xref: /openbsd-src/lib/libc/stdlib/malloc.c (revision de8cc8edbc71bd3e3bc7fbffa27ba0e564c37d8b)
1 /*	$OpenBSD: malloc.c,v 1.268 2021/02/25 15:20:18 otto Exp $	*/
2 /*
3  * Copyright (c) 2008, 2010, 2011, 2016 Otto Moerbeek <otto@drijf.net>
4  * Copyright (c) 2012 Matthew Dempsky <matthew@openbsd.org>
5  * Copyright (c) 2008 Damien Miller <djm@openbsd.org>
6  * Copyright (c) 2000 Poul-Henning Kamp <phk@FreeBSD.org>
7  *
8  * Permission to use, copy, modify, and distribute this software for any
9  * purpose with or without fee is hereby granted, provided that the above
10  * copyright notice and this permission notice appear in all copies.
11  *
12  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
13  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
14  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
15  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
16  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
17  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
18  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
19  */
20 
21 /*
22  * If we meet some day, and you think this stuff is worth it, you
23  * can buy me a beer in return. Poul-Henning Kamp
24  */
25 
26 /* #define MALLOC_STATS */
27 
28 #include <sys/types.h>
29 #include <sys/queue.h>
30 #include <sys/mman.h>
31 #include <sys/sysctl.h>
32 #include <uvm/uvmexp.h>
33 #include <errno.h>
34 #include <stdarg.h>
35 #include <stdint.h>
36 #include <stdio.h>
37 #include <stdlib.h>
38 #include <string.h>
39 #include <unistd.h>
40 
41 #ifdef MALLOC_STATS
42 #include <sys/tree.h>
43 #include <fcntl.h>
44 #endif
45 
46 #include "thread_private.h"
47 #include <tib.h>
48 
49 #define MALLOC_PAGESHIFT	_MAX_PAGE_SHIFT
50 
51 #define MALLOC_MINSHIFT		4
52 #define MALLOC_MAXSHIFT		(MALLOC_PAGESHIFT - 1)
53 #define MALLOC_PAGESIZE		(1UL << MALLOC_PAGESHIFT)
54 #define MALLOC_MINSIZE		(1UL << MALLOC_MINSHIFT)
55 #define MALLOC_PAGEMASK		(MALLOC_PAGESIZE - 1)
56 #define MASK_POINTER(p)		((void *)(((uintptr_t)(p)) & ~MALLOC_PAGEMASK))
57 
58 #define MALLOC_MAXCHUNK		(1 << MALLOC_MAXSHIFT)
59 #define MALLOC_MAXCACHE		256
60 #define MALLOC_DELAYED_CHUNK_MASK	15
61 #ifdef MALLOC_STATS
62 #define MALLOC_INITIAL_REGIONS	512
63 #else
64 #define MALLOC_INITIAL_REGIONS	(MALLOC_PAGESIZE / sizeof(struct region_info))
65 #endif
66 #define MALLOC_DEFAULT_CACHE	64
67 #define MALLOC_CHUNK_LISTS	4
68 #define CHUNK_CHECK_LENGTH	32
69 
70 /*
71  * We move allocations between half a page and a whole page towards the end,
72  * subject to alignment constraints. This is the extra headroom we allow.
73  * Set to zero to be the most strict.
74  */
75 #define MALLOC_LEEWAY		0
76 #define MALLOC_MOVE_COND(sz)	((sz) - mopts.malloc_guard < 		\
77 				    MALLOC_PAGESIZE - MALLOC_LEEWAY)
78 #define MALLOC_MOVE(p, sz)  	(((char *)(p)) +			\
79 				    ((MALLOC_PAGESIZE - MALLOC_LEEWAY -	\
80 			    	    ((sz) - mopts.malloc_guard)) & 	\
81 				    ~(MALLOC_MINSIZE - 1)))
82 
83 #define PAGEROUND(x)  (((x) + (MALLOC_PAGEMASK)) & ~MALLOC_PAGEMASK)
84 
85 /*
86  * What to use for Junk.  This is the byte value we use to fill with
87  * when the 'J' option is enabled. Use SOME_JUNK right after alloc,
88  * and SOME_FREEJUNK right before free.
89  */
90 #define SOME_JUNK		0xdb	/* deadbeef */
91 #define SOME_FREEJUNK		0xdf	/* dead, free */
92 #define SOME_FREEJUNK_ULL	0xdfdfdfdfdfdfdfdfULL
93 
94 #define MMAP(sz,f)	mmap(NULL, (sz), PROT_READ | PROT_WRITE, \
95     MAP_ANON | MAP_PRIVATE | (f), -1, 0)
96 
97 #define MMAPNONE(sz,f)	mmap(NULL, (sz), PROT_NONE, \
98     MAP_ANON | MAP_PRIVATE | (f), -1, 0)
99 
100 #define MMAPA(a,sz,f)	mmap((a), (sz), PROT_READ | PROT_WRITE, \
101     MAP_ANON | MAP_PRIVATE | (f), -1, 0)
102 
103 #define MQUERY(a,sz,f)	mquery((a), (sz), PROT_READ | PROT_WRITE, \
104     MAP_ANON | MAP_PRIVATE | MAP_FIXED | (f), -1, 0)
105 
106 struct region_info {
107 	void *p;		/* page; low bits used to mark chunks */
108 	uintptr_t size;		/* size for pages, or chunk_info pointer */
109 #ifdef MALLOC_STATS
110 	void *f;		/* where allocated from */
111 #endif
112 };
113 
114 LIST_HEAD(chunk_head, chunk_info);
115 
116 struct dir_info {
117 	u_int32_t canary1;
118 	int active;			/* status of malloc */
119 	struct region_info *r;		/* region slots */
120 	size_t regions_total;		/* number of region slots */
121 	size_t regions_free;		/* number of free slots */
122 	size_t free_regions_size;	/* free pages cached */
123 	size_t rbytesused;		/* random bytes used */
124 	char *func;			/* current function */
125 	u_int malloc_cache;		/* # of free pages we cache */
126 	int malloc_junk;		/* junk fill? */
127 	int mmap_flag;			/* extra flag for mmap */
128 	u_int rotor;
129 	int mutex;
130 					/* lists of free chunk info structs */
131 	struct chunk_head chunk_info_list[MALLOC_MAXSHIFT + 1];
132 					/* lists of chunks with free slots */
133 	struct chunk_head chunk_dir[MALLOC_MAXSHIFT + 1][MALLOC_CHUNK_LISTS];
134 					/* free pages cache */
135 	struct region_info free_regions[MALLOC_MAXCACHE];
136 					/* delayed free chunk slots */
137 	void *delayed_chunks[MALLOC_DELAYED_CHUNK_MASK + 1];
138 	u_char rbytes[32];		/* random bytes */
139 #ifdef MALLOC_STATS
140 	size_t inserts;
141 	size_t insert_collisions;
142 	size_t finds;
143 	size_t find_collisions;
144 	size_t deletes;
145 	size_t delete_moves;
146 	size_t cheap_realloc_tries;
147 	size_t cheap_reallocs;
148 	size_t malloc_used;		/* bytes allocated */
149 	size_t malloc_guarded;		/* bytes used for guards */
150 	size_t pool_searches;		/* searches for pool */
151 	size_t other_pool;		/* searches in other pool */
152 #define STATS_ADD(x,y)	((x) += (y))
153 #define STATS_SUB(x,y)	((x) -= (y))
154 #define STATS_INC(x)	((x)++)
155 #define STATS_ZERO(x)	((x) = 0)
156 #define STATS_SETF(x,y)	((x)->f = (y))
157 #else
158 #define STATS_ADD(x,y)	/* nothing */
159 #define STATS_SUB(x,y)	/* nothing */
160 #define STATS_INC(x)	/* nothing */
161 #define STATS_ZERO(x)	/* nothing */
162 #define STATS_SETF(x,y)	/* nothing */
163 #endif /* MALLOC_STATS */
164 	u_int32_t canary2;
165 };
166 #define DIR_INFO_RSZ	((sizeof(struct dir_info) + MALLOC_PAGEMASK) & \
167 			~MALLOC_PAGEMASK)
168 
169 /*
170  * This structure describes a page worth of chunks.
171  *
172  * How many bits per u_short in the bitmap
173  */
174 #define MALLOC_BITS		(NBBY * sizeof(u_short))
175 struct chunk_info {
176 	LIST_ENTRY(chunk_info) entries;
177 	void *page;			/* pointer to the page */
178 	u_short canary;
179 	u_short size;			/* size of this page's chunks */
180 	u_short shift;			/* how far to shift for this size */
181 	u_short free;			/* how many free chunks */
182 	u_short total;			/* how many chunks */
183 	u_short offset;			/* requested size table offset */
184 	u_short bits[1];		/* which chunks are free */
185 };
186 
187 struct malloc_readonly {
188 					/* Main bookkeeping information */
189 	struct dir_info *malloc_pool[_MALLOC_MUTEXES];
190 	u_int	malloc_mutexes;		/* how much in actual use? */
191 	int	malloc_mt;		/* multi-threaded mode? */
192 	int	malloc_freecheck;	/* Extensive double free check */
193 	int	malloc_freeunmap;	/* mprotect free pages PROT_NONE? */
194 	int	def_malloc_junk;	/* junk fill? */
195 	int	malloc_realloc;		/* always realloc? */
196 	int	malloc_xmalloc;		/* xmalloc behaviour? */
197 	u_int	chunk_canaries;		/* use canaries after chunks? */
198 	int	internal_funcs;		/* use better recallocarray/freezero? */
199 	u_int	def_malloc_cache;	/* free pages we cache */
200 	size_t	malloc_guard;		/* use guard pages after allocations? */
201 #ifdef MALLOC_STATS
202 	int	malloc_stats;		/* dump statistics at end */
203 #endif
204 	u_int32_t malloc_canary;	/* Matched against ones in malloc_pool */
205 };
206 
207 /* This object is mapped PROT_READ after initialisation to prevent tampering */
208 static union {
209 	struct malloc_readonly mopts;
210 	u_char _pad[MALLOC_PAGESIZE];
211 } malloc_readonly __attribute__((aligned(MALLOC_PAGESIZE)));
212 #define mopts	malloc_readonly.mopts
213 
214 char		*malloc_options;	/* compile-time options */
215 
216 static __dead void wrterror(struct dir_info *d, char *msg, ...)
