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> | |
e43606c2 | 11 | #include <inttypes.h> |
d24c33a4 JA |
12 | #include <sys/types.h> |
13 | #include <limits.h> | |
3a8600b4 | 14 | #include <fcntl.h> |
d24c33a4 | 15 | |
6548f47f | 16 | #include "mutex.h" |
b3268b92 | 17 | #include "arch/arch.h" |
3a8600b4 | 18 | #include "os/os.h" |
10aa136b | 19 | #include "smalloc.h" |
b0f0326a | 20 | #include "log.h" |
d24c33a4 | 21 | |
55f6491d | 22 | #define SMALLOC_REDZONE /* define to detect memory corruption */ |
d24c33a4 | 23 | |
ec996e9c JA |
24 | #define SMALLOC_BPB 32 /* block size, bytes-per-bit in bitmap */ |
25 | #define SMALLOC_BPI (sizeof(unsigned int) * 8) | |
26 | #define SMALLOC_BPL (SMALLOC_BPB * SMALLOC_BPI) | |
27 | ||
23bd40f9 | 28 | #define INITIAL_SIZE 16*1024*1024 /* new pool size */ |
5f9454a2 JA |
29 | #define INITIAL_POOLS 8 /* maximum number of pools to setup */ |
30 | ||
31 | #define MAX_POOLS 16 | |
d24c33a4 | 32 | |
55f6491d JA |
33 | #define SMALLOC_PRE_RED 0xdeadbeefU |
34 | #define SMALLOC_POST_RED 0x5aa55aa5U | |
55f6491d | 35 | |
2b386d25 | 36 | unsigned int smalloc_pool_size = INITIAL_SIZE; |
aa1af5fd | 37 | #ifdef SMALLOC_REDZONE |
10aa136b | 38 | static const int int_mask = sizeof(int) - 1; |
aa1af5fd | 39 | #endif |
2b386d25 | 40 | |
d24c33a4 | 41 | struct pool { |
6548f47f | 42 | struct fio_mutex *lock; /* protects this pool */ |
d24c33a4 | 43 | void *map; /* map of blocks */ |
ec996e9c | 44 | unsigned int *bitmap; /* blocks free/busy map */ |
a3ebe7e0 JA |
45 | size_t free_blocks; /* free blocks */ |
46 | size_t nr_blocks; /* total blocks */ | |
47 | size_t next_non_full; | |
48 | size_t mmap_size; | |
ec996e9c JA |
49 | }; |
50 | ||
51 | struct block_hdr { | |
a3ebe7e0 | 52 | size_t size; |
ec996e9c JA |
53 | #ifdef SMALLOC_REDZONE |
54 | unsigned int prered; | |
55 | #endif | |
d24c33a4 JA |
56 | }; |
57 | ||
58 | static struct pool mp[MAX_POOLS]; | |
59 | static unsigned int nr_pools; | |
60 | static unsigned int last_pool; | |
d24c33a4 | 61 | |
d24c33a4 JA |
62 | static inline int ptr_valid(struct pool *pool, void *ptr) |
63 | { | |
dcb69098 | 64 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
65 | |
66 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
67 | } |
68 | ||
a3ebe7e0 | 69 | static inline size_t size_to_blocks(size_t size) |
808e9ea8 JA |
70 | { |
71 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
72 | } | |
73 | ||
dcb69098 | 74 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 75 | unsigned int idx, size_t nr_blocks, |
ec996e9c | 76 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 77 | { |
dcb69098 | 78 | |
ec996e9c JA |
79 | while (nr_blocks) { |
80 | unsigned int this_blocks, mask; | |
dcb69098 JA |
81 | unsigned int *map; |
82 | ||
83 | if (pool_idx >= pool->nr_blocks) | |
84 | return 0; | |
85 | ||
86 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
87 | |
88 | this_blocks = nr_blocks; | |
89 | if (this_blocks + idx > SMALLOC_BPI) { | |
90 | this_blocks = SMALLOC_BPI - idx; | |
91 | idx = SMALLOC_BPI - this_blocks; | |
