tools/perf/builtin-report.c

   1 /*
   2  * builtin-report.c
   3  *
   4  * Builtin report command: Analyze the perf.data input file,
   5  * look up and read DSOs and symbol information and display
   6  * a histogram of results, along various sorting keys.
   7  */
   8 #include "builtin.h"
   9
  10 #include "util/util.h"
  11 #include "util/config.h"
  12
  13 #include "util/annotate.h"
  14 #include "util/color.h"
  15 #include <linux/list.h>
  16 #include <linux/rbtree.h>
  17 #include "util/symbol.h"
  18 #include "util/callchain.h"
  19 #include "util/values.h"
  20
  21 #include "perf.h"
  22 #include "util/debug.h"
  23 #include "util/evlist.h"
  24 #include "util/evsel.h"
  25 #include "util/header.h"
  26 #include "util/session.h"
  27 #include "util/tool.h"
  28
  29 #include <subcmd/parse-options.h>
  30 #include <subcmd/exec-cmd.h>
  31 #include "util/parse-events.h"
  32
  33 #include "util/thread.h"
  34 #include "util/sort.h"
  35 #include "util/hist.h"
  36 #include "util/data.h"
  37 #include "arch/common.h"
  38 #include "util/time-utils.h"
  39 #include "util/auxtrace.h"
  40 #include "util/units.h"
  41 #include "util/branch.h"
  42
  43 #include <dlfcn.h>
  44 #include <errno.h>
  45 #include <inttypes.h>
  46 #include <regex.h>
  47 #include <signal.h>
  48 #include <linux/bitmap.h>
  49 #include <linux/stringify.h>
  50 #include <sys/types.h>
  51 #include <sys/stat.h>
  52 #include <unistd.h>
  53
  54 struct report {
  55         struct perf_tool        tool;
  56         struct perf_session     *session;
  57         bool                    use_tui, use_gtk, use_stdio;
  58         bool                    show_full_info;
  59         bool                    show_threads;
  60         bool                    inverted_callchain;
  61         bool                    mem_mode;
  62         bool                    header;
  63         bool                    header_only;
  64         bool                    nonany_branch_mode;
  65         int                     max_stack;
  66         struct perf_read_values show_threads_values;
  67         const char              *pretty_printing_style;
  68         const char              *cpu_list;
  69         const char              *symbol_filter_str;
  70         const char              *time_str;
  71         struct perf_time_interval ptime;
  72         float                   min_percent;
  73         u64                     nr_entries;
  74         u64                     queue_size;
  75         int                     socket_filter;
  76         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  77         struct branch_type_stat brtype_stat;
  78 };
  79
  80 static int report__config(const char *var, const char *value, void *cb)
  81 {
  82         struct report *rep = cb;
  83
  84         if (!strcmp(var, "report.group")) {
  85                 symbol_conf.event_group = perf_config_bool(var, value);
  86                 return 0;
  87         }
  88         if (!strcmp(var, "report.percent-limit")) {
  89                 double pcnt = strtof(value, NULL);
  90
  91                 rep->min_percent = pcnt;
  92                 callchain_param.min_percent = pcnt;
  93                 return 0;
  94         }
  95         if (!strcmp(var, "report.children")) {
  96                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
  97                 return 0;
  98         }
  99         if (!strcmp(var, "report.queue-size"))
 100                 return perf_config_u64(&rep->queue_size, var, value);
 101
 102         if (!strcmp(var, "report.sort_order")) {
 103                 default_sort_order = strdup(value);
 104                 return 0;
 105         }
 106
 107         return 0;
 108 }
 109
 110 static int hist_iter__report_callback(struct hist_entry_iter *iter,
 111                                       struct addr_location *al, bool single,
 112                                       void *arg)
 113 {
 114         int err = 0;
 115         struct report *rep = arg;
 116         struct hist_entry *he = iter->he;
 117         struct perf_evsel *evsel = iter->evsel;
 118         struct perf_sample *sample = iter->sample;
 119         struct mem_info *mi;
 120         struct branch_info *bi;
 121
 122         if (!ui__has_annotation())
 123                 return 0;
 124
 125         hist__account_cycles(sample->branch_stack, al, sample,
 126                              rep->nonany_branch_mode);
 127
 128         if (sort__mode == SORT_MODE__BRANCH) {
 129                 bi = he->branch_info;
 130                 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel->idx);
 131                 if (err)
 132                         goto out;
 133
 134                 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel->idx);
 135
 136         } else if (rep->mem_mode) {
 137                 mi = he->mem_info;
 138                 err = addr_map_symbol__inc_samples(&mi->daddr, sample, evsel->idx);
 139                 if (err)
 140                         goto out;
 141
 142                 err = hist_entry__inc_addr_samples(he, sample, evsel->idx, al->addr);
 143
 144         } else if (symbol_conf.cumulate_callchain) {
