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 perf_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         int                     max_stack;
  53         struct perf_read_values show_threads_values;
  54         const char              *pretty_printing_style;
  55         const char              *cpu_list;
  56         const char              *symbol_filter_str;
  57         float                   min_percent;
  58         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  59 };
  60
  61 static int perf_report_config(const char *var, const char *value, void *cb)
  62 {
  63         if (!strcmp(var, "report.group")) {
  64                 symbol_conf.event_group = perf_config_bool(var, value);
  65                 return 0;
  66         }
  67         if (!strcmp(var, "report.percent-limit")) {
  68                 struct perf_report *rep = cb;
  69                 rep->min_percent = strtof(value, NULL);
  70                 return 0;
  71         }
  72
  73         return perf_default_config(var, value, cb);
  74 }
  75
  76 static int perf_report__add_mem_hist_entry(struct perf_tool *tool,
  77                                            struct addr_location *al,
  78                                            struct perf_sample *sample,
  79                                            struct perf_evsel *evsel,
  80                                            struct machine *machine,
  81                                            union perf_event *event)
  82 {
  83         struct perf_report *rep = container_of(tool, struct perf_report, tool);
  84         struct symbol *parent = NULL;
  85         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
  86         int err = 0;
  87         struct hist_entry *he;
  88         struct mem_info *mi, *mx;
  89         uint64_t cost;
  90
  91         if ((sort__has_parent || symbol_conf.use_callchain) &&
  92             sample->callchain) {
  93                 err = machine__resolve_callchain(machine, evsel, al->thread,
  94                                                  sample, &parent, al,
  95                                                  rep->max_stack);
  96                 if (err)
  97                         return err;
  98         }
  99
 100         mi = machine__resolve_mem(machine, al->thread, sample, cpumode);
 101         if (!mi)
 102                 return -ENOMEM;
 103
 104         if (rep->hide_unresolved && !al->sym)
 105                 return 0;
 106
 107         cost = sample->weight;
 108         if (!cost)
 109                 cost = 1;
 110
 111         /*
 112          * must pass period=weight in order to get the correct
 113          * sorting from hists__collapse_resort() which is solely
 114          * based on periods. We want sorting be done on nr_events * weight
 115          * and this is indirectly achieved by passing period=weight here
 116          * and the he_stat__add_period() function.
 117          */
 118         he = __hists__add_mem_entry(&evsel->hists, al, parent, mi, cost, cost);
 119         if (!he)
 120                 return -ENOMEM;
 121
 122         /*
 123          * In the TUI browser, we are doing integrated annotation,
 124          * so we don't allocate the extra space needed because the stdio
 125          * code will not use it.
 126          */
 127         if (sort__has_sym && he->ms.sym && use_browser > 0) {
 128                 struct annotation *notes = symbol__annotation(he->ms.sym);
 129
 130                 assert(evsel != NULL);
 131
 132                 if (notes->src == NULL && symbol__alloc_hist(he->ms.sym) < 0)
 133                         goto out;
 134
 135                 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 136                 if (err)
 137                         goto out;
 138         }
 139
 140         if (sort__has_sym && he->mem_info->daddr.sym && use_browser > 0) {
 141                 struct annotation *notes;
 142
 143                 mx = he->mem_info;
 144
 145                 notes = symbol__annotation(mx->daddr.sym);
 146                 if (notes->src == NULL && symbol__alloc_hist(mx->daddr.sym) < 0)
 147                         goto out;
 148
 149                 err = symbol__inc_addr_samples(mx->daddr.sym,
 150                                                mx->daddr.map,
 151                                                evsel->idx,
 152                                                mx->daddr.al_addr);
 153                 if (err)
 154                         goto out;
 155         }
 156
 157         evsel->hists.stats.total_period += cost;
 158         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 159         err = 0;
 160
 161         if (symbol_conf.use_callchain) {
 162                 err = callchain_append(he->callchain,
 163                                        &callchain_cursor,
 164                                        sample->period);
 165         }
 166 out:
 167         return err;
 168 }
 169
 170 static int perf_report__add_branch_hist_entry(struct perf_tool *tool,
 171                                         struct addr_location *al,
 172                                         struct perf_sample *sample,
 173                                         struct perf_evsel *evsel,
 174                                       struct machine *machine)
 175 {
 176         struct perf_report *rep = container_of(tool, struct perf_report, tool);
 177         struct symbol *parent = NULL;
 178         int err = 0;
 179         unsigned i;
 180         struct hist_entry *he;
 181         struct branch_info *bi, *bx;
 182
 183         if ((sort__has_parent || symbol_conf.use_callchain)
 184             && sample->callchain) {
 185                 err = machine__resolve_callchain(machine, evsel, al->thread,
 186                                                  sample, &parent, al,
 187                                                  rep->max_stack);
 188                 if (err)
 189                         return err;
 190         }
 191
 192         bi = machine__resolve_bstack(machine, al->thread,
 193                                      sample->branch_stack);
 194         if (!bi)
 195                 return -ENOMEM;
 196
 197         for (i = 0; i < sample->branch_stack->nr; i++) {
 198                 if (rep->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
 199                         continue;
 200
 201                 err = -ENOMEM;
 202
 203                 /*
 204                  * The report shows the percentage of total branches captured
 205                  * and not events sampled. Thus we use a pseudo period of 1.
