xref: /openbsd-src/lib/libc/stdlib/malloc.c (revision 9a36959cd9c6cca283ad6b41d8d84a03e03b9530)
1 /*	$OpenBSD: malloc.c,v 1.285 2023/06/04 06:58:33 otto Exp $	*/
2 /*
3  * Copyright (c) 2008, 2010, 2011, 2016, 2023 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 #ifndef MALLOC_SMALL
27 #define MALLOC_STATS
28 #endif
29 
30 #include <sys/types.h>
31 #include <sys/queue.h>
32 #include <sys/mman.h>
33 #include <sys/sysctl.h>
34 #include <uvm/uvmexp.h>
35 #include <errno.h>
36 #include <stdarg.h>
37 #include <stdint.h>
38 #include <stdio.h>
39 #include <stdlib.h>
40 #include <string.h>
41 #include <unistd.h>
42 
43 #ifdef MALLOC_STATS
44 #include <sys/tree.h>
45 #include <sys/ktrace.h>
46 #include <dlfcn.h>
47 #endif
48 
49 #include "thread_private.h"
50 #include <tib.h>
51 
52 #define MALLOC_PAGESHIFT	_MAX_PAGE_SHIFT
53 
54 #define MALLOC_MINSHIFT		4
55 #define MALLOC_MAXSHIFT		(MALLOC_PAGESHIFT - 1)
56 #define MALLOC_PAGESIZE		(1UL << MALLOC_PAGESHIFT)
57 #define MALLOC_MINSIZE		(1UL << MALLOC_MINSHIFT)
58 #define MALLOC_PAGEMASK		(MALLOC_PAGESIZE - 1)
59 #define MASK_POINTER(p)		((void *)(((uintptr_t)(p)) & ~MALLOC_PAGEMASK))
60 
61 #define MALLOC_MAXCHUNK		(1 << MALLOC_MAXSHIFT)
62 #define MALLOC_MAXCACHE		256
63 #define MALLOC_DELAYED_CHUNK_MASK	15
64 #ifdef MALLOC_STATS
65 #define MALLOC_INITIAL_REGIONS	512
66 #else
67 #define MALLOC_INITIAL_REGIONS	(MALLOC_PAGESIZE / sizeof(struct region_info))
68 #endif
69 #define MALLOC_DEFAULT_CACHE	64
70 #define MALLOC_CHUNK_LISTS	4
71 #define CHUNK_CHECK_LENGTH	32
72 
73 #define B2SIZE(b)		((b) * MALLOC_MINSIZE)
74 #define B2ALLOC(b)		((b) == 0 ? MALLOC_MINSIZE : \
75 				    (b) * MALLOC_MINSIZE)
76 #define BUCKETS 		(MALLOC_MAXCHUNK / MALLOC_MINSIZE)
77 
78 /*
79  * We move allocations between half a page and a whole page towards the end,
80  * subject to alignment constraints. This is the extra headroom we allow.
81  * Set to zero to be the most strict.
82  */
83 #define MALLOC_LEEWAY		0
84 #define MALLOC_MOVE_COND(sz)	((sz) - mopts.malloc_guard < 		\
85 				    MALLOC_PAGESIZE - MALLOC_LEEWAY)
86 #define MALLOC_MOVE(p, sz)  	(((char *)(p)) +			\
87 				    ((MALLOC_PAGESIZE - MALLOC_LEEWAY -	\
88 			    	    ((sz) - mopts.malloc_guard)) & 	\
89 				    ~(MALLOC_MINSIZE - 1)))
90 
91 #define PAGEROUND(x)  (((x) + (MALLOC_PAGEMASK)) & ~MALLOC_PAGEMASK)
92 
93 /*
94  * What to use for Junk.  This is the byte value we use to fill with
95  * when the 'J' option is enabled. Use SOME_JUNK right after alloc,
96  * and SOME_FREEJUNK right before free.
97  */
98 #define SOME_JUNK		0xdb	/* deadbeef */
99 #define SOME_FREEJUNK		0xdf	/* dead, free */
100 #define SOME_FREEJUNK_ULL	0xdfdfdfdfdfdfdfdfULL
101 
102 #define MMAP(sz,f)	mmap(NULL, (sz), PROT_READ | PROT_WRITE, \
103     MAP_ANON | MAP_PRIVATE | (f), -1, 0)
104 
105 #define MMAPNONE(sz,f)	mmap(NULL, (sz), PROT_NONE, \
106     MAP_ANON | MAP_PRIVATE | (f), -1, 0)
107 
108 #define MMAPA(a,sz,f)	mmap((a), (sz), PROT_READ | PROT_WRITE, \
109     MAP_ANON | MAP_PRIVATE | (f), -1, 0)
110 
111 struct region_info {
112 	void *p;		/* page; low bits used to mark chunks */
113 	uintptr_t size;		/* size for pages, or chunk_info pointer */
114 #ifdef MALLOC_STATS
115 	void *f;		/* where allocated from */
116 #endif
117 };
118 
119 LIST_HEAD(chunk_head, chunk_info);
120 
121 /*
122  * Two caches, one for "small" regions, one for "big".
123  * Small cache is an array per size, big cache is one array with different
124  * sized regions
125  */
126 #define MAX_SMALLCACHEABLE_SIZE	32
127 #define MAX_BIGCACHEABLE_SIZE	512
128 /* If the total # of pages is larger than this, evict before inserting */
129 #define BIGCACHE_FILL(sz)	(MAX_BIGCACHEABLE_SIZE * (sz) / 4)
130 
131 struct smallcache {
132 	void **pages;
133 	ushort length;
134 	ushort max;
135 };
136 
137 struct bigcache {
138 	void *page;
139 	size_t psize;
140 };
141 
142 struct dir_info {
143 	u_int32_t canary1;
144 	int active;			/* status of malloc */
145 	struct region_info *r;		/* region slots */
146 	size_t regions_total;		/* number of region slots */
147 	size_t regions_free;		/* number of free slots */
148 	size_t rbytesused;		/* random bytes used */
149 	char *func;			/* current function */
150 	int malloc_junk;		/* junk fill? */
151 	int mmap_flag;			/* extra flag for mmap */
152 	int mutex;
153 	int malloc_mt;			/* multi-threaded mode? */
154 					/* lists of free chunk info structs */
155 	struct chunk_head chunk_info_list[BUCKETS + 1];
156 					/* lists of chunks with free slots */
157 	struct chunk_head chunk_dir[BUCKETS + 1][MALLOC_CHUNK_LISTS];
158 					/* delayed free chunk slots */
159 	void *delayed_chunks[MALLOC_DELAYED_CHUNK_MASK + 1];
160 	u_char rbytes[32];		/* random bytes */
161 					/* free pages cache */
162 	struct smallcache smallcache[MAX_SMALLCACHEABLE_SIZE];
163 	size_t bigcache_used;
164 	size_t bigcache_size;
165 	struct bigcache *bigcache;
166 	void *chunk_pages;
167 	size_t chunk_pages_used;
168 #ifdef MALLOC_STATS
169 	size_t inserts;
170 	size_t insert_collisions;
171 	size_t finds;
172 	size_t find_collisions;
173 	size_t deletes;
174 	size_t delete_moves;
175 	size_t cheap_realloc_tries;
176 	size_t cheap_reallocs;
177 	size_t malloc_used;		/* bytes allocated */
178 	size_t malloc_guarded;		/* bytes used for guards */
179 	size_t pool_searches;		/* searches for pool */
180 	size_t other_pool;		/* searches in other pool */
181 #define STATS_ADD(x,y)	((x) += (y))
182 #define STATS_SUB(x,y)	((x) -= (y))
183 #define STATS_INC(x)	((x)++)
184 #define STATS_ZERO(x)	((x) = 0)
185 #define STATS_SETF(x,y)	((x)->f = (y))
186 #else
187 #define STATS_ADD(x,y)	/* nothing */
188 #define STATS_SUB(x,y)	/* nothing */
189 #define STATS_INC(x)	/* nothing */
190 #define STATS_ZERO(x)	/* nothing */
191 #define STATS_SETF(x,y)	/* nothing */
192 #endif /* MALLOC_STATS */
193 	u_int32_t canary2;
194 };
195 
196 static void unmap(struct dir_info *d, void *p, size_t sz, size_t clear);
197 
198 /*
199  * This structure describes a page worth of chunks.
200  *
201  * How many bits per u_short in the bitmap
202  */
203 #define MALLOC_BITS		(NBBY * sizeof(u_short))
204 struct chunk_info {
205 	LIST_ENTRY(chunk_info) entries;
206 	void *page;			/* pointer to the page */
207 	u_short canary;
208 	u_short bucket;
209 	u_short free;			/* how many free chunks */
210 	u_short total;			/* how many chunks */
211 	u_short offset;			/* requested size table offset */
212 	u_short bits[1];		/* which chunks are free */
213 };
214 
215 struct malloc_readonly {
216 					/* Main bookkeeping information */
217 	struct dir_info *malloc_pool[_MALLOC_MUTEXES];
218 	u_int	malloc_mutexes;		/* how much in actual use? */
219 	int	malloc_freecheck;	/* Extensive double free check */
220 	int	malloc_freeunmap;	/* mprotect free pages PROT_NONE? */
221 	int	def_malloc_junk;	/* junk fill? */
222 	int	malloc_realloc;		/* always realloc? */
223 	int	malloc_xmalloc;		/* xmalloc behaviour? */
224 	u_int	chunk_canaries;		/* use canaries after chunks? */
225 	int	internal_funcs;		/* use better recallocarray/freezero? */
226 	u_int	def_maxcache;		/* free pages we cache */
227 	u_int	junk_loc;		/* variation in location of junk */
228 	size_t	malloc_guard;		/* use guard pages after allocations? */
229 #ifdef MALLOC_STATS
230 	int	malloc_stats;		/* dump leak report at end */
231 	int	malloc_verbose;		/* dump verbose statistics at end */
232 #define	DO_STATS	mopts.malloc_stats
233 #else
234 #define	DO_STATS	0
235 #endif
236 	u_int32_t malloc_canary;	/* Matched against ones in pool */
237 };
238 
239 
240 /* This object is mapped PROT_READ after initialisation to prevent tampering */
241 static union {
242 	struct malloc_readonly mopts;
243 	u_char _pad[MALLOC_PAGESIZE];
244 } malloc_readonly __attribute__((aligned(MALLOC_PAGESIZE)))
245 		__attribute__((section(".openbsd.mutable")));
246 #define mopts	malloc_readonly.mopts
247 
248 char		*malloc_options;	/* compile-time options */
249 
250 static __dead void wrterror(struct dir_info *d, char *msg, ...)
251     __attribute__((__format__ (printf, 2, 3)));
252 
253 #ifdef MALLOC_STATS
254 void malloc_dump(void);
255 PROTO_NORMAL(malloc_dump);
256 static void malloc_exit(void);
257 #endif
258 #define CALLER	(DO_STATS ? __builtin_return_address(0) : NULL)
259 
260 /* low bits of r->p determine size: 0 means >= page size and r->size holding
261  * real size, otherwise low bits is the bucket + 1
262  */
263 #define REALSIZE(sz, r)						\
264 	(sz) = (uintptr_t)(r)->p & MALLOC_PAGEMASK,		\
265 	(sz) = ((sz) == 0 ? (r)->size : B2SIZE((sz) - 1))
266 
267 static inline size_t
268 hash(void *p)
269 {
270 	size_t sum;
271 	uintptr_t u;
272 
273 	u = (uintptr_t)p >> MALLOC_PAGESHIFT;
274 	sum = u;
275 	sum = (sum << 7) - sum + (u >> 16);
276 #ifdef __LP64__
277 	sum = (sum << 7) - sum + (u >> 32);
278 	sum = (sum << 7) - sum + (u >> 48);
279 #endif
280 	return sum;
281 }
282 
283 static inline struct dir_info *
284 getpool(void)
285 {
286 	if (mopts.malloc_pool[1] == NULL || !mopts.malloc_pool[1]->malloc_mt)
287 		return mopts.malloc_pool[1];
288 	else	/* first one reserved for special pool */
289 		return mopts.malloc_pool[1 + TIB_GET()->tib_tid %
290 		    (mopts.malloc_mutexes - 1)];
291 }
292 
293 static __dead void
294 wrterror(struct dir_info *d, char *msg, ...)
