tools/perf/builtin-top.c

   1 /*
   2  * builtin-top.c
   3  *
   4  * Builtin top command: Display a continuously updated profile of
   5  * any workload, CPU or specific PID.
   6  *
   7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   9  *
  10  * Improvements and fixes by:
  11  *
  12  *   Arjan van de Ven <arjan@linux.intel.com>
  13  *   Yanmin Zhang <yanmin.zhang@intel.com>
  14  *   Wu Fengguang <fengguang.wu@intel.com>
  15  *   Mike Galbraith <efault@gmx.de>
  16  *   Paul Mackerras <paulus@samba.org>
  17  *
  18  * Released under the GPL v2. (and only v2, not any later version)
  19  */
  20 #include "builtin.h"
  21
  22 #include "perf.h"
  23
  24 #include "util/annotate.h"
  25 #include "util/cache.h"
  26 #include "util/color.h"
  27 #include "util/evlist.h"
  28 #include "util/evsel.h"
  29 #include "util/session.h"
  30 #include "util/symbol.h"
  31 #include "util/thread.h"
  32 #include "util/thread_map.h"
  33 #include "util/top.h"
  34 #include "util/util.h"
  35 #include <linux/rbtree.h>
  36 #include "util/parse-options.h"
  37 #include "util/parse-events.h"
  38 #include "util/cpumap.h"
  39 #include "util/xyarray.h"
  40 #include "util/sort.h"
  41
  42 #include "util/debug.h"
  43
  44 #include <assert.h>
  45 #include <fcntl.h>
  46
  47 #include <stdio.h>
  48 #include <termios.h>
  49 #include <unistd.h>
  50 #include <inttypes.h>
  51
  52 #include <errno.h>
  53 #include <time.h>
  54 #include <sched.h>
  55
  56 #include <sys/syscall.h>
  57 #include <sys/ioctl.h>
  58 #include <sys/poll.h>
  59 #include <sys/prctl.h>
  60 #include <sys/wait.h>
  61 #include <sys/uio.h>
  62 #include <sys/mman.h>
  63
  64 #include <linux/unistd.h>
  65 #include <linux/types.h>
  66
  67 static struct perf_top top = {
  68         .count_filter           = 5,
  69         .delay_secs             = 2,
  70         .target_pid             = -1,
  71         .target_tid             = -1,
  72         .freq                   = 1000, /* 1 KHz */
  73 };
  74
  75 static bool                     system_wide                     =  false;
  76
  77 static bool                     use_tui, use_stdio;
  78
  79 static bool                     sort_has_symbols;
  80
  81 static bool                     dont_use_callchains;
  82 static char                     callchain_default_opt[]         = "fractal,0.5,callee";
  83
  84
  85 static int                      default_interval                =      0;
  86
  87 static bool                     kptr_restrict_warned;
  88 static bool                     vmlinux_warned;
  89 static bool                     inherit                         =  false;
  90 static int                      realtime_prio                   =      0;
  91 static bool                     group                           =  false;
  92 static bool                     sample_id_all_avail             =   true;
  93 static unsigned int             mmap_pages                      =    128;
  94
  95 static bool                     dump_symtab                     =  false;
  96
  97 static struct winsize           winsize;
  98
  99 static const char               *sym_filter                     =   NULL;
 100 static int                      sym_pcnt_filter                 =      5;
 101
 102 /*
 103  * Source functions
 104  */
 105
 106 void get_term_dimensions(struct winsize *ws)
 107 {
 108         char *s = getenv("LINES");
 109
 110         if (s != NULL) {
 111                 ws->ws_row = atoi(s);
 112                 s = getenv("COLUMNS");
 113                 if (s != NULL) {
 114                         ws->ws_col = atoi(s);
 115                         if (ws->ws_row && ws->ws_col)
 116                                 return;
 117                 }
 118         }
 119 #ifdef TIOCGWINSZ
 120         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 121             ws->ws_row && ws->ws_col)
 122                 return;
 123 #endif
 124         ws->ws_row = 25;
 125         ws->ws_col = 80;
 126 }
 127
 128 static void update_print_entries(struct winsize *ws)
 129 {
 130         top.print_entries = ws->ws_row;
 131
 132         if (top.print_entries > 9)
 133                 top.print_entries -= 9;
 134 }
 135
 136 static void sig_winch_handler(int sig __used)
 137 {
 138         get_term_dimensions(&winsize);
 139         update_print_entries(&winsize);
 140 }
 141
 142 static int parse_source(struct hist_entry *he)
 143 {
 144         struct symbol *sym;
 145         struct annotation *notes;
 146         struct map *map;
 147         int err = -1;
 148
 149         if (!he || !he->ms.sym)
 150                 return -1;
 151
 152         sym = he->ms.sym;
 153         map = he->ms.map;
 154
 155         /*
 156          * We can't annotate with just /proc/kallsyms
 157          */
 158         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
 159                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 160                        "path\n", sym->name);
 161                 sleep(1);
 162                 return -1;
 163         }
 164
 165         notes = symbol__annotation(sym);
 166         if (notes->src != NULL) {
 167                 pthread_mutex_lock(&notes->lock);
 168                 goto out_assign;
 169         }
 170
 171         pthread_mutex_lock(&notes->lock);
 172
 173         if (symbol__alloc_hist(sym) < 0) {
 174                 pthread_mutex_unlock(&notes->lock);
 175                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 176                        sym->name);
 177                 sleep(1);
 178                 return err;
 179         }
 180
 181         err = symbol__annotate(sym, map, 0);
