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
  44 #include "util/debug.h"
  45
  46 #include <assert.h>
  47 #include <elf.h>
  48 #include <fcntl.h>
  49
  50 #include <stdio.h>
  51 #include <termios.h>
  52 #include <unistd.h>
  53 #include <inttypes.h>
  54
  55 #include <errno.h>
  56 #include <time.h>
  57 #include <sched.h>
  58
  59 #include <sys/syscall.h>
  60 #include <sys/ioctl.h>
  61 #include <sys/poll.h>
  62 #include <sys/prctl.h>
  63 #include <sys/wait.h>
  64 #include <sys/uio.h>
  65 #include <sys/utsname.h>
  66 #include <sys/mman.h>
  67
  68 #include <linux/unistd.h>
  69 #include <linux/types.h>
  70
  71 static volatile int done;
  72
  73 static void perf_top__update_print_entries(struct perf_top *top)
  74 {
  75         if (top->print_entries > 9)
  76                 top->print_entries -= 9;
  77 }
  78
  79 static void perf_top__sig_winch(int sig __maybe_unused,
  80                                 siginfo_t *info __maybe_unused, void *arg)
  81 {
  82         struct perf_top *top = arg;
  83
  84         get_term_dimensions(&top->winsize);
  85         if (!top->print_entries
  86             || (top->print_entries+4) > top->winsize.ws_row) {
  87                 top->print_entries = top->winsize.ws_row;
  88         } else {
  89                 top->print_entries += 4;
  90                 top->winsize.ws_row = top->print_entries;
  91         }
  92         perf_top__update_print_entries(top);
  93 }
  94
  95 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  96 {
  97         struct symbol *sym;
  98         struct annotation *notes;
  99         struct map *map;
 100         int err = -1;
 101
 102         if (!he || !he->ms.sym)
 103                 return -1;
 104
 105         sym = he->ms.sym;
 106         map = he->ms.map;
 107
 108         /*
 109          * We can't annotate with just /proc/kallsyms
 110          */
 111         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
 112                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 113                        "path\n", sym->name);
 114                 sleep(1);
 115                 return -1;
 116         }
 117
 118         notes = symbol__annotation(sym);
 119         if (notes->src != NULL) {
 120                 pthread_mutex_lock(&notes->lock);
 121                 goto out_assign;
 122         }
 123
 124         pthread_mutex_lock(&notes->lock);
 125
 126         if (symbol__alloc_hist(sym) < 0) {
 127                 pthread_mutex_unlock(&notes->lock);
 128                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 129                        sym->name);
 130                 sleep(1);
 131                 return err;
 132         }
 133
 134         err = symbol__annotate(sym, map, 0);
 135         if (err == 0) {
 136 out_assign:
 137                 top->sym_filter_entry = he;
 138         }
 139
 140         pthread_mutex_unlock(&notes->lock);
 141         return err;
 142 }
 143
 144 static void __zero_source_counters(struct hist_entry *he)
 145 {
 146         struct symbol *sym = he->ms.sym;
 147         symbol__annotate_zero_histograms(sym);
 148 }
 149
 150 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 151 {
 152         struct utsname uts;
 153         int err = uname(&uts);
 154
 155         ui__warning("Out of bounds address found:\n\n"
 156                     "Addr:   %" PRIx64 "\n"
 157                     "DSO:    %s %c\n"
 158                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 159                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 160                     "Arch:   %s\n"
 161                     "Kernel: %s\n"
 162                     "Tools:  %s\n\n"
 163                     "Not all samples will be on the annotation output.\n\n"
 164                     "Please report to linux-kernel@vger.kernel.org\n",
 165                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 166                     map->start, map->end, sym->start, sym->end,
 167                     sym->binding == STB_GLOBAL ? 'g' :
 168                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 169                     err ? "[unknown]" : uts.machine,
 170                     err ? "[unknown]" : uts.release, perf_version_string);
 171         if (use_browser <= 0)
 172                 sleep(5);
 173
 174         map->erange_warned = true;
 175 }
 176
 177 static void perf_top__record_precise_ip(struct perf_top *top,
 178                                         struct hist_entry *he,
 179                                         int counter, u64 ip)
 180 {
 181         struct annotation *notes;
 182         struct symbol *sym;
 183         int err;
 184
 185         if (he == NULL || he->ms.sym == NULL ||
 186             ((top->sym_filter_entry == NULL ||
 187               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 188                 return;
 189
 190         sym = he->ms.sym;
 191         notes = symbol__annotation(sym);
 192
 193         if (pthread_mutex_trylock(&notes->lock))
 194                 return;
 195
 196         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
 197                 pthread_mutex_unlock(&notes->lock);
 198                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 199                        sym->name);
 200                 sleep(1);
 201                 return;
 202         }
 203
 204         ip = he->ms.map->map_ip(he->ms.map, ip);
 205         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
 206
 207         pthread_mutex_unlock(&notes->lock);
 208
