Quelle  builtin-top.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0-only
/*
 * builtin-top.c
 *
 * Builtin top command: Display a continuously updated profile of
 * any workload, CPU or specific PID.
 *
 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
 *  2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
 *
 * Improvements and fixes by:
 *
 *   Arjan van de Ven <arjan@linux.intel.com>
 *   Yanmin Zhang <yanmin.zhang@intel.com>
 *   Wu Fengguang <fengguang.wu@intel.com>
 *   Mike Galbraith <efault@gmx.de>
 *   Paul Mackerras <paulus@samba.org>
 */
#include "builtin.h"

#include "perf.h"

#include "util/annotate.h"
#include "util/bpf-event.h"
#include "util/cgroup.h"
#include "util/config.h"
#include "util/color.h"
#include "util/dso.h"
#include "util/evlist.h"
#include "util/evsel.h"
#include "util/evsel_config.h"
#include "util/event.h"
#include "util/machine.h"
#include "util/map.h"
#include "util/mmap.h"
#include "util/session.h"
#include "util/thread.h"
#include "util/stat.h"
#include "util/symbol.h"
#include "util/synthetic-events.h"
#include "util/top.h"
#include "util/util.h"
#include <linux/rbtree.h>
#include <subcmd/parse-options.h>
#include "util/parse-events.h"
#include "util/callchain.h"
#include "util/cpumap.h"
#include "util/sort.h"
#include "util/string2.h"
#include "util/term.h"
#include "util/intlist.h"
#include "util/parse-branch-options.h"
#include "arch/common.h"
#include "ui/ui.h"

#include "util/debug.h"
#include "util/ordered-events.h"
#include "util/pfm.h"

#include <assert.h>
#include <elf.h>
#include <fcntl.h>

#include <stdio.h>
#include <termios.h>
#include <unistd.h>
#include <inttypes.h>

#include <errno.h>
#include <time.h>
#include <sched.h>
#include <signal.h>

#include <sys/syscall.h>
#include <sys/ioctl.h>
#include <poll.h>
#include <sys/prctl.h>
#include <sys/wait.h>
#include <sys/uio.h>
#include <sys/utsname.h>
#include <sys/mman.h>

#include <linux/stringify.h>
#include <linux/time64.h>
#include <linux/types.h>
#include <linux/err.h>

#include <linux/ctype.h>
#include <perf/mmap.h>

static volatile sig_atomic_t done;
static volatile sig_atomic_t resize;

#define HEADER_LINE_NR  5

static void perf_top__update_print_entries(struct perf_top *top)
{
 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
}

static void winch_sig(int sig __maybe_unused)
{
 resize = 1;
}

static void perf_top__resize(struct perf_top *top)
{
 get_term_dimensions(&top->winsize);
 perf_top__update_print_entries(top);
}

static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
{
 struct evsel *evsel;
 struct symbol *sym;
 struct annotation *notes;
 struct map *map;
 struct dso *dso;
 int err = -1;

 if (!he || !he->ms.sym)
  return -1;

 evsel = hists_to_evsel(he->hists);

 sym = he->ms.sym;
 map = he->ms.map;
 dso = map__dso(map);

 /*
 * We can't annotate with just /proc/kallsyms
 */
 if (dso__symtab_type(dso) == DSO_BINARY_TYPE__KALLSYMS && !dso__is_kcore(dso)) {
  pr_err("Can't annotate %s: No vmlinux file was found in the "
         "path\n", sym->name);
  sleep(1);
  return -1;
 }

 notes = symbol__annotation(sym);
 annotation__lock(notes);

 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
  annotation__unlock(notes);
  pr_err("Not enough memory for annotating '%s' symbol!\n",
         sym->name);
  sleep(1);
  return err;
 }

 err = symbol__annotate(&he->ms, evsel, NULL);
 if (err == 0) {
  top->sym_filter_entry = he;
 } else {
  char msg[BUFSIZ];
  symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
  pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
 }

 annotation__unlock(notes);
 return err;
}

static void __zero_source_counters(struct hist_entry *he)
{
 struct symbol *sym = he->ms.sym;
 symbol__annotate_zero_histograms(sym);
}

static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
{
 struct utsname uts;
 int err = uname(&uts);
 struct dso *dso = map__dso(map);

 ui__warning("Out of bounds address found:\n\n"
      "Addr: %" PRIx64 "\n"
      "DSO: %s %c\n"
      "Map: %" PRIx64 "-%" PRIx64 "\n"
      "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
      "Arch: %s\n"
      "Kernel: %s\n"
      "Tools: %s\n\n"
      "Not all samples will be on the annotation output.\n\n"
      "Please report to linux-kernel@vger.kernel.org\n",
      ip, dso__long_name(dso), dso__symtab_origin(dso),
      map__start(map), map__end(map), sym->start, sym->end,
      sym->binding == STB_GLOBAL ? 'g' :
      sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
      err ? "[unknown]" : uts.machine,
      err ? "[unknown]" : uts.release, perf_version_string);
 if (use_browser <= 0)
  sleep(5);

 map__set_erange_warned(map);
}

static void perf_top__record_precise_ip(struct perf_top *top,
     struct hist_entry *he,
     struct perf_sample *sample,
     struct evsel *evsel, u64 ip)
 EXCLUSIVE_LOCKS_REQUIRED(he->hists->lock)
{
 struct annotation *notes;
 struct symbol *sym = he->ms.sym;
 int err = 0;

 if (sym == NULL || (use_browser == 0 &&
       (top->sym_filter_entry == NULL ||
        top->sym_filter_entry->ms.sym != sym)))
  return;

 notes = symbol__annotation(sym);

 if (!annotation__trylock(notes))
  return;

