Loading...
1/*
2 * builtin-report.c
3 *
4 * Builtin report command: Analyze the perf.data input file,
5 * look up and read DSOs and symbol information and display
6 * a histogram of results, along various sorting keys.
7 */
8#include "builtin.h"
9
10#include "util/util.h"
11#include "util/cache.h"
12
13#include "util/annotate.h"
14#include "util/color.h"
15#include <linux/list.h>
16#include <linux/rbtree.h>
17#include "util/symbol.h"
18#include "util/callchain.h"
19#include "util/strlist.h"
20#include "util/values.h"
21
22#include "perf.h"
23#include "util/debug.h"
24#include "util/evlist.h"
25#include "util/evsel.h"
26#include "util/header.h"
27#include "util/session.h"
28#include "util/tool.h"
29
30#include "util/parse-options.h"
31#include "util/parse-events.h"
32
33#include "util/thread.h"
34#include "util/sort.h"
35#include "util/hist.h"
36#include "util/data.h"
37#include "arch/common.h"
38
39#include <dlfcn.h>
40#include <linux/bitmap.h>
41
42struct report {
43 struct perf_tool tool;
44 struct perf_session *session;
45 bool force, use_tui, use_gtk, use_stdio;
46 bool hide_unresolved;
47 bool dont_use_callchains;
48 bool show_full_info;
49 bool show_threads;
50 bool inverted_callchain;
51 bool mem_mode;
52 bool header;
53 bool header_only;
54 int max_stack;
55 struct perf_read_values show_threads_values;
56 const char *pretty_printing_style;
57 const char *cpu_list;
58 const char *symbol_filter_str;
59 float min_percent;
60 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
61};
62
63static int report__config(const char *var, const char *value, void *cb)
64{
65 if (!strcmp(var, "report.group")) {
66 symbol_conf.event_group = perf_config_bool(var, value);
67 return 0;
68 }
69 if (!strcmp(var, "report.percent-limit")) {
70 struct report *rep = cb;
71 rep->min_percent = strtof(value, NULL);
72 return 0;
73 }
74
75 return perf_default_config(var, value, cb);
76}
77
78static int report__add_mem_hist_entry(struct report *rep, struct addr_location *al,
79 struct perf_sample *sample, struct perf_evsel *evsel)
80{
81 struct symbol *parent = NULL;
82 struct hist_entry *he;
83 struct mem_info *mi, *mx;
84 uint64_t cost;
85 int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);
86
87 if (err)
88 return err;
89
90 mi = sample__resolve_mem(sample, al);
91 if (!mi)
92 return -ENOMEM;
93
94 if (rep->hide_unresolved && !al->sym)
95 return 0;
96
97 cost = sample->weight;
98 if (!cost)
99 cost = 1;
100
101 /*
102 * must pass period=weight in order to get the correct
103 * sorting from hists__collapse_resort() which is solely
104 * based on periods. We want sorting be done on nr_events * weight
105 * and this is indirectly achieved by passing period=weight here
106 * and the he_stat__add_period() function.
107 */
108 he = __hists__add_entry(&evsel->hists, al, parent, NULL, mi,
109 cost, cost, 0);
110 if (!he)
111 return -ENOMEM;
112
113 if (ui__has_annotation()) {
114 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
115 if (err)
116 goto out;
117
118 mx = he->mem_info;
119 err = addr_map_symbol__inc_samples(&mx->daddr, evsel->idx);
120 if (err)
121 goto out;
122 }
123
124 evsel->hists.stats.total_period += cost;
125 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
126 err = hist_entry__append_callchain(he, sample);
127out:
128 return err;
129}
130
131static int report__add_branch_hist_entry(struct report *rep, struct addr_location *al,
132 struct perf_sample *sample, struct perf_evsel *evsel)
133{
134 struct symbol *parent = NULL;
135 unsigned i;
136 struct hist_entry *he;
137 struct branch_info *bi, *bx;
138 int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);
139
140 if (err)
141 return err;
142
143 bi = sample__resolve_bstack(sample, al);
144 if (!bi)
145 return -ENOMEM;
146
147 for (i = 0; i < sample->branch_stack->nr; i++) {
148 if (rep->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
149 continue;
150
151 err = -ENOMEM;
152
153 /* overwrite the 'al' to branch-to info */
154 al->map = bi[i].to.map;
155 al->sym = bi[i].to.sym;
156 al->addr = bi[i].to.addr;
157 /*
158 * The report shows the percentage of total branches captured
159 * and not events sampled. Thus we use a pseudo period of 1.
160 */
161 he = __hists__add_entry(&evsel->hists, al, parent, &bi[i], NULL,
162 1, 1, 0);
163 if (he) {
164 if (ui__has_annotation()) {
165 bx = he->branch_info;
166 err = addr_map_symbol__inc_samples(&bx->from,
167 evsel->idx);
168 if (err)
169 goto out;
170
171 err = addr_map_symbol__inc_samples(&bx->to,
172 evsel->idx);
173 if (err)
174 goto out;
175 }
176
177 evsel->hists.stats.total_period += 1;
178 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
179 } else
180 goto out;
181 }
182 err = 0;
183out:
184 free(bi);
185 return err;
186}
187
188static int report__add_hist_entry(struct report *rep, struct perf_evsel *evsel,
189 struct addr_location *al, struct perf_sample *sample)
190{
191 struct symbol *parent = NULL;
192 struct hist_entry *he;
193 int err = sample__resolve_callchain(sample, &parent, evsel, al, rep->max_stack);
194
195 if (err)
196 return err;
197
198 he = __hists__add_entry(&evsel->hists, al, parent, NULL, NULL,
199 sample->period, sample->weight,
200 sample->transaction);
201 if (he == NULL)
202 return -ENOMEM;
203
204 err = hist_entry__append_callchain(he, sample);
205 if (err)
206 goto out;
207
208 if (ui__has_annotation())
209 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
210
211 evsel->hists.stats.total_period += sample->period;
212 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
213out:
214 return err;
215}
216
217
218static int process_sample_event(struct perf_tool *tool,
219 union perf_event *event,
220 struct perf_sample *sample,
221 struct perf_evsel *evsel,
222 struct machine *machine)
223{
224 struct report *rep = container_of(tool, struct report, tool);
225 struct addr_location al;
226 int ret;
227
228 if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
229 pr_debug("problem processing %d event, skipping it.\n",
230 event->header.type);
231 return -1;
232 }
233
234 if (rep->hide_unresolved && al.sym == NULL)
235 return 0;
236
237 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
238 return 0;
239
240 if (sort__mode == SORT_MODE__BRANCH) {
241 ret = report__add_branch_hist_entry(rep, &al, sample, evsel);
242 if (ret < 0)
243 pr_debug("problem adding lbr entry, skipping event\n");
244 } else if (rep->mem_mode == 1) {
245 ret = report__add_mem_hist_entry(rep, &al, sample, evsel);
246 if (ret < 0)
247 pr_debug("problem adding mem entry, skipping event\n");
248 } else {
249 if (al.map != NULL)
250 al.map->dso->hit = 1;
251
252 ret = report__add_hist_entry(rep, evsel, &al, sample);
253 if (ret < 0)
254 pr_debug("problem incrementing symbol period, skipping event\n");
255 }
256 return ret;
257}
258
259static int process_read_event(struct perf_tool *tool,
260 union perf_event *event,
261 struct perf_sample *sample __maybe_unused,
262 struct perf_evsel *evsel,
263 struct machine *machine __maybe_unused)
264{
265 struct report *rep = container_of(tool, struct report, tool);
266
267 if (rep->show_threads) {
268 const char *name = evsel ? perf_evsel__name(evsel) : "unknown";
269 perf_read_values_add_value(&rep->show_threads_values,
270 event->read.pid, event->read.tid,
271 event->read.id,
272 name,
273 event->read.value);
274 }
275
276 dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
277 evsel ? perf_evsel__name(evsel) : "FAIL",
278 event->read.value);
279
280 return 0;
281}
282
283/* For pipe mode, sample_type is not currently set */
284static int report__setup_sample_type(struct report *rep)
285{
286 struct perf_session *session = rep->session;
287 u64 sample_type = perf_evlist__combined_sample_type(session->evlist);
288 bool is_pipe = perf_data_file__is_pipe(session->file);
289
290 if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
291 if (sort__has_parent) {
292 ui__error("Selected --sort parent, but no "
293 "callchain data. Did you call "
294 "'perf record' without -g?\n");
295 return -EINVAL;
296 }
297 if (symbol_conf.use_callchain) {
298 ui__error("Selected -g but no callchain data. Did "
299 "you call 'perf record' without -g?\n");
300 return -1;
301 }
302 } else if (!rep->dont_use_callchains &&
303 callchain_param.mode != CHAIN_NONE &&
304 !symbol_conf.use_callchain) {
305 symbol_conf.use_callchain = true;
306 if (callchain_register_param(&callchain_param) < 0) {
307 ui__error("Can't register callchain params.\n");
308 return -EINVAL;
309 }
310 }
311
312 if (sort__mode == SORT_MODE__BRANCH) {
313 if (!is_pipe &&
314 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
315 ui__error("Selected -b but no branch data. "
316 "Did you call perf record without -b?\n");
317 return -1;
318 }
319 }
320
321 return 0;
322}
323
324static void sig_handler(int sig __maybe_unused)
325{
326 session_done = 1;
327}
328
329static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
330 const char *evname, FILE *fp)
331{
332 size_t ret;
333 char unit;
334 unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
335 u64 nr_events = hists->stats.total_period;
336 struct perf_evsel *evsel = hists_to_evsel(hists);
337 char buf[512];
338 size_t size = sizeof(buf);
339
340 if (perf_evsel__is_group_event(evsel)) {
341 struct perf_evsel *pos;
342
343 perf_evsel__group_desc(evsel, buf, size);
344 evname = buf;
345
346 for_each_group_member(pos, evsel) {
347 nr_samples += pos->hists.stats.nr_events[PERF_RECORD_SAMPLE];
348 nr_events += pos->hists.stats.total_period;
349 }
350 }
351
352 nr_samples = convert_unit(nr_samples, &unit);
353 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
354 if (evname != NULL)
355 ret += fprintf(fp, " of event '%s'", evname);
356
357 if (rep->mem_mode) {
358 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
359 ret += fprintf(fp, "\n# Sort order : %s", sort_order);
360 } else
361 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
362 return ret + fprintf(fp, "\n#\n");
363}
364
365static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
366 struct report *rep,
367 const char *help)
368{
369 struct perf_evsel *pos;
370
371 evlist__for_each(evlist, pos) {
372 struct hists *hists = &pos->hists;
373 const char *evname = perf_evsel__name(pos);
374
375 if (symbol_conf.event_group &&
376 !perf_evsel__is_group_leader(pos))
377 continue;
378
379 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
380 hists__fprintf(hists, true, 0, 0, rep->min_percent, stdout);
381 fprintf(stdout, "\n\n");
382 }
383
384 if (sort_order == default_sort_order &&
385 parent_pattern == default_parent_pattern) {
386 fprintf(stdout, "#\n# (%s)\n#\n", help);
387
388 if (rep->show_threads) {
389 bool style = !strcmp(rep->pretty_printing_style, "raw");
390 perf_read_values_display(stdout, &rep->show_threads_values,
391 style);
392 perf_read_values_destroy(&rep->show_threads_values);
393 }
394 }
395
396 return 0;
397}
398
399static void report__warn_kptr_restrict(const struct report *rep)
400{
401 struct map *kernel_map = rep->session->machines.host.vmlinux_maps[MAP__FUNCTION];
402 struct kmap *kernel_kmap = map__kmap(kernel_map);
403
404 if (kernel_map == NULL ||
405 (kernel_map->dso->hit &&
406 (kernel_kmap->ref_reloc_sym == NULL ||
407 kernel_kmap->ref_reloc_sym->addr == 0))) {
408 const char *desc =
409 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
410 "can't be resolved.";
411
412 if (kernel_map) {
413 const struct dso *kdso = kernel_map->dso;
414 if (!RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION])) {
415 desc = "If some relocation was applied (e.g. "
416 "kexec) symbols may be misresolved.";
417 }
418 }
419
420 ui__warning(
421"Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
422"Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
423"Samples in kernel modules can't be resolved as well.\n\n",
424 desc);
425 }
426}
427
428static int report__gtk_browse_hists(struct report *rep, const char *help)
429{
430 int (*hist_browser)(struct perf_evlist *evlist, const char *help,
431 struct hist_browser_timer *timer, float min_pcnt);
432
433 hist_browser = dlsym(perf_gtk_handle, "perf_evlist__gtk_browse_hists");
434
435 if (hist_browser == NULL) {
436 ui__error("GTK browser not found!\n");
437 return -1;
438 }
439
440 return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
441}
442
443static int report__browse_hists(struct report *rep)
444{
445 int ret;
446 struct perf_session *session = rep->session;
447 struct perf_evlist *evlist = session->evlist;
448 const char *help = "For a higher level overview, try: perf report --sort comm,dso";
449
450 switch (use_browser) {
451 case 1:
452 ret = perf_evlist__tui_browse_hists(evlist, help, NULL,
453 rep->min_percent,
454 &session->header.env);
455 /*
456 * Usually "ret" is the last pressed key, and we only
457 * care if the key notifies us to switch data file.
