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 *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 262                                                      struct addr_location *al,
 263                                                      struct perf_sample *sample)
 264 {
 265         struct hist_entry *he;
 266
 267         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 268         if (he == NULL)
 269                 return NULL;
 270
 271         evsel->hists.stats.total_period += sample->period;
 272         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 273         return he;
 274 }
 275
 276 static void print_sym_table(void)
 277 {
 278         char bf[160];
 279         int printed = 0;
 280         const int win_width = winsize.ws_col - 1;
 281
 282         puts(CONSOLE_CLEAR);
 283
 284         perf_top__header_snprintf(&top, bf, sizeof(bf));
 285         printf("%s\n", bf);
 286
 287         perf_top__reset_sample_counters(&top);
 288
 289         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 290
 291         if (top.sym_evsel->hists.stats.nr_lost_warned !=
 292             top.sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 293                 top.sym_evsel->hists.stats.nr_lost_warned =
 294                         top.sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 295                 color_fprintf(stdout, PERF_COLOR_RED,
 296                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 297                               top.sym_evsel->hists.stats.nr_lost_warned);
 298                 ++printed;
 299         }
 300
 301         if (top.sym_filter_entry) {
 302                 show_details(top.sym_filter_entry);
 303                 return;
 304         }
 305
 306         hists__collapse_resort_threaded(&top.sym_evsel->hists);
 307         hists__output_resort_threaded(&top.sym_evsel->hists);
 308         hists__decay_entries_threaded(&top.sym_evsel->hists,
 309                                       top.hide_user_symbols,
 310                                       top.hide_kernel_symbols);
 311         hists__output_recalc_col_len(&top.sym_evsel->hists, winsize.ws_row - 3);
 312         putchar('\n');
 313         hists__fprintf(&top.sym_evsel->hists, NULL, false, false,
 314                        winsize.ws_row - 4 - printed, win_width, stdout);
 315 }
 316
 317 static void prompt_integer(int *target, const char *msg)
 318 {
 319         char *buf = malloc(0), *p;
 320         size_t dummy = 0;
 321         int tmp;
 322
 323         fprintf(stdout, "\n%s: ", msg);
 324         if (getline(&buf, &dummy, stdin) < 0)
 325                 return;
 326
 327         p = strchr(buf, '\n');
 328         if (p)
 329                 *p = 0;
 330
 331         p = buf;
 332         while(*p) {
 333                 if (!isdigit(*p))
 334                         goto out_free;
 335                 p++;
 336         }
 337         tmp = strtoul(buf, NULL, 10);
 338         *target = tmp;
 339 out_free:
 340         free(buf);
 341 }
 342
 343 static void prompt_percent(int *target, const char *msg)
 344 {
 345         int tmp = 0;
 346
 347         prompt_integer(&tmp, msg);
 348         if (tmp >= 0 && tmp <= 100)
 349                 *target = tmp;
 350 }
 351
 352 static void prompt_symbol(struct hist_entry **target, const char *msg)
 353 {
 354         char *buf = malloc(0), *p;
 355         struct hist_entry *syme = *target, *n, *found = NULL;
 356         struct rb_node *next;
 357         size_t dummy = 0;
 358
 359         /* zero counters of active symbol */
 360         if (syme) {
 361                 __zero_source_counters(syme);
 362                 *target = NULL;
 363         }
 364
 365         fprintf(stdout, "\n%s: ", msg);
 366         if (getline(&buf, &dummy, stdin) < 0)
 367                 goto out_free;
 368
 369         p = strchr(buf, '\n');
 370         if (p)
 371                 *p = 0;
 372
 373         next = rb_first(&top.sym_evsel->hists.entries);
 374         while (next) {
 375                 n = rb_entry(next, struct hist_entry, rb_node);
 376                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 377                         found = n;
 378                         break;
 379                 }
 380                 next = rb_next(&n->rb_node);
 381         }
 382
 383         if (!found) {
 384                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 385                 sleep(1);
 386                 return;
 387         } else
 388                 parse_source(found);
 389
 390 out_free:
 391         free(buf);
 392 }
 393
 394 static void print_mapped_keys(void)
 395 {
 396         char *name = NULL;
 397
 398         if (top.sym_filter_entry) {
 399                 struct symbol *sym = top.sym_filter_entry->ms.sym;