 206                  */
 207                 he = __hists__add_branch_entry(&evsel->hists, al, parent,
 208                                 &bi[i], 1, 1);
 209                 if (he) {
 210                         struct annotation *notes;
 211                         bx = he->branch_info;
 212                         if (bx->from.sym && use_browser == 1 && sort__has_sym) {
 213                                 notes = symbol__annotation(bx->from.sym);
 214                                 if (!notes->src
 215                                     && symbol__alloc_hist(bx->from.sym) < 0)
 216                                         goto out;
 217
 218                                 err = symbol__inc_addr_samples(bx->from.sym,
 219                                                                bx->from.map,
 220                                                                evsel->idx,
 221                                                                bx->from.al_addr);
 222                                 if (err)
 223                                         goto out;
 224                         }
 225
 226                         if (bx->to.sym && use_browser == 1 && sort__has_sym) {
 227                                 notes = symbol__annotation(bx->to.sym);
 228                                 if (!notes->src
 229                                     && symbol__alloc_hist(bx->to.sym) < 0)
 230                                         goto out;
 231
 232                                 err = symbol__inc_addr_samples(bx->to.sym,
 233                                                                bx->to.map,
 234                                                                evsel->idx,
 235                                                                bx->to.al_addr);
 236                                 if (err)
 237                                         goto out;
 238                         }
 239                         evsel->hists.stats.total_period += 1;
 240                         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 241                 } else
 242                         goto out;
 243         }
 244         err = 0;
 245 out:
 246         free(bi);
 247         return err;
 248 }
 249
 250 static int perf_evsel__add_hist_entry(struct perf_tool *tool,
 251                                       struct perf_evsel *evsel,
 252                                       struct addr_location *al,
 253                                       struct perf_sample *sample,
 254                                       struct machine *machine)
 255 {
 256         struct perf_report *rep = container_of(tool, struct perf_report, tool);
 257         struct symbol *parent = NULL;
 258         int err = 0;
 259         struct hist_entry *he;
 260
 261         if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
 262                 err = machine__resolve_callchain(machine, evsel, al->thread,
 263                                                  sample, &parent, al,
 264                                                  rep->max_stack);
 265                 if (err)
 266                         return err;
 267         }
 268
 269         he = __hists__add_entry(&evsel->hists, al, parent, sample->period,
 270                                 sample->weight, sample->transaction);
 271         if (he == NULL)
 272                 return -ENOMEM;
 273
 274         if (symbol_conf.use_callchain) {
 275                 err = callchain_append(he->callchain,
 276                                        &callchain_cursor,
 277                                        sample->period);
 278                 if (err)
 279                         return err;
 280         }
 281         /*
 282          * Only in the TUI browser we are doing integrated annotation,
 283          * so we don't allocated the extra space needed because the stdio
 284          * code will not use it.
