]> git.saurik.com Git - redis.git/blob - deps/jemalloc/src/prof.c
04964ef7ca3a3602a32c8f823ae4e2de1d46e4d2
[redis.git] / deps / jemalloc / src / prof.c
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
17 malloc_tsd_data(, prof_tdata, prof_tdata_t *, NULL)
18
19 bool opt_prof = false;
20 bool opt_prof_active = true;
21 size_t opt_lg_prof_sample = LG_PROF_SAMPLE_DEFAULT;
22 ssize_t opt_lg_prof_interval = LG_PROF_INTERVAL_DEFAULT;
23 bool opt_prof_gdump = false;
24 bool opt_prof_final = true;
25 bool opt_prof_leak = false;
26 bool opt_prof_accum = false;
27 char opt_prof_prefix[PATH_MAX + 1];
28
29 uint64_t prof_interval;
30 bool prof_promote;
31
32 /*
33 * Table of mutexes that are shared among ctx's. These are leaf locks, so
34 * there is no problem with using them for more than one ctx at the same time.
35 * The primary motivation for this sharing though is that ctx's are ephemeral,
36 * and destroying mutexes causes complications for systems that allocate when
37 * creating/destroying mutexes.
38 */
39 static malloc_mutex_t *ctx_locks;
40 static unsigned cum_ctxs; /* Atomic counter. */
41
42 /*
43 * Global hash of (prof_bt_t *)-->(prof_ctx_t *). This is the master data
44 * structure that knows about all backtraces currently captured.
45 */
46 static ckh_t bt2ctx;
47 static malloc_mutex_t bt2ctx_mtx;
48
49 static malloc_mutex_t prof_dump_seq_mtx;
50 static uint64_t prof_dump_seq;
51 static uint64_t prof_dump_iseq;
52 static uint64_t prof_dump_mseq;
53 static uint64_t prof_dump_useq;
54
55 /*
56 * This buffer is rather large for stack allocation, so use a single buffer for
57 * all profile dumps. The buffer is implicitly protected by bt2ctx_mtx, since
58 * it must be locked anyway during dumping.
59 */
60 static char prof_dump_buf[PROF_DUMP_BUFSIZE];
61 static unsigned prof_dump_buf_end;
62 static int prof_dump_fd;
63
64 /* Do not dump any profiles until bootstrapping is complete. */
65 static bool prof_booted = false;
66
67 /******************************************************************************/
68 /* Function prototypes for non-inline static functions. */
69
70 static prof_bt_t *bt_dup(prof_bt_t *bt);
71 static void bt_destroy(prof_bt_t *bt);
72 #ifdef JEMALLOC_PROF_LIBGCC
73 static _Unwind_Reason_Code prof_unwind_init_callback(
74 struct _Unwind_Context *context, void *arg);
75 static _Unwind_Reason_Code prof_unwind_callback(
76 struct _Unwind_Context *context, void *arg);
77 #endif
78 static bool prof_flush(bool propagate_err);
79 static bool prof_write(bool propagate_err, const char *s);
80 static bool prof_printf(bool propagate_err, const char *format, ...)
81 JEMALLOC_ATTR(format(printf, 2, 3));
82 static void prof_ctx_sum(prof_ctx_t *ctx, prof_cnt_t *cnt_all,
83 size_t *leak_nctx);
84 static void prof_ctx_destroy(prof_ctx_t *ctx);
85 static void prof_ctx_merge(prof_ctx_t *ctx, prof_thr_cnt_t *cnt);
86 static bool prof_dump_ctx(bool propagate_err, prof_ctx_t *ctx,
87 prof_bt_t *bt);
88 static bool prof_dump_maps(bool propagate_err);
89 static bool prof_dump(bool propagate_err, const char *filename,
90 bool leakcheck);
91 static void prof_dump_filename(char *filename, char v, int64_t vseq);
92 static void prof_fdump(void);
93 static void prof_bt_hash(const void *key, unsigned minbits, size_t *hash1,
94 size_t *hash2);
95 static bool prof_bt_keycomp(const void *k1, const void *k2);
96 static malloc_mutex_t *prof_ctx_mutex_choose(void);
97
98 /******************************************************************************/
99
100 void
101 bt_init(prof_bt_t *bt, void **vec)
102 {
103
104 cassert(config_prof);
105
106 bt->vec = vec;
107 bt->len = 0;
108 }
109
110 static void
111 bt_destroy(prof_bt_t *bt)
112 {
113
114 cassert(config_prof);
115
116 idalloc(bt);
117 }
118
119 static prof_bt_t *
120 bt_dup(prof_bt_t *bt)
121 {
122 prof_bt_t *ret;
123
124 cassert(config_prof);
125
126 /*
127 * Create a single allocation that has space for vec immediately
128 * following the prof_bt_t structure. The backtraces that get
129 * stored in the backtrace caches are copied from stack-allocated
130 * temporary variables, so size is known at creation time. Making this
131 * a contiguous object improves cache locality.
