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 static struct perf_top top = {
  68         .count_filter           = 5,
  69         .delay_secs             = 2,
  70         .target_pid             = -1,
  71         .target_tid             = -1,
  72         .freq                   = 1000, /* 1 KHz */
  73 };
  74
  75 static bool                     system_wide                     =  false;
  76
  77 static bool                     use_tui, use_stdio;
  78
  79 static bool                     sort_has_symbols;
  80
  81 static bool                     dont_use_callchains;
  82 static char                     callchain_default_opt[]         = "fractal,0.5,callee";
  83
  84
  85 static int                      default_interval                =      0;
  86
  87 static bool                     kptr_restrict_warned;
  88 static bool                     vmlinux_warned;
  89 static bool                     inherit                         =  false;
  90 static int                      realtime_prio                   =      0;
  91 static bool                     group                           =  false;
  92 static unsigned int             mmap_pages                      =    128;
  93
  94 static bool                     dump_symtab                     =  false;
  95
  96 static struct winsize           winsize;
  97
  98 static const char               *sym_filter                     =   NULL;
  99 static int                      sym_pcnt_filter                 =      5;
 100
 101 /*
 102  * Source functions
 103  */
 104
 105 void get_term_dimensions(struct winsize *ws)
 106 {
 107         char *s = getenv("LINES");
 108
 109         if (s != NULL) {
 110                 ws->ws_row = atoi(s);
 111                 s = getenv("COLUMNS");
 112                 if (s != NULL) {
 113                         ws->ws_col = atoi(s);
 114                         if (ws->ws_row && ws->ws_col)
 115                                 return;
 116                 }
 117         }
 118 #ifdef TIOCGWINSZ
 119         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 120             ws->ws_row && ws->ws_col)
 121                 return;
 122 #endif
 123         ws->ws_row = 25;
 124         ws->ws_col = 80;
 125 }
 126
 127 static void update_print_entries(struct winsize *ws)
 128 {
 129         top.print_entries = ws->ws_row;
 130
 131         if (top.print_entries > 9)
 132                 top.print_entries -= 9;
 133 }
 134
 135 static void sig_winch_handler(int sig __used)
 136 {
 137         get_term_dimensions(&winsize);
 138         update_print_entries(&winsize);
 139 }
 140
 141 static int parse_source(struct hist_entry *he)
 142 {
 143         struct symbol *sym;
 144         struct annotation *notes;
 145         struct map *map;
 146         int err = -1;
 147
 148         if (!he || !he->ms.sym)
 149                 return -1;
 150
 151         sym = he->ms.sym;
 152         map = he->ms.map;
 153
 154         /*
 155          * We can't annotate with just /proc/kallsyms
 156          */
 157         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
 158                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 159                        "path\n", sym->name);
 160                 sleep(1);
 161                 return -1;
 162         }
 163
 164         notes = symbol__annotation(sym);
 165         if (notes->src != NULL) {
 166                 pthread_mutex_lock(&notes->lock);
 167                 goto out_assign;
 168         }
 169
 170         pthread_mutex_lock(&notes->lock);
 171
 172         if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
 173                 pthread_mutex_unlock(&notes->lock);
 174                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 175                        sym->name);
 176                 sleep(1);
 177                 return err;
 178         }
 179
 180         err = symbol__annotate(sym, map, 0);
 181         if (err == 0) {
 182 out_assign:
 183                 top.sym_filter_entry = he;
 184         }
 185
 186         pthread_mutex_unlock(&notes->lock);
 187         return err;
 188 }
 189
 190 static void __zero_source_counters(struct hist_entry *he)
 191 {
 192         struct symbol *sym = he->ms.sym;
 193         symbol__annotate_zero_histograms(sym);
 194 }
 195
 196 static void record_precise_ip(struct hist_entry *he, int counter, u64 ip)
 197 {
 198         struct annotation *notes;
 199         struct symbol *sym;
 200
 201         if (he == NULL || he->ms.sym == NULL ||
 202             (he != top.sym_filter_entry && use_browser != 1))
 203                 return;
 204
 205         sym = he->ms.sym;
 206         notes = symbol__annotation(sym);
 207
 208         if (pthread_mutex_trylock(&notes->lock))
 209                 return;
 210
 211         if (notes->src == NULL &&
 212             symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
 213                 pthread_mutex_unlock(&notes->lock);
 214                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 215                        sym->name);
 216                 sleep(1);
 217                 return;
 218         }
 219
 220         ip = he->ms.map->map_ip(he->ms.map, ip);
 221         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 222
 223         pthread_mutex_unlock(&notes->lock);
 224 }
 225
 226 static void show_details(struct hist_entry *he)
 227 {
 228         struct annotation *notes;
 229         struct symbol *symbol;
 230         int more;
 231
 232         if (!he)
 233                 return;
 234
 235         symbol = he->ms.sym;
 236         notes = symbol__annotation(symbol);
 237
 238         pthread_mutex_lock(&notes->lock);
 239
 240         if (notes->src == NULL)
 241                 goto out_unlock;
 242