 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);

 annotation__unlock(notes);

 if (unlikely(err)) {
  /*
 * This function is now called with he->hists->lock held.
 * Release it before going to sleep.
 */
  mutex_unlock(&he->hists->lock);

  if (err == -ERANGE && !map__erange_warned(he->ms.map))
   ui__warn_map_erange(he->ms.map, sym, ip);
  else if (err == -ENOMEM) {
   pr_err("Not enough memory for annotating '%s' symbol!\n",
          sym->name);
   sleep(1);
  }

  mutex_lock(&he->hists->lock);
 }
}

static void perf_top__show_details(struct perf_top *top)
{
 struct hist_entry *he = top->sym_filter_entry;
 struct evsel *evsel;
 struct annotation *notes;
 struct symbol *symbol;
 int more;

 if (!he)
  return;

 evsel = hists_to_evsel(he->hists);

 symbol = he->ms.sym;
 notes = symbol__annotation(symbol);

 annotation__lock(notes);

 symbol__calc_percent(symbol, evsel);

 if (notes->src == NULL)
  goto out_unlock;

 printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
 printf(" Events Pcnt (>=%d%%)\n", annotate_opts.min_pcnt);

 more = hist_entry__annotate_printf(he, top->sym_evsel);

 if (top->evlist->enabled) {
  if (top->zero)
   symbol__annotate_zero_histogram(symbol, top->sym_evsel);
  else
   symbol__annotate_decay_histogram(symbol, top->sym_evsel);
 }
 if (more != 0)
  printf("%d lines not displayed, maybe increase display entries [e]\n", more);
out_unlock:
 annotation__unlock(notes);
}

static void perf_top__resort_hists(struct perf_top *t)
{
 struct evlist *evlist = t->evlist;
 struct evsel *pos;

 evlist__for_each_entry(evlist, pos) {
  struct hists *hists = evsel__hists(pos);

  /*
 * unlink existing entries so that they can be linked
 * in a correct order in hists__match() below.
 */
  hists__unlink(hists);

  if (evlist->enabled) {
   if (t->zero) {
    hists__delete_entries(hists);
   } else {
    hists__decay_entries(hists, t->hide_user_symbols,
           t->hide_kernel_symbols);
   }
  }

  hists__collapse_resort(hists, NULL);

  /* Non-group events are considered as leader */
  if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
   struct hists *leader_hists = evsel__hists(evsel__leader(pos));

   hists__match(leader_hists, hists);
   hists__link(leader_hists, hists);
  }
 }

 evlist__for_each_entry(evlist, pos) {
  evsel__output_resort(pos, NULL);
 }
}

static void perf_top__print_sym_table(struct perf_top *top)
{
 char bf[160];
 int printed = 0;
 const int win_width = top->winsize.ws_col - 1;
 struct evsel *evsel = top->sym_evsel;
 struct hists *hists = evsel__hists(evsel);

 puts(CONSOLE_CLEAR);

 perf_top__header_snprintf(top, bf, sizeof(bf));
 printf("%s\n", bf);

 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);

 if (!top->record_opts.overwrite &&
     (top->evlist->stats.nr_lost_warned !=
      top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
  top->evlist->stats.nr_lost_warned =
         top->evlist->stats.nr_events[PERF_RECORD_LOST];
  color_fprintf(stdout, PERF_COLOR_RED,
         "WARNING: LOST %d chunks, Check IO/CPU overload",
         top->evlist->stats.nr_lost_warned);
  ++printed;
 }

 if (top->sym_filter_entry) {
  perf_top__show_details(top);
  return;
 }

 perf_top__resort_hists(top);

 hists__output_recalc_col_len(hists, top->print_entries - printed);
 putchar('\n');
 hists__fprintf(hists, false, top->print_entries - printed, win_width,
         top->min_percent, stdout, !symbol_conf.use_callchain);
}

static void prompt_integer(int *target, const char *msg)
{
 char *buf = NULL, *p;
 size_t dummy = 0;
 int tmp;

 fprintf(stdout, "\n%s: ", msg);
 if (getline(&buf, &dummy, stdin) < 0)
  return;

 p = strchr(buf, '\n');
 if (p)
  *p = 0;

 p = buf;
 while(*p) {
  if (!isdigit(*p))
   goto out_free;
  p++;
 }
 tmp = strtoul(buf, NULL, 10);
 *target = tmp;
out_free:
 free(buf);
}

static void prompt_percent(int *target, const char *msg)
{
 int tmp = 0;

 prompt_integer(&tmp, msg);
 if (tmp >= 0 && tmp <= 100)
  *target = tmp;
}

static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
{
 char *buf = NULL, *p;
 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
 struct hists *hists = evsel__hists(top->sym_evsel);
 struct rb_node *next;
 size_t dummy = 0;

 /* zero counters of active symbol */
 if (syme) {
  __zero_source_counters(syme);
  top->sym_filter_entry = NULL;
 }

 fprintf(stdout, "\n%s: ", msg);
 if (getline(&buf, &dummy, stdin) < 0)
  goto out_free;

 p = strchr(buf, '\n');
 if (p)
  *p = 0;

 next = rb_first_cached(&hists->entries);
 while (next) {
  n = rb_entry(next, struct hist_entry, rb_node);
  if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
   found = n;
   break;
  }
  next = rb_next(&n->rb_node);
 }

 if (!found) {
  fprintf(stderr, "Sorry, %s is not active.\n", buf);
  sleep(1);
 } else
  perf_top__parse_source(top, found);

out_free:
 free(buf);
}

static void perf_top__print_mapped_keys(struct perf_top *top)
{
 char *name = NULL;

 if (top->sym_filter_entry) {
  struct symbol *sym = top->sym_filter_entry->ms.sym;
  name = sym->name;
 }

 fprintf(stdout, "\nMapped keys:\n");
 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);

 if (top->evlist->core.nr_entries > 1)
  fprintf(stdout, "\t[E] active event counter. \t(%s)\n", evsel__name(top->sym_evsel));

 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);

 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", annotate_opts.min_pcnt);
 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
 fprintf(stdout, "\t[S] stop annotation.\n");

 fprintf(stdout,
  "\t[K] hide kernel symbols. \t(%s)\n",
  top->hide_kernel_symbols ? "yes" : "no");
 fprintf(stdout,
  "\t[U] hide user symbols. \t(%s)\n",
  top->hide_user_symbols ? "yes" : "no");
 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
 fprintf(stdout, "\t[qQ] quit.\n");
}

static int perf_top__key_mapped(struct perf_top *top, int c)
{
 switch (c) {
  case 'd':
  case 'e':
  case 'f':
  case 'z':
  case 'q':
  case 'Q':
  case 'K':
  case 'U':
  case 'F':
  case 's':
  case 'S':
   return 1;
  case 'E':
   return top->evlist->core.nr_entries > 1 ? 1 : 0;
  default:
   break;
 }

 return 0;
}

static bool perf_top__handle_keypress(struct perf_top *top, int c)
{
 bool ret = true;

