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> | |
11 | #include <sys/types.h> | |
12 | #include <limits.h> | |
13 | ||
14 | #include "mutex.h" | |
15 | ||
1f37f636 | 16 | #define MP_SAFE /* define to make thread safe */ |
55f6491d | 17 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
d24c33a4 | 18 | |
ec996e9c JA |
19 | #define SMALLOC_BPB 32 /* block size, bytes-per-bit in bitmap */ |
20 | #define SMALLOC_BPI (sizeof(unsigned int) * 8) | |
21 | #define SMALLOC_BPL (SMALLOC_BPB * SMALLOC_BPI) | |
22 | ||
23 | #define INITIAL_SIZE 1024*1024 /* new pool size */ | |
c5dda9e3 | 24 | #define MAX_POOLS 4 /* maximum number of pools to setup */ |
d24c33a4 | 25 | |
55f6491d JA |
26 | #define SMALLOC_PRE_RED 0xdeadbeefU |
27 | #define SMALLOC_POST_RED 0x5aa55aa5U | |
55f6491d | 28 | |
2b386d25 JA |
29 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
30 | ||
d24c33a4 JA |
31 | struct pool { |
32 | struct fio_mutex *lock; /* protects this pool */ | |
33 | void *map; /* map of blocks */ | |
ec996e9c JA |
34 | unsigned int *bitmap; /* blocks free/busy map */ |
35 | unsigned int free_blocks; /* free blocks */ | |
36 | unsigned int nr_blocks; /* total blocks */ | |
37 | unsigned int next_non_full; | |
d24c33a4 JA |
38 | int fd; /* memory backing fd */ |
39 | char file[PATH_MAX]; /* filename for fd */ | |
ec996e9c JA |
40 | unsigned int mmap_size; |
41 | }; | |
42 | ||
43 | struct block_hdr { | |
44 | unsigned int size; | |
45 | #ifdef SMALLOC_REDZONE | |
46 | unsigned int prered; | |
47 | #endif | |
d24c33a4 JA |
48 | }; |
49 | ||
50 | static struct pool mp[MAX_POOLS]; | |
51 | static unsigned int nr_pools; | |
52 | static unsigned int last_pool; | |
53 | static struct fio_mutex *lock; | |
54 | ||
d24c33a4 JA |
55 | static inline void pool_lock(struct pool *pool) |
56 | { | |
57 | if (pool->lock) | |
58 | fio_mutex_down(pool->lock); | |
59 | } | |
60 | ||
61 | static inline void pool_unlock(struct pool *pool) | |
62 | { | |
63 | if (pool->lock) | |
64 | fio_mutex_up(pool->lock); | |
65 | } | |
66 | ||
65864cf7 | 67 | static inline void global_read_lock(void) |
d24c33a4 JA |
68 | { |
69 | if (lock) | |
65864cf7 | 70 | fio_mutex_down_read(lock); |
d24c33a4 JA |
71 | } |
72 | ||
65864cf7 | 73 | static inline void global_read_unlock(void) |
d24c33a4 JA |
74 | { |
75 | if (lock) | |
65864cf7 JA |
76 | fio_mutex_up_read(lock); |
77 | } | |
78 | ||
79 | static inline void global_write_lock(void) | |
80 | { | |
81 | if (lock) | |
82 | fio_mutex_down_write(lock); | |
83 | } | |
84 | ||
85 | static inline void global_write_unlock(void) | |
86 | { | |
87 | if (lock) | |
88 | fio_mutex_up_write(lock); | |
d24c33a4 JA |
89 | } |
90 | ||
d24c33a4 JA |
91 | static inline int ptr_valid(struct pool *pool, void *ptr) |
92 | { | |
ec996e9c JA |
93 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
94 | ||
95 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
96 | } |
97 | ||
808e9ea8 JA |
98 | static inline unsigned int size_to_blocks(unsigned int size) |
99 | { | |
100 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
