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 void get_term_dimensions(struct winsize *ws)
  68 {
  69         char *s = getenv("LINES");
  70
  71         if (s != NULL) {
  72                 ws->ws_row = atoi(s);
  73                 s = getenv("COLUMNS");
  74                 if (s != NULL) {
  75                         ws->ws_col = atoi(s);
  76                         if (ws->ws_row && ws->ws_col)
  77                                 return;
  78                 }
  79         }
  80 #ifdef TIOCGWINSZ
  81         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
  82             ws->ws_row && ws->ws_col)
  83                 return;
  84 #endif
  85         ws->ws_row = 25;
  86         ws->ws_col = 80;
  87 }
  88
  89 static void perf_top__update_print_entries(struct perf_top *top)
  90 {
  91         if (top->print_entries > 9)
  92                 top->print_entries -= 9;
  93 }
  94
  95 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
  96 {
  97         struct perf_top *top = arg;
  98
  99         get_term_dimensions(&top->winsize);
 100         if (!top->print_entries
 101             || (top->print_entries+4) > top->winsize.ws_row) {
 102                 top->print_entries = top->winsize.ws_row;
 103         } else {
 104                 top->print_entries += 4;
 105                 top->winsize.ws_row = top->print_entries;
 106         }
 107         perf_top__update_print_entries(top);
 108 }
 109
 110 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 111 {
 112         struct symbol *sym;
 113         struct annotation *notes;
 114         struct map *map;
 115         int err = -1;
 116
 117         if (!he || !he->ms.sym)
 118                 return -1;
 119
 120         sym = he->ms.sym;
 121         map = he->ms.map;
 122
 123         /*
 124          * We can't annotate with just /proc/kallsyms
 125          */
 126         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
 127                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 128                        "path\n", sym->name);
 129                 sleep(1);
 130                 return -1;
 131         }
 132
 133         notes = symbol__annotation(sym);
 134         if (notes->src != NULL) {
 135                 pthread_mutex_lock(&notes->lock);
 136                 goto out_assign;
 137         }
 138
 139         pthread_mutex_lock(&notes->lock);
 140
 141         if (symbol__alloc_hist(sym) < 0) {
 142                 pthread_mutex_unlock(&notes->lock);
 143                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 144                        sym->name);
 145                 sleep(1);
 146                 return err;
 147         }
 148
 149         err = symbol__annotate(sym, map, 0);
 150         if (err == 0) {
 151 out_assign:
 152                 top->sym_filter_entry = he;
 153         }
 154
 155         pthread_mutex_unlock(&notes->lock);
 156         return err;
 157 }
 158
 159 static void __zero_source_counters(struct hist_entry *he)
 160 {
 161         struct symbol *sym = he->ms.sym;
 162         symbol__annotate_zero_histograms(sym);
 163 }
 164
 165 static void perf_top__record_precise_ip(struct perf_top *top,
 166                                         struct hist_entry *he,
 167                                         int counter, u64 ip)
 168 {
 169         struct annotation *notes;
 170         struct symbol *sym;
 171
 172         if (he == NULL || he->ms.sym == NULL ||
 173             ((top->sym_filter_entry == NULL ||
 174               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 175                 return;
 176
 177         sym = he->ms.sym;
 178         notes = symbol__annotation(sym);
 179
 180         if (pthread_mutex_trylock(&notes->lock))
 181                 return;
 182
 183         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 184                 pthread_mutex_unlock(&notes->lock);
 185                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 186                        sym->name);
 187                 sleep(1);
 188                 return;
 189         }
 190
 191         ip = he->ms.map->map_ip(he->ms.map, ip);
 192         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 193
 194         pthread_mutex_unlock(&notes->lock);
 195 }
 196
 197 static void perf_top__show_details(struct perf_top *top)
 198 {
 199         struct hist_entry *he = top->sym_filter_entry;
 200         struct annotation *notes;
 201         struct symbol *symbol;
 202         int more;
 203
 204         if (!he)
 205                 return;
 206
 207         symbol = he->ms.sym;
 208         notes = symbol__annotation(symbol);
 209
 210         pthread_mutex_lock(&notes->lock);
 211
 212         if (notes->src == NULL)
 213                 goto out_unlock;
 214
 215         printf("Showing %s for %s\n", event_name(top->sym_evsel), symbol->name);
 216         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 217
 218         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
 219                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 220         if (top->zero)
 221                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 222         else
 223                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 224         if (more != 0)
 225                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 226 out_unlock:
 227         pthread_mutex_unlock(&notes->lock);
 228 }
 229
 230 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 231
 232 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 233                                                      struct addr_location *al,
 234                                                      struct perf_sample *sample)
 235 {
 236         struct hist_entry *he;
 237
 238         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 239         if (he == NULL)
 240                 return NULL;
 241
 242         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 243         return he;
 244 }
 245
 246 static void perf_top__print_sym_table(struct perf_top *top)
