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/cache.h"
  26 #include "util/color.h"
  27 #include "util/evlist.h"
  28 #include "util/evsel.h"
  29 #include "util/machine.h"
  30 #include "util/session.h"
  31 #include "util/symbol.h"
  32 #include "util/thread.h"
  33 #include "util/thread_map.h"
  34 #include "util/top.h"
  35 #include "util/util.h"
  36 #include <linux/rbtree.h>
  37 #include "util/parse-options.h"
  38 #include "util/parse-events.h"
  39 #include "util/cpumap.h"
  40 #include "util/xyarray.h"
  41 #include "util/sort.h"
  42 #include "util/intlist.h"
  43
  44 #include "util/debug.h"
  45
  46 #include <assert.h>
  47 #include <elf.h>
  48 #include <fcntl.h>
  49
  50 #include <stdio.h>
  51 #include <termios.h>
  52 #include <unistd.h>
  53 #include <inttypes.h>
  54
  55 #include <errno.h>
  56 #include <time.h>
  57 #include <sched.h>
  58
  59 #include <sys/syscall.h>
  60 #include <sys/ioctl.h>
  61 #include <sys/poll.h>
  62 #include <sys/prctl.h>
  63 #include <sys/wait.h>
  64 #include <sys/uio.h>
  65 #include <sys/utsname.h>
  66 #include <sys/mman.h>
  67
  68 #include <linux/unistd.h>
  69 #include <linux/types.h>
  70
  71 static void perf_top__update_print_entries(struct perf_top *top)
  72 {
  73         if (top->print_entries > 9)
  74                 top->print_entries -= 9;
  75 }
  76
  77 static void perf_top__sig_winch(int sig __maybe_unused,
  78                                 siginfo_t *info __maybe_unused, void *arg)
  79 {
  80         struct perf_top *top = arg;
  81
  82         get_term_dimensions(&top->winsize);
  83         if (!top->print_entries
  84             || (top->print_entries+4) > top->winsize.ws_row) {
  85                 top->print_entries = top->winsize.ws_row;
  86         } else {
  87                 top->print_entries += 4;
  88                 top->winsize.ws_row = top->print_entries;
  89         }
  90         perf_top__update_print_entries(top);
  91 }
  92
  93 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  94 {
  95         struct symbol *sym;
  96         struct annotation *notes;
  97         struct map *map;
  98         int err = -1;
  99
 100         if (!he || !he->ms.sym)
 101                 return -1;
 102
 103         sym = he->ms.sym;
 104         map = he->ms.map;
 105
 106         /*
 107          * We can't annotate with just /proc/kallsyms
 108          */
 109         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
 110                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 111                        "path\n", sym->name);
 112                 sleep(1);
 113                 return -1;
 114         }
 115
 116         notes = symbol__annotation(sym);
 117         if (notes->src != NULL) {
 118                 pthread_mutex_lock(&notes->lock);
 119                 goto out_assign;
 120         }
 121
 122         pthread_mutex_lock(&notes->lock);
 123
 124         if (symbol__alloc_hist(sym) < 0) {
 125                 pthread_mutex_unlock(&notes->lock);
 126                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 127                        sym->name);
 128                 sleep(1);
 129                 return err;
 130         }
 131
 132         err = symbol__annotate(sym, map, 0);
 133         if (err == 0) {
 134 out_assign:
 135                 top->sym_filter_entry = he;
 136         }
 137
 138         pthread_mutex_unlock(&notes->lock);
 139         return err;
 140 }
 141
 142 static void __zero_source_counters(struct hist_entry *he)
 143 {
 144         struct symbol *sym = he->ms.sym;
 145         symbol__annotate_zero_histograms(sym);
 146 }
 147
 148 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 149 {
 150         struct utsname uts;
 151         int err = uname(&uts);
 152
 153         ui__warning("Out of bounds address found:\n\n"
 154                     "Addr:   %" PRIx64 "\n"
 155                     "DSO:    %s %c\n"
 156                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 157                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 158                     "Arch:   %s\n"
 159                     "Kernel: %s\n"
 160                     "Tools:  %s\n\n"
 161                     "Not all samples will be on the annotation output.\n\n"
 162                     "Please report to linux-kernel@vger.kernel.org\n",
 163                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 164                     map->start, map->end, sym->start, sym->end,
 165                     sym->binding == STB_GLOBAL ? 'g' :
 166                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 167                     err ? "[unknown]" : uts.machine,
 168                     err ? "[unknown]" : uts.release, perf_version_string);
 169         if (use_browser <= 0)
 170                 sleep(5);
 171
 172         map->erange_warned = true;
 173 }
 174
 175 static void perf_top__record_precise_ip(struct perf_top *top,
 176                                         struct hist_entry *he,
 177                                         int counter, u64 ip)
 178 {
 179         struct annotation *notes;
 180         struct symbol *sym;
 181         int err;
 182
 183         if (he == NULL || he->ms.sym == NULL ||
 184             ((top->sym_filter_entry == NULL ||
 185               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 186                 return;
 187
