1 // SPDX-License-Identifier: GPL-2.0
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.
11 #include "util/config.h"
13 #include "util/annotate.h"
14 #include "util/color.h"
16 #include <linux/list.h>
17 #include <linux/rbtree.h>
18 #include <linux/err.h>
19 #include <linux/zalloc.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"
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"
38 #include <subcmd/parse-options.h>
39 #include <subcmd/exec-cmd.h>
40 #include "util/parse-events.h"
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/branch.h"
51 #include "util/util.h" // perf_tip()
53 #include "ui/progress.h"
59 #include <linux/ctype.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>
68 #include <linux/mman.h>
71 struct perf_tool tool;
72 struct perf_session *session;
73 struct evswitch evswitch;
74 bool use_tui, use_gtk, use_stdio;
77 bool inverted_callchain;
84 bool nonany_branch_mode;
87 struct perf_read_values show_threads_values;
88 struct annotation_options annotation_opts;
89 const char *pretty_printing_style;
91 const char *symbol_filter_str;
93 struct perf_time_interval *ptime_range;
100 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
101 struct branch_type_stat brtype_stat;
105 static int report__config(const char *var, const char *value, void *cb)
107 struct report *rep = cb;
109 if (!strcmp(var, "report.group")) {
110 symbol_conf.event_group = perf_config_bool(var, value);
113 if (!strcmp(var, "report.percent-limit")) {
114 double pcnt = strtof(value, NULL);
116 rep->min_percent = pcnt;
117 callchain_param.min_percent = pcnt;
120 if (!strcmp(var, "report.children")) {
121 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
124 if (!strcmp(var, "report.queue-size"))
125 return perf_config_u64(&rep->queue_size, var, value);
127 if (!strcmp(var, "report.sort_order")) {
128 default_sort_order = strdup(value);
135 static int hist_iter__report_callback(struct hist_entry_iter *iter,
136 struct addr_location *al, bool single,
140 struct report *rep = arg;
141 struct hist_entry *he = iter->he;
142 struct evsel *evsel = iter->evsel;
143 struct perf_sample *sample = iter->sample;
145 struct branch_info *bi;
147 if (!ui__has_annotation() && !rep->symbol_ipc)
150 if (sort__mode == SORT_MODE__BRANCH) {
151 bi = he->branch_info;
152 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
156 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
158 } else if (rep->mem_mode) {
160 err = addr_map_symbol__inc_samples(&mi->daddr, sample, evsel);
164 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
166 } else if (symbol_conf.cumulate_callchain) {
168 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
170 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
177 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
178 struct addr_location *al __maybe_unused,
179 bool single __maybe_unused,
182 struct hist_entry *he = iter->he;
183 struct report *rep = arg;
184 struct branch_info *bi;
185 struct perf_sample *sample = iter->sample;
186 struct evsel *evsel = iter->evsel;
189 if (!ui__has_annotation() && !rep->symbol_ipc)
192 bi = he->branch_info;
193 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
197 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
199 branch_type_count(&rep->brtype_stat, &bi->flags,
200 bi->from.addr, bi->to.addr);
206 static void setup_forced_leader(struct report *report,
207 struct evlist *evlist)
209 if (report->group_set)
210 perf_evlist__force_leader(evlist);
213 static int process_feature_event(struct perf_session *session,
214 union perf_event *event)
216 struct report *rep = container_of(session->tool, struct report, tool);
218 if (event->feat.feat_id < HEADER_LAST_FEATURE)
219 return perf_event__process_feature(session, event);
221 if (event->feat.feat_id != HEADER_LAST_FEATURE) {
222 pr_err("failed: wrong feature ID: %" PRI_lu64 "\n",
223 event->feat.feat_id);
228 * (feat_id = HEADER_LAST_FEATURE) is the end marker which
229 * means all features are received, now we can force the
232 setup_forced_leader(rep, session->evlist);
236 static int process_sample_event(struct perf_tool *tool,
237 union perf_event *event,
238 struct perf_sample *sample,
240 struct machine *machine)
242 struct report *rep = container_of(tool, struct report, tool);
243 struct addr_location al;
244 struct hist_entry_iter iter = {
247 .hide_unresolved = symbol_conf.hide_unresolved,
248 .add_entry_cb = hist_iter__report_callback,
252 if (perf_time__ranges_skip_sample(rep->ptime_range, rep->range_num,
257 if (evswitch__discard(&rep->evswitch, evsel))
260 if (machine__resolve(machine, &al, sample) < 0) {
261 pr_debug("problem processing %d event, skipping it.\n",
266 if (symbol_conf.hide_unresolved && al.sym == NULL)
269 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
272 if (sort__mode == SORT_MODE__BRANCH) {
274 * A non-synthesized event might not have a branch stack if
275 * branch stacks have been synthesized (using itrace options).