 247 {
 248         char bf[160];
 249         int printed = 0;
 250         const int win_width = top->winsize.ws_col - 1;
 251
 252         puts(CONSOLE_CLEAR);
 253
 254         perf_top__header_snprintf(top, bf, sizeof(bf));
 255         printf("%s\n", bf);
 256
 257         perf_top__reset_sample_counters(top);
 258
 259         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 260
 261         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 262             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 263                 top->sym_evsel->hists.stats.nr_lost_warned =
 264                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 265                 color_fprintf(stdout, PERF_COLOR_RED,
 266                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 267                               top->sym_evsel->hists.stats.nr_lost_warned);
 268                 ++printed;
 269         }
 270
 271         if (top->sym_filter_entry) {
 272                 perf_top__show_details(top);
 273                 return;
 274         }
 275
 276         hists__collapse_resort_threaded(&top->sym_evsel->hists);
 277         hists__output_resort_threaded(&top->sym_evsel->hists);
 278         hists__decay_entries_threaded(&top->sym_evsel->hists,
 279                                       top->hide_user_symbols,
 280                                       top->hide_kernel_symbols);
 281         hists__output_recalc_col_len(&top->sym_evsel->hists,
 282                                      top->winsize.ws_row - 3);
 283         putchar('\n');
 284         hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
 285                        top->winsize.ws_row - 4 - printed, win_width, stdout);
 286 }
 287
 288 static void prompt_integer(int *target, const char *msg)
 289 {
 290         char *buf = malloc(0), *p;
 291         size_t dummy = 0;
 292         int tmp;
 293
 294         fprintf(stdout, "\n%s: ", msg);
 295         if (getline(&buf, &dummy, stdin) < 0)
 296                 return;
 297
 298         p = strchr(buf, '\n');
 299         if (p)
 300                 *p = 0;
 301
 302         p = buf;
 303         while(*p) {
 304                 if (!isdigit(*p))
 305                         goto out_free;
 306                 p++;
 307         }
 308         tmp = strtoul(buf, NULL, 10);
 309         *target = tmp;
 310 out_free:
 311         free(buf);
 312 }
 313
 314 static void prompt_percent(int *target, const char *msg)
 315 {
 316         int tmp = 0;
 317
 318         prompt_integer(&tmp, msg);
 319         if (tmp >= 0 && tmp <= 100)
 320                 *target = tmp;
 321 }
 322
 323 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 324 {
 325         char *buf = malloc(0), *p;
 326         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 327         struct rb_node *next;
 328         size_t dummy = 0;
 329
 330         /* zero counters of active symbol */
 331         if (syme) {
 332                 __zero_source_counters(syme);
 333                 top->sym_filter_entry = NULL;
 334         }
 335
 336         fprintf(stdout, "\n%s: ", msg);
 337         if (getline(&buf, &dummy, stdin) < 0)
 338                 goto out_free;
 339
 340         p = strchr(buf, '\n');
 341         if (p)
 342                 *p = 0;
 343
 344         next = rb_first(&top->sym_evsel->hists.entries);
 345         while (next) {
 346                 n = rb_entry(next, struct hist_entry, rb_node);
 347                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 348                         found = n;
 349                         break;
 350                 }
 351                 next = rb_next(&n->rb_node);
 352         }
 353
 354         if (!found) {
 355                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 356                 sleep(1);
 357         } else
 358                 perf_top__parse_source(top, found);
 359
 360 out_free:
 361         free(buf);
 362 }
 363
 364 static void perf_top__print_mapped_keys(struct perf_top *top)
 365 {
 366         char *name = NULL;
 367
 368         if (top->sym_filter_entry) {
 369                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 370                 name = sym->name;
 371         }
 372
 373         fprintf(stdout, "\nMapped keys:\n");
 374         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 375         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 376
 377         if (top->evlist->nr_entries > 1)
 378                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top->sym_evsel));
 379
 380         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 381
 382         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 383         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 384         fprintf(stdout, "\t[S]     stop annotation.\n");
 385
 386         fprintf(stdout,
 387                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 388                 top->hide_kernel_symbols ? "yes" : "no");
 389         fprintf(stdout,
 390                 "\t[U]     hide user symbols.               \t(%s)\n",
 391                 top->hide_user_symbols ? "yes" : "no");
 392         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 393         fprintf(stdout, "\t[qQ]    quit.\n");
 394 }
 395
 396 static int perf_top__key_mapped(struct perf_top *top, int c)
 397 {
 398         switch (c) {
 399                 case 'd':
 400                 case 'e':
 401                 case 'f':
 402                 case 'z':
 403                 case 'q':
 404                 case 'Q':
 405                 case 'K':
 406                 case 'U':
 407                 case 'F':
 408                 case 's':
 409                 case 'S':
 410                         return 1;
 411                 case 'E':
 412                         return top->evlist->nr_entries > 1 ? 1 : 0;
 413                 default:
 414                         break;
 415         }
 416
 417         return 0;
 418 }
 419
 420 static void perf_top__handle_keypress(struct perf_top *top, int c)
 421 {
