xref: /freebsd/contrib/jemalloc/src/jemalloc.c (revision 63d1fd5970ec814904aa0f4580b10a0d302d08b2)
1 #define	JEMALLOC_C_
2 #include "jemalloc/internal/jemalloc_internal.h"
3 
4 /******************************************************************************/
5 /* Data. */
6 
7 /* Work around <http://llvm.org/bugs/show_bug.cgi?id=12623>: */
8 const char	*__malloc_options_1_0 = NULL;
9 __sym_compat(_malloc_options, __malloc_options_1_0, FBSD_1.0);
10 
11 /* Runtime configuration options. */
12 const char	*je_malloc_conf
13 #ifndef _WIN32
14     JEMALLOC_ATTR(weak)
15 #endif
16     ;
17 bool	opt_abort =
18 #ifdef JEMALLOC_DEBUG
19     true
20 #else
21     false
22 #endif
23     ;
24 const char	*opt_junk =
25 #if (defined(JEMALLOC_DEBUG) && defined(JEMALLOC_FILL))
26     "true"
27 #else
28     "false"
29 #endif
30     ;
31 bool	opt_junk_alloc =
32 #if (defined(JEMALLOC_DEBUG) && defined(JEMALLOC_FILL))
33     true
34 #else
35     false
36 #endif
37     ;
38 bool	opt_junk_free =
39 #if (defined(JEMALLOC_DEBUG) && defined(JEMALLOC_FILL))
40     true
41 #else
42     false
43 #endif
44     ;
45 
46 size_t	opt_quarantine = ZU(0);
47 bool	opt_redzone = false;
48 bool	opt_utrace = false;
49 bool	opt_xmalloc = false;
50 bool	opt_zero = false;
51 unsigned	opt_narenas = 0;
52 
53 /* Initialized to true if the process is running inside Valgrind. */
54 bool	in_valgrind;
55 
56 unsigned	ncpus;
57 
58 /* Protects arenas initialization. */
59 static malloc_mutex_t	arenas_lock;
60 /*
61  * Arenas that are used to service external requests.  Not all elements of the
62  * arenas array are necessarily used; arenas are created lazily as needed.
63  *
64  * arenas[0..narenas_auto) are used for automatic multiplexing of threads and
65  * arenas.  arenas[narenas_auto..narenas_total) are only used if the application
66  * takes some action to create them and allocate from them.
67  */
68 arena_t			**arenas;
69 static unsigned		narenas_total; /* Use narenas_total_*(). */
70 static arena_t		*a0; /* arenas[0]; read-only after initialization. */
71 unsigned		narenas_auto; /* Read-only after initialization. */
72 
73 typedef enum {
74 	malloc_init_uninitialized	= 3,
75 	malloc_init_a0_initialized	= 2,
76 	malloc_init_recursible		= 1,
77 	malloc_init_initialized		= 0 /* Common case --> jnz. */
78 } malloc_init_t;
79 static malloc_init_t	malloc_init_state = malloc_init_uninitialized;
80 
81 /* False should be the common case.  Set to true to trigger initialization. */
82 static bool	malloc_slow = true;
83 
84 /* When malloc_slow is true, set the corresponding bits for sanity check. */
85 enum {
86 	flag_opt_junk_alloc	= (1U),
87 	flag_opt_junk_free	= (1U << 1),
88 	flag_opt_quarantine	= (1U << 2),
89 	flag_opt_zero		= (1U << 3),
90 	flag_opt_utrace		= (1U << 4),
91 	flag_in_valgrind	= (1U << 5),
92 	flag_opt_xmalloc	= (1U << 6)
93 };
94 static uint8_t	malloc_slow_flags;
95 
96 JEMALLOC_ALIGNED(CACHELINE)
97 const size_t	pind2sz_tab[NPSIZES] = {
98 #define	PSZ_yes(lg_grp, ndelta, lg_delta)				\
99 	(((ZU(1)<<lg_grp) + (ZU(ndelta)<<lg_delta))),
100 #define	PSZ_no(lg_grp, ndelta, lg_delta)
101 #define	SC(index, lg_grp, lg_delta, ndelta, psz, bin, lg_delta_lookup)	\
102 	PSZ_##psz(lg_grp, ndelta, lg_delta)
103 	SIZE_CLASSES
104 #undef PSZ_yes
105 #undef PSZ_no
106 #undef SC
107 };
108 
109 JEMALLOC_ALIGNED(CACHELINE)
110 const size_t	index2size_tab[NSIZES] = {
111 #define	SC(index, lg_grp, lg_delta, ndelta, psz, bin, lg_delta_lookup)	\
112 	((ZU(1)<<lg_grp) + (ZU(ndelta)<<lg_delta)),
113 	SIZE_CLASSES
114 #undef SC
115 };
116 
117 JEMALLOC_ALIGNED(CACHELINE)
118 const uint8_t	size2index_tab[] = {
119 #if LG_TINY_MIN == 0
120 #warning "Dangerous LG_TINY_MIN"
121 #define	S2B_0(i)	i,
122 #elif LG_TINY_MIN == 1
123 #warning "Dangerous LG_TINY_MIN"
124 #define	S2B_1(i)	i,
125 #elif LG_TINY_MIN == 2
126 #warning "Dangerous LG_TINY_MIN"
127 #define	S2B_2(i)	i,
128 #elif LG_TINY_MIN == 3
129 #define	S2B_3(i)	i,
130 #elif LG_TINY_MIN == 4
131 #define	S2B_4(i)	i,
132 #elif LG_TINY_MIN == 5
133 #define	S2B_5(i)	i,
134 #elif LG_TINY_MIN == 6
135 #define	S2B_6(i)	i,
136 #elif LG_TINY_MIN == 7
137 #define	S2B_7(i)	i,
138 #elif LG_TINY_MIN == 8
139 #define	S2B_8(i)	i,
140 #elif LG_TINY_MIN == 9
141 #define	S2B_9(i)	i,
142 #elif LG_TINY_MIN == 10
143 #define	S2B_10(i)	i,
144 #elif LG_TINY_MIN == 11
145 #define	S2B_11(i)	i,
146 #else
147 #error "Unsupported LG_TINY_MIN"
148 #endif
149 #if LG_TINY_MIN < 1
150 #define	S2B_1(i)	S2B_0(i) S2B_0(i)
151 #endif
152 #if LG_TINY_MIN < 2
153 #define	S2B_2(i)	S2B_1(i) S2B_1(i)
154 #endif
155 #if LG_TINY_MIN < 3
156 #define	S2B_3(i)	S2B_2(i) S2B_2(i)
157 #endif
158 #if LG_TINY_MIN < 4
159 #define	S2B_4(i)	S2B_3(i) S2B_3(i)
160 #endif
161 #if LG_TINY_MIN < 5
162 #define	S2B_5(i)	S2B_4(i) S2B_4(i)
163 #endif
164 #if LG_TINY_MIN < 6
165 #define	S2B_6(i)	S2B_5(i) S2B_5(i)
166 #endif
167 #if LG_TINY_MIN < 7
168 #define	S2B_7(i)	S2B_6(i) S2B_6(i)
169 #endif
170 #if LG_TINY_MIN < 8
171 #define	S2B_8(i)	S2B_7(i) S2B_7(i)
172 #endif
173 #if LG_TINY_MIN < 9
174 #define	S2B_9(i)	S2B_8(i) S2B_8(i)
175 #endif
176 #if LG_TINY_MIN < 10
177 #define	S2B_10(i)	S2B_9(i) S2B_9(i)
178 #endif
179 #if LG_TINY_MIN < 11
180 #define	S2B_11(i)	S2B_10(i) S2B_10(i)
181 #endif
182 #define	S2B_no(i)
183 #define	SC(index, lg_grp, lg_delta, ndelta, psz, bin, lg_delta_lookup)	\
184 	S2B_##lg_delta_lookup(index)
185 	SIZE_CLASSES
186 #undef S2B_3
187 #undef S2B_4
188 #undef S2B_5
189 #undef S2B_6
190 #undef S2B_7
191 #undef S2B_8
192 #undef S2B_9
193 #undef S2B_10
194 #undef S2B_11
195 #undef S2B_no
196 #undef SC
197 };
198 
199 #ifdef JEMALLOC_THREADED_INIT
200 /* Used to let the initializing thread recursively allocate. */
201 #  define NO_INITIALIZER	((unsigned long)0)
202 #  define INITIALIZER		pthread_self()
203 #  define IS_INITIALIZER	(malloc_initializer == pthread_self())
204 static pthread_t		malloc_initializer = NO_INITIALIZER;
205 #else
206 #  define NO_INITIALIZER	false
207 #  define INITIALIZER		true
208 #  define IS_INITIALIZER	malloc_initializer
209 static bool			malloc_initializer = NO_INITIALIZER;
210 #endif
211 
212 /* Used to avoid initialization races. */
213 #ifdef _WIN32
214 #if _WIN32_WINNT >= 0x0600
215 static malloc_mutex_t	init_lock = SRWLOCK_INIT;
216 #else
217 static malloc_mutex_t	init_lock;
218 static bool init_lock_initialized = false;
219 
220 JEMALLOC_ATTR(constructor)
221 static void WINAPI
222 _init_init_lock(void)
223 {
224 
225 	/* If another constructor in the same binary is using mallctl to
226 	 * e.g. setup chunk hooks, it may end up running before this one,
227 	 * and malloc_init_hard will crash trying to lock the uninitialized
228 	 * lock. So we force an initialization of the lock in
229 	 * malloc_init_hard as well. We don't try to care about atomicity
230 	 * of the accessed to the init_lock_initialized boolean, since it
231 	 * really only matters early in the process creation, before any
232 	 * separate thread normally starts doing anything. */
233 	if (!init_lock_initialized)
234 		malloc_mutex_init(&init_lock, "init", WITNESS_RANK_INIT);
235 	init_lock_initialized = true;
236 }
237 
238 #ifdef _MSC_VER
239 #  pragma section(".CRT$XCU", read)
240 JEMALLOC_SECTION(".CRT$XCU") JEMALLOC_ATTR(used)
241 static const void (WINAPI *init_init_lock)(void) = _init_init_lock;
242 #endif
243 #endif
244 #else
245 static malloc_mutex_t	init_lock = MALLOC_MUTEX_INITIALIZER;
246 #endif
247 
248 typedef struct {
249 	void	*p;	/* Input pointer (as in realloc(p, s)). */
250 	size_t	s;	/* Request size. */
251 	void	*r;	/* Result pointer. */
252 } malloc_utrace_t;
253 
254 #ifdef JEMALLOC_UTRACE
255 #  define UTRACE(a, b, c) do {						\
256 	if (unlikely(opt_utrace)) {					\
257 		int utrace_serrno = errno;				\
258 		malloc_utrace_t ut;					\
259 		ut.p = (a);						\
260 		ut.s = (b);						\
261 		ut.r = (c);						\
262 		utrace(&ut, sizeof(ut));				\
263 		errno = utrace_serrno;					\
264 	}								\
265 } while (0)
266 #else
267 #  define UTRACE(a, b, c)
268 #endif
269 
270 /******************************************************************************/
271 /*
272  * Function prototypes for static functions that are referenced prior to
273  * definition.
274  */
275 
276 static bool	malloc_init_hard_a0(void);
277 static bool	malloc_init_hard(void);
278 
279 /******************************************************************************/
280 /*
281  * Begin miscellaneous support functions.
282  */
283 
284 JEMALLOC_ALWAYS_INLINE_C bool
285 malloc_initialized(void)
286 {
287 
288 	return (malloc_init_state == malloc_init_initialized);
289 }
290 
291 JEMALLOC_ALWAYS_INLINE_C void
292 malloc_thread_init(void)
293 {
294 
295 	/*
296 	 * TSD initialization can't be safely done as a side effect of
297 	 * deallocation, because it is possible for a thread to do nothing but
298 	 * deallocate its TLS data via free(), in which case writing to TLS
299 	 * would cause write-after-free memory corruption.  The quarantine
300 	 * facility *only* gets used as a side effect of deallocation, so make
301 	 * a best effort attempt at initializing its TSD by hooking all
302 	 * allocation events.
303 	 */
304 	if (config_fill && unlikely(opt_quarantine))
305 		quarantine_alloc_hook();
306 }
307 
308 JEMALLOC_ALWAYS_INLINE_C bool
309 malloc_init_a0(void)
310 {
311 
312 	if (unlikely(malloc_init_state == malloc_init_uninitialized))
313 		return (malloc_init_hard_a0());
314 	return (false);
315 }
316 
317 JEMALLOC_ALWAYS_INLINE_C bool
318 malloc_init(void)
319 {
320 
321 	if (unlikely(!malloc_initialized()) && malloc_init_hard())
322 		return (true);
323 	malloc_thread_init();
324 
325 	return (false);
326 }
327 
328 /*
329  * The a0*() functions are used instead of i{d,}alloc() in situations that
330  * cannot tolerate TLS variable access.
331  */
332 
333 static void *
334 a0ialloc(size_t size, bool zero, bool is_metadata)
335 {
336 
337 	if (unlikely(malloc_init_a0()))
338 		return (NULL);
339 
340 	return (iallocztm(TSDN_NULL, size, size2index(size), zero, NULL,
341 	    is_metadata, arena_get(TSDN_NULL, 0, true), true));
342 }
343 
344 static void
345 a0idalloc(void *ptr, bool is_metadata)
346 {
347 
348 	idalloctm(TSDN_NULL, ptr, false, is_metadata, true);
349 }
350 
351 arena_t *
352 a0get(void)
353 {
354 
355 	return (a0);
356 }
357 
358 void *
359 a0malloc(size_t size)
360 {
361 
362 	return (a0ialloc(size, false, true));
363 }
364 
365 void
366 a0dalloc(void *ptr)
367 {
368 
369 	a0idalloc(ptr, true);
370 }
371 
372 /*
373  * FreeBSD's libc uses the bootstrap_*() functions in bootstrap-senstive
374  * situations that cannot tolerate TLS variable access (TLS allocation and very
375  * early internal data structure initialization).