 145                 if (single)
 146                         err = hist_entry__inc_addr_samples(he, sample, evsel->idx,
 147                                                            al->addr);
 148         } else {
 149                 err = hist_entry__inc_addr_samples(he, sample, evsel->idx, al->addr);
 150         }
 151
 152 out:
 153         return err;
 154 }
 155
 156 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
 157                                       struct addr_location *al __maybe_unused,
 158                                       bool single __maybe_unused,
 159                                       void *arg)
 160 {
 161         struct hist_entry *he = iter->he;
 162         struct report *rep = arg;
 163         struct branch_info *bi;
 164
 165         bi = he->branch_info;
 166         branch_type_count(&rep->brtype_stat, &bi->flags,
 167                           bi->from.addr, bi->to.addr);
 168
 169         return 0;
 170 }
 171
 172 static int process_sample_event(struct perf_tool *tool,
 173                                 union perf_event *event,
 174                                 struct perf_sample *sample,
 175                                 struct perf_evsel *evsel,
 176                                 struct machine *machine)
 177 {
 178         struct report *rep = container_of(tool, struct report, tool);
 179         struct addr_location al;
 180         struct hist_entry_iter iter = {
 181                 .evsel                  = evsel,
 182                 .sample                 = sample,
 183                 .hide_unresolved        = symbol_conf.hide_unresolved,
 184                 .add_entry_cb           = hist_iter__report_callback,
 185         };
 186         int ret = 0;
 187
 188         if (perf_time__skip_sample(&rep->ptime, sample->time))
 189                 return 0;
 190
 191         if (machine__resolve(machine, &al, sample) < 0) {
 192                 pr_debug("problem processing %d event, skipping it.\n",
 193                          event->header.type);
 194                 return -1;
 195         }
 196
 197         if (symbol_conf.hide_unresolved && al.sym == NULL)
 198                 goto out_put;
 199
 200         if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
 201                 goto out_put;
 202
 203         if (sort__mode == SORT_MODE__BRANCH) {
 204                 /*
 205                  * A non-synthesized event might not have a branch stack if
 206                  * branch stacks have been synthesized (using itrace options).
 207                  */
 208                 if (!sample->branch_stack)
 209                         goto out_put;
 210
 211                 iter.add_entry_cb = hist_iter__branch_callback;
 212                 iter.ops = &hist_iter_branch;
 213         } else if (rep->mem_mode) {
 214                 iter.ops = &hist_iter_mem;
 215         } else if (symbol_conf.cumulate_callchain) {
 216                 iter.ops = &hist_iter_cumulative;
 217         } else {
 218                 iter.ops = &hist_iter_normal;
 219         }
 220
 221         if (al.map != NULL)
 222                 al.map->dso->hit = 1;
 223
 224         ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
 225         if (ret < 0)
 226                 pr_debug("problem adding hist entry, skipping event\n");
 227 out_put:
 228         addr_location__put(&al);
 229         return ret;
 230 }
 231
 232 static int process_read_event(struct perf_tool *tool,
 233                               union perf_event *event,
 234                               struct perf_sample *sample __maybe_unused,
 235                               struct perf_evsel *evsel,
 236                               struct machine *machine __maybe_unused)
 237 {
 238         struct report *rep = container_of(tool, struct report, tool);
 239
 240         if (rep->show_threads) {
 241                 const char *name = evsel ? perf_evsel__name(evsel) : "unknown";
 242                 int err = perf_read_values_add_value(&rep->show_threads_values,
 243                                            event->read.pid, event->read.tid,
 244                                            evsel->idx,
 245                                            name,
 246                                            event->read.value);
 247
 248                 if (err)
 249                         return err;
 250         }
 251
 252         return 0;
 253 }
 254
 255 /* For pipe mode, sample_type is not currently set */
 256 static int report__setup_sample_type(struct report *rep)
 257 {
 258         struct perf_session *session = rep->session;
 259         u64 sample_type = perf_evlist__combined_sample_type(session->evlist);
 260         bool is_pipe = perf_data_file__is_pipe(session->file);
 261
 262         if (session->itrace_synth_opts->callchain ||
 263             (!is_pipe &&
 264              perf_header__has_feat(&session->header, HEADER_AUXTRACE) &&
 265              !session->itrace_synth_opts->set))
 266                 sample_type |= PERF_SAMPLE_CALLCHAIN;
 267
 268         if (session->itrace_synth_opts->last_branch)
