xref: /netbsd-src/lib/libc/stdlib/jemalloc.c (revision 62a8debe1dc62962e18a1c918def78666141273b)
1 /*	$NetBSD: jemalloc.c,v 1.21 2010/03/04 22:48:31 enami Exp $	*/
2 
3 /*-
4  * Copyright (C) 2006,2007 Jason Evans <jasone@FreeBSD.org>.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice(s), this list of conditions and the following disclaimer as
12  *    the first lines of this file unmodified other than the possible
13  *    addition of one or more copyright notices.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice(s), this list of conditions and the following disclaimer in
16  *    the documentation and/or other materials provided with the
17  *    distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
20  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE
23  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
26  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
28  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
29  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  *
31  *******************************************************************************
32  *
33  * This allocator implementation is designed to provide scalable performance
34  * for multi-threaded programs on multi-processor systems.  The following
35  * features are included for this purpose:
36  *
37  *   + Multiple arenas are used if there are multiple CPUs, which reduces lock
38  *     contention and cache sloshing.
39  *
40  *   + Cache line sharing between arenas is avoided for internal data
41  *     structures.
42  *
43  *   + Memory is managed in chunks and runs (chunks can be split into runs),
44  *     rather than as individual pages.  This provides a constant-time
45  *     mechanism for associating allocations with particular arenas.
46  *
47  * Allocation requests are rounded up to the nearest size class, and no record
48  * of the original request size is maintained.  Allocations are broken into
49  * categories according to size class.  Assuming runtime defaults, 4 kB pages
50  * and a 16 byte quantum, the size classes in each category are as follows:
51  *
52  *   |=====================================|
53  *   | Category | Subcategory    |    Size |
54  *   |=====================================|
55  *   | Small    | Tiny           |       2 |
56  *   |          |                |       4 |
57  *   |          |                |       8 |
58  *   |          |----------------+---------|
59  *   |          | Quantum-spaced |      16 |
60  *   |          |                |      32 |
61  *   |          |                |      48 |
62  *   |          |                |     ... |
63  *   |          |                |     480 |
64  *   |          |                |     496 |
65  *   |          |                |     512 |
66  *   |          |----------------+---------|
67  *   |          | Sub-page       |    1 kB |
68  *   |          |                |    2 kB |
69  *   |=====================================|
70  *   | Large                     |    4 kB |
71  *   |                           |    8 kB |
72  *   |                           |   12 kB |
73  *   |                           |     ... |
74  *   |                           | 1012 kB |
75  *   |                           | 1016 kB |
76  *   |                           | 1020 kB |
77  *   |=====================================|
78  *   | Huge                      |    1 MB |
79  *   |                           |    2 MB |
80  *   |                           |    3 MB |
81  *   |                           |     ... |
82  *   |=====================================|
83  *
84  * A different mechanism is used for each category:
85  *
86  *   Small : Each size class is segregated into its own set of runs.  Each run
87  *           maintains a bitmap of which regions are free/allocated.
88  *
89  *   Large : Each allocation is backed by a dedicated run.  Metadata are stored
90  *           in the associated arena chunk header maps.
91  *
92  *   Huge : Each allocation is backed by a dedicated contiguous set of chunks.
93  *          Metadata are stored in a separate red-black tree.
94  *
95  *******************************************************************************
96  */
97 
98 /* LINTLIBRARY */
99 
100 #ifdef __NetBSD__
101 #  define xutrace(a, b)		utrace("malloc", (a), (b))
102 #  define __DECONST(x, y)	((x)__UNCONST(y))
103 #  define NO_TLS
104 #else
105 #  define xutrace(a, b)		utrace((a), (b))
106 #endif	/* __NetBSD__ */
107 
108 /*
109  * MALLOC_PRODUCTION disables assertions and statistics gathering.  It also
110  * defaults the A and J runtime options to off.  These settings are appropriate
111  * for production systems.
112  */
113 #define MALLOC_PRODUCTION
114 
115 #ifndef MALLOC_PRODUCTION
116 #  define MALLOC_DEBUG
117 #endif
118 
119 #include <sys/cdefs.h>
120 /* __FBSDID("$FreeBSD: src/lib/libc/stdlib/malloc.c,v 1.147 2007/06/15 22:00:16 jasone Exp $"); */
121 __RCSID("$NetBSD: jemalloc.c,v 1.21 2010/03/04 22:48:31 enami Exp $");
122 
123 #ifdef __FreeBSD__
124 #include "libc_private.h"
125 #ifdef MALLOC_DEBUG
126 #  define _LOCK_DEBUG
127 #endif
128 #include "spinlock.h"
129 #endif
130 #include "namespace.h"
131 #include <sys/mman.h>
132 #include <sys/param.h>
133 #ifdef __FreeBSD__
134 #include <sys/stddef.h>
135 #endif
136 #include <sys/time.h>
137 #include <sys/types.h>
138 #include <sys/sysctl.h>
139 #include <sys/tree.h>
140 #include <sys/uio.h>
141 #include <sys/ktrace.h> /* Must come after several other sys/ includes. */
142 
143 #ifdef __FreeBSD__
144 #include <machine/atomic.h>
145 #include <machine/cpufunc.h>
146 #endif
147 #include <machine/vmparam.h>
148 
149 #include <errno.h>
150 #include <limits.h>
151 #include <pthread.h>
152 #include <sched.h>
153 #include <stdarg.h>
154 #include <stdbool.h>
155 #include <stdio.h>
156 #include <stdint.h>
157 #include <stdlib.h>
158 #include <string.h>
159 #include <strings.h>
160 #include <unistd.h>
161 
162 #ifdef __NetBSD__
163 #  include <reentrant.h>
164 #  include "extern.h"
165 
166 #define STRERROR_R(a, b, c)	__strerror_r(a, b, c);
167 /*
168  * A non localized version of strerror, that avoids bringing in
169  * stdio and the locale code. All the malloc messages are in English
170  * so why bother?
171  */
172 static int
173 __strerror_r(int e, char *s, size_t l)
174 {
175 	int rval;
176 	size_t slen;
177 
178 	if (e >= 0 && e < sys_nerr) {
179 		slen = strlcpy(s, sys_errlist[e], l);
180 		rval = 0;
181 	} else {
182 		slen = snprintf_ss(s, l, "Unknown error %u", e);
183 		rval = EINVAL;
184 	}
185 	return slen >= l ? ERANGE : rval;
186 }
187 #endif
188 
189 #ifdef __FreeBSD__
190 #define STRERROR_R(a, b, c)	strerror_r(a, b, c);
191 #include "un-namespace.h"
192 #endif
193 
194 /* MALLOC_STATS enables statistics calculation. */
195 #ifndef MALLOC_PRODUCTION
196 #  define MALLOC_STATS
197 #endif
198 
199 #ifdef MALLOC_DEBUG
200 #  ifdef NDEBUG
201 #    undef NDEBUG
202 #  endif
203 #else
204 #  ifndef NDEBUG
205 #    define NDEBUG
206 #  endif
207 #endif
208 #include <assert.h>
209 
210 #ifdef MALLOC_DEBUG
211    /* Disable inlining to make debugging easier. */
212 #  define inline
213 #endif
214 
215 /* Size of stack-allocated buffer passed to strerror_r(). */
216 #define	STRERROR_BUF		64
217 
218 /* Minimum alignment of allocations is 2^QUANTUM_2POW_MIN bytes. */
219 #ifdef __i386__
220 #  define QUANTUM_2POW_MIN	4
221 #  define SIZEOF_PTR_2POW	2
222 #  define USE_BRK
223 #endif
224 #ifdef __ia64__
225 #  define QUANTUM_2POW_MIN	4
226 #  define SIZEOF_PTR_2POW	3
227 #endif
228 #ifdef __alpha__
229 #  define QUANTUM_2POW_MIN	4
230 #  define SIZEOF_PTR_2POW	3
231 #  define NO_TLS
232 #endif
233 #ifdef __sparc64__
234 #  define QUANTUM_2POW_MIN	4
235 #  define SIZEOF_PTR_2POW	3
236 #  define NO_TLS
237 #endif
238 #ifdef __amd64__
239 #  define QUANTUM_2POW_MIN	4
240 #  define SIZEOF_PTR_2POW	3
241 #endif
242 #ifdef __arm__
243 #  define QUANTUM_2POW_MIN	3
244 #  define SIZEOF_PTR_2POW	2
245 #  define USE_BRK
246 #  define NO_TLS
247 #endif
248 #ifdef __powerpc__
249 #  define QUANTUM_2POW_MIN	4
250 #  define SIZEOF_PTR_2POW	2
251 #  define USE_BRK
252 #endif
253 #if defined(__sparc__) && !defined(__sparc64__)
254 #  define QUANTUM_2POW_MIN	4
255 #  define SIZEOF_PTR_2POW	2
256 #  define USE_BRK
257 #endif
258 #ifdef __vax__
259 #  define QUANTUM_2POW_MIN	4
260 #  define SIZEOF_PTR_2POW	2
261 #  define USE_BRK
262 #endif
263 #ifdef __sh__
264 #  define QUANTUM_2POW_MIN	4
265 #  define SIZEOF_PTR_2POW	2
266 #  define USE_BRK
267 #endif
268 #ifdef __m68k__
269 #  define QUANTUM_2POW_MIN	4
270 #  define SIZEOF_PTR_2POW	2
271 #  define USE_BRK
272 #endif
273 #ifdef __mips__
274 #  define QUANTUM_2POW_MIN	4
275 #  define SIZEOF_PTR_2POW	2
276 #  define USE_BRK
277 #endif
278 #ifdef __hppa__
279 #  define QUANTUM_2POW_MIN     4
280 #  define SIZEOF_PTR_2POW      2
281 #  define USE_BRK
282 #endif
283 
284 #define	SIZEOF_PTR		(1 << SIZEOF_PTR_2POW)
285 
286 /* sizeof(int) == (1 << SIZEOF_INT_2POW). */
287 #ifndef SIZEOF_INT_2POW
288 #  define SIZEOF_INT_2POW	2
289 #endif
290 
291 /* We can't use TLS in non-PIC programs, since TLS relies on loader magic. */
292 #if (!defined(PIC) && !defined(NO_TLS))
293 #  define NO_TLS
294 #endif
295 
296 /*
297  * Size and alignment of memory chunks that are allocated by the OS's virtual
298  * memory system.
299  */
300 #define	CHUNK_2POW_DEFAULT	20
301 
302 /*
303  * Maximum size of L1 cache line.  This is used to avoid cache line aliasing,
304  * so over-estimates are okay (up to a point), but under-estimates will
305  * negatively affect performance.
306  */
307 #define	CACHELINE_2POW		6
308 #define	CACHELINE		((size_t)(1 << CACHELINE_2POW))
309 
310 /* Smallest size class to support. */
311 #define	TINY_MIN_2POW		1
312 
313 /*
314  * Maximum size class that is a multiple of the quantum, but not (necessarily)
315  * a power of 2.  Above this size, allocations are rounded up to the nearest
316  * power of 2.
317  */
318 #define	SMALL_MAX_2POW_DEFAULT	9
319 #define	SMALL_MAX_DEFAULT	(1 << SMALL_MAX_2POW_DEFAULT)
320 
321 /*
322  * Maximum desired run header overhead.  Runs are sized as small as possible
323  * such that this setting is still honored, without violating other constraints.
324  * The goal is to make runs as small as possible without exceeding a per run
325  * external fragmentation threshold.
326  *
327  * Note that it is possible to set this low enough that it cannot be honored
328  * for some/all object sizes, since there is one bit of header overhead per
329  * object (plus a constant).  In such cases, this constraint is relaxed.
330  *
331  * RUN_MAX_OVRHD_RELAX specifies the maximum number of bits per region of
332  * overhead for which RUN_MAX_OVRHD is relaxed.
333  */
334 #define RUN_MAX_OVRHD		0.015
335 #define RUN_MAX_OVRHD_RELAX	1.5
336 
337 /* Put a cap on small object run size.  This overrides RUN_MAX_OVRHD. */
338 #define RUN_MAX_SMALL_2POW	15
339 #define RUN_MAX_SMALL		(1 << RUN_MAX_SMALL_2POW)
340 
341 /******************************************************************************/
342 
343 #ifdef __FreeBSD__
344 /*
345  * Mutexes based on spinlocks.  We can't use normal pthread mutexes, because
346  * they require malloc()ed memory.
347  */
348 typedef struct {
349 	spinlock_t	lock;
350 } malloc_mutex_t;
351 
352 /* Set to true once the allocator has been initialized. */
353 static bool malloc_initialized = false;
354 
355 /* Used to avoid initialization races. */
356 static malloc_mutex_t init_lock = {_SPINLOCK_INITIALIZER};
357 #else
358 #define	malloc_mutex_t	mutex_t
359 
360 /* Set to true once the allocator has been initialized. */
361 static bool malloc_initialized = false;
362 
363 /* Used to avoid initialization races. */
364 static mutex_t init_lock = MUTEX_INITIALIZER;
365 #endif
366 
367 /******************************************************************************/
368 /*
369  * Statistics data structures.
370  */
371 
372 #ifdef MALLOC_STATS
373 
374 typedef struct malloc_bin_stats_s malloc_bin_stats_t;
375 struct malloc_bin_stats_s {
376 	/*
377 	 * Number of allocation requests that corresponded to the size of this
378 	 * bin.
379 	 */
380 	uint64_t	nrequests;
381 
382 	/* Total number of runs created for this bin's size class. */
383 	uint64_t	nruns;
384 
385 	/*
386 	 * Total number of runs reused by extracting them from the runs tree for
387 	 * this bin's size class.
388 	 */
389 	uint64_t	reruns;
390 
391 	/* High-water mark for this bin. */
392 	unsigned long	highruns;
393 
394 	/* Current number of runs in this bin. */
395 	unsigned long	curruns;
396 };
397 
398 typedef struct arena_stats_s arena_stats_t;
399 struct arena_stats_s {
400 	/* Number of bytes currently mapped. */
401 	size_t		mapped;
402 
403 	/* Per-size-category statistics. */
404 	size_t		allocated_small;
405 	uint64_t	nmalloc_small;
406 	uint64_t	ndalloc_small;
407 
408 	size_t		allocated_large;
409 	uint64_t	nmalloc_large;
410 	uint64_t	ndalloc_large;
411 };
412 
413 typedef struct chunk_stats_s chunk_stats_t;
414 struct chunk_stats_s {
415 	/* Number of chunks that were allocated. */
416 	uint64_t	nchunks;
417 
418 	/* High-water mark for number of chunks allocated. */
419 	unsigned long	highchunks;
420 
421 	/*
422 	 * Current number of chunks allocated.  This value isn't maintained for
423 	 * any other purpose, so keep track of it in order to be able to set
424 	 * highchunks.
425 	 */
426 	unsigned long	curchunks;
427 };
428 
429 #endif /* #ifdef MALLOC_STATS */
430 
431 /******************************************************************************/
432 /*
433  * Chunk data structures.
434  */
435 
436 /* Tree of chunks. */
437 typedef struct chunk_node_s chunk_node_t;
438 struct chunk_node_s {
439 	/* Linkage for the chunk tree. */
440 	RB_ENTRY(chunk_node_s) link;
441 
442 	/*
443 	 * Pointer to the chunk that this tree node is responsible for.  In some
444 	 * (but certainly not all) cases, this data structure is placed at the
445 	 * beginning of the corresponding chunk, so this field may point to this
446 	 * node.
447 	 */
448 	void	*chunk;
449 
450 	/* Total chunk size. */
451 	size_t	size;
452 };
453 typedef struct chunk_tree_s chunk_tree_t;
454 RB_HEAD(chunk_tree_s, chunk_node_s);
455 
456 /******************************************************************************/
457 /*
458  * Arena data structures.
459  */
460 
461 typedef struct arena_s arena_t;
462 typedef struct arena_bin_s arena_bin_t;
463 
464 typedef struct arena_chunk_map_s arena_chunk_map_t;
465 struct arena_chunk_map_s {
466 	/* Number of pages in run. */
467 	uint32_t	npages;
468 	/*
469 	 * Position within run.  For a free run, this is POS_FREE for the first
470 	 * and last pages.  The POS_FREE special value makes it possible to
471 	 * quickly coalesce free runs.
472 	 *
473 	 * This is the limiting factor for chunksize; there can be at most 2^31
474 	 * pages in a run.
475 	 */
476 #define POS_FREE ((uint32_t)0xffffffffU)
477 	uint32_t	pos;
478 };
479 
480 /* Arena chunk header. */
481 typedef struct arena_chunk_s arena_chunk_t;
482 struct arena_chunk_s {
483 	/* Arena that owns the chunk. */
484 	arena_t *arena;
485 
486 	/* Linkage for the arena's chunk tree. */
487 	RB_ENTRY(arena_chunk_s) link;
488 
489 	/*
490 	 * Number of pages in use.  This is maintained in order to make
491 	 * detection of empty chunks fast.
492 	 */
493 	uint32_t pages_used;
494 
495 	/*
496 	 * Every time a free run larger than this value is created/coalesced,
497 	 * this value is increased.  The only way that the value decreases is if
498 	 * arena_run_alloc() fails to find a free run as large as advertised by
499 	 * this value.
500 	 */
501 	uint32_t max_frun_npages;
502 
503 	/*
504 	 * Every time a free run that starts at an earlier page than this value
505 	 * is created/coalesced, this value is decreased.  It is reset in a
506 	 * similar fashion to max_frun_npages.
507 	 */
508 	uint32_t min_frun_ind;
509 
510 	/*
511 	 * Map of pages within chunk that keeps track of free/large/small.  For
512 	 * free runs, only the map entries for the first and last pages are
513 	 * kept up to date, so that free runs can be quickly coalesced.