 400                 name = sym->name;
 401         }
 402
 403         fprintf(stdout, "\nMapped keys:\n");
 404         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
 405         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
 406
 407         if (top.evlist->nr_entries > 1)
 408                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
 409
 410         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
 411
 412         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 413         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 414         fprintf(stdout, "\t[S]     stop annotation.\n");
 415
 416         fprintf(stdout,
 417                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 418                 top.hide_kernel_symbols ? "yes" : "no");
 419         fprintf(stdout,
 420                 "\t[U]     hide user symbols.               \t(%s)\n",
 421                 top.hide_user_symbols ? "yes" : "no");
 422         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
 423         fprintf(stdout, "\t[qQ]    quit.\n");
 424 }
 425
 426 static int key_mapped(int c)
 427 {
 428         switch (c) {
 429                 case 'd':
 430                 case 'e':
 431                 case 'f':
 432                 case 'z':
 433                 case 'q':
 434                 case 'Q':
 435                 case 'K':
 436                 case 'U':
 437                 case 'F':
 438                 case 's':
 439                 case 'S':
 440                         return 1;
 441                 case 'E':
 442                         return top.evlist->nr_entries > 1 ? 1 : 0;
 443                 default:
 444                         break;
 445         }
 446
 447         return 0;
 448 }
 449
 450 static void handle_keypress(int c)
 451 {
 452         if (!key_mapped(c)) {
 453                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 454                 struct termios tc, save;
 455
 456                 print_mapped_keys();
 457                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 458                 fflush(stdout);
 459
 460                 tcgetattr(0, &save);
 461                 tc = save;
 462                 tc.c_lflag &= ~(ICANON | ECHO);
 463                 tc.c_cc[VMIN] = 0;
 464                 tc.c_cc[VTIME] = 0;
 465                 tcsetattr(0, TCSANOW, &tc);
 466
 467                 poll(&stdin_poll, 1, -1);
 468                 c = getc(stdin);
 469
 470                 tcsetattr(0, TCSAFLUSH, &save);
 471                 if (!key_mapped(c))
 472                         return;
 473         }
 474
 475         switch (c) {
 476                 case 'd':
 477                         prompt_integer(&top.delay_secs, "Enter display delay");
 478                         if (top.delay_secs < 1)
 479                                 top.delay_secs = 1;
 480                         break;
 481                 case 'e':
 482                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
 483                         if (top.print_entries == 0) {
 484                                 sig_winch_handler(SIGWINCH);
 485                                 signal(SIGWINCH, sig_winch_handler);
 486                         } else
 487                                 signal(SIGWINCH, SIG_DFL);
 488                         break;
 489                 case 'E':
 490                         if (top.evlist->nr_entries > 1) {
 491                                 /* Select 0 as the default event: */
 492                                 int counter = 0;
 493
 494                                 fprintf(stderr, "\nAvailable events:");
 495
 496                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 497                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 498
 499                                 prompt_integer(&counter, "Enter details event counter");
 500
 501                                 if (counter >= top.evlist->nr_entries) {
 502                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 503                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 504                                         sleep(1);
 505                                         break;
 506                                 }
 507                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 508                                         if (top.sym_evsel->idx == counter)
 509                                                 break;
 510                         } else
 511                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 512                         break;
 513                 case 'f':
 514                         prompt_integer(&top.count_filter, "Enter display event count filter");
 515                         break;
