tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9  *
  10  * Improvements and fixes by:
  11  *
  12  *   Arjan van de Ven <arjan@linux.intel.com>
  13  *   Yanmin Zhang <yanmin.zhang@intel.com>
  14  *   Wu Fengguang <fengguang.wu@intel.com>
  15  *   Mike Galbraith <efault@gmx.de>
  16  *   Paul Mackerras <paulus@samba.org>
  17  *
  18  * Released under the GPL v2. (and only v2, not any later version)
  19  */
  20 #include "builtin.h"
  21
  22 #include "perf.h"
  23
  24 #include "util/annotate.h"
  25 #include "util/config.h"
  26 #include "util/color.h"
  27 #include "util/drv_configs.h"
  28 #include "util/evlist.h"
  29 #include "util/evsel.h"
  30 #include "util/event.h"
  31 #include "util/machine.h"
  32 #include "util/session.h"
  33 #include "util/symbol.h"
  34 #include "util/thread.h"
  35 #include "util/thread_map.h"
  36 #include "util/top.h"
  37 #include <linux/rbtree.h>
  38 #include <subcmd/parse-options.h>
  39 #include "util/parse-events.h"
  40 #include "util/cpumap.h"
  41 #include "util/xyarray.h"
  42 #include "util/sort.h"
  43 #include "util/term.h"
  44 #include "util/intlist.h"
  45 #include "util/parse-branch-options.h"
  46 #include "arch/common.h"
  47
  48 #include "util/debug.h"
  49
  50 #include <assert.h>
  51 #include <elf.h>
  52 #include <fcntl.h>
  53
  54 #include <stdio.h>
  55 #include <termios.h>
  56 #include <unistd.h>
  57 #include <inttypes.h>
  58
  59 #include <errno.h>
  60 #include <time.h>
  61 #include <sched.h>
  62 #include <signal.h>
  63
  64 #include <sys/syscall.h>
  65 #include <sys/ioctl.h>
  66 #include <poll.h>
  67 #include <sys/prctl.h>
  68 #include <sys/wait.h>
  69 #include <sys/uio.h>
  70 #include <sys/utsname.h>
  71 #include <sys/mman.h>
  72
  73 #include <linux/stringify.h>
  74 #include <linux/time64.h>
  75 #include <linux/types.h>
  76
  77 #include "sane_ctype.h"
  78
  79 static volatile int done;
  80 static volatile int resize;
  81
  82 #define HEADER_LINE_NR  5
  83
  84 static void perf_top__update_print_entries(struct perf_top *top)
  85 {
  86         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  87 }
  88
  89 static void winch_sig(int sig __maybe_unused)
  90 {
  91         resize = 1;
  92 }
  93
  94 static void perf_top__resize(struct perf_top *top)
  95 {
  96         get_term_dimensions(&top->winsize);
  97         perf_top__update_print_entries(top);
  98 }
  99
 100 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 101 {
 102         struct perf_evsel *evsel;
 103         struct symbol *sym;
 104         struct annotation *notes;
 105         struct map *map;
 106         int err = -1;
 107
 108         if (!he || !he->ms.sym)
 109                 return -1;
 110
 111         evsel = hists_to_evsel(he->hists);
 112
 113         sym = he->ms.sym;
 114         map = he->ms.map;
 115
 116         /*
 117          * We can't annotate with just /proc/kallsyms
 118          */
 119         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 120             !dso__is_kcore(map->dso)) {
 121                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 122                        "path\n", sym->name);
 123                 sleep(1);
 124                 return -1;
 125         }
 126
 127         notes = symbol__annotation(sym);
 128         pthread_mutex_lock(&notes->lock);
 129
 130         if (!symbol__hists(sym, top->evlist->nr_entries)) {
 131                 pthread_mutex_unlock(&notes->lock);
 132                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 133                        sym->name);
 134                 sleep(1);
 135                 return err;
 136         }
 137
 138         err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
 139         if (err == 0) {
 140                 top->sym_filter_entry = he;
 141         } else {
 142                 char msg[BUFSIZ];
 143                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
 144                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 145         }
 146
 147         pthread_mutex_unlock(&notes->lock);
 148         return err;
 149 }
 150
 151 static void __zero_source_counters(struct hist_entry *he)
 152 {
 153         struct symbol *sym = he->ms.sym;
 154         symbol__annotate_zero_histograms(sym);
 155 }
 156
 157 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 158 {
 159         struct utsname uts;
 160         int err = uname(&uts);
 161
 162         ui__warning("Out of bounds address found:\n\n"
 163                     "Addr:   %" PRIx64 "\n"
 164                     "DSO:    %s %c\n"
 165                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 166                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 167                     "Arch:   %s\n"
 168                     "Kernel: %s\n"
 169                     "Tools:  %s\n\n"
 170                     "Not all samples will be on the annotation output.\n\n"
 171                     "Please report to linux-kernel@vger.kernel.org\n",
 172                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 173                     map->start, map->end, sym->start, sym->end,
 174                     sym->binding == STB_GLOBAL ? 'g' :
 175                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 176                     err ? "[unknown]" : uts.machine,
 177                     err ? "[unknown]" : uts.release, perf_version_string);
 178         if (use_browser <= 0)
 179                 sleep(5);
 180
 181         map->erange_warned = true;
 182 }
 183
 184 static void perf_top__record_precise_ip(struct perf_top *top,
 185                                         struct hist_entry *he,
 186                                         struct perf_sample *sample,
 187                                         struct perf_evsel *evsel, u64 ip)
 188 {
 189         struct annotation *notes;
 190         struct symbol *sym = he->ms.sym;
 191         int err = 0;
 192
 193         if (sym == NULL || (use_browser == 0 &&
 194                             (top->sym_filter_entry == NULL ||
 195                              top->sym_filter_entry->ms.sym != sym)))
 196                 return;
 197
 198         notes = symbol__annotation(sym);
 199
 200         if (pthread_mutex_trylock(&notes->lock))
 201                 return;
 202
 203         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 204
 205         pthread_mutex_unlock(&notes->lock);
 206
 207         if (unlikely(err)) {
 208                 /*
 209                  * This function is now called with he->hists->lock held.