 285          */
 286         if (he->ms.sym != NULL && use_browser == 1 && sort__has_sym) {
 287                 struct annotation *notes = symbol__annotation(he->ms.sym);
 288
 289                 assert(evsel != NULL);
 290
 291                 err = -ENOMEM;
 292                 if (notes->src == NULL && symbol__alloc_hist(he->ms.sym) < 0)
 293                         goto out;
 294
 295                 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
 296         }
 297
 298         evsel->hists.stats.total_period += sample->period;
 299         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 300 out:
 301         return err;
 302 }
 303
 304
 305 static int process_sample_event(struct perf_tool *tool,
 306                                 union perf_event *event,
 307                                 struct perf_sample *sample,
 308                                 struct perf_evsel *evsel,
 309                                 struct machine *machine)
 310 {
 311         struct perf_report *rep = container_of(tool, struct perf_report, tool);
 312         struct addr_location al;
 313         int ret;
 314
 315         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
 316                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 317                         event->header.type);
 318                 return -1;
 319         }
 320
 321         if (al.filtered || (rep->hide_unresolved && al.sym == NULL))
 322                 return 0;
 323
 324         if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
 325                 return 0;
 326
 327         if (sort__mode == SORT_MODE__BRANCH) {
 328                 ret = perf_report__add_branch_hist_entry(tool, &al, sample,
 329                                                          evsel, machine);
 330                 if (ret < 0)
 331                         pr_debug("problem adding lbr entry, skipping event\n");
 332         } else if (rep->mem_mode == 1) {
 333                 ret = perf_report__add_mem_hist_entry(tool, &al, sample,
 334                                                       evsel, machine, event);
 335                 if (ret < 0)
 336                         pr_debug("problem adding mem entry, skipping event\n");
 337         } else {
 338                 if (al.map != NULL)
 339                         al.map->dso->hit = 1;
 340
 341                 ret = perf_evsel__add_hist_entry(tool, evsel, &al, sample,
 342                                                  machine);
 343                 if (ret < 0)
 344                         pr_debug("problem incrementing symbol period, skipping event\n");
 345         }
 346         return ret;
 347 }
 348
 349 static int process_read_event(struct perf_tool *tool,
 350                               union perf_event *event,
 351                               struct perf_sample *sample __maybe_unused,
 352                               struct perf_evsel *evsel,
 353                               struct machine *machine __maybe_unused)
 354 {
 355         struct perf_report *rep = container_of(tool, struct perf_report, tool);
 356
 357         if (rep->show_threads) {
 358                 const char *name = evsel ? perf_evsel__name(evsel) : "unknown";
 359                 perf_read_values_add_value(&rep->show_threads_values,
 360                                            event->read.pid, event->read.tid,
 361                                            event->read.id,
 362                                            name,
 363                                            event->read.value);
 364         }
 365
 366         dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 367                     evsel ? perf_evsel__name(evsel) : "FAIL",
 368                     event->read.value);
 369
 370         return 0;
 371 }
 372
 373 /* For pipe mode, sample_type is not currently set */
 374 static int perf_report__setup_sample_type(struct perf_report *rep)
 375 {
 376         struct perf_session *self = rep->session;
 377         u64 sample_type = perf_evlist__combined_sample_type(self->evlist);
 378         bool is_pipe = perf_data_file__is_pipe(self->file);
 379
 380         if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
 381                 if (sort__has_parent) {
 382                         ui__error("Selected --sort parent, but no "
 383                                     "callchain data. Did you call "
 384                                     "'perf record' without -g?\n");
 385                         return -EINVAL;
 386                 }
 387                 if (symbol_conf.use_callchain) {
 388                         ui__error("Selected -g but no callchain data. Did "
 389                                     "you call 'perf record' without -g?\n");
 390                         return -1;
 391                 }
 392         } else if (!rep->dont_use_callchains &&
 393                    callchain_param.mode != CHAIN_NONE &&
 394                    !symbol_conf.use_callchain) {
 395                         symbol_conf.use_callchain = true;
 396                         if (callchain_register_param(&callchain_param) < 0) {
 397                                 ui__error("Can't register callchain params.\n");
 398                                 return -EINVAL;