 422         if (!perf_top__key_mapped(top, c)) {
 423                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 424                 struct termios tc, save;
 425
 426                 perf_top__print_mapped_keys(top);
 427                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 428                 fflush(stdout);
 429
 430                 tcgetattr(0, &save);
 431                 tc = save;
 432                 tc.c_lflag &= ~(ICANON | ECHO);
 433                 tc.c_cc[VMIN] = 0;
 434                 tc.c_cc[VTIME] = 0;
 435                 tcsetattr(0, TCSANOW, &tc);
 436
 437                 poll(&stdin_poll, 1, -1);
 438                 c = getc(stdin);
 439
 440                 tcsetattr(0, TCSAFLUSH, &save);
 441                 if (!perf_top__key_mapped(top, c))
 442                         return;
 443         }
 444
 445         switch (c) {
 446                 case 'd':
 447                         prompt_integer(&top->delay_secs, "Enter display delay");
 448                         if (top->delay_secs < 1)
 449                                 top->delay_secs = 1;
 450                         break;
 451                 case 'e':
 452                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 453                         if (top->print_entries == 0) {
 454                                 struct sigaction act = {
 455                                         .sa_sigaction = perf_top__sig_winch,
 456                                         .sa_flags     = SA_SIGINFO,
 457                                 };
 458                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 459                                 sigaction(SIGWINCH, &act, NULL);
 460                         } else {
 461                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 462                                 signal(SIGWINCH, SIG_DFL);
 463                         }
 464                         break;
 465                 case 'E':
 466                         if (top->evlist->nr_entries > 1) {
 467                                 /* Select 0 as the default event: */
 468                                 int counter = 0;
 469
 470                                 fprintf(stderr, "\nAvailable events:");
 471
 472                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 473                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
 474
 475                                 prompt_integer(&counter, "Enter details event counter");
 476
 477                                 if (counter >= top->evlist->nr_entries) {
 478                                         top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
 479                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
 480                                         sleep(1);
 481                                         break;
 482                                 }
 483                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 484                                         if (top->sym_evsel->idx == counter)
 485                                                 break;
 486                         } else
 487                                 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
 488                         break;
 489                 case 'f':
 490                         prompt_integer(&top->count_filter, "Enter display event count filter");
 491                         break;
 492                 case 'F':
 493                         prompt_percent(&top->sym_pcnt_filter,
 494                                        "Enter details display event filter (percent)");
 495                         break;
 496                 case 'K':
 497                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 498                         break;
 499                 case 'q':
 500                 case 'Q':
 501                         printf("exiting.\n");
 502                         if (top->dump_symtab)
 503                                 perf_session__fprintf_dsos(top->session, stderr);
 504                         exit(0);
 505                 case 's':
 506                         perf_top__prompt_symbol(top, "Enter details symbol");
 507                         break;
 508                 case 'S':
 509                         if (!top->sym_filter_entry)
 510                                 break;
 511                         else {
 512                                 struct hist_entry *syme = top->sym_filter_entry;
 513
 514                                 top->sym_filter_entry = NULL;
 515                                 __zero_source_counters(syme);
 516                         }
 517                         break;
 518                 case 'U':
 519                         top->hide_user_symbols = !top->hide_user_symbols;
 520                         break;
 521                 case 'z':
 522                         top->zero = !top->zero;
 523                         break;
 524                 default:
 525                         break;
 526         }
 527 }
 528
 529 static void perf_top__sort_new_samples(void *arg)
 530 {
 531         struct perf_top *t = arg;
 532         perf_top__reset_sample_counters(t);
 533
 534         if (t->evlist->selected != NULL)
 535                 t->sym_evsel = t->evlist->selected;
 536
 537         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 538         hists__output_resort_threaded(&t->sym_evsel->hists);
 539         hists__decay_entries_threaded(&t->sym_evsel->hists,
 540                                       t->hide_user_symbols,
 541                                       t->hide_kernel_symbols);
 542 }
 543
 544 static void *display_thread_tui(void *arg)
 545 {
 546         struct perf_evsel *pos;
 547         struct perf_top *top = arg;
 548         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 549
 550         perf_top__sort_new_samples(top);
 551
 552         /*
 553          * Initialize the uid_filter_str, in the future the TUI will allow
 554          * Zooming in/out UIDs. For now juse use whatever the user passed
 555          * via --uid.