 if (!perf_top__key_mapped(top, c)) {
  struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
  struct termios save;

  perf_top__print_mapped_keys(top);
  fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
  fflush(stdout);

  set_term_quiet_input(&save);

  poll(&stdin_poll, 1, -1);
  c = getc(stdin);

  tcsetattr(0, TCSAFLUSH, &save);
  if (!perf_top__key_mapped(top, c))
   return ret;
 }

 switch (c) {
  case 'd':
   prompt_integer(&top->delay_secs, "Enter display delay");
   if (top->delay_secs < 1)
    top->delay_secs = 1;
   break;
  case 'e':
   prompt_integer(&top->print_entries, "Enter display entries (lines)");
   if (top->print_entries == 0) {
    perf_top__resize(top);
    signal(SIGWINCH, winch_sig);
   } else {
    signal(SIGWINCH, SIG_DFL);
   }
   break;
  case 'E':
   if (top->evlist->core.nr_entries > 1) {
    /* Select 0 as the default event: */
    int counter = 0;

    fprintf(stderr, "\nAvailable events:");

    evlist__for_each_entry(top->evlist, top->sym_evsel)
     fprintf(stderr, "\n\t%d %s", top->sym_evsel->core.idx, evsel__name(top->sym_evsel));

    prompt_integer(&counter, "Enter details event counter");

    if (counter >= top->evlist->core.nr_entries) {
     top->sym_evsel = evlist__first(top->evlist);
     fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
     sleep(1);
     break;
    }
    evlist__for_each_entry(top->evlist, top->sym_evsel)
     if (top->sym_evsel->core.idx == counter)
      break;
   } else
    top->sym_evsel = evlist__first(top->evlist);
   break;
  case 'f':
   prompt_integer(&top->count_filter, "Enter display event count filter");
   break;
  case 'F':
   prompt_percent(&annotate_opts.min_pcnt,
           "Enter details display event filter (percent)");
   break;
  case 'K':
   top->hide_kernel_symbols = !top->hide_kernel_symbols;
   break;
  case 'q':
  case 'Q':
   printf("exiting.\n");
   if (top->dump_symtab)
    perf_session__fprintf_dsos(top->session, stderr);
   ret = false;
   break;
  case 's':
   perf_top__prompt_symbol(top, "Enter details symbol");
   break;
  case 'S':
   if (!top->sym_filter_entry)
    break;
   else {
    struct hist_entry *syme = top->sym_filter_entry;

    top->sym_filter_entry = NULL;
    __zero_source_counters(syme);
   }
   break;
  case 'U':
   top->hide_user_symbols = !top->hide_user_symbols;
   break;
  case 'z':
   top->zero = !top->zero;
   break;
  default:
   break;
 }

 return ret;
}

static void perf_top__sort_new_samples(void *arg)
{
 struct perf_top *t = arg;

 if (t->evlist->selected != NULL)
  t->sym_evsel = t->evlist->selected;

 perf_top__resort_hists(t);

 if (t->lost || t->drop)
  pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
}

static void stop_top(void)
{
 session_done = 1;
 done = 1;
}

static void *display_thread_tui(void *arg)
{
 struct evsel *pos;
 struct perf_top *top = arg;
 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
 struct hist_browser_timer hbt = {
  .timer  = perf_top__sort_new_samples,
  .arg  = top,
  .refresh = top->delay_secs,
 };
 int ret;

 /* In order to read symbols from other namespaces perf to  needs to call
 * setns(2).  This isn't permitted if the struct_fs has multiple users.
 * unshare(2) the fs so that we may continue to setns into namespaces
 * that we're observing.
 */
 unshare(CLONE_FS);

 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);

repeat:
 perf_top__sort_new_samples(top);

 /*
 * Initialize the uid_filter_str, in the future the TUI will allow
 * Zooming in/out UIDs. For now just use whatever the user passed
 * via --uid.
 */
 evlist__for_each_entry(top->evlist, pos) {
  struct hists *hists = evsel__hists(pos);
  hists->uid_filter_str = top->uid_str;
 }

 ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
           perf_session__env(top->session),
           !top->record_opts.overwrite);
 if (ret == K_RELOAD) {
  top->zero = true;
  goto repeat;
 } else
  stop_top();

 return NULL;
}

static void display_sig(int sig __maybe_unused)
{
 stop_top();
}

static void display_setup_sig(void)
{
 signal(SIGSEGV, sighandler_dump_stack);
 signal(SIGFPE, sighandler_dump_stack);
 signal(SIGINT,  display_sig);
 signal(SIGQUIT, display_sig);
 signal(SIGTERM, display_sig);
}

static void *display_thread(void *arg)
{
 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
 struct termios save;
 struct perf_top *top = arg;
 int delay_msecs, c;

 /* In order to read symbols from other namespaces perf to  needs to call
 * setns(2).  This isn't permitted if the struct_fs has multiple users.
 * unshare(2) the fs so that we may continue to setns into namespaces
 * that we're observing.
 */
 unshare(CLONE_FS);

 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);

 display_setup_sig();
 pthread__unblock_sigwinch();
repeat:
 delay_msecs = top->delay_secs * MSEC_PER_SEC;
 set_term_quiet_input(&save);
 /* trash return*/
 clearerr(stdin);
 if (poll(&stdin_poll, 1, 0) > 0)
  getc(stdin);

 while (!done) {
  perf_top__print_sym_table(top);
  /*
 * Either timeout expired or we got an EINTR due to SIGWINCH,
 * refresh screen in both cases.
 */
  switch (poll(&stdin_poll, 1, delay_msecs)) {
  case 0:
   continue;
  case -1:
   if (errno == EINTR)
    continue;
   fallthrough;
  default:
   c = getc(stdin);
   tcsetattr(0, TCSAFLUSH, &save);

   if (perf_top__handle_keypress(top, c))
    goto repeat;
   stop_top();
  }
 }

 tcsetattr(0, TCSAFLUSH, &save);
 return NULL;
}

static int hist_iter__top_callback(struct hist_entry_iter *iter,
       struct addr_location *al, bool single,
       void *arg)
 EXCLUSIVE_LOCKS_REQUIRED(iter->he->hists->lock)
{
 struct perf_top *top = arg;
 struct evsel *evsel = iter->evsel;

 if (perf_hpp_list.sym && single)
  perf_top__record_precise_ip(top, iter->he, iter->sample, evsel, al->addr);

