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 "arch/common.h"
  44
  45 #include "util/debug.h"
  46
  47 #include <assert.h>
  48 #include <elf.h>
  49 #include <fcntl.h>
  50
  51 #include <stdio.h>
  52 #include <termios.h>
  53 #include <unistd.h>
  54 #include <inttypes.h>
  55
  56 #include <errno.h>
  57 #include <time.h>
  58 #include <sched.h>
  59
  60 #include <sys/syscall.h>
  61 #include <sys/ioctl.h>
  62 #include <poll.h>
  63 #include <sys/prctl.h>
  64 #include <sys/wait.h>
  65 #include <sys/uio.h>
  66 #include <sys/utsname.h>
  67 #include <sys/mman.h>
  68
  69 #include <linux/types.h>
  70
  71 static volatile int done;
  72
  73 #define HEADER_LINE_NR  5
  74
  75 static void perf_top__update_print_entries(struct perf_top *top)
  76 {
  77         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  78 }
  79
  80 static void perf_top__sig_winch(int sig __maybe_unused,
  81                                 siginfo_t *info __maybe_unused, void *arg)
  82 {
  83         struct perf_top *top = arg;
  84
  85         get_term_dimensions(&top->winsize);
  86         perf_top__update_print_entries(top);
  87 }
  88
  89 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
  90 {
  91         struct symbol *sym;
  92         struct annotation *notes;
  93         struct map *map;
  94         int err = -1;
  95
  96         if (!he || !he->ms.sym)
  97                 return -1;
  98
  99         sym = he->ms.sym;
 100         map = he->ms.map;
 101
 102         /*
 103          * We can't annotate with just /proc/kallsyms
 104          */
 105         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 106             !dso__is_kcore(map->dso)) {
 107                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 108                        "path\n", sym->name);
 109                 sleep(1);
 110                 return -1;
 111         }
 112
 113         notes = symbol__annotation(sym);
 114         if (notes->src != NULL) {
 115                 pthread_mutex_lock(&notes->lock);
 116                 goto out_assign;
 117         }
 118
 119         pthread_mutex_lock(&notes->lock);
 120
 121         if (symbol__alloc_hist(sym) < 0) {
 122                 pthread_mutex_unlock(&notes->lock);
 123                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 124                        sym->name);
 125                 sleep(1);
 126                 return err;
 127         }
 128
 129         err = symbol__annotate(sym, map, 0);
 130         if (err == 0) {
 131 out_assign:
 132                 top->sym_filter_entry = he;
 133         }
 134
 135         pthread_mutex_unlock(&notes->lock);
 136         return err;
 137 }
 138
 139 static void __zero_source_counters(struct hist_entry *he)
 140 {
 141         struct symbol *sym = he->ms.sym;
 142         symbol__annotate_zero_histograms(sym);
 143 }
 144
 145 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 146 {
 147         struct utsname uts;
 148         int err = uname(&uts);
 149
 150         ui__warning("Out of bounds address found:\n\n"
 151                     "Addr:   %" PRIx64 "\n"
 152                     "DSO:    %s %c\n"
 153                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 154                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 155                     "Arch:   %s\n"
 156                     "Kernel: %s\n"
 157                     "Tools:  %s\n\n"
 158                     "Not all samples will be on the annotation output.\n\n"
 159                     "Please report to linux-kernel@vger.kernel.org\n",
 160                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 161                     map->start, map->end, sym->start, sym->end,
 162                     sym->binding == STB_GLOBAL ? 'g' :
 163                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 164                     err ? "[unknown]" : uts.machine,
 165                     err ? "[unknown]" : uts.release, perf_version_string);
 166         if (use_browser <= 0)
 167                 sleep(5);
 168
 169         map->erange_warned = true;
 170 }
 171
 172 static void perf_top__record_precise_ip(struct perf_top *top,
 173                                         struct hist_entry *he,
 174                                         int counter, u64 ip)
 175 {
 176         struct annotation *notes;
 177         struct symbol *sym;
 178         int err = 0;
 179
 180         if (he == NULL || he->ms.sym == NULL ||
 181             ((top->sym_filter_entry == NULL ||
 182               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
 183                 return;
 184
 185         sym = he->ms.sym;
 186         notes = symbol__annotation(sym);
 187
 188         if (pthread_mutex_trylock(&notes->lock))
 189                 return;
 190
 191         ip = he->ms.map->map_ip(he->ms.map, ip);
 192
 193         if (ui__has_annotation())
 194                 err = hist_entry__inc_addr_samples(he, counter, ip);
 195
 196         pthread_mutex_unlock(&notes->lock);
 197
 198         /*
 199          * This function is now called with he->hists->lock held.
 200          * Release it before going to sleep.
