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 | ||
6548f47f | 14 | #include "mutex.h" |
b3268b92 | 15 | #include "arch/arch.h" |
d24c33a4 | 16 | |
6548f47f | 17 | #define MP_SAFE /* define to make thread safe */ |
55f6491d | 18 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
d24c33a4 | 19 | |
ec996e9c JA |
20 | #define SMALLOC_BPB 32 /* block size, bytes-per-bit in bitmap */ |
21 | #define SMALLOC_BPI (sizeof(unsigned int) * 8) | |
22 | #define SMALLOC_BPL (SMALLOC_BPB * SMALLOC_BPI) | |
23 | ||
24 | #define INITIAL_SIZE 1024*1024 /* new pool size */ | |
68857686 | 25 | #define MAX_POOLS 128 /* maximum number of pools to setup */ |
d24c33a4 | 26 | |
55f6491d JA |
27 | #define SMALLOC_PRE_RED 0xdeadbeefU |
28 | #define SMALLOC_POST_RED 0x5aa55aa5U | |
55f6491d | 29 | |
2b386d25 JA |
30 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
31 | ||
d24c33a4 | 32 | struct pool { |
6548f47f | 33 | struct fio_mutex *lock; /* protects this pool */ |
d24c33a4 | 34 | void *map; /* map of blocks */ |
ec996e9c JA |
35 | unsigned int *bitmap; /* blocks free/busy map */ |
36 | unsigned int free_blocks; /* free blocks */ | |
37 | unsigned int nr_blocks; /* total blocks */ | |
38 | unsigned int next_non_full; | |
d24c33a4 JA |
39 | int fd; /* memory backing fd */ |
40 | char file[PATH_MAX]; /* filename for fd */ | |
ec996e9c JA |
41 | unsigned int mmap_size; |
42 | }; | |
43 | ||
44 | struct block_hdr { | |
45 | unsigned int size; | |
46 | #ifdef SMALLOC_REDZONE | |
47 | unsigned int prered; | |
48 | #endif | |
d24c33a4 JA |
49 | }; |
50 | ||
51 | static struct pool mp[MAX_POOLS]; | |
52 | static unsigned int nr_pools; | |
53 | static unsigned int last_pool; | |
6548f47f | 54 | static struct fio_mutex *lock; |
d24c33a4 | 55 | |
d24c33a4 JA |
56 | static inline void pool_lock(struct pool *pool) |
57 | { | |
6548f47f JA |
58 | if (pool->lock) |
59 | fio_mutex_down(pool->lock); | |
d24c33a4 JA |
60 | } |
61 | ||
62 | static inline void pool_unlock(struct pool *pool) | |
63 | { | |
6548f47f JA |
64 | if (pool->lock) |
65 | fio_mutex_up(pool->lock); | |
d24c33a4 JA |
66 | } |
67 | ||
65864cf7 | 68 | static inline void global_read_lock(void) |
d24c33a4 | 69 | { |
6548f47f JA |
70 | if (lock) |
71 | fio_mutex_down_read(lock); | |
d24c33a4 JA |
72 | } |
73 | ||
65864cf7 | 74 | static inline void global_read_unlock(void) |
d24c33a4 | 75 | { |
6548f47f JA |
76 | if (lock) |
77 | fio_mutex_up_read(lock); | |
65864cf7 JA |
78 | } |
79 | ||
80 | static inline void global_write_lock(void) | |
81 | { | |
6548f47f JA |
82 | if (lock) |
83 | fio_mutex_down_write(lock); | |
65864cf7 JA |
84 | } |
85 | ||
86 | static inline void global_write_unlock(void) | |
87 | { | |
6548f47f JA |
88 | if (lock) |
89 | fio_mutex_up_write(lock); | |
d24c33a4 JA |
90 | } |
91 | ||
d24c33a4 JA |
92 | static inline int ptr_valid(struct pool *pool, void *ptr) |
93 | { | |
dcb69098 | 94 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
95 | |
96 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
97 | } |
98 | ||
808e9ea8 JA |
99 | static inline unsigned int size_to_blocks(unsigned int size) |
100 | { | |
101 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
102 | } | |
103 | ||
dcb69098 JA |
104 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
105 | unsigned int idx, unsigned int nr_blocks, | |
ec996e9c | 106 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 107 | { |
dcb69098 | 108 | |
ec996e9c JA |
109 | while (nr_blocks) { |
110 | unsigned int this_blocks, mask; | |
dcb69098 JA |
111 | unsigned int *map; |
112 | ||
113 | if (pool_idx >= pool->nr_blocks) | |
114 | return 0; | |
115 | ||
116 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
117 | |
118 | this_blocks = nr_blocks; | |
