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/cache.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/strlist.h"
  20 #include "util/values.h"
  21
  22 #include "perf.h"
  23 #include "util/debug.h"
  24 #include "util/evlist.h"
  25 #include "util/evsel.h"
  26 #include "util/header.h"
  27 #include "util/session.h"
  28 #include "util/tool.h"
  29
  30 #include "util/parse-options.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
  39 #include <dlfcn.h>
  40 #include <linux/bitmap.h>
  41
  42 struct report {
  43         struct perf_tool        tool;
  44         struct perf_session     *session;
  45         bool                    force, use_tui, use_gtk, use_stdio;
  46         bool                    hide_unresolved;
  47         bool                    dont_use_callchains;
  48         bool                    show_full_info;
  49         bool                    show_threads;
  50         bool                    inverted_callchain;
  51         bool                    mem_mode;
  52         bool                    header;
  53         bool                    header_only;
  54         int                     max_stack;
  55         struct perf_read_values show_threads_values;
  56         const char              *pretty_printing_style;
  57         const char              *cpu_list;
  58         const char              *symbol_filter_str;
  59         float                   min_percent;
  60         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  61 };
  62
  63 static int report__config(const char *var, const char *value, void *cb)
  64 {
  65         if (!strcmp(var, "report.group")) {
  66                 symbol_conf.event_group = perf_config_bool(var, value);
  67                 return 0;
  68         }
  69         if (!strcmp(var, "report.percent-limit")) {
  70                 struct report *rep = cb;
  71                 rep->min_percent = strtof(value, NULL);
  72                 return 0;
  73         }
  74
  75         return perf_default_config(var, value, cb);
  76 }
  77
  78 static int report__resolve_callchain(struct report *rep, struct symbol **parent,
  79                                      struct perf_evsel *evsel, struct addr_location *al,
  80                                      struct perf_sample *sample, struct machine *machine)
  81 {
  82         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
  83                 return machine__resolve_callchain(machine, evsel, al->thread, sample,
  84                                                   parent, al, rep->max_stack);
  85         }
  86         return 0;
  87 }
  88
  89 static int hist_entry__append_callchain(struct hist_entry *he, struct perf_sample *sample)
  90 {
  91         if (!symbol_conf.use_callchain)
  92                 return 0;
  93         return callchain_append(he->callchain, &callchain_cursor, sample->period);
  94 }
  95
  96 static int report__add_mem_hist_entry(struct perf_tool *tool, struct addr_location *al,
  97                                       struct perf_sample *sample, struct perf_evsel *evsel,
  98                                       struct machine *machine, union perf_event *event)
  99 {
 100         struct report *rep = container_of(tool, struct report, tool);
 101         struct symbol *parent = NULL;
 102         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 103         struct hist_entry *he;
 104         struct mem_info *mi, *mx;
 105         uint64_t cost;
 106         int err = report__resolve_callchain(rep, &parent, evsel, al, sample, machine);
 107
 108         if (err)
 109                 return err;
 110
 111         mi = machine__resolve_mem(machine, al->thread, sample, cpumode);
 112         if (!mi)
 113                 return -ENOMEM;
 114
 115         if (rep->hide_unresolved && !al->sym)
 116                 return 0;
 117
 118         cost = sample->weight;
 119         if (!cost)
 120                 cost = 1;
 121
 122         /*
 123          * must pass period=weight in order to get the correct
 124          * sorting from hists__collapse_resort() which is solely
 125          * based on periods. We want sorting be done on nr_events * weight
 126          * and this is indirectly achieved by passing period=weight here
 127          * and the he_stat__add_period() function.