217     __attribute__((__format__ (printf, 2, 3)));
218 
219 #ifdef MALLOC_STATS
220 void malloc_dump(int, int, struct dir_info *);
221 PROTO_NORMAL(malloc_dump);
222 void malloc_gdump(int);
223 PROTO_NORMAL(malloc_gdump);
224 static void malloc_exit(void);
225 #define CALLER	__builtin_return_address(0)
226 #else
227 #define CALLER	NULL
228 #endif
229 
230 /* low bits of r->p determine size: 0 means >= page size and r->size holding
231  * real size, otherwise low bits are a shift count, or 1 for malloc(0)
232  */
233 #define REALSIZE(sz, r)						\
234 	(sz) = (uintptr_t)(r)->p & MALLOC_PAGEMASK,		\
235 	(sz) = ((sz) == 0 ? (r)->size : ((sz) == 1 ? 0 : (1 << ((sz)-1))))
236 
237 static inline void
238 _MALLOC_LEAVE(struct dir_info *d)
239 {
240 	if (mopts.malloc_mt) {
241 		d->active--;
242 		_MALLOC_UNLOCK(d->mutex);
243 	}
244 }
245 
246 static inline void
247 _MALLOC_ENTER(struct dir_info *d)
248 {
249 	if (mopts.malloc_mt) {
250 		_MALLOC_LOCK(d->mutex);
251 		d->active++;
252 	}
253 }
254 
255 static inline size_t
256 hash(void *p)
257 {
258 	size_t sum;
259 	uintptr_t u;
260 
261 	u = (uintptr_t)p >> MALLOC_PAGESHIFT;
262 	sum = u;
263 	sum = (sum << 7) - sum + (u >> 16);
264 #ifdef __LP64__
265 	sum = (sum << 7) - sum + (u >> 32);
266 	sum = (sum << 7) - sum + (u >> 48);
267 #endif
268 	return sum;
269 }
270 
271 static inline struct dir_info *
272 getpool(void)
273 {
274 	if (!mopts.malloc_mt)
275 		return mopts.malloc_pool[1];
276 	else	/* first one reserved for special pool */
277 		return mopts.malloc_pool[1 + TIB_GET()->tib_tid %
278 		    (mopts.malloc_mutexes - 1)];
279 }
280 
281 static __dead void
282 wrterror(struct dir_info *d, char *msg, ...)
283 {
284 	int		saved_errno = errno;
285 	va_list		ap;
286 
287 	dprintf(STDERR_FILENO, "%s(%d) in %s(): ", __progname,
288 	    getpid(), (d != NULL && d->func) ? d->func : "unknown");
289 	va_start(ap, msg);
290 	vdprintf(STDERR_FILENO, msg, ap);
291 	va_end(ap);
292 	dprintf(STDERR_FILENO, "\n");
293 
294 #ifdef MALLOC_STATS
295 	if (mopts.malloc_stats)
296 		malloc_gdump(STDERR_FILENO);
297 #endif /* MALLOC_STATS */
298 
299 	errno = saved_errno;
300 
301 	abort();
302 }
303 
304 static void
305 rbytes_init(struct dir_info *d)
306 {
307 	arc4random_buf(d->rbytes, sizeof(d->rbytes));
308 	/* add 1 to account for using d->rbytes[0] */
309 	d->rbytesused = 1 + d->rbytes[0] % (sizeof(d->rbytes) / 2);
310 }
311 
312 static inline u_char
313 getrbyte(struct dir_info *d)
314 {
315 	u_char x;
316 
317 	if (d->rbytesused >= sizeof(d->rbytes))
318 		rbytes_init(d);
319 	x = d->rbytes[d->rbytesused++];
320 	return x;
321 }
322 
323 static void
324 omalloc_parseopt(char opt)
325 {
326 	switch (opt) {
327 	case '+':
328 		mopts.malloc_mutexes <<= 1;
329 		if (mopts.malloc_mutexes > _MALLOC_MUTEXES)
330 			mopts.malloc_mutexes = _MALLOC_MUTEXES;
331 		break;
332 	case '-':
333 		mopts.malloc_mutexes >>= 1;
334 		if (mopts.malloc_mutexes < 2)
335 			mopts.malloc_mutexes = 2;
336 		break;
337 	case '>':
338 		mopts.def_malloc_cache <<= 1;
339 		if (mopts.def_malloc_cache > MALLOC_MAXCACHE)
340 			mopts.def_malloc_cache = MALLOC_MAXCACHE;
341 		break;
342 	case '<':
343 		mopts.def_malloc_cache >>= 1;
344 		break;
345 	case 'c':
346 		mopts.chunk_canaries = 0;
347 		break;
348 	case 'C':
349 		mopts.chunk_canaries = 1;
350 		break;
351 #ifdef MALLOC_STATS
352 	case 'd':
353 		mopts.malloc_stats = 0;
354 		break;
355 	case 'D':
356 		mopts.malloc_stats = 1;
357 		break;
358 #endif /* MALLOC_STATS */
359 	case 'f':
360 		mopts.malloc_freecheck = 0;
361 		mopts.malloc_freeunmap = 0;
362 		break;
363 	case 'F':
364 		mopts.malloc_freecheck = 1;
365 		mopts.malloc_freeunmap = 1;
366 		break;
367 	case 'g':
368 		mopts.malloc_guard = 0;
369 		break;
370 	case 'G':
371 		mopts.malloc_guard = MALLOC_PAGESIZE;
372 		break;
373 	case 'j':
374 		if (mopts.def_malloc_junk > 0)
375 			mopts.def_malloc_junk--;
376 		break;
377 	case 'J':
378 		if (mopts.def_malloc_junk < 2)
379 			mopts.def_malloc_junk++;
380 		break;
381 	case 'r':
382 		mopts.malloc_realloc = 0;
383 		break;
384 	case 'R':
385 		mopts.malloc_realloc = 1;
386 		break;
387 	case 'u':
388 		mopts.malloc_freeunmap = 0;
389 		break;
390 	case 'U':
391 		mopts.malloc_freeunmap = 1;
392 		break;
393 	case 'x':
394 		mopts.malloc_xmalloc = 0;
395 		break;
396 	case 'X':
397 		mopts.malloc_xmalloc = 1;
398 		break;
399 	default:
400 		dprintf(STDERR_FILENO, "malloc() warning: "
401                     "unknown char in MALLOC_OPTIONS\n");
402 		break;
403 	}
404 }
405 
406 static void
407 omalloc_init(void)
408 {
409 	char *p, *q, b[16];
410 	int i, j;
411 	const int mib[2] = { CTL_VM, VM_MALLOC_CONF };
412 	size_t sb;
413 
414 	/*
415 	 * Default options
416 	 */
417 	mopts.malloc_mutexes = 8;
418 	mopts.def_malloc_junk = 1;
419 	mopts.def_malloc_cache = MALLOC_DEFAULT_CACHE;
420 
421 	for (i = 0; i < 3; i++) {
422 		switch (i) {
423 		case 0:
424 			sb = sizeof(b);
425 			j = sysctl(mib, 2, b, &sb, NULL, 0);
426 			if (j != 0)
427 				continue;
428 			p = b;
429 			break;
430 		case 1:
431 			if (issetugid() == 0)
432 				p = getenv("MALLOC_OPTIONS");
433 			else
434 				continue;
435 			break;
436 		case 2:
437 			p = malloc_options;
438 			break;
439 		default:
440 			p = NULL;
441 		}
442 
443 		for (; p != NULL && *p != '\0'; p++) {
444 			switch (*p) {
445 			case 'S':
446 				for (q = "CFGJ"; *q != '\0'; q++)
447 					omalloc_parseopt(*q);
448 				mopts.def_malloc_cache = 0;
449 				break;
450 			case 's':
451 				for (q = "cfgj"; *q != '\0'; q++)
452 					omalloc_parseopt(*q);
453 				mopts.def_malloc_cache = MALLOC_DEFAULT_CACHE;
454 				break;
455 			default:
456 				omalloc_parseopt(*p);
457 				break;
458 			}
459 		}
460 	}
461 
462 #ifdef MALLOC_STATS
463 	if (mopts.malloc_stats && (atexit(malloc_exit) == -1)) {
464 		dprintf(STDERR_FILENO, "malloc() warning: atexit(2) failed."
465 		    " Will not be able to dump stats on exit\n");
466 	}
467 #endif /* MALLOC_STATS */
468 
469 	while ((mopts.malloc_canary = arc4random()) == 0)
470 		;
471 	if (mopts.chunk_canaries)
472 		do {
473 			mopts.chunk_canaries = arc4random();
474 		} while ((u_char)mopts.chunk_canaries == 0 ||
475 		    (u_char)mopts.chunk_canaries == SOME_FREEJUNK);
476 }
477 
478 static void
479 omalloc_poolinit(struct dir_info **dp, int mmap_flag)
480 {
481 	char *p;
482 	size_t d_avail, regioninfo_size;
483 	struct dir_info *d;
484 	int i, j;
485 
486 	/*
487 	 * Allocate dir_info with a guard page on either side. Also
488 	 * randomise offset inside the page at which the dir_info
489 	 * lies (subject to alignment by 1 << MALLOC_MINSHIFT)
490 	 */
491 	if ((p = MMAPNONE(DIR_INFO_RSZ + (MALLOC_PAGESIZE * 2), mmap_flag)) ==
492 	    MAP_FAILED)
493 		wrterror(NULL, "malloc init mmap failed");
494 	mprotect(p + MALLOC_PAGESIZE, DIR_INFO_RSZ, PROT_READ | PROT_WRITE);
495 	d_avail = (DIR_INFO_RSZ - sizeof(*d)) >> MALLOC_MINSHIFT;
496 	d = (struct dir_info *)(p + MALLOC_PAGESIZE +
497 	    (arc4random_uniform(d_avail) << MALLOC_MINSHIFT));
498 
499 	rbytes_init(d);
500 	d->regions_free = d->regions_total = MALLOC_INITIAL_REGIONS;
501 	regioninfo_size = d->regions_total * sizeof(struct region_info);
502 	d->r = MMAP(regioninfo_size, mmap_flag);
503 	if (d->r == MAP_FAILED) {
504 		d->regions_total = 0;
505 		wrterror(NULL, "malloc init mmap failed");
506 	}
507 	for (i = 0; i <= MALLOC_MAXSHIFT; i++) {
508 		LIST_INIT(&d->chunk_info_list[i]);
509 		for (j = 0; j < MALLOC_CHUNK_LISTS; j++)
510 			LIST_INIT(&d->chunk_dir[i][j]);
511 	}
512 	STATS_ADD(d->malloc_used, regioninfo_size + 3 * MALLOC_PAGESIZE);
513 	d->mmap_flag = mmap_flag;
514 	d->malloc_junk = mopts.def_malloc_junk;
515 	d->malloc_cache = mopts.def_malloc_cache;
516 	d->canary1 = mopts.malloc_canary ^ (u_int32_t)(uintptr_t)d;
517 	d->canary2 = ~d->canary1;
518 
519 	*dp = d;
520 }
521 
522 static int
523 omalloc_grow(struct dir_info *d)
524 {
525 	size_t newtotal;
526 	size_t newsize;
527 	size_t mask;
528 	size_t i;
529 	struct region_info *p;
530 
531 	if (d->regions_total > SIZE_MAX / sizeof(struct region_info) / 2)
532 		return 1;
533 
534 	newtotal = d->regions_total * 2;
535 	newsize = newtotal * sizeof(struct region_info);
536 	mask = newtotal - 1;
537 
538 	p = MMAP(newsize, d->mmap_flag);
539 	if (p == MAP_FAILED)
540 		return 1;
541 
542 	STATS_ADD(d->malloc_used, newsize);
543 	STATS_ZERO(d->inserts);
544 	STATS_ZERO(d->insert_collisions);
545 	for (i = 0; i < d->regions_total; i++) {
546 		void *q = d->r[i].p;
547 		if (q != NULL) {
548 			size_t index = hash(q) & mask;
549 			STATS_INC(d->inserts);
550 			while (p[index].p != NULL) {
551 				index = (index - 1) & mask;
552 				STATS_INC(d->insert_collisions);
553 			}
554 			p[index] = d->r[i];
555 		}
556 	}
557 	/* avoid pages containing meta info to end up in cache */
558 	if (munmap(d->r, d->regions_total * sizeof(struct region_info)))
559 		wrterror(d, "munmap %p", (void *)d->r);
560 	else
561 		STATS_SUB(d->malloc_used,
562 		    d->regions_total * sizeof(struct region_info));
563 	d->regions_free = d->regions_free + d->regions_total;
564 	d->regions_total = newtotal;
565 	d->r = p;
566 	return 0;
567 }
568 
569 /*
570  * The hashtable uses the assumption that p is never NULL. This holds since
571  * non-MAP_FIXED mappings with hint 0 start at BRKSIZ.
