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 static struct perf_top top = {
  66         .count_filter           = 5,
  67         .delay_secs             = 2,
  68         .display_weighted       = -1,
  69         .target_pid             = -1,
  70         .target_tid             = -1,
  71         .active_symbols         = LIST_HEAD_INIT(top.active_symbols),
  72         .active_symbols_lock    = PTHREAD_MUTEX_INITIALIZER,
  73         .active_symbols_cond    = PTHREAD_COND_INITIALIZER,
  74         .freq                   = 1000, /* 1 KHz */
  75 };
  76
  77 static bool                     system_wide                     =  false;
  78
  79 static bool                     use_tui, use_stdio;
  80
  81 static int                      default_interval                =      0;
  82
  83 static bool                     kptr_restrict_warned;
  84 static bool                     vmlinux_warned;
  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->symtab_type == SYMTAB__KALLSYMS) {
 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                                 /* Select 0 as the default event: */
 519                                 int counter = 0;
 520
 521                                 fprintf(stderr, "\nAvailable events:");
 522
 523                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 524                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 525
 526                                 prompt_integer(&counter, "Enter details event counter");
 527
 528                                 if (counter >= top.evlist->nr_entries) {
 529                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 530                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 531                                         sleep(1);
 532                                         break;
 533                                 }
 534                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 535                                         if (top.sym_evsel->idx == counter)
 536                                                 break;
 537                         } else
 538                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 539                         break;
 540                 case 'f':
 541                         prompt_integer(&top.count_filter, "Enter display event count filter");
 542                         break;
 543                 case 'F':
 544                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 545                         break;
 546                 case 'K':
 547                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 548                         break;
 549                 case 'q':
 550                 case 'Q':
 551                         printf("exiting.\n");
 552                         if (dump_symtab)
 553                                 perf_session__fprintf_dsos(session, stderr);
 554                         exit(0);
 555                 case 's':
 556                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 557                         break;
 558                 case 'S':
 559                         if (!top.sym_filter_entry)
 560                                 break;
 561                         else {
 562                                 struct sym_entry *syme = top.sym_filter_entry;
 563
 564                                 top.sym_filter_entry = NULL;
 565                                 __zero_source_counters(syme);
 566                         }
 567                         break;
 568                 case 'U':
 569                         top.hide_user_symbols = !top.hide_user_symbols;
 570                         break;
 571                 case 'w':
 572                         top.display_weighted = ~top.display_weighted;
 573                         break;
 574                 case 'z':
 575                         top.zero = !top.zero;
 576                         break;
 577                 default:
 578                         break;
 579         }
 580 }
 581
 582 static void *display_thread_tui(void *arg __used)
 583 {
 584         int err = 0;
 585         pthread_mutex_lock(&top.active_symbols_lock);
 586         while (list_empty(&top.active_symbols)) {
 587                 err = pthread_cond_wait(&top.active_symbols_cond,
 588                                         &top.active_symbols_lock);
 589                 if (err)
 590                         break;
 591         }
 592         pthread_mutex_unlock(&top.active_symbols_lock);
 593         if (!err)
 594                 perf_top__tui_browser(&top);
 595         exit_browser(0);
 596         exit(0);
 597         return NULL;
 598 }
 599
 600 static void *display_thread(void *arg __used)
 601 {
 602         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 603         struct termios tc, save;
 604         int delay_msecs, c;
 605         struct perf_session *session = (struct perf_session *) arg;
 606
 607         tcgetattr(0, &save);
 608         tc = save;
 609         tc.c_lflag &= ~(ICANON | ECHO);
 610         tc.c_cc[VMIN] = 0;
 611         tc.c_cc[VTIME] = 0;
 612
 613 repeat:
 614         delay_msecs = top.delay_secs * 1000;
 615         tcsetattr(0, TCSANOW, &tc);
 616         /* trash return*/
 617         getc(stdin);
 618
 619         do {
 620                 print_sym_table(session);
 621         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 622
 623         c = getc(stdin);
 624         tcsetattr(0, TCSAFLUSH, &save);
 625
 626         handle_keypress(session, c);
 627         goto repeat;
 628
 629         return NULL;
 630 }
 631
 632 /* Tag samples to be skipped. */
 633 static const char *skip_symbols[] = {
 634         "default_idle",
 635         "native_safe_halt",
 636         "cpu_idle",
 637         "enter_idle",
 638         "exit_idle",
 639         "mwait_idle",
 640         "mwait_idle_with_hints",
 641         "poll_idle",
 642         "ppc64_runlatch_off",
 643         "pseries_dedicated_idle_sleep",
 644         NULL
 645 };
 646
 647 static int symbol_filter(struct map *map, struct symbol *sym)
 648 {
 649         struct sym_entry *syme;
 650         const char *name = sym->name;
 651         int i;
 652
 653         /*
 654          * ppc64 uses function descriptors and appends a '.' to the
 655          * start of every instruction address. Remove it.