 201          */
 202         pthread_mutex_unlock(&he->hists->lock);
 203
 204         if (err == -ERANGE && !he->ms.map->erange_warned)
 205                 ui__warn_map_erange(he->ms.map, sym, ip);
 206         else if (err == -ENOMEM) {
 207                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 208                        sym->name);
 209                 sleep(1);
 210         }
 211
 212         pthread_mutex_lock(&he->hists->lock);
 213 }
 214
 215 static void perf_top__show_details(struct perf_top *top)
 216 {
 217         struct hist_entry *he = top->sym_filter_entry;
 218         struct annotation *notes;
 219         struct symbol *symbol;
 220         int more;
 221
 222         if (!he)
 223                 return;
 224
 225         symbol = he->ms.sym;
 226         notes = symbol__annotation(symbol);
 227
 228         pthread_mutex_lock(&notes->lock);
 229
 230         if (notes->src == NULL)
 231                 goto out_unlock;
 232
 233         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 234         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
 235
 236         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
 237                                        0, top->sym_pcnt_filter, top->print_entries, 4);
 238         if (top->zero)
 239                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 240         else
 241                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 242         if (more != 0)
 243                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 244 out_unlock:
 245         pthread_mutex_unlock(&notes->lock);
 246 }
 247
 248 static void perf_top__print_sym_table(struct perf_top *top)
 249 {
 250         char bf[160];
 251         int printed = 0;
 252         const int win_width = top->winsize.ws_col - 1;
 253         struct hists *hists = evsel__hists(top->sym_evsel);
 254
 255         puts(CONSOLE_CLEAR);
 256
 257         perf_top__header_snprintf(top, bf, sizeof(bf));
 258         printf("%s\n", bf);
 259
 260         perf_top__reset_sample_counters(top);
 261
 262         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 263
 264         if (hists->stats.nr_lost_warned !=
 265             hists->stats.nr_events[PERF_RECORD_LOST]) {
 266                 hists->stats.nr_lost_warned =
 267                               hists->stats.nr_events[PERF_RECORD_LOST];
 268                 color_fprintf(stdout, PERF_COLOR_RED,
 269                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 270                               hists->stats.nr_lost_warned);
 271                 ++printed;
 272         }
 273
 274         if (top->sym_filter_entry) {
 275                 perf_top__show_details(top);
 276                 return;
 277         }
 278
 279         if (top->zero) {
 280                 hists__delete_entries(hists);
 281         } else {
 282                 hists__decay_entries(hists, top->hide_user_symbols,
 283                                      top->hide_kernel_symbols);
 284         }
 285
 286         hists__collapse_resort(hists, NULL);
 287         hists__output_resort(hists, NULL);
 288
 289         hists__output_recalc_col_len(hists, top->print_entries - printed);
 290         putchar('\n');
 291         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 292                        top->min_percent, stdout);
 293 }
 294
 295 static void prompt_integer(int *target, const char *msg)
 296 {
 297         char *buf = malloc(0), *p;
 298         size_t dummy = 0;
 299         int tmp;
 300
 301         fprintf(stdout, "\n%s: ", msg);
 302         if (getline(&buf, &dummy, stdin) < 0)
 303                 return;
 304
 305         p = strchr(buf, '\n');
 306         if (p)
 307                 *p = 0;
 308
 309         p = buf;
 310         while(*p) {
 311                 if (!isdigit(*p))
 312                         goto out_free;
 313                 p++;
 314         }
 315         tmp = strtoul(buf, NULL, 10);
 316         *target = tmp;
 317 out_free:
 318         free(buf);
 319 }
 320
 321 static void prompt_percent(int *target, const char *msg)
 322 {
 323         int tmp = 0;
 324
 325         prompt_integer(&tmp, msg);
 326         if (tmp >= 0 && tmp <= 100)
 327                 *target = tmp;
 328 }
 329
 330 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 331 {
 332         char *buf = malloc(0), *p;
 333         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 334         struct hists *hists = evsel__hists(top->sym_evsel);
 335         struct rb_node *next;
 336         size_t dummy = 0;
 337
 338         /* zero counters of active symbol */
 339         if (syme) {
 340                 __zero_source_counters(syme);
 341                 top->sym_filter_entry = NULL;
 342         }
 343
 344         fprintf(stdout, "\n%s: ", msg);
 345         if (getline(&buf, &dummy, stdin) < 0)
 346                 goto out_free;
 347
 348         p = strchr(buf, '\n');
 349         if (p)
 350                 *p = 0;
 351
 352         next = rb_first(&hists->entries);
 353         while (next) {
 354                 n = rb_entry(next, struct hist_entry, rb_node);
 355                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 356                         found = n;
 357                         break;
 358                 }
 359                 next = rb_next(&n->rb_node);
 360         }
 361
 362         if (!found) {
 363                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 364                 sleep(1);
 365         } else