 243         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
 244         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 245
 246         more = symbol__annotate_printf(symbol, he->ms.map, top.sym_evsel->idx,
 247                                        0, sym_pcnt_filter, top.print_entries, 4);
 248         if (top.zero)
 249                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
 250         else
 251                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
 252         if (more != 0)
 253                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 254 out_unlock:
 255         pthread_mutex_unlock(&notes->lock);
 256 }
 257
 258 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 259
 260 static struct hist_entry *
 261         perf_session__add_hist_entry(struct perf_session *session,
 262                                      struct addr_location *al,
 263                                      struct perf_sample *sample,
 264                                      struct perf_evsel *evsel)
 265 {
 266         struct hist_entry *he;
 267
 268         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 269         if (he == NULL)
 270                 return NULL;
 271
 272         session->hists.stats.total_period += sample->period;
 273         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 274         return he;
 275 }
 276
 277 static void print_sym_table(void)
 278 {
 279         char bf[160];
 280         int printed = 0;
 281         const int win_width = winsize.ws_col - 1;
 282
 283         puts(CONSOLE_CLEAR);
 284
 285         perf_top__header_snprintf(&top, bf, sizeof(bf));
 286         printf("%s\n", bf);
 287
 288         perf_top__reset_sample_counters(&top);
 289
 290         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 291
 292         if (top.total_lost_warned != top.session->hists.stats.total_lost) {
 293                 top.total_lost_warned = top.session->hists.stats.total_lost;
 294                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
 295                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
 296                        top.total_lost_warned);
 297                 ++printed;
 298         }
 299
 300         if (top.sym_filter_entry) {
 301                 show_details(top.sym_filter_entry);
 302                 return;
 303         }
 304
 305         hists__collapse_resort_threaded(&top.sym_evsel->hists);
 306         hists__output_resort_threaded(&top.sym_evsel->hists);
 307         hists__decay_entries_threaded(&top.sym_evsel->hists,
 308                                       top.hide_user_symbols,
 309                                       top.hide_kernel_symbols);
 310         hists__output_recalc_col_len(&top.sym_evsel->hists, winsize.ws_row - 3);
 311         putchar('\n');
 312         hists__fprintf(&top.sym_evsel->hists, NULL, false, false,
 313                        winsize.ws_row - 4 - printed, win_width, stdout);
 314 }
 315
 316 static void prompt_integer(int *target, const char *msg)
 317 {
 318         char *buf = malloc(0), *p;
 319         size_t dummy = 0;
 320         int tmp;
 321
 322         fprintf(stdout, "\n%s: ", msg);
 323         if (getline(&buf, &dummy, stdin) < 0)
 324                 return;
 325
 326         p = strchr(buf, '\n');
 327         if (p)
 328                 *p = 0;
 329
 330         p = buf;
 331         while(*p) {
 332                 if (!isdigit(*p))
 333                         goto out_free;
 334                 p++;
 335         }
 336         tmp = strtoul(buf, NULL, 10);
 337         *target = tmp;
 338 out_free:
 339         free(buf);
 340 }
 341
 342 static void prompt_percent(int *target, const char *msg)
 343 {
 344         int tmp = 0;
 345
 346         prompt_integer(&tmp, msg);
 347         if (tmp >= 0 && tmp <= 100)
 348                 *target = tmp;
 349 }
 350
 351 static void prompt_symbol(struct hist_entry **target, const char *msg)
 352 {
 353         char *buf = malloc(0), *p;
 354         struct hist_entry *syme = *target, *n, *found = NULL;
 355         struct rb_node *next;
 356         size_t dummy = 0;
 357
 358         /* zero counters of active symbol */
 359         if (syme) {
 360                 __zero_source_counters(syme);
 361                 *target = NULL;
 362         }
 363
 364         fprintf(stdout, "\n%s: ", msg);
 365         if (getline(&buf, &dummy, stdin) < 0)
 366                 goto out_free;
 367
 368         p = strchr(buf, '\n');
 369         if (p)
 370                 *p = 0;
 371
 372         next = rb_first(&top.sym_evsel->hists.entries);
 373         while (next) {
 374                 n = rb_entry(next, struct hist_entry, rb_node);
 375                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 376                         found = n;
 377                         break;
 378                 }
 379                 next = rb_next(&n->rb_node);
 380         }
 381
 382         if (!found) {
 383                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 384                 sleep(1);
 385                 return;
 386         } else
 387                 parse_source(found);
 388
 389 out_free:
 390         free(buf);
 391 }
 392
 393 static void print_mapped_keys(void)
 394 {
 395         char *name = NULL;
 396
 397         if (top.sym_filter_entry) {
 398                 struct symbol *sym = top.sym_filter_entry->ms.sym;
 399                 name = sym->name;
 400         }
 401
 402         fprintf(stdout, "\nMapped keys:\n");
 403         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
 404         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
 405
 406         if (top.evlist->nr_entries > 1)
 407                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