 182         if (err == 0) {
 183 out_assign:
 184                 top.sym_filter_entry = he;
 185         }
 186
 187         pthread_mutex_unlock(&notes->lock);
 188         return err;
 189 }
 190
 191 static void __zero_source_counters(struct hist_entry *he)
 192 {
 193         struct symbol *sym = he->ms.sym;
 194         symbol__annotate_zero_histograms(sym);
 195 }
 196
 197 static void record_precise_ip(struct hist_entry *he, int counter, u64 ip)
 198 {
 199         struct annotation *notes;
 200         struct symbol *sym;
 201
 202         if (he == NULL || he->ms.sym == NULL ||
 203             ((top.sym_filter_entry == NULL ||
 204               top.sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 205                 return;
 206
 207         sym = he->ms.sym;
 208         notes = symbol__annotation(sym);
 209
 210         if (pthread_mutex_trylock(&notes->lock))
 211                 return;
 212
 213         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 214                 pthread_mutex_unlock(&notes->lock);
 215                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 216                        sym->name);
 217                 sleep(1);
 218                 return;
 219         }
 220
 221         ip = he->ms.map->map_ip(he->ms.map, ip);
 222         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 223
 224         pthread_mutex_unlock(&notes->lock);
 225 }
 226
 227 static void show_details(struct hist_entry *he)
 228 {
 229         struct annotation *notes;
 230         struct symbol *symbol;
 231         int more;
 232
 233         if (!he)
 234                 return;
 235
 236         symbol = he->ms.sym;
 237         notes = symbol__annotation(symbol);
 238
 239         pthread_mutex_lock(&notes->lock);
 240
 241         if (notes->src == NULL)
 242                 goto out_unlock;
 243
 244         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
 245         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 246
 247         more = symbol__annotate_printf(symbol, he->ms.map, top.sym_evsel->idx,
 248                                        0, sym_pcnt_filter, top.print_entries, 4);
 249         if (top.zero)
 250                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
 251         else
 252                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
 253         if (more != 0)
 254                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 255 out_unlock:
 256         pthread_mutex_unlock(&notes->lock);
 257 }
 258
 259 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 260
 261 static struct hist_entry *
 262         perf_session__add_hist_entry(struct perf_session *session,
 263                                      struct addr_location *al,
 264                                      struct perf_sample *sample,
 265                                      struct perf_evsel *evsel)
 266 {
 267         struct hist_entry *he;
 268
 269         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 270         if (he == NULL)
 271                 return NULL;
 272
 273         session->hists.stats.total_period += sample->period;
 274         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 275         return he;
 276 }
 277
 278 static void print_sym_table(void)
 279 {
 280         char bf[160];
 281         int printed = 0;
 282         const int win_width = winsize.ws_col - 1;
 283
 284         puts(CONSOLE_CLEAR);
 285
 286         perf_top__header_snprintf(&top, bf, sizeof(bf));
 287         printf("%s\n", bf);
 288
 289         perf_top__reset_sample_counters(&top);
 290
 291         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 292
 293         if (top.sym_evsel->hists.stats.nr_lost_warned !=
 294             top.sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 295                 top.sym_evsel->hists.stats.nr_lost_warned =
 296                         top.sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 297                 color_fprintf(stdout, PERF_COLOR_RED,
 298                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 299                               top.sym_evsel->hists.stats.nr_lost_warned);
 300                 ++printed;
 301         }
 302
 303         if (top.sym_filter_entry) {
 304                 show_details(top.sym_filter_entry);
 305                 return;
 306         }
 307
 308         hists__collapse_resort_threaded(&top.sym_evsel->hists);
 309         hists__output_resort_threaded(&top.sym_evsel->hists);
 310         hists__decay_entries_threaded(&top.sym_evsel->hists,
 311                                       top.hide_user_symbols,
 312                                       top.hide_kernel_symbols);
 313         hists__output_recalc_col_len(&top.sym_evsel->hists, winsize.ws_row - 3);
 314         putchar('\n');
 315         hists__fprintf(&top.sym_evsel->hists, NULL, false, false,
 316                        winsize.ws_row - 4 - printed, win_width, stdout);
 317 }
 318
 319 static void prompt_integer(int *target, const char *msg)
 320 {
 321         char *buf = malloc(0), *p;
 322         size_t dummy = 0;
 323         int tmp;
 324
 325         fprintf(stdout, "\n%s: ", msg);
 326         if (getline(&buf, &dummy, stdin) < 0)
 327                 return;
 328
 329         p = strchr(buf, '\n');
 330         if (p)
 331                 *p = 0;
 332
 333         p = buf;
 334         while(*p) {
 335                 if (!isdigit(*p))
 336                         goto out_free;
 337                 p++;
 338         }
 339         tmp = strtoul(buf, NULL, 10);
 340         *target = tmp;
 341 out_free:
 342         free(buf);
 343 }
 344