 269                 sample_type |= PERF_SAMPLE_BRANCH_STACK;
 270
 271         if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
 272                 if (perf_hpp_list.parent) {
 273                         ui__error("Selected --sort parent, but no "
 274                                     "callchain data. Did you call "
 275                                     "'perf record' without -g?\n");
 276                         return -EINVAL;
 277                 }
 278                 if (symbol_conf.use_callchain &&
 279                         !symbol_conf.show_branchflag_count) {
 280                         ui__error("Selected -g or --branch-history.\n"
 281                                   "But no callchain or branch data.\n"
 282                                   "Did you call 'perf record' without -g or -b?\n");
 283                         return -1;
 284                 }
 285         } else if (!callchain_param.enabled &&
 286                    callchain_param.mode != CHAIN_NONE &&
 287                    !symbol_conf.use_callchain) {
 288                         symbol_conf.use_callchain = true;
 289                         if (callchain_register_param(&callchain_param) < 0) {
 290                                 ui__error("Can't register callchain params.\n");
 291                                 return -EINVAL;
 292                         }
 293         }
 294
 295         if (symbol_conf.cumulate_callchain) {
 296                 /* Silently ignore if callchain is missing */
 297                 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
 298                         symbol_conf.cumulate_callchain = false;
 299                         perf_hpp__cancel_cumulate();
 300                 }
 301         }
 302
 303         if (sort__mode == SORT_MODE__BRANCH) {
 304                 if (!is_pipe &&
 305                     !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
 306                         ui__error("Selected -b but no branch data. "
 307                                   "Did you call perf record without -b?\n");
 308                         return -1;
 309                 }
 310         }
 311
 312         if (symbol_conf.use_callchain || symbol_conf.cumulate_callchain) {
 313                 if ((sample_type & PERF_SAMPLE_REGS_USER) &&
 314                     (sample_type & PERF_SAMPLE_STACK_USER))
 315                         callchain_param.record_mode = CALLCHAIN_DWARF;
 316                 else if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 317                         callchain_param.record_mode = CALLCHAIN_LBR;
 318                 else
 319                         callchain_param.record_mode = CALLCHAIN_FP;
 320         }
 321
 322         /* ??? handle more cases than just ANY? */
 323         if (!(perf_evlist__combined_branch_type(session->evlist) &
 324                                 PERF_SAMPLE_BRANCH_ANY))
 325                 rep->nonany_branch_mode = true;
 326
 327         return 0;
 328 }
 329
 330 static void sig_handler(int sig __maybe_unused)
 331 {
 332         session_done = 1;
 333 }
 334
 335 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
 336                                               const char *evname, FILE *fp)
 337 {
 338         size_t ret;
 339         char unit;
 340         unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 341         u64 nr_events = hists->stats.total_period;
 342         struct perf_evsel *evsel = hists_to_evsel(hists);
 343         char buf[512];
 344         size_t size = sizeof(buf);
 345         int socked_id = hists->socket_filter;
 346
 347         if (quiet)
 348                 return 0;
 349
 350         if (symbol_conf.filter_relative) {
 351                 nr_samples = hists->stats.nr_non_filtered_samples;
 352                 nr_events = hists->stats.total_non_filtered_period;
 353         }
 354
 355         if (perf_evsel__is_group_event(evsel)) {
 356                 struct perf_evsel *pos;
 357
 358                 perf_evsel__group_desc(evsel, buf, size);
 359                 evname = buf;
 360
 361                 for_each_group_member(pos, evsel) {
 362                         const struct hists *pos_hists = evsel__hists(pos);
 363
 364                         if (symbol_conf.filter_relative) {
 365                                 nr_samples += pos_hists->stats.nr_non_filtered_samples;
 366                                 nr_events += pos_hists->stats.total_non_filtered_period;
 367                         } else {
 368                                 nr_samples += pos_hists->stats.nr_events[PERF_RECORD_SAMPLE];
 369                                 nr_events += pos_hists->stats.total_period;
 370                         }
 371                 }
 372         }
 373
 374         nr_samples = convert_unit(nr_samples, &unit);
 375         ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
 376         if (evname != NULL)
 377                 ret += fprintf(fp, " of event '%s'", evname);
 378
 379         if (symbol_conf.show_ref_callgraph &&
 380             strstr(evname, "call-graph=no")) {
 381                 ret += fprintf(fp, ", show reference callgraph");
