]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - tools/perf/util/hist.c
Merge branch 'for-rafael' of https://git.kernel.org/pub/scm/linux/kernel/git/mzx...
[mirror_ubuntu-artful-kernel.git] / tools / perf / util / hist.c
1 #include "util.h"
2 #include "build-id.h"
3 #include "hist.h"
4 #include "session.h"
5 #include "sort.h"
6 #include "evlist.h"
7 #include "evsel.h"
8 #include "annotate.h"
9 #include "ui/progress.h"
10 #include <math.h>
11
12 static bool hists__filter_entry_by_dso(struct hists *hists,
13 struct hist_entry *he);
14 static bool hists__filter_entry_by_thread(struct hists *hists,
15 struct hist_entry *he);
16 static bool hists__filter_entry_by_symbol(struct hists *hists,
17 struct hist_entry *he);
18
19 u16 hists__col_len(struct hists *hists, enum hist_column col)
20 {
21 return hists->col_len[col];
22 }
23
24 void hists__set_col_len(struct hists *hists, enum hist_column col, u16 len)
25 {
26 hists->col_len[col] = len;
27 }
28
29 bool hists__new_col_len(struct hists *hists, enum hist_column col, u16 len)
30 {
31 if (len > hists__col_len(hists, col)) {
32 hists__set_col_len(hists, col, len);
33 return true;
34 }
35 return false;
36 }
37
38 void hists__reset_col_len(struct hists *hists)
39 {
40 enum hist_column col;
41
42 for (col = 0; col < HISTC_NR_COLS; ++col)
43 hists__set_col_len(hists, col, 0);
44 }
45
46 static void hists__set_unres_dso_col_len(struct hists *hists, int dso)
47 {
48 const unsigned int unresolved_col_width = BITS_PER_LONG / 4;
49
50 if (hists__col_len(hists, dso) < unresolved_col_width &&
51 !symbol_conf.col_width_list_str && !symbol_conf.field_sep &&
52 !symbol_conf.dso_list)
53 hists__set_col_len(hists, dso, unresolved_col_width);
54 }
55
56 void hists__calc_col_len(struct hists *hists, struct hist_entry *h)
57 {
58 const unsigned int unresolved_col_width = BITS_PER_LONG / 4;
59 int symlen;
60 u16 len;
61
62 /*
63 * +4 accounts for '[x] ' priv level info
64 * +2 accounts for 0x prefix on raw addresses
65 * +3 accounts for ' y ' symtab origin info
66 */
67 if (h->ms.sym) {
68 symlen = h->ms.sym->namelen + 4;
69 if (verbose)
70 symlen += BITS_PER_LONG / 4 + 2 + 3;
71 hists__new_col_len(hists, HISTC_SYMBOL, symlen);
72 } else {
73 symlen = unresolved_col_width + 4 + 2;
74 hists__new_col_len(hists, HISTC_SYMBOL, symlen);
75 hists__set_unres_dso_col_len(hists, HISTC_DSO);
76 }
77
78 len = thread__comm_len(h->thread);
79 if (hists__new_col_len(hists, HISTC_COMM, len))
80 hists__set_col_len(hists, HISTC_THREAD, len + 6);
81
82 if (h->ms.map) {
83 len = dso__name_len(h->ms.map->dso);
84 hists__new_col_len(hists, HISTC_DSO, len);
85 }
86
87 if (h->parent)
88 hists__new_col_len(hists, HISTC_PARENT, h->parent->namelen);
89
90 if (h->branch_info) {
91 if (h->branch_info->from.sym) {
92 symlen = (int)h->branch_info->from.sym->namelen + 4;
93 if (verbose)
94 symlen += BITS_PER_LONG / 4 + 2 + 3;
95 hists__new_col_len(hists, HISTC_SYMBOL_FROM, symlen);
96
97 symlen = dso__name_len(h->branch_info->from.map->dso);
98 hists__new_col_len(hists, HISTC_DSO_FROM, symlen);
99 } else {
100 symlen = unresolved_col_width + 4 + 2;
101 hists__new_col_len(hists, HISTC_SYMBOL_FROM, symlen);
102 hists__set_unres_dso_col_len(hists, HISTC_DSO_FROM);
103 }
104
105 if (h->branch_info->to.sym) {
106 symlen = (int)h->branch_info->to.sym->namelen + 4;
107 if (verbose)
108 symlen += BITS_PER_LONG / 4 + 2 + 3;
109 hists__new_col_len(hists, HISTC_SYMBOL_TO, symlen);
110
111 symlen = dso__name_len(h->branch_info->to.map->dso);
112 hists__new_col_len(hists, HISTC_DSO_TO, symlen);
113 } else {
114 symlen = unresolved_col_width + 4 + 2;
115 hists__new_col_len(hists, HISTC_SYMBOL_TO, symlen);
116 hists__set_unres_dso_col_len(hists, HISTC_DSO_TO);
117 }
118 }
119
120 if (h->mem_info) {
121 if (h->mem_info->daddr.sym) {
122 symlen = (int)h->mem_info->daddr.sym->namelen + 4
123 + unresolved_col_width + 2;
124 hists__new_col_len(hists, HISTC_MEM_DADDR_SYMBOL,
125 symlen);
126 hists__new_col_len(hists, HISTC_MEM_DCACHELINE,
127 symlen + 1);
128 } else {
129 symlen = unresolved_col_width + 4 + 2;
130 hists__new_col_len(hists, HISTC_MEM_DADDR_SYMBOL,
131 symlen);
132 }
133 if (h->mem_info->daddr.map) {
134 symlen = dso__name_len(h->mem_info->daddr.map->dso);
135 hists__new_col_len(hists, HISTC_MEM_DADDR_DSO,
136 symlen);
137 } else {
138 symlen = unresolved_col_width + 4 + 2;
139 hists__set_unres_dso_col_len(hists, HISTC_MEM_DADDR_DSO);
140 }
141 } else {
142 symlen = unresolved_col_width + 4 + 2;
143 hists__new_col_len(hists, HISTC_MEM_DADDR_SYMBOL, symlen);
144 hists__set_unres_dso_col_len(hists, HISTC_MEM_DADDR_DSO);
145 }
146
147 hists__new_col_len(hists, HISTC_MEM_LOCKED, 6);
148 hists__new_col_len(hists, HISTC_MEM_TLB, 22);
149 hists__new_col_len(hists, HISTC_MEM_SNOOP, 12);