 345 static void prompt_percent(int *target, const char *msg)
 346 {
 347         int tmp = 0;
 348
 349         prompt_integer(&tmp, msg);
 350         if (tmp >= 0 && tmp <= 100)
 351                 *target = tmp;
 352 }
 353
 354 static void prompt_symbol(struct hist_entry **target, const char *msg)
 355 {
 356         char *buf = malloc(0), *p;
 357         struct hist_entry *syme = *target, *n, *found = NULL;
 358         struct rb_node *next;
 359         size_t dummy = 0;
 360
 361         /* zero counters of active symbol */
 362         if (syme) {
 363                 __zero_source_counters(syme);
 364                 *target = NULL;
 365         }
 366
 367         fprintf(stdout, "\n%s: ", msg);
 368         if (getline(&buf, &dummy, stdin) < 0)
 369                 goto out_free;
 370
 371         p = strchr(buf, '\n');
 372         if (p)
 373                 *p = 0;
 374
 375         next = rb_first(&top.sym_evsel->hists.entries);
 376         while (next) {
 377                 n = rb_entry(next, struct hist_entry, rb_node);
 378                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 379                         found = n;
 380                         break;
 381                 }
 382                 next = rb_next(&n->rb_node);
 383         }
 384
 385         if (!found) {
 386                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 387                 sleep(1);
 388                 return;
 389         } else
 390                 parse_source(found);
 391
 392 out_free:
 393         free(buf);
 394 }
 395
 396 static void print_mapped_keys(void)
 397 {
 398         char *name = NULL;
 399
 400         if (top.sym_filter_entry) {
 401                 struct symbol *sym = top.sym_filter_entry->ms.sym;
 402                 name = sym->name;
 403         }
 404
 405         fprintf(stdout, "\nMapped keys:\n");
 406         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
 407         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
 408
 409         if (top.evlist->nr_entries > 1)
 410                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
 411
 412         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
 413
 414         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 415         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 416         fprintf(stdout, "\t[S]     stop annotation.\n");
 417
 418         fprintf(stdout,
 419                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 420                 top.hide_kernel_symbols ? "yes" : "no");
 421         fprintf(stdout,
 422                 "\t[U]     hide user symbols.               \t(%s)\n",
 423                 top.hide_user_symbols ? "yes" : "no");
 424         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
 425         fprintf(stdout, "\t[qQ]    quit.\n");
 426 }
 427
 428 static int key_mapped(int c)
 429 {
 430         switch (c) {
 431                 case 'd':
 432                 case 'e':
 433                 case 'f':
 434                 case 'z':
 435                 case 'q':
 436                 case 'Q':
 437                 case 'K':
 438                 case 'U':
 439                 case 'F':
 440                 case 's':
 441                 case 'S':
 442                         return 1;
 443                 case 'E':
 444                         return top.evlist->nr_entries > 1 ? 1 : 0;
 445                 default:
 446                         break;
 447         }
 448
 449         return 0;
 450 }
 451
 452 static void handle_keypress(int c)
 453 {
 454         if (!key_mapped(c)) {
 455                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 456                 struct termios tc, save;
 457
 458                 print_mapped_keys();
 459                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 460                 fflush(stdout);
 461
 462                 tcgetattr(0, &save);
 463                 tc = save;
 464                 tc.c_lflag &= ~(ICANON | ECHO);
 465                 tc.c_cc[VMIN] = 0;
 466                 tc.c_cc[VTIME] = 0;
 467                 tcsetattr(0, TCSANOW, &tc);
 468
 469                 poll(&stdin_poll, 1, -1);
 470                 c = getc(stdin);
 471
 472                 tcsetattr(0, TCSAFLUSH, &save);
 473                 if (!key_mapped(c))
 474                         return;
 475         }
 476
 477         switch (c) {
 478                 case 'd':
 479                         prompt_integer(&top.delay_secs, "Enter display delay");
 480                         if (top.delay_secs < 1)
 481                                 top.delay_secs = 1;
 482                         break;
 483                 case 'e':
 484                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
 485                         if (top.print_entries == 0) {
 486                                 sig_winch_handler(SIGWINCH);
 487                                 signal(SIGWINCH, sig_winch_handler);
 488                         } else
 489                                 signal(SIGWINCH, SIG_DFL);
 490                         break;
 491                 case 'E':
 492                         if (top.evlist->nr_entries > 1) {
 493                                 /* Select 0 as the default event: */
 494                                 int counter = 0;
 495
 496                                 fprintf(stderr, "\nAvailable events:");
 497
 498                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 499                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 500
 501                                 prompt_integer(&counter, "Enter details event counter");
 502