 366                 perf_top__parse_source(top, found);
 367
 368 out_free:
 369         free(buf);
 370 }
 371
 372 static void perf_top__print_mapped_keys(struct perf_top *top)
 373 {
 374         char *name = NULL;
 375
 376         if (top->sym_filter_entry) {
 377                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 378                 name = sym->name;
 379         }
 380
 381         fprintf(stdout, "\nMapped keys:\n");
 382         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 383         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 384
 385         if (top->evlist->nr_entries > 1)
 386                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 387
 388         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 389
 390         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
 391         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 392         fprintf(stdout, "\t[S]     stop annotation.\n");
 393
 394         fprintf(stdout,
 395                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
 396                 top->hide_kernel_symbols ? "yes" : "no");
 397         fprintf(stdout,
 398                 "\t[U]     hide user symbols.               \t(%s)\n",
 399                 top->hide_user_symbols ? "yes" : "no");
 400         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 401         fprintf(stdout, "\t[qQ]    quit.\n");
 402 }
 403
 404 static int perf_top__key_mapped(struct perf_top *top, int c)
 405 {
 406         switch (c) {
 407                 case 'd':
 408                 case 'e':
 409                 case 'f':
 410                 case 'z':
 411                 case 'q':
 412                 case 'Q':
 413                 case 'K':
 414                 case 'U':
 415                 case 'F':
 416                 case 's':
 417                 case 'S':
 418                         return 1;
 419                 case 'E':
 420                         return top->evlist->nr_entries > 1 ? 1 : 0;
 421                 default:
 422                         break;
 423         }
 424
 425         return 0;
 426 }
 427
 428 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 429 {
 430         bool ret = true;
 431
 432         if (!perf_top__key_mapped(top, c)) {
 433                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 434                 struct termios save;
 435
 436                 perf_top__print_mapped_keys(top);
 437                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 438                 fflush(stdout);
 439
 440                 set_term_quiet_input(&save);
 441
 442                 poll(&stdin_poll, 1, -1);
 443                 c = getc(stdin);
 444
 445                 tcsetattr(0, TCSAFLUSH, &save);
 446                 if (!perf_top__key_mapped(top, c))
 447                         return ret;
 448         }
 449
 450         switch (c) {
 451                 case 'd':
 452                         prompt_integer(&top->delay_secs, "Enter display delay");
 453                         if (top->delay_secs < 1)
 454                                 top->delay_secs = 1;
 455                         break;
 456                 case 'e':
 457                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 458                         if (top->print_entries == 0) {
 459                                 struct sigaction act = {
 460                                         .sa_sigaction = perf_top__sig_winch,
 461                                         .sa_flags     = SA_SIGINFO,
 462                                 };
 463                                 perf_top__sig_winch(SIGWINCH, NULL, top);
 464                                 sigaction(SIGWINCH, &act, NULL);
 465                         } else {
 466                                 signal(SIGWINCH, SIG_DFL);
 467                         }
 468                         break;
 469                 case 'E':
 470                         if (top->evlist->nr_entries > 1) {
 471                                 /* Select 0 as the default event: */
 472                                 int counter = 0;
 473
 474                                 fprintf(stderr, "\nAvailable events:");
 475
 476                                 evlist__for_each(top->evlist, top->sym_evsel)
 477                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 478
 479                                 prompt_integer(&counter, "Enter details event counter");
 480
 481                                 if (counter >= top->evlist->nr_entries) {
 482                                         top->sym_evsel = perf_evlist__first(top->evlist);
 483                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 484                                         sleep(1);
 485                                         break;
 486                                 }
 487                                 evlist__for_each(top->evlist, top->sym_evsel)
 488                                         if (top->sym_evsel->idx == counter)
 489                                                 break;
 490                         } else
 491                                 top->sym_evsel = perf_evlist__first(top->evlist);
 492                         break;
 493                 case 'f':
 494                         prompt_integer(&top->count_filter, "Enter display event count filter");
 495                         break;
 496                 case 'F':
 497                         prompt_percent(&top->sym_pcnt_filter,
 498                                        "Enter details display event filter (percent)");