572  */
573 static int
574 insert(struct dir_info *d, void *p, size_t sz, void *f)
575 {
576 	size_t index;
577 	size_t mask;
578 	void *q;
579 
580 	if (d->regions_free * 4 < d->regions_total) {
581 		if (omalloc_grow(d))
582 			return 1;
583 	}
584 	mask = d->regions_total - 1;
585 	index = hash(p) & mask;
586 	q = d->r[index].p;
587 	STATS_INC(d->inserts);
588 	while (q != NULL) {
589 		index = (index - 1) & mask;
590 		q = d->r[index].p;
591 		STATS_INC(d->insert_collisions);
592 	}
593 	d->r[index].p = p;
594 	d->r[index].size = sz;
595 #ifdef MALLOC_STATS
596 	d->r[index].f = f;
597 #endif
598 	d->regions_free--;
599 	return 0;
600 }
601 
602 static struct region_info *
603 find(struct dir_info *d, void *p)
604 {
605 	size_t index;
606 	size_t mask = d->regions_total - 1;
607 	void *q, *r;
608 
609 	if (mopts.malloc_canary != (d->canary1 ^ (u_int32_t)(uintptr_t)d) ||
610 	    d->canary1 != ~d->canary2)
611 		wrterror(d, "internal struct corrupt");
612 	p = MASK_POINTER(p);
613 	index = hash(p) & mask;
614 	r = d->r[index].p;
615 	q = MASK_POINTER(r);
616 	STATS_INC(d->finds);
617 	while (q != p && r != NULL) {
618 		index = (index - 1) & mask;
619 		r = d->r[index].p;
620 		q = MASK_POINTER(r);
621 		STATS_INC(d->find_collisions);
622 	}
623 	return (q == p && r != NULL) ? &d->r[index] : NULL;
624 }
625 
626 static void
627 delete(struct dir_info *d, struct region_info *ri)
628 {
629 	/* algorithm R, Knuth Vol III section 6.4 */
630 	size_t mask = d->regions_total - 1;
631 	size_t i, j, r;
632 
633 	if (d->regions_total & (d->regions_total - 1))
634 		wrterror(d, "regions_total not 2^x");
635 	d->regions_free++;
636 	STATS_INC(d->deletes);
637 
638 	i = ri - d->r;
639 	for (;;) {
640 		d->r[i].p = NULL;
641 		d->r[i].size = 0;
642 		j = i;
643 		for (;;) {
644 			i = (i - 1) & mask;
645 			if (d->r[i].p == NULL)
646 				return;
647 			r = hash(d->r[i].p) & mask;
648 			if ((i <= r && r < j) || (r < j && j < i) ||
649 			    (j < i && i <= r))
650 				continue;
651 			d->r[j] = d->r[i];
652 			STATS_INC(d->delete_moves);
653 			break;
654 		}
655 
656 	}
657 }
658 
659 static inline void
660 junk_free(int junk, void *p, size_t sz)
661 {
662 	size_t i, step = 1;
663 	uint64_t *lp = p;
664 
665 	if (junk == 0 || sz == 0)
666 		return;
667 	sz /= sizeof(uint64_t);
668 	if (junk == 1) {
669 		if (sz > MALLOC_PAGESIZE / sizeof(uint64_t))
670 			sz = MALLOC_PAGESIZE / sizeof(uint64_t);
671 		step = sz / 4;
672 		if (step == 0)
673 			step = 1;
674 	}
675 	for (i = 0; i < sz; i += step)
676 		lp[i] = SOME_FREEJUNK_ULL;
677 }
678 
679 static inline void
680 validate_junk(struct dir_info *pool, void *p, size_t sz)
681 {
682 	size_t i, step = 1;
683 	uint64_t *lp = p;
684 
685 	if (pool->malloc_junk == 0 || sz == 0)
686 		return;
687 	sz /= sizeof(uint64_t);
688 	if (pool->malloc_junk == 1) {
689 		if (sz > MALLOC_PAGESIZE / sizeof(uint64_t))
690 			sz = MALLOC_PAGESIZE / sizeof(uint64_t);
691 		step = sz / 4;
692 		if (step == 0)
693 			step = 1;
694 	}
695 	for (i = 0; i < sz; i += step) {
696 		if (lp[i] != SOME_FREEJUNK_ULL)
697 			wrterror(pool, "write after free %p", p);
698 	}
699 }
700 
701 
702 /*
703  * Cache maintenance. We keep at most malloc_cache pages cached.
704  * If the cache is becoming full, unmap pages in the cache for real,
705  * and then add the region to the cache
706  * Opposed to the regular region data structure, the sizes in the
707  * cache are in MALLOC_PAGESIZE units.
708  */
709 static void
710 unmap(struct dir_info *d, void *p, size_t sz, size_t clear)
711 {
712 	size_t psz = sz >> MALLOC_PAGESHIFT;
713 	size_t rsz;
714 	struct region_info *r;
715 	u_int i, offset, mask;
716 
717 	if (sz != PAGEROUND(sz))
718 		wrterror(d, "munmap round");
719 
720 	rsz = d->malloc_cache - d->free_regions_size;
721 
722 	/*
723 	 * normally the cache holds recently freed regions, but if the region
724 	 * to unmap is larger than the cache size or we're clearing and the
725 	 * cache is full, just munmap
726 	 */
727 	if (psz > d->malloc_cache || (clear > 0 && rsz == 0)) {
728 		i = munmap(p, sz);
729 		if (i)
730 			wrterror(d, "munmap %p", p);
731 		STATS_SUB(d->malloc_used, sz);
732 		return;
733 	}
734 	offset = getrbyte(d);
735 	mask = d->malloc_cache - 1;
736 	if (psz > rsz) {
737 		size_t tounmap = psz - rsz;
738 		for (i = 0; ; i++) {
739 			r = &d->free_regions[(i + offset) & mask];
740 			if (r->p != NULL) {
741 				rsz = r->size << MALLOC_PAGESHIFT;
742 				if (!mopts.malloc_freeunmap)
743 					validate_junk(d, r->p, rsz);
744 				if (munmap(r->p, rsz))
745 					wrterror(d, "munmap %p", r->p);
746 				r->p = NULL;
747 				if (tounmap > r->size)
748 					tounmap -= r->size;
749 				else
750 					tounmap = 0;
751 				d->free_regions_size -= r->size;
752 				STATS_SUB(d->malloc_used, rsz);
753 				if (tounmap == 0) {
754 					offset = i;
755 					break;
756 				}
757 			}
758 		}
759 	}
760 	for (i = 0; ; i++) {
761 		r = &d->free_regions[(i + offset) & mask];
762 		if (r->p == NULL) {
763 			if (clear > 0)
764 				memset(p, 0, clear);
765 			if (mopts.malloc_freeunmap)
766 				mprotect(p, sz, PROT_NONE);
767 			else
768 				junk_free(d->malloc_junk, p,
769 				    psz << MALLOC_PAGESHIFT);
770 			r->p = p;
771 			r->size = psz;
772 			d->free_regions_size += psz;
773 			break;
774 		}
775 	}
776 	if (d->free_regions_size > d->malloc_cache)
777 		wrterror(d, "malloc cache overflow");
778 }
779 
780 static void *
781 map(struct dir_info *d, size_t sz, int zero_fill)
782 {
783 	size_t psz = sz >> MALLOC_PAGESHIFT;
784 	struct region_info *r, *big = NULL;
785 	u_int i;
786 	void *p;
787 
788 	if (mopts.malloc_canary != (d->canary1 ^ (u_int32_t)(uintptr_t)d) ||
789 	    d->canary1 != ~d->canary2)
790 		wrterror(d, "internal struct corrupt");
791 	if (sz != PAGEROUND(sz))
792 		wrterror(d, "map round");
793 
794 	if (psz > d->free_regions_size) {
795 		_MALLOC_LEAVE(d);
796 		p = MMAP(sz, d->mmap_flag);
797 		_MALLOC_ENTER(d);
798 		if (p != MAP_FAILED)
799 			STATS_ADD(d->malloc_used, sz);
800 		/* zero fill not needed */
801 		return p;
802 	}
803 	for (i = 0; i < d->malloc_cache; i++) {
804 		r = &d->free_regions[(i + d->rotor) & (d->malloc_cache - 1)];
805 		if (r->p != NULL) {
806 			if (r->size == psz) {
807 				p = r->p;
808 				if (!mopts.malloc_freeunmap)
809 					validate_junk(d, p,
810 					    psz << MALLOC_PAGESHIFT);
811 				r->p = NULL;
812 				d->free_regions_size -= psz;
813 				if (mopts.malloc_freeunmap)
814 					mprotect(p, sz, PROT_READ | PROT_WRITE);
815 				if (zero_fill)
816 					memset(p, 0, sz);
817 				else if (mopts.malloc_freeunmap)
818 					junk_free(d->malloc_junk, p, sz);
819 				d->rotor += i + 1;
820 				return p;
821 			} else if (r->size > psz)
822 				big = r;
823 		}
824 	}
825 	if (big != NULL) {
826 		r = big;
827 		p = r->p;
828 		if (!mopts.malloc_freeunmap)
829 			validate_junk(d, p, r->size << MALLOC_PAGESHIFT);
830 		r->p = (char *)r->p + (psz << MALLOC_PAGESHIFT);
831 		r->size -= psz;
832 		d->free_regions_size -= psz;
833 		if (mopts.malloc_freeunmap)
834 			mprotect(p, sz, PROT_READ | PROT_WRITE);
835 		else
836 			junk_free(d->malloc_junk, r->p,
837 			    r->size << MALLOC_PAGESHIFT);
838 		if (zero_fill)
839 			memset(p, 0, sz);
840 		else if (mopts.malloc_freeunmap)
841 			junk_free(d->malloc_junk, p, sz);
842 		return p;
843 	}
844 	if (d->free_regions_size > d->malloc_cache)
845 		wrterror(d, "malloc cache");
846 	_MALLOC_LEAVE(d);
847 	p = MMAP(sz, d->mmap_flag);
848 	_MALLOC_ENTER(d);
849 	if (p != MAP_FAILED)
850 		STATS_ADD(d->malloc_used, sz);
851 	/* zero fill not needed */
852 	return p;
853 }
854 
855 static void
856 init_chunk_info(struct dir_info *d, struct chunk_info *p, int bits)
857 {
858 	int i;
859 
860 	if (bits == 0) {
861 		p->shift = MALLOC_MINSHIFT;
862 		p->total = p->free = MALLOC_PAGESIZE >> p->shift;
863 		p->size = 0;
864 		p->offset = 0xdead;
865 	} else {
866 		p->shift = bits;
867 		p->total = p->free = MALLOC_PAGESIZE >> p->shift;
868 		p->size = 1U << bits;
869 		p->offset = howmany(p->total, MALLOC_BITS);
870 	}
871 	p->canary = (u_short)d->canary1;
872 
873 	/* set all valid bits in the bitmap */
874  	i = p->total - 1;
875 	memset(p->bits, 0xff, sizeof(p->bits[0]) * (i / MALLOC_BITS));
876 	p->bits[i / MALLOC_BITS] = (2U << (i % MALLOC_BITS)) - 1;
877 }
878 
879 static struct chunk_info *
880 alloc_chunk_info(struct dir_info *d, int bits)
881 {
882 	struct chunk_info *p;
883 
884 	if (LIST_EMPTY(&d->chunk_info_list[bits])) {
885 		size_t size, count, i;
886 		char *q;
887 
888 		if (bits == 0)
889 			count = MALLOC_PAGESIZE / MALLOC_MINSIZE;
890 		else
891 			count = MALLOC_PAGESIZE >> bits;
892 
893 		size = howmany(count, MALLOC_BITS);
894 		size = sizeof(struct chunk_info) + (size - 1) * sizeof(u_short);
895 		if (mopts.chunk_canaries)
896 			size += count * sizeof(u_short);
