1#define JEMALLOC_PROF_C_
2#include "jemalloc/internal/jemalloc_internal.h"
3/******************************************************************************/
4
5#ifdef JEMALLOC_PROF_LIBUNWIND
6#define UNW_LOCAL_ONLY
7#include <libunwind.h>
8#endif
9
10#ifdef JEMALLOC_PROF_LIBGCC
11#include <unwind.h>
12#endif
13
14/******************************************************************************/
15/* Data. */
16
17bool opt_prof = false;
18bool opt_prof_active = true;
19bool opt_prof_thread_active_init = true;
20size_t opt_lg_prof_sample = LG_PROF_SAMPLE_DEFAULT;
21ssize_t opt_lg_prof_interval = LG_PROF_INTERVAL_DEFAULT;
22bool opt_prof_gdump = false;
23bool opt_prof_final = false;
24bool opt_prof_leak = false;
25bool opt_prof_accum = false;
26char opt_prof_prefix[
27 /* Minimize memory bloat for non-prof builds. */
28#ifdef JEMALLOC_PROF
29 PATH_MAX +
30#endif
31 1];
32
33/*
34 * Initialized as opt_prof_active, and accessed via
35 * prof_active_[gs]et{_unlocked,}().
36 */
37bool prof_active;
38static malloc_mutex_t prof_active_mtx;
39
40/*
41 * Initialized as opt_prof_thread_active_init, and accessed via
42 * prof_thread_active_init_[gs]et().
43 */
44static bool prof_thread_active_init;
45static malloc_mutex_t prof_thread_active_init_mtx;
46
47/*
48 * Initialized as opt_prof_gdump, and accessed via
49 * prof_gdump_[gs]et{_unlocked,}().
50 */
51bool prof_gdump_val;
52static malloc_mutex_t prof_gdump_mtx;
53
54uint64_t prof_interval = 0;
55
56size_t lg_prof_sample;
57
58/*
59 * Table of mutexes that are shared among gctx's. These are leaf locks, so
60 * there is no problem with using them for more than one gctx at the same time.
61 * The primary motivation for this sharing though is that gctx's are ephemeral,
62 * and destroying mutexes causes complications for systems that allocate when
63 * creating/destroying mutexes.
64 */
65static malloc_mutex_t *gctx_locks;
66static unsigned cum_gctxs; /* Atomic counter. */
67
68/*
69 * Table of mutexes that are shared among tdata's. No operations require
70 * holding multiple tdata locks, so there is no problem with using them for more
71 * than one tdata at the same time, even though a gctx lock may be acquired
72 * while holding a tdata lock.
73 */
74static malloc_mutex_t *tdata_locks;
75
76/*
77 * Global hash of (prof_bt_t *)-->(prof_gctx_t *). This is the master data
78 * structure that knows about all backtraces currently captured.
79 */
80static ckh_t bt2gctx;
81static malloc_mutex_t bt2gctx_mtx;
82
83/*
84 * Tree of all extant prof_tdata_t structures, regardless of state,
85 * {attached,detached,expired}.
86 */
87static prof_tdata_tree_t tdatas;
88static malloc_mutex_t tdatas_mtx;
89
90static uint64_t next_thr_uid;
91static malloc_mutex_t next_thr_uid_mtx;
92
93static malloc_mutex_t prof_dump_seq_mtx;
94static uint64_t prof_dump_seq;
95static uint64_t prof_dump_iseq;
96static uint64_t prof_dump_mseq;
97static uint64_t prof_dump_useq;
98
99/*
100 * This buffer is rather large for stack allocation, so use a single buffer for
101 * all profile dumps.
102 */
103static malloc_mutex_t prof_dump_mtx;
104static char prof_dump_buf[
105 /* Minimize memory bloat for non-prof builds. */
106#ifdef JEMALLOC_PROF
107 PROF_DUMP_BUFSIZE
108#else
109 1
110#endif
111];
112static size_t prof_dump_buf_end;
113static int prof_dump_fd;
114
115/* Do not dump any profiles until bootstrapping is complete. */
116static bool prof_booted = false;
117
118/******************************************************************************/
119/*
120 * Function prototypes for static functions that are referenced prior to
121 * definition.
122 */
123
124static bool prof_tctx_should_destroy(tsdn_t *tsdn, prof_tctx_t *tctx);
125static void prof_tctx_destroy(tsd_t *tsd, prof_tctx_t *tctx);
126static bool prof_tdata_should_destroy(tsdn_t *tsdn, prof_tdata_t *tdata,
127 bool even_if_attached);
128static void prof_tdata_destroy(tsdn_t *tsdn, prof_tdata_t *tdata,
129 bool even_if_attached);
130static char *prof_thread_name_alloc(tsdn_t *tsdn, const char *thread_name);
131
132/******************************************************************************/
133/* Red-black trees. */
134
135JEMALLOC_INLINE_C int
136prof_tctx_comp(const prof_tctx_t *a, const prof_tctx_t *b)
137{
138 uint64_t a_thr_uid = a->thr_uid;
139 uint64_t b_thr_uid = b->thr_uid;
140 int ret = (a_thr_uid > b_thr_uid) - (a_thr_uid < b_thr_uid);
141 if (ret == 0) {
142 uint64_t a_thr_discrim = a->thr_discrim;
143 uint64_t b_thr_discrim = b->thr_discrim;
144 ret = (a_thr_discrim > b_thr_discrim) - (a_thr_discrim <
145 b_thr_discrim);
146 if (ret == 0) {
147 uint64_t a_tctx_uid = a->tctx_uid;
148 uint64_t b_tctx_uid = b->tctx_uid;
149 ret = (a_tctx_uid > b_tctx_uid) - (a_tctx_uid <
150 b_tctx_uid);
151 }
152 }
153 return (ret);
154}
155
156rb_gen(static UNUSED, tctx_tree_, prof_tctx_tree_t, prof_tctx_t,
157 tctx_link, prof_tctx_comp)
158
159JEMALLOC_INLINE_C int
160prof_gctx_comp(const prof_gctx_t *a, const prof_gctx_t *b)
161{
162 unsigned a_len = a->bt.len;
163 unsigned b_len = b->bt.len;
164 unsigned comp_len = (a_len < b_len) ? a_len : b_len;
165 int ret = memcmp(a->bt.vec, b->bt.vec, comp_len * sizeof(void *));
166 if (ret == 0)
167 ret = (a_len > b_len) - (a_len < b_len);
168 return (ret);
169}
170
171rb_gen(static UNUSED, gctx_tree_, prof_gctx_tree_t, prof_gctx_t, dump_link,
172 prof_gctx_comp)
173
174JEMALLOC_INLINE_C int
175prof_tdata_comp(const prof_tdata_t *a, const prof_tdata_t *b)
176{
177 int ret;
178 uint64_t a_uid = a->thr_uid;
179 uint64_t b_uid = b->thr_uid;
180
181 ret = ((a_uid > b_uid) - (a_uid < b_uid));
182 if (ret == 0) {
183 uint64_t a_discrim = a->thr_discrim;
184 uint64_t b_discrim = b->thr_discrim;
185
186 ret = ((a_discrim > b_discrim) - (a_discrim < b_discrim));
187 }
188 return (ret);
189}
190
191rb_gen(static UNUSED, tdata_tree_, prof_tdata_tree_t, prof_tdata_t, tdata_link,
192 prof_tdata_comp)
193
194/******************************************************************************/
195
196void
197prof_alloc_rollback(tsd_t *tsd, prof_tctx_t *tctx, bool updated)
198{
199 prof_tdata_t *tdata;
200
201 cassert(config_prof);
202
203 if (updated) {
204 /*
205 * Compute a new sample threshold. This isn't very important in
206 * practice, because this function is rarely executed, so the
207 * potential for sample bias is minimal except in contrived
208 * programs.