277 if (!sample->branch_stack)
280 iter.add_entry_cb = hist_iter__branch_callback;
281 iter.ops = &hist_iter_branch;
282 } else if (rep->mem_mode) {
283 iter.ops = &hist_iter_mem;
284 } else if (symbol_conf.cumulate_callchain) {
285 iter.ops = &hist_iter_cumulative;
287 iter.ops = &hist_iter_normal;
291 al.map->dso->hit = 1;
293 if (ui__has_annotation() || rep->symbol_ipc) {
294 hist__account_cycles(sample->branch_stack, &al, sample,
295 rep->nonany_branch_mode);
298 ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
300 pr_debug("problem adding hist entry, skipping event\n");
302 addr_location__put(&al);
306 static int process_read_event(struct perf_tool *tool,
307 union perf_event *event,
308 struct perf_sample *sample __maybe_unused,
310 struct machine *machine __maybe_unused)
312 struct report *rep = container_of(tool, struct report, tool);
314 if (rep->show_threads) {
315 const char *name = perf_evsel__name(evsel);
316 int err = perf_read_values_add_value(&rep->show_threads_values,
317 event->read.pid, event->read.tid,
329 /* For pipe mode, sample_type is not currently set */
330 static int report__setup_sample_type(struct report *rep)
332 struct perf_session *session = rep->session;
333 u64 sample_type = perf_evlist__combined_sample_type(session->evlist);
334 bool is_pipe = perf_data__is_pipe(session->data);
336 if (session->itrace_synth_opts->callchain ||
338 perf_header__has_feat(&session->header, HEADER_AUXTRACE) &&
339 !session->itrace_synth_opts->set))
340 sample_type |= PERF_SAMPLE_CALLCHAIN;
342 if (session->itrace_synth_opts->last_branch)
343 sample_type |= PERF_SAMPLE_BRANCH_STACK;
345 if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
346 if (perf_hpp_list.parent) {
347 ui__error("Selected --sort parent, but no "
348 "callchain data. Did you call "
349 "'perf record' without -g?\n");
352 if (symbol_conf.use_callchain &&
353 !symbol_conf.show_branchflag_count) {
354 ui__error("Selected -g or --branch-history.\n"
355 "But no callchain or branch data.\n"
356 "Did you call 'perf record' without -g or -b?\n");
359 } else if (!callchain_param.enabled &&
360 callchain_param.mode != CHAIN_NONE &&
361 !symbol_conf.use_callchain) {
362 symbol_conf.use_callchain = true;
363 if (callchain_register_param(&callchain_param) < 0) {
364 ui__error("Can't register callchain params.\n");
369 if (symbol_conf.cumulate_callchain) {
370 /* Silently ignore if callchain is missing */
371 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
372 symbol_conf.cumulate_callchain = false;
373 perf_hpp__cancel_cumulate();
377 if (sort__mode == SORT_MODE__BRANCH) {
379 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
380 ui__error("Selected -b but no branch data. "
381 "Did you call perf record without -b?\n");
386 if (symbol_conf.use_callchain || symbol_conf.cumulate_callchain) {
387 if ((sample_type & PERF_SAMPLE_REGS_USER) &&
388 (sample_type & PERF_SAMPLE_STACK_USER)) {
389 callchain_param.record_mode = CALLCHAIN_DWARF;
390 dwarf_callchain_users = true;
391 } else if (sample_type & PERF_SAMPLE_BRANCH_STACK)
392 callchain_param.record_mode = CALLCHAIN_LBR;
394 callchain_param.record_mode = CALLCHAIN_FP;
397 /* ??? handle more cases than just ANY? */
398 if (!(perf_evlist__combined_branch_type(session->evlist) &
399 PERF_SAMPLE_BRANCH_ANY))
400 rep->nonany_branch_mode = true;
402 #ifndef HAVE_LIBUNWIND_SUPPORT
403 if (dwarf_callchain_users) {
404 ui__warning("Please install libunwind development packages "
405 "during the perf build.\n");
412 static void sig_handler(int sig __maybe_unused)
417 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
418 const char *evname, FILE *fp)
422 unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
423 u64 nr_events = hists->stats.total_period;
424 struct evsel *evsel = hists_to_evsel(hists);
426 size_t size = sizeof(buf);
427 int socked_id = hists->socket_filter;
432 if (symbol_conf.filter_relative) {
433 nr_samples = hists->stats.nr_non_filtered_samples;
434 nr_events = hists->stats.total_non_filtered_period;