 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
        !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
        NULL, evsel);
 return 0;
}

static void perf_event__process_sample(const struct perf_tool *tool,
           const union perf_event *event,
           struct evsel *evsel,
           struct perf_sample *sample,
           struct machine *machine)
{
 struct perf_top *top = container_of(tool, struct perf_top, tool);
 struct addr_location al;

 if (!machine && perf_guest) {
  static struct intlist *seen;

  if (!seen)
   seen = intlist__new(NULL);

  if (!intlist__has_entry(seen, sample->pid)) {
   pr_err("Can't find guest [%d]'s kernel information\n",
    sample->pid);
   intlist__add(seen, sample->pid);
  }
  return;
 }

 if (!machine) {
  pr_err("%u unprocessable samples recorded.\r",
         top->session->evlist->stats.nr_unprocessable_samples++);
  return;
 }

 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
  top->exact_samples++;

 addr_location__init(&al);
 if (machine__resolve(machine, &al, sample) < 0)
  goto out;

 if (top->stitch_lbr)
  thread__set_lbr_stitch_enable(al.thread, true);

 if (!machine->kptr_restrict_warned &&
     symbol_conf.kptr_restrict &&
     al.cpumode == PERF_RECORD_MISC_KERNEL) {
  if (!evlist__exclude_kernel(top->session->evlist)) {
   ui__warning(
"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
"Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
"Kernel%s samples will not be resolved.\n",
     al.map && map__has_symbols(al.map) ?
     " modules" : "");
   if (use_browser <= 0)
    sleep(5);
  }
  machine->kptr_restrict_warned = true;
 }

 if (al.sym == NULL && al.map != NULL) {
  const char *msg = "Kernel samples will not be resolved.\n";
  /*
 * As we do lazy loading of symtabs we only will know if the
 * specified vmlinux file is invalid when we actually have a
 * hit in kernel space and then try to load it. So if we get
 * here and there are _no_ symbols in the DSO backing the
 * kernel map, bail out.
 *
 * We may never get here, for instance, if we use -K/
 * --hide-kernel-symbols, even if the user specifies an
 * invalid --vmlinux ;-)
 */
  if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
      __map__is_kernel(al.map) && !map__has_symbols(al.map)) {
   if (symbol_conf.vmlinux_name) {
    char serr[256];

    dso__strerror_load(map__dso(al.map), serr, sizeof(serr));
    ui__warning("The %s file can't be used: %s\n%s",
         symbol_conf.vmlinux_name, serr, msg);
   } else {
    ui__warning("A vmlinux file was not found.\n%s",
         msg);
   }

   if (use_browser <= 0)
    sleep(5);
   top->vmlinux_warned = true;
  }
 }

 if (al.sym == NULL || !al.sym->idle) {
  struct hists *hists = evsel__hists(evsel);
  struct hist_entry_iter iter = {
   .evsel  = evsel,
   .sample  = sample,
   .add_entry_cb  = hist_iter__top_callback,
  };

  if (symbol_conf.cumulate_callchain)
   iter.ops = &hist_iter_cumulative;
  else
   iter.ops = &hist_iter_normal;

  mutex_lock(&hists->lock);

  if (hist_entry_iter__add(&iter, &al, top->max_stack, top) < 0)
   pr_err("Problem incrementing symbol period, skipping event\n");

  mutex_unlock(&hists->lock);
 }

out:
 addr_location__exit(&al);
}

static void
perf_top__process_lost(struct perf_top *top, union perf_event *event,
         struct evsel *evsel)
{
 top->lost += event->lost.lost;
 top->lost_total += event->lost.lost;
 evsel->evlist->stats.total_lost += event->lost.lost;
}

static void
perf_top__process_lost_samples(struct perf_top *top,
          union perf_event *event,
          struct evsel *evsel)
{
 top->lost += event->lost_samples.lost;
 top->lost_total += event->lost_samples.lost;
 evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
}

static u64 last_timestamp;

static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
{
 struct record_opts *opts = &top->record_opts;
 struct evlist *evlist = top->evlist;
 struct mmap *md;
 union perf_event *event;

 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 if (perf_mmap__read_init(&md->core) < 0)
  return;

 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
  int ret;

  ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
  if (ret && ret != -1)
   break;

  ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0, NULL);
  if (ret)
   break;

  perf_mmap__consume(&md->core);

  if (top->qe.rotate) {
   mutex_lock(&top->qe.mutex);
   top->qe.rotate = false;
   cond_signal(&top->qe.cond);
   mutex_unlock(&top->qe.mutex);
  }
 }

 perf_mmap__read_done(&md->core);
}

static void perf_top__mmap_read(struct perf_top *top)
{
 bool overwrite = top->record_opts.overwrite;
 struct evlist *evlist = top->evlist;
 int i;

 if (overwrite)
  evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);

 for (i = 0; i < top->evlist->core.nr_mmaps; i++)
  perf_top__mmap_read_idx(top, i);

 if (overwrite) {
  evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
  evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 }
}

/*
 * Check per-event overwrite term.
 * perf top should support consistent term for all events.
 * - All events don't have per-event term
 *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
 *   Nothing change, return 0.
 * - All events have same per-event term
 *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
 *   Using the per-event setting to replace the opts->overwrite if
 *   they are different, then return 0.
 * - Events have different per-event term
 *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
 *   Return -1
 * - Some of the event set per-event term, but some not.
 *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
 *   Return -1
 */
static int perf_top__overwrite_check(struct perf_top *top)
{
 struct record_opts *opts = &top->record_opts;
 struct evlist *evlist = top->evlist;
 struct evsel_config_term *term;
 struct list_head *config_terms;
 struct evsel *evsel;
 int set, overwrite = -1;

 evlist__for_each_entry(evlist, evsel) {
  set = -1;
  config_terms = &evsel->config_terms;
  list_for_each_entry(term, config_terms, list) {
   if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
    set = term->val.overwrite ? 1 : 0;
  }

  /* no term for current and previous event (likely) */
  if ((overwrite < 0) && (set < 0))
   continue;

  /* has term for both current and previous event, compare */
  if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
   return -1;

  /* no term for current event but has term for previous one */
  if ((overwrite >= 0) && (set < 0))
   return -1;