458 */
459 if (ret != K_SWITCH_INPUT_DATA)
460 ret = 0;
461 break;
462 case 2:
463 ret = report__gtk_browse_hists(rep, help);
464 break;
465 default:
466 ret = perf_evlist__tty_browse_hists(evlist, rep, help);
467 break;
468 }
469
470 return ret;
471}
472
473static u64 report__collapse_hists(struct report *rep)
474{
475 struct ui_progress prog;
476 struct perf_evsel *pos;
477 u64 nr_samples = 0;
478 /*
479 * Count number of histogram entries to use when showing progress,
480 * reusing nr_samples variable.
481 */
482 evlist__for_each(rep->session->evlist, pos)
483 nr_samples += pos->hists.nr_entries;
484
485 ui_progress__init(&prog, nr_samples, "Merging related events...");
486 /*
487 * Count total number of samples, will be used to check if this
488 * session had any.
489 */
490 nr_samples = 0;
491
492 evlist__for_each(rep->session->evlist, pos) {
493 struct hists *hists = &pos->hists;
494
495 if (pos->idx == 0)
496 hists->symbol_filter_str = rep->symbol_filter_str;
497
498 hists__collapse_resort(hists, &prog);
499 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
500
501 /* Non-group events are considered as leader */
502 if (symbol_conf.event_group &&
503 !perf_evsel__is_group_leader(pos)) {
504 struct hists *leader_hists = &pos->leader->hists;
505
506 hists__match(leader_hists, hists);
507 hists__link(leader_hists, hists);
508 }
509 }
510
511 ui_progress__finish();
512
513 return nr_samples;
514}
515
516static int __cmd_report(struct report *rep)
517{
518 int ret;
519 u64 nr_samples;
520 struct perf_session *session = rep->session;
521 struct perf_evsel *pos;
522 struct perf_data_file *file = session->file;
523
524 signal(SIGINT, sig_handler);
525
526 if (rep->cpu_list) {
527 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
528 rep->cpu_bitmap);
529 if (ret)
530 return ret;
531 }
532
533 if (rep->show_threads)
534 perf_read_values_init(&rep->show_threads_values);
535
536 ret = report__setup_sample_type(rep);
537 if (ret)
538 return ret;
539
540 ret = perf_session__process_events(session, &rep->tool);
541 if (ret)
542 return ret;
543
544 report__warn_kptr_restrict(rep);
545
546 if (use_browser == 0) {
547 if (verbose > 3)
548 perf_session__fprintf(session, stdout);
549
550 if (verbose > 2)
551 perf_session__fprintf_dsos(session, stdout);
552
553 if (dump_trace) {
554 perf_session__fprintf_nr_events(session, stdout);
555 return 0;
556 }
557 }
558
559 nr_samples = report__collapse_hists(rep);
560
561 if (session_done())
562 return 0;
563
564 if (nr_samples == 0) {
565 ui__error("The %s file has no samples!\n", file->path);
566 return 0;
567 }
568
569 evlist__for_each(session->evlist, pos)
570 hists__output_resort(&pos->hists);
571
572 return report__browse_hists(rep);
573}
574
575static int
576parse_callchain_opt(const struct option *opt, const char *arg, int unset)
577{
578 struct report *rep = (struct report *)opt->value;
579 char *tok, *tok2;
580 char *endptr;
581
582 /*
583 * --no-call-graph
584 */
585 if (unset) {
586 rep->dont_use_callchains = true;
587 return 0;
588 }
589
590 symbol_conf.use_callchain = true;
591
592 if (!arg)
593 return 0;
594
595 tok = strtok((char *)arg, ",");
596 if (!tok)
597 return -1;
598
599 /* get the output mode */
600 if (!strncmp(tok, "graph", strlen(arg)))
601 callchain_param.mode = CHAIN_GRAPH_ABS;
602
603 else if (!strncmp(tok, "flat", strlen(arg)))
604 callchain_param.mode = CHAIN_FLAT;
605
606 else if (!strncmp(tok, "fractal", strlen(arg)))
607 callchain_param.mode = CHAIN_GRAPH_REL;
608
609 else if (!strncmp(tok, "none", strlen(arg))) {
610 callchain_param.mode = CHAIN_NONE;
611 symbol_conf.use_callchain = false;
612
613 return 0;
614 }
615
616 else
617 return -1;
618
619 /* get the min percentage */
620 tok = strtok(NULL, ",");
621 if (!tok)
622 goto setup;
623
624 callchain_param.min_percent = strtod(tok, &endptr);
625 if (tok == endptr)
626 return -1;
627
628 /* get the print limit */
629 tok2 = strtok(NULL, ",");
630 if (!tok2)
631 goto setup;
632
633 if (tok2[0] != 'c') {
634 callchain_param.print_limit = strtoul(tok2, &endptr, 0);
635 tok2 = strtok(NULL, ",");
636 if (!tok2)
637 goto setup;
638 }
639
640 /* get the call chain order */
641 if (!strncmp(tok2, "caller", strlen("caller")))
642 callchain_param.order = ORDER_CALLER;
643 else if (!strncmp(tok2, "callee", strlen("callee")))
644 callchain_param.order = ORDER_CALLEE;
645 else
646 return -1;
647
648 /* Get the sort key */
649 tok2 = strtok(NULL, ",");
650 if (!tok2)
651 goto setup;
652 if (!strncmp(tok2, "function", strlen("function")))
653 callchain_param.key = CCKEY_FUNCTION;
654 else if (!strncmp(tok2, "address", strlen("address")))
655 callchain_param.key = CCKEY_ADDRESS;
656 else
657 return -1;
658setup:
659 if (callchain_register_param(&callchain_param) < 0) {
660 pr_err("Can't register callchain params\n");
661 return -1;
662 }
663 return 0;
664}
665
666int
667report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
668 const char *arg, int unset __maybe_unused)
669{
670 if (arg) {
671 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
672 if (err) {
673 char buf[BUFSIZ];
674 regerror(err, &ignore_callees_regex, buf, sizeof(buf));
675 pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
676 return -1;
677 }
678 have_ignore_callees = 1;
679 }
680
681 return 0;
682}
683
684static int
685parse_branch_mode(const struct option *opt __maybe_unused,
686 const char *str __maybe_unused, int unset)
687{
688 int *branch_mode = opt->value;
689
690 *branch_mode = !unset;
691 return 0;
692}
693
694static int
695parse_percent_limit(const struct option *opt, const char *str,
696 int unset __maybe_unused)
697{
698 struct report *rep = opt->value;
699
700 rep->min_percent = strtof(str, NULL);
701 return 0;
702}
703
704int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
705{
706 struct perf_session *session;
707 struct stat st;
708 bool has_br_stack = false;
709 int branch_mode = -1;
710 int ret = -1;
711 char callchain_default_opt[] = "fractal,0.5,callee";
712 const char * const report_usage[] = {
713 "perf report [<options>]",
714 NULL
715 };
716 struct report report = {
717 .tool = {
718 .sample = process_sample_event,
719 .mmap = perf_event__process_mmap,
720 .mmap2 = perf_event__process_mmap2,
721 .comm = perf_event__process_comm,
722 .exit = perf_event__process_exit,
723 .fork = perf_event__process_fork,
724 .lost = perf_event__process_lost,
725 .read = process_read_event,
726 .attr = perf_event__process_attr,
727 .tracing_data = perf_event__process_tracing_data,
728 .build_id = perf_event__process_build_id,
729 .ordered_samples = true,
730 .ordering_requires_timestamps = true,
731 },
732 .max_stack = PERF_MAX_STACK_DEPTH,
733 .pretty_printing_style = "normal",
734 };
735 const struct option options[] = {
736 OPT_STRING('i', "input", &input_name, "file",
737 "input file name"),
738 OPT_INCR('v', "verbose", &verbose,
739 "be more verbose (show symbol address, etc)"),
740 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
741 "dump raw trace in ASCII"),
742 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
743 "file", "vmlinux pathname"),
744 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
745 "file", "kallsyms pathname"),
746 OPT_BOOLEAN('f', "force", &report.force, "don't complain, do it"),
747 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
748 "load module symbols - WARNING: use only with -k and LIVE kernel"),
749 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
750 "Show a column with the number of samples"),
751 OPT_BOOLEAN('T', "threads", &report.show_threads,
752 "Show per-thread event counters"),
753 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
754 "pretty printing style key: normal raw"),
755 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
756 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
757 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
758 "Use the stdio interface"),
759 OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
760 OPT_BOOLEAN(0, "header-only", &report.header_only,
761 "Show only data header."),
762 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
763 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline,"
764 " dso_to, dso_from, symbol_to, symbol_from, mispredict,"
765 " weight, local_weight, mem, symbol_daddr, dso_daddr, tlb, "
766 "snoop, locked, abort, in_tx, transaction"),
767 OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
768 "Show sample percentage for different cpu modes"),
769 OPT_STRING('p', "parent", &parent_pattern, "regex",
770 "regex filter to identify parent, see: '--sort parent'"),
771 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
772 "Only display entries with parent-match"),
773 OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
774 "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). "
775 "Default: fractal,0.5,callee,function", &parse_callchain_opt, callchain_default_opt),
776 OPT_INTEGER(0, "max-stack", &report.max_stack,
777 "Set the maximum stack depth when parsing the callchain, "
778 "anything beyond the specified depth will be ignored. "
779 "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
780 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
781 "alias for inverted call graph"),
782 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
783 "ignore callees of these functions in call graphs",
784 report_parse_ignore_callees_opt),
785 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
786 "only consider symbols in these dsos"),
787 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
788 "only consider symbols in these comms"),
789 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
790 "only consider these symbols"),
791 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
792 "only show symbols that (partially) match with this filter"),
793 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
794 "width[,width...]",
795 "don't try to adjust column width, use these fixed values"),
796 OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
797 "separator for columns, no spaces will be added between "
798 "columns '.' is reserved."),
799 OPT_BOOLEAN('U', "hide-unresolved", &report.hide_unresolved,
800 "Only display entries resolved to a symbol"),
801 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
802 "Look for files with symbols relative to this directory"),
803 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
804 "list of cpus to profile"),
805 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
806 "Display extended information about perf.data file"),
807 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
808 "Interleave source code with assembly code (default)"),
809 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
810 "Display raw encoding of assembly instructions (default)"),
811 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
812 "Specify disassembler style (e.g. -M intel for intel syntax)"),
813 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
814 "Show a column with the sum of periods"),
815 OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
