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
  50 #include <assert.h>
  51 #include <elf.h>
  52 #include <fcntl.h>
  53
  54 #include <stdio.h>
  55 #include <termios.h>
  56 #include <unistd.h>
  57 #include <inttypes.h>
  58
  59 #include <errno.h>
  60 #include <time.h>
  61 #include <sched.h>
  62 #include <signal.h>
  63
  64 #include <sys/syscall.h>
  65 #include <sys/ioctl.h>
  66 #include <poll.h>
  67 #include <sys/prctl.h>
  68 #include <sys/wait.h>
  69 #include <sys/uio.h>
  70 #include <sys/utsname.h>
  71 #include <sys/mman.h>
  72
  73 #include <linux/stringify.h>
  74 #include <linux/time64.h>
  75 #include <linux/types.h>
  76
  77 #include "sane_ctype.h"
  78
  79 static volatile int done;
  80 static volatile int resize;
  81
  82 #define HEADER_LINE_NR  5
  83
  84 static void perf_top__update_print_entries(struct perf_top *top)
  85 {
  86         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  87 }
  88
  89 static void winch_sig(int sig __maybe_unused)
  90 {
  91         resize = 1;
  92 }
  93
  94 static void perf_top__resize(struct perf_top *top)
  95 {
  96         get_term_dimensions(&top->winsize);
  97         perf_top__update_print_entries(top);
  98 }
  99
 100 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 101 {
 102         struct perf_evsel *evsel = hists_to_evsel(he->hists);
 103         struct symbol *sym;
 104         struct annotation *notes;
 105         struct map *map;
 106         int err = -1;
 107
 108         if (!he || !he->ms.sym)
 109                 return -1;
 110
 111         sym = he->ms.sym;
 112         map = he->ms.map;
 113
 114         /*
 115          * We can't annotate with just /proc/kallsyms
 116          */
 117         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
 118             !dso__is_kcore(map->dso)) {
 119                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 120                        "path\n", sym->name);
 121                 sleep(1);
 122                 return -1;
 123         }
 124
 125         notes = symbol__annotation(sym);
 126         pthread_mutex_lock(&notes->lock);
 127
 128         if (!symbol__hists(sym, top->evlist->nr_entries)) {
 129                 pthread_mutex_unlock(&notes->lock);
 130                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 131                        sym->name);
 132                 sleep(1);
 133                 return err;
 134         }
 135
 136         err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
 137         if (err == 0) {
 138                 top->sym_filter_entry = he;
 139         } else {
 140                 char msg[BUFSIZ];
 141                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
 142                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 143         }
 144
 145         pthread_mutex_unlock(&notes->lock);
 146         return err;
 147 }
 148
 149 static void __zero_source_counters(struct hist_entry *he)
 150 {
 151         struct symbol *sym = he->ms.sym;
 152         symbol__annotate_zero_histograms(sym);
 153 }
 154
 155 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 156 {
 157         struct utsname uts;
 158         int err = uname(&uts);
 159
 160         ui__warning("Out of bounds address found:\n\n"
 161                     "Addr:   %" PRIx64 "\n"
 162                     "DSO:    %s %c\n"
 163                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 164                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 165                     "Arch:   %s\n"
 166                     "Kernel: %s\n"
 167                     "Tools:  %s\n\n"
 168                     "Not all samples will be on the annotation output.\n\n"
 169                     "Please report to linux-kernel@vger.kernel.org\n",
 170                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
 171                     map->start, map->end, sym->start, sym->end,
 172                     sym->binding == STB_GLOBAL ? 'g' :
 173                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 174                     err ? "[unknown]" : uts.machine,
 175                     err ? "[unknown]" : uts.release, perf_version_string);
 176         if (use_browser <= 0)
 177                 sleep(5);
 178
 179         map->erange_warned = true;
 180 }
 181
 182 static void perf_top__record_precise_ip(struct perf_top *top,
 183                                         struct hist_entry *he,
 184                                         struct perf_sample *sample,
 185                                         struct perf_evsel *evsel, u64 ip)
 186 {
 187         struct annotation *notes;
 188         struct symbol *sym = he->ms.sym;
 189         int err = 0;
 190
 191         if (sym == NULL || (use_browser == 0 &&
 192                             (top->sym_filter_entry == NULL ||
 193                              top->sym_filter_entry->ms.sym != sym)))
 194                 return;
 195
 196         notes = symbol__annotation(sym);
 197
 198         if (pthread_mutex_trylock(&notes->lock))
 199                 return;
 200
 201         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 202
 203         pthread_mutex_unlock(&notes->lock);
 204
 205         if (unlikely(err)) {
 206                 /*
 207                  * This function is now called with he->hists->lock held.
 208                  * Release it before going to sleep.