 656          */
 657         if (name[0] == '.')
 658                 name++;
 659
 660         if (!strcmp(name, "_text") ||
 661             !strcmp(name, "_etext") ||
 662             !strcmp(name, "_sinittext") ||
 663             !strncmp("init_module", name, 11) ||
 664             !strncmp("cleanup_module", name, 14) ||
 665             strstr(name, "_text_start") ||
 666             strstr(name, "_text_end"))
 667                 return 1;
 668
 669         syme = symbol__priv(sym);
 670         syme->map = map;
 671         symbol__annotate_init(map, sym);
 672
 673         if (!top.sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
 674                 /* schedule initial sym_filter_entry setup */
 675                 sym_filter_entry_sched = syme;
 676                 sym_filter = NULL;
 677         }
 678
 679         for (i = 0; skip_symbols[i]; i++) {
 680                 if (!strcmp(skip_symbols[i], name)) {
 681                         sym->ignore = true;
 682                         break;
 683                 }
 684         }
 685
 686         return 0;
 687 }
 688
 689 static void perf_event__process_sample(const union perf_event *event,
 690                                        struct perf_sample *sample,
 691                                        struct perf_session *session)
 692 {
 693         u64 ip = event->ip.ip;
 694         struct sym_entry *syme;
 695         struct addr_location al;
 696         struct machine *machine;
 697         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 698
 699         ++top.samples;
 700
 701         switch (origin) {
 702         case PERF_RECORD_MISC_USER:
 703                 ++top.us_samples;
 704                 if (top.hide_user_symbols)
 705                         return;
 706                 machine = perf_session__find_host_machine(session);
 707                 break;
 708         case PERF_RECORD_MISC_KERNEL:
 709                 ++top.kernel_samples;
 710                 if (top.hide_kernel_symbols)
 711                         return;
 712                 machine = perf_session__find_host_machine(session);
 713                 break;
 714         case PERF_RECORD_MISC_GUEST_KERNEL:
 715                 ++top.guest_kernel_samples;
 716                 machine = perf_session__find_machine(session, event->ip.pid);
 717                 break;
 718         case PERF_RECORD_MISC_GUEST_USER:
 719                 ++top.guest_us_samples;
 720                 /*
 721                  * TODO: we don't process guest user from host side
 722                  * except simple counting.
 723                  */
 724                 return;
 725         default:
 726                 return;
 727         }
 728
 729         if (!machine && perf_guest) {
 730                 pr_err("Can't find guest [%d]'s kernel information\n",
 731                         event->ip.pid);
 732                 return;
 733         }
 734
 735         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 736                 top.exact_samples++;
 737
 738         if (perf_event__preprocess_sample(event, session, &al, sample,
 739                                           symbol_filter) < 0 ||
 740             al.filtered)
 741                 return;
 742
 743         if (!kptr_restrict_warned &&
 744             symbol_conf.kptr_restrict &&
 745             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 746                 ui__warning(
 747 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 748 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 749 "Kernel%s samples will not be resolved.\n",
 750                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 751                           " modules" : "");
 752                 if (use_browser <= 0)
 753                         sleep(5);
 754                 kptr_restrict_warned = true;
 755         }
 756
 757         if (al.sym == NULL) {
 758                 const char *msg = "Kernel samples will not be resolved.\n";
 759                 /*
 760                  * As we do lazy loading of symtabs we only will know if the
 761                  * specified vmlinux file is invalid when we actually have a
 762                  * hit in kernel space and then try to load it. So if we get
 763                  * here and there are _no_ symbols in the DSO backing the
 764                  * kernel map, bail out.