 209         if (err == -ERANGE && !he->ms.map->erange_warned)
 210                 ui__warn_map_erange(he->ms.map, sym, ip);
 211 }
 212
 213 static void perf_top__show_details(struct perf_top *top)
 214 {
 215         struct hist_entry *he = top->sym_filter_entry;
 216         struct annotation *notes;
 217         struct symbol *symbol;
 218         int more;
 219
 220         if (!he)
 221                 return;
 222
 223         symbol = he->ms.sym;
 224         notes = symbol__annotation(symbol);
 225
 226         pthread_mutex_lock(&notes->lock);
 227
 228         if (notes->src == NULL)
 229                 goto out_unlock;
 230
 231         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 232         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 233
 234         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 235                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 236         if (top->zero)
 237                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 238         else
 239                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 240         if (more != 0)
 241                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 242 out_unlock:
 243         pthread_mutex_unlock(&notes->lock);
 244 }
 245
 246 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
 247
 248 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
 249                                                      struct addr_location *al,
 250                                                      struct perf_sample *sample)
 251 {
 252         struct hist_entry *he;
 253
 254         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period,
 255                                 sample->weight);
 256         if (he == NULL)
 257                 return NULL;
 258
 259         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 260         return he;
 261 }
 262
 263 static void perf_top__print_sym_table(struct perf_top *top)
 264 {
 265         char bf[160];
 266         int printed = 0;
 267         const int win_width = top->winsize.ws_col - 1;
 268
 269         puts(CONSOLE_CLEAR);
 270
 271         perf_top__header_snprintf(top, bf, sizeof(bf));
 272         printf("%s\n", bf);
 273
 274         perf_top__reset_sample_counters(top);
 275
 276         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 277
 278         if (top->sym_evsel->hists.stats.nr_lost_warned !=
 279             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
 280                 top->sym_evsel->hists.stats.nr_lost_warned =
 281                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
 282                 color_fprintf(stdout, PERF_COLOR_RED,
 283                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 284                               top->sym_evsel->hists.stats.nr_lost_warned);
 285                 ++printed;
 286         }
 287
 288         if (top->sym_filter_entry) {
 289                 perf_top__show_details(top);
 290                 return;
 291         }
 292
 293         hists__collapse_resort_threaded(&top->sym_evsel->hists);
 294         hists__output_resort_threaded(&top->sym_evsel->hists);
 295         hists__decay_entries_threaded(&top->sym_evsel->hists,
 296                                       top->hide_user_symbols,
 297                                       top->hide_kernel_symbols);
 298         hists__output_recalc_col_len(&top->sym_evsel->hists,
 299                                      top->winsize.ws_row - 3);
 300         putchar('\n');
 301         hists__fprintf(&top->sym_evsel->hists, false,
 302                        top->winsize.ws_row - 4 - printed, win_width, stdout);
 303 }
 304
 305 static void prompt_integer(int *target, const char *msg)
 306 {
 307         char *buf = malloc(0), *p;
 308         size_t dummy = 0;
 309         int tmp;
 310
 311         fprintf(stdout, "\n%s: ", msg);
 312         if (getline(&buf, &dummy, stdin) < 0)
 313                 return;
 314
 315         p = strchr(buf, '\n');
 316         if (p)
 317                 *p = 0;
 318
 319         p = buf;
 320         while(*p) {
 321                 if (!isdigit(*p))
 322                         goto out_free;
 323                 p++;
 324         }
 325         tmp = strtoul(buf, NULL, 10);
 326         *target = tmp;
 327 out_free:
 328         free(buf);
 329 }
 330
 331 static void prompt_percent(int *target, const char *msg)
 332 {
 333         int tmp = 0;
 334
 335         prompt_integer(&tmp, msg);
 336         if (tmp >= 0 && tmp <= 100)
 337                 *target = tmp;
 338 }
 339
 340 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 341 {
 342         char *buf = malloc(0), *p;
 343         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 344         struct rb_node *next;
 345         size_t dummy = 0;
 346
 347         /* zero counters of active symbol */
 348         if (syme) {
 349                 __zero_source_counters(syme);
 350                 top->sym_filter_entry = NULL;
 351         }
 352
 353         fprintf(stdout, "\n%s: ", msg);
 354         if (getline(&buf, &dummy, stdin) < 0)
 355                 goto out_free;
 356
 357         p = strchr(buf, '\n');
 358         if (p)
 359                 *p = 0;
 360
 361         next = rb_first(&top->sym_evsel->hists.entries);
 362         while (next) {
 363                 n = rb_entry(next, struct hist_entry, rb_node);
 364                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 365                         found = n;
 366                         break;
 367                 }
