4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10 * Improvements and fixes by:
12 * Arjan van de Ven <arjan@linux.intel.com>
13 * Yanmin Zhang <yanmin.zhang@intel.com>
14 * Wu Fengguang <fengguang.wu@intel.com>
15 * Mike Galbraith <efault@gmx.de>
16 * Paul Mackerras <paulus@samba.org>
18 * Released under the GPL v2. (and only v2, not any later version)
24 #include "util/annotate.h"
25 #include "util/cache.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/session.h"
30 #include "util/symbol.h"
31 #include "util/thread.h"
32 #include "util/thread_map.h"
34 #include "util/util.h"
35 #include <linux/rbtree.h>
36 #include "util/parse-options.h"
37 #include "util/parse-events.h"
38 #include "util/cpumap.h"
39 #include "util/xyarray.h"
40 #include "util/sort.h"
42 #include "util/debug.h"
56 #include <sys/syscall.h>
57 #include <sys/ioctl.h>
59 #include <sys/prctl.h>
64 #include <linux/unistd.h>
65 #include <linux/types.h>
67 void get_term_dimensions(struct winsize *ws)
69 char *s = getenv("LINES");
73 s = getenv("COLUMNS");
76 if (ws->ws_row && ws->ws_col)
81 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
82 ws->ws_row && ws->ws_col)
89 static void perf_top__update_print_entries(struct perf_top *top)
91 if (top->print_entries > 9)
92 top->print_entries -= 9;
95 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
97 struct perf_top *top = arg;
99 get_term_dimensions(&top->winsize);
100 if (!top->print_entries
101 || (top->print_entries+4) > top->winsize.ws_row) {
102 top->print_entries = top->winsize.ws_row;
104 top->print_entries += 4;
105 top->winsize.ws_row = top->print_entries;
107 perf_top__update_print_entries(top);
110 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
113 struct annotation *notes;
117 if (!he || !he->ms.sym)
124 * We can't annotate with just /proc/kallsyms
126 if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
127 pr_err("Can't annotate %s: No vmlinux file was found in the "
128 "path\n", sym->name);
133 notes = symbol__annotation(sym);
134 if (notes->src != NULL) {
135 pthread_mutex_lock(¬es->lock);
139 pthread_mutex_lock(¬es->lock);
141 if (symbol__alloc_hist(sym) < 0) {
142 pthread_mutex_unlock(¬es->lock);
143 pr_err("Not enough memory for annotating '%s' symbol!\n",
149 err = symbol__annotate(sym, map, 0);
152 top->sym_filter_entry = he;
155 pthread_mutex_unlock(¬es->lock);
159 static void __zero_source_counters(struct hist_entry *he)
161 struct symbol *sym = he->ms.sym;
162 symbol__annotate_zero_histograms(sym);
165 static void perf_top__record_precise_ip(struct perf_top *top,
166 struct hist_entry *he,
169 struct annotation *notes;
172 if (he == NULL || he->ms.sym == NULL ||
173 ((top->sym_filter_entry == NULL ||
174 top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
178 notes = symbol__annotation(sym);
180 if (pthread_mutex_trylock(¬es->lock))
183 if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
184 pthread_mutex_unlock(¬es->lock);
185 pr_err("Not enough memory for annotating '%s' symbol!\n",
191 ip = he->ms.map->map_ip(he->ms.map, ip);
192 symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
194 pthread_mutex_unlock(¬es->lock);
197 static void perf_top__show_details(struct perf_top *top)
199 struct hist_entry *he = top->sym_filter_entry;
200 struct annotation *notes;
201 struct symbol *symbol;
208 notes = symbol__annotation(symbol);
210 pthread_mutex_lock(¬es->lock);
212 if (notes->src == NULL)
215 printf("Showing %s for %s\n", event_name(top->sym_evsel), symbol->name);
216 printf(" Events Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
218 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
219 0, top->sym_pcnt_filter, top->print_entries, 4);
221 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
223 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
225 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
