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/session.h"
  30 #include "util/symbol.h"
  31 #include "util/thread.h"
  32 #include "util/thread_map.h"
  33 #include "util/top.h"
  34 #include "util/util.h"
  35 #include <linux/rbtree.h>
  36 #include "util/parse-options.h"
  37 #include "util/parse-events.h"
  38 #include "util/cpumap.h"
  39 #include "util/xyarray.h"
  40 #include "util/sort.h"
  41
  42 #include "util/debug.h"
  43
  44 #include <assert.h>
  45 #include <fcntl.h>
  46
  47 #include <stdio.h>
  48 #include <termios.h>
  49 #include <unistd.h>
  50 #include <inttypes.h>
  51
  52 #include <errno.h>
  53 #include <time.h>
  54 #include <sched.h>
  55
  56 #include <sys/syscall.h>
  57 #include <sys/ioctl.h>
  58 #include <sys/poll.h>
  59 #include <sys/prctl.h>
  60 #include <sys/wait.h>
  61 #include <sys/uio.h>
  62 #include <sys/mman.h>
  63
  64 #include <linux/unistd.h>
  65 #include <linux/types.h>
  66
  67
  68 void get_term_dimensions(struct winsize *ws)
  69 {
  70         char *s = getenv("LINES");
  71
  72         if (s != NULL) {
  73                 ws->ws_row = atoi(s);
  74                 s = getenv("COLUMNS");
  75                 if (s != NULL) {
  76                         ws->ws_col = atoi(s);
  77                         if (ws->ws_row && ws->ws_col)
  78                                 return;
  79                 }
  80         }
  81 #ifdef TIOCGWINSZ
  82         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
  83             ws->ws_row && ws->ws_col)
  84                 return;
  85 #endif
  86         ws->ws_row = 25;
  87         ws->ws_col = 80;
  88 }
  89
  90 static void perf_top__update_print_entries(struct perf_top *top)
  91 {
  92         if (top->print_entries > 9)
  93                 top->print_entries -= 9;
  94 }
  95
  96 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
  97 {
  98         struct perf_top *top = arg;
  99
 100         get_term_dimensions(&top->winsize);
 101         if (!top->print_entries
 102             || (top->print_entries+4) > top->winsize.ws_row) {
 103                 top->print_entries = top->winsize.ws_row;
 104         } else {
 105                 top->print_entries += 4;
 106                 top->winsize.ws_row = top->print_entries;
 107         }
 108         perf_top__update_print_entries(top);
 109 }
 110
 111 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 112 {
 113         struct symbol *sym;
 114         struct annotation *notes;
 115         struct map *map;
 116         int err = -1;
 117
 118         if (!he || !he->ms.sym)
 119                 return -1;
 120
 121         sym = he->ms.sym;
 122         map = he->ms.map;
 123
 124         /*
 125          * We can't annotate with just /proc/kallsyms
 126          */
 127         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
 128                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 129                        "path\n", sym->name);
 130                 sleep(1);
 131                 return -1;
 132         }
 133
 134         notes = symbol__annotation(sym);
 135         if (notes->src != NULL) {
 136                 pthread_mutex_lock(&notes->lock);
 137                 goto out_assign;
 138         }
 139
 140         pthread_mutex_lock(&notes->lock);
 141
 142         if (symbol__alloc_hist(sym) < 0) {
 143                 pthread_mutex_unlock(&notes->lock);
 144                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 145                        sym->name);
 146                 sleep(1);
 147                 return err;
 148         }
 149
 150         err = symbol__annotate(sym, map, 0);
 151         if (err == 0) {
 152 out_assign:
 153                 top->sym_filter_entry = he;
 154         }
 155
 156         pthread_mutex_unlock(&notes->lock);
 157         return err;
 158 }
 159
 160 static void __zero_source_counters(struct hist_entry *he)
 161 {
 162         struct symbol *sym = he->ms.sym;
 163         symbol__annotate_zero_histograms(sym);
 164 }
 165
 166 static void perf_top__record_precise_ip(struct perf_top *top,
 167                                         struct hist_entry *he,
 168                                         int counter, u64 ip)
 169 {
 170         struct annotation *notes;
 171         struct symbol *sym;
 172
 173         if (he == NULL || he->ms.sym == NULL ||
 174             ((top->sym_filter_entry == NULL ||
 175               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 176                 return;
 177
 178         sym = he->ms.sym;
 179         notes = symbol__annotation(sym);
 180
 181         if (pthread_mutex_trylock(&notes->lock))
 182                 return;
 183
 184         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 185                 pthread_mutex_unlock(&notes->lock);
 186                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 187                        sym->name);
 188                 sleep(1);
 189                 return;
 190         }
 191
 192         ip = he->ms.map->map_ip(he->ms.map, ip);
 193         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 194
 195         pthread_mutex_unlock(&notes->lock);
 196 }
 197
 198 static void perf_top__show_details(struct perf_top *top)
 199 {
 200         struct hist_entry *he = top->sym_filter_entry;
 201         struct annotation *notes;
 202         struct symbol *symbol;
 203         int more;
 204
 205         if (!he)
 206                 return;
 207
 208         symbol = he->ms.sym;
 209         notes = symbol__annotation(symbol);
 210
 211         pthread_mutex_lock(&notes->lock);
 212
 213         if (notes->src == NULL)
 214                 goto out_unlock;
 215
 216         printf("Showing %s for %s\n", event_name(top->sym_evsel), symbol->name);
 217         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 218
 219         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
 220                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 221         if (top->zero)
 222                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 223         else
