10 #include "pthread_impl.h"
11 #include "malloc_impl.h"
13 #if defined(__GNUC__) && defined(__PIC__)
14 #define inline inline __attribute__((always_inline))
18 volatile uint64_t binmap;
20 volatile int free_lock[2];
23 int __malloc_replaced;
25 /* Synchronization tools */
27 static inline void lock(volatile int *lk)
29 if (libc.threads_minus_1)
30 while(a_swap(lk, 1)) __wait(lk, lk+1, 1, 1);
33 static inline void unlock(volatile int *lk)
37 if (lk[1]) __wake(lk, 1, 1);
41 static inline void lock_bin(int i)
43 lock(mal.bins[i].lock);
44 if (!mal.bins[i].head)
45 mal.bins[i].head = mal.bins[i].tail = BIN_TO_CHUNK(i);
48 static inline void unlock_bin(int i)
50 unlock(mal.bins[i].lock);
53 static int first_set(uint64_t x)
58 static const char debruijn64[64] = {
59 0, 1, 2, 53, 3, 7, 54, 27, 4, 38, 41, 8, 34, 55, 48, 28,
60 62, 5, 39, 46, 44, 42, 22, 9, 24, 35, 59, 56, 49, 18, 29, 11,
61 63, 52, 6, 26, 37, 40, 33, 47, 61, 45, 43, 21, 23, 58, 17, 10,
62 51, 25, 36, 32, 60, 20, 57, 16, 50, 31, 19, 15, 30, 14, 13, 12
64 static const char debruijn32[32] = {
65 0, 1, 23, 2, 29, 24, 19, 3, 30, 27, 25, 11, 20, 8, 4, 13,
66 31, 22, 28, 18, 26, 10, 7, 12, 21, 17, 9, 6, 16, 5, 15, 14
68 if (sizeof(long) < 8) {
72 return 32 + debruijn32[(y&-y)*0x076be629 >> 27];
74 return debruijn32[(y&-y)*0x076be629 >> 27];
76 return debruijn64[(x&-x)*0x022fdd63cc95386dull >> 58];
80 static const unsigned char bin_tab[60] = {
81 32,33,34,35,36,36,37,37,38,38,39,39,
82 40,40,40,40,41,41,41,41,42,42,42,42,43,43,43,43,
83 44,44,44,44,44,44,44,44,45,45,45,45,45,45,45,45,
84 46,46,46,46,46,46,46,46,47,47,47,47,47,47,47,47,
87 static int bin_index(size_t x)
89 x = x / SIZE_ALIGN - 1;
90 if (x <= 32) return x;
91 if (x < 512) return bin_tab[x/8-4];
92 if (x > 0x1c00) return 63;
93 return bin_tab[x/128-4] + 16;
96 static int bin_index_up(size_t x)
98 x = x / SIZE_ALIGN - 1;
99 if (x <= 32) return x;
101 if (x < 512) return bin_tab[x/8-4] + 1;
102 return bin_tab[x/128-4] + 17;
106 void __dump_heap(int x)
110 for (c = (void *)mal.heap; CHUNK_SIZE(c); c = NEXT_CHUNK(c))
111 fprintf(stderr, "base %p size %zu (%d) flags %d/%d\n",
112 c, CHUNK_SIZE(c), bin_index(CHUNK_SIZE(c)),
114 NEXT_CHUNK(c)->psize & 15);
115 for (i=0; i<64; i++) {
116 if (mal.bins[i].head != BIN_TO_CHUNK(i) && mal.bins[i].head) {
117 fprintf(stderr, "bin %d: %p\n", i, mal.bins[i].head);
118 if (!(mal.binmap & 1ULL<<i))
119 fprintf(stderr, "missing from binmap!\n");
120 } else if (mal.binmap & 1ULL<<i)
121 fprintf(stderr, "binmap wrongly contains %d!\n", i);
126 void *__expand_heap(size_t *);
128 static struct chunk *expand_heap(size_t n)
130 static int heap_lock[2];
135 /* The argument n already accounts for the caller's chunk
136 * overhead needs, but if the heap can't be extended in-place,
137 * we need room for an extra zero-sized sentinel chunk. */
142 p = __expand_heap(&n);
148 /* If not just expanding existing space, we need to make a
149 * new sentinel chunk below the allocated space. */
151 /* Valid/safe because of the prologue increment. */
153 p = (char *)p + SIZE_ALIGN;
155 w->psize = 0 | C_INUSE;
158 /* Record new heap end and fill in footer. */
160 w = MEM_TO_CHUNK(end);
161 w->psize = n | C_INUSE;
162 w->csize = 0 | C_INUSE;
164 /* Fill in header, which may be new or may be replacing a
165 * zero-size sentinel header at the old end-of-heap. */
167 w->csize = n | C_INUSE;
174 static int adjust_size(size_t *n)
176 /* Result of pointer difference must fit in ptrdiff_t. */
177 if (*n-1 > PTRDIFF_MAX - SIZE_ALIGN - PAGE_SIZE) {
186 *n = (*n + OVERHEAD + SIZE_ALIGN - 1) & SIZE_MASK;
190 static void unbin(struct chunk *c, int i)
192 if (c->prev == c->next)
193 a_and_64(&mal.binmap, ~(1ULL<<i));
194 c->prev->next = c->next;
195 c->next->prev = c->prev;
197 NEXT_CHUNK(c)->psize |= C_INUSE;
200 static int alloc_fwd(struct chunk *c)
204 while (!((k=c->csize) & C_INUSE)) {
217 static int alloc_rev(struct chunk *c)
221 while (!((k=c->psize) & C_INUSE)) {
225 unbin(PREV_CHUNK(c), i);
235 /* pretrim - trims a chunk _prior_ to removing it from its bin.
