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> | |
0ffccc21 BVA |
8 | #ifdef CONFIG_VALGRIND_DEV |
9 | #include <valgrind/valgrind.h> | |
10 | #else | |
11 | #define RUNNING_ON_VALGRIND 0 | |
12 | #define VALGRIND_MALLOCLIKE_BLOCK(addr, size, rzB, is_zeroed) do { } while (0) | |
13 | #define VALGRIND_FREELIKE_BLOCK(addr, rzB) do { } while (0) | |
14 | #endif | |
d24c33a4 | 15 | |
248c9436 | 16 | #include "fio.h" |
971caeb1 | 17 | #include "fio_sem.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 { |
971caeb1 | 42 | struct fio_sem *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 | ||
0ffccc21 BVA |
51 | #ifdef SMALLOC_REDZONE |
52 | #define REDZONE_SIZE sizeof(unsigned int) | |
53 | #else | |
54 | #define REDZONE_SIZE 0 | |
55 | #endif | |
56 | ||
ec996e9c | 57 | struct block_hdr { |
a3ebe7e0 | 58 | size_t size; |
ec996e9c JA |
59 | #ifdef SMALLOC_REDZONE |
60 | unsigned int prered; | |
61 | #endif | |
d24c33a4 JA |
62 | }; |
63 | ||
64 | static struct pool mp[MAX_POOLS]; | |
65 | static unsigned int nr_pools; | |
66 | static unsigned int last_pool; | |
d24c33a4 | 67 | |
d24c33a4 JA |
68 | static inline int ptr_valid(struct pool *pool, void *ptr) |
69 | { | |
dcb69098 | 70 | unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL; |
ec996e9c JA |
71 | |
72 | return (ptr >= pool->map) && (ptr < pool->map + pool_size); | |
d24c33a4 JA |
73 | } |
74 | ||
a3ebe7e0 | 75 | static inline size_t size_to_blocks(size_t size) |
808e9ea8 JA |
76 | { |
77 | return (size + SMALLOC_BPB - 1) / SMALLOC_BPB; | |
78 | } | |
79 | ||
dcb69098 | 80 | static int blocks_iter(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 81 | unsigned int idx, size_t nr_blocks, |
ec996e9c | 82 | int (*func)(unsigned int *map, unsigned int mask)) |
d24c33a4 | 83 | { |
dcb69098 | 84 | |
ec996e9c JA |
85 | while (nr_blocks) { |
86 | unsigned int this_blocks, mask; | |
dcb69098 JA |
87 | unsigned int *map; |
88 | ||
89 | if (pool_idx >= pool->nr_blocks) | |
90 | return 0; | |
91 | ||
92 | map = &pool->bitmap[pool_idx]; | |
ec996e9c JA |
93 | |
94 | this_blocks = nr_blocks; | |
95 | if (this_blocks + idx > SMALLOC_BPI) { | |
96 | this_blocks = SMALLOC_BPI - idx; | |
97 | idx = SMALLOC_BPI - this_blocks; | |
98 | } | |
99 | ||
100 | if (this_blocks == SMALLOC_BPI) | |
101 | mask = -1U; | |
102 | else | |
103 | mask = ((1U << this_blocks) - 1) << idx; | |
104 | ||
105 | if (!func(map, mask)) | |
106 | return 0; | |
107 | ||
108 | nr_blocks -= this_blocks; | |
109 | idx = 0; | |
dcb69098 | 110 | pool_idx++; |
ec996e9c JA |
111 | } |
112 | ||
113 | return 1; | |
d24c33a4 JA |
114 | } |
115 | ||
ec996e9c | 116 | static int mask_cmp(unsigned int *map, unsigned int mask) |
d24c33a4 | 117 | { |
ec996e9c | 118 | return !(*map & mask); |
d24c33a4 JA |
119 | } |
120 | ||