150 hists__new_col_len(hists, HISTC_MEM_LVL, 21 + 3);
151 hists__new_col_len(hists, HISTC_LOCAL_WEIGHT, 12);
152 hists__new_col_len(hists, HISTC_GLOBAL_WEIGHT, 12);
153
154 if (h->srcline)
155 hists__new_col_len(hists, HISTC_SRCLINE, strlen(h->srcline));
156
157 if (h->srcfile)
158 hists__new_col_len(hists, HISTC_SRCFILE, strlen(h->srcfile));
159
160 if (h->transaction)
161 hists__new_col_len(hists, HISTC_TRANSACTION,
162 hist_entry__transaction_len());
163 }
164
165 void hists__output_recalc_col_len(struct hists *hists, int max_rows)
166 {
167 struct rb_node *next = rb_first(&hists->entries);
168 struct hist_entry *n;
169 int row = 0;
170
171 hists__reset_col_len(hists);
172
173 while (next && row++ < max_rows) {
174 n = rb_entry(next, struct hist_entry, rb_node);
175 if (!n->filtered)
176 hists__calc_col_len(hists, n);
177 next = rb_next(&n->rb_node);
178 }
179 }
180
181 static void he_stat__add_cpumode_period(struct he_stat *he_stat,
182 unsigned int cpumode, u64 period)
183 {
184 switch (cpumode) {
185 case PERF_RECORD_MISC_KERNEL:
186 he_stat->period_sys += period;
187 break;
188 case PERF_RECORD_MISC_USER:
189 he_stat->period_us += period;
190 break;
191 case PERF_RECORD_MISC_GUEST_KERNEL:
192 he_stat->period_guest_sys += period;
193 break;
194 case PERF_RECORD_MISC_GUEST_USER:
195 he_stat->period_guest_us += period;
196 break;
197 default:
198 break;
199 }
200 }
201
202 static void he_stat__add_period(struct he_stat *he_stat, u64 period,
203 u64 weight)
204 {
205
206 he_stat->period += period;
207 he_stat->weight += weight;
208 he_stat->nr_events += 1;
209 }
210
211 static void he_stat__add_stat(struct he_stat *dest, struct he_stat *src)
212 {
213 dest->period += src->period;
214 dest->period_sys += src->period_sys;
215 dest->period_us += src->period_us;
216 dest->period_guest_sys += src->period_guest_sys;
217 dest->period_guest_us += src->period_guest_us;
218 dest->nr_events += src->nr_events;
219 dest->weight += src->weight;
220 }
221
222 static void he_stat__decay(struct he_stat *he_stat)
223 {
224 he_stat->period = (he_stat->period * 7) / 8;
225 he_stat->nr_events = (he_stat->nr_events * 7) / 8;
226 /* XXX need decay for weight too? */
227 }
228
229 static bool hists__decay_entry(struct hists *hists, struct hist_entry *he)
230 {
231 u64 prev_period = he->stat.period;
232 u64 diff;
233
234 if (prev_period == 0)
235 return true;
236
237 he_stat__decay(&he->stat);
238 if (symbol_conf.cumulate_callchain)
239 he_stat__decay(he->stat_acc);
240
241 diff = prev_period - he->stat.period;
242
243 hists->stats.total_period -= diff;
244 if (!he->filtered)
245 hists->stats.total_non_filtered_period -= diff;
246
247 return he->stat.period == 0;
248 }
249
250 static void hists__delete_entry(struct hists *hists, struct hist_entry *he)
251 {
252 rb_erase(&he->rb_node, &hists->entries);
253
254 if (sort__need_collapse)
255 rb_erase(&he->rb_node_in, &hists->entries_collapsed);
256
257 --hists->nr_entries;
258 if (!he->filtered)
259 --hists->nr_non_filtered_entries;
260
261 hist_entry__delete(he);
262 }
263
264 void hists__decay_entries(struct hists *hists, bool zap_user, bool zap_kernel)
265 {
266 struct rb_node *next = rb_first(&hists->entries);
267 struct hist_entry *n;
268
269 while (next) {
270 n = rb_entry(next, struct hist_entry, rb_node);
271 next = rb_next(&n->rb_node);
272 if (((zap_user && n->level == '.') ||
273 (zap_kernel && n->level != '.') ||
274 hists__decay_entry(hists, n))) {
275 hists__delete_entry(hists, n);
276 }
277 }
278 }
279
280 void hists__delete_entries(struct hists *hists)
281 {
282 struct rb_node *next = rb_first(&hists->entries);
283 struct hist_entry *n;
284
285 while (next) {
286 n = rb_entry(next, struct hist_entry, rb_node);
287 next = rb_next(&n->rb_node);
288
289 hists__delete_entry(hists, n);
290 }
291 }
292
293 /*
294 * histogram, sorted on item, collects periods
295 */
296
297 static struct hist_entry *hist_entry__new(struct hist_entry *template,
298 bool sample_self)
299 {
300 size_t callchain_size = 0;
301 struct hist_entry *he;
302
303 if (symbol_conf.use_callchain)
304 callchain_size = sizeof(struct callchain_root);
305
306 he = zalloc(sizeof(*he) + callchain_size);
307
308 if (he != NULL) {
309 *he = *template;
310
311 if (symbol_conf.cumulate_callchain) {
312 he->stat_acc = malloc(sizeof(he->stat));
313 if (he->stat_acc == NULL) {
314 free(he);
315 return NULL;
316 }
317 memcpy(he->stat_acc, &he->stat, sizeof(he->stat));
318 if (!sample_self)
319 memset(&he->stat, 0, sizeof(he->stat));
320 }
321
322 map__get(he->ms.map);
323
324 if (he->branch_info) {
325 /*
326 * This branch info is (a part of) allocated from
327 * sample__resolve_bstack() and will be freed after
328 * adding new entries. So we need to save a copy.