209 */
210 tdata = prof_tdata_get(tsd, true);
211 if (tdata != NULL)
212 prof_sample_threshold_update(tdata);
213 }
214
215 if ((uintptr_t)tctx > (uintptr_t)1U) {
216 malloc_mutex_lock(tsd_tsdn(tsd), tctx->tdata->lock);
217 tctx->prepared = false;
218 if (prof_tctx_should_destroy(tsd_tsdn(tsd), tctx))
219 prof_tctx_destroy(tsd, tctx);
220 else
221 malloc_mutex_unlock(tsd_tsdn(tsd), tctx->tdata->lock);
222 }
223}
224
225void
226prof_malloc_sample_object(tsdn_t *tsdn, const void *ptr, size_t usize,
227 prof_tctx_t *tctx)
228{
229
230 prof_tctx_set(tsdn, ptr, usize, tctx);
231
232 malloc_mutex_lock(tsdn, tctx->tdata->lock);
233 tctx->cnts.curobjs++;
234 tctx->cnts.curbytes += usize;
235 if (opt_prof_accum) {
236 tctx->cnts.accumobjs++;
237 tctx->cnts.accumbytes += usize;
238 }
239 tctx->prepared = false;
240 malloc_mutex_unlock(tsdn, tctx->tdata->lock);
241}
242
243void
244prof_free_sampled_object(tsd_t *tsd, size_t usize, prof_tctx_t *tctx)
245{
246
247 malloc_mutex_lock(tsd_tsdn(tsd), tctx->tdata->lock);
248 assert(tctx->cnts.curobjs > 0);
249 assert(tctx->cnts.curbytes >= usize);
250 tctx->cnts.curobjs--;
251 tctx->cnts.curbytes -= usize;
252
253 if (prof_tctx_should_destroy(tsd_tsdn(tsd), tctx))
254 prof_tctx_destroy(tsd, tctx);
255 else
256 malloc_mutex_unlock(tsd_tsdn(tsd), tctx->tdata->lock);
257}
258
259void
260bt_init(prof_bt_t *bt, void **vec)
261{
262
263 cassert(config_prof);
264
265 bt->vec = vec;
266 bt->len = 0;
267}
268
269JEMALLOC_INLINE_C void
270prof_enter(tsd_t *tsd, prof_tdata_t *tdata)
271{
272
273 cassert(config_prof);
274 assert(tdata == prof_tdata_get(tsd, false));
275
276 if (tdata != NULL) {
277 assert(!tdata->enq);
278 tdata->enq = true;
279 }
280
281 malloc_mutex_lock(tsd_tsdn(tsd), &bt2gctx_mtx);
282}
283
284JEMALLOC_INLINE_C void
285prof_leave(tsd_t *tsd, prof_tdata_t *tdata)
286{
287
288 cassert(config_prof);
289 assert(tdata == prof_tdata_get(tsd, false));
290
291 malloc_mutex_unlock(tsd_tsdn(tsd), &bt2gctx_mtx);
292
293 if (tdata != NULL) {
294 bool idump, gdump;
295
296 assert(tdata->enq);
297 tdata->enq = false;
298 idump = tdata->enq_idump;
299 tdata->enq_idump = false;
300 gdump = tdata->enq_gdump;
301 tdata->enq_gdump = false;
302
303 if (idump)
304 prof_idump(tsd_tsdn(tsd));
305 if (gdump)
306 prof_gdump(tsd_tsdn(tsd));
307 }
308}
309
310#ifdef JEMALLOC_PROF_LIBUNWIND
311void
312prof_backtrace(prof_bt_t *bt)
313{
314 int nframes;
315
316 cassert(config_prof);
317 assert(bt->len == 0);
318 assert(bt->vec != NULL);
319
320 nframes = unw_backtrace(bt->vec, PROF_BT_MAX);
321 if (nframes <= 0)
322 return;
323 bt->len = nframes;
324}
325#elif (defined(JEMALLOC_PROF_LIBGCC))
326static _Unwind_Reason_Code
327prof_unwind_init_callback(struct _Unwind_Context *context, void *arg)
328{
329
330 cassert(config_prof);
331
332 return (_URC_NO_REASON);
333}
334
335static _Unwind_Reason_Code
336prof_unwind_callback(struct _Unwind_Context *context, void *arg)
337{
338 prof_unwind_data_t *data = (prof_unwind_data_t *)arg;
339 void *ip;
340
341 cassert(config_prof);
342
343 ip = (void *)_Unwind_GetIP(context);
344 if (ip == NULL)
345 return (_URC_END_OF_STACK);
346 data->bt->vec[data->bt->len] = ip;
347 data->bt->len++;
348 if (data->bt->len == data->max)
349 return (_URC_END_OF_STACK);
350
351 return (_URC_NO_REASON);
352}
353
354void
355prof_backtrace(prof_bt_t *bt)
356{
357 prof_unwind_data_t data = {bt, PROF_BT_MAX};
358
359 cassert(config_prof);
360
361 _Unwind_Backtrace(prof_unwind_callback, &data);
362}
363#elif (defined(JEMALLOC_PROF_GCC))
364void
365prof_backtrace(prof_bt_t *bt)
366{
367#define BT_FRAME(i) \
368 if ((i) < PROF_BT_MAX) { \
369 void *p; \
370 if (__builtin_frame_address(i) == 0) \
371 return; \
372 p = __builtin_return_address(i); \
373 if (p == NULL) \
374 return; \
375 bt->vec[(i)] = p; \
376 bt->len = (i) + 1; \
377 } else \
378 return;
379
380 cassert(config_prof);
381
382 BT_FRAME(0)
383 BT_FRAME(1)
384 BT_FRAME(2)
385 BT_FRAME(3)
386 BT_FRAME(4)
387 BT_FRAME(5)
388 BT_FRAME(6)
389 BT_FRAME(7)
390 BT_FRAME(8)
391 BT_FRAME(9)
392
393 BT_FRAME(10)
394 BT_FRAME(11)
395 BT_FRAME(12)
396 BT_FRAME(13)
397 BT_FRAME(14)
398 BT_FRAME(15)
399 BT_FRAME(16)
400 BT_FRAME(17)
401 BT_FRAME(18)
402 BT_FRAME(19)
403
404 BT_FRAME(20)
405 BT_FRAME(21)
406 BT_FRAME(22)
407 BT_FRAME(23)
408 BT_FRAME(24)
409 BT_FRAME(25)
410 BT_FRAME(26)
411 BT_FRAME(27)
412 BT_FRAME(28)
413 BT_FRAME(29)
414
415 BT_FRAME(30)
416 BT_FRAME(31)
417 BT_FRAME(32)
418 BT_FRAME(33)
419 BT_FRAME(34)
420 BT_FRAME(35)
421 BT_FRAME(36)
422 BT_FRAME(37)
423 BT_FRAME(38)
424 BT_FRAME(39)
425
426 BT_FRAME(40)
427 BT_FRAME(41)
428 BT_FRAME(42)
429 BT_FRAME(43)
430 BT_FRAME(44)
431 BT_FRAME(45)
432 BT_FRAME(46)
433 BT_FRAME(47)
434 BT_FRAME(48)
435 BT_FRAME(49)
436
437 BT_FRAME(50)
438 BT_FRAME(51)
439 BT_FRAME(52)
440 BT_FRAME(53)
441 BT_FRAME(54)
442 BT_FRAME(55)
443 BT_FRAME(56)
444 BT_FRAME(57)
445 BT_FRAME(58)
446 BT_FRAME(59)
447
448 BT_FRAME(60)
449 BT_FRAME(61)
450 BT_FRAME(62)
451 BT_FRAME(63)
452 BT_FRAME(64)
453 BT_FRAME(65)
454 BT_FRAME(66)
455 BT_FRAME(67)
456 BT_FRAME(68)
457 BT_FRAME(69)
458
459 BT_FRAME(70)
460 BT_FRAME(71)
461 BT_FRAME(72)
462 BT_FRAME(73)
463 BT_FRAME(74)
464 BT_FRAME(75)
465 BT_FRAME(76)
466 BT_FRAME(77)
467 BT_FRAME(78)
468 BT_FRAME(79)
469
470 BT_FRAME(80)
471 BT_FRAME(81)
472 BT_FRAME(82)
473 BT_FRAME(83)
474 BT_FRAME(84)
475 BT_FRAME(85)
476 BT_FRAME(86)
477 BT_FRAME(87)
478 BT_FRAME(88)
479 BT_FRAME(89)
480
481 BT_FRAME(90)
482 BT_FRAME(91)
483 BT_FRAME(92)
484 BT_FRAME(93)
485 BT_FRAME(94)
486 BT_FRAME(95)
487 BT_FRAME(96)
488 BT_FRAME(97)
489 BT_FRAME(98)
490 BT_FRAME(99)
491
492 BT_FRAME(100)
493 BT_FRAME(101)
494 BT_FRAME(102)
495 BT_FRAME(103)
496 BT_FRAME(104)
497 BT_FRAME(105)
498 BT_FRAME(106)
499 BT_FRAME(107)
500 BT_FRAME(108)
501 BT_FRAME(109)
502
503 BT_FRAME(110)
504 BT_FRAME(111)
505 BT_FRAME(112)
506 BT_FRAME(113)
507 BT_FRAME(114)
508 BT_FRAME(115)
509 BT_FRAME(116)
510 BT_FRAME(117)
511 BT_FRAME(118)
512 BT_FRAME(119)
513
514 BT_FRAME(120)
515 BT_FRAME(121)
516 BT_FRAME(122)
517 BT_FRAME(123)
518 BT_FRAME(124)
519 BT_FRAME(125)
520 BT_FRAME(126)
521 BT_FRAME(127)
522#undef BT_FRAME
523}
524#else
525void
526prof_backtrace(prof_bt_t *bt)
527{
528
529 cassert(config_prof);
530 not_reached();
531}
532#endif
533
534static malloc_mutex_t *
535prof_gctx_mutex_choose(void)
536{
537 unsigned ngctxs = atomic_add_u(&cum_gctxs, 1);
538
539 return (&gctx_locks[(ngctxs - 1) % PROF_NCTX_LOCKS]);
540}
541
542static malloc_mutex_t *
543prof_tdata_mutex_choose(uint64_t thr_uid)
544{
545
546 return (&tdata_locks[thr_uid % PROF_NTDATA_LOCKS]);
547}
548
549static prof_gctx_t *
550prof_gctx_create(tsdn_t *tsdn, prof_bt_t *bt)
551{
552 /*
553 * Create a single allocation that has space for vec of length bt->len.
554 */
555 size_t size = offsetof(prof_gctx_t, vec) + (bt->len * sizeof(void *));
556 prof_gctx_t *gctx = (prof_gctx_t *)iallocztm(tsdn, size,
557 size2index(size), false, NULL, true, arena_get(TSDN_NULL, 0, true),
558 true);
559 if (gctx == NULL)
560 return (NULL);
561 gctx->lock = prof_gctx_mutex_choose();
562 /*
563 * Set nlimbo to 1, in order to avoid a race condition with
564 * prof_tctx_destroy()/prof_gctx_try_destroy().
565 */
566 gctx->nlimbo = 1;
567 tctx_tree_new(&gctx->tctxs);
568 /* Duplicate bt. */
569 memcpy(gctx->vec, bt->vec, bt->len * sizeof(void *));
570 gctx->bt.vec = gctx->vec;
571 gctx->bt.len = bt->len;
572 return (gctx);
573}
574
575static void
576prof_gctx_try_destroy(tsd_t *tsd, prof_tdata_t *tdata_self, prof_gctx_t *gctx,
577 prof_tdata_t *tdata)
578{
579
580 cassert(config_prof);
581
582 /*
583 * Check that gctx is still unused by any thread cache before destroying
584 * it. prof_lookup() increments gctx->nlimbo in order to avoid a race
585 * condition with this function, as does prof_tctx_destroy() in order to
586 * avoid a race between the main body of prof_tctx_destroy() and entry
587 * into this function.
588 */
589 prof_enter(tsd, tdata_self);
590 malloc_mutex_lock(tsd_tsdn(tsd), gctx->lock);
591 assert(gctx->nlimbo != 0);
592 if (tctx_tree_empty(&gctx->tctxs) && gctx->nlimbo == 1) {
593 /* Remove gctx from bt2gctx. */
594 if (ckh_remove(tsd_tsdn(tsd), &bt2gctx, &gctx->bt, NULL, NULL))
595 not_reached();
596 prof_leave(tsd, tdata_self);
597 /* Destroy gctx. */
598 malloc_mutex_unlock(tsd_tsdn(tsd), gctx->lock);
599 idalloctm(tsd_tsdn(tsd), gctx, NULL, true, true);
600 } else {
601 /*
602 * Compensate for increment in prof_tctx_destroy() or
603 * prof_lookup().
604 */
605 gctx->nlimbo--;
606 malloc_mutex_unlock(tsd_tsdn(tsd), gctx->lock);
607 prof_leave(tsd, tdata_self);
608 }
609}
610
611static bool
612prof_tctx_should_destroy(tsdn_t *tsdn, prof_tctx_t *tctx)
613{
614
615 malloc_mutex_assert_owner(tsdn, tctx->tdata->lock);
616
617 if (opt_prof_accum)
618 return (false);
619 if (tctx->cnts.curobjs != 0)
620 return (false);
621 if (tctx->prepared)
622 return (false);
623 return (true);
624}
625
626static bool
627prof_gctx_should_destroy(prof_gctx_t *gctx)
628{
629
630 if (opt_prof_accum)
631 return (false);
632 if (!tctx_tree_empty(&gctx->tctxs))
633 return (false);
634 if (gctx->nlimbo != 0)
635 return (false);
636 return (true);
637}
638
639static void
640prof_tctx_destroy(tsd_t *tsd, prof_tctx_t *tctx)
641{
642 prof_tdata_t *tdata = tctx->tdata;
643 prof_gctx_t *gctx = tctx->gctx;
644 bool destroy_tdata, destroy_tctx, destroy_gctx;
645
646 malloc_mutex_assert_owner(tsd_tsdn(tsd), tctx->tdata->lock);
647
648 assert(tctx->cnts.curobjs == 0);
649 assert(tctx->cnts.curbytes == 0);
650 assert(!opt_prof_accum);
651 assert(tctx->cnts.accumobjs == 0);
652 assert(tctx->cnts.accumbytes == 0);
653
654 ckh_remove(tsd_tsdn(tsd), &tdata->bt2tctx, &gctx->bt, NULL, NULL);
655 destroy_tdata = prof_tdata_should_destroy(tsd_tsdn(tsd), tdata, false);
656 malloc_mutex_unlock(tsd_tsdn(tsd), tdata->lock);
657
658 malloc_mutex_lock(tsd_tsdn(tsd), gctx->lock);
659 switch (tctx->state) {
660 case prof_tctx_state_nominal:
661 tctx_tree_remove(&gctx->tctxs, tctx);
662 destroy_tctx = true;
663 if (prof_gctx_should_destroy(gctx)) {
664 /*
665 * Increment gctx->nlimbo in order to keep another
666 * thread from winning the race to destroy gctx while
667 * this one has gctx->lock dropped. Without this, it
668 * would be possible for another thread to:
669 *
670 * 1) Sample an allocation associated with gctx.