295 {
296 	int		saved_errno = errno;
297 	va_list		ap;
298 
299 	dprintf(STDERR_FILENO, "%s(%d) in %s(): ", __progname,
300 	    getpid(), (d != NULL && d->func) ? d->func : "unknown");
301 	va_start(ap, msg);
302 	vdprintf(STDERR_FILENO, msg, ap);
303 	va_end(ap);
304 	dprintf(STDERR_FILENO, "\n");
305 
306 #ifdef MALLOC_STATS
307 	if (DO_STATS && mopts.malloc_verbose)
308 		malloc_dump();
309 #endif
310 
311 	errno = saved_errno;
312 
313 	abort();
314 }
315 
316 static void
317 rbytes_init(struct dir_info *d)
318 {
319 	arc4random_buf(d->rbytes, sizeof(d->rbytes));
320 	/* add 1 to account for using d->rbytes[0] */
321 	d->rbytesused = 1 + d->rbytes[0] % (sizeof(d->rbytes) / 2);
322 }
323 
324 static inline u_char
325 getrbyte(struct dir_info *d)
326 {
327 	u_char x;
328 
329 	if (d->rbytesused >= sizeof(d->rbytes))
330 		rbytes_init(d);
331 	x = d->rbytes[d->rbytesused++];
332 	return x;
333 }
334 
335 static void
336 omalloc_parseopt(char opt)
337 {
338 	switch (opt) {
339 	case '+':
340 		mopts.malloc_mutexes <<= 1;
341 		if (mopts.malloc_mutexes > _MALLOC_MUTEXES)
342 			mopts.malloc_mutexes = _MALLOC_MUTEXES;
343 		break;
344 	case '-':
345 		mopts.malloc_mutexes >>= 1;
346 		if (mopts.malloc_mutexes < 2)
347 			mopts.malloc_mutexes = 2;
348 		break;
349 	case '>':
350 		mopts.def_maxcache <<= 1;
351 		if (mopts.def_maxcache > MALLOC_MAXCACHE)
352 			mopts.def_maxcache = MALLOC_MAXCACHE;
353 		break;
354 	case '<':
355 		mopts.def_maxcache >>= 1;
356 		break;
357 	case 'c':
358 		mopts.chunk_canaries = 0;
359 		break;
360 	case 'C':
361 		mopts.chunk_canaries = 1;
362 		break;
363 #ifdef MALLOC_STATS
364 	case 'd':
365 		mopts.malloc_stats = 0;
366 		break;
367 	case 'D':
368 		mopts.malloc_stats = 1;
369 		break;
370 #endif /* MALLOC_STATS */
371 	case 'f':
372 		mopts.malloc_freecheck = 0;
373 		mopts.malloc_freeunmap = 0;
374 		break;
375 	case 'F':
376 		mopts.malloc_freecheck = 1;
377 		mopts.malloc_freeunmap = 1;
378 		break;
379 	case 'g':
380 		mopts.malloc_guard = 0;
381 		break;
382 	case 'G':
383 		mopts.malloc_guard = MALLOC_PAGESIZE;
384 		break;
385 	case 'j':
386 		if (mopts.def_malloc_junk > 0)
387 			mopts.def_malloc_junk--;
388 		break;
389 	case 'J':
390 		if (mopts.def_malloc_junk < 2)
391 			mopts.def_malloc_junk++;
392 		break;
393 	case 'r':
394 		mopts.malloc_realloc = 0;
395 		break;
396 	case 'R':
397 		mopts.malloc_realloc = 1;
398 		break;
399 	case 'u':
400 		mopts.malloc_freeunmap = 0;
401 		break;
402 	case 'U':
403 		mopts.malloc_freeunmap = 1;
404 		break;
405 #ifdef MALLOC_STATS
406 	case 'v':
407 		mopts.malloc_verbose = 0;
408 		break;
409 	case 'V':
410 		mopts.malloc_verbose = 1;
411 		break;
412 #endif /* MALLOC_STATS */
413 	case 'x':
414 		mopts.malloc_xmalloc = 0;
415 		break;
416 	case 'X':
417 		mopts.malloc_xmalloc = 1;
418 		break;
419 	default:
420 		dprintf(STDERR_FILENO, "malloc() warning: "
421                     "unknown char in MALLOC_OPTIONS\n");
422 		break;
423 	}
424 }
425 
426 static void
427 omalloc_init(void)
428 {
429 	char *p, *q, b[16];
430 	int i, j;
431 	const int mib[2] = { CTL_VM, VM_MALLOC_CONF };
432 	size_t sb;
433 
434 	/*
435 	 * Default options
436 	 */
437 	mopts.malloc_mutexes = 8;
438 	mopts.def_malloc_junk = 1;
439 	mopts.def_maxcache = MALLOC_DEFAULT_CACHE;
440 
441 	for (i = 0; i < 3; i++) {
442 		switch (i) {
443 		case 0:
444 			sb = sizeof(b);
445 			j = sysctl(mib, 2, b, &sb, NULL, 0);
446 			if (j != 0)
447 				continue;
448 			p = b;
449 			break;
450 		case 1:
451 			if (issetugid() == 0)
452 				p = getenv("MALLOC_OPTIONS");
453 			else
454 				continue;
455 			break;
456 		case 2:
457 			p = malloc_options;
458 			break;
459 		default:
460 			p = NULL;
461 		}
462 
463 		for (; p != NULL && *p != '\0'; p++) {
464 			switch (*p) {
465 			case 'S':
466 				for (q = "CFGJ"; *q != '\0'; q++)
467 					omalloc_parseopt(*q);
468 				mopts.def_maxcache = 0;
469 				break;
470 			case 's':
471 				for (q = "cfgj"; *q != '\0'; q++)
472 					omalloc_parseopt(*q);
473 				mopts.def_maxcache = MALLOC_DEFAULT_CACHE;
474 				break;
475 			default:
476 				omalloc_parseopt(*p);
477 				break;
478 			}
479 		}
480 	}
481 
482 #ifdef MALLOC_STATS
483 	if (DO_STATS && (atexit(malloc_exit) == -1)) {
484 		dprintf(STDERR_FILENO, "malloc() warning: atexit(2) failed."
485 		    " Will not be able to dump stats on exit\n");
486 	}
487 #endif
488 
489 	while ((mopts.malloc_canary = arc4random()) == 0)
490 		;
491 	mopts.junk_loc = arc4random();
492 	if (mopts.chunk_canaries)
493 		do {
494 			mopts.chunk_canaries = arc4random();
495 		} while ((u_char)mopts.chunk_canaries == 0 ||
496 		    (u_char)mopts.chunk_canaries == SOME_FREEJUNK);
497 }
498 
499 static void
500 omalloc_poolinit(struct dir_info *d, int mmap_flag)
501 {
502 	int i, j;
503 
504 	d->r = NULL;
505 	d->rbytesused = sizeof(d->rbytes);
506 	d->regions_free = d->regions_total = 0;
507 	for (i = 0; i <= BUCKETS; 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 	d->mmap_flag = mmap_flag;
513 	d->malloc_junk = mopts.def_malloc_junk;
514 	d->canary1 = mopts.malloc_canary ^ (u_int32_t)(uintptr_t)d;
515 	d->canary2 = ~d->canary1;
516 }
517 
518 static int
519 omalloc_grow(struct dir_info *d)
520 {
521 	size_t newtotal;
522 	size_t newsize;
523 	size_t mask;
524 	size_t i, oldpsz;
525 	struct region_info *p;
526 
527 	if (d->regions_total > SIZE_MAX / sizeof(struct region_info) / 2)
528 		return 1;
529 
530 	newtotal = d->regions_total == 0 ? MALLOC_INITIAL_REGIONS :
531 	    d->regions_total * 2;
532 	newsize = PAGEROUND(newtotal * sizeof(struct region_info));
533 	mask = newtotal - 1;
534 
535 	/* Don't use cache here, we don't want user uaf touch this */
536 	p = MMAP(newsize, d->mmap_flag);
537 	if (p == MAP_FAILED)
538 		return 1;
539 
540 	STATS_ADD(d->malloc_used, newsize);
541 	STATS_ZERO(d->inserts);
542 	STATS_ZERO(d->insert_collisions);
543 	for (i = 0; i < d->regions_total; i++) {
544 		void *q = d->r[i].p;
545 		if (q != NULL) {
546 			size_t index = hash(q) & mask;
547 			STATS_INC(d->inserts);
548 			while (p[index].p != NULL) {
549 				index = (index - 1) & mask;
550 				STATS_INC(d->insert_collisions);
551 			}
552 			p[index] = d->r[i];
553 		}
554 	}
555 
556 	if (d->regions_total > 0) {
557 		oldpsz = PAGEROUND(d->regions_total * sizeof(struct region_info));
558 		/* clear to avoid meta info ending up in the cache */
559 		unmap(d, d->r, oldpsz, oldpsz);
560 	}
561 	d->regions_free += newtotal - d->regions_total;
562 	d->regions_total = newtotal;
563 	d->r = p;
564 	return 0;
565 }
566 
567 /*
568  * The hashtable uses the assumption that p is never NULL. This holds since
569  * non-MAP_FIXED mappings with hint 0 start at BRKSIZ.
570  */
571 static int
572 insert(struct dir_info *d, void *p, size_t sz, void *f)
573 {
574 	size_t index;
575 	size_t mask;
576 	void *q;
577 
578 	if (d->regions_free * 4 < d->regions_total || d->regions_total == 0) {
579 		if (omalloc_grow(d))
580 			return 1;
581 	}
582 	mask = d->regions_total - 1;
583 	index = hash(p) & mask;
584 	q = d->r[index].p;
585 	STATS_INC(d->inserts);
586 	while (q != NULL) {
587 		index = (index - 1) & mask;
588 		q = d->r[index].p;
589 		STATS_INC(d->insert_collisions);
590 	}
591 	d->r[index].p = p;
592 	d->r[index].size = sz;
593 	STATS_SETF(&d->r[index], f);
594 	d->regions_free--;
595 	return 0;
596 }
597 
598 static struct region_info *
599 find(struct dir_info *d, void *p)
600 {
601 	size_t index;
602 	size_t mask = d->regions_total - 1;
603 	void *q, *r;
604 
605 	if (mopts.malloc_canary != (d->canary1 ^ (u_int32_t)(uintptr_t)d) ||
606 	    d->canary1 != ~d->canary2)
607 		wrterror(d, "internal struct corrupt");
608 	if (d->r == NULL)
609 		return NULL;
610 	p = MASK_POINTER(p);
611 	index = hash(p) & mask;
612 	r = d->r[index].p;
613 	q = MASK_POINTER(r);
614 	STATS_INC(d->finds);
615 	while (q != p && r != NULL) {
616 		index = (index - 1) & mask;
617 		r = d->r[index].p;
618 		q = MASK_POINTER(r);
619 		STATS_INC(d->find_collisions);
620 	}
621 	return (q == p && r != NULL) ? &d->r[index] : NULL;
622 }
623 
624 static void
625 delete(struct dir_info *d, struct region_info *ri)
626 {
627 	/* algorithm R, Knuth Vol III section 6.4 */
628 	size_t mask = d->regions_total - 1;
629 	size_t i, j, r;
630 
631 	if (d->regions_total & (d->regions_total - 1))
632 		wrterror(d, "regions_total not 2^x");
633 	d->regions_free++;
634 	STATS_INC(d->deletes);
635 
636 	i = ri - d->r;
637 	for (;;) {
638 		d->r[i].p = NULL;
639 		d->r[i].size = 0;
640 		j = i;
641 		for (;;) {
642 			i = (i - 1) & mask;
643 			if (d->r[i].p == NULL)
644 				return;
645 			r = hash(d->r[i].p) & mask;
646 			if ((i <= r && r < j) || (r < j && j < i) ||
647 			    (j < i && i <= r))
648 				continue;
649 			d->r[j] = d->r[i];
650 			STATS_INC(d->delete_moves);
651 			break;
652 		}
653 
654 	}
655 }
656 
657 static inline void
658 junk_free(int junk, void *p, size_t sz)
659 {
660 	size_t i, step = 1;
661 	uint64_t *lp = p;
662 
663 	if (junk == 0 || sz == 0)
664 		return;
665 	sz /= sizeof(uint64_t);
666 	if (junk == 1) {
667 		if (sz > MALLOC_PAGESIZE / sizeof(uint64_t))
668 			sz = MALLOC_PAGESIZE / sizeof(uint64_t);
669 		step = sz / 4;
670 		if (step == 0)
671 			step = 1;
672 	}
673 	/* Do not always put the free junk bytes in the same spot.