514 	 */
515 	arena_chunk_map_t map[1]; /* Dynamically sized. */
516 };
517 typedef struct arena_chunk_tree_s arena_chunk_tree_t;
518 RB_HEAD(arena_chunk_tree_s, arena_chunk_s);
519 
520 typedef struct arena_run_s arena_run_t;
521 struct arena_run_s {
522 	/* Linkage for run trees. */
523 	RB_ENTRY(arena_run_s) link;
524 
525 #ifdef MALLOC_DEBUG
526 	uint32_t	magic;
527 #  define ARENA_RUN_MAGIC 0x384adf93
528 #endif
529 
530 	/* Bin this run is associated with. */
531 	arena_bin_t	*bin;
532 
533 	/* Index of first element that might have a free region. */
534 	unsigned	regs_minelm;
535 
536 	/* Number of free regions in run. */
537 	unsigned	nfree;
538 
539 	/* Bitmask of in-use regions (0: in use, 1: free). */
540 	unsigned	regs_mask[1]; /* Dynamically sized. */
541 };
542 typedef struct arena_run_tree_s arena_run_tree_t;
543 RB_HEAD(arena_run_tree_s, arena_run_s);
544 
545 struct arena_bin_s {
546 	/*
547 	 * Current run being used to service allocations of this bin's size
548 	 * class.
549 	 */
550 	arena_run_t	*runcur;
551 
552 	/*
553 	 * Tree of non-full runs.  This tree is used when looking for an
554 	 * existing run when runcur is no longer usable.  We choose the
555 	 * non-full run that is lowest in memory; this policy tends to keep
556 	 * objects packed well, and it can also help reduce the number of
557 	 * almost-empty chunks.
558 	 */
559 	arena_run_tree_t runs;
560 
561 	/* Size of regions in a run for this bin's size class. */
562 	size_t		reg_size;
563 
564 	/* Total size of a run for this bin's size class. */
565 	size_t		run_size;
566 
567 	/* Total number of regions in a run for this bin's size class. */
568 	uint32_t	nregs;
569 
570 	/* Number of elements in a run's regs_mask for this bin's size class. */
571 	uint32_t	regs_mask_nelms;
572 
573 	/* Offset of first region in a run for this bin's size class. */
574 	uint32_t	reg0_offset;
575 
576 #ifdef MALLOC_STATS
577 	/* Bin statistics. */
578 	malloc_bin_stats_t stats;
579 #endif
580 };
581 
582 struct arena_s {
583 #ifdef MALLOC_DEBUG
584 	uint32_t		magic;
585 #  define ARENA_MAGIC 0x947d3d24
586 #endif
587 
588 	/* All operations on this arena require that mtx be locked. */
589 	malloc_mutex_t		mtx;
590 
591 #ifdef MALLOC_STATS
592 	arena_stats_t		stats;
593 #endif
594 
595 	/*
596 	 * Tree of chunks this arena manages.
597 	 */
598 	arena_chunk_tree_t	chunks;
599 
600 	/*
601 	 * In order to avoid rapid chunk allocation/deallocation when an arena
602 	 * oscillates right on the cusp of needing a new chunk, cache the most
603 	 * recently freed chunk.  This caching is disabled by opt_hint.
604 	 *
605 	 * There is one spare chunk per arena, rather than one spare total, in
606 	 * order to avoid interactions between multiple threads that could make
607 	 * a single spare inadequate.
608 	 */
609 	arena_chunk_t *spare;
610 
611 	/*
612 	 * bins is used to store rings of free regions of the following sizes,
613 	 * assuming a 16-byte quantum, 4kB pagesize, and default MALLOC_OPTIONS.
614 	 *
615 	 *   bins[i] | size |
616 	 *   --------+------+
617 	 *        0  |    2 |
618 	 *        1  |    4 |
619 	 *        2  |    8 |
620 	 *   --------+------+
621 	 *        3  |   16 |
622 	 *        4  |   32 |
623 	 *        5  |   48 |
624 	 *        6  |   64 |
625 	 *           :      :
626 	 *           :      :
627 	 *       33  |  496 |
628 	 *       34  |  512 |
629 	 *   --------+------+
630 	 *       35  | 1024 |
631 	 *       36  | 2048 |
632 	 *   --------+------+
633 	 */
634 	arena_bin_t		bins[1]; /* Dynamically sized. */
635 };
636 
637 /******************************************************************************/
638 /*
639  * Data.
640  */
641 
642 /* Number of CPUs. */
643 static unsigned		ncpus;
644 
645 /* VM page size. */
646 static size_t		pagesize;
647 static size_t		pagesize_mask;
648 static int		pagesize_2pow;
649 
650 /* Various bin-related settings. */
651 static size_t		bin_maxclass; /* Max size class for bins. */
652 static unsigned		ntbins; /* Number of (2^n)-spaced tiny bins. */
653 static unsigned		nqbins; /* Number of quantum-spaced bins. */
654 static unsigned		nsbins; /* Number of (2^n)-spaced sub-page bins. */
655 static size_t		small_min;
656 static size_t		small_max;
657 
658 /* Various quantum-related settings. */
659 static size_t		quantum;
660 static size_t		quantum_mask; /* (quantum - 1). */
661 
662 /* Various chunk-related settings. */
663 static size_t		chunksize;
664 static size_t		chunksize_mask; /* (chunksize - 1). */
665 static int		chunksize_2pow;
666 static unsigned		chunk_npages;
667 static unsigned		arena_chunk_header_npages;
668 static size_t		arena_maxclass; /* Max size class for arenas. */
669 
670 /********/
671 /*
672  * Chunks.
673  */
674 
675 /* Protects chunk-related data structures. */
676 static malloc_mutex_t	chunks_mtx;
677 
678 /* Tree of chunks that are stand-alone huge allocations. */
679 static chunk_tree_t	huge;
680 
681 #ifdef USE_BRK
682 /*
683  * Try to use brk for chunk-size allocations, due to address space constraints.
684  */
685 /*
686  * Protects sbrk() calls.  This must be separate from chunks_mtx, since
687  * base_pages_alloc() also uses sbrk(), but cannot lock chunks_mtx (doing so
688  * could cause recursive lock acquisition).
689  */
690 static malloc_mutex_t	brk_mtx;
691 /* Result of first sbrk(0) call. */
692 static void		*brk_base;
693 /* Current end of brk, or ((void *)-1) if brk is exhausted. */
694 static void		*brk_prev;
695 /* Current upper limit on brk addresses. */
696 static void		*brk_max;
697 #endif
698 
699 #ifdef MALLOC_STATS
700 /* Huge allocation statistics. */
701 static uint64_t		huge_nmalloc;
702 static uint64_t		huge_ndalloc;
703 static uint64_t		huge_nralloc;
704 static size_t		huge_allocated;
705 #endif
706 
707 /*
708  * Tree of chunks that were previously allocated.  This is used when allocating
709  * chunks, in an attempt to re-use address space.
710  */
711 static chunk_tree_t	old_chunks;
712 
713 /****************************/
714 /*
715  * base (internal allocation).
716  */
717 
718 /*
719  * Current pages that are being used for internal memory allocations.  These
720  * pages are carved up in cacheline-size quanta, so that there is no chance of
721  * false cache line sharing.
722  */
723 static void		*base_pages;
724 static void		*base_next_addr;
725 static void		*base_past_addr; /* Addr immediately past base_pages. */
726 static chunk_node_t	*base_chunk_nodes; /* LIFO cache of chunk nodes. */
727 static malloc_mutex_t	base_mtx;
728 #ifdef MALLOC_STATS
729 static size_t		base_mapped;
730 #endif
731 
732 /********/
733 /*
734  * Arenas.
735  */
736 
737 /*
738  * Arenas that are used to service external requests.  Not all elements of the
739  * arenas array are necessarily used; arenas are created lazily as needed.
740  */
741 static arena_t		**arenas;
742 static unsigned		narenas;
743 static unsigned		next_arena;
744 static malloc_mutex_t	arenas_mtx; /* Protects arenas initialization. */
745 
746 #ifndef NO_TLS
747 /*
748  * Map of pthread_self() --> arenas[???], used for selecting an arena to use
749  * for allocations.
750  */
751 static __thread arena_t	*arenas_map;
752 #define	get_arenas_map()	(arenas_map)
753 #define	set_arenas_map(x)	(arenas_map = x)
754 #else
755 static thread_key_t arenas_map_key;
756 #define	get_arenas_map()	thr_getspecific(arenas_map_key)
757 #define	set_arenas_map(x)	thr_setspecific(arenas_map_key, x)
758 #endif
759 
760 #ifdef MALLOC_STATS
761 /* Chunk statistics. */
762 static chunk_stats_t	stats_chunks;
763 #endif
764 
765 /*******************************/
766 /*
767  * Runtime configuration options.
768  */
769 const char	*_malloc_options;
770 
771 #ifndef MALLOC_PRODUCTION
772 static bool	opt_abort = true;
773 static bool	opt_junk = true;
774 #else
775 static bool	opt_abort = false;
776 static bool	opt_junk = false;
777 #endif
778 static bool	opt_hint = false;
779 static bool	opt_print_stats = false;
780 static int	opt_quantum_2pow = QUANTUM_2POW_MIN;
781 static int	opt_small_max_2pow = SMALL_MAX_2POW_DEFAULT;
782 static int	opt_chunk_2pow = CHUNK_2POW_DEFAULT;
783 static bool	opt_utrace = false;
784 static bool	opt_sysv = false;
785 static bool	opt_xmalloc = false;
786 static bool	opt_zero = false;
787 static int32_t	opt_narenas_lshift = 0;
788 
789 typedef struct {
790 	void	*p;
791 	size_t	s;
792 	void	*r;
793 } malloc_utrace_t;
794 
795 #define	UTRACE(a, b, c)							\
796 	if (opt_utrace) {						\
797 		malloc_utrace_t ut;					\
798 		ut.p = a;						\
799 		ut.s = b;						\
800 		ut.r = c;						\
801 		xutrace(&ut, sizeof(ut));				\
802 	}
803 
804 /******************************************************************************/
805 /*
806  * Begin function prototypes for non-inline static functions.
807  */
808 
809 static void	wrtmessage(const char *p1, const char *p2, const char *p3,
810 		const char *p4);
811 #ifdef MALLOC_STATS
812 static void	malloc_printf(const char *format, ...);
813 #endif
814 static char	*umax2s(uintmax_t x, char *s);
815 static bool	base_pages_alloc(size_t minsize);
816 static void	*base_alloc(size_t size);
817 static chunk_node_t *base_chunk_node_alloc(void);
818 static void	base_chunk_node_dealloc(chunk_node_t *node);
819 #ifdef MALLOC_STATS
820 static void	stats_print(arena_t *arena);
821 #endif
822 static void	*pages_map(void *addr, size_t size);
823 static void	*pages_map_align(void *addr, size_t size, int align);
824 static void	pages_unmap(void *addr, size_t size);
825 static void	*chunk_alloc(size_t size);
826 static void	chunk_dealloc(void *chunk, size_t size);
827 static void	arena_run_split(arena_t *arena, arena_run_t *run, size_t size);
828 static arena_chunk_t *arena_chunk_alloc(arena_t *arena);
829 static void	arena_chunk_dealloc(arena_t *arena, arena_chunk_t *chunk);
830 static arena_run_t *arena_run_alloc(arena_t *arena, size_t size);
831 static void	arena_run_dalloc(arena_t *arena, arena_run_t *run, size_t size);
832 static arena_run_t *arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin);
833 static void *arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin);
834 static size_t arena_bin_run_size_calc(arena_bin_t *bin, size_t min_run_size);
835 static void	*arena_malloc(arena_t *arena, size_t size);
836 static void	*arena_palloc(arena_t *arena, size_t alignment, size_t size,
837     size_t alloc_size);
838 static size_t	arena_salloc(const void *ptr);
839 static void	*arena_ralloc(void *ptr, size_t size, size_t oldsize);
840 static void	arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr);
841 static bool	arena_new(arena_t *arena);
842 static arena_t	*arenas_extend(unsigned ind);
843 static void	*huge_malloc(size_t size);
844 static void	*huge_palloc(size_t alignment, size_t size);
845 static void	*huge_ralloc(void *ptr, size_t size, size_t oldsize);
846 static void	huge_dalloc(void *ptr);
847 static void	*imalloc(size_t size);
848 static void	*ipalloc(size_t alignment, size_t size);
849 static void	*icalloc(size_t size);
850 static size_t	isalloc(const void *ptr);
851 static void	*iralloc(void *ptr, size_t size);
852 static void	idalloc(void *ptr);
853 static void	malloc_print_stats(void);
854 static bool	malloc_init_hard(void);
855 
856 /*
857  * End function prototypes.
858  */
859 /******************************************************************************/
860 /*
861  * Begin mutex.
862  */
863 
864 #ifdef __NetBSD__
865 #define	malloc_mutex_init(m)	mutex_init(m, NULL)
866 #define	malloc_mutex_lock(m)	mutex_lock(m)
867 #define	malloc_mutex_unlock(m)	mutex_unlock(m)
868 #else	/* __NetBSD__ */
869 static inline void
870 malloc_mutex_init(malloc_mutex_t *a_mutex)
871 {
872 	static const spinlock_t lock = _SPINLOCK_INITIALIZER;
873 
874 	a_mutex->lock = lock;
875 }
876 
877 static inline void
878 malloc_mutex_lock(malloc_mutex_t *a_mutex)
879 {
880 
881 	if (__isthreaded)
882 		_SPINLOCK(&a_mutex->lock);
883 }
884 
885 static inline void
886 malloc_mutex_unlock(malloc_mutex_t *a_mutex)
887 {
888 
889 	if (__isthreaded)
890 		_SPINUNLOCK(&a_mutex->lock);
891 }
892 #endif	/* __NetBSD__ */
893 
894 /*
895  * End mutex.
896  */
897 /******************************************************************************/
898 /*
899  * Begin Utility functions/macros.
900  */
901 
902 /* Return the chunk address for allocation address a. */
903 #define	CHUNK_ADDR2BASE(a)						\
904 	((void *)((uintptr_t)(a) & ~chunksize_mask))
905 
906 /* Return the chunk offset of address a. */
907 #define	CHUNK_ADDR2OFFSET(a)						\
908 	((size_t)((uintptr_t)(a) & chunksize_mask))
909 
910 /* Return the smallest chunk multiple that is >= s. */
911 #define	CHUNK_CEILING(s)						\
912 	(((s) + chunksize_mask) & ~chunksize_mask)
913 
914 /* Return the smallest cacheline multiple that is >= s. */
915 #define	CACHELINE_CEILING(s)						\
916 	(((s) + (CACHELINE - 1)) & ~(CACHELINE - 1))
917 
918 /* Return the smallest quantum multiple that is >= a. */
919 #define	QUANTUM_CEILING(a)						\
920 	(((a) + quantum_mask) & ~quantum_mask)
921 
922 /* Return the smallest pagesize multiple that is >= s. */
923 #define	PAGE_CEILING(s)							\
924 	(((s) + pagesize_mask) & ~pagesize_mask)
925 
926 /* Compute the smallest power of 2 that is >= x. */
927 static inline size_t
928 pow2_ceil(size_t x)
929 {
930 
931 	x--;
932 	x |= x >> 1;
933 	x |= x >> 2;
934 	x |= x >> 4;
935 	x |= x >> 8;
936 	x |= x >> 16;
937 #if (SIZEOF_PTR == 8)
938 	x |= x >> 32;
939 #endif
940 	x++;
941 	return (x);
942 }
943 
944 static void
945 wrtmessage(const char *p1, const char *p2, const char *p3, const char *p4)
946 {
947 
948 	write(STDERR_FILENO, p1, strlen(p1));
949 	write(STDERR_FILENO, p2, strlen(p2));
950 	write(STDERR_FILENO, p3, strlen(p3));
951 	write(STDERR_FILENO, p4, strlen(p4));
952 }
953 
954 void	(*_malloc_message)(const char *p1, const char *p2, const char *p3,
955 	    const char *p4) = wrtmessage;
956 
957 #ifdef MALLOC_STATS
958 /*
959  * Print to stderr in such a way as to (hopefully) avoid memory allocation.
960  */
961 static void
962 malloc_printf(const char *format, ...)
963 {
964 	char buf[4096];
965 	va_list ap;
966 
967 	va_start(ap, format);
968 	vsnprintf(buf, sizeof(buf), format, ap);
969 	va_end(ap);
970 	_malloc_message(buf, "", "", "");
971 }
972 #endif
973 
974 /*
975  * We don't want to depend on vsnprintf() for production builds, since that can
976  * cause unnecessary bloat for static binaries.  umax2s() provides minimal
977  * integer printing functionality, so that malloc_printf() use can be limited to
978  * MALLOC_STATS code.
979  */
980 #define UMAX2S_BUFSIZE	21
981 static char *
982 umax2s(uintmax_t x, char *s)
983 {
984 	unsigned i;
985 
986 	/* Make sure UMAX2S_BUFSIZE is large enough. */
987 	/* LINTED */
988 	assert(sizeof(uintmax_t) <= 8);
989 
990 	i = UMAX2S_BUFSIZE - 1;
991 	s[i] = '\0';
992 	do {
993 		i--;
994 		s[i] = "0123456789"[(int)x % 10];
995 		x /= (uintmax_t)10LL;
996 	} while (x > 0);
997 
998 	return (&s[i]);
999 }
1000 
1001 /******************************************************************************/
1002 
1003 static bool
1004 base_pages_alloc(size_t minsize)
1005 {
1006 	size_t csize = 0;
1007 
1008 #ifdef USE_BRK
1009 	/*
1010 	 * Do special brk allocation here, since base allocations don't need to
1011 	 * be chunk-aligned.
1012 	 */
1013 	if (brk_prev != (void *)-1) {
1014 		void *brk_cur;
1015 		intptr_t incr;
1016 
1017 		if (minsize != 0)
1018 			csize = CHUNK_CEILING(minsize);
1019 
1020 		malloc_mutex_lock(&brk_mtx);
1021 		do {
1022 			/* Get the current end of brk. */
1023 			brk_cur = sbrk(0);
1024 
1025 			/*
1026 			 * Calculate how much padding is necessary to
1027 			 * chunk-align the end of brk.  Don't worry about
1028 			 * brk_cur not being chunk-aligned though.