92 | } | |
93 | ||
94 | if (this_blocks == SMALLOC_BPI) | |
95 | mask = -1U; | |
96 | else | |
97 | mask = ((1U << this_blocks) - 1) << idx; | |
98 | ||
99 | if (!func(map, mask)) | |
100 | return 0; | |
101 | ||
102 | nr_blocks -= this_blocks; | |
103 | idx = 0; | |
dcb69098 | 104 | pool_idx++; |
ec996e9c JA |
105 | } |
106 | ||
107 | return 1; | |
d24c33a4 JA |
108 | } |
109 | ||
ec996e9c | 110 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 111 | { |
ec996e9c | 112 | return !(*map & mask); |
d24c33a4 JA |
113 | } |
114 | ||
ec996e9c | 115 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 116 | { |
dcb69098 | 117 | assert((*map & mask) == mask); |
ec996e9c JA |
118 | *map &= ~mask; |
119 | return 1; | |
d24c33a4 JA |
120 | } |
121 | ||
ec996e9c | 122 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 123 | { |
dcb69098 | 124 | assert(!(*map & mask)); |
ec996e9c JA |
125 | *map |= mask; |
126 | return 1; | |
d24c33a4 JA |
127 | } |
128 | ||
dcb69098 | 129 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 130 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 131 | { |
dcb69098 | 132 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
133 | } |
134 | ||
dcb69098 | 135 | static void set_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 136 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 137 | { |
dcb69098 | 138 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set); |
d24c33a4 JA |
139 | } |
140 | ||
dcb69098 | 141 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 142 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 143 | { |
dcb69098 | 144 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
145 | } |
146 | ||
ec996e9c JA |
147 | static int find_next_zero(int word, int start) |
148 | { | |
149 | assert(word != -1U); | |
271067a6 JH |
150 | word >>= start; |
151 | return ffz(word) + start; | |
d24c33a4 JA |
152 | } |
153 | ||
5f9454a2 | 154 | static bool add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 155 | { |
8d5844e9 | 156 | int bitmap_blocks; |
c8931876 | 157 | int mmap_flags; |
b8a6582e | 158 | void *ptr; |
ec996e9c | 159 | |
5f9454a2 JA |
160 | if (nr_pools == MAX_POOLS) |
161 | return false; | |
162 | ||
55f6491d | 163 | #ifdef SMALLOC_REDZONE |
ec996e9c | 164 | alloc_size += sizeof(unsigned int); |
55f6491d | 165 | #endif |
ec996e9c JA |
166 | alloc_size += sizeof(struct block_hdr); |
167 | if (alloc_size < INITIAL_SIZE) | |
168 | alloc_size = INITIAL_SIZE; | |
169 | ||
170 | /* round up to nearest full number of blocks */ | |
171 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
172 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
173 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
174 | pool->mmap_size = alloc_size; | |
0b9d69ec | 175 | |
ec996e9c JA |
176 | pool->nr_blocks = bitmap_blocks; |
177 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 178 | |
c8931876 JA |
179 | mmap_flags = OS_MAP_ANON; |
180 | #ifdef CONFIG_ESX | |
181 | mmap_flags |= MAP_PRIVATE; | |
182 | #else | |
183 | mmap_flags |= MAP_SHARED; | |