 503                                 if (counter >= top.evlist->nr_entries) {
 504                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 505                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 506                                         sleep(1);
 507                                         break;
 508                                 }
 509                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 510                                         if (top.sym_evsel->idx == counter)
 511                                                 break;
 512                         } else
 513                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 514                         break;
 515                 case 'f':
 516                         prompt_integer(&top.count_filter, "Enter display event count filter");
 517                         break;
 518                 case 'F':
 519                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 520                         break;
 521                 case 'K':
 522                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 523                         break;
 524                 case 'q':
 525                 case 'Q':
 526                         printf("exiting.\n");
 527                         if (dump_symtab)
 528                                 perf_session__fprintf_dsos(top.session, stderr);
 529                         exit(0);
 530                 case 's':
 531                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 532                         break;
 533                 case 'S':
 534                         if (!top.sym_filter_entry)
 535                                 break;
 536                         else {
 537                                 struct hist_entry *syme = top.sym_filter_entry;
 538
 539                                 top.sym_filter_entry = NULL;
 540                                 __zero_source_counters(syme);
 541                         }
 542                         break;
 543                 case 'U':
 544                         top.hide_user_symbols = !top.hide_user_symbols;
 545                         break;
 546                 case 'z':
 547                         top.zero = !top.zero;
 548                         break;
 549                 default:
 550                         break;
 551         }
 552 }
 553
 554 static void perf_top__sort_new_samples(void *arg)
 555 {
 556         struct perf_top *t = arg;
 557         perf_top__reset_sample_counters(t);
 558
 559         if (t->evlist->selected != NULL)
 560                 t->sym_evsel = t->evlist->selected;
 561
 562         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 563         hists__output_resort_threaded(&t->sym_evsel->hists);
 564         hists__decay_entries_threaded(&t->sym_evsel->hists,
 565                                       top.hide_user_symbols,
 566                                       top.hide_kernel_symbols);
 567 }
 568
 569 static void *display_thread_tui(void *arg __used)
 570 {
 571         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 572
 573         perf_top__sort_new_samples(&top);
 574         perf_evlist__tui_browse_hists(top.evlist, help,
 575                                       perf_top__sort_new_samples,
 576                                       &top, top.delay_secs);
 577
 578         exit_browser(0);
 579         exit(0);
 580         return NULL;
 581 }
 582
 583 static void *display_thread(void *arg __used)
 584 {
 585         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 586         struct termios tc, save;
 587         int delay_msecs, c;
 588
 589         tcgetattr(0, &save);
 590         tc = save;
 591         tc.c_lflag &= ~(ICANON | ECHO);
 592         tc.c_cc[VMIN] = 0;
 593         tc.c_cc[VTIME] = 0;
 594
 595         pthread__unblock_sigwinch();
 596 repeat:
 597         delay_msecs = top.delay_secs * 1000;
 598         tcsetattr(0, TCSANOW, &tc);
 599         /* trash return*/
 600         getc(stdin);
 601
 602         while (1) {
 603                 print_sym_table();
 604                 /*
 605                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 606                  * refresh screen in both cases.
 607                  */
 608                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 609                 case 0:
 610                         continue;
 611                 case -1:
 612                         if (errno == EINTR)
 613                                 continue;
 614                         /* Fall trhu */
 615                 default:
 616                         goto process_hotkey;
 617                 }
 618         }
 619 process_hotkey:
 620         c = getc(stdin);
 621         tcsetattr(0, TCSAFLUSH, &save);
 622
 623         handle_keypress(c);
 624         goto repeat;
 625
 626         return NULL;
 627 }
 628
 629 /* Tag samples to be skipped. */
 630 static const char *skip_symbols[] = {
 631         "default_idle",
 632         "native_safe_halt",
 633         "cpu_idle",
 634         "enter_idle",
 635         "exit_idle",
 636         "mwait_idle",
 637         "mwait_idle_with_hints",
 638         "poll_idle",
 639         "ppc64_runlatch_off",
 640         "pseries_dedicated_idle_sleep",
 641         NULL
 642 };
 643
 644 static int symbol_filter(struct map *map __used, struct symbol *sym)
 645 {
 646         const char *name = sym->name;
 647         int i;
 648
 649         /*
 650          * ppc64 uses function descriptors and appends a '.' to the
 651          * start of every instruction address. Remove it.
 652          */
 653         if (name[0] == '.')