 224                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 225         if (more != 0)
 226                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 227 out_unlock:
 228         pthread_mutex_unlock(&notes->lock);
 229 }
 230
 231 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 232
 233 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 234                                                      struct addr_location *al,
 235                                                      struct perf_sample *sample)
 236 {
 237         struct hist_entry *he;
 238
 239         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 240         if (he == NULL)
 241                 return NULL;
 242
 243         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 244         return he;
 245 }
 246
 247 static void perf_top__print_sym_table(struct perf_top *top)
 248 {
 249         char bf[160];
 250         int printed = 0;
 251         const int win_width = top->winsize.ws_col - 1;
 252
 253         puts(CONSOLE_CLEAR);
 254
 255         perf_top__header_snprintf(top, bf, sizeof(bf));
 256         printf("%s\n", bf);
 257
 258         perf_top__reset_sample_counters(top);
 259
 260         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 261
 262         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 263             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 264                 top->sym_evsel->hists.stats.nr_lost_warned =
 265                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 266                 color_fprintf(stdout, PERF_COLOR_RED,
 267                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 268                               top->sym_evsel->hists.stats.nr_lost_warned);
 269                 ++printed;
 270         }
 271
 272         if (top->sym_filter_entry) {
 273                 perf_top__show_details(top);
 274                 return;
 275         }
 276
 277         hists__collapse_resort_threaded(&top->sym_evsel->hists);
 278         hists__output_resort_threaded(&top->sym_evsel->hists);
 279         hists__decay_entries_threaded(&top->sym_evsel->hists,
 280                                       top->hide_user_symbols,
 281                                       top->hide_kernel_symbols);
 282         hists__output_recalc_col_len(&top->sym_evsel->hists,
 283                                      top->winsize.ws_row - 3);
 284         putchar('\n');
 285         hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
 286                        top->winsize.ws_row - 4 - printed, win_width, stdout);
 287 }
 288
 289 static void prompt_integer(int *target, const char *msg)
 290 {
 291         char *buf = malloc(0), *p;
 292         size_t dummy = 0;
 293         int tmp;
 294
 295         fprintf(stdout, "\n%s: ", msg);
 296         if (getline(&buf, &dummy, stdin) < 0)
 297                 return;
 298
 299         p = strchr(buf, '\n');
 300         if (p)
 301                 *p = 0;
 302
 303         p = buf;
 304         while(*p) {
 305                 if (!isdigit(*p))
 306                         goto out_free;
 307                 p++;
 308         }
 309         tmp = strtoul(buf, NULL, 10);
 310         *target = tmp;
 311 out_free:
 312         free(buf);
 313 }
 314
 315 static void prompt_percent(int *target, const char *msg)
 316 {
 317         int tmp = 0;
 318
 319         prompt_integer(&tmp, msg);
 320         if (tmp >= 0 && tmp <= 100)
 321                 *target = tmp;
 322 }
 323
 324 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 325 {
 326         char *buf = malloc(0), *p;
 327         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 328         struct rb_node *next;
 329         size_t dummy = 0;
 330
 331         /* zero counters of active symbol */
 332         if (syme) {
 333                 __zero_source_counters(syme);
 334                 top->sym_filter_entry = NULL;
 335         }
 336
 337         fprintf(stdout, "\n%s: ", msg);
 338         if (getline(&buf, &dummy, stdin) < 0)
 339                 goto out_free;
 340
 341         p = strchr(buf, '\n');
 342         if (p)
 343                 *p = 0;
 344
 345         next = rb_first(&top->sym_evsel->hists.entries);
 346         while (next) {
 347                 n = rb_entry(next, struct hist_entry, rb_node);
 348                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 349                         found = n;
 350                         break;
 351                 }
 352                 next = rb_next(&n->rb_node);
 353         }
 354
 355         if (!found) {
 356                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 357                 sleep(1);
 358         } else
 359                 perf_top__parse_source(top, found);
 360
 361 out_free:
 362         free(buf);
 363 }
 364
 365 static void perf_top__print_mapped_keys(struct perf_top *top)
 366 {
 367         char *name = NULL;
 368
 369         if (top->sym_filter_entry) {
 370                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 371                 name = sym->name;
 372         }
 373
 374         fprintf(stdout, "\nMapped keys:\n");
 375         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 376         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 377
 378         if (top->evlist->nr_entries > 1)
 379                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top->sym_evsel));
 380
 381         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 382
 383         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 384         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 385         fprintf(stdout, "\t[S]     stop annotation.\n");
 386
 387         fprintf(stdout,
 388                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 389                 top->hide_kernel_symbols ? "yes" : "no");
 390         fprintf(stdout,
 391                 "\t[U]     hide user symbols.               \t(%s)\n",
 392                 top->hide_user_symbols ? "yes" : "no");
 393         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 394         fprintf(stdout, "\t[qQ]    quit.\n");