816 "Show event group information together"),
817 OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
818 "use branch records for histogram filling", parse_branch_mode),
819 OPT_STRING(0, "objdump", &objdump_path, "path",
820 "objdump binary to use for disassembly and annotations"),
821 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
822 "Disable symbol demangling"),
823 OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
824 OPT_CALLBACK(0, "percent-limit", &report, "percent",
825 "Don't show entries under that percent", parse_percent_limit),
826 OPT_END()
827 };
828 struct perf_data_file file = {
829 .mode = PERF_DATA_MODE_READ,
830 };
831
832 perf_config(report__config, &report);
833
834 argc = parse_options(argc, argv, options, report_usage, 0);
835
836 if (report.use_stdio)
837 use_browser = 0;
838 else if (report.use_tui)
839 use_browser = 1;
840 else if (report.use_gtk)
841 use_browser = 2;
842
843 if (report.inverted_callchain)
844 callchain_param.order = ORDER_CALLER;
845
846 if (!input_name || !strlen(input_name)) {
847 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
848 input_name = "-";
849 else
850 input_name = "perf.data";
851 }
852
853 file.path = input_name;
854 file.force = report.force;
855
856repeat:
857 session = perf_session__new(&file, false, &report.tool);
858 if (session == NULL)
859 return -ENOMEM;
860
861 report.session = session;
862
863 has_br_stack = perf_header__has_feat(&session->header,
864 HEADER_BRANCH_STACK);
865
866 if (branch_mode == -1 && has_br_stack)
867 sort__mode = SORT_MODE__BRANCH;
868
869 /* sort__mode could be NORMAL if --no-branch-stack */
870 if (sort__mode == SORT_MODE__BRANCH) {
871 /*
872 * if no sort_order is provided, then specify
873 * branch-mode specific order
874 */
875 if (sort_order == default_sort_order)
876 sort_order = "comm,dso_from,symbol_from,"
877 "dso_to,symbol_to";
878
879 }
880 if (report.mem_mode) {
881 if (sort__mode == SORT_MODE__BRANCH) {
882 pr_err("branch and mem mode incompatible\n");
883 goto error;
884 }
885 sort__mode = SORT_MODE__MEMORY;
886
887 /*
888 * if no sort_order is provided, then specify
889 * branch-mode specific order
890 */
891 if (sort_order == default_sort_order)
892 sort_order = "local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked";
893 }
894
895 if (setup_sorting() < 0) {
896 parse_options_usage(report_usage, options, "s", 1);
897 goto error;
898 }
899
900 if (parent_pattern != default_parent_pattern) {
901 if (sort_dimension__add("parent") < 0)
902 goto error;
903 }
904
905 /* Force tty output for header output. */
906 if (report.header || report.header_only)
907 use_browser = 0;
908
909 if (strcmp(input_name, "-") != 0)
910 setup_browser(true);
911 else {
912 use_browser = 0;
913 perf_hpp__init();
914 }
915
916 if (report.header || report.header_only) {
917 perf_session__fprintf_info(session, stdout,
918 report.show_full_info);
919 if (report.header_only)
920 return 0;
921 } else if (use_browser == 0) {
922 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
923 stdout);
924 }
925
926 /*
927 * Only in the TUI browser we are doing integrated annotation,
928 * so don't allocate extra space that won't be used in the stdio
929 * implementation.
930 */
931 if (ui__has_annotation()) {
932 symbol_conf.priv_size = sizeof(struct annotation);
933 machines__set_symbol_filter(&session->machines,
934 symbol__annotate_init);
935 /*
936 * For searching by name on the "Browse map details".
937 * providing it only in verbose mode not to bloat too
938 * much struct symbol.
939 */
940 if (verbose) {
941 /*
942 * XXX: Need to provide a less kludgy way to ask for
943 * more space per symbol, the u32 is for the index on
944 * the ui browser.
945 * See symbol__browser_index.
946 */
947 symbol_conf.priv_size += sizeof(u32);
948 symbol_conf.sort_by_name = true;
949 }
950 }
951
952 if (symbol__init() < 0)
953 goto error;
954
955 if (argc) {
956 /*
957 * Special case: if there's an argument left then assume that
958 * it's a symbol filter:
959 */
960 if (argc > 1)
961 usage_with_options(report_usage, options);
962
963 report.symbol_filter_str = argv[0];
964 }
965
966 sort__setup_elide(stdout);
967
968 ret = __cmd_report(&report);
969 if (ret == K_SWITCH_INPUT_DATA) {
970 perf_session__delete(session);
971 goto repeat;
972 } else
973 ret = 0;
974
975error:
976 perf_session__delete(session);
977 return ret;
978}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * builtin-report.c
4 *
5 * Builtin report command: Analyze the perf.data input file,
6 * look up and read DSOs and symbol information and display
7 * a histogram of results, along various sorting keys.
8 */
9#include "builtin.h"
10
11#include "util/config.h"
12
13#include "util/annotate.h"
14#include "util/color.h"
15#include "util/dso.h"
16#include <linux/list.h>
17#include <linux/rbtree.h>
18#include <linux/err.h>
19#include <linux/zalloc.h>
20#include "util/map.h"
21#include "util/symbol.h"
22#include "util/map_symbol.h"
23#include "util/mem-events.h"
24#include "util/branch.h"
25#include "util/callchain.h"
26#include "util/values.h"
27
28#include "perf.h"
29#include "util/debug.h"
30#include "util/evlist.h"
31#include "util/evsel.h"
32#include "util/evswitch.h"
33#include "util/header.h"
34#include "util/session.h"
35#include "util/srcline.h"
36#include "util/tool.h"
37
38#include <subcmd/parse-options.h>
39#include <subcmd/exec-cmd.h>
40#include "util/parse-events.h"
41
42#include "util/thread.h"
43#include "util/sort.h"
44#include "util/hist.h"
45#include "util/data.h"
46#include "arch/common.h"
47#include "util/time-utils.h"
48#include "util/auxtrace.h"
49#include "util/units.h"
50#include "util/util.h" // perf_tip()
51#include "ui/ui.h"
52#include "ui/progress.h"
53#include "util/block-info.h"
54
55#include <dlfcn.h>
56#include <errno.h>
57#include <inttypes.h>
58#include <regex.h>
59#include <linux/ctype.h>
60#include <signal.h>
61#include <linux/bitmap.h>
62#include <linux/string.h>
63#include <linux/stringify.h>
64#include <linux/time64.h>
65#include <sys/types.h>
66#include <sys/stat.h>
67#include <unistd.h>
68#include <linux/mman.h>
69
70#ifdef HAVE_LIBTRACEEVENT
71#include <traceevent/event-parse.h>
72#endif
73
74struct report {
75 struct perf_tool tool;
76 struct perf_session *session;
77 struct evswitch evswitch;
78#ifdef HAVE_SLANG_SUPPORT
79 bool use_tui;
80#endif
81#ifdef HAVE_GTK2_SUPPORT
82 bool use_gtk;
83#endif
84 bool use_stdio;
85 bool show_full_info;
86 bool show_threads;
87 bool inverted_callchain;
88 bool mem_mode;
89 bool stats_mode;
90 bool tasks_mode;
91 bool mmaps_mode;
92 bool header;
93 bool header_only;
94 bool nonany_branch_mode;
95 bool group_set;
96 bool stitch_lbr;
97 bool disable_order;
98 bool skip_empty;
99 bool data_type;
100 int max_stack;
101 struct perf_read_values show_threads_values;
102 const char *pretty_printing_style;
103 const char *cpu_list;
104 const char *symbol_filter_str;
105 const char *time_str;
106 struct perf_time_interval *ptime_range;
107 int range_size;
108 int range_num;
109 float min_percent;
110 u64 nr_entries;
111 u64 queue_size;
112 u64 total_cycles;
113 int socket_filter;
114 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
115 struct branch_type_stat brtype_stat;
116 bool symbol_ipc;
117 bool total_cycles_mode;
118 struct block_report *block_reports;
119 int nr_block_reports;
120};
121
122static int report__config(const char *var, const char *value, void *cb)
123{
124 struct report *rep = cb;
125
126 if (!strcmp(var, "report.group")) {
127 symbol_conf.event_group = perf_config_bool(var, value);
128 return 0;
129 }
130 if (!strcmp(var, "report.percent-limit")) {
131 double pcnt = strtof(value, NULL);
132
133 rep->min_percent = pcnt;
134 callchain_param.min_percent = pcnt;
135 return 0;
136 }
137 if (!strcmp(var, "report.children")) {
138 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
139 return 0;
140 }
141 if (!strcmp(var, "report.queue-size"))
142 return perf_config_u64(&rep->queue_size, var, value);
143
144 if (!strcmp(var, "report.sort_order")) {
145 default_sort_order = strdup(value);
146 if (!default_sort_order) {
147 pr_err("Not enough memory for report.sort_order\n");
148 return -1;
149 }
150 return 0;
151 }
152
153 if (!strcmp(var, "report.skip-empty")) {
154 rep->skip_empty = perf_config_bool(var, value);
155 return 0;
156 }
157
158 pr_debug("%s variable unknown, ignoring...", var);
159 return 0;
160}
161
162static int hist_iter__report_callback(struct hist_entry_iter *iter,
163 struct addr_location *al, bool single,
164 void *arg)
165{
166 int err = 0;
167 struct report *rep = arg;
168 struct hist_entry *he = iter->he;
169 struct evsel *evsel = iter->evsel;
170 struct perf_sample *sample = iter->sample;
171 struct mem_info *mi;
172 struct branch_info *bi;
173
174 if (!ui__has_annotation() && !rep->symbol_ipc && !rep->data_type)
175 return 0;
176
177 if (sort__mode == SORT_MODE__BRANCH) {
178 bi = he->branch_info;
179 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
180 if (err)
181 goto out;
182
183 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
184
185 } else if (rep->mem_mode) {
186 mi = he->mem_info;
187 err = addr_map_symbol__inc_samples(&mi->daddr, sample, evsel);
188 if (err)
189 goto out;
190
191 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
192
193 } else if (symbol_conf.cumulate_callchain) {
194 if (single)
195 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
196 } else {
197 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
198 }
199
200out:
201 return err;
202}
203
204static int hist_iter__branch_callback(struct hist_entry_iter *iter,
205 struct addr_location *al __maybe_unused,
206 bool single __maybe_unused,
207 void *arg)
208{
209 struct hist_entry *he = iter->he;
210 struct report *rep = arg;
211 struct branch_info *bi = he->branch_info;
212 struct perf_sample *sample = iter->sample;
213 struct evsel *evsel = iter->evsel;
214 int err;
215
216 branch_type_count(&rep->brtype_stat, &bi->flags,
217 bi->from.addr, bi->to.addr);
218
219 if (!ui__has_annotation() && !rep->symbol_ipc)
220 return 0;
221
222 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
223 if (err)
224 goto out;
225
226 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
227
228out:
229 return err;
230}
231
232static void setup_forced_leader(struct report *report,
233 struct evlist *evlist)
234{
235 if (report->group_set)
236 evlist__force_leader(evlist);
237}
238
239static int process_feature_event(struct perf_session *session,
240 union perf_event *event)
241{
242 struct report *rep = container_of(session->tool, struct report, tool);
243
244 if (event->feat.feat_id < HEADER_LAST_FEATURE)
245 return perf_event__process_feature(session, event);
246
247 if (event->feat.feat_id != HEADER_LAST_FEATURE) {
248 pr_err("failed: wrong feature ID: %" PRI_lu64 "\n",
249 event->feat.feat_id);
250 return -1;
251 } else if (rep->header_only) {
252 session_done = 1;
253 }
254
255 /*
256 * (feat_id = HEADER_LAST_FEATURE) is the end marker which
257 * means all features are received, now we can force the
258 * group if needed.