 516                 case 'F':
 517                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 518                         break;
 519                 case 'K':
 520                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 521                         break;
 522                 case 'q':
 523                 case 'Q':
 524                         printf("exiting.\n");
 525                         if (dump_symtab)
 526                                 perf_session__fprintf_dsos(top.session, stderr);
 527                         exit(0);
 528                 case 's':
 529                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 530                         break;
 531                 case 'S':
 532                         if (!top.sym_filter_entry)
 533                                 break;
 534                         else {
 535                                 struct hist_entry *syme = top.sym_filter_entry;
 536
 537                                 top.sym_filter_entry = NULL;
 538                                 __zero_source_counters(syme);
 539                         }
 540                         break;
 541                 case 'U':
 542                         top.hide_user_symbols = !top.hide_user_symbols;
 543                         break;
 544                 case 'z':
 545                         top.zero = !top.zero;
 546                         break;
 547                 default:
 548                         break;
 549         }
 550 }
 551
 552 static void perf_top__sort_new_samples(void *arg)
 553 {
 554         struct perf_top *t = arg;
 555         perf_top__reset_sample_counters(t);
 556
 557         if (t->evlist->selected != NULL)
 558                 t->sym_evsel = t->evlist->selected;
 559
 560         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 561         hists__output_resort_threaded(&t->sym_evsel->hists);
 562         hists__decay_entries_threaded(&t->sym_evsel->hists,
 563                                       top.hide_user_symbols,
 564                                       top.hide_kernel_symbols);
 565 }
 566
 567 static void *display_thread_tui(void *arg __used)
 568 {
 569         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 570
 571         perf_top__sort_new_samples(&top);
 572         perf_evlist__tui_browse_hists(top.evlist, help,
 573                                       perf_top__sort_new_samples,
 574                                       &top, top.delay_secs);
 575
 576         exit_browser(0);
 577         exit(0);
 578         return NULL;
 579 }
 580
 581 static void *display_thread(void *arg __used)
 582 {
 583         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 584         struct termios tc, save;
 585         int delay_msecs, c;
 586
 587         tcgetattr(0, &save);
 588         tc = save;
 589         tc.c_lflag &= ~(ICANON | ECHO);
 590         tc.c_cc[VMIN] = 0;
 591         tc.c_cc[VTIME] = 0;
 592
 593         pthread__unblock_sigwinch();
 594 repeat:
 595         delay_msecs = top.delay_secs * 1000;
 596         tcsetattr(0, TCSANOW, &tc);
 597         /* trash return*/
 598         getc(stdin);
 599
 600         while (1) {
 601                 print_sym_table();
 602                 /*
 603                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 604                  * refresh screen in both cases.
 605                  */
 606                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 607                 case 0:
 608                         continue;
 609                 case -1:
 610                         if (errno == EINTR)
 611                                 continue;
 612                         /* Fall trhu */
 613                 default:
 614                         goto process_hotkey;
 615                 }
 616         }
 617 process_hotkey:
 618         c = getc(stdin);
 619         tcsetattr(0, TCSAFLUSH, &save);
 620
 621         handle_keypress(c);
 622         goto repeat;
 623
 624         return NULL;
 625 }
 626
 627 /* Tag samples to be skipped. */
 628 static const char *skip_symbols[] = {
 629         "default_idle",
 630         "native_safe_halt",
 631         "cpu_idle",
 632         "enter_idle",
 633         "exit_idle",
 634         "mwait_idle",
 635         "mwait_idle_with_hints",
 636         "poll_idle",
 637         "ppc64_runlatch_off",
 638         "pseries_dedicated_idle_sleep",
 639         NULL
 640 };
 641
 642 static int symbol_filter(struct map *map __used, struct symbol *sym)
 643 {
 644         const char *name = sym->name;
 645         int i;
 646
 647         /*
 648          * ppc64 uses function descriptors and appends a '.' to the
 649          * start of every instruction address. Remove it.
 650          */
 651         if (name[0] == '.')