 408
 409         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
 410
 411         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 412         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 413         fprintf(stdout, "\t[S]     stop annotation.\n");
 414
 415         fprintf(stdout,
 416                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 417                 top.hide_kernel_symbols ? "yes" : "no");
 418         fprintf(stdout,
 419                 "\t[U]     hide user symbols.               \t(%s)\n",
 420                 top.hide_user_symbols ? "yes" : "no");
 421         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
 422         fprintf(stdout, "\t[qQ]    quit.\n");
 423 }
 424
 425 static int key_mapped(int c)
 426 {
 427         switch (c) {
 428                 case 'd':
 429                 case 'e':
 430                 case 'f':
 431                 case 'z':
 432                 case 'q':
 433                 case 'Q':
 434                 case 'K':
 435                 case 'U':
 436                 case 'F':
 437                 case 's':
 438                 case 'S':
 439                         return 1;
 440                 case 'E':
 441                         return top.evlist->nr_entries > 1 ? 1 : 0;
 442                 default:
 443                         break;
 444         }
 445
 446         return 0;
 447 }
 448
 449 static void handle_keypress(int c)
 450 {
 451         if (!key_mapped(c)) {
 452                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 453                 struct termios tc, save;
 454
 455                 print_mapped_keys();
 456                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 457                 fflush(stdout);
 458
 459                 tcgetattr(0, &save);
 460                 tc = save;
 461                 tc.c_lflag &= ~(ICANON | ECHO);
 462                 tc.c_cc[VMIN] = 0;
 463                 tc.c_cc[VTIME] = 0;
 464                 tcsetattr(0, TCSANOW, &tc);
 465
 466                 poll(&stdin_poll, 1, -1);
 467                 c = getc(stdin);
 468
 469                 tcsetattr(0, TCSAFLUSH, &save);
 470                 if (!key_mapped(c))
 471                         return;
 472         }
 473
 474         switch (c) {
 475                 case 'd':
 476                         prompt_integer(&top.delay_secs, "Enter display delay");
 477                         if (top.delay_secs < 1)
 478                                 top.delay_secs = 1;
 479                         break;
 480                 case 'e':
 481                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
 482                         if (top.print_entries == 0) {
 483                                 sig_winch_handler(SIGWINCH);
 484                                 signal(SIGWINCH, sig_winch_handler);
 485                         } else
 486                                 signal(SIGWINCH, SIG_DFL);
 487                         break;
 488                 case 'E':
 489                         if (top.evlist->nr_entries > 1) {
 490                                 /* Select 0 as the default event: */
 491                                 int counter = 0;
 492
 493                                 fprintf(stderr, "\nAvailable events:");
 494
 495                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 496                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 497
 498                                 prompt_integer(&counter, "Enter details event counter");
 499
 500                                 if (counter >= top.evlist->nr_entries) {
 501                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 502                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 503                                         sleep(1);
 504                                         break;
 505                                 }
 506                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 507                                         if (top.sym_evsel->idx == counter)
 508                                                 break;
 509                         } else
 510                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 511                         break;
 512                 case 'f':
 513                         prompt_integer(&top.count_filter, "Enter display event count filter");
 514                         break;
 515                 case 'F':
 516                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 517                         break;
 518                 case 'K':
 519                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 520                         break;
 521                 case 'q':
 522                 case 'Q':
 523                         printf("exiting.\n");
 524                         if (dump_symtab)
 525                                 perf_session__fprintf_dsos(top.session, stderr);
 526                         exit(0);
 527                 case 's':
 528                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 529                         break;
 530                 case 'S':
 531                         if (!top.sym_filter_entry)
 532                                 break;
 533                         else {
 534                                 struct hist_entry *syme = top.sym_filter_entry;
 535
 536                                 top.sym_filter_entry = NULL;
 537                                 __zero_source_counters(syme);
 538                         }
 539                         break;
 540                 case 'U':
 541                         top.hide_user_symbols = !top.hide_user_symbols;
 542                         break;