 654                 name++;
 655
 656         if (!strcmp(name, "_text") ||
 657             !strcmp(name, "_etext") ||
 658             !strcmp(name, "_sinittext") ||
 659             !strncmp("init_module", name, 11) ||
 660             !strncmp("cleanup_module", name, 14) ||
 661             strstr(name, "_text_start") ||
 662             strstr(name, "_text_end"))
 663                 return 1;
 664
 665         for (i = 0; skip_symbols[i]; i++) {
 666                 if (!strcmp(skip_symbols[i], name)) {
 667                         sym->ignore = true;
 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_evsel *evsel,
 677                                        struct perf_sample *sample,
 678                                        struct perf_session *session)
 679 {
 680         struct symbol *parent = NULL;
 681         u64 ip = event->ip.ip;
 682         struct addr_location al;
 683         struct machine *machine;
 684         int err;
 685         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 686
 687         ++top.samples;
 688
 689         switch (origin) {
 690         case PERF_RECORD_MISC_USER:
 691                 ++top.us_samples;
 692                 if (top.hide_user_symbols)
 693                         return;
 694                 machine = perf_session__find_host_machine(session);
 695                 break;
 696         case PERF_RECORD_MISC_KERNEL:
 697                 ++top.kernel_samples;
 698                 if (top.hide_kernel_symbols)
 699                         return;
 700                 machine = perf_session__find_host_machine(session);
 701                 break;
 702         case PERF_RECORD_MISC_GUEST_KERNEL:
 703                 ++top.guest_kernel_samples;
 704                 machine = perf_session__find_machine(session, event->ip.pid);
 705                 break;
 706         case PERF_RECORD_MISC_GUEST_USER:
 707                 ++top.guest_us_samples;
 708                 /*
 709                  * TODO: we don't process guest user from host side
 710                  * except simple counting.
 711                  */
 712                 return;
 713         default:
 714                 return;
 715         }
 716
 717         if (!machine && perf_guest) {
 718                 pr_err("Can't find guest [%d]'s kernel information\n",
 719                         event->ip.pid);
 720                 return;
 721         }
 722
 723         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 724                 top.exact_samples++;
 725
 726         if (perf_event__preprocess_sample(event, session, &al, sample,
 727                                           symbol_filter) < 0 ||
 728             al.filtered)
 729                 return;
 730
 731         if (!kptr_restrict_warned &&
 732             symbol_conf.kptr_restrict &&
 733             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 734                 ui__warning(
 735 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 736 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 737 "Kernel%s samples will not be resolved.\n",
 738                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 739                           " modules" : "");
 740                 if (use_browser <= 0)
 741                         sleep(5);
 742                 kptr_restrict_warned = true;
 743         }
 744
 745         if (al.sym == NULL) {
 746                 const char *msg = "Kernel samples will not be resolved.\n";
 747                 /*
 748                  * As we do lazy loading of symtabs we only will know if the
 749                  * specified vmlinux file is invalid when we actually have a
 750                  * hit in kernel space and then try to load it. So if we get
 751                  * here and there are _no_ symbols in the DSO backing the
 752                  * kernel map, bail out.
 753                  *
 754                  * We may never get here, for instance, if we use -K/
 755                  * --hide-kernel-symbols, even if the user specifies an
 756                  * invalid --vmlinux ;-)
 757                  */
 758                 if (!kptr_restrict_warned && !vmlinux_warned &&
 759                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 760                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 761                         if (symbol_conf.vmlinux_name) {
 762                                 ui__warning("The %s file can't be used.\n%s",
 763                                             symbol_conf.vmlinux_name, msg);
 764                         } else {
 765                                 ui__warning("A vmlinux file was not found.\n%s",
 766                                             msg);
 767                         }
 768
 769                         if (use_browser <= 0)
 770                                 sleep(5);
 771                         vmlinux_warned = true;
 772                 }
 773         }
 774
 775         if (al.sym == NULL || !al.sym->ignore) {
 776                 struct hist_entry *he;
 777
 778                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 779                     sample->callchain) {
 780                         err = perf_session__resolve_callchain(session, evsel, al.thread,
 781                                                               sample->callchain, &parent);
 782                         if (err)
 783                                 return;
 784                 }
 785
 786                 he = perf_session__add_hist_entry(session, &al, sample, evsel);
 787                 if (he == NULL) {
 788                         pr_err("Problem incrementing symbol period, skipping event\n");
 789                         return;
 790                 }
 791
 792                 if (symbol_conf.use_callchain) {
 793                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
 794                                                sample->period);
 795                         if (err)
 796                                 return;
 797                 }
 798
 799                 if (sort_has_symbols)
 800                         record_precise_ip(he, evsel->idx, ip);
 801         }
 802
 803         return;
 804 }
 805
 806 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
 807 {
 808         struct perf_sample sample;
 809         struct perf_evsel *evsel;
 810         union perf_event *event;