 399                         }
 400         }
 401
 402         if (sort__mode == SORT_MODE__BRANCH) {
 403                 if (!is_pipe &&
 404                     !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
 405                         ui__error("Selected -b but no branch data. "
 406                                   "Did you call perf record without -b?\n");
 407                         return -1;
 408                 }
 409         }
 410
 411         return 0;
 412 }
 413
 414 static void sig_handler(int sig __maybe_unused)
 415 {
 416         session_done = 1;
 417 }
 418
 419 static size_t hists__fprintf_nr_sample_events(struct perf_report *rep,
 420                                               struct hists *self,
 421                                               const char *evname, FILE *fp)
 422 {
 423         size_t ret;
 424         char unit;
 425         unsigned long nr_samples = self->stats.nr_events[PERF_RECORD_SAMPLE];
 426         u64 nr_events = self->stats.total_period;
 427         struct perf_evsel *evsel = hists_to_evsel(self);
 428         char buf[512];
 429         size_t size = sizeof(buf);
 430
 431         if (perf_evsel__is_group_event(evsel)) {
 432                 struct perf_evsel *pos;
 433
 434                 perf_evsel__group_desc(evsel, buf, size);
 435                 evname = buf;
 436
 437                 for_each_group_member(pos, evsel) {
 438                         nr_samples += pos->hists.stats.nr_events[PERF_RECORD_SAMPLE];
 439                         nr_events += pos->hists.stats.total_period;
 440                 }
 441         }
 442
 443         nr_samples = convert_unit(nr_samples, &unit);
 444         ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
 445         if (evname != NULL)
 446                 ret += fprintf(fp, " of event '%s'", evname);
 447
 448         if (rep->mem_mode) {
 449                 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
 450                 ret += fprintf(fp, "\n# Sort order   : %s", sort_order);
 451         } else
 452                 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
 453         return ret + fprintf(fp, "\n#\n");
 454 }
 455
 456 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
 457                                          struct perf_report *rep,
 458                                          const char *help)
 459 {
 460         struct perf_evsel *pos;
 461
 462         list_for_each_entry(pos, &evlist->entries, node) {
 463                 struct hists *hists = &pos->hists;
 464                 const char *evname = perf_evsel__name(pos);
 465
 466                 if (symbol_conf.event_group &&
 467                     !perf_evsel__is_group_leader(pos))
 468                         continue;
 469
 470                 hists__fprintf_nr_sample_events(rep, hists, evname, stdout);
 471                 hists__fprintf(hists, true, 0, 0, rep->min_percent, stdout);
 472                 fprintf(stdout, "\n\n");
 473         }
 474
 475         if (sort_order == default_sort_order &&
 476             parent_pattern == default_parent_pattern) {
 477                 fprintf(stdout, "#\n# (%s)\n#\n", help);
 478
 479                 if (rep->show_threads) {
 480                         bool style = !strcmp(rep->pretty_printing_style, "raw");
 481                         perf_read_values_display(stdout, &rep->show_threads_values,
 482                                                  style);
 483                         perf_read_values_destroy(&rep->show_threads_values);
 484                 }
 485         }
 486
 487         return 0;
 488 }
 489
 490 static int __cmd_report(struct perf_report *rep)
 491 {
 492         int ret = -EINVAL;
 493         u64 nr_samples;
 494         struct perf_session *session = rep->session;
 495         struct perf_evsel *pos;
 496         struct map *kernel_map;
 497         struct kmap *kernel_kmap;
 498         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 499         struct perf_data_file *file = session->file;
 500
 501         signal(SIGINT, sig_handler);
 502
 503         if (rep->cpu_list) {
 504                 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
 505                                                rep->cpu_bitmap);
 506                 if (ret)
 507                         return ret;
 508         }
 509
 510         if (use_browser <= 0)
 511                 perf_session__fprintf_info(session, stdout, rep->show_full_info);
 512
 513         if (rep->show_threads)
 514                 perf_read_values_init(&rep->show_threads_values);
 515
 516         ret = perf_report__setup_sample_type(rep);
 517         if (ret)
 518                 return ret;
 519
 520         ret = perf_session__process_events(session, &rep->tool);
 521         if (ret)
 522                 return ret;
 523
 524         kernel_map = session->machines.host.vmlinux_maps[MAP__FUNCTION];
 525         kernel_kmap = map__kmap(kernel_map);
 526         if (kernel_map == NULL ||
 527             (kernel_map->dso->hit &&
 528              (kernel_kmap->ref_reloc_sym == NULL ||
 529               kernel_kmap->ref_reloc_sym->addr == 0))) {
 530                 const char *desc =
 531                     "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