 543                 case 'z':
 544                         top.zero = !top.zero;
 545                         break;
 546                 default:
 547                         break;
 548         }
 549 }
 550
 551 static void perf_top__sort_new_samples(void *arg)
 552 {
 553         struct perf_top *t = arg;
 554         perf_top__reset_sample_counters(t);
 555
 556         if (t->evlist->selected != NULL)
 557                 t->sym_evsel = t->evlist->selected;
 558
 559         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 560         hists__output_resort_threaded(&t->sym_evsel->hists);
 561         hists__decay_entries_threaded(&t->sym_evsel->hists,
 562                                       top.hide_user_symbols,
 563                                       top.hide_kernel_symbols);
 564         hists__output_recalc_col_len(&t->sym_evsel->hists, winsize.ws_row - 3);
 565 }
 566
 567 static void *display_thread_tui(void *arg __used)
 568 {
 569         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 570
 571         perf_top__sort_new_samples(&top);
 572         perf_evlist__tui_browse_hists(top.evlist, help,
 573                                       perf_top__sort_new_samples,
 574                                       &top, top.delay_secs);
 575
 576         exit_browser(0);
 577         exit(0);
 578         return NULL;
 579 }
 580
 581 static void *display_thread(void *arg __used)
 582 {
 583         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 584         struct termios tc, save;
 585         int delay_msecs, c;
 586
 587         tcgetattr(0, &save);
 588         tc = save;
 589         tc.c_lflag &= ~(ICANON | ECHO);
 590         tc.c_cc[VMIN] = 0;
 591         tc.c_cc[VTIME] = 0;
 592
 593         pthread__unblock_sigwinch();
 594 repeat:
 595         delay_msecs = top.delay_secs * 1000;
 596         tcsetattr(0, TCSANOW, &tc);
 597         /* trash return*/
 598         getc(stdin);
 599
 600         while (1) {
 601                 print_sym_table();
 602                 /*
 603                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 604                  * refresh screen in both cases.
 605                  */
 606                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 607                 case 0:
 608                         continue;
 609                 case -1:
 610                         if (errno == EINTR)
 611                                 continue;
 612                         /* Fall trhu */
 613                 default:
 614                         goto process_hotkey;
 615                 }
 616         }
 617 process_hotkey:
 618         c = getc(stdin);
 619         tcsetattr(0, TCSAFLUSH, &save);
 620
 621         handle_keypress(c);
 622         goto repeat;
 623
 624         return NULL;
 625 }
 626
 627 /* Tag samples to be skipped. */
 628 static const char *skip_symbols[] = {
 629         "default_idle",
 630         "native_safe_halt",
 631         "cpu_idle",
 632         "enter_idle",
 633         "exit_idle",
 634         "mwait_idle",
 635         "mwait_idle_with_hints",
 636         "poll_idle",
 637         "ppc64_runlatch_off",
 638         "pseries_dedicated_idle_sleep",
 639         NULL
 640 };
 641
 642 static int symbol_filter(struct map *map __used, struct symbol *sym)
 643 {
 644         const char *name = sym->name;
 645         int i;
 646
 647         /*
 648          * ppc64 uses function descriptors and appends a '.' to the
 649          * start of every instruction address. Remove it.
 650          */
 651         if (name[0] == '.')
 652                 name++;
 653
 654         if (!strcmp(name, "_text") ||
 655             !strcmp(name, "_etext") ||
 656             !strcmp(name, "_sinittext") ||
 657             !strncmp("init_module", name, 11) ||
 658             !strncmp("cleanup_module", name, 14) ||
 659             strstr(name, "_text_start") ||
 660             strstr(name, "_text_end"))
 661                 return 1;
 662
 663         for (i = 0; skip_symbols[i]; i++) {
 664                 if (!strcmp(skip_symbols[i], name)) {
 665                         sym->ignore = true;
 666                         break;
 667                 }
 668         }
 669
 670         return 0;
 671 }
 672
 673 static void perf_event__process_sample(const union perf_event *event,
 674                                        struct perf_sample *sample,
 675                                        struct perf_session *session)
 676 {
 677         struct symbol *parent = NULL;
 678         u64 ip = event->ip.ip;
 679         struct addr_location al;
 680         struct machine *machine;
 681         int err;
 682         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 683
 684         ++top.samples;
 685
 686         switch (origin) {
 687         case PERF_RECORD_MISC_USER:
 688                 ++top.us_samples;
 689                 if (top.hide_user_symbols)
 690                         return;
 691                 machine = perf_session__find_host_machine(session);
 692                 break;
 693         case PERF_RECORD_MISC_KERNEL:
 694                 ++top.kernel_samples;
 695                 if (top.hide_kernel_symbols)
 696                         return;
 697                 machine = perf_session__find_host_machine(session);
 698                 break;
 699         case PERF_RECORD_MISC_GUEST_KERNEL:
 700                 ++top.guest_kernel_samples;
 701                 machine = perf_session__find_machine(session, event->ip.pid);
 702                 break;
 703         case PERF_RECORD_MISC_GUEST_USER:
 704                 ++top.guest_us_samples;
 705                 /*
 706                  * TODO: we don't process guest user from host side
 707                  * except simple counting.
