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 | ||
ec996e9c JA |
98 | static int blocks_iter(unsigned int *map, unsigned int idx, |
99 | unsigned int nr_blocks, | |
100 | int (*func)(unsigned int *map, unsigned int mask)) | |
d24c33a4 | 101 | { |
ec996e9c JA |
102 | while (nr_blocks) { |
103 | unsigned int this_blocks, mask; | |
104 | ||
105 | this_blocks = nr_blocks; | |
106 | if (this_blocks + idx > SMALLOC_BPI) { | |
107 | this_blocks = SMALLOC_BPI - idx; | |
108 | idx = SMALLOC_BPI - this_blocks; | |
109 | } | |
110 | ||
111 | if (this_blocks == SMALLOC_BPI) | |
112 | mask = -1U; | |
113 | else | |
114 | mask = ((1U << this_blocks) - 1) << idx; | |
115 | ||
116 | if (!func(map, mask)) | |
117 | return 0; | |
118 | ||
119 | nr_blocks -= this_blocks; | |
120 | idx = 0; | |
121 | map++; | |
122 | } | |
123 | ||
124 | return 1; | |
d24c33a4 JA |
125 | } |
126 | ||
ec996e9c | 127 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 128 | { |
ec996e9c | 129 | return !(*map & mask); |
d24c33a4 JA |
130 | } |
131 | ||
ec996e9c | 132 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 133 | { |
ec996e9c JA |
134 | *map &= ~mask; |
135 | return 1; | |
d24c33a4 JA |
136 | } |
137 | ||
ec996e9c | 138 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 139 | { |
ec996e9c JA |
140 | *map |= mask; |
141 | return 1; | |
d24c33a4 JA |
142 | } |
143 | ||
ec996e9c JA |
144 | static int blocks_free(unsigned int *map, unsigned int idx, |
145 | unsigned int nr_blocks) | |
d24c33a4 | 146 | { |
ec996e9c | 147 | return blocks_iter(map, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
148 | } |
149 | ||
ec996e9c JA |
150 | static void set_blocks(unsigned int *map, unsigned int idx, |
151 | unsigned int nr_blocks) | |
d24c33a4 | 152 | { |
ec996e9c | 153 | blocks_iter(map, idx, nr_blocks, mask_set); |
d24c33a4 JA |
154 | } |
155 | ||
ec996e9c JA |
156 | static void clear_blocks(unsigned int *map, unsigned int idx, |
157 | unsigned int nr_blocks) | |
d24c33a4 | 158 | { |
ec996e9c | 159 | blocks_iter(map, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
160 | } |
161 | ||
ec996e9c | 162 | static inline int __ffs(int word) |
d24c33a4 | 163 | { |
ec996e9c | 164 | int r = 0; |
d24c33a4 | 165 | |
ec996e9c JA |
166 | if (!(word & 0xffff)) { |
167 | word >>= 16; | |
168 | r += 16; | |
169 | } | |
170 | if (!(word & 0xff)) { | |
171 | word >>= 8; | |
172 | r += 8; | |
173 | } | |
174 | if (!(word & 0xf)) { | |
175 | word >>= 4; | |
176 | r += 4; | |
177 | } | |
178 | if (!(word & 3)) { | |
179 | word >>= 2; | |
180 | r += 2; | |
181 | } | |
182 | if (!(word & 1)) { | |
183 | word >>= 1; | |
184 | r += 1; | |
d24c33a4 JA |
185 | } |
186 | ||
ec996e9c JA |
187 | return r; |
188 | } | |
189 | ||
190 | static int find_next_zero(int word, int start) | |
191 | { | |
192 | assert(word != -1U); | |
193 | word >>= (start + 1); | |
194 | return __ffs(~word) + start + 1; | |
d24c33a4 JA |
195 | } |
196 | ||
adf57099 | 197 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 198 | { |
d24c33a4 | 199 | void *ptr; |
ec996e9c JA |
200 | int fd, bitmap_blocks; |
201 | ||
d24c33a4 JA |
202 | strcpy(pool->file, "/tmp/.fio_smalloc.XXXXXX"); |
203 | fd = mkstemp(pool->file); | |
204 | if (fd < 0) | |
205 | goto out_close; | |
206 | ||
55f6491d | 207 | #ifdef SMALLOC_REDZONE |
ec996e9c | 208 | alloc_size += sizeof(unsigned int); |
55f6491d | 209 | #endif |
ec996e9c JA |
210 | alloc_size += sizeof(struct block_hdr); |
211 | if (alloc_size < INITIAL_SIZE) | |
212 | alloc_size = INITIAL_SIZE; | |
213 | ||
214 | /* round up to nearest full number of blocks */ | |
215 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
216 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
217 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
218 | pool->mmap_size = alloc_size; | |
55f6491d | 219 | |
ec996e9c JA |
220 | pool->nr_blocks = bitmap_blocks; |
221 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 222 | |
ec996e9c | 223 | if (ftruncate(fd, alloc_size) < 0) |
d24c33a4 JA |
224 | goto out_unlink; |
225 | ||
ec996e9c | 226 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); |
d24c33a4 JA |
227 | if (ptr == MAP_FAILED) |
228 | goto out_unlink; | |
229 | ||
ec996e9c JA |
230 | memset(ptr, 0, alloc_size); |
231 | pool->map = ptr; | |
232 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 JA |
