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  *
   9  * Improvements and fixes by:
  10  *
  11  *   Arjan van de Ven <arjan@linux.intel.com>
  12  *   Yanmin Zhang <yanmin.zhang@intel.com>
  13  *   Wu Fengguang <fengguang.wu@intel.com>
  14  *   Mike Galbraith <efault@gmx.de>
  15  *   Paul Mackerras <paulus@samba.org>
  16  *
  17  * Released under the GPL v2. (and only v2, not any later version)
  18  */
  19 #include "builtin.h"
  20
  21 #include "perf.h"
  22
  23 #include "util/annotate.h"
  24 #include "util/cache.h"
  25 #include "util/color.h"
  26 #include "util/evlist.h"
  27 #include "util/evsel.h"
  28 #include "util/session.h"
  29 #include "util/symbol.h"
  30 #include "util/thread.h"
  31 #include "util/thread_map.h"
  32 #include "util/top.h"
  33 #include "util/util.h"
  34 #include <linux/rbtree.h>
  35 #include "util/parse-options.h"
  36 #include "util/parse-events.h"
  37 #include "util/cpumap.h"
  38 #include "util/xyarray.h"
  39
  40 #include "util/debug.h"
  41
  42 #include <assert.h>
  43 #include <fcntl.h>
  44
  45 #include <stdio.h>
  46 #include <termios.h>
  47 #include <unistd.h>
  48 #include <inttypes.h>
  49
  50 #include <errno.h>
  51 #include <time.h>
  52 #include <sched.h>
  53
  54 #include <sys/syscall.h>
  55 #include <sys/ioctl.h>
  56 #include <sys/poll.h>
  57 #include <sys/prctl.h>
  58 #include <sys/wait.h>
  59 #include <sys/uio.h>
  60 #include <sys/mman.h>
  61
  62 #include <linux/unistd.h>
  63 #include <linux/types.h>
  64
  65 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  66
  67 static struct perf_top top = {
  68         .count_filter           = 5,
  69         .delay_secs             = 2,
  70         .display_weighted       = -1,
  71         .target_pid             = -1,
  72         .target_tid             = -1,
  73         .active_symbols         = LIST_HEAD_INIT(top.active_symbols),
  74         .active_symbols_lock    = PTHREAD_MUTEX_INITIALIZER,
  75         .active_symbols_cond    = PTHREAD_COND_INITIALIZER,
  76         .freq                   = 1000, /* 1 KHz */
  77 };
  78
  79 static bool                     system_wide                     =  false;
  80
  81 static bool                     use_tui, use_stdio;
  82
  83 static int                      default_interval                =      0;
  84
  85 static bool                     inherit                         =  false;
  86 static int                      realtime_prio                   =      0;
  87 static bool                     group                           =  false;
  88 static unsigned int             page_size;
  89 static unsigned int             mmap_pages                      =    128;
  90
  91 static bool                     dump_symtab                     =  false;
  92
  93 static struct winsize           winsize;
  94
  95 static const char               *sym_filter                     =   NULL;
  96 struct sym_entry                *sym_filter_entry_sched         =   NULL;
  97 static int                      sym_pcnt_filter                 =      5;
  98
  99 /*
 100  * Source functions
 101  */
 102
 103 void get_term_dimensions(struct winsize *ws)
 104 {
 105         char *s = getenv("LINES");
 106
 107         if (s != NULL) {
 108                 ws->ws_row = atoi(s);
 109                 s = getenv("COLUMNS");
 110                 if (s != NULL) {
 111                         ws->ws_col = atoi(s);
 112                         if (ws->ws_row && ws->ws_col)
 113                                 return;
 114                 }
 115         }
 116 #ifdef TIOCGWINSZ
 117         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 118             ws->ws_row && ws->ws_col)
 119                 return;
 120 #endif
 121         ws->ws_row = 25;
 122         ws->ws_col = 80;
 123 }
 124
 125 static void update_print_entries(struct winsize *ws)
 126 {
 127         top.print_entries = ws->ws_row;
 128
 129         if (top.print_entries > 9)
 130                 top.print_entries -= 9;
 131 }
 132
 133 static void sig_winch_handler(int sig __used)
 134 {
 135         get_term_dimensions(&winsize);
 136         update_print_entries(&winsize);
 137 }
 138
 139 static int parse_source(struct sym_entry *syme)
 140 {
 141         struct symbol *sym;
 142         struct annotation *notes;
 143         struct map *map;
 144         int err = -1;
 145
 146         if (!syme)
 147                 return -1;
 148
 149         sym = sym_entry__symbol(syme);
 150         map = syme->map;
 151
 152         /*
 153          * We can't annotate with just /proc/kallsyms
 154          */
 155         if (map->dso->origin == DSO__ORIG_KERNEL) {
 156                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 157                        "path\n", sym->name);
 158                 sleep(1);
 159                 return -1;
 160         }
 161
 162         notes = symbol__annotation(sym);
 163         if (notes->src != NULL) {
 164                 pthread_mutex_lock(&notes->lock);
 165                 goto out_assign;
 166         }
 167
 168         pthread_mutex_lock(&notes->lock);
