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         annotation__lock(notes);
 141
 142         if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
 143                 annotation__unlock(notes);
 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, 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         annotation__unlock(notes);
 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 (!annotation__trylock(notes))
 215                 return;
 216
 217         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
 218
 219         annotation__unlock(notes);
 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         annotation__lock(notes);
 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", annotate_opts.min_pcnt);
 265
 266         more = symbol__annotate_printf(&he->ms, top->sym_evsel);
 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         annotation__unlock(notes);
 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 = NULL, *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 = NULL, *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", annotate_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(&annotate_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         if (ret == K_RELOAD) {
 651                 top->zero = true;
 652                 goto repeat;
 653         } else
 654                 stop_top();
 655
 656         return NULL;
 657 }
 658
 659 static void display_sig(int sig __maybe_unused)
 660 {
 661         stop_top();
 662 }
 663
 664 static void display_setup_sig(void)
 665 {
 666         signal(SIGSEGV, sighandler_dump_stack);
 667         signal(SIGFPE, sighandler_dump_stack);
 668         signal(SIGINT,  display_sig);
 669         signal(SIGQUIT, display_sig);
 670         signal(SIGTERM, display_sig);
 671 }
 672
 673 static void *display_thread(void *arg)
 674 {
 675         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 676         struct termios save;
 677         struct perf_top *top = arg;
 678         int delay_msecs, c;
 679
 680         /* In order to read symbols from other namespaces perf to  needs to call
 681          * setns(2).  This isn't permitted if the struct_fs has multiple users.
 682          * unshare(2) the fs so that we may continue to setns into namespaces
 683          * that we're observing.
 684          */
 685         unshare(CLONE_FS);
 686
 687         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
 688
 689         display_setup_sig();
 690         pthread__unblock_sigwinch();
 691 repeat:
 692         delay_msecs = top->delay_secs * MSEC_PER_SEC;
 693         set_term_quiet_input(&save);
 694         /* trash return*/
 695         clearerr(stdin);
 696         if (poll(&stdin_poll, 1, 0) > 0)
 697                 getc(stdin);
 698
 699         while (!done) {
 700                 perf_top__print_sym_table(top);
 701                 /*
 702                  * Either timeout expired or we got an EINTR due to SIGWINCH,
 703                  * refresh screen in both cases.
 704                  */
 705                 switch (poll(&stdin_poll, 1, delay_msecs)) {
 706                 case 0:
 707                         continue;
 708                 case -1:
 709                         if (errno == EINTR)
 710                                 continue;
 711                         fallthrough;
 712                 default:
 713                         c = getc(stdin);
 714                         tcsetattr(0, TCSAFLUSH, &save);
 715
 716                         if (perf_top__handle_keypress(top, c))
 717                                 goto repeat;
 718                         stop_top();
 719                 }
 720         }
 721
 722         tcsetattr(0, TCSAFLUSH, &save);
 723         return NULL;
 724 }
 725
 726 static int hist_iter__top_callback(struct hist_entry_iter *iter,
 727                                    struct addr_location *al, bool single,
 728                                    void *arg)
 729         EXCLUSIVE_LOCKS_REQUIRED(iter->he->hists->lock)
 730 {
 731         struct perf_top *top = arg;
 732         struct evsel *evsel = iter->evsel;
 733
 734         if (perf_hpp_list.sym && single)
 735                 perf_top__record_precise_ip(top, iter->he, iter->sample, evsel, al->addr);
 736
 737         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
 738                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
 739                      NULL);
 740         return 0;
 741 }
 742
 743 static void perf_event__process_sample(struct perf_tool *tool,
 744                                        const union perf_event *event,
 745                                        struct evsel *evsel,
 746                                        struct perf_sample *sample,
 747                                        struct machine *machine)
 748 {
 749         struct perf_top *top = container_of(tool, struct perf_top, tool);
 750         struct addr_location al;
 751
 752         if (!machine && perf_guest) {
 753                 static struct intlist *seen;
 754
 755                 if (!seen)
 756                         seen = intlist__new(NULL);
 757
 758                 if (!intlist__has_entry(seen, sample->pid)) {
 759                         pr_err("Can't find guest [%d]'s kernel information\n",
 760                                 sample->pid);
 761                         intlist__add(seen, sample->pid);
 762                 }
 763                 return;
 764         }
 765
 766         if (!machine) {
 767                 pr_err("%u unprocessable samples recorded.\r",
 768                        top->session->evlist->stats.nr_unprocessable_samples++);
 769                 return;
 770         }
 771
 772         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
 773                 top->exact_samples++;
 774
 775         addr_location__init(&al);
 776         if (machine__resolve(machine, &al, sample) < 0)
 777                 goto out;
 778
 779         if (top->stitch_lbr)
 780                 thread__set_lbr_stitch_enable(al.thread, 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 out:
 852         addr_location__exit(&al);
 853 }
 854
 855 static void
 856 perf_top__process_lost(struct perf_top *top, union perf_event *event,
 857                        struct evsel *evsel)
 858 {
 859         top->lost += event->lost.lost;
 860         top->lost_total += event->lost.lost;
 861         evsel->evlist->stats.total_lost += event->lost.lost;
 862 }
 863
 864 static void