 811         int ret;
 812
 813         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
 814                 ret = perf_session__parse_sample(self, event, &sample);
 815                 if (ret) {
 816                         pr_err("Can't parse sample, err = %d\n", ret);
 817                         continue;
 818                 }
 819
 820                 evsel = perf_evlist__id2evsel(self->evlist, sample.id);
 821                 assert(evsel != NULL);
 822
 823                 if (event->header.type == PERF_RECORD_SAMPLE)
 824                         perf_event__process_sample(event, evsel, &sample, self);
 825                 else if (event->header.type < PERF_RECORD_MAX) {
 826                         hists__inc_nr_events(&evsel->hists, event->header.type);
 827                         perf_event__process(event, &sample, self);
 828                 } else
 829                         ++self->hists.stats.nr_unknown_events;
 830         }
 831 }
 832
 833 static void perf_session__mmap_read(struct perf_session *self)
 834 {
 835         int i;
 836
 837         for (i = 0; i < top.evlist->nr_mmaps; i++)
 838                 perf_session__mmap_read_idx(self, i);
 839 }
 840
 841 static void start_counters(struct perf_evlist *evlist)
 842 {
 843         struct perf_evsel *counter, *first;
 844
 845         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 846
 847         list_for_each_entry(counter, &evlist->entries, node) {
 848                 struct perf_event_attr *attr = &counter->attr;
 849                 struct xyarray *group_fd = NULL;
 850
 851                 if (group && counter != first)
 852                         group_fd = first->fd;
 853
 854                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 855
 856                 if (top.freq) {
 857                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 858                         attr->freq        = 1;
 859                         attr->sample_freq = top.freq;
 860                 }
 861
 862                 if (evlist->nr_entries > 1) {
 863                         attr->sample_type |= PERF_SAMPLE_ID;
 864                         attr->read_format |= PERF_FORMAT_ID;
 865                 }
 866
 867                 if (symbol_conf.use_callchain)
 868                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
 869
 870                 attr->mmap = 1;
 871                 attr->comm = 1;
 872                 attr->inherit = inherit;
 873 retry_sample_id:
 874                 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
 875 try_again:
 876                 if (perf_evsel__open(counter, top.evlist->cpus,
 877                                      top.evlist->threads, group,
 878                                      group_fd) < 0) {
 879                         int err = errno;
 880
 881                         if (err == EPERM || err == EACCES) {
 882                                 ui__error_paranoid();
 883                                 goto out_err;
 884                         } else if (err == EINVAL && sample_id_all_avail) {
 885                                 /*
 886                                  * Old kernel, no attr->sample_id_type_all field
 887                                  */
 888                                 sample_id_all_avail = false;
 889                                 goto retry_sample_id;
 890                         }
 891                         /*
 892                          * If it's cycles then fall back to hrtimer
 893                          * based cpu-clock-tick sw counter, which
 894                          * is always available even if no PMU support:
 895                          */
 896                         if (attr->type == PERF_TYPE_HARDWARE &&
 897                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 898                                 if (verbose)
 899                                         ui__warning("Cycles event not supported,\n"
 900                                                     "trying to fall back to cpu-clock-ticks\n");
 901
 902                                 attr->type = PERF_TYPE_SOFTWARE;
 903                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 904                                 goto try_again;
 905                         }
 906
 907                         if (err == ENOENT) {
 908                                 ui__warning("The %s event is not supported.\n",
 909                                             event_name(counter));
 910                                 goto out_err;
 911                         }
 912
 913                         ui__warning("The sys_perf_event_open() syscall "
 914                                     "returned with %d (%s).  /bin/dmesg "
 915                                     "may provide additional information.\n"
 916                                     "No CONFIG_PERF_EVENTS=y kernel support "
 917                                     "configured?\n", err, strerror(err));
 918                         goto out_err;
 919                 }
 920         }
 921
 922         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
 923                 ui__warning("Failed to mmap with %d (%s)\n",
 924                             errno, strerror(errno));
 925                 goto out_err;
 926         }
 927
 928         return;
 929
 930 out_err:
 931         exit_browser(0);
 932         exit(0);
 933 }
 934
 935 static int setup_sample_type(void)
 936 {
 937         if (!sort_has_symbols) {
 938                 if (symbol_conf.use_callchain) {
 939                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
 940                         return -EINVAL;
 941                 }
 942         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
 943                 if (callchain_register_param(&callchain_param) < 0) {
 944                         ui__warning("Can't register callchain params.\n");
 945                         return -EINVAL;
 946                 }
 947         }
 948
 949         return 0;
 950 }
 951
 952 static int __cmd_top(void)
 953 {
 954         pthread_t thread;
 955         int ret;
 956         /*
 957          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 958          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 959          */
 960         top.session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 961         if (top.session == NULL)