 210                  * Release it before going to sleep.
 211                  */
 212                 pthread_mutex_unlock(&he->hists->lock);
 213
 214                 if (err == -ERANGE && !he->ms.map->erange_warned)
 215                         ui__warn_map_erange(he->ms.map, sym, ip);
 216                 else if (err == -ENOMEM) {
 217                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 218                                sym->name);
 219                         sleep(1);
 220                 }
 221
 222                 pthread_mutex_lock(&he->hists->lock);
 223         }
 224 }
 225
 226 static void perf_top__show_details(struct perf_top *top)
 227 {
 228         struct hist_entry *he = top->sym_filter_entry;
 229         struct perf_evsel *evsel;
 230         struct annotation *notes;
 231         struct symbol *symbol;
 232         int more;
 233
 234         if (!he)
 235                 return;
 236
 237         evsel = hists_to_evsel(he->hists);
 238
 239         symbol = he->ms.sym;
 240         notes = symbol__annotation(symbol);
 241
 242         pthread_mutex_lock(&notes->lock);
 243
 244         symbol__calc_percent(symbol, evsel);
 245
 246         if (notes->src == NULL)
 247                 goto out_unlock;
 248
 249         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 250         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 251
 252         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
 253
 254         if (top->evlist->enabled) {
 255                 if (top->zero)
 256                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 257                 else
 258                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 259         }
 260         if (more != 0)
 261                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 262 out_unlock:
 263         pthread_mutex_unlock(&notes->lock);
 264 }
 265
 266 static void perf_top__print_sym_table(struct perf_top *top)
 267 {
 268         char bf[160];
 269         int printed = 0;
 270         const int win_width = top->winsize.ws_col - 1;
 271         struct perf_evsel *evsel = top->sym_evsel;
 272         struct hists *hists = evsel__hists(evsel);
 273
 274         puts(CONSOLE_CLEAR);
 275
 276         perf_top__header_snprintf(top, bf, sizeof(bf));
 277         printf("%s\n", bf);
 278
 279         perf_top__reset_sample_counters(top);
 280
 281         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 282
 283         if (!top->record_opts.overwrite &&
 284             (hists->stats.nr_lost_warned !=
 285             hists->stats.nr_events[PERF_RECORD_LOST])) {
 286                 hists->stats.nr_lost_warned =
 287                               hists->stats.nr_events[PERF_RECORD_LOST];
 288                 color_fprintf(stdout, PERF_COLOR_RED,
 289                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 290                               hists->stats.nr_lost_warned);
 291                 ++printed;
 292         }
 293
 294         if (top->sym_filter_entry) {
 295                 perf_top__show_details(top);
 296                 return;
 297         }
 298
 299         if (top->evlist->enabled) {
 300                 if (top->zero) {
 301                         hists__delete_entries(hists);
 302                 } else {
 303                         hists__decay_entries(hists, top->hide_user_symbols,
 304                                              top->hide_kernel_symbols);
 305                 }
 306         }
 307
 308         hists__collapse_resort(hists, NULL);
 309         perf_evsel__output_resort(evsel, NULL);
 310
 311         hists__output_recalc_col_len(hists, top->print_entries - printed);
 312         putchar('\n');
 313         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 314                        top->min_percent, stdout, !symbol_conf.use_callchain);
 315 }
 316
 317 static void prompt_integer(int *target, const char *msg)
 318 {
 319         char *buf = malloc(0), *p;
 320         size_t dummy = 0;
 321         int tmp;
 322
 323         fprintf(stdout, "\n%s: ", msg);
 324         if (getline(&buf, &dummy, stdin) < 0)
 325                 return;
 326
 327         p = strchr(buf, '\n');
 328         if (p)
 329                 *p = 0;
 330
 331         p = buf;
 332         while(*p) {
 333                 if (!isdigit(*p))
 334                         goto out_free;
 335                 p++;
 336         }
 337         tmp = strtoul(buf, NULL, 10);
 338         *target = tmp;
 339 out_free:
 340         free(buf);
 341 }
 342
 343 static void prompt_percent(int *target, const char *msg)
 344 {
 345         int tmp = 0;
 346
 347         prompt_integer(&tmp, msg);
 348         if (tmp >= 0 && tmp <= 100)
 349                 *target = tmp;
 350 }
 351
 352 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 353 {
 354         char *buf = malloc(0), *p;
 355         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 356         struct hists *hists = evsel__hists(top->sym_evsel);
 357         struct rb_node *next;
 358         size_t dummy = 0;
 359
 360         /* zero counters of active symbol */
 361         if (syme) {
 362                 __zero_source_counters(syme);
 363                 top->sym_filter_entry = NULL;
 364         }
 365
 366         fprintf(stdout, "\n%s: ", msg);
 367         if (getline(&buf, &dummy, stdin) < 0)
 368                 goto out_free;
 369
 370         p = strchr(buf, '\n');
 371         if (p)
 372                 *p = 0;
 373
 374         next = rb_first(&hists->entries);
 375         while (next) {
 376                 n = rb_entry(next, struct hist_entry, rb_node);
 377                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 378                         found = n;
 379                         break;
 380                 }
 381                 next = rb_next(&n->rb_node);
 382         }
 383
 384         if (!found) {
 385                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 386                 sleep(1);
 387         } else
 388                 perf_top__parse_source(top, found);
 389
 390 out_free:
 391         free(buf);
 392 }
 393
 394 static void perf_top__print_mapped_keys(struct perf_top *top)
 395 {
 396         char *name = NULL;
 397
 398         if (top->sym_filter_entry) {
 399                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 400                 name = sym->name;
 401         }
 402
 403         fprintf(stdout, "\nMapped keys:\n");
 404         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 405         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 406
 407         if (top->evlist->nr_entries > 1)
 408                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 409
 410         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 411
 412         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 413         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 414         fprintf(stdout, "\t[S]     stop annotation.\n");
 415
 416         fprintf(stdout,
 417                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 418                 top->hide_kernel_symbols ? "yes" : "no");
 419         fprintf(stdout,