 169
 170         if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
 171                 pthread_mutex_unlock(&notes->lock);
 172                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 173                        sym->name);
 174                 sleep(1);
 175                 return err;
 176         }
 177
 178         err = symbol__annotate(sym, syme->map, 0);
 179         if (err == 0) {
 180 out_assign:
 181                 top.sym_filter_entry = syme;
 182         }
 183
 184         pthread_mutex_unlock(&notes->lock);
 185         return err;
 186 }
 187
 188 static void __zero_source_counters(struct sym_entry *syme)
 189 {
 190         struct symbol *sym = sym_entry__symbol(syme);
 191         symbol__annotate_zero_histograms(sym);
 192 }
 193
 194 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
 195 {
 196         struct annotation *notes;
 197         struct symbol *sym;
 198
 199         if (syme != top.sym_filter_entry)
 200                 return;
 201
 202         sym = sym_entry__symbol(syme);
 203         notes = symbol__annotation(sym);
 204
 205         if (pthread_mutex_trylock(&notes->lock))
 206                 return;
 207
 208         ip = syme->map->map_ip(syme->map, ip);
 209         symbol__inc_addr_samples(sym, syme->map, counter, ip);
 210
 211         pthread_mutex_unlock(&notes->lock);
 212 }
 213
 214 static void show_details(struct sym_entry *syme)
 215 {
 216         struct annotation *notes;
 217         struct symbol *symbol;
 218         int more;
 219
 220         if (!syme)
 221                 return;
 222
 223         symbol = sym_entry__symbol(syme);
 224         notes = symbol__annotation(symbol);
 225
 226         pthread_mutex_lock(&notes->lock);
 227
 228         if (notes->src == NULL)
 229                 goto out_unlock;
 230
 231         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
 232         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 233
 234         more = symbol__annotate_printf(symbol, syme->map, top.sym_evsel->idx,
 235                                        0, sym_pcnt_filter, top.print_entries, 4);
 236         if (top.zero)
 237                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
 238         else
 239                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
 240         if (more != 0)
 241                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 242 out_unlock:
 243         pthread_mutex_unlock(&notes->lock);
 244 }
 245
 246 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 247
 248 static void __list_insert_active_sym(struct sym_entry *syme)
 249 {
 250         list_add(&syme->node, &top.active_symbols);
 251 }
 252
 253 static void print_sym_table(struct perf_session *session)
 254 {
 255         char bf[160];
 256         int printed = 0;
 257         struct rb_node *nd;
 258         struct sym_entry *syme;
 259         struct rb_root tmp = RB_ROOT;
 260         const int win_width = winsize.ws_col - 1;
 261         int sym_width, dso_width, dso_short_width;
 262         float sum_ksamples = perf_top__decay_samples(&top, &tmp);
 263
 264         puts(CONSOLE_CLEAR);
 265
 266         perf_top__header_snprintf(&top, bf, sizeof(bf));
 267         printf("%s\n", bf);
 268
 269         perf_top__reset_sample_counters(&top);
 270
 271         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 272
 273         if (session->hists.stats.total_lost != 0) {
 274                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
 275                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
 276                        session->hists.stats.total_lost);
 277         }
 278
 279         if (top.sym_filter_entry) {
 280                 show_details(top.sym_filter_entry);
 281                 return;
 282         }
 283
 284         perf_top__find_widths(&top, &tmp, &dso_width, &dso_short_width,
 285                               &sym_width);
 286
 287         if (sym_width + dso_width > winsize.ws_col - 29) {
 288                 dso_width = dso_short_width;
 289                 if (sym_width + dso_width > winsize.ws_col - 29)
 290                         sym_width = winsize.ws_col - dso_width - 29;
 291         }
 292         putchar('\n');
 293         if (top.evlist->nr_entries == 1)
 294                 printf("             samples  pcnt");
 295         else
 296                 printf("   weight    samples  pcnt");
 297
 298         if (verbose)
 299                 printf("         RIP       ");
 300         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
 301         printf("   %s    _______ _____",
 302                top.evlist->nr_entries == 1 ? "      " : "______");
 303         if (verbose)
 304                 printf(" ________________");
 305         printf(" %-*.*s", sym_width, sym_width, graph_line);
 306         printf(" %-*.*s", dso_width, dso_width, graph_line);
 307         puts("\n");
 308
 309         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 310                 struct symbol *sym;
 311                 double pcnt;