329 */
330 he->branch_info = malloc(sizeof(*he->branch_info));
331 if (he->branch_info == NULL) {
332 map__zput(he->ms.map);
333 free(he->stat_acc);
334 free(he);
335 return NULL;
336 }
337
338 memcpy(he->branch_info, template->branch_info,
339 sizeof(*he->branch_info));
340
341 map__get(he->branch_info->from.map);
342 map__get(he->branch_info->to.map);
343 }
344
345 if (he->mem_info) {
346 map__get(he->mem_info->iaddr.map);
347 map__get(he->mem_info->daddr.map);
348 }
349
350 if (symbol_conf.use_callchain)
351 callchain_init(he->callchain);
352
353 INIT_LIST_HEAD(&he->pairs.node);
354 thread__get(he->thread);
355 }
356
357 return he;
358 }
359
360 static u8 symbol__parent_filter(const struct symbol *parent)
361 {
362 if (symbol_conf.exclude_other && parent == NULL)
363 return 1 << HIST_FILTER__PARENT;
364 return 0;
365 }
366
367 static struct hist_entry *hists__findnew_entry(struct hists *hists,
368 struct hist_entry *entry,
369 struct addr_location *al,
370 bool sample_self)
371 {
372 struct rb_node **p;
373 struct rb_node *parent = NULL;
374 struct hist_entry *he;
375 int64_t cmp;
376 u64 period = entry->stat.period;
377 u64 weight = entry->stat.weight;
378
379 p = &hists->entries_in->rb_node;
380
381 while (*p != NULL) {
382 parent = *p;
383 he = rb_entry(parent, struct hist_entry, rb_node_in);
384
385 /*
386 * Make sure that it receives arguments in a same order as
387 * hist_entry__collapse() so that we can use an appropriate
388 * function when searching an entry regardless which sort
389 * keys were used.
390 */
391 cmp = hist_entry__cmp(he, entry);
392
393 if (!cmp) {
394 if (sample_self)
395 he_stat__add_period(&he->stat, period, weight);
396 if (symbol_conf.cumulate_callchain)
397 he_stat__add_period(he->stat_acc, period, weight);
398
399 /*
400 * This mem info was allocated from sample__resolve_mem
401 * and will not be used anymore.
402 */
403 zfree(&entry->mem_info);
404
405 /* If the map of an existing hist_entry has
406 * become out-of-date due to an exec() or
407 * similar, update it. Otherwise we will
408 * mis-adjust symbol addresses when computing
409 * the history counter to increment.
410 */
411 if (he->ms.map != entry->ms.map) {
412 map__put(he->ms.map);
413 he->ms.map = map__get(entry->ms.map);
414 }
415 goto out;
416 }
417
418 if (cmp < 0)
419 p = &(*p)->rb_left;
420 else
421 p = &(*p)->rb_right;
422 }
423
424 he = hist_entry__new(entry, sample_self);
425 if (!he)
426 return NULL;
427
428 hists->nr_entries++;
429
430 rb_link_node(&he->rb_node_in, parent, p);
431 rb_insert_color(&he->rb_node_in, hists->entries_in);
432 out:
433 if (sample_self)
434 he_stat__add_cpumode_period(&he->stat, al->cpumode, period);
435 if (symbol_conf.cumulate_callchain)
436 he_stat__add_cpumode_period(he->stat_acc, al->cpumode, period);
437 return he;
438 }
439
440 struct hist_entry *__hists__add_entry(struct hists *hists,
441 struct addr_location *al,
442 struct symbol *sym_parent,
443 struct branch_info *bi,
444 struct mem_info *mi,
445 u64 period, u64 weight, u64 transaction,
446 bool sample_self)
447 {
448 struct hist_entry entry = {
449 .thread = al->thread,
450 .comm = thread__comm(al->thread),
451 .ms = {
452 .map = al->map,
453 .sym = al->sym,
454 },
455 .cpu = al->cpu,
456 .cpumode = al->cpumode,
457 .ip = al->addr,
458 .level = al->level,
459 .stat = {
460 .nr_events = 1,
461 .period = period,
462 .weight = weight,
463 },
464 .parent = sym_parent,
465 .filtered = symbol__parent_filter(sym_parent) | al->filtered,
466 .hists = hists,
467 .branch_info = bi,
468 .mem_info = mi,
469 .transaction = transaction,
470 };
471
472 return hists__findnew_entry(hists, &entry, al, sample_self);
473 }
474
475 static int
476 iter_next_nop_entry(struct hist_entry_iter *iter __maybe_unused,
477 struct addr_location *al __maybe_unused)
478 {
479 return 0;
480 }
481
482 static int
483 iter_add_next_nop_entry(struct hist_entry_iter *iter __maybe_unused,
484 struct addr_location *al __maybe_unused)
485 {
486 return 0;
487 }
488
489 static int
490 iter_prepare_mem_entry(struct hist_entry_iter *iter, struct addr_location *al)
491 {
492 struct perf_sample *sample = iter->sample;
493 struct mem_info *mi;
494
495 mi = sample__resolve_mem(sample, al);
496 if (mi == NULL)
497 return -ENOMEM;
498
499 iter->priv = mi;
500 return 0;
501 }
502
503 static int
504 iter_add_single_mem_entry(struct hist_entry_iter *iter, struct addr_location *al)
505 {
506 u64 cost;
507 struct mem_info *mi = iter->priv;
508 struct hists *hists = evsel__hists(iter->evsel);
509 struct hist_entry *he;
510
511 if (mi == NULL)
512 return -EINVAL;
513
514 cost = iter->sample->weight;
515 if (!cost)
516 cost = 1;
517
518 /*
519 * must pass period=weight in order to get the correct
520 * sorting from hists__collapse_resort() which is solely
521 * based on periods. We want sorting be done on nr_events * weight
522 * and this is indirectly achieved by passing period=weight here
523 * and the he_stat__add_period() function.
524 */
525 he = __hists__add_entry(hists, al, iter->parent, NULL, mi,
526 cost, cost, 0, true);
527 if (!he)
528 return -ENOMEM;
529
530 iter->he = he;
531 return 0;
532 }
533
534 static int
535 iter_finish_mem_entry(struct hist_entry_iter *iter,
536 struct addr_location *al __maybe_unused)
537 {
538 struct perf_evsel *evsel = iter->evsel;
539 struct hists *hists = evsel__hists(evsel);
540 struct hist_entry *he = iter->he;
541 int err = -EINVAL;
542
543 if (he == NULL)
544 goto out;
545
546 hists__inc_nr_samples(hists, he->filtered);
547
548 err = hist_entry__append_callchain(he, iter->sample);
549
550 out:
551 /*
552 * We don't need to free iter->priv (mem_info) here since the mem info
553 * was either already freed in hists__findnew_entry() or passed to a
554 * new hist entry by hist_entry__new().