437 if (perf_evsel__is_group_event(evsel)) {
440 perf_evsel__group_desc(evsel, buf, size);
443 for_each_group_member(pos, evsel) {
444 const struct hists *pos_hists = evsel__hists(pos);
446 if (symbol_conf.filter_relative) {
447 nr_samples += pos_hists->stats.nr_non_filtered_samples;
448 nr_events += pos_hists->stats.total_non_filtered_period;
450 nr_samples += pos_hists->stats.nr_events[PERF_RECORD_SAMPLE];
451 nr_events += pos_hists->stats.total_period;
456 nr_samples = convert_unit(nr_samples, &unit);
457 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
458 if (evname != NULL) {
459 ret += fprintf(fp, " of event%s '%s'",
460 evsel->core.nr_members > 1 ? "s" : "", evname);
464 ret += fprintf(fp, " (time slices: %s)", rep->time_str);
466 if (symbol_conf.show_ref_callgraph &&
467 strstr(evname, "call-graph=no")) {
468 ret += fprintf(fp, ", show reference callgraph");
472 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
473 ret += fprintf(fp, "\n# Sort order : %s", sort_order ? : default_mem_sort_order);
475 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
478 ret += fprintf(fp, "\n# Processor Socket: %d", socked_id);
480 return ret + fprintf(fp, "\n#\n");
483 static int perf_evlist__tty_browse_hists(struct evlist *evlist,
490 fprintf(stdout, "#\n# Total Lost Samples: %" PRIu64 "\n#\n",
491 evlist->stats.total_lost_samples);
494 evlist__for_each_entry(evlist, pos) {
495 struct hists *hists = evsel__hists(pos);
496 const char *evname = perf_evsel__name(pos);
498 if (symbol_conf.event_group &&
499 !perf_evsel__is_group_leader(pos))
502 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
503 hists__fprintf(hists, !quiet, 0, 0, rep->min_percent, stdout,
504 !(symbol_conf.use_callchain ||
505 symbol_conf.show_branchflag_count));
506 fprintf(stdout, "\n\n");
510 fprintf(stdout, "#\n# (%s)\n#\n", help);
512 if (rep->show_threads) {
513 bool style = !strcmp(rep->pretty_printing_style, "raw");
514 perf_read_values_display(stdout, &rep->show_threads_values,
516 perf_read_values_destroy(&rep->show_threads_values);
519 if (sort__mode == SORT_MODE__BRANCH)
520 branch_type_stat_display(stdout, &rep->brtype_stat);
525 static void report__warn_kptr_restrict(const struct report *rep)
527 struct map *kernel_map = machine__kernel_map(&rep->session->machines.host);
528 struct kmap *kernel_kmap = kernel_map ? map__kmap(kernel_map) : NULL;
530 if (perf_evlist__exclude_kernel(rep->session->evlist))
533 if (kernel_map == NULL ||
534 (kernel_map->dso->hit &&
535 (kernel_kmap->ref_reloc_sym == NULL ||
536 kernel_kmap->ref_reloc_sym->addr == 0))) {
538 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
539 "can't be resolved.";
541 if (kernel_map && map__has_symbols(kernel_map)) {
542 desc = "If some relocation was applied (e.g. "
543 "kexec) symbols may be misresolved.";
547 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
548 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
549 "Samples in kernel modules can't be resolved as well.\n\n",
554 static int report__gtk_browse_hists(struct report *rep, const char *help)
556 int (*hist_browser)(struct evlist *evlist, const char *help,
557 struct hist_browser_timer *timer, float min_pcnt);
559 hist_browser = dlsym(perf_gtk_handle, "perf_evlist__gtk_browse_hists");
561 if (hist_browser == NULL) {
562 ui__error("GTK browser not found!\n");
566 return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
569 static int report__browse_hists(struct report *rep)
572 struct perf_session *session = rep->session;
573 struct evlist *evlist = session->evlist;
574 const char *help = perf_tip(system_path(TIPDIR));
577 /* fallback for people who don't install perf ;-) */
578 help = perf_tip(DOCDIR);
580 help = "Cannot load tips.txt file, please install perf!";
583 switch (use_browser) {
585 ret = perf_evlist__tui_browse_hists(evlist, help, NULL,
587 &session->header.env,
588 true, &rep->annotation_opts);
590 * Usually "ret" is the last pressed key, and we only
591 * care if the key notifies us to switch data file.