376  */
377 
378 void *
379 bootstrap_malloc(size_t size)
380 {
381 
382 	if (unlikely(size == 0))
383 		size = 1;
384 
385 	return (a0ialloc(size, false, false));
386 }
387 
388 void *
389 bootstrap_calloc(size_t num, size_t size)
390 {
391 	size_t num_size;
392 
393 	num_size = num * size;
394 	if (unlikely(num_size == 0)) {
395 		assert(num == 0 || size == 0);
396 		num_size = 1;
397 	}
398 
399 	return (a0ialloc(num_size, true, false));
400 }
401 
402 void
403 bootstrap_free(void *ptr)
404 {
405 
406 	if (unlikely(ptr == NULL))
407 		return;
408 
409 	a0idalloc(ptr, false);
410 }
411 
412 static void
413 arena_set(unsigned ind, arena_t *arena)
414 {
415 
416 	atomic_write_p((void **)&arenas[ind], arena);
417 }
418 
419 static void
420 narenas_total_set(unsigned narenas)
421 {
422 
423 	atomic_write_u(&narenas_total, narenas);
424 }
425 
426 static void
427 narenas_total_inc(void)
428 {
429 
430 	atomic_add_u(&narenas_total, 1);
431 }
432 
433 unsigned
434 narenas_total_get(void)
435 {
436 
437 	return (atomic_read_u(&narenas_total));
438 }
439 
440 /* Create a new arena and insert it into the arenas array at index ind. */
441 static arena_t *
442 arena_init_locked(tsdn_t *tsdn, unsigned ind)
443 {
444 	arena_t *arena;
445 
446 	assert(ind <= narenas_total_get());
447 	if (ind > MALLOCX_ARENA_MAX)
448 		return (NULL);
449 	if (ind == narenas_total_get())
450 		narenas_total_inc();
451 
452 	/*
453 	 * Another thread may have already initialized arenas[ind] if it's an
454 	 * auto arena.
455 	 */
456 	arena = arena_get(tsdn, ind, false);
457 	if (arena != NULL) {
458 		assert(ind < narenas_auto);
459 		return (arena);
460 	}
461 
462 	/* Actually initialize the arena. */
463 	arena = arena_new(tsdn, ind);
464 	arena_set(ind, arena);
465 	return (arena);
466 }
467 
468 arena_t *
469 arena_init(tsdn_t *tsdn, unsigned ind)
470 {
471 	arena_t *arena;
472 
473 	malloc_mutex_lock(tsdn, &arenas_lock);
474 	arena = arena_init_locked(tsdn, ind);
475 	malloc_mutex_unlock(tsdn, &arenas_lock);
476 	return (arena);
477 }
478 
479 static void
480 arena_bind(tsd_t *tsd, unsigned ind, bool internal)
481 {
482 	arena_t *arena;
483 
484 	if (!tsd_nominal(tsd))
485 		return;
486 
487 	arena = arena_get(tsd_tsdn(tsd), ind, false);
488 	arena_nthreads_inc(arena, internal);
489 
490 	if (internal)
491 		tsd_iarena_set(tsd, arena);
492 	else
493 		tsd_arena_set(tsd, arena);
494 }
495 
496 void
497 arena_migrate(tsd_t *tsd, unsigned oldind, unsigned newind)
498 {
499 	arena_t *oldarena, *newarena;
500 
501 	oldarena = arena_get(tsd_tsdn(tsd), oldind, false);
502 	newarena = arena_get(tsd_tsdn(tsd), newind, false);
503 	arena_nthreads_dec(oldarena, false);
504 	arena_nthreads_inc(newarena, false);
505 	tsd_arena_set(tsd, newarena);
506 }
507 
508 static void
509 arena_unbind(tsd_t *tsd, unsigned ind, bool internal)
510 {
511 	arena_t *arena;
512 
513 	arena = arena_get(tsd_tsdn(tsd), ind, false);
514 	arena_nthreads_dec(arena, internal);
515 	if (internal)
516 		tsd_iarena_set(tsd, NULL);
517 	else
518 		tsd_arena_set(tsd, NULL);
519 }
520 
521 arena_tdata_t *
522 arena_tdata_get_hard(tsd_t *tsd, unsigned ind)
523 {
524 	arena_tdata_t *tdata, *arenas_tdata_old;
525 	arena_tdata_t *arenas_tdata = tsd_arenas_tdata_get(tsd);
526 	unsigned narenas_tdata_old, i;
527 	unsigned narenas_tdata = tsd_narenas_tdata_get(tsd);
528 	unsigned narenas_actual = narenas_total_get();
529 
530 	/*
531 	 * Dissociate old tdata array (and set up for deallocation upon return)
532 	 * if it's too small.
533 	 */
534 	if (arenas_tdata != NULL && narenas_tdata < narenas_actual) {
535 		arenas_tdata_old = arenas_tdata;
536 		narenas_tdata_old = narenas_tdata;
537 		arenas_tdata = NULL;
538 		narenas_tdata = 0;
539 		tsd_arenas_tdata_set(tsd, arenas_tdata);
540 		tsd_narenas_tdata_set(tsd, narenas_tdata);
541 	} else {
542 		arenas_tdata_old = NULL;
543 		narenas_tdata_old = 0;
544 	}
545 
546 	/* Allocate tdata array if it's missing. */
547 	if (arenas_tdata == NULL) {
548 		bool *arenas_tdata_bypassp = tsd_arenas_tdata_bypassp_get(tsd);
549 		narenas_tdata = (ind < narenas_actual) ? narenas_actual : ind+1;
550 
551 		if (tsd_nominal(tsd) && !*arenas_tdata_bypassp) {
552 			*arenas_tdata_bypassp = true;
553 			arenas_tdata = (arena_tdata_t *)a0malloc(
554 			    sizeof(arena_tdata_t) * narenas_tdata);
555 			*arenas_tdata_bypassp = false;
556 		}
557 		if (arenas_tdata == NULL) {
558 			tdata = NULL;
559 			goto label_return;
560 		}
561 		assert(tsd_nominal(tsd) && !*arenas_tdata_bypassp);
562 		tsd_arenas_tdata_set(tsd, arenas_tdata);
563 		tsd_narenas_tdata_set(tsd, narenas_tdata);
564 	}
565 
566 	/*
567 	 * Copy to tdata array.  It's possible that the actual number of arenas
568 	 * has increased since narenas_total_get() was called above, but that
569 	 * causes no correctness issues unless two threads concurrently execute
570 	 * the arenas.extend mallctl, which we trust mallctl synchronization to
571 	 * prevent.
572 	 */
573 
574 	/* Copy/initialize tickers. */
575 	for (i = 0; i < narenas_actual; i++) {
576 		if (i < narenas_tdata_old) {
577 			ticker_copy(&arenas_tdata[i].decay_ticker,
578 			    &arenas_tdata_old[i].decay_ticker);
579 		} else {
580 			ticker_init(&arenas_tdata[i].decay_ticker,
581 			    DECAY_NTICKS_PER_UPDATE);
582 		}
583 	}
584 	if (narenas_tdata > narenas_actual) {
585 		memset(&arenas_tdata[narenas_actual], 0, sizeof(arena_tdata_t)
586 		    * (narenas_tdata - narenas_actual));
587 	}
588 
589 	/* Read the refreshed tdata array. */
590 	tdata = &arenas_tdata[ind];
591 label_return:
592 	if (arenas_tdata_old != NULL)
593 		a0dalloc(arenas_tdata_old);
594 	return (tdata);
595 }
596 
597 /* Slow path, called only by arena_choose(). */
598 arena_t *
599 arena_choose_hard(tsd_t *tsd, bool internal)
600 {
601 	arena_t *ret JEMALLOC_CC_SILENCE_INIT(NULL);
602 
603 	if (narenas_auto > 1) {
604 		unsigned i, j, choose[2], first_null;
605 
606 		/*
607 		 * Determine binding for both non-internal and internal
608 		 * allocation.
609 		 *
610 		 *   choose[0]: For application allocation.
611 		 *   choose[1]: For internal metadata allocation.
612 		 */
613 
614 		for (j = 0; j < 2; j++)
615 			choose[j] = 0;
616 
617 		first_null = narenas_auto;
618 		malloc_mutex_lock(tsd_tsdn(tsd), &arenas_lock);
619 		assert(arena_get(tsd_tsdn(tsd), 0, false) != NULL);
620 		for (i = 1; i < narenas_auto; i++) {
621 			if (arena_get(tsd_tsdn(tsd), i, false) != NULL) {
622 				/*
623 				 * Choose the first arena that has the lowest
624 				 * number of threads assigned to it.
625 				 */
626 				for (j = 0; j < 2; j++) {
627 					if (arena_nthreads_get(arena_get(
628 					    tsd_tsdn(tsd), i, false), !!j) <
629 					    arena_nthreads_get(arena_get(
630 					    tsd_tsdn(tsd), choose[j], false),
631 					    !!j))
632 						choose[j] = i;
633 				}
634 			} else if (first_null == narenas_auto) {
635 				/*
636 				 * Record the index of the first uninitialized
637 				 * arena, in case all extant arenas are in use.
638 				 *
639 				 * NB: It is possible for there to be
640 				 * discontinuities in terms of initialized
641 				 * versus uninitialized arenas, due to the
642 				 * "thread.arena" mallctl.
643 				 */
644 				first_null = i;
645 			}
646 		}
647 
648 		for (j = 0; j < 2; j++) {
649 			if (arena_nthreads_get(arena_get(tsd_tsdn(tsd),
650 			    choose[j], false), !!j) == 0 || first_null ==
651 			    narenas_auto) {
652 				/*
653 				 * Use an unloaded arena, or the least loaded
654 				 * arena if all arenas are already initialized.
655 				 */
656 				if (!!j == internal) {
657 					ret = arena_get(tsd_tsdn(tsd),
658 					    choose[j], false);
659 				}
660 			} else {
661 				arena_t *arena;
662 
663 				/* Initialize a new arena. */
664 				choose[j] = first_null;
665 				arena = arena_init_locked(tsd_tsdn(tsd),
666 				    choose[j]);
667 				if (arena == NULL) {
668 					malloc_mutex_unlock(tsd_tsdn(tsd),
669 					    &arenas_lock);
670 					return (NULL);
671 				}
672 				if (!!j == internal)
673 					ret = arena;
674 			}
675 			arena_bind(tsd, choose[j], !!j);
676 		}
677 		malloc_mutex_unlock(tsd_tsdn(tsd), &arenas_lock);
678 	} else {
679 		ret = arena_get(tsd_tsdn(tsd), 0, false);
680 		arena_bind(tsd, 0, false);
681 		arena_bind(tsd, 0, true);
682 	}
683 
684 	return (ret);
685 }
686 
687 void
688 thread_allocated_cleanup(tsd_t *tsd)
689 {
690 
691 	/* Do nothing. */
692 }
693 
694 void
695 thread_deallocated_cleanup(tsd_t *tsd)
696 {
697 
698 	/* Do nothing. */
699 }
700 
701 void
702 iarena_cleanup(tsd_t *tsd)
703 {
704 	arena_t *iarena;
705 
706 	iarena = tsd_iarena_get(tsd);
707 	if (iarena != NULL)
708 		arena_unbind(tsd, iarena->ind, true);
709 }
710 
711 void
712 arena_cleanup(tsd_t *tsd)
713 {
714 	arena_t *arena;
715 
716 	arena = tsd_arena_get(tsd);
717 	if (arena != NULL)
718 		arena_unbind(tsd, arena->ind, false);
719 }
720 
721 void
722 arenas_tdata_cleanup(tsd_t *tsd)
723 {
724 	arena_tdata_t *arenas_tdata;
725 
726 	/* Prevent tsd->arenas_tdata from being (re)created. */
727 	*tsd_arenas_tdata_bypassp_get(tsd) = true;
728 
729 	arenas_tdata = tsd_arenas_tdata_get(tsd);
730 	if (arenas_tdata != NULL) {
731 		tsd_arenas_tdata_set(tsd, NULL);
732 		a0dalloc(arenas_tdata);
733 	}
734 }
735 
736 void
737 narenas_tdata_cleanup(tsd_t *tsd)
738 {
739 
740 	/* Do nothing. */
741 }
742 
743 void
744 arenas_tdata_bypass_cleanup(tsd_t *tsd)
745 {
746 
747 	/* Do nothing. */
748 }
749 
750 static void
751 stats_print_atexit(void)
752 {
753 
754 	if (config_tcache && config_stats) {
755 		tsdn_t *tsdn;
756 		unsigned narenas, i;
757 
758 		tsdn = tsdn_fetch();
759 
760 		/*
761 		 * Merge stats from extant threads.  This is racy, since
762 		 * individual threads do not lock when recording tcache stats
763 		 * events.  As a consequence, the final stats may be slightly
764 		 * out of date by the time they are reported, if other threads
765 		 * continue to allocate.
766 		 */
767 		for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
768 			arena_t *arena = arena_get(tsdn, i, false);
769 			if (arena != NULL) {
770 				tcache_t *tcache;
771 
772 				/*
773 				 * tcache_stats_merge() locks bins, so if any
774 				 * code is introduced that acquires both arena
775 				 * and bin locks in the opposite order,
776 				 * deadlocks may result.
777 				 */
778 				malloc_mutex_lock(tsdn, &arena->lock);
779 				ql_foreach(tcache, &arena->tcache_ql, link) {
780 					tcache_stats_merge(tsdn, tcache, arena);
781 				}
782 				malloc_mutex_unlock(tsdn, &arena->lock);
783 			}
784 		}
785 	}
786 	je_malloc_stats_print(NULL, NULL, NULL);
787 }
788 
789 /*
790  * End miscellaneous support functions.
791  */
792 /******************************************************************************/
793 /*
794  * Begin initialization functions.
795  */
796 
797 #ifndef JEMALLOC_HAVE_SECURE_GETENV
798 static char *
799 secure_getenv(const char *name)
800 {
801 
802 #  ifdef JEMALLOC_HAVE_ISSETUGID
803 	if (issetugid() != 0)
804 		return (NULL);
805 #  endif
806 	return (getenv(name));
807 }
808 #endif
809 
810 static unsigned
811 malloc_ncpus(void)
812 {
813 	long result;
814 
815 #ifdef _WIN32
816 	SYSTEM_INFO si;
817 	GetSystemInfo(&si);
818 	result = si.dwNumberOfProcessors;
819 #elif defined(JEMALLOC_GLIBC_MALLOC_HOOK) && defined(CPU_COUNT)
820 	/*
821 	 * glibc >= 2.6 has the CPU_COUNT macro.
822 	 *
823 	 * glibc's sysconf() uses isspace().  glibc allocates for the first time
824 	 * *before* setting up the isspace tables.  Therefore we need a
825 	 * different method to get the number of CPUs.
