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/machine.h"
  30 #include "util/session.h"
  31 #include "util/symbol.h"
  32 #include "util/thread.h"
  33 #include "util/thread_map.h"
  34 #include "util/top.h"
  35 #include "util/util.h"
  36 #include <linux/rbtree.h>
  37 #include "util/parse-options.h"
  38 #include "util/parse-events.h"
  39 #include "util/cpumap.h"
  40 #include "util/xyarray.h"
  41 #include "util/sort.h"
  42 #include "util/intlist.h"
  43 #include "util/parse-branch-options.h"
  44 #include "arch/common.h"
  45
  46 #include "util/debug.h"
  47
  48 #include <assert.h>
  49 #include <elf.h>
  50 #include <fcntl.h>
  51
  52 #include <stdio.h>
  53 #include <termios.h>
  54 #include <unistd.h>
  55 #include <inttypes.h>
  56
  57 #include <errno.h>
  58 #include <time.h>
  59 #include <sched.h>
  60
  61 #include <sys/syscall.h>
  62 #include <sys/ioctl.h>
  63 #include <poll.h>
  64 #include <sys/prctl.h>
  65 #include <sys/wait.h>
  66 #include <sys/uio.h>
  67 #include <sys/utsname.h>
  68 #include <sys/mman.h>
  69
  70 #include <linux/types.h>
  71
  72 static volatile int done;
  73
  74 #define HEADER_LINE_NR  5
  75
  76 static void perf_top__update_print_entries(struct perf_top *top)
  77 {
  78         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  79 }
  80
  81 static void perf_top__sig_winch(int sig __maybe_unused,
  82                                 siginfo_t *info __maybe_unused, void *arg)
  83 {
  84         struct perf_top *top = arg;
  85
  86         get_term_dimensions(&top->winsize);
  87         perf_top__update_print_entries(top);
  88 }
  89
  90 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  91 {
  92         struct symbol *sym;
  93         struct annotation *notes;
  94         struct map *map;
  95         int err = -1;
  96
  97         if (!he || !he->ms.sym)
  98                 return -1;
  99
 100         sym = he->ms.sym;
 101         map = he->ms.map;
 102
 103         /*
 104          * We can't annotate with just /proc/kallsyms
 105          */
 106         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 107             !dso__is_kcore(map->dso)) {
 108                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 109                        "path\n", sym->name);
 110                 sleep(1);
 111                 return -1;
 112         }
 113
 114         notes = symbol__annotation(sym);
 115         if (notes->src != NULL) {
 116                 pthread_mutex_lock(&notes->lock);
 117                 goto out_assign;
 118         }
 119
 120         pthread_mutex_lock(&notes->lock);
 121
 122         if (symbol__alloc_hist(sym) < 0) {
 123                 pthread_mutex_unlock(&notes->lock);
 124                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 125                        sym->name);
 126                 sleep(1);
 127                 return err;
 128         }
 129
 130         err = symbol__annotate(sym, map, 0);
 131         if (err == 0) {
 132 out_assign:
 133                 top->sym_filter_entry = he;
 134         }
 135
 136         pthread_mutex_unlock(&notes->lock);
 137         return err;
 138 }
 139
 140 static void __zero_source_counters(struct hist_entry *he)
 141 {
 142         struct symbol *sym = he->ms.sym;
 143         symbol__annotate_zero_histograms(sym);
 144 }
 145
 146 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 147 {
 148         struct utsname uts;
 149         int err = uname(&uts);
 150
 151         ui__warning("Out of bounds address found:\n\n"
 152                     "Addr:   %" PRIx64 "\n"
 153                     "DSO:    %s %c\n"
 154                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 155                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 156                     "Arch:   %s\n"
 157                     "Kernel: %s\n"
 158                     "Tools:  %s\n\n"
 159                     "Not all samples will be on the annotation output.\n\n"
 160                     "Please report to linux-kernel@vger.kernel.org\n",
 161                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 162                     map->start, map->end, sym->start, sym->end,
 163                     sym->binding == STB_GLOBAL ? 'g' :
 164                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 165                     err ? "[unknown]" : uts.machine,
 166                     err ? "[unknown]" : uts.release, perf_version_string);
 167         if (use_browser <= 0)
 168                 sleep(5);
 169
 170         map->erange_warned = true;
 171 }
 172
 173 static void perf_top__record_precise_ip(struct perf_top *top,
 174                                         struct hist_entry *he,
 175                                         int counter, u64 ip)
 176 {
 177         struct annotation *notes;
 178         struct symbol *sym;
 179         int err = 0;
 180
 181         if (he == NULL || he->ms.sym == NULL ||
 182             ((top->sym_filter_entry == NULL ||
 183               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 184                 return;
 185
 186         sym = he->ms.sym;
 187         notes = symbol__annotation(sym);
 188
 189         if (pthread_mutex_trylock(&notes->lock))
 190                 return;
 191
 192         ip = he->ms.map->map_ip(he->ms.map, ip);
 193
 194         if (ui__has_annotation())
 195                 err = hist_entry__inc_addr_samples(he, counter, ip);
 196
 197         pthread_mutex_unlock(&notes->lock);
 198
 199         /*
 200          * This function is now called with he->hists->lock held.