 865 perf_top__process_lost_samples(struct perf_top *top,
 866                                union perf_event *event,
 867                                struct evsel *evsel)
 868 {
 869         top->lost += event->lost_samples.lost;
 870         top->lost_total += event->lost_samples.lost;
 871         evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
 872 }
 873
 874 static u64 last_timestamp;
 875
 876 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 877 {
 878         struct record_opts *opts = &top->record_opts;
 879         struct evlist *evlist = top->evlist;
 880         struct mmap *md;
 881         union perf_event *event;
 882
 883         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 884         if (perf_mmap__read_init(&md->core) < 0)
 885                 return;
 886
 887         while ((event = perf_mmap__read_event(&md->core)) != NULL) {
 888                 int ret;
 889
 890                 ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
 891                 if (ret && ret != -1)
 892                         break;
 893
 894                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
 895                 if (ret)
 896                         break;
 897
 898                 perf_mmap__consume(&md->core);
 899
 900                 if (top->qe.rotate) {
 901                         mutex_lock(&top->qe.mutex);
 902                         top->qe.rotate = false;
 903                         cond_signal(&top->qe.cond);
 904                         mutex_unlock(&top->qe.mutex);
 905                 }
 906         }
 907
 908         perf_mmap__read_done(&md->core);
 909 }
 910
 911 static void perf_top__mmap_read(struct perf_top *top)
 912 {
 913         bool overwrite = top->record_opts.overwrite;
 914         struct evlist *evlist = top->evlist;
 915         int i;
 916
 917         if (overwrite)
 918                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
 919
 920         for (i = 0; i < top->evlist->core.nr_mmaps; i++)
 921                 perf_top__mmap_read_idx(top, i);
 922
 923         if (overwrite) {
 924                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 925                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 926         }
 927 }
 928
 929 /*
 930  * Check per-event overwrite term.
 931  * perf top should support consistent term for all events.
 932  * - All events don't have per-event term
 933  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 934  *   Nothing change, return 0.
 935  * - All events have same per-event term
 936  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 937  *   Using the per-event setting to replace the opts->overwrite if
 938  *   they are different, then return 0.
 939  * - Events have different per-event term
 940  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 941  *   Return -1
 942  * - Some of the event set per-event term, but some not.
 943  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 944  *   Return -1
 945  */
 946 static int perf_top__overwrite_check(struct perf_top *top)
 947 {
 948         struct record_opts *opts = &top->record_opts;
 949         struct evlist *evlist = top->evlist;
 950         struct evsel_config_term *term;
 951         struct list_head *config_terms;
 952         struct evsel *evsel;
 953         int set, overwrite = -1;
 954
 955         evlist__for_each_entry(evlist, evsel) {
 956                 set = -1;
 957                 config_terms = &evsel->config_terms;
 958                 list_for_each_entry(term, config_terms, list) {
 959                         if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
 960                                 set = term->val.overwrite ? 1 : 0;
 961                 }
 962
 963                 /* no term for current and previous event (likely) */
 964                 if ((overwrite < 0) && (set < 0))
 965                         continue;
 966
 967                 /* has term for both current and previous event, compare */
 968                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
 969                         return -1;
 970
 971                 /* no term for current event but has term for previous one */
 972                 if ((overwrite >= 0) && (set < 0))
 973                         return -1;
 974
 975                 /* has term for current event */
 976                 if ((overwrite < 0) && (set >= 0)) {
 977                         /* if it's first event, set overwrite */
 978                         if (evsel == evlist__first(evlist))
 979                                 overwrite = set;
 980                         else
 981                                 return -1;
 982                 }
 983         }
 984
 985         if ((overwrite >= 0) && (opts->overwrite != overwrite))
 986                 opts->overwrite = overwrite;
 987
 988         return 0;
 989 }
 990
 991 static int perf_top_overwrite_fallback(struct perf_top *top,
 992                                        struct evsel *evsel)
 993 {
 994         struct record_opts *opts = &top->record_opts;
 995         struct evlist *evlist = top->evlist;
 996         struct evsel *counter;
 997
 998         if (!opts->overwrite)
 999                 return 0;
1000
1001         /* only fall back when first event fails */
1002         if (evsel != evlist__first(evlist))
1003                 return 0;
1004
1005         evlist__for_each_entry(evlist, counter)
1006                 counter->core.attr.write_backward = false;
1007         opts->overwrite = false;
1008         pr_debug2("fall back to non-overwrite mode\n");
1009         return 1;
1010 }
1011
1012 static int perf_top__start_counters(struct perf_top *top)
1013 {
1014         char msg[BUFSIZ];
1015         struct evsel *counter;
1016         struct evlist *evlist = top->evlist;
1017         struct record_opts *opts = &top->record_opts;
1018
1019         if (perf_top__overwrite_check(top)) {
1020                 ui__error("perf top only support consistent per-event "
1021                           "overwrite setting for all events\n");
1022                 goto out_err;
1023         }
1024
1025         evlist__config(evlist, opts, &callchain_param);
1026
1027         evlist__for_each_entry(evlist, counter) {
1028 try_again:
1029                 if (evsel__open(counter, counter->core.cpus,
1030                                 counter->core.threads) < 0) {
1031
1032                         /*
1033                          * Specially handle overwrite fall back.