 188         sym = he->ms.sym;
 189         notes = symbol__annotation(sym);
 190
 191         if (pthread_mutex_trylock(&notes->lock))
 192                 return;
 193
 194         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 195                 pthread_mutex_unlock(&notes->lock);
 196                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 197                        sym->name);
 198                 sleep(1);
 199                 return;
 200         }
 201
 202         ip = he->ms.map->map_ip(he->ms.map, ip);
 203         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 204
 205         pthread_mutex_unlock(&notes->lock);
 206
 207         if (err == -ERANGE && !he->ms.map->erange_warned)
 208                 ui__warn_map_erange(he->ms.map, sym, ip);
 209 }
 210
 211 static void perf_top__show_details(struct perf_top *top)
 212 {
 213         struct hist_entry *he = top->sym_filter_entry;
 214         struct annotation *notes;
 215         struct symbol *symbol;
 216         int more;
 217
 218         if (!he)
 219                 return;
 220
 221         symbol = he->ms.sym;
 222         notes = symbol__annotation(symbol);
 223
 224         pthread_mutex_lock(&notes->lock);
 225
 226         if (notes->src == NULL)
 227                 goto out_unlock;
 228
 229         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 230         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 231
 232         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
 233                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 234         if (top->zero)
 235                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 236         else
 237                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 238         if (more != 0)
 239                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 240 out_unlock:
 241         pthread_mutex_unlock(&notes->lock);
 242 }
 243
 244 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 245
 246 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 247                                                      struct addr_location *al,
 248                                                      struct perf_sample *sample)
 249 {
 250         struct hist_entry *he;
 251
 252         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 253         if (he == NULL)
 254                 return NULL;
 255
 256         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 257         return he;
 258 }
 259
 260 static void perf_top__print_sym_table(struct perf_top *top)
 261 {
 262         char bf[160];
 263         int printed = 0;
 264         const int win_width = top->winsize.ws_col - 1;
 265
 266         puts(CONSOLE_CLEAR);
 267
 268         perf_top__header_snprintf(top, bf, sizeof(bf));
 269         printf("%s\n", bf);
 270
 271         perf_top__reset_sample_counters(top);
 272
 273         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 274
 275         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 276             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 277                 top->sym_evsel->hists.stats.nr_lost_warned =
 278                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 279                 color_fprintf(stdout, PERF_COLOR_RED,
 280                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 281                               top->sym_evsel->hists.stats.nr_lost_warned);
 282                 ++printed;
 283         }
 284
 285         if (top->sym_filter_entry) {
 286                 perf_top__show_details(top);
 287                 return;
 288         }
 289
 290         hists__collapse_resort_threaded(&top->sym_evsel->hists);
 291         hists__output_resort_threaded(&top->sym_evsel->hists);
 292         hists__decay_entries_threaded(&top->sym_evsel->hists,
 293                                       top->hide_user_symbols,
 294                                       top->hide_kernel_symbols);
 295         hists__output_recalc_col_len(&top->sym_evsel->hists,
 296                                      top->winsize.ws_row - 3);
 297         putchar('\n');
 298         hists__fprintf(&top->sym_evsel->hists, false,
 299                        top->winsize.ws_row - 4 - printed, win_width, stdout);
 300 }
 301
 302 static void prompt_integer(int *target, const char *msg)
 303 {
 304         char *buf = malloc(0), *p;
 305         size_t dummy = 0;
 306         int tmp;
 307
 308         fprintf(stdout, "\n%s: ", msg);
 309         if (getline(&buf, &dummy, stdin) < 0)
 310                 return;
 311
 312         p = strchr(buf, '\n');
 313         if (p)
 314                 *p = 0;
 315
 316         p = buf;
 317         while(*p) {
 318                 if (!isdigit(*p))
 319                         goto out_free;
 320                 p++;
 321         }
 322         tmp = strtoul(buf, NULL, 10);
 323         *target = tmp;
 324 out_free:
 325         free(buf);
 326 }
 327
 328 static void prompt_percent(int *target, const char *msg)
 329 {
 330         int tmp = 0;
 331
 332         prompt_integer(&tmp, msg);
 333         if (tmp >= 0 && tmp <= 100)
 334                 *target = tmp;
 335 }
 336
 337 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 338 {
 339         char *buf = malloc(0), *p;
 340         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 341         struct rb_node *next;