826 	 */
827 	{
828 		cpu_set_t set;
829 
830 		pthread_getaffinity_np(pthread_self(), sizeof(set), &set);
831 		result = CPU_COUNT(&set);
832 	}
833 #else
834 	result = sysconf(_SC_NPROCESSORS_ONLN);
835 #endif
836 	return ((result == -1) ? 1 : (unsigned)result);
837 }
838 
839 static bool
840 malloc_conf_next(char const **opts_p, char const **k_p, size_t *klen_p,
841     char const **v_p, size_t *vlen_p)
842 {
843 	bool accept;
844 	const char *opts = *opts_p;
845 
846 	*k_p = opts;
847 
848 	for (accept = false; !accept;) {
849 		switch (*opts) {
850 		case 'A': case 'B': case 'C': case 'D': case 'E': case 'F':
851 		case 'G': case 'H': case 'I': case 'J': case 'K': case 'L':
852 		case 'M': case 'N': case 'O': case 'P': case 'Q': case 'R':
853 		case 'S': case 'T': case 'U': case 'V': case 'W': case 'X':
854 		case 'Y': case 'Z':
855 		case 'a': case 'b': case 'c': case 'd': case 'e': case 'f':
856 		case 'g': case 'h': case 'i': case 'j': case 'k': case 'l':
857 		case 'm': case 'n': case 'o': case 'p': case 'q': case 'r':
858 		case 's': case 't': case 'u': case 'v': case 'w': case 'x':
859 		case 'y': case 'z':
860 		case '0': case '1': case '2': case '3': case '4': case '5':
861 		case '6': case '7': case '8': case '9':
862 		case '_':
863 			opts++;
864 			break;
865 		case ':':
866 			opts++;
867 			*klen_p = (uintptr_t)opts - 1 - (uintptr_t)*k_p;
868 			*v_p = opts;
869 			accept = true;
870 			break;
871 		case '\0':
872 			if (opts != *opts_p) {
873 				malloc_write("<jemalloc>: Conf string ends "
874 				    "with key\n");
875 			}
876 			return (true);
877 		default:
878 			malloc_write("<jemalloc>: Malformed conf string\n");
879 			return (true);
880 		}
881 	}
882 
883 	for (accept = false; !accept;) {
884 		switch (*opts) {
885 		case ',':
886 			opts++;
887 			/*
888 			 * Look ahead one character here, because the next time
889 			 * this function is called, it will assume that end of
890 			 * input has been cleanly reached if no input remains,
891 			 * but we have optimistically already consumed the
892 			 * comma if one exists.
893 			 */
894 			if (*opts == '\0') {
895 				malloc_write("<jemalloc>: Conf string ends "
896 				    "with comma\n");
897 			}
898 			*vlen_p = (uintptr_t)opts - 1 - (uintptr_t)*v_p;
899 			accept = true;
900 			break;
901 		case '\0':
902 			*vlen_p = (uintptr_t)opts - (uintptr_t)*v_p;
903 			accept = true;
904 			break;
905 		default:
906 			opts++;
907 			break;
908 		}
909 	}
910 
911 	*opts_p = opts;
912 	return (false);
913 }
914 
915 static void
916 malloc_conf_error(const char *msg, const char *k, size_t klen, const char *v,
917     size_t vlen)
918 {
919 
920 	malloc_printf("<jemalloc>: %s: %.*s:%.*s\n", msg, (int)klen, k,
921 	    (int)vlen, v);
922 }
923 
924 static void
925 malloc_slow_flag_init(void)
926 {
927 	/*
928 	 * Combine the runtime options into malloc_slow for fast path.  Called
929 	 * after processing all the options.
930 	 */
931 	malloc_slow_flags |= (opt_junk_alloc ? flag_opt_junk_alloc : 0)
932 	    | (opt_junk_free ? flag_opt_junk_free : 0)
933 	    | (opt_quarantine ? flag_opt_quarantine : 0)
934 	    | (opt_zero ? flag_opt_zero : 0)
935 	    | (opt_utrace ? flag_opt_utrace : 0)
936 	    | (opt_xmalloc ? flag_opt_xmalloc : 0);
937 
938 	if (config_valgrind)
939 		malloc_slow_flags |= (in_valgrind ? flag_in_valgrind : 0);
940 
941 	malloc_slow = (malloc_slow_flags != 0);
942 }
943 
944 static void
945 malloc_conf_init(void)
946 {
947 	unsigned i;
948 	char buf[PATH_MAX + 1];
949 	const char *opts, *k, *v;
950 	size_t klen, vlen;
951 
952 	/*
953 	 * Automatically configure valgrind before processing options.  The
954 	 * valgrind option remains in jemalloc 3.x for compatibility reasons.
955 	 */
956 	if (config_valgrind) {
957 		in_valgrind = (RUNNING_ON_VALGRIND != 0) ? true : false;
958 		if (config_fill && unlikely(in_valgrind)) {
959 			opt_junk = "false";
960 			opt_junk_alloc = false;
961 			opt_junk_free = false;
962 			assert(!opt_zero);
963 			opt_quarantine = JEMALLOC_VALGRIND_QUARANTINE_DEFAULT;
964 			opt_redzone = true;
965 		}
966 		if (config_tcache && unlikely(in_valgrind))
967 			opt_tcache = false;
968 	}
969 
970 	for (i = 0; i < 4; i++) {
971 		/* Get runtime configuration. */
972 		switch (i) {
973 		case 0:
974 			opts = config_malloc_conf;
975 			break;
976 		case 1:
977 			if (je_malloc_conf != NULL) {
978 				/*
979 				 * Use options that were compiled into the
980 				 * program.
981 				 */
982 				opts = je_malloc_conf;
983 			} else {
984 				/* No configuration specified. */
985 				buf[0] = '\0';
986 				opts = buf;
987 			}
988 			break;
989 		case 2: {
990 			ssize_t linklen = 0;
991 #ifndef _WIN32
992 			int saved_errno = errno;
993 			const char *linkname =
994 #  ifdef JEMALLOC_PREFIX
995 			    "/etc/"JEMALLOC_PREFIX"malloc.conf"
996 #  else
997 			    "/etc/malloc.conf"
998 #  endif
999 			    ;
1000 
1001 			/*
1002 			 * Try to use the contents of the "/etc/malloc.conf"
1003 			 * symbolic link's name.
1004 			 */
1005 			linklen = readlink(linkname, buf, sizeof(buf) - 1);
1006 			if (linklen == -1) {
1007 				/* No configuration specified. */
1008 				linklen = 0;
1009 				/* Restore errno. */
1010 				set_errno(saved_errno);
1011 			}
1012 #endif
1013 			buf[linklen] = '\0';
1014 			opts = buf;
1015 			break;
1016 		} case 3: {
1017 			const char *envname =
1018 #ifdef JEMALLOC_PREFIX
1019 			    JEMALLOC_CPREFIX"MALLOC_CONF"
1020 #else
1021 			    "MALLOC_CONF"
1022 #endif
1023 			    ;
1024 
1025 			if ((opts = secure_getenv(envname)) != NULL) {
1026 				/*
1027 				 * Do nothing; opts is already initialized to
1028 				 * the value of the MALLOC_CONF environment
1029 				 * variable.
1030 				 */
1031 			} else {
1032 				/* No configuration specified. */
1033 				buf[0] = '\0';
1034 				opts = buf;
1035 			}
1036 			break;
1037 		} default:
1038 			not_reached();
1039 			buf[0] = '\0';
1040 			opts = buf;
1041 		}
1042 
1043 		while (*opts != '\0' && !malloc_conf_next(&opts, &k, &klen, &v,
1044 		    &vlen)) {
1045 #define	CONF_MATCH(n)							\
1046 	(sizeof(n)-1 == klen && strncmp(n, k, klen) == 0)
1047 #define	CONF_MATCH_VALUE(n)						\
1048 	(sizeof(n)-1 == vlen && strncmp(n, v, vlen) == 0)
1049 #define	CONF_HANDLE_BOOL(o, n, cont)					\
1050 			if (CONF_MATCH(n)) {				\
1051 				if (CONF_MATCH_VALUE("true"))		\
1052 					o = true;			\
1053 				else if (CONF_MATCH_VALUE("false"))	\
1054 					o = false;			\
1055 				else {					\
1056 					malloc_conf_error(		\
1057 					    "Invalid conf value",	\
1058 					    k, klen, v, vlen);		\
1059 				}					\
1060 				if (cont)				\
1061 					continue;			\
1062 			}
1063 #define	CONF_MIN_no(um, min)	false
1064 #define	CONF_MIN_yes(um, min)	((um) < (min))
1065 #define	CONF_MAX_no(um, max)	false
1066 #define	CONF_MAX_yes(um, max)	((um) > (max))
1067 #define	CONF_HANDLE_T_U(t, o, n, min, max, check_min, check_max, clip)	\
1068 			if (CONF_MATCH(n)) {				\
1069 				uintmax_t um;				\
1070 				char *end;				\
1071 									\
1072 				set_errno(0);				\
1073 				um = malloc_strtoumax(v, &end, 0);	\
1074 				if (get_errno() != 0 || (uintptr_t)end -\
1075 				    (uintptr_t)v != vlen) {		\
1076 					malloc_conf_error(		\
1077 					    "Invalid conf value",	\
1078 					    k, klen, v, vlen);		\
1079 				} else if (clip) {			\
1080 					if (CONF_MIN_##check_min(um,	\
1081 					    (min)))			\
1082 						o = (t)(min);		\
1083 					else if (CONF_MAX_##check_max(	\
1084 					    um, (max)))			\
1085 						o = (t)(max);		\
1086 					else				\
1087 						o = (t)um;		\
1088 				} else {				\
1089 					if (CONF_MIN_##check_min(um,	\
1090 					    (min)) ||			\
1091 					    CONF_MAX_##check_max(um,	\
1092 					    (max))) {			\
1093 						malloc_conf_error(	\
1094 						    "Out-of-range "	\
1095 						    "conf value",	\
1096 						    k, klen, v, vlen);	\
1097 					} else				\
1098 						o = (t)um;		\
1099 				}					\
1100 				continue;				\
1101 			}
1102 #define	CONF_HANDLE_UNSIGNED(o, n, min, max, check_min, check_max,	\
1103     clip)								\
1104 			CONF_HANDLE_T_U(unsigned, o, n, min, max,	\
1105 			    check_min, check_max, clip)
1106 #define	CONF_HANDLE_SIZE_T(o, n, min, max, check_min, check_max, clip)	\
1107 			CONF_HANDLE_T_U(size_t, o, n, min, max,		\
1108 			    check_min, check_max, clip)
1109 #define	CONF_HANDLE_SSIZE_T(o, n, min, max)				\
1110 			if (CONF_MATCH(n)) {				\
1111 				long l;					\
1112 				char *end;				\
1113 									\
1114 				set_errno(0);				\
1115 				l = strtol(v, &end, 0);			\
1116 				if (get_errno() != 0 || (uintptr_t)end -\
1117 				    (uintptr_t)v != vlen) {		\
1118 					malloc_conf_error(		\
1119 					    "Invalid conf value",	\
1120 					    k, klen, v, vlen);		\
1121 				} else if (l < (ssize_t)(min) || l >	\
1122 				    (ssize_t)(max)) {			\
1123 					malloc_conf_error(		\
1124 					    "Out-of-range conf value",	\
1125 					    k, klen, v, vlen);		\
1126 				} else					\
1127 					o = l;				\
1128 				continue;				\
1129 			}
1130 #define	CONF_HANDLE_CHAR_P(o, n, d)					\
1131 			if (CONF_MATCH(n)) {				\
1132 				size_t cpylen = (vlen <=		\
1133 				    sizeof(o)-1) ? vlen :		\
1134 				    sizeof(o)-1;			\
1135 				strncpy(o, v, cpylen);			\
1136 				o[cpylen] = '\0';			\
1137 				continue;				\
1138 			}
1139 
1140 			CONF_HANDLE_BOOL(opt_abort, "abort", true)
1141 			/*
1142 			 * Chunks always require at least one header page,
1143 			 * as many as 2^(LG_SIZE_CLASS_GROUP+1) data pages, and
1144 			 * possibly an additional page in the presence of
1145 			 * redzones.  In order to simplify options processing,
1146 			 * use a conservative bound that accommodates all these
1147 			 * constraints.