1034                          * Because perf top is the only tool which has
1035                          * overwrite mode by default, support
1036                          * both overwrite and non-overwrite mode, and
1037                          * require consistent mode for all events.
1038                          *
1039                          * May move it to generic code with more tools
1040                          * have similar attribute.
1041                          */
1042                         if (perf_missing_features.write_backward &&
1043                             perf_top_overwrite_fallback(top, counter))
1044                                 goto try_again;
1045
1046                         if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
1047                                 if (verbose > 0)
1048                                         ui__warning("%s\n", msg);
1049                                 goto try_again;
1050                         }
1051
1052                         evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1053                         ui__error("%s\n", msg);
1054                         goto out_err;
1055                 }
1056         }
1057
1058         if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1059                 ui__error("Failed to mmap with %d (%s)\n",
1060                             errno, str_error_r(errno, msg, sizeof(msg)));
1061                 goto out_err;
1062         }
1063
1064         return 0;
1065
1066 out_err:
1067         return -1;
1068 }
1069
1070 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1071 {
1072         if (callchain->mode != CHAIN_NONE) {
1073                 if (callchain_register_param(callchain) < 0) {
1074                         ui__error("Can't register callchain params.\n");
1075                         return -EINVAL;
1076                 }
1077         }
1078
1079         return 0;
1080 }
1081
1082 static struct ordered_events *rotate_queues(struct perf_top *top)
1083 {
1084         struct ordered_events *in = top->qe.in;
1085
1086         if (top->qe.in == &top->qe.data[1])
1087                 top->qe.in = &top->qe.data[0];
1088         else
1089                 top->qe.in = &top->qe.data[1];
1090
1091         return in;
1092 }
1093
1094 static void *process_thread(void *arg)
1095 {
1096         struct perf_top *top = arg;
1097
1098         while (!done) {
1099                 struct ordered_events *out, *in = top->qe.in;
1100
1101                 if (!in->nr_events) {
1102                         usleep(100);
1103                         continue;
1104                 }
1105
1106                 out = rotate_queues(top);
1107
1108                 mutex_lock(&top->qe.mutex);
1109                 top->qe.rotate = true;
1110                 cond_wait(&top->qe.cond, &top->qe.mutex);
1111                 mutex_unlock(&top->qe.mutex);
1112
1113                 if (ordered_events__flush(out, OE_FLUSH__TOP))
1114                         pr_err("failed to process events\n");
1115         }
1116
1117         return NULL;
1118 }
1119
1120 /*
1121  * Allow only 'top->delay_secs' seconds behind samples.