132 */
133 ret = (prof_bt_t *)imalloc(QUANTUM_CEILING(sizeof(prof_bt_t)) +
134 (bt->len * sizeof(void *)));
135 if (ret == NULL)
136 return (NULL);
137 ret->vec = (void **)((uintptr_t)ret +
138 QUANTUM_CEILING(sizeof(prof_bt_t)));
139 memcpy(ret->vec, bt->vec, bt->len * sizeof(void *));
140 ret->len = bt->len;
141
142 return (ret);
143 }
144
145 static inline void
146 prof_enter(prof_tdata_t *prof_tdata)
147 {
148
149 cassert(config_prof);
150
151 assert(prof_tdata->enq == false);
152 prof_tdata->enq = true;
153
154 malloc_mutex_lock(&bt2ctx_mtx);
155 }
156
157 static inline void
158 prof_leave(prof_tdata_t *prof_tdata)
159 {
160 bool idump, gdump;
161
162 cassert(config_prof);
163
164 malloc_mutex_unlock(&bt2ctx_mtx);
165
166 assert(prof_tdata->enq);
167 prof_tdata->enq = false;
168 idump = prof_tdata->enq_idump;
169 prof_tdata->enq_idump = false;
170 gdump = prof_tdata->enq_gdump;
171 prof_tdata->enq_gdump = false;
172
173 if (idump)
174 prof_idump();
175 if (gdump)
176 prof_gdump();
177 }
178
179 #ifdef JEMALLOC_PROF_LIBUNWIND
180 void
181 prof_backtrace(prof_bt_t *bt, unsigned nignore)
182 {
183 unw_context_t uc;
184 unw_cursor_t cursor;
185 unsigned i;
186 int err;
187
188 cassert(config_prof);
189 assert(bt->len == 0);
190 assert(bt->vec != NULL);
191
192 unw_getcontext(&uc);
193 unw_init_local(&cursor, &uc);
194
195 /* Throw away (nignore+1) stack frames, if that many exist. */
196 for (i = 0; i < nignore + 1; i++) {
197 err = unw_step(&cursor);
198 if (err <= 0)
199 return;
200 }
201
202 /*
203 * Iterate over stack frames until there are no more, or until no space
204 * remains in bt.
205 */
206 for (i = 0; i < PROF_BT_MAX; i++) {
207 unw_get_reg(&cursor, UNW_REG_IP, (unw_word_t *)&bt->vec[i]);
208 bt->len++;
209 err = unw_step(&cursor);
210 if (err <= 0)
211 break;
212 }
213 }
214 #elif (defined(JEMALLOC_PROF_LIBGCC))
215 static _Unwind_Reason_Code
216 prof_unwind_init_callback(struct _Unwind_Context *context, void *arg)
217 {
218
219 cassert(config_prof);
220
221 return (_URC_NO_REASON);
222 }
223
224 static _Unwind_Reason_Code
225 prof_unwind_callback(struct _Unwind_Context *context, void *arg)
226 {
227 prof_unwind_data_t *data = (prof_unwind_data_t *)arg;
228
229 cassert(config_prof);
230
231 if (data->nignore > 0)
232 data->nignore--;
233 else {
234 data->bt->vec[data->bt->len] = (void *)_Unwind_GetIP(context);
235 data->bt->len++;
236 if (data->bt->len == data->max)
237 return (_URC_END_OF_STACK);
238 }
239
240 return (_URC_NO_REASON);
241 }
242
243 void
244 prof_backtrace(prof_bt_t *bt, unsigned nignore)
245 {
246 prof_unwind_data_t data = {bt, nignore, PROF_BT_MAX};
247
248 cassert(config_prof);
249
250 _Unwind_Backtrace(prof_unwind_callback, &data);
251 }
252 #elif (defined(JEMALLOC_PROF_GCC))
253 void
254 prof_backtrace(prof_bt_t *bt, unsigned nignore)
255 {
256 #define BT_FRAME(i) \
257 if ((i) < nignore + PROF_BT_MAX) { \
258 void *p; \
259 if (__builtin_frame_address(i) == 0) \
260 return; \
261 p = __builtin_return_address(i); \
262 if (p == NULL) \
263 return; \
264 if (i >= nignore) { \
265 bt->vec[(i) - nignore] = p; \
266 bt->len = (i) - nignore + 1; \
267 } \
268 } else \
269 return;
270
271 cassert(config_prof);
272 assert(nignore <= 3);
273
274 BT_FRAME(0)
275 BT_FRAME(1)
276 BT_FRAME(2)
277 BT_FRAME(3)
278 BT_FRAME(4)
279 BT_FRAME(5)
280 BT_FRAME(6)
281 BT_FRAME(7)
282 BT_FRAME(8)
283 BT_FRAME(9)
284
285 BT_FRAME(10)
286 BT_FRAME(11)
287 BT_FRAME(12)
288 BT_FRAME(13)
289 BT_FRAME(14)
290 BT_FRAME(15)
291 BT_FRAME(16)
292 BT_FRAME(17)
293 BT_FRAME(18)
294 BT_FRAME(19)
295
296 BT_FRAME(20)
297 BT_FRAME(21)
298 BT_FRAME(22)
299 BT_FRAME(23)
300 BT_FRAME(24)
301 BT_FRAME(25)
302 BT_FRAME(26)
303 BT_FRAME(27)
304 BT_FRAME(28)
305 BT_FRAME(29)
306
307 BT_FRAME(30)
308 BT_FRAME(31)
309 BT_FRAME(32)
310 BT_FRAME(33)
311 BT_FRAME(34)
312 BT_FRAME(35)
313 BT_FRAME(36)