1029 			 */
1030 			incr = (intptr_t)chunksize
1031 			    - (intptr_t)CHUNK_ADDR2OFFSET(brk_cur);
1032 			assert(incr >= 0);
1033 			if ((size_t)incr < minsize)
1034 				incr += csize;
1035 
1036 			brk_prev = sbrk(incr);
1037 			if (brk_prev == brk_cur) {
1038 				/* Success. */
1039 				malloc_mutex_unlock(&brk_mtx);
1040 				base_pages = brk_cur;
1041 				base_next_addr = base_pages;
1042 				base_past_addr = (void *)((uintptr_t)base_pages
1043 				    + incr);
1044 #ifdef MALLOC_STATS
1045 				base_mapped += incr;
1046 #endif
1047 				return (false);
1048 			}
1049 		} while (brk_prev != (void *)-1);
1050 		malloc_mutex_unlock(&brk_mtx);
1051 	}
1052 	if (minsize == 0) {
1053 		/*
1054 		 * Failure during initialization doesn't matter, so avoid
1055 		 * falling through to the mmap-based page mapping code.
1056 		 */
1057 		return (true);
1058 	}
1059 #endif
1060 	assert(minsize != 0);
1061 	csize = PAGE_CEILING(minsize);
1062 	base_pages = pages_map(NULL, csize);
1063 	if (base_pages == NULL)
1064 		return (true);
1065 	base_next_addr = base_pages;
1066 	base_past_addr = (void *)((uintptr_t)base_pages + csize);
1067 #ifdef MALLOC_STATS
1068 	base_mapped += csize;
1069 #endif
1070 	return (false);
1071 }
1072 
1073 static void *
1074 base_alloc(size_t size)
1075 {
1076 	void *ret;
1077 	size_t csize;
1078 
1079 	/* Round size up to nearest multiple of the cacheline size. */
1080 	csize = CACHELINE_CEILING(size);
1081 
1082 	malloc_mutex_lock(&base_mtx);
1083 
1084 	/* Make sure there's enough space for the allocation. */
1085 	if ((uintptr_t)base_next_addr + csize > (uintptr_t)base_past_addr) {
1086 		if (base_pages_alloc(csize)) {
1087 			ret = NULL;
1088 			goto RETURN;
1089 		}
1090 	}
1091 
1092 	/* Allocate. */
1093 	ret = base_next_addr;
1094 	base_next_addr = (void *)((uintptr_t)base_next_addr + csize);
1095 
1096 RETURN:
1097 	malloc_mutex_unlock(&base_mtx);
1098 	return (ret);
1099 }
1100 
1101 static chunk_node_t *
1102 base_chunk_node_alloc(void)
1103 {
1104 	chunk_node_t *ret;
1105 
1106 	malloc_mutex_lock(&base_mtx);
1107 	if (base_chunk_nodes != NULL) {
1108 		ret = base_chunk_nodes;
1109 		/* LINTED */
1110 		base_chunk_nodes = *(chunk_node_t **)ret;
1111 		malloc_mutex_unlock(&base_mtx);
1112 	} else {
1113 		malloc_mutex_unlock(&base_mtx);
1114 		ret = (chunk_node_t *)base_alloc(sizeof(chunk_node_t));
1115 	}
1116 
1117 	return (ret);
1118 }
1119 
1120 static void
1121 base_chunk_node_dealloc(chunk_node_t *node)
1122 {
1123 
1124 	malloc_mutex_lock(&base_mtx);
1125 	/* LINTED */
1126 	*(chunk_node_t **)node = base_chunk_nodes;
1127 	base_chunk_nodes = node;
1128 	malloc_mutex_unlock(&base_mtx);
1129 }
1130 
1131 /******************************************************************************/
1132 
1133 #ifdef MALLOC_STATS
1134 static void
1135 stats_print(arena_t *arena)
1136 {
1137 	unsigned i;
1138 	int gap_start;
1139 
1140 	malloc_printf(
1141 	    "          allocated/mapped            nmalloc      ndalloc\n");
1142 
1143 	malloc_printf("small: %12zu %-12s %12llu %12llu\n",
1144 	    arena->stats.allocated_small, "", arena->stats.nmalloc_small,
1145 	    arena->stats.ndalloc_small);
1146 	malloc_printf("large: %12zu %-12s %12llu %12llu\n",
1147 	    arena->stats.allocated_large, "", arena->stats.nmalloc_large,
1148 	    arena->stats.ndalloc_large);
1149 	malloc_printf("total: %12zu/%-12zu %12llu %12llu\n",
1150 	    arena->stats.allocated_small + arena->stats.allocated_large,
1151 	    arena->stats.mapped,
1152 	    arena->stats.nmalloc_small + arena->stats.nmalloc_large,
1153 	    arena->stats.ndalloc_small + arena->stats.ndalloc_large);
1154 
1155 	malloc_printf("bins:     bin   size regs pgs  requests   newruns"
1156 	    "    reruns maxruns curruns\n");
1157 	for (i = 0, gap_start = -1; i < ntbins + nqbins + nsbins; i++) {
1158 		if (arena->bins[i].stats.nrequests == 0) {
1159 			if (gap_start == -1)
1160 				gap_start = i;
1161 		} else {
1162 			if (gap_start != -1) {
1163 				if (i > gap_start + 1) {
1164 					/* Gap of more than one size class. */
1165 					malloc_printf("[%u..%u]\n",
1166 					    gap_start, i - 1);
1167 				} else {
1168 					/* Gap of one size class. */
1169 					malloc_printf("[%u]\n", gap_start);
1170 				}
1171 				gap_start = -1;
1172 			}
1173 			malloc_printf(
1174 			    "%13u %1s %4u %4u %3u %9llu %9llu"
1175 			    " %9llu %7lu %7lu\n",
1176 			    i,
1177 			    i < ntbins ? "T" : i < ntbins + nqbins ? "Q" : "S",
1178 			    arena->bins[i].reg_size,
1179 			    arena->bins[i].nregs,
1180 			    arena->bins[i].run_size >> pagesize_2pow,
1181 			    arena->bins[i].stats.nrequests,
1182 			    arena->bins[i].stats.nruns,
1183 			    arena->bins[i].stats.reruns,
1184 			    arena->bins[i].stats.highruns,
1185 			    arena->bins[i].stats.curruns);
1186 		}
1187 	}
1188 	if (gap_start != -1) {
1189 		if (i > gap_start + 1) {
1190 			/* Gap of more than one size class. */
1191 			malloc_printf("[%u..%u]\n", gap_start, i - 1);
1192 		} else {
1193 			/* Gap of one size class. */
1194 			malloc_printf("[%u]\n", gap_start);
1195 		}
1196 	}
1197 }
1198 #endif
1199 
1200 /*
1201  * End Utility functions/macros.
1202  */
1203 /******************************************************************************/
1204 /*
1205  * Begin chunk management functions.
1206  */
1207 
1208 #ifndef lint
1209 static inline int
1210 chunk_comp(chunk_node_t *a, chunk_node_t *b)
1211 {
1212 
1213 	assert(a != NULL);
1214 	assert(b != NULL);
1215 
1216 	if ((uintptr_t)a->chunk < (uintptr_t)b->chunk)
1217 		return (-1);
1218 	else if (a->chunk == b->chunk)
1219 		return (0);
1220 	else
1221 		return (1);
1222 }
1223 
1224 /* Generate red-black tree code for chunks. */
1225 RB_GENERATE_STATIC(chunk_tree_s, chunk_node_s, link, chunk_comp);
1226 #endif
1227 
1228 static void *
1229 pages_map_align(void *addr, size_t size, int align)
1230 {
1231 	void *ret;
1232 
1233 	/*
1234 	 * We don't use MAP_FIXED here, because it can cause the *replacement*
1235 	 * of existing mappings, and we only want to create new mappings.
1236 	 */
1237 	ret = mmap(addr, size, PROT_READ | PROT_WRITE,
1238 	    MAP_PRIVATE | MAP_ANON | MAP_ALIGNED(align), -1, 0);
1239 	assert(ret != NULL);
1240 
1241 	if (ret == MAP_FAILED)
1242 		ret = NULL;
1243 	else if (addr != NULL && ret != addr) {
1244 		/*
1245 		 * We succeeded in mapping memory, but not in the right place.
1246 		 */
1247 		if (munmap(ret, size) == -1) {
1248 			char buf[STRERROR_BUF];
1249 
1250 			STRERROR_R(errno, buf, sizeof(buf));
1251 			_malloc_message(getprogname(),
1252 			    ": (malloc) Error in munmap(): ", buf, "\n");
1253 			if (opt_abort)
1254 				abort();
1255 		}
1256 		ret = NULL;
1257 	}
1258 
1259 	assert(ret == NULL || (addr == NULL && ret != addr)
1260 	    || (addr != NULL && ret == addr));
1261 	return (ret);
1262 }
1263 
1264 static void *
1265 pages_map(void *addr, size_t size)
1266 {
1267 
1268 	return pages_map_align(addr, size, 0);
1269 }
1270 
1271 static void
1272 pages_unmap(void *addr, size_t size)
1273 {
1274 
1275 	if (munmap(addr, size) == -1) {
1276 		char buf[STRERROR_BUF];
1277 
1278 		STRERROR_R(errno, buf, sizeof(buf));
1279 		_malloc_message(getprogname(),
1280 		    ": (malloc) Error in munmap(): ", buf, "\n");
1281 		if (opt_abort)
1282 			abort();
1283 	}
1284 }
1285 
1286 static void *
1287 chunk_alloc(size_t size)
1288 {
1289 	void *ret, *chunk;
1290 	chunk_node_t *tchunk, *delchunk;
1291 
1292 	assert(size != 0);
1293 	assert((size & chunksize_mask) == 0);
1294 
1295 	malloc_mutex_lock(&chunks_mtx);
1296 
1297 	if (size == chunksize) {
1298 		/*
1299 		 * Check for address ranges that were previously chunks and try
1300 		 * to use them.
1301 		 */
1302 
1303 		/* LINTED */
1304 		tchunk = RB_MIN(chunk_tree_s, &old_chunks);
1305 		while (tchunk != NULL) {
1306 			/* Found an address range.  Try to recycle it. */
1307 
1308 			chunk = tchunk->chunk;
1309 			delchunk = tchunk;
1310 			/* LINTED */
1311 			tchunk = RB_NEXT(chunk_tree_s, &old_chunks, delchunk);
1312 
1313 			/* Remove delchunk from the tree. */
1314 			/* LINTED */
1315 			RB_REMOVE(chunk_tree_s, &old_chunks, delchunk);
1316 			base_chunk_node_dealloc(delchunk);
1317 
1318 #ifdef USE_BRK
1319 			if ((uintptr_t)chunk >= (uintptr_t)brk_base
1320 			    && (uintptr_t)chunk < (uintptr_t)brk_max) {
1321 				/* Re-use a previously freed brk chunk. */
1322 				ret = chunk;
1323 				goto RETURN;
1324 			}
1325 #endif
1326 			if ((ret = pages_map(chunk, size)) != NULL) {
1327 				/* Success. */
1328 				goto RETURN;
1329 			}
1330 		}
1331 	}
1332 
1333 	/*
1334 	 * Try to over-allocate, but allow the OS to place the allocation
1335 	 * anywhere.  Beware of size_t wrap-around.
1336 	 */
1337 	if (size + chunksize > size) {
1338 		if ((ret = pages_map_align(NULL, size, chunksize_2pow))
1339 		    != NULL) {
1340 			goto RETURN;
1341 		}
1342 	}
1343 
1344 #ifdef USE_BRK
1345 	/*
1346 	 * Try to create allocations in brk, in order to make full use of
1347 	 * limited address space.
1348 	 */
1349 	if (brk_prev != (void *)-1) {
1350 		void *brk_cur;
1351 		intptr_t incr;
1352 
1353 		/*
1354 		 * The loop is necessary to recover from races with other
1355 		 * threads that are using brk for something other than malloc.
1356 		 */
1357 		malloc_mutex_lock(&brk_mtx);
1358 		do {
1359 			/* Get the current end of brk. */
1360 			brk_cur = sbrk(0);
1361 
1362 			/*
1363 			 * Calculate how much padding is necessary to
1364 			 * chunk-align the end of brk.
1365 			 */
1366 			incr = (intptr_t)size
1367 			    - (intptr_t)CHUNK_ADDR2OFFSET(brk_cur);
1368 			if (incr == (intptr_t)size) {
1369 				ret = brk_cur;
1370 			} else {
1371 				ret = (void *)((intptr_t)brk_cur + incr);
1372 				incr += size;
1373 			}
1374 
1375 			brk_prev = sbrk(incr);
1376 			if (brk_prev == brk_cur) {
1377 				/* Success. */
1378 				malloc_mutex_unlock(&brk_mtx);
1379 				brk_max = (void *)((intptr_t)ret + size);
1380 				goto RETURN;
1381 			}
1382 		} while (brk_prev != (void *)-1);
1383 		malloc_mutex_unlock(&brk_mtx);
1384 	}
1385 #endif
1386 
1387 	/* All strategies for allocation failed. */
1388 	ret = NULL;
1389 RETURN:
1390 	if (ret != NULL) {
1391 		chunk_node_t key;
1392 		/*
1393 		 * Clean out any entries in old_chunks that overlap with the
1394 		 * memory we just allocated.
1395 		 */
1396 		key.chunk = ret;
1397 		/* LINTED */
1398 		tchunk = RB_NFIND(chunk_tree_s, &old_chunks, &key);
1399 		while (tchunk != NULL
1400 		    && (uintptr_t)tchunk->chunk >= (uintptr_t)ret
1401 		    && (uintptr_t)tchunk->chunk < (uintptr_t)ret + size) {
1402 			delchunk = tchunk;
1403 			/* LINTED */
1404 			tchunk = RB_NEXT(chunk_tree_s, &old_chunks, delchunk);
1405 			/* LINTED */
1406 			RB_REMOVE(chunk_tree_s, &old_chunks, delchunk);
1407 			base_chunk_node_dealloc(delchunk);
1408 		}
1409 
1410 	}
1411 #ifdef MALLOC_STATS
1412 	if (ret != NULL) {
1413 		stats_chunks.nchunks += (size / chunksize);
1414 		stats_chunks.curchunks += (size / chunksize);
1415 	}
1416 	if (stats_chunks.curchunks > stats_chunks.highchunks)
1417 		stats_chunks.highchunks = stats_chunks.curchunks;
1418 #endif
1419 	malloc_mutex_unlock(&chunks_mtx);
1420 
1421 	assert(CHUNK_ADDR2BASE(ret) == ret);
1422 	return (ret);
1423 }
1424 
1425 static void
1426 chunk_dealloc(void *chunk, size_t size)
1427 {
1428 	chunk_node_t *node;
1429 
1430 	assert(chunk != NULL);
1431 	assert(CHUNK_ADDR2BASE(chunk) == chunk);
1432 	assert(size != 0);
1433 	assert((size & chunksize_mask) == 0);
1434 
1435 	malloc_mutex_lock(&chunks_mtx);
1436 
1437 #ifdef USE_BRK
1438 	if ((uintptr_t)chunk >= (uintptr_t)brk_base
1439 	    && (uintptr_t)chunk < (uintptr_t)brk_max) {
1440 		void *brk_cur;
1441 
1442 		malloc_mutex_lock(&brk_mtx);
1443 		/* Get the current end of brk. */
1444 		brk_cur = sbrk(0);
1445 
1446 		/*
1447 		 * Try to shrink the data segment if this chunk is at the end
1448 		 * of the data segment.  The sbrk() call here is subject to a
1449 		 * race condition with threads that use brk(2) or sbrk(2)
1450 		 * directly, but the alternative would be to leak memory for
1451 		 * the sake of poorly designed multi-threaded programs.
1452 		 */
1453 		if (brk_cur == brk_max
1454 		    && (void *)((uintptr_t)chunk + size) == brk_max
1455 		    && sbrk(-(intptr_t)size) == brk_max) {
1456 			malloc_mutex_unlock(&brk_mtx);
1457 			if (brk_prev == brk_max) {
1458 				/* Success. */
1459 				brk_prev = (void *)((intptr_t)brk_max
1460 				    - (intptr_t)size);
1461 				brk_max = brk_prev;
1462 			}
1463 		} else {
1464 			size_t offset;
1465 
1466 			malloc_mutex_unlock(&brk_mtx);
1467 			madvise(chunk, size, MADV_FREE);
1468 
1469 			/*
1470 			 * Iteratively create records of each chunk-sized
1471 			 * memory region that 'chunk' is comprised of, so that
1472 			 * the address range can be recycled if memory usage
1473 			 * increases later on.
1474 			 */
1475 			for (offset = 0; offset < size; offset += chunksize) {
1476 				node = base_chunk_node_alloc();
1477 				if (node == NULL)
1478 					break;
1479 
1480 				node->chunk = (void *)((uintptr_t)chunk
1481 				    + (uintptr_t)offset);
1482 				node->size = chunksize;
1483 				/* LINTED */
1484 				RB_INSERT(chunk_tree_s, &old_chunks, node);
1485 			}
1486 		}
1487 	} else {
1488 #endif
1489 		pages_unmap(chunk, size);
1490 
1491 		/*
1492 		 * Make a record of the chunk's address, so that the address
1493 		 * range can be recycled if memory usage increases later on.
1494 		 * Don't bother to create entries if (size > chunksize), since
1495 		 * doing so could cause scalability issues for truly gargantuan
1496 		 * objects (many gigabytes or larger).
1497 		 */
1498 		if (size == chunksize) {
1499 			node = base_chunk_node_alloc();
1500 			if (node != NULL) {
1501 				node->chunk = (void *)(uintptr_t)chunk;
1502 				node->size = chunksize;
1503 				/* LINTED */
1504 				RB_INSERT(chunk_tree_s, &old_chunks, node);
1505 			}
1506 		}
1507 #ifdef USE_BRK
1508 	}
1509 #endif
1510 
1511 #ifdef MALLOC_STATS
1512 	stats_chunks.curchunks -= (size / chunksize);
1513 #endif
1514 	malloc_mutex_unlock(&chunks_mtx);
1515 }
1516 
1517 /*
1518  * End chunk management functions.