1122  */
1123 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1124 {
1125         union perf_event *event = qevent->event;
1126         u64 delay_timestamp;
1127
1128         if (event->header.type != PERF_RECORD_SAMPLE)
1129                 return false;
1130
1131         delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1132         return delay_timestamp < last_timestamp;
1133 }
1134
1135 static int deliver_event(struct ordered_events *qe,
1136                          struct ordered_event *qevent)
1137 {
1138         struct perf_top *top = qe->data;
1139         struct evlist *evlist = top->evlist;
1140         struct perf_session *session = top->session;
1141         union perf_event *event = qevent->event;
1142         struct perf_sample sample;
1143         struct evsel *evsel;
1144         struct machine *machine;
1145         int ret = -1;
1146
1147         if (should_drop(qevent, top)) {
1148                 top->drop++;
1149                 top->drop_total++;
1150                 return 0;
1151         }
1152
1153         ret = evlist__parse_sample(evlist, event, &sample);
1154         if (ret) {
1155                 pr_err("Can't parse sample, err = %d\n", ret);
1156                 goto next_event;
1157         }
1158
1159         evsel = evlist__id2evsel(session->evlist, sample.id);
1160         assert(evsel != NULL);
1161
1162         if (event->header.type == PERF_RECORD_SAMPLE) {
1163                 if (evswitch__discard(&top->evswitch, evsel))
1164                         return 0;
1165                 ++top->samples;
1166         }
1167
1168         switch (sample.cpumode) {
1169         case PERF_RECORD_MISC_USER:
1170                 ++top->us_samples;
1171                 if (top->hide_user_symbols)
1172                         goto next_event;
1173                 machine = &session->machines.host;
1174                 break;
1175         case PERF_RECORD_MISC_KERNEL:
1176                 ++top->kernel_samples;
1177                 if (top->hide_kernel_symbols)
1178                         goto next_event;
1179                 machine = &session->machines.host;
1180                 break;
1181         case PERF_RECORD_MISC_GUEST_KERNEL:
1182                 ++top->guest_kernel_samples;
1183                 machine = perf_session__find_machine(session,
1184                                                      sample.pid);
1185                 break;
1186         case PERF_RECORD_MISC_GUEST_USER:
1187                 ++top->guest_us_samples;
1188                 /*
1189                  * TODO: we don't process guest user from host side
1190                  * except simple counting.
1191                  */
1192                 goto next_event;
1193         default:
1194                 if (event->header.type == PERF_RECORD_SAMPLE)
1195                         goto next_event;
1196                 machine = &session->machines.host;
1197                 break;
1198         }
1199
1200         if (event->header.type == PERF_RECORD_SAMPLE) {
1201                 perf_event__process_sample(&top->tool, event, evsel,
1202                                            &sample, machine);
1203         } else if (event->header.type == PERF_RECORD_LOST) {
1204                 perf_top__process_lost(top, event, evsel);
1205         } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1206                 perf_top__process_lost_samples(top, event, evsel);
1207         } else if (event->header.type < PERF_RECORD_MAX) {
1208                 events_stats__inc(&session->evlist->stats, event->header.type);
1209                 machine__process_event(machine, event, &sample);
1210         } else
1211                 ++session->evlist->stats.nr_unknown_events;
1212
1213         ret = 0;
1214 next_event:
1215         return ret;
1216 }
1217
1218 static void init_process_thread(struct perf_top *top)
1219 {
1220         ordered_events__init(&top->qe.data[0], deliver_event, top);
1221         ordered_events__init(&top->qe.data[1], deliver_event, top);
1222         ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1223         ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1224         top->qe.in = &top->qe.data[0];
1225         mutex_init(&top->qe.mutex);
1226         cond_init(&top->qe.cond);
1227 }
1228
1229 static void exit_process_thread(struct perf_top *top)
1230 {
1231         ordered_events__free(&top->qe.data[0]);
1232         ordered_events__free(&top->qe.data[1]);
1233         mutex_destroy(&top->qe.mutex);
1234         cond_destroy(&top->qe.cond);
1235 }
1236
1237 static int __cmd_top(struct perf_top *top)
1238 {
1239         struct record_opts *opts = &top->record_opts;
1240         pthread_t thread, thread_process;
1241         int ret;
1242
1243         if (!annotate_opts.objdump_path) {
1244                 ret = perf_env__lookup_objdump(&top->session->header.env,
1245                                                &annotate_opts.objdump_path);
1246                 if (ret)
1247                         return ret;
1248         }
1249
1250         ret = callchain_param__setup_sample_type(&callchain_param);
1251         if (ret)
1252                 return ret;
1253
1254         if (perf_session__register_idle_thread(top->session) < 0)
1255                 return ret;
1256
1257         if (top->nr_threads_synthesize > 1)
1258                 perf_set_multithreaded();
1259
1260         init_process_thread(top);
1261
1262         if (opts->record_namespaces)
1263                 top->tool.namespace_events = true;
1264         if (opts->record_cgroup) {
1265 #ifdef HAVE_FILE_HANDLE
1266                 top->tool.cgroup_events = true;
1267 #else
1268                 pr_err("cgroup tracking is not supported.\n");
1269                 return -1;
1270 #endif
1271         }
1272
1273         ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1274                                                 &top->session->machines.host,
1275                                                 &top->record_opts);
1276         if (ret < 0)
1277                 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1278
1279         ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1280                                              &top->session->machines.host);
1281         if (ret < 0)
1282                 pr_debug("Couldn't synthesize cgroup events.\n");
1283
1284         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1285                                     top->evlist->core.threads, true, false,
1286                                     top->nr_threads_synthesize);
1287
1288         perf_set_multithreaded();
1289
1290         if (perf_hpp_list.socket) {
1291                 ret = perf_env__read_cpu_topology_map(&perf_env);
1292                 if (ret < 0) {
1293                         char errbuf[BUFSIZ];
1294                         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1295
1296                         ui__error("Could not read the CPU topology map: %s\n", err);
1297                         return ret;
1298                 }
1299         }
1300
1301         evlist__uniquify_name(top->evlist);
1302         ret = perf_top__start_counters(top);
1303         if (ret)
1304                 return ret;
1305
1306         top->session->evlist = top->evlist;
1307         perf_session__set_id_hdr_size(top->session);
1308
1309         /*
1310          * When perf is starting the traced process, all the events (apart from
1311          * group members) have enable_on_exec=1 set, so don't spoil it by
1312          * prematurely enabling them.