314 BT_FRAME(37)
315 BT_FRAME(38)
316 BT_FRAME(39)
317
318 BT_FRAME(40)
319 BT_FRAME(41)
320 BT_FRAME(42)
321 BT_FRAME(43)
322 BT_FRAME(44)
323 BT_FRAME(45)
324 BT_FRAME(46)
325 BT_FRAME(47)
326 BT_FRAME(48)
327 BT_FRAME(49)
328
329 BT_FRAME(50)
330 BT_FRAME(51)
331 BT_FRAME(52)
332 BT_FRAME(53)
333 BT_FRAME(54)
334 BT_FRAME(55)
335 BT_FRAME(56)
336 BT_FRAME(57)
337 BT_FRAME(58)
338 BT_FRAME(59)
339
340 BT_FRAME(60)
341 BT_FRAME(61)
342 BT_FRAME(62)
343 BT_FRAME(63)
344 BT_FRAME(64)
345 BT_FRAME(65)
346 BT_FRAME(66)
347 BT_FRAME(67)
348 BT_FRAME(68)
349 BT_FRAME(69)
350
351 BT_FRAME(70)
352 BT_FRAME(71)
353 BT_FRAME(72)
354 BT_FRAME(73)
355 BT_FRAME(74)
356 BT_FRAME(75)
357 BT_FRAME(76)
358 BT_FRAME(77)
359 BT_FRAME(78)
360 BT_FRAME(79)
361
362 BT_FRAME(80)
363 BT_FRAME(81)
364 BT_FRAME(82)
365 BT_FRAME(83)
366 BT_FRAME(84)
367 BT_FRAME(85)
368 BT_FRAME(86)
369 BT_FRAME(87)
370 BT_FRAME(88)
371 BT_FRAME(89)
372
373 BT_FRAME(90)
374 BT_FRAME(91)
375 BT_FRAME(92)
376 BT_FRAME(93)
377 BT_FRAME(94)
378 BT_FRAME(95)
379 BT_FRAME(96)
380 BT_FRAME(97)
381 BT_FRAME(98)
382 BT_FRAME(99)
383
384 BT_FRAME(100)
385 BT_FRAME(101)
386 BT_FRAME(102)
387 BT_FRAME(103)
388 BT_FRAME(104)
389 BT_FRAME(105)
390 BT_FRAME(106)
391 BT_FRAME(107)
392 BT_FRAME(108)
393 BT_FRAME(109)
394
395 BT_FRAME(110)
396 BT_FRAME(111)
397 BT_FRAME(112)
398 BT_FRAME(113)
399 BT_FRAME(114)
400 BT_FRAME(115)
401 BT_FRAME(116)
402 BT_FRAME(117)
403 BT_FRAME(118)
404 BT_FRAME(119)
405
406 BT_FRAME(120)
407 BT_FRAME(121)
408 BT_FRAME(122)
409 BT_FRAME(123)
410 BT_FRAME(124)
411 BT_FRAME(125)
412 BT_FRAME(126)
413 BT_FRAME(127)
414
415 /* Extras to compensate for nignore. */
416 BT_FRAME(128)
417 BT_FRAME(129)
418 BT_FRAME(130)
419 #undef BT_FRAME
420 }
421 #else
422 void
423 prof_backtrace(prof_bt_t *bt, unsigned nignore)
424 {
425
426 cassert(config_prof);
427 assert(false);
428 }
429 #endif
430
431 prof_thr_cnt_t *
432 prof_lookup(prof_bt_t *bt)
433 {
434 union {
435 prof_thr_cnt_t *p;
436 void *v;
437 } ret;
438 prof_tdata_t *prof_tdata;
439
440 cassert(config_prof);
441
442 prof_tdata = prof_tdata_get();
443 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
444 return (NULL);
445
446 if (ckh_search(&prof_tdata->bt2cnt, bt, NULL, &ret.v)) {
447 union {
448 prof_bt_t *p;
449 void *v;
450 } btkey;
451 union {
452 prof_ctx_t *p;
453 void *v;
454 } ctx;
455 bool new_ctx;
456
457 /*
458 * This thread's cache lacks bt. Look for it in the global
459 * cache.
460 */
461 prof_enter(prof_tdata);
462 if (ckh_search(&bt2ctx, bt, &btkey.v, &ctx.v)) {
463 /* bt has never been seen before. Insert it. */
464 ctx.v = imalloc(sizeof(prof_ctx_t));
465 if (ctx.v == NULL) {
466 prof_leave(prof_tdata);
467 return (NULL);
468 }
469 btkey.p = bt_dup(bt);
470 if (btkey.v == NULL) {
471 prof_leave(prof_tdata);
472 idalloc(ctx.v);
473 return (NULL);
474 }
475 ctx.p->bt = btkey.p;
476 ctx.p->lock = prof_ctx_mutex_choose();
477 /*
478 * Set nlimbo to 1, in order to avoid a race condition
479 * with prof_ctx_merge()/prof_ctx_destroy().
480 */
481 ctx.p->nlimbo = 1;
482 memset(&ctx.p->cnt_merged, 0, sizeof(prof_cnt_t));
483 ql_new(&ctx.p->cnts_ql);
484 if (ckh_insert(&bt2ctx, btkey.v, ctx.v)) {
485 /* OOM. */
486 prof_leave(prof_tdata);
487 idalloc(btkey.v);
488 idalloc(ctx.v);
489 return (NULL);
490 }
491 new_ctx = true;
492 } else {
493 /*
494 * Increment nlimbo, in order to avoid a race condition
495 * with prof_ctx_merge()/prof_ctx_destroy().
496 */
497 malloc_mutex_lock(ctx.p->lock);
498 ctx.p->nlimbo++;
499 malloc_mutex_unlock(ctx.p->lock);
500 new_ctx = false;
501 }
502 prof_leave(prof_tdata);
503
504 /* Link a prof_thd_cnt_t into ctx for this thread. */
505 if (ckh_count(&prof_tdata->bt2cnt) == PROF_TCMAX) {
506 assert(ckh_count(&prof_tdata->bt2cnt) > 0);
507 /*
508 * Flush the least recently used cnt in order to keep
509 * bt2cnt from becoming too large.