 312
 313                 syme = rb_entry(nd, struct sym_entry, rb_node);
 314                 sym = sym_entry__symbol(syme);
 315                 if (++printed > top.print_entries ||
 316                     (int)syme->snap_count < top.count_filter)
 317                         continue;
 318
 319                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 320                                          sum_ksamples));
 321
 322                 if (top.evlist->nr_entries == 1 || !top.display_weighted)
 323                         printf("%20.2f ", syme->weight);
 324                 else
 325                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 326
 327                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 328                 if (verbose)
 329                         printf(" %016" PRIx64, sym->start);
 330                 printf(" %-*.*s", sym_width, sym_width, sym->name);
 331                 printf(" %-*.*s\n", dso_width, dso_width,
 332                        dso_width >= syme->map->dso->long_name_len ?
 333                                         syme->map->dso->long_name :
 334                                         syme->map->dso->short_name);
 335         }
 336 }
 337
 338 static void prompt_integer(int *target, const char *msg)
 339 {
 340         char *buf = malloc(0), *p;
 341         size_t dummy = 0;
 342         int tmp;
 343
 344         fprintf(stdout, "\n%s: ", msg);
 345         if (getline(&buf, &dummy, stdin) < 0)
 346                 return;
 347
 348         p = strchr(buf, '\n');
 349         if (p)
 350                 *p = 0;
 351
 352         p = buf;
 353         while(*p) {
 354                 if (!isdigit(*p))
 355                         goto out_free;
 356                 p++;
 357         }
 358         tmp = strtoul(buf, NULL, 10);
 359         *target = tmp;
 360 out_free:
 361         free(buf);
 362 }
 363
 364 static void prompt_percent(int *target, const char *msg)
 365 {
 366         int tmp = 0;
 367
 368         prompt_integer(&tmp, msg);
 369         if (tmp >= 0 && tmp <= 100)
 370                 *target = tmp;
 371 }
 372
 373 static void prompt_symbol(struct sym_entry **target, const char *msg)
 374 {
 375         char *buf = malloc(0), *p;
 376         struct sym_entry *syme = *target, *n, *found = NULL;
 377         size_t dummy = 0;
 378
 379         /* zero counters of active symbol */
 380         if (syme) {
 381                 __zero_source_counters(syme);
 382                 *target = NULL;
 383         }
 384
 385         fprintf(stdout, "\n%s: ", msg);
 386         if (getline(&buf, &dummy, stdin) < 0)
 387                 goto out_free;
 388
 389         p = strchr(buf, '\n');
 390         if (p)
 391                 *p = 0;
 392
 393         pthread_mutex_lock(&top.active_symbols_lock);
 394         syme = list_entry(top.active_symbols.next, struct sym_entry, node);
 395         pthread_mutex_unlock(&top.active_symbols_lock);
 396
 397         list_for_each_entry_safe_from(syme, n, &top.active_symbols, node) {
 398                 struct symbol *sym = sym_entry__symbol(syme);
 399
 400                 if (!strcmp(buf, sym->name)) {
 401                         found = syme;
 402                         break;
 403                 }
 404         }
 405
 406         if (!found) {
 407                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 408                 sleep(1);
 409                 return;
 410         } else
 411                 parse_source(found);
 412
 413 out_free:
 414         free(buf);
 415 }
 416
 417 static void print_mapped_keys(void)
 418 {
 419         char *name = NULL;
 420
 421         if (top.sym_filter_entry) {
 422                 struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
 423                 name = sym->name;
 424         }
 425
 426         fprintf(stdout, "\nMapped keys:\n");
 427         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
 428         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
 429
 430         if (top.evlist->nr_entries > 1)
 431                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
 432
 433         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
 434
 435         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 436         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 437         fprintf(stdout, "\t[S]     stop annotation.\n");
 438
 439         if (top.evlist->nr_entries > 1)
 440                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", top.display_weighted ? 1 : 0);
 441
 442         fprintf(stdout,
 443                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 444                 top.hide_kernel_symbols ? "yes" : "no");
 445         fprintf(stdout,
 446                 "\t[U]     hide user symbols.               \t(%s)\n",
 447                 top.hide_user_symbols ? "yes" : "no");
 448         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
 449         fprintf(stdout, "\t[qQ]    quit.\n");
 450 }
 451
 452 static int key_mapped(int c)
 453 {
 454         switch (c) {
