1 // SPDX-License-Identifier: GPL-2.0
5 * Builtin annotate 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/color.h"
12 #include <linux/list.h>
13 #include "util/cache.h"
14 #include <linux/rbtree.h>
15 #include <linux/zalloc.h>
16 #include "util/symbol.h"
18 #include "util/debug.h"
20 #include "util/evlist.h"
21 #include "util/evsel.h"
22 #include "util/annotate.h"
23 #include "util/annotate-data.h"
24 #include "util/event.h"
25 #include <subcmd/parse-options.h>
26 #include "util/parse-events.h"
27 #include "util/sort.h"
28 #include "util/hist.h"
30 #include "util/machine.h"
32 #include "util/session.h"
33 #include "util/tool.h"
34 #include "util/data.h"
35 #include "arch/common.h"
36 #include "util/block-range.h"
37 #include "util/map_symbol.h"
38 #include "util/branch.h"
39 #include "util/util.h"
40 #include "ui/progress.h"
44 #include <linux/bitmap.h>
45 #include <linux/err.h>
48 struct perf_annotate {
49 struct perf_tool tool;
50 struct perf_session *session;
51 #ifdef HAVE_SLANG_SUPPORT
54 bool use_stdio, use_stdio2;
55 #ifdef HAVE_GTK2_SUPPORT
65 const char *sym_hist_filter;
67 const char *target_data_type;
68 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
72 * Given one basic block:
82 * where the horizontal are the branches and the vertical is the executed
83 * block of instructions.
85 * We count, for each 'instruction', the number of blocks that covered it as
86 * well as count the ratio each branch is taken.
88 * We can do this without knowing the actual instruction stream by keeping
89 * track of the address ranges. We break down ranges such that there is no
90 * overlap and iterate from the start until the end.
92 * @acme: once we parse the objdump output _before_ processing the samples,
93 * we can easily fold the branch.cycles IPC bits in.
95 static void process_basic_block(struct addr_map_symbol *start,
96 struct addr_map_symbol *end,
97 struct branch_flags *flags)
99 struct symbol *sym = start->ms.sym;
100 struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
101 struct block_range_iter iter;
102 struct block_range *entry;
103 struct annotated_branch *branch;
106 * Sanity; NULL isn't executable and the CPU cannot execute backwards
108 if (!start->addr || start->addr > end->addr)
111 iter = block_range__create(start->addr, end->addr);
112 if (!block_range_iter__valid(&iter))
115 branch = annotation__get_branch(notes);
118 * First block in range is a branch target.
120 entry = block_range_iter(&iter);
121 assert(entry->is_target);
125 entry = block_range_iter(&iter);
131 branch->max_coverage = max(branch->max_coverage, entry->coverage);
133 } while (block_range_iter__next(&iter));
136 * Last block in rage is a branch.
138 entry = block_range_iter(&iter);
139 assert(entry->is_branch);
141 if (flags->predicted)
145 static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
146 struct perf_sample *sample)
148 struct addr_map_symbol *prev = NULL;
149 struct branch_info *bi;
155 bi = sample__resolve_bstack(sample, al);
159 for (i = bs->nr - 1; i >= 0; i--) {
161 * XXX filter against symbol
164 process_basic_block(prev, &bi[i].from, &bi[i].flags);
171 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
172 struct addr_location *al __maybe_unused,
173 bool single __maybe_unused,
174 void *arg __maybe_unused)
176 struct hist_entry *he = iter->he;
177 struct branch_info *bi;
178 struct perf_sample *sample = iter->sample;
179 struct evsel *evsel = iter->evsel;
182 bi = he->branch_info;
183 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
188 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
194 static int process_branch_callback(struct evsel *evsel,
195 struct perf_sample *sample,
196 struct addr_location *al,
197 struct perf_annotate *ann,
198 struct machine *machine)
200 struct hist_entry_iter iter = {
203 .add_entry_cb = hist_iter__branch_callback,
204 .hide_unresolved = symbol_conf.hide_unresolved,
205 .ops = &hist_iter_branch,
207 struct addr_location a;
210 addr_location__init(&a);
211 if (machine__resolve(machine, &a, sample) < 0) {
222 dso__set_hit(map__dso(a.map));
224 hist__account_cycles(sample->branch_stack, al, sample, false, NULL);
226 ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
228 addr_location__exit(&a);
232 static bool has_annotation(struct perf_annotate *ann)
234 return ui__has_annotation() || ann->use_stdio2;
237 static int evsel__add_sample(struct evsel *evsel, struct perf_sample *sample,
238 struct addr_location *al, struct perf_annotate *ann,
239 struct machine *machine)
241 struct hists *hists = evsel__hists(evsel);
242 struct hist_entry *he;
245 if ((!ann->has_br_stack || !has_annotation(ann)) &&
246 ann->sym_hist_filter != NULL &&
248 strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
249 /* We're only interested in a symbol named sym_hist_filter */
251 * FIXME: why isn't this done in the symbol_filter when loading
254 if (al->sym != NULL) {
255 struct dso *dso = map__dso(al->map);
257 rb_erase_cached(&al->sym->rb_node, dso__symbols(dso));
258 symbol__delete(al->sym);
259 dso__reset_find_symbol_cache(dso);
265 * XXX filtered samples can still have branch entries pointing into our
266 * symbol and are missed.