897 		size = _ALIGN(size);
898 
899 		q = MMAP(MALLOC_PAGESIZE, d->mmap_flag);
900 		if (q == MAP_FAILED)
901 			return NULL;
902 		STATS_ADD(d->malloc_used, MALLOC_PAGESIZE);
903 		count = MALLOC_PAGESIZE / size;
904 
905 		for (i = 0; i < count; i++, q += size) {
906 			p = (struct chunk_info *)q;
907 			LIST_INSERT_HEAD(&d->chunk_info_list[bits], p, entries);
908 		}
909 	}
910 	p = LIST_FIRST(&d->chunk_info_list[bits]);
911 	LIST_REMOVE(p, entries);
912 	if (p->shift == 0)
913 		init_chunk_info(d, p, bits);
914 	return p;
915 }
916 
917 /*
918  * Allocate a page of chunks
919  */
920 static struct chunk_info *
921 omalloc_make_chunks(struct dir_info *d, int bits, int listnum)
922 {
923 	struct chunk_info *bp;
924 	void *pp;
925 
926 	/* Allocate a new bucket */
927 	pp = map(d, MALLOC_PAGESIZE, 0);
928 	if (pp == MAP_FAILED)
929 		return NULL;
930 
931 	/* memory protect the page allocated in the malloc(0) case */
932 	if (bits == 0 && mprotect(pp, MALLOC_PAGESIZE, PROT_NONE) == -1)
933 		goto err;
934 
935 	bp = alloc_chunk_info(d, bits);
936 	if (bp == NULL)
937 		goto err;
938 	bp->page = pp;
939 
940 	if (insert(d, (void *)((uintptr_t)pp | (bits + 1)), (uintptr_t)bp,
941 	    NULL))
942 		goto err;
943 	LIST_INSERT_HEAD(&d->chunk_dir[bits][listnum], bp, entries);
944 	return bp;
945 
946 err:
947 	unmap(d, pp, MALLOC_PAGESIZE, 0);
948 	return NULL;
949 }
950 
951 static int
952 find_chunksize(size_t size)
953 {
954 	int r;
955 
956 	/* malloc(0) is special */
957 	if (size == 0)
958 		return 0;
959 
960 	if (size < MALLOC_MINSIZE)
961 		size = MALLOC_MINSIZE;
962 	size--;
963 
964 	r = MALLOC_MINSHIFT;
965 	while (size >> r)
966 		r++;
967 	return r;
968 }
969 
970 static void
971 fill_canary(char *ptr, size_t sz, size_t allocated)
972 {
973 	size_t check_sz = allocated - sz;
974 
975 	if (check_sz > CHUNK_CHECK_LENGTH)
976 		check_sz = CHUNK_CHECK_LENGTH;
977 	memset(ptr + sz, mopts.chunk_canaries, check_sz);
978 }
979 
980 /*
981  * Allocate a chunk
982  */
983 static void *
984 malloc_bytes(struct dir_info *d, size_t size, void *f)
985 {
986 	u_int i, r;
987 	int j, listnum;
988 	size_t k;
989 	u_short	*lp;
990 	struct chunk_info *bp;
991 	void *p;
992 
993 	if (mopts.malloc_canary != (d->canary1 ^ (u_int32_t)(uintptr_t)d) ||
994 	    d->canary1 != ~d->canary2)
995 		wrterror(d, "internal struct corrupt");
996 
997 	j = find_chunksize(size);
998 
999 	r = ((u_int)getrbyte(d) << 8) | getrbyte(d);
1000 	listnum = r % MALLOC_CHUNK_LISTS;
1001 	/* If it's empty, make a page more of that size chunks */
1002 	if ((bp = LIST_FIRST(&d->chunk_dir[j][listnum])) == NULL) {
1003 		bp = omalloc_make_chunks(d, j, listnum);
1004 		if (bp == NULL)
1005 			return NULL;
1006 	}
1007 
1008 	if (bp->canary != (u_short)d->canary1)
1009 		wrterror(d, "chunk info corrupted");
1010 
1011 	i = (r / MALLOC_CHUNK_LISTS) & (bp->total - 1);
1012 
1013 	/* start somewhere in a short */
1014 	lp = &bp->bits[i / MALLOC_BITS];
1015 	if (*lp) {
1016 		j = i % MALLOC_BITS;
1017 		k = ffs(*lp >> j);
1018 		if (k != 0) {
1019 			k += j - 1;
1020 			goto found;
1021 		}
1022 	}
1023 	/* no bit halfway, go to next full short */
1024 	i /= MALLOC_BITS;
1025 	for (;;) {
1026 		if (++i >= bp->total / MALLOC_BITS)
1027 			i = 0;
1028 		lp = &bp->bits[i];
1029 		if (*lp) {
1030 			k = ffs(*lp) - 1;
1031 			break;
1032 		}
1033 	}
1034 found:
1035 #ifdef MALLOC_STATS
1036 	if (i == 0 && k == 0) {
1037 		struct region_info *r = find(d, bp->page);
1038 		r->f = f;
1039 	}
1040 #endif
1041 
1042 	*lp ^= 1 << k;
1043 
1044 	/* If there are no more free, remove from free-list */
1045 	if (--bp->free == 0)
1046 		LIST_REMOVE(bp, entries);
1047 
1048 	/* Adjust to the real offset of that chunk */
1049 	k += (lp - bp->bits) * MALLOC_BITS;
1050 
1051 	if (mopts.chunk_canaries && size > 0)
1052 		bp->bits[bp->offset + k] = size;
1053 
1054 	k <<= bp->shift;
1055 
1056 	p = (char *)bp->page + k;
1057 	if (bp->size > 0) {
1058 		if (d->malloc_junk == 2)
1059 			memset(p, SOME_JUNK, bp->size);
1060 		else if (mopts.chunk_canaries)
1061 			fill_canary(p, size, bp->size);
1062 	}
1063 	return p;
1064 }
1065 
1066 static void
1067 validate_canary(struct dir_info *d, u_char *ptr, size_t sz, size_t allocated)
1068 {
1069 	size_t check_sz = allocated - sz;
1070 	u_char *p, *q;
1071 
1072 	if (check_sz > CHUNK_CHECK_LENGTH)
1073 		check_sz = CHUNK_CHECK_LENGTH;
1074 	p = ptr + sz;
1075 	q = p + check_sz;
1076 
1077 	while (p < q) {
1078 		if (*p != (u_char)mopts.chunk_canaries && *p != SOME_JUNK) {
1079 			wrterror(d, "chunk canary corrupted %p %#tx@%#zx%s",
1080 			    ptr, p - ptr, sz,
1081 			    *p == SOME_FREEJUNK ? " (double free?)" : "");
1082 		}
1083 		p++;
1084 	}
1085 }
1086 
1087 static uint32_t
1088 find_chunknum(struct dir_info *d, struct chunk_info *info, void *ptr, int check)
1089 {
1090 	uint32_t chunknum;
1091 
1092 	if (info->canary != (u_short)d->canary1)
1093 		wrterror(d, "chunk info corrupted");
1094 
1095 	/* Find the chunk number on the page */
1096 	chunknum = ((uintptr_t)ptr & MALLOC_PAGEMASK) >> info->shift;
1097 
1098 	if ((uintptr_t)ptr & ((1U << (info->shift)) - 1))
1099 		wrterror(d, "modified chunk-pointer %p", ptr);
1100 	if (info->bits[chunknum / MALLOC_BITS] &
1101 	    (1U << (chunknum % MALLOC_BITS)))
1102 		wrterror(d, "chunk is already free %p", ptr);
1103 	if (check && info->size > 0) {
1104 		validate_canary(d, ptr, info->bits[info->offset + chunknum],
1105 		    info->size);
1106 	}
1107 	return chunknum;
1108 }
1109 
1110 /*
1111  * Free a chunk, and possibly the page it's on, if the page becomes empty.
1112  */
1113 static void
1114 free_bytes(struct dir_info *d, struct region_info *r, void *ptr)
1115 {
1116 	struct chunk_head *mp;
1117 	struct chunk_info *info;
1118 	uint32_t chunknum;
1119 	int listnum;
1120 
1121 	info = (struct chunk_info *)r->size;
1122 	chunknum = find_chunknum(d, info, ptr, 0);
1123 
1124 	info->bits[chunknum / MALLOC_BITS] |= 1U << (chunknum % MALLOC_BITS);
1125 	info->free++;
1126 
1127 	if (info->free == 1) {
1128 		/* Page became non-full */
1129 		listnum = getrbyte(d) % MALLOC_CHUNK_LISTS;
1130 		if (info->size != 0)
1131 			mp = &d->chunk_dir[info->shift][listnum];
1132 		else
1133 			mp = &d->chunk_dir[0][listnum];
1134 
1135 		LIST_INSERT_HEAD(mp, info, entries);
1136 		return;
1137 	}
1138 
1139 	if (info->free != info->total)
1140 		return;
1141 
1142 	LIST_REMOVE(info, entries);
1143 
1144 	if (info->size == 0 && !mopts.malloc_freeunmap)
1145 		mprotect(info->page, MALLOC_PAGESIZE, PROT_READ | PROT_WRITE);
1146 	unmap(d, info->page, MALLOC_PAGESIZE, 0);
1147 
1148 	delete(d, r);
1149 	if (info->size != 0)
1150 		mp = &d->chunk_info_list[info->shift];
1151 	else
1152 		mp = &d->chunk_info_list[0];
1153 	LIST_INSERT_HEAD(mp, info, entries);
1154 }
1155 
1156 
1157 
1158 static void *
1159 omalloc(struct dir_info *pool, size_t sz, int zero_fill, void *f)
1160 {
1161 	void *p;
1162 	size_t psz;
1163 
1164 	if (sz > MALLOC_MAXCHUNK) {
1165 		if (sz >= SIZE_MAX - mopts.malloc_guard - MALLOC_PAGESIZE) {
1166 			errno = ENOMEM;
1167 			return NULL;
1168 		}
1169 		sz += mopts.malloc_guard;
1170 		psz = PAGEROUND(sz);
1171 		p = map(pool, psz, zero_fill);
1172 		if (p == MAP_FAILED) {
1173 			errno = ENOMEM;
1174 			return NULL;
1175 		}
1176 		if (insert(pool, p, sz, f)) {
1177 			unmap(pool, p, psz, 0);
1178 			errno = ENOMEM;
1179 			return NULL;
1180 		}
1181 		if (mopts.malloc_guard) {
1182 			if (mprotect((char *)p + psz - mopts.malloc_guard,
1183 			    mopts.malloc_guard, PROT_NONE))
1184 				wrterror(pool, "mprotect");
1185 			STATS_ADD(pool->malloc_guarded, mopts.malloc_guard);
1186 		}
1187 
1188 		if (MALLOC_MOVE_COND(sz)) {
1189 			/* fill whole allocation */
1190 			if (pool->malloc_junk == 2)
1191 				memset(p, SOME_JUNK, psz - mopts.malloc_guard);
1192 			/* shift towards the end */
1193 			p = MALLOC_MOVE(p, sz);
1194 			/* fill zeros if needed and overwritten above */
1195 			if (zero_fill && pool->malloc_junk == 2)
1196 				memset(p, 0, sz - mopts.malloc_guard);
1197 		} else {
1198 			if (pool->malloc_junk == 2) {
1199 				if (zero_fill)
1200 					memset((char *)p + sz - mopts.malloc_guard,
1201 					    SOME_JUNK, psz - sz);
1202 				else
1203 					memset(p, SOME_JUNK,
1204 					    psz - mopts.malloc_guard);
1205 			} else if (mopts.chunk_canaries)
1206 				fill_canary(p, sz - mopts.malloc_guard,
1207 				    psz - mopts.malloc_guard);
1208 		}
1209 
1210 	} else {
1211 		/* takes care of SOME_JUNK */
1212 		p = malloc_bytes(pool, sz, f);
1213 		if (zero_fill && p != NULL && sz > 0)
1214 			memset(p, 0, sz);
1215 	}
1216 
1217 	return p;
1218 }
1219 
1220 /*
1221  * Common function for handling recursion.  Only
1222  * print the error message once, to avoid making the problem
1223  * potentially worse.