 455                 case 'd':
 456                 case 'e':
 457                 case 'f':
 458                 case 'z':
 459                 case 'q':
 460                 case 'Q':
 461                 case 'K':
 462                 case 'U':
 463                 case 'F':
 464                 case 's':
 465                 case 'S':
 466                         return 1;
 467                 case 'E':
 468                 case 'w':
 469                         return top.evlist->nr_entries > 1 ? 1 : 0;
 470                 default:
 471                         break;
 472         }
 473
 474         return 0;
 475 }
 476
 477 static void handle_keypress(struct perf_session *session, int c)
 478 {
 479         if (!key_mapped(c)) {
 480                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 481                 struct termios tc, save;
 482
 483                 print_mapped_keys();
 484                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 485                 fflush(stdout);
 486
 487                 tcgetattr(0, &save);
 488                 tc = save;
 489                 tc.c_lflag &= ~(ICANON | ECHO);
 490                 tc.c_cc[VMIN] = 0;
 491                 tc.c_cc[VTIME] = 0;
 492                 tcsetattr(0, TCSANOW, &tc);
 493
 494                 poll(&stdin_poll, 1, -1);
 495                 c = getc(stdin);
 496
 497                 tcsetattr(0, TCSAFLUSH, &save);
 498                 if (!key_mapped(c))
 499                         return;
 500         }
 501
 502         switch (c) {
 503                 case 'd':
 504                         prompt_integer(&top.delay_secs, "Enter display delay");
 505                         if (top.delay_secs < 1)
 506                                 top.delay_secs = 1;
 507                         break;
 508                 case 'e':
 509                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
 510                         if (top.print_entries == 0) {
 511                                 sig_winch_handler(SIGWINCH);
 512                                 signal(SIGWINCH, sig_winch_handler);
 513                         } else
 514                                 signal(SIGWINCH, SIG_DFL);
 515                         break;
 516                 case 'E':
 517                         if (top.evlist->nr_entries > 1) {
 518                                 fprintf(stderr, "\nAvailable events:");
 519
 520                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 521                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 522
 523                                 prompt_integer(&top.sym_counter, "Enter details event counter");
 524
 525                                 if (top.sym_counter >= top.evlist->nr_entries) {
 526                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 527                                         top.sym_counter = 0;
 528                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 529                                         sleep(1);
 530                                         break;
 531                                 }
 532                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 533                                         if (top.sym_evsel->idx == top.sym_counter)
 534                                                 break;
 535                         } else top.sym_counter = 0;
 536                         break;
 537                 case 'f':
 538                         prompt_integer(&top.count_filter, "Enter display event count filter");
 539                         break;
 540                 case 'F':
 541                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 542                         break;
 543                 case 'K':
 544                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 545                         break;
 546                 case 'q':
 547                 case 'Q':
 548                         printf("exiting.\n");
 549                         if (dump_symtab)
 550                                 perf_session__fprintf_dsos(session, stderr);
 551                         exit(0);
 552                 case 's':
 553                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 554                         break;
 555                 case 'S':
 556                         if (!top.sym_filter_entry)
 557                                 break;
 558                         else {
 559                                 struct sym_entry *syme = top.sym_filter_entry;
 560
 561                                 top.sym_filter_entry = NULL;
 562                                 __zero_source_counters(syme);
 563                         }
 564                         break;
 565                 case 'U':
 566                         top.hide_user_symbols = !top.hide_user_symbols;
 567                         break;
 568                 case 'w':
 569                         top.display_weighted = ~top.display_weighted;
 570                         break;
 571                 case 'z':
