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 | |
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 */ | |
68857686 | 24 | #define MAX_POOLS 128 /* 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 | 31 | struct pool { |
6548f47f | 32 | struct fio_mutex *lock; /* protects this pool */ |
d24c33a4 | 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 | 38 | int fd; /* memory backing fd */ |
ec996e9c JA |
39 | unsigned int mmap_size; |
40 | }; | |
41 | ||
42 | struct block_hdr { | |
43 | unsigned int size; | |
44 | #ifdef SMALLOC_REDZONE | |
45 | unsigned int prered; | |
46 | #endif | |
d24c33a4 JA |
47 | }; |
48 | ||
49 | static struct pool mp[MAX_POOLS]; | |
50 | static unsigned int nr_pools; | |
51 | static unsigned int last_pool; | |
6548f47f | 52 | static struct fio_mutex *lock; |
d24c33a4 | 53 | |
d24c33a4 JA |
54 | static inline void pool_lock(struct pool *pool) |
55 | { | |
2e3e31e3 | 56 | fio_mutex_down(pool->lock); |
d24c33a4 JA |
57 | } |
58 | ||
59 | static inline void pool_unlock(struct pool *pool) | |
60 | { | |
2e3e31e3 | 61 | fio_mutex_up(pool->lock); |
d24c33a4 JA |
62 | } |
63 | ||
65864cf7 | 64 | static inline void global_read_lock(void) |
d24c33a4 | 65 | { |
2e3e31e3 | 66 | fio_mutex_down_read(lock); |
d24c33a4 JA |
67 | } |
68 | ||
65864cf7 | 69 | static inline void global_read_unlock(void) |
d24c33a4 | 70 | { |
2e3e31e3 | 71 | fio_mutex_up_read(lock); |
65864cf7 JA |
72 | } |
73 | ||
74 | static inline void global_write_lock(void) | |
75 | { | |
2e3e31e3 | 76 | fio_mutex_down_write(lock); |
65864cf7 JA |
77 | } |
78 | ||
79 | static inline void global_write_unlock(void) | |
80 | { | |
2e3e31e3 | 81 | fio_mutex_up_write(lock); |
d24c33a4 JA |
82 | } |
83 | ||
d24c33a4 JA |
84 | static inline int ptr_valid(struct pool *pool, void *ptr) |
85 | { | |
dcb69098 | 86 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
87 | |
88 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
89 | } |
90 | ||
808e9ea8 JA |
91 | static inline unsigned int size_to_blocks(unsigned int size) |
92 | { | |
93 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
94 | } | |
95 | ||
dcb69098 JA |
96 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
97 | unsigned int idx, unsigned int nr_blocks, | |
ec996e9c | 98 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 99 | { |
dcb69098 | 100 | |
ec996e9c JA |
101 | while (nr_blocks) { |
102 | unsigned int this_blocks, mask; | |
dcb69098 JA |
103 | unsigned int *map; |
104 | ||
105 | if (pool_idx >= pool->nr_blocks) | |
106 | return 0; | |
107 | ||
108 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
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; | |
dcb69098 | 126 | pool_idx++; |
ec996e9c JA |
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 | { |
dcb69098 | 139 | assert((*map & mask) == mask); |
ec996e9c JA |
140 | *map &= ~mask; |
141 | return 1; | |
d24c33a4 JA |
142 | } |
143 | ||
ec996e9c | 144 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 145 | { |
dcb69098 | 146 | assert(!(*map & mask)); |
ec996e9c JA |
147 | *map |= mask; |
148 | return 1; | |
d24c33a4 JA |
149 | } |
150 | ||
dcb69098 JA |
151 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
152 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 153 | { |
dcb69098 | 154 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
155 | } |
156 | ||
dcb69098 JA |