184 | #endif | |
185 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, mmap_flags, -1, 0); | |
186 | ||
d24c33a4 | 187 | if (ptr == MAP_FAILED) |
8d5844e9 | 188 | goto out_fail; |
d24c33a4 | 189 | |
ec996e9c JA |
190 | pool->map = ptr; |
191 | pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL); | |
9c3e13e3 | 192 | memset(pool->bitmap, 0, bitmap_blocks * sizeof(unsigned int)); |
d24c33a4 | 193 | |
521da527 | 194 | pool->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED); |
d24c33a4 | 195 | if (!pool->lock) |
8d5844e9 | 196 | goto out_fail; |
d24c33a4 | 197 | |
d24c33a4 | 198 | nr_pools++; |
5f9454a2 | 199 | return true; |
8d5844e9 | 200 | out_fail: |
b0f0326a | 201 | log_err("smalloc: failed adding pool\n"); |
d24c33a4 | 202 | if (pool->map) |
ec996e9c | 203 | munmap(pool->map, pool->mmap_size); |
5f9454a2 | 204 | return false; |
d24c33a4 JA |
205 | } |
206 | ||
207 | void sinit(void) | |
208 | { | |
5f9454a2 JA |
209 | bool ret; |
210 | int i; | |
d24c33a4 | 211 | |
5f9454a2 JA |
212 | for (i = 0; i < INITIAL_POOLS; i++) { |
213 | ret = add_pool(&mp[nr_pools], smalloc_pool_size); | |
214 | if (!ret) | |
85492cb8 JA |
215 | break; |
216 | } | |
217 | ||
218 | /* | |
219 | * If we added at least one pool, we should be OK for most | |
220 | * cases. | |
221 | */ | |
222 | assert(i); | |
d24c33a4 JA |
223 | } |
224 | ||
225 | static void cleanup_pool(struct pool *pool) | |
226 | { | |
443bb114 JA |
227 | /* |
228 | * This will also remove the temporary file we used as a backing | |
229 | * store, it was already unlinked | |
230 | */ | |
ec996e9c | 231 | munmap(pool->map, pool->mmap_size); |
6548f47f JA |
232 | |
233 | if (pool->lock) | |
234 | fio_mutex_remove(pool->lock); | |
d24c33a4 JA |
235 | } |
236 | ||
237 | void scleanup(void) | |
238 | { | |
239 | unsigned int i; | |
240 | ||
241 | for (i = 0; i < nr_pools; i++) | |
242 | cleanup_pool(&mp[i]); | |
d24c33a4 JA |
243 | } |
244 | ||
89da54e8 | 245 | #ifdef SMALLOC_REDZONE |
cf98708d JA |
246 | static void *postred_ptr(struct block_hdr *hdr) |
247 | { | |
e43606c2 | 248 | uintptr_t ptr; |
cf98708d | 249 | |
e43606c2 | 250 | ptr = (uintptr_t) hdr + hdr->size - sizeof(unsigned int); |
cf98708d JA |
251 | ptr = (ptr + int_mask) & ~int_mask; |
252 | ||
253 | return (void *) ptr; | |
254 | } | |
255 | ||
ec996e9c | 256 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 257 | { |
cf98708d | 258 | unsigned int *postred = postred_ptr(hdr); |
55f6491d | 259 | |
ec996e9c JA |
260 | hdr->prered = SMALLOC_PRE_RED; |
261 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 262 | } |
55f6491d | 263 | |
ec996e9c JA |
264 | static void sfree_check_redzone(struct block_hdr *hdr) |
265 | { | |
cf98708d | 266 | unsigned int *postred = postred_ptr(hdr); |
ec996e9c JA |
267 | |
268 | if (hdr->prered != SMALLOC_PRE_RED) { | |
b0f0326a JA |
269 | log_err("smalloc pre redzone destroyed!\n" |
270 | " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 271 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
272 | assert(0); |
273 | } | |
