tools/perf/builtin-top.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-top.c
   4  *
   5  * Builtin top command: Display a continuously updated profile of
   6  * any workload, CPU or specific PID.
   7  *
   8  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
   9  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  10  *
  11  * Improvements and fixes by:
  12  *
  13  *   Arjan van de Ven <arjan@linux.intel.com>
  14  *   Yanmin Zhang <yanmin.zhang@intel.com>
  15  *   Wu Fengguang <fengguang.wu@intel.com>
  16  *   Mike Galbraith <efault@gmx.de>
  17  *   Paul Mackerras <paulus@samba.org>
  18  */
  19 #include "builtin.h"
  20
  21 #include "perf.h"
  22
  23 #include "util/annotate.h"
  24 #include "util/bpf-event.h"
  25 #include "util/cgroup.h"
  26 #include "util/config.h"
  27 #include "util/color.h"
  28 #include "util/dso.h"
  29 #include "util/evlist.h"
  30 #include "util/evsel.h"
  31 #include "util/evsel_config.h"
  32 #include "util/event.h"
  33 #include "util/machine.h"
  34 #include "util/map.h"
  35 #include "util/mmap.h"
  36 #include "util/session.h"
  37 #include "util/thread.h"
  38 #include "util/symbol.h"
  39 #include "util/synthetic-events.h"
  40 #include "util/top.h"
  41 #include "util/util.h"
  42 #include <linux/rbtree.h>
  43 #include <subcmd/parse-options.h>
  44 #include "util/parse-events.h"
  45 #include "util/callchain.h"
  46 #include "util/cpumap.h"
  47 #include "util/sort.h"
  48 #include "util/string2.h"
  49 #include "util/term.h"
  50 #include "util/intlist.h"
  51 #include "util/parse-branch-options.h"
  52 #include "arch/common.h"
  53 #include "ui/ui.h"
  54
  55 #include "util/debug.h"
  56 #include "util/ordered-events.h"
  57 #include "util/pfm.h"
  58
  59 #include <assert.h>
  60 #include <elf.h>
  61 #include <fcntl.h>
  62
  63 #include <stdio.h>
  64 #include <termios.h>
  65 #include <unistd.h>
  66 #include <inttypes.h>
  67
  68 #include <errno.h>
  69 #include <time.h>
  70 #include <sched.h>
  71 #include <signal.h>
  72
  73 #include <sys/syscall.h>
  74 #include <sys/ioctl.h>
  75 #include <poll.h>
  76 #include <sys/prctl.h>
  77 #include <sys/wait.h>
  78 #include <sys/uio.h>
  79 #include <sys/utsname.h>
  80 #include <sys/mman.h>
  81
  82 #include <linux/stringify.h>
  83 #include <linux/time64.h>
  84 #include <linux/types.h>
  85 #include <linux/err.h>
  86
  87 #include <linux/ctype.h>
  88 #include <perf/mmap.h>
  89
  90 static volatile sig_atomic_t done;
  91 static volatile sig_atomic_t resize;
  92
  93 #define HEADER_LINE_NR  5
  94
  95 static void perf_top__update_print_entries(struct perf_top *top)
  96 {
  97         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
  98 }
  99
 100 static void winch_sig(int sig __maybe_unused)
 101 {
 102         resize = 1;
 103 }
 104
 105 static void perf_top__resize(struct perf_top *top)
 106 {
 107         get_term_dimensions(&top->winsize);
 108         perf_top__update_print_entries(top);
 109 }
 110
 111 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
 112 {
 113         struct evsel *evsel;
 114         struct symbol *sym;
 115         struct annotation *notes;
 116         struct map *map;
 117         struct dso *dso;
 118         int err = -1;
 119
 120         if (!he || !he->ms.sym)
 121                 return -1;
 122
 123         evsel = hists_to_evsel(he->hists);
 124
 125         sym = he->ms.sym;
 126         map = he->ms.map;
 127         dso = map__dso(map);
 128
 129         /*
 130          * We can't annotate with just /proc/kallsyms
 131          */
 132         if (dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && !dso__is_kcore(dso)) {
 133                 pr_err("Can't annotate %s: No vmlinux file was found in the "
 134                        "path\n", sym->name);
 135                 sleep(1);
 136                 return -1;
 137         }
 138
 139         notes = symbol__annotation(sym);
 140         mutex_lock(&notes->lock);
 141
 142         if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 143                 mutex_unlock(&notes->lock);
 144                 pr_err("Not enough memory for annotating '%s' symbol!\n",
 145                        sym->name);
 146                 sleep(1);
 147                 return err;
 148         }
 149
 150         err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
 151         if (err == 0) {
 152                 top->sym_filter_entry = he;
 153         } else {
 154                 char msg[BUFSIZ];
 155                 symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
 156                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 157         }
 158
 159         mutex_unlock(&notes->lock);
 160         return err;
 161 }
 162
 163 static void __zero_source_counters(struct hist_entry *he)
 164 {
 165         struct symbol *sym = he->ms.sym;
 166         symbol__annotate_zero_histograms(sym);
 167 }
 168
 169 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
 170 {
 171         struct utsname uts;
 172         int err = uname(&uts);
 173         struct dso *dso = map__dso(map);
 174
 175         ui__warning("Out of bounds address found:\n\n"
 176                     "Addr:   %" PRIx64 "\n"
 177                     "DSO:    %s %c\n"
 178                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
 179                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
 180                     "Arch:   %s\n"
 181                     "Kernel: %s\n"
 182                     "Tools:  %s\n\n"
 183                     "Not all samples will be on the annotation output.\n\n"
 184                     "Please report to linux-kernel@vger.kernel.org\n",
 185                     ip, dso->long_name, dso__symtab_origin(dso),
 186                     map__start(map), map__end(map), sym->start, sym->end,
 187                     sym->binding == STB_GLOBAL ? 'g' :
 188                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
 189                     err ? "[unknown]" : uts.machine,
 190                     err ? "[unknown]" : uts.release, perf_version_string);
 191         if (use_browser <= 0)
 192                 sleep(5);
 193
 194         map__set_erange_warned(map, true);
 195 }
 196
 197 static void perf_top__record_precise_ip(struct perf_top *top,
 198                                         struct hist_entry *he,
 199                                         struct perf_sample *sample,
 200                                         struct evsel *evsel, u64 ip)
 201         EXCLUSIVE_LOCKS_REQUIRED(he->hists->lock)
 202 {
 203         struct annotation *notes;
 204         struct symbol *sym = he->ms.sym;
 205         int err = 0;
 206
 207         if (sym == NULL || (use_browser == 0 &&
 208                             (top->sym_filter_entry == NULL ||
 209                              top->sym_filter_entry->ms.sym != sym)))
 210                 return;
 211
 212         notes = symbol__annotation(sym);
 213
 214         if (!mutex_trylock(&notes->lock))
 215                 return;
 216
 217         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 218
 219         mutex_unlock(&notes->lock);
 220
 221         if (unlikely(err)) {
 222                 /*
 223                  * This function is now called with he->hists->lock held.
 224                  * Release it before going to sleep.
