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> | |
d24c33a4 JA |
6 | #include <assert.h> |
7 | #include <string.h> | |
d24c33a4 | 8 | |
248c9436 | 9 | #include "fio.h" |
971caeb1 | 10 | #include "fio_sem.h" |
3a8600b4 | 11 | #include "os/os.h" |
10aa136b | 12 | #include "smalloc.h" |
b0f0326a | 13 | #include "log.h" |
d24c33a4 | 14 | |
55f6491d | 15 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
d24c33a4 | 16 | |
ec996e9c JA |
17 | #define SMALLOC_BPB 32 /* block size, bytes-per-bit in bitmap */ |
18 | #define SMALLOC_BPI (sizeof(unsigned int) * 8) | |
19 | #define SMALLOC_BPL (SMALLOC_BPB * SMALLOC_BPI) | |
20 | ||
23bd40f9 | 21 | #define INITIAL_SIZE 16*1024*1024 /* new pool size */ |
5f9454a2 JA |
22 | #define INITIAL_POOLS 8 /* maximum number of pools to setup */ |
23 | ||
24 | #define MAX_POOLS 16 | |
d24c33a4 | 25 | |
55f6491d JA |
26 | #define SMALLOC_PRE_RED 0xdeadbeefU |
27 | #define SMALLOC_POST_RED 0x5aa55aa5U | |
55f6491d | 28 | |
2b386d25 | 29 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
aa1af5fd | 30 | #ifdef SMALLOC_REDZONE |
10aa136b | 31 | static const int int_mask = sizeof(int) - 1; |
aa1af5fd | 32 | #endif |
2b386d25 | 33 | |
d24c33a4 | 34 | struct pool { |
971caeb1 | 35 | struct fio_sem *lock; /* protects this pool */ |
d24c33a4 | 36 | void *map; /* map of blocks */ |
ec996e9c | 37 | unsigned int *bitmap; /* blocks free/busy map */ |
a3ebe7e0 JA |
38 | size_t free_blocks; /* free blocks */ |
39 | size_t nr_blocks; /* total blocks */ | |
40 | size_t next_non_full; | |
41 | size_t mmap_size; | |
ec996e9c JA |
42 | }; |
43 | ||
44 | struct block_hdr { | |
a3ebe7e0 | 45 | size_t size; |
ec996e9c JA |
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; | |
d24c33a4 | 54 | |
d24c33a4 JA |
55 | static inline int ptr_valid(struct pool *pool, void *ptr) |
56 | { | |
dcb69098 | 57 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
58 | |
59 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
60 | } |
61 | ||
a3ebe7e0 | 62 | static inline size_t size_to_blocks(size_t size) |
808e9ea8 JA |
63 | { |
64 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
65 | } | |
66 | ||
dcb69098 | 67 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 68 | unsigned int idx, size_t nr_blocks, |
ec996e9c | 69 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 70 | { |
dcb69098 | 71 | |
ec996e9c JA |
72 | while (nr_blocks) { |
73 | unsigned int this_blocks, mask; | |
dcb69098 JA |
74 | unsigned int *map; |
75 | ||
76 | if (pool_idx >= pool->nr_blocks) | |
77 | return 0; | |
78 | ||
79 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
80 | |
81 | this_blocks = nr_blocks; | |
82 | if (this_blocks + idx > SMALLOC_BPI) { | |
83 | this_blocks = SMALLOC_BPI - idx; | |
84 | idx = SMALLOC_BPI - this_blocks; | |
85 | } | |
86 | ||
87 | if (this_blocks == SMALLOC_BPI) | |
88 | mask = -1U; | |
89 | else | |
90 | mask = ((1U << this_blocks) - 1) << idx; | |
91 | ||
92 | if (!func(map, mask)) | |
93 | return 0; | |
94 | ||
95 | nr_blocks -= this_blocks; | |
96 | idx = 0; | |
dcb69098 | 97 | pool_idx++; |
ec996e9c JA |
98 | } |
99 | ||
100 | return 1; | |
d24c33a4 JA |
101 | } |
102 | ||
ec996e9c | 103 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 104 | { |
ec996e9c | 105 | return !(*map & mask); |
d24c33a4 JA |
106 | } |
107 | ||
ec996e9c | 108 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 109 | { |
dcb69098 | 110 | assert((*map & mask) == mask); |
ec996e9c JA |
111 | *map &= ~mask; |
112 | return 1; | |
d24c33a4 JA |
113 | } |
114 | ||
ec996e9c | 115 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 116 | { |
dcb69098 | 117 | assert(!(*map & mask)); |
ec996e9c JA |
118 | *map |= mask; |
119 | return 1; | |
d24c33a4 JA |
120 | } |
121 | ||
