• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 #define _GNU_SOURCE
2 #include <stdlib.h>
3 #include <string.h>
4 #include <limits.h>
5 #include <stdint.h>
6 #include <errno.h>
7 #include <sys/mman.h>
8 #include "libc.h"
9 #include "atomic.h"
10 #include "pthread_impl.h"
11 #include "malloc_impl.h"
12 #include "fork_impl.h"
13 
14 #define malloc __libc_malloc_impl
15 #define realloc __libc_realloc
16 #define free __libc_free
17 
18 #if defined(__GNUC__) && defined(__PIC__)
19 #define inline inline __attribute__((always_inline))
20 #endif
21 
22 static struct {
23 	volatile uint64_t binmap;
24 	struct bin bins[64];
25 	volatile int split_merge_lock[2];
26 } mal;
27 
28 /* Synchronization tools */
29 
lock(volatile int * lk)30 static inline void lock(volatile int *lk)
31 {
32 	int need_locks = libc.need_locks;
33 	if (need_locks) {
34 		while(a_swap(lk, 1)) __wait(lk, lk+1, 1, 1);
35 		if (need_locks < 0) libc.need_locks = 0;
36 	}
37 }
38 
unlock(volatile int * lk)39 static inline void unlock(volatile int *lk)
40 {
41 	if (lk[0]) {
42 		a_store(lk, 0);
43 		if (lk[1]) __wake(lk, 1, 1);
44 	}
45 }
46 
lock_bin(int i)47 static inline void lock_bin(int i)
48 {
49 	lock(mal.bins[i].lock);
50 	if (!mal.bins[i].head)
51 		mal.bins[i].head = mal.bins[i].tail = BIN_TO_CHUNK(i);
52 }
53 
unlock_bin(int i)54 static inline void unlock_bin(int i)
55 {
56 	unlock(mal.bins[i].lock);
57 }
58 
first_set(uint64_t x)59 static int first_set(uint64_t x)
60 {
61 #if 1
62 	return a_ctz_64(x);
63 #else
64 	static const char debruijn64[64] = {
65 		0, 1, 2, 53, 3, 7, 54, 27, 4, 38, 41, 8, 34, 55, 48, 28,
66 		62, 5, 39, 46, 44, 42, 22, 9, 24, 35, 59, 56, 49, 18, 29, 11,
67 		63, 52, 6, 26, 37, 40, 33, 47, 61, 45, 43, 21, 23, 58, 17, 10,
68 		51, 25, 36, 32, 60, 20, 57, 16, 50, 31, 19, 15, 30, 14, 13, 12
69 	};
70 	static const char debruijn32[32] = {
71 		0, 1, 23, 2, 29, 24, 19, 3, 30, 27, 25, 11, 20, 8, 4, 13,
72 		31, 22, 28, 18, 26, 10, 7, 12, 21, 17, 9, 6, 16, 5, 15, 14
73 	};
74 	if (sizeof(long) < 8) {
75 		uint32_t y = x;
76 		if (!y) {
77 			y = x>>32;
78 			return 32 + debruijn32[(y&-y)*0x076be629 >> 27];
79 		}
80 		return debruijn32[(y&-y)*0x076be629 >> 27];
81 	}
82 	return debruijn64[(x&-x)*0x022fdd63cc95386dull >> 58];
83 #endif
84 }
85 
86 static const unsigned char bin_tab[60] = {
87 	            32,33,34,35,36,36,37,37,38,38,39,39,
88 	40,40,40,40,41,41,41,41,42,42,42,42,43,43,43,43,
89 	44,44,44,44,44,44,44,44,45,45,45,45,45,45,45,45,
90 	46,46,46,46,46,46,46,46,47,47,47,47,47,47,47,47,
91 };
92 
bin_index(size_t x)93 static int bin_index(size_t x)
94 {
95 	x = x / SIZE_ALIGN - 1;
96 	if (x <= 32) return x;
97 	if (x < 512) return bin_tab[x/8-4];
98 	if (x > 0x1c00) return 63;
99 	return bin_tab[x/128-4] + 16;
100 }
101 
bin_index_up(size_t x)102 static int bin_index_up(size_t x)
103 {
104 	x = x / SIZE_ALIGN - 1;
105 	if (x <= 32) return x;
106 	x--;
107 	if (x < 512) return bin_tab[x/8-4] + 1;
108 	return bin_tab[x/128-4] + 17;
109 }
110 
111 #if 0
112 void __dump_heap(int x)
113 {
114 	struct chunk *c;
115 	int i;
116 	for (c = (void *)mal.heap; CHUNK_SIZE(c); c = NEXT_CHUNK(c))