227 pthread_mutex_unlock(¬es->lock);
230 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
232 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
233 struct addr_location *al,
234 struct perf_sample *sample)
236 struct hist_entry *he;
238 he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
242 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
246 static void perf_top__print_sym_table(struct perf_top *top)
250 const int win_width = top->winsize.ws_col - 1;
254 perf_top__header_snprintf(top, bf, sizeof(bf));
257 perf_top__reset_sample_counters(top);
259 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
261 if (top->sym_evsel->hists.stats.nr_lost_warned !=
262 top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
263 top->sym_evsel->hists.stats.nr_lost_warned =
264 top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
265 color_fprintf(stdout, PERF_COLOR_RED,
266 "WARNING: LOST %d chunks, Check IO/CPU overload",
267 top->sym_evsel->hists.stats.nr_lost_warned);
271 if (top->sym_filter_entry) {
272 perf_top__show_details(top);
276 hists__collapse_resort_threaded(&top->sym_evsel->hists);
277 hists__output_resort_threaded(&top->sym_evsel->hists);
278 hists__decay_entries_threaded(&top->sym_evsel->hists,
279 top->hide_user_symbols,
280 top->hide_kernel_symbols);
281 hists__output_recalc_col_len(&top->sym_evsel->hists,
282 top->winsize.ws_row - 3);
284 hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
285 top->winsize.ws_row - 4 - printed, win_width, stdout);
288 static void prompt_integer(int *target, const char *msg)
290 char *buf = malloc(0), *p;
294 fprintf(stdout, "\n%s: ", msg);
295 if (getline(&buf, &dummy, stdin) < 0)
298 p = strchr(buf, '\n');
308 tmp = strtoul(buf, NULL, 10);
314 static void prompt_percent(int *target, const char *msg)
318 prompt_integer(&tmp, msg);
319 if (tmp >= 0 && tmp <= 100)
323 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
325 char *buf = malloc(0), *p;
326 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
327 struct rb_node *next;
330 /* zero counters of active symbol */
332 __zero_source_counters(syme);
333 top->sym_filter_entry = NULL;
336 fprintf(stdout, "\n%s: ", msg);
337 if (getline(&buf, &dummy, stdin) < 0)
340 p = strchr(buf, '\n');
344 next = rb_first(&top->sym_evsel->hists.entries);
346 n = rb_entry(next, struct hist_entry, rb_node);
347 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
351 next = rb_next(&n->rb_node);
355 fprintf(stderr, "Sorry, %s is not active.\n", buf);
358 perf_top__parse_source(top, found);
364 static void perf_top__print_mapped_keys(struct perf_top *top)
368 if (top->sym_filter_entry) {
369 struct symbol *sym = top->sym_filter_entry->ms.sym;
373 fprintf(stdout, "\nMapped keys:\n");
374 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
375 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
377 if (top->evlist->nr_entries > 1)
378 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(top->sym_evsel));
380 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
382 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
383 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
384 fprintf(stdout, "\t[S] stop annotation.\n");
387 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
388 top->hide_kernel_symbols ? "yes" : "no");
390 "\t[U] hide user symbols. \t(%s)\n",
391 top->hide_user_symbols ? "yes" : "no");
392 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
393 fprintf(stdout, "\t[qQ] quit.\n");
396 static int perf_top__key_mapped(struct perf_top *top, int c)
412 return top->evlist->nr_entries > 1 ? 1 : 0;
420 static void perf_top__handle_keypress(struct perf_top *top, int c)
422 if (!perf_top__key_mapped(top, c)) {
423 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
424 struct termios tc, save;
426 perf_top__print_mapped_keys(top);