274 | if (*postred != SMALLOC_POST_RED) { | |
b0f0326a JA |
275 | log_err("smalloc post redzone destroyed!\n" |
276 | " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 277 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
278 | assert(0); |
279 | } | |
89da54e8 JA |
280 | } |
281 | #else | |
282 | static void fill_redzone(struct block_hdr *hdr) | |
283 | { | |
55f6491d JA |
284 | } |
285 | ||
89da54e8 JA |
286 | static void sfree_check_redzone(struct block_hdr *hdr) |
287 | { | |
288 | } | |
289 | #endif | |
290 | ||
d24c33a4 JA |
291 | static void sfree_pool(struct pool *pool, void *ptr) |
292 | { | |
ec996e9c | 293 | struct block_hdr *hdr; |
179446e0 | 294 | unsigned int i, idx; |
ec996e9c | 295 | unsigned long offset; |
d24c33a4 JA |
296 | |
297 | if (!ptr) | |
298 | return; | |
299 | ||
ec996e9c JA |
300 | ptr -= sizeof(*hdr); |
301 | hdr = ptr; | |
55f6491d | 302 | |
d24c33a4 JA |
303 | assert(ptr_valid(pool, ptr)); |
304 | ||
ec996e9c | 305 | sfree_check_redzone(hdr); |
d24c33a4 | 306 | |
ec996e9c JA |
307 | offset = ptr - pool->map; |
308 | i = offset / SMALLOC_BPL; | |
309 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 310 | |
0127c57b | 311 | fio_mutex_down(pool->lock); |
dcb69098 | 312 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
313 | if (i < pool->next_non_full) |
314 | pool->next_non_full = i; | |
179446e0 | 315 | pool->free_blocks += size_to_blocks(hdr->size); |
0127c57b | 316 | fio_mutex_up(pool->lock); |
d24c33a4 JA |
317 | } |
318 | ||
319 | void sfree(void *ptr) | |
320 | { | |
321 | struct pool *pool = NULL; | |
322 | unsigned int i; | |
323 | ||
8e5732e5 JA |
324 | if (!ptr) |
325 | return; | |
326 | ||
d24c33a4 JA |
327 | for (i = 0; i < nr_pools; i++) { |
328 | if (ptr_valid(&mp[i], ptr)) { | |
329 | pool = &mp[i]; | |
330 | break; | |
331 | } | |
332 | } | |
333 | ||
45a65144 JA |
334 | if (pool) { |
335 | sfree_pool(pool, ptr); | |
336 | return; | |
337 | } | |
338 | ||
339 | log_err("smalloc: ptr %p not from smalloc pool\n", ptr); | |
d24c33a4 JA |
340 | } |
341 | ||
a3ebe7e0 | 342 | static void *__smalloc_pool(struct pool *pool, size_t size) |
d24c33a4 | 343 | { |
a3ebe7e0 | 344 | size_t nr_blocks; |
ec996e9c JA |
345 | unsigned int i; |
346 | unsigned int offset; | |
347 | unsigned int last_idx; | |
348 | void *ret = NULL; | |
d24c33a4 | 349 | |
0127c57b | 350 | fio_mutex_down(pool->lock); |
179446e0 JA |
351 | |
352 | nr_blocks = size_to_blocks(size); | |
ec996e9c | 353 | if (nr_blocks > pool->free_blocks) |
8e5732e5 | 354 | goto fail; |
5ec10eaa | 355 | |
ec996e9c JA |
356 | i = pool->next_non_full; |
357 | last_idx = 0; | |
358 | offset = -1U; | |
359 | while (i < pool->nr_blocks) { | |
360 | unsigned int idx; | |
d24c33a4 | 361 | |
ec996e9c JA |
362 | if (pool->bitmap[i] == -1U) { |
363 | i++; | |
364 | pool->next_non_full = i; | |
365 | last_idx = 0; | |
366 | continue; | |
367 | } | |
d24c33a4 | 368 | |
ec996e9c | 369 | idx = find_next_zero(pool->bitmap[i], last_idx); |
dcb69098 | 370 | if (!blocks_free(pool, i, idx, nr_blocks)) { |
ec996e9c JA |
371 | idx += nr_blocks; |