268 process_branch_stack(sample->branch_stack, al, sample);
270 if (ann->has_br_stack && has_annotation(ann))
271 return process_branch_callback(evsel, sample, al, ann, machine);
273 he = hists__add_entry(hists, al, NULL, NULL, NULL, NULL, sample, true);
277 ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
278 hists__inc_nr_samples(hists, true);
282 static int process_sample_event(struct perf_tool *tool,
283 union perf_event *event,
284 struct perf_sample *sample,
286 struct machine *machine)
288 struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
289 struct addr_location al;
292 addr_location__init(&al);
293 if (machine__resolve(machine, &al, sample) < 0) {
294 pr_warning("problem processing %d event, skipping it.\n",
300 if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
304 evsel__add_sample(evsel, sample, &al, ann, machine)) {
305 pr_warning("problem incrementing symbol count, "
310 addr_location__exit(&al);
314 static int process_feature_event(struct perf_session *session,
315 union perf_event *event)
317 if (event->feat.feat_id < HEADER_LAST_FEATURE)
318 return perf_event__process_feature(session, event);
322 static int hist_entry__tty_annotate(struct hist_entry *he,
324 struct perf_annotate *ann)
326 if (!ann->use_stdio2)
327 return symbol__tty_annotate(&he->ms, evsel);
329 return symbol__tty_annotate2(&he->ms, evsel);
332 static void print_annotate_data_stat(struct annotated_data_stat *s)
334 #define PRINT_STAT(fld) if (s->fld) printf("%10d : %s\n", s->fld, #fld)
336 int bad = s->no_sym +
347 int ok = s->total - bad;
349 printf("Annotate data type stats:\n");
350 printf("total %d, ok %d (%.1f%%), bad %d (%.1f%%)\n",
351 s->total, ok, 100.0 * ok / (s->total ?: 1), bad, 100.0 * bad / (s->total ?: 1));
352 printf("-----------------------------------------------------------\n");
355 PRINT_STAT(no_insn_ops);
356 PRINT_STAT(no_mem_ops);
358 PRINT_STAT(no_dbginfo);
359 PRINT_STAT(no_cuinfo);
361 PRINT_STAT(no_typeinfo);
362 PRINT_STAT(invalid_size);
363 PRINT_STAT(bad_offset);
364 PRINT_STAT(insn_track);
370 static void print_annotate_item_stat(struct list_head *head, const char *title)
372 struct annotated_item_stat *istat, *pos, *iter;
373 int total_good, total_bad, total;
377 /* sort the list by count */
378 list_splice_init(head, &tmp);
379 total_good = total_bad = 0;
381 list_for_each_entry_safe(istat, pos, &tmp, list) {
382 total_good += istat->good;
383 total_bad += istat->bad;
384 sum1 = istat->good + istat->bad;
386 list_for_each_entry(iter, head, list) {
387 sum2 = iter->good + iter->bad;
391 list_move_tail(&istat->list, &iter->list);
393 total = total_good + total_bad;
395 printf("Annotate %s stats\n", title);
396 printf("total %d, ok %d (%.1f%%), bad %d (%.1f%%)\n\n", total,
397 total_good, 100.0 * total_good / (total ?: 1),
398 total_bad, 100.0 * total_bad / (total ?: 1));
399 printf(" %-10s: %5s %5s\n", "Name", "Good", "Bad");
400 printf("-----------------------------------------------------------\n");
401 list_for_each_entry(istat, head, list)
402 printf(" %-10s: %5d %5d\n", istat->name, istat->good, istat->bad);
406 static void hists__find_annotations(struct hists *hists,
408 struct perf_annotate *ann)
410 struct rb_node *nd = rb_first_cached(&hists->entries), *next;