 382         }
 383
 384         if (rep->mem_mode) {
 385                 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
 386                 ret += fprintf(fp, "\n# Sort order   : %s", sort_order ? : default_mem_sort_order);
 387         } else
 388                 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
 389
 390         if (socked_id > -1)
 391                 ret += fprintf(fp, "\n# Processor Socket: %d", socked_id);
 392
 393         return ret + fprintf(fp, "\n#\n");
 394 }
 395
 396 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 397                                          struct report *rep,
 398                                          const char *help)
 399 {
 400         struct perf_evsel *pos;
 401
 402         if (!quiet) {
 403                 fprintf(stdout, "#\n# Total Lost Samples: %" PRIu64 "\n#\n",
 404                         evlist->stats.total_lost_samples);
 405         }
 406
 407         evlist__for_each_entry(evlist, pos) {
 408                 struct hists *hists = evsel__hists(pos);
 409                 const char *evname = perf_evsel__name(pos);
 410
 411                 if (symbol_conf.event_group &&
 412                     !perf_evsel__is_group_leader(pos))
 413                         continue;
 414
 415                 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
 416                 hists__fprintf(hists, !quiet, 0, 0, rep->min_percent, stdout,
 417                                symbol_conf.use_callchain ||
 418                                symbol_conf.show_branchflag_count);
 419                 fprintf(stdout, "\n\n");
 420         }
 421
 422         if (!quiet)
 423                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 424
 425         if (rep->show_threads) {
 426                 bool style = !strcmp(rep->pretty_printing_style, "raw");
 427                 perf_read_values_display(stdout, &rep->show_threads_values,
 428                                          style);
 429                 perf_read_values_destroy(&rep->show_threads_values);
 430         }
 431
 432         if (sort__mode == SORT_MODE__BRANCH)
 433                 branch_type_stat_display(stdout, &rep->brtype_stat);
 434
 435         return 0;
 436 }
 437
 438 static void report__warn_kptr_restrict(const struct report *rep)
 439 {
 440         struct map *kernel_map = machine__kernel_map(&rep->session->machines.host);
 441         struct kmap *kernel_kmap = kernel_map ? map__kmap(kernel_map) : NULL;
 442
 443         if (kernel_map == NULL ||
 444             (kernel_map->dso->hit &&
 445              (kernel_kmap->ref_reloc_sym == NULL ||
 446               kernel_kmap->ref_reloc_sym->addr == 0))) {
 447                 const char *desc =
 448                     "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 449                     "can't be resolved.";
 450
 451                 if (kernel_map) {
 452                         const struct dso *kdso = kernel_map->dso;
 453                         if (!RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION])) {
 454                                 desc = "If some relocation was applied (e.g. "
 455                                        "kexec) symbols may be misresolved.";
 456                         }
 457                 }
 458
 459                 ui__warning(
 460 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 461 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 462 "Samples in kernel modules can't be resolved as well.\n\n",
 463                 desc);
 464         }
 465 }
 466
 467 static int report__gtk_browse_hists(struct report *rep, const char *help)
 468 {
 469         int (*hist_browser)(struct perf_evlist *evlist, const char *help,
 470                             struct hist_browser_timer *timer, float min_pcnt);
 471
 472         hist_browser = dlsym(perf_gtk_handle, "perf_evlist__gtk_browse_hists");
 473
 474         if (hist_browser == NULL) {
 475                 ui__error("GTK browser not found!\n");
 476                 return -1;
 477         }
 478
 479         return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
 480 }
 481
 482 static int report__browse_hists(struct report *rep)
 483 {
 484         int ret;
 485         struct perf_session *session = rep->session;
 486         struct perf_evlist *evlist = session->evlist;
 487         const char *help = perf_tip(system_path(TIPDIR));
 488
 489         if (help == NULL) {
 490                 /* fallback for people who don't install perf ;-) */
 491                 help = perf_tip(DOCDIR);
 492                 if (help == NULL)
 493                         help = "Cannot load tips.txt file, please install perf!";
 494         }
 495
 496         switch (use_browser) {
 497         case 1:
 498                 ret = perf_evlist__tui_browse_hists(evlist, help, NULL,
 499                                                     rep->min_percent,
 500                                                     &session->header.env);
 501                 /*
 502                  * Usually "ret" is the last pressed key, and we only
 503                  * care if the key notifies us to switch data file.