 652                 name++;
 653
 654         if (!strcmp(name, "_text") ||
 655             !strcmp(name, "_etext") ||
 656             !strcmp(name, "_sinittext") ||
 657             !strncmp("init_module", name, 11) ||
 658             !strncmp("cleanup_module", name, 14) ||
 659             strstr(name, "_text_start") ||
 660             strstr(name, "_text_end"))
 661                 return 1;
 662
 663         for (i = 0; skip_symbols[i]; i++) {
 664                 if (!strcmp(skip_symbols[i], name)) {
 665                         sym->ignore = true;
 666                         break;
 667                 }
 668         }
 669
 670         return 0;
 671 }
 672
 673 static void perf_event__process_sample(const union perf_event *event,
 674                                        struct perf_evsel *evsel,
 675                                        struct perf_sample *sample,
 676                                        struct machine *machine)
 677 {
 678         struct symbol *parent = NULL;
 679         u64 ip = event->ip.ip;
 680         struct addr_location al;
 681         int err;
 682
 683         if (!machine && perf_guest) {
 684                 pr_err("Can't find guest [%d]'s kernel information\n",
 685                         event->ip.pid);
 686                 return;
 687         }
 688
 689         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 690                 top.exact_samples++;
 691
 692         if (perf_event__preprocess_sample(event, machine, &al, sample,
 693                                           symbol_filter) < 0 ||
 694             al.filtered)
 695                 return;
 696
 697         if (!kptr_restrict_warned &&
 698             symbol_conf.kptr_restrict &&
 699             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 700                 ui__warning(
 701 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 702 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 703 "Kernel%s samples will not be resolved.\n",
 704                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 705                           " modules" : "");
 706                 if (use_browser <= 0)
 707                         sleep(5);
 708                 kptr_restrict_warned = true;
 709         }
 710
 711         if (al.sym == NULL) {
 712                 const char *msg = "Kernel samples will not be resolved.\n";
 713                 /*
 714                  * As we do lazy loading of symtabs we only will know if the
 715                  * specified vmlinux file is invalid when we actually have a
 716                  * hit in kernel space and then try to load it. So if we get
 717                  * here and there are _no_ symbols in the DSO backing the
 718                  * kernel map, bail out.
 719                  *
 720                  * We may never get here, for instance, if we use -K/
 721                  * --hide-kernel-symbols, even if the user specifies an
 722                  * invalid --vmlinux ;-)
 723                  */
 724                 if (!kptr_restrict_warned && !vmlinux_warned &&
 725                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 726                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 727                         if (symbol_conf.vmlinux_name) {
 728                                 ui__warning("The %s file can't be used.\n%s",
 729                                             symbol_conf.vmlinux_name, msg);
 730                         } else {
 731                                 ui__warning("A vmlinux file was not found.\n%s",
 732                                             msg);
 733                         }
 734
 735                         if (use_browser <= 0)
 736                                 sleep(5);
 737                         vmlinux_warned = true;
 738                 }
 739         }
 740
 741         if (al.sym == NULL || !al.sym->ignore) {
 742                 struct hist_entry *he;
 743
 744                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 745                     sample->callchain) {
 746                         err = machine__resolve_callchain(machine, evsel, al.thread,
 747                                                          sample->callchain, &parent);
 748                         if (err)
 749                                 return;
 750                 }
 751
 752                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 753                 if (he == NULL) {
 754                         pr_err("Problem incrementing symbol period, skipping event\n");
 755                         return;
 756                 }
 757
 758                 if (symbol_conf.use_callchain) {
 759                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
 760                                                sample->period);
 761                         if (err)
 762                                 return;
 763                 }
 764
 765                 if (sort_has_symbols)
 766                         record_precise_ip(he, evsel->idx, ip);
 767         }
 768
 769         return;
 770 }
 771
 772 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
 773 {
 774         struct perf_sample sample;
 775         struct perf_evsel *evsel;
 776         union perf_event *event;
 777         struct machine *machine;
 778         u8 origin;
 779         int ret;
 780
 781         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
 782                 ret = perf_session__parse_sample(self, event, &sample);
 783                 if (ret) {
 784                         pr_err("Can't parse sample, err = %d\n", ret);
 785                         continue;
 786                 }
 787
 788                 evsel = perf_evlist__id2evsel(self->evlist, sample.id);
 789                 assert(evsel != NULL);
 790
 791                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 792
 793                 if (event->header.type == PERF_RECORD_SAMPLE)
 794                         ++top.samples;
 795
 796                 switch (origin) {
 797                 case PERF_RECORD_MISC_USER:
 798                         ++top.us_samples;
 799                         if (top.hide_user_symbols)
 800                                 continue;
 801                         machine = perf_session__find_host_machine(self);
 802                         break;
 803                 case PERF_RECORD_MISC_KERNEL:
 804                         ++top.kernel_samples;
 805                         if (top.hide_kernel_symbols)
 806                                 continue;
 807                         machine = perf_session__find_host_machine(self);
 808                         break;
 809                 case PERF_RECORD_MISC_GUEST_KERNEL:
 810                         ++top.guest_kernel_samples;
 811                         machine = perf_session__find_machine(self, event->ip.pid);
 812                         break;
 813                 case PERF_RECORD_MISC_GUEST_USER:
 814                         ++top.guest_us_samples;
 815                         /*
 816                          * TODO: we don't process guest user from host side
 817                          * except simple counting.