427 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
432 tc.c_lflag &= ~(ICANON | ECHO);
435 tcsetattr(0, TCSANOW, &tc);
437 poll(&stdin_poll, 1, -1);
440 tcsetattr(0, TCSAFLUSH, &save);
441 if (!perf_top__key_mapped(top, c))
447 prompt_integer(&top->delay_secs, "Enter display delay");
448 if (top->delay_secs < 1)
452 prompt_integer(&top->print_entries, "Enter display entries (lines)");
453 if (top->print_entries == 0) {
454 struct sigaction act = {
455 .sa_sigaction = perf_top__sig_winch,
456 .sa_flags = SA_SIGINFO,
458 perf_top__sig_winch(SIGWINCH, NULL, top);
459 sigaction(SIGWINCH, &act, NULL);
461 perf_top__sig_winch(SIGWINCH, NULL, top);
462 signal(SIGWINCH, SIG_DFL);
466 if (top->evlist->nr_entries > 1) {
467 /* Select 0 as the default event: */
470 fprintf(stderr, "\nAvailable events:");
472 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
473 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
475 prompt_integer(&counter, "Enter details event counter");
477 if (counter >= top->evlist->nr_entries) {
478 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
479 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
483 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
484 if (top->sym_evsel->idx == counter)
487 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
490 prompt_integer(&top->count_filter, "Enter display event count filter");
493 prompt_percent(&top->sym_pcnt_filter,
494 "Enter details display event filter (percent)");
497 top->hide_kernel_symbols = !top->hide_kernel_symbols;
501 printf("exiting.\n");
502 if (top->dump_symtab)
503 perf_session__fprintf_dsos(top->session, stderr);
506 perf_top__prompt_symbol(top, "Enter details symbol");
509 if (!top->sym_filter_entry)
512 struct hist_entry *syme = top->sym_filter_entry;
514 top->sym_filter_entry = NULL;
515 __zero_source_counters(syme);
519 top->hide_user_symbols = !top->hide_user_symbols;
522 top->zero = !top->zero;
529 static void perf_top__sort_new_samples(void *arg)
531 struct perf_top *t = arg;
532 perf_top__reset_sample_counters(t);
534 if (t->evlist->selected != NULL)
535 t->sym_evsel = t->evlist->selected;
537 hists__collapse_resort_threaded(&t->sym_evsel->hists);
538 hists__output_resort_threaded(&t->sym_evsel->hists);
539 hists__decay_entries_threaded(&t->sym_evsel->hists,
540 t->hide_user_symbols,
541 t->hide_kernel_symbols);
544 static void *display_thread_tui(void *arg)
546 struct perf_evsel *pos;
547 struct perf_top *top = arg;
548 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
550 perf_top__sort_new_samples(top);
553 * Initialize the uid_filter_str, in the future the TUI will allow
554 * Zooming in/out UIDs. For now juse use whatever the user passed
557 list_for_each_entry(pos, &top->evlist->entries, node)
558 pos->hists.uid_filter_str = top->uid_str;
560 perf_evlist__tui_browse_hists(top->evlist, help,
561 perf_top__sort_new_samples,
562 top, top->delay_secs);
569 static void *display_thread(void *arg)
571 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
572 struct termios tc, save;
573 struct perf_top *top = arg;
578 tc.c_lflag &= ~(ICANON | ECHO);
582 pthread__unblock_sigwinch();
584 delay_msecs = top->delay_secs * 1000;
585 tcsetattr(0, TCSANOW, &tc);
590 perf_top__print_sym_table(top);
592 * Either timeout expired or we got an EINTR due to SIGWINCH,
593 * refresh screen in both cases.
595 switch (poll(&stdin_poll, 1, delay_msecs)) {
608 tcsetattr(0, TCSAFLUSH, &save);
610 perf_top__handle_keypress(top, c);
616 /* Tag samples to be skipped. */
617 static const char *skip_symbols[] = {
624 "mwait_idle_with_hints",
626 "ppc64_runlatch_off",
627 "pseries_dedicated_idle_sleep",
631 static int symbol_filter(struct map *map __used, struct symbol *sym)
633 const char *name = sym->name;
637 * ppc64 uses function descriptors and appends a '.' to the
638 * start of every instruction address. Remove it.