674 	   There is modulo bias here, but we ignore that. */
675 	for (i = mopts.junk_loc % step; 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 	/* see junk_free */
696 	for (i = mopts.junk_loc % step; i < sz; i += step) {
697 		if (lp[i] != SOME_FREEJUNK_ULL)
698 			wrterror(pool, "write after free %p", p);
699 	}
700 }
701 
702 
703 /*
704  * Cache maintenance.
705  * Opposed to the regular region data structure, the sizes in the
706  * cache are in MALLOC_PAGESIZE units.
707  */
708 static void
709 unmap(struct dir_info *d, void *p, size_t sz, size_t clear)
710 {
711 	size_t psz = sz >> MALLOC_PAGESHIFT;
712 	void *r;
713 	u_short i;
714 	struct smallcache *cache;
715 
716 	if (sz != PAGEROUND(sz) || psz == 0)
717 		wrterror(d, "munmap round");
718 
719 	if (d->bigcache_size > 0 && psz > MAX_SMALLCACHEABLE_SIZE &&
720 	    psz <= MAX_BIGCACHEABLE_SIZE) {
721 		u_short base = getrbyte(d);
722 		u_short j;
723 
724 		/* don't look through all slots */
725 		for (j = 0; j < d->bigcache_size / 4; j++) {
726 			i = (base + j) & (d->bigcache_size - 1);
727 			if (d->bigcache_used <
728 			    BIGCACHE_FILL(d->bigcache_size))  {
729 				if (d->bigcache[i].psize == 0)
730 					break;
731 			} else {
732 				if (d->bigcache[i].psize != 0)
733 					break;
734 			}
735 		}
736 		/* if we didn't find a preferred slot, use random one */
737 		if (d->bigcache[i].psize != 0) {
738 			size_t tmp;
739 
740 			r = d->bigcache[i].page;
741 			d->bigcache_used -= d->bigcache[i].psize;
742 			tmp = d->bigcache[i].psize << MALLOC_PAGESHIFT;
743 			if (!mopts.malloc_freeunmap)
744 				validate_junk(d, r, tmp);
745 			if (munmap(r, tmp))
746 				 wrterror(d, "munmap %p", r);
747 			STATS_SUB(d->malloc_used, tmp);
748 		}
749 
750 		if (clear > 0)
751 			explicit_bzero(p, clear);
752 		if (mopts.malloc_freeunmap) {
753 			if (mprotect(p, sz, PROT_NONE))
754 				wrterror(d, "mprotect %p", r);
755 		} else
756 			junk_free(d->malloc_junk, p, sz);
757 		d->bigcache[i].page = p;
758 		d->bigcache[i].psize = psz;
759 		d->bigcache_used += psz;
760 		return;
761 	}
762 	if (psz > MAX_SMALLCACHEABLE_SIZE || d->smallcache[psz - 1].max == 0) {
763 		if (munmap(p, sz))
764 			wrterror(d, "munmap %p", p);
765 		STATS_SUB(d->malloc_used, sz);
766 		return;
767 	}
768 	cache = &d->smallcache[psz - 1];
769 	if (cache->length == cache->max) {
770 		int fresh;
771 		/* use a random slot */
772 		i = getrbyte(d) & (cache->max - 1);
773 		r = cache->pages[i];
774 		fresh = (uintptr_t)r & 1;
775 		*(uintptr_t*)&r &= ~1ULL;
776 		if (!fresh && !mopts.malloc_freeunmap)
777 			validate_junk(d, r, sz);
778 		if (munmap(r, sz))
779 			wrterror(d, "munmap %p", r);
780 		STATS_SUB(d->malloc_used, sz);
781 		cache->length--;
782 	} else
783 		i = cache->length;
784 
785 	/* fill slot */
786 	if (clear > 0)
787 		explicit_bzero(p, clear);
788 	if (mopts.malloc_freeunmap)
789 		mprotect(p, sz, PROT_NONE);
790 	else
791 		junk_free(d->malloc_junk, p, sz);
792 	cache->pages[i] = p;
793 	cache->length++;
794 }
795 
796 static void *
797 map(struct dir_info *d, size_t sz, int zero_fill)
798 {
799 	size_t psz = sz >> MALLOC_PAGESHIFT;
800 	u_short i;
801 	void *p;
802 	struct smallcache *cache;
803 
804 	if (mopts.malloc_canary != (d->canary1 ^ (u_int32_t)(uintptr_t)d) ||
805 	    d->canary1 != ~d->canary2)
806 		wrterror(d, "internal struct corrupt");
807 	if (sz != PAGEROUND(sz) || psz == 0)
808 		wrterror(d, "map round");
809 
810 
811 	if (d->bigcache_size > 0 && psz > MAX_SMALLCACHEABLE_SIZE &&
812 	    psz <= MAX_BIGCACHEABLE_SIZE) {
813 		size_t base = getrbyte(d);
814 		size_t cached = d->bigcache_used;
815 		ushort j;
816 
817 		for (j = 0; j < d->bigcache_size && cached >= psz; j++) {
818 			i = (j + base) & (d->bigcache_size - 1);
819 			if (d->bigcache[i].psize == psz) {
820 				p = d->bigcache[i].page;
821 				d->bigcache_used -= psz;
822 				d->bigcache[i].page = NULL;
823 				d->bigcache[i].psize = 0;
824 
825 				if (!mopts.malloc_freeunmap)
826 					validate_junk(d, p, sz);
827 				if (mopts.malloc_freeunmap)
828 					mprotect(p, sz, PROT_READ | PROT_WRITE);
829 				if (zero_fill)
830 					memset(p, 0, sz);
831 				else if (mopts.malloc_freeunmap)
832 					junk_free(d->malloc_junk, p, sz);
833 				return p;
834 			}
835 			cached -= d->bigcache[i].psize;
836 		}
837 	}
838 	if (psz <= MAX_SMALLCACHEABLE_SIZE && d->smallcache[psz - 1].max > 0) {
839 		cache = &d->smallcache[psz - 1];
840 		if (cache->length > 0) {
841 			int fresh;
842 			if (cache->length == 1)
843 				p = cache->pages[--cache->length];
844 			else {
845 				i = getrbyte(d) % cache->length;
846 				p = cache->pages[i];
847 				cache->pages[i] = cache->pages[--cache->length];
848 			}
849 			/* check if page was not junked, i.e. "fresh
850 			   we use the lsb of the pointer for that */
851 			fresh = (uintptr_t)p & 1UL;
852 			*(uintptr_t*)&p &= ~1UL;
853 			if (!fresh && !mopts.malloc_freeunmap)
854 				validate_junk(d, p, sz);
855 			if (mopts.malloc_freeunmap)
856 				mprotect(p, sz, PROT_READ | PROT_WRITE);
857 			if (zero_fill)
858 				memset(p, 0, sz);
859 			else if (mopts.malloc_freeunmap)
860 				junk_free(d->malloc_junk, p, sz);
861 			return p;
862 		}
863 		if (psz <= 1) {
864 			p = MMAP(cache->max * sz, d->mmap_flag);
865 			if (p != MAP_FAILED) {
866 				STATS_ADD(d->malloc_used, cache->max * sz);
867 				cache->length = cache->max - 1;
868 				for (i = 0; i < cache->max - 1; i++) {
869 					void *q = (char*)p + i * sz;
870 					cache->pages[i] = q;
871 					/* mark pointer in slot as not junked */
872 					*(uintptr_t*)&cache->pages[i] |= 1UL;
873 				}
874 				if (mopts.malloc_freeunmap)
875 					mprotect(p, (cache->max - 1) * sz,
876 					    PROT_NONE);
877 				p = (char*)p + (cache->max - 1) * sz;
878 				/* zero fill not needed, freshly mmapped */
879 				return p;
880 			}
881 		}
882 
883 	}
884 	p = MMAP(sz, d->mmap_flag);
885 	if (p != MAP_FAILED)
886 		STATS_ADD(d->malloc_used, sz);
887 	/* zero fill not needed */
888 	return p;
889 }
890 
891 static void
892 init_chunk_info(struct dir_info *d, struct chunk_info *p, u_int bucket)
893 {
894 	u_int i;
895 
896 	p->bucket = bucket;
897 	p->total = p->free = MALLOC_PAGESIZE / B2ALLOC(bucket);
898 	p->offset = bucket == 0 ? 0xdead : howmany(p->total, MALLOC_BITS);
899 	p->canary = (u_short)d->canary1;
900 
901 	/* set all valid bits in the bitmap */
902  	i = p->total - 1;
903 	memset(p->bits, 0xff, sizeof(p->bits[0]) * (i / MALLOC_BITS));
904 	p->bits[i / MALLOC_BITS] = (2U << (i % MALLOC_BITS)) - 1;
905 }
906 
907 static struct chunk_info *
908 alloc_chunk_info(struct dir_info *d, u_int bucket)
909 {
910 	struct chunk_info *p;
911 
912 	if (LIST_EMPTY(&d->chunk_info_list[bucket])) {
913 		const size_t chunk_pages = 64;
914 		size_t size, count, i;
915 		char *q;
916 
917 		count = MALLOC_PAGESIZE / B2ALLOC(bucket);
918 
919 		size = howmany(count, MALLOC_BITS);
920 		size = sizeof(struct chunk_info) + (size - 1) * sizeof(u_short);
921 		if (mopts.chunk_canaries)
922 			size += count * sizeof(u_short);
923 		size = _ALIGN(size);
924 		count = MALLOC_PAGESIZE / size;
925 
926 		/* Don't use cache here, we don't want user uaf touch this */
927 		if (d->chunk_pages_used == chunk_pages ||
928 		     d->chunk_pages == NULL) {
929 			q = MMAP(MALLOC_PAGESIZE * chunk_pages, d->mmap_flag);
930 			if (q == MAP_FAILED)
931 				return NULL;
932 			d->chunk_pages = q;
933 			d->chunk_pages_used = 0;
934 			STATS_ADD(d->malloc_used, MALLOC_PAGESIZE *
935 			    chunk_pages);
936 		}
937 		q = (char *)d->chunk_pages + d->chunk_pages_used *
938 		    MALLOC_PAGESIZE;
939 		d->chunk_pages_used++;
940 
941 		for (i = 0; i < count; i++, q += size) {
942 			p = (struct chunk_info *)q;
943 			LIST_INSERT_HEAD(&d->chunk_info_list[bucket], p, entries);
944 		}
945 	}
946 	p = LIST_FIRST(&d->chunk_info_list[bucket]);
947 	LIST_REMOVE(p, entries);
948 	if (p->total == 0)
949 		init_chunk_info(d, p, bucket);
950 	return p;
951 }
952 
953 /*
954  * Allocate a page of chunks
955  */
956 static struct chunk_info *
957 omalloc_make_chunks(struct dir_info *d, u_int bucket, u_int listnum)
958 {
959 	struct chunk_info *bp;
960 	void *pp;
961 
962 	/* Allocate a new bucket */
963 	pp = map(d, MALLOC_PAGESIZE, 0);
964 	if (pp == MAP_FAILED)
965 		return NULL;
966 
967 	/* memory protect the page allocated in the malloc(0) case */
968 	if (bucket == 0 && mprotect(pp, MALLOC_PAGESIZE, PROT_NONE) == -1)
969 		goto err;
970 
971 	bp = alloc_chunk_info(d, bucket);
972 	if (bp == NULL)
973 		goto err;
974 	bp->page = pp;
975 
976 	if (insert(d, (void *)((uintptr_t)pp | (bucket + 1)), (uintptr_t)bp,