 556          */
 557         list_for_each_entry(pos, &top->evlist->entries, node)
 558                 pos->hists.uid_filter_str = top->uid_str;
 559
 560         perf_evlist__tui_browse_hists(top->evlist, help,
 561                                       perf_top__sort_new_samples,
 562                                       top, top->delay_secs);
 563
 564         exit_browser(0);
 565         exit(0);
 566         return NULL;
 567 }
 568
 569 static void *display_thread(void *arg)
 570 {
 571         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 572         struct termios tc, save;
 573         struct perf_top *top = arg;
 574         int delay_msecs, c;
 575
 576         tcgetattr(0, &save);
 577         tc = save;
 578         tc.c_lflag &= ~(ICANON | ECHO);
 579         tc.c_cc[VMIN] = 0;
 580         tc.c_cc[VTIME] = 0;
 581
 582         pthread__unblock_sigwinch();
 583 repeat:
 584         delay_msecs = top->delay_secs * 1000;
 585         tcsetattr(0, TCSANOW, &tc);
 586         /* trash return*/
 587         getc(stdin);
 588
 589         while (1) {
 590                 perf_top__print_sym_table(top);
 591                 /*
 592                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 593                  * refresh screen in both cases.
 594                  */
 595                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 596                 case 0:
 597                         continue;
 598                 case -1:
 599                         if (errno == EINTR)
 600                                 continue;
 601                         /* Fall trhu */
 602                 default:
 603                         goto process_hotkey;
 604                 }
 605         }
 606 process_hotkey:
 607         c = getc(stdin);
 608         tcsetattr(0, TCSAFLUSH, &save);
 609
 610         perf_top__handle_keypress(top, c);
 611         goto repeat;
 612
 613         return NULL;
 614 }
 615
 616 /* Tag samples to be skipped. */
 617 static const char *skip_symbols[] = {
 618         "intel_idle",
 619         "default_idle",
 620         "native_safe_halt",
 621         "cpu_idle",
 622         "enter_idle",
 623         "exit_idle",
 624         "mwait_idle",
 625         "mwait_idle_with_hints",
 626         "poll_idle",
 627         "ppc64_runlatch_off",
 628         "pseries_dedicated_idle_sleep",
 629         NULL
 630 };
 631
 632 static int symbol_filter(struct map *map __used, struct symbol *sym)
 633 {
 634         const char *name = sym->name;
 635         int i;
 636
 637         /*
 638          * ppc64 uses function descriptors and appends a '.' to the
 639          * start of every instruction address. Remove it.
 640          */
 641         if (name[0] == '.')
 642                 name++;
 643
 644         if (!strcmp(name, "_text") ||
 645             !strcmp(name, "_etext") ||
 646             !strcmp(name, "_sinittext") ||
 647             !strncmp("init_module", name, 11) ||
 648             !strncmp("cleanup_module", name, 14) ||
 649             strstr(name, "_text_start") ||
 650             strstr(name, "_text_end"))
 651                 return 1;
 652
 653         for (i = 0; skip_symbols[i]; i++) {
 654                 if (!strcmp(skip_symbols[i], name)) {
 655                         sym->ignore = true;
 656                         break;
 657                 }
 658         }
 659
 660         return 0;
 661 }
 662
 663 static void perf_event__process_sample(struct perf_tool *tool,
 664                                        const union perf_event *event,
 665                                        struct perf_evsel *evsel,
 666                                        struct perf_sample *sample,
 667                                        struct machine *machine)
 668 {
 669         struct perf_top *top = container_of(tool, struct perf_top, tool);
 670         struct symbol *parent = NULL;
 671         u64 ip = event->ip.ip;
 672         struct addr_location al;
 673         int err;
 674
 675         if (!machine && perf_guest) {
 676                 pr_err("Can't find guest [%d]'s kernel information\n",
 677                         event->ip.pid);
 678                 return;
 679         }
 680
 681         if (!machine) {
 682                 pr_err("%u unprocessable samples recorded.",
 683                        top->session->hists.stats.nr_unprocessable_samples++);
 684                 return;
 685         }
 686
 687         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 688                 top->exact_samples++;
 689
 690         if (perf_event__preprocess_sample(event, machine, &al, sample,
 691                                           symbol_filter) < 0 ||
 692             al.filtered)
 693                 return;
 694
 695         if (!top->kptr_restrict_warned &&
 696             symbol_conf.kptr_restrict &&
 697             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 698                 ui__warning(
 699 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 700 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 701 "Kernel%s samples will not be resolved.\n",
 702                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 703                           " modules" : "");
 704                 if (use_browser <= 0)
 705                         sleep(5);
 706                 top->kptr_restrict_warned = true;
 707         }
 708
 709         if (al.sym == NULL) {
 710                 const char *msg = "Kernel samples will not be resolved.\n";
 711                 /*
 712                  * As we do lazy loading of symtabs we only will know if the
 713                  * specified vmlinux file is invalid when we actually have a
 714                  * hit in kernel space and then try to load it. So if we get
 715                  * here and there are _no_ symbols in the DSO backing the
 716                  * kernel map, bail out.