 532                     "can't be resolved.";
 533
 534                 if (kernel_map) {
 535                         const struct dso *kdso = kernel_map->dso;
 536                         if (!RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION])) {
 537                                 desc = "If some relocation was applied (e.g. "
 538                                        "kexec) symbols may be misresolved.";
 539                         }
 540                 }
 541
 542                 ui__warning(
 543 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
 544 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
 545 "Samples in kernel modules can't be resolved as well.\n\n",
 546                 desc);
 547         }
 548
 549         if (verbose > 3)
 550                 perf_session__fprintf(session, stdout);
 551
 552         if (verbose > 2)
 553                 perf_session__fprintf_dsos(session, stdout);
 554
 555         if (dump_trace) {
 556                 perf_session__fprintf_nr_events(session, stdout);
 557                 return 0;
 558         }
 559
 560         nr_samples = 0;
 561         list_for_each_entry(pos, &session->evlist->entries, node) {
 562                 struct hists *hists = &pos->hists;
 563
 564                 if (pos->idx == 0)
 565                         hists->symbol_filter_str = rep->symbol_filter_str;
 566
 567                 hists__collapse_resort(hists);
 568                 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
 569
 570                 /* Non-group events are considered as leader */
 571                 if (symbol_conf.event_group &&
 572                     !perf_evsel__is_group_leader(pos)) {
 573                         struct hists *leader_hists = &pos->leader->hists;
 574
 575                         hists__match(leader_hists, hists);
 576                         hists__link(leader_hists, hists);
 577                 }
 578         }
 579
 580         if (session_done())
 581                 return 0;
 582
 583         if (nr_samples == 0) {
 584                 ui__error("The %s file has no samples!\n", file->path);
 585                 return 0;
 586         }
 587
 588         list_for_each_entry(pos, &session->evlist->entries, node)
 589                 hists__output_resort(&pos->hists);
 590
 591         if (use_browser > 0) {
 592                 if (use_browser == 1) {
 593                         ret = perf_evlist__tui_browse_hists(session->evlist,
 594                                                         help, NULL,
 595                                                         rep->min_percent,
 596                                                         &session->header.env);
 597                         /*
 598                          * Usually "ret" is the last pressed key, and we only
 599                          * care if the key notifies us to switch data file.
 600                          */
 601                         if (ret != K_SWITCH_INPUT_DATA)
 602                                 ret = 0;
 603
 604                 } else if (use_browser == 2) {
 605                         int (*hist_browser)(struct perf_evlist *,
 606                                             const char *,
 607                                             struct hist_browser_timer *,
 608                                             float min_pcnt);
 609
 610                         hist_browser = dlsym(perf_gtk_handle,
 611                                              "perf_evlist__gtk_browse_hists");
 612                         if (hist_browser == NULL) {
 613                                 ui__error("GTK browser not found!\n");
 614                                 return ret;
 615                         }
 616                         hist_browser(session->evlist, help, NULL,
 617                                      rep->min_percent);
 618                 }
 619         } else
 620                 perf_evlist__tty_browse_hists(session->evlist, rep, help);
 621
 622         return ret;
 623 }
 624
 625 static int
 626 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
 627 {
 628         struct perf_report *rep = (struct perf_report *)opt->value;
 629         char *tok, *tok2;
 630         char *endptr;
 631
 632         /*
 633          * --no-call-graph
 634          */
 635         if (unset) {
 636                 rep->dont_use_callchains = true;
 637                 return 0;
 638         }
 639
 640         symbol_conf.use_callchain = true;
 641
 642         if (!arg)
 643                 return 0;
 644
 645         tok = strtok((char *)arg, ",");
 646         if (!tok)
 647                 return -1;
 648
 649         /* get the output mode */
 650         if (!strncmp(tok, "graph", strlen(arg)))
 651                 callchain_param.mode = CHAIN_GRAPH_ABS;
 652
 653         else if (!strncmp(tok, "flat", strlen(arg)))
 654                 callchain_param.mode = CHAIN_FLAT;
 655
 656         else if (!strncmp(tok, "fractal", strlen(arg)))
 657                 callchain_param.mode = CHAIN_GRAPH_REL;
 658
 659         else if (!strncmp(tok, "none", strlen(arg))) {
 660                 callchain_param.mode = CHAIN_NONE;
 661                 symbol_conf.use_callchain = false;