 201          * Release it before going to sleep.
 202          */
 203         pthread_mutex_unlock(&he->hists->lock);
 204
 205         if (err == -ERANGE && !he->ms.map->erange_warned)
 206                 ui__warn_map_erange(he->ms.map, sym, ip);
 207         else if (err == -ENOMEM) {
 208                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 209                        sym->name);
 210                 sleep(1);
 211         }
 212
 213         pthread_mutex_lock(&he->hists->lock);
 214 }
 215
 216 static void perf_top__show_details(struct perf_top *top)
 217 {
 218         struct hist_entry *he = top->sym_filter_entry;
 219         struct annotation *notes;
 220         struct symbol *symbol;
 221         int more;
 222
 223         if (!he)
 224                 return;
 225
 226         symbol = he->ms.sym;
 227         notes = symbol__annotation(symbol);
 228
 229         pthread_mutex_lock(&notes->lock);
 230
 231         if (notes->src == NULL)
 232                 goto out_unlock;
 233
 234         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 235         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 236
 237         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 238                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 239
 240         if (top->evlist->enabled) {
 241                 if (top->zero)
 242                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 243                 else
 244                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 245         }
 246         if (more != 0)
 247                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 248 out_unlock:
 249         pthread_mutex_unlock(&notes->lock);
 250 }
 251
 252 static void perf_top__print_sym_table(struct perf_top *top)
 253 {
 254         char bf[160];
 255         int printed = 0;
 256         const int win_width = top->winsize.ws_col - 1;
 257         struct hists *hists = evsel__hists(top->sym_evsel);
 258
 259         puts(CONSOLE_CLEAR);
 260
 261         perf_top__header_snprintf(top, bf, sizeof(bf));
 262         printf("%s\n", bf);
 263
 264         perf_top__reset_sample_counters(top);
 265
 266         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 267
 268         if (hists->stats.nr_lost_warned !=
 269             hists->stats.nr_events[PERF_RECORD_LOST]) {
 270                 hists->stats.nr_lost_warned =
 271                               hists->stats.nr_events[PERF_RECORD_LOST];
 272                 color_fprintf(stdout, PERF_COLOR_RED,
 273                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 274                               hists->stats.nr_lost_warned);
 275                 ++printed;
 276         }
 277
 278         if (top->sym_filter_entry) {
 279                 perf_top__show_details(top);
 280                 return;
 281         }
 282
 283         if (top->evlist->enabled) {
 284                 if (top->zero) {
 285                         hists__delete_entries(hists);
 286                 } else {
 287                         hists__decay_entries(hists, top->hide_user_symbols,
 288                                              top->hide_kernel_symbols);
 289                 }
 290         }
 291
 292         hists__collapse_resort(hists, NULL);
 293         hists__output_resort(hists, NULL);
 294
 295         hists__output_recalc_col_len(hists, top->print_entries - printed);
 296         putchar('\n');
 297         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 298                        top->min_percent, stdout);
 299 }
 300
 301 static void prompt_integer(int *target, const char *msg)
 302 {
 303         char *buf = malloc(0), *p;
 304         size_t dummy = 0;
 305         int tmp;
 306
 307         fprintf(stdout, "\n%s: ", msg);
 308         if (getline(&buf, &dummy, stdin) < 0)
 309                 return;
 310
 311         p = strchr(buf, '\n');
 312         if (p)
 313                 *p = 0;
 314
 315         p = buf;
 316         while(*p) {
 317                 if (!isdigit(*p))
 318                         goto out_free;
 319                 p++;
 320         }
 321         tmp = strtoul(buf, NULL, 10);
 322         *target = tmp;
 323 out_free:
 324         free(buf);
 325 }
 326
 327 static void prompt_percent(int *target, const char *msg)
 328 {
 329         int tmp = 0;
 330
 331         prompt_integer(&tmp, msg);
 332         if (tmp >= 0 && tmp <= 100)
 333                 *target = tmp;
 334 }
 335
 336 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 337 {
 338         char *buf = malloc(0), *p;
 339         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 340         struct hists *hists = evsel__hists(top->sym_evsel);
 341         struct rb_node *next;
 342         size_t dummy = 0;
 343
 344         /* zero counters of active symbol */
 345         if (syme) {
 346                 __zero_source_counters(syme);
 347                 top->sym_filter_entry = NULL;
 348         }
 349
 350         fprintf(stdout, "\n%s: ", msg);
 351         if (getline(&buf, &dummy, stdin) < 0)
 352                 goto out_free;
 353
 354         p = strchr(buf, '\n');
 355         if (p)
 356                 *p = 0;
 357
 358         next = rb_first(&hists->entries);
 359         while (next) {
 360                 n = rb_entry(next, struct hist_entry, rb_node);
 361                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 362                         found = n;
 363                         break;
 364                 }
 365                 next = rb_next(&n->rb_node);
 366         }
 367
 368         if (!found) {
 369                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 370                 sleep(1);