 717                  *
 718                  * We may never get here, for instance, if we use -K/
 719                  * --hide-kernel-symbols, even if the user specifies an
 720                  * invalid --vmlinux ;-)
 721                  */
 722                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 723                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 724                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 725                         if (symbol_conf.vmlinux_name) {
 726                                 ui__warning("The %s file can't be used.\n%s",
 727                                             symbol_conf.vmlinux_name, msg);
 728                         } else {
 729                                 ui__warning("A vmlinux file was not found.\n%s",
 730                                             msg);
 731                         }
 732
 733                         if (use_browser <= 0)
 734                                 sleep(5);
 735                         top->vmlinux_warned = true;
 736                 }
 737         }
 738
 739         if (al.sym == NULL || !al.sym->ignore) {
 740                 struct hist_entry *he;
 741
 742                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 743                     sample->callchain) {
 744                         err = machine__resolve_callchain(machine, evsel, al.thread,
 745                                                          sample->callchain, &parent);
 746                         if (err)
 747                                 return;
 748                 }
 749
 750                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 751                 if (he == NULL) {
 752                         pr_err("Problem incrementing symbol period, skipping event\n");
 753                         return;
 754                 }
 755
 756                 if (symbol_conf.use_callchain) {
 757                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
 758                                                sample->period);
 759                         if (err)
 760                                 return;
 761                 }
 762
 763                 if (top->sort_has_symbols)
 764                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 765         }
 766
 767         return;
 768 }
 769
 770 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 771 {
 772         struct perf_sample sample;
 773         struct perf_evsel *evsel;
 774         struct perf_session *session = top->session;
 775         union perf_event *event;
 776         struct machine *machine;
 777         u8 origin;
 778         int ret;
 779
 780         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 781                 ret = perf_session__parse_sample(session, event, &sample);
 782                 if (ret) {
 783                         pr_err("Can't parse sample, err = %d\n", ret);
 784                         continue;
 785                 }
 786
 787                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 788                 assert(evsel != NULL);
 789
 790                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 791
 792                 if (event->header.type == PERF_RECORD_SAMPLE)
 793                         ++top->samples;
 794
 795                 switch (origin) {
 796                 case PERF_RECORD_MISC_USER:
 797                         ++top->us_samples;
 798                         if (top->hide_user_symbols)
 799                                 continue;
 800                         machine = perf_session__find_host_machine(session);
 801                         break;
 802                 case PERF_RECORD_MISC_KERNEL:
 803                         ++top->kernel_samples;
 804                         if (top->hide_kernel_symbols)
 805                                 continue;
 806                         machine = perf_session__find_host_machine(session);
 807                         break;
 808                 case PERF_RECORD_MISC_GUEST_KERNEL:
 809                         ++top->guest_kernel_samples;
 810                         machine = perf_session__find_machine(session, event->ip.pid);
 811                         break;
 812                 case PERF_RECORD_MISC_GUEST_USER:
 813                         ++top->guest_us_samples;
 814                         /*
 815                          * TODO: we don't process guest user from host side
 816                          * except simple counting.
 817                          */
 818                         /* Fall thru */
 819                 default:
 820                         continue;
 821                 }
 822
 823
 824                 if (event->header.type == PERF_RECORD_SAMPLE) {
 825                         perf_event__process_sample(&top->tool, event, evsel,
 826                                                    &sample, machine);
 827                 } else if (event->header.type < PERF_RECORD_MAX) {
 828                         hists__inc_nr_events(&evsel->hists, event->header.type);
 829                         perf_event__process(&top->tool, event, &sample, machine);
 830                 } else
 831                         ++session->hists.stats.nr_unknown_events;
 832         }
 833 }
 834
 835 static void perf_top__mmap_read(struct perf_top *top)
 836 {
 837         int i;
 838
 839         for (i = 0; i < top->evlist->nr_mmaps; i++)
 840                 perf_top__mmap_read_idx(top, i);
 841 }
 842
 843 static void perf_top__start_counters(struct perf_top *top)
 844 {
 845         struct perf_evsel *counter, *first;
 846         struct perf_evlist *evlist = top->evlist;
 847
 848         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 849
 850         list_for_each_entry(counter, &evlist->entries, node) {
 851                 struct perf_event_attr *attr = &counter->attr;
 852                 struct xyarray *group_fd = NULL;
 853
 854                 if (top->group && counter != first)
 855                         group_fd = first->fd;
 856
 857                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 858
 859                 if (top->freq) {
 860                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 861                         attr->freq        = 1;
 862                         attr->sample_freq = top->freq;