671 * 2) Deallocate the sampled object.
672 * 3) Successfully prof_gctx_try_destroy(gctx).
673 *
674 * The result would be that gctx no longer exists by the
675 * time this thread accesses it in
676 * prof_gctx_try_destroy().
677 */
678 gctx->nlimbo++;
679 destroy_gctx = true;
680 } else
681 destroy_gctx = false;
682 break;
683 case prof_tctx_state_dumping:
684 /*
685 * A dumping thread needs tctx to remain valid until dumping
686 * has finished. Change state such that the dumping thread will
687 * complete destruction during a late dump iteration phase.
688 */
689 tctx->state = prof_tctx_state_purgatory;
690 destroy_tctx = false;
691 destroy_gctx = false;
692 break;
693 default:
694 not_reached();
695 destroy_tctx = false;
696 destroy_gctx = false;
697 }
698 malloc_mutex_unlock(tsd_tsdn(tsd), gctx->lock);
699 if (destroy_gctx) {
700 prof_gctx_try_destroy(tsd, prof_tdata_get(tsd, false), gctx,
701 tdata);
702 }
703
704 malloc_mutex_assert_not_owner(tsd_tsdn(tsd), tctx->tdata->lock);
705
706 if (destroy_tdata)
707 prof_tdata_destroy(tsd_tsdn(tsd), tdata, false);
708
709 if (destroy_tctx)
710 idalloctm(tsd_tsdn(tsd), tctx, NULL, true, true);
711}
712
713static bool
714prof_lookup_global(tsd_t *tsd, prof_bt_t *bt, prof_tdata_t *tdata,
715 void **p_btkey, prof_gctx_t **p_gctx, bool *p_new_gctx)
716{
717 union {
718 prof_gctx_t *p;
719 void *v;
720 } gctx;
721 union {
722 prof_bt_t *p;
723 void *v;
724 } btkey;
725 bool new_gctx;
726
727 prof_enter(tsd, tdata);
728 if (ckh_search(&bt2gctx, bt, &btkey.v, &gctx.v)) {
729 /* bt has never been seen before. Insert it. */
730 gctx.p = prof_gctx_create(tsd_tsdn(tsd), bt);
731 if (gctx.v == NULL) {
732 prof_leave(tsd, tdata);
733 return (true);
734 }
735 btkey.p = &gctx.p->bt;
736 if (ckh_insert(tsd_tsdn(tsd), &bt2gctx, btkey.v, gctx.v)) {
737 /* OOM. */
738 prof_leave(tsd, tdata);
739 idalloctm(tsd_tsdn(tsd), gctx.v, NULL, true, true);
740 return (true);
741 }
742 new_gctx = true;
743 } else {
744 /*
745 * Increment nlimbo, in order to avoid a race condition with
746 * prof_tctx_destroy()/prof_gctx_try_destroy().
747 */
748 malloc_mutex_lock(tsd_tsdn(tsd), gctx.p->lock);
749 gctx.p->nlimbo++;
750 malloc_mutex_unlock(tsd_tsdn(tsd), gctx.p->lock);
751 new_gctx = false;
752 }
753 prof_leave(tsd, tdata);
754
755 *p_btkey = btkey.v;
756 *p_gctx = gctx.p;
757 *p_new_gctx = new_gctx;
758 return (false);
759}
760
761prof_tctx_t *
762prof_lookup(tsd_t *tsd, prof_bt_t *bt)
763{
764 union {
765 prof_tctx_t *p;
766 void *v;
767 } ret;
768 prof_tdata_t *tdata;
769 bool not_found;
770
771 cassert(config_prof);
772
773 tdata = prof_tdata_get(tsd, false);
774 if (tdata == NULL)
775 return (NULL);
776
777 malloc_mutex_lock(tsd_tsdn(tsd), tdata->lock);
778 not_found = ckh_search(&tdata->bt2tctx, bt, NULL, &ret.v);
779 if (!not_found) /* Note double negative! */
780 ret.p->prepared = true;
781 malloc_mutex_unlock(tsd_tsdn(tsd), tdata->lock);
782 if (not_found) {
783 void *btkey;
784 prof_gctx_t *gctx;
785 bool new_gctx, error;
786
787 /*
788 * This thread's cache lacks bt. Look for it in the global
789 * cache.
790 */
791 if (prof_lookup_global(tsd, bt, tdata, &btkey, &gctx,
792 &new_gctx))
793 return (NULL);
794
795 /* Link a prof_tctx_t into gctx for this thread. */
796 ret.v = iallocztm(tsd_tsdn(tsd), sizeof(prof_tctx_t),
797 size2index(sizeof(prof_tctx_t)), false, NULL, true,
798 arena_ichoose(tsd_tsdn(tsd), NULL), true);
799 if (ret.p == NULL) {
800 if (new_gctx)
801 prof_gctx_try_destroy(tsd, tdata, gctx, tdata);
802 return (NULL);
803 }
804 ret.p->tdata = tdata;
805 ret.p->thr_uid = tdata->thr_uid;
806 ret.p->thr_discrim = tdata->thr_discrim;
807 memset(&ret.p->cnts, 0, sizeof(prof_cnt_t));
808 ret.p->gctx = gctx;
809 ret.p->tctx_uid = tdata->tctx_uid_next++;
810 ret.p->prepared = true;
811 ret.p->state = prof_tctx_state_initializing;
812 malloc_mutex_lock(tsd_tsdn(tsd), tdata->lock);
813 error = ckh_insert(tsd_tsdn(tsd), &tdata->bt2tctx, btkey,
814 ret.v);
815 malloc_mutex_unlock(tsd_tsdn(tsd), tdata->lock);
816 if (error) {
817 if (new_gctx)
818 prof_gctx_try_destroy(tsd, tdata, gctx, tdata);
819 idalloctm(tsd_tsdn(tsd), ret.v, NULL, true, true);
820 return (NULL);
821 }
822 malloc_mutex_lock(tsd_tsdn(tsd), gctx->lock);
823 ret.p->state = prof_tctx_state_nominal;
824 tctx_tree_insert(&gctx->tctxs, ret.p);
825 gctx->nlimbo--;
826 malloc_mutex_unlock(tsd_tsdn(tsd), gctx->lock);
827 }
828
829 return (ret.p);
830}
831
832/*
833 * The bodies of this function and prof_leakcheck() are compiled out unless heap
834 * profiling is enabled, so that it is possible to compile jemalloc with
835 * floating point support completely disabled. Avoiding floating point code is
836 * important on memory-constrained systems, but it also enables a workaround for
837 * versions of glibc that don't properly save/restore floating point registers
838 * during dynamic lazy symbol loading (which internally calls into whatever
839 * malloc implementation happens to be integrated into the application). Note
840 * that some compilers (e.g. gcc 4.8) may use floating point registers for fast
841 * memory moves, so jemalloc must be compiled with such optimizations disabled
842 * (e.g.
843 * -mno-sse) in order for the workaround to be complete.
844 */
845void
846prof_sample_threshold_update(prof_tdata_t *tdata)
847{
848#ifdef JEMALLOC_PROF
849 uint64_t r;
850 double u;
851
852 if (!config_prof)
853 return;
854
855 if (lg_prof_sample == 0) {
856 tdata->bytes_until_sample = 0;
857 return;
858 }
859
860 /*
861 * Compute sample interval as a geometrically distributed random
862 * variable with mean (2^lg_prof_sample).