1313          *
1314          * XXX 'top' still doesn't start workloads like record, trace, but should,
1315          * so leave the check here.
1316          */
1317         if (!target__none(&opts->target))
1318                 evlist__enable(top->evlist);
1319
1320         ret = -1;
1321         if (pthread_create(&thread_process, NULL, process_thread, top)) {
1322                 ui__error("Could not create process thread.\n");
1323                 return ret;
1324         }
1325
1326         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1327                                                             display_thread), top)) {
1328                 ui__error("Could not create display thread.\n");
1329                 goto out_join_thread;
1330         }
1331
1332         if (top->realtime_prio) {
1333                 struct sched_param param;
1334
1335                 param.sched_priority = top->realtime_prio;
1336                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1337                         ui__error("Could not set realtime priority.\n");
1338                         goto out_join;
1339                 }
1340         }
1341
1342         /* Wait for a minimal set of events before starting the snapshot */
1343         evlist__poll(top->evlist, 100);
1344
1345         perf_top__mmap_read(top);
1346
1347         while (!done) {
1348                 u64 hits = top->samples;
1349
1350                 perf_top__mmap_read(top);
1351
1352                 if (opts->overwrite || (hits == top->samples))
1353                         ret = evlist__poll(top->evlist, 100);
1354
1355                 if (resize) {
1356                         perf_top__resize(top);
1357                         resize = 0;
1358                 }
1359         }
1360
1361         ret = 0;
1362 out_join:
1363         pthread_join(thread, NULL);
1364 out_join_thread:
1365         cond_signal(&top->qe.cond);
1366         pthread_join(thread_process, NULL);
1367         perf_set_singlethreaded();
1368         exit_process_thread(top);
1369         return ret;
1370 }
1371
1372 static int
1373 callchain_opt(const struct option *opt, const char *arg, int unset)
1374 {
1375         symbol_conf.use_callchain = true;
1376         return record_callchain_opt(opt, arg, unset);
1377 }
1378
1379 static int
1380 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1381 {
1382         struct callchain_param *callchain = opt->value;
1383
1384         callchain->enabled = !unset;
1385         callchain->record_mode = CALLCHAIN_FP;
1386
1387         /*
1388          * --no-call-graph
1389          */
1390         if (unset) {
1391                 symbol_conf.use_callchain = false;
1392                 callchain->record_mode = CALLCHAIN_NONE;
1393                 return 0;
1394         }
1395
1396         return parse_callchain_top_opt(arg);
1397 }
1398
1399 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1400 {
1401         if (!strcmp(var, "top.call-graph")) {
1402                 var = "call-graph.record-mode";
1403                 return perf_default_config(var, value, cb);
1404         }
1405         if (!strcmp(var, "top.children")) {
1406                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1407                 return 0;
1408         }
1409
1410         return 0;
1411 }
1412
1413 static int
1414 parse_percent_limit(const struct option *opt, const char *arg,
1415                     int unset __maybe_unused)
1416 {
1417         struct perf_top *top = opt->value;
1418
1419         top->min_percent = strtof(arg, NULL);
1420         return 0;
1421 }
1422
1423 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1424         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1425
1426 int cmd_top(int argc, const char **argv)
1427 {
1428         char errbuf[BUFSIZ];
1429         struct perf_top top = {
1430                 .count_filter        = 5,
1431                 .delay_secs          = 2,
1432                 .record_opts = {
1433                         .mmap_pages     = UINT_MAX,
1434                         .user_freq      = UINT_MAX,
1435                         .user_interval  = ULLONG_MAX,
1436                         .freq           = 4000, /* 4 KHz */
1437                         .target         = {
1438                                 .uses_mmap   = true,
1439                         },
1440                         /*
1441                          * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1442                          * when we pause, fix that and reenable. Probably using a
1443                          * separate evlist with a dummy event, i.e. a non-overwrite
1444                          * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1445                          * stays in overwrite mode. -acme
1446                          * */
1447                         .overwrite      = 0,
1448                         .sample_time    = true,
1449                         .sample_time_set = true,
1450                 },
1451                 .max_stack           = sysctl__max_stack(),
1452                 .nr_threads_synthesize = UINT_MAX,
1453         };
1454         struct parse_events_option_args parse_events_option_args = {
1455                 .evlistp = &top.evlist,
1456         };
1457         bool branch_call_mode = false;
1458         struct record_opts *opts = &top.record_opts;
1459         struct target *target = &opts->target;
1460         const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
1461         const struct option options[] = {
1462         OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1463                      "event selector. use 'perf list' to list available events",
1464                      parse_events_option),
1465         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1466         OPT_STRING('p', "pid", &target->pid, "pid",
1467                     "profile events on existing process id"),
1468         OPT_STRING('t', "tid", &target->tid, "tid",