 342         size_t dummy = 0;
 343
 344         /* zero counters of active symbol */
 345         if (syme) {
 346                 __zero_source_counters(syme);
 347                 top->sym_filter_entry = NULL;
 348         }
 349
 350         fprintf(stdout, "\n%s: ", msg);
 351         if (getline(&buf, &dummy, stdin) < 0)
 352                 goto out_free;
 353
 354         p = strchr(buf, '\n');
 355         if (p)
 356                 *p = 0;
 357
 358         next = rb_first(&top->sym_evsel->hists.entries);
 359         while (next) {
 360                 n = rb_entry(next, struct hist_entry, rb_node);
 361                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 362                         found = n;
 363                         break;
 364                 }
 365                 next = rb_next(&n->rb_node);
 366         }
 367
 368         if (!found) {
 369                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 370                 sleep(1);
 371         } else
 372                 perf_top__parse_source(top, found);
 373
 374 out_free:
 375         free(buf);
 376 }
 377
 378 static void perf_top__print_mapped_keys(struct perf_top *top)
 379 {
 380         char *name = NULL;
 381
 382         if (top->sym_filter_entry) {
 383                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 384                 name = sym->name;
 385         }
 386
 387         fprintf(stdout, "\nMapped keys:\n");
 388         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 389         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 390
 391         if (top->evlist->nr_entries > 1)
 392                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 393
 394         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 395
 396         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 397         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 398         fprintf(stdout, "\t[S]     stop annotation.\n");
 399
 400         fprintf(stdout,
 401                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 402                 top->hide_kernel_symbols ? "yes" : "no");
 403         fprintf(stdout,
 404                 "\t[U]     hide user symbols.               \t(%s)\n",
 405                 top->hide_user_symbols ? "yes" : "no");
 406         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 407         fprintf(stdout, "\t[qQ]    quit.\n");
 408 }
 409
 410 static int perf_top__key_mapped(struct perf_top *top, int c)
 411 {
 412         switch (c) {
 413                 case 'd':
 414                 case 'e':
 415                 case 'f':
 416                 case 'z':
 417                 case 'q':
 418                 case 'Q':
 419                 case 'K':
 420                 case 'U':
 421                 case 'F':
 422                 case 's':
 423                 case 'S':
 424                         return 1;
 425                 case 'E':
 426                         return top->evlist->nr_entries > 1 ? 1 : 0;
 427                 default:
 428                         break;
 429         }
 430
 431         return 0;
 432 }
 433
 434 static void perf_top__handle_keypress(struct perf_top *top, int c)
 435 {
 436         if (!perf_top__key_mapped(top, c)) {
 437                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 438                 struct termios tc, save;
 439
 440                 perf_top__print_mapped_keys(top);
 441                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 442                 fflush(stdout);
 443
 444                 tcgetattr(0, &save);
 445                 tc = save;
 446                 tc.c_lflag &= ~(ICANON | ECHO);
 447                 tc.c_cc[VMIN] = 0;
 448                 tc.c_cc[VTIME] = 0;
 449                 tcsetattr(0, TCSANOW, &tc);
 450
 451                 poll(&stdin_poll, 1, -1);
 452                 c = getc(stdin);
 453
 454                 tcsetattr(0, TCSAFLUSH, &save);
 455                 if (!perf_top__key_mapped(top, c))
 456                         return;
 457         }
 458
 459         switch (c) {
 460                 case 'd':
 461                         prompt_integer(&top->delay_secs, "Enter display delay");
 462                         if (top->delay_secs < 1)
 463                                 top->delay_secs = 1;
 464                         break;
 465                 case 'e':
 466                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 467                         if (top->print_entries == 0) {
 468                                 struct sigaction act = {
 469                                         .sa_sigaction = perf_top__sig_winch,
 470                                         .sa_flags     = SA_SIGINFO,
 471                                 };
 472                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 473                                 sigaction(SIGWINCH, &act, NULL);
 474                         } else {
 475                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 476                                 signal(SIGWINCH, SIG_DFL);
 477                         }
 478                         break;
 479                 case 'E':
 480                         if (top->evlist->nr_entries > 1) {
 481                                 /* Select 0 as the default event: */
 482                                 int counter = 0;