 209                  */
 210                 pthread_mutex_unlock(&he->hists->lock);
 211
 212                 if (err == -ERANGE && !he->ms.map->erange_warned)
 213                         ui__warn_map_erange(he->ms.map, sym, ip);
 214                 else if (err == -ENOMEM) {
 215                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 216                                sym->name);
 217                         sleep(1);
 218                 }
 219
 220                 pthread_mutex_lock(&he->hists->lock);
 221         }
 222 }
 223
 224 static void perf_top__show_details(struct perf_top *top)
 225 {
 226         struct hist_entry *he = top->sym_filter_entry;
 227         struct perf_evsel *evsel = hists_to_evsel(he->hists);
 228         struct annotation *notes;
 229         struct symbol *symbol;
 230         int more;
 231
 232         if (!he)
 233                 return;
 234
 235         symbol = he->ms.sym;
 236         notes = symbol__annotation(symbol);
 237
 238         pthread_mutex_lock(&notes->lock);
 239
 240         symbol__calc_percent(symbol, evsel);
 241
 242         if (notes->src == NULL)
 243                 goto out_unlock;
 244
 245         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
 246         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 247
 248         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
 249
 250         if (top->evlist->enabled) {
 251                 if (top->zero)
 252                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
 253                 else
 254                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
 255         }
 256         if (more != 0)
 257                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 258 out_unlock:
 259         pthread_mutex_unlock(&notes->lock);
 260 }
 261
 262 static void perf_top__print_sym_table(struct perf_top *top)
 263 {
 264         char bf[160];
 265         int printed = 0;
 266         const int win_width = top->winsize.ws_col - 1;
 267         struct perf_evsel *evsel = top->sym_evsel;
 268         struct hists *hists = evsel__hists(evsel);
 269
 270         puts(CONSOLE_CLEAR);
 271
 272         perf_top__header_snprintf(top, bf, sizeof(bf));
 273         printf("%s\n", bf);
 274
 275         perf_top__reset_sample_counters(top);
 276
 277         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 278
 279         if (!top->record_opts.overwrite &&
 280             (hists->stats.nr_lost_warned !=
 281             hists->stats.nr_events[PERF_RECORD_LOST])) {
 282                 hists->stats.nr_lost_warned =
 283                               hists->stats.nr_events[PERF_RECORD_LOST];
 284                 color_fprintf(stdout, PERF_COLOR_RED,
 285                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 286                               hists->stats.nr_lost_warned);
 287                 ++printed;
 288         }
 289
 290         if (top->sym_filter_entry) {
 291                 perf_top__show_details(top);
 292                 return;
 293         }
 294
 295         if (top->evlist->enabled) {
 296                 if (top->zero) {
 297                         hists__delete_entries(hists);
 298                 } else {
 299                         hists__decay_entries(hists, top->hide_user_symbols,
 300                                              top->hide_kernel_symbols);
 301                 }
 302         }
 303
 304         hists__collapse_resort(hists, NULL);
 305         perf_evsel__output_resort(evsel, NULL);
 306
 307         hists__output_recalc_col_len(hists, top->print_entries - printed);
 308         putchar('\n');
 309         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 310                        top->min_percent, stdout, symbol_conf.use_callchain);
 311 }
 312
 313 static void prompt_integer(int *target, const char *msg)
 314 {
 315         char *buf = malloc(0), *p;
 316         size_t dummy = 0;
 317         int tmp;
 318
 319         fprintf(stdout, "\n%s: ", msg);
 320         if (getline(&buf, &dummy, stdin) < 0)
 321                 return;
 322
 323         p = strchr(buf, '\n');
 324         if (p)
 325                 *p = 0;
 326
 327         p = buf;
 328         while(*p) {
 329                 if (!isdigit(*p))
 330                         goto out_free;
 331                 p++;
 332         }
 333         tmp = strtoul(buf, NULL, 10);
 334         *target = tmp;
 335 out_free:
 336         free(buf);
 337 }
 338
 339 static void prompt_percent(int *target, const char *msg)
 340 {
 341         int tmp = 0;
 342
 343         prompt_integer(&tmp, msg);
 344         if (tmp >= 0 && tmp <= 100)
 345                 *target = tmp;
 346 }
 347
 348 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 349 {
 350         char *buf = malloc(0), *p;
 351         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 352         struct hists *hists = evsel__hists(top->sym_evsel);
 353         struct rb_node *next;
 354         size_t dummy = 0;
 355
 356         /* zero counters of active symbol */
 357         if (syme) {
 358                 __zero_source_counters(syme);
 359                 top->sym_filter_entry = NULL;
 360         }
 361
 362         fprintf(stdout, "\n%s: ", msg);
 363         if (getline(&buf, &dummy, stdin) < 0)
 364                 goto out_free;
 365
 366         p = strchr(buf, '\n');
 367         if (p)
 368                 *p = 0;
 369
 370         next = rb_first(&hists->entries);
 371         while (next) {
 372                 n = rb_entry(next, struct hist_entry, rb_node);
 373                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 374                         found = n;
 375                         break;
 376                 }
 377                 next = rb_next(&n->rb_node);
 378         }
 379
 380         if (!found) {
 381                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 382                 sleep(1);
 383         } else
 384                 perf_top__parse_source(top, found);
 385
 386 out_free:
 387         free(buf);
 388 }
 389
 390 static void perf_top__print_mapped_keys(struct perf_top *top)
 391 {
 392         char *name = NULL;
 393
 394         if (top->sym_filter_entry) {
 395                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 396                 name = sym->name;
 397         }
 398
 399         fprintf(stdout, "\nMapped keys:\n");
 400         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 401         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 402
 403         if (top->evlist->nr_entries > 1)
 404                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
 405
 406         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 407
 408         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 409         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 410         fprintf(stdout, "\t[S]     stop annotation.\n");
 411
 412         fprintf(stdout,
 413                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 414                 top->hide_kernel_symbols ? "yes" : "no");
 415         fprintf(stdout,
 416                 "\t[U]     hide user symbols.               \t(%s)\n",
 417                 top->hide_user_symbols ? "yes" : "no");
 418         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 419         fprintf(stdout, "\t[qQ]    quit.\n");
 420 }
 421
 422 static int perf_top__key_mapped(struct perf_top *top, int c)
 423 {
 424         switch (c) {
 425                 case 'd':
 426                 case 'e':
 427                 case 'f':
 428                 case 'z':
 429                 case 'q':
 430                 case 'Q':
 431                 case 'K':
 432                 case 'U':
 433                 case 'F':
 434                 case 's':
 435                 case 'S':
 436                         return 1;
 437                 case 'E':
 438                         return top->evlist->nr_entries > 1 ? 1 : 0;
 439                 default:
 440                         break;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 447 {
 448         bool ret = true;
 449
 450         if (!perf_top__key_mapped(top, c)) {
 451                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 452                 struct termios save;
 453
 454                 perf_top__print_mapped_keys(top);
 455                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 456                 fflush(stdout);
 457
 458                 set_term_quiet_input(&save);
 459
 460                 poll(&stdin_poll, 1, -1);
 461                 c = getc(stdin);
 462
 463                 tcsetattr(0, TCSAFLUSH, &save);
 464                 if (!perf_top__key_mapped(top, c))
 465                         return ret;
 466         }
 467
 468         switch (c) {
 469                 case 'd':
 470                         prompt_integer(&top->delay_secs, "Enter display delay");
 471                         if (top->delay_secs < 1)
 472                                 top->delay_secs = 1;
 473                         break;
 474                 case 'e':
 475                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 476                         if (top->print_entries == 0) {
 477                                 perf_top__resize(top);
 478                                 signal(SIGWINCH, winch_sig);
 479                         } else {
 480                                 signal(SIGWINCH, SIG_DFL);
 481                         }
 482                         break;
 483                 case 'E':
 484                         if (top->evlist->nr_entries > 1) {
 485                                 /* Select 0 as the default event: */
 486                                 int counter = 0;
 487
 488                                 fprintf(stderr, "\nAvailable events:");
 489
 490                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 491                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
 492
 493                                 prompt_integer(&counter, "Enter details event counter");
 494
 495                                 if (counter >= top->evlist->nr_entries) {
 496                                         top->sym_evsel = perf_evlist__first(top->evlist);
 497                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
 498                                         sleep(1);
 499                                         break;
 500                                 }
 501                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 502                                         if (top->sym_evsel->idx == counter)
 503                                                 break;
 504                         } else
 505                                 top->sym_evsel = perf_evlist__first(top->evlist);
 506                         break;
 507                 case 'f':
 508                         prompt_integer(&top->count_filter, "Enter display event count filter");
 509                         break;
 510                 case 'F':
 511                         prompt_percent(&top->annotation_opts.min_pcnt,
 512                                        "Enter details display event filter (percent)");
 513                         break;
 514                 case 'K':
 515                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 516                         break;
 517                 case 'q':
 518                 case 'Q':
 519                         printf("exiting.\n");
 520                         if (top->dump_symtab)
 521                                 perf_session__fprintf_dsos(top->session, stderr);
 522                         ret = false;
 523                         break;
 524                 case 's':
 525                         perf_top__prompt_symbol(top, "Enter details symbol");
 526                         break;
 527                 case 'S':
 528                         if (!top->sym_filter_entry)
 529                                 break;
 530                         else {
 531                                 struct hist_entry *syme = top->sym_filter_entry;
 532
 533                                 top->sym_filter_entry = NULL;
 534                                 __zero_source_counters(syme);
 535                         }
 536                         break;
 537                 case 'U':
 538                         top->hide_user_symbols = !top->hide_user_symbols;
 539                         break;
 540                 case 'z':
 541                         top->zero = !top->zero;
 542                         break;
 543                 default:
 544                         break;
 545         }
 546
 547         return ret;
 548 }
 549
 550 static void perf_top__sort_new_samples(void *arg)
 551 {
 552         struct perf_top *t = arg;
 553         struct perf_evsel *evsel = t->sym_evsel;
 554         struct hists *hists;
 555
 556         perf_top__reset_sample_counters(t);
 557
 558         if (t->evlist->selected != NULL)
 559                 t->sym_evsel = t->evlist->selected;
 560
 561         hists = evsel__hists(evsel);
 562
 563         if (t->evlist->enabled) {
 564                 if (t->zero) {
 565                         hists__delete_entries(hists);
 566                 } else {
 567                         hists__decay_entries(hists, t->hide_user_symbols,
 568                                              t->hide_kernel_symbols);
 569                 }
 570         }
 571
 572         hists__collapse_resort(hists, NULL);
 573         perf_evsel__output_resort(evsel, NULL);
 574 }
 575
 576 static void *display_thread_tui(void *arg)
 577 {
 578         struct perf_evsel *pos;
 579         struct perf_top *top = arg;
 580         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 581         struct hist_browser_timer hbt = {
 582                 .timer          = perf_top__sort_new_samples,
 583                 .arg            = top,
 584                 .refresh        = top->delay_secs,
 585         };
 586
 587         /* In order to read symbols from other namespaces perf to  needs to call
 588          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 589          * unshare(2) the fs so that we may continue to setns into namespaces
 590          * that we're observing.
 591          */
 592         unshare(CLONE_FS);
 593
 594         perf_top__sort_new_samples(top);
 595
 596         /*
 597          * Initialize the uid_filter_str, in the future the TUI will allow
 598          * Zooming in/out UIDs. For now juse use whatever the user passed
 599          * via --uid.
 600          */
 601         evlist__for_each_entry(top->evlist, pos) {
 602                 struct hists *hists = evsel__hists(pos);
 603                 hists->uid_filter_str = top->record_opts.target.uid_str;
 604         }
 605
 606         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
 607                                       top->min_percent,
 608                                       &top->session->header.env,
 609                                       !top->record_opts.overwrite,
 610                                       &top->annotation_opts);
 611
 612         done = 1;
 613         return NULL;
 614 }
 615
 616 static void display_sig(int sig __maybe_unused)
 617 {
 618         done = 1;
 619 }
 620
 621 static void display_setup_sig(void)
 622 {
 623         signal(SIGSEGV, sighandler_dump_stack);
 624         signal(SIGFPE, sighandler_dump_stack);
 625         signal(SIGINT,  display_sig);
 626         signal(SIGQUIT, display_sig);
 627         signal(SIGTERM, display_sig);
 628 }
 629
 630 static void *display_thread(void *arg)
 631 {
 632         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 633         struct termios save;
 634         struct perf_top *top = arg;
 635         int delay_msecs, c;
 636
 637         /* In order to read symbols from other namespaces perf to  needs to call
 638          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 639          * unshare(2) the fs so that we may continue to setns into namespaces
 640          * that we're observing.