119 | if (this_blocks + idx > SMALLOC_BPI) { | |
120 | this_blocks = SMALLOC_BPI - idx; | |
121 | idx = SMALLOC_BPI - this_blocks; | |
122 | } | |
123 | ||
124 | if (this_blocks == SMALLOC_BPI) | |
125 | mask = -1U; | |
126 | else | |
127 | mask = ((1U << this_blocks) - 1) << idx; | |
128 | ||
129 | if (!func(map, mask)) | |
130 | return 0; | |
131 | ||
132 | nr_blocks -= this_blocks; | |
133 | idx = 0; | |
dcb69098 | 134 | pool_idx++; |
ec996e9c JA |
135 | } |
136 | ||
137 | return 1; | |
d24c33a4 JA |
138 | } |
139 | ||
ec996e9c | 140 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 141 | { |
ec996e9c | 142 | return !(*map & mask); |
d24c33a4 JA |
143 | } |
144 | ||
ec996e9c | 145 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 146 | { |
dcb69098 | 147 | assert((*map & mask) == mask); |
ec996e9c JA |
148 | *map &= ~mask; |
149 | return 1; | |
d24c33a4 JA |
150 | } |
151 | ||
ec996e9c | 152 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 153 | { |
dcb69098 | 154 | assert(!(*map & mask)); |
ec996e9c JA |
155 | *map |= mask; |
156 | return 1; | |
d24c33a4 JA |
157 | } |
158 | ||
dcb69098 JA |
159 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
160 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 161 | { |
dcb69098 | 162 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
163 | } |
164 | ||
dcb69098 JA |
165 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
166 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 167 | { |
dcb69098 | 168 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
169 | } |
170 | ||
dcb69098 JA |
171 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
172 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 173 | { |
dcb69098 | 174 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
175 | } |
176 | ||
ec996e9c JA |
177 | static int find_next_zero(int word, int start) |
178 | { | |
179 | assert(word != -1U); | |
180 | word >>= (start + 1); | |
b3268b92 | 181 | return ffz(word) + start + 1; |
d24c33a4 JA |
182 | } |
183 | ||
adf57099 | 184 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 185 | { |
d24c33a4 | 186 | void *ptr; |
ec996e9c JA |
187 | int fd, bitmap_blocks; |
188 | ||
d24c33a4 JA |
189 | strcpy(pool->file, "/tmp/.fio_smalloc.XXXXXX"); |
190 | fd = mkstemp(pool->file); | |
191 | if (fd < 0) | |
192 | goto out_close; | |
193 | ||
55f6491d | 194 | #ifdef SMALLOC_REDZONE |
ec996e9c | 195 | alloc_size += sizeof(unsigned int); |
55f6491d | 196 | #endif |
ec996e9c JA |
197 | alloc_size += sizeof(struct block_hdr); |
198 | if (alloc_size < INITIAL_SIZE) | |
199 | alloc_size = INITIAL_SIZE; | |
200 | ||
201 | /* round up to nearest full number of blocks */ | |
202 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
203 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
204 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
205 | pool->mmap_size = alloc_size; | |
55f6491d | 206 | |
ec996e9c JA |
207 | pool->nr_blocks = bitmap_blocks; |
208 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 209 | |
ec996e9c | 210 | if (ftruncate(fd, alloc_size) < 0) |
d24c33a4 JA |
211 | goto out_unlink; |
212 | ||
ec996e9c | 213 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); |
d24c33a4 JA |
214 | if (ptr == MAP_FAILED) |
215 | goto out_unlink; | |
216 | ||
ec996e9c JA |
217 | memset(ptr, 0, alloc_size); |
218 | pool->map = ptr; | |
219 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 | 220 | |
6548f47f JA |
221 | #ifdef MP_SAFE |
222 | pool->lock = fio_mutex_init(1); | |
d24c33a4 JA |
223 | if (!pool->lock) |
224 | goto out_unlink; | |
6548f47f | 225 | #endif |
d24c33a4 JA |
226 | |
227 | pool->fd = fd; | |
228 | ||
d24c33a4 JA |
229 | nr_pools++; |
230 | return 0; | |
231 | out_unlink: | |