 504                  */
 505                 if (ret != K_SWITCH_INPUT_DATA)
 506                         ret = 0;
 507                 break;
 508         case 2:
 509                 ret = report__gtk_browse_hists(rep, help);
 510                 break;
 511         default:
 512                 ret = perf_evlist__tty_browse_hists(evlist, rep, help);
 513                 break;
 514         }
 515
 516         return ret;
 517 }
 518
 519 static int report__collapse_hists(struct report *rep)
 520 {
 521         struct ui_progress prog;
 522         struct perf_evsel *pos;
 523         int ret = 0;
 524
 525         ui_progress__init(&prog, rep->nr_entries, "Merging related events...");
 526
 527         evlist__for_each_entry(rep->session->evlist, pos) {
 528                 struct hists *hists = evsel__hists(pos);
 529
 530                 if (pos->idx == 0)
 531                         hists->symbol_filter_str = rep->symbol_filter_str;
 532
 533                 hists->socket_filter = rep->socket_filter;
 534
 535                 ret = hists__collapse_resort(hists, &prog);
 536                 if (ret < 0)
 537                         break;
 538
 539                 /* Non-group events are considered as leader */
 540                 if (symbol_conf.event_group &&
 541                     !perf_evsel__is_group_leader(pos)) {
 542                         struct hists *leader_hists = evsel__hists(pos->leader);
 543
 544                         hists__match(leader_hists, hists);
 545                         hists__link(leader_hists, hists);
 546                 }
 547         }
 548
 549         ui_progress__finish();
 550         return ret;
 551 }
 552
 553 static void report__output_resort(struct report *rep)
 554 {
 555         struct ui_progress prog;
 556         struct perf_evsel *pos;
 557
 558         ui_progress__init(&prog, rep->nr_entries, "Sorting events for output...");
 559
 560         evlist__for_each_entry(rep->session->evlist, pos)
 561                 perf_evsel__output_resort(pos, &prog);
 562
 563         ui_progress__finish();
 564 }
 565
 566 static int __cmd_report(struct report *rep)
 567 {
 568         int ret;
 569         struct perf_session *session = rep->session;
 570         struct perf_evsel *pos;
 571         struct perf_data_file *file = session->file;
 572
 573         signal(SIGINT, sig_handler);
 574
 575         if (rep->cpu_list) {
 576                 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
 577                                                rep->cpu_bitmap);
 578                 if (ret) {
 579                         ui__error("failed to set cpu bitmap\n");
 580                         return ret;
 581                 }
 582                 session->itrace_synth_opts->cpu_bitmap = rep->cpu_bitmap;
 583         }
 584
 585         if (rep->show_threads) {
 586                 ret = perf_read_values_init(&rep->show_threads_values);
 587                 if (ret)
 588                         return ret;
 589         }
 590
 591         ret = report__setup_sample_type(rep);
 592         if (ret) {
 593                 /* report__setup_sample_type() already showed error message */
 594                 return ret;
 595         }
 596
 597         ret = perf_session__process_events(session);
 598         if (ret) {
 599                 ui__error("failed to process sample\n");
 600                 return ret;
 601         }
 602
 603         report__warn_kptr_restrict(rep);
 604
 605         evlist__for_each_entry(session->evlist, pos)
 606                 rep->nr_entries += evsel__hists(pos)->nr_entries;
 607
 608         if (use_browser == 0) {
 609                 if (verbose > 3)
 610                         perf_session__fprintf(session, stdout);
 611
 612                 if (verbose > 2)
 613                         perf_session__fprintf_dsos(session, stdout);
 614
 615                 if (dump_trace) {
 616                         perf_session__fprintf_nr_events(session, stdout);
 617                         perf_evlist__fprintf_nr_events(session->evlist, stdout);
 618                         return 0;
 619                 }
 620         }
 621
 622         ret = report__collapse_hists(rep);
 623         if (ret) {
 624                 ui__error("failed to process hist entry\n");
 625                 return ret;
 626         }
 627
 628         if (session_done())
 629                 return 0;
 630
 631         /*
 632          * recalculate number of entries after collapsing since it
 633          * might be changed during the collapse phase.
 634          */
 635         rep->nr_entries = 0;
 636         evlist__for_each_entry(session->evlist, pos)
 637                 rep->nr_entries += evsel__hists(pos)->nr_entries;
 638
 639         if (rep->nr_entries == 0) {
 640                 ui__error("The %s file has no samples!\n", file->path);
 641                 return 0;
 642         }
 643
 644         report__output_resort(rep);
 645
 646         return report__browse_hists(rep);
 647 }
 648
 649 static int
 650 report_parse_callchain_opt(const struct option *opt, const char *arg, int unset)
 651 {
 652         struct callchain_param *callchain = opt->value;
 653
 654         callchain->enabled = !unset;
 655         /*
 656          * --no-call-graph
 657          */
 658         if (unset) {
 659                 symbol_conf.use_callchain = false;
 660                 callchain->mode = CHAIN_NONE;
 661                 return 0;
 662         }
 663
 664         return parse_callchain_report_opt(arg);
 665 }
 666
 667 int
 668 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
 669                                 const char *arg, int unset __maybe_unused)
 670 {
 671         if (arg) {
 672                 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
 673                 if (err) {
 674                         char buf[BUFSIZ];
 675                         regerror(err, &ignore_callees_regex, buf, sizeof(buf));
 676                         pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
 677                         return -1;
 678                 }
 679                 have_ignore_callees = 1;
 680         }
 681
 682         return 0;
 683 }
 684
 685 static int
 686 parse_branch_mode(const struct option *opt,
 687                   const char *str __maybe_unused, int unset)
 688 {
 689         int *branch_mode = opt->value;
 690
 691         *branch_mode = !unset;
 692         return 0;
 693 }
 694
 695 static int
 696 parse_percent_limit(const struct option *opt, const char *str,
 697                     int unset __maybe_unused)
 698 {
 699         struct report *rep = opt->value;
 700         double pcnt = strtof(str, NULL);
 701
 702         rep->min_percent = pcnt;
 703         callchain_param.min_percent = pcnt;
 704         return 0;
 705 }
 706
 707 #define CALLCHAIN_DEFAULT_OPT  "graph,0.5,caller,function,percent"
 708
 709 const char report_callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
 710                                      CALLCHAIN_REPORT_HELP
 711                                      "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
 712
 713 int cmd_report(int argc, const char **argv)