 371         } else
 372                 perf_top__parse_source(top, found);
 373
 374 out_free:
 375         free(buf);
 376 }
 377
 378 static void perf_top__print_mapped_keys(struct perf_top *top)
 379 {
 380         char *name = NULL;
 381
 382         if (top->sym_filter_entry) {
 383                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 384                 name = sym->name;
 385         }
 386
 387         fprintf(stdout, "\nMapped keys:\n");
 388         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 389         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 390
 391         if (top->evlist->nr_entries > 1)
 392                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 393
 394         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 395
 396         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 397         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 398         fprintf(stdout, "\t[S]     stop annotation.\n");
 399
 400         fprintf(stdout,
 401                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 402                 top->hide_kernel_symbols ? "yes" : "no");
 403         fprintf(stdout,
 404                 "\t[U]     hide user symbols.               \t(%s)\n",
 405                 top->hide_user_symbols ? "yes" : "no");
 406         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 407         fprintf(stdout, "\t[qQ]    quit.\n");
 408 }
 409
 410 static int perf_top__key_mapped(struct perf_top *top, int c)
 411 {
 412         switch (c) {
 413                 case 'd':
 414                 case 'e':
 415                 case 'f':
 416                 case 'z':
 417                 case 'q':
 418                 case 'Q':
 419                 case 'K':
 420                 case 'U':
 421                 case 'F':
 422                 case 's':
 423                 case 'S':
 424                         return 1;
 425                 case 'E':
 426                         return top->evlist->nr_entries > 1 ? 1 : 0;
 427                 default:
 428                         break;
 429         }
 430
 431         return 0;
 432 }
 433
 434 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 435 {
 436         bool ret = true;
 437
 438         if (!perf_top__key_mapped(top, c)) {
 439                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 440                 struct termios save;
 441
 442                 perf_top__print_mapped_keys(top);
 443                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 444                 fflush(stdout);
 445
 446                 set_term_quiet_input(&save);
 447
 448                 poll(&stdin_poll, 1, -1);
 449                 c = getc(stdin);
 450
 451                 tcsetattr(0, TCSAFLUSH, &save);
 452                 if (!perf_top__key_mapped(top, c))
 453                         return ret;
 454         }
 455
 456         switch (c) {
 457                 case 'd':
 458                         prompt_integer(&top->delay_secs, "Enter display delay");
 459                         if (top->delay_secs < 1)
 460                                 top->delay_secs = 1;
 461                         break;
 462                 case 'e':
 463                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 464                         if (top->print_entries == 0) {
 465                                 struct sigaction act = {
 466                                         .sa_sigaction = perf_top__sig_winch,
 467                                         .sa_flags     = SA_SIGINFO,
 468                                 };
 469                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 470                                 sigaction(SIGWINCH, &act, NULL);
 471                         } else {
 472                                 signal(SIGWINCH, SIG_DFL);
 473                         }
 474                         break;
 475                 case 'E':
 476                         if (top->evlist->nr_entries > 1) {
 477                                 /* Select 0 as the default event: */
 478                                 int counter = 0;
 479
 480                                 fprintf(stderr, "\nAvailable events:");
 481
 482                                 evlist__for_each(top->evlist, top->sym_evsel)
 483                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 484
 485                                 prompt_integer(&counter, "Enter details event counter");
 486
 487                                 if (counter >= top->evlist->nr_entries) {
 488                                         top->sym_evsel = perf_evlist__first(top->evlist);
 489                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 490                                         sleep(1);
 491                                         break;
 492                                 }
 493                                 evlist__for_each(top->evlist, top->sym_evsel)
 494                                         if (top->sym_evsel->idx == counter)
 495                                                 break;
 496                         } else
 497                                 top->sym_evsel = perf_evlist__first(top->evlist);
 498                         break;
 499                 case 'f':
 500                         prompt_integer(&top->count_filter, "Enter display event count filter");
 501                         break;
 502                 case 'F':
 503                         prompt_percent(&top->sym_pcnt_filter,
 504                                        "Enter details display event filter (percent)");
 505                         break;
