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/util.h" // perf_tip()
52 #include "ui/progress.h"
53 #include "util/block-info.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;
88 struct perf_read_values show_threads_values;
89 struct annotation_options annotation_opts;
90 const char *pretty_printing_style;
92 const char *symbol_filter_str;
94 struct perf_time_interval *ptime_range;
102 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
103 struct branch_type_stat brtype_stat;
105 bool total_cycles_mode;
106 struct block_report *block_reports;
107 int nr_block_reports;
110 static int report__config(const char *var, const char *value, void *cb)
112 struct report *rep = cb;
114 if (!strcmp(var, "report.group")) {
115 symbol_conf.event_group = perf_config_bool(var, value);
118 if (!strcmp(var, "report.percent-limit")) {
119 double pcnt = strtof(value, NULL);
121 rep->min_percent = pcnt;
122 callchain_param.min_percent = pcnt;
125 if (!strcmp(var, "report.children")) {
126 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
129 if (!strcmp(var, "report.queue-size"))
130 return perf_config_u64(&rep->queue_size, var, value);
132 if (!strcmp(var, "report.sort_order")) {
133 default_sort_order = strdup(value);
140 static int hist_iter__report_callback(struct hist_entry_iter *iter,
141 struct addr_location *al, bool single,
145 struct report *rep = arg;
146 struct hist_entry *he = iter->he;
147 struct evsel *evsel = iter->evsel;
148 struct perf_sample *sample = iter->sample;
150 struct branch_info *bi;
152 if (!ui__has_annotation() && !rep->symbol_ipc)
155 if (sort__mode == SORT_MODE__BRANCH) {
156 bi = he->branch_info;
157 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
161 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
163 } else if (rep->mem_mode) {
165 err = addr_map_symbol__inc_samples(&mi->daddr, sample, evsel);
169 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
171 } else if (symbol_conf.cumulate_callchain) {
173 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
175 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
182 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
183 struct addr_location *al __maybe_unused,
184 bool single __maybe_unused,
187 struct hist_entry *he = iter->he;
188 struct report *rep = arg;
189 struct branch_info *bi = he->branch_info;
190 struct perf_sample *sample = iter->sample;
191 struct evsel *evsel = iter->evsel;
194 branch_type_count(&rep->brtype_stat, &bi->flags,
195 bi->from.addr, bi->to.addr);
197 if (!ui__has_annotation() && !rep->symbol_ipc)
200 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
204 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
210 static void setup_forced_leader(struct report *report,
211 struct evlist *evlist)
213 if (report->group_set)
214 perf_evlist__force_leader(evlist);
217 static int process_feature_event(struct perf_session *session,
218 union perf_event *event)
220 struct report *rep = container_of(session->tool, struct report, tool);
222 if (event->feat.feat_id < HEADER_LAST_FEATURE)
223 return perf_event__process_feature(session, event);
225 if (event->feat.feat_id != HEADER_LAST_FEATURE) {
226 pr_err("failed: wrong feature ID: %" PRI_lu64 "\n",
227 event->feat.feat_id);
232 * (feat_id = HEADER_LAST_FEATURE) is the end marker which
233 * means all features are received, now we can force the
236 setup_forced_leader(rep, session->evlist);
240 static int process_sample_event(struct perf_tool *tool,
241 union perf_event *event,
242 struct perf_sample *sample,
244 struct machine *machine)
246 struct report *rep = container_of(tool, struct report, tool);
247 struct addr_location al;
248 struct hist_entry_iter iter = {
251 .hide_unresolved = symbol_conf.hide_unresolved,
252 .add_entry_cb = hist_iter__report_callback,
256 if (perf_time__ranges_skip_sample(rep->ptime_range, rep->range_num,
261 if (evswitch__discard(&rep->evswitch, evsel))
264 if (machine__resolve(machine, &al, sample) < 0) {
265 pr_debug("problem processing %d event, skipping it.\n",
271 al.thread->lbr_stitch_enable = true;
273 if (symbol_conf.hide_unresolved && al.sym == NULL)
276 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
279 if (sort__mode == SORT_MODE__BRANCH) {
281 * A non-synthesized event might not have a branch stack if
282 * branch stacks have been synthesized (using itrace options).
