Merge tag 'fpga-for-6.0-final' of git://git.kernel.org/pub/scm/linux/kernel/git/fpga...
[platform/kernel/linux-starfive.git] / tools / perf / builtin-annotate.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * builtin-annotate.c
4  *
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.
8  */
9 #include "builtin.h"
10
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"
17
18 #include "perf.h"
19 #include "util/debug.h"
20
21 #include "util/evlist.h"
22 #include "util/evsel.h"
23 #include "util/annotate.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"
29 #include "util/dso.h"
30 #include "util/machine.h"
31 #include "util/map.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
40 #include <dlfcn.h>
41 #include <errno.h>
42 #include <linux/bitmap.h>
43 #include <linux/err.h>
44
45 struct perf_annotate {
46         struct perf_tool tool;
47         struct perf_session *session;
48         struct annotation_options opts;
49 #ifdef HAVE_SLANG_SUPPORT
50         bool       use_tui;
51 #endif
52         bool       use_stdio, use_stdio2;
53 #ifdef HAVE_GTK2_SUPPORT
54         bool       use_gtk;
55 #endif
56         bool       skip_missing;
57         bool       has_br_stack;
58         bool       group_set;
59         float      min_percent;
60         const char *sym_hist_filter;
61         const char *cpu_list;
62         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
63 };
64
65 /*
66  * Given one basic block:
67  *
68  *      from    to              branch_i
69  *      * ----> *
70  *              |
71  *              | block
72  *              v
73  *              * ----> *
74  *              from    to      branch_i+1
75  *
76  * where the horizontal are the branches and the vertical is the executed
77  * block of instructions.
78  *
79  * We count, for each 'instruction', the number of blocks that covered it as
80  * well as count the ratio each branch is taken.
81  *
82  * We can do this without knowing the actual instruction stream by keeping
83  * track of the address ranges. We break down ranges such that there is no
84  * overlap and iterate from the start until the end.
85  *
86  * @acme: once we parse the objdump output _before_ processing the samples,
87  * we can easily fold the branch.cycles IPC bits in.
88  */
89 static void process_basic_block(struct addr_map_symbol *start,
90                                 struct addr_map_symbol *end,
91                                 struct branch_flags *flags)
92 {
93         struct symbol *sym = start->ms.sym;
94         struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
95         struct block_range_iter iter;
96         struct block_range *entry;
97
98         /*
99          * Sanity; NULL isn't executable and the CPU cannot execute backwards
100          */
101         if (!start->addr || start->addr > end->addr)
102                 return;
103
104         iter = block_range__create(start->addr, end->addr);
105         if (!block_range_iter__valid(&iter))
106                 return;
107
108         /*
109          * First block in range is a branch target.
110          */
111         entry = block_range_iter(&iter);
112         assert(entry->is_target);
113         entry->entry++;
114
115         do {
116                 entry = block_range_iter(&iter);
117
118                 entry->coverage++;
119                 entry->sym = sym;
120
121                 if (notes)
122                         notes->max_coverage = max(notes->max_coverage, entry->coverage);
123
124         } while (block_range_iter__next(&iter));
125
126         /*
127          * Last block in rage is a branch.