555 */
556 iter->priv = NULL;
557
558 iter->he = NULL;
559 return err;
560 }
561
562 static int
563 iter_prepare_branch_entry(struct hist_entry_iter *iter, struct addr_location *al)
564 {
565 struct branch_info *bi;
566 struct perf_sample *sample = iter->sample;
567
568 bi = sample__resolve_bstack(sample, al);
569 if (!bi)
570 return -ENOMEM;
571
572 iter->curr = 0;
573 iter->total = sample->branch_stack->nr;
574
575 iter->priv = bi;
576 return 0;
577 }
578
579 static int
580 iter_add_single_branch_entry(struct hist_entry_iter *iter __maybe_unused,
581 struct addr_location *al __maybe_unused)
582 {
583 /* to avoid calling callback function */
584 iter->he = NULL;
585
586 return 0;
587 }
588
589 static int
590 iter_next_branch_entry(struct hist_entry_iter *iter, struct addr_location *al)
591 {
592 struct branch_info *bi = iter->priv;
593 int i = iter->curr;
594
595 if (bi == NULL)
596 return 0;
597
598 if (iter->curr >= iter->total)
599 return 0;
600
601 al->map = bi[i].to.map;
602 al->sym = bi[i].to.sym;
603 al->addr = bi[i].to.addr;
604 return 1;
605 }
606
607 static int
608 iter_add_next_branch_entry(struct hist_entry_iter *iter, struct addr_location *al)
609 {
610 struct branch_info *bi;
611 struct perf_evsel *evsel = iter->evsel;
612 struct hists *hists = evsel__hists(evsel);
613 struct hist_entry *he = NULL;
614 int i = iter->curr;
615 int err = 0;
616
617 bi = iter->priv;
618
619 if (iter->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
620 goto out;
621
622 /*
623 * The report shows the percentage of total branches captured
624 * and not events sampled. Thus we use a pseudo period of 1.
625 */
626 he = __hists__add_entry(hists, al, iter->parent, &bi[i], NULL,
627 1, bi->flags.cycles ? bi->flags.cycles : 1,
628 0, true);
629 if (he == NULL)
630 return -ENOMEM;
631
632 hists__inc_nr_samples(hists, he->filtered);
633
634 out:
635 iter->he = he;
636 iter->curr++;
637 return err;
638 }
639
640 static int
641 iter_finish_branch_entry(struct hist_entry_iter *iter,
642 struct addr_location *al __maybe_unused)
643 {
644 zfree(&iter->priv);
645 iter->he = NULL;
646
647 return iter->curr >= iter->total ? 0 : -1;
648 }
649
650 static int
651 iter_prepare_normal_entry(struct hist_entry_iter *iter __maybe_unused,
652 struct addr_location *al __maybe_unused)
653 {
654 return 0;
655 }
656
657 static int
658 iter_add_single_normal_entry(struct hist_entry_iter *iter, struct addr_location *al)
659 {
660 struct perf_evsel *evsel = iter->evsel;
661 struct perf_sample *sample = iter->sample;
662 struct hist_entry *he;
663
664 he = __hists__add_entry(evsel__hists(evsel), al, iter->parent, NULL, NULL,
665 sample->period, sample->weight,
666 sample->transaction, true);
667 if (he == NULL)
668 return -ENOMEM;
669
670 iter->he = he;
671 return 0;
672 }
673
674 static int
675 iter_finish_normal_entry(struct hist_entry_iter *iter,
676 struct addr_location *al __maybe_unused)
677 {
678 struct hist_entry *he = iter->he;
679 struct perf_evsel *evsel = iter->evsel;
680 struct perf_sample *sample = iter->sample;
681
682 if (he == NULL)
683 return 0;
684
685 iter->he = NULL;
686
687 hists__inc_nr_samples(evsel__hists(evsel), he->filtered);
688
689 return hist_entry__append_callchain(he, sample);
690 }
691
692 static int
693 iter_prepare_cumulative_entry(struct hist_entry_iter *iter __maybe_unused,
694 struct addr_location *al __maybe_unused)
695 {
696 struct hist_entry **he_cache;
697
698 callchain_cursor_commit(&callchain_cursor);
699
700 /*
701 * This is for detecting cycles or recursions so that they're
702 * cumulated only one time to prevent entries more than 100%
703 * overhead.
704 */
705 he_cache = malloc(sizeof(*he_cache) * (PERF_MAX_STACK_DEPTH + 1));
706 if (he_cache == NULL)
707 return -ENOMEM;
708
709 iter->priv = he_cache;
710 iter->curr = 0;
711
712 return 0;
713 }
714
715 static int
716 iter_add_single_cumulative_entry(struct hist_entry_iter *iter,
717 struct addr_location *al)
718 {
719 struct perf_evsel *evsel = iter->evsel;
720 struct hists *hists = evsel__hists(evsel);
721 struct perf_sample *sample = iter->sample;
722 struct hist_entry **he_cache = iter->priv;
723 struct hist_entry *he;
724 int err = 0;
725
726 he = __hists__add_entry(hists, al, iter->parent, NULL, NULL,
727 sample->period, sample->weight,
728 sample->transaction, true);
729 if (he == NULL)
730 return -ENOMEM;
731
732 iter->he = he;
733 he_cache[iter->curr++] = he;
734
735 hist_entry__append_callchain(he, sample);
736
737 /*
738 * We need to re-initialize the cursor since callchain_append()
739 * advanced the cursor to the end.