 128          */
 129         he = __hists__add_entry(&evsel->hists, al, parent, NULL, mi,
 130                                 cost, cost, 0);
 131         if (!he)
 132                 return -ENOMEM;
 133
 134         err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 135         if (err)
 136                 goto out;
 137
 138         mx = he->mem_info;
 139         err = addr_map_symbol__inc_samples(&mx->daddr, evsel->idx);
 140         if (err)
 141                 goto out;
 142
 143         evsel->hists.stats.total_period += cost;
 144         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 145         err = hist_entry__append_callchain(he, sample);
 146 out:
 147         return err;
 148 }
 149
 150 static int report__add_branch_hist_entry(struct perf_tool *tool, struct addr_location *al,
 151                                          struct perf_sample *sample, struct perf_evsel *evsel,
 152                                          struct machine *machine)
 153 {
 154         struct report *rep = container_of(tool, struct report, tool);
 155         struct symbol *parent = NULL;
 156         unsigned i;
 157         struct hist_entry *he;
 158         struct branch_info *bi, *bx;
 159         int err = report__resolve_callchain(rep, &parent, evsel, al, sample, machine);
 160
 161         if (err)
 162                 return err;
 163
 164         bi = machine__resolve_bstack(machine, al->thread,
 165                                      sample->branch_stack);
 166         if (!bi)
 167                 return -ENOMEM;
 168
 169         for (i = 0; i < sample->branch_stack->nr; i++) {
 170                 if (rep->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
 171                         continue;
 172
 173                 err = -ENOMEM;
 174
 175                 /* overwrite the 'al' to branch-to info */
 176                 al->map = bi[i].to.map;
 177                 al->sym = bi[i].to.sym;
 178                 al->addr = bi[i].to.addr;
 179                 /*
 180                  * The report shows the percentage of total branches captured
 181                  * and not events sampled. Thus we use a pseudo period of 1.
 182                  */
 183                 he = __hists__add_entry(&evsel->hists, al, parent, &bi[i], NULL,
 184                                         1, 1, 0);
 185                 if (he) {
 186                         bx = he->branch_info;
 187                         err = addr_map_symbol__inc_samples(&bx->from, evsel->idx);
 188                         if (err)
 189                                 goto out;
 190
 191                         err = addr_map_symbol__inc_samples(&bx->to, evsel->idx);
 192                         if (err)
 193                                 goto out;
 194
 195                         evsel->hists.stats.total_period += 1;
 196                         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 197                 } else
 198                         goto out;
 199         }
 200         err = 0;
 201 out:
 202         free(bi);
 203         return err;
 204 }
 205
 206 static int report__add_hist_entry(struct perf_tool *tool, struct perf_evsel *evsel,
 207                                   struct addr_location *al, struct perf_sample *sample,
 208                                   struct machine *machine)
 209 {
 210         struct report *rep = container_of(tool, struct report, tool);
 211         struct symbol *parent = NULL;
 212         struct hist_entry *he;
 213         int err = report__resolve_callchain(rep, &parent, evsel, al, sample, machine);
 214
 215         if (err)
 216                 return err;
 217
 218         he = __hists__add_entry(&evsel->hists, al, parent, NULL, NULL,
 219                                 sample->period, sample->weight,
 220                                 sample->transaction);
 221         if (he == NULL)
 222                 return -ENOMEM;
 223
 224         err = hist_entry__append_callchain(he, sample);
 225         if (err)
 226                 goto out;
 227
 228         err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 229         evsel->hists.stats.total_period += sample->period;
 230         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 231 out:
 232         return err;
 233 }
 234
 235
 236 static int process_sample_event(struct perf_tool *tool,
 237                                 union perf_event *event,
 238                                 struct perf_sample *sample,
 239                                 struct perf_evsel *evsel,
 240                                 struct machine *machine)
 241 {
 242         struct report *rep = container_of(tool, struct report, tool);
 243         struct addr_location al;
 244         int ret;
 245
 246         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
 247                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 248                         event->header.type);
 249                 return -1;
 250         }
 251
 252         if (al.filtered || (rep->hide_unresolved && al.sym == NULL))
 253                 return 0;
 254