101 | } | |
102 | ||
ec996e9c JA |
103 | static int blocks_iter(unsigned int *map, unsigned int idx, |
104 | unsigned int nr_blocks, | |
105 | int (*func)(unsigned int *map, unsigned int mask)) | |
d24c33a4 | 106 | { |
ec996e9c JA |
107 | while (nr_blocks) { |
108 | unsigned int this_blocks, mask; | |
109 | ||
110 | this_blocks = nr_blocks; | |
111 | if (this_blocks + idx > SMALLOC_BPI) { | |
112 | this_blocks = SMALLOC_BPI - idx; | |
113 | idx = SMALLOC_BPI - this_blocks; | |
114 | } | |
115 | ||
116 | if (this_blocks == SMALLOC_BPI) | |
117 | mask = -1U; | |
118 | else | |
119 | mask = ((1U << this_blocks) - 1) << idx; | |
120 | ||
121 | if (!func(map, mask)) | |
122 | return 0; | |
123 | ||
124 | nr_blocks -= this_blocks; | |
125 | idx = 0; | |
126 | map++; | |
127 | } | |
128 | ||
129 | return 1; | |
d24c33a4 JA |
130 | } |
131 | ||
ec996e9c | 132 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 133 | { |
ec996e9c | 134 | return !(*map & mask); |
d24c33a4 JA |
135 | } |
136 | ||
ec996e9c | 137 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 138 | { |
ec996e9c JA |
139 | *map &= ~mask; |
140 | return 1; | |
d24c33a4 JA |
141 | } |
142 | ||
ec996e9c | 143 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 144 | { |
ec996e9c JA |
145 | *map |= mask; |
146 | return 1; | |
d24c33a4 JA |
147 | } |
148 | ||
ec996e9c JA |
149 | static int blocks_free(unsigned int *map, unsigned int idx, |
150 | unsigned int nr_blocks) | |
d24c33a4 | 151 | { |
ec996e9c | 152 | return blocks_iter(map, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
153 | } |
154 | ||
ec996e9c JA |
155 | static void set_blocks(unsigned int *map, unsigned int idx, |
156 | unsigned int nr_blocks) | |
d24c33a4 | 157 | { |
ec996e9c | 158 | blocks_iter(map, idx, nr_blocks, mask_set); |
d24c33a4 JA |
159 | } |
160 | ||
ec996e9c JA |
161 | static void clear_blocks(unsigned int *map, unsigned int idx, |
162 | unsigned int nr_blocks) | |
d24c33a4 | 163 | { |
ec996e9c | 164 | blocks_iter(map, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
165 | } |
166 | ||
ec996e9c | 167 | static inline int __ffs(int word) |
d24c33a4 | 168 | { |
ec996e9c | 169 | int r = 0; |
d24c33a4 | 170 | |
ec996e9c JA |
171 | if (!(word & 0xffff)) { |
172 | word >>= 16; | |
173 | r += 16; | |
174 | } | |
175 | if (!(word & 0xff)) { | |
176 | word >>= 8; | |
177 | r += 8; | |
178 | } | |
179 | if (!(word & 0xf)) { | |
180 | word >>= 4; | |
181 | r += 4; | |
182 | } | |
183 | if (!(word & 3)) { | |
184 | word >>= 2; | |
185 | r += 2; | |
186 | } | |
187 | if (!(word & 1)) { | |
188 | word >>= 1; | |
189 | r += 1; | |
d24c33a4 JA |
190 | } |
191 | ||
ec996e9c JA |
192 | return r; |
193 | } | |
194 | ||
195 | static int find_next_zero(int word, int start) | |
196 | { | |
197 | assert(word != -1U); | |
198 | word >>= (start + 1); | |
199 | return __ffs(~word) + start + 1; | |
d24c33a4 JA |
200 | } |
201 | ||
adf57099 | 202 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 203 | { |
d24c33a4 | 204 | void *ptr; |
ec996e9c JA |
