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 unsigned int             mmap_pages                      =    128;
  93
  94 static bool                     dump_symtab                     =  false;
  95
  96 static struct winsize           winsize;
  97
  98 static const char               *sym_filter                     =   NULL;
  99 static int                      sym_pcnt_filter                 =      5;
 100
 101 /*
 102  * Source functions
 103  */
 104
 105 void get_term_dimensions(struct winsize *ws)
 106 {
 107         char *s = getenv("LINES");
 108
 109         if (s != NULL) {
 110                 ws->ws_row = atoi(s);
 111                 s = getenv("COLUMNS");
 112                 if (s != NULL) {
 113                         ws->ws_col = atoi(s);
 114                         if (ws->ws_row && ws->ws_col)
 115                                 return;
 116                 }
 117         }
 118 #ifdef TIOCGWINSZ
 119         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
 120             ws->ws_row && ws->ws_col)
 121                 return;
 122 #endif
 123         ws->ws_row = 25;
 124         ws->ws_col = 80;
 125 }
 126
 127 static void update_print_entries(struct winsize *ws)
 128 {
 129         top.print_entries = ws->ws_row;
 130
 131         if (top.print_entries > 9)
 132                 top.print_entries -= 9;
 133 }
 134
 135 static void sig_winch_handler(int sig __used)
 136 {
 137         get_term_dimensions(&winsize);
 138         update_print_entries(&winsize);
 139 }
 140
 141 static int parse_source(struct hist_entry *he)
 142 {
 143         struct symbol *sym;
 144         struct annotation *notes;
 145         struct map *map;
 146         int err = -1;
 147
 148         if (!he || !he->ms.sym)
 149                 return -1;
 150
 151         sym = he->ms.sym;
 152         map = he->ms.map;
 153
 154         /*
 155          * We can't annotate with just /proc/kallsyms
 156          */
 157         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
 158                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 159                        "path\n", sym->name);
 160                 sleep(1);
 161                 return -1;
 162         }
 163
 164         notes = symbol__annotation(sym);
 165         if (notes->src != NULL) {
 166                 pthread_mutex_lock(&notes->lock);
 167                 goto out_assign;
 168         }
 169
 170         pthread_mutex_lock(&notes->lock);
 171
 172         if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
 173                 pthread_mutex_unlock(&notes->lock);
 174                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 175                        sym->name);
 176                 sleep(1);
 177                 return err;
 178         }
 179
 180         err = symbol__annotate(sym, map, 0);
 181         if (err == 0) {
 182 out_assign:
 183                 top.sym_filter_entry = he;
 184         }
 185
 186         pthread_mutex_unlock(&notes->lock);
 187         return err;
 188 }
 189
 190 static void __zero_source_counters(struct hist_entry *he)
 191 {
 192         struct symbol *sym = he->ms.sym;
 193         symbol__annotate_zero_histograms(sym);
 194 }
 195
 196 static void record_precise_ip(struct hist_entry *he, int counter, u64 ip)
 197 {
 198         struct annotation *notes;
 199         struct symbol *sym;
 200
 201         if (he == NULL || he->ms.sym == NULL ||
 202             (he != top.sym_filter_entry && use_browser != 1))
 203                 return;
 204
 205         sym = he->ms.sym;
 206         notes = symbol__annotation(sym);
 207
 208         if (pthread_mutex_trylock(&notes->lock))
 209                 return;
 210
 211         if (notes->src == NULL &&
 212             symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
 213                 pthread_mutex_unlock(&notes->lock);
 214                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 215                        sym->name);
 216                 sleep(1);
 217                 return;
 218         }
 219
 220         ip = he->ms.map->map_ip(he->ms.map, ip);
 221         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 222
 223         pthread_mutex_unlock(&notes->lock);
 224 }
 225
 226 static void show_details(struct hist_entry *he)
 227 {
 228         struct annotation *notes;
 229         struct symbol *symbol;
 230         int more;
 231
 232         if (!he)
 233                 return;
 234
 235         symbol = he->ms.sym;
 236         notes = symbol__annotation(symbol);
 237
 238         pthread_mutex_lock(&notes->lock);
 239
 240         if (notes->src == NULL)
 241                 goto out_unlock;
 242
 243         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
 244         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
 245
 246         more = symbol__annotate_printf(symbol, he->ms.map, top.sym_evsel->idx,
 247                                        0, sym_pcnt_filter, top.print_entries, 4);
 248         if (top.zero)
 249                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
 250         else
 251                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
 252         if (more != 0)
 253                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 254 out_unlock:
 255         pthread_mutex_unlock(&notes->lock);
 256 }
 257
 258 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 259
