Return EILSEQ on verify errors instead of EIO
[fio.git] / smalloc.c
... / ...
CommitLineData
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
14#include "mutex.h"
15#include "arch/arch.h"
16
17#define SMALLOC_REDZONE /* define to detect memory corruption */
18
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 */
24#define MAX_POOLS 128 /* maximum number of pools to setup */
25
26#define SMALLOC_PRE_RED 0xdeadbeefU
27#define SMALLOC_POST_RED 0x5aa55aa5U
28
29unsigned int smalloc_pool_size = INITIAL_SIZE;
30
31struct pool {
32 struct fio_mutex *lock; /* protects this pool */
33 void *map; /* map of blocks */
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;
38 int fd; /* memory backing fd */
39 unsigned int mmap_size;
40};
41
42struct block_hdr {
43 unsigned int size;
44#ifdef SMALLOC_REDZONE
45 unsigned int prered;
46#endif
47};
48
49static struct pool mp[MAX_POOLS];
50static unsigned int nr_pools;
51static unsigned int last_pool;
52static struct fio_mutex *lock;
53
54static inline void pool_lock(struct pool *pool)
55{
56 fio_mutex_down(pool->lock);
57}
58
59static inline void pool_unlock(struct pool *pool)
60{
61 fio_mutex_up(pool->lock);
62}
63
64static inline void global_read_lock(void)
65{
66 fio_mutex_down_read(lock);
67}
68
69static inline void global_read_unlock(void)
70{
71 fio_mutex_up_read(lock);
72}
73
74static inline void global_write_lock(void)
75{
76 fio_mutex_down_write(lock);
77}
78
79static inline void global_write_unlock(void)
80{
81 fio_mutex_up_write(lock);
82}
83
84static inline int ptr_valid(struct pool *pool, void *ptr)
85{
86 unsigned int pool_size = pool->nr_blocks * SMALLOC_BPL;
87
88 return (ptr >= pool->map) && (ptr < pool->map + pool_size);
89}
90
91static inline unsigned int size_to_blocks(unsigned int size)
92{
93 return (size + SMALLOC_BPB - 1) / SMALLOC_BPB;
94}
95
96static int blocks_iter(struct pool *pool, unsigned int pool_idx,
97 unsigned int idx, unsigned int nr_blocks,
98 int (*func)(unsigned int *map, unsigned int mask))
99{
100
101 while (nr_blocks) {
102 unsigned int this_blocks, mask;
103 unsigned int *map;
104
105 if (pool_idx >= pool->nr_blocks)
106 return 0;
107
108 map = &pool->bitmap[pool_idx];
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;
126 pool_idx++;
127 }
128
129 return 1;
130}
131
132static int mask_cmp(unsigned int *map, unsigned int mask)
133{
134 return !(*map & mask);
135}
136
137static int mask_clear(unsigned int *map, unsigned int mask)
138{
139 assert((*map & mask) == mask);
140 *map &= ~mask;
141 return 1;
142}
143
144static int mask_set(unsigned int *map, unsigned int mask)
145{
146 assert(!(*map & mask));
147 *map |= mask;
148 return 1;
149}
150
151static int blocks_free(struct pool *pool, unsigned int pool_idx,
152 unsigned int idx, unsigned int nr_blocks)
153{
154 return blocks_iter(pool, pool_idx, idx, nr_blocks, mask_cmp);
155}
156
157static void set_blocks(struct pool *pool, unsigned int pool_idx,
158 unsigned int idx, unsigned int nr_blocks)
159{
160 blocks_iter(pool, pool_idx, idx, nr_blocks, mask_set);
161}
162
163static void clear_blocks(struct pool *pool, unsigned int pool_idx,
164 unsigned int idx, unsigned int nr_blocks)
165{
166 blocks_iter(pool, pool_idx, idx, nr_blocks, mask_clear);
167}
168
169static int find_next_zero(int word, int start)
170{
171 assert(word != -1U);
172 word >>= (start + 1);
173 return ffz(word) + start + 1;
174}
175
176static int add_pool(struct pool *pool, unsigned int alloc_size)
177{
178 int fd, bitmap_blocks;
179 char file[] = "/tmp/.fio_smalloc.XXXXXX";
180 void *ptr;
181
182 fd = mkstemp(file);
183 if (fd < 0)
184 goto out_close;
185
186#ifdef SMALLOC_REDZONE
187 alloc_size += sizeof(unsigned int);
188#endif
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;
198
199 pool->nr_blocks = bitmap_blocks;
200 pool->free_blocks = bitmap_blocks * SMALLOC_BPB;
201
202 if (ftruncate(fd, alloc_size) < 0)
203 goto out_unlink;
204
205 ptr = mmap(NULL, alloc_size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
206 if (ptr == MAP_FAILED)
207 goto out_unlink;
208
209 memset(ptr, 0, alloc_size);
210 pool->map = ptr;
211 pool->bitmap = (void *) ptr + (pool->nr_blocks * SMALLOC_BPL);
212
213 pool->lock = fio_mutex_init(1);
214 if (!pool->lock)
215 goto out_unlink;
216
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 */
222 unlink(file);
223 pool->fd = fd;
224
225 nr_pools++;
226 return 0;
227out_unlink:
228 fprintf(stderr, "smalloc: failed adding pool\n");