1519  */
1520 /******************************************************************************/
1521 /*
1522  * Begin arena.
1523  */
1524 
1525 /*
1526  * Choose an arena based on a per-thread and (optimistically) per-CPU value.
1527  *
1528  * We maintain at least one block of arenas.  Usually there are more.
1529  * The blocks are $ncpu arenas in size.  Whole blocks are 'hashed'
1530  * amongst threads.  To accomplish this, next_arena advances only in
1531  * ncpu steps.
1532  */
1533 static __noinline arena_t *
1534 choose_arena_hard(void)
1535 {
1536 	unsigned i, curcpu;
1537 	arena_t **map;
1538 
1539 	/* Initialize the current block of arenas and advance to next. */
1540 	malloc_mutex_lock(&arenas_mtx);
1541 	assert(next_arena % ncpus == 0);
1542 	assert(narenas % ncpus == 0);
1543 	map = &arenas[next_arena];
1544 	set_arenas_map(map);
1545 	for (i = 0; i < ncpus; i++) {
1546 		if (arenas[next_arena] == NULL)
1547 			arenas_extend(next_arena);
1548 		next_arena = (next_arena + 1) % narenas;
1549 	}
1550 	malloc_mutex_unlock(&arenas_mtx);
1551 
1552 	/*
1553 	 * If we were unable to allocate an arena above, then default to
1554 	 * the first arena, which is always present.
1555 	 */
1556 	curcpu = thr_curcpu();
1557 	if (map[curcpu] != NULL)
1558 		return map[curcpu];
1559 	return arenas[0];
1560 }
1561 
1562 static inline arena_t *
1563 choose_arena(void)
1564 {
1565 	unsigned curcpu;
1566 	arena_t **map;
1567 
1568 	map = get_arenas_map();
1569 	curcpu = thr_curcpu();
1570 	if (__predict_true(map != NULL && map[curcpu] != NULL))
1571 		return map[curcpu];
1572 
1573         return choose_arena_hard();
1574 }
1575 
1576 #ifndef lint
1577 static inline int
1578 arena_chunk_comp(arena_chunk_t *a, arena_chunk_t *b)
1579 {
1580 
1581 	assert(a != NULL);
1582 	assert(b != NULL);
1583 
1584 	if ((uintptr_t)a < (uintptr_t)b)
1585 		return (-1);
1586 	else if (a == b)
1587 		return (0);
1588 	else
1589 		return (1);
1590 }
1591 
1592 /* Generate red-black tree code for arena chunks. */
1593 RB_GENERATE_STATIC(arena_chunk_tree_s, arena_chunk_s, link, arena_chunk_comp);
1594 #endif
1595 
1596 #ifndef lint
1597 static inline int
1598 arena_run_comp(arena_run_t *a, arena_run_t *b)
1599 {
1600 
1601 	assert(a != NULL);
1602 	assert(b != NULL);
1603 
1604 	if ((uintptr_t)a < (uintptr_t)b)
1605 		return (-1);
1606 	else if (a == b)
1607 		return (0);
1608 	else
1609 		return (1);
1610 }
1611 
1612 /* Generate red-black tree code for arena runs. */
1613 RB_GENERATE_STATIC(arena_run_tree_s, arena_run_s, link, arena_run_comp);
1614 #endif
1615 
1616 static inline void *
1617 arena_run_reg_alloc(arena_run_t *run, arena_bin_t *bin)
1618 {
1619 	void *ret;
1620 	unsigned i, mask, bit, regind;
1621 
1622 	assert(run->magic == ARENA_RUN_MAGIC);
1623 	assert(run->regs_minelm < bin->regs_mask_nelms);
1624 
1625 	/*
1626 	 * Move the first check outside the loop, so that run->regs_minelm can
1627 	 * be updated unconditionally, without the possibility of updating it
1628 	 * multiple times.
1629 	 */
1630 	i = run->regs_minelm;
1631 	mask = run->regs_mask[i];
1632 	if (mask != 0) {
1633 		/* Usable allocation found. */
1634 		bit = ffs((int)mask) - 1;
1635 
1636 		regind = ((i << (SIZEOF_INT_2POW + 3)) + bit);
1637 		ret = (void *)(((uintptr_t)run) + bin->reg0_offset
1638 		    + (bin->reg_size * regind));
1639 
1640 		/* Clear bit. */
1641 		mask ^= (1 << bit);
1642 		run->regs_mask[i] = mask;
1643 
1644 		return (ret);
1645 	}
1646 
1647 	for (i++; i < bin->regs_mask_nelms; i++) {
1648 		mask = run->regs_mask[i];
1649 		if (mask != 0) {
1650 			/* Usable allocation found. */
1651 			bit = ffs((int)mask) - 1;
1652 
1653 			regind = ((i << (SIZEOF_INT_2POW + 3)) + bit);
1654 			ret = (void *)(((uintptr_t)run) + bin->reg0_offset
1655 			    + (bin->reg_size * regind));
1656 
1657 			/* Clear bit. */
1658 			mask ^= (1 << bit);
1659 			run->regs_mask[i] = mask;
1660 
1661 			/*
1662 			 * Make a note that nothing before this element
1663 			 * contains a free region.
1664 			 */
1665 			run->regs_minelm = i; /* Low payoff: + (mask == 0); */
1666 
1667 			return (ret);
1668 		}
1669 	}
1670 	/* Not reached. */
1671 	/* LINTED */
1672 	assert(0);
1673 	return (NULL);
1674 }
1675 
1676 static inline void
1677 arena_run_reg_dalloc(arena_run_t *run, arena_bin_t *bin, void *ptr, size_t size)
1678 {
1679 	/*
1680 	 * To divide by a number D that is not a power of two we multiply
1681 	 * by (2^21 / D) and then right shift by 21 positions.
1682 	 *
1683 	 *   X / D
1684 	 *
1685 	 * becomes
1686 	 *
1687 	 *   (X * size_invs[(D >> QUANTUM_2POW_MIN) - 3]) >> SIZE_INV_SHIFT
1688 	 */
1689 #define SIZE_INV_SHIFT 21
1690 #define SIZE_INV(s) (((1 << SIZE_INV_SHIFT) / (s << QUANTUM_2POW_MIN)) + 1)
1691 	static const unsigned size_invs[] = {
1692 	    SIZE_INV(3),
1693 	    SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7),
1694 	    SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11),
1695 	    SIZE_INV(12),SIZE_INV(13), SIZE_INV(14), SIZE_INV(15),
1696 	    SIZE_INV(16),SIZE_INV(17), SIZE_INV(18), SIZE_INV(19),
1697 	    SIZE_INV(20),SIZE_INV(21), SIZE_INV(22), SIZE_INV(23),
1698 	    SIZE_INV(24),SIZE_INV(25), SIZE_INV(26), SIZE_INV(27),
1699 	    SIZE_INV(28),SIZE_INV(29), SIZE_INV(30), SIZE_INV(31)
1700 #if (QUANTUM_2POW_MIN < 4)
1701 	    ,
1702 	    SIZE_INV(32), SIZE_INV(33), SIZE_INV(34), SIZE_INV(35),
1703 	    SIZE_INV(36), SIZE_INV(37), SIZE_INV(38), SIZE_INV(39),
1704 	    SIZE_INV(40), SIZE_INV(41), SIZE_INV(42), SIZE_INV(43),
1705 	    SIZE_INV(44), SIZE_INV(45), SIZE_INV(46), SIZE_INV(47),
1706 	    SIZE_INV(48), SIZE_INV(49), SIZE_INV(50), SIZE_INV(51),
1707 	    SIZE_INV(52), SIZE_INV(53), SIZE_INV(54), SIZE_INV(55),
1708 	    SIZE_INV(56), SIZE_INV(57), SIZE_INV(58), SIZE_INV(59),
1709 	    SIZE_INV(60), SIZE_INV(61), SIZE_INV(62), SIZE_INV(63)
1710 #endif
1711 	};
1712 	unsigned diff, regind, elm, bit;
1713 
1714 	/* LINTED */
1715 	assert(run->magic == ARENA_RUN_MAGIC);
1716 	assert(((sizeof(size_invs)) / sizeof(unsigned)) + 3
1717 	    >= (SMALL_MAX_DEFAULT >> QUANTUM_2POW_MIN));
1718 
1719 	/*
1720 	 * Avoid doing division with a variable divisor if possible.  Using
1721 	 * actual division here can reduce allocator throughput by over 20%!
1722 	 */
1723 	diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - bin->reg0_offset);
1724 	if ((size & (size - 1)) == 0) {
1725 		/*
1726 		 * log2_table allows fast division of a power of two in the
1727 		 * [1..128] range.
1728 		 *
1729 		 * (x / divisor) becomes (x >> log2_table[divisor - 1]).
1730 		 */
1731 		static const unsigned char log2_table[] = {
1732 		    0, 1, 0, 2, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 4,
1733 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5,
1734 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1735 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6,
1736 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1737 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1738 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1739 		    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7
1740 		};
1741 
1742 		if (size <= 128)
1743 			regind = (diff >> log2_table[size - 1]);
1744 		else if (size <= 32768)
1745 			regind = diff >> (8 + log2_table[(size >> 8) - 1]);
1746 		else {
1747 			/*
1748 			 * The page size is too large for us to use the lookup
1749 			 * table.  Use real division.
1750 			 */
1751 			regind = (unsigned)(diff / size);
1752 		}
1753 	} else if (size <= ((sizeof(size_invs) / sizeof(unsigned))
1754 	    << QUANTUM_2POW_MIN) + 2) {
1755 		regind = size_invs[(size >> QUANTUM_2POW_MIN) - 3] * diff;
1756 		regind >>= SIZE_INV_SHIFT;
1757 	} else {
1758 		/*
1759 		 * size_invs isn't large enough to handle this size class, so
1760 		 * calculate regind using actual division.  This only happens
1761 		 * if the user increases small_max via the 'S' runtime
1762 		 * configuration option.
1763 		 */
1764 		regind = (unsigned)(diff / size);
1765 	};
1766 	assert(diff == regind * size);
1767 	assert(regind < bin->nregs);
1768 
1769 	elm = regind >> (SIZEOF_INT_2POW + 3);
1770 	if (elm < run->regs_minelm)
1771 		run->regs_minelm = elm;
1772 	bit = regind - (elm << (SIZEOF_INT_2POW + 3));
1773 	assert((run->regs_mask[elm] & (1 << bit)) == 0);
1774 	run->regs_mask[elm] |= (1 << bit);
1775 #undef SIZE_INV
1776 #undef SIZE_INV_SHIFT
1777 }
1778 
1779 static void
1780 arena_run_split(arena_t *arena, arena_run_t *run, size_t size)
1781 {
1782 	arena_chunk_t *chunk;
1783 	unsigned run_ind, map_offset, total_pages, need_pages, rem_pages;
1784 	unsigned i;
1785 
1786 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
1787 	run_ind = (unsigned)(((uintptr_t)run - (uintptr_t)chunk)
1788 	    >> pagesize_2pow);
1789 	total_pages = chunk->map[run_ind].npages;
1790 	need_pages = (unsigned)(size >> pagesize_2pow);
1791 	assert(need_pages <= total_pages);
1792 	rem_pages = total_pages - need_pages;
1793 
1794 	/* Split enough pages from the front of run to fit allocation size. */
1795 	map_offset = run_ind;
1796 	for (i = 0; i < need_pages; i++) {
1797 		chunk->map[map_offset + i].npages = need_pages;
1798 		chunk->map[map_offset + i].pos = i;
1799 	}
1800 
1801 	/* Keep track of trailing unused pages for later use. */
1802 	if (rem_pages > 0) {
1803 		/* Update map for trailing pages. */
1804 		map_offset += need_pages;
1805 		chunk->map[map_offset].npages = rem_pages;
1806 		chunk->map[map_offset].pos = POS_FREE;
1807 		chunk->map[map_offset + rem_pages - 1].npages = rem_pages;
1808 		chunk->map[map_offset + rem_pages - 1].pos = POS_FREE;
1809 	}
1810 
1811 	chunk->pages_used += need_pages;
1812 }
1813 
1814 static arena_chunk_t *
1815 arena_chunk_alloc(arena_t *arena)
1816 {
1817 	arena_chunk_t *chunk;
1818 
1819 	if (arena->spare != NULL) {
1820 		chunk = arena->spare;
1821 		arena->spare = NULL;
1822 
1823 		/* LINTED */
1824 		RB_INSERT(arena_chunk_tree_s, &arena->chunks, chunk);
1825 	} else {
1826 		chunk = (arena_chunk_t *)chunk_alloc(chunksize);
1827 		if (chunk == NULL)
1828 			return (NULL);
1829 #ifdef MALLOC_STATS
1830 		arena->stats.mapped += chunksize;
1831 #endif
1832 
1833 		chunk->arena = arena;
1834 
1835 		/* LINTED */
1836 		RB_INSERT(arena_chunk_tree_s, &arena->chunks, chunk);
1837 
1838 		/*
1839 		 * Claim that no pages are in use, since the header is merely
1840 		 * overhead.
1841 		 */
1842 		chunk->pages_used = 0;
1843 
1844 		chunk->max_frun_npages = chunk_npages -
1845 		    arena_chunk_header_npages;
1846 		chunk->min_frun_ind = arena_chunk_header_npages;
1847 
1848 		/*
1849 		 * Initialize enough of the map to support one maximal free run.
1850 		 */
1851 		chunk->map[arena_chunk_header_npages].npages = chunk_npages -
1852 		    arena_chunk_header_npages;
1853 		chunk->map[arena_chunk_header_npages].pos = POS_FREE;
1854 		chunk->map[chunk_npages - 1].npages = chunk_npages -
1855 		    arena_chunk_header_npages;
1856 		chunk->map[chunk_npages - 1].pos = POS_FREE;
1857 	}
1858 
1859 	return (chunk);
1860 }
1861 
1862 static void
1863 arena_chunk_dealloc(arena_t *arena, arena_chunk_t *chunk)
1864 {
1865 
1866 	/*
1867 	 * Remove chunk from the chunk tree, regardless of whether this chunk
1868 	 * will be cached, so that the arena does not use it.
1869 	 */
1870 	/* LINTED */
1871 	RB_REMOVE(arena_chunk_tree_s, &chunk->arena->chunks, chunk);
1872 
1873 	if (opt_hint == false) {
1874 		if (arena->spare != NULL) {
1875 			chunk_dealloc((void *)arena->spare, chunksize);
1876 #ifdef MALLOC_STATS
1877 			arena->stats.mapped -= chunksize;
1878 #endif
1879 		}
1880 		arena->spare = chunk;
1881 	} else {
1882 		assert(arena->spare == NULL);
1883 		chunk_dealloc((void *)chunk, chunksize);
1884 #ifdef MALLOC_STATS
1885 		arena->stats.mapped -= chunksize;
1886 #endif
1887 	}
1888 }
1889 
1890 static arena_run_t *
1891 arena_run_alloc(arena_t *arena, size_t size)
1892 {
1893 	arena_chunk_t *chunk;
1894 	arena_run_t *run;
1895 	unsigned need_npages, limit_pages, compl_need_npages;
1896 
1897 	assert(size <= (chunksize - (arena_chunk_header_npages <<
1898 	    pagesize_2pow)));
1899 	assert((size & pagesize_mask) == 0);
1900 
1901 	/*
1902 	 * Search through arena's chunks in address order for a free run that is
1903 	 * large enough.  Look for the first fit.
1904 	 */
1905 	need_npages = (unsigned)(size >> pagesize_2pow);
1906 	limit_pages = chunk_npages - arena_chunk_header_npages;
1907 	compl_need_npages = limit_pages - need_npages;
1908 	/* LINTED */
1909 	RB_FOREACH(chunk, arena_chunk_tree_s, &arena->chunks) {
1910 		/*
1911 		 * Avoid searching this chunk if there are not enough
1912 		 * contiguous free pages for there to possibly be a large
1913 		 * enough free run.
1914 		 */
1915 		if (chunk->pages_used <= compl_need_npages &&
1916 		    need_npages <= chunk->max_frun_npages) {
1917 			arena_chunk_map_t *mapelm;
1918 			unsigned i;
1919 			unsigned max_frun_npages = 0;
1920 			unsigned min_frun_ind = chunk_npages;
1921 
1922 			assert(chunk->min_frun_ind >=
1923 			    arena_chunk_header_npages);
1924 			for (i = chunk->min_frun_ind; i < chunk_npages;) {
1925 				mapelm = &chunk->map[i];
1926 				if (mapelm->pos == POS_FREE) {
1927 					if (mapelm->npages >= need_npages) {
1928 						run = (arena_run_t *)
1929 						    ((uintptr_t)chunk + (i <<
1930 						    pagesize_2pow));
1931 						/* Update page map. */
1932 						arena_run_split(arena, run,
1933 						    size);
1934 						return (run);
1935 					}
1936 					if (mapelm->npages >
1937 					    max_frun_npages) {
1938 						max_frun_npages =
1939 						    mapelm->npages;
1940 					}
1941 					if (i < min_frun_ind) {
1942 						min_frun_ind = i;
1943 						if (i < chunk->min_frun_ind)
1944 							chunk->min_frun_ind = i;
1945 					}
1946 				}
1947 				i += mapelm->npages;
1948 			}
1949 			/*
1950 			 * Search failure.  Reset cached chunk->max_frun_npages.
1951 			 * chunk->min_frun_ind was already reset above (if
1952 			 * necessary).
1953 			 */
1954 			chunk->max_frun_npages = max_frun_npages;
1955 		}
1956 	}
1957 
1958 	/*
1959 	 * No usable runs.  Create a new chunk from which to allocate the run.