  /* has term for current event */
  if ((overwrite < 0) && (set >= 0)) {
   /* if it's first event, set overwrite */
   if (evsel == evlist__first(evlist))
    overwrite = set;
   else
    return -1;
  }
 }

 if ((overwrite >= 0) && (opts->overwrite != overwrite))
  opts->overwrite = overwrite;

 return 0;
}

static int perf_top_overwrite_fallback(struct perf_top *top,
           struct evsel *evsel)
{
 struct record_opts *opts = &top->record_opts;
 struct evlist *evlist = top->evlist;
 struct evsel *counter;

 if (!opts->overwrite)
  return 0;

 /* only fall back when first event fails */
 if (evsel != evlist__first(evlist))
  return 0;

 evlist__for_each_entry(evlist, counter)
  counter->core.attr.write_backward = false;
 opts->overwrite = false;
 pr_debug2("fall back to non-overwrite mode\n");
 return 1;
}

static int perf_top__start_counters(struct perf_top *top)
{
 char msg[BUFSIZ];
 struct evsel *counter;
 struct evlist *evlist = top->evlist;
 struct record_opts *opts = &top->record_opts;

 if (perf_top__overwrite_check(top)) {
  ui__error("perf top only support consistent per-event "
     "overwrite setting for all events\n");
  goto out_err;
 }

 evlist__config(evlist, opts, &callchain_param);

 evlist__for_each_entry(evlist, counter) {
try_again:
  if (evsel__open(counter, counter->core.cpus,
    counter->core.threads) < 0) {

   /*
 * Specially handle overwrite fall back.
 * Because perf top is the only tool which has
 * overwrite mode by default, support
 * both overwrite and non-overwrite mode, and
 * require consistent mode for all events.
 *
 * May move it to generic code with more tools
 * have similar attribute.
 */
   if (perf_missing_features.write_backward &&
       perf_top_overwrite_fallback(top, counter))
    goto try_again;

   if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
    if (verbose > 0)
     ui__warning("%s\n", msg);
    goto try_again;
   }

   evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
   ui__error("%s\n", msg);
   goto out_err;
  }
 }

 if (evlist__apply_filters(evlist, &counter, &opts->target)) {
  pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
   counter->filter ?: "BPF", evsel__name(counter), errno,
   str_error_r(errno, msg, sizeof(msg)));
  goto out_err;
 }

 if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
  ui__error("Failed to mmap with %d (%s)\n",
       errno, str_error_r(errno, msg, sizeof(msg)));
  goto out_err;
 }

 return 0;

out_err:
 return -1;
}

static int callchain_param__setup_sample_type(struct callchain_param *callchain)
{
 if (callchain->mode != CHAIN_NONE) {
  if (callchain_register_param(callchain) < 0) {
   ui__error("Can't register callchain params.\n");
   return -EINVAL;
  }
 }

 return 0;
}

static struct ordered_events *rotate_queues(struct perf_top *top)
{
 struct ordered_events *in = top->qe.in;

 if (top->qe.in == &top->qe.data[1])
  top->qe.in = &top->qe.data[0];
 else
  top->qe.in = &top->qe.data[1];

 return in;
}

static void *process_thread(void *arg)
{
 struct perf_top *top = arg;

 while (!done) {
  struct ordered_events *out, *in = top->qe.in;

  if (!in->nr_events) {
   usleep(100);
   continue;
  }

  out = rotate_queues(top);

  mutex_lock(&top->qe.mutex);
  top->qe.rotate = true;
  cond_wait(&top->qe.cond, &top->qe.mutex);
  mutex_unlock(&top->qe.mutex);

  if (ordered_events__flush(out, OE_FLUSH__TOP))
   pr_err("failed to process events\n");
 }

 return NULL;
}

/*
 * Allow only 'top->delay_secs' seconds behind samples.
 */
static int should_drop(struct ordered_event *qevent, struct perf_top *top)
{
 union perf_event *event = qevent->event;
 u64 delay_timestamp;

 if (event->header.type != PERF_RECORD_SAMPLE)
  return false;

 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
 return delay_timestamp < last_timestamp;
}

static int deliver_event(struct ordered_events *qe,
    struct ordered_event *qevent)
{
 struct perf_top *top = qe->data;
 struct evlist *evlist = top->evlist;
 struct perf_session *session = top->session;
 union perf_event *event = qevent->event;
 struct perf_sample sample;
 struct evsel *evsel;
 struct machine *machine;
 int ret = -1;

 if (should_drop(qevent, top)) {
  top->drop++;
  top->drop_total++;
  return 0;
 }

 perf_sample__init(&sample, /*all=*/false);
 ret = evlist__parse_sample(evlist, event, &sample);
 if (ret) {
  pr_err("Can't parse sample, err = %d\n", ret);
  goto next_event;
 }

 evsel = evlist__id2evsel(session->evlist, sample.id);
 assert(evsel != NULL);

 if (event->header.type == PERF_RECORD_SAMPLE) {
  if (evswitch__discard(&top->evswitch, evsel)) {
   ret = 0;
   goto next_event;
  }
  ++top->samples;
 }

 switch (sample.cpumode) {
 case PERF_RECORD_MISC_USER:
  ++top->us_samples;
  if (top->hide_user_symbols)
   goto next_event;
  machine = &session->machines.host;
  break;
 case PERF_RECORD_MISC_KERNEL:
  ++top->kernel_samples;
  if (top->hide_kernel_symbols)
   goto next_event;
  machine = &session->machines.host;
  break;
 case PERF_RECORD_MISC_GUEST_KERNEL:
  ++top->guest_kernel_samples;
  machine = perf_session__find_machine(session,
           sample.pid);
  break;
 case PERF_RECORD_MISC_GUEST_USER:
  ++top->guest_us_samples;
  /*
 * TODO: we don't process guest user from host side
 * except simple counting.
 */
  goto next_event;
 default:
  if (event->header.type == PERF_RECORD_SAMPLE)
   goto next_event;
  machine = &session->machines.host;
  break;
 }

 if (event->header.type == PERF_RECORD_SAMPLE) {
  perf_event__process_sample(&top->tool, event, evsel,
        &sample, machine);
 } else if (event->header.type == PERF_RECORD_LOST) {
  perf_top__process_lost(top, event, evsel);
 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
  perf_top__process_lost_samples(top, event, evsel);
 } else if (event->header.type < PERF_RECORD_MAX) {
  events_stats__inc(&session->evlist->stats, event->header.type);
  machine__process_event(machine, event, &sample);
 } else
  ++session->evlist->stats.nr_unknown_events;