128          */
129         entry = block_range_iter(&iter);
130         assert(entry->is_branch);
131         entry->taken++;
132         if (flags->predicted)
133                 entry->pred++;
134 }
135
136 static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
137                                  struct perf_sample *sample)
138 {
139         struct addr_map_symbol *prev = NULL;
140         struct branch_info *bi;
141         int i;
142
143         if (!bs || !bs->nr)
144                 return;
145
146         bi = sample__resolve_bstack(sample, al);
147         if (!bi)
148                 return;
149
150         for (i = bs->nr - 1; i >= 0; i--) {
151                 /*
152                  * XXX filter against symbol
153                  */
154                 if (prev)
155                         process_basic_block(prev, &bi[i].from, &bi[i].flags);
156                 prev = &bi[i].to;
157         }
158
159         free(bi);
160 }
161
162 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
163                                       struct addr_location *al __maybe_unused,
164                                       bool single __maybe_unused,
165                                       void *arg __maybe_unused)
166 {
167         struct hist_entry *he = iter->he;
168         struct branch_info *bi;
169         struct perf_sample *sample = iter->sample;
170         struct evsel *evsel = iter->evsel;
171         int err;
172
173         bi = he->branch_info;
174         err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
175
176         if (err)
177                 goto out;
178
179         err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
180
181 out:
182         return err;
183 }
184
185 static int process_branch_callback(struct evsel *evsel,
186                                    struct perf_sample *sample,
187                                    struct addr_location *al __maybe_unused,
188                                    struct perf_annotate *ann,
189                                    struct machine *machine)
190 {
191         struct hist_entry_iter iter = {
192                 .evsel          = evsel,
193                 .sample         = sample,
194                 .add_entry_cb   = hist_iter__branch_callback,
195                 .hide_unresolved        = symbol_conf.hide_unresolved,
196                 .ops            = &hist_iter_branch,
197         };
198
199         struct addr_location a;
200
201         if (machine__resolve(machine, &a, sample) < 0)
202                 return -1;
203
204         if (a.sym == NULL)
205                 return 0;
206
207         if (a.map != NULL)
208                 a.map->dso->hit = 1;
209
210         hist__account_cycles(sample->branch_stack, al, sample, false, NULL);
211
212         return hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
213 }
214
215 static bool has_annotation(struct perf_annotate *ann)
216 {
217         return ui__has_annotation() || ann->use_stdio2;
218 }
219
220 static int evsel__add_sample(struct evsel *evsel, struct perf_sample *sample,
221                              struct addr_location *al, struct perf_annotate *ann,
222                              struct machine *machine)
223 {
224         struct hists *hists = evsel__hists(evsel);
225         struct hist_entry *he;
226         int ret;
227
228         if ((!ann->has_br_stack || !has_annotation(ann)) &&
229             ann->sym_hist_filter != NULL &&
230             (al->sym == NULL ||
231              strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
232                 /* We're only interested in a symbol named sym_hist_filter */
233                 /*
234                  * FIXME: why isn't this done in the symbol_filter when loading
235                  * the DSO?
236                  */
237                 if (al->sym != NULL) {
238                         rb_erase_cached(&al->sym->rb_node,
239                                  &al->map->dso->symbols);
240                         symbol__delete(al->sym);
241                         dso__reset_find_symbol_cache(al->map->dso);
242                 }
243                 return 0;
244         }
245
246         /*
247          * XXX filtered samples can still have branch entries pointing into our
248          * symbol and are missed.
249          */
250         process_branch_stack(sample->branch_stack, al, sample);
251
252         if (ann->has_br_stack && has_annotation(ann))
253                 return process_branch_callback(evsel, sample, al, ann, machine);
254
255         he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
256         if (he == NULL)
257                 return -ENOMEM;
258
259         ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
260         hists__inc_nr_samples(hists, true);
261         return ret;
262 }
263
264 static int process_sample_event(struct perf_tool *tool,
265                                 union perf_event *event,
266                                 struct perf_sample *sample,
267                                 struct evsel *evsel,
268                                 struct machine *machine)
269 {
270         struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
271         struct addr_location al;
272         int ret = 0;
273
274         if (machine__resolve(machine, &al, sample) < 0) {
275                 pr_warning("problem processing %d event, skipping it.\n",
276                            event->header.type);
277                 return -1;
278         }
279
280         if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
281                 goto out_put;
282
283         if (!al.filtered &&
284             evsel__add_sample(evsel, sample, &al, ann, machine)) {
285                 pr_warning("problem incrementing symbol count, "
286                            "skipping event\n");