593 if (ret != K_SWITCH_INPUT_DATA)
597 ret = report__gtk_browse_hists(rep, help);
600 ret = perf_evlist__tty_browse_hists(evlist, rep, help);
607 static int report__collapse_hists(struct report *rep)
609 struct ui_progress prog;
613 ui_progress__init(&prog, rep->nr_entries, "Merging related events...");
615 evlist__for_each_entry(rep->session->evlist, pos) {
616 struct hists *hists = evsel__hists(pos);
619 hists->symbol_filter_str = rep->symbol_filter_str;
621 hists->socket_filter = rep->socket_filter;
623 ret = hists__collapse_resort(hists, &prog);
627 /* Non-group events are considered as leader */
628 if (symbol_conf.event_group &&
629 !perf_evsel__is_group_leader(pos)) {
630 struct hists *leader_hists = evsel__hists(pos->leader);
632 hists__match(leader_hists, hists);
633 hists__link(leader_hists, hists);
637 ui_progress__finish();
641 static int hists__resort_cb(struct hist_entry *he, void *arg)
643 struct report *rep = arg;
644 struct symbol *sym = he->ms.sym;
646 if (rep->symbol_ipc && sym && !sym->annotate2) {
647 struct evsel *evsel = hists_to_evsel(he->hists);
649 symbol__annotate2(sym, he->ms.map, evsel,
650 &annotation__default_options, NULL);
656 static void report__output_resort(struct report *rep)
658 struct ui_progress prog;
661 ui_progress__init(&prog, rep->nr_entries, "Sorting events for output...");
663 evlist__for_each_entry(rep->session->evlist, pos) {
664 perf_evsel__output_resort_cb(pos, &prog,
665 hists__resort_cb, rep);
668 ui_progress__finish();
671 static void stats_setup(struct report *rep)
673 memset(&rep->tool, 0, sizeof(rep->tool));
674 rep->tool.no_warn = true;
677 static int stats_print(struct report *rep)
679 struct perf_session *session = rep->session;
681 perf_session__fprintf_nr_events(session, stdout);
685 static void tasks_setup(struct report *rep)
687 memset(&rep->tool, 0, sizeof(rep->tool));
688 rep->tool.ordered_events = true;
689 if (rep->mmaps_mode) {
690 rep->tool.mmap = perf_event__process_mmap;
691 rep->tool.mmap2 = perf_event__process_mmap2;
693 rep->tool.comm = perf_event__process_comm;
694 rep->tool.exit = perf_event__process_exit;
695 rep->tool.fork = perf_event__process_fork;
696 rep->tool.no_warn = true;
700 struct thread *thread;
701 struct list_head list;
702 struct list_head children;
705 static struct task *tasks_list(struct task *task, struct machine *machine)
707 struct thread *parent_thread, *thread = task->thread;
708 struct task *parent_task;
710 /* Already listed. */
711 if (!list_empty(&task->list))
714 /* Last one in the chain. */
715 if (thread->ppid == -1)
718 parent_thread = machine__find_thread(machine, -1, thread->ppid);
720 return ERR_PTR(-ENOENT);
722 parent_task = thread__priv(parent_thread);
723 list_add_tail(&task->list, &parent_task->children);
724 return tasks_list(parent_task, machine);
727 static size_t maps__fprintf_task(struct maps *maps, int indent, FILE *fp)
732 for (nd = rb_first(&maps->entries); nd; nd = rb_next(nd)) {
733 struct map *map = rb_entry(nd, struct map, rb_node);
735 printed += fprintf(fp, "%*s %" PRIx64 "-%" PRIx64 " %c%c%c%c %08" PRIx64 " %" PRIu64 " %s\n",
736 indent, "", map->start, map->end,
737 map->prot & PROT_READ ? 'r' : '-',
738 map->prot & PROT_WRITE ? 'w' : '-',
739 map->prot & PROT_EXEC ? 'x' : '-',
740 map->flags & MAP_SHARED ? 's' : 'p',
742 map->ino, map->dso->name);
748 static int map_groups__fprintf_task(struct map_groups *mg, int indent, FILE *fp)
750 return maps__fprintf_task(&mg->maps, indent, fp);
753 static void task__print_level(struct task *task, FILE *fp, int level)
755 struct thread *thread = task->thread;
757 int comm_indent = fprintf(fp, " %8d %8d %8d |%*s",
758 thread->pid_, thread->tid, thread->ppid,
761 fprintf(fp, "%s\n", thread__comm_str(thread));
763 map_groups__fprintf_task(thread->mg, comm_indent, fp);
765 if (!list_empty(&task->children)) {
766 list_for_each_entry(child, &task->children, list)
767 task__print_level(child, fp, level + 1);
771 static int tasks_print(struct report *rep, FILE *fp)
773 struct perf_session *session = rep->session;
774 struct machine *machine = &session->machines.host;
775 struct task *tasks, *task;
776 unsigned int nr = 0, itask = 0, i;
781 * No locking needed while accessing machine->threads,
782 * because --tasks is single threaded command.
785 /* Count all the threads. */
786 for (i = 0; i < THREADS__TABLE_SIZE; i++)
787 nr += machine->threads[i].nr;
789 tasks = malloc(sizeof(*tasks) * nr);
793 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
794 struct threads *threads = &machine->threads[i];
796 for (nd = rb_first_cached(&threads->entries); nd;
798 task = tasks + itask++;
800 task->thread = rb_entry(nd, struct thread, rb_node);
801 INIT_LIST_HEAD(&task->children);
802 INIT_LIST_HEAD(&task->list);
803 thread__set_priv(task->thread, task);
808 * Iterate every task down to the unprocessed parent
809 * and link all in task children list. Task with no
810 * parent is added into 'list'.