 863                 }
 864
 865                 if (evlist->nr_entries > 1) {
 866                         attr->sample_type |= PERF_SAMPLE_ID;
 867                         attr->read_format |= PERF_FORMAT_ID;
 868                 }
 869
 870                 if (symbol_conf.use_callchain)
 871                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
 872
 873                 attr->mmap = 1;
 874                 attr->comm = 1;
 875                 attr->inherit = top->inherit;
 876 fallback_missing_features:
 877                 if (top->exclude_guest_missing)
 878                         attr->exclude_guest = attr->exclude_host = 0;
 879 retry_sample_id:
 880                 attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
 881 try_again:
 882                 if (perf_evsel__open(counter, top->evlist->cpus,
 883                                      top->evlist->threads, top->group,
 884                                      group_fd) < 0) {
 885                         int err = errno;
 886
 887                         if (err == EPERM || err == EACCES) {
 888                                 ui__error_paranoid();
 889                                 goto out_err;
 890                         } else if (err == EINVAL) {
 891                                 if (!top->exclude_guest_missing &&
 892                                     (attr->exclude_guest || attr->exclude_host)) {
 893                                         pr_debug("Old kernel, cannot exclude "
 894                                                  "guest or host samples.\n");
 895                                         top->exclude_guest_missing = true;
 896                                         goto fallback_missing_features;
 897                                 } else if (!top->sample_id_all_missing) {
 898                                         /*
 899                                          * Old kernel, no attr->sample_id_type_all field
 900                                          */
 901                                         top->sample_id_all_missing = true;
 902                                         goto retry_sample_id;
 903                                 }
 904                         }
 905                         /*
 906                          * If it's cycles then fall back to hrtimer
 907                          * based cpu-clock-tick sw counter, which
 908                          * is always available even if no PMU support:
 909                          */
 910                         if (attr->type == PERF_TYPE_HARDWARE &&
 911                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 912                                 if (verbose)
 913                                         ui__warning("Cycles event not supported,\n"
 914                                                     "trying to fall back to cpu-clock-ticks\n");
 915
 916                                 attr->type = PERF_TYPE_SOFTWARE;
 917                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 918                                 goto try_again;
 919                         }
 920
 921                         if (err == ENOENT) {
 922                                 ui__warning("The %s event is not supported.\n",
 923                                             event_name(counter));
 924                                 goto out_err;
 925                         } else if (err == EMFILE) {
 926                                 ui__warning("Too many events are opened.\n"
 927                                             "Try again after reducing the number of events\n");
 928                                 goto out_err;
 929                         }
 930
 931                         ui__warning("The sys_perf_event_open() syscall "
 932                                     "returned with %d (%s).  /bin/dmesg "
 933                                     "may provide additional information.\n"
 934                                     "No CONFIG_PERF_EVENTS=y kernel support "
 935                                     "configured?\n", err, strerror(err));
 936                         goto out_err;
 937                 }
 938         }
 939
 940         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
 941                 ui__warning("Failed to mmap with %d (%s)\n",
 942                             errno, strerror(errno));
 943                 goto out_err;
 944         }
 945
 946         return;
 947
 948 out_err:
 949         exit_browser(0);
 950         exit(0);
 951 }
 952
 953 static int perf_top__setup_sample_type(struct perf_top *top)
 954 {
 955         if (!top->sort_has_symbols) {
 956                 if (symbol_conf.use_callchain) {
 957                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
 958                         return -EINVAL;
 959                 }
 960         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
 961                 if (callchain_register_param(&callchain_param) < 0) {
 962                         ui__warning("Can't register callchain params.\n");
 963                         return -EINVAL;
 964                 }
 965         }
 966
 967         return 0;
 968 }
 969
 970 static int __cmd_top(struct perf_top *top)
 971 {
 972         pthread_t thread;
 973         int ret;
 974         /*
 975          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 976          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 977          */
 978         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 979         if (top->session == NULL)
 980                 return -ENOMEM;
 981
 982         ret = perf_top__setup_sample_type(top);
 983         if (ret)
 984                 goto out_delete;
 985
 986         if (top->target_tid || top->uid != UINT_MAX)
 987                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
 988                                                   perf_event__process,
 989                                                   &top->session->host_machine);
 990         else
 991                 perf_event__synthesize_threads(&top->tool, perf_event__process,
 992                                                &top->session->host_machine);
 993         perf_top__start_counters(top);
 994         top->session->evlist = top->evlist;
 995         perf_session__update_sample_type(top->session);