 255         if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
 256                 return 0;
 257
 258         if (sort__mode == SORT_MODE__BRANCH) {
 259                 ret = report__add_branch_hist_entry(tool, &al, sample, evsel, machine);
 260                 if (ret < 0)
 261                         pr_debug("problem adding lbr entry, skipping event\n");
 262         } else if (rep->mem_mode == 1) {
 263                 ret = report__add_mem_hist_entry(tool, &al, sample, evsel, machine, event);
 264                 if (ret < 0)
 265                         pr_debug("problem adding mem entry, skipping event\n");
 266         } else {
 267                 if (al.map != NULL)
 268                         al.map->dso->hit = 1;
 269
 270                 ret = report__add_hist_entry(tool, evsel, &al, sample, machine);
 271                 if (ret < 0)
 272                         pr_debug("problem incrementing symbol period, skipping event\n");
 273         }
 274         return ret;
 275 }
 276
 277 static int process_read_event(struct perf_tool *tool,
 278                               union perf_event *event,
 279                               struct perf_sample *sample __maybe_unused,
 280                               struct perf_evsel *evsel,
 281                               struct machine *machine __maybe_unused)
 282 {
 283         struct report *rep = container_of(tool, struct report, tool);
 284
 285         if (rep->show_threads) {
 286                 const char *name = evsel ? perf_evsel__name(evsel) : "unknown";
 287                 perf_read_values_add_value(&rep->show_threads_values,
 288                                            event->read.pid, event->read.tid,
 289                                            event->read.id,
 290                                            name,
 291                                            event->read.value);
 292         }
 293
 294         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 295                     evsel ? perf_evsel__name(evsel) : "FAIL",
 296                     event->read.value);
 297
 298         return 0;
 299 }
 300
 301 /* For pipe mode, sample_type is not currently set */
 302 static int report__setup_sample_type(struct report *rep)
 303 {
 304         struct perf_session *session = rep->session;
 305         u64 sample_type = perf_evlist__combined_sample_type(session->evlist);
 306         bool is_pipe = perf_data_file__is_pipe(session->file);
 307
 308         if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
 309                 if (sort__has_parent) {
 310                         ui__error("Selected --sort parent, but no "
 311                                     "callchain data. Did you call "
 312                                     "'perf record' without -g?\n");
 313                         return -EINVAL;
 314                 }
 315                 if (symbol_conf.use_callchain) {
 316                         ui__error("Selected -g but no callchain data. Did "
 317                                     "you call 'perf record' without -g?\n");
 318                         return -1;
 319                 }
 320         } else if (!rep->dont_use_callchains &&
 321                    callchain_param.mode != CHAIN_NONE &&
 322                    !symbol_conf.use_callchain) {
 323                         symbol_conf.use_callchain = true;
 324                         if (callchain_register_param(&callchain_param) < 0) {
 325                                 ui__error("Can't register callchain params.\n");
 326                                 return -EINVAL;
 327                         }
 328         }
 329
 330         if (sort__mode == SORT_MODE__BRANCH) {
 331                 if (!is_pipe &&
 332                     !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
 333                         ui__error("Selected -b but no branch data. "
 334                                   "Did you call perf record without -b?\n");
 335                         return -1;
 336                 }
 337         }
 338
 339         return 0;
 340 }
 341
 342 static void sig_handler(int sig __maybe_unused)
 343 {
 344         session_done = 1;
 345 }
 346
 347 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
 348                                               const char *evname, FILE *fp)
 349 {
 350         size_t ret;
 351         char unit;
 352         unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
 353         u64 nr_events = hists->stats.total_period;
 354         struct perf_evsel *evsel = hists_to_evsel(hists);
 355         char buf[512];
 356         size_t size = sizeof(buf);
 357
 358         if (perf_evsel__is_group_event(evsel)) {
 359                 struct perf_evsel *pos;
 360
 361                 perf_evsel__group_desc(evsel, buf, size);
 362                 evname = buf;
 363
 364                 for_each_group_member(pos, evsel) {
 365                         nr_samples += pos->hists.stats.nr_events[PERF_RECORD_SAMPLE];
 366                         nr_events += pos->hists.stats.total_period;
 367                 }
 368         }
 369
 370         nr_samples = convert_unit(nr_samples, &unit);