236 * Must be called with i as the ideal bin for size n, j the bin
237 * for the _free_ chunk self, and bin j locked. */
238 static int pretrim(struct chunk *self, size_t n, int i, int j)
241 struct chunk *next, *split;
243 /* We cannot pretrim if it would require re-binning. */
244 if (j < 40) return 0;
246 if (j != 63) return 0;
247 n1 = CHUNK_SIZE(self);
248 if (n1-n <= MMAP_THRESHOLD) return 0;
250 n1 = CHUNK_SIZE(self);
252 if (bin_index(n1-n) != j) return 0;
254 next = NEXT_CHUNK(self);
255 split = (void *)((char *)self + n);
257 split->prev = self->prev;
258 split->next = self->next;
259 split->prev->next = split;
260 split->next->prev = split;
261 split->psize = n | C_INUSE;
264 self->csize = n | C_INUSE;
268 static void trim(struct chunk *self, size_t n)
270 size_t n1 = CHUNK_SIZE(self);
271 struct chunk *next, *split;
273 if (n >= n1 - DONTCARE) return;
275 next = NEXT_CHUNK(self);
276 split = (void *)((char *)self + n);
278 split->psize = n | C_INUSE;
279 split->csize = n1-n | C_INUSE;
280 next->psize = n1-n | C_INUSE;
281 self->csize = n | C_INUSE;
286 void *malloc(size_t n)
291 if (adjust_size(&n) < 0) return 0;
293 if (n > MMAP_THRESHOLD) {
294 size_t len = n + OVERHEAD + PAGE_SIZE - 1 & -PAGE_SIZE;
295 char *base = __mmap(0, len, PROT_READ|PROT_WRITE,
296 MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
297 if (base == (void *)-1) return 0;
298 c = (void *)(base + SIZE_ALIGN - OVERHEAD);
299 c->csize = len - (SIZE_ALIGN - OVERHEAD);
300 c->psize = SIZE_ALIGN - OVERHEAD;
301 return CHUNK_TO_MEM(c);
306 uint64_t mask = mal.binmap & -(1ULL<<i);
313 NEXT_CHUNK(x)->psize = c->csize =
314 x->csize + CHUNK_SIZE(c);
320 c = mal.bins[j].head;
321 if (c != BIN_TO_CHUNK(j)) {
322 if (!pretrim(c, n, i, j)) unbin(c, j);
329 /* Now patch up in case we over-allocated */
332 return CHUNK_TO_MEM(c);
335 static size_t mal0_clear(char *p, size_t pagesz, size_t n)
338 typedef uint64_t __attribute__((__may_alias__)) T;
340 typedef unsigned char T;
343 size_t i = (uintptr_t)pp & (pagesz - 1);
345 pp = memset(pp - i, 0, i);
346 if (pp - p < pagesz) return pp - p;
347 for (i = pagesz; i; i -= 2*sizeof(T), pp -= 2*sizeof(T))
348 if (((T *)pp)[-1] | ((T *)pp)[-2])
353 void *calloc(size_t m, size_t n)
355 if (n && m > (size_t)-1/n) {
362 if (!__malloc_replaced) {
363 if (IS_MMAPPED(MEM_TO_CHUNK(p)))
366 n = mal0_clear(p, PAGE_SIZE, n);
368 return memset(p, 0, n);
371 void *realloc(void *p, size_t n)
373 struct chunk *self, *next;
377 if (!p) return malloc(n);
379 if (adjust_size(&n) < 0) return 0;
381 self = MEM_TO_CHUNK(p);
382 n1 = n0 = CHUNK_SIZE(self);
384 if (IS_MMAPPED(self)) {
385 size_t extra = self->psize;
386 char *base = (char *)self - extra;
387 size_t oldlen = n0 + extra;
388 size_t newlen = n + extra;
389 /* Crash on realloc of freed chunk */
390 if (extra & 1) a_crash();
391 if (newlen < PAGE_SIZE && (new = malloc(n-OVERHEAD))) {
395 newlen = (newlen + PAGE_SIZE-1) & -PAGE_SIZE;
396 if (oldlen == newlen) return p;
397 base = __mremap(base, oldlen, newlen, MREMAP_MAYMOVE);
398 if (base == (void *)-1)
400 self = (void *)(base + extra);
401 self->csize = newlen - extra;
402 return CHUNK_TO_MEM(self);
405 next = NEXT_CHUNK(self);
407 /* Crash on corrupted footer (likely from buffer overflow) */