 260 static struct hist_entry *
 261         perf_session__add_hist_entry(struct perf_session *session,
 262                                      struct addr_location *al,
 263                                      struct perf_sample *sample,
 264                                      struct perf_evsel *evsel)
 265 {
 266         struct hist_entry *he;
 267
 268         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
 269         if (he == NULL)
 270                 return NULL;
 271
 272         session->hists.stats.total_period += sample->period;
 273         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 274         return he;
 275 }
 276
 277 static void print_sym_table(void)
 278 {
 279         char bf[160];
 280         int printed = 0;
 281         const int win_width = winsize.ws_col - 1;
 282
 283         puts(CONSOLE_CLEAR);
 284
 285         perf_top__header_snprintf(&top, bf, sizeof(bf));
 286         printf("%s\n", bf);
 287
 288         perf_top__reset_sample_counters(&top);
 289
 290         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 291
 292         if (top.total_lost_warned != top.session->hists.stats.total_lost) {
 293                 top.total_lost_warned = top.session->hists.stats.total_lost;
 294                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
 295                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
 296                        top.total_lost_warned);
 297                 ++printed;
 298         }
 299
 300         if (top.sym_filter_entry) {
 301                 show_details(top.sym_filter_entry);
 302                 return;
 303         }
 304
 305         hists__collapse_resort_threaded(&top.sym_evsel->hists);
 306         hists__output_resort_threaded(&top.sym_evsel->hists);
 307         hists__decay_entries_threaded(&top.sym_evsel->hists);
 308         hists__output_recalc_col_len(&top.sym_evsel->hists, winsize.ws_row - 3);
 309         putchar('\n');
 310         hists__fprintf(&top.sym_evsel->hists, NULL, false, false,
 311                        winsize.ws_row - 4 - printed, win_width, stdout);
 312 }
 313
 314 static void prompt_integer(int *target, const char *msg)
 315 {
 316         char *buf = malloc(0), *p;
 317         size_t dummy = 0;
 318         int tmp;
 319
 320         fprintf(stdout, "\n%s: ", msg);
 321         if (getline(&buf, &dummy, stdin) < 0)
 322                 return;
 323
 324         p = strchr(buf, '\n');
 325         if (p)
 326                 *p = 0;
 327
 328         p = buf;
 329         while(*p) {
 330                 if (!isdigit(*p))
 331                         goto out_free;
 332                 p++;
 333         }
 334         tmp = strtoul(buf, NULL, 10);
 335         *target = tmp;
 336 out_free:
 337         free(buf);
 338 }
 339
 340 static void prompt_percent(int *target, const char *msg)
 341 {
 342         int tmp = 0;
 343
 344         prompt_integer(&tmp, msg);
 345         if (tmp >= 0 && tmp <= 100)
 346                 *target = tmp;
 347 }
 348
 349 static void prompt_symbol(struct hist_entry **target, const char *msg)
 350 {
 351         char *buf = malloc(0), *p;
 352         struct hist_entry *syme = *target, *n, *found = NULL;
 353         struct rb_node *next;
 354         size_t dummy = 0;
 355
 356         /* zero counters of active symbol */
 357         if (syme) {
 358                 __zero_source_counters(syme);
 359                 *target = NULL;
 360         }
 361
 362         fprintf(stdout, "\n%s: ", msg);
 363         if (getline(&buf, &dummy, stdin) < 0)
 364                 goto out_free;
 365
 366         p = strchr(buf, '\n');
 367         if (p)
 368                 *p = 0;
 369
 370         next = rb_first(&top.sym_evsel->hists.entries);
 371         while (next) {
 372                 n = rb_entry(next, struct hist_entry, rb_node);
 373                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 374                         found = n;
 375                         break;
 376                 }
 377                 next = rb_next(&n->rb_node);
 378         }
 379
 380         if (!found) {
 381                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 382                 sleep(1);
 383                 return;
 384         } else
 385                 parse_source(found);
 386
 387 out_free:
 388         free(buf);
 389 }
 390
 391 static void print_mapped_keys(void)
 392 {
 393         char *name = NULL;
 394
 395         if (top.sym_filter_entry) {
 396                 struct symbol *sym = top.sym_filter_entry->ms.sym;
 397                 name = sym->name;
 398         }
 399
 400         fprintf(stdout, "\nMapped keys:\n");
 401         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
 402         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
 403
 404         if (top.evlist->nr_entries > 1)
 405                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
 406
 407         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
 408
 409         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
 410         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 411         fprintf(stdout, "\t[S]     stop annotation.\n");
 412
 413         fprintf(stdout,
 414                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 415                 top.hide_kernel_symbols ? "yes" : "no");
 416         fprintf(stdout,