 420                 "\t[U]     hide user symbols.               \t(%s)\n",
 421                 top->hide_user_symbols ? "yes" : "no");
 422         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 423         fprintf(stdout, "\t[qQ]    quit.\n");
 424 }
 425
 426 static int perf_top__key_mapped(struct perf_top *top, int c)
 427 {
 428         switch (c) {
 429                 case 'd':
 430                 case 'e':
 431                 case 'f':
 432                 case 'z':
 433                 case 'q':
 434                 case 'Q':
 435                 case 'K':
 436                 case 'U':
 437                 case 'F':
 438                 case 's':
 439                 case 'S':
 440                         return 1;
 441                 case 'E':
 442                         return top->evlist->nr_entries > 1 ? 1 : 0;
 443                 default:
 444                         break;
 445         }
 446
 447         return 0;
 448 }
 449
 450 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 451 {
 452         bool ret = true;
 453
 454         if (!perf_top__key_mapped(top, c)) {
 455                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 456                 struct termios save;
 457
 458                 perf_top__print_mapped_keys(top);
 459                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 460                 fflush(stdout);
 461
 462                 set_term_quiet_input(&save);
 463
 464                 poll(&stdin_poll, 1, -1);
 465                 c = getc(stdin);
 466
 467                 tcsetattr(0, TCSAFLUSH, &save);
 468                 if (!perf_top__key_mapped(top, c))
 469                         return ret;
 470         }
 471
 472         switch (c) {
 473                 case 'd':
 474                         prompt_integer(&top->delay_secs, "Enter display delay");
 475                         if (top->delay_secs < 1)
 476                                 top->delay_secs = 1;
 477                         break;
 478                 case 'e':
 479                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 480                         if (top->print_entries == 0) {
 481                                 perf_top__resize(top);
 482                                 signal(SIGWINCH, winch_sig);
 483                         } else {
 484                                 signal(SIGWINCH, SIG_DFL);
 485                         }
 486                         break;
 487                 case 'E':
 488                         if (top->evlist->nr_entries > 1) {
 489                                 /* Select 0 as the default event: */
 490                                 int counter = 0;
 491
 492                                 fprintf(stderr, "\nAvailable events:");
 493
 494                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 495                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 496
 497                                 prompt_integer(&counter, "Enter details event counter");
 498
 499                                 if (counter >= top->evlist->nr_entries) {
 500                                         top->sym_evsel = perf_evlist__first(top->evlist);
 501                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 502                                         sleep(1);
 503                                         break;
 504                                 }
 505                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 506                                         if (top->sym_evsel->idx == counter)
 507                                                 break;
 508                         } else
 509                                 top->sym_evsel = perf_evlist__first(top->evlist);
 510                         break;
 511                 case 'f':
 512                         prompt_integer(&top->count_filter, "Enter display event count filter");
 513                         break;
 514                 case 'F':
 515                         prompt_percent(&top->annotation_opts.min_pcnt,
 516                                        "Enter details display event filter (percent)");
 517                         break;
 518                 case 'K':
 519                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 520                         break;
 521                 case 'q':
 522                 case 'Q':
 523                         printf("exiting.\n");
 524                         if (top->dump_symtab)
 525                                 perf_session__fprintf_dsos(top->session, stderr);
 526                         ret = false;
 527                         break;
 528                 case 's':
 529                         perf_top__prompt_symbol(top, "Enter details symbol");
 530                         break;
 531                 case 'S':
 532                         if (!top->sym_filter_entry)
 533                                 break;
 534                         else {
 535                                 struct hist_entry *syme = top->sym_filter_entry;
 536
 537                                 top->sym_filter_entry = NULL;
 538                                 __zero_source_counters(syme);
 539                         }
 540                         break;
 541                 case 'U':
 542                         top->hide_user_symbols = !top->hide_user_symbols;
 543                         break;
 544                 case 'z':
 545                         top->zero = !top->zero;
 546                         break;
 547                 default:
 548                         break;
 549         }
 550
 551         return ret;
 552 }
 553
 554 static void perf_top__sort_new_samples(void *arg)
 555 {
 556         struct perf_top *t = arg;
 557         struct perf_evsel *evsel = t->sym_evsel;
 558         struct hists *hists;
 559
 560         perf_top__reset_sample_counters(t);
 561
 562         if (t->evlist->selected != NULL)
 563                 t->sym_evsel = t->evlist->selected;
 564
 565         hists = evsel__hists(evsel);
 566
 567         if (t->evlist->enabled) {
 568                 if (t->zero) {
 569                         hists__delete_entries(hists);
 570                 } else {
 571                         hists__decay_entries(hists, t->hide_user_symbols,
 572                                              t->hide_kernel_symbols);
 573                 }
 574         }
 575
 576         hists__collapse_resort(hists, NULL);
 577         perf_evsel__output_resort(evsel, NULL);
 578 }
 579
 580 static void *display_thread_tui(void *arg)
 581 {
 582         struct perf_evsel *pos;
 583         struct perf_top *top = arg;
 584         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 585         struct hist_browser_timer hbt = {
 586                 .timer          = perf_top__sort_new_samples,
 587                 .arg            = top,
 588                 .refresh        = top->delay_secs,
 589         };
 590
 591         /* In order to read symbols from other namespaces perf to  needs to call
 592          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 593          * unshare(2) the fs so that we may continue to setns into namespaces
 594          * that we're observing.