287                 ret = -1;
288         }
289 out_put:
290         addr_location__put(&al);
291         return ret;
292 }
293
294 static int process_feature_event(struct perf_session *session,
295                                  union perf_event *event)
296 {
297         if (event->feat.feat_id < HEADER_LAST_FEATURE)
298                 return perf_event__process_feature(session, event);
299         return 0;
300 }
301
302 static int hist_entry__tty_annotate(struct hist_entry *he,
303                                     struct evsel *evsel,
304                                     struct perf_annotate *ann)
305 {
306         if (!ann->use_stdio2)
307                 return symbol__tty_annotate(&he->ms, evsel, &ann->opts);
308
309         return symbol__tty_annotate2(&he->ms, evsel, &ann->opts);
310 }
311
312 static void hists__find_annotations(struct hists *hists,
313                                     struct evsel *evsel,
314                                     struct perf_annotate *ann)
315 {
316         struct rb_node *nd = rb_first_cached(&hists->entries), *next;
317         int key = K_RIGHT;
318
319         while (nd) {
320                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
321                 struct annotation *notes;
322
323                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
324                         goto find_next;
325
326                 if (ann->sym_hist_filter &&
327                     (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
328                         goto find_next;
329
330                 if (ann->min_percent) {
331                         float percent = 0;
332                         u64 total = hists__total_period(hists);
333
334                         if (total)
335                                 percent = 100.0 * he->stat.period / total;
336
337                         if (percent < ann->min_percent)
338                                 goto find_next;
339                 }
340
341                 notes = symbol__annotation(he->ms.sym);
342                 if (notes->src == NULL) {
343 find_next:
344                         if (key == K_LEFT)
345                                 nd = rb_prev(nd);
346                         else
347                                 nd = rb_next(nd);
348                         continue;
349                 }
350
351                 if (use_browser == 2) {
352                         int ret;
353                         int (*annotate)(struct hist_entry *he,
354                                         struct evsel *evsel,
355                                         struct hist_browser_timer *hbt);
356
357                         annotate = dlsym(perf_gtk_handle,
358                                          "hist_entry__gtk_annotate");
359                         if (annotate == NULL) {
360                                 ui__error("GTK browser not found!\n");
361                                 return;
362                         }
363
364                         ret = annotate(he, evsel, NULL);
365                         if (!ret || !ann->skip_missing)
366                                 return;
367
368                         /* skip missing symbols */
369                         nd = rb_next(nd);
370                 } else if (use_browser == 1) {
371                         key = hist_entry__tui_annotate(he, evsel, NULL, &ann->opts);
372
373                         switch (key) {
374                         case -1:
375                                 if (!ann->skip_missing)
376                                         return;
377                                 /* fall through */
378                         case K_RIGHT:
379                                 next = rb_next(nd);
380                                 break;
381                         case K_LEFT:
382                                 next = rb_prev(nd);
383                                 break;
384                         default:
385                                 return;
386                         }
387
388                         if (next != NULL)
389                                 nd = next;
390                 } else {
391                         hist_entry__tty_annotate(he, evsel, ann);
392                         nd = rb_next(nd);
393                 }
394         }
395 }
396
397 static int __cmd_annotate(struct perf_annotate *ann)
398 {
399         int ret;
400         struct perf_session *session = ann->session;
401         struct evsel *pos;
402         u64 total_nr_samples;
403
404         if (ann->cpu_list) {
405                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
406                                                ann->cpu_bitmap);
407                 if (ret)
408                         goto out;
409         }
410
411         if (!ann->opts.objdump_path) {
412                 ret = perf_env__lookup_objdump(&session->header.env,
413                                                &ann->opts.objdump_path);
414                 if (ret)
415                         goto out;
416         }
417
418         ret = perf_session__process_events(session);
419         if (ret)
420                 goto out;
421
422         if (dump_trace) {
423                 perf_session__fprintf_nr_events(session, stdout, false);
424                 evlist__fprintf_nr_events(session->evlist, stdout, false);
425                 goto out;
426         }
427
428         if (verbose > 3)
429                 perf_session__fprintf(session, stdout);
430
431         if (verbose > 2)
432                 perf_session__fprintf_dsos(session, stdout);
433
434         total_nr_samples = 0;
435         evlist__for_each_entry(session->evlist, pos) {
436                 struct hists *hists = evsel__hists(pos);
437                 u32 nr_samples = hists->stats.nr_samples;
438