1224  */
1225 static void
1226 malloc_recurse(struct dir_info *d)
1227 {
1228 	static int noprint;
1229 
1230 	if (noprint == 0) {
1231 		noprint = 1;
1232 		wrterror(d, "recursive call");
1233 	}
1234 	d->active--;
1235 	_MALLOC_UNLOCK(d->mutex);
1236 	errno = EDEADLK;
1237 }
1238 
1239 void
1240 _malloc_init(int from_rthreads)
1241 {
1242 	u_int i, nmutexes;
1243 	struct dir_info *d;
1244 
1245 	_MALLOC_LOCK(1);
1246 	if (!from_rthreads && mopts.malloc_pool[1]) {
1247 		_MALLOC_UNLOCK(1);
1248 		return;
1249 	}
1250 	if (!mopts.malloc_canary)
1251 		omalloc_init();
1252 
1253 	nmutexes = from_rthreads ? mopts.malloc_mutexes : 2;
1254 	if (((uintptr_t)&malloc_readonly & MALLOC_PAGEMASK) == 0)
1255 		mprotect(&malloc_readonly, sizeof(malloc_readonly),
1256 		    PROT_READ | PROT_WRITE);
1257 	for (i = 0; i < nmutexes; i++) {
1258 		if (mopts.malloc_pool[i])
1259 			continue;
1260 		if (i == 0) {
1261 			omalloc_poolinit(&d, MAP_CONCEAL);
1262 			d->malloc_junk = 2;
1263 			d->malloc_cache = 0;
1264 		} else {
1265 			omalloc_poolinit(&d, 0);
1266 			d->malloc_junk = mopts.def_malloc_junk;
1267 			d->malloc_cache = mopts.def_malloc_cache;
1268 		}
1269 		d->mutex = i;
1270 		mopts.malloc_pool[i] = d;
1271 	}
1272 
1273 	if (from_rthreads)
1274 		mopts.malloc_mt = 1;
1275 	else
1276 		mopts.internal_funcs = 1;
1277 
1278 	/*
1279 	 * Options have been set and will never be reset.
1280 	 * Prevent further tampering with them.
1281 	 */
1282 	if (((uintptr_t)&malloc_readonly & MALLOC_PAGEMASK) == 0)
1283 		mprotect(&malloc_readonly, sizeof(malloc_readonly), PROT_READ);
1284 	_MALLOC_UNLOCK(1);
1285 }
1286 DEF_STRONG(_malloc_init);
1287 
1288 #define PROLOGUE(p, fn)			\
1289 	d = (p); 			\
1290 	if (d == NULL) { 		\
1291 		_malloc_init(0);	\
1292 		d = (p);		\
1293 	}				\
1294 	_MALLOC_LOCK(d->mutex);		\
1295 	d->func = fn;			\
1296 	if (d->active++) {		\
1297 		malloc_recurse(d);	\
1298 		return NULL;		\
1299 	}				\
1300 
1301 #define EPILOGUE()				\
1302 	d->active--;				\
1303 	_MALLOC_UNLOCK(d->mutex);		\
1304 	if (r == NULL && mopts.malloc_xmalloc)	\
1305 		wrterror(d, "out of memory");	\
1306 	if (r != NULL)				\
1307 		errno = saved_errno;		\
1308 
1309 void *
1310 malloc(size_t size)
1311 {
1312 	void *r;
1313 	struct dir_info *d;
1314 	int saved_errno = errno;
1315 
1316 	PROLOGUE(getpool(), "malloc")
1317 	r = omalloc(d, size, 0, CALLER);
1318 	EPILOGUE()
1319 	return r;
1320 }
1321 /*DEF_STRONG(malloc);*/
1322 
1323 void *
1324 malloc_conceal(size_t size)
1325 {
1326 	void *r;
1327 	struct dir_info *d;
1328 	int saved_errno = errno;
1329 
1330 	PROLOGUE(mopts.malloc_pool[0], "malloc_conceal")
1331 	r = omalloc(d, size, 0, CALLER);
1332 	EPILOGUE()
1333 	return r;
1334 }
1335 DEF_WEAK(malloc_conceal);
1336 
1337 static struct region_info *
1338 findpool(void *p, struct dir_info *argpool, struct dir_info **foundpool,
1339     char **saved_function)
1340 {
1341 	struct dir_info *pool = argpool;
1342 	struct region_info *r = find(pool, p);
1343 
1344 	STATS_INC(pool->pool_searches);
1345 	if (r == NULL) {
1346 		u_int i, nmutexes;
1347 
1348 		nmutexes = mopts.malloc_mt ? mopts.malloc_mutexes : 2;
1349 		STATS_INC(pool->other_pool);
1350 		for (i = 1; i < nmutexes; i++) {
1351 			u_int j = (argpool->mutex + i) & (nmutexes - 1);
1352 
1353 			pool->active--;
1354 			_MALLOC_UNLOCK(pool->mutex);
1355 			pool = mopts.malloc_pool[j];
1356 			_MALLOC_LOCK(pool->mutex);
1357 			pool->active++;
1358 			r = find(pool, p);
1359 			if (r != NULL) {
1360 				*saved_function = pool->func;
1361 				pool->func = argpool->func;
1362 				break;
1363 			}
1364 		}
1365 		if (r == NULL)
1366 			wrterror(argpool, "bogus pointer (double free?) %p", p);
1367 	}
1368 	*foundpool = pool;
1369 	return r;
1370 }
1371 
1372 static void
1373 ofree(struct dir_info **argpool, void *p, int clear, int check, size_t argsz)
1374 {
1375 	struct region_info *r;
1376 	struct dir_info *pool;
1377 	char *saved_function;
1378 	size_t sz;
1379 
1380 	r = findpool(p, *argpool, &pool, &saved_function);
1381 
1382 	REALSIZE(sz, r);
1383 	if (pool->mmap_flag) {
1384 		clear = 1;
1385 		if (!check)
1386 			argsz = sz;
1387 	}
1388 	if (check) {
1389 		if (sz <= MALLOC_MAXCHUNK) {
1390 			if (mopts.chunk_canaries && sz > 0) {
1391 				struct chunk_info *info =
1392 				    (struct chunk_info *)r->size;
1393 				uint32_t chunknum =
1394 				    find_chunknum(pool, info, p, 0);
1395 
1396 				if (info->bits[info->offset + chunknum] < argsz)
1397 					wrterror(pool, "recorded size %hu"
1398 					    " < %zu",
1399 					    info->bits[info->offset + chunknum],
1400 					    argsz);
1401 			} else {
1402 				if (sz < argsz)
1403 					wrterror(pool, "chunk size %zu < %zu",
1404 					    sz, argsz);
1405 			}
1406 		} else if (sz - mopts.malloc_guard < argsz) {
1407 			wrterror(pool, "recorded size %zu < %zu",
1408 			    sz - mopts.malloc_guard, argsz);
1409 		}
1410 	}
1411 	if (sz > MALLOC_MAXCHUNK) {
1412 		if (!MALLOC_MOVE_COND(sz)) {
1413 			if (r->p != p)
1414 				wrterror(pool, "bogus pointer %p", p);
1415 			if (mopts.chunk_canaries)
1416 				validate_canary(pool, p,
1417 				    sz - mopts.malloc_guard,
1418 				    PAGEROUND(sz - mopts.malloc_guard));
1419 		} else {
1420 			/* shifted towards the end */
1421 			if (p != MALLOC_MOVE(r->p, sz))
1422 				wrterror(pool, "bogus moved pointer %p", p);
1423 			p = r->p;
1424 		}
1425 		if (mopts.malloc_guard) {
1426 			if (sz < mopts.malloc_guard)
1427 				wrterror(pool, "guard size");
1428 			if (!mopts.malloc_freeunmap) {
1429 				if (mprotect((char *)p + PAGEROUND(sz) -
1430 				    mopts.malloc_guard, mopts.malloc_guard,
1431 				    PROT_READ | PROT_WRITE))
1432 					wrterror(pool, "mprotect");
1433 			}
1434 			STATS_SUB(pool->malloc_guarded, mopts.malloc_guard);
1435 		}
1436 		unmap(pool, p, PAGEROUND(sz), clear ? argsz : 0);
1437 		delete(pool, r);
1438 	} else {
1439 		/* Validate and optionally canary check */
1440 		struct chunk_info *info = (struct chunk_info *)r->size;
1441 		find_chunknum(pool, info, p, mopts.chunk_canaries);
1442 		if (!clear) {
1443 			void *tmp;
1444 			int i;
1445 
1446 			if (mopts.malloc_freecheck) {
1447 				for (i = 0; i <= MALLOC_DELAYED_CHUNK_MASK; i++)
1448 					if (p == pool->delayed_chunks[i])
1449 						wrterror(pool,
1450 						    "double free %p", p);
1451 			}
1452 			junk_free(pool->malloc_junk, p, sz);
1453 			i = getrbyte(pool) & MALLOC_DELAYED_CHUNK_MASK;
1454 			tmp = p;
1455 			p = pool->delayed_chunks[i];
1456 			if (tmp == p)
1457 				wrterror(pool, "double free %p", tmp);
1458 			pool->delayed_chunks[i] = tmp;
1459 			if (p != NULL) {
1460 				r = find(pool, p);
1461 				REALSIZE(sz, r);
1462 				if (r != NULL)
1463 					validate_junk(pool, p, sz);
1464 			}
1465 		} else if (argsz > 0) {
1466 			r = find(pool, p);
1467 			memset(p, 0, argsz);
1468 		}
1469 		if (p != NULL) {
1470 			if (r == NULL)
1471 				wrterror(pool,
1472 				    "bogus pointer (double free?) %p", p);
1473 			free_bytes(pool, r, p);
1474 		}
1475 	}
1476 
1477 	if (*argpool != pool) {
1478 		pool->func = saved_function;
1479 		*argpool = pool;
1480 	}
1481 }
1482 
1483 void
1484 free(void *ptr)
1485 {
1486 	struct dir_info *d;
1487 	int saved_errno = errno;
1488 
1489 	/* This is legal. */
1490 	if (ptr == NULL)
1491 		return;
1492 
1493 	d = getpool();
1494 	if (d == NULL)
1495 		wrterror(d, "free() called before allocation");
1496 	_MALLOC_LOCK(d->mutex);
1497 	d->func = "free";
1498 	if (d->active++) {
1499 		malloc_recurse(d);
1500 		return;
1501 	}
1502 	ofree(&d, ptr, 0, 0, 0);
1503 	d->active--;
1504 	_MALLOC_UNLOCK(d->mutex);
1505 	errno = saved_errno;
1506 }
1507 /*DEF_STRONG(free);*/
1508 
1509 static void
1510 freezero_p(void *ptr, size_t sz)
1511 {
1512 	explicit_bzero(ptr, sz);
1513 	free(ptr);
1514 }
1515 
1516 void
1517 freezero(void *ptr, size_t sz)
1518 {
1519 	struct dir_info *d;
1520 	int saved_errno = errno;