 506                 case 'K':
 507                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 508                         break;
 509                 case 'q':
 510                 case 'Q':
 511                         printf("exiting.\n");
 512                         if (top->dump_symtab)
 513                                 perf_session__fprintf_dsos(top->session, stderr);
 514                         ret = false;
 515                         break;
 516                 case 's':
 517                         perf_top__prompt_symbol(top, "Enter details symbol");
 518                         break;
 519                 case 'S':
 520                         if (!top->sym_filter_entry)
 521                                 break;
 522                         else {
 523                                 struct hist_entry *syme = top->sym_filter_entry;
 524
 525                                 top->sym_filter_entry = NULL;
 526                                 __zero_source_counters(syme);
 527                         }
 528                         break;
 529                 case 'U':
 530                         top->hide_user_symbols = !top->hide_user_symbols;
 531                         break;
 532                 case 'z':
 533                         top->zero = !top->zero;
 534                         break;
 535                 default:
 536                         break;
 537         }
 538
 539         return ret;
 540 }
 541
 542 static void perf_top__sort_new_samples(void *arg)
 543 {
 544         struct perf_top *t = arg;
 545         struct hists *hists;
 546
 547         perf_top__reset_sample_counters(t);
 548
 549         if (t->evlist->selected != NULL)
 550                 t->sym_evsel = t->evlist->selected;
 551
 552         hists = evsel__hists(t->sym_evsel);
 553
 554         if (t->evlist->enabled) {
 555                 if (t->zero) {
 556                         hists__delete_entries(hists);
 557                 } else {
 558                         hists__decay_entries(hists, t->hide_user_symbols,
 559                                              t->hide_kernel_symbols);
 560                 }
 561         }
 562
 563         hists__collapse_resort(hists, NULL);
 564         hists__output_resort(hists, NULL);
 565 }
 566
 567 static void *display_thread_tui(void *arg)
 568 {
 569         struct perf_evsel *pos;
 570         struct perf_top *top = arg;
 571         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 572         struct hist_browser_timer hbt = {
 573                 .timer          = perf_top__sort_new_samples,
 574                 .arg            = top,
 575                 .refresh        = top->delay_secs,
 576         };
 577
 578         perf_top__sort_new_samples(top);
 579
 580         /*
 581          * Initialize the uid_filter_str, in the future the TUI will allow
 582          * Zooming in/out UIDs. For now juse use whatever the user passed
 583          * via --uid.
 584          */
 585         evlist__for_each(top->evlist, pos) {
 586                 struct hists *hists = evsel__hists(pos);
 587                 hists->uid_filter_str = top->record_opts.target.uid_str;
 588         }
 589
 590         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 591                                       top->min_percent,
 592                                       &top->session->header.env);
 593
 594         done = 1;
 595         return NULL;
 596 }
 597
 598 static void display_sig(int sig __maybe_unused)
 599 {
 600         done = 1;
 601 }
 602
 603 static void display_setup_sig(void)
 604 {
 605         signal(SIGSEGV, sighandler_dump_stack);
 606         signal(SIGFPE, sighandler_dump_stack);
 607         signal(SIGINT,  display_sig);
 608         signal(SIGQUIT, display_sig);
 609         signal(SIGTERM, display_sig);
 610 }
 611
 612 static void *display_thread(void *arg)
 613 {
 614         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 615         struct termios save;
 616         struct perf_top *top = arg;
 617         int delay_msecs, c;
 618
 619         display_setup_sig();
 620         pthread__unblock_sigwinch();
 621 repeat:
 622         delay_msecs = top->delay_secs * 1000;
 623         set_term_quiet_input(&save);
 624         /* trash return*/
 625         getc(stdin);
 626
 627         while (!done) {
 628                 perf_top__print_sym_table(top);
 629                 /*
 630                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 631                  * refresh screen in both cases.
 632                  */
 633                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 634                 case 0:
 635                         continue;
 636                 case -1:
 637                         if (errno == EINTR)
 638                                 continue;
 639                         /* Fall trhu */
 640                 default:
 641                         c = getc(stdin);
 642                         tcsetattr(0, TCSAFLUSH, &save);
 643
 644                         if (perf_top__handle_keypress(top, c))
 645                                 goto repeat;
 646                         done = 1;
 647                 }
 648         }
 649
 650         tcsetattr(0, TCSAFLUSH, &save);
 651         return NULL;
 652 }
 653
 654 static int symbol_filter(struct map *map, struct symbol *sym)
 655 {
 656         const char *name = sym->name;
 657
 658         if (!__map__is_kernel(map))
 659                 return 0;
 660         /*
 661          * ppc64 uses function descriptors and appends a '.' to the
 662          * start of every instruction address. Remove it.
 663          */
 664         if (name[0] == '.')