740 */
741 callchain_cursor_commit(&callchain_cursor);
742
743 hists__inc_nr_samples(hists, he->filtered);
744
745 return err;
746 }
747
748 static int
749 iter_next_cumulative_entry(struct hist_entry_iter *iter,
750 struct addr_location *al)
751 {
752 struct callchain_cursor_node *node;
753
754 node = callchain_cursor_current(&callchain_cursor);
755 if (node == NULL)
756 return 0;
757
758 return fill_callchain_info(al, node, iter->hide_unresolved);
759 }
760
761 static int
762 iter_add_next_cumulative_entry(struct hist_entry_iter *iter,
763 struct addr_location *al)
764 {
765 struct perf_evsel *evsel = iter->evsel;
766 struct perf_sample *sample = iter->sample;
767 struct hist_entry **he_cache = iter->priv;
768 struct hist_entry *he;
769 struct hist_entry he_tmp = {
770 .hists = evsel__hists(evsel),
771 .cpu = al->cpu,
772 .thread = al->thread,
773 .comm = thread__comm(al->thread),
774 .ip = al->addr,
775 .ms = {
776 .map = al->map,
777 .sym = al->sym,
778 },
779 .parent = iter->parent,
780 };
781 int i;
782 struct callchain_cursor cursor;
783
784 callchain_cursor_snapshot(&cursor, &callchain_cursor);
785
786 callchain_cursor_advance(&callchain_cursor);
787
788 /*
789 * Check if there's duplicate entries in the callchain.
790 * It's possible that it has cycles or recursive calls.
791 */
792 for (i = 0; i < iter->curr; i++) {
793 if (hist_entry__cmp(he_cache[i], &he_tmp) == 0) {
794 /* to avoid calling callback function */
795 iter->he = NULL;
796 return 0;
797 }
798 }
799
800 he = __hists__add_entry(evsel__hists(evsel), al, iter->parent, NULL, NULL,
801 sample->period, sample->weight,
802 sample->transaction, false);
803 if (he == NULL)
804 return -ENOMEM;
805
806 iter->he = he;
807 he_cache[iter->curr++] = he;
808
809 if (symbol_conf.use_callchain)
810 callchain_append(he->callchain, &cursor, sample->period);
811 return 0;
812 }
813
814 static int
815 iter_finish_cumulative_entry(struct hist_entry_iter *iter,
816 struct addr_location *al __maybe_unused)
817 {
818 zfree(&iter->priv);
819 iter->he = NULL;
820
821 return 0;
822 }
823
824 const struct hist_iter_ops hist_iter_mem = {
825 .prepare_entry = iter_prepare_mem_entry,
826 .add_single_entry = iter_add_single_mem_entry,
827 .next_entry = iter_next_nop_entry,
828 .add_next_entry = iter_add_next_nop_entry,
829 .finish_entry = iter_finish_mem_entry,
830 };
831
832 const struct hist_iter_ops hist_iter_branch = {
833 .prepare_entry = iter_prepare_branch_entry,
834 .add_single_entry = iter_add_single_branch_entry,
835 .next_entry = iter_next_branch_entry,
836 .add_next_entry = iter_add_next_branch_entry,
837 .finish_entry = iter_finish_branch_entry,
838 };
839
840 const struct hist_iter_ops hist_iter_normal = {
841 .prepare_entry = iter_prepare_normal_entry,
842 .add_single_entry = iter_add_single_normal_entry,
843 .next_entry = iter_next_nop_entry,
844 .add_next_entry = iter_add_next_nop_entry,
845 .finish_entry = iter_finish_normal_entry,
846 };
847
848 const struct hist_iter_ops hist_iter_cumulative = {
849 .prepare_entry = iter_prepare_cumulative_entry,
850 .add_single_entry = iter_add_single_cumulative_entry,
851 .next_entry = iter_next_cumulative_entry,
852 .add_next_entry = iter_add_next_cumulative_entry,
853 .finish_entry = iter_finish_cumulative_entry,
854 };
855
856 int hist_entry_iter__add(struct hist_entry_iter *iter, struct addr_location *al,
857 int max_stack_depth, void *arg)
858 {
859 int err, err2;
860
861 err = sample__resolve_callchain(iter->sample, &iter->parent,
862 iter->evsel, al, max_stack_depth);
863 if (err)
864 return err;
865
866 err = iter->ops->prepare_entry(iter, al);
867 if (err)
868 goto out;
869
870 err = iter->ops->add_single_entry(iter, al);
871 if (err)
872 goto out;
873
874 if (iter->he && iter->add_entry_cb) {
875 err = iter->add_entry_cb(iter, al, true, arg);
876 if (err)
877 goto out;
878 }
879
880 while (iter->ops->next_entry(iter, al)) {
881 err = iter->ops->add_next_entry(iter, al);
882 if (err)
883 break;
884
885 if (iter->he && iter->add_entry_cb) {
886 err = iter->add_entry_cb(iter, al, false, arg);
887 if (err)
888 goto out;
889 }
890 }
891
892 out:
893 err2 = iter->ops->finish_entry(iter, al);
894 if (!err)
895 err = err2;
896
897 return err;
898 }
899
900 int64_t
901 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
902 {
903 struct perf_hpp_fmt *fmt;
904 int64_t cmp = 0;
905
906 perf_hpp__for_each_sort_list(fmt) {
907 if (perf_hpp__should_skip(fmt))
908 continue;
909
910 cmp = fmt->cmp(fmt, left, right);
911 if (cmp)
912 break;
913 }
914
915 return cmp;
916 }
917
918 int64_t
919 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
920 {
921 struct perf_hpp_fmt *fmt;
922 int64_t cmp = 0;
923
924 perf_hpp__for_each_sort_list(fmt) {
925 if (perf_hpp__should_skip(fmt))
926 continue;
927
928 cmp = fmt->collapse(fmt, left, right);
929 if (cmp)
930 break;
931 }
932
933 return cmp;
934 }
935
936 void hist_entry__delete(struct hist_entry *he)
937 {
938 thread__zput(he->thread);
939 map__zput(he->ms.map);
940
941 if (he->branch_info) {
942 map__zput(he->branch_info->from.map);
943 map__zput(he->branch_info->to.map);