863 *
864 * __ __
865 * | log(u) | 1
866 * tdata->bytes_until_sample = | -------- |, where p = ---------------
867 * | log(1-p) | lg_prof_sample
868 * 2
869 *
870 * For more information on the math, see:
871 *
872 * Non-Uniform Random Variate Generation
873 * Luc Devroye
874 * Springer-Verlag, New York, 1986
875 * pp 500
876 * (http://luc.devroye.org/rnbookindex.html)
877 */
878 r = prng_lg_range(&tdata->prng_state, 53);
879 u = (double)r * (1.0/9007199254740992.0L);
880 tdata->bytes_until_sample = (uint64_t)(log(u) /
881 log(1.0 - (1.0 / (double)((uint64_t)1U << lg_prof_sample))))
882 + (uint64_t)1U;
883#endif
884}
885
886#ifdef JEMALLOC_JET
887static prof_tdata_t *
888prof_tdata_count_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
889{
890 size_t *tdata_count = (size_t *)arg;
891
892 (*tdata_count)++;
893
894 return (NULL);
895}
896
897size_t
898prof_tdata_count(void)
899{
900 size_t tdata_count = 0;
901 tsdn_t *tsdn;
902
903 tsdn = tsdn_fetch();
904 malloc_mutex_lock(tsdn, &tdatas_mtx);
905 tdata_tree_iter(&tdatas, NULL, prof_tdata_count_iter,
906 (void *)&tdata_count);
907 malloc_mutex_unlock(tsdn, &tdatas_mtx);
908
909 return (tdata_count);
910}
911#endif
912
913#ifdef JEMALLOC_JET
914size_t
915prof_bt_count(void)
916{
917 size_t bt_count;
918 tsd_t *tsd;
919 prof_tdata_t *tdata;
920
921 tsd = tsd_fetch();
922 tdata = prof_tdata_get(tsd, false);
923 if (tdata == NULL)
924 return (0);
925
926 malloc_mutex_lock(tsd_tsdn(tsd), &bt2gctx_mtx);
927 bt_count = ckh_count(&bt2gctx);
928 malloc_mutex_unlock(tsd_tsdn(tsd), &bt2gctx_mtx);
929
930 return (bt_count);
931}
932#endif
933
934#ifdef JEMALLOC_JET
935#undef prof_dump_open
936#define prof_dump_open JEMALLOC_N(prof_dump_open_impl)
937#endif
938static int
939prof_dump_open(bool propagate_err, const char *filename)
940{
941 int fd;
942
943 fd = creat(filename, 0644);
944 if (fd == -1 && !propagate_err) {
945 malloc_printf("<jemalloc>: creat(\"%s\"), 0644) failed\n",
946 filename);
947 if (opt_abort)
948 abort();
949 }
950
951 return (fd);
952}
953#ifdef JEMALLOC_JET
954#undef prof_dump_open
955#define prof_dump_open JEMALLOC_N(prof_dump_open)
956prof_dump_open_t *prof_dump_open = JEMALLOC_N(prof_dump_open_impl);
957#endif
958
959static bool
960prof_dump_flush(bool propagate_err)
961{
962 bool ret = false;
963 ssize_t err;
964
965 cassert(config_prof);
966
967 err = write(prof_dump_fd, prof_dump_buf, prof_dump_buf_end);
968 if (err == -1) {
969 if (!propagate_err) {
970 malloc_write("<jemalloc>: write() failed during heap "
971 "profile flush\n");
972 if (opt_abort)
973 abort();
974 }
975 ret = true;
976 }
977 prof_dump_buf_end = 0;
978
979 return (ret);
980}
981
982static bool
983prof_dump_close(bool propagate_err)
984{
985 bool ret;
986
987 assert(prof_dump_fd != -1);
988 ret = prof_dump_flush(propagate_err);
989 close(prof_dump_fd);
990 prof_dump_fd = -1;
991
992 return (ret);
993}
994
995static bool
996prof_dump_write(bool propagate_err, const char *s)
997{
998 size_t i, slen, n;
999
1000 cassert(config_prof);
1001
1002 i = 0;
1003 slen = strlen(s);
1004 while (i < slen) {
1005 /* Flush the buffer if it is full. */
1006 if (prof_dump_buf_end == PROF_DUMP_BUFSIZE)
1007 if (prof_dump_flush(propagate_err) && propagate_err)
1008 return (true);
1009
1010 if (prof_dump_buf_end + slen <= PROF_DUMP_BUFSIZE) {
1011 /* Finish writing. */
1012 n = slen - i;
1013 } else {
1014 /* Write as much of s as will fit. */
1015 n = PROF_DUMP_BUFSIZE - prof_dump_buf_end;
1016 }
1017 memcpy(&prof_dump_buf[prof_dump_buf_end], &s[i], n);
1018 prof_dump_buf_end += n;
1019 i += n;
1020 }
1021
1022 return (false);
1023}
1024
1025JEMALLOC_FORMAT_PRINTF(2, 3)
1026static bool
1027prof_dump_printf(bool propagate_err, const char *format, ...)
1028{
1029 bool ret;
1030 va_list ap;
1031 char buf[PROF_PRINTF_BUFSIZE];
1032
1033 va_start(ap, format);
1034 malloc_vsnprintf(buf, sizeof(buf), format, ap);
1035 va_end(ap);
1036 ret = prof_dump_write(propagate_err, buf);
1037
1038 return (ret);
1039}
1040
1041static void
1042prof_tctx_merge_tdata(tsdn_t *tsdn, prof_tctx_t *tctx, prof_tdata_t *tdata)
1043{
1044
1045 malloc_mutex_assert_owner(tsdn, tctx->tdata->lock);
1046
1047 malloc_mutex_lock(tsdn, tctx->gctx->lock);
1048
1049 switch (tctx->state) {
1050 case prof_tctx_state_initializing:
1051 malloc_mutex_unlock(tsdn, tctx->gctx->lock);
1052 return;
1053 case prof_tctx_state_nominal:
1054 tctx->state = prof_tctx_state_dumping;
1055 malloc_mutex_unlock(tsdn, tctx->gctx->lock);
1056
1057 memcpy(&tctx->dump_cnts, &tctx->cnts, sizeof(prof_cnt_t));
1058
1059 tdata->cnt_summed.curobjs += tctx->dump_cnts.curobjs;
1060 tdata->cnt_summed.curbytes += tctx->dump_cnts.curbytes;
1061 if (opt_prof_accum) {
1062 tdata->cnt_summed.accumobjs +=
1063 tctx->dump_cnts.accumobjs;
1064 tdata->cnt_summed.accumbytes +=
1065 tctx->dump_cnts.accumbytes;
1066 }
1067 break;
1068 case prof_tctx_state_dumping:
1069 case prof_tctx_state_purgatory:
1070 not_reached();
1071 }
1072}
1073
1074static void
1075prof_tctx_merge_gctx(tsdn_t *tsdn, prof_tctx_t *tctx, prof_gctx_t *gctx)
1076{
1077
1078 malloc_mutex_assert_owner(tsdn, gctx->lock);
1079
1080 gctx->cnt_summed.curobjs += tctx->dump_cnts.curobjs;
1081 gctx->cnt_summed.curbytes += tctx->dump_cnts.curbytes;
1082 if (opt_prof_accum) {
1083 gctx->cnt_summed.accumobjs += tctx->dump_cnts.accumobjs;
1084 gctx->cnt_summed.accumbytes += tctx->dump_cnts.accumbytes;
1085 }
1086}
1087
1088static prof_tctx_t *
1089prof_tctx_merge_iter(prof_tctx_tree_t *tctxs, prof_tctx_t *tctx, void *arg)
1090{
1091 tsdn_t *tsdn = (tsdn_t *)arg;
1092
1093 malloc_mutex_assert_owner(tsdn, tctx->gctx->lock);
1094
1095 switch (tctx->state) {
1096 case prof_tctx_state_nominal:
1097 /* New since dumping started; ignore. */
1098 break;
1099 case prof_tctx_state_dumping:
1100 case prof_tctx_state_purgatory:
1101 prof_tctx_merge_gctx(tsdn, tctx, tctx->gctx);
1102 break;
1103 default:
1104 not_reached();
1105 }
1106
1107 return (NULL);
1108}
1109
1110struct prof_tctx_dump_iter_arg_s {
1111 tsdn_t *tsdn;
1112 bool propagate_err;
1113};
1114
1115static prof_tctx_t *
1116prof_tctx_dump_iter(prof_tctx_tree_t *tctxs, prof_tctx_t *tctx, void *opaque)
1117{
1118 struct prof_tctx_dump_iter_arg_s *arg =
1119 (struct prof_tctx_dump_iter_arg_s *)opaque;
1120
1121 malloc_mutex_assert_owner(arg->tsdn, tctx->gctx->lock);
1122
1123 switch (tctx->state) {
1124 case prof_tctx_state_initializing:
1125 case prof_tctx_state_nominal:
1126 /* Not captured by this dump. */
1127 break;
1128 case prof_tctx_state_dumping:
1129 case prof_tctx_state_purgatory:
1130 if (prof_dump_printf(arg->propagate_err,
1131 " t%"FMTu64": %"FMTu64": %"FMTu64" [%"FMTu64": "
1132 "%"FMTu64"]\n", tctx->thr_uid, tctx->dump_cnts.curobjs,
1133 tctx->dump_cnts.curbytes, tctx->dump_cnts.accumobjs,
1134 tctx->dump_cnts.accumbytes))
1135 return (tctx);
1136 break;
1137 default:
1138 not_reached();
1139 }
1140 return (NULL);
1141}
1142
1143static prof_tctx_t *
1144prof_tctx_finish_iter(prof_tctx_tree_t *tctxs, prof_tctx_t *tctx, void *arg)
1145{
1146 tsdn_t *tsdn = (tsdn_t *)arg;
1147 prof_tctx_t *ret;
1148
1149 malloc_mutex_assert_owner(tsdn, tctx->gctx->lock);
1150
1151 switch (tctx->state) {
1152 case prof_tctx_state_nominal:
1153 /* New since dumping started; ignore. */
1154 break;
1155 case prof_tctx_state_dumping:
1156 tctx->state = prof_tctx_state_nominal;
1157 break;
1158 case prof_tctx_state_purgatory:
1159 ret = tctx;
1160 goto label_return;
1161 default:
1162 not_reached();
1163 }
1164
1165 ret = NULL;
1166label_return:
1167 return (ret);
1168}
1169
1170static void
1171prof_dump_gctx_prep(tsdn_t *tsdn, prof_gctx_t *gctx, prof_gctx_tree_t *gctxs)
1172{
1173
1174 cassert(config_prof);
1175
1176 malloc_mutex_lock(tsdn, gctx->lock);
1177
1178 /*
1179 * Increment nlimbo so that gctx won't go away before dump.
1180 * Additionally, link gctx into the dump list so that it is included in
1181 * prof_dump()'s second pass.
1182 */
1183 gctx->nlimbo++;
1184 gctx_tree_insert(gctxs, gctx);
1185
1186 memset(&gctx->cnt_summed, 0, sizeof(prof_cnt_t));
1187
1188 malloc_mutex_unlock(tsdn, gctx->lock);
1189}
1190
1191struct prof_gctx_merge_iter_arg_s {
1192 tsdn_t *tsdn;
1193 size_t leak_ngctx;
1194};
1195
1196static prof_gctx_t *
1197prof_gctx_merge_iter(prof_gctx_tree_t *gctxs, prof_gctx_t *gctx, void *opaque)
1198{
1199 struct prof_gctx_merge_iter_arg_s *arg =
1200 (struct prof_gctx_merge_iter_arg_s *)opaque;
1201
1202 malloc_mutex_lock(arg->tsdn, gctx->lock);
1203 tctx_tree_iter(&gctx->tctxs, NULL, prof_tctx_merge_iter,
1204 (void *)arg->tsdn);
1205 if (gctx->cnt_summed.curobjs != 0)
1206 arg->leak_ngctx++;
1207 malloc_mutex_unlock(arg->tsdn, gctx->lock);
1208
1209 return (NULL);
1210}
1211
1212static void
1213prof_gctx_finish(tsd_t *tsd, prof_gctx_tree_t *gctxs)
1214{
1215 prof_tdata_t *tdata = prof_tdata_get(tsd, false);
1216 prof_gctx_t *gctx;
1217
1218 /*
1219 * Standard tree iteration won't work here, because as soon as we
1220 * decrement gctx->nlimbo and unlock gctx, another thread can
1221 * concurrently destroy it, which will corrupt the tree. Therefore,
1222 * tear down the tree one node at a time during iteration.