812 for (itask = 0; itask < nr; itask++) {
813 task = tasks + itask;
815 if (!list_empty(&task->list))
818 task = tasks_list(task, machine);
820 pr_err("Error: failed to process tasks\n");
822 return PTR_ERR(task);
826 list_add_tail(&task->list, &list);
829 fprintf(fp, "# %8s %8s %8s %s\n", "pid", "tid", "ppid", "comm");
831 list_for_each_entry(task, &list, list)
832 task__print_level(task, fp, 0);
838 static int __cmd_report(struct report *rep)
841 struct perf_session *session = rep->session;
843 struct perf_data *data = session->data;
845 signal(SIGINT, sig_handler);
848 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
851 ui__error("failed to set cpu bitmap\n");
854 session->itrace_synth_opts->cpu_bitmap = rep->cpu_bitmap;
857 if (rep->show_threads) {
858 ret = perf_read_values_init(&rep->show_threads_values);
863 ret = report__setup_sample_type(rep);
865 /* report__setup_sample_type() already showed error message */
875 ret = perf_session__process_events(session);
877 ui__error("failed to process sample\n");
882 return stats_print(rep);
885 return tasks_print(rep, stdout);
887 report__warn_kptr_restrict(rep);
889 evlist__for_each_entry(session->evlist, pos)
890 rep->nr_entries += evsel__hists(pos)->nr_entries;
892 if (use_browser == 0) {
894 perf_session__fprintf(session, stdout);
897 perf_session__fprintf_dsos(session, stdout);
900 perf_session__fprintf_nr_events(session, stdout);
901 perf_evlist__fprintf_nr_events(session->evlist, stdout);
906 ret = report__collapse_hists(rep);
908 ui__error("failed to process hist entry\n");
916 * recalculate number of entries after collapsing since it
917 * might be changed during the collapse phase.
920 evlist__for_each_entry(session->evlist, pos)
921 rep->nr_entries += evsel__hists(pos)->nr_entries;
923 if (rep->nr_entries == 0) {
924 ui__error("The %s data has no samples!\n", data->path);
928 report__output_resort(rep);
930 return report__browse_hists(rep);
934 report_parse_callchain_opt(const struct option *opt, const char *arg, int unset)
936 struct callchain_param *callchain = opt->value;
938 callchain->enabled = !unset;
943 symbol_conf.use_callchain = false;
944 callchain->mode = CHAIN_NONE;
948 return parse_callchain_report_opt(arg);
952 parse_time_quantum(const struct option *opt, const char *arg,
953 int unset __maybe_unused)
955 unsigned long *time_q = opt->value;
958 *time_q = strtoul(arg, &end, 0);
962 pr_err("time quantum cannot be 0");
965 end = skip_spaces(end);
968 if (!strcmp(end, "s")) {
969 *time_q *= NSEC_PER_SEC;
972 if (!strcmp(end, "ms")) {
973 *time_q *= NSEC_PER_MSEC;
976 if (!strcmp(end, "us")) {
977 *time_q *= NSEC_PER_USEC;
980 if (!strcmp(end, "ns"))
983 pr_err("Cannot parse time quantum `%s'\n", arg);
988 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
989 const char *arg, int unset __maybe_unused)
992 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
995 regerror(err, &ignore_callees_regex, buf, sizeof(buf));
996 pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
999 have_ignore_callees = 1;
1006 parse_branch_mode(const struct option *opt,
1007 const char *str __maybe_unused, int unset)
1009 int *branch_mode = opt->value;
1011 *branch_mode = !unset;
1016 parse_percent_limit(const struct option *opt, const char *str,
1017 int unset __maybe_unused)
1019 struct report *rep = opt->value;
1020 double pcnt = strtof(str, NULL);
1022 rep->min_percent = pcnt;
1023 callchain_param.min_percent = pcnt;
1027 int cmd_report(int argc, const char **argv)
1029 struct perf_session *session;
1030 struct itrace_synth_opts itrace_synth_opts = { .set = 0, };
1032 bool has_br_stack = false;
1033 int branch_mode = -1;
1034 bool branch_call_mode = false;
1035 #define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent"
1036 static const char report_callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
1037 CALLCHAIN_REPORT_HELP
1038 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
1039 char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
1040 const char * const report_usage[] = {
1041 "perf report [<options>]",
1044 struct report report = {
1046 .sample = process_sample_event,
1047 .mmap = perf_event__process_mmap,
1048 .mmap2 = perf_event__process_mmap2,
1049 .comm = perf_event__process_comm,
1050 .namespaces = perf_event__process_namespaces,
1051 .exit = perf_event__process_exit,
1052 .fork = perf_event__process_fork,
1053 .lost = perf_event__process_lost,
1054 .read = process_read_event,
1055 .attr = perf_event__process_attr,
1056 .tracing_data = perf_event__process_tracing_data,
1057 .build_id = perf_event__process_build_id,
1058 .id_index = perf_event__process_id_index,
1059 .auxtrace_info = perf_event__process_auxtrace_info,
1060 .auxtrace = perf_event__process_auxtrace,
1061 .event_update = perf_event__process_event_update,
1062 .feature = process_feature_event,
1063 .ordered_events = true,
1064 .ordering_requires_timestamps = true,
1066 .max_stack = PERF_MAX_STACK_DEPTH,
1067 .pretty_printing_style = "normal",
1068 .socket_filter = -1,
1069 .annotation_opts = annotation__default_options,
1071 const struct option options[] = {