372 | if (idx < SMALLOC_BPI) | |
373 | last_idx = idx; | |
374 | else { | |
375 | last_idx = 0; | |
376 | while (idx >= SMALLOC_BPI) { | |
377 | i++; | |
378 | idx -= SMALLOC_BPI; | |
379 | } | |
380 | } | |
381 | continue; | |
d24c33a4 | 382 | } |
dcb69098 | 383 | set_blocks(pool, i, idx, nr_blocks); |
ec996e9c JA |
384 | offset = i * SMALLOC_BPL + idx * SMALLOC_BPB; |
385 | break; | |
386 | } | |
387 | ||
388 | if (i < pool->nr_blocks) { | |
389 | pool->free_blocks -= nr_blocks; | |
390 | ret = pool->map + offset; | |
d24c33a4 | 391 | } |
ec996e9c | 392 | fail: |
0127c57b | 393 | fio_mutex_up(pool->lock); |
ec996e9c | 394 | return ret; |
d24c33a4 JA |
395 | } |
396 | ||
a3ebe7e0 | 397 | static void *smalloc_pool(struct pool *pool, size_t size) |
55f6491d | 398 | { |
a3ebe7e0 | 399 | size_t alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
400 | void *ptr; |
401 | ||
cf98708d | 402 | /* |
122426da JA |
403 | * Round to int alignment, so that the postred pointer will |
404 | * be naturally aligned as well. | |
cf98708d | 405 | */ |
ec996e9c | 406 | #ifdef SMALLOC_REDZONE |
122426da JA |
407 | alloc_size += sizeof(unsigned int); |
408 | alloc_size = (alloc_size + int_mask) & ~int_mask; | |
ec996e9c JA |
409 | #endif |
410 | ||
411 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
412 | if (ptr) { |
413 | struct block_hdr *hdr = ptr; | |
55f6491d | 414 | |
89da54e8 JA |
415 | hdr->size = alloc_size; |
416 | fill_redzone(hdr); | |
55f6491d | 417 | |
89da54e8 JA |
418 | ptr += sizeof(*hdr); |
419 | memset(ptr, 0, size); | |
420 | } | |
ec996e9c | 421 | |
55f6491d | 422 | return ptr; |
55f6491d JA |
423 | } |
424 | ||
7982aa7d | 425 | void *smalloc(size_t size) |
d24c33a4 | 426 | { |
85492cb8 | 427 | unsigned int i, end_pool; |
d24c33a4 | 428 | |
7982aa7d JA |
429 | if (size != (unsigned int) size) |
430 | return NULL; | |
431 | ||
d24c33a4 | 432 | i = last_pool; |
85492cb8 | 433 | end_pool = nr_pools; |
d24c33a4 JA |
434 | |
435 | do { | |
85492cb8 | 436 | for (; i < end_pool; i++) { |
d24c33a4 JA |
437 | void *ptr = smalloc_pool(&mp[i], size); |
438 | ||
439 | if (ptr) { | |
440 | last_pool = i; | |
d24c33a4 JA |
441 | return ptr; |
442 | } | |
443 | } | |
444 | if (last_pool) { | |
85492cb8 JA |
445 | end_pool = last_pool; |
446 | last_pool = i = 0; | |
d24c33a4 JA |
447 | continue; |
448 | } | |
449 | ||
85492cb8 | 450 | break; |
d24c33a4 JA |
451 | } while (1); |
452 | ||
81b3c86f JA |
453 | log_err("smalloc: OOM. Consider using --alloc-size to increase the " |
454 | "shared memory available.\n"); | |
d24c33a4 JA |
455 | return NULL; |
456 | } | |
457 | ||
544992f7 JA |
458 | void *scalloc(size_t nmemb, size_t size) |
459 | { | |
a640ed36 | 460 | return smalloc(nmemb * size); |
544992f7 JA |
461 | } |
462 | ||
d24c33a4 JA |
463 | char *smalloc_strdup(const char *str) |
464 | { | |
2894a2d4 | 465 | char *ptr = NULL; |
d24c33a4 JA |
466 | |
467 | ptr = smalloc(strlen(str) + 1); | |
2894a2d4 CE |
468 | if (ptr) |
469 | strcpy(ptr, str); | |
d24c33a4 JA |
470 | return ptr; |
471 | } |