117 		fprintf(stderr, "base %p size %zu (%d) flags %d/%d\n",
118 			c, CHUNK_SIZE(c), bin_index(CHUNK_SIZE(c)),
119 			c->csize & 15,
120 			NEXT_CHUNK(c)->psize & 15);
121 	for (i=0; i<64; i++) {
122 		if (mal.bins[i].head != BIN_TO_CHUNK(i) && mal.bins[i].head) {
123 			fprintf(stderr, "bin %d: %p\n", i, mal.bins[i].head);
124 			if (!(mal.binmap & 1ULL<<i))
125 				fprintf(stderr, "missing from binmap!\n");
126 		} else if (mal.binmap & 1ULL<<i)
127 			fprintf(stderr, "binmap wrongly contains %d!\n", i);
128 	}
129 }
130 #endif
131 
132 /* This function returns true if the interval [old,new]
133  * intersects the 'len'-sized interval below &libc.auxv
134  * (interpreted as the main-thread stack) or below &b
135  * (the current stack). It is used to defend against
136  * buggy brk implementations that can cross the stack. */
137 
traverses_stack_p(uintptr_t old,uintptr_t new)138 static int traverses_stack_p(uintptr_t old, uintptr_t new)
139 {
140 	const uintptr_t len = 8<<20;
141 	uintptr_t a, b;
142 
143 	b = (uintptr_t)libc.auxv;
144 	a = b > len ? b-len : 0;
145 	if (new>a && old<b) return 1;
146 
147 	b = (uintptr_t)&b;
148 	a = b > len ? b-len : 0;
149 	if (new>a && old<b) return 1;
150 
151 	return 0;
152 }
153 
154 /* Expand the heap in-place if brk can be used, or otherwise via mmap,
155  * using an exponential lower bound on growth by mmap to make
156  * fragmentation asymptotically irrelevant. The size argument is both
157  * an input and an output, since the caller needs to know the size
158  * allocated, which will be larger than requested due to page alignment
159  * and mmap minimum size rules. The caller is responsible for locking
160  * to prevent concurrent calls. */
161 
__expand_heap(size_t * pn)162 static void *__expand_heap(size_t *pn)
163 {
164 	static uintptr_t brk;
165 	static unsigned mmap_step;
166 	size_t n = *pn;
167 
168 	if (n > SIZE_MAX/2 - PAGE_SIZE) {
169 		errno = ENOMEM;
170 		return 0;
171 	}
172 	n += -n & PAGE_SIZE-1;
173 
174 	if (!brk) {
175 		brk = __syscall(SYS_brk, 0);
176 		brk += -brk & PAGE_SIZE-1;
177 	}
178 
179 	if (n < SIZE_MAX-brk && !traverses_stack_p(brk, brk+n)
180 	    && __syscall(SYS_brk, brk+n)==brk+n) {
181 		*pn = n;
182 		brk += n;
183 		return (void *)(brk-n);
184 	}
185 
186 	size_t min = (size_t)PAGE_SIZE << mmap_step/2;
187 	if (n < min) n = min;
188 	void *area = __mmap(0, n, PROT_READ|PROT_WRITE,
189 		MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
190 	if (area == MAP_FAILED) return 0;
191 	*pn = n;
192 	mmap_step++;
193 	return area;
194 }
195 
expand_heap(size_t n)196 static struct chunk *expand_heap(size_t n)
197 {
198 	static void *end;
199 	void *p;
200 	struct chunk *w;
201 
202 	/* The argument n already accounts for the caller's chunk
203 	 * overhead needs, but if the heap can't be extended in-place,
204 	 * we need room for an extra zero-sized sentinel chunk. */
205 	n += SIZE_ALIGN;
206 
207 	p = __expand_heap(&n);
208 	if (!p) return 0;
209 
210 	/* If not just expanding existing space, we need to make a
211 	 * new sentinel chunk below the allocated space. */
212 	if (p != end) {
213 		/* Valid/safe because of the prologue increment. */