510 */
511 ret.p = ql_last(&prof_tdata->lru_ql, lru_link);
512 assert(ret.v != NULL);
513 if (ckh_remove(&prof_tdata->bt2cnt, ret.p->ctx->bt,
514 NULL, NULL))
515 assert(false);
516 ql_remove(&prof_tdata->lru_ql, ret.p, lru_link);
517 prof_ctx_merge(ret.p->ctx, ret.p);
518 /* ret can now be re-used. */
519 } else {
520 assert(ckh_count(&prof_tdata->bt2cnt) < PROF_TCMAX);
521 /* Allocate and partially initialize a new cnt. */
522 ret.v = imalloc(sizeof(prof_thr_cnt_t));
523 if (ret.p == NULL) {
524 if (new_ctx)
525 prof_ctx_destroy(ctx.p);
526 return (NULL);
527 }
528 ql_elm_new(ret.p, cnts_link);
529 ql_elm_new(ret.p, lru_link);
530 }
531 /* Finish initializing ret. */
532 ret.p->ctx = ctx.p;
533 ret.p->epoch = 0;
534 memset(&ret.p->cnts, 0, sizeof(prof_cnt_t));
535 if (ckh_insert(&prof_tdata->bt2cnt, btkey.v, ret.v)) {
536 if (new_ctx)
537 prof_ctx_destroy(ctx.p);
538 idalloc(ret.v);
539 return (NULL);
540 }
541 ql_head_insert(&prof_tdata->lru_ql, ret.p, lru_link);
542 malloc_mutex_lock(ctx.p->lock);
543 ql_tail_insert(&ctx.p->cnts_ql, ret.p, cnts_link);
544 ctx.p->nlimbo--;
545 malloc_mutex_unlock(ctx.p->lock);
546 } else {
547 /* Move ret to the front of the LRU. */
548 ql_remove(&prof_tdata->lru_ql, ret.p, lru_link);
549 ql_head_insert(&prof_tdata->lru_ql, ret.p, lru_link);
550 }
551
552 return (ret.p);
553 }
554
555 static bool
556 prof_flush(bool propagate_err)
557 {
558 bool ret = false;
559 ssize_t err;
560
561 cassert(config_prof);
562
563 err = write(prof_dump_fd, prof_dump_buf, prof_dump_buf_end);
564 if (err == -1) {
565 if (propagate_err == false) {
566 malloc_write("<jemalloc>: write() failed during heap "
567 "profile flush\n");
568 if (opt_abort)
569 abort();
570 }
571 ret = true;
572 }
573 prof_dump_buf_end = 0;
574
575 return (ret);
576 }
577
578 static bool
579 prof_write(bool propagate_err, const char *s)
580 {
581 unsigned i, slen, n;
582
583 cassert(config_prof);
584
585 i = 0;
586 slen = strlen(s);
587 while (i < slen) {
588 /* Flush the buffer if it is full. */
589 if (prof_dump_buf_end == PROF_DUMP_BUFSIZE)
590 if (prof_flush(propagate_err) && propagate_err)
591 return (true);
592
593 if (prof_dump_buf_end + slen <= PROF_DUMP_BUFSIZE) {
594 /* Finish writing. */
595 n = slen - i;
596 } else {
597 /* Write as much of s as will fit. */
598 n = PROF_DUMP_BUFSIZE - prof_dump_buf_end;
599 }
600 memcpy(&prof_dump_buf[prof_dump_buf_end], &s[i], n);
601 prof_dump_buf_end += n;
602 i += n;
603 }
604
605 return (false);
606 }
607
608 JEMALLOC_ATTR(format(printf, 2, 3))
609 static bool
610 prof_printf(bool propagate_err, const char *format, ...)
611 {
612 bool ret;
613 va_list ap;
614 char buf[PROF_PRINTF_BUFSIZE];
615
616 va_start(ap, format);
617 malloc_vsnprintf(buf, sizeof(buf), format, ap);
618 va_end(ap);
619 ret = prof_write(propagate_err, buf);
620
621 return (ret);
622 }
623
624 static void
625 prof_ctx_sum(prof_ctx_t *ctx, prof_cnt_t *cnt_all, size_t *leak_nctx)
626 {
627 prof_thr_cnt_t *thr_cnt;
628 prof_cnt_t tcnt;
629
630 cassert(config_prof);
631
632 malloc_mutex_lock(ctx->lock);
633
634 memcpy(&ctx->cnt_summed, &ctx->cnt_merged, sizeof(prof_cnt_t));
635 ql_foreach(thr_cnt, &ctx->cnts_ql, cnts_link) {
636 volatile unsigned *epoch = &thr_cnt->epoch;
637
638 while (true) {
639 unsigned epoch0 = *epoch;
640
641 /* Make sure epoch is even. */
642 if (epoch0 & 1U)
643 continue;
644
645 memcpy(&tcnt, &thr_cnt->cnts, sizeof(prof_cnt_t));
646
647 /* Terminate if epoch didn't change while reading. */
648 if (*epoch == epoch0)
649 break;
650 }
651
652 ctx->cnt_summed.curobjs += tcnt.curobjs;
653 ctx->cnt_summed.curbytes += tcnt.curbytes;
654 if (opt_prof_accum) {
655 ctx->cnt_summed.accumobjs += tcnt.accumobjs;
656 ctx->cnt_summed.accumbytes += tcnt.accumbytes;
657 }
658 }
659
660 if (ctx->cnt_summed.curobjs != 0)
661 (*leak_nctx)++;
662
663 /* Add to cnt_all. */
664 cnt_all->curobjs += ctx->cnt_summed.curobjs;
665 cnt_all->curbytes += ctx->cnt_summed.curbytes;
666 if (opt_prof_accum) {
667 cnt_all->accumobjs += ctx->cnt_summed.accumobjs;
668 cnt_all->accumbytes += ctx->cnt_summed.accumbytes;
669 }
670
671 malloc_mutex_unlock(ctx->lock);
672 }
673
674 static void
675 prof_ctx_destroy(prof_ctx_t *ctx)
676 {
677 prof_tdata_t *prof_tdata;
678
679 cassert(config_prof);
680
681 /*
682 * Check that ctx is still unused by any thread cache before destroying
683 * it. prof_lookup() increments ctx->nlimbo in order to avoid a race
684 * condition with this function, as does prof_ctx_merge() in order to
685 * avoid a race between the main body of prof_ctx_merge() and entry
686 * into this function.
