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, struct map *map,
 195                               int counter, u64 ip)
 196 {
 197         struct annotation *notes;
 198         struct symbol *sym;
 199
 200         if (syme != top.sym_filter_entry)
 201                 return;
 202
 203         sym = sym_entry__symbol(syme);
 204         notes = symbol__annotation(sym);
 205
 206         if (pthread_mutex_trylock(&notes->lock))
 207                 return;
 208
 209         ip = map->map_ip(map, ip);
 210         symbol__inc_addr_samples(sym, map, counter, ip);
 211
 212         pthread_mutex_unlock(&notes->lock);
 213 }
 214
 215 static void show_details(struct sym_entry *syme)
 216 {
 217         struct annotation *notes;
 218         struct symbol *symbol;
 219         int more;
 220
 221         if (!syme)
 222                 return;
 223
 224         symbol = sym_entry__symbol(syme);
 225         notes = symbol__annotation(symbol);
 226
 227         pthread_mutex_lock(&notes->lock);
 228
 229         if (notes->src == NULL)
 230                 goto out_unlock;
 231
 232         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
 233         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 234
 235         more = symbol__annotate_printf(symbol, syme->map, top.sym_evsel->idx,
 236                                        0, sym_pcnt_filter, top.print_entries, 4);
 237         if (top.zero)
 238                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
 239         else
 240                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
 241         if (more != 0)
 242                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 243 out_unlock:
 244         pthread_mutex_unlock(&notes->lock);
 245 }
 246
 247 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 248
 249 static void __list_insert_active_sym(struct sym_entry *syme)
 250 {
 251         list_add(&syme->node, &top.active_symbols);
 252 }
 253
 254 static void print_sym_table(struct perf_session *session)
 255 {
 256         char bf[160];
 257         int printed = 0;
 258         struct rb_node *nd;
 259         struct sym_entry *syme;
 260         struct rb_root tmp = RB_ROOT;
 261         const int win_width = winsize.ws_col - 1;
 262         int sym_width, dso_width, dso_short_width;
 263         float sum_ksamples = perf_top__decay_samples(&top, &tmp);
 264
 265         puts(CONSOLE_CLEAR);
 266
 267         perf_top__header_snprintf(&top, bf, sizeof(bf));
 268         printf("%s\n", bf);
 269
 270         perf_top__reset_sample_counters(&top);
 271
 272         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 273
 274         if (session->hists.stats.total_lost != 0) {
 275                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
 276                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
 277                        session->hists.stats.total_lost);
 278         }
 279
 280         if (top.sym_filter_entry) {
 281                 show_details(top.sym_filter_entry);
 282                 return;
 283         }
 284
 285         perf_top__find_widths(&top, &tmp, &dso_width, &dso_short_width,
 286                               &sym_width);
 287
 288         if (sym_width + dso_width > winsize.ws_col - 29) {
 289                 dso_width = dso_short_width;
 290                 if (sym_width + dso_width > winsize.ws_col - 29)
 291                         sym_width = winsize.ws_col - dso_width - 29;
 292         }
 293         putchar('\n');
 294         if (top.evlist->nr_entries == 1)
 295                 printf("             samples  pcnt");
 296         else
 297                 printf("   weight    samples  pcnt");
 298
 299         if (verbose)
 300                 printf("         RIP       ");
 301         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
 302         printf("   %s    _______ _____",
 303                top.evlist->nr_entries == 1 ? "      " : "______");
 304         if (verbose)
 305                 printf(" ________________");
 306         printf(" %-*.*s", sym_width, sym_width, graph_line);
 307         printf(" %-*.*s", dso_width, dso_width, graph_line);
 308         puts("\n");
 309
 310         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
 311                 struct symbol *sym;
 312                 double pcnt;
 313
 314                 syme = rb_entry(nd, struct sym_entry, rb_node);
 315                 sym = sym_entry__symbol(syme);
 316                 if (++printed > top.print_entries ||
 317                     (int)syme->snap_count < top.count_filter)
 318                         continue;
 319
 320                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
 321                                          sum_ksamples));
 322
 323                 if (top.evlist->nr_entries == 1 || !top.display_weighted)
 324                         printf("%20.2f ", syme->weight);
 325                 else
 326                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
 327
 328                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
 329                 if (verbose)
 330                         printf(" %016" PRIx64, sym->start);
 331                 printf(" %-*.*s", sym_width, sym_width, sym->name);
 332                 printf(" %-*.*s\n", dso_width, dso_width,
 333                        dso_width >= syme->map->dso->long_name_len ?