 225                  */
 226                 mutex_unlock(&he->hists->lock);
 227
 228                 if (err == -ERANGE && !map__erange_warned(he->ms.map))
 229                         ui__warn_map_erange(he->ms.map, sym, ip);
 230                 else if (err == -ENOMEM) {
 231                         pr_err("Not enough memory for annotating '%s' symbol!\n",
 232                                sym->name);
 233                         sleep(1);
 234                 }
 235
 236                 mutex_lock(&he->hists->lock);
 237         }
 238 }
 239
 240 static void perf_top__show_details(struct perf_top *top)
 241 {
 242         struct hist_entry *he = top->sym_filter_entry;
 243         struct evsel *evsel;
 244         struct annotation *notes;
 245         struct symbol *symbol;
 246         int more;
 247
 248         if (!he)
 249                 return;
 250
 251         evsel = hists_to_evsel(he->hists);
 252
 253         symbol = he->ms.sym;
 254         notes = symbol__annotation(symbol);
 255
 256         mutex_lock(&notes->lock);
 257
 258         symbol__calc_percent(symbol, evsel);
 259
 260         if (notes->src == NULL)
 261                 goto out_unlock;
 262
 263         printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
 264         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
 265
 266         more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
 267
 268         if (top->evlist->enabled) {
 269                 if (top->zero)
 270                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->core.idx);
 271                 else
 272                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->core.idx);
 273         }
 274         if (more != 0)
 275                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
 276 out_unlock:
 277         mutex_unlock(&notes->lock);
 278 }
 279
 280 static void perf_top__resort_hists(struct perf_top *t)
 281 {
 282         struct evlist *evlist = t->evlist;
 283         struct evsel *pos;
 284
 285         evlist__for_each_entry(evlist, pos) {
 286                 struct hists *hists = evsel__hists(pos);
 287
 288                 /*
 289                  * unlink existing entries so that they can be linked
 290                  * in a correct order in hists__match() below.
 291                  */
 292                 hists__unlink(hists);
 293
 294                 if (evlist->enabled) {
 295                         if (t->zero) {
 296                                 hists__delete_entries(hists);
 297                         } else {
 298                                 hists__decay_entries(hists, t->hide_user_symbols,
 299                                                      t->hide_kernel_symbols);
 300                         }
 301                 }
 302
 303                 hists__collapse_resort(hists, NULL);
 304
 305                 /* Non-group events are considered as leader */
 306                 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
 307                         struct hists *leader_hists = evsel__hists(evsel__leader(pos));
 308
 309                         hists__match(leader_hists, hists);
 310                         hists__link(leader_hists, hists);
 311                 }
 312         }
 313
 314         evlist__for_each_entry(evlist, pos) {
 315                 evsel__output_resort(pos, NULL);
 316         }
 317 }
 318
 319 static void perf_top__print_sym_table(struct perf_top *top)
 320 {
 321         char bf[160];
 322         int printed = 0;
 323         const int win_width = top->winsize.ws_col - 1;
 324         struct evsel *evsel = top->sym_evsel;
 325         struct hists *hists = evsel__hists(evsel);
 326
 327         puts(CONSOLE_CLEAR);
 328
 329         perf_top__header_snprintf(top, bf, sizeof(bf));
 330         printf("%s\n", bf);
 331
 332         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
 333
 334         if (!top->record_opts.overwrite &&
 335             (top->evlist->stats.nr_lost_warned !=
 336              top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
 337                 top->evlist->stats.nr_lost_warned =
 338                               top->evlist->stats.nr_events[PERF_RECORD_LOST];
 339                 color_fprintf(stdout, PERF_COLOR_RED,
 340                               "WARNING: LOST %d chunks, Check IO/CPU overload",
 341                               top->evlist->stats.nr_lost_warned);
 342                 ++printed;
 343         }
 344
 345         if (top->sym_filter_entry) {
 346                 perf_top__show_details(top);
 347                 return;
 348         }
 349
 350         perf_top__resort_hists(top);
 351
 352         hists__output_recalc_col_len(hists, top->print_entries - printed);
 353         putchar('\n');
 354         hists__fprintf(hists, false, top->print_entries - printed, win_width,
 355                        top->min_percent, stdout, !symbol_conf.use_callchain);
 356 }
 357
 358 static void prompt_integer(int *target, const char *msg)
 359 {
 360         char *buf = malloc(0), *p;
 361         size_t dummy = 0;
 362         int tmp;
 363
 364         fprintf(stdout, "\n%s: ", msg);
 365         if (getline(&buf, &dummy, stdin) < 0)
 366                 return;
 367
 368         p = strchr(buf, '\n');
 369         if (p)
 370                 *p = 0;
 371
 372         p = buf;
 373         while(*p) {
 374                 if (!isdigit(*p))
 375                         goto out_free;
 376                 p++;
 377         }
 378         tmp = strtoul(buf, NULL, 10);
 379         *target = tmp;
 380 out_free:
 381         free(buf);
 382 }
 383
 384 static void prompt_percent(int *target, const char *msg)
 385 {
 386         int tmp = 0;
 387
 388         prompt_integer(&tmp, msg);
 389         if (tmp >= 0 && tmp <= 100)
 390                 *target = tmp;
 391 }
 392
 393 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
 394 {
 395         char *buf = malloc(0), *p;
 396         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 397         struct hists *hists = evsel__hists(top->sym_evsel);
 398         struct rb_node *next;
 399         size_t dummy = 0;
 400
 401         /* zero counters of active symbol */
 402         if (syme) {
 403                 __zero_source_counters(syme);
 404                 top->sym_filter_entry = NULL;
 405         }
 406
 407         fprintf(stdout, "\n%s: ", msg);
 408         if (getline(&buf, &dummy, stdin) < 0)
 409                 goto out_free;
 410
 411         p = strchr(buf, '\n');
 412         if (p)
 413                 *p = 0;
 414
 415         next = rb_first_cached(&hists->entries);
 416         while (next) {
 417                 n = rb_entry(next, struct hist_entry, rb_node);
 418                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
 419                         found = n;
 420                         break;
 421                 }
 422                 next = rb_next(&n->rb_node);
 423         }
 424
 425         if (!found) {
 426                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
 427                 sleep(1);
 428         } else
 429                 perf_top__parse_source(top, found);
 430
 431 out_free:
 432         free(buf);
 433 }
 434
 435 static void perf_top__print_mapped_keys(struct perf_top *top)
 436 {
 437         char *name = NULL;
 438
 439         if (top->sym_filter_entry) {
 440                 struct symbol *sym = top->sym_filter_entry->ms.sym;
 441                 name = sym->name;
 442         }
 443
 444         fprintf(stdout, "\nMapped keys:\n");
 445         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
 446         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
 447
 448         if (top->evlist->core.nr_entries > 1)
 449                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
 450
 451         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
 452
 453         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
 454         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
 455         fprintf(stdout, "\t[S]     stop annotation.\n");
 456
 457         fprintf(stdout,
 458                 "\t[K]     hide kernel symbols.             \t(%s)\n",
 459                 top->hide_kernel_symbols ? "yes" : "no");
 460         fprintf(stdout,
 461                 "\t[U]     hide user symbols.               \t(%s)\n",
 462                 top->hide_user_symbols ? "yes" : "no");
 463         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
 464         fprintf(stdout, "\t[qQ]    quit.\n");
 465 }
 466
 467 static int perf_top__key_mapped(struct perf_top *top, int c)
 468 {
 469         switch (c) {
 470                 case 'd':
 471                 case 'e':
 472                 case 'f':
 473                 case 'z':
 474                 case 'q':
 475                 case 'Q':
 476                 case 'K':
 477                 case 'U':