 665                 name++;
 666
 667         if (!strcmp(name, "_text") ||
 668             !strcmp(name, "_etext") ||
 669             !strcmp(name, "_sinittext") ||
 670             !strncmp("init_module", name, 11) ||
 671             !strncmp("cleanup_module", name, 14) ||
 672             strstr(name, "_text_start") ||
 673             strstr(name, "_text_end"))
 674                 return 1;
 675
 676         if (symbol__is_idle(sym))
 677                 sym->ignore = true;
 678
 679         return 0;
 680 }
 681
 682 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 683                                    struct addr_location *al, bool single,
 684                                    void *arg)
 685 {
 686         struct perf_top *top = arg;
 687         struct hist_entry *he = iter->he;
 688         struct perf_evsel *evsel = iter->evsel;
 689
 690         if (sort__has_sym && single) {
 691                 u64 ip = al->addr;
 692
 693                 if (al->map)
 694                         ip = al->map->unmap_ip(al->map, ip);
 695
 696                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 697         }
 698
 699         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 700                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 701         return 0;
 702 }
 703
 704 static void perf_event__process_sample(struct perf_tool *tool,
 705                                        const union perf_event *event,
 706                                        struct perf_evsel *evsel,
 707                                        struct perf_sample *sample,
 708                                        struct machine *machine)
 709 {
 710         struct perf_top *top = container_of(tool, struct perf_top, tool);
 711         struct addr_location al;
 712         int err;
 713
 714         if (!machine && perf_guest) {
 715                 static struct intlist *seen;
 716
 717                 if (!seen)
 718                         seen = intlist__new(NULL);
 719
 720                 if (!intlist__has_entry(seen, sample->pid)) {
 721                         pr_err("Can't find guest [%d]'s kernel information\n",
 722                                 sample->pid);
 723                         intlist__add(seen, sample->pid);
 724                 }
 725                 return;
 726         }
 727
 728         if (!machine) {
 729                 pr_err("%u unprocessable samples recorded.\r",
 730                        top->session->evlist->stats.nr_unprocessable_samples++);
 731                 return;
 732         }
 733
 734         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 735                 top->exact_samples++;
 736
 737         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 738                 return;
 739
 740         if (!top->kptr_restrict_warned &&
 741             symbol_conf.kptr_restrict &&
 742             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 743                 ui__warning(
 744 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 745 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 746 "Kernel%s samples will not be resolved.\n",
 747                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 748                           " modules" : "");
 749                 if (use_browser <= 0)
 750                         sleep(5);
 751                 top->kptr_restrict_warned = true;
 752         }
 753
 754         if (al.sym == NULL) {
 755                 const char *msg = "Kernel samples will not be resolved.\n";
 756                 /*
 757                  * As we do lazy loading of symtabs we only will know if the
 758                  * specified vmlinux file is invalid when we actually have a
 759                  * hit in kernel space and then try to load it. So if we get
 760                  * here and there are _no_ symbols in the DSO backing the
 761                  * kernel map, bail out.
 762                  *
 763                  * We may never get here, for instance, if we use -K/
 764                  * --hide-kernel-symbols, even if the user specifies an
 765                  * invalid --vmlinux ;-)
 766                  */
 767                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 768                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 769                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 770                         if (symbol_conf.vmlinux_name) {
 771                                 char serr[256];
 772                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 773                                 ui__warning("The %s file can't be used: %s\n%s",
 774                                             symbol_conf.vmlinux_name, serr, msg);
 775                         } else {
 776                                 ui__warning("A vmlinux file was not found.\n%s",
 777                                             msg);
 778                         }
 779
 780                         if (use_browser <= 0)
 781                                 sleep(5);
 782                         top->vmlinux_warned = true;
 783                 }
 784         }
 785
 786         if (al.sym == NULL || !al.sym->ignore) {
 787                 struct hists *hists = evsel__hists(evsel);
 788                 struct hist_entry_iter iter = {
 789                         .evsel          = evsel,
 790                         .sample         = sample,
 791                         .add_entry_cb   = hist_iter__top_callback,
 792                 };
 793
 794                 if (symbol_conf.cumulate_callchain)
 795                         iter.ops = &hist_iter_cumulative;
 796                 else
 797                         iter.ops = &hist_iter_normal;
 798
 799                 pthread_mutex_lock(&hists->lock);
 800
 801                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 802                 if (err < 0)
 803                         pr_err("Problem incrementing symbol period, skipping event\n");
 804
 805                 pthread_mutex_unlock(&hists->lock);
 806         }
 807
 808         addr_location__put(&al);
 809 }
 810
 811 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 812 {
 813         struct perf_sample sample;
 814         struct perf_evsel *evsel;
 815         struct perf_session *session = top->session;
 816         union perf_event *event;
 817         struct machine *machine;
 818         u8 origin;
 819         int ret;
 820
 821         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 822                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 823                 if (ret) {
 824                         pr_err("Can't parse sample, err = %d\n", ret);
 825                         goto next_event;
 826                 }
 827
 828                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 829                 assert(evsel != NULL);
 830
 831                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 832
 833                 if (event->header.type == PERF_RECORD_SAMPLE)
 834                         ++top->samples;
 835
 836                 switch (origin) {
 837                 case PERF_RECORD_MISC_USER:
 838                         ++top->us_samples;
 839                         if (top->hide_user_symbols)
 840                                 goto next_event;
 841                         machine = &session->machines.host;
 842                         break;
 843                 case PERF_RECORD_MISC_KERNEL:
 844                         ++top->kernel_samples;
 845                         if (top->hide_kernel_symbols)
 846                                 goto next_event;
 847                         machine = &session->machines.host;
 848                         break;
 849                 case PERF_RECORD_MISC_GUEST_KERNEL:
 850                         ++top->guest_kernel_samples;
 851                         machine = perf_session__find_machine(session,
 852                                                              sample.pid);
 853                         break;
 854                 case PERF_RECORD_MISC_GUEST_USER:
 855                         ++top->guest_us_samples;
 856                         /*
 857                          * TODO: we don't process guest user from host side
 858                          * except simple counting.