ec996e9c | 232 | fprintf(stderr, "smalloc: failed adding pool\n"); |
d24c33a4 | 233 | if (pool->map) |
ec996e9c | 234 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
235 | unlink(pool->file); |
236 | out_close: | |
237 | if (fd >= 0) | |
238 | close(fd); | |
239 | return 1; | |
240 | } | |
241 | ||
242 | void sinit(void) | |
243 | { | |
4d4e80f2 | 244 | int ret; |
d24c33a4 | 245 | |
6548f47f JA |
246 | #ifdef MP_SAFE |
247 | lock = fio_mutex_rw_init(); | |
248 | #endif | |
adf57099 | 249 | ret = add_pool(&mp[0], INITIAL_SIZE); |
d24c33a4 JA |
250 | assert(!ret); |
251 | } | |
252 | ||
253 | static void cleanup_pool(struct pool *pool) | |
254 | { | |
255 | unlink(pool->file); | |
256 | close(pool->fd); | |
ec996e9c | 257 | munmap(pool->map, pool->mmap_size); |
6548f47f JA |
258 | |
259 | if (pool->lock) | |
260 | fio_mutex_remove(pool->lock); | |
d24c33a4 JA |
261 | } |
262 | ||
263 | void scleanup(void) | |
264 | { | |
265 | unsigned int i; | |
266 | ||
267 | for (i = 0; i < nr_pools; i++) | |
268 | cleanup_pool(&mp[i]); | |
269 | ||
6548f47f JA |
270 | if (lock) |
271 | fio_mutex_remove(lock); | |
d24c33a4 JA |
272 | } |
273 | ||
89da54e8 | 274 | #ifdef SMALLOC_REDZONE |
ec996e9c | 275 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 276 | { |
ec996e9c | 277 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
55f6491d | 278 | |
ec996e9c JA |
279 | hdr->prered = SMALLOC_PRE_RED; |
280 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 281 | } |
55f6491d | 282 | |
ec996e9c JA |
283 | static void sfree_check_redzone(struct block_hdr *hdr) |
284 | { | |
ec996e9c JA |
285 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
286 | ||
287 | if (hdr->prered != SMALLOC_PRE_RED) { | |
55f6491d JA |
288 | fprintf(stderr, "smalloc pre redzone destroyed!\n"); |
289 | fprintf(stderr, " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 290 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
291 | assert(0); |
292 | } | |
293 | if (*postred != SMALLOC_POST_RED) { | |
294 | fprintf(stderr, "smalloc post redzone destroyed!\n"); | |
295 | fprintf(stderr, " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 296 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
297 | assert(0); |
298 | } | |
89da54e8 JA |
299 | } |
300 | #else | |
301 | static void fill_redzone(struct block_hdr *hdr) | |
302 | { | |
55f6491d JA |
303 | } |
304 | ||
89da54e8 JA |
305 | static void sfree_check_redzone(struct block_hdr *hdr) |
306 | { | |
307 | } | |
308 | #endif | |
309 | ||
d24c33a4 JA |
310 | static void sfree_pool(struct pool *pool, void *ptr) |
311 | { | |
ec996e9c | 312 | struct block_hdr *hdr; |
179446e0 | 313 | unsigned int i, idx; |
ec996e9c | 314 | unsigned long offset; |
d24c33a4 JA |
315 | |
316 | if (!ptr) | |
317 | return; | |
318 | ||
ec996e9c JA |
319 | ptr -= sizeof(*hdr); |
320 | hdr = ptr; | |
55f6491d | 321 | |
d24c33a4 JA |
322 | assert(ptr_valid(pool, ptr)); |
323 | ||
ec996e9c | 324 | sfree_check_redzone(hdr); |
d24c33a4 | 325 | |
ec996e9c JA |
326 | offset = ptr - pool->map; |
327 | i = offset / SMALLOC_BPL; | |
328 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 329 | |
ec996e9c | 330 | pool_lock(pool); |
dcb69098 | 331 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
332 | if (i < pool->next_non_full) |
333 | pool->next_non_full = i; | |
179446e0 | 334 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
335 | pool_unlock(pool); |
336 | } | |
337 | ||
338 | void sfree(void *ptr) | |
339 | { | |
340 | struct pool *pool = NULL; | |
341 | unsigned int i; | |
342 | ||
8e5732e5 JA |
343 | if (!ptr) |
344 | return; | |
345 | ||
65864cf7 | 346 | global_read_lock(); |
d24c33a4 JA |
347 | |
348 | for (i = 0; i < nr_pools; i++) { | |
349 | if (ptr_valid(&mp[i], ptr)) { | |