1223 */
1224 while ((gctx = gctx_tree_first(gctxs)) != NULL) {
1225 gctx_tree_remove(gctxs, gctx);
1226 malloc_mutex_lock(tsd_tsdn(tsd), gctx->lock);
1227 {
1228 prof_tctx_t *next;
1229
1230 next = NULL;
1231 do {
1232 prof_tctx_t *to_destroy =
1233 tctx_tree_iter(&gctx->tctxs, next,
1234 prof_tctx_finish_iter,
1235 (void *)tsd_tsdn(tsd));
1236 if (to_destroy != NULL) {
1237 next = tctx_tree_next(&gctx->tctxs,
1238 to_destroy);
1239 tctx_tree_remove(&gctx->tctxs,
1240 to_destroy);
1241 idalloctm(tsd_tsdn(tsd), to_destroy,
1242 NULL, true, true);
1243 } else
1244 next = NULL;
1245 } while (next != NULL);
1246 }
1247 gctx->nlimbo--;
1248 if (prof_gctx_should_destroy(gctx)) {
1249 gctx->nlimbo++;
1250 malloc_mutex_unlock(tsd_tsdn(tsd), gctx->lock);
1251 prof_gctx_try_destroy(tsd, tdata, gctx, tdata);
1252 } else
1253 malloc_mutex_unlock(tsd_tsdn(tsd), gctx->lock);
1254 }
1255}
1256
1257struct prof_tdata_merge_iter_arg_s {
1258 tsdn_t *tsdn;
1259 prof_cnt_t cnt_all;
1260};
1261
1262static prof_tdata_t *
1263prof_tdata_merge_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata,
1264 void *opaque)
1265{
1266 struct prof_tdata_merge_iter_arg_s *arg =
1267 (struct prof_tdata_merge_iter_arg_s *)opaque;
1268
1269 malloc_mutex_lock(arg->tsdn, tdata->lock);
1270 if (!tdata->expired) {
1271 size_t tabind;
1272 union {
1273 prof_tctx_t *p;
1274 void *v;
1275 } tctx;
1276
1277 tdata->dumping = true;
1278 memset(&tdata->cnt_summed, 0, sizeof(prof_cnt_t));
1279 for (tabind = 0; !ckh_iter(&tdata->bt2tctx, &tabind, NULL,
1280 &tctx.v);)
1281 prof_tctx_merge_tdata(arg->tsdn, tctx.p, tdata);
1282
1283 arg->cnt_all.curobjs += tdata->cnt_summed.curobjs;
1284 arg->cnt_all.curbytes += tdata->cnt_summed.curbytes;
1285 if (opt_prof_accum) {
1286 arg->cnt_all.accumobjs += tdata->cnt_summed.accumobjs;
1287 arg->cnt_all.accumbytes += tdata->cnt_summed.accumbytes;
1288 }
1289 } else
1290 tdata->dumping = false;
1291 malloc_mutex_unlock(arg->tsdn, tdata->lock);
1292
1293 return (NULL);
1294}
1295
1296static prof_tdata_t *
1297prof_tdata_dump_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
1298{
1299 bool propagate_err = *(bool *)arg;
1300
1301 if (!tdata->dumping)
1302 return (NULL);
1303
1304 if (prof_dump_printf(propagate_err,
1305 " t%"FMTu64": %"FMTu64": %"FMTu64" [%"FMTu64": %"FMTu64"]%s%s\n",
1306 tdata->thr_uid, tdata->cnt_summed.curobjs,
1307 tdata->cnt_summed.curbytes, tdata->cnt_summed.accumobjs,
1308 tdata->cnt_summed.accumbytes,
1309 (tdata->thread_name != NULL) ? " " : "",
1310 (tdata->thread_name != NULL) ? tdata->thread_name : ""))
1311 return (tdata);
1312 return (NULL);
1313}
1314
1315#ifdef JEMALLOC_JET
1316#undef prof_dump_header
1317#define prof_dump_header JEMALLOC_N(prof_dump_header_impl)
1318#endif
1319static bool
1320prof_dump_header(tsdn_t *tsdn, bool propagate_err, const prof_cnt_t *cnt_all)
1321{
1322 bool ret;
1323
1324 if (prof_dump_printf(propagate_err,
1325 "heap_v2/%"FMTu64"\n"
1326 " t*: %"FMTu64": %"FMTu64" [%"FMTu64": %"FMTu64"]\n",
1327 ((uint64_t)1U << lg_prof_sample), cnt_all->curobjs,
1328 cnt_all->curbytes, cnt_all->accumobjs, cnt_all->accumbytes))
1329 return (true);
1330
1331 malloc_mutex_lock(tsdn, &tdatas_mtx);
1332 ret = (tdata_tree_iter(&tdatas, NULL, prof_tdata_dump_iter,
1333 (void *)&propagate_err) != NULL);
1334 malloc_mutex_unlock(tsdn, &tdatas_mtx);
1335 return (ret);
1336}
1337#ifdef JEMALLOC_JET
1338#undef prof_dump_header
1339#define prof_dump_header JEMALLOC_N(prof_dump_header)
1340prof_dump_header_t *prof_dump_header = JEMALLOC_N(prof_dump_header_impl);
1341#endif
1342
1343static bool
1344prof_dump_gctx(tsdn_t *tsdn, bool propagate_err, prof_gctx_t *gctx,
1345 const prof_bt_t *bt, prof_gctx_tree_t *gctxs)
1346{
1347 bool ret;
1348 unsigned i;
1349 struct prof_tctx_dump_iter_arg_s prof_tctx_dump_iter_arg;
1350
1351 cassert(config_prof);
1352 malloc_mutex_assert_owner(tsdn, gctx->lock);
1353
1354 /* Avoid dumping such gctx's that have no useful data. */
1355 if ((!opt_prof_accum && gctx->cnt_summed.curobjs == 0) ||
1356 (opt_prof_accum && gctx->cnt_summed.accumobjs == 0)) {
1357 assert(gctx->cnt_summed.curobjs == 0);
1358 assert(gctx->cnt_summed.curbytes == 0);
1359 assert(gctx->cnt_summed.accumobjs == 0);
1360 assert(gctx->cnt_summed.accumbytes == 0);
1361 ret = false;
1362 goto label_return;
1363 }
1364
1365 if (prof_dump_printf(propagate_err, "@")) {
1366 ret = true;
1367 goto label_return;
1368 }
1369 for (i = 0; i < bt->len; i++) {
1370 if (prof_dump_printf(propagate_err, " %#"FMTxPTR,
1371 (uintptr_t)bt->vec[i])) {
1372 ret = true;
1373 goto label_return;
1374 }
1375 }
1376
1377 if (prof_dump_printf(propagate_err,
1378 "\n"
1379 " t*: %"FMTu64": %"FMTu64" [%"FMTu64": %"FMTu64"]\n",
1380 gctx->cnt_summed.curobjs, gctx->cnt_summed.curbytes,
1381 gctx->cnt_summed.accumobjs, gctx->cnt_summed.accumbytes)) {
1382 ret = true;
1383 goto label_return;
1384 }
1385
1386 prof_tctx_dump_iter_arg.tsdn = tsdn;
1387 prof_tctx_dump_iter_arg.propagate_err = propagate_err;
1388 if (tctx_tree_iter(&gctx->tctxs, NULL, prof_tctx_dump_iter,
1389 (void *)&prof_tctx_dump_iter_arg) != NULL) {
1390 ret = true;
1391 goto label_return;
1392 }
1393
1394 ret = false;
1395label_return:
1396 return (ret);
1397}
1398
1399#ifndef _WIN32
1400JEMALLOC_FORMAT_PRINTF(1, 2)
1401static int
1402prof_open_maps(const char *format, ...)
1403{
1404 int mfd;
1405 va_list ap;
1406 char filename[PATH_MAX + 1];
1407
1408 va_start(ap, format);
1409 malloc_vsnprintf(filename, sizeof(filename), format, ap);
1410 va_end(ap);
1411 mfd = open(filename, O_RDONLY);
1412
1413 return (mfd);
1414}
1415#endif
1416
1417static int
1418prof_getpid(void)
1419{
1420
1421#ifdef _WIN32
1422 return (GetCurrentProcessId());
1423#else
1424 return (getpid());
1425#endif
1426}
1427
1428static bool
1429prof_dump_maps(bool propagate_err)
1430{
1431 bool ret;
1432 int mfd;
1433
1434 cassert(config_prof);
1435#ifdef __FreeBSD__
1436 mfd = prof_open_maps("/proc/curproc/map");
1437#elif defined(_WIN32)
1438 mfd = -1; // Not implemented
1439#else
1440 {
1441 int pid = prof_getpid();
1442
1443 mfd = prof_open_maps("/proc/%d/task/%d/maps", pid, pid);
1444 if (mfd == -1)
1445 mfd = prof_open_maps("/proc/%d/maps", pid);
1446 }
1447#endif
1448 if (mfd != -1) {
1449 ssize_t nread;
1450
1451 if (prof_dump_write(propagate_err, "\nMAPPED_LIBRARIES:\n") &&
1452 propagate_err) {
1453 ret = true;
1454 goto label_return;
1455 }
1456 nread = 0;
1457 do {
1458 prof_dump_buf_end += nread;
1459 if (prof_dump_buf_end == PROF_DUMP_BUFSIZE) {
1460 /* Make space in prof_dump_buf before read(). */
1461 if (prof_dump_flush(propagate_err) &&
1462 propagate_err) {
1463 ret = true;
1464 goto label_return;
1465 }
1466 }
1467 nread = read(mfd, &prof_dump_buf[prof_dump_buf_end],
1468 PROF_DUMP_BUFSIZE - prof_dump_buf_end);
1469 } while (nread > 0);
1470 } else {
1471 ret = true;
1472 goto label_return;
1473 }
1474
1475 ret = false;
1476label_return:
1477 if (mfd != -1)
1478 close(mfd);
1479 return (ret);
1480}
1481
1482/*
1483 * See prof_sample_threshold_update() comment for why the body of this function
1484 * is conditionally compiled.
1485 */
1486static void
1487prof_leakcheck(const prof_cnt_t *cnt_all, size_t leak_ngctx,
1488 const char *filename)
1489{
1490
1491#ifdef JEMALLOC_PROF
1492 /*
1493 * Scaling is equivalent AdjustSamples() in jeprof, but the result may
1494 * differ slightly from what jeprof reports, because here we scale the
1495 * summary values, whereas jeprof scales each context individually and
1496 * reports the sums of the scaled values.