977 	    NULL))
978 		goto err;
979 	LIST_INSERT_HEAD(&d->chunk_dir[bucket][listnum], bp, entries);
980 
981 	if (bucket > 0 && d->malloc_junk != 0)
982 		memset(pp, SOME_FREEJUNK, MALLOC_PAGESIZE);
983 
984 	return bp;
985 
986 err:
987 	unmap(d, pp, MALLOC_PAGESIZE, 0);
988 	return NULL;
989 }
990 
991 static inline unsigned int
992 lb(u_int x)
993 {
994 	/* I need an extension just for integer-length (: */
995 	return (sizeof(int) * CHAR_BIT - 1) - __builtin_clz(x);
996 }
997 
998 /* https://pvk.ca/Blog/2015/06/27/linear-log-bucketing-fast-versatile-simple/
999    via Tony Finch */
1000 static inline unsigned int
1001 bin_of(unsigned int size)
1002 {
1003 	const unsigned int linear = 6;
1004 	const unsigned int subbin = 2;
1005 
1006 	unsigned int mask, range, rounded, sub_index, rounded_size;
1007 	unsigned int n_bits, shift;
1008 
1009 	n_bits = lb(size | (1U << linear));
1010 	shift = n_bits - subbin;
1011 	mask = (1ULL << shift) - 1;
1012 	rounded = size + mask; /* XXX: overflow. */
1013 	sub_index = rounded >> shift;
1014 	range = n_bits - linear;
1015 
1016 	rounded_size = rounded & ~mask;
1017 	return rounded_size;
1018 }
1019 
1020 static inline u_short
1021 find_bucket(u_short size)
1022 {
1023 	/* malloc(0) is special */
1024 	if (size == 0)
1025 		return 0;
1026 	if (size < MALLOC_MINSIZE)
1027 		size = MALLOC_MINSIZE;
1028 	if (mopts.def_maxcache != 0)
1029 		size = bin_of(size);
1030 	return howmany(size, MALLOC_MINSIZE);
1031 }
1032 
1033 static void
1034 fill_canary(char *ptr, size_t sz, size_t allocated)
1035 {
1036 	size_t check_sz = allocated - sz;
1037 
1038 	if (check_sz > CHUNK_CHECK_LENGTH)
1039 		check_sz = CHUNK_CHECK_LENGTH;
1040 	memset(ptr + sz, mopts.chunk_canaries, check_sz);
1041 }
1042 
1043 /*
1044  * Allocate a chunk
1045  */
1046 static void *
1047 malloc_bytes(struct dir_info *d, size_t size, void *f)
1048 {
1049 	u_int i, r, bucket, listnum;
1050 	size_t k;
1051 	u_short	*lp;
1052 	struct chunk_info *bp;
1053 	void *p;
1054 
1055 	if (mopts.malloc_canary != (d->canary1 ^ (u_int32_t)(uintptr_t)d) ||
1056 	    d->canary1 != ~d->canary2)
1057 		wrterror(d, "internal struct corrupt");
1058 
1059 	bucket = find_bucket(size);
1060 
1061 	r = ((u_int)getrbyte(d) << 8) | getrbyte(d);
1062 	listnum = r % MALLOC_CHUNK_LISTS;
1063 
1064 	/* If it's empty, make a page more of that size chunks */
1065 	if ((bp = LIST_FIRST(&d->chunk_dir[bucket][listnum])) == NULL) {
1066 		bp = omalloc_make_chunks(d, bucket, listnum);
1067 		if (bp == NULL)
1068 			return NULL;
1069 	}
1070 
1071 	if (bp->canary != (u_short)d->canary1)
1072 		wrterror(d, "chunk info corrupted");
1073 
1074 	/* bias, as bp->total is not a power of 2 */
1075 	i = (r / MALLOC_CHUNK_LISTS) % bp->total;
1076 
1077 	/* potentially start somewhere in a short */
1078 	lp = &bp->bits[i / MALLOC_BITS];
1079 	if (*lp) {
1080 		int j = i % MALLOC_BITS; /* j must be signed */
1081 		k = ffs(*lp >> j);
1082 		if (k != 0) {
1083 			k += j - 1;
1084 			goto found;
1085 		}
1086 	}
1087 	/* no bit halfway, go to next full short */
1088 	i /= MALLOC_BITS;
1089 	for (;;) {
1090 		if (++i >= howmany(bp->total, MALLOC_BITS))
1091 			i = 0;
1092 		lp = &bp->bits[i];
1093 		if (*lp) {
1094 			k = ffs(*lp) - 1;
1095 			break;
1096 		}
1097 	}
1098 found:
1099 	if (i == 0 && k == 0 && DO_STATS) {
1100 		struct region_info *r = find(d, bp->page);
1101 		STATS_SETF(r, f);
1102 	}
1103 
1104 	*lp ^= 1 << k;
1105 
1106 	/* If there are no more free, remove from free-list */
1107 	if (--bp->free == 0)
1108 		LIST_REMOVE(bp, entries);
1109 
1110 	/* Adjust to the real offset of that chunk */
1111 	k += (lp - bp->bits) * MALLOC_BITS;
1112 
1113 	if (mopts.chunk_canaries && size > 0)
1114 		bp->bits[bp->offset + k] = size;
1115 
1116 	k *= B2ALLOC(bp->bucket);
1117 
1118 	p = (char *)bp->page + k;
1119 	if (bp->bucket > 0) {
1120 		validate_junk(d, p, B2SIZE(bp->bucket));
1121 		if (mopts.chunk_canaries)
1122 			fill_canary(p, size, B2SIZE(bp->bucket));
1123 	}
1124 	return p;
1125 }
1126 
1127 static void
1128 validate_canary(struct dir_info *d, u_char *ptr, size_t sz, size_t allocated)
1129 {
1130 	size_t check_sz = allocated - sz;
1131 	u_char *p, *q;
1132 
1133 	if (check_sz > CHUNK_CHECK_LENGTH)
1134 		check_sz = CHUNK_CHECK_LENGTH;
1135 	p = ptr + sz;
1136 	q = p + check_sz;
1137 
1138 	while (p < q) {
1139 		if (*p != (u_char)mopts.chunk_canaries && *p != SOME_JUNK) {
1140 			wrterror(d, "canary corrupted %p %#tx@%#zx%s",
1141 			    ptr, p - ptr, sz,
1142 			    *p == SOME_FREEJUNK ? " (double free?)" : "");
1143 		}
1144 		p++;
1145 	}
1146 }
1147 
1148 static uint32_t
1149 find_chunknum(struct dir_info *d, struct chunk_info *info, void *ptr, int check)
1150 {
1151 	uint32_t chunknum;
1152 
1153 	if (info->canary != (u_short)d->canary1)
1154 		wrterror(d, "chunk info corrupted");
1155 
1156 	/* Find the chunk number on the page */
1157 	chunknum = ((uintptr_t)ptr & MALLOC_PAGEMASK) / B2ALLOC(info->bucket);
1158 
1159 	if ((uintptr_t)ptr & (MALLOC_MINSIZE - 1))
1160 		wrterror(d, "modified chunk-pointer %p", ptr);
1161 	if (info->bits[chunknum / MALLOC_BITS] &
1162 	    (1U << (chunknum % MALLOC_BITS)))
1163 		wrterror(d, "double free %p", ptr);
1164 	if (check && info->bucket > 0) {
1165 		validate_canary(d, ptr, info->bits[info->offset + chunknum],
1166 		    B2SIZE(info->bucket));
1167 	}
1168 	return chunknum;
1169 }
1170 
1171 /*
1172  * Free a chunk, and possibly the page it's on, if the page becomes empty.
1173  */
1174 static void
1175 free_bytes(struct dir_info *d, struct region_info *r, void *ptr)
1176 {
1177 	struct chunk_head *mp;
1178 	struct chunk_info *info;
1179 	uint32_t chunknum;
1180 	uint32_t listnum;
1181 
1182 	info = (struct chunk_info *)r->size;
1183 	chunknum = find_chunknum(d, info, ptr, 0);
1184 
1185 	if (chunknum == 0)
1186 		STATS_SETF(r, NULL);
1187 
1188 	info->bits[chunknum / MALLOC_BITS] |= 1U << (chunknum % MALLOC_BITS);
1189 	info->free++;
1190 
1191 	if (info->free == 1) {
1192 		/* Page became non-full */
1193 		listnum = getrbyte(d) % MALLOC_CHUNK_LISTS;
1194 		mp = &d->chunk_dir[info->bucket][listnum];
1195 		LIST_INSERT_HEAD(mp, info, entries);
1196 		return;
1197 	}
1198 
1199 	if (info->free != info->total)
1200 		return;
1201 
1202 	LIST_REMOVE(info, entries);
1203 
1204 	if (info->bucket == 0 && !mopts.malloc_freeunmap)
1205 		mprotect(info->page, MALLOC_PAGESIZE, PROT_READ | PROT_WRITE);
1206 	unmap(d, info->page, MALLOC_PAGESIZE, 0);
1207 
1208 	delete(d, r);
1209 	mp = &d->chunk_info_list[info->bucket];
1210 	LIST_INSERT_HEAD(mp, info, entries);
1211 }
1212 
1213 
1214 
1215 static void *
1216 omalloc(struct dir_info *pool, size_t sz, int zero_fill, void *f)
1217 {
1218 	void *p;
1219 	size_t psz;
1220 
1221 	if (sz > MALLOC_MAXCHUNK) {
1222 		if (sz >= SIZE_MAX - mopts.malloc_guard - MALLOC_PAGESIZE) {
1223 			errno = ENOMEM;
1224 			return NULL;
1225 		}
1226 		sz += mopts.malloc_guard;
1227 		psz = PAGEROUND(sz);
1228 		p = map(pool, psz, zero_fill);
1229 		if (p == MAP_FAILED) {
1230 			errno = ENOMEM;
1231 			return NULL;
1232 		}
1233 		if (insert(pool, p, sz, f)) {
1234 			unmap(pool, p, psz, 0);
1235 			errno = ENOMEM;
1236 			return NULL;
1237 		}
1238 		if (mopts.malloc_guard) {
1239 			if (mprotect((char *)p + psz - mopts.malloc_guard,
1240 			    mopts.malloc_guard, PROT_NONE))
1241 				wrterror(pool, "mprotect");
1242 			STATS_ADD(pool->malloc_guarded, mopts.malloc_guard);
1243 		}
1244 
1245 		if (MALLOC_MOVE_COND(sz)) {
1246 			/* fill whole allocation */
1247 			if (pool->malloc_junk == 2)
1248 				memset(p, SOME_JUNK, psz - mopts.malloc_guard);
1249 			/* shift towards the end */
1250 			p = MALLOC_MOVE(p, sz);
1251 			/* fill zeros if needed and overwritten above */
1252 			if (zero_fill && pool->malloc_junk == 2)
1253 				memset(p, 0, sz - mopts.malloc_guard);
1254 		} else {
1255 			if (pool->malloc_junk == 2) {
1256 				if (zero_fill)
1257 					memset((char *)p + sz -
1258 					    mopts.malloc_guard, SOME_JUNK,
1259 					    psz - sz);
1260 				else
1261 					memset(p, SOME_JUNK,
1262 					    psz - mopts.malloc_guard);
1263 			} else if (mopts.chunk_canaries)
1264 				fill_canary(p, sz - mopts.malloc_guard,
1265 				    psz - mopts.malloc_guard);
1266 		}
1267 
1268 	} else {
1269 		/* takes care of SOME_JUNK */
1270 		p = malloc_bytes(pool, sz, f);
1271 		if (zero_fill && p != NULL && sz > 0)
1272 			memset(p, 0, sz);
1273 	}
1274 
1275 	return p;
1276 }
1277 
1278 /*
1279  * Common function for handling recursion.  Only
1280  * print the error message once, to avoid making the problem
1281  * potentially worse.