 478                 case 'F':
 479                 case 's':
 480                 case 'S':
 481                         return 1;
 482                 case 'E':
 483                         return top->evlist->core.nr_entries > 1 ? 1 : 0;
 484                 default:
 485                         break;
 486         }
 487
 488         return 0;
 489 }
 490
 491 static bool perf_top__handle_keypress(struct perf_top *top, int c)
 492 {
 493         bool ret = true;
 494
 495         if (!perf_top__key_mapped(top, c)) {
 496                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 497                 struct termios save;
 498
 499                 perf_top__print_mapped_keys(top);
 500                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
 501                 fflush(stdout);
 502
 503                 set_term_quiet_input(&save);
 504
 505                 poll(&stdin_poll, 1, -1);
 506                 c = getc(stdin);
 507
 508                 tcsetattr(0, TCSAFLUSH, &save);
 509                 if (!perf_top__key_mapped(top, c))
 510                         return ret;
 511         }
 512
 513         switch (c) {
 514                 case 'd':
 515                         prompt_integer(&top->delay_secs, "Enter display delay");
 516                         if (top->delay_secs < 1)
 517                                 top->delay_secs = 1;
 518                         break;
 519                 case 'e':
 520                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
 521                         if (top->print_entries == 0) {
 522                                 perf_top__resize(top);
 523                                 signal(SIGWINCH, winch_sig);
 524                         } else {
 525                                 signal(SIGWINCH, SIG_DFL);
 526                         }
 527                         break;
 528                 case 'E':
 529                         if (top->evlist->core.nr_entries > 1) {
 530                                 /* Select 0 as the default event: */
 531                                 int counter = 0;
 532
 533                                 fprintf(stderr, "\nAvailable events:");
 534
 535                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 536                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel));
 537
 538                                 prompt_integer(&counter, "Enter details event counter");
 539
 540                                 if (counter >= top->evlist->core.nr_entries) {
 541                                         top->sym_evsel = evlist__first(top->evlist);
 542                                         fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
 543                                         sleep(1);
 544                                         break;
 545                                 }
 546                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
 547                                         if (top->sym_evsel->core.idx == counter)
 548                                                 break;
 549                         } else
 550                                 top->sym_evsel = evlist__first(top->evlist);
 551                         break;
 552                 case 'f':
 553                         prompt_integer(&top->count_filter, "Enter display event count filter");
 554                         break;
 555                 case 'F':
 556                         prompt_percent(&top->annotation_opts.min_pcnt,
 557                                        "Enter details display event filter (percent)");
 558                         break;
 559                 case 'K':
 560                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
 561                         break;
 562                 case 'q':
 563                 case 'Q':
 564                         printf("exiting.\n");
 565                         if (top->dump_symtab)
 566                                 perf_session__fprintf_dsos(top->session, stderr);
 567                         ret = false;
 568                         break;
 569                 case 's':
 570                         perf_top__prompt_symbol(top, "Enter details symbol");
 571                         break;
 572                 case 'S':
 573                         if (!top->sym_filter_entry)
 574                                 break;
 575                         else {
 576                                 struct hist_entry *syme = top->sym_filter_entry;
 577
 578                                 top->sym_filter_entry = NULL;
 579                                 __zero_source_counters(syme);
 580                         }
 581                         break;
 582                 case 'U':
 583                         top->hide_user_symbols = !top->hide_user_symbols;
 584                         break;
 585                 case 'z':
 586                         top->zero = !top->zero;
 587                         break;
 588                 default:
 589                         break;
 590         }
 591
 592         return ret;
 593 }
 594
 595 static void perf_top__sort_new_samples(void *arg)
 596 {
 597         struct perf_top *t = arg;
 598
 599         if (t->evlist->selected != NULL)
 600                 t->sym_evsel = t->evlist->selected;
 601
 602         perf_top__resort_hists(t);
 603
 604         if (t->lost || t->drop)
 605                 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
 606 }
 607
 608 static void stop_top(void)
 609 {
 610         session_done = 1;
 611         done = 1;
 612 }
 613
 614 static void *display_thread_tui(void *arg)
 615 {
 616         struct evsel *pos;
 617         struct perf_top *top = arg;
 618         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 619         struct hist_browser_timer hbt = {
 620                 .timer          = perf_top__sort_new_samples,
 621                 .arg            = top,
 622                 .refresh        = top->delay_secs,
 623         };
 624         int ret;
 625
 626         /* In order to read symbols from other namespaces perf to  needs to call
 627          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 628          * unshare(2) the fs so that we may continue to setns into namespaces
 629          * that we're observing.
 630          */
 631         unshare(CLONE_FS);
 632
 633         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 634
 635 repeat:
 636         perf_top__sort_new_samples(top);
 637
 638         /*
 639          * Initialize the uid_filter_str, in the future the TUI will allow
 640          * Zooming in/out UIDs. For now just use whatever the user passed
 641          * via --uid.
 642          */
 643         evlist__for_each_entry(top->evlist, pos) {
 644                 struct hists *hists = evsel__hists(pos);
 645                 hists->uid_filter_str = top->record_opts.target.uid_str;
 646         }
 647
 648         ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
 649                                        &top->session->header.env, !top->record_opts.overwrite,
 650                                        &top->annotation_opts);
 651         if (ret == K_RELOAD) {
 652                 top->zero = true;
 653                 goto repeat;
 654         } else
 655                 stop_top();
 656
 657         return NULL;
 658 }
 659
 660 static void display_sig(int sig __maybe_unused)
 661 {
 662         stop_top();
 663 }
 664
 665 static void display_setup_sig(void)
 666 {
 667         signal(SIGSEGV, sighandler_dump_stack);
 668         signal(SIGFPE, sighandler_dump_stack);
 669         signal(SIGINT,  display_sig);
 670         signal(SIGQUIT, display_sig);
 671         signal(SIGTERM, display_sig);
 672 }
 673
 674 static void *display_thread(void *arg)
 675 {
 676         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 677         struct termios save;
 678         struct perf_top *top = arg;
 679         int delay_msecs, c;
 680
 681         /* In order to read symbols from other namespaces perf to  needs to call
 682          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 683          * unshare(2) the fs so that we may continue to setns into namespaces
 684          * that we're observing.
 685          */
 686         unshare(CLONE_FS);
 687
 688         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 689
 690         display_setup_sig();
 691         pthread__unblock_sigwinch();
 692 repeat:
 693         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 694         set_term_quiet_input(&save);
 695         /* trash return*/
 696         clearerr(stdin);
 697         if (poll(&stdin_poll, 1, 0) > 0)
 698                 getc(stdin);
 699
 700         while (!done) {
 701                 perf_top__print_sym_table(top);
 702                 /*
 703                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 704                  * refresh screen in both cases.