1521 
1522 	/* This is legal. */
1523 	if (ptr == NULL)
1524 		return;
1525 
1526 	if (!mopts.internal_funcs) {
1527 		freezero_p(ptr, sz);
1528 		return;
1529 	}
1530 
1531 	d = getpool();
1532 	if (d == NULL)
1533 		wrterror(d, "freezero() called before allocation");
1534 	_MALLOC_LOCK(d->mutex);
1535 	d->func = "freezero";
1536 	if (d->active++) {
1537 		malloc_recurse(d);
1538 		return;
1539 	}
1540 	ofree(&d, ptr, 1, 1, sz);
1541 	d->active--;
1542 	_MALLOC_UNLOCK(d->mutex);
1543 	errno = saved_errno;
1544 }
1545 DEF_WEAK(freezero);
1546 
1547 static void *
1548 orealloc(struct dir_info **argpool, void *p, size_t newsz, void *f)
1549 {
1550 	struct region_info *r;
1551 	struct dir_info *pool;
1552 	char *saved_function;
1553 	struct chunk_info *info;
1554 	size_t oldsz, goldsz, gnewsz;
1555 	void *q, *ret;
1556 	uint32_t chunknum;
1557 	int forced;
1558 
1559 	if (p == NULL)
1560 		return omalloc(*argpool, newsz, 0, f);
1561 
1562 	if (newsz >= SIZE_MAX - mopts.malloc_guard - MALLOC_PAGESIZE) {
1563 		errno = ENOMEM;
1564 		return  NULL;
1565 	}
1566 
1567 	r = findpool(p, *argpool, &pool, &saved_function);
1568 
1569 	REALSIZE(oldsz, r);
1570 	if (mopts.chunk_canaries && oldsz <= MALLOC_MAXCHUNK) {
1571 		info = (struct chunk_info *)r->size;
1572 		chunknum = find_chunknum(pool, info, p, 0);
1573 	}
1574 
1575 	goldsz = oldsz;
1576 	if (oldsz > MALLOC_MAXCHUNK) {
1577 		if (oldsz < mopts.malloc_guard)
1578 			wrterror(pool, "guard size");
1579 		oldsz -= mopts.malloc_guard;
1580 	}
1581 
1582 	gnewsz = newsz;
1583 	if (gnewsz > MALLOC_MAXCHUNK)
1584 		gnewsz += mopts.malloc_guard;
1585 
1586 	forced = mopts.malloc_realloc || pool->mmap_flag;
1587 	if (newsz > MALLOC_MAXCHUNK && oldsz > MALLOC_MAXCHUNK && !forced) {
1588 		/* First case: from n pages sized allocation to m pages sized
1589 		   allocation, m > n */
1590 		size_t roldsz = PAGEROUND(goldsz);
1591 		size_t rnewsz = PAGEROUND(gnewsz);
1592 
1593 		if (rnewsz < roldsz && rnewsz > roldsz / 2 &&
1594 		    roldsz - rnewsz < pool->malloc_cache * MALLOC_PAGESIZE &&
1595 		    !mopts.malloc_guard) {
1596 
1597 			ret = p;
1598 			goto done;
1599 		}
1600 
1601 		if (rnewsz > roldsz) {
1602 			/* try to extend existing region */
1603 			if (!mopts.malloc_guard) {
1604 				void *hint = (char *)r->p + roldsz;
1605 				size_t needed = rnewsz - roldsz;
1606 
1607 				STATS_INC(pool->cheap_realloc_tries);
1608 				q = MQUERY(hint, needed, pool->mmap_flag);
1609 				if (q == hint)
1610 					q = MMAPA(hint, needed, pool->mmap_flag);
1611 				else
1612 					q = MAP_FAILED;
1613 				if (q == hint) {
1614 					STATS_ADD(pool->malloc_used, needed);
1615 					if (pool->malloc_junk == 2)
1616 						memset(q, SOME_JUNK, needed);
1617 					r->size = gnewsz;
1618 					if (r->p != p) {
1619 						/* old pointer is moved */
1620 						memmove(r->p, p, oldsz);
1621 						p = r->p;
1622 					}
1623 					if (mopts.chunk_canaries)
1624 						fill_canary(p, newsz,
1625 						    PAGEROUND(newsz));
1626 					STATS_SETF(r, f);
1627 					STATS_INC(pool->cheap_reallocs);
1628 					ret = p;
1629 					goto done;
1630 				} else if (q != MAP_FAILED) {
1631 					if (munmap(q, needed))
1632 						wrterror(pool, "munmap %p", q);
1633 				}
1634 			}
1635 		} else if (rnewsz < roldsz) {
1636 			/* shrink number of pages */
1637 			if (mopts.malloc_guard) {
1638 				if (mprotect((char *)r->p + rnewsz -
1639 				    mopts.malloc_guard, mopts.malloc_guard,
1640 				    PROT_NONE))
1641 					wrterror(pool, "mprotect");
1642 			}
1643 			if (munmap((char *)r->p + rnewsz, roldsz - rnewsz))
1644 				wrterror(pool, "munmap %p", (char *)r->p + rnewsz);
1645 			r->size = gnewsz;
1646 			if (MALLOC_MOVE_COND(gnewsz)) {
1647 				void *pp = MALLOC_MOVE(r->p, gnewsz);
1648 				memmove(pp, p, newsz);
1649 				p = pp;
1650 			} else if (mopts.chunk_canaries)
1651 				fill_canary(p, newsz, PAGEROUND(newsz));
1652 			STATS_SETF(r, f);
1653 			ret = p;
1654 			goto done;
1655 		} else {
1656 			/* number of pages remains the same */
1657 			void *pp = r->p;
1658 
1659 			r->size = gnewsz;
1660 			if (MALLOC_MOVE_COND(gnewsz))
1661 				pp = MALLOC_MOVE(r->p, gnewsz);
1662 			if (p != pp) {
1663 				memmove(pp, p, oldsz < newsz ? oldsz : newsz);
1664 				p = pp;
1665 			}
1666 			if (p == r->p) {
1667 				if (newsz > oldsz && pool->malloc_junk == 2)
1668 					memset((char *)p + newsz, SOME_JUNK,
1669 					    rnewsz - mopts.malloc_guard -
1670 					    newsz);
1671 				if (mopts.chunk_canaries)
1672 					fill_canary(p, newsz, PAGEROUND(newsz));
1673 			}
1674 			STATS_SETF(r, f);
1675 			ret = p;
1676 			goto done;
1677 		}
1678 	}
1679 	if (oldsz <= MALLOC_MAXCHUNK && oldsz > 0 &&
1680 	    newsz <= MALLOC_MAXCHUNK && newsz > 0 &&
1681 	    1 << find_chunksize(newsz) == oldsz && !forced) {
1682 		/* do not reallocate if new size fits good in existing chunk */
1683 		if (pool->malloc_junk == 2)
1684 			memset((char *)p + newsz, SOME_JUNK, oldsz - newsz);
1685 		if (mopts.chunk_canaries) {
1686 			info->bits[info->offset + chunknum] = newsz;
1687 			fill_canary(p, newsz, info->size);
1688 		}
1689 		STATS_SETF(r, f);
1690 		ret = p;
1691 	} else if (newsz != oldsz || forced) {
1692 		/* create new allocation */
1693 		q = omalloc(pool, newsz, 0, f);
1694 		if (q == NULL) {
1695 			ret = NULL;
1696 			goto done;
1697 		}
1698 		if (newsz != 0 && oldsz != 0)
1699 			memcpy(q, p, oldsz < newsz ? oldsz : newsz);
1700 		ofree(&pool, p, 0, 0, 0);
1701 		ret = q;
1702 	} else {
1703 		/* oldsz == newsz */
1704 		if (newsz != 0)
1705 			wrterror(pool, "realloc internal inconsistency");
1706 		STATS_SETF(r, f);
1707 		ret = p;
1708 	}
1709 done:
1710 	if (*argpool != pool) {
1711 		pool->func = saved_function;
1712 		*argpool = pool;
1713 	}
1714 	return ret;
1715 }
1716 
1717 void *
1718 realloc(void *ptr, size_t size)
1719 {
1720 	struct dir_info *d;
1721 	void *r;
1722 	int saved_errno = errno;
1723 
1724 	PROLOGUE(getpool(), "realloc")
1725 	r = orealloc(&d, ptr, size, CALLER);
1726 	EPILOGUE()
1727 	return r;
1728 }
1729 /*DEF_STRONG(realloc);*/
1730 
1731 /*
1732  * This is sqrt(SIZE_MAX+1), as s1*s2 <= SIZE_MAX
1733  * if both s1 < MUL_NO_OVERFLOW and s2 < MUL_NO_OVERFLOW
1734  */
1735 #define MUL_NO_OVERFLOW	(1UL << (sizeof(size_t) * 4))
1736 
1737 void *
1738 calloc(size_t nmemb, size_t size)
1739 {
1740 	struct dir_info *d;
1741 	void *r;
1742 	int saved_errno = errno;
1743 
1744 	PROLOGUE(getpool(), "calloc")
1745 	if ((nmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1746 	    nmemb > 0 && SIZE_MAX / nmemb < size) {
1747 		d->active--;
1748 		_MALLOC_UNLOCK(d->mutex);
1749 		if (mopts.malloc_xmalloc)
1750 			wrterror(d, "out of memory");
1751 		errno = ENOMEM;
1752 		return NULL;
1753 	}
1754 
1755 	size *= nmemb;
1756 	r = omalloc(d, size, 1, CALLER);
1757 	EPILOGUE()
1758 	return r;
1759 }
1760 /*DEF_STRONG(calloc);*/
1761 
1762 void *
1763 calloc_conceal(size_t nmemb, size_t size)
1764 {
1765 	struct dir_info *d;
1766 	void *r;
1767 	int saved_errno = errno;
1768 
1769 	PROLOGUE(mopts.malloc_pool[0], "calloc_conceal")
1770 	if ((nmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1771 	    nmemb > 0 && SIZE_MAX / nmemb < size) {
1772 		d->active--;
1773 		_MALLOC_UNLOCK(d->mutex);
1774 		if (mopts.malloc_xmalloc)
1775 			wrterror(d, "out of memory");
1776 		errno = ENOMEM;
1777 		return NULL;
1778 	}
1779 
1780 	size *= nmemb;
1781 	r = omalloc(d, size, 1, CALLER);
1782 	EPILOGUE()
1783 	return r;
1784 }
1785 DEF_WEAK(calloc_conceal);
1786 
1787 static void *
1788 orecallocarray(struct dir_info **argpool, void *p, size_t oldsize,
1789     size_t newsize, void *f)
1790 {
1791 	struct region_info *r;
1792 	struct dir_info *pool;
1793 	char *saved_function;
1794 	void *newptr;
1795 	size_t sz;
1796 
1797 	if (p == NULL)
1798 		return omalloc(*argpool, newsize, 1, f);
1799 
1800 	if (oldsize == newsize)
1801 		return p;
1802 