1282  */
1283 static void
1284 malloc_recurse(struct dir_info *d)
1285 {
1286 	static int noprint;
1287 
1288 	if (noprint == 0) {
1289 		noprint = 1;
1290 		wrterror(d, "recursive call");
1291 	}
1292 	d->active--;
1293 	_MALLOC_UNLOCK(d->mutex);
1294 	errno = EDEADLK;
1295 }
1296 
1297 void
1298 _malloc_init(int from_rthreads)
1299 {
1300 	u_int i, j, nmutexes;
1301 	struct dir_info *d;
1302 
1303 	_MALLOC_LOCK(1);
1304 	if (!from_rthreads && mopts.malloc_pool[1]) {
1305 		_MALLOC_UNLOCK(1);
1306 		return;
1307 	}
1308 	if (!mopts.malloc_canary) {
1309 		char *p;
1310 		size_t sz, d_avail;
1311 
1312 		omalloc_init();
1313 		/*
1314 		 * Allocate dir_infos with a guard page on either side. Also
1315 		 * randomise offset inside the page at which the dir_infos
1316 		 * lay (subject to alignment by 1 << MALLOC_MINSHIFT)
1317 		 */
1318 		sz = mopts.malloc_mutexes * sizeof(*d) + 2 * MALLOC_PAGESIZE;
1319 		if ((p = MMAPNONE(sz, 0)) == MAP_FAILED)
1320 			wrterror(NULL, "malloc_init mmap1 failed");
1321 		if (mprotect(p + MALLOC_PAGESIZE, mopts.malloc_mutexes * sizeof(*d),
1322 		    PROT_READ | PROT_WRITE))
1323 			wrterror(NULL, "malloc_init mprotect1 failed");
1324 		if (mimmutable(p, sz))
1325 			wrterror(NULL, "malloc_init mimmutable1 failed");
1326 		d_avail = (((mopts.malloc_mutexes * sizeof(*d) + MALLOC_PAGEMASK) &
1327 		    ~MALLOC_PAGEMASK) - (mopts.malloc_mutexes * sizeof(*d))) >>
1328 		    MALLOC_MINSHIFT;
1329 		d = (struct dir_info *)(p + MALLOC_PAGESIZE +
1330 		    (arc4random_uniform(d_avail) << MALLOC_MINSHIFT));
1331 		STATS_ADD(d[1].malloc_used, sz);
1332 		for (i = 0; i < mopts.malloc_mutexes; i++)
1333 			mopts.malloc_pool[i] = &d[i];
1334 		mopts.internal_funcs = 1;
1335 		if (((uintptr_t)&malloc_readonly & MALLOC_PAGEMASK) == 0) {
1336 			if (mprotect(&malloc_readonly, sizeof(malloc_readonly),
1337 			    PROT_READ))
1338 				wrterror(NULL, "malloc_init mprotect r/o failed");
1339 			if (mimmutable(&malloc_readonly, sizeof(malloc_readonly)))
1340 				wrterror(NULL, "malloc_init mimmutable r/o failed");
1341 		}
1342 	}
1343 
1344 	nmutexes = from_rthreads ? mopts.malloc_mutexes : 2;
1345 	for (i = 0; i < nmutexes; i++) {
1346 		d = mopts.malloc_pool[i];
1347 		d->malloc_mt = from_rthreads;
1348 		if (d->canary1 == ~d->canary2)
1349 			continue;
1350 		if (i == 0) {
1351 			omalloc_poolinit(d, MAP_CONCEAL);
1352 			d->malloc_junk = 2;
1353 			d->bigcache_size = 0;
1354 			for (j = 0; j < MAX_SMALLCACHEABLE_SIZE; j++)
1355 				d->smallcache[j].max = 0;
1356 		} else {
1357 			size_t sz = 0;
1358 
1359 			omalloc_poolinit(d, 0);
1360 			d->malloc_junk = mopts.def_malloc_junk;
1361 			d->bigcache_size = mopts.def_maxcache;
1362 			for (j = 0; j < MAX_SMALLCACHEABLE_SIZE; j++) {
1363 				d->smallcache[j].max =
1364 				    mopts.def_maxcache >> (j / 8);
1365 				sz += d->smallcache[j].max * sizeof(void *);
1366 			}
1367 			sz += d->bigcache_size * sizeof(struct bigcache);
1368 			if (sz > 0) {
1369 				void *p = MMAP(sz, 0);
1370 				if (p == MAP_FAILED)
1371 					wrterror(NULL,
1372 					    "malloc_init mmap2 failed");
1373 				if (mimmutable(p, sz))
1374 					wrterror(NULL, "malloc_init mimmutable2 failed");
1375 				for (j = 0; j < MAX_SMALLCACHEABLE_SIZE; j++) {
1376 					d->smallcache[j].pages = p;
1377 					p = (char *)p + d->smallcache[j].max *
1378 					    sizeof(void *);
1379 				}
1380 				d->bigcache = p;
1381 			}
1382 		}
1383 		d->mutex = i;
1384 	}
1385 
1386 	_MALLOC_UNLOCK(1);
1387 }
1388 DEF_STRONG(_malloc_init);
1389 
1390 #define PROLOGUE(p, fn)			\
1391 	d = (p); 			\
1392 	if (d == NULL) { 		\
1393 		_malloc_init(0);	\
1394 		d = (p);		\
1395 	}				\
1396 	_MALLOC_LOCK(d->mutex);		\
1397 	d->func = fn;			\
1398 	if (d->active++) {		\
1399 		malloc_recurse(d);	\
1400 		return NULL;		\
1401 	}				\
1402 
1403 #define EPILOGUE()				\
1404 	d->active--;				\
1405 	_MALLOC_UNLOCK(d->mutex);		\
1406 	if (r == NULL && mopts.malloc_xmalloc)	\
1407 		wrterror(d, "out of memory");	\
1408 	if (r != NULL)				\
1409 		errno = saved_errno;		\
1410 
1411 void *
1412 malloc(size_t size)
1413 {
1414 	void *r;
1415 	struct dir_info *d;
1416 	int saved_errno = errno;
1417 
1418 	PROLOGUE(getpool(), "malloc")
1419 	r = omalloc(d, size, 0, CALLER);
1420 	EPILOGUE()
1421 	return r;
1422 }
1423 DEF_STRONG(malloc);
1424 
1425 void *
1426 malloc_conceal(size_t size)
1427 {
1428 	void *r;
1429 	struct dir_info *d;
1430 	int saved_errno = errno;
1431 
1432 	PROLOGUE(mopts.malloc_pool[0], "malloc_conceal")
1433 	r = omalloc(d, size, 0, CALLER);
1434 	EPILOGUE()
1435 	return r;
1436 }
1437 DEF_WEAK(malloc_conceal);
1438 
1439 static struct region_info *
1440 findpool(void *p, struct dir_info *argpool, struct dir_info **foundpool,
1441     char **saved_function)
1442 {
1443 	struct dir_info *pool = argpool;
1444 	struct region_info *r = find(pool, p);
1445 
1446 	STATS_INC(pool->pool_searches);
1447 	if (r == NULL) {
1448 		u_int i, nmutexes;
1449 
1450 		nmutexes = mopts.malloc_pool[1]->malloc_mt ? mopts.malloc_mutexes : 2;
1451 		STATS_INC(pool->other_pool);
1452 		for (i = 1; i < nmutexes; i++) {
1453 			u_int j = (argpool->mutex + i) & (nmutexes - 1);
1454 
1455 			pool->active--;
1456 			_MALLOC_UNLOCK(pool->mutex);
1457 			pool = mopts.malloc_pool[j];
1458 			_MALLOC_LOCK(pool->mutex);
1459 			pool->active++;
1460 			r = find(pool, p);
1461 			if (r != NULL) {
1462 				*saved_function = pool->func;
1463 				pool->func = argpool->func;
1464 				break;
1465 			}
1466 		}
1467 		if (r == NULL)
1468 			wrterror(argpool, "bogus pointer (double free?) %p", p);
1469 	}
1470 	*foundpool = pool;
1471 	return r;
1472 }
1473 
1474 static void
1475 ofree(struct dir_info **argpool, void *p, int clear, int check, size_t argsz)
1476 {
1477 	struct region_info *r;
1478 	struct dir_info *pool;
1479 	char *saved_function;
1480 	size_t sz;
1481 
1482 	r = findpool(p, *argpool, &pool, &saved_function);
1483 
1484 	REALSIZE(sz, r);
1485 	if (pool->mmap_flag) {
1486 		clear = 1;
1487 		if (!check) {
1488 			argsz = sz;
1489 			if (sz > MALLOC_MAXCHUNK)
1490 				argsz -= mopts.malloc_guard;
1491 		}
1492 	}
1493 	if (check) {
1494 		if (sz <= MALLOC_MAXCHUNK) {
1495 			if (mopts.chunk_canaries && sz > 0) {
1496 				struct chunk_info *info =
1497 				    (struct chunk_info *)r->size;
1498 				uint32_t chunknum =
1499 				    find_chunknum(pool, info, p, 0);
1500 
1501 				if (info->bits[info->offset + chunknum] < argsz)
1502 					wrterror(pool, "recorded size %hu"
1503 					    " < %zu",
1504 					    info->bits[info->offset + chunknum],
1505 					    argsz);
1506 			} else {
1507 				if (sz < argsz)
1508 					wrterror(pool, "chunk size %zu < %zu",
1509 					    sz, argsz);
1510 			}
1511 		} else if (sz - mopts.malloc_guard < argsz) {
1512 			wrterror(pool, "recorded size %zu < %zu",
1513 			    sz - mopts.malloc_guard, argsz);
1514 		}
1515 	}
1516 	if (sz > MALLOC_MAXCHUNK) {
1517 		if (!MALLOC_MOVE_COND(sz)) {
1518 			if (r->p != p)
1519 				wrterror(pool, "bogus pointer %p", p);
1520 			if (mopts.chunk_canaries)
1521 				validate_canary(pool, p,
1522 				    sz - mopts.malloc_guard,
1523 				    PAGEROUND(sz - mopts.malloc_guard));
1524 		} else {
1525 			/* shifted towards the end */
1526 			if (p != MALLOC_MOVE(r->p, sz))
1527 				wrterror(pool, "bogus moved pointer %p", p);
1528 			p = r->p;
1529 		}
1530 		if (mopts.malloc_guard) {
1531 			if (sz < mopts.malloc_guard)
1532 				wrterror(pool, "guard size");
1533 			if (!mopts.malloc_freeunmap) {
1534 				if (mprotect((char *)p + PAGEROUND(sz) -
1535 				    mopts.malloc_guard, mopts.malloc_guard,
1536 				    PROT_READ | PROT_WRITE))
1537 					wrterror(pool, "mprotect");
1538 			}
1539 			STATS_SUB(pool->malloc_guarded, mopts.malloc_guard);
1540 		}
1541 		unmap(pool, p, PAGEROUND(sz), clear ? argsz : 0);
1542 		delete(pool, r);
1543 	} else {
1544 		void *tmp;
1545 		u_int i;
1546 
1547 		/* Validate and optionally canary check */
1548 		struct chunk_info *info = (struct chunk_info *)r->size;
1549 		if (B2SIZE(info->bucket) != sz)
1550 			wrterror(pool, "internal struct corrupt");
1551 		find_chunknum(pool, info, p, mopts.chunk_canaries);
1552 
1553 		if (mopts.malloc_freecheck) {
1554 			for (i = 0; i <= MALLOC_DELAYED_CHUNK_MASK; i++) {
1555 				tmp = pool->delayed_chunks[i];
1556 				if (tmp == p)
1557 					wrterror(pool,
1558 					    "double free %p", p);
1559 				if (tmp != NULL) {
1560 					size_t tmpsz;
1561 
1562 					r = find(pool, tmp);
1563 					if (r == NULL)
1564 						wrterror(pool,
1565 						    "bogus pointer ("
1566 						    "double free?) %p", tmp);
1567 					REALSIZE(tmpsz, r);
1568 					validate_junk(pool, tmp, tmpsz);
1569 				}
1570 			}
1571 		}
1572 
1573 		if (clear && argsz > 0)
1574 			explicit_bzero(p, argsz);
1575 		junk_free(pool->malloc_junk, p, sz);
1576 
1577 		i = getrbyte(pool) & MALLOC_DELAYED_CHUNK_MASK;
1578 		tmp = p;
1579 		p = pool->delayed_chunks[i];
1580 		if (tmp == p)
1581 			wrterror(pool, "double free %p", p);
1582 		pool->delayed_chunks[i] = tmp;
1583 		if (p != NULL) {
1584 			r = find(pool, p);
1585 			if (r == NULL)
1586 				wrterror(pool,
1587 				    "bogus pointer (double free?) %p", p);
1588 			if (!mopts.malloc_freecheck) {
1589 				REALSIZE(sz, r);
1590 				validate_junk(pool, p, sz);
1591 			}
1592 			free_bytes(pool, r, p);
1593 		}
1594 	}
1595 
1596 	if (*argpool != pool) {
1597 		pool->func = saved_function;
1598 		*argpool = pool;
1599 	}
1600 }
1601 
1602 void
1603 free(void *ptr)
1604 {
1605 	struct dir_info *d;
1606 	int saved_errno = errno;
1607 
1608 	/* This is legal. */
1609 	if (ptr == NULL)
1610 		return;
1611 
1612 	d = getpool();
1613 	if (d == NULL)
1614 		wrterror(d, "free() called before allocation");
1615 	_MALLOC_LOCK(d->mutex);
1616 	d->func = "free";
1617 	if (d->active++) {
1618 		malloc_recurse(d);
1619 		return;
1620 	}
1621 	ofree(&d, ptr, 0, 0, 0);
1622 	d->active--;
1623 	_MALLOC_UNLOCK(d->mutex);
1624 	errno = saved_errno;
1625 }
1626 DEF_STRONG(free);
1627 
1628 static void
1629 freezero_p(void *ptr, size_t sz)
1630 {
1631 	explicit_bzero(ptr, sz);
1632 	free(ptr);
1633 }
1634 
1635 void
1636 freezero(void *ptr, size_t sz)