259 */
260 setup_forced_leader(rep, session->evlist);
261 return 0;
262}
263
264static int process_sample_event(struct perf_tool *tool,
265 union perf_event *event,
266 struct perf_sample *sample,
267 struct evsel *evsel,
268 struct machine *machine)
269{
270 struct report *rep = container_of(tool, struct report, tool);
271 struct addr_location al;
272 struct hist_entry_iter iter = {
273 .evsel = evsel,
274 .sample = sample,
275 .hide_unresolved = symbol_conf.hide_unresolved,
276 .add_entry_cb = hist_iter__report_callback,
277 };
278 int ret = 0;
279
280 if (perf_time__ranges_skip_sample(rep->ptime_range, rep->range_num,
281 sample->time)) {
282 return 0;
283 }
284
285 if (evswitch__discard(&rep->evswitch, evsel))
286 return 0;
287
288 addr_location__init(&al);
289 if (machine__resolve(machine, &al, sample) < 0) {
290 pr_debug("problem processing %d event, skipping it.\n",
291 event->header.type);
292 ret = -1;
293 goto out_put;
294 }
295
296 if (rep->stitch_lbr)
297 thread__set_lbr_stitch_enable(al.thread, true);
298
299 if (symbol_conf.hide_unresolved && al.sym == NULL)
300 goto out_put;
301
302 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
303 goto out_put;
304
305 if (sort__mode == SORT_MODE__BRANCH) {
306 /*
307 * A non-synthesized event might not have a branch stack if
308 * branch stacks have been synthesized (using itrace options).
309 */
310 if (!sample->branch_stack)
311 goto out_put;
312
313 iter.add_entry_cb = hist_iter__branch_callback;
314 iter.ops = &hist_iter_branch;
315 } else if (rep->mem_mode) {
316 iter.ops = &hist_iter_mem;
317 } else if (symbol_conf.cumulate_callchain) {
318 iter.ops = &hist_iter_cumulative;
319 } else {
320 iter.ops = &hist_iter_normal;
321 }
322
323 if (al.map != NULL)
324 map__dso(al.map)->hit = 1;
325
326 if (ui__has_annotation() || rep->symbol_ipc || rep->total_cycles_mode) {
327 hist__account_cycles(sample->branch_stack, &al, sample,
328 rep->nonany_branch_mode,
329 &rep->total_cycles);
330 }
331
332 ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
333 if (ret < 0)
334 pr_debug("problem adding hist entry, skipping event\n");
335out_put:
336 addr_location__exit(&al);
337 return ret;
338}
339
340static int process_read_event(struct perf_tool *tool,
341 union perf_event *event,
342 struct perf_sample *sample __maybe_unused,
343 struct evsel *evsel,
344 struct machine *machine __maybe_unused)
345{
346 struct report *rep = container_of(tool, struct report, tool);
347
348 if (rep->show_threads) {
349 const char *name = evsel__name(evsel);
350 int err = perf_read_values_add_value(&rep->show_threads_values,
351 event->read.pid, event->read.tid,
352 evsel->core.idx,
353 name,
354 event->read.value);
355
356 if (err)
357 return err;
358 }
359
360 return 0;
361}
362
363/* For pipe mode, sample_type is not currently set */
364static int report__setup_sample_type(struct report *rep)
365{
366 struct perf_session *session = rep->session;
367 u64 sample_type = evlist__combined_sample_type(session->evlist);
368 bool is_pipe = perf_data__is_pipe(session->data);
369 struct evsel *evsel;
370
371 if (session->itrace_synth_opts->callchain ||
372 session->itrace_synth_opts->add_callchain ||
373 (!is_pipe &&
374 perf_header__has_feat(&session->header, HEADER_AUXTRACE) &&
375 !session->itrace_synth_opts->set))
376 sample_type |= PERF_SAMPLE_CALLCHAIN;
377
378 if (session->itrace_synth_opts->last_branch ||
379 session->itrace_synth_opts->add_last_branch)
380 sample_type |= PERF_SAMPLE_BRANCH_STACK;
381
382 if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
383 if (perf_hpp_list.parent) {
384 ui__error("Selected --sort parent, but no "
385 "callchain data. Did you call "
386 "'perf record' without -g?\n");
387 return -EINVAL;
388 }
389 if (symbol_conf.use_callchain &&
390 !symbol_conf.show_branchflag_count) {
391 ui__error("Selected -g or --branch-history.\n"
392 "But no callchain or branch data.\n"
393 "Did you call 'perf record' without -g or -b?\n");
394 return -1;
395 }
396 } else if (!callchain_param.enabled &&
397 callchain_param.mode != CHAIN_NONE &&
398 !symbol_conf.use_callchain) {
399 symbol_conf.use_callchain = true;
400 if (callchain_register_param(&callchain_param) < 0) {
401 ui__error("Can't register callchain params.\n");
402 return -EINVAL;
403 }
404 }
405
406 if (symbol_conf.cumulate_callchain) {
407 /* Silently ignore if callchain is missing */
408 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
409 symbol_conf.cumulate_callchain = false;
410 perf_hpp__cancel_cumulate();
411 }
412 }
413
414 if (sort__mode == SORT_MODE__BRANCH) {
415 if (!is_pipe &&
416 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
417 ui__error("Selected -b but no branch data. "
418 "Did you call perf record without -b?\n");
419 return -1;
420 }
421 }
422
423 if (sort__mode == SORT_MODE__MEMORY) {
424 /*
425 * FIXUP: prior to kernel 5.18, Arm SPE missed to set
426 * PERF_SAMPLE_DATA_SRC bit in sample type. For backward
427 * compatibility, set the bit if it's an old perf data file.