 371         ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
 372         if (evname != NULL)
 373                 ret += fprintf(fp, " of event '%s'", evname);
 374
 375         if (rep->mem_mode) {
 376                 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
 377                 ret += fprintf(fp, "\n# Sort order   : %s", sort_order);
 378         } else
 379                 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
 380         return ret + fprintf(fp, "\n#\n");
 381 }
 382
 383 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 384                                          struct report *rep,
 385                                          const char *help)
 386 {
 387         struct perf_evsel *pos;
 388
 389         list_for_each_entry(pos, &evlist->entries, node) {
 390                 struct hists *hists = &pos->hists;
 391                 const char *evname = perf_evsel__name(pos);
 392
 393                 if (symbol_conf.event_group &&
 394                     !perf_evsel__is_group_leader(pos))
 395                         continue;
 396
 397                 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
 398                 hists__fprintf(hists, true, 0, 0, rep->min_percent, stdout);
 399                 fprintf(stdout, "\n\n");
 400         }
 401
 402         if (sort_order == default_sort_order &&
 403             parent_pattern == default_parent_pattern) {
 404                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 405
 406                 if (rep->show_threads) {
 407                         bool style = !strcmp(rep->pretty_printing_style, "raw");
 408                         perf_read_values_display(stdout, &rep->show_threads_values,
 409                                                  style);
 410                         perf_read_values_destroy(&rep->show_threads_values);
 411                 }
 412         }
 413
 414         return 0;
 415 }
 416
 417 static int __cmd_report(struct report *rep)
 418 {
 419         int ret = -EINVAL;
 420         u64 nr_samples;
 421         struct perf_session *session = rep->session;
 422         struct perf_evsel *pos;
 423         struct map *kernel_map;
 424         struct kmap *kernel_kmap;
 425         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 426         struct ui_progress prog;
 427         struct perf_data_file *file = session->file;
 428
 429         signal(SIGINT, sig_handler);
 430
 431         if (rep->cpu_list) {
 432                 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
 433                                                rep->cpu_bitmap);
 434                 if (ret)
 435                         return ret;
 436         }
 437
 438         if (rep->show_threads)
 439                 perf_read_values_init(&rep->show_threads_values);
 440
 441         ret = report__setup_sample_type(rep);
 442         if (ret)
 443                 return ret;
 444
 445         ret = perf_session__process_events(session, &rep->tool);
 446         if (ret)
 447                 return ret;
 448
 449         kernel_map = session->machines.host.vmlinux_maps[MAP__FUNCTION];
 450         kernel_kmap = map__kmap(kernel_map);
 451         if (kernel_map == NULL ||
 452             (kernel_map->dso->hit &&
 453              (kernel_kmap->ref_reloc_sym == NULL ||
 454               kernel_kmap->ref_reloc_sym->addr == 0))) {
 455                 const char *desc =
 456                     "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 457                     "can't be resolved.";
 458
 459                 if (kernel_map) {
 460                         const struct dso *kdso = kernel_map->dso;
 461                         if (!RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION])) {
 462                                 desc = "If some relocation was applied (e.g. "
 463                                        "kexec) symbols may be misresolved.";
 464                         }
 465                 }
 466
 467                 ui__warning(
 468 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 469 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 470 "Samples in kernel modules can't be resolved as well.\n\n",
 471                 desc);
 472         }
 473
 474         if (verbose > 3)
 475                 perf_session__fprintf(session, stdout);
 476
 477         if (verbose > 2)
 478                 perf_session__fprintf_dsos(session, stdout);
 479
 480         if (dump_trace) {
 481                 perf_session__fprintf_nr_events(session, stdout);
 482                 return 0;
 483         }
 484
 485         nr_samples = 0;
 486         list_for_each_entry(pos, &session->evlist->entries, node)
 487                 nr_samples += pos->hists.nr_entries;
 488
 489         ui_progress__init(&prog, nr_samples, "Merging related events...");
 490
 491         nr_samples = 0;
 492         list_for_each_entry(pos, &session->evlist->entries, node) {
 493                 struct hists *hists = &pos->hists;
 494
 495                 if (pos->idx == 0)
 496                         hists->symbol_filter_str = rep->symbol_filter_str;