944 zfree(&he->branch_info);
945 }
946
947 if (he->mem_info) {
948 map__zput(he->mem_info->iaddr.map);
949 map__zput(he->mem_info->daddr.map);
950 zfree(&he->mem_info);
951 }
952
953 zfree(&he->stat_acc);
954 free_srcline(he->srcline);
955 if (he->srcfile && he->srcfile[0])
956 free(he->srcfile);
957 free_callchain(he->callchain);
958 free(he);
959 }
960
961 /*
962 * collapse the histogram
963 */
964
965 static bool hists__collapse_insert_entry(struct hists *hists __maybe_unused,
966 struct rb_root *root,
967 struct hist_entry *he)
968 {
969 struct rb_node **p = &root->rb_node;
970 struct rb_node *parent = NULL;
971 struct hist_entry *iter;
972 int64_t cmp;
973
974 while (*p != NULL) {
975 parent = *p;
976 iter = rb_entry(parent, struct hist_entry, rb_node_in);
977
978 cmp = hist_entry__collapse(iter, he);
979
980 if (!cmp) {
981 he_stat__add_stat(&iter->stat, &he->stat);
982 if (symbol_conf.cumulate_callchain)
983 he_stat__add_stat(iter->stat_acc, he->stat_acc);
984
985 if (symbol_conf.use_callchain) {
986 callchain_cursor_reset(&callchain_cursor);
987 callchain_merge(&callchain_cursor,
988 iter->callchain,
989 he->callchain);
990 }
991 hist_entry__delete(he);
992 return false;
993 }
994
995 if (cmp < 0)
996 p = &(*p)->rb_left;
997 else
998 p = &(*p)->rb_right;
999 }
1000 hists->nr_entries++;
1001
1002 rb_link_node(&he->rb_node_in, parent, p);
1003 rb_insert_color(&he->rb_node_in, root);
1004 return true;
1005 }
1006
1007 static struct rb_root *hists__get_rotate_entries_in(struct hists *hists)
1008 {
1009 struct rb_root *root;
1010
1011 pthread_mutex_lock(&hists->lock);
1012
1013 root = hists->entries_in;
1014 if (++hists->entries_in > &hists->entries_in_array[1])
1015 hists->entries_in = &hists->entries_in_array[0];
1016
1017 pthread_mutex_unlock(&hists->lock);
1018
1019 return root;
1020 }
1021
1022 static void hists__apply_filters(struct hists *hists, struct hist_entry *he)
1023 {
1024 hists__filter_entry_by_dso(hists, he);
1025 hists__filter_entry_by_thread(hists, he);
1026 hists__filter_entry_by_symbol(hists, he);
1027 }
1028
1029 void hists__collapse_resort(struct hists *hists, struct ui_progress *prog)
1030 {
1031 struct rb_root *root;
1032 struct rb_node *next;
1033 struct hist_entry *n;
1034
1035 if (!sort__need_collapse)
1036 return;
1037
1038 hists->nr_entries = 0;
1039
1040 root = hists__get_rotate_entries_in(hists);
1041
1042 next = rb_first(root);
1043
1044 while (next) {
1045 if (session_done())
1046 break;
1047 n = rb_entry(next, struct hist_entry, rb_node_in);
1048 next = rb_next(&n->rb_node_in);
1049
1050 rb_erase(&n->rb_node_in, root);
1051 if (hists__collapse_insert_entry(hists, &hists->entries_collapsed, n)) {
1052 /*
1053 * If it wasn't combined with one of the entries already
1054 * collapsed, we need to apply the filters that may have
1055 * been set by, say, the hist_browser.
1056 */
1057 hists__apply_filters(hists, n);
1058 }
1059 if (prog)
1060 ui_progress__update(prog, 1);
1061 }
1062 }
1063
1064 static int hist_entry__sort(struct hist_entry *a, struct hist_entry *b)
1065 {
1066 struct perf_hpp_fmt *fmt;
1067 int64_t cmp = 0;
1068
1069 perf_hpp__for_each_sort_list(fmt) {
1070 if (perf_hpp__should_skip(fmt))
1071 continue;
1072
1073 cmp = fmt->sort(fmt, a, b);
1074 if (cmp)
1075 break;
1076 }
1077
1078 return cmp;
1079 }
1080
1081 static void hists__reset_filter_stats(struct hists *hists)
1082 {
1083 hists->nr_non_filtered_entries = 0;
1084 hists->stats.total_non_filtered_period = 0;
1085 }
1086
1087 void hists__reset_stats(struct hists *hists)
1088 {
1089 hists->nr_entries = 0;
1090 hists->stats.total_period = 0;
1091
1092 hists__reset_filter_stats(hists);
1093 }
1094
1095 static void hists__inc_filter_stats(struct hists *hists, struct hist_entry *h)
1096 {
1097 hists->nr_non_filtered_entries++;
1098 hists->stats.total_non_filtered_period += h->stat.period;
1099 }
1100
1101 void hists__inc_stats(struct hists *hists, struct hist_entry *h)
1102 {
1103 if (!h->filtered)
1104 hists__inc_filter_stats(hists, h);
1105
1106 hists->nr_entries++;
1107 hists->stats.total_period += h->stat.period;
1108 }
1109
1110 static void __hists__insert_output_entry(struct rb_root *entries,
1111 struct hist_entry *he,
1112 u64 min_callchain_hits,
1113 bool use_callchain)
1114 {
1115 struct rb_node **p = &entries->rb_node;
1116 struct rb_node *parent = NULL;
1117 struct hist_entry *iter;
1118
1119 if (use_callchain)
1120 callchain_param.sort(&he->sorted_chain, he->callchain,
1121 min_callchain_hits, &callchain_param);
1122
1123 while (*p != NULL) {
1124 parent = *p;
1125 iter = rb_entry(parent, struct hist_entry, rb_node);
1126
1127 if (hist_entry__sort(he, iter) > 0)
1128 p = &(*p)->rb_left;
1129 else
1130 p = &(*p)->rb_right;
1131 }
1132
1133 rb_link_node(&he->rb_node, parent, p);
1134 rb_insert_color(&he->rb_node, entries);
1135 }
1136
1137 void hists__output_resort(struct hists *hists, struct ui_progress *prog)
1138 {
1139 struct rb_root *root;
1140 struct rb_node *next;
1141 struct hist_entry *n;
1142 u64 min_callchain_hits;
1143 struct perf_evsel *evsel = hists_to_evsel(hists);
1144 bool use_callchain;