414 print_annotate_data_stat(&ann_data_stat);
416 print_annotate_item_stat(&ann_insn_stat, "Instruction");
419 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
420 struct annotation *notes;
422 if (he->ms.sym == NULL || dso__annotate_warned(map__dso(he->ms.map)))
425 if (ann->sym_hist_filter &&
426 (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
429 if (ann->min_percent) {
431 u64 total = hists__total_period(hists);
434 percent = 100.0 * he->stat.period / total;
436 if (percent < ann->min_percent)
440 notes = symbol__annotation(he->ms.sym);
441 if (notes->src == NULL) {
443 if (key == K_LEFT || key == '<')
450 if (ann->data_type) {
451 /* skip unknown type */
452 if (he->mem_type->histograms == NULL)
455 if (ann->target_data_type) {
456 const char *type_name = he->mem_type->self.type_name;
458 /* skip 'struct ' prefix in the type name */
459 if (strncmp(ann->target_data_type, "struct ", 7) &&
460 !strncmp(type_name, "struct ", 7))
463 /* skip 'union ' prefix in the type name */
464 if (strncmp(ann->target_data_type, "union ", 6) &&
465 !strncmp(type_name, "union ", 6))
468 if (strcmp(ann->target_data_type, type_name))
472 if (use_browser == 1)
473 key = hist_entry__annotate_data_tui(he, evsel, NULL);
475 key = hist_entry__annotate_data_tty(he, evsel);
479 if (!ann->skip_missing)
494 if (use_browser == 0 || next != NULL)
500 if (use_browser == 2) {
502 int (*annotate)(struct hist_entry *he,
504 struct hist_browser_timer *hbt);
506 annotate = dlsym(perf_gtk_handle,
507 "hist_entry__gtk_annotate");
508 if (annotate == NULL) {
509 ui__error("GTK browser not found!\n");
513 ret = annotate(he, evsel, NULL);
514 if (!ret || !ann->skip_missing)
517 /* skip missing symbols */
519 } else if (use_browser == 1) {
520 key = hist_entry__tui_annotate(he, evsel, NULL);
524 if (!ann->skip_missing)
542 hist_entry__tty_annotate(he, evsel, ann);
548 static int __cmd_annotate(struct perf_annotate *ann)
551 struct perf_session *session = ann->session;
553 u64 total_nr_samples;
556 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
562 if (!annotate_opts.objdump_path) {
563 ret = perf_env__lookup_objdump(&session->header.env,
564 &annotate_opts.objdump_path);
569 ret = perf_session__process_events(session);
574 perf_session__fprintf_nr_events(session, stdout, false);
575 evlist__fprintf_nr_events(session->evlist, stdout, false);
580 perf_session__fprintf(session, stdout);
583 perf_session__fprintf_dsos(session, stdout);
585 total_nr_samples = 0;
586 evlist__for_each_entry(session->evlist, pos) {
587 struct hists *hists = evsel__hists(pos);
588 u32 nr_samples = hists->stats.nr_samples;
589 struct ui_progress prog;
591 if (nr_samples > 0) {
592 total_nr_samples += nr_samples;
594 ui_progress__init(&prog, nr_samples,
595 "Merging related events...");
596 hists__collapse_resort(hists, &prog);
597 ui_progress__finish();
599 /* Don't sort callchain */
600 evsel__reset_sample_bit(pos, CALLCHAIN);
602 ui_progress__init(&prog, nr_samples,
603 "Sorting events for output...");
604 evsel__output_resort(pos, &prog);
605 ui_progress__finish();
608 * An event group needs to display other events too.
609 * Let's delay printing until other events are processed.