 499                         break;
 500                 case 'K':
 501                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 502                         break;
 503                 case 'q':
 504                 case 'Q':
 505                         printf("exiting.\n");
 506                         if (top->dump_symtab)
 507                                 perf_session__fprintf_dsos(top->session, stderr);
 508                         ret = false;
 509                         break;
 510                 case 's':
 511                         perf_top__prompt_symbol(top, "Enter details symbol");
 512                         break;
 513                 case 'S':
 514                         if (!top->sym_filter_entry)
 515                                 break;
 516                         else {
 517                                 struct hist_entry *syme = top->sym_filter_entry;
 518
 519                                 top->sym_filter_entry = NULL;
 520                                 __zero_source_counters(syme);
 521                         }
 522                         break;
 523                 case 'U':
 524                         top->hide_user_symbols = !top->hide_user_symbols;
 525                         break;
 526                 case 'z':
 527                         top->zero = !top->zero;
 528                         break;
 529                 default:
 530                         break;
 531         }
 532
 533         return ret;
 534 }
 535
 536 static void perf_top__sort_new_samples(void *arg)
 537 {
 538         struct perf_top *t = arg;
 539         struct hists *hists;
 540
 541         perf_top__reset_sample_counters(t);
 542
 543         if (t->evlist->selected != NULL)
 544                 t->sym_evsel = t->evlist->selected;
 545
 546         hists = evsel__hists(t->sym_evsel);
 547
 548         if (t->zero) {
 549                 hists__delete_entries(hists);
 550         } else {
 551                 hists__decay_entries(hists, t->hide_user_symbols,
 552                                      t->hide_kernel_symbols);
 553         }
 554
 555         hists__collapse_resort(hists, NULL);
 556         hists__output_resort(hists, NULL);
 557 }
 558
 559 static void *display_thread_tui(void *arg)
 560 {
 561         struct perf_evsel *pos;
 562         struct perf_top *top = arg;
 563         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 564         struct hist_browser_timer hbt = {
 565                 .timer          = perf_top__sort_new_samples,
 566                 .arg            = top,
 567                 .refresh        = top->delay_secs,
 568         };
 569
 570         perf_top__sort_new_samples(top);
 571
 572         /*
 573          * Initialize the uid_filter_str, in the future the TUI will allow
 574          * Zooming in/out UIDs. For now juse use whatever the user passed
 575          * via --uid.
 576          */
 577         evlist__for_each(top->evlist, pos) {
 578                 struct hists *hists = evsel__hists(pos);
 579                 hists->uid_filter_str = top->record_opts.target.uid_str;
 580         }
 581
 582         perf_evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 583                                       &top->session->header.env);
 584
 585         done = 1;
 586         return NULL;
 587 }
 588
 589 static void display_sig(int sig __maybe_unused)
 590 {
 591         done = 1;
 592 }
 593
 594 static void display_setup_sig(void)
 595 {
 596         signal(SIGSEGV, display_sig);
 597         signal(SIGFPE,  display_sig);
 598         signal(SIGINT,  display_sig);
 599         signal(SIGQUIT, display_sig);
 600         signal(SIGTERM, display_sig);
 601 }
 602
 603 static void *display_thread(void *arg)
 604 {
 605         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 606         struct termios save;
 607         struct perf_top *top = arg;
 608         int delay_msecs, c;
 609
 610         display_setup_sig();
 611         pthread__unblock_sigwinch();
 612 repeat:
 613         delay_msecs = top->delay_secs * 1000;
 614         set_term_quiet_input(&save);
 615         /* trash return*/
 616         getc(stdin);
 617
 618         while (!done) {
 619                 perf_top__print_sym_table(top);
 620                 /*
 621                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 622                  * refresh screen in both cases.
 623                  */
 624                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 625                 case 0:
 626                         continue;
 627                 case -1:
 628                         if (errno == EINTR)
 629                                 continue;
 630                         /* Fall trhu */
 631                 default:
 632                         c = getc(stdin);
 633                         tcsetattr(0, TCSAFLUSH, &save);
 634
 635                         if (perf_top__handle_keypress(top, c))
 636                                 goto repeat;
 637                         done = 1;
 638                 }
 639         }
 640
 641         tcsetattr(0, TCSAFLUSH, &save);
 642         return NULL;
 643 }
 644
 645 static int symbol_filter(struct map *map, struct symbol *sym)
 646 {
 647         const char *name = sym->name;
 648
 649         if (!map->dso->kernel)
 650                 return 0;
 651         /*
 652          * ppc64 uses function descriptors and appends a '.' to the
 653          * start of every instruction address. Remove it.
 654          */
 655         if (name[0] == '.')