284 if (!sample->branch_stack)
287 iter.add_entry_cb = hist_iter__branch_callback;
288 iter.ops = &hist_iter_branch;
289 } else if (rep->mem_mode) {
290 iter.ops = &hist_iter_mem;
291 } else if (symbol_conf.cumulate_callchain) {
292 iter.ops = &hist_iter_cumulative;
294 iter.ops = &hist_iter_normal;
298 al.map->dso->hit = 1;
300 if (ui__has_annotation() || rep->symbol_ipc || rep->total_cycles_mode) {
301 hist__account_cycles(sample->branch_stack, &al, sample,
302 rep->nonany_branch_mode,
306 ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
308 pr_debug("problem adding hist entry, skipping event\n");
310 addr_location__put(&al);
314 static int process_read_event(struct perf_tool *tool,
315 union perf_event *event,
316 struct perf_sample *sample __maybe_unused,
318 struct machine *machine __maybe_unused)
320 struct report *rep = container_of(tool, struct report, tool);
322 if (rep->show_threads) {
323 const char *name = evsel__name(evsel);
324 int err = perf_read_values_add_value(&rep->show_threads_values,
325 event->read.pid, event->read.tid,
337 /* For pipe mode, sample_type is not currently set */
338 static int report__setup_sample_type(struct report *rep)
340 struct perf_session *session = rep->session;
341 u64 sample_type = evlist__combined_sample_type(session->evlist);
342 bool is_pipe = perf_data__is_pipe(session->data);
345 if (session->itrace_synth_opts->callchain ||
346 session->itrace_synth_opts->add_callchain ||
348 perf_header__has_feat(&session->header, HEADER_AUXTRACE) &&
349 !session->itrace_synth_opts->set))
350 sample_type |= PERF_SAMPLE_CALLCHAIN;
352 if (session->itrace_synth_opts->last_branch ||
353 session->itrace_synth_opts->add_last_branch)
354 sample_type |= PERF_SAMPLE_BRANCH_STACK;
356 if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
357 if (perf_hpp_list.parent) {
358 ui__error("Selected --sort parent, but no "
359 "callchain data. Did you call "
360 "'perf record' without -g?\n");
363 if (symbol_conf.use_callchain &&
364 !symbol_conf.show_branchflag_count) {
365 ui__error("Selected -g or --branch-history.\n"
366 "But no callchain or branch data.\n"
367 "Did you call 'perf record' without -g or -b?\n");
370 } else if (!callchain_param.enabled &&
371 callchain_param.mode != CHAIN_NONE &&
372 !symbol_conf.use_callchain) {
373 symbol_conf.use_callchain = true;
374 if (callchain_register_param(&callchain_param) < 0) {
375 ui__error("Can't register callchain params.\n");
380 if (symbol_conf.cumulate_callchain) {
381 /* Silently ignore if callchain is missing */
382 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
383 symbol_conf.cumulate_callchain = false;
384 perf_hpp__cancel_cumulate();
388 if (sort__mode == SORT_MODE__BRANCH) {
390 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
391 ui__error("Selected -b but no branch data. "
392 "Did you call perf record without -b?\n");
397 if (sort__mode == SORT_MODE__MEMORY) {
399 * FIXUP: prior to kernel 5.18, Arm SPE missed to set
400 * PERF_SAMPLE_DATA_SRC bit in sample type. For backward
401 * compatibility, set the bit if it's an old perf data file.
403 evlist__for_each_entry(session->evlist, evsel) {
404 if (strstr(evsel->name, "arm_spe") &&
405 !(sample_type & PERF_SAMPLE_DATA_SRC)) {
406 evsel->core.attr.sample_type |= PERF_SAMPLE_DATA_SRC;
407 sample_type |= PERF_SAMPLE_DATA_SRC;
411 if (!is_pipe && !(sample_type & PERF_SAMPLE_DATA_SRC)) {
412 ui__error("Selected --mem-mode but no mem data. "
413 "Did you call perf record without -d?\n");
418 callchain_param_setup(sample_type);
420 if (rep->stitch_lbr && (callchain_param.record_mode != CALLCHAIN_LBR)) {
421 ui__warning("Can't find LBR callchain. Switch off --stitch-lbr.\n"
422 "Please apply --call-graph lbr when recording.\n");
423 rep->stitch_lbr = false;
426 /* ??? handle more cases than just ANY? */
427 if (!(evlist__combined_branch_type(session->evlist) & PERF_SAMPLE_BRANCH_ANY))
428 rep->nonany_branch_mode = true;
430 #if !defined(HAVE_LIBUNWIND_SUPPORT) && !defined(HAVE_DWARF_SUPPORT)
431 if (dwarf_callchain_users) {
432 ui__warning("Please install libunwind or libdw "
433 "development packages during the perf build.\n");
440 static void sig_handler(int sig __maybe_unused)
445 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
446 const char *evname, FILE *fp)
450 unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
451 u64 nr_events = hists->stats.total_period;
452 struct evsel *evsel = hists_to_evsel(hists);
454 size_t size = sizeof(buf);
455 int socked_id = hists->socket_filter;
460 if (symbol_conf.filter_relative) {
461 nr_samples = hists->stats.nr_non_filtered_samples;