1148 			 */
1149 			CONF_HANDLE_SIZE_T(opt_lg_chunk, "lg_chunk", LG_PAGE +
1150 			    LG_SIZE_CLASS_GROUP + (config_fill ? 2 : 1),
1151 			    (sizeof(size_t) << 3) - 1, yes, yes, true)
1152 			if (strncmp("dss", k, klen) == 0) {
1153 				int i;
1154 				bool match = false;
1155 				for (i = 0; i < dss_prec_limit; i++) {
1156 					if (strncmp(dss_prec_names[i], v, vlen)
1157 					    == 0) {
1158 						if (chunk_dss_prec_set(i)) {
1159 							malloc_conf_error(
1160 							    "Error setting dss",
1161 							    k, klen, v, vlen);
1162 						} else {
1163 							opt_dss =
1164 							    dss_prec_names[i];
1165 							match = true;
1166 							break;
1167 						}
1168 					}
1169 				}
1170 				if (!match) {
1171 					malloc_conf_error("Invalid conf value",
1172 					    k, klen, v, vlen);
1173 				}
1174 				continue;
1175 			}
1176 			CONF_HANDLE_UNSIGNED(opt_narenas, "narenas", 1,
1177 			    UINT_MAX, yes, no, false)
1178 			if (strncmp("purge", k, klen) == 0) {
1179 				int i;
1180 				bool match = false;
1181 				for (i = 0; i < purge_mode_limit; i++) {
1182 					if (strncmp(purge_mode_names[i], v,
1183 					    vlen) == 0) {
1184 						opt_purge = (purge_mode_t)i;
1185 						match = true;
1186 						break;
1187 					}
1188 				}
1189 				if (!match) {
1190 					malloc_conf_error("Invalid conf value",
1191 					    k, klen, v, vlen);
1192 				}
1193 				continue;
1194 			}
1195 			CONF_HANDLE_SSIZE_T(opt_lg_dirty_mult, "lg_dirty_mult",
1196 			    -1, (sizeof(size_t) << 3) - 1)
1197 			CONF_HANDLE_SSIZE_T(opt_decay_time, "decay_time", -1,
1198 			    NSTIME_SEC_MAX);
1199 			CONF_HANDLE_BOOL(opt_stats_print, "stats_print", true)
1200 			if (config_fill) {
1201 				if (CONF_MATCH("junk")) {
1202 					if (CONF_MATCH_VALUE("true")) {
1203 						if (config_valgrind &&
1204 						    unlikely(in_valgrind)) {
1205 							malloc_conf_error(
1206 							"Deallocation-time "
1207 							"junk filling cannot "
1208 							"be enabled while "
1209 							"running inside "
1210 							"Valgrind", k, klen, v,
1211 							vlen);
1212 						} else {
1213 							opt_junk = "true";
1214 							opt_junk_alloc = true;
1215 							opt_junk_free = true;
1216 						}
1217 					} else if (CONF_MATCH_VALUE("false")) {
1218 						opt_junk = "false";
1219 						opt_junk_alloc = opt_junk_free =
1220 						    false;
1221 					} else if (CONF_MATCH_VALUE("alloc")) {
1222 						opt_junk = "alloc";
1223 						opt_junk_alloc = true;
1224 						opt_junk_free = false;
1225 					} else if (CONF_MATCH_VALUE("free")) {
1226 						if (config_valgrind &&
1227 						    unlikely(in_valgrind)) {
1228 							malloc_conf_error(
1229 							"Deallocation-time "
1230 							"junk filling cannot "
1231 							"be enabled while "
1232 							"running inside "
1233 							"Valgrind", k, klen, v,
1234 							vlen);
1235 						} else {
1236 							opt_junk = "free";
1237 							opt_junk_alloc = false;
1238 							opt_junk_free = true;
1239 						}
1240 					} else {
1241 						malloc_conf_error(
1242 						    "Invalid conf value", k,
1243 						    klen, v, vlen);
1244 					}
1245 					continue;
1246 				}
1247 				CONF_HANDLE_SIZE_T(opt_quarantine, "quarantine",
1248 				    0, SIZE_T_MAX, no, no, false)
1249 				CONF_HANDLE_BOOL(opt_redzone, "redzone", true)
1250 				CONF_HANDLE_BOOL(opt_zero, "zero", true)
1251 			}
1252 			if (config_utrace) {
1253 				CONF_HANDLE_BOOL(opt_utrace, "utrace", true)
1254 			}
1255 			if (config_xmalloc) {
1256 				CONF_HANDLE_BOOL(opt_xmalloc, "xmalloc", true)
1257 			}
1258 			if (config_tcache) {
1259 				CONF_HANDLE_BOOL(opt_tcache, "tcache",
1260 				    !config_valgrind || !in_valgrind)
1261 				if (CONF_MATCH("tcache")) {
1262 					assert(config_valgrind && in_valgrind);
1263 					if (opt_tcache) {
1264 						opt_tcache = false;
1265 						malloc_conf_error(
1266 						"tcache cannot be enabled "
1267 						"while running inside Valgrind",
1268 						k, klen, v, vlen);
1269 					}
1270 					continue;
1271 				}
1272 				CONF_HANDLE_SSIZE_T(opt_lg_tcache_max,
1273 				    "lg_tcache_max", -1,
1274 				    (sizeof(size_t) << 3) - 1)
1275 			}
1276 			if (config_prof) {
1277 				CONF_HANDLE_BOOL(opt_prof, "prof", true)
1278 				CONF_HANDLE_CHAR_P(opt_prof_prefix,
1279 				    "prof_prefix", "jeprof")
1280 				CONF_HANDLE_BOOL(opt_prof_active, "prof_active",
1281 				    true)
1282 				CONF_HANDLE_BOOL(opt_prof_thread_active_init,
1283 				    "prof_thread_active_init", true)
1284 				CONF_HANDLE_SIZE_T(opt_lg_prof_sample,
1285 				    "lg_prof_sample", 0, (sizeof(uint64_t) << 3)
1286 				    - 1, no, yes, true)
1287 				CONF_HANDLE_BOOL(opt_prof_accum, "prof_accum",
1288 				    true)
1289 				CONF_HANDLE_SSIZE_T(opt_lg_prof_interval,
1290 				    "lg_prof_interval", -1,
1291 				    (sizeof(uint64_t) << 3) - 1)
1292 				CONF_HANDLE_BOOL(opt_prof_gdump, "prof_gdump",
1293 				    true)
1294 				CONF_HANDLE_BOOL(opt_prof_final, "prof_final",
1295 				    true)
1296 				CONF_HANDLE_BOOL(opt_prof_leak, "prof_leak",
1297 				    true)
1298 			}
1299 			malloc_conf_error("Invalid conf pair", k, klen, v,
1300 			    vlen);
1301 #undef CONF_MATCH
1302 #undef CONF_MATCH_VALUE
1303 #undef CONF_HANDLE_BOOL
1304 #undef CONF_MIN_no
1305 #undef CONF_MIN_yes
1306 #undef CONF_MAX_no
1307 #undef CONF_MAX_yes
1308 #undef CONF_HANDLE_T_U
1309 #undef CONF_HANDLE_UNSIGNED
1310 #undef CONF_HANDLE_SIZE_T
1311 #undef CONF_HANDLE_SSIZE_T
1312 #undef CONF_HANDLE_CHAR_P
1313 		}
1314 	}
1315 }
1316 
1317 static bool
1318 malloc_init_hard_needed(void)
1319 {
1320 
1321 	if (malloc_initialized() || (IS_INITIALIZER && malloc_init_state ==
1322 	    malloc_init_recursible)) {
1323 		/*
1324 		 * Another thread initialized the allocator before this one
1325 		 * acquired init_lock, or this thread is the initializing
1326 		 * thread, and it is recursively allocating.
1327 		 */
1328 		return (false);
1329 	}
1330 #ifdef JEMALLOC_THREADED_INIT
1331 	if (malloc_initializer != NO_INITIALIZER && !IS_INITIALIZER) {
1332 		spin_t spinner;
1333 
1334 		/* Busy-wait until the initializing thread completes. */
1335 		spin_init(&spinner);
1336 		do {
1337 			malloc_mutex_unlock(TSDN_NULL, &init_lock);
1338 			spin_adaptive(&spinner);
1339 			malloc_mutex_lock(TSDN_NULL, &init_lock);
1340 		} while (!malloc_initialized());
1341 		return (false);
1342 	}
1343 #endif
1344 	return (true);
1345 }
1346 
1347 static bool
1348 malloc_init_hard_a0_locked()
1349 {
1350 
1351 	malloc_initializer = INITIALIZER;
1352 
1353 	if (config_prof)
1354 		prof_boot0();
1355 	malloc_conf_init();
1356 	if (opt_stats_print) {
1357 		/* Print statistics at exit. */
1358 		if (atexit(stats_print_atexit) != 0) {
1359 			malloc_write("<jemalloc>: Error in atexit()\n");
1360 			if (opt_abort)
1361 				abort();
1362 		}
1363 	}
1364 	pages_boot();
1365 	if (base_boot())
1366 		return (true);
1367 	if (chunk_boot())
1368 		return (true);
1369 	if (ctl_boot())
1370 		return (true);
1371 	if (config_prof)
1372 		prof_boot1();
1373 	arena_boot();
1374 	if (config_tcache && tcache_boot(TSDN_NULL))
1375 		return (true);
1376 	if (malloc_mutex_init(&arenas_lock, "arenas", WITNESS_RANK_ARENAS))
1377 		return (true);
1378 	/*
1379 	 * Create enough scaffolding to allow recursive allocation in
1380 	 * malloc_ncpus().
1381 	 */
1382 	narenas_auto = 1;
1383 	narenas_total_set(narenas_auto);
1384 	arenas = &a0;
1385 	memset(arenas, 0, sizeof(arena_t *) * narenas_auto);
1386 	/*
1387 	 * Initialize one arena here.  The rest are lazily created in
1388 	 * arena_choose_hard().
1389 	 */
1390 	if (arena_init(TSDN_NULL, 0) == NULL)
1391 		return (true);
1392 
1393 	malloc_init_state = malloc_init_a0_initialized;
1394 
1395 	return (false);
1396 }
1397 
1398 static bool
1399 malloc_init_hard_a0(void)
1400 {
1401 	bool ret;
1402 
1403 	malloc_mutex_lock(TSDN_NULL, &init_lock);
1404 	ret = malloc_init_hard_a0_locked();
1405 	malloc_mutex_unlock(TSDN_NULL, &init_lock);
1406 	return (ret);
1407 }
1408 
1409 /* Initialize data structures which may trigger recursive allocation. */
1410 static bool
1411 malloc_init_hard_recursible(void)
1412 {
1413 
1414 	malloc_init_state = malloc_init_recursible;
1415 
1416 	ncpus = malloc_ncpus();
1417 
1418 #if (defined(JEMALLOC_HAVE_PTHREAD_ATFORK) && !defined(JEMALLOC_MUTEX_INIT_CB) \
1419     && !defined(JEMALLOC_ZONE) && !defined(_WIN32) && \
1420     !defined(__native_client__))
1421 	/* LinuxThreads' pthread_atfork() allocates. */
1422 	if (pthread_atfork(jemalloc_prefork, jemalloc_postfork_parent,
1423 	    jemalloc_postfork_child) != 0) {
1424 		malloc_write("<jemalloc>: Error in pthread_atfork()\n");
1425 		if (opt_abort)
1426 			abort();
1427 		return (true);
1428 	}
1429 #endif
1430 
1431 	return (false);
1432 }
1433 
1434 static bool
1435 malloc_init_hard_finish(tsdn_t *tsdn)
1436 {
1437 
1438 	if (malloc_mutex_boot())
1439 		return (true);
1440 
1441 	if (opt_narenas == 0) {
1442 		/*
1443 		 * For SMP systems, create more than one arena per CPU by
1444 		 * default.
1445 		 */
1446 		if (ncpus > 1)
1447 			opt_narenas = ncpus << 2;
1448 		else
1449 			opt_narenas = 1;
1450 	}
1451 	narenas_auto = opt_narenas;
1452 	/*
1453 	 * Limit the number of arenas to the indexing range of MALLOCX_ARENA().
1454 	 */
1455 	if (narenas_auto > MALLOCX_ARENA_MAX) {
1456 		narenas_auto = MALLOCX_ARENA_MAX;
1457 		malloc_printf("<jemalloc>: Reducing narenas to limit (%d)\n",
1458 		    narenas_auto);
1459 	}
1460 	narenas_total_set(narenas_auto);
1461 
1462 	/* Allocate and initialize arenas. */
1463 	arenas = (arena_t **)base_alloc(tsdn, sizeof(arena_t *) *
1464 	    (MALLOCX_ARENA_MAX+1));
1465 	if (arenas == NULL)
1466 		return (true);
1467 	/* Copy the pointer to the one arena that was already initialized. */
1468 	arena_set(0, a0);
1469 
1470 	malloc_init_state = malloc_init_initialized;
1471 	malloc_slow_flag_init();
1472 
1473 	return (false);
1474 }
1475 
1476 static bool
1477 malloc_init_hard(void)
1478 {
1479 	tsd_t *tsd;
1480 
1481 #if defined(_WIN32) && _WIN32_WINNT < 0x0600
1482 	_init_init_lock();
1483 #endif
1484 	malloc_mutex_lock(TSDN_NULL, &init_lock);
1485 	if (!malloc_init_hard_needed()) {
1486 		malloc_mutex_unlock(TSDN_NULL, &init_lock);
1487 		return (false);
1488 	}
1489 
1490 	if (malloc_init_state != malloc_init_a0_initialized &&
1491 	    malloc_init_hard_a0_locked()) {
1492 		malloc_mutex_unlock(TSDN_NULL, &init_lock);
1493 		return (true);
1494 	}
1495 
1496 	malloc_mutex_unlock(TSDN_NULL, &init_lock);
1497 	/* Recursive allocation relies on functional tsd. */
1498 	tsd = malloc_tsd_boot0();
1499 	if (tsd == NULL)
1500 		return (true);
1501 	if (malloc_init_hard_recursible())
1502 		return (true);
1503 	malloc_mutex_lock(tsd_tsdn(tsd), &init_lock);
1504 
1505 	if (config_prof && prof_boot2(tsd)) {
1506 		malloc_mutex_unlock(tsd_tsdn(tsd), &init_lock);
1507 		return (true);
1508 	}
1509 
1510 	if (malloc_init_hard_finish(tsd_tsdn(tsd))) {
1511 		malloc_mutex_unlock(tsd_tsdn(tsd), &init_lock);
1512 		return (true);
1513 	}
1514 
1515 	malloc_mutex_unlock(tsd_tsdn(tsd), &init_lock);
1516 	malloc_tsd_boot1();
1517 	return (false);
1518 }
1519 
1520 /*
1521  * End initialization functions.
1522  */
1523 /******************************************************************************/
1524 /*
1525  * Begin malloc(3)-compatible functions.
1526  */
1527 
1528 static void *
1529 ialloc_prof_sample(tsd_t *tsd, size_t usize, szind_t ind, bool zero,
1530     prof_tctx_t *tctx, bool slow_path)
1531 {
1532 	void *p;
1533 
1534 	if (tctx == NULL)
1535 		return (NULL);
1536 	if (usize <= SMALL_MAXCLASS) {
1537 		szind_t ind_large = size2index(LARGE_MINCLASS);
1538 		p = ialloc(tsd, LARGE_MINCLASS, ind_large, zero, slow_path);
1539 		if (p == NULL)
1540 			return (NULL);
1541 		arena_prof_promoted(tsd_tsdn(tsd), p, usize);
1542 	} else
1543 		p = ialloc(tsd, usize, ind, zero, slow_path);
1544 
1545 	return (p);
1546 }
1547 
1548 JEMALLOC_ALWAYS_INLINE_C void *
1549 ialloc_prof(tsd_t *tsd, size_t usize, szind_t ind, bool zero, bool slow_path)
1550 {
1551 	void *p;
1552 	prof_tctx_t *tctx;
1553 
1554 	tctx = prof_alloc_prep(tsd, usize, prof_active_get_unlocked(), true);
1555 	if (unlikely((uintptr_t)tctx != (uintptr_t)1U))
1556 		p = ialloc_prof_sample(tsd, usize, ind, zero, tctx, slow_path);
1557 	else
1558 		p = ialloc(tsd, usize, ind, zero, slow_path);
1559 	if (unlikely(p == NULL)) {
1560 		prof_alloc_rollback(tsd, tctx, true);
1561 		return (NULL);
1562 	}
1563 	prof_malloc(tsd_tsdn(tsd), p, usize, tctx);
1564 
1565 	return (p);
1566 }
1567 
1568 /*
1569  * ialloc_body() is inlined so that fast and slow paths are generated separately
1570  * with statically known slow_path.
1571  *
1572  * This function guarantees that *tsdn is non-NULL on success.