 395 }
 396
 397 static int perf_top__key_mapped(struct perf_top *top, int c)
 398 {
 399         switch (c) {
 400                 case 'd':
 401                 case 'e':
 402                 case 'f':
 403                 case 'z':
 404                 case 'q':
 405                 case 'Q':
 406                 case 'K':
 407                 case 'U':
 408                 case 'F':
 409                 case 's':
 410                 case 'S':
 411                         return 1;
 412                 case 'E':
 413                         return top->evlist->nr_entries > 1 ? 1 : 0;
 414                 default:
 415                         break;
 416         }
 417
 418         return 0;
 419 }
 420
 421 static void perf_top__handle_keypress(struct perf_top *top, int c)
 422 {
 423         if (!perf_top__key_mapped(top, c)) {
 424                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 425                 struct termios tc, save;
 426
 427                 perf_top__print_mapped_keys(top);
 428                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 429                 fflush(stdout);
 430
 431                 tcgetattr(0, &save);
 432                 tc = save;
 433                 tc.c_lflag &= ~(ICANON | ECHO);
 434                 tc.c_cc[VMIN] = 0;
 435                 tc.c_cc[VTIME] = 0;
 436                 tcsetattr(0, TCSANOW, &tc);
 437
 438                 poll(&stdin_poll, 1, -1);
 439                 c = getc(stdin);
 440
 441                 tcsetattr(0, TCSAFLUSH, &save);
 442                 if (!perf_top__key_mapped(top, c))
 443                         return;
 444         }
 445
 446         switch (c) {
 447                 case 'd':
 448                         prompt_integer(&top->delay_secs, "Enter display delay");
 449                         if (top->delay_secs < 1)
 450                                 top->delay_secs = 1;
 451                         break;
 452                 case 'e':
 453                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 454                         if (top->print_entries == 0) {
 455                                 struct sigaction act = {
 456                                         .sa_sigaction = perf_top__sig_winch,
 457                                         .sa_flags     = SA_SIGINFO,
 458                                 };
 459                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 460                                 sigaction(SIGWINCH, &act, NULL);
 461                         } else {
 462                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 463                                 signal(SIGWINCH, SIG_DFL);
 464                         }
 465                         break;
 466                 case 'E':
 467                         if (top->evlist->nr_entries > 1) {
 468                                 /* Select 0 as the default event: */
 469                                 int counter = 0;
 470
 471                                 fprintf(stderr, "\nAvailable events:");
 472
 473                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 474                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
 475
 476                                 prompt_integer(&counter, "Enter details event counter");
 477
 478                                 if (counter >= top->evlist->nr_entries) {
 479                                         top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
 480                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
 481                                         sleep(1);
 482                                         break;
 483                                 }
 484                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 485                                         if (top->sym_evsel->idx == counter)
 486                                                 break;
 487                         } else
 488                                 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
 489                         break;
 490                 case 'f':
 491                         prompt_integer(&top->count_filter, "Enter display event count filter");
 492                         break;
 493                 case 'F':
 494                         prompt_percent(&top->sym_pcnt_filter,
 495                                        "Enter details display event filter (percent)");
 496                         break;
 497                 case 'K':
 498                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 499                         break;
 500                 case 'q':
 501                 case 'Q':
 502                         printf("exiting.\n");
 503                         if (top->dump_symtab)
 504                                 perf_session__fprintf_dsos(top->session, stderr);
 505                         exit(0);
 506                 case 's':
 507                         perf_top__prompt_symbol(top, "Enter details symbol");
 508                         break;
 509                 case 'S':
 510                         if (!top->sym_filter_entry)
 511                                 break;
 512                         else {
 513                                 struct hist_entry *syme = top->sym_filter_entry;
 514
 515                                 top->sym_filter_entry = NULL;
 516                                 __zero_source_counters(syme);
 517                         }
 518                         break;
 519                 case 'U':
 520                         top->hide_user_symbols = !top->hide_user_symbols;
 521                         break;
 522                 case 'z':
 523                         top->zero = !top->zero;
 524                         break;
 525                 default:
 526                         break;
 527         }
 528 }
 529
 530 static void perf_top__sort_new_samples(void *arg)
 531 {
 532         struct perf_top *t = arg;
 533         perf_top__reset_sample_counters(t);
 534
 535         if (t->evlist->selected != NULL)