1469                     "profile events on existing thread id"),
1470         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1471                             "system-wide collection from all CPUs"),
1472         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1473                     "list of cpus to monitor"),
1474         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1475                    "file", "vmlinux pathname"),
1476         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1477                     "don't load vmlinux even if found"),
1478         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1479                    "file", "kallsyms pathname"),
1480         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1481                     "hide kernel symbols"),
1482         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1483                      "number of mmap data pages", evlist__parse_mmap_pages),
1484         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1485                     "collect data with this RT SCHED_FIFO priority"),
1486         OPT_INTEGER('d', "delay", &top.delay_secs,
1487                     "number of seconds to delay between refreshes"),
1488         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1489                             "dump the symbol table used for profiling"),
1490         OPT_INTEGER('f', "count-filter", &top.count_filter,
1491                     "only display functions with more events than this"),
1492         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1493                     "child tasks do not inherit counters"),
1494         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1495                     "symbol to annotate"),
1496         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1497         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1498                      "profile at this frequency",
1499                       record__parse_freq),
1500         OPT_INTEGER('E', "entries", &top.print_entries,
1501                     "display this many functions"),
1502         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1503                     "hide user symbols"),
1504 #ifdef HAVE_SLANG_SUPPORT
1505         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1506 #endif
1507         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1508         OPT_INCR('v', "verbose", &verbose,
1509                     "be more verbose (show counter open errors, etc)"),
1510         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1511                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1512                    " Please refer the man page for the complete list."),
1513         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1514                    "output field(s): overhead, period, sample plus all of sort keys"),
1515         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1516                     "Show a column with the number of samples"),
1517         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1518                            NULL, "enables call-graph recording and display",
1519                            &callchain_opt),
1520         OPT_CALLBACK(0, "call-graph", &callchain_param,
1521                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1522                      top_callchain_help, &parse_callchain_opt),
1523         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1524                     "Accumulate callchains of children and show total overhead as well"),
1525         OPT_INTEGER(0, "max-stack", &top.max_stack,
1526                     "Set the maximum stack depth when parsing the callchain. "
1527                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1528         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1529                    "ignore callees of these functions in call graphs",
1530                    report_parse_ignore_callees_opt),
1531         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1532                     "Show a column with the sum of periods"),
1533         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1534                    "only consider symbols in these dsos"),
1535         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1536                    "only consider symbols in these comms"),
1537         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1538                    "only consider these symbols"),
1539         OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
1540                     "Interleave source code with assembly code (default)"),
1541         OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
1542                     "Display raw encoding of assembly instructions (default)"),
1543         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1544                     "Enable kernel symbol demangling"),
1545         OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1546         OPT_STRING(0, "objdump", &objdump_path, "path",
1547                     "objdump binary to use for disassembly and annotations"),
1548         OPT_STRING(0, "addr2line", &addr2line_path, "path",
1549                    "addr2line binary to use for line numbers"),
1550         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1551                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1552         OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
1553                     "Add prefix to source file path names in programs (with --prefix-strip)"),
1554         OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
1555                     "Strip first N entries of source file path name in programs (with --prefix)"),
1556         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1557         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1558                      "Don't show entries under that percent", parse_percent_limit),