 483
 484                                 fprintf(stderr, "\nAvailable events:");
 485
 486                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 487                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 488
 489                                 prompt_integer(&counter, "Enter details event counter");
 490
 491                                 if (counter >= top->evlist->nr_entries) {
 492                                         top->sym_evsel = perf_evlist__first(top->evlist);
 493                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 494                                         sleep(1);
 495                                         break;
 496                                 }
 497                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 498                                         if (top->sym_evsel->idx == counter)
 499                                                 break;
 500                         } else
 501                                 top->sym_evsel = perf_evlist__first(top->evlist);
 502                         break;
 503                 case 'f':
 504                         prompt_integer(&top->count_filter, "Enter display event count filter");
 505                         break;
 506                 case 'F':
 507                         prompt_percent(&top->sym_pcnt_filter,
 508                                        "Enter details display event filter (percent)");
 509                         break;
 510                 case 'K':
 511                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 512                         break;
 513                 case 'q':
 514                 case 'Q':
 515                         printf("exiting.\n");
 516                         if (top->dump_symtab)
 517                                 perf_session__fprintf_dsos(top->session, stderr);
 518                         exit(0);
 519                 case 's':
 520                         perf_top__prompt_symbol(top, "Enter details symbol");
 521                         break;
 522                 case 'S':
 523                         if (!top->sym_filter_entry)
 524                                 break;
 525                         else {
 526                                 struct hist_entry *syme = top->sym_filter_entry;
 527
 528                                 top->sym_filter_entry = NULL;
 529                                 __zero_source_counters(syme);
 530                         }
 531                         break;
 532                 case 'U':
 533                         top->hide_user_symbols = !top->hide_user_symbols;
 534                         break;
 535                 case 'z':
 536                         top->zero = !top->zero;
 537                         break;
 538                 default:
 539                         break;
 540         }
 541 }
 542
 543 static void perf_top__sort_new_samples(void *arg)
 544 {
 545         struct perf_top *t = arg;
 546         perf_top__reset_sample_counters(t);
 547
 548         if (t->evlist->selected != NULL)
 549                 t->sym_evsel = t->evlist->selected;
 550
 551         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 552         hists__output_resort_threaded(&t->sym_evsel->hists);
 553         hists__decay_entries_threaded(&t->sym_evsel->hists,
 554                                       t->hide_user_symbols,
 555                                       t->hide_kernel_symbols);
 556 }
 557
 558 static void *display_thread_tui(void *arg)
 559 {
 560         struct perf_evsel *pos;
 561         struct perf_top *top = arg;
 562         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 563         struct hist_browser_timer hbt = {
 564                 .timer          = perf_top__sort_new_samples,
 565                 .arg            = top,
 566                 .refresh        = top->delay_secs,
 567         };
 568
 569         perf_top__sort_new_samples(top);
 570
 571         /*
 572          * Initialize the uid_filter_str, in the future the TUI will allow
 573          * Zooming in/out UIDs. For now juse use whatever the user passed
 574          * via --uid.
 575          */
 576         list_for_each_entry(pos, &top->evlist->entries, node)
 577                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 578
 579         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 580                                       &top->session->header.env);
 581
 582         exit_browser(0);
 583         exit(0);
 584         return NULL;
 585 }
 586
 587 static void *display_thread(void *arg)
 588 {
 589         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 590         struct termios tc, save;
 591         struct perf_top *top = arg;
 592         int delay_msecs, c;
 593
 594         tcgetattr(0, &save);
 595         tc = save;
 596         tc.c_lflag &= ~(ICANON | ECHO);
 597         tc.c_cc[VMIN] = 0;
 598         tc.c_cc[VTIME] = 0;
 599
 600         pthread__unblock_sigwinch();
 601 repeat:
 602         delay_msecs = top->delay_secs * 1000;
 603         tcsetattr(0, TCSANOW, &tc);
 604         /* trash return*/
 605         getc(stdin);
 606
 607         while (1) {
 608                 perf_top__print_sym_table(top);
 609                 /*
 610                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 611                  * refresh screen in both cases.