1637 {
1638 	struct dir_info *d;
1639 	int saved_errno = errno;
1640 
1641 	/* This is legal. */
1642 	if (ptr == NULL)
1643 		return;
1644 
1645 	if (!mopts.internal_funcs) {
1646 		freezero_p(ptr, sz);
1647 		return;
1648 	}
1649 
1650 	d = getpool();
1651 	if (d == NULL)
1652 		wrterror(d, "freezero() called before allocation");
1653 	_MALLOC_LOCK(d->mutex);
1654 	d->func = "freezero";
1655 	if (d->active++) {
1656 		malloc_recurse(d);
1657 		return;
1658 	}
1659 	ofree(&d, ptr, 1, 1, sz);
1660 	d->active--;
1661 	_MALLOC_UNLOCK(d->mutex);
1662 	errno = saved_errno;
1663 }
1664 DEF_WEAK(freezero);
1665 
1666 static void *
1667 orealloc(struct dir_info **argpool, void *p, size_t newsz, void *f)
1668 {
1669 	struct region_info *r;
1670 	struct dir_info *pool;
1671 	char *saved_function;
1672 	struct chunk_info *info;
1673 	size_t oldsz, goldsz, gnewsz;
1674 	void *q, *ret;
1675 	uint32_t chunknum;
1676 	int forced;
1677 
1678 	if (p == NULL)
1679 		return omalloc(*argpool, newsz, 0, f);
1680 
1681 	if (newsz >= SIZE_MAX - mopts.malloc_guard - MALLOC_PAGESIZE) {
1682 		errno = ENOMEM;
1683 		return  NULL;
1684 	}
1685 
1686 	r = findpool(p, *argpool, &pool, &saved_function);
1687 
1688 	REALSIZE(oldsz, r);
1689 	if (oldsz <= MALLOC_MAXCHUNK) {
1690 		if (DO_STATS || mopts.chunk_canaries) {
1691 			info = (struct chunk_info *)r->size;
1692 			chunknum = find_chunknum(pool, info, p, 0);
1693 		}
1694 	}
1695 
1696 	goldsz = oldsz;
1697 	if (oldsz > MALLOC_MAXCHUNK) {
1698 		if (oldsz < mopts.malloc_guard)
1699 			wrterror(pool, "guard size");
1700 		oldsz -= mopts.malloc_guard;
1701 	}
1702 
1703 	gnewsz = newsz;
1704 	if (gnewsz > MALLOC_MAXCHUNK)
1705 		gnewsz += mopts.malloc_guard;
1706 
1707 	forced = mopts.malloc_realloc || pool->mmap_flag;
1708 	if (newsz > MALLOC_MAXCHUNK && oldsz > MALLOC_MAXCHUNK && !forced) {
1709 		/* First case: from n pages sized allocation to m pages sized
1710 		   allocation, m > n */
1711 		size_t roldsz = PAGEROUND(goldsz);
1712 		size_t rnewsz = PAGEROUND(gnewsz);
1713 
1714 		if (rnewsz < roldsz && rnewsz > roldsz / 2 &&
1715 		    roldsz - rnewsz < mopts.def_maxcache * MALLOC_PAGESIZE &&
1716 		    !mopts.malloc_guard) {
1717 
1718 			ret = p;
1719 			goto done;
1720 		}
1721 
1722 		if (rnewsz > roldsz) {
1723 			/* try to extend existing region */
1724 			if (!mopts.malloc_guard) {
1725 				void *hint = (char *)r->p + roldsz;
1726 				size_t needed = rnewsz - roldsz;
1727 
1728 				STATS_INC(pool->cheap_realloc_tries);
1729 				q = MMAPA(hint, needed, MAP_FIXED | __MAP_NOREPLACE | pool->mmap_flag);
1730 				if (q == hint) {
1731 					STATS_ADD(pool->malloc_used, needed);
1732 					if (pool->malloc_junk == 2)
1733 						memset(q, SOME_JUNK, needed);
1734 					r->size = gnewsz;
1735 					if (r->p != p) {
1736 						/* old pointer is moved */
1737 						memmove(r->p, p, oldsz);
1738 						p = r->p;
1739 					}
1740 					if (mopts.chunk_canaries)
1741 						fill_canary(p, newsz,
1742 						    PAGEROUND(newsz));
1743 					STATS_SETF(r, f);
1744 					STATS_INC(pool->cheap_reallocs);
1745 					ret = p;
1746 					goto done;
1747 				}
1748 			}
1749 		} else if (rnewsz < roldsz) {
1750 			/* shrink number of pages */
1751 			if (mopts.malloc_guard) {
1752 				if (mprotect((char *)r->p + rnewsz -
1753 				    mopts.malloc_guard, mopts.malloc_guard,
1754 				    PROT_NONE))
1755 					wrterror(pool, "mprotect");
1756 			}
1757 			if (munmap((char *)r->p + rnewsz, roldsz - rnewsz))
1758 				wrterror(pool, "munmap %p", (char *)r->p +
1759 				    rnewsz);
1760 			STATS_SUB(pool->malloc_used, roldsz - rnewsz);
1761 			r->size = gnewsz;
1762 			if (MALLOC_MOVE_COND(gnewsz)) {
1763 				void *pp = MALLOC_MOVE(r->p, gnewsz);
1764 				memmove(pp, p, newsz);
1765 				p = pp;
1766 			} else if (mopts.chunk_canaries)
1767 				fill_canary(p, newsz, PAGEROUND(newsz));
1768 			STATS_SETF(r, f);
1769 			ret = p;
1770 			goto done;
1771 		} else {
1772 			/* number of pages remains the same */
1773 			void *pp = r->p;
1774 
1775 			r->size = gnewsz;
1776 			if (MALLOC_MOVE_COND(gnewsz))
1777 				pp = MALLOC_MOVE(r->p, gnewsz);
1778 			if (p != pp) {
1779 				memmove(pp, p, oldsz < newsz ? oldsz : newsz);
1780 				p = pp;
1781 			}
1782 			if (p == r->p) {
1783 				if (newsz > oldsz && pool->malloc_junk == 2)
1784 					memset((char *)p + newsz, SOME_JUNK,
1785 					    rnewsz - mopts.malloc_guard -
1786 					    newsz);
1787 				if (mopts.chunk_canaries)
1788 					fill_canary(p, newsz, PAGEROUND(newsz));
1789 			}
1790 			STATS_SETF(r, f);
1791 			ret = p;
1792 			goto done;
1793 		}
1794 	}
1795 	if (oldsz <= MALLOC_MAXCHUNK && oldsz > 0 &&
1796 	    newsz <= MALLOC_MAXCHUNK && newsz > 0 &&
1797 	    !forced && find_bucket(newsz) == find_bucket(oldsz)) {
1798 		/* do not reallocate if new size fits good in existing chunk */
1799 		if (pool->malloc_junk == 2)
1800 			memset((char *)p + newsz, SOME_JUNK, oldsz - newsz);
1801 		if (mopts.chunk_canaries) {
1802 			info->bits[info->offset + chunknum] = newsz;
1803 			fill_canary(p, newsz, B2SIZE(info->bucket));
1804 		}
1805 		if (DO_STATS && chunknum == 0)
1806 			STATS_SETF(r, f);
1807 		ret = p;
1808 	} else if (newsz != oldsz || forced) {
1809 		/* create new allocation */
1810 		q = omalloc(pool, newsz, 0, f);
1811 		if (q == NULL) {
1812 			ret = NULL;
1813 			goto done;
1814 		}
1815 		if (newsz != 0 && oldsz != 0)
1816 			memcpy(q, p, oldsz < newsz ? oldsz : newsz);
1817 		ofree(&pool, p, 0, 0, 0);
1818 		ret = q;
1819 	} else {
1820 		/* oldsz == newsz */
1821 		if (newsz != 0)
1822 			wrterror(pool, "realloc internal inconsistency");
1823 		if (DO_STATS && chunknum == 0)
1824 			STATS_SETF(r, f);
1825 		ret = p;
1826 	}
1827 done:
1828 	if (*argpool != pool) {
1829 		pool->func = saved_function;
1830 		*argpool = pool;
1831 	}
1832 	return ret;
1833 }
1834 
1835 void *
1836 realloc(void *ptr, size_t size)
1837 {
1838 	struct dir_info *d;
1839 	void *r;
1840 	int saved_errno = errno;
1841 
1842 	PROLOGUE(getpool(), "realloc")
1843 	r = orealloc(&d, ptr, size, CALLER);
1844 	EPILOGUE()
1845 	return r;
1846 }
1847 DEF_STRONG(realloc);
1848 
1849 /*
1850  * This is sqrt(SIZE_MAX+1), as s1*s2 <= SIZE_MAX
1851  * if both s1 < MUL_NO_OVERFLOW and s2 < MUL_NO_OVERFLOW
1852  */
1853 #define MUL_NO_OVERFLOW	(1UL << (sizeof(size_t) * 4))
1854 
1855 void *
1856 calloc(size_t nmemb, size_t size)
1857 {
1858 	struct dir_info *d;
1859 	void *r;
1860 	int saved_errno = errno;
1861 
1862 	PROLOGUE(getpool(), "calloc")
1863 	if ((nmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1864 	    nmemb > 0 && SIZE_MAX / nmemb < size) {
1865 		d->active--;
1866 		_MALLOC_UNLOCK(d->mutex);
1867 		if (mopts.malloc_xmalloc)
1868 			wrterror(d, "out of memory");
1869 		errno = ENOMEM;
1870 		return NULL;
1871 	}
1872 
1873 	size *= nmemb;
1874 	r = omalloc(d, size, 1, CALLER);
1875 	EPILOGUE()
1876 	return r;
1877 }
1878 DEF_STRONG(calloc);
1879 
1880 void *
1881 calloc_conceal(size_t nmemb, size_t size)
1882 {
1883 	struct dir_info *d;
1884 	void *r;
1885 	int saved_errno = errno;
1886 
1887 	PROLOGUE(mopts.malloc_pool[0], "calloc_conceal")
1888 	if ((nmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1889 	    nmemb > 0 && SIZE_MAX / nmemb < size) {
1890 		d->active--;
1891 		_MALLOC_UNLOCK(d->mutex);
1892 		if (mopts.malloc_xmalloc)
1893 			wrterror(d, "out of memory");
1894 		errno = ENOMEM;
1895 		return NULL;
1896 	}
1897 
1898 	size *= nmemb;
1899 	r = omalloc(d, size, 1, CALLER);
1900 	EPILOGUE()
1901 	return r;
1902 }
1903 DEF_WEAK(calloc_conceal);
1904 
1905 static void *
1906 orecallocarray(struct dir_info **argpool, void *p, size_t oldsize,
1907     size_t newsize, void *f)
1908 {
1909 	struct region_info *r;
1910 	struct dir_info *pool;
1911 	char *saved_function;
1912 	void *newptr;
1913 	size_t sz;
1914 
1915 	if (p == NULL)
1916 		return omalloc(*argpool, newsize, 1, f);
1917 
1918 	if (oldsize == newsize)
1919 		return p;
1920 
1921 	r = findpool(p, *argpool, &pool, &saved_function);
1922 
1923 	REALSIZE(sz, r);
1924 	if (sz <= MALLOC_MAXCHUNK) {
1925 		if (mopts.chunk_canaries && sz > 0) {
1926 			struct chunk_info *info = (struct chunk_info *)r->size;
1927 			uint32_t chunknum = find_chunknum(pool, info, p, 0);
1928 
1929 			if (info->bits[info->offset + chunknum] != oldsize)
1930 				wrterror(pool, "recorded size %hu != %zu",
1931 				    info->bits[info->offset + chunknum],
1932 				    oldsize);
1933 		} else {
1934 			if (sz < oldsize)
1935 				wrterror(pool, "chunk size %zu < %zu",
1936 				    sz, oldsize);
1937 		}
1938 	} else {
1939 		if (sz - mopts.malloc_guard < oldsize)
1940 			wrterror(pool, "recorded size %zu < %zu",
1941 			    sz - mopts.malloc_guard, oldsize);
1942 		if (oldsize < (sz - mopts.malloc_guard) / 2)
1943 			wrterror(pool, "recorded size %zu inconsistent with %zu",
1944 			    sz - mopts.malloc_guard, oldsize);
1945 	}
1946 
1947 	newptr = omalloc(pool, newsize, 0, f);
1948 	if (newptr == NULL)
1949 		goto done;
1950 
1951 	if (newsize > oldsize) {
1952 		memcpy(newptr, p, oldsize);
1953 		memset((char *)newptr + oldsize, 0, newsize - oldsize);
1954 	} else
1955 		memcpy(newptr, p, newsize);
1956 
1957 	ofree(&pool, p, 1, 0, oldsize);
1958 
1959 done:
1960 	if (*argpool != pool) {
1961 		pool->func = saved_function;
1962 		*argpool = pool;
1963 	}
1964 
1965 	return newptr;
1966 }
1967 
1968 static void *
1969 recallocarray_p(void *ptr, size_t oldnmemb, size_t newnmemb, size_t size)
1970 {
1971 	size_t oldsize, newsize;
1972 	void *newptr;
1973 
1974 	if (ptr == NULL)
1975 		return calloc(newnmemb, size);
1976 
1977 	if ((newnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1978 	    newnmemb > 0 && SIZE_MAX / newnmemb < size) {
1979 		errno = ENOMEM;
1980 		return NULL;
1981 	}
1982 	newsize = newnmemb * size;
1983 
1984 	if ((oldnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
1985 	    oldnmemb > 0 && SIZE_MAX / oldnmemb < size) {
1986 		errno = EINVAL;
1987 		return NULL;
1988 	}
1989 	oldsize = oldnmemb * size;
1990 
1991 	/*
1992 	 * Don't bother too much if we're shrinking just a bit,
1993 	 * we do not shrink for series of small steps, oh well.