 368                 next = rb_next(&n->rb_node);
 369         }
 370
 371         if (!found) {
 372                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 373                 sleep(1);
 374         } else
 375                 perf_top__parse_source(top, found);
 376
 377 out_free:
 378         free(buf);
 379 }
 380
 381 static void perf_top__print_mapped_keys(struct perf_top *top)
 382 {
 383         char *name = NULL;
 384
 385         if (top->sym_filter_entry) {
 386                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 387                 name = sym->name;
 388         }
 389
 390         fprintf(stdout, "\nMapped keys:\n");
 391         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 392         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 393
 394         if (top->evlist->nr_entries > 1)
 395                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 396
 397         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 398
 399         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 400         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 401         fprintf(stdout, "\t[S]     stop annotation.\n");
 402
 403         fprintf(stdout,
 404                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 405                 top->hide_kernel_symbols ? "yes" : "no");
 406         fprintf(stdout,
 407                 "\t[U]     hide user symbols.               \t(%s)\n",
 408                 top->hide_user_symbols ? "yes" : "no");
 409         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 410         fprintf(stdout, "\t[qQ]    quit.\n");
 411 }
 412
 413 static int perf_top__key_mapped(struct perf_top *top, int c)
 414 {
 415         switch (c) {
 416                 case 'd':
 417                 case 'e':
 418                 case 'f':
 419                 case 'z':
 420                 case 'q':
 421                 case 'Q':
 422                 case 'K':
 423                 case 'U':
 424                 case 'F':
 425                 case 's':
 426                 case 'S':
 427                         return 1;
 428                 case 'E':
 429                         return top->evlist->nr_entries > 1 ? 1 : 0;
 430                 default:
 431                         break;
 432         }
 433
 434         return 0;
 435 }
 436
 437 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 438 {
 439         bool ret = true;
 440
 441         if (!perf_top__key_mapped(top, c)) {
 442                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 443                 struct termios tc, save;
 444
 445                 perf_top__print_mapped_keys(top);
 446                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 447                 fflush(stdout);
 448
 449                 tcgetattr(0, &save);
 450                 tc = save;
 451                 tc.c_lflag &= ~(ICANON | ECHO);
 452                 tc.c_cc[VMIN] = 0;
 453                 tc.c_cc[VTIME] = 0;
 454                 tcsetattr(0, TCSANOW, &tc);
 455
 456                 poll(&stdin_poll, 1, -1);
 457                 c = getc(stdin);
 458
 459                 tcsetattr(0, TCSAFLUSH, &save);
 460                 if (!perf_top__key_mapped(top, c))
 461                         return ret;
 462         }
 463
 464         switch (c) {
 465                 case 'd':
 466                         prompt_integer(&top->delay_secs, "Enter display delay");
 467                         if (top->delay_secs < 1)
 468                                 top->delay_secs = 1;
 469                         break;
 470                 case 'e':
 471                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 472                         if (top->print_entries == 0) {
 473                                 struct sigaction act = {
 474                                         .sa_sigaction = perf_top__sig_winch,
 475                                         .sa_flags     = SA_SIGINFO,
 476                                 };
 477                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 478                                 sigaction(SIGWINCH, &act, NULL);
 479                         } else {
 480                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 481                                 signal(SIGWINCH, SIG_DFL);
 482                         }
 483                         break;
 484                 case 'E':
 485                         if (top->evlist->nr_entries > 1) {
 486                                 /* Select 0 as the default event: */
 487                                 int counter = 0;
 488
 489                                 fprintf(stderr, "\nAvailable events:");
 490
 491                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 492                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 493
 494                                 prompt_integer(&counter, "Enter details event counter");
 495
 496                                 if (counter >= top->evlist->nr_entries) {
 497                                         top->sym_evsel = perf_evlist__first(top->evlist);
 498                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 499                                         sleep(1);
 500                                         break;
 501                                 }
 502                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
 503                                         if (top->sym_evsel->idx == counter)
 504                                                 break;
 505                         } else