 765                  *
 766                  * We may never get here, for instance, if we use -K/
 767                  * --hide-kernel-symbols, even if the user specifies an
 768                  * invalid --vmlinux ;-)
 769                  */
 770                 if (!kptr_restrict_warned && !vmlinux_warned &&
 771                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 772                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 773                         if (symbol_conf.vmlinux_name) {
 774                                 ui__warning("The %s file can't be used.\n%s",
 775                                             symbol_conf.vmlinux_name, msg);
 776                         } else {
 777                                 ui__warning("A vmlinux file was not found.\n%s",
 778                                             msg);
 779                         }
 780
 781                         if (use_browser <= 0)
 782                                 sleep(5);
 783                         vmlinux_warned = true;
 784                 }
 785
 786                 return;
 787         }
 788
 789         /* let's see, whether we need to install initial sym_filter_entry */
 790         if (sym_filter_entry_sched) {
 791                 top.sym_filter_entry = sym_filter_entry_sched;
 792                 sym_filter_entry_sched = NULL;
 793                 if (parse_source(top.sym_filter_entry) < 0) {
 794                         struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
 795
 796                         pr_err("Can't annotate %s", sym->name);
 797                         if (top.sym_filter_entry->map->dso->symtab_type == SYMTAB__KALLSYMS) {
 798                                 pr_err(": No vmlinux file was found in the path:\n");
 799                                 machine__fprintf_vmlinux_path(machine, stderr);
 800                         } else
 801                                 pr_err(".\n");
 802                         exit(1);
 803                 }
 804         }
 805
 806         syme = symbol__priv(al.sym);
 807         if (!al.sym->ignore) {
 808                 struct perf_evsel *evsel;
 809
 810                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
 811                 assert(evsel != NULL);
 812                 syme->count[evsel->idx]++;
 813                 record_precise_ip(syme, evsel->idx, ip);
 814                 pthread_mutex_lock(&top.active_symbols_lock);
 815                 if (list_empty(&syme->node) || !syme->node.next) {
 816                         static bool first = true;
 817                         __list_insert_active_sym(syme);
 818                         if (first) {
 819                                 pthread_cond_broadcast(&top.active_symbols_cond);
 820                                 first = false;
 821                         }
 822                 }
 823                 pthread_mutex_unlock(&top.active_symbols_lock);
 824         }
 825 }
 826
 827 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
 828 {
 829         struct perf_sample sample;
 830         union perf_event *event;
 831         int ret;
 832
 833         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
 834                 ret = perf_session__parse_sample(self, event, &sample);
 835                 if (ret) {
 836                         pr_err("Can't parse sample, err = %d\n", ret);
 837                         continue;
 838                 }
 839
 840                 if (event->header.type == PERF_RECORD_SAMPLE)
 841                         perf_event__process_sample(event, &sample, self);
 842                 else
 843                         perf_event__process(event, &sample, self);
 844         }
 845 }
 846
 847 static void perf_session__mmap_read(struct perf_session *self)
 848 {
 849         int i;
 850
 851         for (i = 0; i < top.evlist->nr_mmaps; i++)
 852                 perf_session__mmap_read_idx(self, i);
 853 }
 854
 855 static void start_counters(struct perf_evlist *evlist)
 856 {
 857         struct perf_evsel *counter;
 858
 859         list_for_each_entry(counter, &evlist->entries, node) {
 860                 struct perf_event_attr *attr = &counter->attr;
 861
 862                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 863
 864                 if (top.freq) {
 865                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 866                         attr->freq        = 1;
 867                         attr->sample_freq = top.freq;
 868                 }
 869
 870                 if (evlist->nr_entries > 1) {
 871                         attr->sample_type |= PERF_SAMPLE_ID;
 872                         attr->read_format |= PERF_FORMAT_ID;
 873                 }
 874
 875                 attr->mmap = 1;
 876                 attr->inherit = inherit;
 877 try_again:
 878                 if (perf_evsel__open(counter, top.evlist->cpus,
 879                                      top.evlist->threads, group) < 0) {
 880                         int err = errno;
 881
 882                         if (err == EPERM || err == EACCES) {
 883                                 ui__warning_paranoid();
 884                                 goto out_err;
 885                         }
 886                         /*