1145
1146 if (evsel && !symbol_conf.show_ref_callgraph)
1147 use_callchain = evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN;
1148 else
1149 use_callchain = symbol_conf.use_callchain;
1150
1151 min_callchain_hits = hists->stats.total_period * (callchain_param.min_percent / 100);
1152
1153 if (sort__need_collapse)
1154 root = &hists->entries_collapsed;
1155 else
1156 root = hists->entries_in;
1157
1158 next = rb_first(root);
1159 hists->entries = RB_ROOT;
1160
1161 hists__reset_stats(hists);
1162 hists__reset_col_len(hists);
1163
1164 while (next) {
1165 n = rb_entry(next, struct hist_entry, rb_node_in);
1166 next = rb_next(&n->rb_node_in);
1167
1168 __hists__insert_output_entry(&hists->entries, n, min_callchain_hits, use_callchain);
1169 hists__inc_stats(hists, n);
1170
1171 if (!n->filtered)
1172 hists__calc_col_len(hists, n);
1173
1174 if (prog)
1175 ui_progress__update(prog, 1);
1176 }
1177 }
1178
1179 static void hists__remove_entry_filter(struct hists *hists, struct hist_entry *h,
1180 enum hist_filter filter)
1181 {
1182 h->filtered &= ~(1 << filter);
1183 if (h->filtered)
1184 return;
1185
1186 /* force fold unfiltered entry for simplicity */
1187 h->unfolded = false;
1188 h->row_offset = 0;
1189 h->nr_rows = 0;
1190
1191 hists->stats.nr_non_filtered_samples += h->stat.nr_events;
1192
1193 hists__inc_filter_stats(hists, h);
1194 hists__calc_col_len(hists, h);
1195 }
1196
1197
1198 static bool hists__filter_entry_by_dso(struct hists *hists,
1199 struct hist_entry *he)
1200 {
1201 if (hists->dso_filter != NULL &&
1202 (he->ms.map == NULL || he->ms.map->dso != hists->dso_filter)) {
1203 he->filtered |= (1 << HIST_FILTER__DSO);
1204 return true;
1205 }
1206
1207 return false;
1208 }
1209
1210 void hists__filter_by_dso(struct hists *hists)
1211 {
1212 struct rb_node *nd;
1213
1214 hists->stats.nr_non_filtered_samples = 0;
1215
1216 hists__reset_filter_stats(hists);
1217 hists__reset_col_len(hists);
1218
1219 for (nd = rb_first(&hists->entries); nd; nd = rb_next(nd)) {
1220 struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
1221
1222 if (symbol_conf.exclude_other && !h->parent)
1223 continue;
1224
1225 if (hists__filter_entry_by_dso(hists, h))
1226 continue;
1227
1228 hists__remove_entry_filter(hists, h, HIST_FILTER__DSO);
1229 }
1230 }
1231
1232 static bool hists__filter_entry_by_thread(struct hists *hists,
1233 struct hist_entry *he)
1234 {
1235 if (hists->thread_filter != NULL &&
1236 he->thread != hists->thread_filter) {
1237 he->filtered |= (1 << HIST_FILTER__THREAD);
1238 return true;
1239 }
1240
1241 return false;
1242 }
1243
1244 void hists__filter_by_thread(struct hists *hists)
1245 {
1246 struct rb_node *nd;
1247
1248 hists->stats.nr_non_filtered_samples = 0;
1249
1250 hists__reset_filter_stats(hists);
1251 hists__reset_col_len(hists);
1252
1253 for (nd = rb_first(&hists->entries); nd; nd = rb_next(nd)) {
1254 struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
1255
1256 if (hists__filter_entry_by_thread(hists, h))
1257 continue;
1258
1259 hists__remove_entry_filter(hists, h, HIST_FILTER__THREAD);
1260 }
1261 }
1262
1263 static bool hists__filter_entry_by_symbol(struct hists *hists,
1264 struct hist_entry *he)
1265 {
1266 if (hists->symbol_filter_str != NULL &&
1267 (!he->ms.sym || strstr(he->ms.sym->name,
1268 hists->symbol_filter_str) == NULL)) {
1269 he->filtered |= (1 << HIST_FILTER__SYMBOL);
1270 return true;
1271 }
1272
1273 return false;
1274 }
1275
1276 void hists__filter_by_symbol(struct hists *hists)
1277 {
1278 struct rb_node *nd;
1279
1280 hists->stats.nr_non_filtered_samples = 0;
1281
1282 hists__reset_filter_stats(hists);
1283 hists__reset_col_len(hists);
1284
1285 for (nd = rb_first(&hists->entries); nd; nd = rb_next(nd)) {
1286 struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
1287
1288 if (hists__filter_entry_by_symbol(hists, h))
1289 continue;
1290
1291 hists__remove_entry_filter(hists, h, HIST_FILTER__SYMBOL);
1292 }
1293 }
1294
1295 void events_stats__inc(struct events_stats *stats, u32 type)
1296 {
1297 ++stats->nr_events[0];
1298 ++stats->nr_events[type];
1299 }
1300
1301 void hists__inc_nr_events(struct hists *hists, u32 type)
1302 {
1303 events_stats__inc(&hists->stats, type);
1304 }
1305
1306 void hists__inc_nr_samples(struct hists *hists, bool filtered)
1307 {
1308 events_stats__inc(&hists->stats, PERF_RECORD_SAMPLE);
1309 if (!filtered)
1310 hists->stats.nr_non_filtered_samples++;
1311 }
1312
1313 static struct hist_entry *hists__add_dummy_entry(struct hists *hists,
1314 struct hist_entry *pair)
1315 {
1316 struct rb_root *root;
1317 struct rb_node **p;
1318 struct rb_node *parent = NULL;
1319 struct hist_entry *he;
1320 int64_t cmp;
1321
1322 if (sort__need_collapse)
1323 root = &hists->entries_collapsed;
1324 else
1325 root = hists->entries_in;
1326
1327 p = &root->rb_node;
1328
1329 while (*p != NULL) {
1330 parent = *p;
1331 he = rb_entry(parent, struct hist_entry, rb_node_in);
1332
1333 cmp = hist_entry__collapse(he, pair);
1334
1335 if (!cmp)
1336 goto out;
1337
1338 if (cmp < 0)
1339 p = &(*p)->rb_left;
1340 else