1960 	 */
1961 	chunk = arena_chunk_alloc(arena);
1962 	if (chunk == NULL)
1963 		return (NULL);
1964 	run = (arena_run_t *)((uintptr_t)chunk + (arena_chunk_header_npages <<
1965 	    pagesize_2pow));
1966 	/* Update page map. */
1967 	arena_run_split(arena, run, size);
1968 	return (run);
1969 }
1970 
1971 static void
1972 arena_run_dalloc(arena_t *arena, arena_run_t *run, size_t size)
1973 {
1974 	arena_chunk_t *chunk;
1975 	unsigned run_ind, run_pages;
1976 
1977 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run);
1978 
1979 	run_ind = (unsigned)(((uintptr_t)run - (uintptr_t)chunk)
1980 	    >> pagesize_2pow);
1981 	assert(run_ind >= arena_chunk_header_npages);
1982 	assert(run_ind < (chunksize >> pagesize_2pow));
1983 	run_pages = (unsigned)(size >> pagesize_2pow);
1984 	assert(run_pages == chunk->map[run_ind].npages);
1985 
1986 	/* Subtract pages from count of pages used in chunk. */
1987 	chunk->pages_used -= run_pages;
1988 
1989 	/* Mark run as deallocated. */
1990 	assert(chunk->map[run_ind].npages == run_pages);
1991 	chunk->map[run_ind].pos = POS_FREE;
1992 	assert(chunk->map[run_ind + run_pages - 1].npages == run_pages);
1993 	chunk->map[run_ind + run_pages - 1].pos = POS_FREE;
1994 
1995 	/*
1996 	 * Tell the kernel that we don't need the data in this run, but only if
1997 	 * requested via runtime configuration.
1998 	 */
1999 	if (opt_hint)
2000 		madvise(run, size, MADV_FREE);
2001 
2002 	/* Try to coalesce with neighboring runs. */
2003 	if (run_ind > arena_chunk_header_npages &&
2004 	    chunk->map[run_ind - 1].pos == POS_FREE) {
2005 		unsigned prev_npages;
2006 
2007 		/* Coalesce with previous run. */
2008 		prev_npages = chunk->map[run_ind - 1].npages;
2009 		run_ind -= prev_npages;
2010 		assert(chunk->map[run_ind].npages == prev_npages);
2011 		assert(chunk->map[run_ind].pos == POS_FREE);
2012 		run_pages += prev_npages;
2013 
2014 		chunk->map[run_ind].npages = run_pages;
2015 		assert(chunk->map[run_ind].pos == POS_FREE);
2016 		chunk->map[run_ind + run_pages - 1].npages = run_pages;
2017 		assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE);
2018 	}
2019 
2020 	if (run_ind + run_pages < chunk_npages &&
2021 	    chunk->map[run_ind + run_pages].pos == POS_FREE) {
2022 		unsigned next_npages;
2023 
2024 		/* Coalesce with next run. */
2025 		next_npages = chunk->map[run_ind + run_pages].npages;
2026 		run_pages += next_npages;
2027 		assert(chunk->map[run_ind + run_pages - 1].npages ==
2028 		    next_npages);
2029 		assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE);
2030 
2031 		chunk->map[run_ind].npages = run_pages;
2032 		chunk->map[run_ind].pos = POS_FREE;
2033 		chunk->map[run_ind + run_pages - 1].npages = run_pages;
2034 		assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE);
2035 	}
2036 
2037 	if (chunk->map[run_ind].npages > chunk->max_frun_npages)
2038 		chunk->max_frun_npages = chunk->map[run_ind].npages;
2039 	if (run_ind < chunk->min_frun_ind)
2040 		chunk->min_frun_ind = run_ind;
2041 
2042 	/* Deallocate chunk if it is now completely unused. */
2043 	if (chunk->pages_used == 0)
2044 		arena_chunk_dealloc(arena, chunk);
2045 }
2046 
2047 static arena_run_t *
2048 arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin)
2049 {
2050 	arena_run_t *run;
2051 	unsigned i, remainder;
2052 
2053 	/* Look for a usable run. */
2054 	/* LINTED */
2055 	if ((run = RB_MIN(arena_run_tree_s, &bin->runs)) != NULL) {
2056 		/* run is guaranteed to have available space. */
2057 		/* LINTED */
2058 		RB_REMOVE(arena_run_tree_s, &bin->runs, run);
2059 #ifdef MALLOC_STATS
2060 		bin->stats.reruns++;
2061 #endif
2062 		return (run);
2063 	}
2064 	/* No existing runs have any space available. */
2065 
2066 	/* Allocate a new run. */
2067 	run = arena_run_alloc(arena, bin->run_size);
2068 	if (run == NULL)
2069 		return (NULL);
2070 
2071 	/* Initialize run internals. */
2072 	run->bin = bin;
2073 
2074 	for (i = 0; i < bin->regs_mask_nelms; i++)
2075 		run->regs_mask[i] = UINT_MAX;
2076 	remainder = bin->nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1);
2077 	if (remainder != 0) {
2078 		/* The last element has spare bits that need to be unset. */
2079 		run->regs_mask[i] = (UINT_MAX >> ((1 << (SIZEOF_INT_2POW + 3))
2080 		    - remainder));
2081 	}
2082 
2083 	run->regs_minelm = 0;
2084 
2085 	run->nfree = bin->nregs;
2086 #ifdef MALLOC_DEBUG
2087 	run->magic = ARENA_RUN_MAGIC;
2088 #endif
2089 
2090 #ifdef MALLOC_STATS
2091 	bin->stats.nruns++;
2092 	bin->stats.curruns++;
2093 	if (bin->stats.curruns > bin->stats.highruns)
2094 		bin->stats.highruns = bin->stats.curruns;
2095 #endif
2096 	return (run);
2097 }
2098 
2099 /* bin->runcur must have space available before this function is called. */
2100 static inline void *
2101 arena_bin_malloc_easy(arena_t *arena, arena_bin_t *bin, arena_run_t *run)
2102 {
2103 	void *ret;
2104 
2105 	assert(run->magic == ARENA_RUN_MAGIC);
2106 	assert(run->nfree > 0);
2107 
2108 	ret = arena_run_reg_alloc(run, bin);
2109 	assert(ret != NULL);
2110 	run->nfree--;
2111 
2112 	return (ret);
2113 }
2114 
2115 /* Re-fill bin->runcur, then call arena_bin_malloc_easy(). */
2116 static void *
2117 arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin)
2118 {
2119 
2120 	bin->runcur = arena_bin_nonfull_run_get(arena, bin);
2121 	if (bin->runcur == NULL)
2122 		return (NULL);
2123 	assert(bin->runcur->magic == ARENA_RUN_MAGIC);
2124 	assert(bin->runcur->nfree > 0);
2125 
2126 	return (arena_bin_malloc_easy(arena, bin, bin->runcur));
2127 }
2128 
2129 /*
2130  * Calculate bin->run_size such that it meets the following constraints:
2131  *
2132  *   *) bin->run_size >= min_run_size
2133  *   *) bin->run_size <= arena_maxclass
2134  *   *) bin->run_size <= RUN_MAX_SMALL
2135  *   *) run header overhead <= RUN_MAX_OVRHD (or header overhead relaxed).
2136  *
2137  * bin->nregs, bin->regs_mask_nelms, and bin->reg0_offset are
2138  * also calculated here, since these settings are all interdependent.
2139  */
2140 static size_t
2141 arena_bin_run_size_calc(arena_bin_t *bin, size_t min_run_size)
2142 {
2143 	size_t try_run_size, good_run_size;
2144 	unsigned good_nregs, good_mask_nelms, good_reg0_offset;
2145 	unsigned try_nregs, try_mask_nelms, try_reg0_offset;
2146 	float max_ovrhd = RUN_MAX_OVRHD;
2147 
2148 	assert(min_run_size >= pagesize);
2149 	assert(min_run_size <= arena_maxclass);
2150 	assert(min_run_size <= RUN_MAX_SMALL);
2151 
2152 	/*
2153 	 * Calculate known-valid settings before entering the run_size
2154 	 * expansion loop, so that the first part of the loop always copies
2155 	 * valid settings.
2156 	 *
2157 	 * The do..while loop iteratively reduces the number of regions until
2158 	 * the run header and the regions no longer overlap.  A closed formula
2159 	 * would be quite messy, since there is an interdependency between the
2160 	 * header's mask length and the number of regions.
2161 	 */
2162 	try_run_size = min_run_size;
2163 	try_nregs = (unsigned)(((try_run_size - sizeof(arena_run_t)) /
2164 	    bin->reg_size) + 1); /* Counter-act the first line of the loop. */
2165 	do {
2166 		try_nregs--;
2167 		try_mask_nelms = (try_nregs >> (SIZEOF_INT_2POW + 3)) +
2168 		    ((try_nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1)) ? 1 : 0);
2169 		try_reg0_offset = (unsigned)(try_run_size -
2170 		    (try_nregs * bin->reg_size));
2171 	} while (sizeof(arena_run_t) + (sizeof(unsigned) * (try_mask_nelms - 1))
2172 	    > try_reg0_offset);
2173 
2174 	/* run_size expansion loop. */
2175 	do {
2176 		/*
2177 		 * Copy valid settings before trying more aggressive settings.
2178 		 */
2179 		good_run_size = try_run_size;
2180 		good_nregs = try_nregs;
2181 		good_mask_nelms = try_mask_nelms;
2182 		good_reg0_offset = try_reg0_offset;
2183 
2184 		/* Try more aggressive settings. */
2185 		try_run_size += pagesize;
2186 		try_nregs = (unsigned)(((try_run_size - sizeof(arena_run_t)) /
2187 		    bin->reg_size) + 1); /* Counter-act try_nregs-- in loop. */
2188 		do {
2189 			try_nregs--;
2190 			try_mask_nelms = (try_nregs >> (SIZEOF_INT_2POW + 3)) +
2191 			    ((try_nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1)) ?
2192 			    1 : 0);
2193 			try_reg0_offset = (unsigned)(try_run_size - (try_nregs *
2194 			    bin->reg_size));
2195 		} while (sizeof(arena_run_t) + (sizeof(unsigned) *
2196 		    (try_mask_nelms - 1)) > try_reg0_offset);
2197 	} while (try_run_size <= arena_maxclass && try_run_size <= RUN_MAX_SMALL
2198 	    && max_ovrhd > RUN_MAX_OVRHD_RELAX / ((float)(bin->reg_size << 3))
2199 	    && ((float)(try_reg0_offset)) / ((float)(try_run_size)) >
2200 	    max_ovrhd);
2201 
2202 	assert(sizeof(arena_run_t) + (sizeof(unsigned) * (good_mask_nelms - 1))
2203 	    <= good_reg0_offset);
2204 	assert((good_mask_nelms << (SIZEOF_INT_2POW + 3)) >= good_nregs);
2205 
2206 	/* Copy final settings. */
2207 	bin->run_size = good_run_size;
2208 	bin->nregs = good_nregs;
2209 	bin->regs_mask_nelms = good_mask_nelms;
2210 	bin->reg0_offset = good_reg0_offset;
2211 
2212 	return (good_run_size);
2213 }
2214 
2215 static void *
2216 arena_malloc(arena_t *arena, size_t size)
2217 {
2218 	void *ret;
2219 
2220 	assert(arena != NULL);
2221 	assert(arena->magic == ARENA_MAGIC);
2222 	assert(size != 0);
2223 	assert(QUANTUM_CEILING(size) <= arena_maxclass);
2224 
2225 	if (size <= bin_maxclass) {
2226 		arena_bin_t *bin;
2227 		arena_run_t *run;
2228 
2229 		/* Small allocation. */
2230 
2231 		if (size < small_min) {
2232 			/* Tiny. */
2233 			size = pow2_ceil(size);
2234 			bin = &arena->bins[ffs((int)(size >> (TINY_MIN_2POW +
2235 			    1)))];
2236 #if (!defined(NDEBUG) || defined(MALLOC_STATS))
2237 			/*
2238 			 * Bin calculation is always correct, but we may need
2239 			 * to fix size for the purposes of assertions and/or
2240 			 * stats accuracy.
2241 			 */
2242 			if (size < (1 << TINY_MIN_2POW))
2243 				size = (1 << TINY_MIN_2POW);
2244 #endif
2245 		} else if (size <= small_max) {
2246 			/* Quantum-spaced. */
2247 			size = QUANTUM_CEILING(size);
2248 			bin = &arena->bins[ntbins + (size >> opt_quantum_2pow)
2249 			    - 1];
2250 		} else {
2251 			/* Sub-page. */
2252 			size = pow2_ceil(size);
2253 			bin = &arena->bins[ntbins + nqbins
2254 			    + (ffs((int)(size >> opt_small_max_2pow)) - 2)];
2255 		}
2256 		assert(size == bin->reg_size);
2257 
2258 		malloc_mutex_lock(&arena->mtx);
2259 		if ((run = bin->runcur) != NULL && run->nfree > 0)
2260 			ret = arena_bin_malloc_easy(arena, bin, run);
2261 		else
2262 			ret = arena_bin_malloc_hard(arena, bin);
2263 
2264 		if (ret == NULL) {
2265 			malloc_mutex_unlock(&arena->mtx);
2266 			return (NULL);
2267 		}
2268 
2269 #ifdef MALLOC_STATS
2270 		bin->stats.nrequests++;
2271 		arena->stats.nmalloc_small++;
2272 		arena->stats.allocated_small += size;
2273 #endif
2274 	} else {
2275 		/* Large allocation. */
2276 		size = PAGE_CEILING(size);
2277 		malloc_mutex_lock(&arena->mtx);
2278 		ret = (void *)arena_run_alloc(arena, size);
2279 		if (ret == NULL) {
2280 			malloc_mutex_unlock(&arena->mtx);
2281 			return (NULL);
2282 		}
2283 #ifdef MALLOC_STATS
2284 		arena->stats.nmalloc_large++;
2285 		arena->stats.allocated_large += size;
2286 #endif
2287 	}
2288 
2289 	malloc_mutex_unlock(&arena->mtx);
2290 
2291 	if (opt_junk)
2292 		memset(ret, 0xa5, size);
2293 	else if (opt_zero)
2294 		memset(ret, 0, size);
2295 	return (ret);
2296 }
2297 
2298 static inline void
2299 arena_palloc_trim(arena_t *arena, arena_chunk_t *chunk, unsigned pageind,
2300     unsigned npages)
2301 {
2302 	unsigned i;
2303 
2304 	assert(npages > 0);
2305 
2306 	/*
2307 	 * Modifiy the map such that arena_run_dalloc() sees the run as
2308 	 * separately allocated.
2309 	 */
2310 	for (i = 0; i < npages; i++) {
2311 		chunk->map[pageind + i].npages = npages;
2312 		chunk->map[pageind + i].pos = i;
2313 	}
2314 	arena_run_dalloc(arena, (arena_run_t *)((uintptr_t)chunk + (pageind <<
2315 	    pagesize_2pow)), npages << pagesize_2pow);
2316 }
2317 
2318 /* Only handles large allocations that require more than page alignment. */
2319 static void *
2320 arena_palloc(arena_t *arena, size_t alignment, size_t size, size_t alloc_size)
2321 {
2322 	void *ret;
2323 	size_t offset;
2324 	arena_chunk_t *chunk;
2325 	unsigned pageind, i, npages;
2326 
2327 	assert((size & pagesize_mask) == 0);
2328 	assert((alignment & pagesize_mask) == 0);
2329 
2330 	npages = (unsigned)(size >> pagesize_2pow);
2331 
2332 	malloc_mutex_lock(&arena->mtx);
2333 	ret = (void *)arena_run_alloc(arena, alloc_size);
2334 	if (ret == NULL) {
2335 		malloc_mutex_unlock(&arena->mtx);
2336 		return (NULL);
2337 	}
2338 
2339 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ret);
2340 
2341 	offset = (uintptr_t)ret & (alignment - 1);
2342 	assert((offset & pagesize_mask) == 0);
2343 	assert(offset < alloc_size);
2344 	if (offset == 0) {
2345 		pageind = (unsigned)(((uintptr_t)ret - (uintptr_t)chunk) >>
2346 		    pagesize_2pow);
2347 
2348 		/* Update the map for the run to be kept. */
2349 		for (i = 0; i < npages; i++) {
2350 			chunk->map[pageind + i].npages = npages;
2351 			assert(chunk->map[pageind + i].pos == i);
2352 		}
2353 
2354 		/* Trim trailing space. */
2355 		arena_palloc_trim(arena, chunk, pageind + npages,
2356 		    (unsigned)((alloc_size - size) >> pagesize_2pow));
2357 	} else {
2358 		size_t leadsize, trailsize;
2359 
2360 		leadsize = alignment - offset;
2361 		ret = (void *)((uintptr_t)ret + leadsize);
2362 		pageind = (unsigned)(((uintptr_t)ret - (uintptr_t)chunk) >>
2363 		    pagesize_2pow);
2364 
2365 		/* Update the map for the run to be kept. */
2366 		for (i = 0; i < npages; i++) {
2367 			chunk->map[pageind + i].npages = npages;
2368 			chunk->map[pageind + i].pos = i;
2369 		}
2370 
2371 		/* Trim leading space. */
2372 		arena_palloc_trim(arena, chunk,
2373 		    (unsigned)(pageind - (leadsize >> pagesize_2pow)),
2374 		    (unsigned)(leadsize >> pagesize_2pow));
2375 
2376 		trailsize = alloc_size - leadsize - size;
2377 		if (trailsize != 0) {
2378 			/* Trim trailing space. */
2379 			assert(trailsize < alloc_size);
2380 			arena_palloc_trim(arena, chunk, pageind + npages,
2381 			    (unsigned)(trailsize >> pagesize_2pow));
2382 		}
2383 	}
2384 
2385 #ifdef MALLOC_STATS
2386 	arena->stats.nmalloc_large++;
2387 	arena->stats.allocated_large += size;
2388 #endif
2389 	malloc_mutex_unlock(&arena->mtx);
2390 
2391 	if (opt_junk)
2392 		memset(ret, 0xa5, size);
2393 	else if (opt_zero)
2394 		memset(ret, 0, size);
2395 	return (ret);
2396 }
2397 
2398 /* Return the size of the allocation pointed to by ptr. */
2399 static size_t
2400 arena_salloc(const void *ptr)
2401 {
2402 	size_t ret;
2403 	arena_chunk_t *chunk;
2404 	arena_chunk_map_t *mapelm;
2405 	unsigned pageind;
2406 
2407 	assert(ptr != NULL);
2408 	assert(CHUNK_ADDR2BASE(ptr) != ptr);
2409 
2410 	/*
2411 	 * No arena data structures that we query here can change in a way that
2412 	 * affects this function, so we don't need to lock.