462 nr_events = hists->stats.total_non_filtered_period;
465 if (evsel__is_group_event(evsel)) {
468 evsel__group_desc(evsel, buf, size);
471 for_each_group_member(pos, evsel) {
472 const struct hists *pos_hists = evsel__hists(pos);
474 if (symbol_conf.filter_relative) {
475 nr_samples += pos_hists->stats.nr_non_filtered_samples;
476 nr_events += pos_hists->stats.total_non_filtered_period;
478 nr_samples += pos_hists->stats.nr_events[PERF_RECORD_SAMPLE];
479 nr_events += pos_hists->stats.total_period;
484 nr_samples = convert_unit(nr_samples, &unit);
485 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
486 if (evname != NULL) {
487 ret += fprintf(fp, " of event%s '%s'",
488 evsel->core.nr_members > 1 ? "s" : "", evname);
492 ret += fprintf(fp, " (time slices: %s)", rep->time_str);
494 if (symbol_conf.show_ref_callgraph && evname && strstr(evname, "call-graph=no")) {
495 ret += fprintf(fp, ", show reference callgraph");
499 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
500 ret += fprintf(fp, "\n# Sort order : %s", sort_order ? : default_mem_sort_order);
502 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
505 ret += fprintf(fp, "\n# Processor Socket: %d", socked_id);
507 return ret + fprintf(fp, "\n#\n");
510 static int perf_evlist__tui_block_hists_browse(struct evlist *evlist,
516 evlist__for_each_entry(evlist, pos) {
517 ret = report__browse_block_hists(&rep->block_reports[i++].hist,
518 rep->min_percent, pos,
519 &rep->session->header.env,
520 &rep->annotation_opts);
528 static int perf_evlist__tty_browse_hists(struct evlist *evlist,
536 fprintf(stdout, "#\n# Total Lost Samples: %" PRIu64 "\n#\n",
537 evlist->stats.total_lost_samples);
540 evlist__for_each_entry(evlist, pos) {
541 struct hists *hists = evsel__hists(pos);
542 const char *evname = evsel__name(pos);
544 if (symbol_conf.event_group && !evsel__is_group_leader(pos))
547 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
549 if (rep->total_cycles_mode) {
550 report__browse_block_hists(&rep->block_reports[i++].hist,
551 rep->min_percent, pos,
556 hists__fprintf(hists, !quiet, 0, 0, rep->min_percent, stdout,
557 !(symbol_conf.use_callchain ||
558 symbol_conf.show_branchflag_count));
559 fprintf(stdout, "\n\n");
563 fprintf(stdout, "#\n# (%s)\n#\n", help);
565 if (rep->show_threads) {
566 bool style = !strcmp(rep->pretty_printing_style, "raw");
567 perf_read_values_display(stdout, &rep->show_threads_values,
569 perf_read_values_destroy(&rep->show_threads_values);
572 if (sort__mode == SORT_MODE__BRANCH)
573 branch_type_stat_display(stdout, &rep->brtype_stat);
578 static void report__warn_kptr_restrict(const struct report *rep)
580 struct map *kernel_map = machine__kernel_map(&rep->session->machines.host);
581 struct kmap *kernel_kmap = kernel_map ? map__kmap(kernel_map) : NULL;
583 if (perf_evlist__exclude_kernel(rep->session->evlist))
586 if (kernel_map == NULL ||
587 (kernel_map->dso->hit &&
588 (kernel_kmap->ref_reloc_sym == NULL ||
589 kernel_kmap->ref_reloc_sym->addr == 0))) {
591 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
592 "can't be resolved.";
594 if (kernel_map && map__has_symbols(kernel_map)) {
595 desc = "If some relocation was applied (e.g. "
596 "kexec) symbols may be misresolved.";
600 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
601 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
602 "Samples in kernel modules can't be resolved as well.\n\n",
607 static int report__gtk_browse_hists(struct report *rep, const char *help)
609 int (*hist_browser)(struct evlist *evlist, const char *help,
610 struct hist_browser_timer *timer, float min_pcnt);
612 hist_browser = dlsym(perf_gtk_handle, "perf_evlist__gtk_browse_hists");
614 if (hist_browser == NULL) {
615 ui__error("GTK browser not found!\n");
619 return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
622 static int report__browse_hists(struct report *rep)
625 struct perf_session *session = rep->session;
626 struct evlist *evlist = session->evlist;
627 char *help = NULL, *path = NULL;
629 path = system_path(TIPDIR);
630 if (perf_tip(&help, path) || help == NULL) {
631 /* fallback for people who don't install perf ;-) */
633 path = system_path(DOCDIR);
634 if (perf_tip(&help, path) || help == NULL)
635 help = strdup("Cannot load tips.txt file, please install perf!");
639 switch (use_browser) {
641 if (rep->total_cycles_mode) {
642 ret = perf_evlist__tui_block_hists_browse(evlist, rep);
646 ret = perf_evlist__tui_browse_hists(evlist, help, NULL,
648 &session->header.env,
649 true, &rep->annotation_opts);
651 * Usually "ret" is the last pressed key, and we only
652 * care if the key notifies us to switch data file.