1497 */
1498 if (cnt_all->curbytes != 0) {
1499 double sample_period = (double)((uint64_t)1 << lg_prof_sample);
1500 double ratio = (((double)cnt_all->curbytes) /
1501 (double)cnt_all->curobjs) / sample_period;
1502 double scale_factor = 1.0 / (1.0 - exp(-ratio));
1503 uint64_t curbytes = (uint64_t)round(((double)cnt_all->curbytes)
1504 * scale_factor);
1505 uint64_t curobjs = (uint64_t)round(((double)cnt_all->curobjs) *
1506 scale_factor);
1507
1508 malloc_printf("<jemalloc>: Leak approximation summary: ~%"FMTu64
1509 " byte%s, ~%"FMTu64" object%s, >= %zu context%s\n",
1510 curbytes, (curbytes != 1) ? "s" : "", curobjs, (curobjs !=
1511 1) ? "s" : "", leak_ngctx, (leak_ngctx != 1) ? "s" : "");
1512 malloc_printf(
1513 "<jemalloc>: Run jeprof on \"%s\" for leak detail\n",
1514 filename);
1515 }
1516#endif
1517}
1518
1519struct prof_gctx_dump_iter_arg_s {
1520 tsdn_t *tsdn;
1521 bool propagate_err;
1522};
1523
1524static prof_gctx_t *
1525prof_gctx_dump_iter(prof_gctx_tree_t *gctxs, prof_gctx_t *gctx, void *opaque)
1526{
1527 prof_gctx_t *ret;
1528 struct prof_gctx_dump_iter_arg_s *arg =
1529 (struct prof_gctx_dump_iter_arg_s *)opaque;
1530
1531 malloc_mutex_lock(arg->tsdn, gctx->lock);
1532
1533 if (prof_dump_gctx(arg->tsdn, arg->propagate_err, gctx, &gctx->bt,
1534 gctxs)) {
1535 ret = gctx;
1536 goto label_return;
1537 }
1538
1539 ret = NULL;
1540label_return:
1541 malloc_mutex_unlock(arg->tsdn, gctx->lock);
1542 return (ret);
1543}
1544
1545static bool
1546prof_dump(tsd_t *tsd, bool propagate_err, const char *filename, bool leakcheck)
1547{
1548 prof_tdata_t *tdata;
1549 struct prof_tdata_merge_iter_arg_s prof_tdata_merge_iter_arg;
1550 size_t tabind;
1551 union {
1552 prof_gctx_t *p;
1553 void *v;
1554 } gctx;
1555 struct prof_gctx_merge_iter_arg_s prof_gctx_merge_iter_arg;
1556 struct prof_gctx_dump_iter_arg_s prof_gctx_dump_iter_arg;
1557 prof_gctx_tree_t gctxs;
1558
1559 cassert(config_prof);
1560
1561 tdata = prof_tdata_get(tsd, true);
1562 if (tdata == NULL)
1563 return (true);
1564
1565 malloc_mutex_lock(tsd_tsdn(tsd), &prof_dump_mtx);
1566 prof_enter(tsd, tdata);
1567
1568 /*
1569 * Put gctx's in limbo and clear their counters in preparation for
1570 * summing.
1571 */
1572 gctx_tree_new(&gctxs);
1573 for (tabind = 0; !ckh_iter(&bt2gctx, &tabind, NULL, &gctx.v);)
1574 prof_dump_gctx_prep(tsd_tsdn(tsd), gctx.p, &gctxs);
1575
1576 /*
1577 * Iterate over tdatas, and for the non-expired ones snapshot their tctx
1578 * stats and merge them into the associated gctx's.
1579 */
1580 prof_tdata_merge_iter_arg.tsdn = tsd_tsdn(tsd);
1581 memset(&prof_tdata_merge_iter_arg.cnt_all, 0, sizeof(prof_cnt_t));
1582 malloc_mutex_lock(tsd_tsdn(tsd), &tdatas_mtx);
1583 tdata_tree_iter(&tdatas, NULL, prof_tdata_merge_iter,
1584 (void *)&prof_tdata_merge_iter_arg);
1585 malloc_mutex_unlock(tsd_tsdn(tsd), &tdatas_mtx);
1586
1587 /* Merge tctx stats into gctx's. */
1588 prof_gctx_merge_iter_arg.tsdn = tsd_tsdn(tsd);
1589 prof_gctx_merge_iter_arg.leak_ngctx = 0;
1590 gctx_tree_iter(&gctxs, NULL, prof_gctx_merge_iter,
1591 (void *)&prof_gctx_merge_iter_arg);
1592
1593 prof_leave(tsd, tdata);
1594
1595 /* Create dump file. */
1596 if ((prof_dump_fd = prof_dump_open(propagate_err, filename)) == -1)
1597 goto label_open_close_error;
1598
1599 /* Dump profile header. */
1600 if (prof_dump_header(tsd_tsdn(tsd), propagate_err,
1601 &prof_tdata_merge_iter_arg.cnt_all))
1602 goto label_write_error;
1603
1604 /* Dump per gctx profile stats. */
1605 prof_gctx_dump_iter_arg.tsdn = tsd_tsdn(tsd);
1606 prof_gctx_dump_iter_arg.propagate_err = propagate_err;
1607 if (gctx_tree_iter(&gctxs, NULL, prof_gctx_dump_iter,
1608 (void *)&prof_gctx_dump_iter_arg) != NULL)
1609 goto label_write_error;
1610
1611 /* Dump /proc/<pid>/maps if possible. */
1612 if (prof_dump_maps(propagate_err))
1613 goto label_write_error;
1614
1615 if (prof_dump_close(propagate_err))
1616 goto label_open_close_error;
1617
1618 prof_gctx_finish(tsd, &gctxs);
1619 malloc_mutex_unlock(tsd_tsdn(tsd), &prof_dump_mtx);
1620
1621 if (leakcheck) {
1622 prof_leakcheck(&prof_tdata_merge_iter_arg.cnt_all,
1623 prof_gctx_merge_iter_arg.leak_ngctx, filename);
1624 }
1625 return (false);
1626label_write_error:
1627 prof_dump_close(propagate_err);
1628label_open_close_error:
1629 prof_gctx_finish(tsd, &gctxs);
1630 malloc_mutex_unlock(tsd_tsdn(tsd), &prof_dump_mtx);
1631 return (true);
1632}
1633
1634#define DUMP_FILENAME_BUFSIZE (PATH_MAX + 1)
1635#define VSEQ_INVALID UINT64_C(0xffffffffffffffff)
1636static void
1637prof_dump_filename(char *filename, char v, uint64_t vseq)
1638{
1639
1640 cassert(config_prof);
1641
1642 if (vseq != VSEQ_INVALID) {
1643 /* "<prefix>.<pid>.<seq>.v<vseq>.heap" */
1644 malloc_snprintf(filename, DUMP_FILENAME_BUFSIZE,
1645 "%s.%d.%"FMTu64".%c%"FMTu64".heap",
1646 opt_prof_prefix, prof_getpid(), prof_dump_seq, v, vseq);
1647 } else {
1648 /* "<prefix>.<pid>.<seq>.<v>.heap" */
1649 malloc_snprintf(filename, DUMP_FILENAME_BUFSIZE,
1650 "%s.%d.%"FMTu64".%c.heap",
1651 opt_prof_prefix, prof_getpid(), prof_dump_seq, v);
1652 }
1653 prof_dump_seq++;
1654}
1655
1656static void
1657prof_fdump(void)
1658{
1659 tsd_t *tsd;
1660 char filename[DUMP_FILENAME_BUFSIZE];
1661
1662 cassert(config_prof);
1663 assert(opt_prof_final);
1664 assert(opt_prof_prefix[0] != '\0');
1665
1666 if (!prof_booted)
1667 return;
1668 tsd = tsd_fetch();
1669
1670 malloc_mutex_lock(tsd_tsdn(tsd), &prof_dump_seq_mtx);
1671 prof_dump_filename(filename, 'f', VSEQ_INVALID);
1672 malloc_mutex_unlock(tsd_tsdn(tsd), &prof_dump_seq_mtx);
1673 prof_dump(tsd, false, filename, opt_prof_leak);
1674}
1675
1676void
1677prof_idump(tsdn_t *tsdn)
1678{
1679 tsd_t *tsd;
1680 prof_tdata_t *tdata;
1681
1682 cassert(config_prof);
1683
1684 if (!prof_booted || tsdn_null(tsdn))
1685 return;
1686 tsd = tsdn_tsd(tsdn);
1687 tdata = prof_tdata_get(tsd, false);
1688 if (tdata == NULL)
1689 return;
1690 if (tdata->enq) {
1691 tdata->enq_idump = true;
1692 return;
1693 }
1694
1695 if (opt_prof_prefix[0] != '\0') {
1696 char filename[PATH_MAX + 1];
1697 malloc_mutex_lock(tsd_tsdn(tsd), &prof_dump_seq_mtx);
1698 prof_dump_filename(filename, 'i', prof_dump_iseq);
1699 prof_dump_iseq++;
1700 malloc_mutex_unlock(tsd_tsdn(tsd), &prof_dump_seq_mtx);
1701 prof_dump(tsd, false, filename, false);
1702 }
1703}
1704
1705bool
1706prof_mdump(tsd_t *tsd, const char *filename)
1707{
1708 char filename_buf[DUMP_FILENAME_BUFSIZE];
1709
1710 cassert(config_prof);
1711
1712 if (!opt_prof || !prof_booted)
1713 return (true);
1714
1715 if (filename == NULL) {
1716 /* No filename specified, so automatically generate one. */
1717 if (opt_prof_prefix[0] == '\0')
1718 return (true);
1719 malloc_mutex_lock(tsd_tsdn(tsd), &prof_dump_seq_mtx);
1720 prof_dump_filename(filename_buf, 'm', prof_dump_mseq);
1721 prof_dump_mseq++;
1722 malloc_mutex_unlock(tsd_tsdn(tsd), &prof_dump_seq_mtx);
1723 filename = filename_buf;
1724 }
1725 return (prof_dump(tsd, true, filename, false));
1726}
1727
1728void
1729prof_gdump(tsdn_t *tsdn)
1730{
1731 tsd_t *tsd;
1732 prof_tdata_t *tdata;
1733
1734 cassert(config_prof);
1735
1736 if (!prof_booted || tsdn_null(tsdn))
1737 return;
1738 tsd = tsdn_tsd(tsdn);
1739 tdata = prof_tdata_get(tsd, false);
1740 if (tdata == NULL)
1741 return;
1742 if (tdata->enq) {
1743 tdata->enq_gdump = true;
1744 return;
1745 }
1746
1747 if (opt_prof_prefix[0] != '\0') {
1748 char filename[DUMP_FILENAME_BUFSIZE];
1749 malloc_mutex_lock(tsdn, &prof_dump_seq_mtx);
1750 prof_dump_filename(filename, 'u', prof_dump_useq);
1751 prof_dump_useq++;
1752 malloc_mutex_unlock(tsdn, &prof_dump_seq_mtx);
1753 prof_dump(tsd, false, filename, false);
1754 }
1755}
1756
1757static void
1758prof_bt_hash(const void *key, size_t r_hash[2])
1759{
1760 prof_bt_t *bt = (prof_bt_t *)key;
1761
1762 cassert(config_prof);
1763
1764 hash(bt->vec, bt->len * sizeof(void *), 0x94122f33U, r_hash);
1765}
1766
1767static bool
1768prof_bt_keycomp(const void *k1, const void *k2)
1769{
1770 const prof_bt_t *bt1 = (prof_bt_t *)k1;
1771 const prof_bt_t *bt2 = (prof_bt_t *)k2;
1772
1773 cassert(config_prof);
1774
1775 if (bt1->len != bt2->len)
1776 return (false);
1777 return (memcmp(bt1->vec, bt2->vec, bt1->len * sizeof(void *)) == 0);
1778}
1779
1780JEMALLOC_INLINE_C uint64_t
1781prof_thr_uid_alloc(tsdn_t *tsdn)
1782{
1783 uint64_t thr_uid;
1784
1785 malloc_mutex_lock(tsdn, &next_thr_uid_mtx);
1786 thr_uid = next_thr_uid;
1787 next_thr_uid++;
1788 malloc_mutex_unlock(tsdn, &next_thr_uid_mtx);
1789
1790 return (thr_uid);
1791}
1792
1793static prof_tdata_t *
1794prof_tdata_init_impl(tsdn_t *tsdn, uint64_t thr_uid, uint64_t thr_discrim,
1795 char *thread_name, bool active)
1796{
1797 prof_tdata_t *tdata;
1798
1799 cassert(config_prof);
1800
1801 /* Initialize an empty cache for this thread. */
1802 tdata = (prof_tdata_t *)iallocztm(tsdn, sizeof(prof_tdata_t),
1803 size2index(sizeof(prof_tdata_t)), false, NULL, true,
1804 arena_get(TSDN_NULL, 0, true), true);
1805 if (tdata == NULL)
1806 return (NULL);
1807
1808 tdata->lock = prof_tdata_mutex_choose(thr_uid);
1809 tdata->thr_uid = thr_uid;
1810 tdata->thr_discrim = thr_discrim;
1811 tdata->thread_name = thread_name;