 887                          * If it's cycles then fall back to hrtimer
 888                          * based cpu-clock-tick sw counter, which
 889                          * is always available even if no PMU support:
 890                          */
 891                         if (attr->type == PERF_TYPE_HARDWARE &&
 892                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 893                                 if (verbose)
 894                                         ui__warning("Cycles event not supported,\n"
 895                                                     "trying to fall back to cpu-clock-ticks\n");
 896
 897                                 attr->type = PERF_TYPE_SOFTWARE;
 898                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 899                                 goto try_again;
 900                         }
 901
 902                         if (err == ENOENT) {
 903                                 ui__warning("The %s event is not supported.\n",
 904                                             event_name(counter));
 905                                 goto out_err;
 906                         }
 907
 908                         ui__warning("The sys_perf_event_open() syscall "
 909                                     "returned with %d (%s).  /bin/dmesg "
 910                                     "may provide additional information.\n"
 911                                     "No CONFIG_PERF_EVENTS=y kernel support "
 912                                     "configured?\n", err, strerror(err));
 913                         goto out_err;
 914                 }
 915         }
 916
 917         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
 918                 ui__warning("Failed to mmap with %d (%s)\n",
 919                             errno, strerror(errno));
 920                 goto out_err;
 921         }
 922
 923         return;
 924
 925 out_err:
 926         exit_browser(0);
 927         exit(0);
 928 }
 929
 930 static int __cmd_top(void)
 931 {
 932         pthread_t thread;
 933         int ret __used;
 934         /*
 935          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 936          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 937          */
 938         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 939         if (session == NULL)
 940                 return -ENOMEM;
 941
 942         if (top.target_tid != -1)
 943                 perf_event__synthesize_thread_map(top.evlist->threads,
 944                                                   perf_event__process, session);
 945         else
 946                 perf_event__synthesize_threads(perf_event__process, session);
 947
 948         start_counters(top.evlist);
 949         session->evlist = top.evlist;
 950         perf_session__update_sample_type(session);
 951
 952         /* Wait for a minimal set of events before starting the snapshot */
 953         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 954
 955         perf_session__mmap_read(session);
 956
 957         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 958                                                              display_thread), session)) {
 959                 printf("Could not create display thread.\n");
 960                 exit(-1);
 961         }
 962
 963         if (realtime_prio) {
 964                 struct sched_param param;
 965
 966                 param.sched_priority = realtime_prio;
 967                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 968                         printf("Could not set realtime priority.\n");
 969                         exit(-1);
 970                 }
 971         }
 972
 973         while (1) {
 974                 u64 hits = top.samples;
 975
 976                 perf_session__mmap_read(session);
 977
 978                 if (hits == top.samples)
 979                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 980         }
 981
 982         return 0;
 983 }
 984
 985 static const char * const top_usage[] = {
 986         "perf top [<options>]",
 987         NULL
 988 };
 989
 990 static const struct option options[] = {
 991         OPT_CALLBACK('e', "event", &top.evlist, "event",
 992                      "event selector. use 'perf list' to list available events",
 993                      parse_events_option),
 994         OPT_INTEGER('c', "count", &default_interval,
 995                     "event period to sample"),
 996         OPT_INTEGER('p', "pid", &top.target_pid,
 997                     "profile events on existing process id"),
 998         OPT_INTEGER('t', "tid", &top.target_tid,
 999                     "profile events on existing thread id"),
1000         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1001                             "system-wide collection from all CPUs"),
1002         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1003                     "list of cpus to monitor"),
1004         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1005                    "file", "vmlinux pathname"),
1006         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1007                     "hide kernel symbols"),
1008         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1009         OPT_INTEGER('r', "realtime", &realtime_prio,
