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 | ||
b3268b92 | 14 | #include "arch/arch.h" |
3f195910 | 15 | #include "spinlock.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 */ | |
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 | 31 | struct pool { |
3f195910 | 32 | struct fio_spinlock *lock; |
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 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; | |
3f195910 | 53 | static struct fio_spinlock *slock; |
d24c33a4 | 54 | |
d24c33a4 JA |
55 | static inline void pool_lock(struct pool *pool) |
56 | { | |
3f195910 | 57 | fio_spin_lock(pool->lock); |
d24c33a4 JA |
58 | } |
59 | ||
60 | static inline void pool_unlock(struct pool *pool) | |
61 | { | |
3f195910 | 62 | fio_spin_unlock(pool->lock); |
d24c33a4 JA |
63 | } |
64 | ||
65864cf7 | 65 | static inline void global_read_lock(void) |
d24c33a4 | 66 | { |
3f195910 | 67 | fio_spin_lock(slock); |
d24c33a4 JA |
68 | } |
69 | ||
65864cf7 | 70 | static inline void global_read_unlock(void) |
d24c33a4 | 71 | { |
3f195910 | 72 | fio_spin_unlock(slock); |
65864cf7 JA |
73 | } |
74 | ||
75 | static inline void global_write_lock(void) | |
76 | { | |
3f195910 | 77 | fio_spin_lock(slock); |
65864cf7 JA |
78 | } |
79 | ||
80 | static inline void global_write_unlock(void) | |
81 | { | |
3f195910 | 82 | fio_spin_unlock(slock); |
d24c33a4 JA |
83 | } |
84 | ||
d24c33a4 JA |
85 | static inline int ptr_valid(struct pool *pool, void *ptr) |
86 | { | |
dcb69098 | 87 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
88 | |
89 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
90 | } |
91 | ||
808e9ea8 JA |
92 | static inline unsigned int size_to_blocks(unsigned int size) |
93 | { | |
94 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
95 | } | |
96 | ||
dcb69098 JA |
97 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
98 | unsigned int idx, unsigned int nr_blocks, | |
ec996e9c | 99 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 100 | { |
dcb69098 | 101 | |
ec996e9c JA |
102 | while (nr_blocks) { |
103 | unsigned int this_blocks, mask; | |
dcb69098 JA |
104 | unsigned int *map; |
105 | ||
106 | if (pool_idx >= pool->nr_blocks) | |
107 | return 0; | |
108 | ||
109 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
110 | |
111 | this_blocks = nr_blocks; | |
112 | if (this_blocks + idx > SMALLOC_BPI) { | |
113 | this_blocks = SMALLOC_BPI - idx; | |
114 | idx = SMALLOC_BPI - this_blocks; | |
115 | } | |
116 | ||
117 | if (this_blocks == SMALLOC_BPI) | |
118 | mask = -1U; | |
119 | else | |
120 | mask = ((1U << this_blocks) - 1) << idx; | |
121 | ||
122 | if (!func(map, mask)) | |
123 | return 0; | |
124 | ||
125 | nr_blocks -= this_blocks; | |
126 | idx = 0; | |
dcb69098 | 127 | pool_idx++; |
ec996e9c JA |
128 | } |
129 | ||
130 | return 1; | |
d24c33a4 JA |
131 | } |
132 | ||
ec996e9c | 133 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 134 | { |
ec996e9c | 135 | return !(*map & mask); |
d24c33a4 JA |
136 | } |
137 | ||
ec996e9c | 138 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 139 | { |
dcb69098 | 140 | assert((*map & mask) == mask); |
ec996e9c JA |
141 | *map &= ~mask; |
142 | return 1; | |
d24c33a4 JA |
143 | } |
144 | ||
ec996e9c | 145 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 146 | { |
dcb69098 | 147 | assert(!(*map & mask)); |
ec996e9c JA |
148 | *map |= mask; |
149 | return 1; | |