 818                          */
 819                         /* Fall thru */
 820                 default:
 821                         continue;
 822                 }
 823
 824
 825                 if (event->header.type == PERF_RECORD_SAMPLE)
 826                         perf_event__process_sample(event, evsel, &sample, machine);
 827                 else if (event->header.type < PERF_RECORD_MAX) {
 828                         hists__inc_nr_events(&evsel->hists, event->header.type);
 829                         perf_event__process(&top.ops, event, &sample, machine);
 830                 } else
 831                         ++self->hists.stats.nr_unknown_events;
 832         }
 833 }
 834
 835 static void perf_session__mmap_read(struct perf_session *self)
 836 {
 837         int i;
 838
 839         for (i = 0; i < top.evlist->nr_mmaps; i++)
 840                 perf_session__mmap_read_idx(self, i);
 841 }
 842
 843 static void start_counters(struct perf_evlist *evlist)
 844 {
 845         struct perf_evsel *counter, *first;
 846
 847         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 848
 849         list_for_each_entry(counter, &evlist->entries, node) {
 850                 struct perf_event_attr *attr = &counter->attr;
 851                 struct xyarray *group_fd = NULL;
 852
 853                 if (group && counter != first)
 854                         group_fd = first->fd;
 855
 856                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 857
 858                 if (top.freq) {
 859                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 860                         attr->freq        = 1;
 861                         attr->sample_freq = top.freq;
 862                 }
 863
 864                 if (evlist->nr_entries > 1) {
 865                         attr->sample_type |= PERF_SAMPLE_ID;
 866                         attr->read_format |= PERF_FORMAT_ID;
 867                 }
 868
 869                 if (symbol_conf.use_callchain)
 870                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
 871
 872                 attr->mmap = 1;
 873                 attr->comm = 1;
 874                 attr->inherit = inherit;
 875 retry_sample_id:
 876                 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
 877 try_again:
 878                 if (perf_evsel__open(counter, top.evlist->cpus,
 879                                      top.evlist->threads, group,
 880                                      group_fd) < 0) {
 881                         int err = errno;
 882
 883                         if (err == EPERM || err == EACCES) {
 884                                 ui__error_paranoid();
 885                                 goto out_err;
 886                         } else if (err == EINVAL && sample_id_all_avail) {
 887                                 /*
 888                                  * Old kernel, no attr->sample_id_type_all field
 889                                  */
 890                                 sample_id_all_avail = false;
 891                                 goto retry_sample_id;
 892                         }
 893                         /*
 894                          * If it's cycles then fall back to hrtimer
 895                          * based cpu-clock-tick sw counter, which
 896                          * is always available even if no PMU support:
 897                          */
 898                         if (attr->type == PERF_TYPE_HARDWARE &&
 899                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 900                                 if (verbose)
 901                                         ui__warning("Cycles event not supported,\n"
 902                                                     "trying to fall back to cpu-clock-ticks\n");
 903
 904                                 attr->type = PERF_TYPE_SOFTWARE;
 905                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 906                                 goto try_again;
 907                         }
 908
 909                         if (err == ENOENT) {
 910                                 ui__warning("The %s event is not supported.\n",
 911                                             event_name(counter));
 912                                 goto out_err;
 913                         }
 914
 915                         ui__warning("The sys_perf_event_open() syscall "
 916                                     "returned with %d (%s).  /bin/dmesg "
 917                                     "may provide additional information.\n"
 918                                     "No CONFIG_PERF_EVENTS=y kernel support "
 919                                     "configured?\n", err, strerror(err));
 920                         goto out_err;