 497
 498                 hists__collapse_resort(hists, &prog);
 499                 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
 500
 501                 /* Non-group events are considered as leader */
 502                 if (symbol_conf.event_group &&
 503                     !perf_evsel__is_group_leader(pos)) {
 504                         struct hists *leader_hists = &pos->leader->hists;
 505
 506                         hists__match(leader_hists, hists);
 507                         hists__link(leader_hists, hists);
 508                 }
 509         }
 510         ui_progress__finish();
 511
 512         if (session_done())
 513                 return 0;
 514
 515         if (nr_samples == 0) {
 516                 ui__error("The %s file has no samples!\n", file->path);
 517                 return 0;
 518         }
 519
 520         list_for_each_entry(pos, &session->evlist->entries, node)
 521                 hists__output_resort(&pos->hists);
 522
 523         if (use_browser > 0) {
 524                 if (use_browser == 1) {
 525                         ret = perf_evlist__tui_browse_hists(session->evlist,
 526                                                         help, NULL,
 527                                                         rep->min_percent,
 528                                                         &session->header.env);
 529                         /*
 530                          * Usually "ret" is the last pressed key, and we only
 531                          * care if the key notifies us to switch data file.
 532                          */
 533                         if (ret != K_SWITCH_INPUT_DATA)
 534                                 ret = 0;
 535
 536                 } else if (use_browser == 2) {
 537                         int (*hist_browser)(struct perf_evlist *,
 538                                             const char *,
 539                                             struct hist_browser_timer *,
 540                                             float min_pcnt);
 541
 542                         hist_browser = dlsym(perf_gtk_handle,
 543                                              "perf_evlist__gtk_browse_hists");
 544                         if (hist_browser == NULL) {
 545                                 ui__error("GTK browser not found!\n");
 546                                 return ret;
 547                         }
 548                         hist_browser(session->evlist, help, NULL,
 549                                      rep->min_percent);
 550                 }
 551         } else
 552                 perf_evlist__tty_browse_hists(session->evlist, rep, help);
 553
 554         return ret;
 555 }
 556
 557 static int
 558 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
 559 {
 560         struct report *rep = (struct report *)opt->value;
 561         char *tok, *tok2;
 562         char *endptr;
 563
 564         /*
 565          * --no-call-graph
 566          */
 567         if (unset) {
 568                 rep->dont_use_callchains = true;
 569                 return 0;
 570         }
 571
 572         symbol_conf.use_callchain = true;
 573
 574         if (!arg)
 575                 return 0;
 576
 577         tok = strtok((char *)arg, ",");
 578         if (!tok)
 579                 return -1;
 580
 581         /* get the output mode */
 582         if (!strncmp(tok, "graph", strlen(arg)))
 583                 callchain_param.mode = CHAIN_GRAPH_ABS;
 584
 585         else if (!strncmp(tok, "flat", strlen(arg)))
 586                 callchain_param.mode = CHAIN_FLAT;
 587
 588         else if (!strncmp(tok, "fractal", strlen(arg)))
 589                 callchain_param.mode = CHAIN_GRAPH_REL;
 590
 591         else if (!strncmp(tok, "none", strlen(arg))) {
 592                 callchain_param.mode = CHAIN_NONE;
 593                 symbol_conf.use_callchain = false;
 594
 595                 return 0;
 596         }
 597
 598         else
 599                 return -1;
 600
 601         /* get the min percentage */
 602         tok = strtok(NULL, ",");
 603         if (!tok)
 604                 goto setup;
 605
 606         callchain_param.min_percent = strtod(tok, &endptr);
 607         if (tok == endptr)
 608                 return -1;
 609
 610         /* get the print limit */
 611         tok2 = strtok(NULL, ",");
 612         if (!tok2)
 613                 goto setup;
 614
 615         if (tok2[0] != 'c') {
 616                 callchain_param.print_limit = strtoul(tok2, &endptr, 0);
 617                 tok2 = strtok(NULL, ",");
 618                 if (!tok2)
 619                         goto setup;
 620         }
 621
 622         /* get the call chain order */
 623         if (!strncmp(tok2, "caller", strlen("caller")))
 624                 callchain_param.order = ORDER_CALLER;
 625         else if (!strncmp(tok2, "callee", strlen("callee")))
 626                 callchain_param.order = ORDER_CALLEE;
 627         else
 628                 return -1;
 629
 630         /* Get the sort key */
 631         tok2 = strtok(NULL, ",");
 632         if (!tok2)
