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/config.h"
26 #include "util/color.h"
27 #include "util/drv_configs.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/event.h"
31 #include "util/machine.h"
32 #include "util/session.h"
33 #include "util/symbol.h"
34 #include "util/thread.h"
35 #include "util/thread_map.h"
37 #include <linux/rbtree.h>
38 #include <subcmd/parse-options.h>
39 #include "util/parse-events.h"
40 #include "util/cpumap.h"
41 #include "util/xyarray.h"
42 #include "util/sort.h"
43 #include "util/term.h"
44 #include "util/intlist.h"
45 #include "util/parse-branch-options.h"
46 #include "arch/common.h"
48 #include "util/debug.h"
64 #include <sys/syscall.h>
65 #include <sys/ioctl.h>
67 #include <sys/prctl.h>
70 #include <sys/utsname.h>
73 #include <linux/stringify.h>
74 #include <linux/time64.h>
75 #include <linux/types.h>
77 #include "sane_ctype.h"
79 static volatile int done;
81 #define HEADER_LINE_NR 5
83 static void perf_top__update_print_entries(struct perf_top *top)
85 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
88 static void perf_top__sig_winch(int sig __maybe_unused,
89 siginfo_t *info __maybe_unused, void *arg)
91 struct perf_top *top = arg;
93 get_term_dimensions(&top->winsize);
94 perf_top__update_print_entries(top);
97 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
100 struct annotation *notes;
104 if (!he || !he->ms.sym)
111 * We can't annotate with just /proc/kallsyms
113 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
114 !dso__is_kcore(map->dso)) {
115 pr_err("Can't annotate %s: No vmlinux file was found in the "
116 "path\n", sym->name);
121 notes = symbol__annotation(sym);
122 if (notes->src != NULL) {
123 pthread_mutex_lock(¬es->lock);
127 pthread_mutex_lock(¬es->lock);
129 if (symbol__alloc_hist(sym) < 0) {
130 pthread_mutex_unlock(¬es->lock);
131 pr_err("Not enough memory for annotating '%s' symbol!\n",
137 err = symbol__disassemble(sym, map, NULL, 0, NULL, NULL);
140 top->sym_filter_entry = he;
143 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
144 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
147 pthread_mutex_unlock(¬es->lock);
151 static void __zero_source_counters(struct hist_entry *he)
153 struct symbol *sym = he->ms.sym;
154 symbol__annotate_zero_histograms(sym);
157 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
160 int err = uname(&uts);
162 ui__warning("Out of bounds address found:\n\n"
163 "Addr: %" PRIx64 "\n"
165 "Map: %" PRIx64 "-%" PRIx64 "\n"
166 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
170 "Not all samples will be on the annotation output.\n\n"
171 "Please report to linux-kernel@vger.kernel.org\n",
172 ip, map->dso->long_name, dso__symtab_origin(map->dso),
173 map->start, map->end, sym->start, sym->end,
174 sym->binding == STB_GLOBAL ? 'g' :
175 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
176 err ? "[unknown]" : uts.machine,
177 err ? "[unknown]" : uts.release, perf_version_string);
178 if (use_browser <= 0)
181 map->erange_warned = true;
184 static void perf_top__record_precise_ip(struct perf_top *top,
185 struct hist_entry *he,
186 struct perf_sample *sample,
189 struct annotation *notes;
190 struct symbol *sym = he->ms.sym;
193 if (sym == NULL || (use_browser == 0 &&
194 (top->sym_filter_entry == NULL ||
195 top->sym_filter_entry->ms.sym != sym)))
198 notes = symbol__annotation(sym);
200 if (pthread_mutex_trylock(¬es->lock))
203 err = hist_entry__inc_addr_samples(he, sample, counter, ip);
205 pthread_mutex_unlock(¬es->lock);
209 * This function is now called with he->hists->lock held.
210 * Release it before going to sleep.