643 if (!strcmp(name, "_text") ||
644 !strcmp(name, "_etext") ||
645 !strcmp(name, "_sinittext") ||
646 !strncmp("init_module", name, 11) ||
647 !strncmp("cleanup_module", name, 14) ||
648 strstr(name, "_text_start") ||
649 strstr(name, "_text_end"))
652 for (i = 0; skip_symbols[i]; i++) {
653 if (!strcmp(skip_symbols[i], name)) {
662 static void perf_event__process_sample(struct perf_tool *tool,
663 const union perf_event *event,
664 struct perf_evsel *evsel,
665 struct perf_sample *sample,
666 struct machine *machine)
668 struct perf_top *top = container_of(tool, struct perf_top, tool);
669 struct symbol *parent = NULL;
670 u64 ip = event->ip.ip;
671 struct addr_location al;
674 if (!machine && perf_guest) {
675 pr_err("Can't find guest [%d]'s kernel information\n",
681 pr_err("%u unprocessable samples recorded.",
682 top->session->hists.stats.nr_unprocessable_samples++);
686 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
687 top->exact_samples++;
689 if (perf_event__preprocess_sample(event, machine, &al, sample,
690 symbol_filter) < 0 ||
694 if (!top->kptr_restrict_warned &&
695 symbol_conf.kptr_restrict &&
696 al.cpumode == PERF_RECORD_MISC_KERNEL) {
698 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
699 "Check /proc/sys/kernel/kptr_restrict.\n\n"
700 "Kernel%s samples will not be resolved.\n",
701 !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
703 if (use_browser <= 0)
705 top->kptr_restrict_warned = true;
708 if (al.sym == NULL) {
709 const char *msg = "Kernel samples will not be resolved.\n";
711 * As we do lazy loading of symtabs we only will know if the
712 * specified vmlinux file is invalid when we actually have a
713 * hit in kernel space and then try to load it. So if we get
714 * here and there are _no_ symbols in the DSO backing the
715 * kernel map, bail out.
717 * We may never get here, for instance, if we use -K/
718 * --hide-kernel-symbols, even if the user specifies an
719 * invalid --vmlinux ;-)
721 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
722 al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
723 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
724 if (symbol_conf.vmlinux_name) {
725 ui__warning("The %s file can't be used.\n%s",
726 symbol_conf.vmlinux_name, msg);
728 ui__warning("A vmlinux file was not found.\n%s",
732 if (use_browser <= 0)
734 top->vmlinux_warned = true;
738 if (al.sym == NULL || !al.sym->ignore) {
739 struct hist_entry *he;
741 if ((sort__has_parent || symbol_conf.use_callchain) &&
743 err = machine__resolve_callchain(machine, evsel, al.thread,
744 sample->callchain, &parent);
749 he = perf_evsel__add_hist_entry(evsel, &al, sample);
751 pr_err("Problem incrementing symbol period, skipping event\n");
755 if (symbol_conf.use_callchain) {
756 err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
762 if (top->sort_has_symbols)
763 perf_top__record_precise_ip(top, he, evsel->idx, ip);
769 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
771 struct perf_sample sample;
772 struct perf_evsel *evsel;
773 struct perf_session *session = top->session;
774 union perf_event *event;
775 struct machine *machine;
779 while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
780 ret = perf_session__parse_sample(session, event, &sample);
782 pr_err("Can't parse sample, err = %d\n", ret);
786 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
787 assert(evsel != NULL);
789 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
791 if (event->header.type == PERF_RECORD_SAMPLE)
795 case PERF_RECORD_MISC_USER:
797 if (top->hide_user_symbols)
799 machine = perf_session__find_host_machine(session);
801 case PERF_RECORD_MISC_KERNEL:
802 ++top->kernel_samples;
803 if (top->hide_kernel_symbols)
805 machine = perf_session__find_host_machine(session);
807 case PERF_RECORD_MISC_GUEST_KERNEL:
808 ++top->guest_kernel_samples;
809 machine = perf_session__find_machine(session, event->ip.pid);
811 case PERF_RECORD_MISC_GUEST_USER:
812 ++top->guest_us_samples;
814 * TODO: we don't process guest user from host side
815 * except simple counting.