1559         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1560                      "How to display percentage of filtered entries", parse_filter_percentage),
1561         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1562                    "width[,width...]",
1563                    "don't try to adjust column width, use these fixed values"),
1564         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1565                         "per thread proc mmap processing timeout in ms"),
1566         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1567                      "branch any", "sample any taken branches",
1568                      parse_branch_stack),
1569         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1570                      "branch filter mask", "branch stack filter modes",
1571                      parse_branch_stack),
1572         OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1573                     "add last branch records to call history"),
1574         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1575                     "Show raw trace event output (do not use print fmt or plugins)"),
1576         OPT_BOOLEAN('H', "hierarchy", &symbol_conf.report_hierarchy,
1577                     "Show entries in a hierarchy"),
1578         OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1579                     "Use a backward ring buffer, default: no"),
1580         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1581         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1582                         "number of thread to run event synthesize"),
1583         OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
1584                      "monitor event in cgroup name only", parse_cgroups),
1585         OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1586                     "Record namespaces events"),
1587         OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1588                     "Record cgroup events"),
1589         OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1590                     "Sort the output by the event at the index n in group. "
1591                     "If n is invalid, sort by the first event. "
1592                     "WARNING: should be used on grouped events."),
1593         OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1594                     "Enable LBR callgraph stitching approach"),
1595 #ifdef HAVE_LIBPFM
1596         OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1597                 "libpfm4 event selector. use 'perf list' to list available events",
1598                 parse_libpfm_events_option),
1599 #endif
1600         OPTS_EVSWITCH(&top.evswitch),
1601         OPT_END()
1602         };
1603         const char * const top_usage[] = {
1604                 "perf top [<options>]",
1605                 NULL
1606         };
1607         int status = hists__init();
1608
1609         if (status < 0)
1610                 return status;
1611
1612         annotation_options__init();
1613
1614         annotate_opts.min_pcnt = 5;
1615         annotate_opts.context  = 4;
1616
1617         top.evlist = evlist__new();
1618         if (top.evlist == NULL)
1619                 return -ENOMEM;
1620
1621         status = perf_config(perf_top_config, &top);
1622         if (status)
1623                 return status;
1624         /*
1625          * Since the per arch annotation init routine may need the cpuid, read
1626          * it here, since we are not getting this from the perf.data header.
1627          */
1628         status = perf_env__read_cpuid(&perf_env);
1629         if (status) {
1630                 /*
1631                  * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1632                  * warn the user explicitly.
1633                  */
1634                 eprintf(status == ENOSYS ? 1 : 0, verbose,
1635                         "Couldn't read the cpuid for this machine: %s\n",
1636                         str_error_r(errno, errbuf, sizeof(errbuf)));
1637         }
1638         top.evlist->env = &perf_env;
1639
1640         argc = parse_options(argc, argv, options, top_usage, 0);
1641         if (argc)
1642                 usage_with_options(top_usage, options);
1643
1644         if (disassembler_style) {
1645                 annotate_opts.disassembler_style = strdup(disassembler_style);
1646                 if (!annotate_opts.disassembler_style)
1647                         return -ENOMEM;
1648         }
1649         if (objdump_path) {
1650                 annotate_opts.objdump_path = strdup(objdump_path);
1651                 if (!annotate_opts.objdump_path)
1652                         return -ENOMEM;
1653         }
1654         if (addr2line_path) {
1655                 symbol_conf.addr2line_path = strdup(addr2line_path);
1656                 if (!symbol_conf.addr2line_path)
1657                         return -ENOMEM;
1658         }
1659
1660         status = symbol__validate_sym_arguments();
1661         if (status)
1662                 goto out_delete_evlist;
1663
1664         if (annotate_check_args() < 0)
1665                 goto out_delete_evlist;
1666
1667         if (!top.evlist->core.nr_entries) {
1668                 bool can_profile_kernel = perf_event_paranoid_check(1);
1669                 int err = parse_event(top.evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");
1670
1671                 if (err)
1672                         goto out_delete_evlist;
1673         }
1674
1675         status = evswitch__init(&top.evswitch, top.evlist, stderr);
1676         if (status)
1677                 goto out_delete_evlist;
1678
1679         if (symbol_conf.report_hierarchy) {
1680                 /* disable incompatible options */
1681                 symbol_conf.event_group = false;
1682                 symbol_conf.cumulate_callchain = false;
1683
1684                 if (field_order) {
1685                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1686                         parse_options_usage(top_usage, options, "fields", 0);