 641          */
 642         unshare(CLONE_FS);
 643
 644         display_setup_sig();
 645         pthread__unblock_sigwinch();
 646 repeat:
 647         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 648         set_term_quiet_input(&save);
 649         /* trash return*/
 650         getc(stdin);
 651
 652         while (!done) {
 653                 perf_top__print_sym_table(top);
 654                 /*
 655                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 656                  * refresh screen in both cases.
 657                  */
 658                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 659                 case 0:
 660                         continue;
 661                 case -1:
 662                         if (errno == EINTR)
 663                                 continue;
 664                         __fallthrough;
 665                 default:
 666                         c = getc(stdin);
 667                         tcsetattr(0, TCSAFLUSH, &save);
 668
 669                         if (perf_top__handle_keypress(top, c))
 670                                 goto repeat;
 671                         done = 1;
 672                 }
 673         }
 674
 675         tcsetattr(0, TCSAFLUSH, &save);
 676         return NULL;
 677 }
 678
 679 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 680                                    struct addr_location *al, bool single,
 681                                    void *arg)
 682 {
 683         struct perf_top *top = arg;
 684         struct hist_entry *he = iter->he;
 685         struct perf_evsel *evsel = iter->evsel;
 686
 687         if (perf_hpp_list.sym && single)
 688                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
 689
 690         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 691                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
 692         return 0;
 693 }
 694
 695 static void perf_event__process_sample(struct perf_tool *tool,
 696                                        const union perf_event *event,
 697                                        struct perf_evsel *evsel,
 698                                        struct perf_sample *sample,
 699                                        struct machine *machine)
 700 {
 701         struct perf_top *top = container_of(tool, struct perf_top, tool);
 702         struct addr_location al;
 703         int err;
 704
 705         if (!machine && perf_guest) {
 706                 static struct intlist *seen;
 707
 708                 if (!seen)
 709                         seen = intlist__new(NULL);
 710
 711                 if (!intlist__has_entry(seen, sample->pid)) {
 712                         pr_err("Can't find guest [%d]'s kernel information\n",
 713                                 sample->pid);
 714                         intlist__add(seen, sample->pid);
 715                 }
 716                 return;
 717         }
 718
 719         if (!machine) {
 720                 pr_err("%u unprocessable samples recorded.\r",
 721                        top->session->evlist->stats.nr_unprocessable_samples++);
 722                 return;
 723         }
 724
 725         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 726                 top->exact_samples++;
 727
 728         if (machine__resolve(machine, &al, sample) < 0)
 729                 return;
 730
 731         if (!machine->kptr_restrict_warned &&
 732             symbol_conf.kptr_restrict &&
 733             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 734                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
 735                         ui__warning(
 736 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 737 "Check /proc/sys/kernel/kptr_restrict.\n\n"
 738 "Kernel%s samples will not be resolved.\n",
 739                           al.map && map__has_symbols(al.map) ?
 740                           " modules" : "");
 741                         if (use_browser <= 0)
 742                                 sleep(5);
 743                 }
 744                 machine->kptr_restrict_warned = true;
 745         }
 746
 747         if (al.sym == NULL && al.map != NULL) {
 748                 const char *msg = "Kernel samples will not be resolved.\n";
 749                 /*
 750                  * As we do lazy loading of symtabs we only will know if the
 751                  * specified vmlinux file is invalid when we actually have a
 752                  * hit in kernel space and then try to load it. So if we get
 753                  * here and there are _no_ symbols in the DSO backing the
 754                  * kernel map, bail out.