 595          */
 596         unshare(CLONE_FS);
 597
 598         perf_top__sort_new_samples(top);
 599
 600         /*
 601          * Initialize the uid_filter_str, in the future the TUI will allow
 602          * Zooming in/out UIDs. For now juse use whatever the user passed
 603          * via --uid.
 604          */
 605         evlist__for_each_entry(top->evlist, pos) {
 606                 struct hists *hists = evsel__hists(pos);
 607                 hists->uid_filter_str = top->record_opts.target.uid_str;
 608         }
 609
 610         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 611                                       top->min_percent,
 612                                       &top->session->header.env,
 613                                       !top->record_opts.overwrite,
 614                                       &top->annotation_opts);
 615
 616         done = 1;
 617         return NULL;
 618 }
 619
 620 static void display_sig(int sig __maybe_unused)
 621 {
 622         done = 1;
 623 }
 624
 625 static void display_setup_sig(void)
 626 {
 627         signal(SIGSEGV, sighandler_dump_stack);
 628         signal(SIGFPE, sighandler_dump_stack);
 629         signal(SIGINT,  display_sig);
 630         signal(SIGQUIT, display_sig);
 631         signal(SIGTERM, display_sig);
 632 }
 633
 634 static void *display_thread(void *arg)
 635 {
 636         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 637         struct termios save;
 638         struct perf_top *top = arg;
 639         int delay_msecs, c;
 640
 641         /* In order to read symbols from other namespaces perf to  needs to call
 642          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 643          * unshare(2) the fs so that we may continue to setns into namespaces
 644          * that we're observing.
 645          */
 646         unshare(CLONE_FS);
 647
 648         display_setup_sig();
 649         pthread__unblock_sigwinch();
 650 repeat:
 651         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 652         set_term_quiet_input(&save);
 653         /* trash return*/
 654         clearerr(stdin);
 655         if (poll(&stdin_poll, 1, 0) > 0)
 656                 getc(stdin);
 657
 658         while (!done) {
 659                 perf_top__print_sym_table(top);
 660                 /*
 661                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 662                  * refresh screen in both cases.
 663                  */
 664                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 665                 case 0:
 666                         continue;
 667                 case -1:
 668                         if (errno == EINTR)
 669                                 continue;
 670                         __fallthrough;
 671                 default:
 672                         c = getc(stdin);
 673                         tcsetattr(0, TCSAFLUSH, &save);
 674
 675                         if (perf_top__handle_keypress(top, c))
 676                                 goto repeat;
 677                         done = 1;
 678                 }
 679         }
 680
 681         tcsetattr(0, TCSAFLUSH, &save);
 682         return NULL;
 683 }
 684
 685 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 686                                    struct addr_location *al, bool single,
 687                                    void *arg)
 688 {
 689         struct perf_top *top = arg;
 690         struct hist_entry *he = iter->he;
 691         struct perf_evsel *evsel = iter->evsel;
 692
 693         if (perf_hpp_list.sym && single)
 694                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 695
 696         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 697                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 698         return 0;
 699 }
 700
 701 static void perf_event__process_sample(struct perf_tool *tool,
 702                                        const union perf_event *event,
 703                                        struct perf_evsel *evsel,
 704                                        struct perf_sample *sample,
 705                                        struct machine *machine)
 706 {
 707         struct perf_top *top = container_of(tool, struct perf_top, tool);
 708         struct addr_location al;
 709         int err;
 710
 711         if (!machine && perf_guest) {
 712                 static struct intlist *seen;
 713
 714                 if (!seen)
 715                         seen = intlist__new(NULL);
 716
 717                 if (!intlist__has_entry(seen, sample->pid)) {
 718                         pr_err("Can't find guest [%d]'s kernel information\n",
 719                                 sample->pid);
 720                         intlist__add(seen, sample->pid);
 721                 }
 722                 return;
 723         }
 724
 725         if (!machine) {
 726                 pr_err("%u unprocessable samples recorded.\r",
 727                        top->session->evlist->stats.nr_unprocessable_samples++);
 728                 return;
 729         }
 730
 731         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 732                 top->exact_samples++;
 733
 734         if (machine__resolve(machine, &al, sample) < 0)
 735                 return;
 736
 737         if (!machine->kptr_restrict_warned &&
 738             symbol_conf.kptr_restrict &&
 739             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 740                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 741                         ui__warning(
 742 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 743 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 744 "Kernel%s samples will not be resolved.\n",
 745                           al.map && map__has_symbols(al.map) ?
 746                           " modules" : "");
 747                         if (use_browser <= 0)
 748                                 sleep(5);
 749                 }
 750                 machine->kptr_restrict_warned = true;
 751         }
 752
 753         if (al.sym == NULL && al.map != NULL) {
 754                 const char *msg = "Kernel samples will not be resolved.\n";
 755                 /*
 756                  * As we do lazy loading of symtabs we only will know if the
 757                  * specified vmlinux file is invalid when we actually have a
 758                  * hit in kernel space and then try to load it. So if we get
 759                  * here and there are _no_ symbols in the DSO backing the
 760                  * kernel map, bail out.