 506                                 top->sym_evsel = perf_evlist__first(top->evlist);
 507                         break;
 508                 case 'f':
 509                         prompt_integer(&top->count_filter, "Enter display event count filter");
 510                         break;
 511                 case 'F':
 512                         prompt_percent(&top->sym_pcnt_filter,
 513                                        "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 (top->dump_symtab)
 522                                 perf_session__fprintf_dsos(top->session, stderr);
 523                         ret = false;
 524                         break;
 525                 case 's':
 526                         perf_top__prompt_symbol(top, "Enter details symbol");
 527                         break;
 528                 case 'S':
 529                         if (!top->sym_filter_entry)
 530                                 break;
 531                         else {
 532                                 struct hist_entry *syme = top->sym_filter_entry;
 533
 534                                 top->sym_filter_entry = NULL;
 535                                 __zero_source_counters(syme);
 536                         }
 537                         break;
 538                 case 'U':
 539                         top->hide_user_symbols = !top->hide_user_symbols;
 540                         break;
 541                 case 'z':
 542                         top->zero = !top->zero;
 543                         break;
 544                 default:
 545                         break;
 546         }
 547
 548         return ret;
 549 }
 550
 551 static void perf_top__sort_new_samples(void *arg)
 552 {
 553         struct perf_top *t = arg;
 554         perf_top__reset_sample_counters(t);
 555
 556         if (t->evlist->selected != NULL)
 557                 t->sym_evsel = t->evlist->selected;
 558
 559         hists__collapse_resort_threaded(&t->sym_evsel->hists);
 560         hists__output_resort_threaded(&t->sym_evsel->hists);
 561         hists__decay_entries_threaded(&t->sym_evsel->hists,
 562                                       t->hide_user_symbols,
 563                                       t->hide_kernel_symbols);
 564 }
 565
 566 static void *display_thread_tui(void *arg)
 567 {
 568         struct perf_evsel *pos;
 569         struct perf_top *top = arg;
 570         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 571         struct hist_browser_timer hbt = {
 572                 .timer          = perf_top__sort_new_samples,
 573                 .arg            = top,
 574                 .refresh        = top->delay_secs,
 575         };
 576
 577         perf_top__sort_new_samples(top);
 578
 579         /*
 580          * Initialize the uid_filter_str, in the future the TUI will allow
 581          * Zooming in/out UIDs. For now juse use whatever the user passed
 582          * via --uid.
 583          */
 584         list_for_each_entry(pos, &top->evlist->entries, node)
 585                 pos->hists.uid_filter_str = top->record_opts.target.uid_str;
 586
 587         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 588                                       &top->session->header.env);
 589
 590         done = 1;
 591         return NULL;
 592 }
 593
 594 static void *display_thread(void *arg)
 595 {
 596         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 597         struct termios tc, save;
 598         struct perf_top *top = arg;
 599         int delay_msecs, c;
 600
 601         tcgetattr(0, &save);
 602         tc = save;
 603         tc.c_lflag &= ~(ICANON | ECHO);
 604         tc.c_cc[VMIN] = 0;
 605         tc.c_cc[VTIME] = 0;
 606
 607         pthread__unblock_sigwinch();
 608 repeat:
 609         delay_msecs = top->delay_secs * 1000;
 610         tcsetattr(0, TCSANOW, &tc);
 611         /* trash return*/
 612         getc(stdin);
 613
 614         while (!done) {
 615                 perf_top__print_sym_table(top);
 616                 /*
 617                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 618                  * refresh screen in both cases.
 619                  */
 620                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 621                 case 0:
 622                         continue;
 623                 case -1:
 624                         if (errno == EINTR)
 625                                 continue;
 626                         /* Fall trhu */
 627                 default:
 628                         c = getc(stdin);
 629                         tcsetattr(0, TCSAFLUSH, &save);
 630
 631                         if (perf_top__handle_keypress(top, c))
 632                                 goto repeat;
 633                         done = 1;
 634                 }
 635         }
 636
 637         return NULL;
 638 }
 639
 640 /* Tag samples to be skipped. */
 641 static const char *skip_symbols[] = {
 642         "intel_idle",
 643         "default_idle",
 644         "native_safe_halt",
 645         "cpu_idle",
 646         "enter_idle",
 647         "exit_idle",
 648         "mwait_idle",
 649         "mwait_idle_with_hints",
 650         "poll_idle",
 651         "ppc64_runlatch_off",
 652         "pseries_dedicated_idle_sleep",
 653         NULL
 654 };
 655
 656 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
 657 {
 658         const char *name = sym->name;
 659         int i;
 660
 661         /*
 662          * ppc64 uses function descriptors and appends a '.' to the
 663          * start of every instruction address. Remove it.