611 if (symbol_conf.event_group) {
612 if (!evsel__is_group_leader(pos)) {
613 struct hists *leader_hists;
615 leader_hists = evsel__hists(evsel__leader(pos));
616 hists__match(leader_hists, hists);
617 hists__link(leader_hists, hists);
622 hists__find_annotations(hists, pos, ann);
626 if (total_nr_samples == 0) {
627 ui__error("The %s data has no samples!\n", session->data->path);
631 /* Display group events together */
632 evlist__for_each_entry(session->evlist, pos) {
633 struct hists *hists = evsel__hists(pos);
634 u32 nr_samples = hists->stats.nr_samples;
639 if (!symbol_conf.event_group || !evsel__is_group_leader(pos))
642 hists__find_annotations(hists, pos, ann);
645 if (use_browser == 2) {
646 void (*show_annotations)(void);
648 show_annotations = dlsym(perf_gtk_handle,
649 "perf_gtk__show_annotations");
650 if (show_annotations == NULL) {
651 ui__error("GTK browser not found!\n");
661 static int parse_percent_limit(const struct option *opt, const char *str,
662 int unset __maybe_unused)
664 struct perf_annotate *ann = opt->value;
665 double pcnt = strtof(str, NULL);
667 ann->min_percent = pcnt;
671 static int parse_data_type(const struct option *opt, const char *str, int unset)
673 struct perf_annotate *ann = opt->value;
675 ann->data_type = !unset;
677 ann->target_data_type = strdup(str);
682 static const char * const annotate_usage[] = {
683 "perf annotate [<options>]",
687 int cmd_annotate(int argc, const char **argv)
689 struct perf_annotate annotate = {
691 .sample = process_sample_event,
692 .mmap = perf_event__process_mmap,
693 .mmap2 = perf_event__process_mmap2,
694 .comm = perf_event__process_comm,
695 .exit = perf_event__process_exit,
696 .fork = perf_event__process_fork,
697 .namespaces = perf_event__process_namespaces,
698 .attr = perf_event__process_attr,
699 .build_id = perf_event__process_build_id,
700 #ifdef HAVE_LIBTRACEEVENT
701 .tracing_data = perf_event__process_tracing_data,
703 .id_index = perf_event__process_id_index,
704 .auxtrace_info = perf_event__process_auxtrace_info,
705 .auxtrace = perf_event__process_auxtrace,
706 .feature = process_feature_event,
707 .ordered_events = true,
708 .ordering_requires_timestamps = true,
711 struct perf_data data = {
712 .mode = PERF_DATA_MODE_READ,
714 struct itrace_synth_opts itrace_synth_opts = {
717 const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
718 struct option options[] = {
719 OPT_STRING('i', "input", &input_name, "file",
721 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
722 "only consider symbols in these dsos"),
723 OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
724 "symbol to annotate"),
725 OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
726 OPT_INCR('v', "verbose", &verbose,
727 "be more verbose (show symbol address, etc)"),
728 OPT_BOOLEAN('q', "quiet", &quiet, "do now show any warnings or messages"),
729 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
730 "dump raw trace in ASCII"),
731 #ifdef HAVE_GTK2_SUPPORT
732 OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
734 #ifdef HAVE_SLANG_SUPPORT
735 OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
737 OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
738 OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
739 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
740 "don't load vmlinux even if found"),
741 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
742 "file", "vmlinux pathname"),
743 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
744 "load module symbols - WARNING: use only with -k and LIVE kernel"),
745 OPT_BOOLEAN('l', "print-line", &annotate_opts.print_lines,
746 "print matching source lines (may be slow)"),
747 OPT_BOOLEAN('P', "full-paths", &annotate_opts.full_path,
748 "Don't shorten the displayed pathnames"),
749 OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
750 "Skip symbols that cannot be annotated"),
751 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
753 "Show event group information together"),
754 OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
755 OPT_CALLBACK(0, "symfs", NULL, "directory",
756 "Look for files with symbols relative to this directory",
757 symbol__config_symfs),
758 OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
759 "Interleave source code with assembly code (default)"),
760 OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
761 "Display raw encoding of assembly instructions (default)"),
762 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
763 "Specify disassembler style (e.g. -M intel for intel syntax)"),