687 */
688 prof_tdata = *prof_tdata_tsd_get();
689 assert((uintptr_t)prof_tdata > (uintptr_t)PROF_TDATA_STATE_MAX);
690 prof_enter(prof_tdata);
691 malloc_mutex_lock(ctx->lock);
692 if (ql_first(&ctx->cnts_ql) == NULL && ctx->cnt_merged.curobjs == 0 &&
693 ctx->nlimbo == 1) {
694 assert(ctx->cnt_merged.curbytes == 0);
695 assert(ctx->cnt_merged.accumobjs == 0);
696 assert(ctx->cnt_merged.accumbytes == 0);
697 /* Remove ctx from bt2ctx. */
698 if (ckh_remove(&bt2ctx, ctx->bt, NULL, NULL))
699 assert(false);
700 prof_leave(prof_tdata);
701 /* Destroy ctx. */
702 malloc_mutex_unlock(ctx->lock);
703 bt_destroy(ctx->bt);
704 idalloc(ctx);
705 } else {
706 /*
707 * Compensate for increment in prof_ctx_merge() or
708 * prof_lookup().
709 */
710 ctx->nlimbo--;
711 malloc_mutex_unlock(ctx->lock);
712 prof_leave(prof_tdata);
713 }
714 }
715
716 static void
717 prof_ctx_merge(prof_ctx_t *ctx, prof_thr_cnt_t *cnt)
718 {
719 bool destroy;
720
721 cassert(config_prof);
722
723 /* Merge cnt stats and detach from ctx. */
724 malloc_mutex_lock(ctx->lock);
725 ctx->cnt_merged.curobjs += cnt->cnts.curobjs;
726 ctx->cnt_merged.curbytes += cnt->cnts.curbytes;
727 ctx->cnt_merged.accumobjs += cnt->cnts.accumobjs;
728 ctx->cnt_merged.accumbytes += cnt->cnts.accumbytes;
729 ql_remove(&ctx->cnts_ql, cnt, cnts_link);
730 if (opt_prof_accum == false && ql_first(&ctx->cnts_ql) == NULL &&
731 ctx->cnt_merged.curobjs == 0 && ctx->nlimbo == 0) {
732 /*
733 * Increment ctx->nlimbo in order to keep another thread from
734 * winning the race to destroy ctx while this one has ctx->lock
735 * dropped. Without this, it would be possible for another
736 * thread to:
737 *
738 * 1) Sample an allocation associated with ctx.
739 * 2) Deallocate the sampled object.
740 * 3) Successfully prof_ctx_destroy(ctx).
741 *
742 * The result would be that ctx no longer exists by the time
743 * this thread accesses it in prof_ctx_destroy().
744 */
745 ctx->nlimbo++;
746 destroy = true;
747 } else
748 destroy = false;
749 malloc_mutex_unlock(ctx->lock);
750 if (destroy)
751 prof_ctx_destroy(ctx);
752 }
753
754 static bool
755 prof_dump_ctx(bool propagate_err, prof_ctx_t *ctx, prof_bt_t *bt)
756 {
757 unsigned i;
758
759 cassert(config_prof);
760
761 /*
762 * Current statistics can sum to 0 as a result of unmerged per thread
763 * statistics. Additionally, interval- and growth-triggered dumps can
764 * occur between the time a ctx is created and when its statistics are
765 * filled in. Avoid dumping any ctx that is an artifact of either
766 * implementation detail.
767 */
768 if ((opt_prof_accum == false && ctx->cnt_summed.curobjs == 0) ||
769 (opt_prof_accum && ctx->cnt_summed.accumobjs == 0)) {
770 assert(ctx->cnt_summed.curobjs == 0);
771 assert(ctx->cnt_summed.curbytes == 0);
772 assert(ctx->cnt_summed.accumobjs == 0);
773 assert(ctx->cnt_summed.accumbytes == 0);
774 return (false);
775 }
776
777 if (prof_printf(propagate_err, "%"PRId64": %"PRId64
778 " [%"PRIu64": %"PRIu64"] @",
779 ctx->cnt_summed.curobjs, ctx->cnt_summed.curbytes,
780 ctx->cnt_summed.accumobjs, ctx->cnt_summed.accumbytes))
781 return (true);
782
783 for (i = 0; i < bt->len; i++) {
784 if (prof_printf(propagate_err, " %#"PRIxPTR,
785 (uintptr_t)bt->vec[i]))
786 return (true);
787 }
788
789 if (prof_write(propagate_err, "\n"))
790 return (true);
791
792 return (false);
793 }
794
795 static bool
796 prof_dump_maps(bool propagate_err)
797 {
798 int mfd;
799 char filename[PATH_MAX + 1];
800
801 cassert(config_prof);
802
803 malloc_snprintf(filename, sizeof(filename), "/proc/%d/maps",
804 (int)getpid());
805 mfd = open(filename, O_RDONLY);
806 if (mfd != -1) {
807 ssize_t nread;
808
809 if (prof_write(propagate_err, "\nMAPPED_LIBRARIES:\n") &&
810 propagate_err)
811 return (true);
812 nread = 0;
813 do {
814 prof_dump_buf_end += nread;
815 if (prof_dump_buf_end == PROF_DUMP_BUFSIZE) {
816 /* Make space in prof_dump_buf before read(). */
817 if (prof_flush(propagate_err) && propagate_err)
818 return (true);
819 }
820 nread = read(mfd, &prof_dump_buf[prof_dump_buf_end],