 417                 "\t[U]     hide user symbols.               \t(%s)\n",
 418                 top.hide_user_symbols ? "yes" : "no");
 419         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
 420         fprintf(stdout, "\t[qQ]    quit.\n");
 421 }
 422
 423 static int key_mapped(int c)
 424 {
 425         switch (c) {
 426                 case 'd':
 427                 case 'e':
 428                 case 'f':
 429                 case 'z':
 430                 case 'q':
 431                 case 'Q':
 432                 case 'K':
 433                 case 'U':
 434                 case 'F':
 435                 case 's':
 436                 case 'S':
 437                         return 1;
 438                 case 'E':
 439                 default:
 440                         break;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static void handle_keypress(int c)
 447 {
 448         if (!key_mapped(c)) {
 449                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 450                 struct termios tc, save;
 451
 452                 print_mapped_keys();
 453                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 454                 fflush(stdout);
 455
 456                 tcgetattr(0, &save);
 457                 tc = save;
 458                 tc.c_lflag &= ~(ICANON | ECHO);
 459                 tc.c_cc[VMIN] = 0;
 460                 tc.c_cc[VTIME] = 0;
 461                 tcsetattr(0, TCSANOW, &tc);
 462
 463                 poll(&stdin_poll, 1, -1);
 464                 c = getc(stdin);
 465
 466                 tcsetattr(0, TCSAFLUSH, &save);
 467                 if (!key_mapped(c))
 468                         return;
 469         }
 470
 471         switch (c) {
 472                 case 'd':
 473                         prompt_integer(&top.delay_secs, "Enter display delay");
 474                         if (top.delay_secs < 1)
 475                                 top.delay_secs = 1;
 476                         break;
 477                 case 'e':
 478                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
 479                         if (top.print_entries == 0) {
 480                                 sig_winch_handler(SIGWINCH);
 481                                 signal(SIGWINCH, sig_winch_handler);
 482                         } else
 483                                 signal(SIGWINCH, SIG_DFL);
 484                         break;
 485                 case 'E':
 486                         if (top.evlist->nr_entries > 1) {
 487                                 /* Select 0 as the default event: */
 488                                 int counter = 0;
 489
 490                                 fprintf(stderr, "\nAvailable events:");
 491
 492                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 493                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
 494
 495                                 prompt_integer(&counter, "Enter details event counter");
 496
 497                                 if (counter >= top.evlist->nr_entries) {
 498                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 499                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
 500                                         sleep(1);
 501                                         break;
 502                                 }
 503                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
 504                                         if (top.sym_evsel->idx == counter)
 505                                                 break;
 506                         } else
 507                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
 508                         break;
 509                 case 'f':
 510                         prompt_integer(&top.count_filter, "Enter display event count filter");
 511                         break;
 512                 case 'F':
 513                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
 514                         break;
 515                 case 'K':
 516                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
 517                         break;
 518                 case 'q':
 519                 case 'Q':
 520                         printf("exiting.\n");
 521                         if (dump_symtab)
 522                                 perf_session__fprintf_dsos(top.session, stderr);
 523                         exit(0);
 524                 case 's':
 525                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
 526                         break;
 527                 case 'S':
 528                         if (!top.sym_filter_entry)
 529                                 break;
 530                         else {
 531                                 struct hist_entry *syme = top.sym_filter_entry;
 532
 533                                 top.sym_filter_entry = NULL;
 534                                 __zero_source_counters(syme);
 535                         }
 536                         break;
 537                 case 'U':
 538                         top.hide_user_symbols = !top.hide_user_symbols;
 539                         break;
 540                 case 'z':
 541                         top.zero = !top.zero;
 542                         break;
 543                 default:
 544                         break;
 545         }
 546 }
 547