764 OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
765 "Add prefix to source file path names in programs (with --prefix-strip)"),
766 OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
767 "Strip first N entries of source file path name in programs (with --prefix)"),
768 OPT_STRING(0, "objdump", &objdump_path, "path",
769 "objdump binary to use for disassembly and annotations"),
770 OPT_STRING(0, "addr2line", &addr2line_path, "path",
771 "addr2line binary to use for line numbers"),
772 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
773 "Enable symbol demangling"),
774 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
775 "Enable kernel symbol demangling"),
776 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
777 "Show a column with the sum of periods"),
778 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
779 "Show a column with the number of samples"),
780 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
781 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
782 stdio__config_color, "always"),
783 OPT_CALLBACK(0, "percent-type", &annotate_opts, "local-period",
784 "Set percent type local/global-period/hits",
785 annotate_parse_percent_type),
786 OPT_CALLBACK(0, "percent-limit", &annotate, "percent",
787 "Don't show entries under that percent", parse_percent_limit),
788 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
789 "Instruction Tracing options\n" ITRACE_HELP,
790 itrace_parse_synth_opts),
791 OPT_CALLBACK_OPTARG(0, "data-type", &annotate, NULL, "name",
792 "Show data type annotate for the memory accesses",
794 OPT_BOOLEAN(0, "type-stat", &annotate.type_stat,
795 "Show stats for the data type annotation"),
796 OPT_BOOLEAN(0, "insn-stat", &annotate.insn_stat,
797 "Show instruction stats for the data type annotation"),
802 set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
803 set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
805 annotation_options__init();
811 annotation_config__init();
813 argc = parse_options(argc, argv, options, annotate_usage, 0);
816 * Special case: if there's an argument left then assume that
817 * it's a symbol filter:
820 usage_with_options(annotate_usage, options);
822 annotate.sym_hist_filter = argv[0];
825 if (disassembler_style) {
826 annotate_opts.disassembler_style = strdup(disassembler_style);
827 if (!annotate_opts.disassembler_style)
831 annotate_opts.objdump_path = strdup(objdump_path);
832 if (!annotate_opts.objdump_path)
835 if (addr2line_path) {
836 symbol_conf.addr2line_path = strdup(addr2line_path);
837 if (!symbol_conf.addr2line_path)
841 if (annotate_check_args() < 0)
844 #ifdef HAVE_GTK2_SUPPORT
845 if (symbol_conf.show_nr_samples && annotate.use_gtk) {
846 pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
851 #ifndef HAVE_DWARF_GETLOCATIONS_SUPPORT
852 if (annotate.data_type) {
853 pr_err("Error: Data type profiling is disabled due to missing DWARF support\n");
858 ret = symbol__validate_sym_arguments();
865 data.path = input_name;
867 annotate.session = perf_session__new(&data, &annotate.tool);
868 if (IS_ERR(annotate.session))
869 return PTR_ERR(annotate.session);
871 annotate.session->itrace_synth_opts = &itrace_synth_opts;
873 annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
874 HEADER_BRANCH_STACK);
876 if (annotate.group_set)
877 evlist__force_leader(annotate.session->evlist);
879 ret = symbol__annotation_init();
883 symbol_conf.try_vmlinux_path = true;
885 ret = symbol__init(&annotate.session->header.env);
889 if (annotate.use_stdio || annotate.use_stdio2)
891 #ifdef HAVE_SLANG_SUPPORT
892 else if (annotate.use_tui)
895 #ifdef HAVE_GTK2_SUPPORT
896 else if (annotate.use_gtk)
900 if (annotate.data_type) {
901 annotate_opts.annotate_src = false;
902 symbol_conf.annotate_data_member = true;
903 symbol_conf.annotate_data_sample = true;
909 * Events of different processes may correspond to the same
910 * symbol, we do not care about the processes in annotate,
911 * set sort order to avoid repeated output.
913 if (annotate.data_type)
914 sort_order = "dso,type";
916 sort_order = "dso,symbol";
919 * Set SORT_MODE__BRANCH so that annotate display IPC/Cycle
920 * if branch info is in perf data in TUI mode.
922 if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack)
923 sort__mode = SORT_MODE__BRANCH;
925 if (setup_sorting(NULL) < 0)
926 usage_with_options(annotate_usage, options);
928 ret = __cmd_annotate(&annotate);
932 * Speed up the exit process by only deleting for debug builds. For
933 * large files this can save time.
936 perf_session__delete(annotate.session);
938 annotation_options__exit();