 962                 return -ENOMEM;
 963
 964         ret = setup_sample_type();
 965         if (ret)
 966                 goto out_delete;
 967
 968         if (top.target_tid != -1)
 969                 perf_event__synthesize_thread_map(top.evlist->threads,
 970                                                   perf_event__process, top.session);
 971         else
 972                 perf_event__synthesize_threads(perf_event__process, top.session);
 973
 974         start_counters(top.evlist);
 975         top.session->evlist = top.evlist;
 976         perf_session__update_sample_type(top.session);
 977
 978         /* Wait for a minimal set of events before starting the snapshot */
 979         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 980
 981         perf_session__mmap_read(top.session);
 982
 983         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 984                                                              display_thread), NULL)) {
 985                 printf("Could not create display thread.\n");
 986                 exit(-1);
 987         }
 988
 989         if (realtime_prio) {
 990                 struct sched_param param;
 991
 992                 param.sched_priority = realtime_prio;
 993                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 994                         printf("Could not set realtime priority.\n");
 995                         exit(-1);
 996                 }
 997         }
 998
 999         while (1) {
1000                 u64 hits = top.samples;
1001
1002                 perf_session__mmap_read(top.session);
1003
1004                 if (hits == top.samples)
1005                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
1006         }
1007
1008 out_delete:
1009         perf_session__delete(top.session);
1010         top.session = NULL;
1011
1012         return 0;
1013 }
1014
1015 static int
1016 parse_callchain_opt(const struct option *opt __used, const char *arg,
1017                     int unset)
1018 {
1019         char *tok, *tok2;
1020         char *endptr;
1021
1022         /*
1023          * --no-call-graph
1024          */
1025         if (unset) {
1026                 dont_use_callchains = true;
1027                 return 0;
1028         }
1029
1030         symbol_conf.use_callchain = true;
1031
1032         if (!arg)
1033                 return 0;
1034
1035         tok = strtok((char *)arg, ",");
1036         if (!tok)
1037                 return -1;
1038
1039         /* get the output mode */
1040         if (!strncmp(tok, "graph", strlen(arg)))
1041                 callchain_param.mode = CHAIN_GRAPH_ABS;
1042
1043         else if (!strncmp(tok, "flat", strlen(arg)))
1044                 callchain_param.mode = CHAIN_FLAT;
1045
1046         else if (!strncmp(tok, "fractal", strlen(arg)))
1047                 callchain_param.mode = CHAIN_GRAPH_REL;
1048
1049         else if (!strncmp(tok, "none", strlen(arg))) {
1050                 callchain_param.mode = CHAIN_NONE;
1051                 symbol_conf.use_callchain = false;
1052
1053                 return 0;
1054         }
1055
1056         else
1057                 return -1;
1058
1059         /* get the min percentage */
1060         tok = strtok(NULL, ",");
1061         if (!tok)
1062                 goto setup;
1063
1064         callchain_param.min_percent = strtod(tok, &endptr);
1065         if (tok == endptr)
1066                 return -1;
1067
1068         /* get the print limit */
1069         tok2 = strtok(NULL, ",");
1070         if (!tok2)
1071                 goto setup;
1072
1073         if (tok2[0] != 'c') {
1074                 callchain_param.print_limit = strtod(tok2, &endptr);
1075                 tok2 = strtok(NULL, ",");
1076                 if (!tok2)
1077                         goto setup;
1078         }
1079
1080         /* get the call chain order */
1081         if (!strcmp(tok2, "caller"))
1082                 callchain_param.order = ORDER_CALLER;
1083         else if (!strcmp(tok2, "callee"))
1084                 callchain_param.order = ORDER_CALLEE;
1085         else
1086                 return -1;
1087 setup:
1088         if (callchain_register_param(&callchain_param) < 0) {
1089                 fprintf(stderr, "Can't register callchain params\n");
1090                 return -1;
1091         }
1092         return 0;
1093 }
1094
1095 static const char * const top_usage[] = {
1096         "perf top [<options>]",
1097         NULL
1098 };
1099
1100 static const struct option options[] = {
1101         OPT_CALLBACK('e', "event", &top.evlist, "event",
1102                      "event selector. use 'perf list' to list available events",
1103                      parse_events_option),
1104         OPT_INTEGER('c', "count", &default_interval,
1105                     "event period to sample"),
1106         OPT_INTEGER('p', "pid", &top.target_pid,
1107                     "profile events on existing process id"),
1108         OPT_INTEGER('t', "tid", &top.target_tid,
1109                     "profile events on existing thread id"),
1110         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1111                             "system-wide collection from all CPUs"),
1112         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1113                     "list of cpus to monitor"),
1114         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1115                    "file", "vmlinux pathname"),
1116         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1117                     "hide kernel symbols"),
1118         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1119         OPT_INTEGER('r', "realtime", &realtime_prio,
1120                     "collect data with this RT SCHED_FIFO priority"),
1121         OPT_INTEGER('d', "delay", &top.delay_secs,
1122                     "number of seconds to delay between refreshes"),
1123         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1124                             "dump the symbol table used for profiling"),
1125         OPT_INTEGER('f', "count-filter", &top.count_filter,
1126                     "only display functions with more events than this"),
1127         OPT_BOOLEAN('g', "group", &group,
1128                             "put the counters into a counter group"),
1129         OPT_BOOLEAN('i', "inherit", &inherit,
