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 | { | |
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 | ||
808e9ea8 JA |
98 | static inline unsigned int size_to_blocks(unsigned int size) |
99 | { | |
100 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
101 | } | |
102 | ||
dcb69098 JA |
103 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
104 | unsigned int idx, unsigned int 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 JA |
158 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
159 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 160 | { |
dcb69098 | 161 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
162 | } |
163 | ||
dcb69098 JA |
164 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
165 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 166 | { |
dcb69098 | 167 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
168 | } |
169 | ||
dcb69098 JA |
170 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
171 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 172 | { |
dcb69098 | 173 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
174 | } |
175 | ||
ec996e9c | 176 | static inline int __ffs(int word) |
d24c33a4 | 177 | { |
ec996e9c | 178 | int r = 0; |
d24c33a4 | 179 | |
ec996e9c JA |
180 | if (!(word & 0xffff)) { |
181 | word >>= 16; | |
182 | r += 16; | |
183 | } | |
184 | if (!(word & 0xff)) { | |
185 | word >>= 8; | |
186 | r += 8; | |
187 | } | |
188 | if (!(word & 0xf)) { | |
189 | word >>= 4; | |
190 | r += 4; | |
191 | } | |
192 | if (!(word & 3)) { | |
193 | word >>= 2; | |
194 | r += 2; | |
195 | } | |
196 | if (!(word & 1)) { | |
197 | word >>= 1; | |
198 | r += 1; | |
d24c33a4 JA |
199 | } |
200 | ||
ec996e9c JA |
201 | return r; |
202 | } | |
203 | ||
204 | static int find_next_zero(int word, int start) | |
205 | { | |
206 | assert(word != -1U); | |
207 | word >>= (start + 1); | |
208 | return __ffs(~word) + start + 1; | |
d24c33a4 JA |
209 | } |
210 | ||
adf57099 | 211 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 212 | { |
d24c33a4 | 213 | void *ptr; |
ec996e9c JA |
214 | int fd, bitmap_blocks; |
215 | ||
d24c33a4 JA |
216 | strcpy(pool->file, "/tmp/.fio_smalloc.XXXXXX"); |
217 | fd = mkstemp(pool->file); | |
218 | if (fd < 0) | |
219 | goto out_close; | |
220 | ||
55f6491d | 221 | #ifdef SMALLOC_REDZONE |
ec996e9c | 222 | alloc_size += sizeof(unsigned int); |
55f6491d | 223 | #endif |
ec996e9c JA |
224 | alloc_size += sizeof(struct block_hdr); |
225 | if (alloc_size < INITIAL_SIZE) | |
226 | alloc_size = INITIAL_SIZE; | |
227 | ||
228 | /* round up to nearest full number of blocks */ | |
229 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
230 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
231 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
232 | pool->mmap_size = alloc_size; | |
55f6491d | 233 | |
ec996e9c JA |
234 | pool->nr_blocks = bitmap_blocks; |
235 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 236 | |
ec996e9c | 237 | if (ftruncate(fd, alloc_size) < 0) |
d24c33a4 JA |
238 | goto out_unlink; |
239 | ||
ec996e9c | 240 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); |
d24c33a4 JA |
241 | if (ptr == MAP_FAILED) |
242 | goto out_unlink; | |
243 | ||
ec996e9c JA |
244 | memset(ptr, 0, alloc_size); |
245 | pool->map = ptr; | |
246 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 JA |
247 | |
248 | #ifdef MP_SAFE | |
249 | pool->lock = fio_mutex_init(1); | |
250 | if (!pool->lock) | |
251 | goto out_unlink; | |
252 | #endif | |
253 | ||
254 | pool->fd = fd; | |
255 | ||
65864cf7 | 256 | global_write_lock(); |
d24c33a4 | 257 | nr_pools++; |
65864cf7 | 258 | global_write_unlock(); |
d24c33a4 JA |
259 | return 0; |
260 | out_unlink: | |