 921                 }
 922         }
 923
 924         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
 925                 ui__warning("Failed to mmap with %d (%s)\n",
 926                             errno, strerror(errno));
 927                 goto out_err;
 928         }
 929
 930         return;
 931
 932 out_err:
 933         exit_browser(0);
 934         exit(0);
 935 }
 936
 937 static int setup_sample_type(void)
 938 {
 939         if (!sort_has_symbols) {
 940                 if (symbol_conf.use_callchain) {
 941                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
 942                         return -EINVAL;
 943                 }
 944         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
 945                 if (callchain_register_param(&callchain_param) < 0) {
 946                         ui__warning("Can't register callchain params.\n");
 947                         return -EINVAL;
 948                 }
 949         }
 950
 951         return 0;
 952 }
 953
 954 static int __cmd_top(void)
 955 {
 956         pthread_t thread;
 957         int ret;
 958         /*
 959          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 960          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 961          */
 962         top.session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 963         if (top.session == NULL)
 964                 return -ENOMEM;
 965
 966         ret = setup_sample_type();
 967         if (ret)
 968                 goto out_delete;
 969
 970         if (top.target_tid != -1)
 971                 perf_event__synthesize_thread_map(&top.ops, top.evlist->threads,
 972                                                   perf_event__process,
 973                                                   &top.session->host_machine);
 974         else
 975                 perf_event__synthesize_threads(&top.ops, perf_event__process,
 976                                                &top.session->host_machine);
 977         start_counters(top.evlist);
 978         top.session->evlist = top.evlist;
 979         perf_session__update_sample_type(top.session);
 980
 981         /* Wait for a minimal set of events before starting the snapshot */
 982         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 983
 984         perf_session__mmap_read(top.session);
 985
 986         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 987                                                              display_thread), NULL)) {
 988                 printf("Could not create display thread.\n");
 989                 exit(-1);
 990         }
 991
 992         if (realtime_prio) {
 993                 struct sched_param param;
 994
 995                 param.sched_priority = realtime_prio;
 996                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 997                         printf("Could not set realtime priority.\n");
 998                         exit(-1);
 999                 }
1000         }
1001
1002         while (1) {
1003                 u64 hits = top.samples;
1004
1005                 perf_session__mmap_read(top.session);
1006
1007                 if (hits == top.samples)
1008                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
1009         }
1010
1011 out_delete:
1012         perf_session__delete(top.session);
1013         top.session = NULL;
1014
1015         return 0;
1016 }
1017
1018 static int
1019 parse_callchain_opt(const struct option *opt __used, const char *arg,
1020                     int unset)
1021 {
1022         char *tok, *tok2;
1023         char *endptr;
1024
1025         /*
1026          * --no-call-graph
1027          */
1028         if (unset) {
1029                 dont_use_callchains = true;
1030                 return 0;
1031         }
1032
1033         symbol_conf.use_callchain = true;
1034
1035         if (!arg)
1036                 return 0;
1037
1038         tok = strtok((char *)arg, ",");
1039         if (!tok)
1040                 return -1;
1041
1042         /* get the output mode */
1043         if (!strncmp(tok, "graph", strlen(arg)))
1044                 callchain_param.mode = CHAIN_GRAPH_ABS;
1045
1046         else if (!strncmp(tok, "flat", strlen(arg)))
1047                 callchain_param.mode = CHAIN_FLAT;
1048
1049         else if (!strncmp(tok, "fractal", strlen(arg)))
1050                 callchain_param.mode = CHAIN_GRAPH_REL;
1051
1052         else if (!strncmp(tok, "none", strlen(arg))) {
1053                 callchain_param.mode = CHAIN_NONE;
1054                 symbol_conf.use_callchain = false;
1055
1056                 return 0;
1057         }
1058
1059         else
1060                 return -1;
1061
1062         /* get the min percentage */