1130                     "child tasks inherit counters"),
1131         OPT_STRING(0, "sym-annotate", &sym_filter, "symbol name",
1132                     "symbol to annotate"),
1133         OPT_BOOLEAN('z', "zero", &top.zero,
1134                     "zero history across updates"),
1135         OPT_INTEGER('F', "freq", &top.freq,
1136                     "profile at this frequency"),
1137         OPT_INTEGER('E', "entries", &top.print_entries,
1138                     "display this many functions"),
1139         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1140                     "hide user symbols"),
1141         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1142         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1143         OPT_INCR('v', "verbose", &verbose,
1144                     "be more verbose (show counter open errors, etc)"),
1145         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1146                    "sort by key(s): pid, comm, dso, symbol, parent"),
1147         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1148                     "Show a column with the number of samples"),
1149         OPT_CALLBACK_DEFAULT('G', "call-graph", NULL, "output_type,min_percent, call_order",
1150                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1151                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1152                      callchain_default_opt),
1153         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1154                     "Show a column with the sum of periods"),
1155         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1156                    "only consider symbols in these dsos"),
1157         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1158                    "only consider symbols in these comms"),
1159         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1160                    "only consider these symbols"),
1161         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1162                     "Interleave source code with assembly code (default)"),
1163         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1164                     "Display raw encoding of assembly instructions (default)"),
1165         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1166                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1167         OPT_END()
1168 };
1169
1170 int cmd_top(int argc, const char **argv, const char *prefix __used)
1171 {
1172         struct perf_evsel *pos;
1173         int status = -ENOMEM;
1174
1175         top.evlist = perf_evlist__new(NULL, NULL);
1176         if (top.evlist == NULL)
1177                 return -ENOMEM;
1178
1179         symbol_conf.exclude_other = false;
1180
1181         argc = parse_options(argc, argv, options, top_usage, 0);
1182         if (argc)
1183                 usage_with_options(top_usage, options);
1184
1185         if (sort_order == default_sort_order)
1186                 sort_order = "dso,symbol";
1187
1188         setup_sorting(top_usage, options);
1189
1190         if (use_stdio)
1191                 use_browser = 0;
1192         else if (use_tui)
1193                 use_browser = 1;
1194
1195         setup_browser(false);
1196
1197         /* CPU and PID are mutually exclusive */
1198         if (top.target_tid > 0 && top.cpu_list) {
1199                 printf("WARNING: PID switch overriding CPU\n");
1200                 sleep(1);
1201                 top.cpu_list = NULL;
1202         }
1203
1204         if (top.target_pid != -1)
1205                 top.target_tid = top.target_pid;
1206
1207         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1208                                      top.target_tid, top.cpu_list) < 0)
1209                 usage_with_options(top_usage, options);
1210
1211         if (!top.evlist->nr_entries &&
1212             perf_evlist__add_default(top.evlist) < 0) {
1213                 pr_err("Not enough memory for event selector list\n");
1214                 return -ENOMEM;
1215         }
1216
1217         symbol_conf.nr_events = top.evlist->nr_entries;
1218
1219         if (top.delay_secs < 1)
1220                 top.delay_secs = 1;
1221
1222         /*
1223          * User specified count overrides default frequency.
1224          */
1225         if (default_interval)
1226                 top.freq = 0;
1227         else if (top.freq) {
1228                 default_interval = top.freq;
1229         } else {
1230                 fprintf(stderr, "frequency and count are zero, aborting\n");
1231                 exit(EXIT_FAILURE);
1232         }
1233
1234         list_for_each_entry(pos, &top.evlist->entries, node) {
1235                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1236                                          top.evlist->threads->nr) < 0)
1237                         goto out_free_fd;
1238                 /*
1239                  * Fill in the ones not specifically initialized via -c:
1240                  */
1241                 if (pos->attr.sample_period)
1242                         continue;
1243
1244                 pos->attr.sample_period = default_interval;
1245         }
1246
1247         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1248             perf_evlist__alloc_mmap(top.evlist) < 0)
1249                 goto out_free_fd;
1250
1251         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1252
1253         symbol_conf.priv_size = sizeof(struct annotation);
1254
1255         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1256         if (symbol__init() < 0)
1257                 return -1;
1258
1259         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1260         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1261         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1262
1263         /*
1264          * Avoid annotation data structures overhead when symbols aren't on the
1265          * sort list.
1266          */
1267         sort_has_symbols = sort_sym.list.next != NULL;
1268
1269         get_term_dimensions(&winsize);
1270         if (top.print_entries == 0) {
1271                 update_print_entries(&winsize);
1272                 signal(SIGWINCH, sig_winch_handler);
1273         }
1274
1275         status = __cmd_top();
1276 out_free_fd:
1277         perf_evlist__delete(top.evlist);
1278
1279         return status;
1280 }