1573  */
1574 JEMALLOC_ALWAYS_INLINE_C void *
1575 ialloc_body(size_t size, bool zero, tsdn_t **tsdn, size_t *usize,
1576     bool slow_path)
1577 {
1578 	tsd_t *tsd;
1579 	szind_t ind;
1580 
1581 	if (slow_path && unlikely(malloc_init())) {
1582 		*tsdn = NULL;
1583 		return (NULL);
1584 	}
1585 
1586 	tsd = tsd_fetch();
1587 	*tsdn = tsd_tsdn(tsd);
1588 	witness_assert_lockless(tsd_tsdn(tsd));
1589 
1590 	ind = size2index(size);
1591 	if (unlikely(ind >= NSIZES))
1592 		return (NULL);
1593 
1594 	if (config_stats || (config_prof && opt_prof) || (slow_path &&
1595 	    config_valgrind && unlikely(in_valgrind))) {
1596 		*usize = index2size(ind);
1597 		assert(*usize > 0 && *usize <= HUGE_MAXCLASS);
1598 	}
1599 
1600 	if (config_prof && opt_prof)
1601 		return (ialloc_prof(tsd, *usize, ind, zero, slow_path));
1602 
1603 	return (ialloc(tsd, size, ind, zero, slow_path));
1604 }
1605 
1606 JEMALLOC_ALWAYS_INLINE_C void
1607 ialloc_post_check(void *ret, tsdn_t *tsdn, size_t usize, const char *func,
1608     bool update_errno, bool slow_path)
1609 {
1610 
1611 	assert(!tsdn_null(tsdn) || ret == NULL);
1612 
1613 	if (unlikely(ret == NULL)) {
1614 		if (slow_path && config_xmalloc && unlikely(opt_xmalloc)) {
1615 			malloc_printf("<jemalloc>: Error in %s(): out of "
1616 			    "memory\n", func);
1617 			abort();
1618 		}
1619 		if (update_errno)
1620 			set_errno(ENOMEM);
1621 	}
1622 	if (config_stats && likely(ret != NULL)) {
1623 		assert(usize == isalloc(tsdn, ret, config_prof));
1624 		*tsd_thread_allocatedp_get(tsdn_tsd(tsdn)) += usize;
1625 	}
1626 	witness_assert_lockless(tsdn);
1627 }
1628 
1629 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
1630 void JEMALLOC_NOTHROW *
1631 JEMALLOC_ATTR(malloc) JEMALLOC_ALLOC_SIZE(1)
1632 je_malloc(size_t size)
1633 {
1634 	void *ret;
1635 	tsdn_t *tsdn;
1636 	size_t usize JEMALLOC_CC_SILENCE_INIT(0);
1637 
1638 	if (size == 0)
1639 		size = 1;
1640 
1641 	if (likely(!malloc_slow)) {
1642 		ret = ialloc_body(size, false, &tsdn, &usize, false);
1643 		ialloc_post_check(ret, tsdn, usize, "malloc", true, false);
1644 	} else {
1645 		ret = ialloc_body(size, false, &tsdn, &usize, true);
1646 		ialloc_post_check(ret, tsdn, usize, "malloc", true, true);
1647 		UTRACE(0, size, ret);
1648 		JEMALLOC_VALGRIND_MALLOC(ret != NULL, tsdn, ret, usize, false);
1649 	}
1650 
1651 	return (ret);
1652 }
1653 
1654 static void *
1655 imemalign_prof_sample(tsd_t *tsd, size_t alignment, size_t usize,
1656     prof_tctx_t *tctx)
1657 {
1658 	void *p;
1659 
1660 	if (tctx == NULL)
1661 		return (NULL);
1662 	if (usize <= SMALL_MAXCLASS) {
1663 		assert(sa2u(LARGE_MINCLASS, alignment) == LARGE_MINCLASS);
1664 		p = ipalloc(tsd, LARGE_MINCLASS, alignment, false);
1665 		if (p == NULL)
1666 			return (NULL);
1667 		arena_prof_promoted(tsd_tsdn(tsd), p, usize);
1668 	} else
1669 		p = ipalloc(tsd, usize, alignment, false);
1670 
1671 	return (p);
1672 }
1673 
1674 JEMALLOC_ALWAYS_INLINE_C void *
1675 imemalign_prof(tsd_t *tsd, size_t alignment, size_t usize)
1676 {
1677 	void *p;
1678 	prof_tctx_t *tctx;
1679 
1680 	tctx = prof_alloc_prep(tsd, usize, prof_active_get_unlocked(), true);
1681 	if (unlikely((uintptr_t)tctx != (uintptr_t)1U))
1682 		p = imemalign_prof_sample(tsd, alignment, usize, tctx);
1683 	else
1684 		p = ipalloc(tsd, usize, alignment, false);
1685 	if (unlikely(p == NULL)) {
1686 		prof_alloc_rollback(tsd, tctx, true);
1687 		return (NULL);
1688 	}
1689 	prof_malloc(tsd_tsdn(tsd), p, usize, tctx);
1690 
1691 	return (p);
1692 }
1693 
1694 JEMALLOC_ATTR(nonnull(1))
1695 static int
1696 imemalign(void **memptr, size_t alignment, size_t size, size_t min_alignment)
1697 {
1698 	int ret;
1699 	tsd_t *tsd;
1700 	size_t usize;
1701 	void *result;
1702 
1703 	assert(min_alignment != 0);
1704 
1705 	if (unlikely(malloc_init())) {
1706 		tsd = NULL;
1707 		result = NULL;
1708 		goto label_oom;
1709 	}
1710 	tsd = tsd_fetch();
1711 	witness_assert_lockless(tsd_tsdn(tsd));
1712 	if (size == 0)
1713 		size = 1;
1714 
1715 	/* Make sure that alignment is a large enough power of 2. */
1716 	if (unlikely(((alignment - 1) & alignment) != 0
1717 	    || (alignment < min_alignment))) {
1718 		if (config_xmalloc && unlikely(opt_xmalloc)) {
1719 			malloc_write("<jemalloc>: Error allocating "
1720 			    "aligned memory: invalid alignment\n");
1721 			abort();
1722 		}
1723 		result = NULL;
1724 		ret = EINVAL;
1725 		goto label_return;
1726 	}
1727 
1728 	usize = sa2u(size, alignment);
1729 	if (unlikely(usize == 0 || usize > HUGE_MAXCLASS)) {
1730 		result = NULL;
1731 		goto label_oom;
1732 	}
1733 
1734 	if (config_prof && opt_prof)
1735 		result = imemalign_prof(tsd, alignment, usize);
1736 	else
1737 		result = ipalloc(tsd, usize, alignment, false);
1738 	if (unlikely(result == NULL))
1739 		goto label_oom;
1740 	assert(((uintptr_t)result & (alignment - 1)) == ZU(0));
1741 
1742 	*memptr = result;
1743 	ret = 0;
1744 label_return:
1745 	if (config_stats && likely(result != NULL)) {
1746 		assert(usize == isalloc(tsd_tsdn(tsd), result, config_prof));
1747 		*tsd_thread_allocatedp_get(tsd) += usize;
1748 	}
1749 	UTRACE(0, size, result);
1750 	JEMALLOC_VALGRIND_MALLOC(result != NULL, tsd_tsdn(tsd), result, usize,
1751 	    false);
1752 	witness_assert_lockless(tsd_tsdn(tsd));
1753 	return (ret);
1754 label_oom:
1755 	assert(result == NULL);
1756 	if (config_xmalloc && unlikely(opt_xmalloc)) {
1757 		malloc_write("<jemalloc>: Error allocating aligned memory: "
1758 		    "out of memory\n");
1759 		abort();
1760 	}
1761 	ret = ENOMEM;
1762 	witness_assert_lockless(tsd_tsdn(tsd));
1763 	goto label_return;
1764 }
1765 
1766 JEMALLOC_EXPORT int JEMALLOC_NOTHROW
1767 JEMALLOC_ATTR(nonnull(1))
1768 je_posix_memalign(void **memptr, size_t alignment, size_t size)
1769 {
1770 	int ret;
1771 
1772 	ret = imemalign(memptr, alignment, size, sizeof(void *));
1773 
1774 	return (ret);
1775 }
1776 
1777 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
1778 void JEMALLOC_NOTHROW *
1779 JEMALLOC_ATTR(malloc) JEMALLOC_ALLOC_SIZE(2)
1780 je_aligned_alloc(size_t alignment, size_t size)
1781 {
1782 	void *ret;
1783 	int err;
1784 
1785 	if (unlikely((err = imemalign(&ret, alignment, size, 1)) != 0)) {
1786 		ret = NULL;
1787 		set_errno(err);
1788 	}
1789 
1790 	return (ret);
1791 }
1792 
1793 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
1794 void JEMALLOC_NOTHROW *
1795 JEMALLOC_ATTR(malloc) JEMALLOC_ALLOC_SIZE2(1, 2)
1796 je_calloc(size_t num, size_t size)
1797 {
1798 	void *ret;
1799 	tsdn_t *tsdn;
1800 	size_t num_size;
1801 	size_t usize JEMALLOC_CC_SILENCE_INIT(0);
1802 
1803 	num_size = num * size;
1804 	if (unlikely(num_size == 0)) {
1805 		if (num == 0 || size == 0)
1806 			num_size = 1;
1807 		else
1808 			num_size = HUGE_MAXCLASS + 1; /* Trigger OOM. */
1809 	/*
1810 	 * Try to avoid division here.  We know that it isn't possible to
1811 	 * overflow during multiplication if neither operand uses any of the
1812 	 * most significant half of the bits in a size_t.
1813 	 */
1814 	} else if (unlikely(((num | size) & (SIZE_T_MAX << (sizeof(size_t) <<
1815 	    2))) && (num_size / size != num)))
1816 		num_size = HUGE_MAXCLASS + 1; /* size_t overflow. */
1817 
1818 	if (likely(!malloc_slow)) {
1819 		ret = ialloc_body(num_size, true, &tsdn, &usize, false);
1820 		ialloc_post_check(ret, tsdn, usize, "calloc", true, false);
1821 	} else {
1822 		ret = ialloc_body(num_size, true, &tsdn, &usize, true);
1823 		ialloc_post_check(ret, tsdn, usize, "calloc", true, true);
1824 		UTRACE(0, num_size, ret);
1825 		JEMALLOC_VALGRIND_MALLOC(ret != NULL, tsdn, ret, usize, true);
1826 	}
1827 
1828 	return (ret);
1829 }
1830 
1831 static void *
1832 irealloc_prof_sample(tsd_t *tsd, void *old_ptr, size_t old_usize, size_t usize,
1833     prof_tctx_t *tctx)
1834 {
1835 	void *p;
1836 
1837 	if (tctx == NULL)
1838 		return (NULL);
1839 	if (usize <= SMALL_MAXCLASS) {
1840 		p = iralloc(tsd, old_ptr, old_usize, LARGE_MINCLASS, 0, false);
1841 		if (p == NULL)
1842 			return (NULL);
1843 		arena_prof_promoted(tsd_tsdn(tsd), p, usize);
1844 	} else
1845 		p = iralloc(tsd, old_ptr, old_usize, usize, 0, false);
1846 
1847 	return (p);
1848 }
1849 
1850 JEMALLOC_ALWAYS_INLINE_C void *
1851 irealloc_prof(tsd_t *tsd, void *old_ptr, size_t old_usize, size_t usize)
1852 {
1853 	void *p;
1854 	bool prof_active;
1855 	prof_tctx_t *old_tctx, *tctx;
1856 
1857 	prof_active = prof_active_get_unlocked();
1858 	old_tctx = prof_tctx_get(tsd_tsdn(tsd), old_ptr);
1859 	tctx = prof_alloc_prep(tsd, usize, prof_active, true);
1860 	if (unlikely((uintptr_t)tctx != (uintptr_t)1U))
1861 		p = irealloc_prof_sample(tsd, old_ptr, old_usize, usize, tctx);
1862 	else
1863 		p = iralloc(tsd, old_ptr, old_usize, usize, 0, false);
1864 	if (unlikely(p == NULL)) {
1865 		prof_alloc_rollback(tsd, tctx, true);
1866 		return (NULL);
1867 	}
1868 	prof_realloc(tsd, p, usize, tctx, prof_active, true, old_ptr, old_usize,
1869 	    old_tctx);
1870 
1871 	return (p);
1872 }
1873 
1874 JEMALLOC_INLINE_C void
1875 ifree(tsd_t *tsd, void *ptr, tcache_t *tcache, bool slow_path)
1876 {
1877 	size_t usize;
1878 	UNUSED size_t rzsize JEMALLOC_CC_SILENCE_INIT(0);
1879 
1880 	witness_assert_lockless(tsd_tsdn(tsd));
1881 
1882 	assert(ptr != NULL);
1883 	assert(malloc_initialized() || IS_INITIALIZER);
1884 
1885 	if (config_prof && opt_prof) {
1886 		usize = isalloc(tsd_tsdn(tsd), ptr, config_prof);
1887 		prof_free(tsd, ptr, usize);
1888 	} else if (config_stats || config_valgrind)
1889 		usize = isalloc(tsd_tsdn(tsd), ptr, config_prof);
1890 	if (config_stats)
1891 		*tsd_thread_deallocatedp_get(tsd) += usize;
1892 
1893 	if (likely(!slow_path))
1894 		iqalloc(tsd, ptr, tcache, false);
1895 	else {
1896 		if (config_valgrind && unlikely(in_valgrind))
1897 			rzsize = p2rz(tsd_tsdn(tsd), ptr);
1898 		iqalloc(tsd, ptr, tcache, true);
1899 		JEMALLOC_VALGRIND_FREE(ptr, rzsize);
1900 	}
1901 }
1902 
1903 JEMALLOC_INLINE_C void
1904 isfree(tsd_t *tsd, void *ptr, size_t usize, tcache_t *tcache, bool slow_path)
1905 {
1906 	UNUSED size_t rzsize JEMALLOC_CC_SILENCE_INIT(0);
1907 
1908 	witness_assert_lockless(tsd_tsdn(tsd));
1909 
1910 	assert(ptr != NULL);
1911 	assert(malloc_initialized() || IS_INITIALIZER);
1912 
1913 	if (config_prof && opt_prof)
1914 		prof_free(tsd, ptr, usize);
1915 	if (config_stats)
1916 		*tsd_thread_deallocatedp_get(tsd) += usize;
1917 	if (config_valgrind && unlikely(in_valgrind))
1918 		rzsize = p2rz(tsd_tsdn(tsd), ptr);
1919 	isqalloc(tsd, ptr, usize, tcache, slow_path);
1920 	JEMALLOC_VALGRIND_FREE(ptr, rzsize);
1921 }
1922 
1923 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
1924 void JEMALLOC_NOTHROW *
1925 JEMALLOC_ALLOC_SIZE(2)
1926 je_realloc(void *ptr, size_t size)
1927 {
1928 	void *ret;
1929 	tsdn_t *tsdn JEMALLOC_CC_SILENCE_INIT(NULL);
1930 	size_t usize JEMALLOC_CC_SILENCE_INIT(0);
1931 	size_t old_usize = 0;
1932 	UNUSED size_t old_rzsize JEMALLOC_CC_SILENCE_INIT(0);
1933 
1934 	if (unlikely(size == 0)) {
1935 		if (ptr != NULL) {
1936 			tsd_t *tsd;
1937 
1938 			/* realloc(ptr, 0) is equivalent to free(ptr). */
1939 			UTRACE(ptr, 0, 0);
1940 			tsd = tsd_fetch();
1941 			ifree(tsd, ptr, tcache_get(tsd, false), true);
1942 			return (NULL);
1943 		}
1944 		size = 1;
1945 	}
1946 
1947 	if (likely(ptr != NULL)) {
1948 		tsd_t *tsd;
1949 
1950 		assert(malloc_initialized() || IS_INITIALIZER);
1951 		malloc_thread_init();
1952 		tsd = tsd_fetch();
1953 
1954 		witness_assert_lockless(tsd_tsdn(tsd));
1955 
1956 		old_usize = isalloc(tsd_tsdn(tsd), ptr, config_prof);
1957 		if (config_valgrind && unlikely(in_valgrind)) {
1958 			old_rzsize = config_prof ? p2rz(tsd_tsdn(tsd), ptr) :
1959 			    u2rz(old_usize);
1960 		}
1961 
1962 		if (config_prof && opt_prof) {
1963 			usize = s2u(size);
1964 			ret = unlikely(usize == 0 || usize > HUGE_MAXCLASS) ?