2413 	 */
2414 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
2415 	pageind = (unsigned)(((uintptr_t)ptr - (uintptr_t)chunk) >>
2416 	    pagesize_2pow);
2417 	mapelm = &chunk->map[pageind];
2418 	if (mapelm->pos != 0 || ptr != (char *)((uintptr_t)chunk) + (pageind <<
2419 	    pagesize_2pow)) {
2420 		arena_run_t *run;
2421 
2422 		pageind -= mapelm->pos;
2423 
2424 		run = (arena_run_t *)((uintptr_t)chunk + (pageind <<
2425 		    pagesize_2pow));
2426 		assert(run->magic == ARENA_RUN_MAGIC);
2427 		ret = run->bin->reg_size;
2428 	} else
2429 		ret = mapelm->npages << pagesize_2pow;
2430 
2431 	return (ret);
2432 }
2433 
2434 static void *
2435 arena_ralloc(void *ptr, size_t size, size_t oldsize)
2436 {
2437 	void *ret;
2438 
2439 	/* Avoid moving the allocation if the size class would not change. */
2440 	if (size < small_min) {
2441 		if (oldsize < small_min &&
2442 		    ffs((int)(pow2_ceil(size) >> (TINY_MIN_2POW + 1)))
2443 		    == ffs((int)(pow2_ceil(oldsize) >> (TINY_MIN_2POW + 1))))
2444 			goto IN_PLACE;
2445 	} else if (size <= small_max) {
2446 		if (oldsize >= small_min && oldsize <= small_max &&
2447 		    (QUANTUM_CEILING(size) >> opt_quantum_2pow)
2448 		    == (QUANTUM_CEILING(oldsize) >> opt_quantum_2pow))
2449 			goto IN_PLACE;
2450 	} else {
2451 		/*
2452 		 * We make no attempt to resize runs here, though it would be
2453 		 * possible to do so.
2454 		 */
2455 		if (oldsize > small_max && PAGE_CEILING(size) == oldsize)
2456 			goto IN_PLACE;
2457 	}
2458 
2459 	/*
2460 	 * If we get here, then size and oldsize are different enough that we
2461 	 * need to use a different size class.  In that case, fall back to
2462 	 * allocating new space and copying.
2463 	 */
2464 	ret = arena_malloc(choose_arena(), size);
2465 	if (ret == NULL)
2466 		return (NULL);
2467 
2468 	/* Junk/zero-filling were already done by arena_malloc(). */
2469 	if (size < oldsize)
2470 		memcpy(ret, ptr, size);
2471 	else
2472 		memcpy(ret, ptr, oldsize);
2473 	idalloc(ptr);
2474 	return (ret);
2475 IN_PLACE:
2476 	if (opt_junk && size < oldsize)
2477 		memset((void *)((uintptr_t)ptr + size), 0x5a, oldsize - size);
2478 	else if (opt_zero && size > oldsize)
2479 		memset((void *)((uintptr_t)ptr + oldsize), 0, size - oldsize);
2480 	return (ptr);
2481 }
2482 
2483 static void
2484 arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr)
2485 {
2486 	unsigned pageind;
2487 	arena_chunk_map_t *mapelm;
2488 	size_t size;
2489 
2490 	assert(arena != NULL);
2491 	assert(arena->magic == ARENA_MAGIC);
2492 	assert(chunk->arena == arena);
2493 	assert(ptr != NULL);
2494 	assert(CHUNK_ADDR2BASE(ptr) != ptr);
2495 
2496 	pageind = (unsigned)(((uintptr_t)ptr - (uintptr_t)chunk) >>
2497 	    pagesize_2pow);
2498 	mapelm = &chunk->map[pageind];
2499 	if (mapelm->pos != 0 || ptr != (char *)((uintptr_t)chunk) + (pageind <<
2500 	    pagesize_2pow)) {
2501 		arena_run_t *run;
2502 		arena_bin_t *bin;
2503 
2504 		/* Small allocation. */
2505 
2506 		pageind -= mapelm->pos;
2507 
2508 		run = (arena_run_t *)((uintptr_t)chunk + (pageind <<
2509 		    pagesize_2pow));
2510 		assert(run->magic == ARENA_RUN_MAGIC);
2511 		bin = run->bin;
2512 		size = bin->reg_size;
2513 
2514 		if (opt_junk)
2515 			memset(ptr, 0x5a, size);
2516 
2517 		malloc_mutex_lock(&arena->mtx);
2518 		arena_run_reg_dalloc(run, bin, ptr, size);
2519 		run->nfree++;
2520 
2521 		if (run->nfree == bin->nregs) {
2522 			/* Deallocate run. */
2523 			if (run == bin->runcur)
2524 				bin->runcur = NULL;
2525 			else if (bin->nregs != 1) {
2526 				/*
2527 				 * This block's conditional is necessary because
2528 				 * if the run only contains one region, then it
2529 				 * never gets inserted into the non-full runs
2530 				 * tree.
2531 				 */
2532 				/* LINTED */
2533 				RB_REMOVE(arena_run_tree_s, &bin->runs, run);
2534 			}
2535 #ifdef MALLOC_DEBUG
2536 			run->magic = 0;
2537 #endif
2538 			arena_run_dalloc(arena, run, bin->run_size);
2539 #ifdef MALLOC_STATS
2540 			bin->stats.curruns--;
2541 #endif
2542 		} else if (run->nfree == 1 && run != bin->runcur) {
2543 			/*
2544 			 * Make sure that bin->runcur always refers to the
2545 			 * lowest non-full run, if one exists.
2546 			 */
2547 			if (bin->runcur == NULL)
2548 				bin->runcur = run;
2549 			else if ((uintptr_t)run < (uintptr_t)bin->runcur) {
2550 				/* Switch runcur. */
2551 				if (bin->runcur->nfree > 0) {
2552 					/* Insert runcur. */
2553 					/* LINTED */
2554 					RB_INSERT(arena_run_tree_s, &bin->runs,
2555 					    bin->runcur);
2556 				}
2557 				bin->runcur = run;
2558 			} else {
2559 				/* LINTED */
2560 				RB_INSERT(arena_run_tree_s, &bin->runs, run);
2561 			}
2562 		}
2563 #ifdef MALLOC_STATS
2564 		arena->stats.allocated_small -= size;
2565 		arena->stats.ndalloc_small++;
2566 #endif
2567 	} else {
2568 		/* Large allocation. */
2569 
2570 		size = mapelm->npages << pagesize_2pow;
2571 		assert((((uintptr_t)ptr) & pagesize_mask) == 0);
2572 
2573 		if (opt_junk)
2574 			memset(ptr, 0x5a, size);
2575 
2576 		malloc_mutex_lock(&arena->mtx);
2577 		arena_run_dalloc(arena, (arena_run_t *)ptr, size);
2578 #ifdef MALLOC_STATS
2579 		arena->stats.allocated_large -= size;
2580 		arena->stats.ndalloc_large++;
2581 #endif
2582 	}
2583 
2584 	malloc_mutex_unlock(&arena->mtx);
2585 }
2586 
2587 static bool
2588 arena_new(arena_t *arena)
2589 {
2590 	unsigned i;
2591 	arena_bin_t *bin;
2592 	size_t prev_run_size;
2593 
2594 	malloc_mutex_init(&arena->mtx);
2595 
2596 #ifdef MALLOC_STATS
2597 	memset(&arena->stats, 0, sizeof(arena_stats_t));
2598 #endif
2599 
2600 	/* Initialize chunks. */
2601 	RB_INIT(&arena->chunks);
2602 	arena->spare = NULL;
2603 
2604 	/* Initialize bins. */
2605 	prev_run_size = pagesize;
2606 
2607 	/* (2^n)-spaced tiny bins. */
2608 	for (i = 0; i < ntbins; i++) {
2609 		bin = &arena->bins[i];
2610 		bin->runcur = NULL;
2611 		RB_INIT(&bin->runs);
2612 
2613 		bin->reg_size = (1 << (TINY_MIN_2POW + i));
2614 		prev_run_size = arena_bin_run_size_calc(bin, prev_run_size);
2615 
2616 #ifdef MALLOC_STATS
2617 		memset(&bin->stats, 0, sizeof(malloc_bin_stats_t));
2618 #endif
2619 	}
2620 
2621 	/* Quantum-spaced bins. */
2622 	for (; i < ntbins + nqbins; i++) {
2623 		bin = &arena->bins[i];
2624 		bin->runcur = NULL;
2625 		RB_INIT(&bin->runs);
2626 
2627 		bin->reg_size = quantum * (i - ntbins + 1);
2628 /*
2629 		pow2_size = pow2_ceil(quantum * (i - ntbins + 1));
2630 */
2631 		prev_run_size = arena_bin_run_size_calc(bin, prev_run_size);
2632 
2633 #ifdef MALLOC_STATS
2634 		memset(&bin->stats, 0, sizeof(malloc_bin_stats_t));
2635 #endif
2636 	}
2637 
2638 	/* (2^n)-spaced sub-page bins. */
2639 	for (; i < ntbins + nqbins + nsbins; i++) {
2640 		bin = &arena->bins[i];
2641 		bin->runcur = NULL;
2642 		RB_INIT(&bin->runs);
2643 
2644 		bin->reg_size = (small_max << (i - (ntbins + nqbins) + 1));
2645 
2646 		prev_run_size = arena_bin_run_size_calc(bin, prev_run_size);
2647 
2648 #ifdef MALLOC_STATS
2649 		memset(&bin->stats, 0, sizeof(malloc_bin_stats_t));
2650 #endif
2651 	}
2652 
2653 #ifdef MALLOC_DEBUG
2654 	arena->magic = ARENA_MAGIC;
2655 #endif
2656 
2657 	return (false);
2658 }
2659 
2660 /* Create a new arena and insert it into the arenas array at index ind. */
2661 static arena_t *
2662 arenas_extend(unsigned ind)
2663 {
2664 	arena_t *ret;
2665 
2666 	/* Allocate enough space for trailing bins. */
2667 	ret = (arena_t *)base_alloc(sizeof(arena_t)
2668 	    + (sizeof(arena_bin_t) * (ntbins + nqbins + nsbins - 1)));
2669 	if (ret != NULL && arena_new(ret) == false) {
2670 		arenas[ind] = ret;
2671 		return (ret);
2672 	}
2673 	/* Only reached if there is an OOM error. */
2674 
2675 	/*
2676 	 * OOM here is quite inconvenient to propagate, since dealing with it
2677 	 * would require a check for failure in the fast path.  Instead, punt
2678 	 * by using arenas[0].  In practice, this is an extremely unlikely
2679 	 * failure.
2680 	 */
2681 	_malloc_message(getprogname(),
2682 	    ": (malloc) Error initializing arena\n", "", "");
2683 	if (opt_abort)
2684 		abort();
2685 
2686 	return (arenas[0]);
2687 }
2688 
2689 /*
2690  * End arena.
2691  */
2692 /******************************************************************************/
2693 /*
2694  * Begin general internal functions.
2695  */
2696 
2697 static void *
2698 huge_malloc(size_t size)
2699 {
2700 	void *ret;
2701 	size_t csize;
2702 	chunk_node_t *node;
2703 
2704 	/* Allocate one or more contiguous chunks for this request. */
2705 
2706 	csize = CHUNK_CEILING(size);
2707 	if (csize == 0) {
2708 		/* size is large enough to cause size_t wrap-around. */
2709 		return (NULL);
2710 	}
2711 
2712 	/* Allocate a chunk node with which to track the chunk. */
2713 	node = base_chunk_node_alloc();
2714 	if (node == NULL)
2715 		return (NULL);
2716 
2717 	ret = chunk_alloc(csize);
2718 	if (ret == NULL) {
2719 		base_chunk_node_dealloc(node);
2720 		return (NULL);
2721 	}
2722 
2723 	/* Insert node into huge. */
2724 	node->chunk = ret;
2725 	node->size = csize;
2726 
2727 	malloc_mutex_lock(&chunks_mtx);
2728 	RB_INSERT(chunk_tree_s, &huge, node);
2729 #ifdef MALLOC_STATS
2730 	huge_nmalloc++;
2731 	huge_allocated += csize;
2732 #endif
2733 	malloc_mutex_unlock(&chunks_mtx);
2734 
2735 	if (opt_junk)
2736 		memset(ret, 0xa5, csize);
2737 	else if (opt_zero)
2738 		memset(ret, 0, csize);
2739 
2740 	return (ret);
2741 }
2742 
2743 /* Only handles large allocations that require more than chunk alignment. */
2744 static void *
2745 huge_palloc(size_t alignment, size_t size)
2746 {
2747 	void *ret;
2748 	size_t alloc_size, chunk_size, offset;
2749 	chunk_node_t *node;
2750 
2751 	/*
2752 	 * This allocation requires alignment that is even larger than chunk
2753 	 * alignment.  This means that huge_malloc() isn't good enough.
2754 	 *
2755 	 * Allocate almost twice as many chunks as are demanded by the size or
2756 	 * alignment, in order to assure the alignment can be achieved, then
2757 	 * unmap leading and trailing chunks.
2758 	 */
2759 	assert(alignment >= chunksize);
2760 
2761 	chunk_size = CHUNK_CEILING(size);
2762 
2763 	if (size >= alignment)
2764 		alloc_size = chunk_size + alignment - chunksize;
2765 	else
2766 		alloc_size = (alignment << 1) - chunksize;
2767 
2768 	/* Allocate a chunk node with which to track the chunk. */
2769 	node = base_chunk_node_alloc();
2770 	if (node == NULL)
2771 		return (NULL);
2772 
2773 	ret = chunk_alloc(alloc_size);
2774 	if (ret == NULL) {
2775 		base_chunk_node_dealloc(node);
2776 		return (NULL);
2777 	}
2778 
2779 	offset = (uintptr_t)ret & (alignment - 1);
2780 	assert((offset & chunksize_mask) == 0);
2781 	assert(offset < alloc_size);
2782 	if (offset == 0) {
2783 		/* Trim trailing space. */
2784 		chunk_dealloc((void *)((uintptr_t)ret + chunk_size), alloc_size
2785 		    - chunk_size);
2786 	} else {
2787 		size_t trailsize;
2788 
2789 		/* Trim leading space. */
2790 		chunk_dealloc(ret, alignment - offset);
2791 
2792 		ret = (void *)((uintptr_t)ret + (alignment - offset));
2793 
2794 		trailsize = alloc_size - (alignment - offset) - chunk_size;
2795 		if (trailsize != 0) {
2796 		    /* Trim trailing space. */
2797 		    assert(trailsize < alloc_size);
2798 		    chunk_dealloc((void *)((uintptr_t)ret + chunk_size),
2799 			trailsize);
2800 		}
2801 	}
2802 
2803 	/* Insert node into huge. */
2804 	node->chunk = ret;
2805 	node->size = chunk_size;
2806 
2807 	malloc_mutex_lock(&chunks_mtx);
2808 	RB_INSERT(chunk_tree_s, &huge, node);
2809 #ifdef MALLOC_STATS
2810 	huge_nmalloc++;
2811 	huge_allocated += chunk_size;
2812 #endif
2813 	malloc_mutex_unlock(&chunks_mtx);
2814 
2815 	if (opt_junk)
2816 		memset(ret, 0xa5, chunk_size);
2817 	else if (opt_zero)
2818 		memset(ret, 0, chunk_size);
2819 
2820 	return (ret);
2821 }
2822 
2823 static void *
2824 huge_ralloc(void *ptr, size_t size, size_t oldsize)
2825 {
2826 	void *ret;
2827 
2828 	/* Avoid moving the allocation if the size class would not change. */
2829 	if (oldsize > arena_maxclass &&
2830 	    CHUNK_CEILING(size) == CHUNK_CEILING(oldsize)) {
2831 		if (opt_junk && size < oldsize) {
2832 			memset((void *)((uintptr_t)ptr + size), 0x5a, oldsize
2833 			    - size);
2834 		} else if (opt_zero && size > oldsize) {
2835 			memset((void *)((uintptr_t)ptr + oldsize), 0, size
2836 			    - oldsize);
2837 		}
2838 		return (ptr);
2839 	}
2840 
2841 	if (CHUNK_ADDR2BASE(ptr) == ptr
2842 #ifdef USE_BRK
2843 	    && ((uintptr_t)ptr < (uintptr_t)brk_base
2844 	    || (uintptr_t)ptr >= (uintptr_t)brk_max)
2845 #endif
2846 	    ) {
2847 		chunk_node_t *node, key;
2848 		void *newptr;
2849 		size_t oldcsize;
2850 		size_t newcsize;
2851 
2852 		newcsize = CHUNK_CEILING(size);
2853 		oldcsize = CHUNK_CEILING(oldsize);
2854 		assert(oldcsize != newcsize);
2855 		if (newcsize == 0) {
2856 			/* size_t wrap-around */
2857 			return (NULL);
2858 		}
2859 
2860 		/*
2861 		 * Remove the old region from the tree now.  If mremap()
2862 		 * returns the region to the system, other thread may
2863 		 * map it for same huge allocation and insert it to the
2864 		 * tree before we acquire the mutex lock again.