408 if (next->psize != self->csize) a_crash();
410 /* Merge adjacent chunks if we need more space. This is not
411 * a waste of time even if we fail to get enough space, because our
412 * subsequent call to free would otherwise have to do the merge. */
413 if (n > n1 && alloc_fwd(next)) {
414 n1 += CHUNK_SIZE(next);
415 next = NEXT_CHUNK(next);
417 /* FIXME: find what's wrong here and reenable it..? */
418 if (0 && n > n1 && alloc_rev(self)) {
419 self = PREV_CHUNK(self);
420 n1 += CHUNK_SIZE(self);
422 self->csize = n1 | C_INUSE;
423 next->psize = n1 | C_INUSE;
425 /* If we got enough space, split off the excess and return */
427 //memmove(CHUNK_TO_MEM(self), p, n0-OVERHEAD);
429 return CHUNK_TO_MEM(self);
433 /* As a last resort, allocate a new chunk and copy to it. */
434 new = malloc(n-OVERHEAD);
437 memcpy(new, p, n0-OVERHEAD);
438 free(CHUNK_TO_MEM(self));
442 void __bin_chunk(struct chunk *self)
444 struct chunk *next = NEXT_CHUNK(self);
445 size_t final_size, new_size, size;
449 final_size = new_size = CHUNK_SIZE(self);
451 /* Crash on corrupted footer (likely from buffer overflow) */
452 if (next->psize != self->csize) a_crash();
455 if (self->psize & next->csize & C_INUSE) {
456 self->csize = final_size | C_INUSE;
457 next->psize = final_size | C_INUSE;
458 i = bin_index(final_size);
461 if (self->psize & next->csize & C_INUSE)
463 unlock(mal.free_lock);
467 if (alloc_rev(self)) {
468 self = PREV_CHUNK(self);
469 size = CHUNK_SIZE(self);
471 if (new_size+size > RECLAIM && (new_size+size^size) > size)
475 if (alloc_fwd(next)) {
476 size = CHUNK_SIZE(next);
478 if (new_size+size > RECLAIM && (new_size+size^size) > size)
480 next = NEXT_CHUNK(next);
484 if (!(mal.binmap & 1ULL<<i))
485 a_or_64(&mal.binmap, 1ULL<<i);
487 self->csize = final_size;
488 next->psize = final_size;
489 unlock(mal.free_lock);
491 self->next = BIN_TO_CHUNK(i);
492 self->prev = mal.bins[i].tail;
493 self->next->prev = self;
494 self->prev->next = self;
496 /* Replace middle of large chunks with fresh zero pages */
498 uintptr_t a = (uintptr_t)self + SIZE_ALIGN+PAGE_SIZE-1 & -PAGE_SIZE;
499 uintptr_t b = (uintptr_t)next - SIZE_ALIGN & -PAGE_SIZE;
501 __madvise((void *)a, b-a, MADV_DONTNEED);
503 __mmap((void *)a, b-a, PROT_READ|PROT_WRITE,
504 MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED, -1, 0);
511 static void unmap_chunk(struct chunk *self)
513 size_t extra = self->psize;
514 char *base = (char *)self - extra;
515 size_t len = CHUNK_SIZE(self) + extra;
516 /* Crash on double free */
517 if (extra & 1) a_crash();
525 struct chunk *self = MEM_TO_CHUNK(p);
527 if (IS_MMAPPED(self))
533 void __malloc_donate(char *start, char *end)
535 size_t align_start_up = (SIZE_ALIGN-1) & (-(uintptr_t)start - OVERHEAD);
536 size_t align_end_down = (SIZE_ALIGN-1) & (uintptr_t)end;
538 /* Getting past this condition ensures that the padding for alignment
539 * and header overhead will not overflow and will leave a nonzero
540 * multiple of SIZE_ALIGN bytes between start and end. */
541 if (end - start <= OVERHEAD + align_start_up + align_end_down)
543 start += align_start_up + OVERHEAD;
544 end -= align_end_down;
546 struct chunk *c = MEM_TO_CHUNK(start), *n = MEM_TO_CHUNK(end);
547 c->psize = n->csize = C_INUSE;
548 c->csize = n->psize = C_INUSE | (end-start);