1803 	r = findpool(p, *argpool, &pool, &saved_function);
1804 
1805 	REALSIZE(sz, r);
1806 	if (sz <= MALLOC_MAXCHUNK) {
1807 		if (mopts.chunk_canaries && sz > 0) {
1808 			struct chunk_info *info = (struct chunk_info *)r->size;
1809 			uint32_t chunknum = find_chunknum(pool, info, p, 0);
1810 
1811 			if (info->bits[info->offset + chunknum] != oldsize)
1812 				wrterror(pool, "recorded old size %hu != %zu",
1813 				    info->bits[info->offset + chunknum],
1814 				    oldsize);
1815 		}
1816 	} else if (oldsize < (sz - mopts.malloc_guard) / 2)
1817 		wrterror(pool, "recorded old size %zu != %zu",
1818 		    sz - mopts.malloc_guard, oldsize);
1819 
1820 	newptr = omalloc(pool, newsize, 0, f);
1821 	if (newptr == NULL)
1822 		goto done;
1823 
1824 	if (newsize > oldsize) {
1825 		memcpy(newptr, p, oldsize);
1826 		memset((char *)newptr + oldsize, 0, newsize - oldsize);
1827 	} else
1828 		memcpy(newptr, p, newsize);
1829 
1830 	ofree(&pool, p, 1, 0, oldsize);
1831 
1832 done:
1833 	if (*argpool != pool) {
1834 		pool->func = saved_function;
1835 		*argpool = pool;
1836 	}
1837 
1838 	return newptr;
1839 }
1840 
1841 static void *
1842 recallocarray_p(void *ptr, size_t oldnmemb, size_t newnmemb, size_t size)
1843 {
1844 	size_t oldsize, newsize;
1845 	void *newptr;
1846 
1847 	if (ptr == NULL)
1848 		return calloc(newnmemb, size);
1849 
1850 	if ((newnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1851 	    newnmemb > 0 && SIZE_MAX / newnmemb < size) {
1852 		errno = ENOMEM;
1853 		return NULL;
1854 	}
1855 	newsize = newnmemb * size;
1856 
1857 	if ((oldnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1858 	    oldnmemb > 0 && SIZE_MAX / oldnmemb < size) {
1859 		errno = EINVAL;
1860 		return NULL;
1861 	}
1862 	oldsize = oldnmemb * size;
1863 
1864 	/*
1865 	 * Don't bother too much if we're shrinking just a bit,
1866 	 * we do not shrink for series of small steps, oh well.
1867 	 */
1868 	if (newsize <= oldsize) {
1869 		size_t d = oldsize - newsize;
1870 
1871 		if (d < oldsize / 2 && d < MALLOC_PAGESIZE) {
1872 			memset((char *)ptr + newsize, 0, d);
1873 			return ptr;
1874 		}
1875 	}
1876 
1877 	newptr = malloc(newsize);
1878 	if (newptr == NULL)
1879 		return NULL;
1880 
1881 	if (newsize > oldsize) {
1882 		memcpy(newptr, ptr, oldsize);
1883 		memset((char *)newptr + oldsize, 0, newsize - oldsize);
1884 	} else
1885 		memcpy(newptr, ptr, newsize);
1886 
1887 	explicit_bzero(ptr, oldsize);
1888 	free(ptr);
1889 
1890 	return newptr;
1891 }
1892 
1893 void *
1894 recallocarray(void *ptr, size_t oldnmemb, size_t newnmemb, size_t size)
1895 {
1896 	struct dir_info *d;
1897 	size_t oldsize = 0, newsize;
1898 	void *r;
1899 	int saved_errno = errno;
1900 
1901 	if (!mopts.internal_funcs)
1902 		return recallocarray_p(ptr, oldnmemb, newnmemb, size);
1903 
1904 	PROLOGUE(getpool(), "recallocarray")
1905 
1906 	if ((newnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1907 	    newnmemb > 0 && SIZE_MAX / newnmemb < size) {
1908 		d->active--;
1909 		_MALLOC_UNLOCK(d->mutex);
1910 		if (mopts.malloc_xmalloc)
1911 			wrterror(d, "out of memory");
1912 		errno = ENOMEM;
1913 		return NULL;
1914 	}
1915 	newsize = newnmemb * size;
1916 
1917 	if (ptr != NULL) {
1918 		if ((oldnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1919 		    oldnmemb > 0 && SIZE_MAX / oldnmemb < size) {
1920 			d->active--;
1921 			_MALLOC_UNLOCK(d->mutex);
1922 			errno = EINVAL;
1923 			return NULL;
1924 		}
1925 		oldsize = oldnmemb * size;
1926 	}
1927 
1928 	r = orecallocarray(&d, ptr, oldsize, newsize, CALLER);
1929 	EPILOGUE()
1930 	return r;
1931 }
1932 DEF_WEAK(recallocarray);
1933 
1934 static void *
1935 mapalign(struct dir_info *d, size_t alignment, size_t sz, int zero_fill)
1936 {
1937 	char *p, *q;
1938 
1939 	if (alignment < MALLOC_PAGESIZE || ((alignment - 1) & alignment) != 0)
1940 		wrterror(d, "mapalign bad alignment");
1941 	if (sz != PAGEROUND(sz))
1942 		wrterror(d, "mapalign round");
1943 
1944 	/* Allocate sz + alignment bytes of memory, which must include a
1945 	 * subrange of size bytes that is properly aligned.  Unmap the
1946 	 * other bytes, and then return that subrange.
1947 	 */
1948 
1949 	/* We need sz + alignment to fit into a size_t. */
1950 	if (alignment > SIZE_MAX - sz)
1951 		return MAP_FAILED;
1952 
1953 	p = map(d, sz + alignment, zero_fill);
1954 	if (p == MAP_FAILED)
1955 		return MAP_FAILED;
1956 	q = (char *)(((uintptr_t)p + alignment - 1) & ~(alignment - 1));
1957 	if (q != p) {
1958 		if (munmap(p, q - p))
1959 			wrterror(d, "munmap %p", p);
1960 	}
1961 	if (munmap(q + sz, alignment - (q - p)))
1962 		wrterror(d, "munmap %p", q + sz);
1963 	STATS_SUB(d->malloc_used, alignment);
1964 
1965 	return q;
1966 }
1967 
1968 static void *
1969 omemalign(struct dir_info *pool, size_t alignment, size_t sz, int zero_fill,
1970     void *f)
1971 {
1972 	size_t psz;
1973 	void *p;
1974 
1975 	/* If between half a page and a page, avoid MALLOC_MOVE. */
1976 	if (sz > MALLOC_MAXCHUNK && sz < MALLOC_PAGESIZE)
1977 		sz = MALLOC_PAGESIZE;
1978 	if (alignment <= MALLOC_PAGESIZE) {
1979 		/*
1980 		 * max(size, alignment) is enough to assure the requested
1981 		 * alignment, since the allocator always allocates
1982 		 * power-of-two blocks.
1983 		 */
1984 		if (sz < alignment)
1985 			sz = alignment;
1986 		return omalloc(pool, sz, zero_fill, f);
1987 	}
1988 
1989 	if (sz >= SIZE_MAX - mopts.malloc_guard - MALLOC_PAGESIZE) {
1990 		errno = ENOMEM;
1991 		return NULL;
1992 	}
1993 
1994 	if (sz < MALLOC_PAGESIZE)
1995 		sz = MALLOC_PAGESIZE;
1996 	sz += mopts.malloc_guard;
1997 	psz = PAGEROUND(sz);
1998 
1999 	p = mapalign(pool, alignment, psz, zero_fill);
2000 	if (p == MAP_FAILED) {
2001 		errno = ENOMEM;
2002 		return NULL;
2003 	}
2004 
2005 	if (insert(pool, p, sz, f)) {
2006 		unmap(pool, p, psz, 0);
2007 		errno = ENOMEM;
2008 		return NULL;
2009 	}
2010 
2011 	if (mopts.malloc_guard) {
2012 		if (mprotect((char *)p + psz - mopts.malloc_guard,
2013 		    mopts.malloc_guard, PROT_NONE))
2014 			wrterror(pool, "mprotect");
2015 		STATS_ADD(pool->malloc_guarded, mopts.malloc_guard);
2016 	}
2017 
2018 	if (pool->malloc_junk == 2) {
2019 		if (zero_fill)
2020 			memset((char *)p + sz - mopts.malloc_guard,
2021 			    SOME_JUNK, psz - sz);
2022 		else
2023 			memset(p, SOME_JUNK, psz - mopts.malloc_guard);
2024 	} else if (mopts.chunk_canaries)
2025 		fill_canary(p, sz - mopts.malloc_guard,
2026 		    psz - mopts.malloc_guard);
2027 
2028 	return p;
2029 }
2030 
2031 int
2032 posix_memalign(void **memptr, size_t alignment, size_t size)
2033 {
2034 	struct dir_info *d;
2035 	int res, saved_errno = errno;
2036 	void *r;
2037 
2038 	/* Make sure that alignment is a large enough power of 2. */
2039 	if (((alignment - 1) & alignment) != 0 || alignment < sizeof(void *))
2040 		return EINVAL;
2041 
2042 	d = getpool();
2043 	if (d == NULL) {
2044 		_malloc_init(0);
2045 		d = getpool();
2046 	}
2047 	_MALLOC_LOCK(d->mutex);
2048 	d->func = "posix_memalign";
2049 	if (d->active++) {
2050 		malloc_recurse(d);
2051 		goto err;
2052 	}
2053 	r = omemalign(d, alignment, size, 0, CALLER);
2054 	d->active--;
2055 	_MALLOC_UNLOCK(d->mutex);
2056 	if (r == NULL) {
2057 		if (mopts.malloc_xmalloc)
2058 			wrterror(d, "out of memory");
2059 		goto err;
2060 	}
2061 	errno = saved_errno;
2062 	*memptr = r;
2063 	return 0;
2064 
2065 err:
2066 	res = errno;
2067 	errno = saved_errno;
2068 	return res;
2069 }
2070 /*DEF_STRONG(posix_memalign);*/
2071 
2072 void *
2073 aligned_alloc(size_t alignment, size_t size)
2074 {
2075 	struct dir_info *d;
2076 	int saved_errno = errno;
2077 	void *r;
2078 
2079 	/* Make sure that alignment is a positive power of 2. */
2080 	if (((alignment - 1) & alignment) != 0 || alignment == 0) {
2081 		errno = EINVAL;
2082 		return NULL;
2083 	};
2084 	/* Per spec, size should be a multiple of alignment */
2085 	if ((size & (alignment - 1)) != 0) {