821 PROF_DUMP_BUFSIZE - prof_dump_buf_end);
822 } while (nread > 0);
823 close(mfd);
824 } else
825 return (true);
826
827 return (false);
828 }
829
830 static bool
831 prof_dump(bool propagate_err, const char *filename, bool leakcheck)
832 {
833 prof_tdata_t *prof_tdata;
834 prof_cnt_t cnt_all;
835 size_t tabind;
836 union {
837 prof_bt_t *p;
838 void *v;
839 } bt;
840 union {
841 prof_ctx_t *p;
842 void *v;
843 } ctx;
844 size_t leak_nctx;
845
846 cassert(config_prof);
847
848 prof_tdata = prof_tdata_get();
849 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
850 return (true);
851 prof_enter(prof_tdata);
852 prof_dump_fd = creat(filename, 0644);
853 if (prof_dump_fd == -1) {
854 if (propagate_err == false) {
855 malloc_printf(
856 "<jemalloc>: creat(\"%s\"), 0644) failed\n",
857 filename);
858 if (opt_abort)
859 abort();
860 }
861 goto label_error;
862 }
863
864 /* Merge per thread profile stats, and sum them in cnt_all. */
865 memset(&cnt_all, 0, sizeof(prof_cnt_t));
866 leak_nctx = 0;
867 for (tabind = 0; ckh_iter(&bt2ctx, &tabind, NULL, &ctx.v) == false;)
868 prof_ctx_sum(ctx.p, &cnt_all, &leak_nctx);
869
870 /* Dump profile header. */
871 if (opt_lg_prof_sample == 0) {
872 if (prof_printf(propagate_err,
873 "heap profile: %"PRId64": %"PRId64
874 " [%"PRIu64": %"PRIu64"] @ heapprofile\n",
875 cnt_all.curobjs, cnt_all.curbytes,
876 cnt_all.accumobjs, cnt_all.accumbytes))
877 goto label_error;
878 } else {
879 if (prof_printf(propagate_err,
880 "heap profile: %"PRId64": %"PRId64
881 " [%"PRIu64": %"PRIu64"] @ heap_v2/%"PRIu64"\n",
882 cnt_all.curobjs, cnt_all.curbytes,
883 cnt_all.accumobjs, cnt_all.accumbytes,
884 ((uint64_t)1U << opt_lg_prof_sample)))
885 goto label_error;
886 }
887
888 /* Dump per ctx profile stats. */
889 for (tabind = 0; ckh_iter(&bt2ctx, &tabind, &bt.v, &ctx.v)
890 == false;) {
891 if (prof_dump_ctx(propagate_err, ctx.p, bt.p))
892 goto label_error;
893 }
894
895 /* Dump /proc/<pid>/maps if possible. */
896 if (prof_dump_maps(propagate_err))
897 goto label_error;
898
899 if (prof_flush(propagate_err))
900 goto label_error;
901 close(prof_dump_fd);
902 prof_leave(prof_tdata);
903
904 if (leakcheck && cnt_all.curbytes != 0) {
905 malloc_printf("<jemalloc>: Leak summary: %"PRId64" byte%s, %"
906 PRId64" object%s, %zu context%s\n",
907 cnt_all.curbytes, (cnt_all.curbytes != 1) ? "s" : "",
908 cnt_all.curobjs, (cnt_all.curobjs != 1) ? "s" : "",
909 leak_nctx, (leak_nctx != 1) ? "s" : "");
910 malloc_printf(
911 "<jemalloc>: Run pprof on \"%s\" for leak detail\n",
912 filename);
913 }
914
915 return (false);
916 label_error:
917 prof_leave(prof_tdata);
918 return (true);
919 }
920
921 #define DUMP_FILENAME_BUFSIZE (PATH_MAX + 1)
922 static void
923 prof_dump_filename(char *filename, char v, int64_t vseq)
924 {
925
926 cassert(config_prof);
927
928 if (vseq != UINT64_C(0xffffffffffffffff)) {
929 /* "<prefix>.<pid>.<seq>.v<vseq>.heap" */
930 malloc_snprintf(filename, DUMP_FILENAME_BUFSIZE,
931 "%s.%d.%"PRIu64".%c%"PRId64".heap",
932 opt_prof_prefix, (int)getpid(), prof_dump_seq, v, vseq);
933 } else {
934 /* "<prefix>.<pid>.<seq>.<v>.heap" */
935 malloc_snprintf(filename, DUMP_FILENAME_BUFSIZE,
936 "%s.%d.%"PRIu64".%c.heap",
937 opt_prof_prefix, (int)getpid(), prof_dump_seq, v);
938 }
939 prof_dump_seq++;
940 }
941
942 static void
943 prof_fdump(void)
944 {
945 char filename[DUMP_FILENAME_BUFSIZE];
946
947 cassert(config_prof);
948
949 if (prof_booted == false)
950 return;
951
952 if (opt_prof_final && opt_prof_prefix[0] != '\0') {
953 malloc_mutex_lock(&prof_dump_seq_mtx);
954 prof_dump_filename(filename, 'f', UINT64_C(0xffffffffffffffff));
955 malloc_mutex_unlock(&prof_dump_seq_mtx);
956 prof_dump(false, filename, opt_prof_leak);
957 }
958 }
959
960 void
961 prof_idump(void)
962 {
963 prof_tdata_t *prof_tdata;
964 char filename[PATH_MAX + 1];
965
966 cassert(config_prof);
967
968 if (prof_booted == false)
969 return;
970 /*
971 * Don't call prof_tdata_get() here, because it could cause recursive
972 * allocation.