dcb69098 | 122 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 123 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 124 | { |
dcb69098 | 125 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
126 | } |
127 | ||
dcb69098 | 128 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 129 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 130 | { |
dcb69098 | 131 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
132 | } |
133 | ||
dcb69098 | 134 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 135 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 136 | { |
dcb69098 | 137 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
138 | } |
139 | ||
ec996e9c JA |
140 | static int find_next_zero(int word, int start) |
141 | { | |
142 | assert(word != -1U); | |
271067a6 JH |
143 | word >>= start; |
144 | return ffz(word) + start; | |
d24c33a4 JA |
145 | } |
146 | ||
5f9454a2 | 147 | static bool add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 148 | { |
8d5844e9 | 149 | int bitmap_blocks; |
c8931876 | 150 | int mmap_flags; |
b8a6582e | 151 | void *ptr; |
ec996e9c | 152 | |
5f9454a2 JA |
153 | if (nr_pools == MAX_POOLS) |
154 | return false; | |
155 | ||
55f6491d | 156 | #ifdef SMALLOC_REDZONE |
ec996e9c | 157 | alloc_size += sizeof(unsigned int); |
55f6491d | 158 | #endif |
ec996e9c JA |
159 | alloc_size += sizeof(struct block_hdr); |
160 | if (alloc_size < INITIAL_SIZE) | |
161 | alloc_size = INITIAL_SIZE; | |
162 | ||
163 | /* round up to nearest full number of blocks */ | |
164 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
165 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
166 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
167 | pool->mmap_size = alloc_size; | |
0b9d69ec | 168 | |
ec996e9c JA |
169 | pool->nr_blocks = bitmap_blocks; |
170 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 171 | |
c8931876 JA |
172 | mmap_flags = OS_MAP_ANON; |
173 | #ifdef CONFIG_ESX | |
174 | mmap_flags |= MAP_PRIVATE; | |
175 | #else | |
176 | mmap_flags |= MAP_SHARED; | |
177 | #endif | |
178 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, mmap_flags, -1, 0); | |
179 | ||
d24c33a4 | 180 | if (ptr == MAP_FAILED) |
8d5844e9 | 181 | goto out_fail; |
d24c33a4 | 182 | |
ec996e9c | 183 | pool->map = ptr; |
17f7fcd0 | 184 | pool->bitmap = (unsigned int *)((char *) ptr + (pool->nr_blocks * SMALLOC_BPL)); |
9c3e13e3 | 185 | memset(pool->bitmap, 0, bitmap_blocks * sizeof(unsigned int)); |
d24c33a4 | 186 | |
971caeb1 | 187 | pool->lock = fio_sem_init(FIO_SEM_UNLOCKED); |
d24c33a4 | 188 | if (!pool->lock) |
8d5844e9 | 189 | goto out_fail; |
d24c33a4 | 190 | |
d24c33a4 | 191 | nr_pools++; |
5f9454a2 | 192 | return true; |
8d5844e9 | 193 | out_fail: |
b0f0326a | 194 | log_err("smalloc: failed adding pool\n"); |
d24c33a4 | 195 | if (pool->map) |
ec996e9c | 196 | munmap(pool->map, pool->mmap_size); |
5f9454a2 | 197 | return false; |
d24c33a4 JA |
198 | } |
199 | ||
200 | void sinit(void) | |
201 | { | |
5f9454a2 JA |
202 | bool ret; |
203 | int i; | |
d24c33a4 | 204 | |
5f9454a2 JA |
205 | for (i = 0; i < INITIAL_POOLS; i++) { |
206 | ret = add_pool(&mp[nr_pools], smalloc_pool_size); | |
207 | if (!ret) | |
85492cb8 JA |
208 | break; |
209 | } | |
210 | ||
211 | /* | |
212 | * If we added at least one pool, we should be OK for most | |
213 | * cases. | |
214 | */ | |
215 | assert(i); | |
d24c33a4 JA |
216 | } |
217 | ||
218 | static void cleanup_pool(struct pool *pool) | |
219 | { | |
443bb114 JA |
220 | /* |
221 | * This will also remove the temporary file we used as a backing | |
222 | * store, it was already unlinked | |
223 | */ | |
ec996e9c | 224 | munmap(pool->map, pool->mmap_size); |
6548f47f JA |
225 | |
226 | if (pool->lock) | |