654 if (ret != K_SWITCH_INPUT_DATA && ret != K_RELOAD)
658 ret = report__gtk_browse_hists(rep, help);
661 ret = perf_evlist__tty_browse_hists(evlist, rep, help);
668 static int report__collapse_hists(struct report *rep)
670 struct ui_progress prog;
674 ui_progress__init(&prog, rep->nr_entries, "Merging related events...");
676 evlist__for_each_entry(rep->session->evlist, pos) {
677 struct hists *hists = evsel__hists(pos);
680 hists->symbol_filter_str = rep->symbol_filter_str;
682 hists->socket_filter = rep->socket_filter;
684 ret = hists__collapse_resort(hists, &prog);
688 /* Non-group events are considered as leader */
689 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
690 struct hists *leader_hists = evsel__hists(pos->leader);
692 hists__match(leader_hists, hists);
693 hists__link(leader_hists, hists);
697 ui_progress__finish();
701 static int hists__resort_cb(struct hist_entry *he, void *arg)
703 struct report *rep = arg;
704 struct symbol *sym = he->ms.sym;
706 if (rep->symbol_ipc && sym && !sym->annotate2) {
707 struct evsel *evsel = hists_to_evsel(he->hists);
709 symbol__annotate2(&he->ms, evsel,
710 &annotation__default_options, NULL);
716 static void report__output_resort(struct report *rep)
718 struct ui_progress prog;
721 ui_progress__init(&prog, rep->nr_entries, "Sorting events for output...");
723 evlist__for_each_entry(rep->session->evlist, pos) {
724 evsel__output_resort_cb(pos, &prog, hists__resort_cb, rep);
727 ui_progress__finish();
730 static void stats_setup(struct report *rep)
732 memset(&rep->tool, 0, sizeof(rep->tool));
733 rep->tool.no_warn = true;
736 static int stats_print(struct report *rep)
738 struct perf_session *session = rep->session;
740 perf_session__fprintf_nr_events(session, stdout);
744 static void tasks_setup(struct report *rep)
746 memset(&rep->tool, 0, sizeof(rep->tool));
747 rep->tool.ordered_events = true;
748 if (rep->mmaps_mode) {
749 rep->tool.mmap = perf_event__process_mmap;
750 rep->tool.mmap2 = perf_event__process_mmap2;
752 rep->tool.comm = perf_event__process_comm;
753 rep->tool.exit = perf_event__process_exit;
754 rep->tool.fork = perf_event__process_fork;
755 rep->tool.no_warn = true;
759 struct thread *thread;
760 struct list_head list;
761 struct list_head children;
764 static struct task *tasks_list(struct task *task, struct machine *machine)
766 struct thread *parent_thread, *thread = task->thread;
767 struct task *parent_task;
769 /* Already listed. */
770 if (!list_empty(&task->list))
773 /* Last one in the chain. */
774 if (thread->ppid == -1)
777 parent_thread = machine__find_thread(machine, -1, thread->ppid);
779 return ERR_PTR(-ENOENT);
781 parent_task = thread__priv(parent_thread);
782 list_add_tail(&task->list, &parent_task->children);
783 return tasks_list(parent_task, machine);
786 static size_t maps__fprintf_task(struct maps *maps, int indent, FILE *fp)
791 maps__for_each_entry(maps, map) {
792 printed += fprintf(fp, "%*s %" PRIx64 "-%" PRIx64 " %c%c%c%c %08" PRIx64 " %" PRIu64 " %s\n",
793 indent, "", map->start, map->end,
794 map->prot & PROT_READ ? 'r' : '-',
795 map->prot & PROT_WRITE ? 'w' : '-',
796 map->prot & PROT_EXEC ? 'x' : '-',
797 map->flags & MAP_SHARED ? 's' : 'p',
799 map->dso->id.ino, map->dso->name);
805 static void task__print_level(struct task *task, FILE *fp, int level)
807 struct thread *thread = task->thread;
809 int comm_indent = fprintf(fp, " %8d %8d %8d |%*s",
810 thread->pid_, thread->tid, thread->ppid,
813 fprintf(fp, "%s\n", thread__comm_str(thread));
815 maps__fprintf_task(thread->maps, comm_indent, fp);
817 if (!list_empty(&task->children)) {
818 list_for_each_entry(child, &task->children, list)
819 task__print_level(child, fp, level + 1);
823 static int tasks_print(struct report *rep, FILE *fp)
825 struct perf_session *session = rep->session;
826 struct machine *machine = &session->machines.host;
827 struct task *tasks, *task;
828 unsigned int nr = 0, itask = 0, i;
833 * No locking needed while accessing machine->threads,
834 * because --tasks is single threaded command.
837 /* Count all the threads. */
838 for (i = 0; i < THREADS__TABLE_SIZE; i++)
839 nr += machine->threads[i].nr;
841 tasks = malloc(sizeof(*tasks) * nr);
845 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
846 struct threads *threads = &machine->threads[i];
848 for (nd = rb_first_cached(&threads->entries); nd;
850 task = tasks + itask++;
852 task->thread = rb_entry(nd, struct thread, rb_node);
853 INIT_LIST_HEAD(&task->children);
854 INIT_LIST_HEAD(&task->list);
855 thread__set_priv(task->thread, task);
860 * Iterate every task down to the unprocessed parent
861 * and link all in task children list. Task with no
862 * parent is added into 'list'.
864 for (itask = 0; itask < nr; itask++) {
865 task = tasks + itask;
867 if (!list_empty(&task->list))
870 task = tasks_list(task, machine);
872 pr_err("Error: failed to process tasks\n");
874 return PTR_ERR(task);
878 list_add_tail(&task->list, &list);
881 fprintf(fp, "# %8s %8s %8s %s\n", "pid", "tid", "ppid", "comm");
883 list_for_each_entry(task, &list, list)
884 task__print_level(task, fp, 0);
890 static int __cmd_report(struct report *rep)
893 struct perf_session *session = rep->session;
895 struct perf_data *data = session->data;
897 signal(SIGINT, sig_handler);
900 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
903 ui__error("failed to set cpu bitmap\n");
906 session->itrace_synth_opts->cpu_bitmap = rep->cpu_bitmap;
909 if (rep->show_threads) {
910 ret = perf_read_values_init(&rep->show_threads_values);
915 ret = report__setup_sample_type(rep);
917 /* report__setup_sample_type() already showed error message */
927 ret = perf_session__process_events(session);
929 ui__error("failed to process sample\n");
934 return stats_print(rep);
937 return tasks_print(rep, stdout);
939 report__warn_kptr_restrict(rep);
941 evlist__for_each_entry(session->evlist, pos)
942 rep->nr_entries += evsel__hists(pos)->nr_entries;
944 if (use_browser == 0) {
946 perf_session__fprintf(session, stdout);
949 perf_session__fprintf_dsos(session, stdout);
952 perf_session__fprintf_nr_events(session, stdout);
953 perf_evlist__fprintf_nr_events(session->evlist, stdout);
958 ret = report__collapse_hists(rep);
960 ui__error("failed to process hist entry\n");
968 * recalculate number of entries after collapsing since it
969 * might be changed during the collapse phase.