 761                  *
 762                  * We may never get here, for instance, if we use -K/
 763                  * --hide-kernel-symbols, even if the user specifies an
 764                  * invalid --vmlinux ;-)
 765                  */
 766                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 767                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 768                         if (symbol_conf.vmlinux_name) {
 769                                 char serr[256];
 770                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 771                                 ui__warning("The %s file can't be used: %s\n%s",
 772                                             symbol_conf.vmlinux_name, serr, msg);
 773                         } else {
 774                                 ui__warning("A vmlinux file was not found.\n%s",
 775                                             msg);
 776                         }
 777
 778                         if (use_browser <= 0)
 779                                 sleep(5);
 780                         top->vmlinux_warned = true;
 781                 }
 782         }
 783
 784         if (al.sym == NULL || !al.sym->idle) {
 785                 struct hists *hists = evsel__hists(evsel);
 786                 struct hist_entry_iter iter = {
 787                         .evsel          = evsel,
 788                         .sample         = sample,
 789                         .add_entry_cb   = hist_iter__top_callback,
 790                 };
 791
 792                 if (symbol_conf.cumulate_callchain)
 793                         iter.ops = &hist_iter_cumulative;
 794                 else
 795                         iter.ops = &hist_iter_normal;
 796
 797                 pthread_mutex_lock(&hists->lock);
 798
 799                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 800                 if (err < 0)
 801                         pr_err("Problem incrementing symbol period, skipping event\n");
 802
 803                 pthread_mutex_unlock(&hists->lock);
 804         }
 805
 806         addr_location__put(&al);
 807 }
 808
 809 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 810 {
 811         struct record_opts *opts = &top->record_opts;
 812         struct perf_evlist *evlist = top->evlist;
 813         struct perf_sample sample;
 814         struct perf_evsel *evsel;
 815         struct perf_mmap *md;
 816         struct perf_session *session = top->session;
 817         union perf_event *event;
 818         struct machine *machine;
 819         int ret;
 820
 821         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 822         if (perf_mmap__read_init(md) < 0)
 823                 return;
 824
 825         while ((event = perf_mmap__read_event(md)) != NULL) {
 826                 ret = perf_evlist__parse_sample(evlist, event, &sample);
 827                 if (ret) {
 828                         pr_err("Can't parse sample, err = %d\n", ret);
 829                         goto next_event;
 830                 }
 831
 832                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 833                 assert(evsel != NULL);
 834
 835                 if (event->header.type == PERF_RECORD_SAMPLE)
 836                         ++top->samples;
 837
 838                 switch (sample.cpumode) {
 839                 case PERF_RECORD_MISC_USER:
 840                         ++top->us_samples;
 841                         if (top->hide_user_symbols)
 842                                 goto next_event;
 843                         machine = &session->machines.host;
 844                         break;
 845                 case PERF_RECORD_MISC_KERNEL:
 846                         ++top->kernel_samples;
 847                         if (top->hide_kernel_symbols)
 848                                 goto next_event;
 849                         machine = &session->machines.host;
 850                         break;
 851                 case PERF_RECORD_MISC_GUEST_KERNEL:
 852                         ++top->guest_kernel_samples;
 853                         machine = perf_session__find_machine(session,
 854                                                              sample.pid);
 855                         break;
 856                 case PERF_RECORD_MISC_GUEST_USER:
 857                         ++top->guest_us_samples;
 858                         /*
 859                          * TODO: we don't process guest user from host side
 860                          * except simple counting.
 861                          */
 862                         goto next_event;
 863                 default:
 864                         if (event->header.type == PERF_RECORD_SAMPLE)
 865                                 goto next_event;
 866                         machine = &session->machines.host;
 867                         break;
 868                 }
 869
 870
 871                 if (event->header.type == PERF_RECORD_SAMPLE) {
 872                         perf_event__process_sample(&top->tool, event, evsel,
 873                                                    &sample, machine);
 874                 } else if (event->header.type < PERF_RECORD_MAX) {
 875                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 876                         machine__process_event(machine, event, &sample);
 877                 } else
 878                         ++session->evlist->stats.nr_unknown_events;
 879 next_event:
 880                 perf_mmap__consume(md);
 881         }
 882
 883         perf_mmap__read_done(md);
 884 }
 885
 886 static void perf_top__mmap_read(struct perf_top *top)
 887 {
 888         bool overwrite = top->record_opts.overwrite;
 889         struct perf_evlist *evlist = top->evlist;
 890         unsigned long long start, end;
 891         int i;
 892
 893         start = rdclock();
 894         if (overwrite)
 895                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 896
 897         for (i = 0; i < top->evlist->nr_mmaps; i++)
 898                 perf_top__mmap_read_idx(top, i);
 899
 900         if (overwrite) {
 901                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 902                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 903         }
 904         end = rdclock();
 905
 906         if ((end - start) > (unsigned long long)top->delay_secs * NSEC_PER_SEC)
 907                 ui__warning("Too slow to read ring buffer.\n"
 908                             "Please try increasing the period (-c) or\n"
 909                             "decreasing the freq (-F) or\n"
 910                             "limiting the number of CPUs (-C)\n");
 911 }
 912
 913 /*
 914  * Check per-event overwrite term.
 915  * perf top should support consistent term for all events.
 916  * - All events don't have per-event term
 917  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 918  *   Nothing change, return 0.
 919  * - All events have same per-event term
 920  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 921  *   Using the per-event setting to replace the opts->overwrite if
 922  *   they are different, then return 0.
 923  * - Events have different per-event term
 924  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 925  *   Return -1
 926  * - Some of the event set per-event term, but some not.