 662
 663                 return 0;
 664         }
 665
 666         else
 667                 return -1;
 668
 669         /* get the min percentage */
 670         tok = strtok(NULL, ",");
 671         if (!tok)
 672                 goto setup;
 673
 674         callchain_param.min_percent = strtod(tok, &endptr);
 675         if (tok == endptr)
 676                 return -1;
 677
 678         /* get the print limit */
 679         tok2 = strtok(NULL, ",");
 680         if (!tok2)
 681                 goto setup;
 682
 683         if (tok2[0] != 'c') {
 684                 callchain_param.print_limit = strtoul(tok2, &endptr, 0);
 685                 tok2 = strtok(NULL, ",");
 686                 if (!tok2)
 687                         goto setup;
 688         }
 689
 690         /* get the call chain order */
 691         if (!strncmp(tok2, "caller", strlen("caller")))
 692                 callchain_param.order = ORDER_CALLER;
 693         else if (!strncmp(tok2, "callee", strlen("callee")))
 694                 callchain_param.order = ORDER_CALLEE;
 695         else
 696                 return -1;
 697
 698         /* Get the sort key */
 699         tok2 = strtok(NULL, ",");
 700         if (!tok2)
 701                 goto setup;
 702         if (!strncmp(tok2, "function", strlen("function")))
 703                 callchain_param.key = CCKEY_FUNCTION;
 704         else if (!strncmp(tok2, "address", strlen("address")))
 705                 callchain_param.key = CCKEY_ADDRESS;
 706         else
 707                 return -1;
 708 setup:
 709         if (callchain_register_param(&callchain_param) < 0) {
 710                 fprintf(stderr, "Can't register callchain params\n");
 711                 return -1;
 712         }
 713         return 0;
 714 }
 715
 716 int
 717 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
 718                                 const char *arg, int unset __maybe_unused)
 719 {
 720         if (arg) {
 721                 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
 722                 if (err) {
 723                         char buf[BUFSIZ];
 724                         regerror(err, &ignore_callees_regex, buf, sizeof(buf));
 725                         pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
 726                         return -1;
 727                 }
 728                 have_ignore_callees = 1;
 729         }
 730
 731         return 0;
 732 }
 733
 734 static int
 735 parse_branch_mode(const struct option *opt __maybe_unused,
 736                   const char *str __maybe_unused, int unset)
 737 {
 738         int *branch_mode = opt->value;
 739
 740         *branch_mode = !unset;
 741         return 0;
 742 }
 743
 744 static int
 745 parse_percent_limit(const struct option *opt, const char *str,
 746                     int unset __maybe_unused)
 747 {
 748         struct perf_report *rep = opt->value;
 749
 750         rep->min_percent = strtof(str, NULL);
 751         return 0;
 752 }
 753
 754 int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
 755 {
 756         struct perf_session *session;
 757         struct stat st;
 758         bool has_br_stack = false;
 759         int branch_mode = -1;
 760         int ret = -1;
 761         char callchain_default_opt[] = "fractal,0.5,callee";
 762         const char * const report_usage[] = {
 763                 "perf report [<options>]",
 764                 NULL
 765         };
 766         struct perf_report report = {
 767                 .tool = {
 768                         .sample          = process_sample_event,
 769                         .mmap            = perf_event__process_mmap,
 770                         .mmap2           = perf_event__process_mmap2,
 771                         .comm            = perf_event__process_comm,
 772                         .exit            = perf_event__process_exit,
 773                         .fork            = perf_event__process_fork,
 774                         .lost            = perf_event__process_lost,
 775                         .read            = process_read_event,
 776                         .attr            = perf_event__process_attr,
 777                         .tracing_data    = perf_event__process_tracing_data,
 778                         .build_id        = perf_event__process_build_id,
 779                         .ordered_samples = true,
 780                         .ordering_requires_timestamps = true,
 781                 },
 782                 .max_stack               = PERF_MAX_STACK_DEPTH,
 783                 .pretty_printing_style   = "normal",
 784         };
 785         const struct option options[] = {
 786         OPT_STRING('i', "input", &input_name, "file",
 787                     "input file name"),
 788         OPT_INCR('v', "verbose", &verbose,
 789                     "be more verbose (show symbol address, etc)"),
 790         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 791                     "dump raw trace in ASCII"),
 792         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 793                    "file", "vmlinux pathname"),
 794         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 795                    "file", "kallsyms pathname"),