1072 OPT_STRING('i', "input", &input_name, "file",
1074 OPT_INCR('v', "verbose", &verbose,
1075 "be more verbose (show symbol address, etc)"),
1076 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
1077 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1078 "dump raw trace in ASCII"),
1079 OPT_BOOLEAN(0, "stats", &report.stats_mode, "Display event stats"),
1080 OPT_BOOLEAN(0, "tasks", &report.tasks_mode, "Display recorded tasks"),
1081 OPT_BOOLEAN(0, "mmaps", &report.mmaps_mode, "Display recorded tasks memory maps"),
1082 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1083 "file", "vmlinux pathname"),
1084 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1085 "don't load vmlinux even if found"),
1086 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1087 "file", "kallsyms pathname"),
1088 OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
1089 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1090 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1091 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1092 "Show a column with the number of samples"),
1093 OPT_BOOLEAN('T', "threads", &report.show_threads,
1094 "Show per-thread event counters"),
1095 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
1096 "pretty printing style key: normal raw"),
1097 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
1098 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
1099 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
1100 "Use the stdio interface"),
1101 OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
1102 OPT_BOOLEAN(0, "header-only", &report.header_only,
1103 "Show only data header."),
1104 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1105 sort_help("sort by key(s):")),
1106 OPT_STRING('F', "fields", &field_order, "key[,keys...]",
1107 sort_help("output field(s): overhead period sample ")),
1108 OPT_BOOLEAN(0, "show-cpu-utilization", &symbol_conf.show_cpu_utilization,
1109 "Show sample percentage for different cpu modes"),
1110 OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
1111 "Show sample percentage for different cpu modes", PARSE_OPT_HIDDEN),
1112 OPT_STRING('p', "parent", &parent_pattern, "regex",
1113 "regex filter to identify parent, see: '--sort parent'"),
1114 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
1115 "Only display entries with parent-match"),
1116 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
1117 "print_type,threshold[,print_limit],order,sort_key[,branch],value",
1118 report_callchain_help, &report_parse_callchain_opt,
1119 callchain_default_opt),
1120 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1121 "Accumulate callchains of children and show total overhead as well"),
1122 OPT_INTEGER(0, "max-stack", &report.max_stack,
1123 "Set the maximum stack depth when parsing the callchain, "
1124 "anything beyond the specified depth will be ignored. "
1125 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1126 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
1127 "alias for inverted call graph"),
1128 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1129 "ignore callees of these functions in call graphs",
1130 report_parse_ignore_callees_opt),
1131 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1132 "only consider symbols in these dsos"),
1133 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1134 "only consider symbols in these comms"),
1135 OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1136 "only consider symbols in these pids"),
1137 OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1138 "only consider symbols in these tids"),
1139 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1140 "only consider these symbols"),
1141 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
1142 "only show symbols that (partially) match with this filter"),
1143 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1145 "don't try to adjust column width, use these fixed values"),
1146 OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1147 "separator for columns, no spaces will be added between "
1148 "columns '.' is reserved."),
1149 OPT_BOOLEAN('U', "hide-unresolved", &symbol_conf.hide_unresolved,
1150 "Only display entries resolved to a symbol"),
1151 OPT_CALLBACK(0, "symfs", NULL, "directory",
1152 "Look for files with symbols relative to this directory",
1153 symbol__config_symfs),
1154 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
1155 "list of cpus to profile"),
1156 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
1157 "Display extended information about perf.data file"),
1158 OPT_BOOLEAN(0, "source", &report.annotation_opts.annotate_src,
1159 "Interleave source code with assembly code (default)"),
1160 OPT_BOOLEAN(0, "asm-raw", &report.annotation_opts.show_asm_raw,
1161 "Display raw encoding of assembly instructions (default)"),
1162 OPT_STRING('M', "disassembler-style", &report.annotation_opts.disassembler_style, "disassembler style",
1163 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1164 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1165 "Show a column with the sum of periods"),