 334                                         syme->map->dso->long_name :
 335                                         syme->map->dso->short_name);
 336         }
 337 }
 338
 339 static void prompt_integer(int *target, const char *msg)
 340 {
 341         char *buf = malloc(0), *p;
 342         size_t dummy = 0;
 343         int tmp;
 344
 345         fprintf(stdout, "\n%s: ", msg);
 346         if (getline(&buf, &dummy, stdin) < 0)
 347                 return;
 348
 349         p = strchr(buf, '\n');
 350         if (p)
 351                 *p = 0;
 352
 353         p = buf;
 354         while(*p) {
 355                 if (!isdigit(*p))
 356                         goto out_free;
 357                 p++;
 358         }
 359         tmp = strtoul(buf, NULL, 10);
 360         *target = tmp;
 361 out_free:
 362         free(buf);
 363 }
 364
 365 static void prompt_percent(int *target, const char *msg)
 366 {
 367         int tmp = 0;
 368
 369         prompt_integer(&tmp, msg);
 370         if (tmp >= 0 && tmp <= 100)
 371                 *target = tmp;
 372 }
 373
 374 static void prompt_symbol(struct sym_entry **target, const char *msg)
 375 {
 376         char *buf = malloc(0), *p;
 377         struct sym_entry *syme = *target, *n, *found = NULL;
 378         size_t dummy = 0;
 379
 380         /* zero counters of active symbol */
 381         if (syme) {
 382                 __zero_source_counters(syme);
 383                 *target = NULL;
 384         }
 385
 386         fprintf(stdout, "\n%s: ", msg);
 387         if (getline(&buf, &dummy, stdin) < 0)
 388                 goto out_free;
 389
 390         p = strchr(buf, '\n');
 391         if (p)
 392                 *p = 0;
 393
 394         pthread_mutex_lock(&top.active_symbols_lock);
 395         syme = list_entry(top.active_symbols.next, struct sym_entry, node);
 396         pthread_mutex_unlock(&top.active_symbols_lock);
 397
 398         list_for_each_entry_safe_from(syme, n, &top.active_symbols, node) {
 399                 struct symbol *sym = sym_entry__symbol(syme);
 400
 401                 if (!strcmp(buf, sym->name)) {
 402                         found = syme;
 403                         break;
 404                 }
 405         }
 406
 407         if (!found) {
 408                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 409                 sleep(1);
 410                 return;
 411         } else
 412                 parse_source(found);
 413
 414 out_free:
 415         free(buf);
 416 }
 417
 418 static void print_mapped_keys(void)
 419 {
 420         char *name = NULL;
 421
 422         if (top.sym_filter_entry) {
 423                 struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
 424                 name = sym->name;
 425         }
 426
 427         fprintf(stdout, "\nMapped keys:\n");
 428         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
 429         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
 430
 431         if (top.evlist->nr_entries > 1)
 432                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
 433
 434         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
 435
 436         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 437         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 438         fprintf(stdout, "\t[S]     stop annotation.\n");
 439
 440         if (top.evlist->nr_entries > 1)
 441                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", top.display_weighted ? 1 : 0);
 442
 443         fprintf(stdout,
 444                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 445                 top.hide_kernel_symbols ? "yes" : "no");
 446         fprintf(stdout,
 447                 "\t[U]     hide user symbols.               \t(%s)\n",
 448                 top.hide_user_symbols ? "yes" : "no");
 449         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
 450         fprintf(stdout, "\t[qQ]    quit.\n");
 451 }
 452
 453 static int key_mapped(int c)
 454 {
 455         switch (c) {
 456                 case 'd':
 457                 case 'e':
 458                 case 'f':
 459                 case 'z':
 460                 case 'q':
 461                 case 'Q':
 462                 case 'K':
 463                 case 'U':
 464                 case 'F':
 465                 case 's':
 466                 case 'S':
 467                         return 1;
 468                 case 'E':
 469                 case 'w':
 470                         return top.evlist->nr_entries > 1 ? 1 : 0;
 471                 default:
 472                         break;
 473         }
 474
 475         return 0;
 476 }
 477
 478 static void handle_keypress(struct perf_session *session, int c)
 479 {
 480         if (!key_mapped(c)) {
 481                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 482                 struct termios tc, save;
 483
 484                 print_mapped_keys();
 485                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 486                 fflush(stdout);
 487
 488                 tcgetattr(0, &save);
 489                 tc = save;
 490                 tc.c_lflag &= ~(ICANON | ECHO);
 491                 tc.c_cc[VMIN] = 0;
 492                 tc.c_cc[VTIME] = 0;
 493                 tcsetattr(0, TCSANOW, &tc);
 494
 495                 poll(&stdin_poll, 1, -1);
 496                 c = getc(stdin);
 497
 498                 tcsetattr(0, TCSAFLUSH, &save);
 499                 if (!key_mapped(c))
 500                         return;
 501         }
 502
 503         switch (c) {
 504                 case 'd':
 505                         prompt_integer(&top.delay_secs, "Enter display delay");
 506                         if (top.delay_secs < 1)
 507                                 top.delay_secs = 1;
 508                         break;