1965 			    NULL : irealloc_prof(tsd, ptr, old_usize, usize);
1966 		} else {
1967 			if (config_stats || (config_valgrind &&
1968 			    unlikely(in_valgrind)))
1969 				usize = s2u(size);
1970 			ret = iralloc(tsd, ptr, old_usize, size, 0, false);
1971 		}
1972 		tsdn = tsd_tsdn(tsd);
1973 	} else {
1974 		/* realloc(NULL, size) is equivalent to malloc(size). */
1975 		if (likely(!malloc_slow))
1976 			ret = ialloc_body(size, false, &tsdn, &usize, false);
1977 		else
1978 			ret = ialloc_body(size, false, &tsdn, &usize, true);
1979 		assert(!tsdn_null(tsdn) || ret == NULL);
1980 	}
1981 
1982 	if (unlikely(ret == NULL)) {
1983 		if (config_xmalloc && unlikely(opt_xmalloc)) {
1984 			malloc_write("<jemalloc>: Error in realloc(): "
1985 			    "out of memory\n");
1986 			abort();
1987 		}
1988 		set_errno(ENOMEM);
1989 	}
1990 	if (config_stats && likely(ret != NULL)) {
1991 		tsd_t *tsd;
1992 
1993 		assert(usize == isalloc(tsdn, ret, config_prof));
1994 		tsd = tsdn_tsd(tsdn);
1995 		*tsd_thread_allocatedp_get(tsd) += usize;
1996 		*tsd_thread_deallocatedp_get(tsd) += old_usize;
1997 	}
1998 	UTRACE(ptr, size, ret);
1999 	JEMALLOC_VALGRIND_REALLOC(maybe, tsdn, ret, usize, maybe, ptr,
2000 	    old_usize, old_rzsize, maybe, false);
2001 	witness_assert_lockless(tsdn);
2002 	return (ret);
2003 }
2004 
2005 JEMALLOC_EXPORT void JEMALLOC_NOTHROW
2006 je_free(void *ptr)
2007 {
2008 
2009 	UTRACE(ptr, 0, 0);
2010 	if (likely(ptr != NULL)) {
2011 		tsd_t *tsd = tsd_fetch();
2012 		witness_assert_lockless(tsd_tsdn(tsd));
2013 		if (likely(!malloc_slow))
2014 			ifree(tsd, ptr, tcache_get(tsd, false), false);
2015 		else
2016 			ifree(tsd, ptr, tcache_get(tsd, false), true);
2017 		witness_assert_lockless(tsd_tsdn(tsd));
2018 	}
2019 }
2020 
2021 /*
2022  * End malloc(3)-compatible functions.
2023  */
2024 /******************************************************************************/
2025 /*
2026  * Begin non-standard override functions.
2027  */
2028 
2029 #ifdef JEMALLOC_OVERRIDE_MEMALIGN
2030 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
2031 void JEMALLOC_NOTHROW *
2032 JEMALLOC_ATTR(malloc)
2033 je_memalign(size_t alignment, size_t size)
2034 {
2035 	void *ret JEMALLOC_CC_SILENCE_INIT(NULL);
2036 	if (unlikely(imemalign(&ret, alignment, size, 1) != 0))
2037 		ret = NULL;
2038 	return (ret);
2039 }
2040 #endif
2041 
2042 #ifdef JEMALLOC_OVERRIDE_VALLOC
2043 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
2044 void JEMALLOC_NOTHROW *
2045 JEMALLOC_ATTR(malloc)
2046 je_valloc(size_t size)
2047 {
2048 	void *ret JEMALLOC_CC_SILENCE_INIT(NULL);
2049 	if (unlikely(imemalign(&ret, PAGE, size, 1) != 0))
2050 		ret = NULL;
2051 	return (ret);
2052 }
2053 #endif
2054 
2055 /*
2056  * is_malloc(je_malloc) is some macro magic to detect if jemalloc_defs.h has
2057  * #define je_malloc malloc
2058  */
2059 #define	malloc_is_malloc 1
2060 #define	is_malloc_(a) malloc_is_ ## a
2061 #define	is_malloc(a) is_malloc_(a)
2062 
2063 #if ((is_malloc(je_malloc) == 1) && defined(JEMALLOC_GLIBC_MALLOC_HOOK))
2064 /*
2065  * glibc provides the RTLD_DEEPBIND flag for dlopen which can make it possible
2066  * to inconsistently reference libc's malloc(3)-compatible functions
2067  * (https://bugzilla.mozilla.org/show_bug.cgi?id=493541).
2068  *
2069  * These definitions interpose hooks in glibc.  The functions are actually
2070  * passed an extra argument for the caller return address, which will be
2071  * ignored.
2072  */
2073 JEMALLOC_EXPORT void (*__free_hook)(void *ptr) = je_free;
2074 JEMALLOC_EXPORT void *(*__malloc_hook)(size_t size) = je_malloc;
2075 JEMALLOC_EXPORT void *(*__realloc_hook)(void *ptr, size_t size) = je_realloc;
2076 # ifdef JEMALLOC_GLIBC_MEMALIGN_HOOK
2077 JEMALLOC_EXPORT void *(*__memalign_hook)(size_t alignment, size_t size) =
2078     je_memalign;
2079 # endif
2080 
2081 #ifdef CPU_COUNT
2082 /*
2083  * To enable static linking with glibc, the libc specific malloc interface must
2084  * be implemented also, so none of glibc's malloc.o functions are added to the
2085  * link.
2086  */
2087 #define	ALIAS(je_fn)	__attribute__((alias (#je_fn), used))
2088 /* To force macro expansion of je_ prefix before stringification. */
2089 #define	PREALIAS(je_fn)  ALIAS(je_fn)
2090 void	*__libc_malloc(size_t size) PREALIAS(je_malloc);
2091 void	__libc_free(void* ptr) PREALIAS(je_free);
2092 void	*__libc_realloc(void* ptr, size_t size) PREALIAS(je_realloc);
2093 void	*__libc_calloc(size_t n, size_t size) PREALIAS(je_calloc);
2094 void	*__libc_memalign(size_t align, size_t s) PREALIAS(je_memalign);
2095 void	*__libc_valloc(size_t size) PREALIAS(je_valloc);
2096 int	__posix_memalign(void** r, size_t a, size_t s)
2097     PREALIAS(je_posix_memalign);
2098 #undef PREALIAS
2099 #undef ALIAS
2100 
2101 #endif
2102 
2103 #endif
2104 
2105 /*
2106  * End non-standard override functions.
2107  */
2108 /******************************************************************************/
2109 /*
2110  * Begin non-standard functions.
2111  */
2112 
2113 JEMALLOC_ALWAYS_INLINE_C bool
2114 imallocx_flags_decode(tsd_t *tsd, size_t size, int flags, size_t *usize,
2115     size_t *alignment, bool *zero, tcache_t **tcache, arena_t **arena)
2116 {
2117 
2118 	if ((flags & MALLOCX_LG_ALIGN_MASK) == 0) {
2119 		*alignment = 0;
2120 		*usize = s2u(size);
2121 	} else {
2122 		*alignment = MALLOCX_ALIGN_GET_SPECIFIED(flags);
2123 		*usize = sa2u(size, *alignment);
2124 	}
2125 	if (unlikely(*usize == 0 || *usize > HUGE_MAXCLASS))
2126 		return (true);
2127 	*zero = MALLOCX_ZERO_GET(flags);
2128 	if ((flags & MALLOCX_TCACHE_MASK) != 0) {
2129 		if ((flags & MALLOCX_TCACHE_MASK) == MALLOCX_TCACHE_NONE)
2130 			*tcache = NULL;
2131 		else
2132 			*tcache = tcaches_get(tsd, MALLOCX_TCACHE_GET(flags));
2133 	} else
2134 		*tcache = tcache_get(tsd, true);
2135 	if ((flags & MALLOCX_ARENA_MASK) != 0) {
2136 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
2137 		*arena = arena_get(tsd_tsdn(tsd), arena_ind, true);
2138 		if (unlikely(*arena == NULL))
2139 			return (true);
2140 	} else
2141 		*arena = NULL;
2142 	return (false);
2143 }
2144 
2145 JEMALLOC_ALWAYS_INLINE_C void *
2146 imallocx_flags(tsdn_t *tsdn, size_t usize, size_t alignment, bool zero,
2147     tcache_t *tcache, arena_t *arena, bool slow_path)
2148 {
2149 	szind_t ind;
2150 
2151 	if (unlikely(alignment != 0))
2152 		return (ipalloct(tsdn, usize, alignment, zero, tcache, arena));
2153 	ind = size2index(usize);
2154 	assert(ind < NSIZES);
2155 	return (iallocztm(tsdn, usize, ind, zero, tcache, false, arena,
2156 	    slow_path));
2157 }
2158 
2159 static void *
2160 imallocx_prof_sample(tsdn_t *tsdn, size_t usize, size_t alignment, bool zero,
2161     tcache_t *tcache, arena_t *arena, bool slow_path)
2162 {
2163 	void *p;
2164 
2165 	if (usize <= SMALL_MAXCLASS) {
2166 		assert(((alignment == 0) ? s2u(LARGE_MINCLASS) :
2167 		    sa2u(LARGE_MINCLASS, alignment)) == LARGE_MINCLASS);
2168 		p = imallocx_flags(tsdn, LARGE_MINCLASS, alignment, zero,
2169 		    tcache, arena, slow_path);
2170 		if (p == NULL)
2171 			return (NULL);
2172 		arena_prof_promoted(tsdn, p, usize);
2173 	} else {
2174 		p = imallocx_flags(tsdn, usize, alignment, zero, tcache, arena,
2175 		    slow_path);
2176 	}
2177 
2178 	return (p);
2179 }
2180 
2181 JEMALLOC_ALWAYS_INLINE_C void *
2182 imallocx_prof(tsd_t *tsd, size_t size, int flags, size_t *usize, bool slow_path)
2183 {
2184 	void *p;
2185 	size_t alignment;
2186 	bool zero;
2187 	tcache_t *tcache;
2188 	arena_t *arena;
2189 	prof_tctx_t *tctx;
2190 
2191 	if (unlikely(imallocx_flags_decode(tsd, size, flags, usize, &alignment,
2192 	    &zero, &tcache, &arena)))
2193 		return (NULL);
2194 	tctx = prof_alloc_prep(tsd, *usize, prof_active_get_unlocked(), true);
2195 	if (likely((uintptr_t)tctx == (uintptr_t)1U)) {
2196 		p = imallocx_flags(tsd_tsdn(tsd), *usize, alignment, zero,
2197 		    tcache, arena, slow_path);
2198 	} else if ((uintptr_t)tctx > (uintptr_t)1U) {
2199 		p = imallocx_prof_sample(tsd_tsdn(tsd), *usize, alignment, zero,
2200 		    tcache, arena, slow_path);
2201 	} else
2202 		p = NULL;
2203 	if (unlikely(p == NULL)) {
2204 		prof_alloc_rollback(tsd, tctx, true);
2205 		return (NULL);
2206 	}
2207 	prof_malloc(tsd_tsdn(tsd), p, *usize, tctx);
2208 
2209 	assert(alignment == 0 || ((uintptr_t)p & (alignment - 1)) == ZU(0));
2210 	return (p);
2211 }
2212 
2213 JEMALLOC_ALWAYS_INLINE_C void *
2214 imallocx_no_prof(tsd_t *tsd, size_t size, int flags, size_t *usize,
2215     bool slow_path)
2216 {
2217 	void *p;
2218 	size_t alignment;
2219 	bool zero;
2220 	tcache_t *tcache;
2221 	arena_t *arena;
2222 
2223 	if (unlikely(imallocx_flags_decode(tsd, size, flags, usize, &alignment,
2224 	    &zero, &tcache, &arena)))
2225 		return (NULL);
2226 	p = imallocx_flags(tsd_tsdn(tsd), *usize, alignment, zero, tcache,
2227 	    arena, slow_path);
2228 	assert(alignment == 0 || ((uintptr_t)p & (alignment - 1)) == ZU(0));
2229 	return (p);
2230 }
2231 
2232 /* This function guarantees that *tsdn is non-NULL on success. */
2233 JEMALLOC_ALWAYS_INLINE_C void *
2234 imallocx_body(size_t size, int flags, tsdn_t **tsdn, size_t *usize,
2235     bool slow_path)
2236 {
2237 	tsd_t *tsd;
2238 
2239 	if (slow_path && unlikely(malloc_init())) {
2240 		*tsdn = NULL;
2241 		return (NULL);
2242 	}
2243 
2244 	tsd = tsd_fetch();
2245 	*tsdn = tsd_tsdn(tsd);
2246 	witness_assert_lockless(tsd_tsdn(tsd));
2247 
2248 	if (likely(flags == 0)) {
2249 		szind_t ind = size2index(size);
2250 		if (unlikely(ind >= NSIZES))
2251 			return (NULL);
2252 		if (config_stats || (config_prof && opt_prof) || (slow_path &&
2253 		    config_valgrind && unlikely(in_valgrind))) {
2254 			*usize = index2size(ind);
2255 			assert(*usize > 0 && *usize <= HUGE_MAXCLASS);
2256 		}
2257 
2258 		if (config_prof && opt_prof) {
2259 			return (ialloc_prof(tsd, *usize, ind, false,
2260 			    slow_path));
2261 		}
2262 
2263 		return (ialloc(tsd, size, ind, false, slow_path));
2264 	}
2265 
2266 	if (config_prof && opt_prof)
2267 		return (imallocx_prof(tsd, size, flags, usize, slow_path));
2268 
2269 	return (imallocx_no_prof(tsd, size, flags, usize, slow_path));
2270 }
2271 
2272 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
2273 void JEMALLOC_NOTHROW *
2274 JEMALLOC_ATTR(malloc) JEMALLOC_ALLOC_SIZE(1)
2275 je_mallocx(size_t size, int flags)
2276 {
2277 	tsdn_t *tsdn;
2278 	void *p;
2279 	size_t usize;
2280 
2281 	assert(size != 0);
2282 
2283 	if (likely(!malloc_slow)) {
2284 		p = imallocx_body(size, flags, &tsdn, &usize, false);
2285 		ialloc_post_check(p, tsdn, usize, "mallocx", false, false);
2286 	} else {
2287 		p = imallocx_body(size, flags, &tsdn, &usize, true);
2288 		ialloc_post_check(p, tsdn, usize, "mallocx", false, true);
2289 		UTRACE(0, size, p);
2290 		JEMALLOC_VALGRIND_MALLOC(p != NULL, tsdn, p, usize,
2291 		    MALLOCX_ZERO_GET(flags));
2292 	}
2293 
2294 	return (p);
2295 }
2296 
2297 static void *
2298 irallocx_prof_sample(tsd_t *tsd, void *old_ptr, size_t old_usize,
2299     size_t usize, size_t alignment, bool zero, tcache_t *tcache, arena_t *arena,
2300     prof_tctx_t *tctx)
2301 {
2302 	void *p;
2303 
2304 	if (tctx == NULL)
2305 		return (NULL);
2306 	if (usize <= SMALL_MAXCLASS) {
2307 		p = iralloct(tsd, old_ptr, old_usize, LARGE_MINCLASS, alignment,
2308 		    zero, tcache, arena);
2309 		if (p == NULL)
2310 			return (NULL);
2311 		arena_prof_promoted(tsd_tsdn(tsd), p, usize);
2312 	} else {
2313 		p = iralloct(tsd, old_ptr, old_usize, usize, alignment, zero,
2314 		    tcache, arena);
2315 	}
2316 
2317 	return (p);
2318 }
2319 
2320 JEMALLOC_ALWAYS_INLINE_C void *
2321 irallocx_prof(tsd_t *tsd, void *old_ptr, size_t old_usize, size_t size,
2322     size_t alignment, size_t *usize, bool zero, tcache_t *tcache,
2323     arena_t *arena)
2324 {
2325 	void *p;
2326 	bool prof_active;
2327 	prof_tctx_t *old_tctx, *tctx;
2328 
2329 	prof_active = prof_active_get_unlocked();
2330 	old_tctx = prof_tctx_get(tsd_tsdn(tsd), old_ptr);
2331 	tctx = prof_alloc_prep(tsd, *usize, prof_active, false);
2332 	if (unlikely((uintptr_t)tctx != (uintptr_t)1U)) {
2333 		p = irallocx_prof_sample(tsd, old_ptr, old_usize, *usize,
2334 		    alignment, zero, tcache, arena, tctx);
2335 	} else {
2336 		p = iralloct(tsd, old_ptr, old_usize, size, alignment, zero,
2337 		    tcache, arena);
2338 	}
2339 	if (unlikely(p == NULL)) {
2340 		prof_alloc_rollback(tsd, tctx, false);
2341 		return (NULL);
2342 	}
2343 
2344 	if (p == old_ptr && alignment != 0) {
2345 		/*
2346 		 * The allocation did not move, so it is possible that the size
2347 		 * class is smaller than would guarantee the requested
2348 		 * alignment, and that the alignment constraint was
2349 		 * serendipitously satisfied.  Additionally, old_usize may not
2350 		 * be the same as the current usize because of in-place large
2351 		 * reallocation.  Therefore, query the actual value of usize.