2086 		errno = EINVAL;
2087 		return NULL;
2088 	}
2089 
2090 	PROLOGUE(getpool(), "aligned_alloc")
2091 	r = omemalign(d, alignment, size, 0, CALLER);
2092 	EPILOGUE()
2093 	return r;
2094 }
2095 /*DEF_STRONG(aligned_alloc);*/
2096 
2097 #ifdef MALLOC_STATS
2098 
2099 struct malloc_leak {
2100 	void *f;
2101 	size_t total_size;
2102 	int count;
2103 };
2104 
2105 struct leaknode {
2106 	RBT_ENTRY(leaknode) entry;
2107 	struct malloc_leak d;
2108 };
2109 
2110 static inline int
2111 leakcmp(const struct leaknode *e1, const struct leaknode *e2)
2112 {
2113 	return e1->d.f < e2->d.f ? -1 : e1->d.f > e2->d.f;
2114 }
2115 
2116 static RBT_HEAD(leaktree, leaknode) leakhead;
2117 RBT_PROTOTYPE(leaktree, leaknode, entry, leakcmp);
2118 RBT_GENERATE(leaktree, leaknode, entry, leakcmp);
2119 
2120 static void
2121 putleakinfo(void *f, size_t sz, int cnt)
2122 {
2123 	struct leaknode key, *p;
2124 	static struct leaknode *page;
2125 	static int used;
2126 
2127 	if (cnt == 0 || page == MAP_FAILED)
2128 		return;
2129 
2130 	key.d.f = f;
2131 	p = RBT_FIND(leaktree, &leakhead, &key);
2132 	if (p == NULL) {
2133 		if (page == NULL ||
2134 		    used >= MALLOC_PAGESIZE / sizeof(struct leaknode)) {
2135 			page = MMAP(MALLOC_PAGESIZE, 0);
2136 			if (page == MAP_FAILED)
2137 				return;
2138 			used = 0;
2139 		}
2140 		p = &page[used++];
2141 		p->d.f = f;
2142 		p->d.total_size = sz * cnt;
2143 		p->d.count = cnt;
2144 		RBT_INSERT(leaktree, &leakhead, p);
2145 	} else {
2146 		p->d.total_size += sz * cnt;
2147 		p->d.count += cnt;
2148 	}
2149 }
2150 
2151 static struct malloc_leak *malloc_leaks;
2152 
2153 static void
2154 dump_leaks(int fd)
2155 {
2156 	struct leaknode *p;
2157 	int i = 0;
2158 
2159 	dprintf(fd, "Leak report\n");
2160 	dprintf(fd, "                 f     sum      #    avg\n");
2161 	/* XXX only one page of summary */
2162 	if (malloc_leaks == NULL)
2163 		malloc_leaks = MMAP(MALLOC_PAGESIZE, 0);
2164 	if (malloc_leaks != MAP_FAILED)
2165 		memset(malloc_leaks, 0, MALLOC_PAGESIZE);
2166 	RBT_FOREACH(p, leaktree, &leakhead) {
2167 		dprintf(fd, "%18p %7zu %6u %6zu\n", p->d.f,
2168 		    p->d.total_size, p->d.count, p->d.total_size / p->d.count);
2169 		if (malloc_leaks == MAP_FAILED ||
2170 		    i >= MALLOC_PAGESIZE / sizeof(struct malloc_leak))
2171 			continue;
2172 		malloc_leaks[i].f = p->d.f;
2173 		malloc_leaks[i].total_size = p->d.total_size;
2174 		malloc_leaks[i].count = p->d.count;
2175 		i++;
2176 	}
2177 }
2178 
2179 static void
2180 dump_chunk(int fd, struct chunk_info *p, void *f, int fromfreelist)
2181 {
2182 	while (p != NULL) {
2183 		dprintf(fd, "chunk %18p %18p %4d %d/%d\n",
2184 		    p->page, ((p->bits[0] & 1) ? NULL : f),
2185 		    p->size, p->free, p->total);
2186 		if (!fromfreelist) {
2187 			if (p->bits[0] & 1)
2188 				putleakinfo(NULL, p->size, p->total - p->free);
2189 			else {
2190 				putleakinfo(f, p->size, 1);
2191 				putleakinfo(NULL, p->size,
2192 				    p->total - p->free - 1);
2193 			}
2194 			break;
2195 		}
2196 		p = LIST_NEXT(p, entries);
2197 		if (p != NULL)
2198 			dprintf(fd, "        ");
2199 	}
2200 }
2201 
2202 static void
2203 dump_free_chunk_info(int fd, struct dir_info *d)
2204 {
2205 	int i, j, count;
2206 	struct chunk_info *p;
2207 
2208 	dprintf(fd, "Free chunk structs:\n");
2209 	for (i = 0; i <= MALLOC_MAXSHIFT; i++) {
2210 		count = 0;
2211 		LIST_FOREACH(p, &d->chunk_info_list[i], entries)
2212 			count++;
2213 		for (j = 0; j < MALLOC_CHUNK_LISTS; j++) {
2214 			p = LIST_FIRST(&d->chunk_dir[i][j]);
2215 			if (p == NULL && count == 0)
2216 				continue;
2217 			dprintf(fd, "%2d) %3d ", i, count);
2218 			if (p != NULL)
2219 				dump_chunk(fd, p, NULL, 1);
2220 			else
2221 				dprintf(fd, "\n");
2222 		}
2223 	}
2224 
2225 }
2226 
2227 static void
2228 dump_free_page_info(int fd, struct dir_info *d)
2229 {
2230 	int i;
2231 
2232 	dprintf(fd, "Free pages cached: %zu\n", d->free_regions_size);
2233 	for (i = 0; i < d->malloc_cache; i++) {
2234 		if (d->free_regions[i].p != NULL) {
2235 			dprintf(fd, "%2d) ", i);
2236 			dprintf(fd, "free at %p: %zu\n",
2237 			    d->free_regions[i].p, d->free_regions[i].size);
2238 		}
2239 	}
2240 }
2241 
2242 static void
2243 malloc_dump1(int fd, int poolno, struct dir_info *d)
2244 {
2245 	size_t i, realsize;
2246 
2247 	dprintf(fd, "Malloc dir of %s pool %d at %p\n", __progname, poolno, d);
2248 	if (d == NULL)
2249 		return;
2250 	dprintf(fd, "J=%d cache=%u Fl=%x\n",
2251 	    d->malloc_junk, d->malloc_cache, d->mmap_flag);
2252 	dprintf(fd, "Region slots free %zu/%zu\n",
2253 		d->regions_free, d->regions_total);
2254 	dprintf(fd, "Finds %zu/%zu\n", d->finds, d->find_collisions);
2255 	dprintf(fd, "Inserts %zu/%zu\n", d->inserts, d->insert_collisions);
2256 	dprintf(fd, "Deletes %zu/%zu\n", d->deletes, d->delete_moves);
2257 	dprintf(fd, "Cheap reallocs %zu/%zu\n",
2258 	    d->cheap_reallocs, d->cheap_realloc_tries);
2259 	dprintf(fd, "Other pool searches %zu/%zu\n",
2260 	    d->other_pool, d->pool_searches);
2261 	dprintf(fd, "In use %zu\n", d->malloc_used);
2262 	dprintf(fd, "Guarded %zu\n", d->malloc_guarded);
2263 	dump_free_chunk_info(fd, d);
2264 	dump_free_page_info(fd, d);
2265 	dprintf(fd,
2266 	    "slot)  hash d  type               page                  f size [free/n]\n");
2267 	for (i = 0; i < d->regions_total; i++) {
2268 		if (d->r[i].p != NULL) {
2269 			size_t h = hash(d->r[i].p) &
2270 			    (d->regions_total - 1);
2271 			dprintf(fd, "%4zx) #%4zx %zd ",
2272 			    i, h, h - i);
2273 			REALSIZE(realsize, &d->r[i]);
2274 			if (realsize > MALLOC_MAXCHUNK) {
2275 				putleakinfo(d->r[i].f, realsize, 1);
2276 				dprintf(fd,
2277 				    "pages %18p %18p %zu\n", d->r[i].p,
2278 				    d->r[i].f, realsize);
2279 			} else
2280 				dump_chunk(fd,
2281 				    (struct chunk_info *)d->r[i].size,
2282 				    d->r[i].f, 0);
2283 		}
2284 	}
2285 	dump_leaks(fd);
2286 	dprintf(fd, "\n");
2287 }
2288 
2289 void
2290 malloc_dump(int fd, int poolno, struct dir_info *pool)
2291 {
2292 	int i;
2293 	void *p;
2294 	struct region_info *r;
2295 	int saved_errno = errno;
2296 
2297 	if (pool == NULL)
2298 		return;
2299 	for (i = 0; i < MALLOC_DELAYED_CHUNK_MASK + 1; i++) {
2300 		p = pool->delayed_chunks[i];
2301 		if (p == NULL)
2302 			continue;
2303 		r = find(pool, p);
2304 		if (r == NULL)
2305 			wrterror(pool, "bogus pointer in malloc_dump %p", p);
2306 		free_bytes(pool, r, p);
2307 		pool->delayed_chunks[i] = NULL;
2308 	}
2309 	/* XXX leak when run multiple times */
2310 	RBT_INIT(leaktree, &leakhead);
2311 	malloc_dump1(fd, poolno, pool);
2312 	errno = saved_errno;
2313 }
2314 DEF_WEAK(malloc_dump);
2315 
2316 void
2317 malloc_gdump(int fd)
2318 {
2319 	int i;
2320 	int saved_errno = errno;
2321 
2322 	for (i = 0; i < mopts.malloc_mutexes; i++)
2323 		malloc_dump(fd, i, mopts.malloc_pool[i]);
2324 
2325 	errno = saved_errno;
2326 }
2327 DEF_WEAK(malloc_gdump);
2328 
2329 static void
2330 malloc_exit(void)
2331 {
2332 	int save_errno = errno, fd, i;
2333 
2334 	fd = open("malloc.out", O_RDWR|O_APPEND);
2335 	if (fd != -1) {
2336 		dprintf(fd, "******** Start dump %s *******\n", __progname);
2337 		dprintf(fd,
2338 		    "MT=%d M=%u I=%d F=%d U=%d J=%d R=%d X=%d C=%d cache=%u G=%zu\n",
2339 		    mopts.malloc_mt, mopts.malloc_mutexes,
2340 		    mopts.internal_funcs, mopts.malloc_freecheck,
2341 		    mopts.malloc_freeunmap, mopts.def_malloc_junk,
2342 		    mopts.malloc_realloc, mopts.malloc_xmalloc,
2343 		    mopts.chunk_canaries, mopts.def_malloc_cache,
2344 		    mopts.malloc_guard);
2345 
2346 		for (i = 0; i < mopts.malloc_mutexes; i++)
2347 			malloc_dump(fd, i, mopts.malloc_pool[i]);
2348 		dprintf(fd, "******** End dump %s *******\n", __progname);
2349 		close(fd);
2350 	} else
2351 		dprintf(STDERR_FILENO,
2352 		    "malloc() warning: Couldn't dump stats\n");
2353 	errno = save_errno;
2354 }
2355 
2356 #endif /* MALLOC_STATS */
2357