439                 if (nr_samples > 0) {
440                         total_nr_samples += nr_samples;
441                         hists__collapse_resort(hists, NULL);
442                         /* Don't sort callchain */
443                         evsel__reset_sample_bit(pos, CALLCHAIN);
444                         evsel__output_resort(pos, NULL);
445
446                         if (symbol_conf.event_group && !evsel__is_group_leader(pos))
447                                 continue;
448
449                         hists__find_annotations(hists, pos, ann);
450                 }
451         }
452
453         if (total_nr_samples == 0) {
454                 ui__error("The %s data has no samples!\n", session->data->path);
455                 goto out;
456         }
457
458         if (use_browser == 2) {
459                 void (*show_annotations)(void);
460
461                 show_annotations = dlsym(perf_gtk_handle,
462                                          "perf_gtk__show_annotations");
463                 if (show_annotations == NULL) {
464                         ui__error("GTK browser not found!\n");
465                         goto out;
466                 }
467                 show_annotations();
468         }
469
470 out:
471         return ret;
472 }
473
474 static int parse_percent_limit(const struct option *opt, const char *str,
475                                int unset __maybe_unused)
476 {
477         struct perf_annotate *ann = opt->value;
478         double pcnt = strtof(str, NULL);
479
480         ann->min_percent = pcnt;
481         return 0;
482 }
483
484 static const char * const annotate_usage[] = {
485         "perf annotate [<options>]",
486         NULL
487 };
488
489 int cmd_annotate(int argc, const char **argv)
490 {
491         struct perf_annotate annotate = {
492                 .tool = {
493                         .sample = process_sample_event,
494                         .mmap   = perf_event__process_mmap,
495                         .mmap2  = perf_event__process_mmap2,
496                         .comm   = perf_event__process_comm,
497                         .exit   = perf_event__process_exit,
498                         .fork   = perf_event__process_fork,
499                         .namespaces = perf_event__process_namespaces,
500                         .attr   = perf_event__process_attr,
501                         .build_id = perf_event__process_build_id,
502                         .tracing_data   = perf_event__process_tracing_data,
503                         .id_index       = perf_event__process_id_index,
504                         .auxtrace_info  = perf_event__process_auxtrace_info,
505                         .auxtrace       = perf_event__process_auxtrace,
506                         .feature        = process_feature_event,
507                         .ordered_events = true,
508                         .ordering_requires_timestamps = true,
509                 },
510                 .opts = annotation__default_options,
511         };
512         struct perf_data data = {
513                 .mode  = PERF_DATA_MODE_READ,
514         };
515         struct itrace_synth_opts itrace_synth_opts = {
516                 .set = 0,
517         };
518         struct option options[] = {
519         OPT_STRING('i', "input", &input_name, "file",
520                     "input file name"),
521         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
522                    "only consider symbols in these dsos"),
523         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
524                     "symbol to annotate"),
525         OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
526         OPT_INCR('v', "verbose", &verbose,
527                     "be more verbose (show symbol address, etc)"),
528         OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
529         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
530                     "dump raw trace in ASCII"),
531 #ifdef HAVE_GTK2_SUPPORT
532         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
533 #endif
534 #ifdef HAVE_SLANG_SUPPORT
535         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
536 #endif
537         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
538         OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
539         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
540                     "don't load vmlinux even if found"),
541         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
542                    "file", "vmlinux pathname"),
543         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
544                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
545         OPT_BOOLEAN('l', "print-line", &annotate.opts.print_lines,
546                     "print matching source lines (may be slow)"),
547         OPT_BOOLEAN('P', "full-paths", &annotate.opts.full_path,
548                     "Don't shorten the displayed pathnames"),
549         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
550                     "Skip symbols that cannot be annotated"),
551         OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
552                         &annotate.group_set,
553                         "Show event group information together"),
554         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
555         OPT_CALLBACK(0, "symfs", NULL, "directory",
556                      "Look for files with symbols relative to this directory",
557                      symbol__config_symfs),
558         OPT_BOOLEAN(0, "source", &annotate.opts.annotate_src,
559                     "Interleave source code with assembly code (default)"),
560         OPT_BOOLEAN(0, "asm-raw", &annotate.opts.show_asm_raw,
561                     "Display raw encoding of assembly instructions (default)"),