 714 {
 715         struct perf_session *session;
 716         struct itrace_synth_opts itrace_synth_opts = { .set = 0, };
 717         struct stat st;
 718         bool has_br_stack = false;
 719         int branch_mode = -1;
 720         bool branch_call_mode = false;
 721         char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
 722         const char * const report_usage[] = {
 723                 "perf report [<options>]",
 724                 NULL
 725         };
 726         struct report report = {
 727                 .tool = {
 728                         .sample          = process_sample_event,
 729                         .mmap            = perf_event__process_mmap,
 730                         .mmap2           = perf_event__process_mmap2,
 731                         .comm            = perf_event__process_comm,
 732                         .namespaces      = perf_event__process_namespaces,
 733                         .exit            = perf_event__process_exit,
 734                         .fork            = perf_event__process_fork,
 735                         .lost            = perf_event__process_lost,
 736                         .read            = process_read_event,
 737                         .attr            = perf_event__process_attr,
 738                         .tracing_data    = perf_event__process_tracing_data,
 739                         .build_id        = perf_event__process_build_id,
 740                         .id_index        = perf_event__process_id_index,
 741                         .auxtrace_info   = perf_event__process_auxtrace_info,
 742                         .auxtrace        = perf_event__process_auxtrace,
 743                         .feature         = perf_event__process_feature,
 744                         .ordered_events  = true,
 745                         .ordering_requires_timestamps = true,
 746                 },
 747                 .max_stack               = PERF_MAX_STACK_DEPTH,
 748                 .pretty_printing_style   = "normal",
 749                 .socket_filter           = -1,
 750         };
 751         const struct option options[] = {
 752         OPT_STRING('i', "input", &input_name, "file",
 753                     "input file name"),
 754         OPT_INCR('v', "verbose", &verbose,
 755                     "be more verbose (show symbol address, etc)"),
 756         OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
 757         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 758                     "dump raw trace in ASCII"),
 759         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 760                    "file", "vmlinux pathname"),
 761         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 762                    "file", "kallsyms pathname"),
 763         OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
 764         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 765                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 766         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 767                     "Show a column with the number of samples"),
 768         OPT_BOOLEAN('T', "threads", &report.show_threads,
 769                     "Show per-thread event counters"),
 770         OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
 771                    "pretty printing style key: normal raw"),
 772         OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
 773         OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
 774         OPT_BOOLEAN(0, "stdio", &report.use_stdio,
 775                     "Use the stdio interface"),
 776         OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
 777         OPT_BOOLEAN(0, "header-only", &report.header_only,
 778                     "Show only data header."),
 779         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 780                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
 781                    " Please refer the man page for the complete list."),
 782         OPT_STRING('F', "fields", &field_order, "key[,keys...]",
 783                    "output field(s): overhead, period, sample plus all of sort keys"),
 784         OPT_BOOLEAN(0, "show-cpu-utilization", &symbol_conf.show_cpu_utilization,
 785                     "Show sample percentage for different cpu modes"),
 786         OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 787                     "Show sample percentage for different cpu modes", PARSE_OPT_HIDDEN),
 788         OPT_STRING('p', "parent", &parent_pattern, "regex",
 789                    "regex filter to identify parent, see: '--sort parent'"),
 790         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 791                     "Only display entries with parent-match"),
 792         OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
 793                              "print_type,threshold[,print_limit],order,sort_key[,branch],value",
 794                              report_callchain_help, &report_parse_callchain_opt,
 795                              callchain_default_opt),
 796         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
 797                     "Accumulate callchains of children and show total overhead as well"),
 798         OPT_INTEGER(0, "max-stack", &report.max_stack,
 799                     "Set the maximum stack depth when parsing the callchain, "
 800                     "anything beyond the specified depth will be ignored. "
 801                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
 802         OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
 803                     "alias for inverted call graph"),
 804         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
 805                    "ignore callees of these functions in call graphs",
 806                    report_parse_ignore_callees_opt),
 807         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 808                    "only consider symbols in these dsos"),
 809         OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 810                    "only consider symbols in these comms"),
 811         OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
 812                    "only consider symbols in these pids"),
 813         OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
 814                    "only consider symbols in these tids"),
 815         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 816                    "only consider these symbols"),
 817         OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
 818                    "only show symbols that (partially) match with this filter"),
 819         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 820                    "width[,width...]",
 821                    "don't try to adjust column width, use these fixed values"),
 822         OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
 823                    "separator for columns, no spaces will be added between "
 824                    "columns '.' is reserved."),
 825         OPT_BOOLEAN('U', "hide-unresolved", &symbol_conf.hide_unresolved,