 705                  */
 706                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 707                 case 0:
 708                         continue;
 709                 case -1:
 710                         if (errno == EINTR)
 711                                 continue;
 712                         fallthrough;
 713                 default:
 714                         c = getc(stdin);
 715                         tcsetattr(0, TCSAFLUSH, &save);
 716
 717                         if (perf_top__handle_keypress(top, c))
 718                                 goto repeat;
 719                         stop_top();
 720                 }
 721         }
 722
 723         tcsetattr(0, TCSAFLUSH, &save);
 724         return NULL;
 725 }
 726
 727 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 728                                    struct addr_location *al, bool single,
 729                                    void *arg)
 730         EXCLUSIVE_LOCKS_REQUIRED(iter->he->hists->lock)
 731 {
 732         struct perf_top *top = arg;
 733         struct evsel *evsel = iter->evsel;
 734
 735         if (perf_hpp_list.sym && single)
 736                 perf_top__record_precise_ip(top, iter->he, iter->sample, evsel, al->addr);
 737
 738         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 739                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
 740                      NULL);
 741         return 0;
 742 }
 743
 744 static void perf_event__process_sample(struct perf_tool *tool,
 745                                        const union perf_event *event,
 746                                        struct evsel *evsel,
 747                                        struct perf_sample *sample,
 748                                        struct machine *machine)
 749 {
 750         struct perf_top *top = container_of(tool, struct perf_top, tool);
 751         struct addr_location al;
 752
 753         if (!machine && perf_guest) {
 754                 static struct intlist *seen;
 755
 756                 if (!seen)
 757                         seen = intlist__new(NULL);
 758
 759                 if (!intlist__has_entry(seen, sample->pid)) {
 760                         pr_err("Can't find guest [%d]'s kernel information\n",
 761                                 sample->pid);
 762                         intlist__add(seen, sample->pid);
 763                 }
 764                 return;
 765         }
 766
 767         if (!machine) {
 768                 pr_err("%u unprocessable samples recorded.\r",
 769                        top->session->evlist->stats.nr_unprocessable_samples++);
 770                 return;
 771         }
 772
 773         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 774                 top->exact_samples++;
 775
 776         if (machine__resolve(machine, &al, sample) < 0)
 777                 return;
 778
 779         if (top->stitch_lbr)
 780                 al.thread->lbr_stitch_enable = true;
 781
 782         if (!machine->kptr_restrict_warned &&
 783             symbol_conf.kptr_restrict &&
 784             al.cpumode == PERF_RECORD_MISC_KERNEL) {
 785                 if (!evlist__exclude_kernel(top->session->evlist)) {
 786                         ui__warning(
 787 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
 788 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
 789 "Kernel%s samples will not be resolved.\n",
 790                           al.map && map__has_symbols(al.map) ?
 791                           " modules" : "");
 792                         if (use_browser <= 0)
 793                                 sleep(5);
 794                 }
 795                 machine->kptr_restrict_warned = true;
 796         }
 797
 798         if (al.sym == NULL && al.map != NULL) {
 799                 const char *msg = "Kernel samples will not be resolved.\n";
 800                 /*
 801                  * As we do lazy loading of symtabs we only will know if the
 802                  * specified vmlinux file is invalid when we actually have a
 803                  * hit in kernel space and then try to load it. So if we get
 804                  * here and there are _no_ symbols in the DSO backing the
 805                  * kernel map, bail out.
 806                  *
 807                  * We may never get here, for instance, if we use -K/
 808                  * --hide-kernel-symbols, even if the user specifies an
 809                  * invalid --vmlinux ;-)
 810                  */
 811                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
 812                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
 813                         if (symbol_conf.vmlinux_name) {
 814                                 char serr[256];
 815
 816                                 dso__strerror_load(map__dso(al.map), serr, sizeof(serr));
 817                                 ui__warning("The %s file can't be used: %s\n%s",
 818                                             symbol_conf.vmlinux_name, serr, msg);
 819                         } else {
 820                                 ui__warning("A vmlinux file was not found.\n%s",
 821                                             msg);
 822                         }
 823
 824                         if (use_browser <= 0)
 825                                 sleep(5);
 826                         top->vmlinux_warned = true;
 827                 }
 828         }
 829
 830         if (al.sym == NULL || !al.sym->idle) {
 831                 struct hists *hists = evsel__hists(evsel);
 832                 struct hist_entry_iter iter = {
 833                         .evsel          = evsel,
 834                         .sample         = sample,
 835                         .add_entry_cb   = hist_iter__top_callback,
 836                 };
 837
 838                 if (symbol_conf.cumulate_callchain)
 839                         iter.ops = &hist_iter_cumulative;
 840                 else
 841                         iter.ops = &hist_iter_normal;
 842
 843                 mutex_lock(&hists->lock);
 844
 845                 if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0)
 846                         pr_err("Problem incrementing symbol period, skipping event\n");
 847
 848                 mutex_unlock(&hists->lock);
 849         }
 850
 851         addr_location__put(&al);
 852 }
 853
 854 static void
 855 perf_top__process_lost(struct perf_top *top, union perf_event *event,
 856                        struct evsel *evsel)
 857 {
 858         top->lost += event->lost.lost;
 859         top->lost_total += event->lost.lost;
 860         evsel->evlist->stats.total_lost += event->lost.lost;
 861 }
 862
 863 static void
 864 perf_top__process_lost_samples(struct perf_top *top,
 865                                union perf_event *event,
 866                                struct evsel *evsel)
 867 {
 868         top->lost += event->lost_samples.lost;
 869         top->lost_total += event->lost_samples.lost;
 870         evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
 871 }
 872
 873 static u64 last_timestamp;
 874
 875 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 876 {
 877         struct record_opts *opts = &top->record_opts;
 878         struct evlist *evlist = top->evlist;
 879         struct mmap *md;
 880         union perf_event *event;
 881
 882         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 883         if (perf_mmap__read_init(&md->core) < 0)
 884                 return;
 885
 886         while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 887                 int ret;
 888
 889                 ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 890                 if (ret && ret != -1)
 891                         break;
 892
 893                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
 894                 if (ret)
 895                         break;
 896
 897                 perf_mmap__consume(&md->core);
 898
 899                 if (top->qe.rotate) {
 900                         mutex_lock(&top->qe.mutex);
 901                         top->qe.rotate = false;
 902                         cond_signal(&top->qe.cond);
 903                         mutex_unlock(&top->qe.mutex);
 904                 }
 905         }
 906
 907         perf_mmap__read_done(&md->core);
 908 }
 909
 910 static void perf_top__mmap_read(struct perf_top *top)
 911 {
 912         bool overwrite = top->record_opts.overwrite;
 913         struct evlist *evlist = top->evlist;
 914         int i;
 915
 916         if (overwrite)
 917                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 918
 919         for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 920                 perf_top__mmap_read_idx(top, i);
 921
 922         if (overwrite) {
 923                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 924                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 925         }
 926 }
 927
 928 /*
 929  * Check per-event overwrite term.
 930  * perf top should support consistent term for all events.
 931  * - All events don't have per-event term
 932  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 933  *   Nothing change, return 0.
 934  * - All events have same per-event term
 935  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 936  *   Using the per-event setting to replace the opts->overwrite if
 937  *   they are different, then return 0.
 938  * - Events have different per-event term
 939  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 940  *   Return -1
 941  * - Some of the event set per-event term, but some not.