214 		n -= SIZE_ALIGN;
215 		p = (char *)p + SIZE_ALIGN;
216 		w = MEM_TO_CHUNK(p);
217 		w->psize = 0 | C_INUSE;
218 	}
219 
220 	/* Record new heap end and fill in footer. */
221 	end = (char *)p + n;
222 	w = MEM_TO_CHUNK(end);
223 	w->psize = n | C_INUSE;
224 	w->csize = 0 | C_INUSE;
225 
226 	/* Fill in header, which may be new or may be replacing a
227 	 * zero-size sentinel header at the old end-of-heap. */
228 	w = MEM_TO_CHUNK(p);
229 	w->csize = n | C_INUSE;
230 
231 	return w;
232 }
233 
adjust_size(size_t * n)234 static int adjust_size(size_t *n)
235 {
236 	/* Result of pointer difference must fit in ptrdiff_t. */
237 	if (*n-1 > PTRDIFF_MAX - SIZE_ALIGN - PAGE_SIZE) {
238 		if (*n) {
239 			errno = ENOMEM;
240 			return -1;
241 		} else {
242 			*n = SIZE_ALIGN;
243 			return 0;
244 		}
245 	}
246 	*n = (*n + OVERHEAD + SIZE_ALIGN - 1) & SIZE_MASK;
247 	return 0;
248 }
249 
unbin(struct chunk * c,int i)250 static void unbin(struct chunk *c, int i)
251 {
252 	if (c->prev == c->next)
253 		a_and_64(&mal.binmap, ~(1ULL<<i));
254 	c->prev->next = c->next;
255 	c->next->prev = c->prev;
256 	c->csize |= C_INUSE;
257 	NEXT_CHUNK(c)->psize |= C_INUSE;
258 }
259 
bin_chunk(struct chunk * self,int i)260 static void bin_chunk(struct chunk *self, int i)
261 {
262 	self->next = BIN_TO_CHUNK(i);
263 	self->prev = mal.bins[i].tail;
264 	self->next->prev = self;
265 	self->prev->next = self;
266 	if (self->prev == BIN_TO_CHUNK(i))
267 		a_or_64(&mal.binmap, 1ULL<<i);
268 }
269 
trim(struct chunk * self,size_t n)270 static void trim(struct chunk *self, size_t n)
271 {
272 	size_t n1 = CHUNK_SIZE(self);
273 	struct chunk *next, *split;
274 
275 	if (n >= n1 - DONTCARE) return;
276 
277 	next = NEXT_CHUNK(self);
278 	split = (void *)((char *)self + n);
279 
280 	split->psize = n | C_INUSE;
281 	split->csize = n1-n;
282 	next->psize = n1-n;
283 	self->csize = n | C_INUSE;
284 
285 	int i = bin_index(n1-n);
286 	lock_bin(i);
287 
288 	bin_chunk(split, i);
289 
290 	unlock_bin(i);
291 }
292 
malloc(size_t n)293 void *malloc(size_t n)
294 {
295 	struct chunk *c;
296 	int i, j;
297 	uint64_t mask;
298 
299 	if (adjust_size(&n) < 0) return 0;
300 
301 	if (n > MMAP_THRESHOLD) {
302 		size_t len = n + OVERHEAD + PAGE_SIZE - 1 & -PAGE_SIZE;
303 		char *base = __mmap(0, len, PROT_READ|PROT_WRITE,
304 			MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
305 		if (base == (void *)-1) return 0;
306 		c = (void *)(base + SIZE_ALIGN - OVERHEAD);
307 		c->csize = len - (SIZE_ALIGN - OVERHEAD);
308 		c->psize = SIZE_ALIGN - OVERHEAD;
309 		return CHUNK_TO_MEM(c);
310 	}
311 
312 	i = bin_index_up(n);
313 	if (i<63 && (mal.binmap & (1ULL<<i))) {
314 		lock_bin(i);
315 		c = mal.bins[i].head;
316 		if (c != BIN_TO_CHUNK(i) && CHUNK_SIZE(c)-n <= DONTCARE) {
317 			unbin(c, i);
318 			unlock_bin(i);
319 			return CHUNK_TO_MEM(c);
320 		}
321 		unlock_bin(i);
322 	}
323 	lock(mal.split_merge_lock);
324 	for (mask = mal.binmap & -(1ULL<<i); mask; mask -= (mask&-mask)) {
325 		j = first_set(mask);
326 		lock_bin(j);
327 		c = mal.bins[j].head;
328 		if (c != BIN_TO_CHUNK(j)) {
329 			unbin(c, j);
330 			unlock_bin(j);
331 			break;
332 		}