ec996e9c | 121 | static int mask_clear(unsigned int *map, unsigned int mask) |
d24c33a4 | 122 | { |
dcb69098 | 123 | assert((*map & mask) == mask); |
ec996e9c JA |
124 | *map &= ~mask; |
125 | return 1; | |
d24c33a4 JA |
126 | } |
127 | ||
ec996e9c | 128 | static int mask_set(unsigned int *map, unsigned int mask) |
d24c33a4 | 129 | { |
dcb69098 | 130 | assert(!(*map & mask)); |
ec996e9c JA |
131 | *map |= mask; |
132 | return 1; | |
d24c33a4 JA |
133 | } |
134 | ||
dcb69098 | 135 | static int blocks_free(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 136 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 137 | { |
dcb69098 | 138 | return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp); |
d24c33a4 JA |
139 | } |
140 | ||
dcb69098 | 141 | static void set_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_set); |
d24c33a4 JA |
145 | } |
146 | ||
dcb69098 | 147 | static void clear_blocks(struct pool *pool, unsigned int pool_idx, |
a3ebe7e0 | 148 | unsigned int idx, size_t nr_blocks) |
d24c33a4 | 149 | { |
dcb69098 | 150 | blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear); |
d24c33a4 JA |
151 | } |
152 | ||
ec996e9c JA |
153 | static int find_next_zero(int word, int start) |
154 | { | |
155 | assert(word != -1U); | |
271067a6 JH |
156 | word >>= start; |
157 | return ffz(word) + start; | |
d24c33a4 JA |
158 | } |
159 | ||
5f9454a2 | 160 | static bool add_pool(struct pool *pool, unsigned int alloc_size) |
d24c33a4 | 161 | { |
8d5844e9 | 162 | int bitmap_blocks; |
c8931876 | 163 | int mmap_flags; |
b8a6582e | 164 | void *ptr; |
ec996e9c | 165 | |
5f9454a2 JA |
166 | if (nr_pools == MAX_POOLS) |
167 | return false; | |
168 | ||
55f6491d | 169 | #ifdef SMALLOC_REDZONE |
ec996e9c | 170 | alloc_size += sizeof(unsigned int); |
55f6491d | 171 | #endif |
ec996e9c JA |
172 | alloc_size += sizeof(struct block_hdr); |
173 | if (alloc_size < INITIAL_SIZE) | |
174 | alloc_size = INITIAL_SIZE; | |
175 | ||
176 | /* round up to nearest full number of blocks */ | |
177 | alloc_size = (alloc_size + SMALLOC_BPL - 1) & ~(SMALLOC_BPL - 1); | |
178 | bitmap_blocks = alloc_size / SMALLOC_BPL; | |
179 | alloc_size += bitmap_blocks * sizeof(unsigned int); | |
180 | pool->mmap_size = alloc_size; | |
0b9d69ec | 181 | |
ec996e9c JA |
182 | pool->nr_blocks = bitmap_blocks; |
183 | pool->free_blocks = bitmap_blocks * SMALLOC_BPB; | |
adf57099 | 184 | |
c8931876 JA |
185 | mmap_flags = OS_MAP_ANON; |
186 | #ifdef CONFIG_ESX | |
187 | mmap_flags |= MAP_PRIVATE; | |
188 | #else | |
189 | mmap_flags |= MAP_SHARED; | |
190 | #endif | |
191 | ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, mmap_flags, -1, 0); | |
192 | ||
d24c33a4 | 193 | if (ptr == MAP_FAILED) |
8d5844e9 | 194 | goto out_fail; |
d24c33a4 | 195 | |
ec996e9c | 196 | pool->map = ptr; |