d24c33a4 JA |
150 | } |
151 | ||
dcb69098 JA |
152 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
153 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 154 | { |
dcb69098 | 155 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
156 | } |
157 | ||
dcb69098 JA |
158 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
159 | unsigned int idx, unsigned int nr_blocks) | |
d24c33a4 | 160 | { |
dcb69098 | 161 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
162 | } |
163 | ||
dcb69098 JA |
164 | static void clear_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_clear); |
d24c33a4 JA |
168 | } |
169 | ||
ec996e9c JA |
170 | static int find_next_zero(int word, int start) |
171 | { | |
172 | assert(word != -1U); | |
173 | word >>= (start + 1); | |
b3268b92 | 174 | return ffz(word) + start + 1; |
d24c33a4 JA |
175 | } |
176 | ||
adf57099 | 177 | static int add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 178 | { |
d24c33a4 | 179 | void *ptr; |
ec996e9c JA |
180 | int fd, bitmap_blocks; |
181 | ||
d24c33a4 JA |
182 | strcpy(pool->file, "/tmp/.fio_smalloc.XXXXXX"); |
183 | fd = mkstemp(pool->file); | |
184 | if (fd < 0) | |
185 | goto out_close; | |
186 | ||
55f6491d | 187 | #ifdef SMALLOC_REDZONE |
ec996e9c | 188 | alloc_size += sizeof(unsigned int); |
55f6491d | 189 | #endif |
ec996e9c JA |
190 | alloc_size += sizeof(struct block_hdr); |
191 | if (alloc_size < INITIAL_SIZE) | |
192 | alloc_size = INITIAL_SIZE; | |
193 | ||
194 | /* round up to nearest full number of blocks */ | |
195 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
196 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
197 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
198 | pool->mmap_size = alloc_size; | |
55f6491d | 199 | |
ec996e9c JA |
200 | pool->nr_blocks = bitmap_blocks; |
201 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 202 | |
ec996e9c | 203 | if (ftruncate(fd, alloc_size) < 0) |
d24c33a4 JA |
204 | goto out_unlink; |
205 | ||
ec996e9c | 206 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0); |
d24c33a4 JA |
207 | if (ptr == MAP_FAILED) |
208 | goto out_unlink; | |
209 | ||
ec996e9c JA |
210 | memset(ptr, 0, alloc_size); |
211 | pool->map = ptr; | |
212 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
d24c33a4 | 213 | |
3f195910 | 214 | pool->lock = fio_spinlock_init(); |
d24c33a4 JA |
215 | if (!pool->lock) |
216 | goto out_unlink; | |
d24c33a4 JA |
217 | |
218 | pool->fd = fd; | |
219 | ||
65864cf7 | 220 | global_write_lock(); |
d24c33a4 | 221 | nr_pools++; |
65864cf7 | 222 | global_write_unlock(); |
d24c33a4 JA |
223 | return 0; |
224 | out_unlink: | |
ec996e9c | 225 | fprintf(stderr, "smalloc: failed adding pool\n"); |
d24c33a4 | 226 | if (pool->map) |
ec996e9c | 227 | munmap(pool->map, pool->mmap_size); |
d24c33a4 JA |
228 | unlink(pool->file); |
229 | out_close: | |
230 | if (fd >= 0) | |
231 | close(fd); | |
232 | return 1; | |
233 | } | |
234 | ||
235 | void sinit(void) | |
236 | { | |
4d4e80f2 | 237 | int ret; |
d24c33a4 | 238 | |
3f195910 JA |
239 | slock = fio_spinlock_init(); |
240 | assert(slock); | |
adf57099 | 241 | ret = add_pool(&mp[0], INITIAL_SIZE); |
d24c33a4 JA |
242 | assert(!ret); |
243 | } | |
244 | ||
245 | static void cleanup_pool(struct pool *pool) | |
246 | { | |
247 | unlink(pool->file); | |
248 | close(pool->fd); | |
ec996e9c | 249 | munmap(pool->map, pool->mmap_size); |