973 */
974 prof_tdata = *prof_tdata_tsd_get();
975 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
976 return;
977 if (prof_tdata->enq) {
978 prof_tdata->enq_idump = true;
979 return;
980 }
981
982 if (opt_prof_prefix[0] != '\0') {
983 malloc_mutex_lock(&prof_dump_seq_mtx);
984 prof_dump_filename(filename, 'i', prof_dump_iseq);
985 prof_dump_iseq++;
986 malloc_mutex_unlock(&prof_dump_seq_mtx);
987 prof_dump(false, filename, false);
988 }
989 }
990
991 bool
992 prof_mdump(const char *filename)
993 {
994 char filename_buf[DUMP_FILENAME_BUFSIZE];
995
996 cassert(config_prof);
997
998 if (opt_prof == false || prof_booted == false)
999 return (true);
1000
1001 if (filename == NULL) {
1002 /* No filename specified, so automatically generate one. */
1003 if (opt_prof_prefix[0] == '\0')
1004 return (true);
1005 malloc_mutex_lock(&prof_dump_seq_mtx);
1006 prof_dump_filename(filename_buf, 'm', prof_dump_mseq);
1007 prof_dump_mseq++;
1008 malloc_mutex_unlock(&prof_dump_seq_mtx);
1009 filename = filename_buf;
1010 }
1011 return (prof_dump(true, filename, false));
1012 }
1013
1014 void
1015 prof_gdump(void)
1016 {
1017 prof_tdata_t *prof_tdata;
1018 char filename[DUMP_FILENAME_BUFSIZE];
1019
1020 cassert(config_prof);
1021
1022 if (prof_booted == false)
1023 return;
1024 /*
1025 * Don't call prof_tdata_get() here, because it could cause recursive
1026 * allocation.
1027 */
1028 prof_tdata = *prof_tdata_tsd_get();
1029 if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
1030 return;
1031 if (prof_tdata->enq) {
1032 prof_tdata->enq_gdump = true;
1033 return;
1034 }
1035
1036 if (opt_prof_prefix[0] != '\0') {
1037 malloc_mutex_lock(&prof_dump_seq_mtx);
1038 prof_dump_filename(filename, 'u', prof_dump_useq);
1039 prof_dump_useq++;
1040 malloc_mutex_unlock(&prof_dump_seq_mtx);
1041 prof_dump(false, filename, false);
1042 }
1043 }
1044
1045 static void
1046 prof_bt_hash(const void *key, unsigned minbits, size_t *hash1, size_t *hash2)
1047 {
1048 size_t ret1, ret2;
1049 uint64_t h;
1050 prof_bt_t *bt = (prof_bt_t *)key;
1051
1052 cassert(config_prof);
1053 assert(minbits <= 32 || (SIZEOF_PTR == 8 && minbits <= 64));
1054 assert(hash1 != NULL);
1055 assert(hash2 != NULL);
1056
1057 h = hash(bt->vec, bt->len * sizeof(void *),
1058 UINT64_C(0x94122f335b332aea));
1059 if (minbits <= 32) {
1060 /*
1061 * Avoid doing multiple hashes, since a single hash provides
1062 * enough bits.
1063 */
1064 ret1 = h & ZU(0xffffffffU);
1065 ret2 = h >> 32;
1066 } else {
1067 ret1 = h;
1068 ret2 = hash(bt->vec, bt->len * sizeof(void *),
1069 UINT64_C(0x8432a476666bbc13));
1070 }
1071
1072 *hash1 = ret1;
1073 *hash2 = ret2;
1074 }
1075
1076 static bool
1077 prof_bt_keycomp(const void *k1, const void *k2)
1078 {
1079 const prof_bt_t *bt1 = (prof_bt_t *)k1;
1080 const prof_bt_t *bt2 = (prof_bt_t *)k2;
1081
1082 cassert(config_prof);
1083
1084 if (bt1->len != bt2->len)
1085 return (false);
1086 return (memcmp(bt1->vec, bt2->vec, bt1->len * sizeof(void *)) == 0);
1087 }
1088
1089 static malloc_mutex_t *
1090 prof_ctx_mutex_choose(void)
1091 {
1092 unsigned nctxs = atomic_add_u(&cum_ctxs, 1);
1093
1094 return (&ctx_locks[(nctxs - 1) % PROF_NCTX_LOCKS]);
1095 }
1096
1097 prof_tdata_t *
1098 prof_tdata_init(void)
1099 {
1100 prof_tdata_t *prof_tdata;
1101
1102 cassert(config_prof);
1103
1104 /* Initialize an empty cache for this thread. */
1105 prof_tdata = (prof_tdata_t *)imalloc(sizeof(prof_tdata_t));
1106 if (prof_tdata == NULL)
1107 return (NULL);
1108
1109 if (ckh_new(&prof_tdata->bt2cnt, PROF_CKH_MINITEMS,
1110 prof_bt_hash, prof_bt_keycomp)) {
1111 idalloc(prof_tdata);
1112 return (NULL);
1113 }
1114 ql_new(&prof_tdata->lru_ql);
1115
1116 prof_tdata->vec = imalloc(sizeof(void *) * PROF_BT_MAX);
1117 if (prof_tdata->vec == NULL) {
1118 ckh_delete(&prof_tdata->bt2cnt);
1119 idalloc(prof_tdata);
1120 return (NULL);
1121 }
1122
1123 prof_tdata->prng_state = 0;
1124 prof_tdata->threshold = 0;
1125 prof_tdata->accum = 0;
1126
1127 prof_tdata->enq = false;
1128 prof_tdata->enq_idump = false;
1129 prof_tdata->enq_gdump = false;
1130
1131 prof_tdata_tsd_set(&prof_tdata);
1132
1133 return (prof_tdata);
1134 }
1135
1136 void
1137 prof_tdata_cleanup(void *arg)
1138 {
1139 prof_thr_cnt_t *cnt;
1140 prof_tdata_t *prof_tdata = *(prof_tdata_t **)arg;
1141
1142 cassert(config_prof);
1143
1144 if (prof_tdata == PROF_TDATA_STATE_REINCARNATED) {
1145 /*
1146 * Another destructor deallocated memory after this destructor
1147 * was called. Reset prof_tdata to PROF_TDATA_STATE_PURGATORY
1148 * in order to receive another callback.