 796         OPT_BOOLEAN('f', "force", &report.force, "don't complain, do it"),
 797         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 798                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
 799         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 800                     "Show a column with the number of samples"),
 801         OPT_BOOLEAN('T', "threads", &report.show_threads,
 802                     "Show per-thread event counters"),
 803         OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
 804                    "pretty printing style key: normal raw"),
 805         OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
 806         OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
 807         OPT_BOOLEAN(0, "stdio", &report.use_stdio,
 808                     "Use the stdio interface"),
 809         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 810                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline,"
 811                    " dso_to, dso_from, symbol_to, symbol_from, mispredict,"
 812                    " weight, local_weight, mem, symbol_daddr, dso_daddr, tlb, "
 813                    "snoop, locked, abort, in_tx, transaction"),
 814         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 815                     "Show sample percentage for different cpu modes"),
 816         OPT_STRING('p', "parent", &parent_pattern, "regex",
 817                    "regex filter to identify parent, see: '--sort parent'"),
 818         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 819                     "Only display entries with parent-match"),
 820         OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
 821                      "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). "
 822                      "Default: fractal,0.5,callee,function", &parse_callchain_opt, callchain_default_opt),
 823         OPT_INTEGER(0, "max-stack", &report.max_stack,
 824                     "Set the maximum stack depth when parsing the callchain, "
 825                     "anything beyond the specified depth will be ignored. "
 826                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
 827         OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
 828                     "alias for inverted call graph"),
 829         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
 830                    "ignore callees of these functions in call graphs",
 831                    report_parse_ignore_callees_opt),
 832         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 833                    "only consider symbols in these dsos"),
 834         OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 835                    "only consider symbols in these comms"),
 836         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 837                    "only consider these symbols"),
 838         OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
 839                    "only show symbols that (partially) match with this filter"),
 840         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 841                    "width[,width...]",
 842                    "don't try to adjust column width, use these fixed values"),
 843         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 844                    "separator for columns, no spaces will be added between "
 845                    "columns '.' is reserved."),
 846         OPT_BOOLEAN('U', "hide-unresolved", &report.hide_unresolved,
 847                     "Only display entries resolved to a symbol"),
 848         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 849                     "Look for files with symbols relative to this directory"),
 850         OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
 851                    "list of cpus to profile"),
 852         OPT_BOOLEAN('I', "show-info", &report.show_full_info,
 853                     "Display extended information about perf.data file"),
 854         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
 855                     "Interleave source code with assembly code (default)"),
 856         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
 857                     "Display raw encoding of assembly instructions (default)"),
 858         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
 859                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
 860         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
 861                     "Show a column with the sum of periods"),
 862         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
 863                     "Show event group information together"),
 864         OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
 865                     "use branch records for histogram filling", parse_branch_mode),
 866         OPT_STRING(0, "objdump", &objdump_path, "path",
 867                    "objdump binary to use for disassembly and annotations"),
 868         OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
 869                     "Disable symbol demangling"),
 870         OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
 871         OPT_CALLBACK(0, "percent-limit", &report, "percent",