333 		unlock_bin(j);
334 	}
335 	if (!mask) {
336 		c = expand_heap(n);
337 		if (!c) {
338 			unlock(mal.split_merge_lock);
339 			return 0;
340 		}
341 	}
342 	trim(c, n);
343 	unlock(mal.split_merge_lock);
344 	return CHUNK_TO_MEM(c);
345 }
346 
__malloc_allzerop(void * p)347 int __malloc_allzerop(void *p)
348 {
349 	return IS_MMAPPED(MEM_TO_CHUNK(p));
350 }
351 
realloc(void * p,size_t n)352 void *realloc(void *p, size_t n)
353 {
354 	struct chunk *self, *next;
355 	size_t n0, n1;
356 	void *new;
357 
358 	if (!p) return malloc(n);
359 
360 	if (adjust_size(&n) < 0) return 0;
361 
362 	self = MEM_TO_CHUNK(p);
363 	n1 = n0 = CHUNK_SIZE(self);
364 
365 	if (n<=n0 && n0-n<=DONTCARE) return p;
366 
367 	if (IS_MMAPPED(self)) {
368 		size_t extra = self->psize;
369 		char *base = (char *)self - extra;
370 		size_t oldlen = n0 + extra;
371 		size_t newlen = n + extra;
372 		/* Crash on realloc of freed chunk */
373 		if (extra & 1) a_crash();
374 		if (newlen < PAGE_SIZE && (new = malloc(n-OVERHEAD))) {
375 			n0 = n;
376 			goto copy_free_ret;
377 		}
378 		newlen = (newlen + PAGE_SIZE-1) & -PAGE_SIZE;
379 		if (oldlen == newlen) return p;
380 		base = __mremap(base, oldlen, newlen, MREMAP_MAYMOVE);
381 		if (base == (void *)-1)
382 			goto copy_realloc;
383 		self = (void *)(base + extra);
384 		self->csize = newlen - extra;
385 		return CHUNK_TO_MEM(self);
386 	}
387 
388 	next = NEXT_CHUNK(self);
389 
390 	/* Crash on corrupted footer (likely from buffer overflow) */
391 	if (next->psize != self->csize) a_crash();
392 
393 	if (n < n0) {
394 		int i = bin_index_up(n);
395 		int j = bin_index(n0);
396 		if (i<j && (mal.binmap & (1ULL << i)))
397 			goto copy_realloc;
398 		struct chunk *split = (void *)((char *)self + n);
399 		self->csize = split->psize = n | C_INUSE;
400 		split->csize = next->psize = n0-n | C_INUSE;
401 		__bin_chunk(split);
402 		return CHUNK_TO_MEM(self);
403 	}
404 
405 	lock(mal.split_merge_lock);
406 
407 	size_t nsize = next->csize & C_INUSE ? 0 : CHUNK_SIZE(next);
408 	if (n0+nsize >= n) {
409 		int i = bin_index(nsize);
410 		lock_bin(i);
411 		if (!(next->csize & C_INUSE)) {
412 			unbin(next, i);
413 			unlock_bin(i);
414 			next = NEXT_CHUNK(next);
415 			self->csize = next->psize = n0+nsize | C_INUSE;
416 			trim(self, n);
417 			unlock(mal.split_merge_lock);
418 			return CHUNK_TO_MEM(self);
419 		}
420 		unlock_bin(i);
421 	}
422 	unlock(mal.split_merge_lock);
423 
424 copy_realloc:
425 	/* As a last resort, allocate a new chunk and copy to it. */
426 	new = malloc(n-OVERHEAD);
427 	if (!new) return 0;
428 copy_free_ret:
429 	memcpy(new, p, (n<n0 ? n : n0) - OVERHEAD);
430 	free(CHUNK_TO_MEM(self));
431 	return new;
432 }
433 
__bin_chunk(struct chunk * self)434 void __bin_chunk(struct chunk *self)
435 {
436 	struct chunk *next = NEXT_CHUNK(self);
437 
438 	/* Crash on corrupted footer (likely from buffer overflow) */
439 	if (next->psize != self->csize) a_crash();
440 
441 	lock(mal.split_merge_lock);
442 
443 	size_t osize = CHUNK_SIZE(self), size = osize;
444 
445 	/* Since we hold split_merge_lock, only transition from free to
446 	 * in-use can race; in-use to free is impossible */