 572                         top.zero = !top.zero;
 573                         break;
 574                 default:
 575                         break;
 576         }
 577 }
 578
 579 static void *display_thread_tui(void *arg __used)
 580 {
 581         int err = 0;
 582         pthread_mutex_lock(&top.active_symbols_lock);
 583         while (list_empty(&top.active_symbols)) {
 584                 err = pthread_cond_wait(&top.active_symbols_cond,
 585                                         &top.active_symbols_lock);
 586                 if (err)
 587                         break;
 588         }
 589         pthread_mutex_unlock(&top.active_symbols_lock);
 590         if (!err)
 591                 perf_top__tui_browser(&top);
 592         exit_browser(0);
 593         exit(0);
 594         return NULL;
 595 }
 596
 597 static void *display_thread(void *arg __used)
 598 {
 599         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 600         struct termios tc, save;
 601         int delay_msecs, c;
 602         struct perf_session *session = (struct perf_session *) arg;
 603
 604         tcgetattr(0, &save);
 605         tc = save;
 606         tc.c_lflag &= ~(ICANON | ECHO);
 607         tc.c_cc[VMIN] = 0;
 608         tc.c_cc[VTIME] = 0;
 609
 610 repeat:
 611         delay_msecs = top.delay_secs * 1000;
 612         tcsetattr(0, TCSANOW, &tc);
 613         /* trash return*/
 614         getc(stdin);
 615
 616         do {
 617                 print_sym_table(session);
 618         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 619
 620         c = getc(stdin);
 621         tcsetattr(0, TCSAFLUSH, &save);
 622
 623         handle_keypress(session, c);
 624         goto repeat;
 625
 626         return NULL;
 627 }
 628
 629 /* Tag samples to be skipped. */
 630 static const char *skip_symbols[] = {
 631         "default_idle",
 632         "native_safe_halt",
 633         "cpu_idle",
 634         "enter_idle",
 635         "exit_idle",
 636         "mwait_idle",
 637         "mwait_idle_with_hints",
 638         "poll_idle",
 639         "ppc64_runlatch_off",
 640         "pseries_dedicated_idle_sleep",
 641         NULL
 642 };
 643
 644 static int symbol_filter(struct map *map, struct symbol *sym)
 645 {
 646         struct sym_entry *syme;
 647         const char *name = sym->name;
 648         int i;
 649
 650         /*
 651          * ppc64 uses function descriptors and appends a '.' to the
 652          * start of every instruction address. Remove it.
 653          */
 654         if (name[0] == '.')
 655                 name++;
 656
 657         if (!strcmp(name, "_text") ||
 658             !strcmp(name, "_etext") ||
 659             !strcmp(name, "_sinittext") ||
 660             !strncmp("init_module", name, 11) ||
 661             !strncmp("cleanup_module", name, 14) ||
 662             strstr(name, "_text_start") ||
 663             strstr(name, "_text_end"))
 664                 return 1;
 665
 666         syme = symbol__priv(sym);
 667         syme->map = map;
 668         symbol__annotate_init(map, sym);
 669
 670         if (!top.sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
 671                 /* schedule initial sym_filter_entry setup */
 672                 sym_filter_entry_sched = syme;
 673                 sym_filter = NULL;
 674         }
 675
 676         for (i = 0; skip_symbols[i]; i++) {
 677                 if (!strcmp(skip_symbols[i], name)) {
 678                         syme->skip = 1;
 679                         break;
 680                 }
 681         }
 682
 683         return 0;
 684 }
 685
 686 static void perf_event__process_sample(const union perf_event *event,
 687                                        struct perf_sample *sample,
 688                                        struct perf_session *session)
 689 {
 690         u64 ip = event->ip.ip;
 691         struct sym_entry *syme;
 692         struct addr_location al;
 693         struct machine *machine;
 694         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 695
 696         ++top.samples;
 697
 698         switch (origin) {
 699         case PERF_RECORD_MISC_USER:
 700                 ++top.us_samples;
 701                 if (top.hide_user_symbols)
 702                         return;
 703                 machine = perf_session__find_host_machine(session);
 704                 break;
 705         case PERF_RECORD_MISC_KERNEL:
 706                 ++top.kernel_samples;
 707                 if (top.hide_kernel_symbols)
 708                         return;
 709                 machine = perf_session__find_host_machine(session);
 710                 break;
 711         case PERF_RECORD_MISC_GUEST_KERNEL:
 712                 ++top.guest_kernel_samples;
 713                 machine = perf_session__find_machine(session, event->ip.pid);
 714                 break;
 715         case PERF_RECORD_MISC_GUEST_USER:
 716                 ++top.guest_us_samples;
 717                 /*
 718                  * TODO: we don't process guest user from host side
 719                  * except simple counting.