17f7fcd0 | 197 | pool->bitmap = (unsigned int *)((char *) ptr + (pool->nr_blocks * SMALLOC_BPL)); |
9c3e13e3 | 198 | memset(pool->bitmap, 0, bitmap_blocks * sizeof(unsigned int)); |
d24c33a4 | 199 | |
971caeb1 | 200 | pool->lock = fio_sem_init(FIO_SEM_UNLOCKED); |
d24c33a4 | 201 | if (!pool->lock) |
8d5844e9 | 202 | goto out_fail; |
d24c33a4 | 203 | |
d24c33a4 | 204 | nr_pools++; |
5f9454a2 | 205 | return true; |
8d5844e9 | 206 | out_fail: |
b0f0326a | 207 | log_err("smalloc: failed adding pool\n"); |
d24c33a4 | 208 | if (pool->map) |
ec996e9c | 209 | munmap(pool->map, pool->mmap_size); |
5f9454a2 | 210 | return false; |
d24c33a4 JA |
211 | } |
212 | ||
213 | void sinit(void) | |
214 | { | |
5f9454a2 JA |
215 | bool ret; |
216 | int i; | |
d24c33a4 | 217 | |
5f9454a2 JA |
218 | for (i = 0; i < INITIAL_POOLS; i++) { |
219 | ret = add_pool(&mp[nr_pools], smalloc_pool_size); | |
220 | if (!ret) | |
85492cb8 JA |
221 | break; |
222 | } | |
223 | ||
224 | /* | |
225 | * If we added at least one pool, we should be OK for most | |
226 | * cases. | |
227 | */ | |
228 | assert(i); | |
d24c33a4 JA |
229 | } |
230 | ||
231 | static void cleanup_pool(struct pool *pool) | |
232 | { | |
443bb114 JA |
233 | /* |
234 | * This will also remove the temporary file we used as a backing | |
235 | * store, it was already unlinked | |
236 | */ | |
ec996e9c | 237 | munmap(pool->map, pool->mmap_size); |
6548f47f JA |
238 | |
239 | if (pool->lock) | |
971caeb1 | 240 | fio_sem_remove(pool->lock); |
d24c33a4 JA |
241 | } |
242 | ||
243 | void scleanup(void) | |
244 | { | |
245 | unsigned int i; | |
246 | ||
247 | for (i = 0; i < nr_pools; i++) | |
248 | cleanup_pool(&mp[i]); | |
d24c33a4 JA |
249 | } |
250 | ||
89da54e8 | 251 | #ifdef SMALLOC_REDZONE |
cf98708d JA |
252 | static void *postred_ptr(struct block_hdr *hdr) |
253 | { | |
e43606c2 | 254 | uintptr_t ptr; |
cf98708d | 255 | |
e43606c2 | 256 | ptr = (uintptr_t) hdr + hdr->size - sizeof(unsigned int); |
248c9436 | 257 | ptr = (uintptr_t) PTR_ALIGN(ptr, int_mask); |
cf98708d JA |
258 | |
259 | return (void *) ptr; | |
260 | } | |
261 | ||
ec996e9c | 262 | static void fill_redzone(struct block_hdr *hdr) |
55f6491d | 263 | { |
cf98708d | 264 | unsigned int *postred = postred_ptr(hdr); |
55f6491d | 265 | |
0ffccc21 BVA |
266 | /* Let Valgrind fill the red zones. */ |
267 | if (RUNNING_ON_VALGRIND) | |
268 | return; | |
269 | ||
ec996e9c JA |
270 | hdr->prered = SMALLOC_PRE_RED; |
271 | *postred = SMALLOC_POST_RED; | |
ec996e9c | 272 | } |
55f6491d | 273 | |
ec996e9c JA |
274 | static void sfree_check_redzone(struct block_hdr *hdr) |
275 | { | |
cf98708d | 276 | unsigned int *postred = postred_ptr(hdr); |
ec996e9c | 277 | |
0ffccc21 BVA |
278 | /* Let Valgrind check the red zones. */ |
279 | if (RUNNING_ON_VALGRIND) | |