350 | pool = &mp[i]; | |
351 | break; | |
352 | } | |
353 | } | |
354 | ||
65864cf7 | 355 | global_read_unlock(); |
d24c33a4 JA |
356 | |
357 | assert(pool); | |
358 | sfree_pool(pool, ptr); | |
359 | } | |
360 | ||
55f6491d | 361 | static void *__smalloc_pool(struct pool *pool, unsigned int size) |
d24c33a4 | 362 | { |
ec996e9c JA |
363 | unsigned int nr_blocks; |
364 | unsigned int i; | |
365 | unsigned int offset; | |
366 | unsigned int last_idx; | |
367 | void *ret = NULL; | |
d24c33a4 | 368 | |
d24c33a4 | 369 | pool_lock(pool); |
179446e0 JA |
370 | |
371 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 372 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 373 | goto fail; |
5ec10eaa | 374 | |
ec996e9c JA |
375 | i = pool->next_non_full; |
376 | last_idx = 0; | |
377 | offset = -1U; | |
378 | while (i < pool->nr_blocks) { | |
379 | unsigned int idx; | |
d24c33a4 | 380 | |
ec996e9c JA |
381 | if (pool->bitmap[i] == -1U) { |
382 | i++; | |
383 | pool->next_non_full = i; | |
384 | last_idx = 0; | |
385 | continue; | |
386 | } | |
d24c33a4 | 387 | |
ec996e9c | 388 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 389 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
390 | idx += nr_blocks; |
391 | if (idx < SMALLOC_BPI) | |
392 | last_idx = idx; | |
393 | else { | |
394 | last_idx = 0; | |
395 | while (idx >= SMALLOC_BPI) { | |
396 | i++; | |
397 | idx -= SMALLOC_BPI; | |
398 | } | |
399 | } | |
400 | continue; | |
d24c33a4 | 401 | } |
dcb69098 | 402 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
403 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
404 | break; | |
405 | } | |
406 | ||
407 | if (i < pool->nr_blocks) { | |
408 | pool->free_blocks -= nr_blocks; | |
409 | ret = pool->map + offset; | |
d24c33a4 | 410 | } |
ec996e9c | 411 | fail: |
d24c33a4 | 412 | pool_unlock(pool); |
ec996e9c | 413 | return ret; |
d24c33a4 JA |
414 | } |
415 | ||
55f6491d JA |
416 | static void *smalloc_pool(struct pool *pool, unsigned int size) |
417 | { | |
89da54e8 | 418 | unsigned int alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
419 | void *ptr; |
420 | ||
ec996e9c JA |
421 | #ifdef SMALLOC_REDZONE |
422 | alloc_size += sizeof(unsigned int); | |
423 | #endif | |
424 | ||
425 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
426 | if (ptr) { |
427 | struct block_hdr *hdr = ptr; | |
55f6491d | 428 | |
89da54e8 JA |
429 | hdr->size = alloc_size; |
430 | fill_redzone(hdr); | |
55f6491d | 431 | |
89da54e8 JA |
432 | ptr += sizeof(*hdr); |
433 | memset(ptr, 0, size); | |
434 | } | |
ec996e9c | 435 | |
55f6491d | 436 | return ptr; |
55f6491d JA |
437 | } |
438 | ||
d24c33a4 JA |
439 | void *smalloc(unsigned int size) |
440 | { | |
441 | unsigned int i; | |
442 | ||
d1271dc1 | 443 | global_write_lock(); |
d24c33a4 JA |
444 | i = last_pool; |
445 | ||
446 | do { | |
447 | for (; i < nr_pools; i++) { | |
448 | void *ptr = smalloc_pool(&mp[i], size); | |
449 | ||
450 | if (ptr) { | |
451 | last_pool = i; | |
d1271dc1 | 452 | global_write_unlock(); |
d24c33a4 JA |
453 | return ptr; |
454 | } | |
455 | } | |
456 | if (last_pool) { | |
457 | last_pool = 0; | |
458 | continue; | |
459 | } | |
460 | ||
ec996e9c | 461 | if (nr_pools + 1 > MAX_POOLS) |
d24c33a4 JA |
462 | break; |
463 | else { | |
464 | i = nr_pools; | |
adf57099 | 465 | if (add_pool(&mp[nr_pools], size)) |
65864cf7 | 466 | goto out; |
d24c33a4 JA |
467 | } |
468 | } while (1); | |
469 | ||
65864cf7 | 470 | out: |
d1271dc1 | 471 | global_write_unlock(); |
d24c33a4 JA |
472 | return NULL; |
473 | } | |
474 | ||
475 | char *smalloc_strdup(const char *str) | |
476 | { | |
477 | char *ptr; | |
478 | ||
479 | ptr = smalloc(strlen(str) + 1); | |
480 | strcpy(ptr, str); | |
481 | return ptr; | |
482 | } |