 536                 t->sym_evsel = t->evlist->selected;
 537
 538         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 539         hists__output_resort_threaded(&t->sym_evsel->hists);
 540         hists__decay_entries_threaded(&t->sym_evsel->hists,
 541                                       t->hide_user_symbols,
 542                                       t->hide_kernel_symbols);
 543 }
 544
 545 static void *display_thread_tui(void *arg)
 546 {
 547         struct perf_top *top = arg;
 548         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 549
 550         perf_top__sort_new_samples(top);
 551         perf_evlist__tui_browse_hists(top->evlist, help,
 552                                       perf_top__sort_new_samples,
 553                                       top, top->delay_secs);
 554
 555         exit_browser(0);
 556         exit(0);
 557         return NULL;
 558 }
 559
 560 static void *display_thread(void *arg)
 561 {
 562         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 563         struct termios tc, save;
 564         struct perf_top *top = arg;
 565         int delay_msecs, c;
 566
 567         tcgetattr(0, &save);
 568         tc = save;
 569         tc.c_lflag &= ~(ICANON | ECHO);
 570         tc.c_cc[VMIN] = 0;
 571         tc.c_cc[VTIME] = 0;
 572
 573         pthread__unblock_sigwinch();
 574 repeat:
 575         delay_msecs = top->delay_secs * 1000;
 576         tcsetattr(0, TCSANOW, &tc);
 577         /* trash return*/
 578         getc(stdin);
 579
 580         while (1) {
 581                 perf_top__print_sym_table(top);
 582                 /*
 583                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 584                  * refresh screen in both cases.
 585                  */
 586                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 587                 case 0:
 588                         continue;
 589                 case -1:
 590                         if (errno == EINTR)
 591                                 continue;
 592                         /* Fall trhu */
 593                 default:
 594                         goto process_hotkey;
 595                 }
 596         }
 597 process_hotkey:
 598         c = getc(stdin);
 599         tcsetattr(0, TCSAFLUSH, &save);
 600
 601         perf_top__handle_keypress(top, c);
 602         goto repeat;
 603
 604         return NULL;
 605 }
 606
 607 /* Tag samples to be skipped. */
 608 static const char *skip_symbols[] = {
 609         "default_idle",
 610         "native_safe_halt",
 611         "cpu_idle",
 612         "enter_idle",
 613         "exit_idle",
 614         "mwait_idle",
 615         "mwait_idle_with_hints",
 616         "poll_idle",
 617         "ppc64_runlatch_off",
 618         "pseries_dedicated_idle_sleep",
 619         NULL
 620 };
 621
 622 static int symbol_filter(struct map *map __used, struct symbol *sym)
 623 {
 624         const char *name = sym->name;
 625         int i;
 626
 627         /*
 628          * ppc64 uses function descriptors and appends a '.' to the
 629          * start of every instruction address. Remove it.
 630          */
 631         if (name[0] == '.')
 632                 name++;
 633
 634         if (!strcmp(name, "_text") ||
 635             !strcmp(name, "_etext") ||
 636             !strcmp(name, "_sinittext") ||
 637             !strncmp("init_module", name, 11) ||
 638             !strncmp("cleanup_module", name, 14) ||
 639             strstr(name, "_text_start") ||
 640             strstr(name, "_text_end"))
 641                 return 1;
 642
 643         for (i = 0; skip_symbols[i]; i++) {
 644                 if (!strcmp(skip_symbols[i], name)) {
 645                         sym->ignore = true;
 646                         break;
 647                 }
 648         }
 649
 650         return 0;
 651 }
 652
 653 static void perf_event__process_sample(struct perf_tool *tool,
 654                                        const union perf_event *event,
 655                                        struct perf_evsel *evsel,
 656                                        struct perf_sample *sample,
 657                                        struct machine *machine)
 658 {
 659         struct perf_top *top = container_of(tool, struct perf_top, tool);
 660         struct symbol *parent = NULL;
 661         u64 ip = event->ip.ip;
 662         struct addr_location al;
 663         int err;
 664
 665         if (!machine && perf_guest) {
 666                 pr_err("Can't find guest [%d]'s kernel information\n",
 667                         event->ip.pid);
 668                 return;
 669         }
 670
 671         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 672                 top->exact_samples++;
 673
 674         if (perf_event__preprocess_sample(event, machine, &al, sample,
 675                                           symbol_filter) < 0 ||
 676             al.filtered)
 677                 return;
 678
 679         if (!top->kptr_restrict_warned &&
 680             symbol_conf.kptr_restrict &&
 681             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 682                 ui__warning(
 683 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 684 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 685 "Kernel%s samples will not be resolved.\n",
 686                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 687                           " modules" : "");
 688                 if (use_browser <= 0)
 689                         sleep(5);
 690                 top->kptr_restrict_warned = true;
 691         }
 692
 693         if (al.sym == NULL) {
 694                 const char *msg = "Kernel samples will not be resolved.\n";
 695                 /*
 696                  * As we do lazy loading of symtabs we only will know if the
 697                  * specified vmlinux file is invalid when we actually have a
 698                  * hit in kernel space and then try to load it. So if we get
 699                  * here and there are _no_ symbols in the DSO backing the
 700                  * kernel map, bail out.