 ret = 0;
next_event:
 perf_sample__exit(&sample);
 return ret;
}

static void init_process_thread(struct perf_top *top)
{
 ordered_events__init(&top->qe.data[0], deliver_event, top);
 ordered_events__init(&top->qe.data[1], deliver_event, top);
 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
 top->qe.in = &top->qe.data[0];
 mutex_init(&top->qe.mutex);
 cond_init(&top->qe.cond);
}

static void exit_process_thread(struct perf_top *top)
{
 ordered_events__free(&top->qe.data[0]);
 ordered_events__free(&top->qe.data[1]);
 mutex_destroy(&top->qe.mutex);
 cond_destroy(&top->qe.cond);
}

static int __cmd_top(struct perf_top *top)
{
 struct record_opts *opts = &top->record_opts;
 pthread_t thread, thread_process;
 int ret;

 if (!annotate_opts.objdump_path) {
  ret = perf_env__lookup_objdump(perf_session__env(top->session),
            &annotate_opts.objdump_path);
  if (ret)
   return ret;
 }

 ret = callchain_param__setup_sample_type(&callchain_param);
 if (ret)
  return ret;

 if (perf_session__register_idle_thread(top->session) < 0)
  return ret;

 if (top->nr_threads_synthesize > 1)
  perf_set_multithreaded();

 init_process_thread(top);

 if (opts->record_namespaces)
  top->tool.namespace_events = true;
 if (opts->record_cgroup) {
#ifdef HAVE_FILE_HANDLE
  top->tool.cgroup_events = true;
#else
  pr_err("cgroup tracking is not supported.\n");
  return -1;
#endif
 }

 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
      &top->session->machines.host,
      &top->record_opts);
 if (ret < 0)
  pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");

 ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
          &top->session->machines.host);
 if (ret < 0)
  pr_debug("Couldn't synthesize cgroup events.\n");

 machine__synthesize_threads(&top->session->machines.host, &opts->target,
        top->evlist->core.threads, true, false,
        top->nr_threads_synthesize);

 perf_set_multithreaded();

 if (perf_hpp_list.socket) {
  ret = perf_env__read_cpu_topology_map(perf_session__env(top->session));
  if (ret < 0) {
   char errbuf[BUFSIZ];
   const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));

   ui__error("Could not read the CPU topology map: %s\n", err);
   return ret;
  }
 }

 /*
 * Use global stat_config that is zero meaning aggr_mode is AGGR_NONE
 * and hybrid_merge is false.
 */
 evlist__uniquify_evsel_names(top->evlist, &stat_config);
 ret = perf_top__start_counters(top);
 if (ret)
  return ret;

 top->session->evlist = top->evlist;
 perf_session__set_id_hdr_size(top->session);

 /*
 * When perf is starting the traced process, all the events (apart from
 * group members) have enable_on_exec=1 set, so don't spoil it by
 * prematurely enabling them.
 *
 * XXX 'top' still doesn't start workloads like record, trace, but should,
 * so leave the check here.
 */
        if (!target__none(&opts->target))
  evlist__enable(top->evlist);

 ret = -1;
 if (pthread_create(&thread_process, NULL, process_thread, top)) {
  ui__error("Could not create process thread.\n");
  return ret;
 }

 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
           display_thread), top)) {
  ui__error("Could not create display thread.\n");
  goto out_join_thread;
 }

 if (top->realtime_prio) {
  struct sched_param param;

  param.sched_priority = top->realtime_prio;
  if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
   ui__error("Could not set realtime priority.\n");
   goto out_join;
  }
 }

 /* Wait for a minimal set of events before starting the snapshot */
 evlist__poll(top->evlist, 100);

 perf_top__mmap_read(top);

 while (!done) {
  u64 hits = top->samples;

  perf_top__mmap_read(top);

  if (opts->overwrite || (hits == top->samples))
   ret = evlist__poll(top->evlist, 100);

  if (resize) {
   perf_top__resize(top);
   resize = 0;
  }
 }

 ret = 0;
out_join:
 pthread_join(thread, NULL);
out_join_thread:
 cond_signal(&top->qe.cond);
 pthread_join(thread_process, NULL);
 perf_set_singlethreaded();
 exit_process_thread(top);
 return ret;
}

static int
callchain_opt(const struct option *opt, const char *arg, int unset)
{
 symbol_conf.use_callchain = true;
 return record_callchain_opt(opt, arg, unset);
}

static int
parse_callchain_opt(const struct option *opt, const char *arg, int unset)
{
 struct callchain_param *callchain = opt->value;

 callchain->enabled = !unset;
 callchain->record_mode = CALLCHAIN_FP;

 /*
 * --no-call-graph
 */
 if (unset) {
  symbol_conf.use_callchain = false;
  callchain->record_mode = CALLCHAIN_NONE;
  return 0;
 }

 return parse_callchain_top_opt(arg);
}

static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
{
 if (!strcmp(var, "top.call-graph")) {
  var = "call-graph.record-mode";
  return perf_default_config(var, value, cb);
 }
 if (!strcmp(var, "top.children")) {
  symbol_conf.cumulate_callchain = perf_config_bool(var, value);
  return 0;
 }

 return 0;
}

static int
parse_percent_limit(const struct option *opt, const char *arg,
      int unset __maybe_unused)
{
 struct perf_top *top = opt->value;

 top->min_percent = strtof(arg, NULL);
 return 0;
}

const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";

