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