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/config.h"
  26 #include "util/color.h"
  27 #include "util/drv_configs.h"
  28 #include "util/evlist.h"
  29 #include "util/evsel.h"
  30 #include "util/event.h"
  31 #include "util/machine.h"
  32 #include "util/session.h"
  33 #include "util/symbol.h"
  34 #include "util/thread.h"
  35 #include "util/thread_map.h"
  36 #include "util/top.h"
  37 #include <linux/rbtree.h>
  38 #include <subcmd/parse-options.h>
  39 #include "util/parse-events.h"
  40 #include "util/cpumap.h"
  41 #include "util/xyarray.h"
  42 #include "util/sort.h"
  43 #include "util/term.h"
  44 #include "util/intlist.h"
  45 #include "util/parse-branch-options.h"
  46 #include "arch/common.h"
  47
  48 #include "util/debug.h"
  49 #include "util/ordered-events.h"
  50
  51 #include <assert.h>
  52 #include <elf.h>
  53 #include <fcntl.h>
  54
  55 #include <stdio.h>
  56 #include <termios.h>
  57 #include <unistd.h>
  58 #include <inttypes.h>
  59
  60 #include <errno.h>
  61 #include <time.h>
  62 #include <sched.h>
  63 #include <signal.h>
  64
  65 #include <sys/syscall.h>
  66 #include <sys/ioctl.h>
  67 #include <poll.h>
  68 #include <sys/prctl.h>
  69 #include <sys/wait.h>
  70 #include <sys/uio.h>
  71 #include <sys/utsname.h>
  72 #include <sys/mman.h>
  73
  74 #include <linux/stringify.h>
  75 #include <linux/time64.h>
  76 #include <linux/types.h>
  77
  78 #include "sane_ctype.h"
  79
  80 static volatile int done;
  81 static volatile int resize;
  82
  83 #define HEADER_LINE_NR  5
  84
  85 static void perf_top__update_print_entries(struct perf_top *top)
  86 {
  87         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  88 }
  89
  90 static void winch_sig(int sig __maybe_unused)
  91 {
  92         resize = 1;
  93 }
  94
  95 static void perf_top__resize(struct perf_top *top)
  96 {
  97         get_term_dimensions(&top->winsize);
  98         perf_top__update_print_entries(top);
  99 }
 100
 101 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 102 {
 103         struct perf_evsel *evsel = hists_to_evsel(he->hists);
 104         struct symbol *sym;
 105         struct annotation *notes;
 106         struct map *map;
 107         int err = -1;
 108
 109         if (!he || !he->ms.sym)
 110                 return -1;
 111
 112         sym = he->ms.sym;
 113         map = he->ms.map;
 114
 115         /*
 116          * We can't annotate with just /proc/kallsyms
 117          */
 118         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 119             !dso__is_kcore(map->dso)) {
 120                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 121                        "path\n", sym->name);
 122                 sleep(1);
 123                 return -1;
 124         }
 125
 126         notes = symbol__annotation(sym);
 127         pthread_mutex_lock(&notes->lock);
 128
 129         if (!symbol__hists(sym, top->evlist->nr_entries)) {
 130                 pthread_mutex_unlock(&notes->lock);
 131                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 132                        sym->name);
 133                 sleep(1);
 134                 return err;
 135         }
 136
 137         err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
 138         if (err == 0) {
 139                 top->sym_filter_entry = he;
 140         } else {
 141                 char msg[BUFSIZ];
 142                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
 143                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 144         }
 145
 146         pthread_mutex_unlock(&notes->lock);
 147         return err;
 148 }
 149
 150 static void __zero_source_counters(struct hist_entry *he)
 151 {
 152         struct symbol *sym = he->ms.sym;
 153         symbol__annotate_zero_histograms(sym);
 154 }
 155
 156 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 157 {
 158         struct utsname uts;
 159         int err = uname(&uts);
 160
 161         ui__warning("Out of bounds address found:\n\n"
 162                     "Addr:   %" PRIx64 "\n"
 163                     "DSO:    %s %c\n"
 164                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 165                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 166                     "Arch:   %s\n"
 167                     "Kernel: %s\n"
 168                     "Tools:  %s\n\n"
 169                     "Not all samples will be on the annotation output.\n\n"
 170                     "Please report to linux-kernel@vger.kernel.org\n",
 171                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 172                     map->start, map->end, sym->start, sym->end,
 173                     sym->binding == STB_GLOBAL ? 'g' :
 174                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 175                     err ? "[unknown]" : uts.machine,
 176                     err ? "[unknown]" : uts.release, perf_version_string);
 177         if (use_browser <= 0)
 178                 sleep(5);
 179
 180         map->erange_warned = true;
 181 }
 182
 183 static void perf_top__record_precise_ip(struct perf_top *top,
 184                                         struct hist_entry *he,
 185                                         struct perf_sample *sample,
 186                                         struct perf_evsel *evsel, u64 ip)
 187 {
 188         struct annotation *notes;
 189         struct symbol *sym = he->ms.sym;
 190         int err = 0;
 191
 192         if (sym == NULL || (use_browser == 0 &&
 193                             (top->sym_filter_entry == NULL ||
 194                              top->sym_filter_entry->ms.sym != sym)))
 195                 return;
 196
 197         notes = symbol__annotation(sym);
 198
 199         if (pthread_mutex_trylock(&notes->lock))
 200                 return;
 201
 202         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 203
 204         pthread_mutex_unlock(&notes->lock);
 205
 206         if (unlikely(err)) {
 207                 /*
 208                  * This function is now called with he->hists->lock held.
 209                  * Release it before going to sleep.