2352 		 */
2353 		*usize = isalloc(tsd_tsdn(tsd), p, config_prof);
2354 	}
2355 	prof_realloc(tsd, p, *usize, tctx, prof_active, false, old_ptr,
2356 	    old_usize, old_tctx);
2357 
2358 	return (p);
2359 }
2360 
2361 JEMALLOC_EXPORT JEMALLOC_ALLOCATOR JEMALLOC_RESTRICT_RETURN
2362 void JEMALLOC_NOTHROW *
2363 JEMALLOC_ALLOC_SIZE(2)
2364 je_rallocx(void *ptr, size_t size, int flags)
2365 {
2366 	void *p;
2367 	tsd_t *tsd;
2368 	size_t usize;
2369 	size_t old_usize;
2370 	UNUSED size_t old_rzsize JEMALLOC_CC_SILENCE_INIT(0);
2371 	size_t alignment = MALLOCX_ALIGN_GET(flags);
2372 	bool zero = flags & MALLOCX_ZERO;
2373 	arena_t *arena;
2374 	tcache_t *tcache;
2375 
2376 	assert(ptr != NULL);
2377 	assert(size != 0);
2378 	assert(malloc_initialized() || IS_INITIALIZER);
2379 	malloc_thread_init();
2380 	tsd = tsd_fetch();
2381 	witness_assert_lockless(tsd_tsdn(tsd));
2382 
2383 	if (unlikely((flags & MALLOCX_ARENA_MASK) != 0)) {
2384 		unsigned arena_ind = MALLOCX_ARENA_GET(flags);
2385 		arena = arena_get(tsd_tsdn(tsd), arena_ind, true);
2386 		if (unlikely(arena == NULL))
2387 			goto label_oom;
2388 	} else
2389 		arena = NULL;
2390 
2391 	if (unlikely((flags & MALLOCX_TCACHE_MASK) != 0)) {
2392 		if ((flags & MALLOCX_TCACHE_MASK) == MALLOCX_TCACHE_NONE)
2393 			tcache = NULL;
2394 		else
2395 			tcache = tcaches_get(tsd, MALLOCX_TCACHE_GET(flags));
2396 	} else
2397 		tcache = tcache_get(tsd, true);
2398 
2399 	old_usize = isalloc(tsd_tsdn(tsd), ptr, config_prof);
2400 	if (config_valgrind && unlikely(in_valgrind))
2401 		old_rzsize = u2rz(old_usize);
2402 
2403 	if (config_prof && opt_prof) {
2404 		usize = (alignment == 0) ? s2u(size) : sa2u(size, alignment);
2405 		if (unlikely(usize == 0 || usize > HUGE_MAXCLASS))
2406 			goto label_oom;
2407 		p = irallocx_prof(tsd, ptr, old_usize, size, alignment, &usize,
2408 		    zero, tcache, arena);
2409 		if (unlikely(p == NULL))
2410 			goto label_oom;
2411 	} else {
2412 		p = iralloct(tsd, ptr, old_usize, size, alignment, zero,
2413 		     tcache, arena);
2414 		if (unlikely(p == NULL))
2415 			goto label_oom;
2416 		if (config_stats || (config_valgrind && unlikely(in_valgrind)))
2417 			usize = isalloc(tsd_tsdn(tsd), p, config_prof);
2418 	}
2419 	assert(alignment == 0 || ((uintptr_t)p & (alignment - 1)) == ZU(0));
2420 
2421 	if (config_stats) {
2422 		*tsd_thread_allocatedp_get(tsd) += usize;
2423 		*tsd_thread_deallocatedp_get(tsd) += old_usize;
2424 	}
2425 	UTRACE(ptr, size, p);
2426 	JEMALLOC_VALGRIND_REALLOC(maybe, tsd_tsdn(tsd), p, usize, no, ptr,
2427 	    old_usize, old_rzsize, no, zero);
2428 	witness_assert_lockless(tsd_tsdn(tsd));
2429 	return (p);
2430 label_oom:
2431 	if (config_xmalloc && unlikely(opt_xmalloc)) {
2432 		malloc_write("<jemalloc>: Error in rallocx(): out of memory\n");
2433 		abort();
2434 	}
2435 	UTRACE(ptr, size, 0);
2436 	witness_assert_lockless(tsd_tsdn(tsd));
2437 	return (NULL);
2438 }
2439 
2440 JEMALLOC_ALWAYS_INLINE_C size_t
2441 ixallocx_helper(tsdn_t *tsdn, void *ptr, size_t old_usize, size_t size,
2442     size_t extra, size_t alignment, bool zero)
2443 {
2444 	size_t usize;
2445 
2446 	if (ixalloc(tsdn, ptr, old_usize, size, extra, alignment, zero))
2447 		return (old_usize);
2448 	usize = isalloc(tsdn, ptr, config_prof);
2449 
2450 	return (usize);
2451 }
2452 
2453 static size_t
2454 ixallocx_prof_sample(tsdn_t *tsdn, void *ptr, size_t old_usize, size_t size,
2455     size_t extra, size_t alignment, bool zero, prof_tctx_t *tctx)
2456 {
2457 	size_t usize;
2458 
2459 	if (tctx == NULL)
2460 		return (old_usize);
2461 	usize = ixallocx_helper(tsdn, ptr, old_usize, size, extra, alignment,
2462 	    zero);
2463 
2464 	return (usize);
2465 }
2466 
2467 JEMALLOC_ALWAYS_INLINE_C size_t
2468 ixallocx_prof(tsd_t *tsd, void *ptr, size_t old_usize, size_t size,
2469     size_t extra, size_t alignment, bool zero)
2470 {
2471 	size_t usize_max, usize;
2472 	bool prof_active;
2473 	prof_tctx_t *old_tctx, *tctx;
2474 
2475 	prof_active = prof_active_get_unlocked();
2476 	old_tctx = prof_tctx_get(tsd_tsdn(tsd), ptr);
2477 	/*
2478 	 * usize isn't knowable before ixalloc() returns when extra is non-zero.
2479 	 * Therefore, compute its maximum possible value and use that in
2480 	 * prof_alloc_prep() to decide whether to capture a backtrace.
2481 	 * prof_realloc() will use the actual usize to decide whether to sample.
2482 	 */
2483 	if (alignment == 0) {
2484 		usize_max = s2u(size+extra);
2485 		assert(usize_max > 0 && usize_max <= HUGE_MAXCLASS);
2486 	} else {
2487 		usize_max = sa2u(size+extra, alignment);
2488 		if (unlikely(usize_max == 0 || usize_max > HUGE_MAXCLASS)) {
2489 			/*
2490 			 * usize_max is out of range, and chances are that
2491 			 * allocation will fail, but use the maximum possible
2492 			 * value and carry on with prof_alloc_prep(), just in
2493 			 * case allocation succeeds.
2494 			 */
2495 			usize_max = HUGE_MAXCLASS;
2496 		}
2497 	}
2498 	tctx = prof_alloc_prep(tsd, usize_max, prof_active, false);
2499 
2500 	if (unlikely((uintptr_t)tctx != (uintptr_t)1U)) {
2501 		usize = ixallocx_prof_sample(tsd_tsdn(tsd), ptr, old_usize,
2502 		    size, extra, alignment, zero, tctx);
2503 	} else {
2504 		usize = ixallocx_helper(tsd_tsdn(tsd), ptr, old_usize, size,
2505 		    extra, alignment, zero);
2506 	}
2507 	if (usize == old_usize) {
2508 		prof_alloc_rollback(tsd, tctx, false);
2509 		return (usize);
2510 	}
2511 	prof_realloc(tsd, ptr, usize, tctx, prof_active, false, ptr, old_usize,
2512 	    old_tctx);
2513 
2514 	return (usize);
2515 }
2516 
2517 JEMALLOC_EXPORT size_t JEMALLOC_NOTHROW
2518 je_xallocx(void *ptr, size_t size, size_t extra, int flags)
2519 {
2520 	tsd_t *tsd;
2521 	size_t usize, old_usize;
2522 	UNUSED size_t old_rzsize JEMALLOC_CC_SILENCE_INIT(0);
2523 	size_t alignment = MALLOCX_ALIGN_GET(flags);
2524 	bool zero = flags & MALLOCX_ZERO;
2525 
2526 	assert(ptr != NULL);
2527 	assert(size != 0);
2528 	assert(SIZE_T_MAX - size >= extra);
2529 	assert(malloc_initialized() || IS_INITIALIZER);
2530 	malloc_thread_init();
2531 	tsd = tsd_fetch();
2532 	witness_assert_lockless(tsd_tsdn(tsd));
2533 
2534 	old_usize = isalloc(tsd_tsdn(tsd), ptr, config_prof);
2535 
2536 	/*
2537 	 * The API explicitly absolves itself of protecting against (size +
2538 	 * extra) numerical overflow, but we may need to clamp extra to avoid
2539 	 * exceeding HUGE_MAXCLASS.
2540 	 *
2541 	 * Ordinarily, size limit checking is handled deeper down, but here we
2542 	 * have to check as part of (size + extra) clamping, since we need the
2543 	 * clamped value in the above helper functions.