823 if (event->header.type == PERF_RECORD_SAMPLE) {
824 perf_event__process_sample(&top->tool, event, evsel,
826 } else if (event->header.type < PERF_RECORD_MAX) {
827 hists__inc_nr_events(&evsel->hists, event->header.type);
828 perf_event__process(&top->tool, event, &sample, machine);
830 ++session->hists.stats.nr_unknown_events;
834 static void perf_top__mmap_read(struct perf_top *top)
838 for (i = 0; i < top->evlist->nr_mmaps; i++)
839 perf_top__mmap_read_idx(top, i);
842 static void perf_top__start_counters(struct perf_top *top)
844 struct perf_evsel *counter, *first;
845 struct perf_evlist *evlist = top->evlist;
847 first = list_entry(evlist->entries.next, struct perf_evsel, node);
849 list_for_each_entry(counter, &evlist->entries, node) {
850 struct perf_event_attr *attr = &counter->attr;
851 struct xyarray *group_fd = NULL;
853 if (top->group && counter != first)
854 group_fd = first->fd;
856 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
859 attr->sample_type |= PERF_SAMPLE_PERIOD;
861 attr->sample_freq = top->freq;
864 if (evlist->nr_entries > 1) {
865 attr->sample_type |= PERF_SAMPLE_ID;
866 attr->read_format |= PERF_FORMAT_ID;
869 if (symbol_conf.use_callchain)
870 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
874 attr->inherit = top->inherit;
875 fallback_missing_features:
876 if (top->exclude_guest_missing)
877 attr->exclude_guest = attr->exclude_host = 0;
879 attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
881 if (perf_evsel__open(counter, top->evlist->cpus,
882 top->evlist->threads, top->group,
886 if (err == EPERM || err == EACCES) {
887 ui__error_paranoid();
889 } else if (err == EINVAL) {
890 if (!top->exclude_guest_missing &&
891 (attr->exclude_guest || attr->exclude_host)) {
892 pr_debug("Old kernel, cannot exclude "
893 "guest or host samples.\n");
894 top->exclude_guest_missing = true;
895 goto fallback_missing_features;
896 } else if (!top->sample_id_all_missing) {
898 * Old kernel, no attr->sample_id_type_all field
900 top->sample_id_all_missing = true;
901 goto retry_sample_id;
905 * If it's cycles then fall back to hrtimer
906 * based cpu-clock-tick sw counter, which
907 * is always available even if no PMU support:
909 if (attr->type == PERF_TYPE_HARDWARE &&
910 attr->config == PERF_COUNT_HW_CPU_CYCLES) {
912 ui__warning("Cycles event not supported,\n"
913 "trying to fall back to cpu-clock-ticks\n");
915 attr->type = PERF_TYPE_SOFTWARE;
916 attr->config = PERF_COUNT_SW_CPU_CLOCK;
921 ui__warning("The %s event is not supported.\n",
922 event_name(counter));
924 } else if (err == EMFILE) {
925 ui__warning("Too many events are opened.\n"
926 "Try again after reducing the number of events\n");
930 ui__warning("The sys_perf_event_open() syscall "
931 "returned with %d (%s). /bin/dmesg "
932 "may provide additional information.\n"
933 "No CONFIG_PERF_EVENTS=y kernel support "
934 "configured?\n", err, strerror(err));
939 if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
940 ui__warning("Failed to mmap with %d (%s)\n",
941 errno, strerror(errno));
952 static int perf_top__setup_sample_type(struct perf_top *top)
954 if (!top->sort_has_symbols) {
955 if (symbol_conf.use_callchain) {
956 ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
959 } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
960 if (callchain_register_param(&callchain_param) < 0) {
961 ui__warning("Can't register callchain params.\n");
969 static int __cmd_top(struct perf_top *top)
974 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
975 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
977 top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
978 if (top->session == NULL)
981 ret = perf_top__setup_sample_type(top);
985 if (top->target_tid || top->uid != UINT_MAX)
986 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
988 &top->session->host_machine);
990 perf_event__synthesize_threads(&top->tool, perf_event__process,
991 &top->session->host_machine);
992 perf_top__start_counters(top);
993 top->session->evlist = top->evlist;
994 perf_session__update_sample_type(top->session);