280 | return; | |
281 | ||
ec996e9c | 282 | if (hdr->prered != SMALLOC_PRE_RED) { |
b0f0326a JA |
283 | log_err("smalloc pre redzone destroyed!\n" |
284 | " ptr=%p, prered=%x, expected %x\n", | |
ec996e9c | 285 | hdr, hdr->prered, SMALLOC_PRE_RED); |
55f6491d JA |
286 | assert(0); |
287 | } | |
288 | if (*postred != SMALLOC_POST_RED) { | |
b0f0326a JA |
289 | log_err("smalloc post redzone destroyed!\n" |
290 | " ptr=%p, postred=%x, expected %x\n", | |
ec996e9c | 291 | hdr, *postred, SMALLOC_POST_RED); |
55f6491d JA |
292 | assert(0); |
293 | } | |
89da54e8 JA |
294 | } |
295 | #else | |
296 | static void fill_redzone(struct block_hdr *hdr) | |
297 | { | |
55f6491d JA |
298 | } |
299 | ||
89da54e8 JA |
300 | static void sfree_check_redzone(struct block_hdr *hdr) |
301 | { | |
302 | } | |
303 | #endif | |
304 | ||
d24c33a4 JA |
305 | static void sfree_pool(struct pool *pool, void *ptr) |
306 | { | |
ec996e9c | 307 | struct block_hdr *hdr; |
179446e0 | 308 | unsigned int i, idx; |
ec996e9c | 309 | unsigned long offset; |
d24c33a4 JA |
310 | |
311 | if (!ptr) | |
312 | return; | |
313 | ||
ec996e9c JA |
314 | ptr -= sizeof(*hdr); |
315 | hdr = ptr; | |
55f6491d | 316 | |
d24c33a4 JA |
317 | assert(ptr_valid(pool, ptr)); |
318 | ||
ec996e9c | 319 | sfree_check_redzone(hdr); |
d24c33a4 | 320 | |
ec996e9c JA |
321 | offset = ptr - pool->map; |
322 | i = offset / SMALLOC_BPL; | |
323 | idx = (offset % SMALLOC_BPL) / SMALLOC_BPB; | |
d24c33a4 | 324 | |
971caeb1 | 325 | fio_sem_down(pool->lock); |
dcb69098 | 326 | clear_blocks(pool, i, idx, size_to_blocks(hdr->size)); |
ec996e9c JA |
327 | if (i < pool->next_non_full) |
328 | pool->next_non_full = i; | |
179446e0 | 329 | pool->free_blocks += size_to_blocks(hdr->size); |
971caeb1 | 330 | fio_sem_up(pool->lock); |
d24c33a4 JA |
331 | } |
332 | ||
333 | void sfree(void *ptr) | |
334 | { | |
335 | struct pool *pool = NULL; | |
336 | unsigned int i; | |
337 | ||
8e5732e5 JA |
338 | if (!ptr) |
339 | return; | |
340 | ||
d24c33a4 JA |
341 | for (i = 0; i < nr_pools; i++) { |
342 | if (ptr_valid(&mp[i], ptr)) { | |
343 | pool = &mp[i]; | |
344 | break; | |
345 | } | |
346 | } | |
347 | ||
45a65144 | 348 | if (pool) { |
0ffccc21 | 349 | VALGRIND_FREELIKE_BLOCK(ptr, REDZONE_SIZE); |
45a65144 JA |
350 | sfree_pool(pool, ptr); |
351 | return; | |
352 | } | |
353 | ||
354 | log_err("smalloc: ptr %p not from smalloc pool\n", ptr); | |
d24c33a4 JA |
355 | } |
356 | ||
a3ebe7e0 | 357 | static void *__smalloc_pool(struct pool *pool, size_t size) |
d24c33a4 | 358 | { |
a3ebe7e0 | 359 | size_t nr_blocks; |
ec996e9c JA |
360 | unsigned int i; |
361 | unsigned int offset; | |
362 | unsigned int last_idx; | |
363 | void *ret = NULL; | |
d24c33a4 | 364 | |
971caeb1 | 365 | fio_sem_down(pool->lock); |