1010                     "collect data with this RT SCHED_FIFO priority"),
1011         OPT_INTEGER('d', "delay", &top.delay_secs,
1012                     "number of seconds to delay between refreshes"),
1013         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1014                             "dump the symbol table used for profiling"),
1015         OPT_INTEGER('f', "count-filter", &top.count_filter,
1016                     "only display functions with more events than this"),
1017         OPT_BOOLEAN('g', "group", &group,
1018                             "put the counters into a counter group"),
1019         OPT_BOOLEAN('i', "inherit", &inherit,
1020                     "child tasks inherit counters"),
1021         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1022                     "symbol to annotate"),
1023         OPT_BOOLEAN('z', "zero", &top.zero,
1024                     "zero history across updates"),
1025         OPT_INTEGER('F', "freq", &top.freq,
1026                     "profile at this frequency"),
1027         OPT_INTEGER('E', "entries", &top.print_entries,
1028                     "display this many functions"),
1029         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1030                     "hide user symbols"),
1031         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1032         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1033         OPT_INCR('v', "verbose", &verbose,
1034                     "be more verbose (show counter open errors, etc)"),
1035         OPT_END()
1036 };
1037
1038 int cmd_top(int argc, const char **argv, const char *prefix __used)
1039 {
1040         struct perf_evsel *pos;
1041         int status = -ENOMEM;
1042
1043         top.evlist = perf_evlist__new(NULL, NULL);
1044         if (top.evlist == NULL)
1045                 return -ENOMEM;
1046
1047         page_size = sysconf(_SC_PAGE_SIZE);
1048
1049         argc = parse_options(argc, argv, options, top_usage, 0);
1050         if (argc)
1051                 usage_with_options(top_usage, options);
1052
1053         /*
1054          * XXX For now start disabled, only using TUI if explicitely asked for.
1055          * Change that when handle_keys equivalent gets written, live annotation
1056          * done, etc.
1057          */
1058         use_browser = 0;
1059
1060         if (use_stdio)
1061                 use_browser = 0;
1062         else if (use_tui)
1063                 use_browser = 1;
1064
1065         setup_browser(false);
1066
1067         /* CPU and PID are mutually exclusive */
1068         if (top.target_tid > 0 && top.cpu_list) {
1069                 printf("WARNING: PID switch overriding CPU\n");
1070                 sleep(1);
1071                 top.cpu_list = NULL;
1072         }
1073
1074         if (top.target_pid != -1)
1075                 top.target_tid = top.target_pid;
1076
1077         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1078                                      top.target_tid, top.cpu_list) < 0)
1079                 usage_with_options(top_usage, options);
1080
1081         if (!top.evlist->nr_entries &&
1082             perf_evlist__add_default(top.evlist) < 0) {
1083                 pr_err("Not enough memory for event selector list\n");
1084                 return -ENOMEM;
1085         }
1086
1087         if (top.delay_secs < 1)
1088                 top.delay_secs = 1;
1089
1090         /*
1091          * User specified count overrides default frequency.
1092          */
1093         if (default_interval)
1094                 top.freq = 0;
1095         else if (top.freq) {
1096                 default_interval = top.freq;
1097         } else {
1098                 fprintf(stderr, "frequency and count are zero, aborting\n");
1099                 exit(EXIT_FAILURE);
1100         }
1101
1102         list_for_each_entry(pos, &top.evlist->entries, node) {
1103                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1104                                          top.evlist->threads->nr) < 0)
1105                         goto out_free_fd;
1106                 /*
1107                  * Fill in the ones not specifically initialized via -c:
1108                  */
1109                 if (pos->attr.sample_period)
1110                         continue;
1111
1112                 pos->attr.sample_period = default_interval;
1113         }
1114
1115         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1116             perf_evlist__alloc_mmap(top.evlist) < 0)
1117                 goto out_free_fd;
1118
1119         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1120
1121         symbol_conf.priv_size = (sizeof(struct sym_entry) + sizeof(struct annotation) +
1122                                  (top.evlist->nr_entries + 1) * sizeof(unsigned long));
1123
1124         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1125         if (symbol__init() < 0)
1126                 return -1;
1127
1128         get_term_dimensions(&winsize);
1129         if (top.print_entries == 0) {
1130                 update_print_entries(&winsize);
1131                 signal(SIGWINCH, sig_winch_handler);
1132         }
1133
1134         status = __cmd_top();
1135 out_free_fd:
1136         perf_evlist__delete(top.evlist);
1137
1138         return status;
1139 }