 996
 997         /* Wait for a minimal set of events before starting the snapshot */
 998         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 999
1000         perf_top__mmap_read(top);
1001
1002         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1003                                                             display_thread), top)) {
1004                 printf("Could not create display thread.\n");
1005                 exit(-1);
1006         }
1007
1008         if (top->realtime_prio) {
1009                 struct sched_param param;
1010
1011                 param.sched_priority = top->realtime_prio;
1012                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1013                         printf("Could not set realtime priority.\n");
1014                         exit(-1);
1015                 }
1016         }
1017
1018         while (1) {
1019                 u64 hits = top->samples;
1020
1021                 perf_top__mmap_read(top);
1022
1023                 if (hits == top->samples)
1024                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1025         }
1026
1027 out_delete:
1028         perf_session__delete(top->session);
1029         top->session = NULL;
1030
1031         return 0;
1032 }
1033
1034 static int
1035 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1036 {
1037         struct perf_top *top = (struct perf_top *)opt->value;
1038         char *tok, *tok2;
1039         char *endptr;
1040
1041         /*
1042          * --no-call-graph
1043          */
1044         if (unset) {
1045                 top->dont_use_callchains = true;
1046                 return 0;
1047         }
1048
1049         symbol_conf.use_callchain = true;
1050
1051         if (!arg)
1052                 return 0;
1053
1054         tok = strtok((char *)arg, ",");
1055         if (!tok)
1056                 return -1;
1057
1058         /* get the output mode */
1059         if (!strncmp(tok, "graph", strlen(arg)))
1060                 callchain_param.mode = CHAIN_GRAPH_ABS;
1061
1062         else if (!strncmp(tok, "flat", strlen(arg)))
1063                 callchain_param.mode = CHAIN_FLAT;
1064
1065         else if (!strncmp(tok, "fractal", strlen(arg)))
1066                 callchain_param.mode = CHAIN_GRAPH_REL;
1067
1068         else if (!strncmp(tok, "none", strlen(arg))) {
1069                 callchain_param.mode = CHAIN_NONE;
1070                 symbol_conf.use_callchain = false;
1071
1072                 return 0;
1073         } else
1074                 return -1;
1075
1076         /* get the min percentage */
1077         tok = strtok(NULL, ",");
1078         if (!tok)
1079                 goto setup;
1080
1081         callchain_param.min_percent = strtod(tok, &endptr);
1082         if (tok == endptr)
1083                 return -1;
1084
1085         /* get the print limit */
1086         tok2 = strtok(NULL, ",");
1087         if (!tok2)
1088                 goto setup;
1089
1090         if (tok2[0] != 'c') {
1091                 callchain_param.print_limit = strtod(tok2, &endptr);
1092                 tok2 = strtok(NULL, ",");
1093                 if (!tok2)
1094                         goto setup;
1095         }
1096
1097         /* get the call chain order */
1098         if (!strcmp(tok2, "caller"))
1099                 callchain_param.order = ORDER_CALLER;
1100         else if (!strcmp(tok2, "callee"))
1101                 callchain_param.order = ORDER_CALLEE;
1102         else
1103                 return -1;
1104 setup:
1105         if (callchain_register_param(&callchain_param) < 0) {
1106                 fprintf(stderr, "Can't register callchain params\n");
1107                 return -1;
1108         }
1109         return 0;
1110 }
1111
1112 static const char * const top_usage[] = {
1113         "perf top [<options>]",
1114         NULL
1115 };
1116
1117 int cmd_top(int argc, const char **argv, const char *prefix __used)
1118 {
1119         struct perf_evsel *pos;
1120         int status = -ENOMEM;
1121         struct perf_top top = {
1122                 .count_filter        = 5,
1123                 .delay_secs          = 2,
1124                 .uid                 = UINT_MAX,
1125                 .freq                = 1000, /* 1 KHz */
1126                 .mmap_pages          = 128,
1127                 .sym_pcnt_filter     = 5,
1128         };
1129         char callchain_default_opt[] = "fractal,0.5,callee";
1130         const struct option options[] = {
1131         OPT_CALLBACK('e', "event", &top.evlist, "event",
1132                      "event selector. use 'perf list' to list available events",
1133                      parse_events_option),
1134         OPT_INTEGER('c', "count", &top.default_interval,
1135                     "event period to sample"),
1136         OPT_STRING('p', "pid", &top.target_pid, "pid",
1137                     "profile events on existing process id"),
1138         OPT_STRING('t', "tid", &top.target_tid, "tid",
1139                     "profile events on existing thread id"),
1140         OPT_BOOLEAN('a', "all-cpus", &top.system_wide,
1141                             "system-wide collection from all CPUs"),
1142         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1143                     "list of cpus to monitor"),
1144         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1145                    "file", "vmlinux pathname"),
1146         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1147                     "hide kernel symbols"),
1148         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1149         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1150                     "collect data with this RT SCHED_FIFO priority"),
1151         OPT_INTEGER('d', "delay", &top.delay_secs,
1152                     "number of seconds to delay between refreshes"),
1153         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1154                             "dump the symbol table used for profiling"),
1155         OPT_INTEGER('f', "count-filter", &top.count_filter,
1156                     "only display functions with more events than this"),
1157         OPT_BOOLEAN('g', "group", &top.group,
1158                             "put the counters into a counter group"),
1159         OPT_BOOLEAN('i', "inherit", &top.inherit,