 927  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 928  *   Return -1
 929  */
 930 static int perf_top__overwrite_check(struct perf_top *top)
 931 {
 932         struct record_opts *opts = &top->record_opts;
 933         struct perf_evlist *evlist = top->evlist;
 934         struct perf_evsel_config_term *term;
 935         struct list_head *config_terms;
 936         struct perf_evsel *evsel;
 937         int set, overwrite = -1;
 938
 939         evlist__for_each_entry(evlist, evsel) {
 940                 set = -1;
 941                 config_terms = &evsel->config_terms;
 942                 list_for_each_entry(term, config_terms, list) {
 943                         if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
 944                                 set = term->val.overwrite ? 1 : 0;
 945                 }
 946
 947                 /* no term for current and previous event (likely) */
 948                 if ((overwrite < 0) && (set < 0))
 949                         continue;
 950
 951                 /* has term for both current and previous event, compare */
 952                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 953                         return -1;
 954
 955                 /* no term for current event but has term for previous one */
 956                 if ((overwrite >= 0) && (set < 0))
 957                         return -1;
 958
 959                 /* has term for current event */
 960                 if ((overwrite < 0) && (set >= 0)) {
 961                         /* if it's first event, set overwrite */
 962                         if (evsel == perf_evlist__first(evlist))
 963                                 overwrite = set;
 964                         else
 965                                 return -1;
 966                 }
 967         }
 968
 969         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 970                 opts->overwrite = overwrite;
 971
 972         return 0;
 973 }
 974
 975 static int perf_top_overwrite_fallback(struct perf_top *top,
 976                                        struct perf_evsel *evsel)
 977 {
 978         struct record_opts *opts = &top->record_opts;
 979         struct perf_evlist *evlist = top->evlist;
 980         struct perf_evsel *counter;
 981
 982         if (!opts->overwrite)
 983                 return 0;
 984
 985         /* only fall back when first event fails */
 986         if (evsel != perf_evlist__first(evlist))
 987                 return 0;
 988
 989         evlist__for_each_entry(evlist, counter)
 990                 counter->attr.write_backward = false;
 991         opts->overwrite = false;
 992         pr_debug2("fall back to non-overwrite mode\n");
 993         return 1;
 994 }
 995
 996 static int perf_top__start_counters(struct perf_top *top)
 997 {
 998         char msg[BUFSIZ];
 999         struct perf_evsel *counter;
1000         struct perf_evlist *evlist = top->evlist;
1001         struct record_opts *opts = &top->record_opts;
1002
1003         if (perf_top__overwrite_check(top)) {
1004                 ui__error("perf top only support consistent per-event "
1005                           "overwrite setting for all events\n");
1006                 goto out_err;
1007         }
1008
1009         perf_evlist__config(evlist, opts, &callchain_param);
1010
1011         evlist__for_each_entry(evlist, counter) {
1012 try_again:
1013                 if (perf_evsel__open(counter, top->evlist->cpus,
1014                                      top->evlist->threads) < 0) {
1015
1016                         /*
1017                          * Specially handle overwrite fall back.
1018                          * Because perf top is the only tool which has
1019                          * overwrite mode by default, support
1020                          * both overwrite and non-overwrite mode, and
1021                          * require consistent mode for all events.
1022                          *
1023                          * May move it to generic code with more tools
1024                          * have similar attribute.
1025                          */
1026                         if (perf_missing_features.write_backward &&
1027                             perf_top_overwrite_fallback(top, counter))
1028                                 goto try_again;
1029
1030                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1031                                 if (verbose > 0)
1032                                         ui__warning("%s\n", msg);
1033                                 goto try_again;
1034                         }
1035
1036                         perf_evsel__open_strerror(counter, &opts->target,
1037                                                   errno, msg, sizeof(msg));
1038                         ui__error("%s\n", msg);
1039                         goto out_err;
1040                 }
1041         }
1042
1043         if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1044                 ui__error("Failed to mmap with %d (%s)\n",
1045                             errno, str_error_r(errno, msg, sizeof(msg)));
1046                 goto out_err;
1047         }
1048
1049         return 0;
1050
1051 out_err:
1052         return -1;
1053 }
1054
1055 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1056 {
1057         if (!perf_hpp_list.sym) {
1058                 if (callchain->enabled) {
1059                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1060                         return -EINVAL;
1061                 }
1062         } else if (callchain->mode != CHAIN_NONE) {
1063                 if (callchain_register_param(callchain) < 0) {
1064                         ui__error("Can't register callchain params.\n");
1065                         return -EINVAL;
1066                 }
1067         }
1068
1069         return 0;
1070 }
1071
1072 static int __cmd_top(struct perf_top *top)
1073 {
1074         char msg[512];
1075         struct perf_evsel *pos;
1076         struct perf_evsel_config_term *err_term;
1077         struct perf_evlist *evlist = top->evlist;
1078         struct record_opts *opts = &top->record_opts;
1079         pthread_t thread;
1080         int ret;
1081
1082         top->session = perf_session__new(NULL, false, NULL);
1083         if (top->session == NULL)
1084                 return -1;
1085
1086         if (!top->annotation_opts.objdump_path) {
1087                 ret = perf_env__lookup_objdump(&top->session->header.env,
1088                                                &top->annotation_opts.objdump_path);
1089                 if (ret)
1090                         goto out_delete;
1091         }
1092
1093         ret = callchain_param__setup_sample_type(&callchain_param);
1094         if (ret)
1095                 goto out_delete;
1096
1097         if (perf_session__register_idle_thread(top->session) < 0)
1098                 goto out_delete;
1099
1100         if (top->nr_threads_synthesize > 1)
1101                 perf_set_multithreaded();
1102
1103         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1104                                     top->evlist->threads, false,
1105                                     opts->proc_map_timeout,
1106                                     top->nr_threads_synthesize);
1107
1108         if (top->nr_threads_synthesize > 1)
1109                 perf_set_singlethreaded();
1110
1111         if (perf_hpp_list.socket) {
1112                 ret = perf_env__read_cpu_topology_map(&perf_env);
1113                 if (ret < 0)
1114                         goto out_err_cpu_topo;
1115         }
1116
1117         ret = perf_top__start_counters(top);
1118         if (ret)
1119                 goto out_delete;
1120
1121         ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1122         if (ret) {
1123                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1124                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1125                         str_error_r(errno, msg, sizeof(msg)));
1126                 goto out_delete;
1127         }
1128
1129         top->session->evlist = top->evlist;
1130         perf_session__set_id_hdr_size(top->session);
1131
1132         /*
1133          * When perf is starting the traced process, all the events (apart from
1134          * group members) have enable_on_exec=1 set, so don't spoil it by
1135          * prematurely enabling them.
1136          *
1137          * XXX 'top' still doesn't start workloads like record, trace, but should,
1138          * so leave the check here.