1994 	 */
1995 	if (newsize <= oldsize) {
1996 		size_t d = oldsize - newsize;
1997 
1998 		if (d < oldsize / 2 && d < MALLOC_PAGESIZE) {
1999 			memset((char *)ptr + newsize, 0, d);
2000 			return ptr;
2001 		}
2002 	}
2003 
2004 	newptr = malloc(newsize);
2005 	if (newptr == NULL)
2006 		return NULL;
2007 
2008 	if (newsize > oldsize) {
2009 		memcpy(newptr, ptr, oldsize);
2010 		memset((char *)newptr + oldsize, 0, newsize - oldsize);
2011 	} else
2012 		memcpy(newptr, ptr, newsize);
2013 
2014 	explicit_bzero(ptr, oldsize);
2015 	free(ptr);
2016 
2017 	return newptr;
2018 }
2019 
2020 void *
2021 recallocarray(void *ptr, size_t oldnmemb, size_t newnmemb, size_t size)
2022 {
2023 	struct dir_info *d;
2024 	size_t oldsize = 0, newsize;
2025 	void *r;
2026 	int saved_errno = errno;
2027 
2028 	if (!mopts.internal_funcs)
2029 		return recallocarray_p(ptr, oldnmemb, newnmemb, size);
2030 
2031 	PROLOGUE(getpool(), "recallocarray")
2032 
2033 	if ((newnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
2034 	    newnmemb > 0 && SIZE_MAX / newnmemb < size) {
2035 		d->active--;
2036 		_MALLOC_UNLOCK(d->mutex);
2037 		if (mopts.malloc_xmalloc)
2038 			wrterror(d, "out of memory");
2039 		errno = ENOMEM;
2040 		return NULL;
2041 	}
2042 	newsize = newnmemb * size;
2043 
2044 	if (ptr != NULL) {
2045 		if ((oldnmemb >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&
2046 		    oldnmemb > 0 && SIZE_MAX / oldnmemb < size) {
2047 			d->active--;
2048 			_MALLOC_UNLOCK(d->mutex);
2049 			errno = EINVAL;
2050 			return NULL;
2051 		}
2052 		oldsize = oldnmemb * size;
2053 	}
2054 
2055 	r = orecallocarray(&d, ptr, oldsize, newsize, CALLER);
2056 	EPILOGUE()
2057 	return r;
2058 }
2059 DEF_WEAK(recallocarray);
2060 
2061 static void *
2062 mapalign(struct dir_info *d, size_t alignment, size_t sz, int zero_fill)
2063 {
2064 	char *p, *q;
2065 
2066 	if (alignment < MALLOC_PAGESIZE || ((alignment - 1) & alignment) != 0)
2067 		wrterror(d, "mapalign bad alignment");
2068 	if (sz != PAGEROUND(sz))
2069 		wrterror(d, "mapalign round");
2070 
2071 	/* Allocate sz + alignment bytes of memory, which must include a
2072 	 * subrange of size bytes that is properly aligned.  Unmap the
2073 	 * other bytes, and then return that subrange.
2074 	 */
2075 
2076 	/* We need sz + alignment to fit into a size_t. */
2077 	if (alignment > SIZE_MAX - sz)
2078 		return MAP_FAILED;
2079 
2080 	p = map(d, sz + alignment, zero_fill);
2081 	if (p == MAP_FAILED)
2082 		return MAP_FAILED;
2083 	q = (char *)(((uintptr_t)p + alignment - 1) & ~(alignment - 1));
2084 	if (q != p) {
2085 		if (munmap(p, q - p))
2086 			wrterror(d, "munmap %p", p);
2087 	}
2088 	if (munmap(q + sz, alignment - (q - p)))
2089 		wrterror(d, "munmap %p", q + sz);
2090 	STATS_SUB(d->malloc_used, alignment);
2091 
2092 	return q;
2093 }
2094 
2095 static void *
2096 omemalign(struct dir_info *pool, size_t alignment, size_t sz, int zero_fill,
2097     void *f)
2098 {
2099 	size_t psz;
2100 	void *p;
2101 
2102 	/* If between half a page and a page, avoid MALLOC_MOVE. */
2103 	if (sz > MALLOC_MAXCHUNK && sz < MALLOC_PAGESIZE)
2104 		sz = MALLOC_PAGESIZE;
2105 	if (alignment <= MALLOC_PAGESIZE) {
2106 		size_t pof2;
2107 		/*
2108 		 * max(size, alignment) rounded up to power of 2 is enough
2109 		 * to assure the requested alignment. Large regions are
2110 		 * always page aligned.
2111 		 */
2112 		if (sz < alignment)
2113 			sz = alignment;
2114 		if (sz < MALLOC_PAGESIZE) {
2115 			pof2 = MALLOC_MINSIZE;
2116 			while (pof2 < sz)
2117 				pof2 <<= 1;
2118 		} else
2119 			pof2 = sz;
2120 		return omalloc(pool, pof2, zero_fill, f);
2121 	}
2122 
2123 	if (sz >= SIZE_MAX - mopts.malloc_guard - MALLOC_PAGESIZE) {
2124 		errno = ENOMEM;
2125 		return NULL;
2126 	}
2127 
2128 	if (sz < MALLOC_PAGESIZE)
2129 		sz = MALLOC_PAGESIZE;
2130 	sz += mopts.malloc_guard;
2131 	psz = PAGEROUND(sz);
2132 
2133 	p = mapalign(pool, alignment, psz, zero_fill);
2134 	if (p == MAP_FAILED) {
2135 		errno = ENOMEM;
2136 		return NULL;
2137 	}
2138 
2139 	if (insert(pool, p, sz, f)) {
2140 		unmap(pool, p, psz, 0);
2141 		errno = ENOMEM;
2142 		return NULL;
2143 	}
2144 
2145 	if (mopts.malloc_guard) {
2146 		if (mprotect((char *)p + psz - mopts.malloc_guard,
2147 		    mopts.malloc_guard, PROT_NONE))
2148 			wrterror(pool, "mprotect");
2149 		STATS_ADD(pool->malloc_guarded, mopts.malloc_guard);
2150 	}
2151 
2152 	if (pool->malloc_junk == 2) {
2153 		if (zero_fill)
2154 			memset((char *)p + sz - mopts.malloc_guard,
2155 			    SOME_JUNK, psz - sz);
2156 		else
2157 			memset(p, SOME_JUNK, psz - mopts.malloc_guard);
2158 	} else if (mopts.chunk_canaries)
2159 		fill_canary(p, sz - mopts.malloc_guard,
2160 		    psz - mopts.malloc_guard);
2161 
2162 	return p;
2163 }
2164 
2165 int
2166 posix_memalign(void **memptr, size_t alignment, size_t size)
2167 {
2168 	struct dir_info *d;
2169 	int res, saved_errno = errno;
2170 	void *r;
2171 
2172 	/* Make sure that alignment is a large enough power of 2. */
2173 	if (((alignment - 1) & alignment) != 0 || alignment < sizeof(void *))
2174 		return EINVAL;
2175 
2176 	d = getpool();
2177 	if (d == NULL) {
2178 		_malloc_init(0);
2179 		d = getpool();
2180 	}
2181 	_MALLOC_LOCK(d->mutex);
2182 	d->func = "posix_memalign";
2183 	if (d->active++) {
2184 		malloc_recurse(d);
2185 		goto err;
2186 	}
2187 	r = omemalign(d, alignment, size, 0, CALLER);
2188 	d->active--;
2189 	_MALLOC_UNLOCK(d->mutex);
2190 	if (r == NULL) {
2191 		if (mopts.malloc_xmalloc)
2192 			wrterror(d, "out of memory");
2193 		goto err;
2194 	}
2195 	errno = saved_errno;
2196 	*memptr = r;
2197 	return 0;
2198 
2199 err:
2200 	res = errno;
2201 	errno = saved_errno;
2202 	return res;
2203 }
2204 DEF_STRONG(posix_memalign);
2205 
2206 void *
2207 aligned_alloc(size_t alignment, size_t size)
2208 {
2209 	struct dir_info *d;
2210 	int saved_errno = errno;
2211 	void *r;
2212 
2213 	/* Make sure that alignment is a positive power of 2. */
2214 	if (((alignment - 1) & alignment) != 0 || alignment == 0) {
2215 		errno = EINVAL;
2216 		return NULL;
2217 	};
2218 	/* Per spec, size should be a multiple of alignment */
2219 	if ((size & (alignment - 1)) != 0) {
2220 		errno = EINVAL;
2221 		return NULL;
2222 	}
2223 
2224 	PROLOGUE(getpool(), "aligned_alloc")
2225 	r = omemalign(d, alignment, size, 0, CALLER);
2226 	EPILOGUE()
2227 	return r;
2228 }
2229 DEF_STRONG(aligned_alloc);
2230 
2231 #ifdef MALLOC_STATS
2232 
2233 static void
2234 ulog(const char *format, ...)