205 | int fd, bitmap_blocks; |
206 | ||
d24c33a4 JA |
207 | strcpy(pool->file, "/tmp/.fio_smalloc.XXXXXX"); |
208 | fd = mkstemp(pool->file); | |
209 | if (fd < 0) | |
210 | goto out_close; | |
211 | ||
55f6491d | 212 | #ifdef SMALLOC_REDZONE |
ec996e9c | 213 | alloc_size += sizeof(unsigned int); |
55f6491d | 214 | #endif |
ec996e9c JA |
215 | alloc_size += sizeof(struct block_hdr); |
216 | if (alloc_size < INITIAL_SIZE) | |
217 | alloc_size = INITIAL_SIZE; | |
218 | ||
219 | /* round up to nearest full number of blocks */ | |
220 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
221 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
222 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
223 | pool->mmap_size = alloc_size; | |
55f6491d | 224 | |
ec996e9c JA |
225 | pool->nr_blocks = bitmap_blocks; |
226 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 227 | |
ec996e9c | 228 | if (ftruncate(fd, alloc_size) < 0) |
d24c33a4 JA |
229 | goto out_unlink; |
230 | ||
ec996e9c | 231 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); |
d24c33a4 JA |
232 | if (ptr == MAP_FAILED) |
233 | goto out_unlink; | |
234 | ||
ec996e9c JA |
235 | memset(ptr, 0, alloc_size); |
236 | pool->map = ptr; | |
237 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 JA |
238 | |
239 | #ifdef MP_SAFE | |
240 | pool->lock = fio_mutex_init(1); | |
241 | if (!pool->lock) | |
242 | goto out_unlink; | |
243 | #endif | |
244 | ||
245 | pool->fd = fd; | |
246 | ||
65864cf7 | 247 | global_write_lock(); |
d24c33a4 | 248 | nr_pools++; |
65864cf7 | 249 | global_write_unlock(); |
d24c33a4 JA |
250 | return 0; |
251 | out_unlink: | |
ec996e9c | 252 | fprintf(stderr, "smalloc: failed adding pool\n"); |
d24c33a4 | 253 | if (pool->map) |
ec996e9c | 254 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
255 | unlink(pool->file); |
256 | out_close: | |
257 | if (fd >= 0) | |
258 | close(fd); | |
259 | return 1; | |
260 | } | |
261 | ||
262 | void sinit(void) | |
263 | { | |
4d4e80f2 | 264 | int ret; |
d24c33a4 JA |
265 | |
266 | #ifdef MP_SAFE | |
9c5b5290 | 267 | lock = fio_mutex_rw_init(); |
d24c33a4 | 268 | #endif |
adf57099 | 269 | ret = add_pool(&mp[0], INITIAL_SIZE); |
d24c33a4 JA |
270 | assert(!ret); |
271 | } | |
272 | ||
273 | static void cleanup_pool(struct pool *pool) | |
274 | { | |
275 | unlink(pool->file); | |
276 | close(pool->fd); | |
ec996e9c | 277 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
278 | |
279 | if (pool->lock) | |
280 | fio_mutex_remove(pool->lock); | |
281 | } | |
282 | ||
283 | void scleanup(void) | |
284 | { | |
285 | unsigned int i; | |
286 | ||
287 | for (i = 0; i < nr_pools; i++) | |
288 | cleanup_pool(&mp[i]); | |
289 | ||
290 | if (lock) | |
291 | fio_mutex_remove(lock); | |
292 | } | |
293 | ||
ec996e9c | 294 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d JA |
295 | { |
296 | #ifdef SMALLOC_REDZONE | |
ec996e9c | 297 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
55f6491d | 298 | |
ec996e9c JA |
299 | hdr->prered = SMALLOC_PRE_RED; |