 942  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 943  *   Return -1
 944  */
 945 static int perf_top__overwrite_check(struct perf_top *top)
 946 {
 947         struct record_opts *opts = &top->record_opts;
 948         struct evlist *evlist = top->evlist;
 949         struct evsel_config_term *term;
 950         struct list_head *config_terms;
 951         struct evsel *evsel;
 952         int set, overwrite = -1;
 953
 954         evlist__for_each_entry(evlist, evsel) {
 955                 set = -1;
 956                 config_terms = &evsel->config_terms;
 957                 list_for_each_entry(term, config_terms, list) {
 958                         if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
 959                                 set = term->val.overwrite ? 1 : 0;
 960                 }
 961
 962                 /* no term for current and previous event (likely) */
 963                 if ((overwrite < 0) && (set < 0))
 964                         continue;
 965
 966                 /* has term for both current and previous event, compare */
 967                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 968                         return -1;
 969
 970                 /* no term for current event but has term for previous one */
 971                 if ((overwrite >= 0) && (set < 0))
 972                         return -1;
 973
 974                 /* has term for current event */
 975                 if ((overwrite < 0) && (set >= 0)) {
 976                         /* if it's first event, set overwrite */
 977                         if (evsel == evlist__first(evlist))
 978                                 overwrite = set;
 979                         else
 980                                 return -1;
 981                 }
 982         }
 983
 984         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 985                 opts->overwrite = overwrite;
 986
 987         return 0;
 988 }
 989
 990 static int perf_top_overwrite_fallback(struct perf_top *top,
 991                                        struct evsel *evsel)
 992 {
 993         struct record_opts *opts = &top->record_opts;
 994         struct evlist *evlist = top->evlist;
 995         struct evsel *counter;
 996
 997         if (!opts->overwrite)
 998                 return 0;
 999
1000         /* only fall back when first event fails */
1001         if (evsel != evlist__first(evlist))
1002                 return 0;
1003
1004         evlist__for_each_entry(evlist, counter)
1005                 counter->core.attr.write_backward = false;
1006         opts->overwrite = false;
1007         pr_debug2("fall back to non-overwrite mode\n");
1008         return 1;
1009 }
1010
1011 static int perf_top__start_counters(struct perf_top *top)
1012 {
1013         char msg[BUFSIZ];
1014         struct evsel *counter;
1015         struct evlist *evlist = top->evlist;
1016         struct record_opts *opts = &top->record_opts;
1017
1018         if (perf_top__overwrite_check(top)) {
1019                 ui__error("perf top only support consistent per-event "
1020                           "overwrite setting for all events\n");
1021                 goto out_err;
1022         }
1023
1024         evlist__config(evlist, opts, &callchain_param);
1025
1026         evlist__for_each_entry(evlist, counter) {
1027 try_again:
1028                 if (evsel__open(counter, top->evlist->core.user_requested_cpus,
1029                                      top->evlist->core.threads) < 0) {
1030
1031                         /*
1032                          * Specially handle overwrite fall back.
1033                          * Because perf top is the only tool which has
1034                          * overwrite mode by default, support
1035                          * both overwrite and non-overwrite mode, and
1036                          * require consistent mode for all events.
1037                          *
1038                          * May move it to generic code with more tools
1039                          * have similar attribute.
1040                          */
1041                         if (perf_missing_features.write_backward &&
1042                             perf_top_overwrite_fallback(top, counter))
1043                                 goto try_again;
1044
1045                         if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
1046                                 if (verbose > 0)
1047                                         ui__warning("%s\n", msg);
1048                                 goto try_again;
1049                         }
1050
1051                         evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1052                         ui__error("%s\n", msg);
1053                         goto out_err;
1054                 }
1055         }
1056
1057         if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1058                 ui__error("Failed to mmap with %d (%s)\n",
1059                             errno, str_error_r(errno, msg, sizeof(msg)));
1060                 goto out_err;
1061         }
1062
1063         return 0;
1064
1065 out_err:
1066         return -1;
1067 }
1068
1069 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1070 {
1071         if (callchain->mode != CHAIN_NONE) {
1072                 if (callchain_register_param(callchain) < 0) {
1073                         ui__error("Can't register callchain params.\n");
1074                         return -EINVAL;
1075                 }
1076         }
1077
1078         return 0;
1079 }
1080
1081 static struct ordered_events *rotate_queues(struct perf_top *top)
1082 {
1083         struct ordered_events *in = top->qe.in;
1084
1085         if (top->qe.in == &top->qe.data[1])
1086                 top->qe.in = &top->qe.data[0];
1087         else
1088                 top->qe.in = &top->qe.data[1];
1089
1090         return in;
1091 }
1092
1093 static void *process_thread(void *arg)
1094 {
1095         struct perf_top *top = arg;
1096
1097         while (!done) {
1098                 struct ordered_events *out, *in = top->qe.in;
1099
1100                 if (!in->nr_events) {
1101                         usleep(100);
1102                         continue;
1103                 }
1104
1105                 out = rotate_queues(top);
1106
1107                 mutex_lock(&top->qe.mutex);
1108                 top->qe.rotate = true;
1109                 cond_wait(&top->qe.cond, &top->qe.mutex);
1110                 mutex_unlock(&top->qe.mutex);
1111
1112                 if (ordered_events__flush(out, OE_FLUSH__TOP))
1113                         pr_err("failed to process events\n");
1114         }
1115
1116         return NULL;
1117 }
1118
1119 /*
1120  * Allow only 'top->delay_secs' seconds behind samples.
1121  */
1122 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1123 {
1124         union perf_event *event = qevent->event;
1125         u64 delay_timestamp;
1126
1127         if (event->header.type != PERF_RECORD_SAMPLE)
1128                 return false;
1129
1130         delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1131         return delay_timestamp < last_timestamp;
1132 }
1133
1134 static int deliver_event(struct ordered_events *qe,
1135                          struct ordered_event *qevent)
1136 {
1137         struct perf_top *top = qe->data;
1138         struct evlist *evlist = top->evlist;
1139         struct perf_session *session = top->session;
1140         union perf_event *event = qevent->event;
1141         struct perf_sample sample;
1142         struct evsel *evsel;
1143         struct machine *machine;
1144         int ret = -1;
1145
1146         if (should_drop(qevent, top)) {
1147                 top->drop++;
1148                 top->drop_total++;
1149                 return 0;
1150         }
1151
1152         ret = evlist__parse_sample(evlist, event, &sample);
1153         if (ret) {
1154                 pr_err("Can't parse sample, err = %d\n", ret);
1155                 goto next_event;
1156         }
1157
1158         evsel = evlist__id2evsel(session->evlist, sample.id);
1159         assert(evsel != NULL);
1160
1161         if (event->header.type == PERF_RECORD_SAMPLE) {
1162                 if (evswitch__discard(&top->evswitch, evsel))
1163                         return 0;
1164                 ++top->samples;
1165         }
1166
1167         switch (sample.cpumode) {
1168         case PERF_RECORD_MISC_USER:
1169                 ++top->us_samples;
1170                 if (top->hide_user_symbols)
1171                         goto next_event;
1172                 machine = &session->machines.host;
1173                 break;
1174         case PERF_RECORD_MISC_KERNEL:
1175                 ++top->kernel_samples;
1176                 if (top->hide_kernel_symbols)
1177                         goto next_event;
1178                 machine = &session->machines.host;
1179                 break;
1180         case PERF_RECORD_MISC_GUEST_KERNEL:
1181                 ++top->guest_kernel_samples;
1182                 machine = perf_session__find_machine(session,
1183                                                      sample.pid);
1184                 break;
1185         case PERF_RECORD_MISC_GUEST_USER:
1186                 ++top->guest_us_samples;
1187                 /*
1188                  * TODO: we don't process guest user from host side
1189                  * except simple counting.