 548 static void perf_top__sort_new_samples(void *arg)
 549 {
 550         struct perf_top *t = arg;
 551         perf_top__reset_sample_counters(t);
 552
 553         if (t->evlist->selected != NULL)
 554                 t->sym_evsel = t->evlist->selected;
 555
 556         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 557         hists__output_resort_threaded(&t->sym_evsel->hists);
 558         hists__decay_entries_threaded(&t->sym_evsel->hists);
 559         hists__output_recalc_col_len(&t->sym_evsel->hists, winsize.ws_row - 3);
 560 }
 561
 562 static void *display_thread_tui(void *arg __used)
 563 {
 564         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 565
 566         perf_top__sort_new_samples(&top);
 567         perf_evlist__tui_browse_hists(top.evlist, help,
 568                                       perf_top__sort_new_samples,
 569                                       &top, top.delay_secs);
 570
 571         exit_browser(0);
 572         exit(0);
 573         return NULL;
 574 }
 575
 576 static void *display_thread(void *arg __used)
 577 {
 578         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 579         struct termios tc, save;
 580         int delay_msecs, c;
 581
 582         tcgetattr(0, &save);
 583         tc = save;
 584         tc.c_lflag &= ~(ICANON | ECHO);
 585         tc.c_cc[VMIN] = 0;
 586         tc.c_cc[VTIME] = 0;
 587
 588         pthread__unblock_sigwinch();
 589 repeat:
 590         delay_msecs = top.delay_secs * 1000;
 591         tcsetattr(0, TCSANOW, &tc);
 592         /* trash return*/
 593         getc(stdin);
 594
 595         while (1) {
 596                 print_sym_table();
 597                 /*
 598                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 599                  * refresh screen in both cases.
 600                  */
 601                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 602                 case 0:
 603                         continue;
 604                 case -1:
 605                         if (errno == EINTR)
 606                                 continue;
 607                         /* Fall trhu */
 608                 default:
 609                         goto process_hotkey;
 610                 }
 611         }
 612 process_hotkey:
 613         c = getc(stdin);
 614         tcsetattr(0, TCSAFLUSH, &save);
 615
 616         handle_keypress(c);
 617         goto repeat;
 618
 619         return NULL;
 620 }
 621
 622 /* Tag samples to be skipped. */
 623 static const char *skip_symbols[] = {
 624         "default_idle",
 625         "native_safe_halt",
 626         "cpu_idle",
 627         "enter_idle",
 628         "exit_idle",
 629         "mwait_idle",
 630         "mwait_idle_with_hints",
 631         "poll_idle",
 632         "ppc64_runlatch_off",
 633         "pseries_dedicated_idle_sleep",
 634         NULL
 635 };
 636
 637 static int symbol_filter(struct map *map __used, struct symbol *sym)
 638 {
 639         const char *name = sym->name;
 640         int i;
 641
 642         /*
 643          * ppc64 uses function descriptors and appends a '.' to the
 644          * start of every instruction address. Remove it.
 645          */
 646         if (name[0] == '.')
 647                 name++;
 648
 649         if (!strcmp(name, "_text") ||
 650             !strcmp(name, "_etext") ||
 651             !strcmp(name, "_sinittext") ||
 652             !strncmp("init_module", name, 11) ||
 653             !strncmp("cleanup_module", name, 14) ||
 654             strstr(name, "_text_start") ||
 655             strstr(name, "_text_end"))
 656                 return 1;
 657
 658         for (i = 0; skip_symbols[i]; i++) {
 659                 if (!strcmp(skip_symbols[i], name)) {
 660                         sym->ignore = true;
 661                         break;
 662                 }
 663         }
 664
 665         return 0;
 666 }
 667
 668 static void perf_event__process_sample(const union perf_event *event,
 669                                        struct perf_sample *sample,
 670                                        struct perf_session *session)
 671 {
 672         struct symbol *parent = NULL;
 673         u64 ip = event->ip.ip;
 674         struct addr_location al;
 675         struct machine *machine;
 676         int err;
 677         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 678
 679         ++top.samples;
 680
 681         switch (origin) {
 682         case PERF_RECORD_MISC_USER:
 683                 ++top.us_samples;
 684                 if (top.hide_user_symbols)
 685                         return;
 686                 machine = perf_session__find_host_machine(session);
 687                 break;
 688         case PERF_RECORD_MISC_KERNEL:
 689                 ++top.kernel_samples;
 690                 if (top.hide_kernel_symbols)
 691                         return;
 692                 machine = perf_session__find_host_machine(session);
 693                 break;
 694         case PERF_RECORD_MISC_GUEST_KERNEL:
 695                 ++top.guest_kernel_samples;
 696                 machine = perf_session__find_machine(session, event->ip.pid);
 697                 break;
 698         case PERF_RECORD_MISC_GUEST_USER:
 699                 ++top.guest_us_samples;
 700                 /*
 701                  * TODO: we don't process guest user from host side
 702                  * except simple counting.