157 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
158 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 159 | { |
dcb69098 | 160 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
161 | } |
162 | ||
dcb69098 JA |
163 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
164 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 165 | { |
dcb69098 | 166 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
167 | } |
168 | ||
ec996e9c JA |
169 | static int find_next_zero(int word, int start) |
170 | { | |
171 | assert(word != -1U); | |
172 | word >>= (start + 1); | |
b3268b92 | 173 | return ffz(word) + start + 1; |
d24c33a4 JA |
174 | } |
175 | ||
adf57099 | 176 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 177 | { |
ec996e9c | 178 | int fd, bitmap_blocks; |
b8a6582e JA |
179 | char file[] = "/tmp/.fio_smalloc.XXXXXX"; |
180 | void *ptr; | |
ec996e9c | 181 | |
b8a6582e | 182 | fd = mkstemp(file); |
d24c33a4 JA |
183 | if (fd < 0) |
184 | goto out_close; | |
185 | ||
55f6491d | 186 | #ifdef SMALLOC_REDZONE |
ec996e9c | 187 | alloc_size += sizeof(unsigned int); |
55f6491d | 188 | #endif |
ec996e9c JA |
189 | alloc_size += sizeof(struct block_hdr); |
190 | if (alloc_size < INITIAL_SIZE) | |
191 | alloc_size = INITIAL_SIZE; | |
192 | ||
193 | /* round up to nearest full number of blocks */ | |
194 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
195 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
196 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
197 | pool->mmap_size = alloc_size; | |
55f6491d | 198 | |
ec996e9c JA |
199 | pool->nr_blocks = bitmap_blocks; |
200 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 201 | |
ec996e9c | 202 | if (ftruncate(fd, alloc_size) < 0) |
d24c33a4 JA |
203 | goto out_unlink; |
204 | ||
ec996e9c | 205 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); |
d24c33a4 JA |
206 | if (ptr == MAP_FAILED) |
207 | goto out_unlink; | |
208 | ||
ec996e9c JA |
209 | memset(ptr, 0, alloc_size); |
210 | pool->map = ptr; | |
211 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 | 212 | |
6548f47f | 213 | pool->lock = fio_mutex_init(1); |
d24c33a4 JA |
214 | if (!pool->lock) |
215 | goto out_unlink; | |
d24c33a4 | 216 | |
443bb114 JA |
217 | /* |
218 | * Unlink pool file now. It wont get deleted until the fd is closed, | |
219 | * which happens both for cleanup or unexpected quit. This way we | |
220 | * don't leave temp files around in case of a crash. | |
221 | */ | |
b8a6582e | 222 | unlink(file); |
d24c33a4 JA |
223 | pool->fd = fd; |
224 | ||
d24c33a4 JA |
225 | nr_pools++; |
226 | return 0; | |
227 | out_unlink: | |
ec996e9c | 228 | fprintf(stderr, "smalloc: failed adding pool\n"); |
d24c33a4 | 229 | if (pool->map) |
ec996e9c | 230 | munmap(pool->map, pool->mmap_size); |
b8a6582e | 231 | unlink(file); |
d24c33a4 | 232 | out_close: |
b8a6582e | 233 | close(fd); |
d24c33a4 JA |
234 | return 1; |
235 | } | |
236 | ||
237 | void sinit(void) | |
238 | { | |
4d4e80f2 | 239 | int ret; |
d24c33a4 | 240 | |
6548f47f | 241 | lock = fio_mutex_rw_init(); |
adf57099 | 242 | ret = add_pool(&mp[0], INITIAL_SIZE); |
d24c33a4 JA |
243 | assert(!ret); |
244 | } | |
245 | ||
246 | static void cleanup_pool(struct pool *pool) | |
247 | { | |
443bb114 JA |
248 | /* |
249 | * This will also remove the temporary file we used as a backing | |
250 | * store, it was already unlinked | |
251 | */ | |
d24c33a4 | 252 | close(pool->fd); |