1139          */
1140         if (!target__none(&opts->target))
1141                 perf_evlist__enable(top->evlist);
1142
1143         /* Wait for a minimal set of events before starting the snapshot */
1144         perf_evlist__poll(top->evlist, 100);
1145
1146         perf_top__mmap_read(top);
1147
1148         ret = -1;
1149         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1150                                                             display_thread), top)) {
1151                 ui__error("Could not create display thread.\n");
1152                 goto out_delete;
1153         }
1154
1155         if (top->realtime_prio) {
1156                 struct sched_param param;
1157
1158                 param.sched_priority = top->realtime_prio;
1159                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1160                         ui__error("Could not set realtime priority.\n");
1161                         goto out_join;
1162                 }
1163         }
1164
1165         while (!done) {
1166                 u64 hits = top->samples;
1167
1168                 perf_top__mmap_read(top);
1169
1170                 if (opts->overwrite || (hits == top->samples))
1171                         ret = perf_evlist__poll(top->evlist, 100);
1172
1173                 if (resize) {
1174                         perf_top__resize(top);
1175                         resize = 0;
1176                 }
1177         }
1178
1179         ret = 0;
1180 out_join:
1181         pthread_join(thread, NULL);
1182 out_delete:
1183         perf_session__delete(top->session);
1184         top->session = NULL;
1185
1186         return ret;
1187
1188 out_err_cpu_topo: {
1189         char errbuf[BUFSIZ];
1190         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1191
1192         ui__error("Could not read the CPU topology map: %s\n", err);
1193         goto out_delete;
1194 }
1195 }
1196
1197 static int
1198 callchain_opt(const struct option *opt, const char *arg, int unset)
1199 {
1200         symbol_conf.use_callchain = true;
1201         return record_callchain_opt(opt, arg, unset);
1202 }
1203
1204 static int
1205 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1206 {
1207         struct callchain_param *callchain = opt->value;
1208
1209         callchain->enabled = !unset;
1210         callchain->record_mode = CALLCHAIN_FP;
1211
1212         /*
1213          * --no-call-graph
1214          */
1215         if (unset) {
1216                 symbol_conf.use_callchain = false;
1217                 callchain->record_mode = CALLCHAIN_NONE;
1218                 return 0;
1219         }
1220
1221         return parse_callchain_top_opt(arg);
1222 }
1223
1224 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1225 {
1226         if (!strcmp(var, "top.call-graph")) {
1227                 var = "call-graph.record-mode";
1228                 return perf_default_config(var, value, cb);
1229         }
1230         if (!strcmp(var, "top.children")) {
1231                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1232                 return 0;
1233         }
1234
1235         return 0;
1236 }
1237
1238 static int
1239 parse_percent_limit(const struct option *opt, const char *arg,
1240                     int unset __maybe_unused)
1241 {
1242         struct perf_top *top = opt->value;
1243
1244         top->min_percent = strtof(arg, NULL);
1245         return 0;
1246 }
1247
1248 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1249         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1250
1251 int cmd_top(int argc, const char **argv)
1252 {
1253         char errbuf[BUFSIZ];
1254         struct perf_top top = {
1255                 .count_filter        = 5,
1256                 .delay_secs          = 2,
1257                 .record_opts = {
1258                         .mmap_pages     = UINT_MAX,
1259                         .user_freq      = UINT_MAX,
1260                         .user_interval  = ULLONG_MAX,
1261                         .freq           = 4000, /* 4 KHz */
1262                         .target         = {
1263                                 .uses_mmap   = true,
1264                         },
1265                         .proc_map_timeout    = 500,
1266                         .overwrite      = 1,
1267                 },
1268                 .max_stack           = sysctl__max_stack(),
1269                 .annotation_opts     = annotation__default_options,
1270                 .nr_threads_synthesize = UINT_MAX,
1271         };
1272         struct record_opts *opts = &top.record_opts;
1273         struct target *target = &opts->target;
1274         const struct option options[] = {
1275         OPT_CALLBACK('e', "event", &top.evlist, "event",
1276                      "event selector. use 'perf list' to list available events",
1277                      parse_events_option),
1278         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1279         OPT_STRING('p', "pid", &target->pid, "pid",
1280                     "profile events on existing process id"),
1281         OPT_STRING('t', "tid", &target->tid, "tid",
1282                     "profile events on existing thread id"),
1283         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1284                             "system-wide collection from all CPUs"),
1285         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1286                     "list of cpus to monitor"),
1287         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1288                    "file", "vmlinux pathname"),
1289         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1290                     "don't load vmlinux even if found"),
1291         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1292                     "hide kernel symbols"),
1293         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1294                      "number of mmap data pages",
1295                      perf_evlist__parse_mmap_pages),
1296         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1297                     "collect data with this RT SCHED_FIFO priority"),
1298         OPT_INTEGER('d', "delay", &top.delay_secs,
1299                     "number of seconds to delay between refreshes"),
1300         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1301                             "dump the symbol table used for profiling"),
1302         OPT_INTEGER('f', "count-filter", &top.count_filter,
1303                     "only display functions with more events than this"),
1304         OPT_BOOLEAN(0, "group", &opts->group,
1305                             "put the counters into a counter group"),
1306         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1307                     "child tasks do not inherit counters"),
1308         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1309                     "symbol to annotate"),
1310         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1311         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1312                      "profile at this frequency",
1313                       record__parse_freq),
1314         OPT_INTEGER('E', "entries", &top.print_entries,
1315                     "display this many functions"),
1316         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1317                     "hide user symbols"),
1318         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1319         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1320         OPT_INCR('v', "verbose", &verbose,
1321                     "be more verbose (show counter open errors, etc)"),
1322         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1323                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1324                    " Please refer the man page for the complete list."),
1325         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1326                    "output field(s): overhead, period, sample plus all of sort keys"),
1327         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1328                     "Show a column with the number of samples"),
1329         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1330                            NULL, "enables call-graph recording and display",
1331                            &callchain_opt),