 703                  */
 704                 return;
 705         default:
 706                 return;
 707         }
 708
 709         if (!machine && perf_guest) {
 710                 pr_err("Can't find guest [%d]'s kernel information\n",
 711                         event->ip.pid);
 712                 return;
 713         }
 714
 715         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 716                 top.exact_samples++;
 717
 718         if (perf_event__preprocess_sample(event, session, &al, sample,
 719                                           symbol_filter) < 0 ||
 720             al.filtered)
 721                 return;
 722
 723         if (!kptr_restrict_warned &&
 724             symbol_conf.kptr_restrict &&
 725             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 726                 ui__warning(
 727 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 728 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 729 "Kernel%s samples will not be resolved.\n",
 730                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 731                           " modules" : "");
 732                 if (use_browser <= 0)
 733                         sleep(5);
 734                 kptr_restrict_warned = true;
 735         }
 736
 737         if (al.sym == NULL) {
 738                 const char *msg = "Kernel samples will not be resolved.\n";
 739                 /*
 740                  * As we do lazy loading of symtabs we only will know if the
 741                  * specified vmlinux file is invalid when we actually have a
 742                  * hit in kernel space and then try to load it. So if we get
 743                  * here and there are _no_ symbols in the DSO backing the
 744                  * kernel map, bail out.
 745                  *
 746                  * We may never get here, for instance, if we use -K/
 747                  * --hide-kernel-symbols, even if the user specifies an
 748                  * invalid --vmlinux ;-)
 749                  */
 750                 if (!kptr_restrict_warned && !vmlinux_warned &&
 751                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 752                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 753                         if (symbol_conf.vmlinux_name) {
 754                                 ui__warning("The %s file can't be used.\n%s",
 755                                             symbol_conf.vmlinux_name, msg);
 756                         } else {
 757                                 ui__warning("A vmlinux file was not found.\n%s",
 758                                             msg);
 759                         }
 760
 761                         if (use_browser <= 0)
 762                                 sleep(5);
 763                         vmlinux_warned = true;
 764                 }
 765         }
 766
 767         if (al.sym == NULL || !al.sym->ignore) {
 768                 struct perf_evsel *evsel;
 769                 struct hist_entry *he;
 770
 771                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
 772                 assert(evsel != NULL);
 773
 774                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 775                     sample->callchain) {
 776                         err = perf_session__resolve_callchain(session, al.thread,
 777                                                               sample->callchain, &parent);
 778                         if (err)
 779                                 return;
 780                 }
 781
 782                 he = perf_session__add_hist_entry(session, &al, sample, evsel);
 783                 if (he == NULL) {
 784                         pr_err("Problem incrementing symbol period, skipping event\n");
 785                         return;
 786                 }
 787
 788                 if (symbol_conf.use_callchain) {
 789                         err = callchain_append(he->callchain, &session->callchain_cursor,
 790                                                sample->period);
 791                         if (err)
 792                                 return;
 793                 }
 794
 795                 if (sort_has_symbols)
 796                         record_precise_ip(he, evsel->idx, ip);
 797         }
 798
 799         return;
 800 }
 801
 802 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
 803 {
 804         struct perf_sample sample;
 805         union perf_event *event;
 806         int ret;
 807
 808         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
 809                 ret = perf_session__parse_sample(self, event, &sample);
 810                 if (ret) {
 811                         pr_err("Can't parse sample, err = %d\n", ret);
 812                         continue;
 813                 }
 814
 815                 if (event->header.type == PERF_RECORD_SAMPLE)
 816                         perf_event__process_sample(event, &sample, self);
 817                 else
 818                         perf_event__process(event, &sample, self);
 819         }
 820 }
 821
 822 static void perf_session__mmap_read(struct perf_session *self)
 823 {
 824         int i;
 825
 826         for (i = 0; i < top.evlist->nr_mmaps; i++)
 827                 perf_session__mmap_read_idx(self, i);
 828 }
 829
 830 static void start_counters(struct perf_evlist *evlist)
 831 {
 832         struct perf_evsel *counter;
 833
 834         list_for_each_entry(counter, &evlist->entries, node) {
 835                 struct perf_event_attr *attr = &counter->attr;
 836
 837                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 838