 859                          */
 860                         goto next_event;
 861                 default:
 862                         if (event->header.type == PERF_RECORD_SAMPLE)
 863                                 goto next_event;
 864                         machine = &session->machines.host;
 865                         break;
 866                 }
 867
 868
 869                 if (event->header.type == PERF_RECORD_SAMPLE) {
 870                         perf_event__process_sample(&top->tool, event, evsel,
 871                                                    &sample, machine);
 872                 } else if (event->header.type < PERF_RECORD_MAX) {
 873                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 874                         machine__process_event(machine, event, &sample);
 875                 } else
 876                         ++session->evlist->stats.nr_unknown_events;
 877 next_event:
 878                 perf_evlist__mmap_consume(top->evlist, idx);
 879         }
 880 }
 881
 882 static void perf_top__mmap_read(struct perf_top *top)
 883 {
 884         int i;
 885
 886         for (i = 0; i < top->evlist->nr_mmaps; i++)
 887                 perf_top__mmap_read_idx(top, i);
 888 }
 889
 890 static int perf_top__start_counters(struct perf_top *top)
 891 {
 892         char msg[512];
 893         struct perf_evsel *counter;
 894         struct perf_evlist *evlist = top->evlist;
 895         struct record_opts *opts = &top->record_opts;
 896
 897         perf_evlist__config(evlist, opts);
 898
 899         evlist__for_each(evlist, counter) {
 900 try_again:
 901                 if (perf_evsel__open(counter, top->evlist->cpus,
 902                                      top->evlist->threads) < 0) {
 903                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 904                                 if (verbose)
 905                                         ui__warning("%s\n", msg);
 906                                 goto try_again;
 907                         }
 908
 909                         perf_evsel__open_strerror(counter, &opts->target,
 910                                                   errno, msg, sizeof(msg));
 911                         ui__error("%s\n", msg);
 912                         goto out_err;
 913                 }
 914         }
 915
 916         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 917                 ui__error("Failed to mmap with %d (%s)\n",
 918                             errno, strerror_r(errno, msg, sizeof(msg)));
 919                 goto out_err;
 920         }
 921
 922         return 0;
 923
 924 out_err:
 925         return -1;
 926 }
 927
 928 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 929 {
 930         if (!sort__has_sym) {
 931                 if (symbol_conf.use_callchain) {
 932                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 933                         return -EINVAL;
 934                 }
 935         } else if (callchain_param.mode != CHAIN_NONE) {
 936                 if (callchain_register_param(&callchain_param) < 0) {
 937                         ui__error("Can't register callchain params.\n");
 938                         return -EINVAL;
 939                 }
 940         }
 941
 942         return 0;
 943 }
 944
 945 static int __cmd_top(struct perf_top *top)
 946 {
 947         struct record_opts *opts = &top->record_opts;
 948         pthread_t thread;
 949         int ret;
 950
 951         top->session = perf_session__new(NULL, false, NULL);
 952         if (top->session == NULL)
 953                 return -1;
 954
 955         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 956
 957         if (!objdump_path) {
 958                 ret = perf_env__lookup_objdump(&top->session->header.env);
 959                 if (ret)
 960                         goto out_delete;
 961         }
 962
 963         ret = perf_top__setup_sample_type(top);
 964         if (ret)
 965                 goto out_delete;
 966
 967         if (perf_session__register_idle_thread(top->session) == NULL)
 968                 goto out_delete;
 969
 970         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 971                                     top->evlist->threads, false, opts->proc_map_timeout);
 972
 973         if (sort__has_socket) {
 974                 ret = perf_env__read_cpu_topology_map(&perf_env);
 975                 if (ret < 0)
 976                         goto out_err_cpu_topo;
 977         }
 978
 979         ret = perf_top__start_counters(top);
 980         if (ret)
 981                 goto out_delete;
 982
 983         top->session->evlist = top->evlist;
 984         perf_session__set_id_hdr_size(top->session);
 985
 986         /*
 987          * When perf is starting the traced process, all the events (apart from
 988          * group members) have enable_on_exec=1 set, so don't spoil it by
 989          * prematurely enabling them.
 990          *
 991          * XXX 'top' still doesn't start workloads like record, trace, but should,
 992          * so leave the check here.