428 */
429 evlist__for_each_entry(session->evlist, evsel) {
430 if (strstr(evsel->name, "arm_spe") &&
431 !(sample_type & PERF_SAMPLE_DATA_SRC)) {
432 evsel->core.attr.sample_type |= PERF_SAMPLE_DATA_SRC;
433 sample_type |= PERF_SAMPLE_DATA_SRC;
434 }
435 }
436
437 if (!is_pipe && !(sample_type & PERF_SAMPLE_DATA_SRC)) {
438 ui__error("Selected --mem-mode but no mem data. "
439 "Did you call perf record without -d?\n");
440 return -1;
441 }
442 }
443
444 callchain_param_setup(sample_type, perf_env__arch(&rep->session->header.env));
445
446 if (rep->stitch_lbr && (callchain_param.record_mode != CALLCHAIN_LBR)) {
447 ui__warning("Can't find LBR callchain. Switch off --stitch-lbr.\n"
448 "Please apply --call-graph lbr when recording.\n");
449 rep->stitch_lbr = false;
450 }
451
452 /* ??? handle more cases than just ANY? */
453 if (!(evlist__combined_branch_type(session->evlist) & PERF_SAMPLE_BRANCH_ANY))
454 rep->nonany_branch_mode = true;
455
456#if !defined(HAVE_LIBUNWIND_SUPPORT) && !defined(HAVE_DWARF_SUPPORT)
457 if (dwarf_callchain_users) {
458 ui__warning("Please install libunwind or libdw "
459 "development packages during the perf build.\n");
460 }
461#endif
462
463 return 0;
464}
465
466static void sig_handler(int sig __maybe_unused)
467{
468 session_done = 1;
469}
470
471static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
472 const char *evname, FILE *fp)
473{
474 size_t ret;
475 char unit;
476 unsigned long nr_samples = hists->stats.nr_samples;
477 u64 nr_events = hists->stats.total_period;
478 struct evsel *evsel = hists_to_evsel(hists);
479 char buf[512];
480 size_t size = sizeof(buf);
481 int socked_id = hists->socket_filter;
482
483 if (quiet)
484 return 0;
485
486 if (symbol_conf.filter_relative) {
487 nr_samples = hists->stats.nr_non_filtered_samples;
488 nr_events = hists->stats.total_non_filtered_period;
489 }
490
491 if (evsel__is_group_event(evsel)) {
492 struct evsel *pos;
493
494 evsel__group_desc(evsel, buf, size);
495 evname = buf;
496
497 for_each_group_member(pos, evsel) {
498 const struct hists *pos_hists = evsel__hists(pos);
499
500 if (symbol_conf.filter_relative) {
501 nr_samples += pos_hists->stats.nr_non_filtered_samples;
502 nr_events += pos_hists->stats.total_non_filtered_period;
503 } else {
504 nr_samples += pos_hists->stats.nr_samples;
505 nr_events += pos_hists->stats.total_period;
506 }
507 }
508 }
509
510 nr_samples = convert_unit(nr_samples, &unit);
511 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
512 if (evname != NULL) {
513 ret += fprintf(fp, " of event%s '%s'",
514 evsel->core.nr_members > 1 ? "s" : "", evname);
515 }
516
517 if (rep->time_str)
518 ret += fprintf(fp, " (time slices: %s)", rep->time_str);
519
520 if (symbol_conf.show_ref_callgraph && evname && strstr(evname, "call-graph=no")) {
521 ret += fprintf(fp, ", show reference callgraph");
522 }
523
524 if (rep->mem_mode) {
525 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
526 ret += fprintf(fp, "\n# Sort order : %s", sort_order ? : default_mem_sort_order);
527 } else
528 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
529
530 if (socked_id > -1)
531 ret += fprintf(fp, "\n# Processor Socket: %d", socked_id);
532
533 return ret + fprintf(fp, "\n#\n");
534}
535
536static int evlist__tui_block_hists_browse(struct evlist *evlist, struct report *rep)
537{
538 struct evsel *pos;
539 int i = 0, ret;
540
541 evlist__for_each_entry(evlist, pos) {
542 ret = report__browse_block_hists(&rep->block_reports[i++].hist,
543 rep->min_percent, pos,
544 &rep->session->header.env);
545 if (ret != 0)
546 return ret;
547 }
548
549 return 0;
550}
551
552static int evlist__tty_browse_hists(struct evlist *evlist, struct report *rep, const char *help)
553{
554 struct evsel *pos;
555 int i = 0;
556
557 if (!quiet) {
558 fprintf(stdout, "#\n# Total Lost Samples: %" PRIu64 "\n#\n",
559 evlist->stats.total_lost_samples);
560 }
561
562 evlist__for_each_entry(evlist, pos) {
563 struct hists *hists = evsel__hists(pos);
564 const char *evname = evsel__name(pos);
565
566 if (symbol_conf.event_group && !evsel__is_group_leader(pos))
567 continue;
568
569 if (rep->skip_empty && !hists->stats.nr_samples)
570 continue;
571
572 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
573
574 if (rep->total_cycles_mode) {
575 report__browse_block_hists(&rep->block_reports[i++].hist,
576 rep->min_percent, pos, NULL);
577 continue;
578 }
579
580 hists__fprintf(hists, !quiet, 0, 0, rep->min_percent, stdout,
581 !(symbol_conf.use_callchain ||
582 symbol_conf.show_branchflag_count));
583 fprintf(stdout, "\n\n");
584 }
585
586 if (!quiet)
587 fprintf(stdout, "#\n# (%s)\n#\n", help);
588
589 if (rep->show_threads) {
590 bool style = !strcmp(rep->pretty_printing_style, "raw");
591 perf_read_values_display(stdout, &rep->show_threads_values,
592 style);
593 perf_read_values_destroy(&rep->show_threads_values);
594 }
595
596 if (sort__mode == SORT_MODE__BRANCH)
597 branch_type_stat_display(stdout, &rep->brtype_stat);
598
599 return 0;
600}
601
602static void report__warn_kptr_restrict(const struct report *rep)
603{
604 struct map *kernel_map = machine__kernel_map(&rep->session->machines.host);
605 struct kmap *kernel_kmap = kernel_map ? map__kmap(kernel_map) : NULL;
606
607 if (evlist__exclude_kernel(rep->session->evlist))
608 return;
609
610 if (kernel_map == NULL ||
611 (map__dso(kernel_map)->hit &&
612 (kernel_kmap->ref_reloc_sym == NULL ||
613 kernel_kmap->ref_reloc_sym->addr == 0))) {
614 const char *desc =
615 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
616 "can't be resolved.";
617
618 if (kernel_map && map__has_symbols(kernel_map)) {
619 desc = "If some relocation was applied (e.g. "
620 "kexec) symbols may be misresolved.";
621 }
622
623 ui__warning(
624"Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
625"Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
626"Samples in kernel modules can't be resolved as well.\n\n",
627 desc);
628 }
629}
630
631static int report__gtk_browse_hists(struct report *rep, const char *help)
632{
633 int (*hist_browser)(struct evlist *evlist, const char *help,
634 struct hist_browser_timer *timer, float min_pcnt);
635
636 hist_browser = dlsym(perf_gtk_handle, "evlist__gtk_browse_hists");
637
638 if (hist_browser == NULL) {
639 ui__error("GTK browser not found!\n");
640 return -1;
641 }
642
643 return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
644}
645
646static int report__browse_hists(struct report *rep)
647{
648 int ret;
649 struct perf_session *session = rep->session;
650 struct evlist *evlist = session->evlist;
651 char *help = NULL, *path = NULL;
652
653 path = system_path(TIPDIR);
654 if (perf_tip(&help, path) || help == NULL) {
655 /* fallback for people who don't install perf ;-) */
656 free(path);
657 path = system_path(DOCDIR);
658 if (perf_tip(&help, path) || help == NULL)
659 help = strdup("Cannot load tips.txt file, please install perf!");
660 }
661 free(path);
662
663 switch (use_browser) {
664 case 1:
665 if (rep->total_cycles_mode) {
666 ret = evlist__tui_block_hists_browse(evlist, rep);
667 break;
668 }
669
670 ret = evlist__tui_browse_hists(evlist, help, NULL, rep->min_percent,
671 &session->header.env, true);
672 /*
673 * Usually "ret" is the last pressed key, and we only
674 * care if the key notifies us to switch data file.
675 */
676 if (ret != K_SWITCH_INPUT_DATA && ret != K_RELOAD)
677 ret = 0;
678 break;
679 case 2:
680 ret = report__gtk_browse_hists(rep, help);
681 break;
682 default:
683 ret = evlist__tty_browse_hists(evlist, rep, help);
684 break;
685 }
686 free(help);
687 return ret;
688}
689
690static int report__collapse_hists(struct report *rep)
691{
692 struct perf_session *session = rep->session;
693 struct evlist *evlist = session->evlist;
694 struct ui_progress prog;
695 struct evsel *pos;
696 int ret = 0;
697
698 /*
699 * The pipe data needs to setup hierarchy hpp formats now, because it
700 * cannot know about evsels in the data before reading the data. The
701 * normal file data saves the event (attribute) info in the header
702 * section, but pipe does not have the luxury.
703 */
704 if (perf_data__is_pipe(session->data)) {
705 if (perf_hpp__setup_hists_formats(&perf_hpp_list, evlist) < 0) {
706 ui__error("Failed to setup hierarchy output formats\n");
707 return -1;
708 }
709 }
710
711 ui_progress__init(&prog, rep->nr_entries, "Merging related events...");
712
713 evlist__for_each_entry(rep->session->evlist, pos) {
714 struct hists *hists = evsel__hists(pos);
715
716 if (pos->core.idx == 0)
717 hists->symbol_filter_str = rep->symbol_filter_str;
718
719 hists->socket_filter = rep->socket_filter;
720
721 ret = hists__collapse_resort(hists, &prog);
722 if (ret < 0)
723 break;
724
725 /* Non-group events are considered as leader */
726 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
727 struct hists *leader_hists = evsel__hists(evsel__leader(pos));
728
729 hists__match(leader_hists, hists);
730 hists__link(leader_hists, hists);
731 }
732 }
733
734 ui_progress__finish();
735 return ret;
736}
737
738static int hists__resort_cb(struct hist_entry *he, void *arg)
739{
740 struct report *rep = arg;
741 struct symbol *sym = he->ms.sym;
742
743 if (rep->symbol_ipc && sym && !sym->annotate2) {
744 struct evsel *evsel = hists_to_evsel(he->hists);
745
746 symbol__annotate2(&he->ms, evsel, NULL);
747 }
748
749 return 0;
750}
751
752static void report__output_resort(struct report *rep)
753{
754 struct ui_progress prog;
755 struct evsel *pos;
756
757 ui_progress__init(&prog, rep->nr_entries, "Sorting events for output...");
758
759 evlist__for_each_entry(rep->session->evlist, pos) {
760 evsel__output_resort_cb(pos, &prog, hists__resort_cb, rep);
761 }
762
763 ui_progress__finish();
764}
765
766static int count_sample_event(struct perf_tool *tool __maybe_unused,
767 union perf_event *event __maybe_unused,
768 struct perf_sample *sample __maybe_unused,
769 struct evsel *evsel,
770 struct machine *machine __maybe_unused)
771{
772 struct hists *hists = evsel__hists(evsel);
773
774 hists__inc_nr_events(hists);
775 return 0;
776}
777
778static int count_lost_samples_event(struct perf_tool *tool,
779 union perf_event *event,
780 struct perf_sample *sample,
781 struct machine *machine __maybe_unused)
782{
783 struct report *rep = container_of(tool, struct report, tool);
784 struct evsel *evsel;
785
786 evsel = evlist__id2evsel(rep->session->evlist, sample->id);
787 if (evsel) {
788 hists__inc_nr_lost_samples(evsel__hists(evsel),
789 event->lost_samples.lost);
790 }
791 return 0;
792}
793
794static int process_attr(struct perf_tool *tool __maybe_unused,
795 union perf_event *event,
796 struct evlist **pevlist);
797
798static void stats_setup(struct report *rep)
799{
800 memset(&rep->tool, 0, sizeof(rep->tool));
801 rep->tool.attr = process_attr;
802 rep->tool.sample = count_sample_event;
803 rep->tool.lost_samples = count_lost_samples_event;
804 rep->tool.no_warn = true;
805}
806
807static int stats_print(struct report *rep)
808{
809 struct perf_session *session = rep->session;
810
811 perf_session__fprintf_nr_events(session, stdout, rep->skip_empty);
812 evlist__fprintf_nr_events(session->evlist, stdout, rep->skip_empty);
813 return 0;
814}
815
816static void tasks_setup(struct report *rep)
817{
818 memset(&rep->tool, 0, sizeof(rep->tool));
819 rep->tool.ordered_events = true;
820 if (rep->mmaps_mode) {
821 rep->tool.mmap = perf_event__process_mmap;
822 rep->tool.mmap2 = perf_event__process_mmap2;
823 }
824 rep->tool.attr = process_attr;
825 rep->tool.comm = perf_event__process_comm;
826 rep->tool.exit = perf_event__process_exit;
827 rep->tool.fork = perf_event__process_fork;
828 rep->tool.no_warn = true;
829}
830
831struct task {
832 struct thread *thread;
833 struct list_head list;
834 struct list_head children;
835};
836
837static struct task *tasks_list(struct task *task, struct machine *machine)
838{
839 struct thread *parent_thread, *thread = task->thread;
840 struct task *parent_task;
841
842 /* Already listed. */
843 if (!list_empty(&task->list))
844 return NULL;
845
846 /* Last one in the chain. */
847 if (thread__ppid(thread) == -1)
848 return task;
849
850 parent_thread = machine__find_thread(machine, -1, thread__ppid(thread));
851 if (!parent_thread)
852 return ERR_PTR(-ENOENT);
853
854 parent_task = thread__priv(parent_thread);
855 thread__put(parent_thread);
856 list_add_tail(&task->list, &parent_task->children);
857 return tasks_list(parent_task, machine);
858}
859
860struct maps__fprintf_task_args {
861 int indent;
862 FILE *fp;
863 size_t printed;
864};
865
866static int maps__fprintf_task_cb(struct map *map, void *data)
867{
868 struct maps__fprintf_task_args *args = data;
869 const struct dso *dso = map__dso(map);
870 u32 prot = map__prot(map);
871 int ret;
872
873 ret = fprintf(args->fp,
874 "%*s %" PRIx64 "-%" PRIx64 " %c%c%c%c %08" PRIx64 " %" PRIu64 " %s\n",
875 args->indent, "", map__start(map), map__end(map),
876 prot & PROT_READ ? 'r' : '-',
877 prot & PROT_WRITE ? 'w' : '-',
878 prot & PROT_EXEC ? 'x' : '-',
879 map__flags(map) ? 's' : 'p',
880 map__pgoff(map),
881 dso->id.ino, dso->name);
882
883 if (ret < 0)
884 return ret;
885
886 args->printed += ret;
887 return 0;
888}
889
890static size_t maps__fprintf_task(struct maps *maps, int indent, FILE *fp)
891{
892 struct maps__fprintf_task_args args = {
893 .indent = indent,
894 .fp = fp,
895 .printed = 0,
896 };
897
898 maps__for_each_map(maps, maps__fprintf_task_cb, &args);
899
900 return args.printed;
901}
902
903static void task__print_level(struct task *task, FILE *fp, int level)
904{
905 struct thread *thread = task->thread;
906 struct task *child;
907 int comm_indent = fprintf(fp, " %8d %8d %8d |%*s",
908 thread__pid(thread), thread__tid(thread),
909 thread__ppid(thread), level, "");
910
911 fprintf(fp, "%s\n", thread__comm_str(thread));
912
913 maps__fprintf_task(thread__maps(thread), comm_indent, fp);
914
915 if (!list_empty(&task->children)) {
916 list_for_each_entry(child, &task->children, list)
917 task__print_level(child, fp, level + 1);
918 }
919}
920
921static int tasks_print(struct report *rep, FILE *fp)
922{
923 struct perf_session *session = rep->session;
924 struct machine *machine = &session->machines.host;
925 struct task *tasks, *task;
926 unsigned int nr = 0, itask = 0, i;
927 struct rb_node *nd;
928 LIST_HEAD(list);
929
930 /*
931 * No locking needed while accessing machine->threads,
932 * because --tasks is single threaded command.