 839                 if (top.freq) {
 840                         attr->sample_type |= PERF_SAMPLE_PERIOD;
 841                         attr->freq        = 1;
 842                         attr->sample_freq = top.freq;
 843                 }
 844
 845                 if (evlist->nr_entries > 1) {
 846                         attr->sample_type |= PERF_SAMPLE_ID;
 847                         attr->read_format |= PERF_FORMAT_ID;
 848                 }
 849
 850                 if (symbol_conf.use_callchain)
 851                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
 852
 853                 attr->mmap = 1;
 854                 attr->comm = 1;
 855                 attr->inherit = inherit;
 856 try_again:
 857                 if (perf_evsel__open(counter, top.evlist->cpus,
 858                                      top.evlist->threads, group) < 0) {
 859                         int err = errno;
 860
 861                         if (err == EPERM || err == EACCES) {
 862                                 ui__warning_paranoid();
 863                                 goto out_err;
 864                         }
 865                         /*
 866                          * If it's cycles then fall back to hrtimer
 867                          * based cpu-clock-tick sw counter, which
 868                          * is always available even if no PMU support:
 869                          */
 870                         if (attr->type == PERF_TYPE_HARDWARE &&
 871                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 872                                 if (verbose)
 873                                         ui__warning("Cycles event not supported,\n"
 874                                                     "trying to fall back to cpu-clock-ticks\n");
 875
 876                                 attr->type = PERF_TYPE_SOFTWARE;
 877                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 878                                 goto try_again;
 879                         }
 880
 881                         if (err == ENOENT) {
 882                                 ui__warning("The %s event is not supported.\n",
 883                                             event_name(counter));
 884                                 goto out_err;
 885                         }
 886
 887                         ui__warning("The sys_perf_event_open() syscall "
 888                                     "returned with %d (%s).  /bin/dmesg "
 889                                     "may provide additional information.\n"
 890                                     "No CONFIG_PERF_EVENTS=y kernel support "
 891                                     "configured?\n", err, strerror(err));
 892                         goto out_err;
 893                 }
 894         }
 895
 896         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
 897                 ui__warning("Failed to mmap with %d (%s)\n",
 898                             errno, strerror(errno));
 899                 goto out_err;
 900         }
 901
 902         return;
 903
 904 out_err:
 905         exit_browser(0);
 906         exit(0);
 907 }
 908
 909 static int setup_sample_type(void)
 910 {
 911         if (!sort_has_symbols) {
 912                 if (symbol_conf.use_callchain) {
 913                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
 914                         return -EINVAL;
 915                 }
 916         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
 917                 if (callchain_register_param(&callchain_param) < 0) {
 918                         ui__warning("Can't register callchain params.\n");
 919                         return -EINVAL;
 920                 }
 921         }
 922
 923         return 0;
 924 }
 925
 926 static int __cmd_top(void)
 927 {
 928         pthread_t thread;
 929         int ret;
 930         /*
 931          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 932          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 933          */
 934         top.session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 935         if (top.session == NULL)
 936                 return -ENOMEM;
 937
 938         ret = setup_sample_type();
 939         if (ret)
 940                 goto out_delete;
 941
 942         if (top.target_tid != -1)
 943                 perf_event__synthesize_thread_map(top.evlist->threads,
 944                                                   perf_event__process, top.session);
 945         else
 946                 perf_event__synthesize_threads(perf_event__process, top.session);
 947
 948         start_counters(top.evlist);
 949         top.session->evlist = top.evlist;
 950         perf_session__update_sample_type(top.session);
 951
 952         /* Wait for a minimal set of events before starting the snapshot */
 953         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 954
 955         perf_session__mmap_read(top.session);
 956
 957         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 958                                                              display_thread), NULL)) {
 959                 printf("Could not create display thread.\n");
 960                 exit(-1);
 961         }
 962
 963         if (realtime_prio) {
 964                 struct sched_param param;
 965
 966                 param.sched_priority = realtime_prio;
 967                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 968                         printf("Could not set realtime priority.\n");
 969                         exit(-1);
 970                 }