 664          */
 665         if (name[0] == '.')
 666                 name++;
 667
 668         if (!strcmp(name, "_text") ||
 669             !strcmp(name, "_etext") ||
 670             !strcmp(name, "_sinittext") ||
 671             !strncmp("init_module", name, 11) ||
 672             !strncmp("cleanup_module", name, 14) ||
 673             strstr(name, "_text_start") ||
 674             strstr(name, "_text_end"))
 675                 return 1;
 676
 677         for (i = 0; skip_symbols[i]; i++) {
 678                 if (!strcmp(skip_symbols[i], name)) {
 679                         sym->ignore = true;
 680                         break;
 681                 }
 682         }
 683
 684         return 0;
 685 }
 686
 687 static void perf_event__process_sample(struct perf_tool *tool,
 688                                        const union perf_event *event,
 689                                        struct perf_evsel *evsel,
 690                                        struct perf_sample *sample,
 691                                        struct machine *machine)
 692 {
 693         struct perf_top *top = container_of(tool, struct perf_top, tool);
 694         struct symbol *parent = NULL;
 695         u64 ip = event->ip.ip;
 696         struct addr_location al;
 697         int err;
 698
 699         if (!machine && perf_guest) {
 700                 static struct intlist *seen;
 701
 702                 if (!seen)
 703                         seen = intlist__new(NULL);
 704
 705                 if (!intlist__has_entry(seen, event->ip.pid)) {
 706                         pr_err("Can't find guest [%d]'s kernel information\n",
 707                                 event->ip.pid);
 708                         intlist__add(seen, event->ip.pid);
 709                 }
 710                 return;
 711         }
 712
 713         if (!machine) {
 714                 pr_err("%u unprocessable samples recorded.\r",
 715                        top->session->stats.nr_unprocessable_samples++);
 716                 return;
 717         }
 718
 719         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 720                 top->exact_samples++;
 721
 722         if (perf_event__preprocess_sample(event, machine, &al, sample,
 723                                           symbol_filter) < 0 ||
 724             al.filtered)
 725                 return;
 726
 727         if (!top->kptr_restrict_warned &&
 728             symbol_conf.kptr_restrict &&
 729             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 730                 ui__warning(
 731 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 732 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 733 "Kernel%s samples will not be resolved.\n",
 734                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 735                           " modules" : "");
 736                 if (use_browser <= 0)
 737                         sleep(5);
 738                 top->kptr_restrict_warned = true;
 739         }
 740
 741         if (al.sym == NULL) {
 742                 const char *msg = "Kernel samples will not be resolved.\n";
 743                 /*
 744                  * As we do lazy loading of symtabs we only will know if the
 745                  * specified vmlinux file is invalid when we actually have a
 746                  * hit in kernel space and then try to load it. So if we get
 747                  * here and there are _no_ symbols in the DSO backing the
 748                  * kernel map, bail out.
 749                  *
 750                  * We may never get here, for instance, if we use -K/
 751                  * --hide-kernel-symbols, even if the user specifies an
 752                  * invalid --vmlinux ;-)
 753                  */
 754                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 755                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 756                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 757                         if (symbol_conf.vmlinux_name) {
 758                                 ui__warning("The %s file can't be used.\n%s",
 759                                             symbol_conf.vmlinux_name, msg);
 760                         } else {
 761                                 ui__warning("A vmlinux file was not found.\n%s",
 762                                             msg);
 763                         }
 764
 765                         if (use_browser <= 0)
 766                                 sleep(5);
 767                         top->vmlinux_warned = true;
 768                 }
 769         }
 770
 771         if (al.sym == NULL || !al.sym->ignore) {
 772                 struct hist_entry *he;
 773
 774                 if ((sort__has_parent || symbol_conf.use_callchain) &&
 775                     sample->callchain) {
 776                         err = machine__resolve_callchain(machine, evsel,
 777                                                          al.thread, sample,
 778                                                          &parent);
 779
 780                         if (err)
 781                                 return;
 782                 }
 783
 784                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
 785                 if (he == NULL) {
 786                         pr_err("Problem incrementing symbol period, skipping event\n");
 787                         return;
 788                 }
 789
 790                 if (symbol_conf.use_callchain) {
 791                         err = callchain_append(he->callchain, &callchain_cursor,
 792                                                sample->period);
 793                         if (err)
 794                                 return;
 795                 }
 796
 797                 if (top->sort_has_symbols)
 798                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
 799         }
 800
 801         return;
 802 }
 803
 804 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 805 {
 806         struct perf_sample sample;
 807         struct perf_evsel *evsel;
 808         struct perf_session *session = top->session;
 809         union perf_event *event;
 810         struct machine *machine;
 811         u8 origin;
 812         int ret;
 813