ec996e9c | 253 | munmap(pool->map, pool->mmap_size); |
6548f47f JA |
254 | |
255 | if (pool->lock) | |
256 | fio_mutex_remove(pool->lock); | |
d24c33a4 JA |
257 | } |
258 | ||
259 | void scleanup(void) | |
260 | { | |
261 | unsigned int i; | |
262 | ||
263 | for (i = 0; i < nr_pools; i++) | |
264 | cleanup_pool(&mp[i]); | |
265 | ||
6548f47f JA |
266 | if (lock) |
267 | fio_mutex_remove(lock); | |
d24c33a4 JA |
268 | } |
269 | ||
89da54e8 | 270 | #ifdef SMALLOC_REDZONE |
ec996e9c | 271 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 272 | { |
ec996e9c | 273 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
55f6491d | 274 | |
ec996e9c JA |
275 | hdr->prered = SMALLOC_PRE_RED; |
276 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 277 | } |
55f6491d | 278 | |
ec996e9c JA |
279 | static void sfree_check_redzone(struct block_hdr *hdr) |
280 | { | |
ec996e9c JA |
281 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
282 | ||
283 | if (hdr->prered != SMALLOC_PRE_RED) { | |
55f6491d JA |
284 | fprintf(stderr, "smalloc pre redzone destroyed!\n"); |
285 | fprintf(stderr, " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 286 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
287 | assert(0); |
288 | } | |
289 | if (*postred != SMALLOC_POST_RED) { | |
290 | fprintf(stderr, "smalloc post redzone destroyed!\n"); | |
291 | fprintf(stderr, " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 292 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
293 | assert(0); |
294 | } | |
89da54e8 JA |
295 | } |
296 | #else | |
297 | static void fill_redzone(struct block_hdr *hdr) | |
298 | { | |
55f6491d JA |
299 | } |
300 | ||
89da54e8 JA |
301 | static void sfree_check_redzone(struct block_hdr *hdr) |
302 | { | |
303 | } | |
304 | #endif | |
305 | ||
d24c33a4 JA |
306 | static void sfree_pool(struct pool *pool, void *ptr) |
307 | { | |
ec996e9c | 308 | struct block_hdr *hdr; |
179446e0 | 309 | unsigned int i, idx; |
ec996e9c | 310 | unsigned long offset; |
d24c33a4 JA |
311 | |
312 | if (!ptr) | |
313 | return; | |
314 | ||
ec996e9c JA |
315 | ptr -= sizeof(*hdr); |
316 | hdr = ptr; | |
55f6491d | 317 | |
d24c33a4 JA |
318 | assert(ptr_valid(pool, ptr)); |
319 | ||
ec996e9c | 320 | sfree_check_redzone(hdr); |
d24c33a4 | 321 | |
ec996e9c JA |
322 | offset = ptr - pool->map; |
323 | i = offset / SMALLOC_BPL; | |
324 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 325 | |
ec996e9c | 326 | pool_lock(pool); |
dcb69098 | 327 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
328 | if (i < pool->next_non_full) |
329 | pool->next_non_full = i; | |
179446e0 | 330 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
331 | pool_unlock(pool); |
332 | } | |
333 | ||
334 | void sfree(void *ptr) | |
335 | { | |
336 | struct pool *pool = NULL; | |
337 | unsigned int i; | |
338 | ||
8e5732e5 JA |
339 | if (!ptr) |
340 | return; | |
341 | ||
65864cf7 | 342 | global_read_lock(); |
d24c33a4 JA |
343 | |
344 | for (i = 0; i < nr_pools; i++) { | |
345 | if (ptr_valid(&mp[i], ptr)) { | |
346 | pool = &mp[i]; | |
347 | break; | |
348 | } | |
349 | } | |
350 | ||
65864cf7 | 351 | global_read_unlock(); |
d24c33a4 JA |
352 | |
353 | assert(pool); | |
354 | sfree_pool(pool, ptr); | |
355 | } | |
356 | ||
55f6491d | 357 | static void *__smalloc_pool(struct pool *pool, unsigned int size) |
d24c33a4 | 358 | { |
ec996e9c JA |