1812 tdata->attached = true;
1813 tdata->expired = false;
1814 tdata->tctx_uid_next = 0;
1815
1816 if (ckh_new(tsdn, &tdata->bt2tctx, PROF_CKH_MINITEMS,
1817 prof_bt_hash, prof_bt_keycomp)) {
1818 idalloctm(tsdn, tdata, NULL, true, true);
1819 return (NULL);
1820 }
1821
1822 tdata->prng_state = (uint64_t)(uintptr_t)tdata;
1823 prof_sample_threshold_update(tdata);
1824
1825 tdata->enq = false;
1826 tdata->enq_idump = false;
1827 tdata->enq_gdump = false;
1828
1829 tdata->dumping = false;
1830 tdata->active = active;
1831
1832 malloc_mutex_lock(tsdn, &tdatas_mtx);
1833 tdata_tree_insert(&tdatas, tdata);
1834 malloc_mutex_unlock(tsdn, &tdatas_mtx);
1835
1836 return (tdata);
1837}
1838
1839prof_tdata_t *
1840prof_tdata_init(tsdn_t *tsdn)
1841{
1842
1843 return (prof_tdata_init_impl(tsdn, prof_thr_uid_alloc(tsdn), 0, NULL,
1844 prof_thread_active_init_get(tsdn)));
1845}
1846
1847static bool
1848prof_tdata_should_destroy_unlocked(prof_tdata_t *tdata, bool even_if_attached)
1849{
1850
1851 if (tdata->attached && !even_if_attached)
1852 return (false);
1853 if (ckh_count(&tdata->bt2tctx) != 0)
1854 return (false);
1855 return (true);
1856}
1857
1858static bool
1859prof_tdata_should_destroy(tsdn_t *tsdn, prof_tdata_t *tdata,
1860 bool even_if_attached)
1861{
1862
1863 malloc_mutex_assert_owner(tsdn, tdata->lock);
1864
1865 return (prof_tdata_should_destroy_unlocked(tdata, even_if_attached));
1866}
1867
1868static void
1869prof_tdata_destroy_locked(tsdn_t *tsdn, prof_tdata_t *tdata,
1870 bool even_if_attached)
1871{
1872
1873 malloc_mutex_assert_owner(tsdn, &tdatas_mtx);
1874
1875 assert(tsdn_null(tsdn) || tsd_prof_tdata_get(tsdn_tsd(tsdn)) != tdata);
1876
1877 tdata_tree_remove(&tdatas, tdata);
1878
1879 assert(prof_tdata_should_destroy_unlocked(tdata, even_if_attached));
1880
1881 if (tdata->thread_name != NULL)
1882 idalloctm(tsdn, tdata->thread_name, NULL, true, true);
1883 ckh_delete(tsdn, &tdata->bt2tctx);
1884 idalloctm(tsdn, tdata, NULL, true, true);
1885}
1886
1887static void
1888prof_tdata_destroy(tsdn_t *tsdn, prof_tdata_t *tdata, bool even_if_attached)
1889{
1890
1891 malloc_mutex_lock(tsdn, &tdatas_mtx);
1892 prof_tdata_destroy_locked(tsdn, tdata, even_if_attached);
1893 malloc_mutex_unlock(tsdn, &tdatas_mtx);
1894}
1895
1896static void
1897prof_tdata_detach(tsd_t *tsd, prof_tdata_t *tdata)
1898{
1899 bool destroy_tdata;
1900
1901 malloc_mutex_lock(tsd_tsdn(tsd), tdata->lock);
1902 if (tdata->attached) {
1903 destroy_tdata = prof_tdata_should_destroy(tsd_tsdn(tsd), tdata,
1904 true);
1905 /*
1906 * Only detach if !destroy_tdata, because detaching would allow
1907 * another thread to win the race to destroy tdata.
1908 */
1909 if (!destroy_tdata)
1910 tdata->attached = false;
1911 tsd_prof_tdata_set(tsd, NULL);
1912 } else
1913 destroy_tdata = false;
1914 malloc_mutex_unlock(tsd_tsdn(tsd), tdata->lock);
1915 if (destroy_tdata)
1916 prof_tdata_destroy(tsd_tsdn(tsd), tdata, true);
1917}
1918
1919prof_tdata_t *
1920prof_tdata_reinit(tsd_t *tsd, prof_tdata_t *tdata)
1921{
1922 uint64_t thr_uid = tdata->thr_uid;
1923 uint64_t thr_discrim = tdata->thr_discrim + 1;
1924 char *thread_name = (tdata->thread_name != NULL) ?
1925 prof_thread_name_alloc(tsd_tsdn(tsd), tdata->thread_name) : NULL;
1926 bool active = tdata->active;
1927
1928 prof_tdata_detach(tsd, tdata);
1929 return (prof_tdata_init_impl(tsd_tsdn(tsd), thr_uid, thr_discrim,
1930 thread_name, active));
1931}
1932
1933static bool
1934prof_tdata_expire(tsdn_t *tsdn, prof_tdata_t *tdata)
1935{
1936 bool destroy_tdata;
1937
1938 malloc_mutex_lock(tsdn, tdata->lock);
1939 if (!tdata->expired) {
1940 tdata->expired = true;
1941 destroy_tdata = tdata->attached ? false :
1942 prof_tdata_should_destroy(tsdn, tdata, false);
1943 } else
1944 destroy_tdata = false;
1945 malloc_mutex_unlock(tsdn, tdata->lock);
1946
1947 return (destroy_tdata);
1948}
1949
1950static prof_tdata_t *
1951prof_tdata_reset_iter(prof_tdata_tree_t *tdatas, prof_tdata_t *tdata, void *arg)
1952{
1953 tsdn_t *tsdn = (tsdn_t *)arg;
1954
1955 return (prof_tdata_expire(tsdn, tdata) ? tdata : NULL);
1956}
1957
1958void
1959prof_reset(tsdn_t *tsdn, size_t lg_sample)
1960{
1961 prof_tdata_t *next;
1962
1963 assert(lg_sample < (sizeof(uint64_t) << 3));
1964
1965 malloc_mutex_lock(tsdn, &prof_dump_mtx);
1966 malloc_mutex_lock(tsdn, &tdatas_mtx);
1967
1968 lg_prof_sample = lg_sample;
1969
1970 next = NULL;
1971 do {
1972 prof_tdata_t *to_destroy = tdata_tree_iter(&tdatas, next,
1973 prof_tdata_reset_iter, (void *)tsdn);
1974 if (to_destroy != NULL) {
1975 next = tdata_tree_next(&tdatas, to_destroy);
1976 prof_tdata_destroy_locked(tsdn, to_destroy, false);
1977 } else
1978 next = NULL;
1979 } while (next != NULL);
1980
1981 malloc_mutex_unlock(tsdn, &tdatas_mtx);
1982 malloc_mutex_unlock(tsdn, &prof_dump_mtx);
1983}
1984
1985void
1986prof_tdata_cleanup(tsd_t *tsd)
1987{
1988 prof_tdata_t *tdata;
1989
1990 if (!config_prof)
1991 return;
1992
1993 tdata = tsd_prof_tdata_get(tsd);
1994 if (tdata != NULL)
1995 prof_tdata_detach(tsd, tdata);
1996}
1997
1998bool
1999prof_active_get(tsdn_t *tsdn)
2000{
2001 bool prof_active_current;
2002
2003 malloc_mutex_lock(tsdn, &prof_active_mtx);
2004 prof_active_current = prof_active;
2005 malloc_mutex_unlock(tsdn, &prof_active_mtx);
2006 return (prof_active_current);
2007}
2008
2009bool
2010prof_active_set(tsdn_t *tsdn, bool active)
2011{
2012 bool prof_active_old;
2013
2014 malloc_mutex_lock(tsdn, &prof_active_mtx);
2015 prof_active_old = prof_active;
2016 prof_active = active;
2017 malloc_mutex_unlock(tsdn, &prof_active_mtx);
2018 return (prof_active_old);
2019}
2020
2021const char *
2022prof_thread_name_get(tsd_t *tsd)
2023{
2024 prof_tdata_t *tdata;
2025
2026 tdata = prof_tdata_get(tsd, true);
2027 if (tdata == NULL)
2028 return ("");
2029 return (tdata->thread_name != NULL ? tdata->thread_name : "");
2030}
2031
2032static char *
2033prof_thread_name_alloc(tsdn_t *tsdn, const char *thread_name)
2034{
2035 char *ret;
2036 size_t size;
2037
2038 if (thread_name == NULL)
2039 return (NULL);
2040
2041 size = strlen(thread_name) + 1;
2042 if (size == 1)
2043 return ("");
2044
2045 ret = iallocztm(tsdn, size, size2index(size), false, NULL, true,
2046 arena_get(TSDN_NULL, 0, true), true);
2047 if (ret == NULL)
2048 return (NULL);
2049 memcpy(ret, thread_name, size);
2050 return (ret);
2051}
2052
2053int
2054prof_thread_name_set(tsd_t *tsd, const char *thread_name)
2055{
2056 prof_tdata_t *tdata;
2057 unsigned i;
2058 char *s;
2059
2060 tdata = prof_tdata_get(tsd, true);
2061 if (tdata == NULL)
2062 return (EAGAIN);
2063
2064 /* Validate input. */
2065 if (thread_name == NULL)
2066 return (EFAULT);
2067 for (i = 0; thread_name[i] != '\0'; i++) {
2068 char c = thread_name[i];
2069 if (!isgraph(c) && !isblank(c))
2070 return (EFAULT);
2071 }
2072
2073 s = prof_thread_name_alloc(tsd_tsdn(tsd), thread_name);
2074 if (s == NULL)
2075 return (EAGAIN);
2076
2077 if (tdata->thread_name != NULL) {
2078 idalloctm(tsd_tsdn(tsd), tdata->thread_name, NULL, true, true);
2079 tdata->thread_name = NULL;
2080 }
2081 if (strlen(s) > 0)
2082 tdata->thread_name = s;
2083 return (0);
2084}
2085
2086bool
2087prof_thread_active_get(tsd_t *tsd)
2088{
2089 prof_tdata_t *tdata;
2090
2091 tdata = prof_tdata_get(tsd, true);
2092 if (tdata == NULL)
2093 return (false);
2094 return (tdata->active);
2095}
2096
2097bool
2098prof_thread_active_set(tsd_t *tsd, bool active)
2099{
2100 prof_tdata_t *tdata;
2101
2102 tdata = prof_tdata_get(tsd, true);
2103 if (tdata == NULL)
2104 return (true);
2105 tdata->active = active;
2106 return (false);
2107}
2108
2109bool
2110prof_thread_active_init_get(tsdn_t *tsdn)
2111{
2112 bool active_init;
2113
2114 malloc_mutex_lock(tsdn, &prof_thread_active_init_mtx);
2115 active_init = prof_thread_active_init;
2116 malloc_mutex_unlock(tsdn, &prof_thread_active_init_mtx);
2117 return (active_init);
2118}
2119
2120bool
2121prof_thread_active_init_set(tsdn_t *tsdn, bool active_init)
2122{
2123 bool active_init_old;
2124
2125 malloc_mutex_lock(tsdn, &prof_thread_active_init_mtx);
2126 active_init_old = prof_thread_active_init;
2127 prof_thread_active_init = active_init;
2128 malloc_mutex_unlock(tsdn, &prof_thread_active_init_mtx);
2129 return (active_init_old);
2130}
2131
2132bool
2133prof_gdump_get(tsdn_t *tsdn)
2134{
2135 bool prof_gdump_current;
2136
2137 malloc_mutex_lock(tsdn, &prof_gdump_mtx);
2138 prof_gdump_current = prof_gdump_val;
2139 malloc_mutex_unlock(tsdn, &prof_gdump_mtx);
2140 return (prof_gdump_current);
2141}
2142
2143bool
2144prof_gdump_set(tsdn_t *tsdn, bool gdump)
2145{
2146 bool prof_gdump_old;
2147
2148 malloc_mutex_lock(tsdn, &prof_gdump_mtx);
2149 prof_gdump_old = prof_gdump_val;
2150 prof_gdump_val = gdump;
2151 malloc_mutex_unlock(tsdn, &prof_gdump_mtx);
2152 return (prof_gdump_old);
2153}
2154
2155void
2156prof_boot0(void)
2157{
2158
2159 cassert(config_prof);
2160
2161 memcpy(opt_prof_prefix, PROF_PREFIX_DEFAULT,
2162 sizeof(PROF_PREFIX_DEFAULT));
2163}
2164
2165void
2166prof_boot1(void)
2167{
2168
2169 cassert(config_prof);
2170
2171 /*
2172 * opt_prof must be in its final state before any arenas are
2173 * initialized, so this function must be executed early.