 872                      "Don't show entries under that percent", parse_percent_limit),
 873         OPT_END()
 874         };
 875         struct perf_data_file file = {
 876                 .mode  = PERF_DATA_MODE_READ,
 877         };
 878
 879         perf_config(perf_report_config, &report);
 880
 881         argc = parse_options(argc, argv, options, report_usage, 0);
 882
 883         if (report.use_stdio)
 884                 use_browser = 0;
 885         else if (report.use_tui)
 886                 use_browser = 1;
 887         else if (report.use_gtk)
 888                 use_browser = 2;
 889
 890         if (report.inverted_callchain)
 891                 callchain_param.order = ORDER_CALLER;
 892
 893         if (!input_name || !strlen(input_name)) {
 894                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 895                         input_name = "-";
 896                 else
 897                         input_name = "perf.data";
 898         }
 899
 900         if (strcmp(input_name, "-") != 0)
 901                 setup_browser(true);
 902         else {
 903                 use_browser = 0;
 904                 perf_hpp__init();
 905         }
 906
 907         file.path  = input_name;
 908         file.force = report.force;
 909
 910 repeat:
 911         session = perf_session__new(&file, false, &report.tool);
 912         if (session == NULL)
 913                 return -ENOMEM;
 914
 915         report.session = session;
 916
 917         has_br_stack = perf_header__has_feat(&session->header,
 918                                              HEADER_BRANCH_STACK);
 919
 920         if (branch_mode == -1 && has_br_stack)
 921                 sort__mode = SORT_MODE__BRANCH;
 922
 923         /* sort__mode could be NORMAL if --no-branch-stack */
 924         if (sort__mode == SORT_MODE__BRANCH) {
 925                 /*
 926                  * if no sort_order is provided, then specify
 927                  * branch-mode specific order
 928                  */
 929                 if (sort_order == default_sort_order)
 930                         sort_order = "comm,dso_from,symbol_from,"
 931                                      "dso_to,symbol_to";
 932
 933         }
 934         if (report.mem_mode) {
 935                 if (sort__mode == SORT_MODE__BRANCH) {
 936                         fprintf(stderr, "branch and mem mode incompatible\n");
 937                         goto error;
 938                 }
 939                 sort__mode = SORT_MODE__MEMORY;
 940
 941                 /*
 942                  * if no sort_order is provided, then specify
 943                  * branch-mode specific order
 944                  */
 945                 if (sort_order == default_sort_order)
 946                         sort_order = "local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked";
 947         }
 948
 949         if (setup_sorting() < 0)
 950                 usage_with_options(report_usage, options);
 951
 952         /*
 953          * Only in the TUI browser we are doing integrated annotation,
 954          * so don't allocate extra space that won't be used in the stdio
 955          * implementation.
 956          */
 957         if (use_browser == 1 && sort__has_sym) {
 958                 symbol_conf.priv_size = sizeof(struct annotation);
 959                 machines__set_symbol_filter(&session->machines,
 960                                             symbol__annotate_init);
 961                 /*
 962                  * For searching by name on the "Browse map details".
 963                  * providing it only in verbose mode not to bloat too
 964                  * much struct symbol.
 965                  */
 966                 if (verbose) {
 967                         /*
 968                          * XXX: Need to provide a less kludgy way to ask for
 969                          * more space per symbol, the u32 is for the index on
 970                          * the ui browser.
 971                          * See symbol__browser_index.
 972                          */
 973                         symbol_conf.priv_size += sizeof(u32);
 974                         symbol_conf.sort_by_name = true;
 975                 }
 976         }
 977
 978         if (symbol__init() < 0)
 979                 goto error;
 980
 981         if (parent_pattern != default_parent_pattern) {
 982                 if (sort_dimension__add("parent") < 0)
 983                         goto error;
 984         }
 985
 986         if (argc) {
 987                 /*
 988                  * Special case: if there's an argument left then assume that
 989                  * it's a symbol filter:
 990                  */
 991                 if (argc > 1)
 992                         usage_with_options(report_usage, options);
 993
 994                 report.symbol_filter_str = argv[0];
 995         }
 996
 997         sort__setup_elide(stdout);
 998
 999         ret = __cmd_report(&report);
1000         if (ret == K_SWITCH_INPUT_DATA) {
1001                 perf_session__delete(session);
1002                 goto repeat;
1003         } else
1004                 ret = 0;
1005
1006 error:
1007         perf_session__delete(session);
1008         return ret;
1009 }