 826                     "Only display entries resolved to a symbol"),
 827         OPT_CALLBACK(0, "symfs", NULL, "directory",
 828                      "Look for files with symbols relative to this directory",
 829                      symbol__config_symfs),
 830         OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
 831                    "list of cpus to profile"),
 832         OPT_BOOLEAN('I', "show-info", &report.show_full_info,
 833                     "Display extended information about perf.data file"),
 834         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 835                     "Interleave source code with assembly code (default)"),
 836         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 837                     "Display raw encoding of assembly instructions (default)"),
 838         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 839                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 840         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 841                     "Show a column with the sum of periods"),
 842         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 843                     "Show event group information together"),
 844         OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
 845                     "use branch records for per branch histogram filling",
 846                     parse_branch_mode),
 847         OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
 848                     "add last branch records to call history"),
 849         OPT_STRING(0, "objdump", &objdump_path, "path",
 850                    "objdump binary to use for disassembly and annotations"),
 851         OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
 852                     "Disable symbol demangling"),
 853         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
 854                     "Enable kernel symbol demangling"),
 855         OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
 856         OPT_CALLBACK(0, "percent-limit", &report, "percent",
 857                      "Don't show entries under that percent", parse_percent_limit),
 858         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
 859                      "how to display percentage of filtered entries", parse_filter_percentage),
 860         OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
 861                             "Instruction Tracing options",
 862                             itrace_parse_synth_opts),
 863         OPT_BOOLEAN(0, "full-source-path", &srcline_full_filename,
 864                         "Show full source file name path for source lines"),
 865         OPT_BOOLEAN(0, "show-ref-call-graph", &symbol_conf.show_ref_callgraph,
 866                     "Show callgraph from reference event"),
 867         OPT_INTEGER(0, "socket-filter", &report.socket_filter,
 868                     "only show processor socket that match with this filter"),
 869         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
 870                     "Show raw trace event output (do not use print fmt or plugins)"),
 871         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
 872                     "Show entries in a hierarchy"),
 873         OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
 874                              "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
 875                              stdio__config_color, "always"),
 876         OPT_STRING(0, "time", &report.time_str, "str",
 877                    "Time span of interest (start,stop)"),
 878         OPT_BOOLEAN(0, "inline", &symbol_conf.inline_name,
 879                     "Show inline function"),
 880         OPT_END()
 881         };
 882         struct perf_data_file file = {
 883                 .mode  = PERF_DATA_MODE_READ,
 884         };
 885         int ret = hists__init();
 886
 887         if (ret < 0)
 888                 return ret;
 889
 890         ret = perf_config(report__config, &report);
 891         if (ret)
 892                 return ret;
 893
 894         argc = parse_options(argc, argv, options, report_usage, 0);
 895         if (argc) {
 896                 /*
 897                  * Special case: if there's an argument left then assume that
 898                  * it's a symbol filter:
 899                  */
 900                 if (argc > 1)
 901                         usage_with_options(report_usage, options);
 902
 903                 report.symbol_filter_str = argv[0];
 904         }
 905
 906         if (quiet)
 907                 perf_quiet_option();
 908
 909         if (symbol_conf.vmlinux_name &&
 910             access(symbol_conf.vmlinux_name, R_OK)) {
 911                 pr_err("Invalid file: %s\n", symbol_conf.vmlinux_name);
 912                 return -EINVAL;
 913         }
 914         if (symbol_conf.kallsyms_name &&
 915             access(symbol_conf.kallsyms_name, R_OK)) {
 916                 pr_err("Invalid file: %s\n", symbol_conf.kallsyms_name);
 917                 return -EINVAL;
 918         }
 919
 920         if (report.use_stdio)
 921                 use_browser = 0;
 922         else if (report.use_tui)
 923                 use_browser = 1;
 924         else if (report.use_gtk)
 925                 use_browser = 2;
 926
 927         if (report.inverted_callchain)
 928                 callchain_param.order = ORDER_CALLER;
 929         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
 930                 callchain_param.order = ORDER_CALLER;
 931
 932         if (itrace_synth_opts.callchain &&
 933             (int)itrace_synth_opts.callchain_sz > report.max_stack)
 934                 report.max_stack = itrace_synth_opts.callchain_sz;
 935
 936         if (!input_name || !strlen(input_name)) {
 937                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 938                         input_name = "-";
 939                 else
 940                         input_name = "perf.data";
 941         }
 942
 943         file.path  = input_name;
 944         file.force = symbol_conf.force;
 945
 946 repeat:
 947         session = perf_session__new(&file, false, &report.tool);
 948         if (session == NULL)
 949                 return -1;
 950
 951         if (report.queue_size) {
 952                 ordered_events__set_alloc_size(&session->ordered_events,
 953                                                report.queue_size);
 954         }
 955
 956         session->itrace_synth_opts = &itrace_synth_opts;
 957
 958         report.session = session;
 959
 960         has_br_stack = perf_header__has_feat(&session->header,
 961                                              HEADER_BRANCH_STACK);
 962
 963         if (itrace_synth_opts.last_branch)
 964                 has_br_stack = true;
 965
 966         if (has_br_stack && branch_call_mode)
 967                 symbol_conf.show_branchflag_count = true;
 968
 969         memset(&report.brtype_stat, 0, sizeof(struct branch_type_stat));
 970
 971         /*
 972          * Branch mode is a tristate:
 973          * -1 means default, so decide based on the file having branch data.