 210                  */
 211                 pthread_mutex_unlock(&he->hists->lock);
 212
 213                 if (err == -ERANGE && !he->ms.map->erange_warned)
 214                         ui__warn_map_erange(he->ms.map, sym, ip);
 215                 else if (err == -ENOMEM) {
 216                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 217                                sym->name);
 218                         sleep(1);
 219                 }
 220
 221                 pthread_mutex_lock(&he->hists->lock);
 222         }
 223 }
 224
 225 static void perf_top__show_details(struct perf_top *top)
 226 {
 227         struct hist_entry *he = top->sym_filter_entry;
 228         struct perf_evsel *evsel = hists_to_evsel(he->hists);
 229         struct annotation *notes;
 230         struct symbol *symbol;
 231         int more;
 232
 233         if (!he)
 234                 return;
 235
 236         symbol = he->ms.sym;
 237         notes = symbol__annotation(symbol);
 238
 239         pthread_mutex_lock(&notes->lock);
 240
 241         symbol__calc_percent(symbol, evsel);
 242
 243         if (notes->src == NULL)
 244                 goto out_unlock;
 245
 246         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 247         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 248
 249         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
 250
 251         if (top->evlist->enabled) {
 252                 if (top->zero)
 253                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 254                 else
 255                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 256         }
 257         if (more != 0)
 258                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 259 out_unlock:
 260         pthread_mutex_unlock(&notes->lock);
 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         struct perf_evsel *evsel = top->sym_evsel;
 269         struct hists *hists = evsel__hists(evsel);
 270
 271         puts(CONSOLE_CLEAR);
 272
 273         perf_top__header_snprintf(top, bf, sizeof(bf));
 274         printf("%s\n", bf);
 275
 276         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 277
 278         if (!top->record_opts.overwrite &&
 279             (hists->stats.nr_lost_warned !=
 280             hists->stats.nr_events[PERF_RECORD_LOST])) {
 281                 hists->stats.nr_lost_warned =
 282                               hists->stats.nr_events[PERF_RECORD_LOST];
 283                 color_fprintf(stdout, PERF_COLOR_RED,
 284                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 285                               hists->stats.nr_lost_warned);
 286                 ++printed;
 287         }
 288
 289         if (top->sym_filter_entry) {
 290                 perf_top__show_details(top);
 291                 return;
 292         }
 293
 294         if (top->evlist->enabled) {
 295                 if (top->zero) {
 296                         hists__delete_entries(hists);
 297                 } else {
 298                         hists__decay_entries(hists, top->hide_user_symbols,
 299                                              top->hide_kernel_symbols);
 300                 }
 301         }
 302
 303         hists__collapse_resort(hists, NULL);
 304         perf_evsel__output_resort(evsel, NULL);
 305
 306         hists__output_recalc_col_len(hists, top->print_entries - printed);
 307         putchar('\n');
 308         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 309                        top->min_percent, stdout, !symbol_conf.use_callchain);
 310 }
 311
 312 static void prompt_integer(int *target, const char *msg)
 313 {
 314         char *buf = malloc(0), *p;
 315         size_t dummy = 0;
 316         int tmp;
 317
 318         fprintf(stdout, "\n%s: ", msg);
 319         if (getline(&buf, &dummy, stdin) < 0)
 320                 return;
 321
 322         p = strchr(buf, '\n');
 323         if (p)
 324                 *p = 0;
 325
 326         p = buf;
 327         while(*p) {
 328                 if (!isdigit(*p))
 329                         goto out_free;
 330                 p++;
 331         }
 332         tmp = strtoul(buf, NULL, 10);
 333         *target = tmp;
 334 out_free:
 335         free(buf);
 336 }
 337
 338 static void prompt_percent(int *target, const char *msg)
 339 {
 340         int tmp = 0;
 341
 342         prompt_integer(&tmp, msg);
 343         if (tmp >= 0 && tmp <= 100)
 344                 *target = tmp;
 345 }
 346
 347 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 348 {
 349         char *buf = malloc(0), *p;
 350         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 351         struct hists *hists = evsel__hists(top->sym_evsel);
 352         struct rb_node *next;
 353         size_t dummy = 0;
 354
 355         /* zero counters of active symbol */
 356         if (syme) {
 357                 __zero_source_counters(syme);
 358                 top->sym_filter_entry = NULL;
 359         }
 360
 361         fprintf(stdout, "\n%s: ", msg);
 362         if (getline(&buf, &dummy, stdin) < 0)
 363                 goto out_free;
 364
 365         p = strchr(buf, '\n');
 366         if (p)
 367                 *p = 0;
 368
 369         next = rb_first(&hists->entries);
 370         while (next) {
 371                 n = rb_entry(next, struct hist_entry, rb_node);
 372                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 373                         found = n;
 374                         break;
 375                 }
 376                 next = rb_next(&n->rb_node);
 377         }
 378
 379         if (!found) {
 380                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 381                 sleep(1);
 382         } else
 383                 perf_top__parse_source(top, found);
 384
 385 out_free:
 386         free(buf);
 387 }
 388
 389 static void perf_top__print_mapped_keys(struct perf_top *top)
 390 {
 391         char *name = NULL;
 392
 393         if (top->sym_filter_entry) {
 394                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 395                 name = sym->name;
 396         }
 397
 398         fprintf(stdout, "\nMapped keys:\n");
 399         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 400         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 401
 402         if (top->evlist->nr_entries > 1)
 403                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 404
 405         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 406
 407         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 408         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 409         fprintf(stdout, "\t[S]     stop annotation.\n");
 410
 411         fprintf(stdout,
 412                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 413                 top->hide_kernel_symbols ? "yes" : "no");
 414         fprintf(stdout,
 415                 "\t[U]     hide user symbols.               \t(%s)\n",
 416                 top->hide_user_symbols ? "yes" : "no");
 417         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 418         fprintf(stdout, "\t[qQ]    quit.\n");
 419 }
 420
 421 static int perf_top__key_mapped(struct perf_top *top, int c)
 422 {
 423         switch (c) {
 424                 case 'd':
 425                 case 'e':
 426                 case 'f':
 427                 case 'z':
 428                 case 'q':
 429                 case 'Q':
 430                 case 'K':
 431                 case 'U':
 432                 case 'F':
 433                 case 's':
 434                 case 'S':
 435                         return 1;
 436                 case 'E':
 437                         return top->evlist->nr_entries > 1 ? 1 : 0;
 438                 default:
 439                         break;
 440         }
 441
 442         return 0;
 443 }
 444
 445 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 446 {
 447         bool ret = true;
 448
 449         if (!perf_top__key_mapped(top, c)) {
 450                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 451                 struct termios save;
 452
 453                 perf_top__print_mapped_keys(top);
 454                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 455                 fflush(stdout);
 456