972 evlist__for_each_entry(session->evlist, pos)
973 rep->nr_entries += evsel__hists(pos)->nr_entries;
975 if (rep->nr_entries == 0) {
976 ui__error("The %s data has no samples!\n", data->path);
980 report__output_resort(rep);
982 if (rep->total_cycles_mode) {
983 int block_hpps[6] = {
984 PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_PCT,
985 PERF_HPP_REPORT__BLOCK_LBR_CYCLES,
986 PERF_HPP_REPORT__BLOCK_CYCLES_PCT,
987 PERF_HPP_REPORT__BLOCK_AVG_CYCLES,
988 PERF_HPP_REPORT__BLOCK_RANGE,
989 PERF_HPP_REPORT__BLOCK_DSO,
992 rep->block_reports = block_info__create_report(session->evlist,
995 &rep->nr_block_reports);
996 if (!rep->block_reports)
1000 return report__browse_hists(rep);
1004 report_parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1006 struct callchain_param *callchain = opt->value;
1008 callchain->enabled = !unset;
1013 symbol_conf.use_callchain = false;
1014 callchain->mode = CHAIN_NONE;
1018 return parse_callchain_report_opt(arg);
1022 parse_time_quantum(const struct option *opt, const char *arg,
1023 int unset __maybe_unused)
1025 unsigned long *time_q = opt->value;
1028 *time_q = strtoul(arg, &end, 0);
1032 pr_err("time quantum cannot be 0");
1035 end = skip_spaces(end);
1038 if (!strcmp(end, "s")) {
1039 *time_q *= NSEC_PER_SEC;
1042 if (!strcmp(end, "ms")) {
1043 *time_q *= NSEC_PER_MSEC;
1046 if (!strcmp(end, "us")) {
1047 *time_q *= NSEC_PER_USEC;
1050 if (!strcmp(end, "ns"))
1053 pr_err("Cannot parse time quantum `%s'\n", arg);
1058 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
1059 const char *arg, int unset __maybe_unused)
1062 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
1065 regerror(err, &ignore_callees_regex, buf, sizeof(buf));
1066 pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
1069 have_ignore_callees = 1;
1076 parse_branch_mode(const struct option *opt,
1077 const char *str __maybe_unused, int unset)
1079 int *branch_mode = opt->value;
1081 *branch_mode = !unset;
1086 parse_percent_limit(const struct option *opt, const char *str,
1087 int unset __maybe_unused)
1089 struct report *rep = opt->value;
1090 double pcnt = strtof(str, NULL);
1092 rep->min_percent = pcnt;
1093 callchain_param.min_percent = pcnt;
1097 static int process_attr(struct perf_tool *tool __maybe_unused,
1098 union perf_event *event,
1099 struct evlist **pevlist)
1104 err = perf_event__process_attr(tool, event, pevlist);
1109 * Check if we need to enable callchains based
1110 * on events sample_type.
1112 sample_type = evlist__combined_sample_type(*pevlist);
1113 callchain_param_setup(sample_type);
1117 int cmd_report(int argc, const char **argv)
1119 struct perf_session *session;
1120 struct itrace_synth_opts itrace_synth_opts = { .set = 0, };
1122 bool has_br_stack = false;
1123 int branch_mode = -1;
1125 bool branch_call_mode = false;
1126 #define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent"
1127 static const char report_callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
1128 CALLCHAIN_REPORT_HELP
1129 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
1130 char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
1131 const char * const report_usage[] = {
1132 "perf report [<options>]",
1135 struct report report = {
1137 .sample = process_sample_event,
1138 .mmap = perf_event__process_mmap,
1139 .mmap2 = perf_event__process_mmap2,
1140 .comm = perf_event__process_comm,
1141 .namespaces = perf_event__process_namespaces,
1142 .cgroup = perf_event__process_cgroup,
1143 .exit = perf_event__process_exit,
1144 .fork = perf_event__process_fork,
1145 .lost = perf_event__process_lost,
1146 .read = process_read_event,
1147 .attr = process_attr,
1148 .tracing_data = perf_event__process_tracing_data,
1149 .build_id = perf_event__process_build_id,
1150 .id_index = perf_event__process_id_index,
1151 .auxtrace_info = perf_event__process_auxtrace_info,
1152 .auxtrace = perf_event__process_auxtrace,
1153 .event_update = perf_event__process_event_update,
1154 .feature = process_feature_event,
1155 .ordered_events = true,
1156 .ordering_requires_timestamps = true,
1158 .max_stack = PERF_MAX_STACK_DEPTH,
1159 .pretty_printing_style = "normal",
1160 .socket_filter = -1,
1161 .annotation_opts = annotation__default_options,
1163 char *sort_order_help = sort_help("sort by key(s):");
1164 char *field_order_help = sort_help("output field(s): overhead period sample ");
1165 const struct option options[] = {
1166 OPT_STRING('i', "input", &input_name, "file",
1168 OPT_INCR('v', "verbose", &verbose,
1169 "be more verbose (show symbol address, etc)"),
1170 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
1171 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1172 "dump raw trace in ASCII"),
1173 OPT_BOOLEAN(0, "stats", &report.stats_mode, "Display event stats"),
1174 OPT_BOOLEAN(0, "tasks", &report.tasks_mode, "Display recorded tasks"),
1175 OPT_BOOLEAN(0, "mmaps", &report.mmaps_mode, "Display recorded tasks memory maps"),
1176 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1177 "file", "vmlinux pathname"),
1178 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1179 "don't load vmlinux even if found"),
1180 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1181 "file", "kallsyms pathname"),