933 */
934
935 /* Count all the threads. */
936 for (i = 0; i < THREADS__TABLE_SIZE; i++)
937 nr += machine->threads[i].nr;
938
939 tasks = malloc(sizeof(*tasks) * nr);
940 if (!tasks)
941 return -ENOMEM;
942
943 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
944 struct threads *threads = &machine->threads[i];
945
946 for (nd = rb_first_cached(&threads->entries); nd;
947 nd = rb_next(nd)) {
948 task = tasks + itask++;
949
950 task->thread = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
951 INIT_LIST_HEAD(&task->children);
952 INIT_LIST_HEAD(&task->list);
953 thread__set_priv(task->thread, task);
954 }
955 }
956
957 /*
958 * Iterate every task down to the unprocessed parent
959 * and link all in task children list. Task with no
960 * parent is added into 'list'.
961 */
962 for (itask = 0; itask < nr; itask++) {
963 task = tasks + itask;
964
965 if (!list_empty(&task->list))
966 continue;
967
968 task = tasks_list(task, machine);
969 if (IS_ERR(task)) {
970 pr_err("Error: failed to process tasks\n");
971 free(tasks);
972 return PTR_ERR(task);
973 }
974
975 if (task)
976 list_add_tail(&task->list, &list);
977 }
978
979 fprintf(fp, "# %8s %8s %8s %s\n", "pid", "tid", "ppid", "comm");
980
981 list_for_each_entry(task, &list, list)
982 task__print_level(task, fp, 0);
983
984 free(tasks);
985 return 0;
986}
987
988static int __cmd_report(struct report *rep)
989{
990 int ret;
991 struct perf_session *session = rep->session;
992 struct evsel *pos;
993 struct perf_data *data = session->data;
994
995 signal(SIGINT, sig_handler);
996
997 if (rep->cpu_list) {
998 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
999 rep->cpu_bitmap);
1000 if (ret) {
1001 ui__error("failed to set cpu bitmap\n");
1002 return ret;
1003 }
1004 session->itrace_synth_opts->cpu_bitmap = rep->cpu_bitmap;
1005 }
1006
1007 if (rep->show_threads) {
1008 ret = perf_read_values_init(&rep->show_threads_values);
1009 if (ret)
1010 return ret;
1011 }
1012
1013 ret = report__setup_sample_type(rep);
1014 if (ret) {
1015 /* report__setup_sample_type() already showed error message */
1016 return ret;
1017 }
1018
1019 if (rep->stats_mode)
1020 stats_setup(rep);
1021
1022 if (rep->tasks_mode)
1023 tasks_setup(rep);
1024
1025 ret = perf_session__process_events(session);
1026 if (ret) {
1027 ui__error("failed to process sample\n");
1028 return ret;
1029 }
1030
1031 evlist__check_mem_load_aux(session->evlist);
1032
1033 if (rep->stats_mode)
1034 return stats_print(rep);
1035
1036 if (rep->tasks_mode)
1037 return tasks_print(rep, stdout);
1038
1039 report__warn_kptr_restrict(rep);
1040
1041 evlist__for_each_entry(session->evlist, pos)
1042 rep->nr_entries += evsel__hists(pos)->nr_entries;
1043
1044 if (use_browser == 0) {
1045 if (verbose > 3)
1046 perf_session__fprintf(session, stdout);
1047
1048 if (verbose > 2)
1049 perf_session__fprintf_dsos(session, stdout);
1050
1051 if (dump_trace) {
1052 perf_session__fprintf_nr_events(session, stdout,
1053 rep->skip_empty);
1054 evlist__fprintf_nr_events(session->evlist, stdout,
1055 rep->skip_empty);
1056 return 0;
1057 }
1058 }
1059
1060 ret = report__collapse_hists(rep);
1061 if (ret) {
1062 ui__error("failed to process hist entry\n");
1063 return ret;
1064 }
1065
1066 if (session_done())
1067 return 0;
1068
1069 /*
1070 * recalculate number of entries after collapsing since it
1071 * might be changed during the collapse phase.
1072 */
1073 rep->nr_entries = 0;
1074 evlist__for_each_entry(session->evlist, pos)
1075 rep->nr_entries += evsel__hists(pos)->nr_entries;
1076
1077 if (rep->nr_entries == 0) {
1078 ui__error("The %s data has no samples!\n", data->path);
1079 return 0;
1080 }
1081
1082 report__output_resort(rep);
1083
1084 if (rep->total_cycles_mode) {
1085 int block_hpps[6] = {
1086 PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_PCT,
1087 PERF_HPP_REPORT__BLOCK_LBR_CYCLES,
1088 PERF_HPP_REPORT__BLOCK_CYCLES_PCT,
1089 PERF_HPP_REPORT__BLOCK_AVG_CYCLES,
1090 PERF_HPP_REPORT__BLOCK_RANGE,
1091 PERF_HPP_REPORT__BLOCK_DSO,
1092 };
1093
1094 rep->block_reports = block_info__create_report(session->evlist,
1095 rep->total_cycles,
1096 block_hpps, 6,
1097 &rep->nr_block_reports);
1098 if (!rep->block_reports)
1099 return -1;
1100 }
1101
1102 return report__browse_hists(rep);
1103}
1104
1105static int
1106report_parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1107{
1108 struct callchain_param *callchain = opt->value;
1109
1110 callchain->enabled = !unset;
1111 /*
1112 * --no-call-graph
1113 */
1114 if (unset) {
1115 symbol_conf.use_callchain = false;
1116 callchain->mode = CHAIN_NONE;
1117 return 0;
1118 }
1119
1120 return parse_callchain_report_opt(arg);
1121}
1122
1123static int
1124parse_time_quantum(const struct option *opt, const char *arg,
1125 int unset __maybe_unused)
1126{
1127 unsigned long *time_q = opt->value;
1128 char *end;
1129
1130 *time_q = strtoul(arg, &end, 0);
1131 if (end == arg)
1132 goto parse_err;
1133 if (*time_q == 0) {
1134 pr_err("time quantum cannot be 0");
1135 return -1;
1136 }
1137 end = skip_spaces(end);
1138 if (*end == 0)
1139 return 0;
1140 if (!strcmp(end, "s")) {
1141 *time_q *= NSEC_PER_SEC;
1142 return 0;
1143 }
1144 if (!strcmp(end, "ms")) {
1145 *time_q *= NSEC_PER_MSEC;
1146 return 0;
1147 }
1148 if (!strcmp(end, "us")) {
1149 *time_q *= NSEC_PER_USEC;
1150 return 0;
1151 }
1152 if (!strcmp(end, "ns"))
1153 return 0;
1154parse_err:
1155 pr_err("Cannot parse time quantum `%s'\n", arg);
1156 return -1;
1157}
1158
1159int
1160report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
1161 const char *arg, int unset __maybe_unused)
1162{
1163 if (arg) {
1164 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
1165 if (err) {
1166 char buf[BUFSIZ];
1167 regerror(err, &ignore_callees_regex, buf, sizeof(buf));
1168 pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
1169 return -1;
1170 }
1171 have_ignore_callees = 1;
1172 }
1173
1174 return 0;
1175}
1176
1177static int
1178parse_branch_mode(const struct option *opt,
1179 const char *str __maybe_unused, int unset)
1180{
1181 int *branch_mode = opt->value;
1182
1183 *branch_mode = !unset;
1184 return 0;
1185}
1186
1187static int
1188parse_percent_limit(const struct option *opt, const char *str,
1189 int unset __maybe_unused)
1190{
1191 struct report *rep = opt->value;
1192 double pcnt = strtof(str, NULL);
1193
1194 rep->min_percent = pcnt;
1195 callchain_param.min_percent = pcnt;
1196 return 0;
1197}
1198
1199static int process_attr(struct perf_tool *tool __maybe_unused,
1200 union perf_event *event,
1201 struct evlist **pevlist)
1202{
1203 u64 sample_type;
1204 int err;
1205
1206 err = perf_event__process_attr(tool, event, pevlist);
1207 if (err)
1208 return err;
1209
1210 /*
1211 * Check if we need to enable callchains based
1212 * on events sample_type.