 720                  */
 721                 return;
 722         default:
 723                 return;
 724         }
 725
 726         if (!machine && perf_guest) {
 727                 pr_err("Can't find guest [%d]'s kernel information\n",
 728                         event->ip.pid);
 729                 return;
 730         }
 731
 732         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 733                 top.exact_samples++;
 734
 735         if (perf_event__preprocess_sample(event, session, &al, sample,
 736                                           symbol_filter) < 0 ||
 737             al.filtered)
 738                 return;
 739
 740         if (al.sym == NULL) {
 741                 /*
 742                  * As we do lazy loading of symtabs we only will know if the
 743                  * specified vmlinux file is invalid when we actually have a
 744                  * hit in kernel space and then try to load it. So if we get
 745                  * here and there are _no_ symbols in the DSO backing the
 746                  * kernel map, bail out.
 747                  *
 748                  * We may never get here, for instance, if we use -K/
 749                  * --hide-kernel-symbols, even if the user specifies an
 750                  * invalid --vmlinux ;-)
 751                  */
 752                 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 753                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 754                         ui__warning("The %s file can't be used\n",
 755                                     symbol_conf.vmlinux_name);
 756                         exit_browser(0);
 757                         exit(1);
 758                 }
 759
 760                 return;
 761         }
 762
 763         /* let's see, whether we need to install initial sym_filter_entry */
 764         if (sym_filter_entry_sched) {
 765                 top.sym_filter_entry = sym_filter_entry_sched;
 766                 sym_filter_entry_sched = NULL;
 767                 if (parse_source(top.sym_filter_entry) < 0) {
 768                         struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
 769
 770                         pr_err("Can't annotate %s", sym->name);
 771                         if (top.sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
 772                                 pr_err(": No vmlinux file was found in the path:\n");
 773                                 machine__fprintf_vmlinux_path(machine, stderr);
 774                         } else
 775                                 pr_err(".\n");
 776                         exit(1);
 777                 }
 778         }
 779
 780         syme = symbol__priv(al.sym);
 781         if (!syme->skip) {
 782                 struct perf_evsel *evsel;
 783
 784                 syme->origin = origin;
 785                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
 786                 assert(evsel != NULL);
 787                 syme->count[evsel->idx]++;
 788                 record_precise_ip(syme, evsel->idx, ip);
 789                 pthread_mutex_lock(&top.active_symbols_lock);
 790                 if (list_empty(&syme->node) || !syme->node.next) {
 791                         static bool first = true;
 792                         __list_insert_active_sym(syme);
 793                         if (first) {
 794                                 pthread_cond_broadcast(&top.active_symbols_cond);
 795                                 first = false;
 796                         }
 797                 }
 798                 pthread_mutex_unlock(&top.active_symbols_lock);
 799         }
 800 }
 801
 802 static void perf_session__mmap_read_cpu(struct perf_session *self, int cpu)
 803 {
 804         struct perf_sample sample;
 805         union perf_event *event;
 806
 807         while ((event = perf_evlist__read_on_cpu(top.evlist, cpu)) != NULL) {
 808                 perf_session__parse_sample(self, event, &sample);
 809
 810                 if (event->header.type == PERF_RECORD_SAMPLE)
 811                         perf_event__process_sample(event, &sample, self);
 812                 else
 813                         perf_event__process(event, &sample, self);
 814         }
 815 }
 816
 817 static void perf_session__mmap_read(struct perf_session *self)
 818 {
 819         int i;
 820
 821         for (i = 0; i < top.evlist->cpus->nr; i++)
 822                 perf_session__mmap_read_cpu(self, i);
 823 }
 824
 825 static void start_counters(struct perf_evlist *evlist)
 826 {
 827         struct perf_evsel *counter;
 828
 829         list_for_each_entry(counter, &evlist->entries, node) {
 830                 struct perf_event_attr *attr = &counter->attr;
 831
 832                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 833
 834                 if (top.freq) {
 835                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 836                         attr->freq        = 1;
 837                         attr->sample_freq = top.freq;
 838                 }
 839
 840                 if (evlist->nr_entries > 1) {
 841                         attr->sample_type |= PERF_SAMPLE_ID;
 842                         attr->read_format |= PERF_FORMAT_ID;
 843                 }
 844