1063         tok = strtok(NULL, ",");
1064         if (!tok)
1065                 goto setup;
1066
1067         callchain_param.min_percent = strtod(tok, &endptr);
1068         if (tok == endptr)
1069                 return -1;
1070
1071         /* get the print limit */
1072         tok2 = strtok(NULL, ",");
1073         if (!tok2)
1074                 goto setup;
1075
1076         if (tok2[0] != 'c') {
1077                 callchain_param.print_limit = strtod(tok2, &endptr);
1078                 tok2 = strtok(NULL, ",");
1079                 if (!tok2)
1080                         goto setup;
1081         }
1082
1083         /* get the call chain order */
1084         if (!strcmp(tok2, "caller"))
1085                 callchain_param.order = ORDER_CALLER;
1086         else if (!strcmp(tok2, "callee"))
1087                 callchain_param.order = ORDER_CALLEE;
1088         else
1089                 return -1;
1090 setup:
1091         if (callchain_register_param(&callchain_param) < 0) {
1092                 fprintf(stderr, "Can't register callchain params\n");
1093                 return -1;
1094         }
1095         return 0;
1096 }
1097
1098 static const char * const top_usage[] = {
1099         "perf top [<options>]",
1100         NULL
1101 };
1102
1103 static const struct option options[] = {
1104         OPT_CALLBACK('e', "event", &top.evlist, "event",
1105                      "event selector. use 'perf list' to list available events",
1106                      parse_events_option),
1107         OPT_INTEGER('c', "count", &default_interval,
1108                     "event period to sample"),
1109         OPT_INTEGER('p', "pid", &top.target_pid,
1110                     "profile events on existing process id"),
1111         OPT_INTEGER('t', "tid", &top.target_tid,
1112                     "profile events on existing thread id"),
1113         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1114                             "system-wide collection from all CPUs"),
1115         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1116                     "list of cpus to monitor"),
1117         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1118                    "file", "vmlinux pathname"),
1119         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1120                     "hide kernel symbols"),
1121         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1122         OPT_INTEGER('r', "realtime", &realtime_prio,
1123                     "collect data with this RT SCHED_FIFO priority"),
1124         OPT_INTEGER('d', "delay", &top.delay_secs,
1125                     "number of seconds to delay between refreshes"),
1126         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1127                             "dump the symbol table used for profiling"),
1128         OPT_INTEGER('f', "count-filter", &top.count_filter,
1129                     "only display functions with more events than this"),
1130         OPT_BOOLEAN('g', "group", &group,
1131                             "put the counters into a counter group"),
1132         OPT_BOOLEAN('i', "inherit", &inherit,
1133                     "child tasks inherit counters"),
1134         OPT_STRING(0, "sym-annotate", &sym_filter, "symbol name",
1135                     "symbol to annotate"),
1136         OPT_BOOLEAN('z', "zero", &top.zero,
1137                     "zero history across updates"),
1138         OPT_INTEGER('F', "freq", &top.freq,
1139                     "profile at this frequency"),
1140         OPT_INTEGER('E', "entries", &top.print_entries,
1141                     "display this many functions"),
1142         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1143                     "hide user symbols"),
1144         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1145         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1146         OPT_INCR('v', "verbose", &verbose,
1147                     "be more verbose (show counter open errors, etc)"),
1148         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1149                    "sort by key(s): pid, comm, dso, symbol, parent"),
1150         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1151                     "Show a column with the number of samples"),
1152         OPT_CALLBACK_DEFAULT('G', "call-graph", NULL, "output_type,min_percent, call_order",
1153                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1154                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1155                      callchain_default_opt),
1156         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1157                     "Show a column with the sum of periods"),
1158         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1159                    "only consider symbols in these dsos"),