2174 */
2175
2176 if (opt_prof_leak && !opt_prof) {
2177 /*
2178 * Enable opt_prof, but in such a way that profiles are never
2179 * automatically dumped.
2180 */
2181 opt_prof = true;
2182 opt_prof_gdump = false;
2183 } else if (opt_prof) {
2184 if (opt_lg_prof_interval >= 0) {
2185 prof_interval = (((uint64_t)1U) <<
2186 opt_lg_prof_interval);
2187 }
2188 }
2189}
2190
2191bool
2192prof_boot2(tsdn_t *tsdn)
2193{
2194
2195 cassert(config_prof);
2196
2197 if (opt_prof) {
2198 unsigned i;
2199
2200 lg_prof_sample = opt_lg_prof_sample;
2201
2202 prof_active = opt_prof_active;
2203 if (malloc_mutex_init(&prof_active_mtx, "prof_active",
2204 WITNESS_RANK_PROF_ACTIVE))
2205 return (true);
2206
2207 prof_gdump_val = opt_prof_gdump;
2208 if (malloc_mutex_init(&prof_gdump_mtx, "prof_gdump",
2209 WITNESS_RANK_PROF_GDUMP))
2210 return (true);
2211
2212 prof_thread_active_init = opt_prof_thread_active_init;
2213 if (malloc_mutex_init(&prof_thread_active_init_mtx,
2214 "prof_thread_active_init",
2215 WITNESS_RANK_PROF_THREAD_ACTIVE_INIT))
2216 return (true);
2217
2218 if (ckh_new(tsdn, &bt2gctx, PROF_CKH_MINITEMS, prof_bt_hash,
2219 prof_bt_keycomp))
2220 return (true);
2221 if (malloc_mutex_init(&bt2gctx_mtx, "prof_bt2gctx",
2222 WITNESS_RANK_PROF_BT2GCTX))
2223 return (true);
2224
2225 tdata_tree_new(&tdatas);
2226 if (malloc_mutex_init(&tdatas_mtx, "prof_tdatas",
2227 WITNESS_RANK_PROF_TDATAS))
2228 return (true);
2229
2230 next_thr_uid = 0;
2231 if (malloc_mutex_init(&next_thr_uid_mtx, "prof_next_thr_uid",
2232 WITNESS_RANK_PROF_NEXT_THR_UID))
2233 return (true);
2234
2235 if (malloc_mutex_init(&prof_dump_seq_mtx, "prof_dump_seq",
2236 WITNESS_RANK_PROF_DUMP_SEQ))
2237 return (true);
2238 if (malloc_mutex_init(&prof_dump_mtx, "prof_dump",
2239 WITNESS_RANK_PROF_DUMP))
2240 return (true);
2241
2242 if (opt_prof_final && opt_prof_prefix[0] != '\0' &&
2243 atexit(prof_fdump) != 0) {
2244 malloc_write("<jemalloc>: Error in atexit()\n");
2245 if (opt_abort)
2246 abort();
2247 }
2248
2249 gctx_locks = (malloc_mutex_t *)base_alloc(tsdn, PROF_NCTX_LOCKS
2250 * sizeof(malloc_mutex_t));
2251 if (gctx_locks == NULL)
2252 return (true);
2253 for (i = 0; i < PROF_NCTX_LOCKS; i++) {
2254 if (malloc_mutex_init(&gctx_locks[i], "prof_gctx",
2255 WITNESS_RANK_PROF_GCTX))
2256 return (true);
2257 }
2258
2259 tdata_locks = (malloc_mutex_t *)base_alloc(tsdn,
2260 PROF_NTDATA_LOCKS * sizeof(malloc_mutex_t));
2261 if (tdata_locks == NULL)
2262 return (true);
2263 for (i = 0; i < PROF_NTDATA_LOCKS; i++) {
2264 if (malloc_mutex_init(&tdata_locks[i], "prof_tdata",
2265 WITNESS_RANK_PROF_TDATA))
2266 return (true);
2267 }
2268 }
2269
2270#ifdef JEMALLOC_PROF_LIBGCC
2271 /*
2272 * Cause the backtracing machinery to allocate its internal state
2273 * before enabling profiling.
2274 */
2275 _Unwind_Backtrace(prof_unwind_init_callback, NULL);
2276#endif
2277
2278 prof_booted = true;
2279
2280 return (false);
2281}
2282
2283void
2284prof_prefork0(tsdn_t *tsdn)
2285{
2286
2287 if (opt_prof) {
2288 unsigned i;
2289
2290 malloc_mutex_prefork(tsdn, &prof_dump_mtx);
2291 malloc_mutex_prefork(tsdn, &bt2gctx_mtx);
2292 malloc_mutex_prefork(tsdn, &tdatas_mtx);
2293 for (i = 0; i < PROF_NTDATA_LOCKS; i++)
2294 malloc_mutex_prefork(tsdn, &tdata_locks[i]);
2295 for (i = 0; i < PROF_NCTX_LOCKS; i++)
2296 malloc_mutex_prefork(tsdn, &gctx_locks[i]);
2297 }
2298}
2299
2300void
2301prof_prefork1(tsdn_t *tsdn)
2302{
2303
2304 if (opt_prof) {
2305 malloc_mutex_prefork(tsdn, &prof_active_mtx);
2306 malloc_mutex_prefork(tsdn, &prof_dump_seq_mtx);
2307 malloc_mutex_prefork(tsdn, &prof_gdump_mtx);
2308 malloc_mutex_prefork(tsdn, &next_thr_uid_mtx);
2309 malloc_mutex_prefork(tsdn, &prof_thread_active_init_mtx);
2310 }
2311}
2312
2313void
2314prof_postfork_parent(tsdn_t *tsdn)
2315{
2316
2317 if (opt_prof) {
2318 unsigned i;
2319
2320 malloc_mutex_postfork_parent(tsdn,
2321 &prof_thread_active_init_mtx);
2322 malloc_mutex_postfork_parent(tsdn, &next_thr_uid_mtx);
2323 malloc_mutex_postfork_parent(tsdn, &prof_gdump_mtx);
2324 malloc_mutex_postfork_parent(tsdn, &prof_dump_seq_mtx);
2325 malloc_mutex_postfork_parent(tsdn, &prof_active_mtx);
2326 for (i = 0; i < PROF_NCTX_LOCKS; i++)
2327 malloc_mutex_postfork_parent(tsdn, &gctx_locks[i]);
2328 for (i = 0; i < PROF_NTDATA_LOCKS; i++)
2329 malloc_mutex_postfork_parent(tsdn, &tdata_locks[i]);
2330 malloc_mutex_postfork_parent(tsdn, &tdatas_mtx);
2331 malloc_mutex_postfork_parent(tsdn, &bt2gctx_mtx);
2332 malloc_mutex_postfork_parent(tsdn, &prof_dump_mtx);
2333 }
2334}
2335
2336void
2337prof_postfork_child(tsdn_t *tsdn)
2338{
2339
2340 if (opt_prof) {
2341 unsigned i;
2342
2343 malloc_mutex_postfork_child(tsdn, &prof_thread_active_init_mtx);
2344 malloc_mutex_postfork_child(tsdn, &next_thr_uid_mtx);
2345 malloc_mutex_postfork_child(tsdn, &prof_gdump_mtx);
2346 malloc_mutex_postfork_child(tsdn, &prof_dump_seq_mtx);
2347 malloc_mutex_postfork_child(tsdn, &prof_active_mtx);
2348 for (i = 0; i < PROF_NCTX_LOCKS; i++)
2349 malloc_mutex_postfork_child(tsdn, &gctx_locks[i]);
2350 for (i = 0; i < PROF_NTDATA_LOCKS; i++)
2351 malloc_mutex_postfork_child(tsdn, &tdata_locks[i]);
2352 malloc_mutex_postfork_child(tsdn, &tdatas_mtx);
2353 malloc_mutex_postfork_child(tsdn, &bt2gctx_mtx);
2354 malloc_mutex_postfork_child(tsdn, &prof_dump_mtx);
2355 }
2356}
2357
2358/******************************************************************************/
2359