 701                  *
 702                  * We may never get here, for instance, if we use -K/
 703                  * --hide-kernel-symbols, even if the user specifies an
 704                  * invalid --vmlinux ;-)
 705                  */
 706                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 707                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 708                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 709                         if (symbol_conf.vmlinux_name) {
 710                                 ui__warning("The %s file can't be used.\n%s",
 711                                             symbol_conf.vmlinux_name, msg);
 712                         } else {
 713                                 ui__warning("A vmlinux file was not found.\n%s",
 714                                             msg);
 715                         }
 716
 717                         if (use_browser <= 0)
 718                                 sleep(5);
 719                         top->vmlinux_warned = true;
 720                 }
 721         }
 722
 723         if (al.sym == NULL || !al.sym->ignore) {
 724                 struct hist_entry *he;
 725
 726                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 727                     sample->callchain) {
 728                         err = machine__resolve_callchain(machine, evsel, al.thread,
 729                                                          sample->callchain, &parent);
 730                         if (err)
 731                                 return;
 732                 }
 733
 734                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 735                 if (he == NULL) {
 736                         pr_err("Problem incrementing symbol period, skipping event\n");
 737                         return;
 738                 }
 739
 740                 if (symbol_conf.use_callchain) {
 741                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
 742                                                sample->period);
 743                         if (err)
 744                                 return;
 745                 }
 746
 747                 if (top->sort_has_symbols)
 748                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 749         }
 750
 751         return;
 752 }
 753
 754 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 755 {
 756         struct perf_sample sample;
 757         struct perf_evsel *evsel;
 758         struct perf_session *session = top->session;
 759         union perf_event *event;
 760         struct machine *machine;
 761         u8 origin;
 762         int ret;
 763
 764         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 765                 ret = perf_session__parse_sample(session, event, &sample);
 766                 if (ret) {
 767                         pr_err("Can't parse sample, err = %d\n", ret);
 768                         continue;
 769                 }
 770
 771                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 772                 assert(evsel != NULL);
 773
 774                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 775
 776                 if (event->header.type == PERF_RECORD_SAMPLE)
 777                         ++top->samples;
 778
 779                 switch (origin) {
 780                 case PERF_RECORD_MISC_USER:
 781                         ++top->us_samples;
 782                         if (top->hide_user_symbols)
 783                                 continue;
 784                         machine = perf_session__find_host_machine(session);
 785                         break;
 786                 case PERF_RECORD_MISC_KERNEL:
 787                         ++top->kernel_samples;
 788                         if (top->hide_kernel_symbols)
 789                                 continue;
 790                         machine = perf_session__find_host_machine(session);
 791                         break;
 792                 case PERF_RECORD_MISC_GUEST_KERNEL:
 793                         ++top->guest_kernel_samples;
 794                         machine = perf_session__find_machine(session, event->ip.pid);
 795                         break;
 796                 case PERF_RECORD_MISC_GUEST_USER:
 797                         ++top->guest_us_samples;
 798                         /*
 799                          * TODO: we don't process guest user from host side
 800                          * except simple counting.