 845                 attr->mmap = 1;
 846 try_again:
 847                 if (perf_evsel__open(counter, top.evlist->cpus,
 848                                      top.evlist->threads, group, inherit) < 0) {
 849                         int err = errno;
 850
 851                         if (err == EPERM || err == EACCES)
 852                                 die("Permission error - are you root?\n"
 853                                         "\t Consider tweaking"
 854                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 855                         /*
 856                          * If it's cycles then fall back to hrtimer
 857                          * based cpu-clock-tick sw counter, which
 858                          * is always available even if no PMU support:
 859                          */
 860                         if (attr->type == PERF_TYPE_HARDWARE &&
 861                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 862
 863                                 if (verbose)
 864                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 865
 866                                 attr->type = PERF_TYPE_SOFTWARE;
 867                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 868                                 goto try_again;
 869                         }
 870                         printf("\n");
 871                         error("sys_perf_event_open() syscall returned with %d "
 872                               "(%s).  /bin/dmesg may provide additional information.\n",
 873                               err, strerror(err));
 874                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 875                         exit(-1);
 876                 }
 877         }
 878
 879         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
 880                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
 881 }
 882
 883 static int __cmd_top(void)
 884 {
 885         pthread_t thread;
 886         int ret __used;
 887         /*
 888          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 889          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 890          */
 891         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 892         if (session == NULL)
 893                 return -ENOMEM;
 894
 895         if (top.target_tid != -1)
 896                 perf_event__synthesize_thread_map(top.evlist->threads,
 897                                                   perf_event__process, session);
 898         else
 899                 perf_event__synthesize_threads(perf_event__process, session);
 900
 901         start_counters(top.evlist);
 902         session->evlist = top.evlist;
 903         perf_session__update_sample_type(session);
 904
 905         /* Wait for a minimal set of events before starting the snapshot */
 906         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 907
 908         perf_session__mmap_read(session);
 909
 910         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 911                                                              display_thread), session)) {
 912                 printf("Could not create display thread.\n");
 913                 exit(-1);
 914         }
 915
 916         if (realtime_prio) {
 917                 struct sched_param param;
 918
 919                 param.sched_priority = realtime_prio;
 920                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 921                         printf("Could not set realtime priority.\n");
 922                         exit(-1);
 923                 }
 924         }
 925
 926         while (1) {
 927                 u64 hits = top.samples;
 928
 929                 perf_session__mmap_read(session);
 930
 931                 if (hits == top.samples)
 932                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 933         }
 934
 935         return 0;
 936 }
 937
 938 static const char * const top_usage[] = {
 939         "perf top [<options>]",
 940         NULL
 941 };
 942
 943 static const struct option options[] = {
 944         OPT_CALLBACK('e', "event", &top.evlist, "event",
 945                      "event selector. use 'perf list' to list available events",
 946                      parse_events),
 947         OPT_INTEGER('c', "count", &default_interval,
 948                     "event period to sample"),
 949         OPT_INTEGER('p', "pid", &top.target_pid,
 950                     "profile events on existing process id"),
 951         OPT_INTEGER('t', "tid", &top.target_tid,
 952                     "profile events on existing thread id"),
 953         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 954                             "system-wide collection from all CPUs"),
 955         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
 956                     "list of cpus to monitor"),
 957         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 958                    "file", "vmlinux pathname"),
 959         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
 960                     "hide kernel symbols"),
 961         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 962         OPT_INTEGER('r', "realtime", &realtime_prio,
 963                     "collect data with this RT SCHED_FIFO priority"),