2235 {
2236 	va_list ap;
2237 	static char* buf;
2238 	static size_t filled;
2239 	int len;
2240 
2241 	if (buf == NULL)
2242 		buf = MMAP(KTR_USER_MAXLEN, 0);
2243 	if (buf == MAP_FAILED)
2244 		return;
2245 
2246 	va_start(ap, format);
2247 	len = vsnprintf(buf + filled, KTR_USER_MAXLEN - filled, format, ap);
2248 	va_end(ap);
2249 	if (len < 0)
2250 		return;
2251 	if (len > KTR_USER_MAXLEN - filled)
2252 		len = KTR_USER_MAXLEN - filled;
2253 	filled += len;
2254 	if (filled > 0) {
2255 		if (filled == KTR_USER_MAXLEN || buf[filled - 1] == '\n') {
2256 			utrace("malloc", buf, filled);
2257 			filled = 0;
2258 		}
2259 	}
2260 }
2261 
2262 struct malloc_leak {
2263 	void *f;
2264 	size_t total_size;
2265 	int count;
2266 };
2267 
2268 struct leaknode {
2269 	RBT_ENTRY(leaknode) entry;
2270 	struct malloc_leak d;
2271 };
2272 
2273 static inline int
2274 leakcmp(const struct leaknode *e1, const struct leaknode *e2)
2275 {
2276 	return e1->d.f < e2->d.f ? -1 : e1->d.f > e2->d.f;
2277 }
2278 
2279 RBT_HEAD(leaktree, leaknode);
2280 RBT_PROTOTYPE(leaktree, leaknode, entry, leakcmp);
2281 RBT_GENERATE(leaktree, leaknode, entry, leakcmp);
2282 
2283 static void
2284 putleakinfo(struct leaktree *leaks, void *f, size_t sz, int cnt)
2285 {
2286 	struct leaknode key, *p;
2287 	static struct leaknode *page;
2288 	static unsigned int used;
2289 
2290 	if (cnt == 0 || page == MAP_FAILED)
2291 		return;
2292 
2293 	key.d.f = f;
2294 	p = RBT_FIND(leaktree, leaks, &key);
2295 	if (p == NULL) {
2296 		if (page == NULL ||
2297 		    used >= MALLOC_PAGESIZE / sizeof(struct leaknode)) {
2298 			page = MMAP(MALLOC_PAGESIZE, 0);
2299 			if (page == MAP_FAILED)
2300 				return;
2301 			used = 0;
2302 		}
2303 		p = &page[used++];
2304 		p->d.f = f;
2305 		p->d.total_size = sz * cnt;
2306 		p->d.count = cnt;
2307 		RBT_INSERT(leaktree, leaks, p);
2308 	} else {
2309 		p->d.total_size += sz * cnt;
2310 		p->d.count += cnt;
2311 	}
2312 }
2313 
2314 static void
2315 dump_leaks(struct leaktree *leaks)
2316 {
2317 	struct leaknode *p;
2318 
2319 	ulog("Leak report:\n");
2320 	ulog("                 f     sum      #    avg\n");
2321 
2322 	RBT_FOREACH(p, leaktree, leaks) {
2323 		Dl_info info;
2324 		const char *caller = p->d.f;
2325 		const char *object = ".";
2326 
2327 		if (caller != NULL) {
2328 			if (dladdr(p->d.f, &info) != 0) {
2329 				caller -= (uintptr_t)info.dli_fbase;
2330 				object = info.dli_fname;
2331 			}
2332 		}
2333 		ulog("%18p %7zu %6u %6zu addr2line -e %s %p\n",
2334 		    p->d.f, p->d.total_size, p->d.count,
2335 		    p->d.total_size / p->d.count,
2336 		    object, caller);
2337 	}
2338 }
2339 
2340 static void
2341 dump_chunk(struct leaktree* leaks, struct chunk_info *p, void *f,
2342     int fromfreelist)
2343 {
2344 	while (p != NULL) {
2345 		if (mopts.malloc_verbose)
2346 			ulog("chunk %18p %18p %4zu %d/%d\n",
2347 			    p->page, ((p->bits[0] & 1) ? NULL : f),
2348 			    B2SIZE(p->bucket), p->free, p->total);
2349 		if (!fromfreelist) {
2350 			size_t sz =  B2SIZE(p->bucket);
2351 			if (p->bits[0] & 1)
2352 				putleakinfo(leaks, NULL, sz, p->total -
2353 				    p->free);
2354 			else {
2355 				putleakinfo(leaks, f, sz, 1);
2356 				putleakinfo(leaks, NULL, sz,
2357 				    p->total - p->free - 1);
2358 			}
2359 			break;
2360 		}
2361 		p = LIST_NEXT(p, entries);
2362 		if (mopts.malloc_verbose && p != NULL)
2363 			ulog("       ->");
2364 	}
2365 }
2366 
2367 static void
2368 dump_free_chunk_info(struct dir_info *d, struct leaktree *leaks)
2369 {
2370 	int i, j, count;
2371 	struct chunk_info *p;
2372 
2373 	ulog("Free chunk structs:\n");
2374 	ulog("Bkt) #CI                     page"
2375 	    "                  f size free/n\n");
2376 	for (i = 0; i <= BUCKETS; i++) {
2377 		count = 0;
2378 		LIST_FOREACH(p, &d->chunk_info_list[i], entries)
2379 			count++;
2380 		for (j = 0; j < MALLOC_CHUNK_LISTS; j++) {
2381 			p = LIST_FIRST(&d->chunk_dir[i][j]);
2382 			if (p == NULL && count == 0)
2383 				continue;
2384 			if (j == 0)
2385 				ulog("%3d) %3d ", i, count);
2386 			else
2387 				ulog("         ");
2388 			if (p != NULL)
2389 				dump_chunk(leaks, p, NULL, 1);
2390 			else
2391 				ulog(".\n");
2392 		}
2393 	}
2394 
2395 }
2396 
2397 static void
2398 dump_free_page_info(struct dir_info *d)
2399 {
2400 	struct smallcache *cache;
2401 	size_t i, total = 0;
2402 
2403 	ulog("Cached in small cache:\n");
2404 	for (i = 0; i < MAX_SMALLCACHEABLE_SIZE; i++) {
2405 		cache = &d->smallcache[i];
2406 		if (cache->length != 0)
2407 			ulog("%zu(%u): %u = %zu\n", i + 1, cache->max,
2408 			    cache->length, cache->length * (i + 1));
2409 		total += cache->length * (i + 1);
2410 	}
2411 
2412 	ulog("Cached in big cache: %zu/%zu\n", d->bigcache_used,
2413 	    d->bigcache_size);
2414 	for (i = 0; i < d->bigcache_size; i++) {
2415 		if (d->bigcache[i].psize != 0)
2416 			ulog("%zu: %zu\n", i, d->bigcache[i].psize);
2417 		total += d->bigcache[i].psize;
2418 	}
2419 	ulog("Free pages cached: %zu\n", total);
2420 }
2421 
2422 static void
2423 malloc_dump1(int poolno, struct dir_info *d, struct leaktree *leaks)
2424 {
2425 	size_t i, realsize;
2426 
2427 	if (mopts.malloc_verbose) {
2428 		ulog("Malloc dir of %s pool %d at %p\n", __progname, poolno, d);
2429 		ulog("MT=%d J=%d Fl=%x\n", d->malloc_mt, d->malloc_junk,
2430 		    d->mmap_flag);
2431 		ulog("Region slots free %zu/%zu\n",
2432 			d->regions_free, d->regions_total);
2433 		ulog("Finds %zu/%zu\n", d->finds, d->find_collisions);
2434 		ulog("Inserts %zu/%zu\n", d->inserts, d->insert_collisions);
2435 		ulog("Deletes %zu/%zu\n", d->deletes, d->delete_moves);
2436 		ulog("Cheap reallocs %zu/%zu\n",
2437 		    d->cheap_reallocs, d->cheap_realloc_tries);
2438 		ulog("Other pool searches %zu/%zu\n",
2439 		    d->other_pool, d->pool_searches);
2440 		ulog("In use %zu\n", d->malloc_used);
2441 		ulog("Guarded %zu\n", d->malloc_guarded);
2442 		dump_free_chunk_info(d, leaks);
2443 		dump_free_page_info(d);
2444 		ulog("Hash table:\n");
2445 		ulog("slot)  hash d  type               page                  "
2446 		    "f size [free/n]\n");
2447 	}
2448 	for (i = 0; i < d->regions_total; i++) {
2449 		if (d->r[i].p != NULL) {
2450 			size_t h = hash(d->r[i].p) &
2451 			    (d->regions_total - 1);
2452 			if (mopts.malloc_verbose)
2453 				ulog("%4zx) #%4zx %zd ",
2454 			        i, h, h - i);
2455 			REALSIZE(realsize, &d->r[i]);
2456 			if (realsize > MALLOC_MAXCHUNK) {
2457 				putleakinfo(leaks, d->r[i].f, realsize, 1);
2458 				if (mopts.malloc_verbose)
2459 					ulog("pages %18p %18p %zu\n", d->r[i].p,
2460 				        d->r[i].f, realsize);
2461 			} else
2462 				dump_chunk(leaks,
2463 				    (struct chunk_info *)d->r[i].size,
2464 				    d->r[i].f, 0);
2465 		}
2466 	}
2467 	if (mopts.malloc_verbose)
2468 		ulog("\n");
2469 }
2470 
2471 static void
2472 malloc_dump0(int poolno, struct dir_info *pool, struct leaktree *leaks)
2473 {
2474 	int i;
2475 	void *p;
2476 	struct region_info *r;
2477 
2478 	if (pool == NULL || pool->r == NULL)
2479 		return;
2480 	for (i = 0; i < MALLOC_DELAYED_CHUNK_MASK + 1; i++) {
2481 		p = pool->delayed_chunks[i];
2482 		if (p == NULL)
2483 			continue;
2484 		r = find(pool, p);
2485 		if (r == NULL)
2486 			wrterror(pool, "bogus pointer in malloc_dump %p", p);
2487 		free_bytes(pool, r, p);
2488 		pool->delayed_chunks[i] = NULL;
2489 	}
2490 	malloc_dump1(poolno, pool, leaks);
2491 }
2492 
2493 void
2494 malloc_dump(void)
2495 {
2496 	int i;
2497 	int saved_errno = errno;
2498 
2499 	/* XXX leak when run multiple times */
2500 	struct leaktree leaks = RBT_INITIALIZER(&leaks);
2501 
2502 	for (i = 0; i < mopts.malloc_mutexes; i++)
2503 		malloc_dump0(i, mopts.malloc_pool[i], &leaks);
2504 
2505 	dump_leaks(&leaks);
2506 	ulog("\n");
2507 	errno = saved_errno;
2508 }
2509 DEF_WEAK(malloc_dump);
2510 
2511 static void
2512 malloc_exit(void)
2513 {
2514 	int save_errno = errno;
2515 
2516 	ulog("******** Start dump %s *******\n", __progname);
2517 	ulog("M=%u I=%d F=%d U=%d J=%d R=%d X=%d C=%d cache=%u "
2518 	    "G=%zu\n",
2519 	    mopts.malloc_mutexes,
2520 	    mopts.internal_funcs, mopts.malloc_freecheck,
2521 	    mopts.malloc_freeunmap, mopts.def_malloc_junk,
2522 	    mopts.malloc_realloc, mopts.malloc_xmalloc,
2523 	    mopts.chunk_canaries, mopts.def_maxcache,
2524 	    mopts.malloc_guard);
2525 
2526 	malloc_dump();
2527 	ulog("******** End dump %s *******\n", __progname);
2528 	errno = save_errno;
2529 }
2530 
2531 #endif /* MALLOC_STATS */
2532