 509                 case 'e':
 510                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
 511                         if (top.print_entries == 0) {
 512                                 sig_winch_handler(SIGWINCH);
 513                                 signal(SIGWINCH, sig_winch_handler);
 514                         } else
 515                                 signal(SIGWINCH, SIG_DFL);
 516                         break;
 517                 case 'E':
 518                         if (top.evlist->nr_entries > 1) {
 519                                 /* Select 0 as the default event: */
 520                                 int counter = 0;
 521
 522                                 fprintf(stderr, "\nAvailable events:");
 523
 524                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 525                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 526
 527                                 prompt_integer(&counter, "Enter details event counter");
 528
 529                                 if (counter >= top.evlist->nr_entries) {
 530                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 531                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 532                                         sleep(1);
 533                                         break;
 534                                 }
 535                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 536                                         if (top.sym_evsel->idx == counter)
 537                                                 break;
 538                         } else
 539                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 540                         break;
 541                 case 'f':
 542                         prompt_integer(&top.count_filter, "Enter display event count filter");
 543                         break;
 544                 case 'F':
 545                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 546                         break;
 547                 case 'K':
 548                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 549                         break;
 550                 case 'q':
 551                 case 'Q':
 552                         printf("exiting.\n");
 553                         if (dump_symtab)
 554                                 perf_session__fprintf_dsos(session, stderr);
 555                         exit(0);
 556                 case 's':
 557                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 558                         break;
 559                 case 'S':
 560                         if (!top.sym_filter_entry)
 561                                 break;
 562                         else {
 563                                 struct sym_entry *syme = top.sym_filter_entry;
 564
 565                                 top.sym_filter_entry = NULL;
 566                                 __zero_source_counters(syme);
 567                         }
 568                         break;
 569                 case 'U':
 570                         top.hide_user_symbols = !top.hide_user_symbols;
 571                         break;
 572                 case 'w':
 573                         top.display_weighted = ~top.display_weighted;
 574                         break;
 575                 case 'z':
 576                         top.zero = !top.zero;
 577                         break;
 578                 default:
 579                         break;
 580         }
 581 }
 582
 583 static void *display_thread_tui(void *arg __used)
 584 {
 585         int err = 0;
 586         pthread_mutex_lock(&top.active_symbols_lock);
 587         while (list_empty(&top.active_symbols)) {
 588                 err = pthread_cond_wait(&top.active_symbols_cond,
 589                                         &top.active_symbols_lock);
 590                 if (err)
 591                         break;
 592         }
 593         pthread_mutex_unlock(&top.active_symbols_lock);
 594         if (!err)
 595                 perf_top__tui_browser(&top);
 596         exit_browser(0);
 597         exit(0);
 598         return NULL;
 599 }
 600
 601 static void *display_thread(void *arg __used)
 602 {
 603         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 604         struct termios tc, save;
 605         int delay_msecs, c;
 606         struct perf_session *session = (struct perf_session *) arg;
 607
 608         tcgetattr(0, &save);
 609         tc = save;
 610         tc.c_lflag &= ~(ICANON | ECHO);
 611         tc.c_cc[VMIN] = 0;
 612         tc.c_cc[VTIME] = 0;
 613
 614 repeat:
 615         delay_msecs = top.delay_secs * 1000;
 616         tcsetattr(0, TCSANOW, &tc);
 617         /* trash return*/
 618         getc(stdin);
 619
 620         do {
 621                 print_sym_table(session);
 622         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
 623
 624         c = getc(stdin);
 625         tcsetattr(0, TCSAFLUSH, &save);
 626
 627         handle_keypress(session, c);
 628         goto repeat;
 629
 630         return NULL;
 631 }
 632
 633 /* Tag samples to be skipped. */
 634 static const char *skip_symbols[] = {
 635         "default_idle",
 636         "native_safe_halt",
 637         "cpu_idle",
 638         "enter_idle",
 639         "exit_idle",
 640         "mwait_idle",
 641         "mwait_idle_with_hints",
 642         "poll_idle",
 643         "ppc64_runlatch_off",
 644         "pseries_dedicated_idle_sleep",
 645         NULL
 646 };
 647
 648 static int symbol_filter(struct map *map, struct symbol *sym)
 649 {
 650         struct sym_entry *syme;
 651         const char *name = sym->name;
 652         int i;
 653
 654         /*
 655          * ppc64 uses function descriptors and appends a '.' to the
 656          * start of every instruction address. Remove it.