2544 	 */
2545 	if (unlikely(size > HUGE_MAXCLASS)) {
2546 		usize = old_usize;
2547 		goto label_not_resized;
2548 	}
2549 	if (unlikely(HUGE_MAXCLASS - size < extra))
2550 		extra = HUGE_MAXCLASS - size;
2551 
2552 	if (config_valgrind && unlikely(in_valgrind))
2553 		old_rzsize = u2rz(old_usize);
2554 
2555 	if (config_prof && opt_prof) {
2556 		usize = ixallocx_prof(tsd, ptr, old_usize, size, extra,
2557 		    alignment, zero);
2558 	} else {
2559 		usize = ixallocx_helper(tsd_tsdn(tsd), ptr, old_usize, size,
2560 		    extra, alignment, zero);
2561 	}
2562 	if (unlikely(usize == old_usize))
2563 		goto label_not_resized;
2564 
2565 	if (config_stats) {
2566 		*tsd_thread_allocatedp_get(tsd) += usize;
2567 		*tsd_thread_deallocatedp_get(tsd) += old_usize;
2568 	}
2569 	JEMALLOC_VALGRIND_REALLOC(no, tsd_tsdn(tsd), ptr, usize, no, ptr,
2570 	    old_usize, old_rzsize, no, zero);
2571 label_not_resized:
2572 	UTRACE(ptr, size, ptr);
2573 	witness_assert_lockless(tsd_tsdn(tsd));
2574 	return (usize);
2575 }
2576 
2577 JEMALLOC_EXPORT size_t JEMALLOC_NOTHROW
2578 JEMALLOC_ATTR(pure)
2579 je_sallocx(const void *ptr, int flags)
2580 {
2581 	size_t usize;
2582 	tsdn_t *tsdn;
2583 
2584 	assert(malloc_initialized() || IS_INITIALIZER);
2585 	malloc_thread_init();
2586 
2587 	tsdn = tsdn_fetch();
2588 	witness_assert_lockless(tsdn);
2589 
2590 	if (config_ivsalloc)
2591 		usize = ivsalloc(tsdn, ptr, config_prof);
2592 	else
2593 		usize = isalloc(tsdn, ptr, config_prof);
2594 
2595 	witness_assert_lockless(tsdn);
2596 	return (usize);
2597 }
2598 
2599 JEMALLOC_EXPORT void JEMALLOC_NOTHROW
2600 je_dallocx(void *ptr, int flags)
2601 {
2602 	tsd_t *tsd;
2603 	tcache_t *tcache;
2604 
2605 	assert(ptr != NULL);
2606 	assert(malloc_initialized() || IS_INITIALIZER);
2607 
2608 	tsd = tsd_fetch();
2609 	witness_assert_lockless(tsd_tsdn(tsd));
2610 	if (unlikely((flags & MALLOCX_TCACHE_MASK) != 0)) {
2611 		if ((flags & MALLOCX_TCACHE_MASK) == MALLOCX_TCACHE_NONE)
2612 			tcache = NULL;
2613 		else
2614 			tcache = tcaches_get(tsd, MALLOCX_TCACHE_GET(flags));
2615 	} else
2616 		tcache = tcache_get(tsd, false);
2617 
2618 	UTRACE(ptr, 0, 0);
2619 	if (likely(!malloc_slow))
2620 		ifree(tsd, ptr, tcache, false);
2621 	else
2622 		ifree(tsd, ptr, tcache, true);
2623 	witness_assert_lockless(tsd_tsdn(tsd));
2624 }
2625 
2626 JEMALLOC_ALWAYS_INLINE_C size_t
2627 inallocx(tsdn_t *tsdn, size_t size, int flags)
2628 {
2629 	size_t usize;
2630 
2631 	witness_assert_lockless(tsdn);
2632 
2633 	if (likely((flags & MALLOCX_LG_ALIGN_MASK) == 0))
2634 		usize = s2u(size);
2635 	else
2636 		usize = sa2u(size, MALLOCX_ALIGN_GET_SPECIFIED(flags));
2637 	witness_assert_lockless(tsdn);
2638 	return (usize);
2639 }
2640 
2641 JEMALLOC_EXPORT void JEMALLOC_NOTHROW
2642 je_sdallocx(void *ptr, size_t size, int flags)
2643 {
2644 	tsd_t *tsd;
2645 	tcache_t *tcache;
2646 	size_t usize;
2647 
2648 	assert(ptr != NULL);
2649 	assert(malloc_initialized() || IS_INITIALIZER);
2650 	tsd = tsd_fetch();
2651 	usize = inallocx(tsd_tsdn(tsd), size, flags);
2652 	assert(usize == isalloc(tsd_tsdn(tsd), ptr, config_prof));
2653 
2654 	witness_assert_lockless(tsd_tsdn(tsd));
2655 	if (unlikely((flags & MALLOCX_TCACHE_MASK) != 0)) {
2656 		if ((flags & MALLOCX_TCACHE_MASK) == MALLOCX_TCACHE_NONE)
2657 			tcache = NULL;
2658 		else
2659 			tcache = tcaches_get(tsd, MALLOCX_TCACHE_GET(flags));
2660 	} else
2661 		tcache = tcache_get(tsd, false);
2662 
2663 	UTRACE(ptr, 0, 0);
2664 	if (likely(!malloc_slow))
2665 		isfree(tsd, ptr, usize, tcache, false);
2666 	else
2667 		isfree(tsd, ptr, usize, tcache, true);
2668 	witness_assert_lockless(tsd_tsdn(tsd));
2669 }
2670 
2671 JEMALLOC_EXPORT size_t JEMALLOC_NOTHROW
2672 JEMALLOC_ATTR(pure)
2673 je_nallocx(size_t size, int flags)
2674 {
2675 	size_t usize;
2676 	tsdn_t *tsdn;
2677 
2678 	assert(size != 0);
2679 
2680 	if (unlikely(malloc_init()))
2681 		return (0);
2682 
2683 	tsdn = tsdn_fetch();
2684 	witness_assert_lockless(tsdn);
2685 
2686 	usize = inallocx(tsdn, size, flags);
2687 	if (unlikely(usize > HUGE_MAXCLASS))
2688 		return (0);
2689 
2690 	witness_assert_lockless(tsdn);
2691 	return (usize);
2692 }
2693 
2694 JEMALLOC_EXPORT int JEMALLOC_NOTHROW
2695 je_mallctl(const char *name, void *oldp, size_t *oldlenp, void *newp,
2696     size_t newlen)
2697 {
2698 	int ret;
2699 	tsd_t *tsd;
2700 
2701 	if (unlikely(malloc_init()))
2702 		return (EAGAIN);
2703 
2704 	tsd = tsd_fetch();
2705 	witness_assert_lockless(tsd_tsdn(tsd));
2706 	ret = ctl_byname(tsd, name, oldp, oldlenp, newp, newlen);
2707 	witness_assert_lockless(tsd_tsdn(tsd));
2708 	return (ret);
2709 }
2710 
2711 JEMALLOC_EXPORT int JEMALLOC_NOTHROW
2712 je_mallctlnametomib(const char *name, size_t *mibp, size_t *miblenp)
2713 {
2714 	int ret;
2715 	tsdn_t *tsdn;
2716 
2717 	if (unlikely(malloc_init()))
2718 		return (EAGAIN);
2719 
2720 	tsdn = tsdn_fetch();
2721 	witness_assert_lockless(tsdn);
2722 	ret = ctl_nametomib(tsdn, name, mibp, miblenp);
2723 	witness_assert_lockless(tsdn);
2724 	return (ret);
2725 }
2726 
2727 JEMALLOC_EXPORT int JEMALLOC_NOTHROW
2728 je_mallctlbymib(const size_t *mib, size_t miblen, void *oldp, size_t *oldlenp,
2729   void *newp, size_t newlen)
2730 {
2731 	int ret;
2732 	tsd_t *tsd;
2733 
2734 	if (unlikely(malloc_init()))
2735 		return (EAGAIN);
2736 
2737 	tsd = tsd_fetch();
2738 	witness_assert_lockless(tsd_tsdn(tsd));
2739 	ret = ctl_bymib(tsd, mib, miblen, oldp, oldlenp, newp, newlen);
2740 	witness_assert_lockless(tsd_tsdn(tsd));
2741 	return (ret);
2742 }
2743 
2744 JEMALLOC_EXPORT void JEMALLOC_NOTHROW
2745 je_malloc_stats_print(void (*write_cb)(void *, const char *), void *cbopaque,
2746     const char *opts)
2747 {
2748 	tsdn_t *tsdn;
2749 
2750 	tsdn = tsdn_fetch();
2751 	witness_assert_lockless(tsdn);
2752 	stats_print(write_cb, cbopaque, opts);
2753 	witness_assert_lockless(tsdn);
2754 }
2755 
2756 JEMALLOC_EXPORT size_t JEMALLOC_NOTHROW
2757 je_malloc_usable_size(JEMALLOC_USABLE_SIZE_CONST void *ptr)
2758 {
2759 	size_t ret;
2760 	tsdn_t *tsdn;
2761 
2762 	assert(malloc_initialized() || IS_INITIALIZER);
2763 	malloc_thread_init();
2764 
2765 	tsdn = tsdn_fetch();
2766 	witness_assert_lockless(tsdn);
2767 
2768 	if (config_ivsalloc)
2769 		ret = ivsalloc(tsdn, ptr, config_prof);
2770 	else
2771 		ret = (ptr == NULL) ? 0 : isalloc(tsdn, ptr, config_prof);
2772 
2773 	witness_assert_lockless(tsdn);
2774 	return (ret);
2775 }
2776 
2777 /*
2778  * End non-standard functions.
2779  */
2780 /******************************************************************************/
2781 /*
2782  * Begin compatibility functions.
2783  */
2784 
2785 #define	ALLOCM_LG_ALIGN(la)	(la)
2786 #define	ALLOCM_ALIGN(a)		(ffsl(a)-1)
2787 #define	ALLOCM_ZERO		((int)0x40)
2788 #define	ALLOCM_NO_MOVE		((int)0x80)
2789 
2790 #define	ALLOCM_SUCCESS		0
2791 #define	ALLOCM_ERR_OOM		1
2792 #define	ALLOCM_ERR_NOT_MOVED	2
2793 
2794 int
2795 je_allocm(void **ptr, size_t *rsize, size_t size, int flags)
2796 {
2797 	void *p;
2798 
2799 	assert(ptr != NULL);
2800 
2801 	p = je_mallocx(size, flags);
2802 	if (p == NULL)
2803 		return (ALLOCM_ERR_OOM);
2804 	if (rsize != NULL)
2805 		*rsize = isalloc(tsdn_fetch(), p, config_prof);
2806 	*ptr = p;
2807 	return (ALLOCM_SUCCESS);
2808 }
2809 
2810 int
2811 je_rallocm(void **ptr, size_t *rsize, size_t size, size_t extra, int flags)
2812 {
2813 	int ret;
2814 	bool no_move = flags & ALLOCM_NO_MOVE;
2815 
2816 	assert(ptr != NULL);
2817 	assert(*ptr != NULL);
2818 	assert(size != 0);
2819 	assert(SIZE_T_MAX - size >= extra);
2820 
2821 	if (no_move) {
2822 		size_t usize = je_xallocx(*ptr, size, extra, flags);
2823 		ret = (usize >= size) ? ALLOCM_SUCCESS : ALLOCM_ERR_NOT_MOVED;
2824 		if (rsize != NULL)
2825 			*rsize = usize;
2826 	} else {
2827 		void *p = je_rallocx(*ptr, size+extra, flags);
2828 		if (p != NULL) {
2829 			*ptr = p;
2830 			ret = ALLOCM_SUCCESS;
2831 		} else
2832 			ret = ALLOCM_ERR_OOM;
2833 		if (rsize != NULL)
2834 			*rsize = isalloc(tsdn_fetch(), *ptr, config_prof);
2835 	}
2836 	return (ret);
2837 }
2838 
2839 int
2840 je_sallocm(const void *ptr, size_t *rsize, int flags)
2841 {
2842 
2843 	assert(rsize != NULL);
2844 	*rsize = je_sallocx(ptr, flags);
2845 	return (ALLOCM_SUCCESS);
2846 }
2847 
2848 int
2849 je_dallocm(void *ptr, int flags)
2850 {
2851 
2852 	je_dallocx(ptr, flags);
2853 	return (ALLOCM_SUCCESS);
2854 }
2855 
2856 int
2857 je_nallocm(size_t *rsize, size_t size, int flags)
2858 {
2859 	size_t usize;
2860 
2861 	usize = je_nallocx(size, flags);
2862 	if (usize == 0)
2863 		return (ALLOCM_ERR_OOM);
2864 	if (rsize != NULL)
2865 		*rsize = usize;
2866 	return (ALLOCM_SUCCESS);
2867 }
2868 
2869 #undef ALLOCM_LG_ALIGN
2870 #undef ALLOCM_ALIGN
2871 #undef ALLOCM_ZERO
2872 #undef ALLOCM_NO_MOVE
2873 
2874 #undef ALLOCM_SUCCESS
2875 #undef ALLOCM_ERR_OOM
2876 #undef ALLOCM_ERR_NOT_MOVED
2877 
2878 /*
2879  * End compatibility functions.
2880  */
2881 /******************************************************************************/
2882 /*
2883  * The following functions are used by threading libraries for protection of
2884  * malloc during fork().
2885  */
2886 
2887 /*
2888  * If an application creates a thread before doing any allocation in the main
2889  * thread, then calls fork(2) in the main thread followed by memory allocation
2890  * in the child process, a race can occur that results in deadlock within the
2891  * child: the main thread may have forked while the created thread had
2892  * partially initialized the allocator.  Ordinarily jemalloc prevents
2893  * fork/malloc races via the following functions it registers during
2894  * initialization using pthread_atfork(), but of course that does no good if
2895  * the allocator isn't fully initialized at fork time.  The following library
2896  * constructor is a partial solution to this problem.  It may still be possible
2897  * to trigger the deadlock described above, but doing so would involve forking
2898  * via a library constructor that runs before jemalloc's runs.
2899  */
2900 #ifndef JEMALLOC_JET
2901 JEMALLOC_ATTR(constructor)
2902 static void
2903 jemalloc_constructor(void)
2904 {
2905 
2906 	malloc_init();
2907 }
2908 #endif
2909 
2910 #ifndef JEMALLOC_MUTEX_INIT_CB
2911 void
2912 jemalloc_prefork(void)
2913 #else
2914 JEMALLOC_EXPORT void
2915 _malloc_prefork(void)
2916 #endif
2917 {
2918 	tsd_t *tsd;
2919 	unsigned i, j, narenas;
2920 	arena_t *arena;
2921 
2922 #ifdef JEMALLOC_MUTEX_INIT_CB
2923 	if (!malloc_initialized())
2924 		return;
2925 #endif
2926 	assert(malloc_initialized());
2927 
2928 	tsd = tsd_fetch();
2929 
2930 	narenas = narenas_total_get();
2931 
2932 	witness_prefork(tsd);
2933 	/* Acquire all mutexes in a safe order. */
2934 	ctl_prefork(tsd_tsdn(tsd));
2935 	malloc_mutex_prefork(tsd_tsdn(tsd), &arenas_lock);
2936 	prof_prefork0(tsd_tsdn(tsd));
2937 	for (i = 0; i < 3; i++) {
2938 		for (j = 0; j < narenas; j++) {
2939 			if ((arena = arena_get(tsd_tsdn(tsd), j, false)) !=
2940 			    NULL) {
2941 				switch (i) {
2942 				case 0:
2943 					arena_prefork0(tsd_tsdn(tsd), arena);
2944 					break;
2945 				case 1:
2946 					arena_prefork1(tsd_tsdn(tsd), arena);
2947 					break;
2948 				case 2:
2949 					arena_prefork2(tsd_tsdn(tsd), arena);
2950 					break;
2951 				default: not_reached();
2952 				}
2953 			}
2954 		}
2955 	}
2956 	base_prefork(tsd_tsdn(tsd));
2957 	for (i = 0; i < narenas; i++) {
2958 		if ((arena = arena_get(tsd_tsdn(tsd), i, false)) != NULL)
2959 			arena_prefork3(tsd_tsdn(tsd), arena);
2960 	}
2961 	prof_prefork1(tsd_tsdn(tsd));
2962 }
2963 
2964 #ifndef JEMALLOC_MUTEX_INIT_CB
2965 void
2966 jemalloc_postfork_parent(void)
2967 #else
2968 JEMALLOC_EXPORT void
2969 _malloc_postfork(void)
2970 #endif
2971 {
2972 	tsd_t *tsd;
2973 	unsigned i, narenas;
2974 
2975 #ifdef JEMALLOC_MUTEX_INIT_CB
2976 	if (!malloc_initialized())
2977 		return;
2978 #endif
2979 	assert(malloc_initialized());
2980 
2981 	tsd = tsd_fetch();
2982 
2983 	witness_postfork_parent(tsd);
2984 	/* Release all mutexes, now that fork() has completed. */
2985 	base_postfork_parent(tsd_tsdn(tsd));
2986 	for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
2987 		arena_t *arena;
2988 
2989 		if ((arena = arena_get(tsd_tsdn(tsd), i, false)) != NULL)
2990 			arena_postfork_parent(tsd_tsdn(tsd), arena);
2991 	}
2992 	prof_postfork_parent(tsd_tsdn(tsd));
2993 	malloc_mutex_postfork_parent(tsd_tsdn(tsd), &arenas_lock);
2994 	ctl_postfork_parent(tsd_tsdn(tsd));
2995 }
2996 
2997 void
2998 jemalloc_postfork_child(void)
2999 {
3000 	tsd_t *tsd;
3001 	unsigned i, narenas;
3002 
3003 	assert(malloc_initialized());
3004 
3005 	tsd = tsd_fetch();
3006 
3007 	witness_postfork_child(tsd);
3008 	/* Release all mutexes, now that fork() has completed. */
3009 	base_postfork_child(tsd_tsdn(tsd));
3010 	for (i = 0, narenas = narenas_total_get(); i < narenas; i++) {
3011 		arena_t *arena;
3012 
3013 		if ((arena = arena_get(tsd_tsdn(tsd), i, false)) != NULL)
3014 			arena_postfork_child(tsd_tsdn(tsd), arena);
3015 	}
3016 	prof_postfork_child(tsd_tsdn(tsd));
3017 	malloc_mutex_postfork_child(tsd_tsdn(tsd), &arenas_lock);
3018 	ctl_postfork_child(tsd_tsdn(tsd));
3019 }
3020 
3021 void
3022 _malloc_first_thread(void)
3023 {
3024 
3025 	(void)malloc_mutex_first_thread();
3026 }
3027 
3028 /******************************************************************************/
3029