2865 		 */
2866 		malloc_mutex_lock(&chunks_mtx);
2867 		key.chunk = __DECONST(void *, ptr);
2868 		/* LINTED */
2869 		node = RB_FIND(chunk_tree_s, &huge, &key);
2870 		assert(node != NULL);
2871 		assert(node->chunk == ptr);
2872 		assert(node->size == oldcsize);
2873 		RB_REMOVE(chunk_tree_s, &huge, node);
2874 		malloc_mutex_unlock(&chunks_mtx);
2875 
2876 		newptr = mremap(ptr, oldcsize, NULL, newcsize,
2877 		    MAP_ALIGNED(chunksize_2pow));
2878 		if (newptr == MAP_FAILED) {
2879 			/* We still own the old region. */
2880 			malloc_mutex_lock(&chunks_mtx);
2881 			RB_INSERT(chunk_tree_s, &huge, node);
2882 			malloc_mutex_unlock(&chunks_mtx);
2883 		} else {
2884 			assert(CHUNK_ADDR2BASE(newptr) == newptr);
2885 
2886 			/* Insert new or resized old region. */
2887 			malloc_mutex_lock(&chunks_mtx);
2888 			node->size = newcsize;
2889 			node->chunk = newptr;
2890 			RB_INSERT(chunk_tree_s, &huge, node);
2891 #ifdef MALLOC_STATS
2892 			huge_nralloc++;
2893 			huge_allocated += newcsize - oldcsize;
2894 			if (newcsize > oldcsize) {
2895 				stats_chunks.curchunks +=
2896 				    (newcsize - oldcsize) / chunksize;
2897 				if (stats_chunks.curchunks >
2898 				    stats_chunks.highchunks)
2899 					stats_chunks.highchunks =
2900 					    stats_chunks.curchunks;
2901 			} else {
2902 				stats_chunks.curchunks -=
2903 				    (oldcsize - newcsize) / chunksize;
2904 			}
2905 #endif
2906 			malloc_mutex_unlock(&chunks_mtx);
2907 
2908 			if (opt_junk && size < oldsize) {
2909 				memset((void *)((uintptr_t)newptr + size), 0x5a,
2910 				    newcsize - size);
2911 			} else if (opt_zero && size > oldsize) {
2912 				memset((void *)((uintptr_t)newptr + oldsize), 0,
2913 				    size - oldsize);
2914 			}
2915 			return (newptr);
2916 		}
2917 	}
2918 
2919 	/*
2920 	 * If we get here, then size and oldsize are different enough that we
2921 	 * need to use a different size class.  In that case, fall back to
2922 	 * allocating new space and copying.
2923 	 */
2924 	ret = huge_malloc(size);
2925 	if (ret == NULL)
2926 		return (NULL);
2927 
2928 	if (CHUNK_ADDR2BASE(ptr) == ptr) {
2929 		/* The old allocation is a chunk. */
2930 		if (size < oldsize)
2931 			memcpy(ret, ptr, size);
2932 		else
2933 			memcpy(ret, ptr, oldsize);
2934 	} else {
2935 		/* The old allocation is a region. */
2936 		assert(oldsize < size);
2937 		memcpy(ret, ptr, oldsize);
2938 	}
2939 	idalloc(ptr);
2940 	return (ret);
2941 }
2942 
2943 static void
2944 huge_dalloc(void *ptr)
2945 {
2946 	chunk_node_t key;
2947 	chunk_node_t *node;
2948 
2949 	malloc_mutex_lock(&chunks_mtx);
2950 
2951 	/* Extract from tree of huge allocations. */
2952 	key.chunk = ptr;
2953 	/* LINTED */
2954 	node = RB_FIND(chunk_tree_s, &huge, &key);
2955 	assert(node != NULL);
2956 	assert(node->chunk == ptr);
2957 	/* LINTED */
2958 	RB_REMOVE(chunk_tree_s, &huge, node);
2959 
2960 #ifdef MALLOC_STATS
2961 	huge_ndalloc++;
2962 	huge_allocated -= node->size;
2963 #endif
2964 
2965 	malloc_mutex_unlock(&chunks_mtx);
2966 
2967 	/* Unmap chunk. */
2968 #ifdef USE_BRK
2969 	if (opt_junk)
2970 		memset(node->chunk, 0x5a, node->size);
2971 #endif
2972 	chunk_dealloc(node->chunk, node->size);
2973 
2974 	base_chunk_node_dealloc(node);
2975 }
2976 
2977 static void *
2978 imalloc(size_t size)
2979 {
2980 	void *ret;
2981 
2982 	assert(size != 0);
2983 
2984 	if (size <= arena_maxclass)
2985 		ret = arena_malloc(choose_arena(), size);
2986 	else
2987 		ret = huge_malloc(size);
2988 
2989 	return (ret);
2990 }
2991 
2992 static void *
2993 ipalloc(size_t alignment, size_t size)
2994 {
2995 	void *ret;
2996 	size_t ceil_size;
2997 
2998 	/*
2999 	 * Round size up to the nearest multiple of alignment.
3000 	 *
3001 	 * This done, we can take advantage of the fact that for each small
3002 	 * size class, every object is aligned at the smallest power of two
3003 	 * that is non-zero in the base two representation of the size.  For
3004 	 * example:
3005 	 *
3006 	 *   Size |   Base 2 | Minimum alignment
3007 	 *   -----+----------+------------------
3008 	 *     96 |  1100000 |  32
3009 	 *    144 | 10100000 |  32
3010 	 *    192 | 11000000 |  64
3011 	 *
3012 	 * Depending on runtime settings, it is possible that arena_malloc()
3013 	 * will further round up to a power of two, but that never causes
3014 	 * correctness issues.
3015 	 */
3016 	ceil_size = (size + (alignment - 1)) & (-alignment);
3017 	/*
3018 	 * (ceil_size < size) protects against the combination of maximal
3019 	 * alignment and size greater than maximal alignment.
3020 	 */
3021 	if (ceil_size < size) {
3022 		/* size_t overflow. */
3023 		return (NULL);
3024 	}
3025 
3026 	if (ceil_size <= pagesize || (alignment <= pagesize
3027 	    && ceil_size <= arena_maxclass))
3028 		ret = arena_malloc(choose_arena(), ceil_size);
3029 	else {
3030 		size_t run_size;
3031 
3032 		/*
3033 		 * We can't achieve sub-page alignment, so round up alignment
3034 		 * permanently; it makes later calculations simpler.
3035 		 */
3036 		alignment = PAGE_CEILING(alignment);
3037 		ceil_size = PAGE_CEILING(size);
3038 		/*
3039 		 * (ceil_size < size) protects against very large sizes within
3040 		 * pagesize of SIZE_T_MAX.
3041 		 *
3042 		 * (ceil_size + alignment < ceil_size) protects against the
3043 		 * combination of maximal alignment and ceil_size large enough
3044 		 * to cause overflow.  This is similar to the first overflow
3045 		 * check above, but it needs to be repeated due to the new
3046 		 * ceil_size value, which may now be *equal* to maximal
3047 		 * alignment, whereas before we only detected overflow if the
3048 		 * original size was *greater* than maximal alignment.
3049 		 */
3050 		if (ceil_size < size || ceil_size + alignment < ceil_size) {
3051 			/* size_t overflow. */
3052 			return (NULL);
3053 		}
3054 
3055 		/*
3056 		 * Calculate the size of the over-size run that arena_palloc()
3057 		 * would need to allocate in order to guarantee the alignment.
3058 		 */
3059 		if (ceil_size >= alignment)
3060 			run_size = ceil_size + alignment - pagesize;
3061 		else {
3062 			/*
3063 			 * It is possible that (alignment << 1) will cause
3064 			 * overflow, but it doesn't matter because we also
3065 			 * subtract pagesize, which in the case of overflow
3066 			 * leaves us with a very large run_size.  That causes
3067 			 * the first conditional below to fail, which means
3068 			 * that the bogus run_size value never gets used for
3069 			 * anything important.
3070 			 */
3071 			run_size = (alignment << 1) - pagesize;
3072 		}
3073 
3074 		if (run_size <= arena_maxclass) {
3075 			ret = arena_palloc(choose_arena(), alignment, ceil_size,
3076 			    run_size);
3077 		} else if (alignment <= chunksize)
3078 			ret = huge_malloc(ceil_size);
3079 		else
3080 			ret = huge_palloc(alignment, ceil_size);
3081 	}
3082 
3083 	assert(((uintptr_t)ret & (alignment - 1)) == 0);
3084 	return (ret);
3085 }
3086 
3087 static void *
3088 icalloc(size_t size)
3089 {
3090 	void *ret;
3091 
3092 	if (size <= arena_maxclass) {
3093 		ret = arena_malloc(choose_arena(), size);
3094 		if (ret == NULL)
3095 			return (NULL);
3096 		memset(ret, 0, size);
3097 	} else {
3098 		/*
3099 		 * The virtual memory system provides zero-filled pages, so
3100 		 * there is no need to do so manually, unless opt_junk is
3101 		 * enabled, in which case huge_malloc() fills huge allocations
3102 		 * with junk.
3103 		 */
3104 		ret = huge_malloc(size);
3105 		if (ret == NULL)
3106 			return (NULL);
3107 
3108 		if (opt_junk)
3109 			memset(ret, 0, size);
3110 #ifdef USE_BRK
3111 		else if ((uintptr_t)ret >= (uintptr_t)brk_base
3112 		    && (uintptr_t)ret < (uintptr_t)brk_max) {
3113 			/*
3114 			 * This may be a re-used brk chunk.  Therefore, zero
3115 			 * the memory.
3116 			 */
3117 			memset(ret, 0, size);
3118 		}
3119 #endif
3120 	}
3121 
3122 	return (ret);
3123 }
3124 
3125 static size_t
3126 isalloc(const void *ptr)
3127 {
3128 	size_t ret;
3129 	arena_chunk_t *chunk;
3130 
3131 	assert(ptr != NULL);
3132 
3133 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
3134 	if (chunk != ptr) {
3135 		/* Region. */
3136 		assert(chunk->arena->magic == ARENA_MAGIC);
3137 
3138 		ret = arena_salloc(ptr);
3139 	} else {
3140 		chunk_node_t *node, key;
3141 
3142 		/* Chunk (huge allocation). */
3143 
3144 		malloc_mutex_lock(&chunks_mtx);
3145 
3146 		/* Extract from tree of huge allocations. */
3147 		key.chunk = __DECONST(void *, ptr);
3148 		/* LINTED */
3149 		node = RB_FIND(chunk_tree_s, &huge, &key);
3150 		assert(node != NULL);
3151 
3152 		ret = node->size;
3153 
3154 		malloc_mutex_unlock(&chunks_mtx);
3155 	}
3156 
3157 	return (ret);
3158 }
3159 
3160 static void *
3161 iralloc(void *ptr, size_t size)
3162 {
3163 	void *ret;
3164 	size_t oldsize;
3165 
3166 	assert(ptr != NULL);
3167 	assert(size != 0);
3168 
3169 	oldsize = isalloc(ptr);
3170 
3171 	if (size <= arena_maxclass)
3172 		ret = arena_ralloc(ptr, size, oldsize);
3173 	else
3174 		ret = huge_ralloc(ptr, size, oldsize);
3175 
3176 	return (ret);
3177 }
3178 
3179 static void
3180 idalloc(void *ptr)
3181 {
3182 	arena_chunk_t *chunk;
3183 
3184 	assert(ptr != NULL);
3185 
3186 	chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
3187 	if (chunk != ptr) {
3188 		/* Region. */
3189 		arena_dalloc(chunk->arena, chunk, ptr);
3190 	} else
3191 		huge_dalloc(ptr);
3192 }
3193 
3194 static void
3195 malloc_print_stats(void)
3196 {
3197 
3198 	if (opt_print_stats) {
3199 		char s[UMAX2S_BUFSIZE];
3200 		_malloc_message("___ Begin malloc statistics ___\n", "", "",
3201 		    "");
3202 		_malloc_message("Assertions ",
3203 #ifdef NDEBUG
3204 		    "disabled",
3205 #else
3206 		    "enabled",
3207 #endif
3208 		    "\n", "");
3209 		_malloc_message("Boolean MALLOC_OPTIONS: ",
3210 		    opt_abort ? "A" : "a",
3211 		    opt_junk ? "J" : "j",
3212 		    opt_hint ? "H" : "h");
3213 		_malloc_message(opt_utrace ? "PU" : "Pu",
3214 		    opt_sysv ? "V" : "v",
3215 		    opt_xmalloc ? "X" : "x",
3216 		    opt_zero ? "Z\n" : "z\n");
3217 
3218 		_malloc_message("CPUs: ", umax2s(ncpus, s), "\n", "");
3219 		_malloc_message("Max arenas: ", umax2s(narenas, s), "\n", "");
3220 		_malloc_message("Pointer size: ", umax2s(sizeof(void *), s),
3221 		    "\n", "");
3222 		_malloc_message("Quantum size: ", umax2s(quantum, s), "\n", "");
3223 		_malloc_message("Max small size: ", umax2s(small_max, s), "\n",
3224 		    "");
3225 
3226 		_malloc_message("Chunk size: ", umax2s(chunksize, s), "", "");
3227 		_malloc_message(" (2^", umax2s(opt_chunk_2pow, s), ")\n", "");
3228 
3229 #ifdef MALLOC_STATS
3230 		{
3231 			size_t allocated, mapped;
3232 			unsigned i;
3233 			arena_t *arena;
3234 
3235 			/* Calculate and print allocated/mapped stats. */
3236 
3237 			/* arenas. */
3238 			for (i = 0, allocated = 0; i < narenas; i++) {
3239 				if (arenas[i] != NULL) {
3240 					malloc_mutex_lock(&arenas[i]->mtx);
3241 					allocated +=
3242 					    arenas[i]->stats.allocated_small;
3243 					allocated +=
3244 					    arenas[i]->stats.allocated_large;
3245 					malloc_mutex_unlock(&arenas[i]->mtx);
3246 				}
3247 			}
3248 
3249 			/* huge/base. */
3250 			malloc_mutex_lock(&chunks_mtx);
3251 			allocated += huge_allocated;
3252 			mapped = stats_chunks.curchunks * chunksize;
3253 			malloc_mutex_unlock(&chunks_mtx);
3254 
3255 			malloc_mutex_lock(&base_mtx);
3256 			mapped += base_mapped;
3257 			malloc_mutex_unlock(&base_mtx);
3258 
3259 			malloc_printf("Allocated: %zu, mapped: %zu\n",
3260 			    allocated, mapped);
3261 
3262 			/* Print chunk stats. */
3263 			{
3264 				chunk_stats_t chunks_stats;
3265 
3266 				malloc_mutex_lock(&chunks_mtx);
3267 				chunks_stats = stats_chunks;
3268 				malloc_mutex_unlock(&chunks_mtx);
3269 
3270 				malloc_printf("chunks: nchunks   "
3271 				    "highchunks    curchunks\n");
3272 				malloc_printf("  %13llu%13lu%13lu\n",
3273 				    chunks_stats.nchunks,
3274 				    chunks_stats.highchunks,
3275 				    chunks_stats.curchunks);
3276 			}
3277 
3278 			/* Print chunk stats. */
3279 			malloc_printf(
3280 			    "huge: nmalloc      ndalloc      "
3281 			    "nralloc    allocated\n");
3282 			malloc_printf(" %12llu %12llu %12llu %12zu\n",
3283 			    huge_nmalloc, huge_ndalloc, huge_nralloc,
3284 			    huge_allocated);
3285 
3286 			/* Print stats for each arena. */
3287 			for (i = 0; i < narenas; i++) {
3288 				arena = arenas[i];
3289 				if (arena != NULL) {
3290 					malloc_printf(
3291 					    "\narenas[%u] @ %p\n", i, arena);
3292 					malloc_mutex_lock(&arena->mtx);
3293 					stats_print(arena);
3294 					malloc_mutex_unlock(&arena->mtx);
3295 				}
3296 			}
3297 		}
3298 #endif /* #ifdef MALLOC_STATS */
3299 		_malloc_message("--- End malloc statistics ---\n", "", "", "");
3300 	}
3301 }
3302 
3303 /*
3304  * FreeBSD's pthreads implementation calls malloc(3), so the malloc
3305  * implementation has to take pains to avoid infinite recursion during
3306  * initialization.
3307  */
3308 static inline bool
3309 malloc_init(void)
3310 {
3311 
3312 	if (malloc_initialized == false)
3313 		return (malloc_init_hard());
3314 
3315 	return (false);
3316 }
3317 
3318 static bool
3319 malloc_init_hard(void)
3320 {
3321 	unsigned i, j;
3322 	ssize_t linklen;
3323 	char buf[PATH_MAX + 1];
3324 	const char *opts = "";
3325 
3326 	malloc_mutex_lock(&init_lock);
3327 	if (malloc_initialized) {
3328 		/*
3329 		 * Another thread initialized the allocator before this one
3330 		 * acquired init_lock.
3331 		 */
3332 		malloc_mutex_unlock(&init_lock);
3333 		return (false);
3334 	}
3335 
3336 	/* Get number of CPUs. */
3337 	{
3338 		int mib[2];
3339 		size_t len;
3340 
3341 		mib[0] = CTL_HW;
3342 		mib[1] = HW_NCPU;
3343 		len = sizeof(ncpus);
3344 		if (sysctl(mib, 2, &ncpus, &len, (void *) 0, 0) == -1) {
3345 			/* Error. */
3346 			ncpus = 1;
3347 		}
3348 	}
3349 
3350 	/* Get page size. */
3351 	{
3352 		long result;
3353 
3354 		result = sysconf(_SC_PAGESIZE);
3355 		assert(result != -1);
3356 		pagesize = (unsigned) result;
3357 
3358 		/*
3359 		 * We assume that pagesize is a power of 2 when calculating
3360 		 * pagesize_mask and pagesize_2pow.
3361 		 */
3362 		assert(((result - 1) & result) == 0);
3363 		pagesize_mask = result - 1;
3364 		pagesize_2pow = ffs((int)result) - 1;
3365 	}
3366 
3367 	for (i = 0; i < 3; i++) {
3368 		/* Get runtime configuration. */
3369 		switch (i) {
3370 		case 0:
3371 			if ((linklen = readlink("/etc/malloc.conf", buf,
3372 						sizeof(buf) - 1)) != -1) {
3373 				/*
3374 				 * Use the contents of the "/etc/malloc.conf"
3375 				 * symbolic link's name.