 457                 set_term_quiet_input(&save);
 458
 459                 poll(&stdin_poll, 1, -1);
 460                 c = getc(stdin);
 461
 462                 tcsetattr(0, TCSAFLUSH, &save);
 463                 if (!perf_top__key_mapped(top, c))
 464                         return ret;
 465         }
 466
 467         switch (c) {
 468                 case 'd':
 469                         prompt_integer(&top->delay_secs, "Enter display delay");
 470                         if (top->delay_secs < 1)
 471                                 top->delay_secs = 1;
 472                         break;
 473                 case 'e':
 474                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 475                         if (top->print_entries == 0) {
 476                                 perf_top__resize(top);
 477                                 signal(SIGWINCH, winch_sig);
 478                         } else {
 479                                 signal(SIGWINCH, SIG_DFL);
 480                         }
 481                         break;
 482                 case 'E':
 483                         if (top->evlist->nr_entries > 1) {
 484                                 /* Select 0 as the default event: */
 485                                 int counter = 0;
 486
 487                                 fprintf(stderr, "\nAvailable events:");
 488
 489                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 490                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 491
 492                                 prompt_integer(&counter, "Enter details event counter");
 493
 494                                 if (counter >= top->evlist->nr_entries) {
 495                                         top->sym_evsel = perf_evlist__first(top->evlist);
 496                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 497                                         sleep(1);
 498                                         break;
 499                                 }
 500                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 501                                         if (top->sym_evsel->idx == counter)
 502                                                 break;
 503                         } else
 504                                 top->sym_evsel = perf_evlist__first(top->evlist);
 505                         break;
 506                 case 'f':
 507                         prompt_integer(&top->count_filter, "Enter display event count filter");
 508                         break;
 509                 case 'F':
 510                         prompt_percent(&top->annotation_opts.min_pcnt,
 511                                        "Enter details display event filter (percent)");
 512                         break;
 513                 case 'K':
 514                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 515                         break;
 516                 case 'q':
 517                 case 'Q':
 518                         printf("exiting.\n");
 519                         if (top->dump_symtab)
 520                                 perf_session__fprintf_dsos(top->session, stderr);
 521                         ret = false;
 522                         break;
 523                 case 's':
 524                         perf_top__prompt_symbol(top, "Enter details symbol");
 525                         break;
 526                 case 'S':
 527                         if (!top->sym_filter_entry)
 528                                 break;
 529                         else {
 530                                 struct hist_entry *syme = top->sym_filter_entry;
 531
 532                                 top->sym_filter_entry = NULL;
 533                                 __zero_source_counters(syme);
 534                         }
 535                         break;
 536                 case 'U':
 537                         top->hide_user_symbols = !top->hide_user_symbols;
 538                         break;
 539                 case 'z':
 540                         top->zero = !top->zero;
 541                         break;
 542                 default:
 543                         break;
 544         }
 545
 546         return ret;
 547 }
 548
 549 static void perf_top__sort_new_samples(void *arg)
 550 {
 551         struct perf_top *t = arg;
 552         struct perf_evsel *evsel = t->sym_evsel;
 553         struct hists *hists;
 554
 555         if (t->evlist->selected != NULL)
 556                 t->sym_evsel = t->evlist->selected;
 557
 558         hists = evsel__hists(evsel);
 559
 560         if (t->evlist->enabled) {
 561                 if (t->zero) {
 562                         hists__delete_entries(hists);
 563                 } else {
 564                         hists__decay_entries(hists, t->hide_user_symbols,
 565                                              t->hide_kernel_symbols);
 566                 }
 567         }
 568
 569         hists__collapse_resort(hists, NULL);
 570         perf_evsel__output_resort(evsel, NULL);
 571
 572         if (t->lost || t->drop)
 573                 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 574 }
 575
 576 static void stop_top(void)
 577 {
 578         session_done = 1;
 579         done = 1;
 580 }
 581
 582 static void *display_thread_tui(void *arg)
 583 {
 584         struct perf_evsel *pos;
 585         struct perf_top *top = arg;
 586         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 587         struct hist_browser_timer hbt = {
 588                 .timer          = perf_top__sort_new_samples,
 589                 .arg            = top,
 590                 .refresh        = top->delay_secs,
 591         };
 592
 593         /* In order to read symbols from other namespaces perf to  needs to call
 594          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 595          * unshare(2) the fs so that we may continue to setns into namespaces
 596          * that we're observing.
 597          */
 598         unshare(CLONE_FS);
 599
 600         perf_top__sort_new_samples(top);
 601
 602         /*
 603          * Initialize the uid_filter_str, in the future the TUI will allow
 604          * Zooming in/out UIDs. For now just use whatever the user passed
 605          * via --uid.
 606          */
 607         evlist__for_each_entry(top->evlist, pos) {
 608                 struct hists *hists = evsel__hists(pos);
 609                 hists->uid_filter_str = top->record_opts.target.uid_str;
 610         }
 611
 612         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 613                                       top->min_percent,
 614                                       &top->session->header.env,
 615                                       !top->record_opts.overwrite,
 616                                       &top->annotation_opts);
 617
 618         stop_top();
 619         return NULL;
 620 }
 621
 622 static void display_sig(int sig __maybe_unused)
 623 {
 624         stop_top();
 625 }
 626
 627 static void display_setup_sig(void)
 628 {
 629         signal(SIGSEGV, sighandler_dump_stack);
 630         signal(SIGFPE, sighandler_dump_stack);
 631         signal(SIGINT,  display_sig);
 632         signal(SIGQUIT, display_sig);
 633         signal(SIGTERM, display_sig);
 634 }
 635
 636 static void *display_thread(void *arg)
 637 {
 638         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 639         struct termios save;
 640         struct perf_top *top = arg;
 641         int delay_msecs, c;
 642
 643         /* In order to read symbols from other namespaces perf to  needs to call
 644          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 645          * unshare(2) the fs so that we may continue to setns into namespaces
 646          * that we're observing.
 647          */
 648         unshare(CLONE_FS);
 649
 650         display_setup_sig();
 651         pthread__unblock_sigwinch();
 652 repeat:
 653         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 654         set_term_quiet_input(&save);
 655         /* trash return*/
 656         getc(stdin);
 657
 658         while (!done) {
 659                 perf_top__print_sym_table(top);
 660                 /*
 661                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 662                  * refresh screen in both cases.