229 if (pool->map)
230 munmap(pool->map, pool->mmap_size);
231 unlink(file);
232out_close:
233 close(fd);
234 return 1;
235}
236
237void sinit(void)
238{
239 int ret;
240
241 lock = fio_mutex_rw_init();
242 ret = add_pool(&mp[0], INITIAL_SIZE);
243 assert(!ret);
244}
245
246static void cleanup_pool(struct pool *pool)
247{
248 /*
249 * This will also remove the temporary file we used as a backing
250 * store, it was already unlinked
251 */
252 close(pool->fd);
253 munmap(pool->map, pool->mmap_size);
254
255 if (pool->lock)
256 fio_mutex_remove(pool->lock);
257}
258
259void scleanup(void)
260{
261 unsigned int i;
262
263 for (i = 0; i < nr_pools; i++)
264 cleanup_pool(&mp[i]);
265
266 if (lock)
267 fio_mutex_remove(lock);
268}
269
270#ifdef SMALLOC_REDZONE
271static void fill_redzone(struct block_hdr *hdr)
272{
273 unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int);
274
275 hdr->prered = SMALLOC_PRE_RED;
276 *postred = SMALLOC_POST_RED;
277}
278
279static void sfree_check_redzone(struct block_hdr *hdr)
280{
281 unsigned int *postred = (void *) hdr + hdr->size - sizeof(unsigned int);
282
283 if (hdr->prered != SMALLOC_PRE_RED) {
284 fprintf(stderr, "smalloc pre redzone destroyed!\n");
285 fprintf(stderr, " ptr=%p, prered=%x, expected %x\n",
286 hdr, hdr->prered, SMALLOC_PRE_RED);
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",
292 hdr, *postred, SMALLOC_POST_RED);
293 assert(0);
294 }
295}
296#else
297static void fill_redzone(struct block_hdr *hdr)
298{
299}
300
301static void sfree_check_redzone(struct block_hdr *hdr)
302{
303}
304#endif
305
306static void sfree_pool(struct pool *pool, void *ptr)
307{
308 struct block_hdr *hdr;
309 unsigned int i, idx;
310 unsigned long offset;
311
312 if (!ptr)
313 return;
314
315 ptr -= sizeof(*hdr);
316 hdr = ptr;
317
318 assert(ptr_valid(pool, ptr));
319
320 sfree_check_redzone(hdr);
321
322 offset = ptr - pool->map;
323 i = offset / SMALLOC_BPL;
324 idx = (offset % SMALLOC_BPL) / SMALLOC_BPB;
325
326 pool_lock(pool);
327 clear_blocks(pool, i, idx, size_to_blocks(hdr->size));
328 if (i < pool->next_non_full)
329 pool->next_non_full = i;
330 pool->free_blocks += size_to_blocks(hdr->size);
331 pool_unlock(pool);
332}
333
334void sfree(void *ptr)
335{
336 struct pool *pool = NULL;
337 unsigned int i;
338
339 if (!ptr)
340 return;
341
342 global_read_lock();
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
351 global_read_unlock();
352
353 assert(pool);
354 sfree_pool(pool, ptr);
355}
356
357static void *__smalloc_pool(struct pool *pool, unsigned int size)
358{
359 unsigned int nr_blocks;
360 unsigned int i;
361 unsigned int offset;
362 unsigned int last_idx;
363 void *ret = NULL;
364
365 pool_lock(pool);
366
367 nr_blocks = size_to_blocks(size);
368 if (nr_blocks > pool->free_blocks)
369 goto fail;
370
371 i = pool->next_non_full;
372 last_idx = 0;
373 offset = -1U;
374 while (i < pool->nr_blocks) {
375 unsigned int idx;
376
377 if (pool->bitmap[i] == -1U) {
378 i++;
379 pool->next_non_full = i;
380 last_idx = 0;
381 continue;
382 }
383
384 idx = find_next_zero(pool->bitmap[i], last_idx);
385 if (!blocks_free(pool, i, idx, nr_blocks)) {
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;
397 }
398 set_blocks(pool, i, idx, nr_blocks);
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;
406 }
407fail:
408 pool_unlock(pool);
409 return ret;
410}
411
412static void *smalloc_pool(struct pool *pool, unsigned int size)
413{
414 unsigned int alloc_size = size + sizeof(struct block_hdr);
415 void *ptr;
416
417#ifdef SMALLOC_REDZONE
418 alloc_size += sizeof(unsigned int);
419#endif
420
421 ptr = __smalloc_pool(pool, alloc_size);
422 if (ptr) {
423 struct block_hdr *hdr = ptr;
424
425 hdr->size = alloc_size;
426 fill_redzone(hdr);
427
428 ptr += sizeof(*hdr);
429 memset(ptr, 0, size);
430 }
431
432 return ptr;
433}
434
435void *smalloc(unsigned int size)
436{
437 unsigned int i;
438
439 global_write_lock();
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;
448 global_write_unlock();
449 return ptr;
450 }
451 }
452 if (last_pool) {
453 last_pool = 0;
454 continue;
455 }
456
457 if (nr_pools + 1 > MAX_POOLS)
458 break;
459 else {
460 i = nr_pools;
461 if (add_pool(&mp[nr_pools], size))
462 goto out;
463 }
464 } while (1);
465
466out:
467 global_write_unlock();
468 return NULL;
469}
470
471char *smalloc_strdup(const char *str)
472{
473 char *ptr;
474
475 ptr = smalloc(strlen(str) + 1);
476 strcpy(ptr, str);
477 return ptr;
478}