1160                     "child tasks inherit counters"),
1161         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1162                     "symbol to annotate"),
1163         OPT_BOOLEAN('z', "zero", &top.zero,
1164                     "zero history across updates"),
1165         OPT_INTEGER('F', "freq", &top.freq,
1166                     "profile at this frequency"),
1167         OPT_INTEGER('E', "entries", &top.print_entries,
1168                     "display this many functions"),
1169         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1170                     "hide user symbols"),
1171         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1172         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1173         OPT_INCR('v', "verbose", &verbose,
1174                     "be more verbose (show counter open errors, etc)"),
1175         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1176                    "sort by key(s): pid, comm, dso, symbol, parent"),
1177         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1178                     "Show a column with the number of samples"),
1179         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1180                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1181                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1182                      callchain_default_opt),
1183         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1184                     "Show a column with the sum of periods"),
1185         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1186                    "only consider symbols in these dsos"),
1187         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1188                    "only consider symbols in these comms"),
1189         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1190                    "only consider these symbols"),
1191         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1192                     "Interleave source code with assembly code (default)"),
1193         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1194                     "Display raw encoding of assembly instructions (default)"),
1195         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1196                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1197         OPT_STRING('u', "uid", &top.uid_str, "user", "user to profile"),
1198         OPT_END()
1199         };
1200
1201         top.evlist = perf_evlist__new(NULL, NULL);
1202         if (top.evlist == NULL)
1203                 return -ENOMEM;
1204
1205         symbol_conf.exclude_other = false;
1206
1207         argc = parse_options(argc, argv, options, top_usage, 0);
1208         if (argc)
1209                 usage_with_options(top_usage, options);
1210
1211         if (sort_order == default_sort_order)
1212                 sort_order = "dso,symbol";
1213
1214         setup_sorting(top_usage, options);
1215
1216         if (top.use_stdio)
1217                 use_browser = 0;
1218         else if (top.use_tui)
1219                 use_browser = 1;
1220
1221         setup_browser(false);
1222
1223         top.uid = parse_target_uid(top.uid_str, top.target_tid, top.target_pid);
1224         if (top.uid_str != NULL && top.uid == UINT_MAX - 1)
1225                 goto out_delete_evlist;
1226
1227         /* CPU and PID are mutually exclusive */
1228         if (top.target_tid && top.cpu_list) {
1229                 printf("WARNING: PID switch overriding CPU\n");
1230                 sleep(1);
1231                 top.cpu_list = NULL;
1232         }
1233
1234         if (top.target_pid)
1235                 top.target_tid = top.target_pid;
1236
1237         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1238                                      top.target_tid, top.uid, top.cpu_list) < 0)
1239                 usage_with_options(top_usage, options);
1240
1241         if (!top.evlist->nr_entries &&
1242             perf_evlist__add_default(top.evlist) < 0) {
1243                 pr_err("Not enough memory for event selector list\n");
1244                 return -ENOMEM;
1245         }
1246
1247         symbol_conf.nr_events = top.evlist->nr_entries;
1248
1249         if (top.delay_secs < 1)
1250                 top.delay_secs = 1;
1251
1252         /*
1253          * User specified count overrides default frequency.
1254          */
1255         if (top.default_interval)
1256                 top.freq = 0;
1257         else if (top.freq) {
1258                 top.default_interval = top.freq;
1259         } else {
1260                 fprintf(stderr, "frequency and count are zero, aborting\n");
1261                 exit(EXIT_FAILURE);
1262         }
1263
1264         list_for_each_entry(pos, &top.evlist->entries, node) {
1265                 /*
1266                  * Fill in the ones not specifically initialized via -c:
1267                  */
1268                 if (!pos->attr.sample_period)
1269                         pos->attr.sample_period = top.default_interval;
1270         }
1271
1272         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1273
1274         symbol_conf.priv_size = sizeof(struct annotation);
1275
1276         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1277         if (symbol__init() < 0)
1278                 return -1;
1279
1280         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1281         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1282         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1283
1284         /*
1285          * Avoid annotation data structures overhead when symbols aren't on the
1286          * sort list.
1287          */
1288         top.sort_has_symbols = sort_sym.list.next != NULL;
1289
1290         get_term_dimensions(&top.winsize);
1291         if (top.print_entries == 0) {
1292                 struct sigaction act = {
1293                         .sa_sigaction = perf_top__sig_winch,
1294                         .sa_flags     = SA_SIGINFO,
1295                 };
1296                 perf_top__update_print_entries(&top);
1297                 sigaction(SIGWINCH, &act, NULL);
1298         }
1299
1300         status = __cmd_top(&top);
1301
1302 out_delete_evlist:
1303         perf_evlist__delete(top.evlist);
1304
1305         return status;
1306 }