 663                  */
 664                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 665                 case 0:
 666                         continue;
 667                 case -1:
 668                         if (errno == EINTR)
 669                                 continue;
 670                         __fallthrough;
 671                 default:
 672                         c = getc(stdin);
 673                         tcsetattr(0, TCSAFLUSH, &save);
 674
 675                         if (perf_top__handle_keypress(top, c))
 676                                 goto repeat;
 677                         stop_top();
 678                 }
 679         }
 680
 681         tcsetattr(0, TCSAFLUSH, &save);
 682         return NULL;
 683 }
 684
 685 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 686                                    struct addr_location *al, bool single,
 687                                    void *arg)
 688 {
 689         struct perf_top *top = arg;
 690         struct hist_entry *he = iter->he;
 691         struct perf_evsel *evsel = iter->evsel;
 692
 693         if (perf_hpp_list.sym && single)
 694                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 695
 696         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 697                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 698         return 0;
 699 }
 700
 701 static void perf_event__process_sample(struct perf_tool *tool,
 702                                        const union perf_event *event,
 703                                        struct perf_evsel *evsel,
 704                                        struct perf_sample *sample,
 705                                        struct machine *machine)
 706 {
 707         struct perf_top *top = container_of(tool, struct perf_top, tool);
 708         struct addr_location al;
 709         int err;
 710
 711         if (!machine && perf_guest) {
 712                 static struct intlist *seen;
 713
 714                 if (!seen)
 715                         seen = intlist__new(NULL);
 716
 717                 if (!intlist__has_entry(seen, sample->pid)) {
 718                         pr_err("Can't find guest [%d]'s kernel information\n",
 719                                 sample->pid);
 720                         intlist__add(seen, sample->pid);
 721                 }
 722                 return;
 723         }
 724
 725         if (!machine) {
 726                 pr_err("%u unprocessable samples recorded.\r",
 727                        top->session->evlist->stats.nr_unprocessable_samples++);
 728                 return;
 729         }
 730
 731         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 732                 top->exact_samples++;
 733
 734         if (machine__resolve(machine, &al, sample) < 0)
 735                 return;
 736
 737         if (!machine->kptr_restrict_warned &&
 738             symbol_conf.kptr_restrict &&
 739             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 740                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 741                         ui__warning(
 742 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 743 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 744 "Kernel%s samples will not be resolved.\n",
 745                           al.map && map__has_symbols(al.map) ?
 746                           " modules" : "");
 747                         if (use_browser <= 0)
 748                                 sleep(5);
 749                 }
 750                 machine->kptr_restrict_warned = true;
 751         }
 752
 753         if (al.sym == NULL && al.map != NULL) {
 754                 const char *msg = "Kernel samples will not be resolved.\n";
 755                 /*
 756                  * As we do lazy loading of symtabs we only will know if the
 757                  * specified vmlinux file is invalid when we actually have a
 758                  * hit in kernel space and then try to load it. So if we get
 759                  * here and there are _no_ symbols in the DSO backing the
 760                  * kernel map, bail out.
 761                  *
 762                  * We may never get here, for instance, if we use -K/
 763                  * --hide-kernel-symbols, even if the user specifies an
 764                  * invalid --vmlinux ;-)
 765                  */
 766                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 767                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 768                         if (symbol_conf.vmlinux_name) {
 769                                 char serr[256];
 770                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 771                                 ui__warning("The %s file can't be used: %s\n%s",
 772                                             symbol_conf.vmlinux_name, serr, msg);
 773                         } else {
 774                                 ui__warning("A vmlinux file was not found.\n%s",
 775                                             msg);
 776                         }
 777
 778                         if (use_browser <= 0)
 779                                 sleep(5);
 780                         top->vmlinux_warned = true;
 781                 }
 782         }
 783
 784         if (al.sym == NULL || !al.sym->idle) {
 785                 struct hists *hists = evsel__hists(evsel);
 786                 struct hist_entry_iter iter = {
 787                         .evsel          = evsel,
 788                         .sample         = sample,
 789                         .add_entry_cb   = hist_iter__top_callback,
 790                 };
 791
 792                 if (symbol_conf.cumulate_callchain)
 793                         iter.ops = &hist_iter_cumulative;
 794                 else
 795                         iter.ops = &hist_iter_normal;
 796
 797                 pthread_mutex_lock(&hists->lock);
 798
 799                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 800                 if (err < 0)
 801                         pr_err("Problem incrementing symbol period, skipping event\n");
 802
 803                 pthread_mutex_unlock(&hists->lock);
 804         }
 805
 806         addr_location__put(&al);
 807 }
 808
 809 static void
 810 perf_top__process_lost(struct perf_top *top, union perf_event *event,
 811                        struct perf_evsel *evsel)
 812 {
 813         struct hists *hists = evsel__hists(evsel);
 814
 815         top->lost += event->lost.lost;
 816         top->lost_total += event->lost.lost;
 817         hists->stats.total_lost += event->lost.lost;
 818 }
 819
 820 static void
 821 perf_top__process_lost_samples(struct perf_top *top,
 822                                union perf_event *event,
 823                                struct perf_evsel *evsel)
 824 {
 825         struct hists *hists = evsel__hists(evsel);
 826
 827         top->lost += event->lost_samples.lost;
 828         top->lost_total += event->lost_samples.lost;
 829         hists->stats.total_lost_samples += event->lost_samples.lost;
 830 }
 831
 832 static u64 last_timestamp;
 833
 834 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 835 {
 836         struct record_opts *opts = &top->record_opts;
 837         struct perf_evlist *evlist = top->evlist;
 838         struct perf_mmap *md;
 839         union perf_event *event;
 840
 841         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 842         if (perf_mmap__read_init(md) < 0)
 843                 return;
 844
 845         while ((event = perf_mmap__read_event(md)) != NULL) {
 846                 int ret;
 847
 848                 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 849                 if (ret && ret != -1)
 850                         break;
 851
 852                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
 853                 if (ret)
 854                         break;
 855
 856                 perf_mmap__consume(md);
 857
 858                 if (top->qe.rotate) {
 859                         pthread_mutex_lock(&top->qe.mutex);
 860                         top->qe.rotate = false;
 861                         pthread_cond_signal(&top->qe.cond);
 862                         pthread_mutex_unlock(&top->qe.mutex);
 863                 }
 864         }
 865
 866         perf_mmap__read_done(md);
 867 }
 868
 869 static void perf_top__mmap_read(struct perf_top *top)
 870 {
 871         bool overwrite = top->record_opts.overwrite;
 872         struct perf_evlist *evlist = top->evlist;
 873         int i;
 874
 875         if (overwrite)
 876                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 877
 878         for (i = 0; i < top->evlist->nr_mmaps; i++)
 879                 perf_top__mmap_read_idx(top, i);
 880
 881         if (overwrite) {
 882                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 883                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 884         }
 885 }
 886
 887 /*
 888  * Check per-event overwrite term.