1166 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group, &report.group_set,
1167 "Show event group information together"),
1168 OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
1169 "use branch records for per branch histogram filling",
1171 OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1172 "add last branch records to call history"),
1173 OPT_STRING(0, "objdump", &report.annotation_opts.objdump_path, "path",
1174 "objdump binary to use for disassembly and annotations"),
1175 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
1176 "Disable symbol demangling"),
1177 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1178 "Enable kernel symbol demangling"),
1179 OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
1180 OPT_INTEGER(0, "samples", &symbol_conf.res_sample,
1181 "Number of samples to save per histogram entry for individual browsing"),
1182 OPT_CALLBACK(0, "percent-limit", &report, "percent",
1183 "Don't show entries under that percent", parse_percent_limit),
1184 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1185 "how to display percentage of filtered entries", parse_filter_percentage),
1186 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
1187 "Instruction Tracing options\n" ITRACE_HELP,
1188 itrace_parse_synth_opts),
1189 OPT_BOOLEAN(0, "full-source-path", &srcline_full_filename,
1190 "Show full source file name path for source lines"),
1191 OPT_BOOLEAN(0, "show-ref-call-graph", &symbol_conf.show_ref_callgraph,
1192 "Show callgraph from reference event"),
1193 OPT_INTEGER(0, "socket-filter", &report.socket_filter,
1194 "only show processor socket that match with this filter"),
1195 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1196 "Show raw trace event output (do not use print fmt or plugins)"),
1197 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1198 "Show entries in a hierarchy"),
1199 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
1200 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
1201 stdio__config_color, "always"),
1202 OPT_STRING(0, "time", &report.time_str, "str",
1203 "Time span of interest (start,stop)"),
1204 OPT_BOOLEAN(0, "inline", &symbol_conf.inline_name,
1205 "Show inline function"),
1206 OPT_CALLBACK(0, "percent-type", &report.annotation_opts, "local-period",
1207 "Set percent type local/global-period/hits",
1208 annotate_parse_percent_type),
1209 OPT_BOOLEAN(0, "ns", &symbol_conf.nanosecs, "Show times in nanosecs"),
1210 OPT_CALLBACK(0, "time-quantum", &symbol_conf.time_quantum, "time (ms|us|ns|s)",
1211 "Set time quantum for time sort key (default 100ms)",
1212 parse_time_quantum),
1213 OPTS_EVSWITCH(&report.evswitch),
1216 struct perf_data data = {
1217 .mode = PERF_DATA_MODE_READ,
1219 int ret = hists__init();
1225 ret = perf_config(report__config, &report);
1229 argc = parse_options(argc, argv, options, report_usage, 0);
1232 * Special case: if there's an argument left then assume that
1233 * it's a symbol filter:
1236 usage_with_options(report_usage, options);
1238 report.symbol_filter_str = argv[0];
1241 if (report.mmaps_mode)
1242 report.tasks_mode = true;
1245 perf_quiet_option();
1247 if (symbol_conf.vmlinux_name &&
1248 access(symbol_conf.vmlinux_name, R_OK)) {
1249 pr_err("Invalid file: %s\n", symbol_conf.vmlinux_name);
1252 if (symbol_conf.kallsyms_name &&
1253 access(symbol_conf.kallsyms_name, R_OK)) {
1254 pr_err("Invalid file: %s\n", symbol_conf.kallsyms_name);
1258 if (report.inverted_callchain)
1259 callchain_param.order = ORDER_CALLER;
1260 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1261 callchain_param.order = ORDER_CALLER;
1263 if (itrace_synth_opts.callchain &&
1264 (int)itrace_synth_opts.callchain_sz > report.max_stack)
1265 report.max_stack = itrace_synth_opts.callchain_sz;
1267 if (!input_name || !strlen(input_name)) {
1268 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1271 input_name = "perf.data";
1274 data.path = input_name;
1275 data.force = symbol_conf.force;
1278 session = perf_session__new(&data, false, &report.tool);
1279 if (IS_ERR(session))
1280 return PTR_ERR(session);
1282 ret = evswitch__init(&report.evswitch, session->evlist, stderr);
1286 if (zstd_init(&(session->zstd_data), 0) < 0)
1287 pr_warning("Decompression initialization failed. Reported data may be incomplete.\n");
1289 if (report.queue_size) {
1290 ordered_events__set_alloc_size(&session->ordered_events,
1294 session->itrace_synth_opts = &itrace_synth_opts;
1296 report.session = session;
1298 has_br_stack = perf_header__has_feat(&session->header,
1299 HEADER_BRANCH_STACK);
1300 if (perf_evlist__combined_sample_type(session->evlist) & PERF_SAMPLE_STACK_USER)
1301 has_br_stack = false;
1303 setup_forced_leader(&report, session->evlist);
1305 if (itrace_synth_opts.last_branch)
1306 has_br_stack = true;
1308 if (has_br_stack && branch_call_mode)
1309 symbol_conf.show_branchflag_count = true;
1311 memset(&report.brtype_stat, 0, sizeof(struct branch_type_stat));
1314 * Branch mode is a tristate:
1315 * -1 means default, so decide based on the file having branch data.