 657          */
 658         if (name[0] == '.')
 659                 name++;
 660
 661         if (!strcmp(name, "_text") ||
 662             !strcmp(name, "_etext") ||
 663             !strcmp(name, "_sinittext") ||
 664             !strncmp("init_module", name, 11) ||
 665             !strncmp("cleanup_module", name, 14) ||
 666             strstr(name, "_text_start") ||
 667             strstr(name, "_text_end"))
 668                 return 1;
 669
 670         syme = symbol__priv(sym);
 671         syme->map = map;
 672         symbol__annotate_init(map, sym);
 673
 674         if (!top.sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
 675                 /* schedule initial sym_filter_entry setup */
 676                 sym_filter_entry_sched = syme;
 677                 sym_filter = NULL;
 678         }
 679
 680         for (i = 0; skip_symbols[i]; i++) {
 681                 if (!strcmp(skip_symbols[i], name)) {
 682                         sym->ignore = true;
 683                         break;
 684                 }
 685         }
 686
 687         return 0;
 688 }
 689
 690 static void perf_event__process_sample(const union perf_event *event,
 691                                        struct perf_sample *sample,
 692                                        struct perf_session *session)
 693 {
 694         u64 ip = event->ip.ip;
 695         struct sym_entry *syme;
 696         struct addr_location al;
 697         struct machine *machine;
 698         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 699
 700         ++top.samples;
 701
 702         switch (origin) {
 703         case PERF_RECORD_MISC_USER:
 704                 ++top.us_samples;
 705                 if (top.hide_user_symbols)
 706                         return;
 707                 machine = perf_session__find_host_machine(session);
 708                 break;
 709         case PERF_RECORD_MISC_KERNEL:
 710                 ++top.kernel_samples;
 711                 if (top.hide_kernel_symbols)
 712                         return;
 713                 machine = perf_session__find_host_machine(session);
 714                 break;
 715         case PERF_RECORD_MISC_GUEST_KERNEL:
 716                 ++top.guest_kernel_samples;
 717                 machine = perf_session__find_machine(session, event->ip.pid);
 718                 break;
 719         case PERF_RECORD_MISC_GUEST_USER:
 720                 ++top.guest_us_samples;
 721                 /*
 722                  * TODO: we don't process guest user from host side
 723                  * except simple counting.
 724                  */
 725                 return;
 726         default:
 727                 return;
 728         }
 729
 730         if (!machine && perf_guest) {
 731                 pr_err("Can't find guest [%d]'s kernel information\n",
 732                         event->ip.pid);
 733                 return;
 734         }
 735
 736         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 737                 top.exact_samples++;
 738
 739         if (perf_event__preprocess_sample(event, session, &al, sample,
 740                                           symbol_filter) < 0 ||
 741             al.filtered)
 742                 return;
 743
 744         if (!kptr_restrict_warned &&
 745             symbol_conf.kptr_restrict &&
 746             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 747                 ui__warning(
 748 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 749 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 750 "Kernel%s samples will not be resolved.\n",
 751                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 752                           " modules" : "");
 753                 if (use_browser <= 0)
 754                         sleep(5);
 755                 kptr_restrict_warned = true;
 756         }
 757
 758         if (al.sym == NULL) {
 759                 const char *msg = "Kernel samples will not be resolved.\n";
 760                 /*
 761                  * As we do lazy loading of symtabs we only will know if the
 762                  * specified vmlinux file is invalid when we actually have a
 763                  * hit in kernel space and then try to load it. So if we get
 764                  * here and there are _no_ symbols in the DSO backing the
 765                  * kernel map, bail out.