 889  * perf top should support consistent term for all events.
 890  * - All events don't have per-event term
 891  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 892  *   Nothing change, return 0.
 893  * - All events have same per-event term
 894  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 895  *   Using the per-event setting to replace the opts->overwrite if
 896  *   they are different, then return 0.
 897  * - Events have different per-event term
 898  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 899  *   Return -1
 900  * - Some of the event set per-event term, but some not.
 901  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 902  *   Return -1
 903  */
 904 static int perf_top__overwrite_check(struct perf_top *top)
 905 {
 906         struct record_opts *opts = &top->record_opts;
 907         struct perf_evlist *evlist = top->evlist;
 908         struct perf_evsel_config_term *term;
 909         struct list_head *config_terms;
 910         struct perf_evsel *evsel;
 911         int set, overwrite = -1;
 912
 913         evlist__for_each_entry(evlist, evsel) {
 914                 set = -1;
 915                 config_terms = &evsel->config_terms;
 916                 list_for_each_entry(term, config_terms, list) {
 917                         if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
 918                                 set = term->val.overwrite ? 1 : 0;
 919                 }
 920
 921                 /* no term for current and previous event (likely) */
 922                 if ((overwrite < 0) && (set < 0))
 923                         continue;
 924
 925                 /* has term for both current and previous event, compare */
 926                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 927                         return -1;
 928
 929                 /* no term for current event but has term for previous one */
 930                 if ((overwrite >= 0) && (set < 0))
 931                         return -1;
 932
 933                 /* has term for current event */
 934                 if ((overwrite < 0) && (set >= 0)) {
 935                         /* if it's first event, set overwrite */
 936                         if (evsel == perf_evlist__first(evlist))
 937                                 overwrite = set;
 938                         else
 939                                 return -1;
 940                 }
 941         }
 942
 943         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 944                 opts->overwrite = overwrite;
 945
 946         return 0;
 947 }
 948
 949 static int perf_top_overwrite_fallback(struct perf_top *top,
 950                                        struct perf_evsel *evsel)
 951 {
 952         struct record_opts *opts = &top->record_opts;
 953         struct perf_evlist *evlist = top->evlist;
 954         struct perf_evsel *counter;
 955
 956         if (!opts->overwrite)
 957                 return 0;
 958
 959         /* only fall back when first event fails */
 960         if (evsel != perf_evlist__first(evlist))
 961                 return 0;
 962
 963         evlist__for_each_entry(evlist, counter)
 964                 counter->attr.write_backward = false;
 965         opts->overwrite = false;
 966         pr_debug2("fall back to non-overwrite mode\n");
 967         return 1;
 968 }
 969
 970 static int perf_top__start_counters(struct perf_top *top)
 971 {
 972         char msg[BUFSIZ];
 973         struct perf_evsel *counter;
 974         struct perf_evlist *evlist = top->evlist;
 975         struct record_opts *opts = &top->record_opts;
 976
 977         if (perf_top__overwrite_check(top)) {
 978                 ui__error("perf top only support consistent per-event "
 979                           "overwrite setting for all events\n");
 980                 goto out_err;
 981         }
 982
 983         perf_evlist__config(evlist, opts, &callchain_param);
 984
 985         evlist__for_each_entry(evlist, counter) {
 986 try_again:
 987                 if (perf_evsel__open(counter, top->evlist->cpus,
 988                                      top->evlist->threads) < 0) {
 989
 990                         /*
 991                          * Specially handle overwrite fall back.
 992                          * Because perf top is the only tool which has
 993                          * overwrite mode by default, support
 994                          * both overwrite and non-overwrite mode, and
 995                          * require consistent mode for all events.
 996                          *
 997                          * May move it to generic code with more tools
 998                          * have similar attribute.
 999                          */
1000                         if (perf_missing_features.write_backward &&
1001                             perf_top_overwrite_fallback(top, counter))
1002                                 goto try_again;
1003
1004                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1005                                 if (verbose > 0)
1006                                         ui__warning("%s\n", msg);
1007                                 goto try_again;
1008                         }
1009
1010                         perf_evsel__open_strerror(counter, &opts->target,
1011                                                   errno, msg, sizeof(msg));
1012                         ui__error("%s\n", msg);
1013                         goto out_err;
1014                 }
1015         }
1016
1017         if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1018                 ui__error("Failed to mmap with %d (%s)\n",
1019                             errno, str_error_r(errno, msg, sizeof(msg)));
1020                 goto out_err;
1021         }
1022
1023         return 0;
1024
1025 out_err:
1026         return -1;
1027 }
1028
1029 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1030 {
1031         if (callchain->mode != CHAIN_NONE) {
1032                 if (callchain_register_param(callchain) < 0) {
1033                         ui__error("Can't register callchain params.\n");
1034                         return -EINVAL;
1035                 }
1036         }
1037
1038         return 0;
1039 }
1040
1041 static struct ordered_events *rotate_queues(struct perf_top *top)
1042 {
1043         struct ordered_events *in = top->qe.in;
1044
1045         if (top->qe.in == &top->qe.data[1])
1046                 top->qe.in = &top->qe.data[0];
1047         else
1048                 top->qe.in = &top->qe.data[1];
1049
1050         return in;
1051 }
1052
1053 static void *process_thread(void *arg)
1054 {
1055         struct perf_top *top = arg;
1056
1057         while (!done) {
1058                 struct ordered_events *out, *in = top->qe.in;
1059
1060                 if (!in->nr_events) {
1061                         usleep(100);
1062                         continue;
1063                 }
1064
1065                 out = rotate_queues(top);
1066
1067                 pthread_mutex_lock(&top->qe.mutex);
1068                 top->qe.rotate = true;
1069                 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1070                 pthread_mutex_unlock(&top->qe.mutex);
1071
1072                 if (ordered_events__flush(out, OE_FLUSH__TOP))
1073                         pr_err("failed to process events\n");
1074         }
1075
1076         return NULL;
1077 }
1078
1079 /*
1080  * Allow only 'top->delay_secs' seconds behind samples.