 755                  *
 756                  * We may never get here, for instance, if we use -K/
 757                  * --hide-kernel-symbols, even if the user specifies an
 758                  * invalid --vmlinux ;-)
 759                  */
 760                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 761                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 762                         if (symbol_conf.vmlinux_name) {
 763                                 char serr[256];
 764                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
 765                                 ui__warning("The %s file can't be used: %s\n%s",
 766                                             symbol_conf.vmlinux_name, serr, msg);
 767                         } else {
 768                                 ui__warning("A vmlinux file was not found.\n%s",
 769                                             msg);
 770                         }
 771
 772                         if (use_browser <= 0)
 773                                 sleep(5);
 774                         top->vmlinux_warned = true;
 775                 }
 776         }
 777
 778         if (al.sym == NULL || !al.sym->idle) {
 779                 struct hists *hists = evsel__hists(evsel);
 780                 struct hist_entry_iter iter = {
 781                         .evsel          = evsel,
 782                         .sample         = sample,
 783                         .add_entry_cb   = hist_iter__top_callback,
 784                 };
 785
 786                 if (symbol_conf.cumulate_callchain)
 787                         iter.ops = &hist_iter_cumulative;
 788                 else
 789                         iter.ops = &hist_iter_normal;
 790
 791                 pthread_mutex_lock(&hists->lock);
 792
 793                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
 794                 if (err < 0)
 795                         pr_err("Problem incrementing symbol period, skipping event\n");
 796
 797                 pthread_mutex_unlock(&hists->lock);
 798         }
 799
 800         addr_location__put(&al);
 801 }
 802
 803 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 804 {
 805         struct record_opts *opts = &top->record_opts;
 806         struct perf_evlist *evlist = top->evlist;
 807         struct perf_sample sample;
 808         struct perf_evsel *evsel;
 809         struct perf_mmap *md;
 810         struct perf_session *session = top->session;
 811         union perf_event *event;
 812         struct machine *machine;
 813         int ret;
 814
 815         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 816         if (perf_mmap__read_init(md) < 0)
 817                 return;
 818
 819         while ((event = perf_mmap__read_event(md)) != NULL) {
 820                 ret = perf_evlist__parse_sample(evlist, event, &sample);
 821                 if (ret) {
 822                         pr_err("Can't parse sample, err = %d\n", ret);
 823                         goto next_event;
 824                 }
 825
 826                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
 827                 assert(evsel != NULL);
 828
 829                 if (event->header.type == PERF_RECORD_SAMPLE)
 830                         ++top->samples;
 831
 832                 switch (sample.cpumode) {
 833                 case PERF_RECORD_MISC_USER:
 834                         ++top->us_samples;
 835                         if (top->hide_user_symbols)
 836                                 goto next_event;
 837                         machine = &session->machines.host;
 838                         break;
 839                 case PERF_RECORD_MISC_KERNEL:
 840                         ++top->kernel_samples;
 841                         if (top->hide_kernel_symbols)
 842                                 goto next_event;
 843                         machine = &session->machines.host;
 844                         break;
 845                 case PERF_RECORD_MISC_GUEST_KERNEL:
 846                         ++top->guest_kernel_samples;
 847                         machine = perf_session__find_machine(session,
 848                                                              sample.pid);
 849                         break;
 850                 case PERF_RECORD_MISC_GUEST_USER:
 851                         ++top->guest_us_samples;
 852                         /*
 853                          * TODO: we don't process guest user from host side
 854                          * except simple counting.
 855                          */
 856                         goto next_event;
 857                 default:
 858                         if (event->header.type == PERF_RECORD_SAMPLE)
 859                                 goto next_event;
 860                         machine = &session->machines.host;
 861                         break;
 862                 }
 863
 864
 865                 if (event->header.type == PERF_RECORD_SAMPLE) {
 866                         perf_event__process_sample(&top->tool, event, evsel,
 867                                                    &sample, machine);
 868                 } else if (event->header.type < PERF_RECORD_MAX) {
 869                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
 870                         machine__process_event(machine, event, &sample);
 871                 } else
 872                         ++session->evlist->stats.nr_unknown_events;
 873 next_event:
 874                 perf_mmap__consume(md);
 875         }
 876
 877         perf_mmap__read_done(md);
 878 }
 879
 880 static void perf_top__mmap_read(struct perf_top *top)
 881 {
 882         bool overwrite = top->record_opts.overwrite;
 883         struct perf_evlist *evlist = top->evlist;
 884         unsigned long long start, end;
 885         int i;
 886
 887         start = rdclock();
 888         if (overwrite)
 889                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 890
 891         for (i = 0; i < top->evlist->nr_mmaps; i++)
 892                 perf_top__mmap_read_idx(top, i);
 893
 894         if (overwrite) {
 895                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 896                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 897         }
 898         end = rdclock();
 899
 900         if ((end - start) > (unsigned long long)top->delay_secs * NSEC_PER_SEC)
 901                 ui__warning("Too slow to read ring buffer.\n"
 902                             "Please try increasing the period (-c) or\n"
 903                             "decreasing the freq (-F) or\n"
 904                             "limiting the number of CPUs (-C)\n");
 905 }
 906
 907 /*
 908  * Check per-event overwrite term.
 909  * perf top should support consistent term for all events.
 910  * - All events don't have per-event term
 911  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 912  *   Nothing change, return 0.
 913  * - All events have same per-event term
 914  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 915  *   Using the per-event setting to replace the opts->overwrite if
 916  *   they are different, then return 0.
 917  * - Events have different per-event term
 918  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 919  *   Return -1
 920  * - Some of the event set per-event term, but some not.
 921  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 922  *   Return -1
 923  */
 924 static int perf_top__overwrite_check(struct perf_top *top)
 925 {
 926         struct record_opts *opts = &top->record_opts;
 927         struct perf_evlist *evlist = top->evlist;
 928         struct perf_evsel_config_term *term;
 929         struct list_head *config_terms;
 930         struct perf_evsel *evsel;
 931         int set, overwrite = -1;
 932
 933         evlist__for_each_entry(evlist, evsel) {
 934                 set = -1;
 935                 config_terms = &evsel->config_terms;
 936                 list_for_each_entry(term, config_terms, list) {
 937                         if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
 938                                 set = term->val.overwrite ? 1 : 0;
 939                 }
 940
 941                 /* no term for current and previous event (likely) */
 942                 if ((overwrite < 0) && (set < 0))
 943                         continue;
 944
 945                 /* has term for both current and previous event, compare */
 946                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 947                         return -1;
 948
 949                 /* no term for current event but has term for previous one */
 950                 if ((overwrite >= 0) && (set < 0))
 951                         return -1;
 952
 953                 /* has term for current event */
 954                 if ((overwrite < 0) && (set >= 0)) {
 955                         /* if it's first event, set overwrite */
 956                         if (evsel == perf_evlist__first(evlist))
 957                                 overwrite = set;
 958                         else
 959                                 return -1;
 960                 }
 961         }
 962
 963         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 964                 opts->overwrite = overwrite;
 965
 966         return 0;
 967 }
 968
 969 static int perf_top_overwrite_fallback(struct perf_top *top,
 970                                        struct perf_evsel *evsel)
 971 {
 972         struct record_opts *opts = &top->record_opts;
 973         struct perf_evlist *evlist = top->evlist;
 974         struct perf_evsel *counter;
 975
 976         if (!opts->overwrite)
 977                 return 0;
 978
 979         /* only fall back when first event fails */
 980         if (evsel != perf_evlist__first(evlist))
 981                 return 0;
 982
 983         evlist__for_each_entry(evlist, counter)
 984                 counter->attr.write_backward = false;
 985         opts->overwrite = false;
 986         pr_debug2("fall back to non-overwrite mode\n");
 987         return 1;
 988 }
 989
 990 static int perf_top__start_counters(struct perf_top *top)
 991 {
 992         char msg[BUFSIZ];
 993         struct perf_evsel *counter;
 994         struct perf_evlist *evlist = top->evlist;
 995         struct record_opts *opts = &top->record_opts;
 996
 997         if (perf_top__overwrite_check(top)) {
 998                 ui__error("perf top only support consistent per-event "
 999                           "overwrite setting for all events\n");
1000                 goto out_err;
1001         }
1002
1003         perf_evlist__config(evlist, opts, &callchain_param);
1004
1005         evlist__for_each_entry(evlist, counter) {
1006 try_again:
1007                 if (perf_evsel__open(counter, top->evlist->cpus,
1008                                      top->evlist->threads) < 0) {
1009
1010                         /*
1011                          * Specially handle overwrite fall back.