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: |
971caeb1 | 408 | fio_sem_up(pool->lock); |
ec996e9c | 409 | return ret; |
d24c33a4 JA |
410 | } |
411 | ||
a3ebe7e0 | 412 | static void *smalloc_pool(struct pool *pool, size_t size) |
55f6491d | 413 | { |
a3ebe7e0 | 414 | size_t alloc_size = size + sizeof(struct block_hdr); |
55f6491d JA |
415 | void *ptr; |
416 | ||
cf98708d | 417 | /* |
122426da JA |
418 | * Round to int alignment, so that the postred pointer will |
419 | * be naturally aligned as well. | |
cf98708d | 420 | */ |
ec996e9c | 421 | #ifdef SMALLOC_REDZONE |
122426da JA |
422 | alloc_size += sizeof(unsigned int); |
423 | alloc_size = (alloc_size + int_mask) & ~int_mask; | |
ec996e9c JA |
424 | #endif |
425 | ||
426 | ptr = __smalloc_pool(pool, alloc_size); | |
89da54e8 JA |
427 | if (ptr) { |
428 | struct block_hdr *hdr = ptr; | |
55f6491d | 429 | |
89da54e8 JA |
430 | hdr->size = alloc_size; |
431 | fill_redzone(hdr); | |
55f6491d | 432 | |
89da54e8 JA |
433 | ptr += sizeof(*hdr); |
434 | memset(ptr, 0, size); | |
435 | } | |
ec996e9c | 436 | |
55f6491d | 437 | return ptr; |
55f6491d JA |
438 | } |
439 | ||
0ffccc21 | 440 | static void *__smalloc(size_t size, bool is_zeroed) |
d24c33a4 | 441 | { |
85492cb8 | 442 | unsigned int i, end_pool; |
d24c33a4 | 443 | |
7982aa7d JA |
444 | if (size != (unsigned int) size) |
445 | return NULL; | |
446 | ||
d24c33a4 | 447 | i = last_pool; |
85492cb8 | 448 | end_pool = nr_pools; |
d24c33a4 JA |
449 | |
450 | do { | |
85492cb8 | 451 | for (; i < end_pool; i++) { |
d24c33a4 JA |
452 | void *ptr = smalloc_pool(&mp[i], size); |
453 | ||
454 | if (ptr) { | |
455 | last_pool = i; | |
0ffccc21 BVA |
456 | VALGRIND_MALLOCLIKE_BLOCK(ptr, size, |
457 | REDZONE_SIZE, | |
458 | is_zeroed); | |
d24c33a4 JA |
459 | return ptr; |
460 | } | |
461 | } | |
462 | if (last_pool) { | |
85492cb8 JA |
463 | end_pool = last_pool; |
464 | last_pool = i = 0; | |
d24c33a4 JA |
465 | continue; |
466 | } | |
467 | ||
85492cb8 | 468 | break; |
d24c33a4 JA |
469 | } while (1); |
470 | ||
81b3c86f JA |
471 | log_err("smalloc: OOM. Consider using --alloc-size to increase the " |
472 | "shared memory available.\n"); | |
d24c33a4 JA |
473 | return NULL; |
474 | } | |
475 | ||
0ffccc21 BVA |
476 | void *smalloc(size_t size) |
477 | { | |
478 | return __smalloc(size, false); | |
479 | } | |
480 | ||
544992f7 JA |
481 | void *scalloc(size_t nmemb, size_t size) |
482 | { | |
0ffccc21 | 483 | return __smalloc(nmemb * size, true); |
544992f7 JA |
484 | } |
485 | ||
d24c33a4 JA |
486 | char *smalloc_strdup(const char *str) |
487 | { | |
2894a2d4 | 488 | char *ptr = NULL; |
d24c33a4 JA |
489 | |
490 | ptr = smalloc(strlen(str) + 1); | |
2894a2d4 CE |
491 | if (ptr) |
492 | strcpy(ptr, str); | |
d24c33a4 JA |
493 | return ptr; |
494 | } |