 766                  *
 767                  * We may never get here, for instance, if we use -K/
 768                  * --hide-kernel-symbols, even if the user specifies an
 769                  * invalid --vmlinux ;-)
 770                  */
 771                 if (!kptr_restrict_warned && !vmlinux_warned &&
 772                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 773                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 774                         if (symbol_conf.vmlinux_name) {
 775                                 ui__warning("The %s file can't be used.\n%s",
 776                                             symbol_conf.vmlinux_name, msg);
 777                         } else {
 778                                 ui__warning("A vmlinux file was not found.\n%s",
 779                                             msg);
 780                         }
 781
 782                         if (use_browser <= 0)
 783                                 sleep(5);
 784                         vmlinux_warned = true;
 785                 }
 786
 787                 return;
 788         }
 789
 790         /* let's see, whether we need to install initial sym_filter_entry */
 791         if (sym_filter_entry_sched) {
 792                 top.sym_filter_entry = sym_filter_entry_sched;
 793                 sym_filter_entry_sched = NULL;
 794                 if (parse_source(top.sym_filter_entry) < 0) {
 795                         struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
 796
 797                         pr_err("Can't annotate %s", sym->name);
 798                         if (top.sym_filter_entry->map->dso->symtab_type == SYMTAB__KALLSYMS) {
 799                                 pr_err(": No vmlinux file was found in the path:\n");
 800                                 machine__fprintf_vmlinux_path(machine, stderr);
 801                         } else
 802                                 pr_err(".\n");
 803                         exit(1);
 804                 }
 805         }
 806
 807         syme = symbol__priv(al.sym);
 808         if (!al.sym->ignore) {
 809                 struct perf_evsel *evsel;
 810
 811                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
 812                 assert(evsel != NULL);
 813                 syme->count[evsel->idx]++;
 814                 record_precise_ip(syme, al.map, evsel->idx, ip);
 815                 pthread_mutex_lock(&top.active_symbols_lock);
 816                 if (list_empty(&syme->node) || !syme->node.next) {
 817                         static bool first = true;
 818                         __list_insert_active_sym(syme);
 819                         if (first) {
 820                                 pthread_cond_broadcast(&top.active_symbols_cond);
 821                                 first = false;
 822                         }
 823                 }
 824                 pthread_mutex_unlock(&top.active_symbols_lock);
 825         }
 826 }
 827
 828 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
 829 {
 830         struct perf_sample sample;
 831         union perf_event *event;
 832         int ret;
 833
 834         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
 835                 ret = perf_session__parse_sample(self, event, &sample);
 836                 if (ret) {
 837                         pr_err("Can't parse sample, err = %d\n", ret);
 838                         continue;
 839                 }
 840
 841                 if (event->header.type == PERF_RECORD_SAMPLE)
 842                         perf_event__process_sample(event, &sample, self);
 843                 else
 844                         perf_event__process(event, &sample, self);
 845         }
 846 }
 847
 848 static void perf_session__mmap_read(struct perf_session *self)
 849 {
 850         int i;
 851
 852         for (i = 0; i < top.evlist->nr_mmaps; i++)
 853                 perf_session__mmap_read_idx(self, i);
 854 }
 855
 856 static void start_counters(struct perf_evlist *evlist)
 857 {
 858         struct perf_evsel *counter;
 859
 860         list_for_each_entry(counter, &evlist->entries, node) {
 861                 struct perf_event_attr *attr = &counter->attr;
 862
 863                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 864
 865                 if (top.freq) {
 866                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 867                         attr->freq        = 1;
 868                         attr->sample_freq = top.freq;
 869                 }
 870
 871                 if (evlist->nr_entries > 1) {
 872                         attr->sample_type |= PERF_SAMPLE_ID;
 873                         attr->read_format |= PERF_FORMAT_ID;
 874                 }
 875
 876                 attr->mmap = 1;
 877                 attr->inherit = inherit;
 878 try_again:
 879                 if (perf_evsel__open(counter, top.evlist->cpus,
 880                                      top.evlist->threads, group) < 0) {
 881                         int err = errno;
 882
 883                         if (err == EPERM || err == EACCES) {
 884                                 ui__warning_paranoid();
 885                                 goto out_err;
 886                         }
 887                         /*