1081  */
1082 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1083 {
1084         union perf_event *event = qevent->event;
1085         u64 delay_timestamp;
1086
1087         if (event->header.type != PERF_RECORD_SAMPLE)
1088                 return false;
1089
1090         delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1091         return delay_timestamp < last_timestamp;
1092 }
1093
1094 static int deliver_event(struct ordered_events *qe,
1095                          struct ordered_event *qevent)
1096 {
1097         struct perf_top *top = qe->data;
1098         struct perf_evlist *evlist = top->evlist;
1099         struct perf_session *session = top->session;
1100         union perf_event *event = qevent->event;
1101         struct perf_sample sample;
1102         struct perf_evsel *evsel;
1103         struct machine *machine;
1104         int ret = -1;
1105
1106         if (should_drop(qevent, top)) {
1107                 top->drop++;
1108                 top->drop_total++;
1109                 return 0;
1110         }
1111
1112         ret = perf_evlist__parse_sample(evlist, event, &sample);
1113         if (ret) {
1114                 pr_err("Can't parse sample, err = %d\n", ret);
1115                 goto next_event;
1116         }
1117
1118         evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1119         assert(evsel != NULL);
1120
1121         if (event->header.type == PERF_RECORD_SAMPLE)
1122                 ++top->samples;
1123
1124         switch (sample.cpumode) {
1125         case PERF_RECORD_MISC_USER:
1126                 ++top->us_samples;
1127                 if (top->hide_user_symbols)
1128                         goto next_event;
1129                 machine = &session->machines.host;
1130                 break;
1131         case PERF_RECORD_MISC_KERNEL:
1132                 ++top->kernel_samples;
1133                 if (top->hide_kernel_symbols)
1134                         goto next_event;
1135                 machine = &session->machines.host;
1136                 break;
1137         case PERF_RECORD_MISC_GUEST_KERNEL:
1138                 ++top->guest_kernel_samples;
1139                 machine = perf_session__find_machine(session,
1140                                                      sample.pid);
1141                 break;
1142         case PERF_RECORD_MISC_GUEST_USER:
1143                 ++top->guest_us_samples;
1144                 /*
1145                  * TODO: we don't process guest user from host side
1146                  * except simple counting.
1147                  */
1148                 goto next_event;
1149         default:
1150                 if (event->header.type == PERF_RECORD_SAMPLE)
1151                         goto next_event;
1152                 machine = &session->machines.host;
1153                 break;
1154         }
1155
1156         if (event->header.type == PERF_RECORD_SAMPLE) {
1157                 perf_event__process_sample(&top->tool, event, evsel,
1158                                            &sample, machine);
1159         } else if (event->header.type == PERF_RECORD_LOST) {
1160                 perf_top__process_lost(top, event, evsel);
1161         } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1162                 perf_top__process_lost_samples(top, event, evsel);
1163         } else if (event->header.type < PERF_RECORD_MAX) {
1164                 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1165                 machine__process_event(machine, event, &sample);
1166         } else
1167                 ++session->evlist->stats.nr_unknown_events;
1168
1169         ret = 0;
1170 next_event:
1171         return ret;
1172 }
1173
1174 static void init_process_thread(struct perf_top *top)
1175 {
1176         ordered_events__init(&top->qe.data[0], deliver_event, top);
1177         ordered_events__init(&top->qe.data[1], deliver_event, top);
1178         ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1179         ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1180         top->qe.in = &top->qe.data[0];
1181         pthread_mutex_init(&top->qe.mutex, NULL);
1182         pthread_cond_init(&top->qe.cond, NULL);
1183 }
1184
1185 static int __cmd_top(struct perf_top *top)
1186 {
1187         char msg[512];
1188         struct perf_evsel *pos;
1189         struct perf_evsel_config_term *err_term;
1190         struct perf_evlist *evlist = top->evlist;
1191         struct record_opts *opts = &top->record_opts;
1192         pthread_t thread, thread_process;
1193         int ret;
1194
1195         top->session = perf_session__new(NULL, false, NULL);
1196         if (top->session == NULL)
1197                 return -1;
1198
1199         if (!top->annotation_opts.objdump_path) {
1200                 ret = perf_env__lookup_objdump(&top->session->header.env,
1201                                                &top->annotation_opts.objdump_path);
1202                 if (ret)
1203                         goto out_delete;
1204         }
1205
1206         ret = callchain_param__setup_sample_type(&callchain_param);
1207         if (ret)
1208                 goto out_delete;
1209
1210         if (perf_session__register_idle_thread(top->session) < 0)
1211                 goto out_delete;
1212
1213         if (top->nr_threads_synthesize > 1)
1214                 perf_set_multithreaded();
1215
1216         init_process_thread(top);
1217
1218         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1219                                     top->evlist->threads, false,
1220                                     top->nr_threads_synthesize);
1221
1222         if (top->nr_threads_synthesize > 1)
1223                 perf_set_singlethreaded();
1224
1225         if (perf_hpp_list.socket) {
1226                 ret = perf_env__read_cpu_topology_map(&perf_env);
1227                 if (ret < 0)
1228                         goto out_err_cpu_topo;
1229         }
1230
1231         ret = perf_top__start_counters(top);
1232         if (ret)
1233                 goto out_delete;
1234
1235         ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1236         if (ret) {
1237                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1238                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1239                         str_error_r(errno, msg, sizeof(msg)));
1240                 goto out_delete;
1241         }
1242
1243         top->session->evlist = top->evlist;
1244         perf_session__set_id_hdr_size(top->session);
1245
1246         /*
1247          * When perf is starting the traced process, all the events (apart from
1248          * group members) have enable_on_exec=1 set, so don't spoil it by
1249          * prematurely enabling them.
1250          *
1251          * XXX 'top' still doesn't start workloads like record, trace, but should,
1252          * so leave the check here.