 974          * 0/1 means the user chose a mode.
 975          */
 976         if (((branch_mode == -1 && has_br_stack) || branch_mode == 1) &&
 977             !branch_call_mode) {
 978                 sort__mode = SORT_MODE__BRANCH;
 979                 symbol_conf.cumulate_callchain = false;
 980         }
 981         if (branch_call_mode) {
 982                 callchain_param.key = CCKEY_ADDRESS;
 983                 callchain_param.branch_callstack = 1;
 984                 symbol_conf.use_callchain = true;
 985                 callchain_register_param(&callchain_param);
 986                 if (sort_order == NULL)
 987                         sort_order = "srcline,symbol,dso";
 988         }
 989
 990         if (report.mem_mode) {
 991                 if (sort__mode == SORT_MODE__BRANCH) {
 992                         pr_err("branch and mem mode incompatible\n");
 993                         goto error;
 994                 }
 995                 sort__mode = SORT_MODE__MEMORY;
 996                 symbol_conf.cumulate_callchain = false;
 997         }
 998
 999         if (symbol_conf.report_hierarchy) {
1000                 /* disable incompatible options */
1001                 symbol_conf.cumulate_callchain = false;
1002
1003                 if (field_order) {
1004                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1005                         parse_options_usage(report_usage, options, "F", 1);
1006                         parse_options_usage(NULL, options, "hierarchy", 0);
1007                         goto error;
1008                 }
1009
1010                 perf_hpp_list.need_collapse = true;
1011         }
1012
1013         /* Force tty output for header output and per-thread stat. */
1014         if (report.header || report.header_only || report.show_threads)
1015                 use_browser = 0;
1016         if (report.header || report.header_only)
1017                 report.tool.show_feat_hdr = SHOW_FEAT_HEADER;
1018         if (report.show_full_info)
1019                 report.tool.show_feat_hdr = SHOW_FEAT_HEADER_FULL_INFO;
1020
1021         if (strcmp(input_name, "-") != 0)
1022                 setup_browser(true);
1023         else
1024                 use_browser = 0;
1025
1026         if (setup_sorting(session->evlist) < 0) {
1027                 if (sort_order)
1028                         parse_options_usage(report_usage, options, "s", 1);
1029                 if (field_order)
1030                         parse_options_usage(sort_order ? NULL : report_usage,
1031                                             options, "F", 1);
1032                 goto error;
1033         }
1034
1035         if ((report.header || report.header_only) && !quiet) {
1036                 perf_session__fprintf_info(session, stdout,
1037                                            report.show_full_info);
1038                 if (report.header_only) {
1039                         ret = 0;
1040                         goto error;
1041                 }
1042         } else if (use_browser == 0 && !quiet) {
1043                 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
1044                       stdout);
1045         }
1046
1047         /*
1048          * Only in the TUI browser we are doing integrated annotation,
1049          * so don't allocate extra space that won't be used in the stdio
1050          * implementation.
1051          */
1052         if (ui__has_annotation()) {
1053                 ret = symbol__annotation_init();
1054                 if (ret < 0)
1055                         goto error;
1056                 /*
1057                  * For searching by name on the "Browse map details".
1058                  * providing it only in verbose mode not to bloat too
1059                  * much struct symbol.
1060                  */
1061                 if (verbose > 0) {
1062                         /*
1063                          * XXX: Need to provide a less kludgy way to ask for
1064                          * more space per symbol, the u32 is for the index on
1065                          * the ui browser.
1066                          * See symbol__browser_index.
1067                          */
1068                         symbol_conf.priv_size += sizeof(u32);
1069                         symbol_conf.sort_by_name = true;
1070                 }
1071         }
1072
1073         if (symbol__init(&session->header.env) < 0)
1074                 goto error;
1075
1076         if (perf_time__parse_str(&report.ptime, report.time_str) != 0) {
1077                 pr_err("Invalid time string\n");
1078                 return -EINVAL;
1079         }
1080
1081         sort__setup_elide(stdout);
1082
1083         ret = __cmd_report(&report);
1084         if (ret == K_SWITCH_INPUT_DATA) {
1085                 perf_session__delete(session);
1086                 goto repeat;
1087         } else
1088                 ret = 0;
1089
1090 error:
1091         perf_session__delete(session);
1092         return ret;
1093 }