 633                 goto setup;
 634         if (!strncmp(tok2, "function", strlen("function")))
 635                 callchain_param.key = CCKEY_FUNCTION;
 636         else if (!strncmp(tok2, "address", strlen("address")))
 637                 callchain_param.key = CCKEY_ADDRESS;
 638         else
 639                 return -1;
 640 setup:
 641         if (callchain_register_param(&callchain_param) < 0) {
 642                 fprintf(stderr, "Can't register callchain params\n");
 643                 return -1;
 644         }
 645         return 0;
 646 }
 647
 648 int
 649 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
 650                                 const char *arg, int unset __maybe_unused)
 651 {
 652         if (arg) {
 653                 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
 654                 if (err) {
 655                         char buf[BUFSIZ];
 656                         regerror(err, &ignore_callees_regex, buf, sizeof(buf));
 657                         pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
 658                         return -1;
 659                 }
 660                 have_ignore_callees = 1;
 661         }
 662
 663         return 0;
 664 }
 665
 666 static int
 667 parse_branch_mode(const struct option *opt __maybe_unused,
 668                   const char *str __maybe_unused, int unset)
 669 {
 670         int *branch_mode = opt->value;
 671
 672         *branch_mode = !unset;
 673         return 0;
 674 }
 675
 676 static int
 677 parse_percent_limit(const struct option *opt, const char *str,
 678                     int unset __maybe_unused)
 679 {
 680         struct report *rep = opt->value;
 681
 682         rep->min_percent = strtof(str, NULL);
 683         return 0;
 684 }
 685
 686 int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
 687 {
 688         struct perf_session *session;
 689         struct stat st;
 690         bool has_br_stack = false;
 691         int branch_mode = -1;
 692         int ret = -1;
 693         char callchain_default_opt[] = "fractal,0.5,callee";
 694         const char * const report_usage[] = {
 695                 "perf report [<options>]",
 696                 NULL
 697         };
 698         struct report report = {
 699                 .tool = {
 700                         .sample          = process_sample_event,
 701                         .mmap            = perf_event__process_mmap,
 702                         .mmap2           = perf_event__process_mmap2,
 703                         .comm            = perf_event__process_comm,
 704                         .exit            = perf_event__process_exit,
 705                         .fork            = perf_event__process_fork,
 706                         .lost            = perf_event__process_lost,
 707                         .read            = process_read_event,
 708                         .attr            = perf_event__process_attr,
 709                         .tracing_data    = perf_event__process_tracing_data,
 710                         .build_id        = perf_event__process_build_id,
 711                         .ordered_samples = true,
 712                         .ordering_requires_timestamps = true,
 713                 },
 714                 .max_stack               = PERF_MAX_STACK_DEPTH,
 715                 .pretty_printing_style   = "normal",
 716         };
 717         const struct option options[] = {
 718         OPT_STRING('i', "input", &input_name, "file",
 719                     "input file name"),
 720         OPT_INCR('v', "verbose", &verbose,
 721                     "be more verbose (show symbol address, etc)"),
 722         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 723                     "dump raw trace in ASCII"),
 724         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 725                    "file", "vmlinux pathname"),
 726         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 727                    "file", "kallsyms pathname"),
 728         OPT_BOOLEAN('f', "force", &report.force, "don't complain, do it"),
 729         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 730                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 731         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 732                     "Show a column with the number of samples"),
 733         OPT_BOOLEAN('T', "threads", &report.show_threads,
 734                     "Show per-thread event counters"),
 735         OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
 736                    "pretty printing style key: normal raw"),
 737         OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
 738         OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
 739         OPT_BOOLEAN(0, "stdio", &report.use_stdio,
 740                     "Use the stdio interface"),
 741         OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
 742         OPT_BOOLEAN(0, "header-only", &report.header_only,
 743                     "Show only data header."),
 744         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 745                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline,"