1182 OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
1183 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1184 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1185 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1186 "Show a column with the number of samples"),
1187 OPT_BOOLEAN('T', "threads", &report.show_threads,
1188 "Show per-thread event counters"),
1189 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
1190 "pretty printing style key: normal raw"),
1191 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
1192 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
1193 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
1194 "Use the stdio interface"),
1195 OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
1196 OPT_BOOLEAN(0, "header-only", &report.header_only,
1197 "Show only data header."),
1198 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1200 OPT_STRING('F', "fields", &field_order, "key[,keys...]",
1202 OPT_BOOLEAN(0, "show-cpu-utilization", &symbol_conf.show_cpu_utilization,
1203 "Show sample percentage for different cpu modes"),
1204 OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
1205 "Show sample percentage for different cpu modes", PARSE_OPT_HIDDEN),
1206 OPT_STRING('p', "parent", &parent_pattern, "regex",
1207 "regex filter to identify parent, see: '--sort parent'"),
1208 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
1209 "Only display entries with parent-match"),
1210 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
1211 "print_type,threshold[,print_limit],order,sort_key[,branch],value",
1212 report_callchain_help, &report_parse_callchain_opt,
1213 callchain_default_opt),
1214 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1215 "Accumulate callchains of children and show total overhead as well. "
1216 "Enabled by default, use --no-children to disable."),
1217 OPT_INTEGER(0, "max-stack", &report.max_stack,
1218 "Set the maximum stack depth when parsing the callchain, "
1219 "anything beyond the specified depth will be ignored. "
1220 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1221 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
1222 "alias for inverted call graph"),
1223 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1224 "ignore callees of these functions in call graphs",
1225 report_parse_ignore_callees_opt),
1226 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1227 "only consider symbols in these dsos"),
1228 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1229 "only consider symbols in these comms"),
1230 OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1231 "only consider symbols in these pids"),
1232 OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1233 "only consider symbols in these tids"),
1234 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1235 "only consider these symbols"),
1236 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
1237 "only show symbols that (partially) match with this filter"),
1238 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1240 "don't try to adjust column width, use these fixed values"),
1241 OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1242 "separator for columns, no spaces will be added between "
1243 "columns '.' is reserved."),
1244 OPT_BOOLEAN('U', "hide-unresolved", &symbol_conf.hide_unresolved,
1245 "Only display entries resolved to a symbol"),
1246 OPT_CALLBACK(0, "symfs", NULL, "directory",
1247 "Look for files with symbols relative to this directory",
1248 symbol__config_symfs),
1249 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
1250 "list of cpus to profile"),
1251 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
1252 "Display extended information about perf.data file"),
1253 OPT_BOOLEAN(0, "source", &report.annotation_opts.annotate_src,
1254 "Interleave source code with assembly code (default)"),
1255 OPT_BOOLEAN(0, "asm-raw", &report.annotation_opts.show_asm_raw,
1256 "Display raw encoding of assembly instructions (default)"),
1257 OPT_STRING('M', "disassembler-style", &report.annotation_opts.disassembler_style, "disassembler style",
1258 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1259 OPT_STRING(0, "prefix", &report.annotation_opts.prefix, "prefix",
1260 "Add prefix to source file path names in programs (with --prefix-strip)"),
1261 OPT_STRING(0, "prefix-strip", &report.annotation_opts.prefix_strip, "N",
1262 "Strip first N entries of source file path name in programs (with --prefix)"),
1263 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1264 "Show a column with the sum of periods"),
1265 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group, &report.group_set,
1266 "Show event group information together"),
1267 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1268 "Sort the output by the event at the index n in group. "
1269 "If n is invalid, sort by the first event. "
1270 "WARNING: should be used on grouped events."),
1271 OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
1272 "use branch records for per branch histogram filling",
1274 OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1275 "add last branch records to call history"),