1253          */
1254         if (!target__none(&opts->target))
1255                 perf_evlist__enable(top->evlist);
1256
1257         ret = -1;
1258         if (pthread_create(&thread_process, NULL, process_thread, top)) {
1259                 ui__error("Could not create process thread.\n");
1260                 goto out_delete;
1261         }
1262
1263         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1264                                                             display_thread), top)) {
1265                 ui__error("Could not create display thread.\n");
1266                 goto out_join_thread;
1267         }
1268
1269         if (top->realtime_prio) {
1270                 struct sched_param param;
1271
1272                 param.sched_priority = top->realtime_prio;
1273                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1274                         ui__error("Could not set realtime priority.\n");
1275                         goto out_join;
1276                 }
1277         }
1278
1279         /* Wait for a minimal set of events before starting the snapshot */
1280         perf_evlist__poll(top->evlist, 100);
1281
1282         perf_top__mmap_read(top);
1283
1284         while (!done) {
1285                 u64 hits = top->samples;
1286
1287                 perf_top__mmap_read(top);
1288
1289                 if (opts->overwrite || (hits == top->samples))
1290                         ret = perf_evlist__poll(top->evlist, 100);
1291
1292                 if (resize) {
1293                         perf_top__resize(top);
1294                         resize = 0;
1295                 }
1296         }
1297
1298         ret = 0;
1299 out_join:
1300         pthread_join(thread, NULL);
1301 out_join_thread:
1302         pthread_cond_signal(&top->qe.cond);
1303         pthread_join(thread_process, NULL);
1304 out_delete:
1305         perf_session__delete(top->session);
1306         top->session = NULL;
1307
1308         return ret;
1309
1310 out_err_cpu_topo: {
1311         char errbuf[BUFSIZ];
1312         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1313
1314         ui__error("Could not read the CPU topology map: %s\n", err);
1315         goto out_delete;
1316 }
1317 }
1318
1319 static int
1320 callchain_opt(const struct option *opt, const char *arg, int unset)
1321 {
1322         symbol_conf.use_callchain = true;
1323         return record_callchain_opt(opt, arg, unset);
1324 }
1325
1326 static int
1327 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1328 {
1329         struct callchain_param *callchain = opt->value;
1330
1331         callchain->enabled = !unset;
1332         callchain->record_mode = CALLCHAIN_FP;
1333
1334         /*
1335          * --no-call-graph
1336          */
1337         if (unset) {
1338                 symbol_conf.use_callchain = false;
1339                 callchain->record_mode = CALLCHAIN_NONE;
1340                 return 0;
1341         }
1342
1343         return parse_callchain_top_opt(arg);
1344 }
1345
1346 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1347 {
1348         if (!strcmp(var, "top.call-graph")) {
1349                 var = "call-graph.record-mode";
1350                 return perf_default_config(var, value, cb);
1351         }
1352         if (!strcmp(var, "top.children")) {
1353                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1354                 return 0;
1355         }
1356
1357         return 0;
1358 }
1359
1360 static int
1361 parse_percent_limit(const struct option *opt, const char *arg,
1362                     int unset __maybe_unused)
1363 {
1364         struct perf_top *top = opt->value;
1365
1366         top->min_percent = strtof(arg, NULL);
1367         return 0;
1368 }
1369
1370 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1371         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1372
1373 int cmd_top(int argc, const char **argv)
1374 {
1375         char errbuf[BUFSIZ];
1376         struct perf_top top = {
1377                 .count_filter        = 5,
1378                 .delay_secs          = 2,
1379                 .record_opts = {
1380                         .mmap_pages     = UINT_MAX,
1381                         .user_freq      = UINT_MAX,
1382                         .user_interval  = ULLONG_MAX,
1383                         .freq           = 4000, /* 4 KHz */
1384                         .target         = {
1385                                 .uses_mmap   = true,
1386                         },
1387                         /*
1388                          * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1389                          * when we pause, fix that and reenable. Probably using a
1390                          * separate evlist with a dummy event, i.e. a non-overwrite
1391                          * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1392                          * stays in overwrite mode. -acme
1393                          * */
1394                         .overwrite      = 0,
1395                         .sample_time    = true,
1396                 },
1397                 .max_stack           = sysctl__max_stack(),
1398                 .annotation_opts     = annotation__default_options,
1399                 .nr_threads_synthesize = UINT_MAX,
1400         };
1401         struct record_opts *opts = &top.record_opts;
1402         struct target *target = &opts->target;
1403         const struct option options[] = {
1404         OPT_CALLBACK('e', "event", &top.evlist, "event",
1405                      "event selector. use 'perf list' to list available events",
1406                      parse_events_option),
1407         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1408         OPT_STRING('p', "pid", &target->pid, "pid",
1409                     "profile events on existing process id"),
1410         OPT_STRING('t', "tid", &target->tid, "tid",
1411                     "profile events on existing thread id"),
1412         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1413                             "system-wide collection from all CPUs"),
1414         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1415                     "list of cpus to monitor"),
1416         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1417                    "file", "vmlinux pathname"),
1418         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1419                     "don't load vmlinux even if found"),
1420         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1421                    "file", "kallsyms pathname"),
1422         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1423                     "hide kernel symbols"),
1424         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1425                      "number of mmap data pages",
1426                      perf_evlist__parse_mmap_pages),
1427         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1428                     "collect data with this RT SCHED_FIFO priority"),
1429         OPT_INTEGER('d', "delay", &top.delay_secs,
1430                     "number of seconds to delay between refreshes"),
1431         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1432                             "dump the symbol table used for profiling"),
1433         OPT_INTEGER('f', "count-filter", &top.count_filter,
1434                     "only display functions with more events than this"),
1435         OPT_BOOLEAN(0, "group", &opts->group,
1436                             "put the counters into a counter group"),
1437         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1438                     "child tasks do not inherit counters"),
1439         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1440                     "symbol to annotate"),
1441         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1442         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1443                      "profile at this frequency",
1444                       record__parse_freq),
1445         OPT_INTEGER('E', "entries", &top.print_entries,
1446                     "display this many functions"),
1447         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1448                     "hide user symbols"),
1449         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1450         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1451         OPT_INCR('v', "verbose", &verbose,
1452                     "be more verbose (show counter open errors, etc)"),
1453         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1454                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1455                    " Please refer the man page for the complete list."),
1456         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1457                    "output field(s): overhead, period, sample plus all of sort keys"),