1213 */
1214 sample_type = evlist__combined_sample_type(*pevlist);
1215 callchain_param_setup(sample_type, perf_env__arch((*pevlist)->env));
1216 return 0;
1217}
1218
1219int cmd_report(int argc, const char **argv)
1220{
1221 struct perf_session *session;
1222 struct itrace_synth_opts itrace_synth_opts = { .set = 0, };
1223 struct stat st;
1224 bool has_br_stack = false;
1225 int branch_mode = -1;
1226 int last_key = 0;
1227 bool branch_call_mode = false;
1228#define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent"
1229 static const char report_callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
1230 CALLCHAIN_REPORT_HELP
1231 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
1232 char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
1233 const char * const report_usage[] = {
1234 "perf report [<options>]",
1235 NULL
1236 };
1237 struct report report = {
1238 .tool = {
1239 .sample = process_sample_event,
1240 .mmap = perf_event__process_mmap,
1241 .mmap2 = perf_event__process_mmap2,
1242 .comm = perf_event__process_comm,
1243 .namespaces = perf_event__process_namespaces,
1244 .cgroup = perf_event__process_cgroup,
1245 .exit = perf_event__process_exit,
1246 .fork = perf_event__process_fork,
1247 .lost = perf_event__process_lost,
1248 .read = process_read_event,
1249 .attr = process_attr,
1250#ifdef HAVE_LIBTRACEEVENT
1251 .tracing_data = perf_event__process_tracing_data,
1252#endif
1253 .build_id = perf_event__process_build_id,
1254 .id_index = perf_event__process_id_index,
1255 .auxtrace_info = perf_event__process_auxtrace_info,
1256 .auxtrace = perf_event__process_auxtrace,
1257 .event_update = perf_event__process_event_update,
1258 .feature = process_feature_event,
1259 .ordered_events = true,
1260 .ordering_requires_timestamps = true,
1261 },
1262 .max_stack = PERF_MAX_STACK_DEPTH,
1263 .pretty_printing_style = "normal",
1264 .socket_filter = -1,
1265 .skip_empty = true,
1266 };
1267 char *sort_order_help = sort_help("sort by key(s):");
1268 char *field_order_help = sort_help("output field(s): overhead period sample ");
1269 const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
1270 const struct option options[] = {
1271 OPT_STRING('i', "input", &input_name, "file",
1272 "input file name"),
1273 OPT_INCR('v', "verbose", &verbose,
1274 "be more verbose (show symbol address, etc)"),
1275 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"),
1276 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1277 "dump raw trace in ASCII"),
1278 OPT_BOOLEAN(0, "stats", &report.stats_mode, "Display event stats"),
1279 OPT_BOOLEAN(0, "tasks", &report.tasks_mode, "Display recorded tasks"),
1280 OPT_BOOLEAN(0, "mmaps", &report.mmaps_mode, "Display recorded tasks memory maps"),
1281 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1282 "file", "vmlinux pathname"),
1283 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1284 "don't load vmlinux even if found"),
1285 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1286 "file", "kallsyms pathname"),
1287 OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
1288 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1289 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1290 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1291 "Show a column with the number of samples"),
1292 OPT_BOOLEAN('T', "threads", &report.show_threads,
1293 "Show per-thread event counters"),
1294 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
1295 "pretty printing style key: normal raw"),
1296#ifdef HAVE_SLANG_SUPPORT
1297 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
1298#endif
1299#ifdef HAVE_GTK2_SUPPORT
1300 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
1301#endif
1302 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
1303 "Use the stdio interface"),
1304 OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
1305 OPT_BOOLEAN(0, "header-only", &report.header_only,
1306 "Show only data header."),
1307 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1308 sort_order_help),
1309 OPT_STRING('F', "fields", &field_order, "key[,keys...]",
1310 field_order_help),
1311 OPT_BOOLEAN(0, "show-cpu-utilization", &symbol_conf.show_cpu_utilization,
1312 "Show sample percentage for different cpu modes"),
1313 OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
1314 "Show sample percentage for different cpu modes", PARSE_OPT_HIDDEN),
1315 OPT_STRING('p', "parent", &parent_pattern, "regex",
1316 "regex filter to identify parent, see: '--sort parent'"),
1317 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
1318 "Only display entries with parent-match"),
1319 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
1320 "print_type,threshold[,print_limit],order,sort_key[,branch],value",
1321 report_callchain_help, &report_parse_callchain_opt,
1322 callchain_default_opt),
1323 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1324 "Accumulate callchains of children and show total overhead as well. "
1325 "Enabled by default, use --no-children to disable."),
1326 OPT_INTEGER(0, "max-stack", &report.max_stack,
1327 "Set the maximum stack depth when parsing the callchain, "
1328 "anything beyond the specified depth will be ignored. "
1329 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1330 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
1331 "alias for inverted call graph"),
1332 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1333 "ignore callees of these functions in call graphs",
1334 report_parse_ignore_callees_opt),
1335 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1336 "only consider symbols in these dsos"),
1337 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1338 "only consider symbols in these comms"),
1339 OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1340 "only consider symbols in these pids"),
1341 OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1342 "only consider symbols in these tids"),
1343 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1344 "only consider these symbols"),
1345 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
1346 "only show symbols that (partially) match with this filter"),
1347 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1348 "width[,width...]",
1349 "don't try to adjust column width, use these fixed values"),
1350 OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1351 "separator for columns, no spaces will be added between "
1352 "columns '.' is reserved."),
1353 OPT_BOOLEAN('U', "hide-unresolved", &symbol_conf.hide_unresolved,
1354 "Only display entries resolved to a symbol"),
1355 OPT_CALLBACK(0, "symfs", NULL, "directory",
1356 "Look for files with symbols relative to this directory",
1357 symbol__config_symfs),
1358 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
1359 "list of cpus to profile"),
1360 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
1361 "Display extended information about perf.data file"),
1362 OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
1363 "Interleave source code with assembly code (default)"),
1364 OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
1365 "Display raw encoding of assembly instructions (default)"),
1366 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1367 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1368 OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
1369 "Add prefix to source file path names in programs (with --prefix-strip)"),
1370 OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
1371 "Strip first N entries of source file path name in programs (with --prefix)"),
1372 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1373 "Show a column with the sum of periods"),
1374 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group, &report.group_set,
1375 "Show event group information together"),
1376 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1377 "Sort the output by the event at the index n in group. "
1378 "If n is invalid, sort by the first event. "
1379 "WARNING: should be used on grouped events."),
1380 OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
1381 "use branch records for per branch histogram filling",
1382 parse_branch_mode),
1383 OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1384 "add last branch records to call history"),
1385 OPT_STRING(0, "objdump", &objdump_path, "path",
1386 "objdump binary to use for disassembly and annotations"),
1387 OPT_STRING(0, "addr2line", &addr2line_path, "path",
1388 "addr2line binary to use for line numbers"),
1389 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
1390 "Disable symbol demangling"),
1391 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1392 "Enable kernel symbol demangling"),
1393 OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
1394 OPT_INTEGER(0, "samples", &symbol_conf.res_sample,
1395 "Number of samples to save per histogram entry for individual browsing"),
1396 OPT_CALLBACK(0, "percent-limit", &report, "percent",
1397 "Don't show entries under that percent", parse_percent_limit),
1398 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1399 "how to display percentage of filtered entries", parse_filter_percentage),
1400 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
1401 "Instruction Tracing options\n" ITRACE_HELP,
1402 itrace_parse_synth_opts),
1403 OPT_BOOLEAN(0, "full-source-path", &srcline_full_filename,
1404 "Show full source file name path for source lines"),
1405 OPT_BOOLEAN(0, "show-ref-call-graph", &symbol_conf.show_ref_callgraph,
1406 "Show callgraph from reference event"),
1407 OPT_BOOLEAN(0, "stitch-lbr", &report.stitch_lbr,
1408 "Enable LBR callgraph stitching approach"),
1409 OPT_INTEGER(0, "socket-filter", &report.socket_filter,
1410 "only show processor socket that match with this filter"),
1411 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1412 "Show raw trace event output (do not use print fmt or plugins)"),
1413 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1414 "Show entries in a hierarchy"),
1415 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
1416 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
1417 stdio__config_color, "always"),
1418 OPT_STRING(0, "time", &report.time_str, "str",
1419 "Time span of interest (start,stop)"),
1420 OPT_BOOLEAN(0, "inline", &symbol_conf.inline_name,
1421 "Show inline function"),
1422 OPT_CALLBACK(0, "percent-type", &annotate_opts, "local-period",
1423 "Set percent type local/global-period/hits",
1424 annotate_parse_percent_type),
1425 OPT_BOOLEAN(0, "ns", &symbol_conf.nanosecs, "Show times in nanosecs"),
1426 OPT_CALLBACK(0, "time-quantum", &symbol_conf.time_quantum, "time (ms|us|ns|s)",
1427 "Set time quantum for time sort key (default 100ms)",
1428 parse_time_quantum),
1429 OPTS_EVSWITCH(&report.evswitch),
1430 OPT_BOOLEAN(0, "total-cycles", &report.total_cycles_mode,
1431 "Sort all blocks by 'Sampled Cycles%'"),
1432 OPT_BOOLEAN(0, "disable-order", &report.disable_order,
1433 "Disable raw trace ordering"),
1434 OPT_BOOLEAN(0, "skip-empty", &report.skip_empty,
1435 "Do not display empty (or dummy) events in the output"),
1436 OPT_END()
1437 };
1438 struct perf_data data = {
1439 .mode = PERF_DATA_MODE_READ,
1440 };
1441 int ret = hists__init();
1442 char sort_tmp[128];
1443
1444 if (ret < 0)
1445 goto exit;
1446
1447 /*
1448 * tasks_mode require access to exited threads to list those that are in
1449 * the data file. Off-cpu events are synthesized after other events and
1450 * reference exited threads.
