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