1316 * 0/1 means the user chose a mode.
1318 if (((branch_mode == -1 && has_br_stack) || branch_mode == 1) &&
1319 !branch_call_mode) {
1320 sort__mode = SORT_MODE__BRANCH;
1321 symbol_conf.cumulate_callchain = false;
1323 if (branch_call_mode) {
1324 callchain_param.key = CCKEY_ADDRESS;
1325 callchain_param.branch_callstack = 1;
1326 symbol_conf.use_callchain = true;
1327 callchain_register_param(&callchain_param);
1328 if (sort_order == NULL)
1329 sort_order = "srcline,symbol,dso";
1332 if (report.mem_mode) {
1333 if (sort__mode == SORT_MODE__BRANCH) {
1334 pr_err("branch and mem mode incompatible\n");
1337 sort__mode = SORT_MODE__MEMORY;
1338 symbol_conf.cumulate_callchain = false;
1341 if (symbol_conf.report_hierarchy) {
1342 /* disable incompatible options */
1343 symbol_conf.cumulate_callchain = false;
1346 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1347 parse_options_usage(report_usage, options, "F", 1);
1348 parse_options_usage(NULL, options, "hierarchy", 0);
1352 perf_hpp_list.need_collapse = true;
1355 if (report.use_stdio)
1357 else if (report.use_tui)
1359 else if (report.use_gtk)
1362 /* Force tty output for header output and per-thread stat. */
1363 if (report.header || report.header_only || report.show_threads)
1365 if (report.header || report.header_only)
1366 report.tool.show_feat_hdr = SHOW_FEAT_HEADER;
1367 if (report.show_full_info)
1368 report.tool.show_feat_hdr = SHOW_FEAT_HEADER_FULL_INFO;
1369 if (report.stats_mode || report.tasks_mode)
1371 if (report.stats_mode && report.tasks_mode) {
1372 pr_err("Error: --tasks and --mmaps can't be used together with --stats\n");
1376 if (strcmp(input_name, "-") != 0)
1377 setup_browser(true);
1381 if (sort_order && strstr(sort_order, "ipc")) {
1382 parse_options_usage(report_usage, options, "s", 1);
1386 if (sort_order && strstr(sort_order, "symbol")) {
1387 if (sort__mode == SORT_MODE__BRANCH) {
1388 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1389 sort_order, "ipc_lbr");
1390 report.symbol_ipc = true;
1392 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1393 sort_order, "ipc_null");
1396 sort_order = sort_tmp;
1399 if (setup_sorting(session->evlist) < 0) {
1401 parse_options_usage(report_usage, options, "s", 1);
1403 parse_options_usage(sort_order ? NULL : report_usage,
1408 if ((report.header || report.header_only) && !quiet) {
1409 perf_session__fprintf_info(session, stdout,
1410 report.show_full_info);
1411 if (report.header_only) {
1415 } else if (use_browser == 0 && !quiet &&
1416 !report.stats_mode && !report.tasks_mode) {
1417 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
1422 * Only in the TUI browser we are doing integrated annotation,
1423 * so don't allocate extra space that won't be used in the stdio
1426 if (ui__has_annotation() || report.symbol_ipc) {
1427 ret = symbol__annotation_init();
1431 * For searching by name on the "Browse map details".
1432 * providing it only in verbose mode not to bloat too
1433 * much struct symbol.
1437 * XXX: Need to provide a less kludgy way to ask for
1438 * more space per symbol, the u32 is for the index on
1440 * See symbol__browser_index.
1442 symbol_conf.priv_size += sizeof(u32);
1443 symbol_conf.sort_by_name = true;
1445 annotation_config__init();
1448 if (symbol__init(&session->header.env) < 0)
1451 if (report.time_str) {
1452 ret = perf_time__parse_for_ranges(report.time_str, session,
1453 &report.ptime_range,
1459 itrace_synth_opts__set_time_range(&itrace_synth_opts,
1464 if (session->tevent.pevent &&
1465 tep_set_function_resolver(session->tevent.pevent,
1466 machine__resolve_kernel_addr,
1467 &session->machines.host) < 0) {
1468 pr_err("%s: failed to set libtraceevent function resolver\n",
1473 sort__setup_elide(stdout);
1475 ret = __cmd_report(&report);
1476 if (ret == K_SWITCH_INPUT_DATA) {
1477 perf_session__delete(session);
1483 if (report.ptime_range) {
1484 itrace_synth_opts__clear_time_range(&itrace_synth_opts);
1485 zfree(&report.ptime_range);
1487 zstd_fini(&(session->zstd_data));
1488 perf_session__delete(session);