1012                          * Because perf top is the only tool which has
1013                          * overwrite mode by default, support
1014                          * both overwrite and non-overwrite mode, and
1015                          * require consistent mode for all events.
1016                          *
1017                          * May move it to generic code with more tools
1018                          * have similar attribute.
1019                          */
1020                         if (perf_missing_features.write_backward &&
1021                             perf_top_overwrite_fallback(top, counter))
1022                                 goto try_again;
1023
1024                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1025                                 if (verbose > 0)
1026                                         ui__warning("%s\n", msg);
1027                                 goto try_again;
1028                         }
1029
1030                         perf_evsel__open_strerror(counter, &opts->target,
1031                                                   errno, msg, sizeof(msg));
1032                         ui__error("%s\n", msg);
1033                         goto out_err;
1034                 }
1035         }
1036
1037         if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1038                 ui__error("Failed to mmap with %d (%s)\n",
1039                             errno, str_error_r(errno, msg, sizeof(msg)));
1040                 goto out_err;
1041         }
1042
1043         return 0;
1044
1045 out_err:
1046         return -1;
1047 }
1048
1049 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1050 {
1051         if (!perf_hpp_list.sym) {
1052                 if (callchain->enabled) {
1053                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1054                         return -EINVAL;
1055                 }
1056         } else if (callchain->mode != CHAIN_NONE) {
1057                 if (callchain_register_param(callchain) < 0) {
1058                         ui__error("Can't register callchain params.\n");
1059                         return -EINVAL;
1060                 }
1061         }
1062
1063         return 0;
1064 }
1065
1066 static int __cmd_top(struct perf_top *top)
1067 {
1068         char msg[512];
1069         struct perf_evsel *pos;
1070         struct perf_evsel_config_term *err_term;
1071         struct perf_evlist *evlist = top->evlist;
1072         struct record_opts *opts = &top->record_opts;
1073         pthread_t thread;
1074         int ret;
1075
1076         top->session = perf_session__new(NULL, false, NULL);
1077         if (top->session == NULL)
1078                 return -1;
1079
1080         if (!top->annotation_opts.objdump_path) {
1081                 ret = perf_env__lookup_objdump(&top->session->header.env,
1082                                                &top->annotation_opts.objdump_path);
1083                 if (ret)
1084                         goto out_delete;
1085         }
1086
1087         ret = callchain_param__setup_sample_type(&callchain_param);
1088         if (ret)
1089                 goto out_delete;
1090
1091         if (perf_session__register_idle_thread(top->session) < 0)
1092                 goto out_delete;
1093
1094         if (top->nr_threads_synthesize > 1)
1095                 perf_set_multithreaded();
1096
1097         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1098                                     top->evlist->threads, false,
1099                                     opts->proc_map_timeout,
1100                                     top->nr_threads_synthesize);
1101
1102         if (top->nr_threads_synthesize > 1)
1103                 perf_set_singlethreaded();
1104
1105         if (perf_hpp_list.socket) {
1106                 ret = perf_env__read_cpu_topology_map(&perf_env);
1107                 if (ret < 0)
1108                         goto out_err_cpu_topo;
1109         }
1110
1111         ret = perf_top__start_counters(top);
1112         if (ret)
1113                 goto out_delete;
1114
1115         ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1116         if (ret) {
1117                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1118                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1119                         str_error_r(errno, msg, sizeof(msg)));
1120                 goto out_delete;
1121         }
1122
1123         top->session->evlist = top->evlist;
1124         perf_session__set_id_hdr_size(top->session);
1125
1126         /*
1127          * When perf is starting the traced process, all the events (apart from
1128          * group members) have enable_on_exec=1 set, so don't spoil it by
1129          * prematurely enabling them.
1130          *
1131          * XXX 'top' still doesn't start workloads like record, trace, but should,
1132          * so leave the check here.