 612                  */
 613                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 614                 case 0:
 615                         continue;
 616                 case -1:
 617                         if (errno == EINTR)
 618                                 continue;
 619                         /* Fall trhu */
 620                 default:
 621                         goto process_hotkey;
 622                 }
 623         }
 624 process_hotkey:
 625         c = getc(stdin);
 626         tcsetattr(0, TCSAFLUSH, &save);
 627
 628         perf_top__handle_keypress(top, c);
 629         goto repeat;
 630
 631         return NULL;
 632 }
 633
 634 /* Tag samples to be skipped. */
 635 static const char *skip_symbols[] = {
 636         "intel_idle",
 637         "default_idle",
 638         "native_safe_halt",
 639         "cpu_idle",
 640         "enter_idle",
 641         "exit_idle",
 642         "mwait_idle",
 643         "mwait_idle_with_hints",
 644         "poll_idle",
 645         "ppc64_runlatch_off",
 646         "pseries_dedicated_idle_sleep",
 647         NULL
 648 };
 649
 650 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 651 {
 652         const char *name = sym->name;
 653         int i;
 654
 655         /*
 656          * ppc64 uses function descriptors and appends a '.' to the
 657          * start of every instruction address. Remove it.
 658          */
 659         if (name[0] == '.')
 660                 name++;
 661
 662         if (!strcmp(name, "_text") ||
 663             !strcmp(name, "_etext") ||
 664             !strcmp(name, "_sinittext") ||
 665             !strncmp("init_module", name, 11) ||
 666             !strncmp("cleanup_module", name, 14) ||
 667             strstr(name, "_text_start") ||
 668             strstr(name, "_text_end"))
 669                 return 1;
 670
 671         for (i = 0; skip_symbols[i]; i++) {
 672                 if (!strcmp(skip_symbols[i], name)) {
 673                         sym->ignore = true;
 674                         break;
 675                 }
 676         }
 677
 678         return 0;
 679 }
 680
 681 static void perf_event__process_sample(struct perf_tool *tool,
 682                                        const union perf_event *event,
 683                                        struct perf_evsel *evsel,
 684                                        struct perf_sample *sample,
 685                                        struct machine *machine)
 686 {
 687         struct perf_top *top = container_of(tool, struct perf_top, tool);
 688         struct symbol *parent = NULL;
 689         u64 ip = event->ip.ip;
 690         struct addr_location al;
 691         int err;
 692
 693         if (!machine && perf_guest) {
 694                 static struct intlist *seen;
 695
 696                 if (!seen)
 697                         seen = intlist__new();
 698
 699                 if (!intlist__has_entry(seen, event->ip.pid)) {
 700                         pr_err("Can't find guest [%d]'s kernel information\n",
 701                                 event->ip.pid);
 702                         intlist__add(seen, event->ip.pid);
 703                 }
 704                 return;
 705         }
 706
 707         if (!machine) {
 708                 pr_err("%u unprocessable samples recorded.\n",
 709                        top->session->stats.nr_unprocessable_samples++);
 710                 return;
 711         }
 712
 713         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 714                 top->exact_samples++;
 715
 716         if (perf_event__preprocess_sample(event, machine, &al, sample,
 717                                           symbol_filter) < 0 ||
 718             al.filtered)
 719                 return;
 720
 721         if (!top->kptr_restrict_warned &&
 722             symbol_conf.kptr_restrict &&
 723             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 724                 ui__warning(
 725 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 726 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 727 "Kernel%s samples will not be resolved.\n",
 728                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 729                           " modules" : "");
 730                 if (use_browser <= 0)
 731                         sleep(5);
 732                 top->kptr_restrict_warned = true;
 733         }
 734
 735         if (al.sym == NULL) {
 736                 const char *msg = "Kernel samples will not be resolved.\n";
 737                 /*
 738                  * As we do lazy loading of symtabs we only will know if the
 739                  * specified vmlinux file is invalid when we actually have a
 740                  * hit in kernel space and then try to load it. So if we get
 741                  * here and there are _no_ symbols in the DSO backing the
 742                  * kernel map, bail out.