1190                  */
1191                 goto next_event;
1192         default:
1193                 if (event->header.type == PERF_RECORD_SAMPLE)
1194                         goto next_event;
1195                 machine = &session->machines.host;
1196                 break;
1197         }
1198
1199         if (event->header.type == PERF_RECORD_SAMPLE) {
1200                 perf_event__process_sample(&top->tool, event, evsel,
1201                                            &sample, machine);
1202         } else if (event->header.type == PERF_RECORD_LOST) {
1203                 perf_top__process_lost(top, event, evsel);
1204         } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1205                 perf_top__process_lost_samples(top, event, evsel);
1206         } else if (event->header.type < PERF_RECORD_MAX) {
1207                 events_stats__inc(&session->evlist->stats, event->header.type);
1208                 machine__process_event(machine, event, &sample);
1209         } else
1210                 ++session->evlist->stats.nr_unknown_events;
1211
1212         ret = 0;
1213 next_event:
1214         return ret;
1215 }
1216
1217 static void init_process_thread(struct perf_top *top)
1218 {
1219         ordered_events__init(&top->qe.data[0], deliver_event, top);
1220         ordered_events__init(&top->qe.data[1], deliver_event, top);
1221         ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1222         ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1223         top->qe.in = &top->qe.data[0];
1224         mutex_init(&top->qe.mutex);
1225         cond_init(&top->qe.cond);
1226 }
1227
1228 static int __cmd_top(struct perf_top *top)
1229 {
1230         struct record_opts *opts = &top->record_opts;
1231         pthread_t thread, thread_process;
1232         int ret;
1233
1234         if (!top->annotation_opts.objdump_path) {
1235                 ret = perf_env__lookup_objdump(&top->session->header.env,
1236                                                &top->annotation_opts.objdump_path);
1237                 if (ret)
1238                         return ret;
1239         }
1240
1241         ret = callchain_param__setup_sample_type(&callchain_param);
1242         if (ret)
1243                 return ret;
1244
1245         if (perf_session__register_idle_thread(top->session) < 0)
1246                 return ret;
1247
1248         if (top->nr_threads_synthesize > 1)
1249                 perf_set_multithreaded();
1250
1251         init_process_thread(top);
1252
1253         if (opts->record_namespaces)
1254                 top->tool.namespace_events = true;
1255         if (opts->record_cgroup) {
1256 #ifdef HAVE_FILE_HANDLE
1257                 top->tool.cgroup_events = true;
1258 #else
1259                 pr_err("cgroup tracking is not supported.\n");
1260                 return -1;
1261 #endif
1262         }
1263
1264         ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1265                                                 &top->session->machines.host,
1266                                                 &top->record_opts);
1267         if (ret < 0)
1268                 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1269
1270         ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1271                                              &top->session->machines.host);
1272         if (ret < 0)
1273                 pr_debug("Couldn't synthesize cgroup events.\n");
1274
1275         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1276                                     top->evlist->core.threads, true, false,
1277                                     top->nr_threads_synthesize);
1278
1279         perf_set_multithreaded();
1280
1281         if (perf_hpp_list.socket) {
1282                 ret = perf_env__read_cpu_topology_map(&perf_env);
1283                 if (ret < 0) {
1284                         char errbuf[BUFSIZ];
1285                         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1286
1287                         ui__error("Could not read the CPU topology map: %s\n", err);
1288                         return ret;
1289                 }
1290         }
1291
1292         ret = perf_top__start_counters(top);
1293         if (ret)
1294                 return ret;
1295
1296         top->session->evlist = top->evlist;
1297         perf_session__set_id_hdr_size(top->session);
1298
1299         /*
1300          * When perf is starting the traced process, all the events (apart from
1301          * group members) have enable_on_exec=1 set, so don't spoil it by
1302          * prematurely enabling them.
1303          *
1304          * XXX 'top' still doesn't start workloads like record, trace, but should,
1305          * so leave the check here.
1306          */
1307         if (!target__none(&opts->target))
1308                 evlist__enable(top->evlist);
1309
1310         ret = -1;
1311         if (pthread_create(&thread_process, NULL, process_thread, top)) {
1312                 ui__error("Could not create process thread.\n");
1313                 return ret;
1314         }
1315
1316         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1317                                                             display_thread), top)) {
1318                 ui__error("Could not create display thread.\n");
1319                 goto out_join_thread;
1320         }
1321
1322         if (top->realtime_prio) {
1323                 struct sched_param param;
1324
1325                 param.sched_priority = top->realtime_prio;
1326                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1327                         ui__error("Could not set realtime priority.\n");
1328                         goto out_join;
1329                 }
1330         }
1331
1332         /* Wait for a minimal set of events before starting the snapshot */
1333         evlist__poll(top->evlist, 100);
1334
1335         perf_top__mmap_read(top);
1336
1337         while (!done) {
1338                 u64 hits = top->samples;
1339
1340                 perf_top__mmap_read(top);
1341
1342                 if (opts->overwrite || (hits == top->samples))
1343                         ret = evlist__poll(top->evlist, 100);
1344
1345                 if (resize) {
1346                         perf_top__resize(top);
1347                         resize = 0;
1348                 }
1349         }
1350
1351         ret = 0;
1352 out_join:
1353         pthread_join(thread, NULL);
1354 out_join_thread:
1355         cond_signal(&top->qe.cond);
1356         pthread_join(thread_process, NULL);
1357         perf_set_singlethreaded();
1358         return ret;
1359 }
1360
1361 static int
1362 callchain_opt(const struct option *opt, const char *arg, int unset)
1363 {
1364         symbol_conf.use_callchain = true;
1365         return record_callchain_opt(opt, arg, unset);
1366 }
1367
1368 static int
1369 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1370 {
1371         struct callchain_param *callchain = opt->value;
1372
1373         callchain->enabled = !unset;
1374         callchain->record_mode = CALLCHAIN_FP;
1375
1376         /*
1377          * --no-call-graph
1378          */
1379         if (unset) {
1380                 symbol_conf.use_callchain = false;
1381                 callchain->record_mode = CALLCHAIN_NONE;
1382                 return 0;
1383         }
1384
1385         return parse_callchain_top_opt(arg);
1386 }
1387
1388 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1389 {
1390         if (!strcmp(var, "top.call-graph")) {
1391                 var = "call-graph.record-mode";
1392                 return perf_default_config(var, value, cb);
1393         }
1394         if (!strcmp(var, "top.children")) {
1395                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1396                 return 0;
1397         }
1398
1399         return 0;
1400 }
1401
1402 static int
1403 parse_percent_limit(const struct option *opt, const char *arg,
1404                     int unset __maybe_unused)
1405 {
1406         struct perf_top *top = opt->value;
1407
1408         top->min_percent = strtof(arg, NULL);
1409         return 0;
1410 }
1411
1412 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1413         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1414
1415 int cmd_top(int argc, const char **argv)
1416 {
1417         char errbuf[BUFSIZ];
1418         struct perf_top top = {
1419                 .count_filter        = 5,
1420                 .delay_secs          = 2,
1421                 .record_opts = {
1422                         .mmap_pages     = UINT_MAX,
1423                         .user_freq      = UINT_MAX,
1424                         .user_interval  = ULLONG_MAX,
1425                         .freq           = 4000, /* 4 KHz */
1426                         .target         = {
1427                                 .uses_mmap   = true,
1428                         },
1429                         /*
1430                          * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1431                          * when we pause, fix that and reenable. Probably using a
1432                          * separate evlist with a dummy event, i.e. a non-overwrite
1433                          * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1434                          * stays in overwrite mode. -acme
1435                          * */
1436                         .overwrite      = 0,
1437                         .sample_time    = true,
1438                         .sample_time_set = true,
1439                 },
1440                 .max_stack           = sysctl__max_stack(),
1441                 .nr_threads_synthesize = UINT_MAX,
1442         };
1443         bool branch_call_mode = false;
1444         struct record_opts *opts = &top.record_opts;
1445         struct target *target = &opts->target;
1446         const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
1447         const struct option options[] = {
1448         OPT_CALLBACK('e', "event", &top.evlist, "event",
1449                      "event selector. use 'perf list' to list available events",