3f195910 | 250 | fio_spinlock_remove(pool->lock); |
d24c33a4 JA |
251 | } |
252 | ||
253 | void scleanup(void) | |
254 | { | |
255 | unsigned int i; | |
256 | ||
257 | for (i = 0; i < nr_pools; i++) | |
258 | cleanup_pool(&mp[i]); | |
259 | ||
3f195910 | 260 | fio_spinlock_remove(slock); |
d24c33a4 JA |
261 | } |
262 | ||
89da54e8 | 263 | #ifdef SMALLOC_REDZONE |
ec996e9c | 264 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 265 | { |
ec996e9c | 266 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
55f6491d | 267 | |
ec996e9c JA |
268 | hdr->prered = SMALLOC_PRE_RED; |
269 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 270 | } |
55f6491d | 271 | |
ec996e9c JA |
272 | static void sfree_check_redzone(struct block_hdr *hdr) |
273 | { | |
ec996e9c JA |
274 | unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int); |
275 | ||
276 | if (hdr->prered != SMALLOC_PRE_RED) { | |
55f6491d JA |
277 | fprintf(stderr, "smalloc pre redzone destroyed!\n"); |
278 | fprintf(stderr, " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 279 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
280 | assert(0); |
281 | } | |
282 | if (*postred != SMALLOC_POST_RED) { | |
283 | fprintf(stderr, "smalloc post redzone destroyed!\n"); | |
284 | fprintf(stderr, " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 285 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
286 | assert(0); |
287 | } | |
89da54e8 JA |
288 | } |
289 | #else | |
290 | static void fill_redzone(struct block_hdr *hdr) | |
291 | { | |
55f6491d JA |
292 | } |
293 | ||
89da54e8 JA |
294 | static void sfree_check_redzone(struct block_hdr *hdr) |
295 | { | |
296 | } | |
297 | #endif | |
298 | ||
d24c33a4 JA |
299 | static void sfree_pool(struct pool *pool, void *ptr) |
300 | { | |
ec996e9c | 301 | struct block_hdr *hdr; |
179446e0 | 302 | unsigned int i, idx; |
ec996e9c | 303 | unsigned long offset; |
d24c33a4 JA |
304 | |
305 | if (!ptr) | |
306 | return; | |
307 | ||
ec996e9c JA |
308 | ptr -= sizeof(*hdr); |
309 | hdr = ptr; | |
55f6491d | 310 | |
d24c33a4 JA |
311 | assert(ptr_valid(pool, ptr)); |
312 | ||
ec996e9c | 313 | sfree_check_redzone(hdr); |
d24c33a4 | 314 | |
ec996e9c JA |
315 | offset = ptr - pool->map; |
316 | i = offset / SMALLOC_BPL; | |
317 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 318 | |
ec996e9c | 319 | pool_lock(pool); |
dcb69098 | 320 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
321 | if (i < pool->next_non_full) |
322 | pool->next_non_full = i; | |
179446e0 | 323 | pool->free_blocks += size_to_blocks(hdr->size); |
d24c33a4 JA |
324 | pool_unlock(pool); |
325 | } | |
326 | ||
327 | void sfree(void *ptr) | |
328 | { | |
329 | struct pool *pool = NULL; | |
330 | unsigned int i; | |
331 | ||
8e5732e5 JA |
332 | if (!ptr) |
333 | return; | |
334 | ||
65864cf7 | 335 | global_read_lock(); |
d24c33a4 JA |
336 | |
337 | for (i = 0; i < nr_pools; i++) { | |
338 | if (ptr_valid(&mp[i], ptr)) { | |
339 | pool = &mp[i]; | |
340 | break; | |
341 | } | |
342 | } | |
343 | ||
65864cf7 | 344 | global_read_unlock(); |
d24c33a4 JA |
345 | |
346 | assert(pool); | |
347 | sfree_pool(pool, ptr); | |
348 | } | |
349 | ||
55f6491d | 350 | static void *__smalloc_pool(struct pool *pool, unsigned int size) |
d24c33a4 | 351 | { |
ec996e9c JA |
352 | unsigned int nr_blocks; |
353 | unsigned int i; | |
354 | unsigned int offset; | |