1687                         parse_options_usage(NULL, options, "hierarchy", 0);
1688                         goto out_delete_evlist;
1689                 }
1690         }
1691
1692         if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1693                 pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1694                 goto out_delete_evlist;
1695         }
1696
1697         if (nr_cgroups > 0 && opts->record_cgroup) {
1698                 pr_err("--cgroup and --all-cgroups cannot be used together\n");
1699                 goto out_delete_evlist;
1700         }
1701
1702         if (branch_call_mode) {
1703                 if (!opts->branch_stack)
1704                         opts->branch_stack = PERF_SAMPLE_BRANCH_ANY;
1705                 symbol_conf.use_callchain = true;
1706                 callchain_param.key = CCKEY_ADDRESS;
1707                 callchain_param.branch_callstack = true;
1708                 callchain_param.enabled = true;
1709                 if (callchain_param.record_mode == CALLCHAIN_NONE)
1710                         callchain_param.record_mode = CALLCHAIN_FP;
1711                 callchain_register_param(&callchain_param);
1712                 if (!sort_order)
1713                         sort_order = "srcline,symbol,dso";
1714         }
1715
1716         if (opts->branch_stack && callchain_param.enabled)
1717                 symbol_conf.show_branchflag_count = true;
1718
1719         sort__mode = SORT_MODE__TOP;
1720         /* display thread wants entries to be collapsed in a different tree */
1721         perf_hpp_list.need_collapse = 1;
1722
1723         if (top.use_stdio)
1724                 use_browser = 0;
1725 #ifdef HAVE_SLANG_SUPPORT
1726         else if (top.use_tui)
1727                 use_browser = 1;
1728 #endif
1729
1730         setup_browser(false);
1731
1732         if (setup_sorting(top.evlist) < 0) {
1733                 if (sort_order)
1734                         parse_options_usage(top_usage, options, "s", 1);
1735                 if (field_order)
1736                         parse_options_usage(sort_order ? NULL : top_usage,
1737                                             options, "fields", 0);
1738                 goto out_delete_evlist;
1739         }
1740
1741         status = target__validate(target);
1742         if (status) {
1743                 target__strerror(target, status, errbuf, BUFSIZ);
1744                 ui__warning("%s\n", errbuf);
1745         }
1746
1747         status = target__parse_uid(target);
1748         if (status) {
1749                 int saved_errno = errno;
1750
1751                 target__strerror(target, status, errbuf, BUFSIZ);
1752                 ui__error("%s\n", errbuf);
1753
1754                 status = -saved_errno;
1755                 goto out_delete_evlist;
1756         }
1757
1758         if (target__none(target))
1759                 target->system_wide = true;
1760
1761         if (evlist__create_maps(top.evlist, target) < 0) {
1762                 ui__error("Couldn't create thread/CPU maps: %s\n",
1763                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1764                 status = -errno;
1765                 goto out_delete_evlist;
1766         }
1767
1768         if (top.delay_secs < 1)
1769                 top.delay_secs = 1;
1770
1771         if (record_opts__config(opts)) {
1772                 status = -EINVAL;
1773                 goto out_delete_evlist;
1774         }
1775
1776         top.sym_evsel = evlist__first(top.evlist);
1777
1778         if (!callchain_param.enabled) {
1779                 symbol_conf.cumulate_callchain = false;
1780                 perf_hpp__cancel_cumulate();
1781         }
1782
1783         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1784                 callchain_param.order = ORDER_CALLER;
1785
1786         status = symbol__annotation_init();
1787         if (status < 0)
1788                 goto out_delete_evlist;
1789
1790         annotation_config__init();
1791
1792         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1793         status = symbol__init(NULL);
1794         if (status < 0)
1795                 goto out_delete_evlist;
1796
1797         sort__setup_elide(stdout);
1798
1799         get_term_dimensions(&top.winsize);
1800         if (top.print_entries == 0) {
1801                 perf_top__update_print_entries(&top);
1802                 signal(SIGWINCH, winch_sig);
1803         }
1804
1805         top.session = perf_session__new(NULL, NULL);
1806         if (IS_ERR(top.session)) {
1807                 status = PTR_ERR(top.session);
1808                 top.session = NULL;
1809                 goto out_delete_evlist;
1810         }
1811
1812 #ifdef HAVE_LIBBPF_SUPPORT
1813         if (!top.record_opts.no_bpf_event) {
1814                 top.sb_evlist = evlist__new();
1815
1816                 if (top.sb_evlist == NULL) {
1817                         pr_err("Couldn't create side band evlist.\n.");
1818                         status = -EINVAL;
1819                         goto out_delete_evlist;
1820                 }
1821
1822                 if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1823                         pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1824                         status = -EINVAL;
1825                         goto out_delete_evlist;
1826                 }
1827         }
1828 #endif
1829
1830         if (evlist__start_sb_thread(top.sb_evlist, target)) {
1831                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1832                 opts->no_bpf_event = true;
1833         }
1834
1835         status = __cmd_top(&top);
1836
1837         if (!opts->no_bpf_event)
1838                 evlist__stop_sb_thread(top.sb_evlist);
1839
1840 out_delete_evlist:
1841         evlist__delete(top.evlist);
1842         perf_session__delete(top.session);
1843         annotation_options__exit();
1844
1845         return status;
1846 }