 888                          * If it's cycles then fall back to hrtimer
 889                          * based cpu-clock-tick sw counter, which
 890                          * is always available even if no PMU support:
 891                          */
 892                         if (attr->type == PERF_TYPE_HARDWARE &&
 893                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 894                                 if (verbose)
 895                                         ui__warning("Cycles event not supported,\n"
 896                                                     "trying to fall back to cpu-clock-ticks\n");
 897
 898                                 attr->type = PERF_TYPE_SOFTWARE;
 899                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 900                                 goto try_again;
 901                         }
 902
 903                         if (err == ENOENT) {
 904                                 ui__warning("The %s event is not supported.\n",
 905                                             event_name(counter));
 906                                 goto out_err;
 907                         }
 908
 909                         ui__warning("The sys_perf_event_open() syscall "
 910                                     "returned with %d (%s).  /bin/dmesg "
 911                                     "may provide additional information.\n"
 912                                     "No CONFIG_PERF_EVENTS=y kernel support "
 913                                     "configured?\n", err, strerror(err));
 914                         goto out_err;
 915                 }
 916         }
 917
 918         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
 919                 ui__warning("Failed to mmap with %d (%s)\n",
 920                             errno, strerror(errno));
 921                 goto out_err;
 922         }
 923
 924         return;
 925
 926 out_err:
 927         exit_browser(0);
 928         exit(0);
 929 }
 930
 931 static int __cmd_top(void)
 932 {
 933         pthread_t thread;
 934         int ret __used;
 935         /*
 936          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 937          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 938          */
 939         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 940         if (session == NULL)
 941                 return -ENOMEM;
 942
 943         if (top.target_tid != -1)
 944                 perf_event__synthesize_thread_map(top.evlist->threads,
 945                                                   perf_event__process, session);
 946         else
 947                 perf_event__synthesize_threads(perf_event__process, session);
 948
 949         start_counters(top.evlist);
 950         session->evlist = top.evlist;
 951         perf_session__update_sample_type(session);
 952
 953         /* Wait for a minimal set of events before starting the snapshot */
 954         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 955
 956         perf_session__mmap_read(session);
 957
 958         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 959                                                              display_thread), session)) {
 960                 printf("Could not create display thread.\n");
 961                 exit(-1);
 962         }
 963
 964         if (realtime_prio) {
 965                 struct sched_param param;
 966
 967                 param.sched_priority = realtime_prio;
 968                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 969                         printf("Could not set realtime priority.\n");
 970                         exit(-1);
 971                 }
 972         }
 973
 974         while (1) {
 975                 u64 hits = top.samples;
 976
 977                 perf_session__mmap_read(session);
 978
 979                 if (hits == top.samples)
 980                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 981         }
 982
 983         return 0;
 984 }
 985
 986 static const char * const top_usage[] = {
 987         "perf top [<options>]",
 988         NULL
 989 };
 990
 991 static const struct option options[] = {
 992         OPT_CALLBACK('e', "event", &top.evlist, "event",
 993                      "event selector. use 'perf list' to list available events",
 994                      parse_events_option),
 995         OPT_INTEGER('c', "count", &default_interval,
 996                     "event period to sample"),
 997         OPT_INTEGER('p', "pid", &top.target_pid,
 998                     "profile events on existing process id"),
 999         OPT_INTEGER('t', "tid", &top.target_tid,
1000                     "profile events on existing thread id"),
1001         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1002                             "system-wide collection from all CPUs"),
1003         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1004                     "list of cpus to monitor"),
1005         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1006                    "file", "vmlinux pathname"),
1007         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1008                     "hide kernel symbols"),
1009         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1010         OPT_INTEGER('r', "realtime", &realtime_prio,