300 | *postred = SMALLOC_POST_RED; | |
301 | #endif | |
302 | } | |
55f6491d | 303 | |
ec996e9c JA |
304 | static void sfree_check_redzone(struct block_hdr *hdr) |
305 | { | |
306 | #ifdef SMALLOC_REDZONE | |
307 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); | |
308 | ||
309 | if (hdr->prered != SMALLOC_PRE_RED) { | |
55f6491d JA |
310 | fprintf(stderr, "smalloc pre redzone destroyed!\n"); |
311 | fprintf(stderr, " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 312 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
313 | assert(0); |
314 | } | |
315 | if (*postred != SMALLOC_POST_RED) { | |
316 | fprintf(stderr, "smalloc post redzone destroyed!\n"); | |
317 | fprintf(stderr, " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 318 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
319 | assert(0); |
320 | } | |
321 | #endif | |
322 | } | |
323 | ||
d24c33a4 JA |
324 | static void sfree_pool(struct pool *pool, void *ptr) |
325 | { | |
ec996e9c | 326 | struct block_hdr *hdr; |
179446e0 | 327 | unsigned int i, idx; |
ec996e9c | 328 | unsigned long offset; |
d24c33a4 JA |
329 | |
330 | if (!ptr) | |
331 | return; | |
332 | ||
ec996e9c JA |
333 | ptr -= sizeof(*hdr); |
334 | hdr = ptr; | |
55f6491d | 335 | |
d24c33a4 JA |
336 | assert(ptr_valid(pool, ptr)); |
337 | ||
ec996e9c | 338 | sfree_check_redzone(hdr); |
d24c33a4 | 339 | |
ec996e9c JA |
340 | offset = ptr - pool->map; |
341 | i = offset / SMALLOC_BPL; | |
342 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 343 | |
ec996e9c | 344 | pool_lock(pool); |
179446e0 | 345 | clear_blocks(&pool->bitmap[i], idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
346 | if (i < pool->next_non_full) |
347 | pool->next_non_full = i; | |
179446e0 | 348 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
349 | pool_unlock(pool); |
350 | } | |
351 | ||
352 | void sfree(void *ptr) | |
353 | { | |
354 | struct pool *pool = NULL; | |
355 | unsigned int i; | |
356 | ||
8e5732e5 JA |
357 | if (!ptr) |
358 | return; | |
359 | ||
65864cf7 | 360 | global_read_lock(); |
d24c33a4 JA |
361 | |
362 | for (i = 0; i < nr_pools; i++) { | |
363 | if (ptr_valid(&mp[i], ptr)) { | |
364 | pool = &mp[i]; | |
365 | break; | |
366 | } | |
367 | } | |
368 | ||
65864cf7 | 369 | global_read_unlock(); |
d24c33a4 JA |
370 | |
371 | assert(pool); | |
372 | sfree_pool(pool, ptr); | |
373 | } | |
374 | ||
55f6491d | 375 | static void *__smalloc_pool(struct pool *pool, unsigned int size) |
d24c33a4 | 376 | { |
ec996e9c JA |
377 | unsigned int nr_blocks; |
378 | unsigned int i; | |
379 | unsigned int offset; | |
380 | unsigned int last_idx; | |
381 | void *ret = NULL; | |
d24c33a4 | 382 | |
d24c33a4 | 383 | pool_lock(pool); |
179446e0 JA |
384 | |
385 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 386 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 387 | goto fail; |
5ec10eaa | 388 | |
ec996e9c JA |
389 | i = pool->next_non_full; |
390 | last_idx = 0; | |
391 | offset = -1U; | |
392 | while (i < pool->nr_blocks) { | |
393 | unsigned int idx; | |
d24c33a4 | 394 | |