 993          */
 994         if (!target__none(&opts->target))
 995                 perf_evlist__enable(top->evlist);
 996
 997         /* Wait for a minimal set of events before starting the snapshot */
 998         perf_evlist__poll(top->evlist, 100);
 999
1000         perf_top__mmap_read(top);
1001
1002         ret = -1;
1003         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1004                                                             display_thread), top)) {
1005                 ui__error("Could not create display thread.\n");
1006                 goto out_delete;
1007         }
1008
1009         if (top->realtime_prio) {
1010                 struct sched_param param;
1011
1012                 param.sched_priority = top->realtime_prio;
1013                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1014                         ui__error("Could not set realtime priority.\n");
1015                         goto out_join;
1016                 }
1017         }
1018
1019         while (!done) {
1020                 u64 hits = top->samples;
1021
1022                 perf_top__mmap_read(top);
1023
1024                 if (hits == top->samples)
1025                         ret = perf_evlist__poll(top->evlist, 100);
1026         }
1027
1028         ret = 0;
1029 out_join:
1030         pthread_join(thread, NULL);
1031 out_delete:
1032         perf_session__delete(top->session);
1033         top->session = NULL;
1034
1035         return ret;
1036
1037 out_err_cpu_topo: {
1038         char errbuf[BUFSIZ];
1039         const char *err = strerror_r(-ret, errbuf, sizeof(errbuf));
1040
1041         ui__error("Could not read the CPU topology map: %s\n", err);
1042         goto out_delete;
1043 }
1044 }
1045
1046 static int
1047 callchain_opt(const struct option *opt, const char *arg, int unset)
1048 {
1049         symbol_conf.use_callchain = true;
1050         return record_callchain_opt(opt, arg, unset);
1051 }
1052
1053 static int
1054 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1055 {
1056         symbol_conf.use_callchain = true;
1057         return record_parse_callchain_opt(opt, arg, unset);
1058 }
1059
1060 static int perf_top_config(const char *var, const char *value, void *cb)
1061 {
1062         if (!strcmp(var, "top.call-graph"))
1063                 var = "call-graph.record-mode"; /* fall-through */
1064         if (!strcmp(var, "top.children")) {
1065                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1066                 return 0;
1067         }
1068
1069         return perf_default_config(var, value, cb);
1070 }
1071
1072 static int
1073 parse_percent_limit(const struct option *opt, const char *arg,
1074                     int unset __maybe_unused)
1075 {
1076         struct perf_top *top = opt->value;
1077
1078         top->min_percent = strtof(arg, NULL);
1079         return 0;
1080 }
1081
1082 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1083 {
1084         char errbuf[BUFSIZ];
1085         struct perf_top top = {
1086                 .count_filter        = 5,
1087                 .delay_secs          = 2,
1088                 .record_opts = {
1089                         .mmap_pages     = UINT_MAX,
1090                         .user_freq      = UINT_MAX,
1091                         .user_interval  = ULLONG_MAX,
1092                         .freq           = 4000, /* 4 KHz */
1093                         .target         = {
1094                                 .uses_mmap   = true,
1095                         },
1096                         .proc_map_timeout    = 500,
1097                 },
1098                 .max_stack           = PERF_MAX_STACK_DEPTH,
1099                 .sym_pcnt_filter     = 5,
1100         };
1101         struct record_opts *opts = &top.record_opts;
1102         struct target *target = &opts->target;
1103         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_U64('c', "count", &opts->user_interval, "event period to sample"),
1108         OPT_STRING('p', "pid", &target->pid, "pid",
1109                     "profile events on existing process id"),
1110         OPT_STRING('t', "tid", &target->tid, "tid",
1111                     "profile events on existing thread id"),
1112         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1113                             "system-wide collection from all CPUs"),
1114         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1115                     "list of cpus to monitor"),
1116         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1117                    "file", "vmlinux pathname"),
1118         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1119                     "don't load vmlinux even if found"),
1120         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1121                     "hide kernel symbols"),
1122         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1123                      "number of mmap data pages",
1124                      perf_evlist__parse_mmap_pages),
1125         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1126                     "collect data with this RT SCHED_FIFO priority"),
1127         OPT_INTEGER('d', "delay", &top.delay_secs,
1128                     "number of seconds to delay between refreshes"),
1129         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1130                             "dump the symbol table used for profiling"),
1131         OPT_INTEGER('f', "count-filter", &top.count_filter,
1132                     "only display functions with more events than this"),
1133         OPT_BOOLEAN(0, "group", &opts->group,
1134                             "put the counters into a counter group"),
1135         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1136                     "child tasks do not inherit counters"),
1137         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1138                     "symbol to annotate"),
1139         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1140         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1141         OPT_INTEGER('E', "entries", &top.print_entries,
1142                     "display this many functions"),
1143         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1144                     "hide user symbols"),
1145         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1146         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1147         OPT_INCR('v', "verbose", &verbose,
1148                     "be more verbose (show counter open errors, etc)"),
1149         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1150                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1151                    " Please refer the man page for the complete list."),
1152         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1153                    "output field(s): overhead, period, sample plus all of sort keys"),
1154         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1155                     "Show a column with the number of samples"),