3376 				 */
3377 				buf[linklen] = '\0';
3378 				opts = buf;
3379 			} else {
3380 				/* No configuration specified. */
3381 				buf[0] = '\0';
3382 				opts = buf;
3383 			}
3384 			break;
3385 		case 1:
3386 			if ((opts = getenv("MALLOC_OPTIONS")) != NULL &&
3387 			    issetugid() == 0) {
3388 				/*
3389 				 * Do nothing; opts is already initialized to
3390 				 * the value of the MALLOC_OPTIONS environment
3391 				 * variable.
3392 				 */
3393 			} else {
3394 				/* No configuration specified. */
3395 				buf[0] = '\0';
3396 				opts = buf;
3397 			}
3398 			break;
3399 		case 2:
3400 			if (_malloc_options != NULL) {
3401 			    /*
3402 			     * Use options that were compiled into the program.
3403 			     */
3404 			    opts = _malloc_options;
3405 			} else {
3406 				/* No configuration specified. */
3407 				buf[0] = '\0';
3408 				opts = buf;
3409 			}
3410 			break;
3411 		default:
3412 			/* NOTREACHED */
3413 			/* LINTED */
3414 			assert(false);
3415 		}
3416 
3417 		for (j = 0; opts[j] != '\0'; j++) {
3418 			switch (opts[j]) {
3419 			case 'a':
3420 				opt_abort = false;
3421 				break;
3422 			case 'A':
3423 				opt_abort = true;
3424 				break;
3425 			case 'h':
3426 				opt_hint = false;
3427 				break;
3428 			case 'H':
3429 				opt_hint = true;
3430 				break;
3431 			case 'j':
3432 				opt_junk = false;
3433 				break;
3434 			case 'J':
3435 				opt_junk = true;
3436 				break;
3437 			case 'k':
3438 				/*
3439 				 * Chunks always require at least one header
3440 				 * page, so chunks can never be smaller than
3441 				 * two pages.
3442 				 */
3443 				if (opt_chunk_2pow > pagesize_2pow + 1)
3444 					opt_chunk_2pow--;
3445 				break;
3446 			case 'K':
3447 				if (opt_chunk_2pow + 1 <
3448 				    (int)(sizeof(size_t) << 3))
3449 					opt_chunk_2pow++;
3450 				break;
3451 			case 'n':
3452 				opt_narenas_lshift--;
3453 				break;
3454 			case 'N':
3455 				opt_narenas_lshift++;
3456 				break;
3457 			case 'p':
3458 				opt_print_stats = false;
3459 				break;
3460 			case 'P':
3461 				opt_print_stats = true;
3462 				break;
3463 			case 'q':
3464 				if (opt_quantum_2pow > QUANTUM_2POW_MIN)
3465 					opt_quantum_2pow--;
3466 				break;
3467 			case 'Q':
3468 				if (opt_quantum_2pow < pagesize_2pow - 1)
3469 					opt_quantum_2pow++;
3470 				break;
3471 			case 's':
3472 				if (opt_small_max_2pow > QUANTUM_2POW_MIN)
3473 					opt_small_max_2pow--;
3474 				break;
3475 			case 'S':
3476 				if (opt_small_max_2pow < pagesize_2pow - 1)
3477 					opt_small_max_2pow++;
3478 				break;
3479 			case 'u':
3480 				opt_utrace = false;
3481 				break;
3482 			case 'U':
3483 				opt_utrace = true;
3484 				break;
3485 			case 'v':
3486 				opt_sysv = false;
3487 				break;
3488 			case 'V':
3489 				opt_sysv = true;
3490 				break;
3491 			case 'x':
3492 				opt_xmalloc = false;
3493 				break;
3494 			case 'X':
3495 				opt_xmalloc = true;
3496 				break;
3497 			case 'z':
3498 				opt_zero = false;
3499 				break;
3500 			case 'Z':
3501 				opt_zero = true;
3502 				break;
3503 			default: {
3504 				char cbuf[2];
3505 
3506 				cbuf[0] = opts[j];
3507 				cbuf[1] = '\0';
3508 				_malloc_message(getprogname(),
3509 				    ": (malloc) Unsupported character in "
3510 				    "malloc options: '", cbuf, "'\n");
3511 			}
3512 			}
3513 		}
3514 	}
3515 
3516 	/* Take care to call atexit() only once. */
3517 	if (opt_print_stats) {
3518 		/* Print statistics at exit. */
3519 		atexit(malloc_print_stats);
3520 	}
3521 
3522 	/* Set variables according to the value of opt_small_max_2pow. */
3523 	if (opt_small_max_2pow < opt_quantum_2pow)
3524 		opt_small_max_2pow = opt_quantum_2pow;
3525 	small_max = (1 << opt_small_max_2pow);
3526 
3527 	/* Set bin-related variables. */
3528 	bin_maxclass = (pagesize >> 1);
3529 	assert(opt_quantum_2pow >= TINY_MIN_2POW);
3530 	ntbins = (unsigned)(opt_quantum_2pow - TINY_MIN_2POW);
3531 	assert(ntbins <= opt_quantum_2pow);
3532 	nqbins = (unsigned)(small_max >> opt_quantum_2pow);
3533 	nsbins = (unsigned)(pagesize_2pow - opt_small_max_2pow - 1);
3534 
3535 	/* Set variables according to the value of opt_quantum_2pow. */
3536 	quantum = (1 << opt_quantum_2pow);
3537 	quantum_mask = quantum - 1;
3538 	if (ntbins > 0)
3539 		small_min = (quantum >> 1) + 1;
3540 	else
3541 		small_min = 1;
3542 	assert(small_min <= quantum);
3543 
3544 	/* Set variables according to the value of opt_chunk_2pow. */
3545 	chunksize = (1LU << opt_chunk_2pow);
3546 	chunksize_mask = chunksize - 1;
3547 	chunksize_2pow = (unsigned)opt_chunk_2pow;
3548 	chunk_npages = (unsigned)(chunksize >> pagesize_2pow);
3549 	{
3550 		unsigned header_size;
3551 
3552 		header_size = (unsigned)(sizeof(arena_chunk_t) +
3553 		    (sizeof(arena_chunk_map_t) * (chunk_npages - 1)));
3554 		arena_chunk_header_npages = (header_size >> pagesize_2pow);
3555 		if ((header_size & pagesize_mask) != 0)
3556 			arena_chunk_header_npages++;
3557 	}
3558 	arena_maxclass = chunksize - (arena_chunk_header_npages <<
3559 	    pagesize_2pow);
3560 
3561 	UTRACE(0, 0, 0);
3562 
3563 #ifdef MALLOC_STATS
3564 	memset(&stats_chunks, 0, sizeof(chunk_stats_t));
3565 #endif
3566 
3567 	/* Various sanity checks that regard configuration. */
3568 	assert(quantum >= sizeof(void *));
3569 	assert(quantum <= pagesize);
3570 	assert(chunksize >= pagesize);
3571 	assert(quantum * 4 <= chunksize);
3572 
3573 	/* Initialize chunks data. */
3574 	malloc_mutex_init(&chunks_mtx);
3575 	RB_INIT(&huge);
3576 #ifdef USE_BRK
3577 	malloc_mutex_init(&brk_mtx);
3578 	brk_base = sbrk(0);
3579 	brk_prev = brk_base;
3580 	brk_max = brk_base;
3581 #endif
3582 #ifdef MALLOC_STATS
3583 	huge_nmalloc = 0;
3584 	huge_ndalloc = 0;
3585 	huge_nralloc = 0;
3586 	huge_allocated = 0;
3587 #endif
3588 	RB_INIT(&old_chunks);
3589 
3590 	/* Initialize base allocation data structures. */
3591 #ifdef MALLOC_STATS
3592 	base_mapped = 0;
3593 #endif
3594 #ifdef USE_BRK
3595 	/*
3596 	 * Allocate a base chunk here, since it doesn't actually have to be
3597 	 * chunk-aligned.  Doing this before allocating any other chunks allows
3598 	 * the use of space that would otherwise be wasted.
3599 	 */
3600 	base_pages_alloc(0);
3601 #endif
3602 	base_chunk_nodes = NULL;
3603 	malloc_mutex_init(&base_mtx);
3604 
3605 	if (ncpus > 1) {
3606 		/*
3607 		 * For SMP systems, create four times as many arenas as there
3608 		 * are CPUs by default.
3609 		 */
3610 		opt_narenas_lshift += 2;
3611 	}
3612 
3613 #ifdef NO_TLS
3614 	/* Initialize arena key. */
3615 	(void)thr_keycreate(&arenas_map_key, NULL);
3616 #endif
3617 
3618 	/* Determine how many arenas to use. */
3619 	narenas = ncpus;
3620 	if (opt_narenas_lshift > 0) {
3621 		if ((narenas << opt_narenas_lshift) > narenas)
3622 			narenas <<= opt_narenas_lshift;
3623 		/*
3624 		 * Make sure not to exceed the limits of what base_malloc()
3625 		 * can handle.
3626 		 */
3627 		if (narenas * sizeof(arena_t *) > chunksize)
3628 			narenas = (unsigned)(chunksize / sizeof(arena_t *));
3629 	} else if (opt_narenas_lshift < 0) {
3630 		if ((narenas << opt_narenas_lshift) < narenas)
3631 			narenas <<= opt_narenas_lshift;
3632 		/* Make sure there is at least one arena. */
3633 		if (narenas == 0)
3634 			narenas = 1;
3635 	}
3636 
3637 	next_arena = 0;
3638 
3639 	/* Allocate and initialize arenas. */
3640 	arenas = (arena_t **)base_alloc(sizeof(arena_t *) * narenas);
3641 	if (arenas == NULL) {
3642 		malloc_mutex_unlock(&init_lock);
3643 		return (true);
3644 	}
3645 	/*
3646 	 * Zero the array.  In practice, this should always be pre-zeroed,
3647 	 * since it was just mmap()ed, but let's be sure.
3648 	 */
3649 	memset(arenas, 0, sizeof(arena_t *) * narenas);
3650 
3651 	/*
3652 	 * Initialize one arena here.  The rest are lazily created in
3653 	 * arena_choose_hard().
3654 	 */
3655 	arenas_extend(0);
3656 	if (arenas[0] == NULL) {
3657 		malloc_mutex_unlock(&init_lock);
3658 		return (true);
3659 	}
3660 
3661 	malloc_mutex_init(&arenas_mtx);
3662 
3663 	malloc_initialized = true;
3664 	malloc_mutex_unlock(&init_lock);
3665 	return (false);
3666 }
3667 
3668 /*
3669  * End general internal functions.
3670  */
3671 /******************************************************************************/
3672 /*
3673  * Begin malloc(3)-compatible functions.
3674  */
3675 
3676 void *
3677 malloc(size_t size)
3678 {
3679 	void *ret;
3680 
3681 	if (malloc_init()) {
3682 		ret = NULL;
3683 		goto RETURN;
3684 	}
3685 
3686 	if (size == 0) {
3687 		if (opt_sysv == false)
3688 			size = 1;
3689 		else {
3690 			ret = NULL;
3691 			goto RETURN;
3692 		}
3693 	}
3694 
3695 	ret = imalloc(size);
3696 
3697 RETURN:
3698 	if (ret == NULL) {
3699 		if (opt_xmalloc) {
3700 			_malloc_message(getprogname(),
3701 			    ": (malloc) Error in malloc(): out of memory\n", "",
3702 			    "");
3703 			abort();
3704 		}
3705 		errno = ENOMEM;
3706 	}
3707 
3708 	UTRACE(0, size, ret);
3709 	return (ret);
3710 }
3711 
3712 int
3713 posix_memalign(void **memptr, size_t alignment, size_t size)
3714 {
3715 	int ret;
3716 	void *result;
3717 
3718 	if (malloc_init())
3719 		result = NULL;
3720 	else {
3721 		/* Make sure that alignment is a large enough power of 2. */
3722 		if (((alignment - 1) & alignment) != 0
3723 		    || alignment < sizeof(void *)) {
3724 			if (opt_xmalloc) {
3725 				_malloc_message(getprogname(),
3726 				    ": (malloc) Error in posix_memalign(): "
3727 				    "invalid alignment\n", "", "");
3728 				abort();
3729 			}
3730 			result = NULL;
3731 			ret = EINVAL;
3732 			goto RETURN;
3733 		}
3734 
3735 		result = ipalloc(alignment, size);
3736 	}
3737 
3738 	if (result == NULL) {
3739 		if (opt_xmalloc) {
3740 			_malloc_message(getprogname(),
3741 			": (malloc) Error in posix_memalign(): out of memory\n",
3742 			"", "");
3743 			abort();
3744 		}
3745 		ret = ENOMEM;
3746 		goto RETURN;
3747 	}
3748 
3749 	*memptr = result;
3750 	ret = 0;
3751 
3752 RETURN:
3753 	UTRACE(0, size, result);
3754 	return (ret);
3755 }
3756 
3757 void *
3758 calloc(size_t num, size_t size)
3759 {
3760 	void *ret;
3761 	size_t num_size;
3762 
3763 	if (malloc_init()) {
3764 		num_size = 0;
3765 		ret = NULL;
3766 		goto RETURN;
3767 	}
3768 
3769 	num_size = num * size;
3770 	if (num_size == 0) {
3771 		if ((opt_sysv == false) && ((num == 0) || (size == 0)))
3772 			num_size = 1;
3773 		else {
3774 			ret = NULL;
3775 			goto RETURN;
3776 		}
3777 	/*
3778 	 * Try to avoid division here.  We know that it isn't possible to
3779 	 * overflow during multiplication if neither operand uses any of the
3780 	 * most significant half of the bits in a size_t.
3781 	 */
3782 	} else if ((unsigned long long)((num | size) &
3783 	   ((unsigned long long)SIZE_T_MAX << (sizeof(size_t) << 2))) &&
3784 	   (num_size / size != num)) {
3785 		/* size_t overflow. */
3786 		ret = NULL;
3787 		goto RETURN;
3788 	}
3789 
3790 	ret = icalloc(num_size);
3791 
3792 RETURN:
3793 	if (ret == NULL) {
3794 		if (opt_xmalloc) {
3795 			_malloc_message(getprogname(),
3796 			    ": (malloc) Error in calloc(): out of memory\n", "",
3797 			    "");
3798 			abort();
3799 		}
3800 		errno = ENOMEM;
3801 	}
3802 
3803 	UTRACE(0, num_size, ret);
3804 	return (ret);
3805 }
3806 
3807 void *
3808 realloc(void *ptr, size_t size)
3809 {
3810 	void *ret;
3811 
3812 	if (size == 0) {
3813 		if (opt_sysv == false)
3814 			size = 1;
3815 		else {
3816 			if (ptr != NULL)
3817 				idalloc(ptr);
3818 			ret = NULL;
3819 			goto RETURN;
3820 		}
3821 	}
3822 
3823 	if (ptr != NULL) {
3824 		assert(malloc_initialized);
3825 
3826 		ret = iralloc(ptr, size);
3827 
3828 		if (ret == NULL) {
3829 			if (opt_xmalloc) {
3830 				_malloc_message(getprogname(),
3831 				    ": (malloc) Error in realloc(): out of "
3832 				    "memory\n", "", "");
3833 				abort();
3834 			}
3835 			errno = ENOMEM;
3836 		}
3837 	} else {
3838 		if (malloc_init())
3839 			ret = NULL;
3840 		else
3841 			ret = imalloc(size);
3842 
3843 		if (ret == NULL) {
3844 			if (opt_xmalloc) {
3845 				_malloc_message(getprogname(),
3846 				    ": (malloc) Error in realloc(): out of "
3847 				    "memory\n", "", "");
3848 				abort();
3849 			}
3850 			errno = ENOMEM;
3851 		}
3852 	}
3853 
3854 RETURN:
3855 	UTRACE(ptr, size, ret);
3856 	return (ret);
3857 }
3858 
3859 void
3860 free(void *ptr)
3861 {
3862 
3863 	UTRACE(ptr, 0, 0);
3864 	if (ptr != NULL) {
3865 		assert(malloc_initialized);
3866 
3867 		idalloc(ptr);
3868 	}
3869 }
3870 
3871 /*
3872  * End malloc(3)-compatible functions.
3873  */
3874 /******************************************************************************/
3875 /*
3876  * Begin non-standard functions.
3877  */
3878 #ifndef __NetBSD__
3879 size_t
3880 malloc_usable_size(const void *ptr)
3881 {
3882 
3883 	assert(ptr != NULL);
3884 
3885 	return (isalloc(ptr));
3886 }
3887 #endif
3888 
3889 /*
3890  * End non-standard functions.
3891  */
3892 /******************************************************************************/
3893 /*
3894  * Begin library-private functions, used by threading libraries for protection
3895  * of malloc during fork().  These functions are only called if the program is
3896  * running in threaded mode, so there is no need to check whether the program
3897  * is threaded here.
3898  */
3899 
3900 void
3901 _malloc_prefork(void)
3902 {
3903 	unsigned i;
3904 
3905 	/* Acquire all mutexes in a safe order. */
3906 
3907 	malloc_mutex_lock(&arenas_mtx);
3908 	for (i = 0; i < narenas; i++) {
3909 		if (arenas[i] != NULL)
3910 			malloc_mutex_lock(&arenas[i]->mtx);
3911 	}
3912 	malloc_mutex_unlock(&arenas_mtx);
3913 
3914 	malloc_mutex_lock(&base_mtx);
3915 
3916 	malloc_mutex_lock(&chunks_mtx);
3917 }
3918 
3919 void
3920 _malloc_postfork(void)
3921 {
3922 	unsigned i;
3923 
3924 	/* Release all mutexes, now that fork() has completed. */
3925 
3926 	malloc_mutex_unlock(&chunks_mtx);
3927 
3928 	malloc_mutex_unlock(&base_mtx);
3929 
3930 	malloc_mutex_lock(&arenas_mtx);
3931 	for (i = 0; i < narenas; i++) {
3932 		if (arenas[i] != NULL)
3933 			malloc_mutex_unlock(&arenas[i]->mtx);
3934 	}
3935 	malloc_mutex_unlock(&arenas_mtx);
3936 }
3937 
3938 /*
3939  * End library-private functions.
3940  */
3941 /******************************************************************************/
3942