 708                  */
 709                 return;
 710         default:
 711                 return;
 712         }
 713
 714         if (!machine && perf_guest) {
 715                 pr_err("Can't find guest [%d]'s kernel information\n",
 716                         event->ip.pid);
 717                 return;
 718         }
 719
 720         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 721                 top.exact_samples++;
 722
 723         if (perf_event__preprocess_sample(event, session, &al, sample,
 724                                           symbol_filter) < 0 ||
 725             al.filtered)
 726                 return;
 727
 728         if (!kptr_restrict_warned &&
 729             symbol_conf.kptr_restrict &&
 730             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 731                 ui__warning(
 732 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 733 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 734 "Kernel%s samples will not be resolved.\n",
 735                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 736                           " modules" : "");
 737                 if (use_browser <= 0)
 738                         sleep(5);
 739                 kptr_restrict_warned = true;
 740         }
 741
 742         if (al.sym == NULL) {
 743                 const char *msg = "Kernel samples will not be resolved.\n";
 744                 /*
 745                  * As we do lazy loading of symtabs we only will know if the
 746                  * specified vmlinux file is invalid when we actually have a
 747                  * hit in kernel space and then try to load it. So if we get
 748                  * here and there are _no_ symbols in the DSO backing the
 749                  * kernel map, bail out.
 750                  *
 751                  * We may never get here, for instance, if we use -K/
 752                  * --hide-kernel-symbols, even if the user specifies an
 753                  * invalid --vmlinux ;-)
 754                  */
 755                 if (!kptr_restrict_warned && !vmlinux_warned &&
 756                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 757                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 758                         if (symbol_conf.vmlinux_name) {
 759                                 ui__warning("The %s file can't be used.\n%s",
 760                                             symbol_conf.vmlinux_name, msg);
 761                         } else {
 762                                 ui__warning("A vmlinux file was not found.\n%s",
 763                                             msg);
 764                         }
 765
 766                         if (use_browser <= 0)
 767                                 sleep(5);
 768                         vmlinux_warned = true;
 769                 }
 770         }
 771
 772         if (al.sym == NULL || !al.sym->ignore) {
 773                 struct perf_evsel *evsel;
 774                 struct hist_entry *he;
 775
 776                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
 777                 assert(evsel != NULL);
 778
 779                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 780                     sample->callchain) {
 781                         err = perf_session__resolve_callchain(session, al.thread,
 782                                                               sample->callchain, &parent);
 783                         if (err)
 784                                 return;
 785                 }
 786
 787                 he = perf_session__add_hist_entry(session, &al, sample, evsel);
 788                 if (he == NULL) {
 789                         pr_err("Problem incrementing symbol period, skipping event\n");
 790                         return;
 791                 }
 792
 793                 if (symbol_conf.use_callchain) {
 794                         err = callchain_append(he->callchain, &session->callchain_cursor,
 795                                                sample->period);
 796                         if (err)
 797                                 return;
 798                 }
 799
 800                 if (sort_has_symbols)
 801                         record_precise_ip(he, evsel->idx, ip);
 802         }
 803
 804         return;
 805 }
 806
 807 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
 808 {
 809         struct perf_sample sample;
 810         union perf_event *event;
 811         int ret;
 812
 813         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
 814                 ret = perf_session__parse_sample(self, event, &sample);
 815                 if (ret) {
 816                         pr_err("Can't parse sample, err = %d\n", ret);
 817                         continue;
 818                 }
 819
 820                 if (event->header.type == PERF_RECORD_SAMPLE)
 821                         perf_event__process_sample(event, &sample, self);
 822                 else
 823                         perf_event__process(event, &sample, self);
 824         }
 825 }
 826
 827 static void perf_session__mmap_read(struct perf_session *self)
 828 {
 829         int i;
 830
 831         for (i = 0; i < top.evlist->nr_mmaps; i++)
 832                 perf_session__mmap_read_idx(self, i);
 833 }
 834
 835 static void start_counters(struct perf_evlist *evlist)
 836 {
 837         struct perf_evsel *counter;
 838
 839         list_for_each_entry(counter, &evlist->entries, node) {
 840                 struct perf_event_attr *attr = &counter->attr;
 841
 842                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 843