1156         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1157                            NULL, "enables call-graph recording",
1158                            &callchain_opt),
1159         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1160                      "mode[,dump_size]", record_callchain_help,
1161                      &parse_callchain_opt),
1162         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1163                     "Accumulate callchains of children and show total overhead as well"),
1164         OPT_INTEGER(0, "max-stack", &top.max_stack,
1165                     "Set the maximum stack depth when parsing the callchain. "
1166                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1167         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1168                    "ignore callees of these functions in call graphs",
1169                    report_parse_ignore_callees_opt),
1170         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1171                     "Show a column with the sum of periods"),
1172         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1173                    "only consider symbols in these dsos"),
1174         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1175                    "only consider symbols in these comms"),
1176         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1177                    "only consider these symbols"),
1178         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1179                     "Interleave source code with assembly code (default)"),
1180         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1181                     "Display raw encoding of assembly instructions (default)"),
1182         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1183                     "Enable kernel symbol demangling"),
1184         OPT_STRING(0, "objdump", &objdump_path, "path",
1185                     "objdump binary to use for disassembly and annotations"),
1186         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1187                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1188         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1189         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1190                      "Don't show entries under that percent", parse_percent_limit),
1191         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1192                      "How to display percentage of filtered entries", parse_filter_percentage),
1193         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1194                    "width[,width...]",
1195                    "don't try to adjust column width, use these fixed values"),
1196         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1197                         "per thread proc mmap processing timeout in ms"),
1198         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1199                      "branch any", "sample any taken branches",
1200                      parse_branch_stack),
1201         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1202                      "branch filter mask", "branch stack filter modes",
1203                      parse_branch_stack),
1204         OPT_END()
1205         };
1206         const char * const top_usage[] = {
1207                 "perf top [<options>]",
1208                 NULL
1209         };
1210         int status = hists__init();
1211
1212         if (status < 0)
1213                 return status;
1214
1215         top.evlist = perf_evlist__new();
1216         if (top.evlist == NULL)
1217                 return -ENOMEM;
1218
1219         perf_config(perf_top_config, &top);
1220
1221         argc = parse_options(argc, argv, options, top_usage, 0);
1222         if (argc)
1223                 usage_with_options(top_usage, options);
1224
1225         sort__mode = SORT_MODE__TOP;
1226         /* display thread wants entries to be collapsed in a different tree */
1227         sort__need_collapse = 1;
1228
1229         if (setup_sorting() < 0) {
1230                 if (sort_order)
1231                         parse_options_usage(top_usage, options, "s", 1);
1232                 if (field_order)
1233                         parse_options_usage(sort_order ? NULL : top_usage,
1234                                             options, "fields", 0);
1235                 goto out_delete_evlist;
1236         }
1237
1238         if (top.use_stdio)
1239                 use_browser = 0;
1240         else if (top.use_tui)
1241                 use_browser = 1;
1242
1243         setup_browser(false);
1244
1245         status = target__validate(target);
1246         if (status) {
1247                 target__strerror(target, status, errbuf, BUFSIZ);
1248                 ui__warning("%s\n", errbuf);
1249         }
1250
1251         status = target__parse_uid(target);
1252         if (status) {
1253                 int saved_errno = errno;
1254
1255                 target__strerror(target, status, errbuf, BUFSIZ);
1256                 ui__error("%s\n", errbuf);
1257
1258                 status = -saved_errno;
1259                 goto out_delete_evlist;
1260         }
1261
1262         if (target__none(target))
1263                 target->system_wide = true;
1264
1265         if (perf_evlist__create_maps(top.evlist, target) < 0)
1266                 usage_with_options(top_usage, options);
1267
1268         if (!top.evlist->nr_entries &&
1269             perf_evlist__add_default(top.evlist) < 0) {
1270                 ui__error("Not enough memory for event selector list\n");
1271                 goto out_delete_evlist;
1272         }
1273
1274         symbol_conf.nr_events = top.evlist->nr_entries;
1275
1276         if (top.delay_secs < 1)
1277                 top.delay_secs = 1;
1278
1279         if (record_opts__config(opts)) {
1280                 status = -EINVAL;
1281                 goto out_delete_evlist;
1282         }
1283
1284         top.sym_evsel = perf_evlist__first(top.evlist);
1285
1286         if (!symbol_conf.use_callchain) {
1287                 symbol_conf.cumulate_callchain = false;
1288                 perf_hpp__cancel_cumulate();
1289         }
1290
1291         symbol_conf.priv_size = sizeof(struct annotation);
1292
1293         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1294         if (symbol__init(NULL) < 0)
1295                 return -1;
1296
1297         sort__setup_elide(stdout);
1298
1299         get_term_dimensions(&top.winsize);
1300         if (top.print_entries == 0) {
1301                 struct sigaction act = {
1302                         .sa_sigaction = perf_top__sig_winch,
1303                         .sa_flags     = SA_SIGINFO,
1304                 };
1305                 perf_top__update_print_entries(&top);
1306                 sigaction(SIGWINCH, &act, NULL);
1307         }
1308
1309         status = __cmd_top(&top);
1310
1311 out_delete_evlist:
1312         perf_evlist__delete(top.evlist);
1313
1314         return status;
1315 }