 814         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 815                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 816                 if (ret) {
 817                         pr_err("Can't parse sample, err = %d\n", ret);
 818                         continue;
 819                 }
 820
 821                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 822                 assert(evsel != NULL);
 823
 824                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 825
 826                 if (event->header.type == PERF_RECORD_SAMPLE)
 827                         ++top->samples;
 828
 829                 switch (origin) {
 830                 case PERF_RECORD_MISC_USER:
 831                         ++top->us_samples;
 832                         if (top->hide_user_symbols)
 833                                 continue;
 834                         machine = &session->machines.host;
 835                         break;
 836                 case PERF_RECORD_MISC_KERNEL:
 837                         ++top->kernel_samples;
 838                         if (top->hide_kernel_symbols)
 839                                 continue;
 840                         machine = &session->machines.host;
 841                         break;
 842                 case PERF_RECORD_MISC_GUEST_KERNEL:
 843                         ++top->guest_kernel_samples;
 844                         machine = perf_session__find_machine(session, event->ip.pid);
 845                         break;
 846                 case PERF_RECORD_MISC_GUEST_USER:
 847                         ++top->guest_us_samples;
 848                         /*
 849                          * TODO: we don't process guest user from host side
 850                          * except simple counting.
 851                          */
 852                         /* Fall thru */
 853                 default:
 854                         continue;
 855                 }
 856
 857
 858                 if (event->header.type == PERF_RECORD_SAMPLE) {
 859                         perf_event__process_sample(&top->tool, event, evsel,
 860                                                    &sample, machine);
 861                 } else if (event->header.type < PERF_RECORD_MAX) {
 862                         hists__inc_nr_events(&evsel->hists, event->header.type);
 863                         machine__process_event(machine, event);
 864                 } else
 865                         ++session->stats.nr_unknown_events;
 866         }
 867 }
 868
 869 static void perf_top__mmap_read(struct perf_top *top)
 870 {
 871         int i;
 872
 873         for (i = 0; i < top->evlist->nr_mmaps; i++)
 874                 perf_top__mmap_read_idx(top, i);
 875 }
 876
 877 static int perf_top__start_counters(struct perf_top *top)
 878 {
 879         char msg[512];
 880         struct perf_evsel *counter;
 881         struct perf_evlist *evlist = top->evlist;
 882         struct perf_record_opts *opts = &top->record_opts;
 883
 884         perf_evlist__config(evlist, opts);
 885
 886         list_for_each_entry(counter, &evlist->entries, node) {
 887 try_again:
 888                 if (perf_evsel__open(counter, top->evlist->cpus,
 889                                      top->evlist->threads) < 0) {
 890                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 891                                 if (verbose)
 892                                         ui__warning("%s\n", msg);
 893                                 goto try_again;
 894                         }
 895
 896                         perf_evsel__open_strerror(counter, &opts->target,
 897                                                   errno, msg, sizeof(msg));
 898                         ui__error("%s\n", msg);
 899                         goto out_err;
 900                 }
 901         }
 902
 903         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 904                 ui__error("Failed to mmap with %d (%s)\n",
 905                             errno, strerror(errno));
 906                 goto out_err;
 907         }
 908
 909         return 0;
 910
 911 out_err:
 912         return -1;
 913 }
 914
 915 static int perf_top__setup_sample_type(struct perf_top *top)
 916 {
 917         if (!top->sort_has_symbols) {
 918                 if (symbol_conf.use_callchain) {
 919                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 920                         return -EINVAL;
 921                 }
 922         } else if (callchain_param.mode != CHAIN_NONE) {
 923                 if (callchain_register_param(&callchain_param) < 0) {
 924                         ui__error("Can't register callchain params.\n");
 925                         return -EINVAL;
 926                 }
 927         }
 928
 929         return 0;
 930 }
 931
 932 static int __cmd_top(struct perf_top *top)
 933 {
 934         struct perf_record_opts *opts = &top->record_opts;
 935         pthread_t thread;
 936         int ret;
 937         /*
 938          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
 939          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
 940          */
 941         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
 942         if (top->session == NULL)
 943                 return -ENOMEM;
 944
 945         ret = perf_top__setup_sample_type(top);
 946         if (ret)
 947                 goto out_delete;
 948
 949         if (perf_target__has_task(&opts->target))
 950                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
 951                                                   perf_event__process,
 952                                                   &top->session->machines.host);
 953         else
 954                 perf_event__synthesize_threads(&top->tool, perf_event__process,
 955                                                &top->session->machines.host);
 956
 957         ret = perf_top__start_counters(top);
 958         if (ret)
 959                 goto out_delete;
 960
 961         top->session->evlist = top->evlist;
 962         perf_session__set_id_hdr_size(top->session);
 963
 964         /*
 965          * When perf is starting the traced process, all the events (apart from
 966          * group members) have enable_on_exec=1 set, so don't spoil it by
 967          * prematurely enabling them.