 844                 if (top.freq) {
 845                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 846                         attr->freq        = 1;
 847                         attr->sample_freq = top.freq;
 848                 }
 849
 850                 if (evlist->nr_entries > 1) {
 851                         attr->sample_type |= PERF_SAMPLE_ID;
 852                         attr->read_format |= PERF_FORMAT_ID;
 853                 }
 854
 855                 if (symbol_conf.use_callchain)
 856                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
 857
 858                 attr->mmap = 1;
 859                 attr->comm = 1;
 860                 attr->inherit = inherit;
 861 try_again:
 862                 if (perf_evsel__open(counter, top.evlist->cpus,
 863                                      top.evlist->threads, group) < 0) {
 864                         int err = errno;
 865
 866                         if (err == EPERM || err == EACCES) {
 867                                 ui__warning_paranoid();
 868                                 goto out_err;
 869                         }
 870                         /*
 871                          * If it's cycles then fall back to hrtimer
 872                          * based cpu-clock-tick sw counter, which
 873                          * is always available even if no PMU support:
 874                          */
 875                         if (attr->type == PERF_TYPE_HARDWARE &&
 876                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 877                                 if (verbose)
 878                                         ui__warning("Cycles event not supported,\n"
 879                                                     "trying to fall back to cpu-clock-ticks\n");
 880
 881                                 attr->type = PERF_TYPE_SOFTWARE;
 882                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 883                                 goto try_again;
 884                         }
 885
 886                         if (err == ENOENT) {
 887                                 ui__warning("The %s event is not supported.\n",
 888                                             event_name(counter));
 889                                 goto out_err;
 890                         }
 891
 892                         ui__warning("The sys_perf_event_open() syscall "
 893                                     "returned with %d (%s).  /bin/dmesg "
 894                                     "may provide additional information.\n"
 895                                     "No CONFIG_PERF_EVENTS=y kernel support "
 896                                     "configured?\n", err, strerror(err));
 897                         goto out_err;
 898                 }
 899         }
 900
 901         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
 902                 ui__warning("Failed to mmap with %d (%s)\n",
 903                             errno, strerror(errno));
 904                 goto out_err;
 905         }
 906
 907         return;
 908
 909 out_err:
 910         exit_browser(0);
 911         exit(0);
 912 }
 913
 914 static int setup_sample_type(void)
 915 {
 916         if (!sort_has_symbols) {
 917                 if (symbol_conf.use_callchain) {
 918                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
 919                         return -EINVAL;
 920                 }
 921         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
 922                 if (callchain_register_param(&callchain_param) < 0) {
 923                         ui__warning("Can't register callchain params.\n");
 924                         return -EINVAL;
 925                 }
 926         }
 927
 928         return 0;
 929 }
 930
 931 static int __cmd_top(void)
 932 {
 933         pthread_t thread;
 934         int ret;
 935         /*
 936          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 937          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 938          */
 939         top.session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 940         if (top.session == NULL)
 941                 return -ENOMEM;
 942
 943         ret = setup_sample_type();
 944         if (ret)
 945                 goto out_delete;
 946
 947         if (top.target_tid != -1)
 948                 perf_event__synthesize_thread_map(top.evlist->threads,
 949                                                   perf_event__process, top.session);
 950         else
 951                 perf_event__synthesize_threads(perf_event__process, top.session);
 952
 953         start_counters(top.evlist);
 954         top.session->evlist = top.evlist;
 955         perf_session__update_sample_type(top.session);
 956
 957         /* Wait for a minimal set of events before starting the snapshot */
 958         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 959
 960         perf_session__mmap_read(top.session);
 961
 962         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 963                                                              display_thread), NULL)) {
 964                 printf("Could not create display thread.\n");
 965                 exit(-1);
 966         }
 967
 968         if (realtime_prio) {
 969                 struct sched_param param;
 970
 971                 param.sched_priority = realtime_prio;
 972                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 973                         printf("Could not set realtime priority.\n");
 974                         exit(-1);
 975                 }