971caeb1 | 227 | fio_sem_remove(pool->lock); |
d24c33a4 JA |
228 | } |
229 | ||
230 | void scleanup(void) | |
231 | { | |
232 | unsigned int i; | |
233 | ||
234 | for (i = 0; i < nr_pools; i++) | |
235 | cleanup_pool(&mp[i]); | |
d24c33a4 JA |
236 | } |
237 | ||
89da54e8 | 238 | #ifdef SMALLOC_REDZONE |
cf98708d JA |
239 | static void *postred_ptr(struct block_hdr *hdr) |
240 | { | |
e43606c2 | 241 | uintptr_t ptr; |
cf98708d | 242 | |
e43606c2 | 243 | ptr = (uintptr_t) hdr + hdr->size - sizeof(unsigned int); |
248c9436 | 244 | ptr = (uintptr_t) PTR_ALIGN(ptr, int_mask); |
cf98708d JA |
245 | |
246 | return (void *) ptr; | |
247 | } | |
248 | ||
ec996e9c | 249 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 250 | { |
cf98708d | 251 | unsigned int *postred = postred_ptr(hdr); |
55f6491d | 252 | |
ec996e9c JA |
253 | hdr->prered = SMALLOC_PRE_RED; |
254 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 255 | } |
55f6491d | 256 | |
ec996e9c JA |
257 | static void sfree_check_redzone(struct block_hdr *hdr) |
258 | { | |
cf98708d | 259 | unsigned int *postred = postred_ptr(hdr); |
ec996e9c JA |
260 | |
261 | if (hdr->prered != SMALLOC_PRE_RED) { | |
b0f0326a JA |
262 | log_err("smalloc pre redzone destroyed!\n" |
263 | " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 264 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
265 | assert(0); |
266 | } | |
267 | if (*postred != SMALLOC_POST_RED) { | |
b0f0326a JA |
268 | log_err("smalloc post redzone destroyed!\n" |
269 | " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 270 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
271 | assert(0); |
272 | } | |
89da54e8 JA |
273 | } |
274 | #else | |
275 | static void fill_redzone(struct block_hdr *hdr) | |
276 | { | |
55f6491d JA |
277 | } |
278 | ||
89da54e8 JA |
279 | static void sfree_check_redzone(struct block_hdr *hdr) |
280 | { | |
281 | } | |
282 | #endif | |
283 | ||
d24c33a4 JA |
284 | static void sfree_pool(struct pool *pool, void *ptr) |
285 | { | |
ec996e9c | 286 | struct block_hdr *hdr; |
179446e0 | 287 | unsigned int i, idx; |
ec996e9c | 288 | unsigned long offset; |
d24c33a4 JA |
289 | |
290 | if (!ptr) | |
291 | return; | |
292 | ||
ec996e9c JA |
293 | ptr -= sizeof(*hdr); |
294 | hdr = ptr; | |
55f6491d | 295 | |
d24c33a4 JA |
296 | assert(ptr_valid(pool, ptr)); |
297 | ||
ec996e9c | 298 | sfree_check_redzone(hdr); |
d24c33a4 | 299 | |
ec996e9c JA |
300 | offset = ptr - pool->map; |
301 | i = offset / SMALLOC_BPL; | |
302 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 303 | |
971caeb1 | 304 | fio_sem_down(pool->lock); |
dcb69098 | 305 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
306 | if (i < pool->next_non_full) |
307 | pool->next_non_full = i; | |
179446e0 | 308 | pool->free_blocks += size_to_blocks(hdr->size); |
971caeb1 | 309 | fio_sem_up(pool->lock); |
d24c33a4 JA |
310 | } |
311 | ||
312 | void sfree(void *ptr) | |
313 | { | |
314 | struct pool *pool = NULL; | |
315 | unsigned int i; | |
316 | ||
8e5732e5 JA |
317 | if (!ptr) |
318 | return; | |
319 | ||
d24c33a4 JA |
320 | for (i = 0; i < nr_pools; i++) { |
321 | if (ptr_valid(&mp[i], ptr)) { | |
322 | pool = &mp[i]; | |
323 | break; | |
324 | } | |
325 | } | |
326 | ||
45a65144 JA |
327 | if (pool) { |
328 | sfree_pool(pool, ptr); | |
329 | return; | |
330 | } | |
331 | ||
332 | log_err("smalloc: ptr %p not from smalloc pool\n", ptr); | |
d24c33a4 JA |
333 | } |
334 | ||
a3ebe7e0 | 335 | static void *__smalloc_pool(struct pool *pool, size_t size) |
d24c33a4 | 336 | { |
a3ebe7e0 | 337 | size_t nr_blocks; |
ec996e9c JA |
338 | unsigned int i; |
339 | unsigned int offset; | |
340 | unsigned int last_idx; | |
341 | void *ret = NULL; | |
d24c33a4 | 342 | |
971caeb1 | 343 | fio_sem_down(pool->lock); |