1450                      parse_events_option),
1451         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1452         OPT_STRING('p', "pid", &target->pid, "pid",
1453                     "profile events on existing process id"),
1454         OPT_STRING('t', "tid", &target->tid, "tid",
1455                     "profile events on existing thread id"),
1456         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1457                             "system-wide collection from all CPUs"),
1458         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1459                     "list of cpus to monitor"),
1460         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1461                    "file", "vmlinux pathname"),
1462         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1463                     "don't load vmlinux even if found"),
1464         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1465                    "file", "kallsyms pathname"),
1466         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1467                     "hide kernel symbols"),
1468         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1469                      "number of mmap data pages", evlist__parse_mmap_pages),
1470         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1471                     "collect data with this RT SCHED_FIFO priority"),
1472         OPT_INTEGER('d', "delay", &top.delay_secs,
1473                     "number of seconds to delay between refreshes"),
1474         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1475                             "dump the symbol table used for profiling"),
1476         OPT_INTEGER('f', "count-filter", &top.count_filter,
1477                     "only display functions with more events than this"),
1478         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1479                     "child tasks do not inherit counters"),
1480         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1481                     "symbol to annotate"),
1482         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1483         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1484                      "profile at this frequency",
1485                       record__parse_freq),
1486         OPT_INTEGER('E', "entries", &top.print_entries,
1487                     "display this many functions"),
1488         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1489                     "hide user symbols"),
1490 #ifdef HAVE_SLANG_SUPPORT
1491         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1492 #endif
1493         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1494         OPT_INCR('v', "verbose", &verbose,
1495                     "be more verbose (show counter open errors, etc)"),
1496         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1497                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1498                    " Please refer the man page for the complete list."),
1499         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1500                    "output field(s): overhead, period, sample plus all of sort keys"),
1501         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1502                     "Show a column with the number of samples"),
1503         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1504                            NULL, "enables call-graph recording and display",
1505                            &callchain_opt),
1506         OPT_CALLBACK(0, "call-graph", &callchain_param,
1507                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1508                      top_callchain_help, &parse_callchain_opt),
1509         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1510                     "Accumulate callchains of children and show total overhead as well"),
1511         OPT_INTEGER(0, "max-stack", &top.max_stack,
1512                     "Set the maximum stack depth when parsing the callchain. "
1513                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1514         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1515                    "ignore callees of these functions in call graphs",
1516                    report_parse_ignore_callees_opt),
1517         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1518                     "Show a column with the sum of periods"),
1519         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1520                    "only consider symbols in these dsos"),
1521         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1522                    "only consider symbols in these comms"),
1523         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1524                    "only consider these symbols"),
1525         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1526                     "Interleave source code with assembly code (default)"),
1527         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1528                     "Display raw encoding of assembly instructions (default)"),
1529         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1530                     "Enable kernel symbol demangling"),
1531         OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1532         OPT_STRING(0, "objdump", &objdump_path, "path",
1533                     "objdump binary to use for disassembly and annotations"),
1534         OPT_STRING(0, "addr2line", &addr2line_path, "path",
1535                    "addr2line binary to use for line numbers"),
1536         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1537                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1538         OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix",
1539                     "Add prefix to source file path names in programs (with --prefix-strip)"),
1540         OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N",
1541                     "Strip first N entries of source file path name in programs (with --prefix)"),
1542         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1543         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1544                      "Don't show entries under that percent", parse_percent_limit),
1545         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1546                      "How to display percentage of filtered entries", parse_filter_percentage),
1547         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1548                    "width[,width...]",
1549                    "don't try to adjust column width, use these fixed values"),
1550         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1551                         "per thread proc mmap processing timeout in ms"),
1552         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1553                      "branch any", "sample any taken branches",
1554                      parse_branch_stack),
1555         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1556                      "branch filter mask", "branch stack filter modes",
1557                      parse_branch_stack),
1558         OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1559                     "add last branch records to call history"),
1560         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1561                     "Show raw trace event output (do not use print fmt or plugins)"),
1562         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1563                     "Show entries in a hierarchy"),
1564         OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1565                     "Use a backward ring buffer, default: no"),
1566         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1567         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1568                         "number of thread to run event synthesize"),
1569         OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
1570                      "monitor event in cgroup name only", parse_cgroups),
1571         OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1572                     "Record namespaces events"),
1573         OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1574                     "Record cgroup events"),
1575         OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1576                     "Sort the output by the event at the index n in group. "
1577                     "If n is invalid, sort by the first event. "
1578                     "WARNING: should be used on grouped events."),
1579         OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1580                     "Enable LBR callgraph stitching approach"),
1581 #ifdef HAVE_LIBPFM
1582         OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1583                 "libpfm4 event selector. use 'perf list' to list available events",
1584                 parse_libpfm_events_option),
1585 #endif
1586         OPTS_EVSWITCH(&top.evswitch),
1587         OPT_END()
1588         };
1589         const char * const top_usage[] = {
1590                 "perf top [<options>]",
1591                 NULL
1592         };
1593         int status = hists__init();
1594
1595         if (status < 0)
1596                 return status;
1597
1598         annotation_options__init(&top.annotation_opts);
1599
1600         top.annotation_opts.min_pcnt = 5;
1601         top.annotation_opts.context  = 4;
1602
1603         top.evlist = evlist__new();
1604         if (top.evlist == NULL)
1605                 return -ENOMEM;
1606
1607         status = perf_config(perf_top_config, &top);
1608         if (status)
1609                 return status;
1610         /*
1611          * Since the per arch annotation init routine may need the cpuid, read
1612          * it here, since we are not getting this from the perf.data header.
1613          */
1614         status = perf_env__read_cpuid(&perf_env);
1615         if (status) {
1616                 /*
1617                  * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1618                  * warn the user explicitly.