ec996e9c JA |
395 | if (pool->bitmap[i] == -1U) { |
396 | i++; | |
397 | pool->next_non_full = i; | |
398 | last_idx = 0; | |
399 | continue; | |
400 | } | |
d24c33a4 | 401 | |
ec996e9c JA |
402 | idx = find_next_zero(pool->bitmap[i], last_idx); |
403 | if (!blocks_free(&pool->bitmap[i], idx, nr_blocks)) { | |
404 | idx += nr_blocks; | |
405 | if (idx < SMALLOC_BPI) | |
406 | last_idx = idx; | |
407 | else { | |
408 | last_idx = 0; | |
409 | while (idx >= SMALLOC_BPI) { | |
410 | i++; | |
411 | idx -= SMALLOC_BPI; | |
412 | } | |
413 | } | |
414 | continue; | |
d24c33a4 | 415 | } |
ec996e9c JA |
416 | set_blocks(&pool->bitmap[i], idx, nr_blocks); |
417 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; | |
418 | break; | |
419 | } | |
420 | ||
421 | if (i < pool->nr_blocks) { | |
422 | pool->free_blocks -= nr_blocks; | |
423 | ret = pool->map + offset; | |
d24c33a4 | 424 | } |
ec996e9c | 425 | fail: |
d24c33a4 | 426 | pool_unlock(pool); |
ec996e9c | 427 | return ret; |
d24c33a4 JA |
428 | } |
429 | ||
55f6491d JA |
430 | static void *smalloc_pool(struct pool *pool, unsigned int size) |
431 | { | |
ec996e9c JA |
432 | struct block_hdr *hdr; |
433 | unsigned int alloc_size; | |
55f6491d JA |
434 | void *ptr; |
435 | ||
ec996e9c JA |
436 | alloc_size = size + sizeof(*hdr); |
437 | #ifdef SMALLOC_REDZONE | |
438 | alloc_size += sizeof(unsigned int); | |
439 | #endif | |
440 | ||
441 | ptr = __smalloc_pool(pool, alloc_size); | |
442 | if (!ptr) { | |
443 | printf("failed allocating %u\n", alloc_size); | |
55f6491d | 444 | return NULL; |
ec996e9c | 445 | } |
55f6491d | 446 | |
ec996e9c JA |
447 | hdr = ptr; |
448 | hdr->size = alloc_size; | |
449 | ptr += sizeof(*hdr); | |
55f6491d | 450 | |
ec996e9c JA |
451 | fill_redzone(hdr); |
452 | ||
453 | memset(ptr, 0, size); | |
55f6491d | 454 | return ptr; |
55f6491d JA |
455 | } |
456 | ||
d24c33a4 JA |
457 | void *smalloc(unsigned int size) |
458 | { | |
459 | unsigned int i; | |
460 | ||
65864cf7 | 461 | global_read_lock(); |
d24c33a4 JA |
462 | i = last_pool; |
463 | ||
464 | do { | |
465 | for (; i < nr_pools; i++) { | |
466 | void *ptr = smalloc_pool(&mp[i], size); | |
467 | ||
468 | if (ptr) { | |
469 | last_pool = i; | |
65864cf7 | 470 | global_read_unlock(); |
d24c33a4 JA |
471 | return ptr; |
472 | } | |
473 | } | |
474 | if (last_pool) { | |
475 | last_pool = 0; | |
476 | continue; | |
477 | } | |
478 | ||
ec996e9c | 479 | if (nr_pools + 1 > MAX_POOLS) |
d24c33a4 JA |
480 | break; |
481 | else { | |
482 | i = nr_pools; | |
65864cf7 | 483 | global_read_unlock(); |
adf57099 | 484 | if (add_pool(&mp[nr_pools], size)) |
65864cf7 JA |
485 | goto out; |
486 | global_read_lock(); | |
d24c33a4 JA |
487 | } |
488 | } while (1); | |
489 | ||
65864cf7 JA |
490 | global_read_unlock(); |
491 | out: | |
d24c33a4 JA |
492 | return NULL; |
493 | } | |
494 | ||
495 | char *smalloc_strdup(const char *str) | |
496 | { | |
497 | char *ptr; | |
498 | ||
499 | ptr = smalloc(strlen(str) + 1); | |
500 | strcpy(ptr, str); | |
501 | return ptr; | |
502 | } |