179446e0 JA |
344 | |
345 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 346 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 347 | goto fail; |
5ec10eaa | 348 | |
ec996e9c JA |
349 | i = pool->next_non_full; |
350 | last_idx = 0; | |
351 | offset = -1U; | |
352 | while (i < pool->nr_blocks) { | |
353 | unsigned int idx; | |
d24c33a4 | 354 | |
ec996e9c JA |
355 | if (pool->bitmap[i] == -1U) { |
356 | i++; | |
357 | pool->next_non_full = i; | |
358 | last_idx = 0; | |
359 | continue; | |
360 | } | |
d24c33a4 | 361 | |
ec996e9c | 362 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 363 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
364 | idx += nr_blocks; |
365 | if (idx < SMALLOC_BPI) | |
366 | last_idx = idx; | |
367 | else { | |
368 | last_idx = 0; | |
369 | while (idx >= SMALLOC_BPI) { | |
370 | i++; | |
371 | idx -= SMALLOC_BPI; | |
372 | } | |
373 | } | |
374 | continue; | |
d24c33a4 | 375 | } |
dcb69098 | 376 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
377 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
378 | break; | |
379 | } | |
380 | ||
381 | if (i < pool->nr_blocks) { | |
382 | pool->free_blocks -= nr_blocks; | |
383 | ret = pool->map + offset; | |
d24c33a4 | 384 | } |
ec996e9c | 385 | fail: |
971caeb1 | 386 | fio_sem_up(pool->lock); |
ec996e9c | 387 | return ret; |
d24c33a4 JA |
388 | } |
389 | ||
a3ebe7e0 | 390 | static void *smalloc_pool(struct pool *pool, size_t size) |
55f6491d | 391 | { |
a3ebe7e0 | 392 | size_t alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
393 | void *ptr; |
394 | ||
cf98708d | 395 | /* |
122426da JA |
396 | * Round to int alignment, so that the postred pointer will |
397 | * be naturally aligned as well. | |
cf98708d | 398 | */ |
ec996e9c | 399 | #ifdef SMALLOC_REDZONE |
122426da JA |
400 | alloc_size += sizeof(unsigned int); |
401 | alloc_size = (alloc_size + int_mask) & ~int_mask; | |
ec996e9c JA |
402 | #endif |
403 | ||
404 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
405 | if (ptr) { |
406 | struct block_hdr *hdr = ptr; | |
55f6491d | 407 | |
89da54e8 JA |
408 | hdr->size = alloc_size; |
409 | fill_redzone(hdr); | |
55f6491d | 410 | |
89da54e8 JA |
411 | ptr += sizeof(*hdr); |
412 | memset(ptr, 0, size); | |
413 | } | |
ec996e9c | 414 | |
55f6491d | 415 | return ptr; |
55f6491d JA |
416 | } |
417 | ||
5457259f | 418 | void *smalloc(size_t size) |
d24c33a4 | 419 | { |
85492cb8 | 420 | unsigned int i, end_pool; |
d24c33a4 | 421 | |
7982aa7d JA |
422 | if (size != (unsigned int) size) |
423 | return NULL; | |
424 | ||
d24c33a4 | 425 | i = last_pool; |
85492cb8 | 426 | end_pool = nr_pools; |
d24c33a4 JA |
427 | |
428 | do { | |
85492cb8 | 429 | for (; i < end_pool; i++) { |
d24c33a4 JA |
430 | void *ptr = smalloc_pool(&mp[i], size); |
431 | ||
432 | if (ptr) { | |
433 | last_pool = i; | |
d24c33a4 JA |
434 | return ptr; |
435 | } | |
436 | } | |
437 | if (last_pool) { | |
85492cb8 JA |
438 | end_pool = last_pool; |
439 | last_pool = i = 0; | |
d24c33a4 JA |
440 | continue; |
441 | } | |
442 | ||
85492cb8 | 443 | break; |
d24c33a4 JA |
444 | } while (1); |
445 | ||
81b3c86f JA |
446 | log_err("smalloc: OOM. Consider using --alloc-size to increase the " |
447 | "shared memory available.\n"); | |
d24c33a4 JA |
448 | return NULL; |
449 | } | |
450 | ||
544992f7 JA |
451 | void *scalloc(size_t nmemb, size_t size) |
452 | { | |
5457259f | 453 | return smalloc(nmemb * size); |
544992f7 JA |
454 | } |
455 | ||
d24c33a4 JA |
456 | char *smalloc_strdup(const char *str) |
457 | { | |
2894a2d4 | 458 | char *ptr = NULL; |
d24c33a4 JA |
459 | |
460 | ptr = smalloc(strlen(str) + 1); | |
2894a2d4 CE |
461 | if (ptr) |
462 | strcpy(ptr, str); | |
d24c33a4 JA |
463 | return ptr; |
464 | } |