ec996e9c | 261 | fprintf(stderr, "smalloc: failed adding pool\n"); |
d24c33a4 | 262 | if (pool->map) |
ec996e9c | 263 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
264 | unlink(pool->file); |
265 | out_close: | |
266 | if (fd >= 0) | |
267 | close(fd); | |
268 | return 1; | |
269 | } | |
270 | ||
271 | void sinit(void) | |
272 | { | |
4d4e80f2 | 273 | int ret; |
d24c33a4 JA |
274 | |
275 | #ifdef MP_SAFE | |
9c5b5290 | 276 | lock = fio_mutex_rw_init(); |
d24c33a4 | 277 | #endif |
adf57099 | 278 | ret = add_pool(&mp[0], INITIAL_SIZE); |
d24c33a4 JA |
279 | assert(!ret); |
280 | } | |
281 | ||
282 | static void cleanup_pool(struct pool *pool) | |
283 | { | |
284 | unlink(pool->file); | |
285 | close(pool->fd); | |
ec996e9c | 286 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
287 | |
288 | if (pool->lock) | |
289 | fio_mutex_remove(pool->lock); | |
290 | } | |
291 | ||
292 | void scleanup(void) | |
293 | { | |
294 | unsigned int i; | |
295 | ||
296 | for (i = 0; i < nr_pools; i++) | |
297 | cleanup_pool(&mp[i]); | |
298 | ||
299 | if (lock) | |
300 | fio_mutex_remove(lock); | |
301 | } | |
302 | ||
89da54e8 | 303 | #ifdef SMALLOC_REDZONE |
ec996e9c | 304 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 305 | { |
ec996e9c | 306 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
55f6491d | 307 | |
ec996e9c JA |
308 | hdr->prered = SMALLOC_PRE_RED; |
309 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 310 | } |
55f6491d | 311 | |
ec996e9c JA |
312 | static void sfree_check_redzone(struct block_hdr *hdr) |
313 | { | |
ec996e9c JA |
314 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
315 | ||
316 | if (hdr->prered != SMALLOC_PRE_RED) { | |
55f6491d JA |
317 | fprintf(stderr, "smalloc pre redzone destroyed!\n"); |
318 | fprintf(stderr, " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 319 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
320 | assert(0); |
321 | } | |
322 | if (*postred != SMALLOC_POST_RED) { | |
323 | fprintf(stderr, "smalloc post redzone destroyed!\n"); | |
324 | fprintf(stderr, " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 325 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
326 | assert(0); |
327 | } | |
89da54e8 JA |
328 | } |
329 | #else | |
330 | static void fill_redzone(struct block_hdr *hdr) | |
331 | { | |
55f6491d JA |
332 | } |
333 | ||
89da54e8 JA |
334 | static void sfree_check_redzone(struct block_hdr *hdr) |
335 | { | |
336 | } | |
337 | #endif | |
338 | ||
d24c33a4 JA |
339 | static void sfree_pool(struct pool *pool, void *ptr) |
340 | { | |
ec996e9c | 341 | struct block_hdr *hdr; |
179446e0 | 342 | unsigned int i, idx; |
ec996e9c | 343 | unsigned long offset; |
d24c33a4 JA |
344 | |
345 | if (!ptr) | |
346 | return; | |
347 | ||
ec996e9c JA |
348 | ptr -= sizeof(*hdr); |
349 | hdr = ptr; | |
55f6491d | 350 | |
d24c33a4 JA |
351 | assert(ptr_valid(pool, ptr)); |
352 | ||
ec996e9c | 353 | sfree_check_redzone(hdr); |
d24c33a4 | 354 | |
ec996e9c JA |
355 | offset = ptr - pool->map; |
356 | i = offset / SMALLOC_BPL; | |
357 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 358 | |
ec996e9c | 359 | pool_lock(pool); |
dcb69098 | 360 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
361 | if (i < pool->next_non_full) |
362 | pool->next_non_full = i; | |
179446e0 | 363 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
364 | pool_unlock(pool); |
365 | } | |
366 | ||
367 | void sfree(void *ptr) | |
368 | { | |
369 | struct pool *pool = NULL; | |
370 | unsigned int i; | |
371 | ||
8e5732e5 JA |
372 | if (!ptr) |
373 | return; | |
374 | ||
65864cf7 | 375 | global_read_lock(); |
d24c33a4 JA |
376 | |
377 | for (i = 0; i < nr_pools; i++) { | |
378 | if (ptr_valid(&mp[i], ptr)) { | |
379 | pool = &mp[i]; | |
380 | break; | |
381 | } | |