996 /* Wait for a minimal set of events before starting the snapshot */
997 poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
999 perf_top__mmap_read(top);
1001 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1002 display_thread), top)) {
1003 printf("Could not create display thread.\n");
1007 if (top->realtime_prio) {
1008 struct sched_param param;
1010 param.sched_priority = top->realtime_prio;
1011 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1012 printf("Could not set realtime priority.\n");
1018 u64 hits = top->samples;
1020 perf_top__mmap_read(top);
1022 if (hits == top->samples)
1023 ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1027 perf_session__delete(top->session);
1028 top->session = NULL;
1034 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1036 struct perf_top *top = (struct perf_top *)opt->value;
1044 top->dont_use_callchains = true;
1048 symbol_conf.use_callchain = true;
1053 tok = strtok((char *)arg, ",");
1057 /* get the output mode */
1058 if (!strncmp(tok, "graph", strlen(arg)))
1059 callchain_param.mode = CHAIN_GRAPH_ABS;
1061 else if (!strncmp(tok, "flat", strlen(arg)))
1062 callchain_param.mode = CHAIN_FLAT;
1064 else if (!strncmp(tok, "fractal", strlen(arg)))
1065 callchain_param.mode = CHAIN_GRAPH_REL;
1067 else if (!strncmp(tok, "none", strlen(arg))) {
1068 callchain_param.mode = CHAIN_NONE;
1069 symbol_conf.use_callchain = false;
1075 /* get the min percentage */
1076 tok = strtok(NULL, ",");
1080 callchain_param.min_percent = strtod(tok, &endptr);
1084 /* get the print limit */
1085 tok2 = strtok(NULL, ",");
1089 if (tok2[0] != 'c') {
1090 callchain_param.print_limit = strtod(tok2, &endptr);
1091 tok2 = strtok(NULL, ",");
1096 /* get the call chain order */
1097 if (!strcmp(tok2, "caller"))
1098 callchain_param.order = ORDER_CALLER;
1099 else if (!strcmp(tok2, "callee"))
1100 callchain_param.order = ORDER_CALLEE;
1104 if (callchain_register_param(&callchain_param) < 0) {
1105 fprintf(stderr, "Can't register callchain params\n");
1111 static const char * const top_usage[] = {
1112 "perf top [<options>]",
1116 int cmd_top(int argc, const char **argv, const char *prefix __used)
1118 struct perf_evsel *pos;
1119 int status = -ENOMEM;
1120 struct perf_top top = {
1124 .freq = 1000, /* 1 KHz */
1126 .sym_pcnt_filter = 5,
1128 char callchain_default_opt[] = "fractal,0.5,callee";
1129 const struct option options[] = {
1130 OPT_CALLBACK('e', "event", &top.evlist, "event",
1131 "event selector. use 'perf list' to list available events",
1132 parse_events_option),
1133 OPT_INTEGER('c', "count", &top.default_interval,
1134 "event period to sample"),
1135 OPT_STRING('p', "pid", &top.target_pid, "pid",
1136 "profile events on existing process id"),
1137 OPT_STRING('t', "tid", &top.target_tid, "tid",
1138 "profile events on existing thread id"),
1139 OPT_BOOLEAN('a', "all-cpus", &top.system_wide,
1140 "system-wide collection from all CPUs"),
1141 OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1142 "list of cpus to monitor"),
1143 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1144 "file", "vmlinux pathname"),
1145 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1146 "hide kernel symbols"),
1147 OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1148 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1149 "collect data with this RT SCHED_FIFO priority"),
1150 OPT_INTEGER('d', "delay", &top.delay_secs,
1151 "number of seconds to delay between refreshes"),
1152 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1153 "dump the symbol table used for profiling"),
1154 OPT_INTEGER('f', "count-filter", &top.count_filter,
1155 "only display functions with more events than this"),
1156 OPT_BOOLEAN('g', "group", &top.group,
1157 "put the counters into a counter group"),
1158 OPT_BOOLEAN('i', "inherit", &top.inherit,
1159 "child tasks inherit counters"),
1160 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1161 "symbol to annotate"),
1162 OPT_BOOLEAN('z', "zero", &top.zero,
1163 "zero history across updates"),
1164 OPT_INTEGER('F', "freq", &top.freq,