 656                 name++;
 657
 658         if (!strcmp(name, "_text") ||
 659             !strcmp(name, "_etext") ||
 660             !strcmp(name, "_sinittext") ||
 661             !strncmp("init_module", name, 11) ||
 662             !strncmp("cleanup_module", name, 14) ||
 663             strstr(name, "_text_start") ||
 664             strstr(name, "_text_end"))
 665                 return 1;
 666
 667         if (symbol__is_idle(sym))
 668                 sym->ignore = true;
 669
 670         return 0;
 671 }
 672
 673 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 674                                    struct addr_location *al, bool single,
 675                                    void *arg)
 676 {
 677         struct perf_top *top = arg;
 678         struct hist_entry *he = iter->he;
 679         struct perf_evsel *evsel = iter->evsel;
 680
 681         if (sort__has_sym && single) {
 682                 u64 ip = al->addr;
 683
 684                 if (al->map)
 685                         ip = al->map->unmap_ip(al->map, ip);
 686
 687                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
 688         }
 689
 690         return 0;
 691 }
 692
 693 static void perf_event__process_sample(struct perf_tool *tool,
 694                                        const union perf_event *event,
 695                                        struct perf_evsel *evsel,
 696                                        struct perf_sample *sample,
 697                                        struct machine *machine)
 698 {
 699         struct perf_top *top = container_of(tool, struct perf_top, tool);
 700         struct addr_location al;
 701         int err;
 702
 703         if (!machine && perf_guest) {
 704                 static struct intlist *seen;
 705
 706                 if (!seen)
 707                         seen = intlist__new(NULL);
 708
 709                 if (!intlist__has_entry(seen, sample->pid)) {
 710                         pr_err("Can't find guest [%d]'s kernel information\n",
 711                                 sample->pid);
 712                         intlist__add(seen, sample->pid);
 713                 }
 714                 return;
 715         }
 716
 717         if (!machine) {
 718                 pr_err("%u unprocessable samples recorded.\r",
 719                        top->session->stats.nr_unprocessable_samples++);
 720                 return;
 721         }
 722
 723         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 724                 top->exact_samples++;
 725
 726         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
 727                 return;
 728
 729         if (!top->kptr_restrict_warned &&
 730             symbol_conf.kptr_restrict &&
 731             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 732                 ui__warning(
 733 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 734 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 735 "Kernel%s samples will not be resolved.\n",
 736                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
 737                           " modules" : "");
 738                 if (use_browser <= 0)
 739                         sleep(5);
 740                 top->kptr_restrict_warned = true;
 741         }
 742
 743         if (al.sym == NULL) {
 744                 const char *msg = "Kernel samples will not be resolved.\n";
 745                 /*
 746                  * As we do lazy loading of symtabs we only will know if the
 747                  * specified vmlinux file is invalid when we actually have a
 748                  * hit in kernel space and then try to load it. So if we get
 749                  * here and there are _no_ symbols in the DSO backing the
 750                  * kernel map, bail out.
 751                  *
 752                  * We may never get here, for instance, if we use -K/
 753                  * --hide-kernel-symbols, even if the user specifies an
 754                  * invalid --vmlinux ;-)
 755                  */
 756                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
 757                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
 758                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
 759                         if (symbol_conf.vmlinux_name) {
 760                                 ui__warning("The %s file can't be used.\n%s",
 761                                             symbol_conf.vmlinux_name, msg);
 762                         } else {
 763                                 ui__warning("A vmlinux file was not found.\n%s",
 764                                             msg);
 765                         }
 766
 767                         if (use_browser <= 0)
 768                                 sleep(5);
 769                         top->vmlinux_warned = true;
 770                 }
 771         }
 772
 773         if (al.sym == NULL || !al.sym->ignore) {
 774                 struct hists *hists = evsel__hists(evsel);
 775                 struct hist_entry_iter iter = {
 776                         .add_entry_cb = hist_iter__top_callback,
 777                 };
 778
 779                 if (symbol_conf.cumulate_callchain)
 780                         iter.ops = &hist_iter_cumulative;
 781                 else
 782                         iter.ops = &hist_iter_normal;
 783
 784                 pthread_mutex_lock(&hists->lock);
 785
 786                 err = hist_entry_iter__add(&iter, &al, evsel, sample,
 787                                            top->max_stack, top);
 788                 if (err < 0)
 789                         pr_err("Problem incrementing symbol period, skipping event\n");
 790
 791                 pthread_mutex_unlock(&hists->lock);
 792         }
 793
 794         return;
 795 }
 796
 797 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 798 {
 799         struct perf_sample sample;
 800         struct perf_evsel *evsel;
 801         struct perf_session *session = top->session;
 802         union perf_event *event;
 803         struct machine *machine;
 804         u8 origin;
 805         int ret;
 806
 807         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
 808                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
 809                 if (ret) {
 810                         pr_err("Can't parse sample, err = %d\n", ret);
 811                         goto next_event;
 812                 }
 813
 814                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 815                 assert(evsel != NULL);
 816
 817                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 818
 819                 if (event->header.type == PERF_RECORD_SAMPLE)
 820                         ++top->samples;
 821
 822                 switch (origin) {
 823                 case PERF_RECORD_MISC_USER:
 824                         ++top->us_samples;
 825                         if (top->hide_user_symbols)
 826                                 goto next_event;
 827                         machine = &session->machines.host;
 828                         break;
 829                 case PERF_RECORD_MISC_KERNEL:
 830                         ++top->kernel_samples;
 831                         if (top->hide_kernel_symbols)
 832                                 goto next_event;
 833                         machine = &session->machines.host;
 834                         break;
 835                 case PERF_RECORD_MISC_GUEST_KERNEL:
 836                         ++top->guest_kernel_samples;
 837                         machine = perf_session__find_machine(session,
 838                                                              sample.pid);
 839                         break;
 840                 case PERF_RECORD_MISC_GUEST_USER:
 841                         ++top->guest_us_samples;
 842                         /*
 843                          * TODO: we don't process guest user from host side
 844                          * except simple counting.