382 | } | |
383 | ||
65864cf7 | 384 | global_read_unlock(); |
d24c33a4 JA |
385 | |
386 | assert(pool); | |
387 | sfree_pool(pool, ptr); | |
388 | } | |
389 | ||
55f6491d | 390 | static void *__smalloc_pool(struct pool *pool, unsigned int size) |
d24c33a4 | 391 | { |
ec996e9c JA |
392 | unsigned int nr_blocks; |
393 | unsigned int i; | |
394 | unsigned int offset; | |
395 | unsigned int last_idx; | |
396 | void *ret = NULL; | |
d24c33a4 | 397 | |
d24c33a4 | 398 | pool_lock(pool); |
179446e0 JA |
399 | |
400 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 401 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 402 | goto fail; |
5ec10eaa | 403 | |
ec996e9c JA |
404 | i = pool->next_non_full; |
405 | last_idx = 0; | |
406 | offset = -1U; | |
407 | while (i < pool->nr_blocks) { | |
408 | unsigned int idx; | |
d24c33a4 | 409 | |
ec996e9c JA |
410 | if (pool->bitmap[i] == -1U) { |
411 | i++; | |
412 | pool->next_non_full = i; | |
413 | last_idx = 0; | |
414 | continue; | |
415 | } | |
d24c33a4 | 416 | |
ec996e9c | 417 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 418 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
419 | idx += nr_blocks; |
420 | if (idx < SMALLOC_BPI) | |
421 | last_idx = idx; | |
422 | else { | |
423 | last_idx = 0; | |
424 | while (idx >= SMALLOC_BPI) { | |
425 | i++; | |
426 | idx -= SMALLOC_BPI; | |
427 | } | |
428 | } | |
429 | continue; | |
d24c33a4 | 430 | } |
dcb69098 | 431 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
432 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
433 | break; | |
434 | } | |
435 | ||
436 | if (i < pool->nr_blocks) { | |
437 | pool->free_blocks -= nr_blocks; | |
438 | ret = pool->map + offset; | |
d24c33a4 | 439 | } |
ec996e9c | 440 | fail: |
d24c33a4 | 441 | pool_unlock(pool); |
ec996e9c | 442 | return ret; |
d24c33a4 JA |
443 | } |
444 | ||
55f6491d JA |
445 | static void *smalloc_pool(struct pool *pool, unsigned int size) |
446 | { | |
89da54e8 | 447 | unsigned int alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
448 | void *ptr; |
449 | ||
ec996e9c JA |
450 | #ifdef SMALLOC_REDZONE |
451 | alloc_size += sizeof(unsigned int); | |
452 | #endif | |
453 | ||
454 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
455 | if (ptr) { |
456 | struct block_hdr *hdr = ptr; | |
55f6491d | 457 | |
89da54e8 JA |
458 | hdr->size = alloc_size; |
459 | fill_redzone(hdr); | |
55f6491d | 460 | |
89da54e8 JA |
461 | ptr += sizeof(*hdr); |
462 | memset(ptr, 0, size); | |
463 | } | |
ec996e9c | 464 | |
55f6491d | 465 | return ptr; |
55f6491d JA |
466 | } |
467 | ||
d24c33a4 JA |
468 | void *smalloc(unsigned int size) |
469 | { | |
470 | unsigned int i; | |
471 | ||
65864cf7 | 472 | global_read_lock(); |
d24c33a4 JA |
473 | i = last_pool; |
474 | ||
475 | do { | |
476 | for (; i < nr_pools; i++) { | |
477 | void *ptr = smalloc_pool(&mp[i], size); | |
478 | ||
479 | if (ptr) { | |
480 | last_pool = i; | |
65864cf7 | 481 | global_read_unlock(); |
d24c33a4 JA |
482 | return ptr; |
483 | } | |
484 | } | |
485 | if (last_pool) { | |
486 | last_pool = 0; | |
487 | continue; | |
488 | } | |
489 | ||
ec996e9c | 490 | if (nr_pools + 1 > MAX_POOLS) |
d24c33a4 JA |
491 | break; |
492 | else { | |
493 | i = nr_pools; | |
65864cf7 | 494 | global_read_unlock(); |
adf57099 | 495 | if (add_pool(&mp[nr_pools], size)) |
65864cf7 JA |
496 | goto out; |
497 | global_read_lock(); | |
d24c33a4 JA |
498 | } |
499 | } while (1); | |
500 | ||
65864cf7 JA |
501 | global_read_unlock(); |
502 | out: | |
d24c33a4 JA |
503 | return NULL; |
504 | } | |
505 | ||
506 | char *smalloc_strdup(const char *str) | |
507 | { | |
508 | char *ptr; | |
509 | ||
510 | ptr = smalloc(strlen(str) + 1); | |
511 | strcpy(ptr, str); | |
512 | return ptr; | |
513 | } |