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