562         OPT_STRING('M', "disassembler-style", &annotate.opts.disassembler_style, "disassembler style",
563                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
564         OPT_STRING(0, "prefix", &annotate.opts.prefix, "prefix",
565                     "Add prefix to source file path names in programs (with --prefix-strip)"),
566         OPT_STRING(0, "prefix-strip", &annotate.opts.prefix_strip, "N",
567                     "Strip first N entries of source file path name in programs (with --prefix)"),
568         OPT_STRING(0, "objdump", &annotate.opts.objdump_path, "path",
569                    "objdump binary to use for disassembly and annotations"),
570         OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
571                     "Enable symbol demangling"),
572         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
573                     "Enable kernel symbol demangling"),
574         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
575                     "Show event group information together"),
576         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
577                     "Show a column with the sum of periods"),
578         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
579                     "Show a column with the number of samples"),
580         OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
581                              "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
582                              stdio__config_color, "always"),
583         OPT_CALLBACK(0, "percent-type", &annotate.opts, "local-period",
584                      "Set percent type local/global-period/hits",
585                      annotate_parse_percent_type),
586         OPT_CALLBACK(0, "percent-limit", &annotate, "percent",
587                      "Don't show entries under that percent", parse_percent_limit),
588         OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
589                             "Instruction Tracing options\n" ITRACE_HELP,
590                             itrace_parse_synth_opts),
591
592         OPT_END()
593         };
594         int ret;
595
596         set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
597         set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
598
599
600         ret = hists__init();
601         if (ret < 0)
602                 return ret;
603
604         annotation_config__init(&annotate.opts);
605
606         argc = parse_options(argc, argv, options, annotate_usage, 0);
607         if (argc) {
608                 /*
609                  * Special case: if there's an argument left then assume that
610                  * it's a symbol filter:
611                  */
612                 if (argc > 1)
613                         usage_with_options(annotate_usage, options);
614
615                 annotate.sym_hist_filter = argv[0];
616         }
617
618         if (annotate_check_args(&annotate.opts) < 0)
619                 return -EINVAL;
620
621 #ifdef HAVE_GTK2_SUPPORT
622         if (symbol_conf.show_nr_samples && annotate.use_gtk) {
623                 pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
624                 return ret;
625         }
626 #endif
627
628         ret = symbol__validate_sym_arguments();
629         if (ret)
630                 return ret;
631
632         if (quiet)
633                 perf_quiet_option();
634
635         data.path = input_name;
636
637         annotate.session = perf_session__new(&data, &annotate.tool);
638         if (IS_ERR(annotate.session))
639                 return PTR_ERR(annotate.session);
640
641         annotate.session->itrace_synth_opts = &itrace_synth_opts;
642
643         annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
644                                                       HEADER_BRANCH_STACK);
645
646         if (annotate.group_set)
647                 evlist__force_leader(annotate.session->evlist);
648
649         ret = symbol__annotation_init();
650         if (ret < 0)
651                 goto out_delete;
652
653         symbol_conf.try_vmlinux_path = true;
654
655         ret = symbol__init(&annotate.session->header.env);
656         if (ret < 0)
657                 goto out_delete;
658
659         if (annotate.use_stdio || annotate.use_stdio2)
660                 use_browser = 0;
661 #ifdef HAVE_SLANG_SUPPORT
662         else if (annotate.use_tui)
663                 use_browser = 1;
664 #endif
665 #ifdef HAVE_GTK2_SUPPORT
666         else if (annotate.use_gtk)
667                 use_browser = 2;
668 #endif
669
670         setup_browser(true);
671
672         /*
673          * Events of different processes may correspond to the same
674          * symbol, we do not care about the processes in annotate,
675          * set sort order to avoid repeated output.
676          */
677         sort_order = "dso,symbol";
678
679         /*
680          * Set SORT_MODE__BRANCH so that annotate display IPC/Cycle
681          * if branch info is in perf data in TUI mode.
682          */
683         if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack)
684                 sort__mode = SORT_MODE__BRANCH;
685
686         if (setup_sorting(NULL) < 0)
687                 usage_with_options(annotate_usage, options);
688
689         ret = __cmd_annotate(&annotate);
690
691 out_delete:
692         /*
693          * Speed up the exit process, for large files this can
694          * take quite a while.
695          *
696          * XXX Enable this when using valgrind or if we ever
697          * librarize this command.
698          *
699          * Also experiment with obstacks to see how much speed
700          * up we'll get here.
701          *
702          * perf_session__delete(session);
703          */
704         return ret;
705 }