1149 */
1150 prof_tdata = PROF_TDATA_STATE_PURGATORY;
1151 prof_tdata_tsd_set(&prof_tdata);
1152 } else if (prof_tdata == PROF_TDATA_STATE_PURGATORY) {
1153 /*
1154 * The previous time this destructor was called, we set the key
1155 * to PROF_TDATA_STATE_PURGATORY so that other destructors
1156 * wouldn't cause re-creation of the prof_tdata. This time, do
1157 * nothing, so that the destructor will not be called again.
1158 */
1159 } else if (prof_tdata != NULL) {
1160 /*
1161 * Delete the hash table. All of its contents can still be
1162 * iterated over via the LRU.
1163 */
1164 ckh_delete(&prof_tdata->bt2cnt);
1165 /*
1166 * Iteratively merge cnt's into the global stats and delete
1167 * them.
1168 */
1169 while ((cnt = ql_last(&prof_tdata->lru_ql, lru_link)) != NULL) {
1170 ql_remove(&prof_tdata->lru_ql, cnt, lru_link);
1171 prof_ctx_merge(cnt->ctx, cnt);
1172 idalloc(cnt);
1173 }
1174 idalloc(prof_tdata->vec);
1175 idalloc(prof_tdata);
1176 prof_tdata = PROF_TDATA_STATE_PURGATORY;
1177 prof_tdata_tsd_set(&prof_tdata);
1178 }
1179 }
1180
1181 void
1182 prof_boot0(void)
1183 {
1184
1185 cassert(config_prof);
1186
1187 memcpy(opt_prof_prefix, PROF_PREFIX_DEFAULT,
1188 sizeof(PROF_PREFIX_DEFAULT));
1189 }
1190
1191 void
1192 prof_boot1(void)
1193 {
1194
1195 cassert(config_prof);
1196
1197 /*
1198 * opt_prof and prof_promote must be in their final state before any
1199 * arenas are initialized, so this function must be executed early.
1200 */
1201
1202 if (opt_prof_leak && opt_prof == false) {
1203 /*
1204 * Enable opt_prof, but in such a way that profiles are never
1205 * automatically dumped.
1206 */
1207 opt_prof = true;
1208 opt_prof_gdump = false;
1209 prof_interval = 0;
1210 } else if (opt_prof) {
1211 if (opt_lg_prof_interval >= 0) {
1212 prof_interval = (((uint64_t)1U) <<
1213 opt_lg_prof_interval);
1214 } else
1215 prof_interval = 0;
1216 }
1217
1218 prof_promote = (opt_prof && opt_lg_prof_sample > LG_PAGE);
1219 }
1220
1221 bool
1222 prof_boot2(void)
1223 {
1224
1225 cassert(config_prof);
1226
1227 if (opt_prof) {
1228 unsigned i;
1229
1230 if (ckh_new(&bt2ctx, PROF_CKH_MINITEMS, prof_bt_hash,
1231 prof_bt_keycomp))
1232 return (true);
1233 if (malloc_mutex_init(&bt2ctx_mtx))
1234 return (true);
1235 if (prof_tdata_tsd_boot()) {
1236 malloc_write(
1237 "<jemalloc>: Error in pthread_key_create()\n");
1238 abort();
1239 }
1240
1241 if (malloc_mutex_init(&prof_dump_seq_mtx))
1242 return (true);
1243
1244 if (atexit(prof_fdump) != 0) {
1245 malloc_write("<jemalloc>: Error in atexit()\n");
1246 if (opt_abort)
1247 abort();
1248 }
1249
1250 ctx_locks = (malloc_mutex_t *)base_alloc(PROF_NCTX_LOCKS *
1251 sizeof(malloc_mutex_t));
1252 if (ctx_locks == NULL)
1253 return (true);
1254 for (i = 0; i < PROF_NCTX_LOCKS; i++) {
1255 if (malloc_mutex_init(&ctx_locks[i]))
1256 return (true);
1257 }
1258 }
1259
1260 #ifdef JEMALLOC_PROF_LIBGCC
1261 /*
1262 * Cause the backtracing machinery to allocate its internal state
1263 * before enabling profiling.
1264 */
1265 _Unwind_Backtrace(prof_unwind_init_callback, NULL);
1266 #endif
1267
1268 prof_booted = true;
1269
1270 return (false);
1271 }
1272
1273 void
1274 prof_prefork(void)
1275 {
1276
1277 if (opt_prof) {
1278 unsigned i;
1279
1280 malloc_mutex_lock(&bt2ctx_mtx);
1281 malloc_mutex_lock(&prof_dump_seq_mtx);
1282 for (i = 0; i < PROF_NCTX_LOCKS; i++)
1283 malloc_mutex_lock(&ctx_locks[i]);
1284 }
1285 }
1286
1287 void
1288 prof_postfork_parent(void)
1289 {
1290
1291 if (opt_prof) {
1292 unsigned i;
1293
1294 for (i = 0; i < PROF_NCTX_LOCKS; i++)
1295 malloc_mutex_postfork_parent(&ctx_locks[i]);
1296 malloc_mutex_postfork_parent(&prof_dump_seq_mtx);
1297 malloc_mutex_postfork_parent(&bt2ctx_mtx);
1298 }
1299 }
1300
1301 void
1302 prof_postfork_child(void)
1303 {
1304
1305 if (opt_prof) {
1306 unsigned i;
1307
1308 for (i = 0; i < PROF_NCTX_LOCKS; i++)
1309 malloc_mutex_postfork_child(&ctx_locks[i]);
1310 malloc_mutex_postfork_child(&prof_dump_seq_mtx);
1311 malloc_mutex_postfork_child(&bt2ctx_mtx);
1312 }
1313 }
1314
1315 /******************************************************************************/