 743                  *
 744                  * We may never get here, for instance, if we use -K/
 745                  * --hide-kernel-symbols, even if the user specifies an
 746                  * invalid --vmlinux ;-)
 747                  */
 748                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 749                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 750                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 751                         if (symbol_conf.vmlinux_name) {
 752                                 ui__warning("The %s file can't be used.\n%s",
 753                                             symbol_conf.vmlinux_name, msg);
 754                         } else {
 755                                 ui__warning("A vmlinux file was not found.\n%s",
 756                                             msg);
 757                         }
 758
 759                         if (use_browser <= 0)
 760                                 sleep(5);
 761                         top->vmlinux_warned = true;
 762                 }
 763         }
 764
 765         if (al.sym == NULL || !al.sym->ignore) {
 766                 struct hist_entry *he;
 767
 768                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 769                     sample->callchain) {
 770                         err = machine__resolve_callchain(machine, evsel,
 771                                                          al.thread, sample,
 772                                                          &parent);
 773
 774                         if (err)
 775                                 return;
 776                 }
 777
 778                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 779                 if (he == NULL) {
 780                         pr_err("Problem incrementing symbol period, skipping event\n");
 781                         return;
 782                 }
 783
 784                 if (symbol_conf.use_callchain) {
 785                         err = callchain_append(he->callchain, &callchain_cursor,
 786                                                sample->period);
 787                         if (err)
 788                                 return;
 789                 }
 790
 791                 if (top->sort_has_symbols)
 792                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 793         }
 794
 795         return;
 796 }
 797
 798 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 799 {
 800         struct perf_sample sample;
 801         struct perf_evsel *evsel;
 802         struct perf_session *session = top->session;
 803         union perf_event *event;
 804         struct machine *machine;
 805         u8 origin;
 806         int ret;
 807
 808         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 809                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 810                 if (ret) {
 811                         pr_err("Can't parse sample, err = %d\n", ret);
 812                         continue;
 813                 }
 814
 815                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 816                 assert(evsel != NULL);
 817
 818                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 819
 820                 if (event->header.type == PERF_RECORD_SAMPLE)
 821                         ++top->samples;
 822
 823                 switch (origin) {
 824                 case PERF_RECORD_MISC_USER:
 825                         ++top->us_samples;
 826                         if (top->hide_user_symbols)
 827                                 continue;
 828                         machine = &session->machines.host;
 829                         break;
 830                 case PERF_RECORD_MISC_KERNEL:
 831                         ++top->kernel_samples;
 832                         if (top->hide_kernel_symbols)
 833                                 continue;
 834                         machine = &session->machines.host;
 835                         break;
 836                 case PERF_RECORD_MISC_GUEST_KERNEL:
 837                         ++top->guest_kernel_samples;
 838                         machine = perf_session__find_machine(session, event->ip.pid);
 839                         break;
 840                 case PERF_RECORD_MISC_GUEST_USER:
 841                         ++top->guest_us_samples;
 842                         /*
 843                          * TODO: we don't process guest user from host side
 844                          * except simple counting.
 845                          */
 846                         /* Fall thru */
 847                 default:
 848                         continue;
 849                 }
 850
 851
 852                 if (event->header.type == PERF_RECORD_SAMPLE) {
 853                         perf_event__process_sample(&top->tool, event, evsel,
 854                                                    &sample, machine);
 855                 } else if (event->header.type < PERF_RECORD_MAX) {
 856                         hists__inc_nr_events(&evsel->hists, event->header.type);
 857                         machine__process_event(machine, event);
 858                 } else
 859                         ++session->stats.nr_unknown_events;
 860         }
 861 }
 862
 863 static void perf_top__mmap_read(struct perf_top *top)
 864 {
 865         int i;
 866
 867         for (i = 0; i < top->evlist->nr_mmaps; i++)
 868                 perf_top__mmap_read_idx(top, i);
 869 }
 870
 871 static void perf_top__start_counters(struct perf_top *top)
 872 {
 873         char msg[512];
 874         struct perf_evsel *counter;
 875         struct perf_evlist *evlist = top->evlist;
 876         struct perf_record_opts *opts = &top->record_opts;
 877
 878         perf_evlist__config(evlist, opts);
 879
 880         list_for_each_entry(counter, &evlist->entries, node) {
 881 try_again:
 882                 if (perf_evsel__open(counter, top->evlist->cpus,
 883                                      top->evlist->threads) < 0) {
 884                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 885                                 if (verbose)
 886                                         ui__warning("%s\n", msg);
 887                                 goto try_again;
 888                         }
 889
 890                         perf_evsel__open_strerror(counter, &opts->target,
 891                                                   errno, msg, sizeof(msg));
 892                         ui__error("%s\n", msg);
 893                         goto out_err;
 894                 }
 895         }
 896
 897         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 898                 ui__error("Failed to mmap with %d (%s)\n",
 899                             errno, strerror(errno));
 900                 goto out_err;
 901         }
 902
 903         return;
 904
 905 out_err:
 906         exit_browser(0);
 907         exit(0);
 908 }
 909
 910 static int perf_top__setup_sample_type(struct perf_top *top)
 911 {
 912         if (!top->sort_has_symbols) {
 913                 if (symbol_conf.use_callchain) {
 914                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 915                         return -EINVAL;
 916                 }
 917         } else if (callchain_param.mode != CHAIN_NONE) {
 918                 if (callchain_register_param(&callchain_param) < 0) {
 919                         ui__error("Can't register callchain params.\n");
 920                         return -EINVAL;
 921                 }
 922         }
 923
 924         return 0;
 925 }
 926
 927 static int __cmd_top(struct perf_top *top)
 928 {
 929         struct perf_record_opts *opts = &top->record_opts;
 930         pthread_t thread;
 931         int ret;
 932         /*
 933          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 934          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 935          */
 936         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 937         if (top->session == NULL)
 938                 return -ENOMEM;
 939
 940         ret = perf_top__setup_sample_type(top);
 941         if (ret)
 942                 goto out_delete;
 943
 944         if (perf_target__has_task(&opts->target))
 945                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
 946                                                   perf_event__process,
 947                                                   &top->session->machines.host);
 948         else
 949                 perf_event__synthesize_threads(&top->tool, perf_event__process,
 950                                                &top->session->machines.host);
 951         perf_top__start_counters(top);
 952         top->session->evlist = top->evlist;
 953         perf_session__set_id_hdr_size(top->session);
 954
 955         /*
 956          * When perf is starting the traced process, all the events (apart from
 957          * group members) have enable_on_exec=1 set, so don't spoil it by
 958          * prematurely enabling them.