 801                          */
 802                         /* Fall thru */
 803                 default:
 804                         continue;
 805                 }
 806
 807
 808                 if (event->header.type == PERF_RECORD_SAMPLE) {
 809                         perf_event__process_sample(&top->tool, event, evsel,
 810                                                    &sample, machine);
 811                 } else if (event->header.type < PERF_RECORD_MAX) {
 812                         hists__inc_nr_events(&evsel->hists, event->header.type);
 813                         perf_event__process(&top->tool, event, &sample, machine);
 814                 } else
 815                         ++session->hists.stats.nr_unknown_events;
 816         }
 817 }
 818
 819 static void perf_top__mmap_read(struct perf_top *top)
 820 {
 821         int i;
 822
 823         for (i = 0; i < top->evlist->nr_mmaps; i++)
 824                 perf_top__mmap_read_idx(top, i);
 825 }
 826
 827 static void perf_top__start_counters(struct perf_top *top)
 828 {
 829         struct perf_evsel *counter, *first;
 830         struct perf_evlist *evlist = top->evlist;
 831
 832         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 833
 834         list_for_each_entry(counter, &evlist->entries, node) {
 835                 struct perf_event_attr *attr = &counter->attr;
 836                 struct xyarray *group_fd = NULL;
 837
 838                 if (top->group && counter != first)
 839                         group_fd = first->fd;
 840
 841                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 842
 843                 if (top->freq) {
 844                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 845                         attr->freq        = 1;
 846                         attr->sample_freq = top->freq;
 847                 }
 848
 849                 if (evlist->nr_entries > 1) {
 850                         attr->sample_type |= PERF_SAMPLE_ID;
 851                         attr->read_format |= PERF_FORMAT_ID;
 852                 }
 853
 854                 if (symbol_conf.use_callchain)
 855                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
 856
 857                 attr->mmap = 1;
 858                 attr->comm = 1;
 859                 attr->inherit = top->inherit;
 860 fallback_missing_features:
 861                 if (top->exclude_guest_missing)
 862                         attr->exclude_guest = attr->exclude_host = 0;
 863 retry_sample_id:
 864                 attr->sample_id_all = top->sample_id_all_avail ? 1 : 0;
 865 try_again:
 866                 if (perf_evsel__open(counter, top->evlist->cpus,
 867                                      top->evlist->threads, top->group,
 868                                      group_fd) < 0) {
 869                         int err = errno;
 870
 871                         if (err == EPERM || err == EACCES) {
 872                                 ui__error_paranoid();
 873                                 goto out_err;
 874                         } else if (err == EINVAL) {
 875                                 if (!top->exclude_guest_missing &&
 876                                     (attr->exclude_guest || attr->exclude_host)) {
 877                                         pr_debug("Old kernel, cannot exclude "
 878                                                  "guest or host samples.\n");
 879                                         top->exclude_guest_missing = true;
 880                                         goto fallback_missing_features;
 881                                 } else if (top->sample_id_all_avail) {
 882                                         /*
 883                                          * Old kernel, no attr->sample_id_type_all field
 884                                          */
 885                                         top->sample_id_all_avail = false;
 886                                         goto retry_sample_id;
 887                                 }
 888                         }
 889                         /*
 890                          * If it's cycles then fall back to hrtimer
 891                          * based cpu-clock-tick sw counter, which
 892                          * is always available even if no PMU support:
 893                          */
 894                         if (attr->type == PERF_TYPE_HARDWARE &&
 895                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 896                                 if (verbose)
 897                                         ui__warning("Cycles event not supported,\n"
 898                                                     "trying to fall back to cpu-clock-ticks\n");
 899
 900                                 attr->type = PERF_TYPE_SOFTWARE;
 901                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 902                                 goto try_again;
 903                         }
 904
 905                         if (err == ENOENT) {
 906                                 ui__warning("The %s event is not supported.\n",
 907                                             event_name(counter));
 908                                 goto out_err;
 909                         } else if (err == EMFILE) {
 910                                 ui__warning("Too many events are opened.\n"
 911                                             "Try again after reducing the number of events\n");
 912                                 goto out_err;
 913                         }
 914
 915                         ui__warning("The sys_perf_event_open() syscall "
 916                                     "returned with %d (%s).  /bin/dmesg "
 917                                     "may provide additional information.\n"
 918                                     "No CONFIG_PERF_EVENTS=y kernel support "
 919                                     "configured?\n", err, strerror(err));
 920                         goto out_err;
 921                 }
 922         }
 923
 924         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
 925                 ui__warning("Failed to mmap with %d (%s)\n",
 926                             errno, strerror(errno));
 927                 goto out_err;
 928         }
 929
 930         return;
 931
 932 out_err:
 933         exit_browser(0);
 934         exit(0);
 935 }
 936
 937 static int perf_top__setup_sample_type(struct perf_top *top)
 938 {
 939         if (!top->sort_has_symbols) {
 940                 if (symbol_conf.use_callchain) {
 941                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
 942                         return -EINVAL;
 943                 }
 944         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
 945                 if (callchain_register_param(&callchain_param) < 0) {
 946                         ui__warning("Can't register callchain params.\n");