1160         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1161                    "only consider symbols in these comms"),
1162         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1163                    "only consider these symbols"),
1164         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1165                     "Interleave source code with assembly code (default)"),
1166         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1167                     "Display raw encoding of assembly instructions (default)"),
1168         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1169                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1170         OPT_END()
1171 };
1172
1173 int cmd_top(int argc, const char **argv, const char *prefix __used)
1174 {
1175         struct perf_evsel *pos;
1176         int status = -ENOMEM;
1177
1178         top.evlist = perf_evlist__new(NULL, NULL);
1179         if (top.evlist == NULL)
1180                 return -ENOMEM;
1181
1182         symbol_conf.exclude_other = false;
1183
1184         argc = parse_options(argc, argv, options, top_usage, 0);
1185         if (argc)
1186                 usage_with_options(top_usage, options);
1187
1188         if (sort_order == default_sort_order)
1189                 sort_order = "dso,symbol";
1190
1191         setup_sorting(top_usage, options);
1192
1193         if (use_stdio)
1194                 use_browser = 0;
1195         else if (use_tui)
1196                 use_browser = 1;
1197
1198         setup_browser(false);
1199
1200         /* CPU and PID are mutually exclusive */
1201         if (top.target_tid > 0 && top.cpu_list) {
1202                 printf("WARNING: PID switch overriding CPU\n");
1203                 sleep(1);
1204                 top.cpu_list = NULL;
1205         }
1206
1207         if (top.target_pid != -1)
1208                 top.target_tid = top.target_pid;
1209
1210         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1211                                      top.target_tid, top.cpu_list) < 0)
1212                 usage_with_options(top_usage, options);
1213
1214         if (!top.evlist->nr_entries &&
1215             perf_evlist__add_default(top.evlist) < 0) {
1216                 pr_err("Not enough memory for event selector list\n");
1217                 return -ENOMEM;
1218         }
1219
1220         symbol_conf.nr_events = top.evlist->nr_entries;
1221
1222         if (top.delay_secs < 1)
1223                 top.delay_secs = 1;
1224
1225         /*
1226          * User specified count overrides default frequency.
1227          */
1228         if (default_interval)
1229                 top.freq = 0;
1230         else if (top.freq) {
1231                 default_interval = top.freq;
1232         } else {
1233                 fprintf(stderr, "frequency and count are zero, aborting\n");
1234                 exit(EXIT_FAILURE);
1235         }
1236
1237         list_for_each_entry(pos, &top.evlist->entries, node) {
1238                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1239                                          top.evlist->threads->nr) < 0)
1240                         goto out_free_fd;
1241                 /*
1242                  * Fill in the ones not specifically initialized via -c:
1243                  */
1244                 if (pos->attr.sample_period)
1245                         continue;
1246
1247                 pos->attr.sample_period = default_interval;
1248         }
1249
1250         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1251             perf_evlist__alloc_mmap(top.evlist) < 0)
1252                 goto out_free_fd;
1253
1254         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1255
1256         symbol_conf.priv_size = sizeof(struct annotation);
1257
1258         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1259         if (symbol__init() < 0)
1260                 return -1;
1261
1262         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1263         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1264         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1265
1266         /*
1267          * Avoid annotation data structures overhead when symbols aren't on the
1268          * sort list.
1269          */
1270         sort_has_symbols = sort_sym.list.next != NULL;
1271
1272         get_term_dimensions(&winsize);
1273         if (top.print_entries == 0) {
1274                 update_print_entries(&winsize);
1275                 signal(SIGWINCH, sig_winch_handler);
1276         }
1277
1278         status = __cmd_top();
1279 out_free_fd:
1280         perf_evlist__delete(top.evlist);
1281
1282         return status;
1283 }