 959          *
 960          * XXX 'top' still doesn't start workloads like record, trace, but should,
 961          * so leave the check here.
 962          */
 963         if (!perf_target__none(&opts->target))
 964                 perf_evlist__enable(top->evlist);
 965
 966         /* Wait for a minimal set of events before starting the snapshot */
 967         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 968
 969         perf_top__mmap_read(top);
 970
 971         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 972                                                             display_thread), top)) {
 973                 ui__error("Could not create display thread.\n");
 974                 exit(-1);
 975         }
 976
 977         if (top->realtime_prio) {
 978                 struct sched_param param;
 979
 980                 param.sched_priority = top->realtime_prio;
 981                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 982                         ui__error("Could not set realtime priority.\n");
 983                         exit(-1);
 984                 }
 985         }
 986
 987         while (1) {
 988                 u64 hits = top->samples;
 989
 990                 perf_top__mmap_read(top);
 991
 992                 if (hits == top->samples)
 993                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 994         }
 995
 996 out_delete:
 997         perf_session__delete(top->session);
 998         top->session = NULL;
 999
1000         return 0;
1001 }
1002
1003 static int
1004 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1005 {
1006         /*
1007          * --no-call-graph
1008          */
1009         if (unset)
1010                 return 0;
1011
1012         symbol_conf.use_callchain = true;
1013
1014         return record_parse_callchain_opt(opt, arg, unset);
1015 }
1016
1017 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1018 {
1019         int status;
1020         char errbuf[BUFSIZ];
1021         struct perf_top top = {
1022                 .count_filter        = 5,
1023                 .delay_secs          = 2,
1024                 .record_opts = {
1025                         .mmap_pages     = UINT_MAX,
1026                         .user_freq      = UINT_MAX,
1027                         .user_interval  = ULLONG_MAX,
1028                         .freq           = 4000, /* 4 KHz */
1029                         .target              = {
1030                                 .uses_mmap   = true,
1031                         },
1032                 },
1033                 .sym_pcnt_filter     = 5,
1034         };
1035         struct perf_record_opts *opts = &top.record_opts;
1036         struct perf_target *target = &opts->target;
1037         const struct option options[] = {
1038         OPT_CALLBACK('e', "event", &top.evlist, "event",
1039                      "event selector. use 'perf list' to list available events",
1040                      parse_events_option),
1041         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1042         OPT_STRING('p', "pid", &target->pid, "pid",
1043                     "profile events on existing process id"),
1044         OPT_STRING('t', "tid", &target->tid, "tid",
1045                     "profile events on existing thread id"),
1046         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1047                             "system-wide collection from all CPUs"),
1048         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1049                     "list of cpus to monitor"),
1050         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1051                    "file", "vmlinux pathname"),
1052         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1053                     "hide kernel symbols"),
1054         OPT_UINTEGER('m', "mmap-pages", &opts->mmap_pages,
1055                      "number of mmap data pages"),
1056         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1057                     "collect data with this RT SCHED_FIFO priority"),
1058         OPT_INTEGER('d', "delay", &top.delay_secs,
1059                     "number of seconds to delay between refreshes"),
1060         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1061                             "dump the symbol table used for profiling"),
1062         OPT_INTEGER('f', "count-filter", &top.count_filter,
1063                     "only display functions with more events than this"),
1064         OPT_BOOLEAN('g', "group", &opts->group,
1065                             "put the counters into a counter group"),
1066         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1067                     "child tasks do not inherit counters"),
1068         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1069                     "symbol to annotate"),
1070         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1071         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1072         OPT_INTEGER('E', "entries", &top.print_entries,
1073                     "display this many functions"),
1074         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1075                     "hide user symbols"),
1076         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1077         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1078         OPT_INCR('v', "verbose", &verbose,
1079                     "be more verbose (show counter open errors, etc)"),
