfree_pool(struct pool *pool, void *ptr) |
322 | { | |
ec996e9c | 323 | struct block_hdr *hdr; |
179446e0 | 324 | unsigned int i, idx; |
ec996e9c | 325 | unsigned long offset; |
d24c33a4 JA |
326 | |
327 | if (!ptr) | |
328 | return; | |
329 | ||
ec996e9c JA |
330 | ptr -= sizeof(*hdr); |
331 | hdr = ptr; | |
55f6491d | 332 | |
d24c33a4 JA |
333 | assert(ptr_valid(pool, ptr)); |
334 | ||
ec996e9c | 335 | sfree_check_redzone(hdr); |
d24c33a4 | 336 | |
ec996e9c JA |
337 | offset = ptr - pool->map; |
338 | i = offset / SMALLOC_BPL; | |
339 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 340 | |
ec996e9c | 341 | pool_lock(pool); |
dcb69098 | 342 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
343 | if (i < pool->next_non_full) |
344 | pool->next_non_full = i; | |
179446e0 | 345 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
346 | pool_unlock(pool); |
347 | } | |
348 | ||
349 | void sfree(void *ptr) | |
350 | { | |
351 | struct pool *pool = NULL; | |
352 | unsigned int i; | |
353 | ||
8e5732e5 JA |
354 | if (!ptr) |
355 | return; | |
356 | ||
65864cf7 | 357 | global_read_lock(); |
d24c33a4 JA |
358 | |
359 | for (i = 0; i < nr_pools; i++) { | |
360 | if (ptr_valid(&mp[i], ptr)) { | |
361 | pool = &mp[i]; | |
362 | break; | |
363 | } | |
364 | } | |
365 | ||
65864cf7 | 366 | global_read_unlock(); |
d24c33a4 | 367 | |
45a65144 JA |
368 | if (pool) { |
369 | sfree_pool(pool, ptr); | |
370 | return; | |
371 | } | |
372 | ||
373 | log_err("smalloc: ptr %p not from smalloc pool\n", ptr); | |
d24c33a4 JA |
374 | } |
375 | ||
a3ebe7e0 | 376 | static void *__smalloc_pool(struct pool *pool, size_t size) |
d24c33a4 | 377 | { |
a3ebe7e0 | 378 | size_t nr_blocks; |
ec996e9c JA |
379 | unsigned int i; |
380 | unsigned int offset; | |
381 | unsigned int last_idx; | |
382 | void *ret = NULL; | |
d24c33a4 | 383 | |
d24c33a4 | 384 | pool_lock(pool); |
179446e0 JA |
385 | |
386 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 387 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 388 | goto fail; |
5ec10eaa | 389 | |
ec996e9c JA |
390 | i = pool->next_non_full; |
391 | last_idx = 0; | |
392 | offset = -1U; | |
393 | while (i < pool->nr_blocks) { | |
394 | unsigned int idx; | |
d24c33a4 | 395 | |
ec996e9c JA |
396 | if (pool->bitmap[i] == -1U) { |
397 | i++; | |
398 | pool->next_non_full = i; | |
399 | last_idx = 0; | |
400 | continue; | |
401 | } | |
d24c33a4 | 402 | |
ec996e9c | 403 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 404 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
405 | idx += nr_blocks; |
406 | if (idx < SMALLOC_BPI) | |
407 | last_idx = idx; | |
408 | else { | |
409 | last_idx = 0; | |
410 | while (idx >= SMALLOC_BPI) { | |
411 | i++; | |
412 | idx -= SMALLOC_BPI; | |
413 | } | |
414 | } | |
415 | continue; | |
d24c33a4 | 416 | } |
dcb69098 | 417 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
418 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
419 | break; | |
420 | } | |
421 | ||
422 | if (i < pool->nr_blocks) { | |
423 | pool->free_blocks -= nr_blocks; | |
424 | ret = pool->map + offset; | |
d24c33a4 | 425 | } |
ec996e9c | 426 | fail: |
d24c33a4 | 427 | pool_unlock(pool); |
ec996e9c | 428 | return ret; |
d24c33a4 JA |
429 | } |
430 | ||
a3ebe7e0 | 431 | static void *smalloc_pool(struct pool *pool, size_t size) |
55f6491d | 432 | { |
a3ebe7e0 | 433 | size_t alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
434 | void *ptr; |
435 | ||
cf98708d | 436 | /* |
122426da JA |
437 | * Round to int alignment, so that the postred pointer will |
438 | * be naturally aligned as well. | |
cf98708d | 439 | */ |
ec996e9c | 440 | #ifdef SMALLOC_REDZONE |
122426da JA |
441 | alloc_size += sizeof(unsigned int); |
442 | alloc_size = (alloc_size + int_mask) & ~int_mask; | |
ec996e9c JA |
443 | #endif |
444 | ||
445 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
446 | if (ptr) { |
447 | struct block_hdr *hdr = ptr; | |
55f6491d | 448 | |
89da54e8 JA |
449 | hdr->size = alloc_size; |
450 | fill_redzone(hdr); | |
55f6491d | 451 | |
89da54e8 JA |
452 | ptr += sizeof(*hdr); |
453 | memset(ptr, 0, size); | |
454 | } | |
ec996e9c | 455 | |
55f6491d | 456 | return ptr; |
55f6491d JA |
457 | } |
458 | ||
7982aa7d | 459 | void *smalloc(size_t size) |
d24c33a4 | 460 | { |
85492cb8 | 461 | unsigned int i, end_pool; |
d24c33a4 | 462 | |
7982aa7d JA |
463 | if (size != (unsigned int) size) |
464 | return NULL; | |
465 | ||
d1271dc1 | 466 | global_write_lock(); |
d24c33a4 | 467 | i = last_pool; |
85492cb8 | 468 | end_pool = nr_pools; |
d24c33a4 JA |
469 | |
470 | do { | |
85492cb8 | 471 | for (; i < end_pool; i++) { |
d24c33a4 JA |
472 | void *ptr = smalloc_pool(&mp[i], size); |
473 | ||
474 | if (ptr) { | |
475 | last_pool = i; | |
d1271dc1 | 476 | global_write_unlock(); |
d24c33a4 JA |
477 | return ptr; |
478 | } | |
479 | } | |
480 | if (last_pool) { | |
85492cb8 JA |
481 | end_pool = last_pool; |
482 | last_pool = i = 0; | |
d24c33a4 JA |
483 | continue; |
484 | } | |
485 | ||
85492cb8 | 486 | break; |
d24c33a4 JA |
487 | } while (1); |
488 | ||
d1271dc1 | 489 | global_write_unlock(); |
d24c33a4 JA |
490 | return NULL; |
491 | } | |
492 | ||
544992f7 JA |
493 | void *scalloc(size_t nmemb, size_t size) |
494 | { | |
495 | void *ret; | |
496 | ||
497 | ret = smalloc(nmemb * size); | |
498 | if (ret) | |
499 | memset(ret, 0, nmemb * size); | |
500 | ||
501 | return ret; | |
502 | } | |
503 | ||
d24c33a4 JA |
504 | char *smalloc_strdup(const char *str) |
505 | { | |
2894a2d4 | 506 | char *ptr = NULL; |
d24c33a4 JA |
507 | |
508 | ptr = smalloc(strlen(str) + 1); | |
2894a2d4 CE |
509 | if (ptr) |
510 | strcpy(ptr, str); | |
d24c33a4 JA |
511 | return ptr; |
512 | } |