 845                          */
 846                         /* Fall thru */
 847                 default:
 848                         goto next_event;
 849                 }
 850
 851
 852                 if (event->header.type == PERF_RECORD_SAMPLE) {
 853                         perf_event__process_sample(&top->tool, event, evsel,
 854                                                    &sample, machine);
 855                 } else if (event->header.type < PERF_RECORD_MAX) {
 856                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 857                         machine__process_event(machine, event, &sample);
 858                 } else
 859                         ++session->stats.nr_unknown_events;
 860 next_event:
 861                 perf_evlist__mmap_consume(top->evlist, idx);
 862         }
 863 }
 864
 865 static void perf_top__mmap_read(struct perf_top *top)
 866 {
 867         int i;
 868
 869         for (i = 0; i < top->evlist->nr_mmaps; i++)
 870                 perf_top__mmap_read_idx(top, i);
 871 }
 872
 873 static int perf_top__start_counters(struct perf_top *top)
 874 {
 875         char msg[512];
 876         struct perf_evsel *counter;
 877         struct perf_evlist *evlist = top->evlist;
 878         struct record_opts *opts = &top->record_opts;
 879
 880         perf_evlist__config(evlist, opts);
 881
 882         evlist__for_each(evlist, counter) {
 883 try_again:
 884                 if (perf_evsel__open(counter, top->evlist->cpus,
 885                                      top->evlist->threads) < 0) {
 886                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 887                                 if (verbose)
 888                                         ui__warning("%s\n", msg);
 889                                 goto try_again;
 890                         }
 891
 892                         perf_evsel__open_strerror(counter, &opts->target,
 893                                                   errno, msg, sizeof(msg));
 894                         ui__error("%s\n", msg);
 895                         goto out_err;
 896                 }
 897         }
 898
 899         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 900                 ui__error("Failed to mmap with %d (%s)\n",
 901                             errno, strerror_r(errno, msg, sizeof(msg)));
 902                 goto out_err;
 903         }
 904
 905         return 0;
 906
 907 out_err:
 908         return -1;
 909 }
 910
 911 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
 912 {
 913         if (!sort__has_sym) {
 914                 if (symbol_conf.use_callchain) {
 915                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
 916                         return -EINVAL;
 917                 }
 918         } else if (callchain_param.mode != CHAIN_NONE) {
 919                 if (callchain_register_param(&callchain_param) < 0) {
 920                         ui__error("Can't register callchain params.\n");
 921                         return -EINVAL;
 922                 }
 923         }
 924
 925         return 0;
 926 }
 927
 928 static int __cmd_top(struct perf_top *top)
 929 {
 930         struct record_opts *opts = &top->record_opts;
 931         pthread_t thread;
 932         int ret;
 933
 934         top->session = perf_session__new(NULL, false, NULL);
 935         if (top->session == NULL)
 936                 return -1;
 937
 938         machines__set_symbol_filter(&top->session->machines, symbol_filter);
 939
 940         if (!objdump_path) {
 941                 ret = perf_session_env__lookup_objdump(&top->session->header.env);
 942                 if (ret)
 943                         goto out_delete;
 944         }
 945
 946         ret = perf_top__setup_sample_type(top);
 947         if (ret)
 948                 goto out_delete;
 949
 950         machine__synthesize_threads(&top->session->machines.host, &opts->target,
 951                                     top->evlist->threads, false);
 952         ret = perf_top__start_counters(top);
 953         if (ret)
 954                 goto out_delete;
 955
 956         top->session->evlist = top->evlist;
 957         perf_session__set_id_hdr_size(top->session);
 958
 959         /*
 960          * When perf is starting the traced process, all the events (apart from
 961          * group members) have enable_on_exec=1 set, so don't spoil it by
 962          * prematurely enabling them.
 963          *
 964          * XXX 'top' still doesn't start workloads like record, trace, but should,
 965          * so leave the check here.