1276 OPT_STRING(0, "objdump", &report.annotation_opts.objdump_path, "path",
1277 "objdump binary to use for disassembly and annotations"),
1278 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
1279 "Disable symbol demangling"),
1280 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1281 "Enable kernel symbol demangling"),
1282 OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
1283 OPT_INTEGER(0, "samples", &symbol_conf.res_sample,
1284 "Number of samples to save per histogram entry for individual browsing"),
1285 OPT_CALLBACK(0, "percent-limit", &report, "percent",
1286 "Don't show entries under that percent", parse_percent_limit),
1287 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1288 "how to display percentage of filtered entries", parse_filter_percentage),
1289 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
1290 "Instruction Tracing options\n" ITRACE_HELP,
1291 itrace_parse_synth_opts),
1292 OPT_BOOLEAN(0, "full-source-path", &srcline_full_filename,
1293 "Show full source file name path for source lines"),
1294 OPT_BOOLEAN(0, "show-ref-call-graph", &symbol_conf.show_ref_callgraph,
1295 "Show callgraph from reference event"),
1296 OPT_BOOLEAN(0, "stitch-lbr", &report.stitch_lbr,
1297 "Enable LBR callgraph stitching approach"),
1298 OPT_INTEGER(0, "socket-filter", &report.socket_filter,
1299 "only show processor socket that match with this filter"),
1300 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1301 "Show raw trace event output (do not use print fmt or plugins)"),
1302 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1303 "Show entries in a hierarchy"),
1304 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
1305 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
1306 stdio__config_color, "always"),
1307 OPT_STRING(0, "time", &report.time_str, "str",
1308 "Time span of interest (start,stop)"),
1309 OPT_BOOLEAN(0, "inline", &symbol_conf.inline_name,
1310 "Show inline function"),
1311 OPT_CALLBACK(0, "percent-type", &report.annotation_opts, "local-period",
1312 "Set percent type local/global-period/hits",
1313 annotate_parse_percent_type),
1314 OPT_BOOLEAN(0, "ns", &symbol_conf.nanosecs, "Show times in nanosecs"),
1315 OPT_CALLBACK(0, "time-quantum", &symbol_conf.time_quantum, "time (ms|us|ns|s)",
1316 "Set time quantum for time sort key (default 100ms)",
1317 parse_time_quantum),
1318 OPTS_EVSWITCH(&report.evswitch),
1319 OPT_BOOLEAN(0, "total-cycles", &report.total_cycles_mode,
1320 "Sort all blocks by 'Sampled Cycles%'"),
1323 struct perf_data data = {
1324 .mode = PERF_DATA_MODE_READ,
1326 int ret = hists__init();
1332 ret = perf_config(report__config, &report);
1336 argc = parse_options(argc, argv, options, report_usage, 0);
1339 * Special case: if there's an argument left then assume that
1340 * it's a symbol filter:
1343 usage_with_options(report_usage, options);
1345 report.symbol_filter_str = argv[0];
1348 if (annotate_check_args(&report.annotation_opts) < 0) {
1353 if (report.mmaps_mode)
1354 report.tasks_mode = true;
1357 report.tool.ordered_events = false;
1360 perf_quiet_option();
1362 if (symbol_conf.vmlinux_name &&
1363 access(symbol_conf.vmlinux_name, R_OK)) {
1364 pr_err("Invalid file: %s\n", symbol_conf.vmlinux_name);
1368 if (symbol_conf.kallsyms_name &&
1369 access(symbol_conf.kallsyms_name, R_OK)) {
1370 pr_err("Invalid file: %s\n", symbol_conf.kallsyms_name);
1375 if (report.inverted_callchain)
1376 callchain_param.order = ORDER_CALLER;
1377 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1378 callchain_param.order = ORDER_CALLER;
1380 if ((itrace_synth_opts.callchain || itrace_synth_opts.add_callchain) &&
1381 (int)itrace_synth_opts.callchain_sz > report.max_stack)
1382 report.max_stack = itrace_synth_opts.callchain_sz;
1384 if (!input_name || !strlen(input_name)) {
1385 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1388 input_name = "perf.data";
1391 data.path = input_name;
1392 data.force = symbol_conf.force;
1395 session = perf_session__new(&data, false, &report.tool);
1396 if (IS_ERR(session)) {
1397 ret = PTR_ERR(session);
1401 ret = evswitch__init(&report.evswitch, session->evlist, stderr);
1405 if (zstd_init(&(session->zstd_data), 0) < 0)
1406 pr_warning("Decompression initialization failed. Reported data may be incomplete.\n");
1408 if (report.queue_size) {
1409 ordered_events__set_alloc_size(&session->ordered_events,
1413 session->itrace_synth_opts = &itrace_synth_opts;
1415 report.session = session;
1417 has_br_stack = perf_header__has_feat(&session->header,
1418 HEADER_BRANCH_STACK);
1419 if (evlist__combined_sample_type(session->evlist) & PERF_SAMPLE_STACK_USER)
1420 has_br_stack = false;
1422 setup_forced_leader(&report, session->evlist);
1424 if (symbol_conf.group_sort_idx && !session->evlist->nr_groups) {
1425 parse_options_usage(NULL, options, "group-sort-idx", 0);
1430 if (itrace_synth_opts.last_branch || itrace_synth_opts.add_last_branch)
1431 has_br_stack = true;
1433 if (has_br_stack && branch_call_mode)
1434 symbol_conf.show_branchflag_count = true;
1436 memset(&report.brtype_stat, 0, sizeof(struct branch_type_stat));
1439 * Branch mode is a tristate:
1440 * -1 means default, so decide based on the file having branch data.