int cmd_top(int argc, const char **argv)
{
 char errbuf[BUFSIZ];
 struct perf_top top = {
  .count_filter      = 5,
  .delay_secs      = 2,
  .record_opts = {
   .mmap_pages = UINT_MAX,
   .user_freq = UINT_MAX,
   .user_interval = ULLONG_MAX,
   .freq  = 4000, /* 4 KHz */
   .target  = {
    .uses_mmap   = true,
   },
   /*
 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
 * when we pause, fix that and reenable. Probably using a
 * separate evlist with a dummy event, i.e. a non-overwrite
 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
 * stays in overwrite mode. -acme
 * */
   .overwrite = 0,
   .sample_time = true,
   .sample_time_set = true,
  },
  .max_stack      = sysctl__max_stack(),
  .nr_threads_synthesize = UINT_MAX,
 };
 struct parse_events_option_args parse_events_option_args = {
  .evlistp = &top.evlist,
 };
 bool branch_call_mode = false;
 struct record_opts *opts = &top.record_opts;
 struct target *target = &opts->target;
 const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
 const struct option options[] = {
 OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
       "event selector. use 'perf list' to list available events",
       parse_events_option),
 OPT_CALLBACK(0, "filter", &top.evlist, "filter",
       "event filter", parse_filter),
 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
 OPT_STRING('p', "pid", &target->pid, "pid",
      "profile events on existing process id"),
 OPT_STRING('t', "tid", &target->tid, "tid",
      "profile events on existing thread id"),
 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
       "system-wide collection from all CPUs"),
 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
      "list of cpus to monitor"),
 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
     "file", "vmlinux pathname"),
 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
      "don't load vmlinux even if found"),
 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
     "file", "kallsyms pathname"),
 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
      "hide kernel symbols"),
 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
       "number of mmap data pages", evlist__parse_mmap_pages),
 OPT_INTEGER('r', "realtime", &top.realtime_prio,
      "collect data with this RT SCHED_FIFO priority"),
 OPT_INTEGER('d', "delay", &top.delay_secs,
      "number of seconds to delay between refreshes"),
 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
       "dump the symbol table used for profiling"),
 OPT_INTEGER('f', "count-filter", &top.count_filter,
      "only display functions with more events than this"),
 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
      "child tasks do not inherit counters"),
 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
      "symbol to annotate"),
 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
       "profile at this frequency",
        record__parse_freq),
 OPT_INTEGER('E', "entries", &top.print_entries,
      "display this many functions"),
 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
      "hide user symbols"),
#ifdef HAVE_SLANG_SUPPORT
 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
#endif
 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
 OPT_INCR('v', "verbose", &verbose,
      "be more verbose (show counter open errors, etc)"),
 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
     "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
     " Please refer the man page for the complete list."),
 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
     "output field(s): overhead, period, sample plus all of sort keys"),
 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
      "Show a column with the number of samples"),
 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
      NULL, "enables call-graph recording and display",
      &callchain_opt),
 OPT_CALLBACK(0, "call-graph", &callchain_param,
       "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
       top_callchain_help, &parse_callchain_opt),
 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
      "Accumulate callchains of children and show total overhead as well"),
 OPT_INTEGER(0, "max-stack", &top.max_stack,
      "Set the maximum stack depth when parsing the callchain. "
      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
     "ignore callees of these functions in call graphs",
     report_parse_ignore_callees_opt),
 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
      "Show a column with the sum of periods"),
 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
     "only consider symbols in these dsos"),
 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
     "only consider symbols in these comms"),
 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
     "only consider these symbols"),
 OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
      "Interleave source code with assembly code (default)"),
 OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
      "Display raw encoding of assembly instructions (default)"),
 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
      "Enable kernel symbol demangling"),
 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
 OPT_STRING(0, "objdump", &objdump_path, "path",
      "objdump binary to use for disassembly and annotations"),
 OPT_STRING(0, "addr2line", &addr2line_path, "path",
     "addr2line binary to use for line numbers"),
 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
     "Specify disassembler style (e.g. -M intel for intel syntax)"),
 OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
      "Add prefix to source file path names in programs (with --prefix-strip)"),
 OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
      "Strip first N entries of source file path name in programs (with --prefix)"),
 OPT_STRING('u', "uid", &top.uid_str, "user", "user to profile"),
 OPT_CALLBACK(0, "percent-limit", &top, "percent",
       "Don't show entries under that percent", parse_percent_limit),
 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
       "How to display percentage of filtered entries", parse_filter_percentage),
 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
     "width[,width...]",
     "don't try to adjust column width, use these fixed values"),
 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
   "per thread proc mmap processing timeout in ms"),
 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
       "branch any", "sample any taken branches",
       parse_branch_stack),
 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
       "branch filter mask", "branch stack filter modes",
       parse_branch_stack),
 OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
      "add last branch records to call history"),
 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
      "Show raw trace event output (do not use print fmt or plugins)"),
 OPT_BOOLEAN('H', "hierarchy", &symbol_conf.report_hierarchy,
      "Show entries in a hierarchy"),
 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
      "Use a backward ring buffer, default: no"),
 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
   "number of thread to run event synthesize"),
 OPT_CALLBACK('G', "cgroup", &top.evlist, "name",
       "monitor event in cgroup name only", parse_cgroups),
 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
      "Record namespaces events"),
 OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
      "Record cgroup events"),
 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
      "Sort the output by the event at the index n in group. "
      "If n is invalid, sort by the first event. "
      "WARNING: should be used on grouped events."),
 OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
      "Enable LBR callgraph stitching approach"),
#ifdef HAVE_LIBPFM
 OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
  "libpfm4 event selector. use 'perf list' to list available events",
  parse_libpfm_events_option),
#endif
 OPTS_EVSWITCH(&top.evswitch),
 OPT_END()
 };
 const char * const top_usage[] = {
  "perf top []",
  NULL
 };
 int status = hists__init();
 struct perf_env host_env;

 if (status < 0)
  return status;

 annotation_options__init();

 annotate_opts.min_pcnt = 5;
 annotate_opts.context  = 4;

 top.evlist = evlist__new();
 if (top.evlist == NULL)
  return -ENOMEM;

 perf_env__init(&host_env);
 status = perf_config(perf_top_config, &top);
 if (status)
  goto out_delete_evlist;
 /*
 * Since the per arch annotation init routine may need the cpuid, read
 * it here, since we are not getting this from the perf.data header.
 */
 status = perf_env__set_cmdline(&host_env, argc, argv);
 if (status)
  goto out_delete_evlist;

 status = perf_env__read_cpuid(&host_env);
 if (status) {
  /*
 * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
 * warn the user explicitly.
 */
  eprintf(status == ENOSYS ? 1 : 0, verbose,
   "Couldn't read the cpuid for this machine: %s\n",
   str_error_r(errno, errbuf, sizeof(errbuf)));
 }

 argc = parse_options(argc, argv, options, top_usage, 0);
 if (argc)
  usage_with_options(top_usage, options);