 976         }
 977
 978         while (1) {
 979                 u64 hits = top.samples;
 980
 981                 perf_session__mmap_read(top.session);
 982
 983                 if (hits == top.samples)
 984                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 985         }
 986
 987 out_delete:
 988         perf_session__delete(top.session);
 989         top.session = NULL;
 990
 991         return 0;
 992 }
 993
 994 static int
 995 parse_callchain_opt(const struct option *opt __used, const char *arg,
 996                     int unset)
 997 {
 998         char *tok, *tok2;
 999         char *endptr;
1000
1001         /*
1002          * --no-call-graph
1003          */
1004         if (unset) {
1005                 dont_use_callchains = true;
1006                 return 0;
1007         }
1008
1009         symbol_conf.use_callchain = true;
1010
1011         if (!arg)
1012                 return 0;
1013
1014         tok = strtok((char *)arg, ",");
1015         if (!tok)
1016                 return -1;
1017
1018         /* get the output mode */
1019         if (!strncmp(tok, "graph", strlen(arg)))
1020                 callchain_param.mode = CHAIN_GRAPH_ABS;
1021
1022         else if (!strncmp(tok, "flat", strlen(arg)))
1023                 callchain_param.mode = CHAIN_FLAT;
1024
1025         else if (!strncmp(tok, "fractal", strlen(arg)))
1026                 callchain_param.mode = CHAIN_GRAPH_REL;
1027
1028         else if (!strncmp(tok, "none", strlen(arg))) {
1029                 callchain_param.mode = CHAIN_NONE;
1030                 symbol_conf.use_callchain = false;
1031
1032                 return 0;
1033         }
1034
1035         else
1036                 return -1;
1037
1038         /* get the min percentage */
1039         tok = strtok(NULL, ",");
1040         if (!tok)
1041                 goto setup;
1042
1043         callchain_param.min_percent = strtod(tok, &endptr);
1044         if (tok == endptr)
1045                 return -1;
1046
1047         /* get the print limit */
1048         tok2 = strtok(NULL, ",");
1049         if (!tok2)
1050                 goto setup;
1051
1052         if (tok2[0] != 'c') {
1053                 callchain_param.print_limit = strtod(tok2, &endptr);
1054                 tok2 = strtok(NULL, ",");
1055                 if (!tok2)
1056                         goto setup;
1057         }
1058
1059         /* get the call chain order */
1060         if (!strcmp(tok2, "caller"))
1061                 callchain_param.order = ORDER_CALLER;
1062         else if (!strcmp(tok2, "callee"))
1063                 callchain_param.order = ORDER_CALLEE;
1064         else
1065                 return -1;
1066 setup:
1067         if (callchain_register_param(&callchain_param) < 0) {
1068                 fprintf(stderr, "Can't register callchain params\n");
1069                 return -1;
1070         }
1071         return 0;
1072 }
1073
1074 static const char * const top_usage[] = {
1075         "perf top [<options>]",
1076         NULL
1077 };
1078
1079 static const struct option options[] = {
1080         OPT_CALLBACK('e', "event", &top.evlist, "event",
1081                      "event selector. use 'perf list' to list available events",
1082                      parse_events_option),
1083         OPT_INTEGER('c', "count", &default_interval,
1084                     "event period to sample"),
1085         OPT_INTEGER('p', "pid", &top.target_pid,
1086                     "profile events on existing process id"),
1087         OPT_INTEGER('t', "tid", &top.target_tid,
1088                     "profile events on existing thread id"),
1089         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1090                             "system-wide collection from all CPUs"),
1091         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1092                     "list of cpus to monitor"),
1093         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1094                    "file", "vmlinux pathname"),
1095         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1096                     "hide kernel symbols"),
1097         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1098         OPT_INTEGER('r', "realtime", &realtime_prio,
1099                     "collect data with this RT SCHED_FIFO priority"),
1100         OPT_INTEGER('d', "delay", &top.delay_secs,
1101                     "number of seconds to delay between refreshes"),
1102         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1103                             "dump the symbol table used for profiling"),
1104         OPT_INTEGER('f', "count-filter", &top.count_filter,
1105                     "only display functions with more events than this"),
1106         OPT_BOOLEAN('g', "group", &group,
1107                             "put the counters into a counter group"),
1108         OPT_BOOLEAN('i', "inherit", &inherit,
1109                     "child tasks inherit counters"),
1110         OPT_STRING(0, "sym-annotate", &sym_filter, "symbol name",
1111                     "symbol to annotate"),
1112         OPT_BOOLEAN('z', "zero", &top.zero,
1113                     "zero history across updates"),
1114         OPT_INTEGER('F', "freq", &top.freq,
1115                     "profile at this frequency"),
1116         OPT_INTEGER('E', "entries", &top.print_entries,
1117                     "display this many functions"),
1118         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1119                     "hide user symbols"),
1120         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1121         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1122         OPT_INCR('v', "verbose", &verbose,