359 | unsigned int nr_blocks; |
360 | unsigned int i; | |
361 | unsigned int offset; | |
362 | unsigned int last_idx; | |
363 | void *ret = NULL; | |
d24c33a4 | 364 | |
d24c33a4 | 365 | pool_lock(pool); |
179446e0 JA |
366 | |
367 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 368 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 369 | goto fail; |
5ec10eaa | 370 | |
ec996e9c JA |
371 | i = pool->next_non_full; |
372 | last_idx = 0; | |
373 | offset = -1U; | |
374 | while (i < pool->nr_blocks) { | |
375 | unsigned int idx; | |
d24c33a4 | 376 | |
ec996e9c JA |
377 | if (pool->bitmap[i] == -1U) { |
378 | i++; | |
379 | pool->next_non_full = i; | |
380 | last_idx = 0; | |
381 | continue; | |
382 | } | |
d24c33a4 | 383 | |
ec996e9c | 384 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 385 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
386 | idx += nr_blocks; |
387 | if (idx < SMALLOC_BPI) | |
388 | last_idx = idx; | |
389 | else { | |
390 | last_idx = 0; | |
391 | while (idx >= SMALLOC_BPI) { | |
392 | i++; | |
393 | idx -= SMALLOC_BPI; | |
394 | } | |
395 | } | |
396 | continue; | |
d24c33a4 | 397 | } |
dcb69098 | 398 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
399 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
400 | break; | |
401 | } | |
402 | ||
403 | if (i < pool->nr_blocks) { | |
404 | pool->free_blocks -= nr_blocks; | |
405 | ret = pool->map + offset; | |
d24c33a4 | 406 | } |
ec996e9c | 407 | fail: |
d24c33a4 | 408 | pool_unlock(pool); |
ec996e9c | 409 | return ret; |
d24c33a4 JA |
410 | } |
411 | ||
55f6491d JA |
412 | static void *smalloc_pool(struct pool *pool, unsigned int size) |
413 | { | |
89da54e8 | 414 | unsigned int alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
415 | void *ptr; |
416 | ||
ec996e9c JA |
417 | #ifdef SMALLOC_REDZONE |
418 | alloc_size += sizeof(unsigned int); | |
419 | #endif | |
420 | ||
421 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
422 | if (ptr) { |
423 | struct block_hdr *hdr = ptr; | |
55f6491d | 424 | |
89da54e8 JA |
425 | hdr->size = alloc_size; |
426 | fill_redzone(hdr); | |
55f6491d | 427 | |
89da54e8 JA |
428 | ptr += sizeof(*hdr); |
429 | memset(ptr, 0, size); | |
430 | } | |
ec996e9c | 431 | |
55f6491d | 432 | return ptr; |
55f6491d JA |
433 | } |
434 | ||
d24c33a4 JA |
435 | void *smalloc(unsigned int size) |
436 | { | |
437 | unsigned int i; | |
438 | ||
d1271dc1 | 439 | global_write_lock(); |
d24c33a4 JA |
440 | i = last_pool; |
441 | ||
442 | do { | |
443 | for (; i < nr_pools; i++) { | |
444 | void *ptr = smalloc_pool(&mp[i], size); | |
445 | ||
446 | if (ptr) { | |
447 | last_pool = i; | |
d1271dc1 | 448 | global_write_unlock(); |
d24c33a4 JA |
449 | return ptr; |
450 | } | |
451 | } | |
452 | if (last_pool) { | |
453 | last_pool = 0; | |
454 | continue; | |
455 | } | |
456 | ||
ec996e9c | 457 | if (nr_pools + 1 > MAX_POOLS) |
d24c33a4 JA |
458 | break; |
459 | else { | |
460 | i = nr_pools; | |
adf57099 | 461 | if (add_pool(&mp[nr_pools], size)) |
65864cf7 | 462 | goto out; |
d24c33a4 JA |
463 | } |
464 | } while (1); | |
465 | ||
65864cf7 | 466 | out: |
d1271dc1 | 467 | global_write_unlock(); |
d24c33a4 JA |
468 | return NULL; |
469 | } | |
470 | ||
471 | char *smalloc_strdup(const char *str) | |
472 | { | |
473 | char *ptr; | |
474 | ||
475 | ptr = smalloc(strlen(str) + 1); | |
476 | strcpy(ptr, str); | |
477 | return ptr; | |
478 | } |