 746                    " dso_to, dso_from, symbol_to, symbol_from, mispredict,"
 747                    " weight, local_weight, mem, symbol_daddr, dso_daddr, tlb, "
 748                    "snoop, locked, abort, in_tx, transaction"),
 749         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 750                     "Show sample percentage for different cpu modes"),
 751         OPT_STRING('p', "parent", &parent_pattern, "regex",
 752                    "regex filter to identify parent, see: '--sort parent'"),
 753         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 754                     "Only display entries with parent-match"),
 755         OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
 756                      "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). "
 757                      "Default: fractal,0.5,callee,function", &parse_callchain_opt, callchain_default_opt),
 758         OPT_INTEGER(0, "max-stack", &report.max_stack,
 759                     "Set the maximum stack depth when parsing the callchain, "
 760                     "anything beyond the specified depth will be ignored. "
 761                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
 762         OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
 763                     "alias for inverted call graph"),
 764         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
 765                    "ignore callees of these functions in call graphs",
 766                    report_parse_ignore_callees_opt),
 767         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 768                    "only consider symbols in these dsos"),
 769         OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 770                    "only consider symbols in these comms"),
 771         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 772                    "only consider these symbols"),
 773         OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
 774                    "only show symbols that (partially) match with this filter"),
 775         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 776                    "width[,width...]",
 777                    "don't try to adjust column width, use these fixed values"),
 778         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 779                    "separator for columns, no spaces will be added between "
 780                    "columns '.' is reserved."),
 781         OPT_BOOLEAN('U', "hide-unresolved", &report.hide_unresolved,
 782                     "Only display entries resolved to a symbol"),
 783         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 784                     "Look for files with symbols relative to this directory"),
 785         OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
 786                    "list of cpus to profile"),
 787         OPT_BOOLEAN('I', "show-info", &report.show_full_info,
 788                     "Display extended information about perf.data file"),
 789         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 790                     "Interleave source code with assembly code (default)"),
 791         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 792                     "Display raw encoding of assembly instructions (default)"),
 793         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 794                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 795         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 796                     "Show a column with the sum of periods"),
 797         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 798                     "Show event group information together"),
 799         OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
 800                     "use branch records for histogram filling", parse_branch_mode),
 801         OPT_STRING(0, "objdump", &objdump_path, "path",
 802                    "objdump binary to use for disassembly and annotations"),
 803         OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
 804                     "Disable symbol demangling"),
 805         OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
 806         OPT_CALLBACK(0, "percent-limit", &report, "percent",
 807                      "Don't show entries under that percent", parse_percent_limit),
 808         OPT_END()
 809         };
 810         struct perf_data_file file = {
 811                 .mode  = PERF_DATA_MODE_READ,
 812         };
 813
 814         perf_config(report__config, &report);
 815
 816         argc = parse_options(argc, argv, options, report_usage, 0);
 817
 818         if (report.use_stdio)
 819                 use_browser = 0;
 820         else if (report.use_tui)
 821                 use_browser = 1;
 822         else if (report.use_gtk)
 823                 use_browser = 2;
 824
 825         if (report.inverted_callchain)
 826                 callchain_param.order = ORDER_CALLER;
 827
 828         if (!input_name || !strlen(input_name)) {