1165 "profile at this frequency"),
1166 OPT_INTEGER('E', "entries", &top.print_entries,
1167 "display this many functions"),
1168 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1169 "hide user symbols"),
1170 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1171 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1172 OPT_INCR('v', "verbose", &verbose,
1173 "be more verbose (show counter open errors, etc)"),
1174 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1175 "sort by key(s): pid, comm, dso, symbol, parent"),
1176 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1177 "Show a column with the number of samples"),
1178 OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1179 "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1180 "Default: fractal,0.5,callee", &parse_callchain_opt,
1181 callchain_default_opt),
1182 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1183 "Show a column with the sum of periods"),
1184 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1185 "only consider symbols in these dsos"),
1186 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1187 "only consider symbols in these comms"),
1188 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1189 "only consider these symbols"),
1190 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1191 "Interleave source code with assembly code (default)"),
1192 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1193 "Display raw encoding of assembly instructions (default)"),
1194 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1195 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1196 OPT_STRING('u', "uid", &top.uid_str, "user", "user to profile"),
1200 top.evlist = perf_evlist__new(NULL, NULL);
1201 if (top.evlist == NULL)
1204 symbol_conf.exclude_other = false;
1206 argc = parse_options(argc, argv, options, top_usage, 0);
1208 usage_with_options(top_usage, options);
1210 if (sort_order == default_sort_order)
1211 sort_order = "dso,symbol";
1213 setup_sorting(top_usage, options);
1217 else if (top.use_tui)
1220 setup_browser(false);
1222 top.uid = parse_target_uid(top.uid_str, top.target_tid, top.target_pid);
1223 if (top.uid_str != NULL && top.uid == UINT_MAX - 1)
1224 goto out_delete_evlist;
1226 /* CPU and PID are mutually exclusive */
1227 if (top.target_tid && top.cpu_list) {
1228 printf("WARNING: PID switch overriding CPU\n");
1230 top.cpu_list = NULL;
1234 top.target_tid = top.target_pid;
1236 if (perf_evlist__create_maps(top.evlist, top.target_pid,
1237 top.target_tid, top.uid, top.cpu_list) < 0)
1238 usage_with_options(top_usage, options);
1240 if (!top.evlist->nr_entries &&
1241 perf_evlist__add_default(top.evlist) < 0) {
1242 pr_err("Not enough memory for event selector list\n");
1246 symbol_conf.nr_events = top.evlist->nr_entries;
1248 if (top.delay_secs < 1)
1252 * User specified count overrides default frequency.
1254 if (top.default_interval)
1256 else if (top.freq) {
1257 top.default_interval = top.freq;
1259 fprintf(stderr, "frequency and count are zero, aborting\n");
1263 list_for_each_entry(pos, &top.evlist->entries, node) {
1265 * Fill in the ones not specifically initialized via -c:
1267 if (!pos->attr.sample_period)
1268 pos->attr.sample_period = top.default_interval;
1271 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1273 symbol_conf.priv_size = sizeof(struct annotation);
1275 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1276 if (symbol__init() < 0)
1279 sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1280 sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1281 sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1284 * Avoid annotation data structures overhead when symbols aren't on the
1287 top.sort_has_symbols = sort_sym.list.next != NULL;
1289 get_term_dimensions(&top.winsize);
1290 if (top.print_entries == 0) {
1291 struct sigaction act = {
1292 .sa_sigaction = perf_top__sig_winch,
1293 .sa_flags = SA_SIGINFO,
1295 perf_top__update_print_entries(&top);
1296 sigaction(SIGWINCH, &act, NULL);
1299 status = __cmd_top(&top);
1302 perf_evlist__delete(top.evlist);