 968          *
 969          * XXX 'top' still doesn't start workloads like record, trace, but should,
 970          * so leave the check here.
 971          */
 972         if (!perf_target__none(&opts->target))
 973                 perf_evlist__enable(top->evlist);
 974
 975         /* Wait for a minimal set of events before starting the snapshot */
 976         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
 977
 978         perf_top__mmap_read(top);
 979
 980         ret = -1;
 981         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 982                                                             display_thread), top)) {
 983                 ui__error("Could not create display thread.\n");
 984                 goto out_delete;
 985         }
 986
 987         if (top->realtime_prio) {
 988                 struct sched_param param;
 989
 990                 param.sched_priority = top->realtime_prio;
 991                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 992                         ui__error("Could not set realtime priority.\n");
 993                         goto out_delete;
 994                 }
 995         }
 996
 997         while (!done) {
 998                 u64 hits = top->samples;
 999
1000                 perf_top__mmap_read(top);
1001
1002                 if (hits == top->samples)
1003                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1004         }
1005
1006         ret = 0;
1007 out_delete:
1008         perf_session__delete(top->session);
1009         top->session = NULL;
1010
1011         return ret;
1012 }
1013
1014 static int
1015 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1016 {
1017         /*
1018          * --no-call-graph
1019          */
1020         if (unset)
1021                 return 0;
1022
1023         symbol_conf.use_callchain = true;
1024
1025         return record_parse_callchain_opt(opt, arg, unset);
1026 }
1027
1028 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1029 {
1030         int status;
1031         char errbuf[BUFSIZ];
1032         struct perf_top top = {
1033                 .count_filter        = 5,
1034                 .delay_secs          = 2,
1035                 .record_opts = {
1036                         .mmap_pages     = UINT_MAX,
1037                         .user_freq      = UINT_MAX,
1038                         .user_interval  = ULLONG_MAX,
1039                         .freq           = 4000, /* 4 KHz */
1040                         .target              = {
1041                                 .uses_mmap   = true,
1042                         },
1043                 },
1044                 .sym_pcnt_filter     = 5,
1045         };
1046         struct perf_record_opts *opts = &top.record_opts;
1047         struct perf_target *target = &opts->target;
1048         const struct option options[] = {
1049         OPT_CALLBACK('e', "event", &top.evlist, "event",
1050                      "event selector. use 'perf list' to list available events",
1051                      parse_events_option),
1052         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1053         OPT_STRING('p', "pid", &target->pid, "pid",
1054                     "profile events on existing process id"),
1055         OPT_STRING('t', "tid", &target->tid, "tid",
1056                     "profile events on existing thread id"),
1057         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1058                             "system-wide collection from all CPUs"),
1059         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1060                     "list of cpus to monitor"),
1061         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1062                    "file", "vmlinux pathname"),
1063         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1064                     "hide kernel symbols"),
1065         OPT_UINTEGER('m', "mmap-pages", &opts->mmap_pages,
1066                      "number of mmap data pages"),
1067         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1068                     "collect data with this RT SCHED_FIFO priority"),
1069         OPT_INTEGER('d', "delay", &top.delay_secs,
1070                     "number of seconds to delay between refreshes"),
1071         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1072                             "dump the symbol table used for profiling"),
1073         OPT_INTEGER('f', "count-filter", &top.count_filter,
1074                     "only display functions with more events than this"),
1075         OPT_BOOLEAN('g', "group", &opts->group,
1076                             "put the counters into a counter group"),
1077         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1078                     "child tasks do not inherit counters"),
1079         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1080                     "symbol to annotate"),
1081         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1082         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1083         OPT_INTEGER('E', "entries", &top.print_entries,
1084                     "display this many functions"),
1085         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1086                     "hide user symbols"),
1087         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1088         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1089         OPT_INCR('v', "verbose", &verbose,
1090                     "be more verbose (show counter open errors, etc)"),