1451 */
1452 symbol_conf.keep_exited_threads = true;
1453
1454 annotation_options__init();
1455
1456 ret = perf_config(report__config, &report);
1457 if (ret)
1458 goto exit;
1459
1460 argc = parse_options(argc, argv, options, report_usage, 0);
1461 if (argc) {
1462 /*
1463 * Special case: if there's an argument left then assume that
1464 * it's a symbol filter:
1465 */
1466 if (argc > 1)
1467 usage_with_options(report_usage, options);
1468
1469 report.symbol_filter_str = argv[0];
1470 }
1471
1472 if (disassembler_style) {
1473 annotate_opts.disassembler_style = strdup(disassembler_style);
1474 if (!annotate_opts.disassembler_style)
1475 return -ENOMEM;
1476 }
1477 if (objdump_path) {
1478 annotate_opts.objdump_path = strdup(objdump_path);
1479 if (!annotate_opts.objdump_path)
1480 return -ENOMEM;
1481 }
1482 if (addr2line_path) {
1483 symbol_conf.addr2line_path = strdup(addr2line_path);
1484 if (!symbol_conf.addr2line_path)
1485 return -ENOMEM;
1486 }
1487
1488 if (annotate_check_args() < 0) {
1489 ret = -EINVAL;
1490 goto exit;
1491 }
1492
1493 if (report.mmaps_mode)
1494 report.tasks_mode = true;
1495
1496 if (dump_trace && report.disable_order)
1497 report.tool.ordered_events = false;
1498
1499 if (quiet)
1500 perf_quiet_option();
1501
1502 ret = symbol__validate_sym_arguments();
1503 if (ret)
1504 goto exit;
1505
1506 if (report.inverted_callchain)
1507 callchain_param.order = ORDER_CALLER;
1508 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1509 callchain_param.order = ORDER_CALLER;
1510
1511 if ((itrace_synth_opts.callchain || itrace_synth_opts.add_callchain) &&
1512 (int)itrace_synth_opts.callchain_sz > report.max_stack)
1513 report.max_stack = itrace_synth_opts.callchain_sz;
1514
1515 if (!input_name || !strlen(input_name)) {
1516 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1517 input_name = "-";
1518 else
1519 input_name = "perf.data";
1520 }
1521
1522 data.path = input_name;
1523 data.force = symbol_conf.force;
1524
1525repeat:
1526 session = perf_session__new(&data, &report.tool);
1527 if (IS_ERR(session)) {
1528 ret = PTR_ERR(session);
1529 goto exit;
1530 }
1531
1532 ret = evswitch__init(&report.evswitch, session->evlist, stderr);
1533 if (ret)
1534 goto exit;
1535
1536 if (zstd_init(&(session->zstd_data), 0) < 0)
1537 pr_warning("Decompression initialization failed. Reported data may be incomplete.\n");
1538
1539 if (report.queue_size) {
1540 ordered_events__set_alloc_size(&session->ordered_events,
1541 report.queue_size);
1542 }
1543
1544 session->itrace_synth_opts = &itrace_synth_opts;
1545
1546 report.session = session;
1547
1548 has_br_stack = perf_header__has_feat(&session->header,
1549 HEADER_BRANCH_STACK);
1550 if (evlist__combined_sample_type(session->evlist) & PERF_SAMPLE_STACK_USER)
1551 has_br_stack = false;
1552
1553 setup_forced_leader(&report, session->evlist);
1554
1555 if (symbol_conf.group_sort_idx && evlist__nr_groups(session->evlist) == 0) {
1556 parse_options_usage(NULL, options, "group-sort-idx", 0);
1557 ret = -EINVAL;
1558 goto error;
1559 }
1560
1561 if (itrace_synth_opts.last_branch || itrace_synth_opts.add_last_branch)
1562 has_br_stack = true;
1563
1564 if (has_br_stack && branch_call_mode)
1565 symbol_conf.show_branchflag_count = true;
1566
1567 memset(&report.brtype_stat, 0, sizeof(struct branch_type_stat));
1568
1569 /*
1570 * Branch mode is a tristate:
1571 * -1 means default, so decide based on the file having branch data.
1572 * 0/1 means the user chose a mode.
1573 */
1574 if (((branch_mode == -1 && has_br_stack) || branch_mode == 1) &&
1575 !branch_call_mode) {
1576 sort__mode = SORT_MODE__BRANCH;
1577 symbol_conf.cumulate_callchain = false;
1578 }
1579 if (branch_call_mode) {
1580 callchain_param.key = CCKEY_ADDRESS;
1581 callchain_param.branch_callstack = true;
1582 symbol_conf.use_callchain = true;
1583 callchain_register_param(&callchain_param);
1584 if (sort_order == NULL)
1585 sort_order = "srcline,symbol,dso";
1586 }
1587
1588 if (report.mem_mode) {
1589 if (sort__mode == SORT_MODE__BRANCH) {
1590 pr_err("branch and mem mode incompatible\n");
1591 goto error;
1592 }
1593 sort__mode = SORT_MODE__MEMORY;
1594 symbol_conf.cumulate_callchain = false;
1595 }
1596
1597 if (symbol_conf.report_hierarchy) {
1598 /* disable incompatible options */
1599 symbol_conf.cumulate_callchain = false;
1600
1601 if (field_order) {
1602 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1603 parse_options_usage(report_usage, options, "F", 1);
1604 parse_options_usage(NULL, options, "hierarchy", 0);
1605 goto error;
1606 }
1607
1608 perf_hpp_list.need_collapse = true;
1609 }
1610
1611 if (report.use_stdio)
1612 use_browser = 0;
1613#ifdef HAVE_SLANG_SUPPORT
1614 else if (report.use_tui)
1615 use_browser = 1;
1616#endif
1617#ifdef HAVE_GTK2_SUPPORT
1618 else if (report.use_gtk)
1619 use_browser = 2;
1620#endif
1621
1622 /* Force tty output for header output and per-thread stat. */
1623 if (report.header || report.header_only || report.show_threads)
1624 use_browser = 0;
1625 if (report.header || report.header_only)
1626 report.tool.show_feat_hdr = SHOW_FEAT_HEADER;
1627 if (report.show_full_info)
1628 report.tool.show_feat_hdr = SHOW_FEAT_HEADER_FULL_INFO;
1629 if (report.stats_mode || report.tasks_mode)
1630 use_browser = 0;
1631 if (report.stats_mode && report.tasks_mode) {
1632 pr_err("Error: --tasks and --mmaps can't be used together with --stats\n");
1633 goto error;
1634 }
1635
1636 if (report.total_cycles_mode) {
1637 if (sort__mode != SORT_MODE__BRANCH)
1638 report.total_cycles_mode = false;
1639 else
1640 sort_order = NULL;
1641 }
1642
1643 if (sort_order && strstr(sort_order, "type")) {
1644 report.data_type = true;
1645 annotate_opts.annotate_src = false;
1646
1647#ifndef HAVE_DWARF_GETLOCATIONS_SUPPORT
1648 pr_err("Error: Data type profiling is disabled due to missing DWARF support\n");
1649 goto error;
1650#endif
1651 }
1652
1653 if (strcmp(input_name, "-") != 0)
1654 setup_browser(true);
1655 else
1656 use_browser = 0;
1657
1658 if (sort_order && strstr(sort_order, "ipc")) {
1659 parse_options_usage(report_usage, options, "s", 1);
1660 goto error;
1661 }
1662
1663 if (sort_order && strstr(sort_order, "symbol")) {
1664 if (sort__mode == SORT_MODE__BRANCH) {
1665 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1666 sort_order, "ipc_lbr");
1667 report.symbol_ipc = true;
1668 } else {
1669 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1670 sort_order, "ipc_null");
1671 }
1672
1673 sort_order = sort_tmp;
1674 }
1675
1676 if ((last_key != K_SWITCH_INPUT_DATA && last_key != K_RELOAD) &&
1677 (setup_sorting(session->evlist) < 0)) {
1678 if (sort_order)
1679 parse_options_usage(report_usage, options, "s", 1);
1680 if (field_order)
1681 parse_options_usage(sort_order ? NULL : report_usage,
1682 options, "F", 1);
1683 goto error;
1684 }
1685
1686 if ((report.header || report.header_only) && !quiet) {
1687 perf_session__fprintf_info(session, stdout,
1688 report.show_full_info);
1689 if (report.header_only) {
1690 if (data.is_pipe) {
1691 /*
1692 * we need to process first few records
1693 * which contains PERF_RECORD_HEADER_FEATURE.
1694 */
1695 perf_session__process_events(session);
1696 }
1697 ret = 0;
1698 goto error;
1699 }
1700 } else if (use_browser == 0 && !quiet &&
1701 !report.stats_mode && !report.tasks_mode) {
1702 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
1703 stdout);
1704 }
1705
1706 /*
1707 * Only in the TUI browser we are doing integrated annotation,
1708 * so don't allocate extra space that won't be used in the stdio
1709 * implementation.
1710 */
1711 if (ui__has_annotation() || report.symbol_ipc || report.data_type ||
1712 report.total_cycles_mode) {
1713 ret = symbol__annotation_init();
1714 if (ret < 0)
1715 goto error;
1716 /*
1717 * For searching by name on the "Browse map details".
1718 * providing it only in verbose mode not to bloat too
1719 * much struct symbol.
1720 */
1721 if (verbose > 0) {
1722 /*
1723 * XXX: Need to provide a less kludgy way to ask for
1724 * more space per symbol, the u32 is for the index on
1725 * the ui browser.
1726 * See symbol__browser_index.
1727 */
1728 symbol_conf.priv_size += sizeof(u32);
1729 }
1730 annotation_config__init();
1731 }
1732
1733 if (symbol__init(&session->header.env) < 0)
1734 goto error;
1735
1736 if (report.time_str) {
1737 ret = perf_time__parse_for_ranges(report.time_str, session,
1738 &report.ptime_range,
1739 &report.range_size,
1740 &report.range_num);
1741 if (ret < 0)
1742 goto error;
1743
1744 itrace_synth_opts__set_time_range(&itrace_synth_opts,
1745 report.ptime_range,
1746 report.range_num);
1747 }
1748
1749#ifdef HAVE_LIBTRACEEVENT
1750 if (session->tevent.pevent &&
1751 tep_set_function_resolver(session->tevent.pevent,
1752 machine__resolve_kernel_addr,
1753 &session->machines.host) < 0) {
1754 pr_err("%s: failed to set libtraceevent function resolver\n",
1755 __func__);
1756 return -1;
1757 }
1758#endif
1759 sort__setup_elide(stdout);
1760
1761 ret = __cmd_report(&report);
1762 if (ret == K_SWITCH_INPUT_DATA || ret == K_RELOAD) {
1763 perf_session__delete(session);
1764 last_key = K_SWITCH_INPUT_DATA;
1765 goto repeat;
1766 } else
1767 ret = 0;
1768
1769error:
1770 if (report.ptime_range) {
1771 itrace_synth_opts__clear_time_range(&itrace_synth_opts);
1772 zfree(&report.ptime_range);
1773 }
1774
1775 if (report.block_reports) {
1776 block_info__free_report(report.block_reports,
1777 report.nr_block_reports);
1778 report.block_reports = NULL;
1779 }
1780
1781 zstd_fini(&(session->zstd_data));
1782 perf_session__delete(session);
1783exit:
1784 annotation_options__exit();
1785 free(sort_order_help);
1786 free(field_order_help);
1787 return ret;
1788}