1341 p = &(*p)->rb_right;
1342 }
1343
1344 he = hist_entry__new(pair, true);
1345 if (he) {
1346 memset(&he->stat, 0, sizeof(he->stat));
1347 he->hists = hists;
1348 rb_link_node(&he->rb_node_in, parent, p);
1349 rb_insert_color(&he->rb_node_in, root);
1350 hists__inc_stats(hists, he);
1351 he->dummy = true;
1352 }
1353 out:
1354 return he;
1355 }
1356
1357 static struct hist_entry *hists__find_entry(struct hists *hists,
1358 struct hist_entry *he)
1359 {
1360 struct rb_node *n;
1361
1362 if (sort__need_collapse)
1363 n = hists->entries_collapsed.rb_node;
1364 else
1365 n = hists->entries_in->rb_node;
1366
1367 while (n) {
1368 struct hist_entry *iter = rb_entry(n, struct hist_entry, rb_node_in);
1369 int64_t cmp = hist_entry__collapse(iter, he);
1370
1371 if (cmp < 0)
1372 n = n->rb_left;
1373 else if (cmp > 0)
1374 n = n->rb_right;
1375 else
1376 return iter;
1377 }
1378
1379 return NULL;
1380 }
1381
1382 /*
1383 * Look for pairs to link to the leader buckets (hist_entries):
1384 */
1385 void hists__match(struct hists *leader, struct hists *other)
1386 {
1387 struct rb_root *root;
1388 struct rb_node *nd;
1389 struct hist_entry *pos, *pair;
1390
1391 if (sort__need_collapse)
1392 root = &leader->entries_collapsed;
1393 else
1394 root = leader->entries_in;
1395
1396 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
1397 pos = rb_entry(nd, struct hist_entry, rb_node_in);
1398 pair = hists__find_entry(other, pos);
1399
1400 if (pair)
1401 hist_entry__add_pair(pair, pos);
1402 }
1403 }
1404
1405 /*
1406 * Look for entries in the other hists that are not present in the leader, if
1407 * we find them, just add a dummy entry on the leader hists, with period=0,
1408 * nr_events=0, to serve as the list header.
1409 */
1410 int hists__link(struct hists *leader, struct hists *other)
1411 {
1412 struct rb_root *root;
1413 struct rb_node *nd;
1414 struct hist_entry *pos, *pair;
1415
1416 if (sort__need_collapse)
1417 root = &other->entries_collapsed;
1418 else
1419 root = other->entries_in;
1420
1421 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
1422 pos = rb_entry(nd, struct hist_entry, rb_node_in);
1423
1424 if (!hist_entry__has_pairs(pos)) {
1425 pair = hists__add_dummy_entry(leader, pos);
1426 if (pair == NULL)
1427 return -1;
1428 hist_entry__add_pair(pos, pair);
1429 }
1430 }
1431
1432 return 0;
1433 }
1434
1435 void hist__account_cycles(struct branch_stack *bs, struct addr_location *al,
1436 struct perf_sample *sample, bool nonany_branch_mode)
1437 {
1438 struct branch_info *bi;
1439
1440 /* If we have branch cycles always annotate them. */
1441 if (bs && bs->nr && bs->entries[0].flags.cycles) {
1442 int i;
1443
1444 bi = sample__resolve_bstack(sample, al);
1445 if (bi) {
1446 struct addr_map_symbol *prev = NULL;
1447
1448 /*
1449 * Ignore errors, still want to process the
1450 * other entries.
1451 *
1452 * For non standard branch modes always
1453 * force no IPC (prev == NULL)
1454 *
1455 * Note that perf stores branches reversed from
1456 * program order!
1457 */
1458 for (i = bs->nr - 1; i >= 0; i--) {
1459 addr_map_symbol__account_cycles(&bi[i].from,
1460 nonany_branch_mode ? NULL : prev,
1461 bi[i].flags.cycles);
1462 prev = &bi[i].to;
1463 }
1464 free(bi);
1465 }
1466 }
1467 }
1468
1469 size_t perf_evlist__fprintf_nr_events(struct perf_evlist *evlist, FILE *fp)
1470 {
1471 struct perf_evsel *pos;
1472 size_t ret = 0;
1473
1474 evlist__for_each(evlist, pos) {
1475 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1476 ret += events_stats__fprintf(&evsel__hists(pos)->stats, fp);
1477 }
1478
1479 return ret;
1480 }
1481
1482
1483 u64 hists__total_period(struct hists *hists)
1484 {
1485 return symbol_conf.filter_relative ? hists->stats.total_non_filtered_period :
1486 hists->stats.total_period;
1487 }
1488
1489 int parse_filter_percentage(const struct option *opt __maybe_unused,
1490 const char *arg, int unset __maybe_unused)
1491 {
1492 if (!strcmp(arg, "relative"))
1493 symbol_conf.filter_relative = true;
1494 else if (!strcmp(arg, "absolute"))
1495 symbol_conf.filter_relative = false;
1496 else
1497 return -1;
1498
1499 return 0;
1500 }
1501
1502 int perf_hist_config(const char *var, const char *value)
1503 {
1504 if (!strcmp(var, "hist.percentage"))
1505 return parse_filter_percentage(NULL, value, 0);
1506
1507 return 0;
1508 }
1509
1510 static int hists_evsel__init(struct perf_evsel *evsel)
1511 {
1512 struct hists *hists = evsel__hists(evsel);
1513
1514 memset(hists, 0, sizeof(*hists));
1515 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
1516 hists->entries_in = &hists->entries_in_array[0];
1517 hists->entries_collapsed = RB_ROOT;
1518 hists->entries = RB_ROOT;
1519 pthread_mutex_init(&hists->lock, NULL);
1520 return 0;
1521 }
1522
1523 /*
1524 * XXX We probably need a hists_evsel__exit() to free the hist_entries
1525 * stored in the rbtree...
1526 */
1527
1528 int hists__init(void)
1529 {
1530 int err = perf_evsel__object_config(sizeof(struct hists_evsel),
1531 hists_evsel__init, NULL);
1532 if (err)
1533 fputs("FATAL ERROR: Couldn't setup hists class\n", stderr);
1534
1535 return err;
1536 }