 947                         return -EINVAL;
 948                 }
 949         }
 950
 951         return 0;
 952 }
 953
 954 static int __cmd_top(struct perf_top *top)
 955 {
 956         pthread_t thread;
 957         int ret;
 958         /*
 959          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 960          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 961          */
 962         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 963         if (top->session == NULL)
 964                 return -ENOMEM;
 965
 966         ret = perf_top__setup_sample_type(top);
 967         if (ret)
 968                 goto out_delete;
 969
 970         if (top->target_tid != -1)
 971                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
 972                                                   perf_event__process,
 973                                                   &top->session->host_machine);
 974         else
 975                 perf_event__synthesize_threads(&top->tool, perf_event__process,
 976                                                &top->session->host_machine);
 977         perf_top__start_counters(top);
 978         top->session->evlist = top->evlist;
 979         perf_session__update_sample_type(top->session);
 980
 981         /* Wait for a minimal set of events before starting the snapshot */
 982         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 983
 984         perf_top__mmap_read(top);
 985
 986         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 987                                                             display_thread), top)) {
 988                 printf("Could not create display thread.\n");
 989                 exit(-1);
 990         }
 991
 992         if (top->realtime_prio) {
 993                 struct sched_param param;
 994
 995                 param.sched_priority = top->realtime_prio;
 996                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 997                         printf("Could not set realtime priority.\n");
 998                         exit(-1);
 999                 }
1000         }
1001
1002         while (1) {
1003                 u64 hits = top->samples;
1004
1005                 perf_top__mmap_read(top);
1006
1007                 if (hits == top->samples)
1008                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1009         }
1010
1011 out_delete:
1012         perf_session__delete(top->session);
1013         top->session = NULL;
1014
1015         return 0;
1016 }
1017
1018 static int
1019 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1020 {
1021         struct perf_top *top = (struct perf_top *)opt->value;
1022         char *tok, *tok2;
1023         char *endptr;
1024
1025         /*
1026          * --no-call-graph
1027          */
1028         if (unset) {
1029                 top->dont_use_callchains = true;
1030                 return 0;
1031         }
1032
1033         symbol_conf.use_callchain = true;
1034
1035         if (!arg)
1036                 return 0;
1037
1038         tok = strtok((char *)arg, ",");
1039         if (!tok)
1040                 return -1;
1041
1042         /* get the output mode */
1043         if (!strncmp(tok, "graph", strlen(arg)))
1044                 callchain_param.mode = CHAIN_GRAPH_ABS;
1045
1046         else if (!strncmp(tok, "flat", strlen(arg)))
1047                 callchain_param.mode = CHAIN_FLAT;
1048
1049         else if (!strncmp(tok, "fractal", strlen(arg)))
1050                 callchain_param.mode = CHAIN_GRAPH_REL;
1051
1052         else if (!strncmp(tok, "none", strlen(arg))) {
1053                 callchain_param.mode = CHAIN_NONE;
1054                 symbol_conf.use_callchain = false;
1055
1056                 return 0;
1057         } else
1058                 return -1;
1059
1060         /* get the min percentage */
1061         tok = strtok(NULL, ",");
1062         if (!tok)
1063                 goto setup;
1064
1065         callchain_param.min_percent = strtod(tok, &endptr);
1066         if (tok == endptr)
1067                 return -1;
1068
1069         /* get the print limit */
1070         tok2 = strtok(NULL, ",");
1071         if (!tok2)
1072                 goto setup;
1073
1074         if (tok2[0] != 'c') {
1075                 callchain_param.print_limit = strtod(tok2, &endptr);
1076                 tok2 = strtok(NULL, ",");
1077                 if (!tok2)
1078                         goto setup;
1079         }
1080
1081         /* get the call chain order */
1082         if (!strcmp(tok2, "caller"))
1083                 callchain_param.order = ORDER_CALLER;
1084         else if (!strcmp(tok2, "callee"))
1085                 callchain_param.order = ORDER_CALLEE;
1086         else
1087                 return -1;
1088 setup:
1089         if (callchain_register_param(&callchain_param) < 0) {
1090                 fprintf(stderr, "Can't register callchain params\n");
1091                 return -1;
1092         }
1093         return 0;
1094 }
1095
1096 static const char * const top_usage[] = {
1097         "perf top [<options>]",
1098         NULL
1099 };
1100
1101 int cmd_top(int argc, const char **argv, const char *prefix __used)
1102 {
1103         struct perf_evsel *pos;
1104         int status = -ENOMEM;
1105         struct perf_top top = {
1106                 .count_filter        = 5,
1107                 .delay_secs          = 2,
1108                 .target_pid          = -1,
1109                 .target_tid          = -1,
1110                 .freq                = 1000, /* 1 KHz */
1111                 .sample_id_all_avail = true,
1112                 .mmap_pages          = 128,
1113                 .sym_pcnt_filter     = 5,
1114         };
1115         char callchain_default_opt[] = "fractal,0.5,callee";
1116         const struct option options[] = {
1117         OPT_CALLBACK('e', "event", &top.evlist, "event",
1118                      "event selector. use 'perf list' to list available events",
1119                      parse_events_option),
1120         OPT_INTEGER('c', "count", &top.default_interval,
1121                     "event period to sample"),
1122         OPT_INTEGER('p', "pid", &top.target_pid,
1123                     "profile events on existing process id"),
1124         OPT_INTEGER('t', "tid", &top.target_tid,
1125                     "profile events on existing thread id"),
1126         OPT_BOOLEAN('a', "all-cpus", &top.system_wide,
1127                             "system-wide collection from all CPUs"),
1128         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1129                     "list of cpus to monitor"),
1130         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1131                    "file", "vmlinux pathname"),
1132         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1133                     "hide kernel symbols"),