447 	size_t psize = self->psize & C_INUSE ? 0 : CHUNK_PSIZE(self);
448 	size_t nsize = next->csize & C_INUSE ? 0 : CHUNK_SIZE(next);
449 
450 	if (psize) {
451 		int i = bin_index(psize);
452 		lock_bin(i);
453 		if (!(self->psize & C_INUSE)) {
454 			struct chunk *prev = PREV_CHUNK(self);
455 			unbin(prev, i);
456 			self = prev;
457 			size += psize;
458 		}
459 		unlock_bin(i);
460 	}
461 	if (nsize) {
462 		int i = bin_index(nsize);
463 		lock_bin(i);
464 		if (!(next->csize & C_INUSE)) {
465 			unbin(next, i);
466 			next = NEXT_CHUNK(next);
467 			size += nsize;
468 		}
469 		unlock_bin(i);
470 	}
471 
472 	int i = bin_index(size);
473 	lock_bin(i);
474 
475 	self->csize = size;
476 	next->psize = size;
477 	bin_chunk(self, i);
478 	unlock(mal.split_merge_lock);
479 
480 	/* Replace middle of large chunks with fresh zero pages */
481 	if (size > RECLAIM && (size^(size-osize)) > size-osize) {
482 		uintptr_t a = (uintptr_t)self + SIZE_ALIGN+PAGE_SIZE-1 & -PAGE_SIZE;
483 		uintptr_t b = (uintptr_t)next - SIZE_ALIGN & -PAGE_SIZE;
484 		int e = errno;
485 #if 1
486 		__madvise((void *)a, b-a, MADV_DONTNEED);
487 #else
488 		__mmap((void *)a, b-a, PROT_READ|PROT_WRITE,
489 			MAP_PRIVATE|MAP_ANONYMOUS|MAP_FIXED, -1, 0);
490 #endif
491 		errno = e;
492 	}
493 
494 	unlock_bin(i);
495 }
496 
unmap_chunk(struct chunk * self)497 static void unmap_chunk(struct chunk *self)
498 {
499 	size_t extra = self->psize;
500 	char *base = (char *)self - extra;
501 	size_t len = CHUNK_SIZE(self) + extra;
502 	/* Crash on double free */
503 	if (extra & 1) a_crash();
504 	int e = errno;
505 	__munmap(base, len);
506 	errno = e;
507 }
508 
free(void * p)509 void free(void *p)
510 {
511 	if (!p) return;
512 
513 	struct chunk *self = MEM_TO_CHUNK(p);
514 
515 	if (IS_MMAPPED(self))
516 		unmap_chunk(self);
517 	else
518 		__bin_chunk(self);
519 }
520 
__malloc_donate(char * start,char * end)521 void __malloc_donate(char *start, char *end)
522 {
523 	size_t align_start_up = (SIZE_ALIGN-1) & (-(uintptr_t)start - OVERHEAD);
524 	size_t align_end_down = (SIZE_ALIGN-1) & (uintptr_t)end;
525 
526 	/* Getting past this condition ensures that the padding for alignment
527 	 * and header overhead will not overflow and will leave a nonzero
528 	 * multiple of SIZE_ALIGN bytes between start and end. */
529 	if (end - start <= OVERHEAD + align_start_up + align_end_down)
530 		return;
531 	start += align_start_up + OVERHEAD;
532 	end   -= align_end_down;
533 
534 	struct chunk *c = MEM_TO_CHUNK(start), *n = MEM_TO_CHUNK(end);
535 	c->psize = n->csize = C_INUSE;
536 	c->csize = n->psize = C_INUSE | (end-start);
537 	__bin_chunk(c);
538 }
539 
__malloc_atfork(int who)540 void __malloc_atfork(int who)
541 {
542 	if (who<0) {
543 		lock(mal.split_merge_lock);
544 		for (int i=0; i<64; i++)
545 			lock(mal.bins[i].lock);
546 	} else if (!who) {
547 		for (int i=0; i<64; i++)
548 			unlock(mal.bins[i].lock);
549 		unlock(mal.split_merge_lock);
550 	} else {
551 		for (int i=0; i<64; i++)
552 			mal.bins[i].lock[0] = mal.bins[i].lock[1] = 0;
553 		mal.split_merge_lock[1] = 0;
554 		mal.split_merge_lock[0] = 0;
555 	}
556 }
557