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