1133          */
1134         if (!target__none(&opts->target))
1135                 perf_evlist__enable(top->evlist);
1136
1137         /* Wait for a minimal set of events before starting the snapshot */
1138         perf_evlist__poll(top->evlist, 100);
1139
1140         perf_top__mmap_read(top);
1141
1142         ret = -1;
1143         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1144                                                             display_thread), top)) {
1145                 ui__error("Could not create display thread.\n");
1146                 goto out_delete;
1147         }
1148
1149         if (top->realtime_prio) {
1150                 struct sched_param param;
1151
1152                 param.sched_priority = top->realtime_prio;
1153                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1154                         ui__error("Could not set realtime priority.\n");
1155                         goto out_join;
1156                 }
1157         }
1158
1159         while (!done) {
1160                 u64 hits = top->samples;
1161
1162                 perf_top__mmap_read(top);
1163
1164                 if (opts->overwrite || (hits == top->samples))
1165                         ret = perf_evlist__poll(top->evlist, 100);
1166
1167                 if (resize) {
1168                         perf_top__resize(top);
1169                         resize = 0;
1170                 }
1171         }
1172
1173         ret = 0;
1174 out_join:
1175         pthread_join(thread, NULL);
1176 out_delete:
1177         perf_session__delete(top->session);
1178         top->session = NULL;
1179
1180         return ret;
1181
1182 out_err_cpu_topo: {
1183         char errbuf[BUFSIZ];
1184         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1185
1186         ui__error("Could not read the CPU topology map: %s\n", err);
1187         goto out_delete;
1188 }
1189 }
1190
1191 static int
1192 callchain_opt(const struct option *opt, const char *arg, int unset)
1193 {
1194         symbol_conf.use_callchain = true;
1195         return record_callchain_opt(opt, arg, unset);
1196 }
1197
1198 static int
1199 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1200 {
1201         struct callchain_param *callchain = opt->value;
1202
1203         callchain->enabled = !unset;
1204         callchain->record_mode = CALLCHAIN_FP;
1205
1206         /*
1207          * --no-call-graph
1208          */
1209         if (unset) {
1210                 symbol_conf.use_callchain = false;
1211                 callchain->record_mode = CALLCHAIN_NONE;
1212                 return 0;
1213         }
1214
1215         return parse_callchain_top_opt(arg);
1216 }
1217
1218 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1219 {
1220         if (!strcmp(var, "top.call-graph")) {
1221                 var = "call-graph.record-mode";
1222                 return perf_default_config(var, value, cb);
1223         }
1224         if (!strcmp(var, "top.children")) {
1225                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1226                 return 0;
1227         }
1228
1229         return 0;
1230 }
1231
1232 static int
1233 parse_percent_limit(const struct option *opt, const char *arg,
1234                     int unset __maybe_unused)
1235 {
1236         struct perf_top *top = opt->value;
1237
1238         top->min_percent = strtof(arg, NULL);
1239         return 0;
1240 }
1241
1242 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1243         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1244
1245 int cmd_top(int argc, const char **argv)
1246 {
1247         char errbuf[BUFSIZ];
1248         struct perf_top top = {
1249                 .count_filter        = 5,
1250                 .delay_secs          = 2,
1251                 .record_opts = {
1252                         .mmap_pages     = UINT_MAX,
1253                         .user_freq      = UINT_MAX,
1254                         .user_interval  = ULLONG_MAX,
1255                         .freq           = 4000, /* 4 KHz */
1256                         .target         = {
1257                                 .uses_mmap   = true,
1258                         },
1259                         .proc_map_timeout    = 500,
1260                         .overwrite      = 1,
1261                 },
1262                 .max_stack           = sysctl__max_stack(),
1263                 .annotation_opts     = annotation__default_options,
1264                 .nr_threads_synthesize = UINT_MAX,
1265         };
1266         struct record_opts *opts = &top.record_opts;
1267         struct target *target = &opts->target;
1268         const struct option options[] = {
1269         OPT_CALLBACK('e', "event", &top.evlist, "event",
1270                      "event selector. use 'perf list' to list available events",
1271                      parse_events_option),
1272         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1273         OPT_STRING('p', "pid", &target->pid, "pid",
1274                     "profile events on existing process id"),
1275         OPT_STRING('t', "tid", &target->tid, "tid",
1276                     "profile events on existing thread id"),
1277         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1278                             "system-wide collection from all CPUs"),
1279         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1280                     "list of cpus to monitor"),
1281         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1282                    "file", "vmlinux pathname"),
1283         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1284                     "don't load vmlinux even if found"),
1285         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1286                     "hide kernel symbols"),
1287         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1288                      "number of mmap data pages",
1289                      perf_evlist__parse_mmap_pages),
1290         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1291                     "collect data with this RT SCHED_FIFO priority"),
1292         OPT_INTEGER('d', "delay", &top.delay_secs,
1293                     "number of seconds to delay between refreshes"),
1294         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1295                             "dump the symbol table used for profiling"),
1296         OPT_INTEGER('f', "count-filter", &top.count_filter,
1297                     "only display functions with more events than this"),
1298         OPT_BOOLEAN(0, "group", &opts->group,
1299                             "put the counters into a counter group"),
1300         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1301                     "child tasks do not inherit counters"),
1302         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1303                     "symbol to annotate"),
1304         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1305         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1306                      "profile at this frequency",
1307                       record__parse_freq),
1308         OPT_INTEGER('E', "entries", &top.print_entries,
1309                     "display this many functions"),
1310         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1311                     "hide user symbols"),
1312         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1313         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1314         OPT_INCR('v', "verbose", &verbose,
1315                     "be more verbose (show counter open errors, etc)"),
1316         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1317                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1318                    " Please refer the man page for the complete list."),
1319         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1320                    "output field(s): overhead, period, sample plus all of sort keys"),
1321         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1322                     "Show a column with the number of samples"),
1323         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1324                            NULL, "enables call-graph recording and display",