 if (disassembler_style) {
  annotate_opts.disassembler_style = strdup(disassembler_style);
  if (!annotate_opts.disassembler_style) {
   status = -ENOMEM;
   goto out_delete_evlist;
  }
 }
 if (objdump_path) {
  annotate_opts.objdump_path = strdup(objdump_path);
  if (!annotate_opts.objdump_path) {
   status = -ENOMEM;
   goto out_delete_evlist;
  }
 }
 if (addr2line_path) {
  symbol_conf.addr2line_path = strdup(addr2line_path);
  if (!symbol_conf.addr2line_path) {
   status = -ENOMEM;
   goto out_delete_evlist;
  }
 }

 status = symbol__validate_sym_arguments();
 if (status)
  goto out_delete_evlist;

 if (annotate_check_args() < 0)
  goto out_delete_evlist;

 if (!top.evlist->core.nr_entries) {
  bool can_profile_kernel = perf_event_paranoid_check(1);
  int err = parse_event(top.evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");

  if (err)
   goto out_delete_evlist;
 }

 status = evswitch__init(&top.evswitch, top.evlist, stderr);
 if (status)
  goto out_delete_evlist;

 if (symbol_conf.report_hierarchy) {
  /* disable incompatible options */
  symbol_conf.event_group = false;
  symbol_conf.cumulate_callchain = false;

  if (field_order) {
   pr_err("Error: --hierarchy and --fields options cannot be used together\n");
   parse_options_usage(top_usage, options, "fields", 0);
   parse_options_usage(NULL, options, "hierarchy", 0);
   goto out_delete_evlist;
  }
 }

 if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
  pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
  goto out_delete_evlist;
 }

 if (nr_cgroups > 0 && opts->record_cgroup) {
  pr_err("--cgroup and --all-cgroups cannot be used together\n");
  goto out_delete_evlist;
 }

 if (branch_call_mode) {
  if (!opts->branch_stack)
   opts->branch_stack = PERF_SAMPLE_BRANCH_ANY;
  symbol_conf.use_callchain = true;
  callchain_param.key = CCKEY_ADDRESS;
  callchain_param.branch_callstack = true;
  callchain_param.enabled = true;
  if (callchain_param.record_mode == CALLCHAIN_NONE)
   callchain_param.record_mode = CALLCHAIN_FP;
  callchain_register_param(&callchain_param);
  if (!sort_order)
   sort_order = "srcline,symbol,dso";
 }

 if (opts->branch_stack && callchain_param.enabled)
  symbol_conf.show_branchflag_count = true;

 if (opts->branch_stack) {
  status = perf_env__read_core_pmu_caps(&host_env);
  if (status) {
   pr_err("PMU capability data is not available\n");
   goto out_delete_evlist;
  }
 }

 sort__mode = SORT_MODE__TOP;
 /* display thread wants entries to be collapsed in a different tree */
 perf_hpp_list.need_collapse = 1;

 if (top.use_stdio)
  use_browser = 0;
#ifdef HAVE_SLANG_SUPPORT
 else if (top.use_tui)
  use_browser = 1;
#endif

 setup_browser(false);

 top.session = __perf_session__new(/*data=*/NULL, /*tool=*/NULL,
       /*trace_event_repipe=*/false,
       &host_env);
 if (IS_ERR(top.session)) {
  status = PTR_ERR(top.session);
  top.session = NULL;
  goto out_delete_evlist;
 }
 top.evlist->session = top.session;

 if (setup_sorting(top.evlist, perf_session__env(top.session)) < 0) {
  if (sort_order)
   parse_options_usage(top_usage, options, "s", 1);
  if (field_order)
   parse_options_usage(sort_order ? NULL : top_usage,
         options, "fields", 0);
  goto out_delete_evlist;
 }

 status = target__validate(target);
 if (status) {
  target__strerror(target, status, errbuf, BUFSIZ);
  ui__warning("%s\n", errbuf);
 }

 if (top.uid_str) {
  uid_t uid = parse_uid(top.uid_str);

  if (uid == UINT_MAX) {
   ui__error("Invalid User: %s", top.uid_str);
   status = -EINVAL;
   goto out_delete_evlist;
  }
  status = parse_uid_filter(top.evlist, uid);
  if (status)
   goto out_delete_evlist;
 }

 if (target__none(target))
  target->system_wide = true;

 if (evlist__create_maps(top.evlist, target) < 0) {
  ui__error("Couldn't create thread/CPU maps: %s\n",
     errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
  status = -errno;
  goto out_delete_evlist;
 }

 if (top.delay_secs < 1)
  top.delay_secs = 1;

 if (record_opts__config(opts)) {
  status = -EINVAL;
  goto out_delete_evlist;
 }

 top.sym_evsel = evlist__first(top.evlist);

 if (!callchain_param.enabled) {
  symbol_conf.cumulate_callchain = false;
  perf_hpp__cancel_cumulate(top.evlist);
 }

 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
  callchain_param.order = ORDER_CALLER;

 status = symbol__annotation_init();
 if (status < 0)
  goto out_delete_evlist;

 annotation_config__init();

 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
 status = symbol__init(NULL);
 if (status < 0)
  goto out_delete_evlist;

 sort__setup_elide(stdout);

 get_term_dimensions(&top.winsize);
 if (top.print_entries == 0) {
  perf_top__update_print_entries(&top);
  signal(SIGWINCH, winch_sig);
 }

 if (!evlist__needs_bpf_sb_event(top.evlist))
  top.record_opts.no_bpf_event = true;

#ifdef HAVE_LIBBPF_SUPPORT
 if (!top.record_opts.no_bpf_event) {
  top.sb_evlist = evlist__new();

  if (top.sb_evlist == NULL) {
   pr_err("Couldn't create side band evlist.\n.");
   status = -EINVAL;
   goto out_delete_evlist;
  }

  if (evlist__add_bpf_sb_event(top.sb_evlist, &host_env)) {
   pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
   status = -EINVAL;
   goto out_delete_evlist;
  }
 }
#endif

 if (evlist__start_sb_thread(top.sb_evlist, target)) {
  pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
  opts->no_bpf_event = true;
 }

 status = __cmd_top(&top);

 if (!opts->no_bpf_event)
  evlist__stop_sb_thread(top.sb_evlist);

out_delete_evlist:
 evlist__delete(top.evlist);
 perf_session__delete(top.session);
 annotation_options__exit();
 perf_env__exit(&host_env);

 return status;
}