1441 * 0/1 means the user chose a mode.
1443 if (((branch_mode == -1 && has_br_stack) || branch_mode == 1) &&
1444 !branch_call_mode) {
1445 sort__mode = SORT_MODE__BRANCH;
1446 symbol_conf.cumulate_callchain = false;
1448 if (branch_call_mode) {
1449 callchain_param.key = CCKEY_ADDRESS;
1450 callchain_param.branch_callstack = true;
1451 symbol_conf.use_callchain = true;
1452 callchain_register_param(&callchain_param);
1453 if (sort_order == NULL)
1454 sort_order = "srcline,symbol,dso";
1457 if (report.mem_mode) {
1458 if (sort__mode == SORT_MODE__BRANCH) {
1459 pr_err("branch and mem mode incompatible\n");
1462 sort__mode = SORT_MODE__MEMORY;
1463 symbol_conf.cumulate_callchain = false;
1466 if (symbol_conf.report_hierarchy) {
1467 /* disable incompatible options */
1468 symbol_conf.cumulate_callchain = false;
1471 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1472 parse_options_usage(report_usage, options, "F", 1);
1473 parse_options_usage(NULL, options, "hierarchy", 0);
1477 perf_hpp_list.need_collapse = true;
1480 if (report.use_stdio)
1482 else if (report.use_tui)
1484 else if (report.use_gtk)
1487 /* Force tty output for header output and per-thread stat. */
1488 if (report.header || report.header_only || report.show_threads)
1490 if (report.header || report.header_only)
1491 report.tool.show_feat_hdr = SHOW_FEAT_HEADER;
1492 if (report.show_full_info)
1493 report.tool.show_feat_hdr = SHOW_FEAT_HEADER_FULL_INFO;
1494 if (report.stats_mode || report.tasks_mode)
1496 if (report.stats_mode && report.tasks_mode) {
1497 pr_err("Error: --tasks and --mmaps can't be used together with --stats\n");
1501 if (report.total_cycles_mode) {
1502 if (sort__mode != SORT_MODE__BRANCH)
1503 report.total_cycles_mode = false;
1508 if (strcmp(input_name, "-") != 0)
1509 setup_browser(true);
1513 if (sort_order && strstr(sort_order, "ipc")) {
1514 parse_options_usage(report_usage, options, "s", 1);
1518 if (sort_order && strstr(sort_order, "symbol")) {
1519 if (sort__mode == SORT_MODE__BRANCH) {
1520 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1521 sort_order, "ipc_lbr");
1522 report.symbol_ipc = true;
1524 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1525 sort_order, "ipc_null");
1528 sort_order = sort_tmp;
1531 if ((last_key != K_SWITCH_INPUT_DATA && last_key != K_RELOAD) &&
1532 (setup_sorting(session->evlist) < 0)) {
1534 parse_options_usage(report_usage, options, "s", 1);
1536 parse_options_usage(sort_order ? NULL : report_usage,
1541 if ((report.header || report.header_only) && !quiet) {
1542 perf_session__fprintf_info(session, stdout,
1543 report.show_full_info);
1544 if (report.header_only) {
1548 } else if (use_browser == 0 && !quiet &&
1549 !report.stats_mode && !report.tasks_mode) {
1550 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
1555 * Only in the TUI browser we are doing integrated annotation,
1556 * so don't allocate extra space that won't be used in the stdio
1559 if (ui__has_annotation() || report.symbol_ipc ||
1560 report.total_cycles_mode) {
1561 ret = symbol__annotation_init();
1565 * For searching by name on the "Browse map details".
1566 * providing it only in verbose mode not to bloat too
1567 * much struct symbol.
1571 * XXX: Need to provide a less kludgy way to ask for
1572 * more space per symbol, the u32 is for the index on
1574 * See symbol__browser_index.
1576 symbol_conf.priv_size += sizeof(u32);
1577 symbol_conf.sort_by_name = true;
1579 annotation_config__init(&report.annotation_opts);
1582 if (symbol__init(&session->header.env) < 0)
1585 if (report.time_str) {
1586 ret = perf_time__parse_for_ranges(report.time_str, session,
1587 &report.ptime_range,
1593 itrace_synth_opts__set_time_range(&itrace_synth_opts,
1598 if (session->tevent.pevent &&
1599 tep_set_function_resolver(session->tevent.pevent,
1600 machine__resolve_kernel_addr,
1601 &session->machines.host) < 0) {
1602 pr_err("%s: failed to set libtraceevent function resolver\n",
1607 sort__setup_elide(stdout);
1609 ret = __cmd_report(&report);
1610 if (ret == K_SWITCH_INPUT_DATA || ret == K_RELOAD) {
1611 perf_session__delete(session);
1612 last_key = K_SWITCH_INPUT_DATA;
1618 if (report.ptime_range) {
1619 itrace_synth_opts__clear_time_range(&itrace_synth_opts);
1620 zfree(&report.ptime_range);
1623 if (report.block_reports) {
1624 block_info__free_report(report.block_reports,
1625 report.nr_block_reports);
1626 report.block_reports = NULL;
1629 zstd_fini(&(session->zstd_data));
1630 perf_session__delete(session);
1632 free(sort_order_help);
1633 free(field_order_help);