 966          */
 967         if (!target__none(&opts->target))
 968                 perf_evlist__enable(top->evlist);
 969
 970         /* Wait for a minimal set of events before starting the snapshot */
 971         perf_evlist__poll(top->evlist, 100);
 972
 973         perf_top__mmap_read(top);
 974
 975         ret = -1;
 976         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 977                                                             display_thread), top)) {
 978                 ui__error("Could not create display thread.\n");
 979                 goto out_delete;
 980         }
 981
 982         if (top->realtime_prio) {
 983                 struct sched_param param;
 984
 985                 param.sched_priority = top->realtime_prio;
 986                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 987                         ui__error("Could not set realtime priority.\n");
 988                         goto out_join;
 989                 }
 990         }
 991
 992         while (!done) {
 993                 u64 hits = top->samples;
 994
 995                 perf_top__mmap_read(top);
 996
 997                 if (hits == top->samples)
 998                         ret = perf_evlist__poll(top->evlist, 100);
 999         }
1000
1001         ret = 0;
1002 out_join:
1003         pthread_join(thread, NULL);
1004 out_delete:
1005         perf_session__delete(top->session);
1006         top->session = NULL;
1007
1008         return ret;
1009 }
1010
1011 static int
1012 callchain_opt(const struct option *opt, const char *arg, int unset)
1013 {
1014         symbol_conf.use_callchain = true;
1015         return record_callchain_opt(opt, arg, unset);
1016 }
1017
1018 static int
1019 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1020 {
1021         symbol_conf.use_callchain = true;
1022         return record_parse_callchain_opt(opt, arg, unset);
1023 }
1024
1025 static int perf_top_config(const char *var, const char *value, void *cb)
1026 {
1027         if (!strcmp(var, "top.call-graph"))
1028                 var = "call-graph.record-mode"; /* fall-through */
1029         if (!strcmp(var, "top.children")) {
1030                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1031                 return 0;
1032         }
1033
1034         return perf_default_config(var, value, cb);
1035 }
1036
1037 static int
1038 parse_percent_limit(const struct option *opt, const char *arg,
1039                     int unset __maybe_unused)
1040 {
1041         struct perf_top *top = opt->value;
1042
1043         top->min_percent = strtof(arg, NULL);
1044         return 0;
1045 }
1046
1047 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1048 {
1049         char errbuf[BUFSIZ];
1050         struct perf_top top = {
1051                 .count_filter        = 5,
1052                 .delay_secs          = 2,
1053                 .record_opts = {
1054                         .mmap_pages     = UINT_MAX,
1055                         .user_freq      = UINT_MAX,
1056                         .user_interval  = ULLONG_MAX,
1057                         .freq           = 4000, /* 4 KHz */
1058                         .target         = {
1059                                 .uses_mmap   = true,
1060                         },
1061                 },
1062                 .max_stack           = PERF_MAX_STACK_DEPTH,
1063                 .sym_pcnt_filter     = 5,
1064         };
1065         struct record_opts *opts = &top.record_opts;
1066         struct target *target = &opts->target;
1067         const struct option options[] = {
1068         OPT_CALLBACK('e', "event", &top.evlist, "event",
1069                      "event selector. use 'perf list' to list available events",
1070                      parse_events_option),
1071         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1072         OPT_STRING('p', "pid", &target->pid, "pid",
1073                     "profile events on existing process id"),
1074         OPT_STRING('t', "tid", &target->tid, "tid",
1075                     "profile events on existing thread id"),
1076         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1077                             "system-wide collection from all CPUs"),
1078         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1079                     "list of cpus to monitor"),
1080         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1081                    "file", "vmlinux pathname"),
1082         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1083                     "don't load vmlinux even if found"),
1084         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1085                     "hide kernel symbols"),
1086         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1087                      "number of mmap data pages",
1088                      perf_evlist__parse_mmap_pages),
1089         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1090                     "collect data with this RT SCHED_FIFO priority"),
1091         OPT_INTEGER('d', "delay", &top.delay_secs,
1092                     "number of seconds to delay between refreshes"),
1093         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1094                             "dump the symbol table used for profiling"),
1095         OPT_INTEGER('f', "count-filter", &top.count_filter,
1096                     "only display functions with more events than this"),
1097         OPT_BOOLEAN(0, "group", &opts->group,
1098                             "put the counters into a counter group"),
1099         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1100                     "child tasks do not inherit counters"),
1101         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1102                     "symbol to annotate"),
1103         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1104         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1105         OPT_INTEGER('E', "entries", &top.print_entries,
1106                     "display this many functions"),
1107         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1108                     "hide user symbols"),
1109         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1110         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1111         OPT_INCR('v', "verbose", &verbose,
1112                     "be more verbose (show counter open errors, etc)"),
1113         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1114                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1115                    " Please refer the man page for the complete list."),
1116         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1117                    "output field(s): overhead, period, sample plus all of sort keys"),