 971         }
 972
 973         while (1) {
 974                 u64 hits = top.samples;
 975
 976                 perf_session__mmap_read(top.session);
 977
 978                 if (hits == top.samples)
 979                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
 980         }
 981
 982 out_delete:
 983         perf_session__delete(top.session);
 984         top.session = NULL;
 985
 986         return 0;
 987 }
 988
 989 static int
 990 parse_callchain_opt(const struct option *opt __used, const char *arg,
 991                     int unset)
 992 {
 993         char *tok, *tok2;
 994         char *endptr;
 995
 996         /*
 997          * --no-call-graph
 998          */
 999         if (unset) {
1000                 dont_use_callchains = true;
1001                 return 0;
1002         }
1003
1004         symbol_conf.use_callchain = true;
1005
1006         if (!arg)
1007                 return 0;
1008
1009         tok = strtok((char *)arg, ",");
1010         if (!tok)
1011                 return -1;
1012
1013         /* get the output mode */
1014         if (!strncmp(tok, "graph", strlen(arg)))
1015                 callchain_param.mode = CHAIN_GRAPH_ABS;
1016
1017         else if (!strncmp(tok, "flat", strlen(arg)))
1018                 callchain_param.mode = CHAIN_FLAT;
1019
1020         else if (!strncmp(tok, "fractal", strlen(arg)))
1021                 callchain_param.mode = CHAIN_GRAPH_REL;
1022
1023         else if (!strncmp(tok, "none", strlen(arg))) {
1024                 callchain_param.mode = CHAIN_NONE;
1025                 symbol_conf.use_callchain = false;
1026
1027                 return 0;
1028         }
1029
1030         else
1031                 return -1;
1032
1033         /* get the min percentage */
1034         tok = strtok(NULL, ",");
1035         if (!tok)
1036                 goto setup;
1037
1038         callchain_param.min_percent = strtod(tok, &endptr);
1039         if (tok == endptr)
1040                 return -1;
1041
1042         /* get the print limit */
1043         tok2 = strtok(NULL, ",");
1044         if (!tok2)
1045                 goto setup;
1046
1047         if (tok2[0] != 'c') {
1048                 callchain_param.print_limit = strtod(tok2, &endptr);
1049                 tok2 = strtok(NULL, ",");
1050                 if (!tok2)
1051                         goto setup;
1052         }
1053
1054         /* get the call chain order */
1055         if (!strcmp(tok2, "caller"))
1056                 callchain_param.order = ORDER_CALLER;
1057         else if (!strcmp(tok2, "callee"))
1058                 callchain_param.order = ORDER_CALLEE;
1059         else
1060                 return -1;
1061 setup:
1062         if (callchain_register_param(&callchain_param) < 0) {
1063                 fprintf(stderr, "Can't register callchain params\n");
1064                 return -1;
1065         }
1066         return 0;
1067 }
1068
1069 static const char * const top_usage[] = {
1070         "perf top [<options>]",
1071         NULL
1072 };
1073
1074 static const struct option options[] = {
1075         OPT_CALLBACK('e', "event", &top.evlist, "event",
1076                      "event selector. use 'perf list' to list available events",
1077                      parse_events_option),
1078         OPT_INTEGER('c', "count", &default_interval,
1079                     "event period to sample"),
1080         OPT_INTEGER('p', "pid", &top.target_pid,
1081                     "profile events on existing process id"),
1082         OPT_INTEGER('t', "tid", &top.target_tid,
1083                     "profile events on existing thread id"),
1084         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1085                             "system-wide collection from all CPUs"),
1086         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1087                     "list of cpus to monitor"),
1088         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1089                    "file", "vmlinux pathname"),
1090         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1091                     "hide kernel symbols"),
1092         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1093         OPT_INTEGER('r', "realtime", &realtime_prio,
1094                     "collect data with this RT SCHED_FIFO priority"),
1095         OPT_INTEGER('d', "delay", &top.delay_secs,
1096                     "number of seconds to delay between refreshes"),
1097         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1098                             "dump the symbol table used for profiling"),
1099         OPT_INTEGER('f', "count-filter", &top.count_filter,
1100                     "only display functions with more events than this"),
1101         OPT_BOOLEAN('g', "group", &group,
1102                             "put the counters into a counter group"),
1103         OPT_BOOLEAN('i', "inherit", &inherit,
1104                     "child tasks inherit counters"),
1105         OPT_STRING(0, "sym-annotate", &sym_filter, "symbol name",
1106                     "symbol to annotate"),
1107         OPT_BOOLEAN('z', "zero", &top.zero,
1108                     "zero history across updates"),
1109         OPT_INTEGER('F', "freq", &top.freq,
1110                     "profile at this frequency"),
1111         OPT_INTEGER('E', "entries", &top.print_entries,
1112                     "display this many functions"),
1113         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1114                     "hide user symbols"),
1115         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1116         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1117         OPT_INCR('v', "verbose", &verbose,