233 | |
234 | #ifdef MP_SAFE | |
235 | pool->lock = fio_mutex_init(1); | |
236 | if (!pool->lock) | |
237 | goto out_unlink; | |
238 | #endif | |
239 | ||
240 | pool->fd = fd; | |
241 | ||
65864cf7 | 242 | global_write_lock(); |
d24c33a4 | 243 | nr_pools++; |
65864cf7 | 244 | global_write_unlock(); |
d24c33a4 JA |
245 | return 0; |
246 | out_unlink: | |
ec996e9c | 247 | fprintf(stderr, "smalloc: failed adding pool\n"); |
d24c33a4 | 248 | if (pool->map) |
ec996e9c | 249 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
250 | unlink(pool->file); |
251 | out_close: | |
252 | if (fd >= 0) | |
253 | close(fd); | |
254 | return 1; | |
255 | } | |
256 | ||
257 | void sinit(void) | |
258 | { | |
4d4e80f2 | 259 | int ret; |
d24c33a4 JA |
260 | |
261 | #ifdef MP_SAFE | |
9c5b5290 | 262 | lock = fio_mutex_rw_init(); |
d24c33a4 | 263 | #endif |
adf57099 | 264 | ret = add_pool(&mp[0], INITIAL_SIZE); |
d24c33a4 JA |
265 | assert(!ret); |
266 | } | |
267 | ||
268 | static void cleanup_pool(struct pool *pool) | |
269 | { | |
270 | unlink(pool->file); | |
271 | close(pool->fd); | |
ec996e9c | 272 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
273 | |
274 | if (pool->lock) | |
275 | fio_mutex_remove(pool->lock); | |
276 | } | |
277 | ||
278 | void scleanup(void) | |
279 | { | |
280 | unsigned int i; | |
281 | ||
282 | for (i = 0; i < nr_pools; i++) | |
283 | cleanup_pool(&mp[i]); | |
284 | ||
285 | if (lock) | |
286 | fio_mutex_remove(lock); | |
287 | } | |
288 | ||
ec996e9c | 289 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d JA |
290 | { |
291 | #ifdef SMALLOC_REDZONE | |
ec996e9c | 292 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
55f6491d | 293 | |
ec996e9c JA |
294 | hdr->prered = SMALLOC_PRE_RED; |
295 | *postred = SMALLOC_POST_RED; | |
296 | #endif | |
297 | } | |
55f6491d | 298 | |
ec996e9c JA |
299 | static void sfree_check_redzone(struct block_hdr *hdr) |
300 | { | |
301 | #ifdef SMALLOC_REDZONE | |
302 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); | |
303 | ||
304 | if (hdr->prered != SMALLOC_PRE_RED) { | |
55f6491d JA |
305 | fprintf(stderr, "smalloc pre redzone destroyed!\n"); |
306 | fprintf(stderr, " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 307 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
308 | assert(0); |
309 | } | |
310 | if (*postred != SMALLOC_POST_RED) { | |
311 | fprintf(stderr, "smalloc post redzone destroyed!\n"); | |
312 | fprintf(stderr, " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 313 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
314 | assert(0); |
315 | } | |
316 | #endif | |
317 | } | |
318 | ||
d24c33a4 JA |
319 | static void sfree_pool(struct pool *pool, void *ptr) |
320 | { | |
ec996e9c | 321 | struct block_hdr *hdr; |
179446e0 | 322 | unsigned int i, idx; |
ec996e9c | 323 | unsigned long offset; |
d24c33a4 JA |
324 | |
325 | if (!ptr) | |
326 | return; | |
327 | ||
ec996e9c JA |
328 | ptr -= sizeof(*hdr); |
329 | hdr = ptr; | |
55f6491d | 330 | |
d24c33a4 JA |
331 | assert(ptr_valid(pool, ptr)); |
332 | ||
ec996e9c | 333 | sfree_check_redzone(hdr); |
d24c33a4 | 334 | |
ec996e9c JA |
335 | offset = ptr - pool->map; |
336 | i = offset / SMALLOC_BPL; | |
337 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 338 | |
ec996e9c | 339 | pool_lock(pool); |
179446e0 | 340 | clear_blocks(&pool->bitmap[i], idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
341 | if (i < pool->next_non_full) |
342 | pool->next_non_full = i; | |
179446e0 | 343 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
344 | pool_unlock(pool); |
345 | } | |
346 | ||
347 | void sfree(void *ptr) | |
348 | { | |
349 | struct pool *pool = NULL; | |
350 | unsigned int i; | |
351 | ||
8e5732e5 JA |
352 | if (!ptr) |
353 | return; | |
354 | ||
65864cf7 | 355 | global_read_lock(); |
d24c33a4 JA |
356 | |
357 | for (i = 0; i < nr_pools; i++) { | |
358 | if (ptr_valid(&mp[i], ptr)) { | |
359 | pool = &mp[i]; | |
360 | break; | |
361 | } | |
362 | } | |
363 | ||
65864cf7 | 364 | global_read_unlock(); |
d24c33a4 JA |
365 | |
366 | assert(pool); | |
367 | sfree_pool(pool, ptr); | |
368 | } | |
369 | ||
179446e0 JA |
370 | static inline unsigned int size_to_blocks(unsigned int size) |
371 | { | |
372 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
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 | } |