1118         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1119                     "Show a column with the number of samples"),
1120         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1121                            NULL, "enables call-graph recording",
1122                            &callchain_opt),
1123         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1124                      "mode[,dump_size]", record_callchain_help,
1125                      &parse_callchain_opt),
1126         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1127                     "Accumulate callchains of children and show total overhead as well"),
1128         OPT_INTEGER(0, "max-stack", &top.max_stack,
1129                     "Set the maximum stack depth when parsing the callchain. "
1130                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1131         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1132                    "ignore callees of these functions in call graphs",
1133                    report_parse_ignore_callees_opt),
1134         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1135                     "Show a column with the sum of periods"),
1136         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1137                    "only consider symbols in these dsos"),
1138         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1139                    "only consider symbols in these comms"),
1140         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1141                    "only consider these symbols"),
1142         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1143                     "Interleave source code with assembly code (default)"),
1144         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1145                     "Display raw encoding of assembly instructions (default)"),
1146         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1147                     "Enable kernel symbol demangling"),
1148         OPT_STRING(0, "objdump", &objdump_path, "path",
1149                     "objdump binary to use for disassembly and annotations"),
1150         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1151                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1152         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1153         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1154                      "Don't show entries under that percent", parse_percent_limit),
1155         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1156                      "How to display percentage of filtered entries", parse_filter_percentage),
1157         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1158                    "width[,width...]",
1159                    "don't try to adjust column width, use these fixed values"),
1160         OPT_END()
1161         };
1162         const char * const top_usage[] = {
1163                 "perf top [<options>]",
1164                 NULL
1165         };
1166         int status = hists__init();
1167
1168         if (status < 0)
1169                 return status;
1170
1171         top.evlist = perf_evlist__new();
1172         if (top.evlist == NULL)
1173                 return -ENOMEM;
1174
1175         perf_config(perf_top_config, &top);
1176
1177         argc = parse_options(argc, argv, options, top_usage, 0);
1178         if (argc)
1179                 usage_with_options(top_usage, options);
1180
1181         sort__mode = SORT_MODE__TOP;
1182         /* display thread wants entries to be collapsed in a different tree */
1183         sort__need_collapse = 1;
1184
1185         if (setup_sorting() < 0) {
1186                 if (sort_order)
1187                         parse_options_usage(top_usage, options, "s", 1);
1188                 if (field_order)
1189                         parse_options_usage(sort_order ? NULL : top_usage,
1190                                             options, "fields", 0);
1191                 goto out_delete_evlist;
1192         }
1193
1194         if (top.use_stdio)
1195                 use_browser = 0;
1196         else if (top.use_tui)
1197                 use_browser = 1;
1198
1199         setup_browser(false);
1200
1201         status = target__validate(target);
1202         if (status) {
1203                 target__strerror(target, status, errbuf, BUFSIZ);
1204                 ui__warning("%s\n", errbuf);
1205         }
1206
1207         status = target__parse_uid(target);
1208         if (status) {
1209                 int saved_errno = errno;
1210
1211                 target__strerror(target, status, errbuf, BUFSIZ);
1212                 ui__error("%s\n", errbuf);
1213
1214                 status = -saved_errno;
1215                 goto out_delete_evlist;
1216         }
1217
1218         if (target__none(target))
1219                 target->system_wide = true;
1220
1221         if (perf_evlist__create_maps(top.evlist, target) < 0)
1222                 usage_with_options(top_usage, options);
1223
1224         if (!top.evlist->nr_entries &&
1225             perf_evlist__add_default(top.evlist) < 0) {
1226                 ui__error("Not enough memory for event selector list\n");
1227                 goto out_delete_evlist;
1228         }
1229
1230         symbol_conf.nr_events = top.evlist->nr_entries;
1231
1232         if (top.delay_secs < 1)
1233                 top.delay_secs = 1;
1234
1235         if (record_opts__config(opts)) {
1236                 status = -EINVAL;
1237                 goto out_delete_evlist;
1238         }
1239
1240         top.sym_evsel = perf_evlist__first(top.evlist);
1241
1242         if (!symbol_conf.use_callchain) {
1243                 symbol_conf.cumulate_callchain = false;
1244                 perf_hpp__cancel_cumulate();
1245         }
1246
1247         symbol_conf.priv_size = sizeof(struct annotation);
1248
1249         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1250         if (symbol__init(NULL) < 0)
1251                 return -1;
1252
1253         sort__setup_elide(stdout);
1254
1255         get_term_dimensions(&top.winsize);
1256         if (top.print_entries == 0) {
1257                 struct sigaction act = {
1258                         .sa_sigaction = perf_top__sig_winch,
1259                         .sa_flags     = SA_SIGINFO,
1260                 };
1261                 perf_top__update_print_entries(&top);
1262                 sigaction(SIGWINCH, &act, NULL);
1263         }
1264
1265         status = __cmd_top(&top);
1266
1267 out_delete_evlist:
1268         perf_evlist__delete(top.evlist);
1269
1270         return status;
1271 }