1118                     "be more verbose (show counter open errors, etc)"),
1119         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1120                    "sort by key(s): pid, comm, dso, symbol, parent"),
1121         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1122                     "Show a column with the number of samples"),
1123         OPT_CALLBACK_DEFAULT('G', "call-graph", NULL, "output_type,min_percent, call_order",
1124                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1125                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1126                      callchain_default_opt),
1127         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1128                     "Show a column with the sum of periods"),
1129         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1130                    "only consider symbols in these dsos"),
1131         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1132                    "only consider symbols in these comms"),
1133         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1134                    "only consider these symbols"),
1135         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1136                     "Interleave source code with assembly code (default)"),
1137         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1138                     "Display raw encoding of assembly instructions (default)"),
1139         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1140                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1141         OPT_END()
1142 };
1143
1144 int cmd_top(int argc, const char **argv, const char *prefix __used)
1145 {
1146         struct perf_evsel *pos;
1147         int status = -ENOMEM;
1148
1149         top.evlist = perf_evlist__new(NULL, NULL);
1150         if (top.evlist == NULL)
1151                 return -ENOMEM;
1152
1153         symbol_conf.exclude_other = false;
1154
1155         argc = parse_options(argc, argv, options, top_usage, 0);
1156         if (argc)
1157                 usage_with_options(top_usage, options);
1158
1159         if (sort_order == default_sort_order)
1160                 sort_order = "dso,symbol";
1161
1162         setup_sorting(top_usage, options);
1163
1164         if (use_stdio)
1165                 use_browser = 0;
1166         else if (use_tui)
1167                 use_browser = 1;
1168
1169         setup_browser(false);
1170
1171         /* CPU and PID are mutually exclusive */
1172         if (top.target_tid > 0 && top.cpu_list) {
1173                 printf("WARNING: PID switch overriding CPU\n");
1174                 sleep(1);
1175                 top.cpu_list = NULL;
1176         }
1177
1178         if (top.target_pid != -1)
1179                 top.target_tid = top.target_pid;
1180
1181         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1182                                      top.target_tid, top.cpu_list) < 0)
1183                 usage_with_options(top_usage, options);
1184
1185         if (!top.evlist->nr_entries &&
1186             perf_evlist__add_default(top.evlist) < 0) {
1187                 pr_err("Not enough memory for event selector list\n");
1188                 return -ENOMEM;
1189         }
1190
1191         if (top.delay_secs < 1)
1192                 top.delay_secs = 1;
1193
1194         /*
1195          * User specified count overrides default frequency.
1196          */
1197         if (default_interval)
1198                 top.freq = 0;
1199         else if (top.freq) {
1200                 default_interval = top.freq;
1201         } else {
1202                 fprintf(stderr, "frequency and count are zero, aborting\n");
1203                 exit(EXIT_FAILURE);
1204         }
1205
1206         list_for_each_entry(pos, &top.evlist->entries, node) {
1207                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1208                                          top.evlist->threads->nr) < 0)
1209                         goto out_free_fd;
1210                 /*
1211                  * Fill in the ones not specifically initialized via -c:
1212                  */
1213                 if (pos->attr.sample_period)
1214                         continue;
1215
1216                 pos->attr.sample_period = default_interval;
1217         }
1218
1219         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1220             perf_evlist__alloc_mmap(top.evlist) < 0)
1221                 goto out_free_fd;
1222
1223         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1224
1225         symbol_conf.priv_size = sizeof(struct annotation);
1226
1227         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1228         if (symbol__init() < 0)
1229                 return -1;
1230
1231         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1232         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1233         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1234
1235         /*
1236          * Avoid annotation data structures overhead when symbols aren't on the
1237          * sort list.
1238          */
1239         sort_has_symbols = sort_sym.list.next != NULL;
1240
1241         get_term_dimensions(&winsize);
1242         if (top.print_entries == 0) {
1243                 update_print_entries(&winsize);
1244                 signal(SIGWINCH, sig_winch_handler);
1245         }
1246
1247         status = __cmd_top();
1248 out_free_fd:
1249         perf_evlist__delete(top.evlist);
1250
1251         return status;
1252 }