 829                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 830                         input_name = "-";
 831                 else
 832                         input_name = "perf.data";
 833         }
 834
 835         file.path  = input_name;
 836         file.force = report.force;
 837
 838 repeat:
 839         session = perf_session__new(&file, false, &report.tool);
 840         if (session == NULL)
 841                 return -ENOMEM;
 842
 843         report.session = session;
 844
 845         has_br_stack = perf_header__has_feat(&session->header,
 846                                              HEADER_BRANCH_STACK);
 847
 848         if (branch_mode == -1 && has_br_stack)
 849                 sort__mode = SORT_MODE__BRANCH;
 850
 851         /* sort__mode could be NORMAL if --no-branch-stack */
 852         if (sort__mode == SORT_MODE__BRANCH) {
 853                 /*
 854                  * if no sort_order is provided, then specify
 855                  * branch-mode specific order
 856                  */
 857                 if (sort_order == default_sort_order)
 858                         sort_order = "comm,dso_from,symbol_from,"
 859                                      "dso_to,symbol_to";
 860
 861         }
 862         if (report.mem_mode) {
 863                 if (sort__mode == SORT_MODE__BRANCH) {
 864                         fprintf(stderr, "branch and mem mode incompatible\n");
 865                         goto error;
 866                 }
 867                 sort__mode = SORT_MODE__MEMORY;
 868
 869                 /*
 870                  * if no sort_order is provided, then specify
 871                  * branch-mode specific order
 872                  */
 873                 if (sort_order == default_sort_order)
 874                         sort_order = "local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked";
 875         }
 876
 877         if (setup_sorting() < 0) {
 878                 parse_options_usage(report_usage, options, "s", 1);
 879                 goto error;
 880         }
 881
 882         if (parent_pattern != default_parent_pattern) {
 883                 if (sort_dimension__add("parent") < 0)
 884                         goto error;
 885         }
 886
 887         /* Force tty output for header output. */
 888         if (report.header || report.header_only)
 889                 use_browser = 0;
 890
 891         if (strcmp(input_name, "-") != 0)
 892                 setup_browser(true);
 893         else {
 894                 use_browser = 0;
 895                 perf_hpp__init();
 896         }
 897
 898         if (report.header || report.header_only) {
 899                 perf_session__fprintf_info(session, stdout,
 900                                            report.show_full_info);
 901                 if (report.header_only)
 902                         return 0;
 903         } else if (use_browser == 0) {
 904                 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
 905                       stdout);
 906         }
 907
 908         /*
 909          * Only in the TUI browser we are doing integrated annotation,
 910          * so don't allocate extra space that won't be used in the stdio
 911          * implementation.
 912          */
 913         if (use_browser == 1 && sort__has_sym) {
 914                 symbol_conf.priv_size = sizeof(struct annotation);
 915                 machines__set_symbol_filter(&session->machines,
 916                                             symbol__annotate_init);
 917                 /*
 918                  * For searching by name on the "Browse map details".
 919                  * providing it only in verbose mode not to bloat too
 920                  * much struct symbol.
 921                  */
 922                 if (verbose) {
 923                         /*
 924                          * XXX: Need to provide a less kludgy way to ask for
 925                          * more space per symbol, the u32 is for the index on
 926                          * the ui browser.
 927                          * See symbol__browser_index.
 928                          */
 929                         symbol_conf.priv_size += sizeof(u32);
 930                         symbol_conf.sort_by_name = true;
 931                 }
 932         }
 933
 934         if (symbol__init() < 0)
 935                 goto error;
 936
 937         if (argc) {
 938                 /*
 939                  * Special case: if there's an argument left then assume that
 940                  * it's a symbol filter:
 941                  */
 942                 if (argc > 1)
 943                         usage_with_options(report_usage, options);
 944
 945                 report.symbol_filter_str = argv[0];
 946         }
 947
 948         sort__setup_elide(stdout);
 949
 950         ret = __cmd_report(&report);
 951         if (ret == K_SWITCH_INPUT_DATA) {
 952                 perf_session__delete(session);
 953                 goto repeat;
 954         } else
 955                 ret = 0;
 956
 957 error:
 958         perf_session__delete(session);
 959         return ret;
 960 }