1091         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1092                    "sort by key(s): pid, comm, dso, symbol, parent, weight, local_weight"),
1093         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1094                     "Show a column with the number of samples"),
1095         OPT_CALLBACK_DEFAULT('G', "call-graph", &top.record_opts,
1096                              "mode[,dump_size]", record_callchain_help,
1097                              &parse_callchain_opt, "fp"),
1098         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1099                     "Show a column with the sum of periods"),
1100         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1101                    "only consider symbols in these dsos"),
1102         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1103                    "only consider symbols in these comms"),
1104         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1105                    "only consider these symbols"),
1106         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1107                     "Interleave source code with assembly code (default)"),
1108         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1109                     "Display raw encoding of assembly instructions (default)"),
1110         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1111                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1112         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1113         OPT_END()
1114         };
1115         const char * const top_usage[] = {
1116                 "perf top [<options>]",
1117                 NULL
1118         };
1119
1120         top.evlist = perf_evlist__new();
1121         if (top.evlist == NULL)
1122                 return -ENOMEM;
1123
1124         symbol_conf.exclude_other = false;
1125
1126         argc = parse_options(argc, argv, options, top_usage, 0);
1127         if (argc)
1128                 usage_with_options(top_usage, options);
1129
1130         if (sort_order == default_sort_order)
1131                 sort_order = "dso,symbol";
1132
1133         if (setup_sorting() < 0)
1134                 usage_with_options(top_usage, options);
1135
1136         if (top.use_stdio)
1137                 use_browser = 0;
1138         else if (top.use_tui)
1139                 use_browser = 1;
1140
1141         setup_browser(false);
1142
1143         status = perf_target__validate(target);
1144         if (status) {
1145                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1146                 ui__warning("%s", errbuf);
1147         }
1148
1149         status = perf_target__parse_uid(target);
1150         if (status) {
1151                 int saved_errno = errno;
1152
1153                 perf_target__strerror(target, status, errbuf, BUFSIZ);
1154                 ui__error("%s", errbuf);
1155
1156                 status = -saved_errno;
1157                 goto out_delete_evlist;
1158         }
1159
1160         if (perf_target__none(target))
1161                 target->system_wide = true;
1162
1163         if (perf_evlist__create_maps(top.evlist, target) < 0)
1164                 usage_with_options(top_usage, options);
1165
1166         if (!top.evlist->nr_entries &&
1167             perf_evlist__add_default(top.evlist) < 0) {
1168                 ui__error("Not enough memory for event selector list\n");
1169                 goto out_delete_maps;
1170         }
1171
1172         symbol_conf.nr_events = top.evlist->nr_entries;
1173
1174         if (top.delay_secs < 1)
1175                 top.delay_secs = 1;
1176
1177         if (opts->user_interval != ULLONG_MAX)
1178                 opts->default_interval = opts->user_interval;
1179         if (opts->user_freq != UINT_MAX)
1180                 opts->freq = opts->user_freq;
1181
1182         /*
1183          * User specified count overrides default frequency.
1184          */
1185         if (opts->default_interval)
1186                 opts->freq = 0;
1187         else if (opts->freq) {
1188                 opts->default_interval = opts->freq;
1189         } else {
1190                 ui__error("frequency and count are zero, aborting\n");
1191                 status = -EINVAL;
1192                 goto out_delete_maps;
1193         }
1194
1195         top.sym_evsel = perf_evlist__first(top.evlist);
1196
1197         symbol_conf.priv_size = sizeof(struct annotation);
1198
1199         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1200         if (symbol__init() < 0)
1201                 return -1;
1202
1203         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1204         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1205         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1206
1207         /*
1208          * Avoid annotation data structures overhead when symbols aren't on the
1209          * sort list.
1210          */
1211         top.sort_has_symbols = sort_sym.list.next != NULL;
1212
1213         get_term_dimensions(&top.winsize);
1214         if (top.print_entries == 0) {
1215                 struct sigaction act = {
1216                         .sa_sigaction = perf_top__sig_winch,
1217                         .sa_flags     = SA_SIGINFO,
1218                 };
1219                 perf_top__update_print_entries(&top);
1220                 sigaction(SIGWINCH, &act, NULL);
1221         }
1222
1223         status = __cmd_top(&top);
1224
1225 out_delete_maps:
1226         perf_evlist__delete_maps(top.evlist);
1227 out_delete_evlist:
1228         perf_evlist__delete(top.evlist);
1229
1230         return status;
1231 }