1080         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1081                    "sort by key(s): pid, comm, dso, symbol, parent"),
1082         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1083                     "Show a column with the number of samples"),
1084         OPT_CALLBACK_DEFAULT('G', "call-graph", &top.record_opts,
1085                              "mode[,dump_size]", record_callchain_help,
1086                              &parse_callchain_opt, "fp"),
1087         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1088                     "Show a column with the sum of periods"),
1089         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1090                    "only consider symbols in these dsos"),
1091         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1092                    "only consider symbols in these comms"),
1093         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1094                    "only consider these symbols"),
1095         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1096                     "Interleave source code with assembly code (default)"),
1097         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1098                     "Display raw encoding of assembly instructions (default)"),
1099         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1100                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1101         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1102         OPT_END()
1103         };
1104         const char * const top_usage[] = {
1105                 "perf top [<options>]",
1106                 NULL
1107         };
1108
1109         top.evlist = perf_evlist__new(NULL, NULL);
1110         if (top.evlist == NULL)
1111                 return -ENOMEM;
1112
1113         symbol_conf.exclude_other = false;
1114
1115         argc = parse_options(argc, argv, options, top_usage, 0);
1116         if (argc)
1117                 usage_with_options(top_usage, options);
1118
1119         if (sort_order == default_sort_order)
1120                 sort_order = "dso,symbol";
1121
1122         setup_sorting(top_usage, options);
1123
1124         if (top.use_stdio)
1125                 use_browser = 0;
1126         else if (top.use_tui)
1127                 use_browser = 1;
1128
1129         setup_browser(false);
1130
1131         status = perf_target__validate(target);
1132         if (status) {
1133                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1134                 ui__warning("%s", errbuf);
1135         }
1136
1137         status = perf_target__parse_uid(target);
1138         if (status) {
1139                 int saved_errno = errno;
1140
1141                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1142                 ui__error("%s", errbuf);
1143
1144                 status = -saved_errno;
1145                 goto out_delete_evlist;
1146         }
1147
1148         if (perf_target__none(target))
1149                 target->system_wide = true;
1150
1151         if (perf_evlist__create_maps(top.evlist, target) < 0)
1152                 usage_with_options(top_usage, options);
1153
1154         if (!top.evlist->nr_entries &&
1155             perf_evlist__add_default(top.evlist) < 0) {
1156                 ui__error("Not enough memory for event selector list\n");
1157                 return -ENOMEM;
1158         }
1159
1160         symbol_conf.nr_events = top.evlist->nr_entries;
1161
1162         if (top.delay_secs < 1)
1163                 top.delay_secs = 1;
1164
1165         if (opts->user_interval != ULLONG_MAX)
1166                 opts->default_interval = opts->user_interval;
1167         if (opts->user_freq != UINT_MAX)
1168                 opts->freq = opts->user_freq;
1169
1170         /*
1171          * User specified count overrides default frequency.
1172          */
1173         if (opts->default_interval)
1174                 opts->freq = 0;
1175         else if (opts->freq) {
1176                 opts->default_interval = opts->freq;
1177         } else {
1178                 ui__error("frequency and count are zero, aborting\n");
1179                 status = -EINVAL;
1180                 goto out_delete_evlist;
1181         }
1182
1183         top.sym_evsel = perf_evlist__first(top.evlist);
1184
1185         symbol_conf.priv_size = sizeof(struct annotation);
1186
1187         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1188         if (symbol__init() < 0)
1189                 return -1;
1190
1191         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1192         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1193         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1194
1195         /*
1196          * Avoid annotation data structures overhead when symbols aren't on the
1197          * sort list.
1198          */
1199         top.sort_has_symbols = sort_sym.list.next != NULL;
1200
1201         get_term_dimensions(&top.winsize);
1202         if (top.print_entries == 0) {
1203                 struct sigaction act = {
1204                         .sa_sigaction = perf_top__sig_winch,
1205                         .sa_flags     = SA_SIGINFO,
1206                 };
1207                 perf_top__update_print_entries(&top);
1208                 sigaction(SIGWINCH, &act, NULL);
1209         }
1210
1211         status = __cmd_top(&top);
1212
1213 out_delete_evlist:
1214         perf_evlist__delete(top.evlist);
1215
1216         return status;
1217 }