1123                     "be more verbose (show counter open errors, etc)"),
1124         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1125                    "sort by key(s): pid, comm, dso, symbol, parent"),
1126         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1127                     "Show a column with the number of samples"),
1128         OPT_CALLBACK_DEFAULT('G', "call-graph", NULL, "output_type,min_percent, call_order",
1129                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1130                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1131                      callchain_default_opt),
1132         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1133                     "Show a column with the sum of periods"),
1134         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1135                    "only consider symbols in these dsos"),
1136         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1137                    "only consider symbols in these comms"),
1138         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1139                    "only consider these symbols"),
1140         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1141                     "Interleave source code with assembly code (default)"),
1142         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1143                     "Display raw encoding of assembly instructions (default)"),
1144         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1145                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1146         OPT_END()
1147 };
1148
1149 int cmd_top(int argc, const char **argv, const char *prefix __used)
1150 {
1151         struct perf_evsel *pos;
1152         int status = -ENOMEM;
1153
1154         top.evlist = perf_evlist__new(NULL, NULL);
1155         if (top.evlist == NULL)
1156                 return -ENOMEM;
1157
1158         symbol_conf.exclude_other = false;
1159
1160         argc = parse_options(argc, argv, options, top_usage, 0);
1161         if (argc)
1162                 usage_with_options(top_usage, options);
1163
1164         if (sort_order == default_sort_order)
1165                 sort_order = "dso,symbol";
1166
1167         setup_sorting(top_usage, options);
1168
1169         if (use_stdio)
1170                 use_browser = 0;
1171         else if (use_tui)
1172                 use_browser = 1;
1173
1174         setup_browser(false);
1175
1176         /* CPU and PID are mutually exclusive */
1177         if (top.target_tid > 0 && top.cpu_list) {
1178                 printf("WARNING: PID switch overriding CPU\n");
1179                 sleep(1);
1180                 top.cpu_list = NULL;
1181         }
1182
1183         if (top.target_pid != -1)
1184                 top.target_tid = top.target_pid;
1185
1186         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1187                                      top.target_tid, top.cpu_list) < 0)
1188                 usage_with_options(top_usage, options);
1189
1190         if (!top.evlist->nr_entries &&
1191             perf_evlist__add_default(top.evlist) < 0) {
1192                 pr_err("Not enough memory for event selector list\n");
1193                 return -ENOMEM;
1194         }
1195
1196         if (top.delay_secs < 1)
1197                 top.delay_secs = 1;
1198
1199         /*
1200          * User specified count overrides default frequency.
1201          */
1202         if (default_interval)
1203                 top.freq = 0;
1204         else if (top.freq) {
1205                 default_interval = top.freq;
1206         } else {
1207                 fprintf(stderr, "frequency and count are zero, aborting\n");
1208                 exit(EXIT_FAILURE);
1209         }
1210
1211         list_for_each_entry(pos, &top.evlist->entries, node) {
1212                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1213                                          top.evlist->threads->nr) < 0)
1214                         goto out_free_fd;
1215                 /*
1216                  * Fill in the ones not specifically initialized via -c:
1217                  */
1218                 if (pos->attr.sample_period)
1219                         continue;
1220
1221                 pos->attr.sample_period = default_interval;
1222         }
1223
1224         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1225             perf_evlist__alloc_mmap(top.evlist) < 0)
1226                 goto out_free_fd;
1227
1228         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1229
1230         symbol_conf.priv_size = sizeof(struct annotation);
1231
1232         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1233         if (symbol__init() < 0)
1234                 return -1;
1235
1236         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1237         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1238         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1239
1240         /*
1241          * Avoid annotation data structures overhead when symbols aren't on the
1242          * sort list.
1243          */
1244         sort_has_symbols = sort_sym.list.next != NULL;
1245
1246         get_term_dimensions(&winsize);
1247         if (top.print_entries == 0) {
1248                 update_print_entries(&winsize);
1249                 signal(SIGWINCH, sig_winch_handler);
1250         }
1251
1252         status = __cmd_top();
1253 out_free_fd:
1254         perf_evlist__delete(top.evlist);
1255
1256         return status;
1257 }