355 | unsigned int last_idx; | |
356 | void *ret = NULL; | |
d24c33a4 | 357 | |
d24c33a4 | 358 | pool_lock(pool); |
179446e0 JA |
359 | |
360 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 361 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 362 | goto fail; |
5ec10eaa | 363 | |
ec996e9c JA |
364 | i = pool->next_non_full; |
365 | last_idx = 0; | |
366 | offset = -1U; | |
367 | while (i < pool->nr_blocks) { | |
368 | unsigned int idx; | |
d24c33a4 | 369 | |
ec996e9c JA |
370 | if (pool->bitmap[i] == -1U) { |
371 | i++; | |
372 | pool->next_non_full = i; | |
373 | last_idx = 0; | |
374 | continue; | |
375 | } | |
d24c33a4 | 376 | |
ec996e9c | 377 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 378 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
379 | idx += nr_blocks; |
380 | if (idx < SMALLOC_BPI) | |
381 | last_idx = idx; | |
382 | else { | |
383 | last_idx = 0; | |
384 | while (idx >= SMALLOC_BPI) { | |
385 | i++; | |
386 | idx -= SMALLOC_BPI; | |
387 | } | |
388 | } | |
389 | continue; | |
d24c33a4 | 390 | } |
dcb69098 | 391 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
392 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
393 | break; | |
394 | } | |
395 | ||
396 | if (i < pool->nr_blocks) { | |
397 | pool->free_blocks -= nr_blocks; | |
398 | ret = pool->map + offset; | |
d24c33a4 | 399 | } |
ec996e9c | 400 | fail: |
d24c33a4 | 401 | pool_unlock(pool); |
ec996e9c | 402 | return ret; |
d24c33a4 JA |
403 | } |
404 | ||
55f6491d JA |
405 | static void *smalloc_pool(struct pool *pool, unsigned int size) |
406 | { | |
89da54e8 | 407 | unsigned int alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
408 | void *ptr; |
409 | ||
ec996e9c JA |
410 | #ifdef SMALLOC_REDZONE |
411 | alloc_size += sizeof(unsigned int); | |
412 | #endif | |
413 | ||
414 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
415 | if (ptr) { |
416 | struct block_hdr *hdr = ptr; | |
55f6491d | 417 | |
89da54e8 JA |
418 | hdr->size = alloc_size; |
419 | fill_redzone(hdr); | |
55f6491d | 420 | |
89da54e8 JA |
421 | ptr += sizeof(*hdr); |
422 | memset(ptr, 0, size); | |
423 | } | |
ec996e9c | 424 | |
55f6491d | 425 | return ptr; |
55f6491d JA |
426 | } |
427 | ||
d24c33a4 JA |
428 | void *smalloc(unsigned int size) |
429 | { | |
430 | unsigned int i; | |
431 | ||
65864cf7 | 432 | global_read_lock(); |
d24c33a4 JA |
433 | i = last_pool; |
434 | ||
435 | do { | |
436 | for (; i < nr_pools; i++) { | |
437 | void *ptr = smalloc_pool(&mp[i], size); | |
438 | ||
439 | if (ptr) { | |
440 | last_pool = i; | |
65864cf7 | 441 | global_read_unlock(); |
d24c33a4 JA |
442 | return ptr; |
443 | } | |
444 | } | |
445 | if (last_pool) { | |
446 | last_pool = 0; | |
447 | continue; | |
448 | } | |
449 | ||
ec996e9c | 450 | if (nr_pools + 1 > MAX_POOLS) |
d24c33a4 JA |
451 | break; |
452 | else { | |
453 | i = nr_pools; | |
65864cf7 | 454 | global_read_unlock(); |
adf57099 | 455 | if (add_pool(&mp[nr_pools], size)) |
65864cf7 JA |
456 | goto out; |
457 | global_read_lock(); | |
d24c33a4 JA |
458 | } |
459 | } while (1); | |
460 | ||
65864cf7 JA |
461 | global_read_unlock(); |
462 | out: | |
d24c33a4 JA |
463 | return NULL; |
464 | } | |
465 | ||
466 | char *smalloc_strdup(const char *str) | |
467 | { | |
468 | char *ptr; | |
469 | ||
470 | ptr = smalloc(strlen(str) + 1); | |
471 | strcpy(ptr, str); | |
472 | return ptr; | |
473 | } |