1134         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1135         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1136                     "collect data with this RT SCHED_FIFO priority"),
1137         OPT_INTEGER('d', "delay", &top.delay_secs,
1138                     "number of seconds to delay between refreshes"),
1139         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1140                             "dump the symbol table used for profiling"),
1141         OPT_INTEGER('f', "count-filter", &top.count_filter,
1142                     "only display functions with more events than this"),
1143         OPT_BOOLEAN('g', "group", &top.group,
1144                             "put the counters into a counter group"),
1145         OPT_BOOLEAN('i', "inherit", &top.inherit,
1146                     "child tasks inherit counters"),
1147         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1148                     "symbol to annotate"),
1149         OPT_BOOLEAN('z', "zero", &top.zero,
1150                     "zero history across updates"),
1151         OPT_INTEGER('F', "freq", &top.freq,
1152                     "profile at this frequency"),
1153         OPT_INTEGER('E', "entries", &top.print_entries,
1154                     "display this many functions"),
1155         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1156                     "hide user symbols"),
1157         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1158         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1159         OPT_INCR('v', "verbose", &verbose,
1160                     "be more verbose (show counter open errors, etc)"),
1161         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1162                    "sort by key(s): pid, comm, dso, symbol, parent"),
1163         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1164                     "Show a column with the number of samples"),
1165         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1166                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1167                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1168                      callchain_default_opt),
1169         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1170                     "Show a column with the sum of periods"),
1171         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1172                    "only consider symbols in these dsos"),
1173         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1174                    "only consider symbols in these comms"),
1175         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1176                    "only consider these symbols"),
1177         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1178                     "Interleave source code with assembly code (default)"),
1179         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1180                     "Display raw encoding of assembly instructions (default)"),
1181         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1182                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1183         OPT_END()
1184         };
1185
1186         top.evlist = perf_evlist__new(NULL, NULL);
1187         if (top.evlist == NULL)
1188                 return -ENOMEM;
1189
1190         symbol_conf.exclude_other = false;
1191
1192         argc = parse_options(argc, argv, options, top_usage, 0);
1193         if (argc)
1194                 usage_with_options(top_usage, options);
1195
1196         if (sort_order == default_sort_order)
1197                 sort_order = "dso,symbol";
1198
1199         setup_sorting(top_usage, options);
1200
1201         if (top.use_stdio)
1202                 use_browser = 0;
1203         else if (top.use_tui)
1204                 use_browser = 1;
1205
1206         setup_browser(false);
1207
1208         /* CPU and PID are mutually exclusive */
1209         if (top.target_tid > 0 && top.cpu_list) {
1210                 printf("WARNING: PID switch overriding CPU\n");
1211                 sleep(1);
1212                 top.cpu_list = NULL;
1213         }
1214
1215         if (top.target_pid != -1)
1216                 top.target_tid = top.target_pid;
1217
1218         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1219                                      top.target_tid, top.cpu_list) < 0)
1220                 usage_with_options(top_usage, options);
1221
1222         if (!top.evlist->nr_entries &&
1223             perf_evlist__add_default(top.evlist) < 0) {
1224                 pr_err("Not enough memory for event selector list\n");
1225                 return -ENOMEM;
1226         }
1227
1228         symbol_conf.nr_events = top.evlist->nr_entries;
1229
1230         if (top.delay_secs < 1)
1231                 top.delay_secs = 1;
1232
1233         /*
1234          * User specified count overrides default frequency.
1235          */
1236         if (top.default_interval)
1237                 top.freq = 0;
1238         else if (top.freq) {
1239                 top.default_interval = top.freq;
1240         } else {
1241                 fprintf(stderr, "frequency and count are zero, aborting\n");
1242                 exit(EXIT_FAILURE);
1243         }
1244
1245         list_for_each_entry(pos, &top.evlist->entries, node) {
1246                 /*
1247                  * Fill in the ones not specifically initialized via -c:
1248                  */
1249                 if (!pos->attr.sample_period)
1250                         pos->attr.sample_period = top.default_interval;
1251         }
1252
1253         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1254
1255         symbol_conf.priv_size = sizeof(struct annotation);
1256
1257         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1258         if (symbol__init() < 0)
1259                 return -1;
1260
1261         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1262         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1263         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1264
1265         /*
1266          * Avoid annotation data structures overhead when symbols aren't on the
1267          * sort list.
1268          */
1269         top.sort_has_symbols = sort_sym.list.next != NULL;
1270
1271         get_term_dimensions(&top.winsize);
1272         if (top.print_entries == 0) {
1273                 struct sigaction act = {
1274                         .sa_sigaction = perf_top__sig_winch,
1275                         .sa_flags     = SA_SIGINFO,
1276                 };
1277                 perf_top__update_print_entries(&top);
1278                 sigaction(SIGWINCH, &act, NULL);
1279         }
1280
1281         status = __cmd_top(&top);
1282
1283         perf_evlist__delete(top.evlist);
1284
1285         return status;
1286 }