1332         OPT_CALLBACK(0, "call-graph", &callchain_param,
1333                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1334                      top_callchain_help, &parse_callchain_opt),
1335         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1336                     "Accumulate callchains of children and show total overhead as well"),
1337         OPT_INTEGER(0, "max-stack", &top.max_stack,
1338                     "Set the maximum stack depth when parsing the callchain. "
1339                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1340         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1341                    "ignore callees of these functions in call graphs",
1342                    report_parse_ignore_callees_opt),
1343         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1344                     "Show a column with the sum of periods"),
1345         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1346                    "only consider symbols in these dsos"),
1347         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1348                    "only consider symbols in these comms"),
1349         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1350                    "only consider these symbols"),
1351         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1352                     "Interleave source code with assembly code (default)"),
1353         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1354                     "Display raw encoding of assembly instructions (default)"),
1355         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1356                     "Enable kernel symbol demangling"),
1357         OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1358                     "objdump binary to use for disassembly and annotations"),
1359         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1360                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1361         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1362         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1363                      "Don't show entries under that percent", parse_percent_limit),
1364         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1365                      "How to display percentage of filtered entries", parse_filter_percentage),
1366         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1367                    "width[,width...]",
1368                    "don't try to adjust column width, use these fixed values"),
1369         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1370                         "per thread proc mmap processing timeout in ms"),
1371         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1372                      "branch any", "sample any taken branches",
1373                      parse_branch_stack),
1374         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1375                      "branch filter mask", "branch stack filter modes",
1376                      parse_branch_stack),
1377         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1378                     "Show raw trace event output (do not use print fmt or plugins)"),
1379         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1380                     "Show entries in a hierarchy"),
1381         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1382         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1383                         "number of thread to run event synthesize"),
1384         OPT_END()
1385         };
1386         const char * const top_usage[] = {
1387                 "perf top [<options>]",
1388                 NULL
1389         };
1390         int status = hists__init();
1391
1392         if (status < 0)
1393                 return status;
1394
1395         top.annotation_opts.min_pcnt = 5;
1396         top.annotation_opts.context  = 4;
1397
1398         top.evlist = perf_evlist__new();
1399         if (top.evlist == NULL)
1400                 return -ENOMEM;
1401
1402         status = perf_config(perf_top_config, &top);
1403         if (status)
1404                 return status;
1405
1406         argc = parse_options(argc, argv, options, top_usage, 0);
1407         if (argc)
1408                 usage_with_options(top_usage, options);
1409
1410         if (!top.evlist->nr_entries &&
1411             perf_evlist__add_default(top.evlist) < 0) {
1412                 pr_err("Not enough memory for event selector list\n");
1413                 goto out_delete_evlist;
1414         }
1415
1416         if (symbol_conf.report_hierarchy) {
1417                 /* disable incompatible options */
1418                 symbol_conf.event_group = false;
1419                 symbol_conf.cumulate_callchain = false;
1420
1421                 if (field_order) {
1422                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1423                         parse_options_usage(top_usage, options, "fields", 0);
1424                         parse_options_usage(NULL, options, "hierarchy", 0);
1425                         goto out_delete_evlist;
1426                 }
1427         }
1428
1429         sort__mode = SORT_MODE__TOP;
1430         /* display thread wants entries to be collapsed in a different tree */
1431         perf_hpp_list.need_collapse = 1;
1432
1433         if (top.use_stdio)
1434                 use_browser = 0;
1435         else if (top.use_tui)
1436                 use_browser = 1;
1437
1438         setup_browser(false);
1439
1440         if (setup_sorting(top.evlist) < 0) {
1441                 if (sort_order)
1442                         parse_options_usage(top_usage, options, "s", 1);
1443                 if (field_order)
1444                         parse_options_usage(sort_order ? NULL : top_usage,
1445                                             options, "fields", 0);
1446                 goto out_delete_evlist;
1447         }
1448
1449         status = target__validate(target);
1450         if (status) {
1451                 target__strerror(target, status, errbuf, BUFSIZ);
1452                 ui__warning("%s\n", errbuf);
1453         }
1454
1455         status = target__parse_uid(target);
1456         if (status) {
1457                 int saved_errno = errno;
1458
1459                 target__strerror(target, status, errbuf, BUFSIZ);
1460                 ui__error("%s\n", errbuf);
1461
1462                 status = -saved_errno;
1463                 goto out_delete_evlist;
1464         }
1465
1466         if (target__none(target))
1467                 target->system_wide = true;
1468
1469         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1470                 ui__error("Couldn't create thread/CPU maps: %s\n",
1471                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1472                 goto out_delete_evlist;
1473         }
1474
1475         if (top.delay_secs < 1)
1476                 top.delay_secs = 1;
1477
1478         if (record_opts__config(opts)) {
1479                 status = -EINVAL;
1480                 goto out_delete_evlist;
1481         }
1482
1483         top.sym_evsel = perf_evlist__first(top.evlist);
1484
1485         if (!callchain_param.enabled) {
1486                 symbol_conf.cumulate_callchain = false;
1487                 perf_hpp__cancel_cumulate();
1488         }
1489
1490         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1491                 callchain_param.order = ORDER_CALLER;
1492
1493         status = symbol__annotation_init();
1494         if (status < 0)
1495                 goto out_delete_evlist;
1496
1497         annotation_config__init();
1498
1499         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1500         status = symbol__init(NULL);
1501         if (status < 0)
1502                 goto out_delete_evlist;
1503
1504         sort__setup_elide(stdout);
1505
1506         get_term_dimensions(&top.winsize);
1507         if (top.print_entries == 0) {
1508                 perf_top__update_print_entries(&top);
1509                 signal(SIGWINCH, winch_sig);
1510         }
1511
1512         status = __cmd_top(&top);
1513
1514 out_delete_evlist:
1515         perf_evlist__delete(top.evlist);
1516
1517         return status;
1518 }