 964         OPT_INTEGER('d', "delay", &top.delay_secs,
 965                     "number of seconds to delay between refreshes"),
 966         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
 967                             "dump the symbol table used for profiling"),
 968         OPT_INTEGER('f', "count-filter", &top.count_filter,
 969                     "only display functions with more events than this"),
 970         OPT_BOOLEAN('g', "group", &group,
 971                             "put the counters into a counter group"),
 972         OPT_BOOLEAN('i', "inherit", &inherit,
 973                     "child tasks inherit counters"),
 974         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
 975                     "symbol to annotate"),
 976         OPT_BOOLEAN('z', "zero", &top.zero,
 977                     "zero history across updates"),
 978         OPT_INTEGER('F', "freq", &top.freq,
 979                     "profile at this frequency"),
 980         OPT_INTEGER('E', "entries", &top.print_entries,
 981                     "display this many functions"),
 982         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
 983                     "hide user symbols"),
 984         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 985         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 986         OPT_INCR('v', "verbose", &verbose,
 987                     "be more verbose (show counter open errors, etc)"),
 988         OPT_END()
 989 };
 990
 991 int cmd_top(int argc, const char **argv, const char *prefix __used)
 992 {
 993         struct perf_evsel *pos;
 994         int status = -ENOMEM;
 995
 996         top.evlist = perf_evlist__new(NULL, NULL);
 997         if (top.evlist == NULL)
 998                 return -ENOMEM;
 999
1000         page_size = sysconf(_SC_PAGE_SIZE);
1001
1002         argc = parse_options(argc, argv, options, top_usage, 0);
1003         if (argc)
1004                 usage_with_options(top_usage, options);
1005
1006         /*
1007          * XXX For now start disabled, only using TUI if explicitely asked for.
1008          * Change that when handle_keys equivalent gets written, live annotation
1009          * done, etc.
1010          */
1011         use_browser = 0;
1012
1013         if (use_stdio)
1014                 use_browser = 0;
1015         else if (use_tui)
1016                 use_browser = 1;
1017
1018         setup_browser(false);
1019
1020         /* CPU and PID are mutually exclusive */
1021         if (top.target_tid > 0 && top.cpu_list) {
1022                 printf("WARNING: PID switch overriding CPU\n");
1023                 sleep(1);
1024                 top.cpu_list = NULL;
1025         }
1026
1027         if (top.target_pid != -1)
1028                 top.target_tid = top.target_pid;
1029
1030         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1031                                      top.target_tid, top.cpu_list) < 0)
1032                 usage_with_options(top_usage, options);
1033
1034         if (!top.evlist->nr_entries &&
1035             perf_evlist__add_default(top.evlist) < 0) {
1036                 pr_err("Not enough memory for event selector list\n");
1037                 return -ENOMEM;
1038         }
1039
1040         if (top.delay_secs < 1)
1041                 top.delay_secs = 1;
1042
1043         /*
1044          * User specified count overrides default frequency.
1045          */
1046         if (default_interval)
1047                 top.freq = 0;
1048         else if (top.freq) {
1049                 default_interval = top.freq;
1050         } else {
1051                 fprintf(stderr, "frequency and count are zero, aborting\n");
1052                 exit(EXIT_FAILURE);
1053         }
1054
1055         list_for_each_entry(pos, &top.evlist->entries, node) {
1056                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1057                                          top.evlist->threads->nr) < 0)
1058                         goto out_free_fd;
1059                 /*
1060                  * Fill in the ones not specifically initialized via -c:
1061                  */
1062                 if (pos->attr.sample_period)
1063                         continue;
1064
1065                 pos->attr.sample_period = default_interval;
1066         }
1067
1068         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1069             perf_evlist__alloc_mmap(top.evlist) < 0)
1070                 goto out_free_fd;
1071
1072         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1073
1074         symbol_conf.priv_size = (sizeof(struct sym_entry) + sizeof(struct annotation) +
1075                                  (top.evlist->nr_entries + 1) * sizeof(unsigned long));
1076
1077         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1078         if (symbol__init() < 0)
1079                 return -1;
1080
1081         get_term_dimensions(&winsize);
1082         if (top.print_entries == 0) {
1083                 update_print_entries(&winsize);
1084                 signal(SIGWINCH, sig_winch_handler);
1085         }
1086
1087         status = __cmd_top();
1088 out_free_fd:
1089         perf_evlist__delete(top.evlist);
1090
1091         return status;
1092 }