1458         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1459                     "Show a column with the number of samples"),
1460         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1461                            NULL, "enables call-graph recording and display",
1462                            &callchain_opt),
1463         OPT_CALLBACK(0, "call-graph", &callchain_param,
1464                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1465                      top_callchain_help, &parse_callchain_opt),
1466         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1467                     "Accumulate callchains of children and show total overhead as well"),
1468         OPT_INTEGER(0, "max-stack", &top.max_stack,
1469                     "Set the maximum stack depth when parsing the callchain. "
1470                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1471         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1472                    "ignore callees of these functions in call graphs",
1473                    report_parse_ignore_callees_opt),
1474         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1475                     "Show a column with the sum of periods"),
1476         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1477                    "only consider symbols in these dsos"),
1478         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1479                    "only consider symbols in these comms"),
1480         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1481                    "only consider these symbols"),
1482         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1483                     "Interleave source code with assembly code (default)"),
1484         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1485                     "Display raw encoding of assembly instructions (default)"),
1486         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1487                     "Enable kernel symbol demangling"),
1488         OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1489                     "objdump binary to use for disassembly and annotations"),
1490         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1491                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1492         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1493         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1494                      "Don't show entries under that percent", parse_percent_limit),
1495         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1496                      "How to display percentage of filtered entries", parse_filter_percentage),
1497         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1498                    "width[,width...]",
1499                    "don't try to adjust column width, use these fixed values"),
1500         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1501                         "per thread proc mmap processing timeout in ms"),
1502         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1503                      "branch any", "sample any taken branches",
1504                      parse_branch_stack),
1505         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1506                      "branch filter mask", "branch stack filter modes",
1507                      parse_branch_stack),
1508         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1509                     "Show raw trace event output (do not use print fmt or plugins)"),
1510         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1511                     "Show entries in a hierarchy"),
1512         OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1513                     "Use a backward ring buffer, default: no"),
1514         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1515         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1516                         "number of thread to run event synthesize"),
1517         OPT_END()
1518         };
1519         const char * const top_usage[] = {
1520                 "perf top [<options>]",
1521                 NULL
1522         };
1523         int status = hists__init();
1524
1525         if (status < 0)
1526                 return status;
1527
1528         top.annotation_opts.min_pcnt = 5;
1529         top.annotation_opts.context  = 4;
1530
1531         top.evlist = perf_evlist__new();
1532         if (top.evlist == NULL)
1533                 return -ENOMEM;
1534
1535         status = perf_config(perf_top_config, &top);
1536         if (status)
1537                 return status;
1538
1539         argc = parse_options(argc, argv, options, top_usage, 0);
1540         if (argc)
1541                 usage_with_options(top_usage, options);
1542
1543         if (!top.evlist->nr_entries &&
1544             perf_evlist__add_default(top.evlist) < 0) {
1545                 pr_err("Not enough memory for event selector list\n");
1546                 goto out_delete_evlist;
1547         }
1548
1549         if (symbol_conf.report_hierarchy) {
1550                 /* disable incompatible options */
1551                 symbol_conf.event_group = false;
1552                 symbol_conf.cumulate_callchain = false;
1553
1554                 if (field_order) {
1555                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1556                         parse_options_usage(top_usage, options, "fields", 0);
1557                         parse_options_usage(NULL, options, "hierarchy", 0);
1558                         goto out_delete_evlist;
1559                 }
1560         }
1561
1562         if (opts->branch_stack && callchain_param.enabled)
1563                 symbol_conf.show_branchflag_count = true;
1564
1565         sort__mode = SORT_MODE__TOP;
1566         /* display thread wants entries to be collapsed in a different tree */
1567         perf_hpp_list.need_collapse = 1;
1568
1569         if (top.use_stdio)
1570                 use_browser = 0;
1571         else if (top.use_tui)
1572                 use_browser = 1;
1573
1574         setup_browser(false);
1575
1576         if (setup_sorting(top.evlist) < 0) {
1577                 if (sort_order)
1578                         parse_options_usage(top_usage, options, "s", 1);
1579                 if (field_order)
1580                         parse_options_usage(sort_order ? NULL : top_usage,
1581                                             options, "fields", 0);
1582                 goto out_delete_evlist;
1583         }
1584
1585         status = target__validate(target);
1586         if (status) {
1587                 target__strerror(target, status, errbuf, BUFSIZ);
1588                 ui__warning("%s\n", errbuf);
1589         }
1590
1591         status = target__parse_uid(target);
1592         if (status) {
1593                 int saved_errno = errno;
1594
1595                 target__strerror(target, status, errbuf, BUFSIZ);
1596                 ui__error("%s\n", errbuf);
1597
1598                 status = -saved_errno;
1599                 goto out_delete_evlist;
1600         }
1601
1602         if (target__none(target))
1603                 target->system_wide = true;
1604
1605         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1606                 ui__error("Couldn't create thread/CPU maps: %s\n",
1607                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1608                 goto out_delete_evlist;
1609         }
1610
1611         if (top.delay_secs < 1)
1612                 top.delay_secs = 1;
1613
1614         if (record_opts__config(opts)) {
1615                 status = -EINVAL;
1616                 goto out_delete_evlist;
1617         }
1618
1619         top.sym_evsel = perf_evlist__first(top.evlist);
1620
1621         if (!callchain_param.enabled) {
1622                 symbol_conf.cumulate_callchain = false;
1623                 perf_hpp__cancel_cumulate();
1624         }
1625
1626         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1627                 callchain_param.order = ORDER_CALLER;
1628
1629         status = symbol__annotation_init();
1630         if (status < 0)
1631                 goto out_delete_evlist;
1632
1633         annotation_config__init();
1634
1635         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1636         if (symbol__init(NULL) < 0)
1637                 return -1;
1638
1639         sort__setup_elide(stdout);
1640
1641         get_term_dimensions(&top.winsize);
1642         if (top.print_entries == 0) {
1643                 perf_top__update_print_entries(&top);
1644                 signal(SIGWINCH, winch_sig);
1645         }
1646
1647         status = __cmd_top(&top);
1648
1649 out_delete_evlist:
1650         perf_evlist__delete(top.evlist);
1651
1652         return status;
1653 }