1011                     "collect data with this RT SCHED_FIFO priority"),
1012         OPT_INTEGER('d', "delay", &top.delay_secs,
1013                     "number of seconds to delay between refreshes"),
1014         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1015                             "dump the symbol table used for profiling"),
1016         OPT_INTEGER('f', "count-filter", &top.count_filter,
1017                     "only display functions with more events than this"),
1018         OPT_BOOLEAN('g', "group", &group,
1019                             "put the counters into a counter group"),
1020         OPT_BOOLEAN('i', "inherit", &inherit,
1021                     "child tasks inherit counters"),
1022         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1023                     "symbol to annotate"),
1024         OPT_BOOLEAN('z', "zero", &top.zero,
1025                     "zero history across updates"),
1026         OPT_INTEGER('F', "freq", &top.freq,
1027                     "profile at this frequency"),
1028         OPT_INTEGER('E', "entries", &top.print_entries,
1029                     "display this many functions"),
1030         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1031                     "hide user symbols"),
1032         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1033         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1034         OPT_INCR('v', "verbose", &verbose,
1035                     "be more verbose (show counter open errors, etc)"),
1036         OPT_END()
1037 };
1038
1039 int cmd_top(int argc, const char **argv, const char *prefix __used)
1040 {
1041         struct perf_evsel *pos;
1042         int status = -ENOMEM;
1043
1044         top.evlist = perf_evlist__new(NULL, NULL);
1045         if (top.evlist == NULL)
1046                 return -ENOMEM;
1047
1048         page_size = sysconf(_SC_PAGE_SIZE);
1049
1050         argc = parse_options(argc, argv, options, top_usage, 0);
1051         if (argc)
1052                 usage_with_options(top_usage, options);
1053
1054         /*
1055          * XXX For now start disabled, only using TUI if explicitely asked for.
1056          * Change that when handle_keys equivalent gets written, live annotation
1057          * done, etc.
1058          */
1059         use_browser = 0;
1060
1061         if (use_stdio)
1062                 use_browser = 0;
1063         else if (use_tui)
1064                 use_browser = 1;
1065
1066         setup_browser(false);
1067
1068         /* CPU and PID are mutually exclusive */
1069         if (top.target_tid > 0 && top.cpu_list) {
1070                 printf("WARNING: PID switch overriding CPU\n");
1071                 sleep(1);
1072                 top.cpu_list = NULL;
1073         }
1074
1075         if (top.target_pid != -1)
1076                 top.target_tid = top.target_pid;
1077
1078         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1079                                      top.target_tid, top.cpu_list) < 0)
1080                 usage_with_options(top_usage, options);
1081
1082         if (!top.evlist->nr_entries &&
1083             perf_evlist__add_default(top.evlist) < 0) {
1084                 pr_err("Not enough memory for event selector list\n");
1085                 return -ENOMEM;
1086         }
1087
1088         if (top.delay_secs < 1)
1089                 top.delay_secs = 1;
1090
1091         /*
1092          * User specified count overrides default frequency.
1093          */
1094         if (default_interval)
1095                 top.freq = 0;
1096         else if (top.freq) {
1097                 default_interval = top.freq;
1098         } else {
1099                 fprintf(stderr, "frequency and count are zero, aborting\n");
1100                 exit(EXIT_FAILURE);
1101         }
1102
1103         list_for_each_entry(pos, &top.evlist->entries, node) {
1104                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1105                                          top.evlist->threads->nr) < 0)
1106                         goto out_free_fd;
1107                 /*
1108                  * Fill in the ones not specifically initialized via -c:
1109                  */
1110                 if (pos->attr.sample_period)
1111                         continue;
1112
1113                 pos->attr.sample_period = default_interval;
1114         }
1115
1116         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1117             perf_evlist__alloc_mmap(top.evlist) < 0)
1118                 goto out_free_fd;
1119
1120         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1121
1122         symbol_conf.priv_size = (sizeof(struct sym_entry) + sizeof(struct annotation) +
1123                                  (top.evlist->nr_entries + 1) * sizeof(unsigned long));
1124
1125         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1126         if (symbol__init() < 0)
1127                 return -1;
1128
1129         get_term_dimensions(&winsize);
1130         if (top.print_entries == 0) {
1131                 update_print_entries(&winsize);
1132                 signal(SIGWINCH, sig_winch_handler);
1133         }
1134
1135         status = __cmd_top();
1136 out_free_fd:
1137         perf_evlist__delete(top.evlist);
1138
1139         return status;
1140 }