1325                            &callchain_opt),
1326         OPT_CALLBACK(0, "call-graph", &callchain_param,
1327                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1328                      top_callchain_help, &parse_callchain_opt),
1329         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1330                     "Accumulate callchains of children and show total overhead as well"),
1331         OPT_INTEGER(0, "max-stack", &top.max_stack,
1332                     "Set the maximum stack depth when parsing the callchain. "
1333                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1334         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1335                    "ignore callees of these functions in call graphs",
1336                    report_parse_ignore_callees_opt),
1337         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1338                     "Show a column with the sum of periods"),
1339         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1340                    "only consider symbols in these dsos"),
1341         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1342                    "only consider symbols in these comms"),
1343         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1344                    "only consider these symbols"),
1345         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1346                     "Interleave source code with assembly code (default)"),
1347         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1348                     "Display raw encoding of assembly instructions (default)"),
1349         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1350                     "Enable kernel symbol demangling"),
1351         OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1352                     "objdump binary to use for disassembly and annotations"),
1353         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1354                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1355         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1356         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1357                      "Don't show entries under that percent", parse_percent_limit),
1358         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1359                      "How to display percentage of filtered entries", parse_filter_percentage),
1360         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1361                    "width[,width...]",
1362                    "don't try to adjust column width, use these fixed values"),
1363         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1364                         "per thread proc mmap processing timeout in ms"),
1365         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1366                      "branch any", "sample any taken branches",
1367                      parse_branch_stack),
1368         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1369                      "branch filter mask", "branch stack filter modes",
1370                      parse_branch_stack),
1371         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1372                     "Show raw trace event output (do not use print fmt or plugins)"),
1373         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1374                     "Show entries in a hierarchy"),
1375         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1376         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1377                         "number of thread to run event synthesize"),
1378         OPT_END()
1379         };
1380         const char * const top_usage[] = {
1381                 "perf top [<options>]",
1382                 NULL
1383         };
1384         int status = hists__init();
1385
1386         if (status < 0)
1387                 return status;
1388
1389         top.annotation_opts.min_pcnt = 5;
1390         top.annotation_opts.context  = 4;
1391
1392         top.evlist = perf_evlist__new();
1393         if (top.evlist == NULL)
1394                 return -ENOMEM;
1395
1396         status = perf_config(perf_top_config, &top);
1397         if (status)
1398                 return status;
1399
1400         argc = parse_options(argc, argv, options, top_usage, 0);
1401         if (argc)
1402                 usage_with_options(top_usage, options);
1403
1404         if (!top.evlist->nr_entries &&
1405             perf_evlist__add_default(top.evlist) < 0) {
1406                 pr_err("Not enough memory for event selector list\n");
1407                 goto out_delete_evlist;
1408         }
1409
1410         if (symbol_conf.report_hierarchy) {
1411                 /* disable incompatible options */
1412                 symbol_conf.event_group = false;
1413                 symbol_conf.cumulate_callchain = false;
1414
1415                 if (field_order) {
1416                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1417                         parse_options_usage(top_usage, options, "fields", 0);
1418                         parse_options_usage(NULL, options, "hierarchy", 0);
1419                         goto out_delete_evlist;
1420                 }
1421         }
1422
1423         sort__mode = SORT_MODE__TOP;
1424         /* display thread wants entries to be collapsed in a different tree */
1425         perf_hpp_list.need_collapse = 1;
1426
1427         if (top.use_stdio)
1428                 use_browser = 0;
1429         else if (top.use_tui)
1430                 use_browser = 1;
1431
1432         setup_browser(false);
1433
1434         if (setup_sorting(top.evlist) < 0) {
1435                 if (sort_order)
1436                         parse_options_usage(top_usage, options, "s", 1);
1437                 if (field_order)
1438                         parse_options_usage(sort_order ? NULL : top_usage,
1439                                             options, "fields", 0);
1440                 goto out_delete_evlist;
1441         }
1442
1443         status = target__validate(target);
1444         if (status) {
1445                 target__strerror(target, status, errbuf, BUFSIZ);
1446                 ui__warning("%s\n", errbuf);
1447         }
1448
1449         status = target__parse_uid(target);
1450         if (status) {
1451                 int saved_errno = errno;
1452
1453                 target__strerror(target, status, errbuf, BUFSIZ);
1454                 ui__error("%s\n", errbuf);
1455
1456                 status = -saved_errno;
1457                 goto out_delete_evlist;
1458         }
1459
1460         if (target__none(target))
1461                 target->system_wide = true;
1462
1463         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1464                 ui__error("Couldn't create thread/CPU maps: %s\n",
1465                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1466                 goto out_delete_evlist;
1467         }
1468
1469         if (top.delay_secs < 1)
1470                 top.delay_secs = 1;
1471
1472         if (record_opts__config(opts)) {
1473                 status = -EINVAL;
1474                 goto out_delete_evlist;
1475         }
1476
1477         top.sym_evsel = perf_evlist__first(top.evlist);
1478
1479         if (!callchain_param.enabled) {
1480                 symbol_conf.cumulate_callchain = false;
1481                 perf_hpp__cancel_cumulate();
1482         }
1483
1484         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1485                 callchain_param.order = ORDER_CALLER;
1486
1487         status = symbol__annotation_init();
1488         if (status < 0)
1489                 goto out_delete_evlist;
1490
1491         annotation_config__init();
1492
1493         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1494         if (symbol__init(NULL) < 0)
1495                 return -1;
1496
1497         sort__setup_elide(stdout);
1498
1499         get_term_dimensions(&top.winsize);
1500         if (top.print_entries == 0) {
1501                 perf_top__update_print_entries(&top);
1502                 signal(SIGWINCH, winch_sig);
1503         }
1504
1505         status = __cmd_top(&top);
1506
1507 out_delete_evlist:
1508         perf_evlist__delete(top.evlist);
1509
1510         return status;
1511 }