1619                  */
1620                 eprintf(status == ENOSYS ? 1 : 0, verbose,
1621                         "Couldn't read the cpuid for this machine: %s\n",
1622                         str_error_r(errno, errbuf, sizeof(errbuf)));
1623         }
1624         top.evlist->env = &perf_env;
1625
1626         argc = parse_options(argc, argv, options, top_usage, 0);
1627         if (argc)
1628                 usage_with_options(top_usage, options);
1629
1630         if (disassembler_style) {
1631                 top.annotation_opts.disassembler_style = strdup(disassembler_style);
1632                 if (!top.annotation_opts.disassembler_style)
1633                         return -ENOMEM;
1634         }
1635         if (objdump_path) {
1636                 top.annotation_opts.objdump_path = strdup(objdump_path);
1637                 if (!top.annotation_opts.objdump_path)
1638                         return -ENOMEM;
1639         }
1640         if (addr2line_path) {
1641                 symbol_conf.addr2line_path = strdup(addr2line_path);
1642                 if (!symbol_conf.addr2line_path)
1643                         return -ENOMEM;
1644         }
1645
1646         status = symbol__validate_sym_arguments();
1647         if (status)
1648                 goto out_delete_evlist;
1649
1650         if (annotate_check_args(&top.annotation_opts) < 0)
1651                 goto out_delete_evlist;
1652
1653         if (!top.evlist->core.nr_entries &&
1654             evlist__add_default(top.evlist) < 0) {
1655                 pr_err("Not enough memory for event selector list\n");
1656                 goto out_delete_evlist;
1657         }
1658
1659         status = evswitch__init(&top.evswitch, top.evlist, stderr);
1660         if (status)
1661                 goto out_delete_evlist;
1662
1663         if (symbol_conf.report_hierarchy) {
1664                 /* disable incompatible options */
1665                 symbol_conf.event_group = false;
1666                 symbol_conf.cumulate_callchain = false;
1667
1668                 if (field_order) {
1669                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1670                         parse_options_usage(top_usage, options, "fields", 0);
1671                         parse_options_usage(NULL, options, "hierarchy", 0);
1672                         goto out_delete_evlist;
1673                 }
1674         }
1675
1676         if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1677                 pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1678                 goto out_delete_evlist;
1679         }
1680
1681         if (nr_cgroups > 0 && opts->record_cgroup) {
1682                 pr_err("--cgroup and --all-cgroups cannot be used together\n");
1683                 goto out_delete_evlist;
1684         }
1685
1686         if (branch_call_mode) {
1687                 if (!opts->branch_stack)
1688                         opts->branch_stack = PERF_SAMPLE_BRANCH_ANY;
1689                 symbol_conf.use_callchain = true;
1690                 callchain_param.key = CCKEY_ADDRESS;
1691                 callchain_param.branch_callstack = true;
1692                 callchain_param.enabled = true;
1693                 if (callchain_param.record_mode == CALLCHAIN_NONE)
1694                         callchain_param.record_mode = CALLCHAIN_FP;
1695                 callchain_register_param(&callchain_param);
1696                 if (!sort_order)
1697                         sort_order = "srcline,symbol,dso";
1698         }
1699
1700         if (opts->branch_stack && callchain_param.enabled)
1701                 symbol_conf.show_branchflag_count = true;
1702
1703         sort__mode = SORT_MODE__TOP;
1704         /* display thread wants entries to be collapsed in a different tree */
1705         perf_hpp_list.need_collapse = 1;
1706
1707         if (top.use_stdio)
1708                 use_browser = 0;
1709 #ifdef HAVE_SLANG_SUPPORT
1710         else if (top.use_tui)
1711                 use_browser = 1;
1712 #endif
1713
1714         setup_browser(false);
1715
1716         if (setup_sorting(top.evlist) < 0) {
1717                 if (sort_order)
1718                         parse_options_usage(top_usage, options, "s", 1);
1719                 if (field_order)
1720                         parse_options_usage(sort_order ? NULL : top_usage,
1721                                             options, "fields", 0);
1722                 goto out_delete_evlist;
1723         }
1724
1725         status = target__validate(target);
1726         if (status) {
1727                 target__strerror(target, status, errbuf, BUFSIZ);
1728                 ui__warning("%s\n", errbuf);
1729         }
1730
1731         status = target__parse_uid(target);
1732         if (status) {
1733                 int saved_errno = errno;
1734
1735                 target__strerror(target, status, errbuf, BUFSIZ);
1736                 ui__error("%s\n", errbuf);
1737
1738                 status = -saved_errno;
1739                 goto out_delete_evlist;
1740         }
1741
1742         if (target__none(target))
1743                 target->system_wide = true;
1744
1745         if (evlist__create_maps(top.evlist, target) < 0) {
1746                 ui__error("Couldn't create thread/CPU maps: %s\n",
1747                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1748                 status = -errno;
1749                 goto out_delete_evlist;
1750         }
1751
1752         if (top.delay_secs < 1)
1753                 top.delay_secs = 1;
1754
1755         if (record_opts__config(opts)) {
1756                 status = -EINVAL;
1757                 goto out_delete_evlist;
1758         }
1759
1760         top.sym_evsel = evlist__first(top.evlist);
1761
1762         if (!callchain_param.enabled) {
1763                 symbol_conf.cumulate_callchain = false;
1764                 perf_hpp__cancel_cumulate();
1765         }
1766
1767         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1768                 callchain_param.order = ORDER_CALLER;
1769
1770         status = symbol__annotation_init();
1771         if (status < 0)
1772                 goto out_delete_evlist;
1773
1774         annotation_config__init(&top.annotation_opts);
1775
1776         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1777         status = symbol__init(NULL);
1778         if (status < 0)
1779                 goto out_delete_evlist;
1780
1781         sort__setup_elide(stdout);
1782
1783         get_term_dimensions(&top.winsize);
1784         if (top.print_entries == 0) {
1785                 perf_top__update_print_entries(&top);
1786                 signal(SIGWINCH, winch_sig);
1787         }
1788
1789         top.session = perf_session__new(NULL, NULL);
1790         if (IS_ERR(top.session)) {
1791                 status = PTR_ERR(top.session);
1792                 goto out_delete_evlist;
1793         }
1794
1795 #ifdef HAVE_LIBBPF_SUPPORT
1796         if (!top.record_opts.no_bpf_event) {
1797                 top.sb_evlist = evlist__new();
1798
1799                 if (top.sb_evlist == NULL) {
1800                         pr_err("Couldn't create side band evlist.\n.");
1801                         status = -EINVAL;
1802                         goto out_delete_evlist;
1803                 }
1804
1805                 if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1806                         pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1807                         status = -EINVAL;
1808                         goto out_delete_evlist;
1809                 }
1810         }
1811 #endif
1812
1813         if (evlist__start_sb_thread(top.sb_evlist, target)) {
1814                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1815                 opts->no_bpf_event = true;
1816         }
1817
1818         status = __cmd_top(&top);
1819
1820         if (!opts->no_bpf_event)
1821                 evlist__stop_sb_thread(top.sb_evlist);
1822
1823 out_delete_evlist:
1824         evlist__delete(top.evlist);
1825         perf_session__delete(top.session);
1826         annotation_options__exit(&top.annotation_opts);
1827
1828         return status;
1829 }