212 pthread_mutex_unlock(&he->hists->lock);
214 if (err == -ERANGE && !he->ms.map->erange_warned)
215 ui__warn_map_erange(he->ms.map, sym, ip);
216 else if (err == -ENOMEM) {
217 pr_err("Not enough memory for annotating '%s' symbol!\n",
222 pthread_mutex_lock(&he->hists->lock);
226 static void perf_top__show_details(struct perf_top *top)
228 struct hist_entry *he = top->sym_filter_entry;
229 struct annotation *notes;
230 struct symbol *symbol;
237 notes = symbol__annotation(symbol);
239 pthread_mutex_lock(¬es->lock);
241 if (notes->src == NULL)
244 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
245 printf(" Events Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
247 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
248 0, top->sym_pcnt_filter, top->print_entries, 4);
250 if (top->evlist->enabled) {
252 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
254 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
257 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
259 pthread_mutex_unlock(¬es->lock);
262 static void perf_top__print_sym_table(struct perf_top *top)
266 const int win_width = top->winsize.ws_col - 1;
267 struct perf_evsel *evsel = top->sym_evsel;
268 struct hists *hists = evsel__hists(evsel);
272 perf_top__header_snprintf(top, bf, sizeof(bf));
275 perf_top__reset_sample_counters(top);
277 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
279 if (hists->stats.nr_lost_warned !=
280 hists->stats.nr_events[PERF_RECORD_LOST]) {
281 hists->stats.nr_lost_warned =
282 hists->stats.nr_events[PERF_RECORD_LOST];
283 color_fprintf(stdout, PERF_COLOR_RED,
284 "WARNING: LOST %d chunks, Check IO/CPU overload",
285 hists->stats.nr_lost_warned);
289 if (top->sym_filter_entry) {
290 perf_top__show_details(top);
294 if (top->evlist->enabled) {
296 hists__delete_entries(hists);
298 hists__decay_entries(hists, top->hide_user_symbols,
299 top->hide_kernel_symbols);
303 hists__collapse_resort(hists, NULL);
304 perf_evsel__output_resort(evsel, NULL);
306 hists__output_recalc_col_len(hists, top->print_entries - printed);
308 hists__fprintf(hists, false, top->print_entries - printed, win_width,
309 top->min_percent, stdout, symbol_conf.use_callchain);
312 static void prompt_integer(int *target, const char *msg)
314 char *buf = malloc(0), *p;
318 fprintf(stdout, "\n%s: ", msg);
319 if (getline(&buf, &dummy, stdin) < 0)
322 p = strchr(buf, '\n');
332 tmp = strtoul(buf, NULL, 10);
338 static void prompt_percent(int *target, const char *msg)
342 prompt_integer(&tmp, msg);
343 if (tmp >= 0 && tmp <= 100)
347 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
349 char *buf = malloc(0), *p;
350 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
351 struct hists *hists = evsel__hists(top->sym_evsel);
352 struct rb_node *next;
355 /* zero counters of active symbol */
357 __zero_source_counters(syme);
358 top->sym_filter_entry = NULL;
361 fprintf(stdout, "\n%s: ", msg);
362 if (getline(&buf, &dummy, stdin) < 0)
365 p = strchr(buf, '\n');
369 next = rb_first(&hists->entries);
371 n = rb_entry(next, struct hist_entry, rb_node);
372 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
376 next = rb_next(&n->rb_node);
380 fprintf(stderr, "Sorry, %s is not active.\n", buf);
383 perf_top__parse_source(top, found);
389 static void perf_top__print_mapped_keys(struct perf_top *top)
393 if (top->sym_filter_entry) {
394 struct symbol *sym = top->sym_filter_entry->ms.sym;
398 fprintf(stdout, "\nMapped keys:\n");
399 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
400 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
402 if (top->evlist->nr_entries > 1)
403 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
405 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
407 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
408 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
409 fprintf(stdout, "\t[S] stop annotation.\n");
412 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
413 top->hide_kernel_symbols ? "yes" : "no");
415 "\t[U] hide user symbols. \t(%s)\n",
416 top->hide_user_symbols ? "yes" : "no");
417 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
418 fprintf(stdout, "\t[qQ] quit.\n");
421 static int perf_top__key_mapped(struct perf_top *top, int c)
437 return top->evlist->nr_entries > 1 ? 1 : 0;
445 static bool perf_top__handle_keypress(struct perf_top *top, int c)
449 if (!perf_top__key_mapped(top, c)) {
450 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
453 perf_top__print_mapped_keys(top);
454 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
457 set_term_quiet_input(&save);
459 poll(&stdin_poll, 1, -1);
462 tcsetattr(0, TCSAFLUSH, &save);
463 if (!perf_top__key_mapped(top, c))
469 prompt_integer(&top->delay_secs, "Enter display delay");
470 if (top->delay_secs < 1)
474 prompt_integer(&top->print_entries, "Enter display entries (lines)");
475 if (top->print_entries == 0) {
476 struct sigaction act = {
477 .sa_sigaction = perf_top__sig_winch,
478 .sa_flags = SA_SIGINFO,
480 perf_top__sig_winch(SIGWINCH, NULL, top);
481 sigaction(SIGWINCH, &act, NULL);
483 signal(SIGWINCH, SIG_DFL);
487 if (top->evlist->nr_entries > 1) {
488 /* Select 0 as the default event: */
491 fprintf(stderr, "\nAvailable events:");
493 evlist__for_each_entry(top->evlist, top->sym_evsel)
494 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
496 prompt_integer(&counter, "Enter details event counter");
498 if (counter >= top->evlist->nr_entries) {
499 top->sym_evsel = perf_evlist__first(top->evlist);
500 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
504 evlist__for_each_entry(top->evlist, top->sym_evsel)
505 if (top->sym_evsel->idx == counter)
508 top->sym_evsel = perf_evlist__first(top->evlist);
511 prompt_integer(&top->count_filter, "Enter display event count filter");
514 prompt_percent(&top->sym_pcnt_filter,
515 "Enter details display event filter (percent)");
518 top->hide_kernel_symbols = !top->hide_kernel_symbols;
522 printf("exiting.\n");
523 if (top->dump_symtab)
524 perf_session__fprintf_dsos(top->session, stderr);
528 perf_top__prompt_symbol(top, "Enter details symbol");
531 if (!top->sym_filter_entry)
534 struct hist_entry *syme = top->sym_filter_entry;
536 top->sym_filter_entry = NULL;
537 __zero_source_counters(syme);
541 top->hide_user_symbols = !top->hide_user_symbols;
544 top->zero = !top->zero;
553 static void perf_top__sort_new_samples(void *arg)
555 struct perf_top *t = arg;
556 struct perf_evsel *evsel = t->sym_evsel;
559 perf_top__reset_sample_counters(t);
561 if (t->evlist->selected != NULL)
562 t->sym_evsel = t->evlist->selected;
564 hists = evsel__hists(evsel);
566 if (t->evlist->enabled) {
568 hists__delete_entries(hists);
570 hists__decay_entries(hists, t->hide_user_symbols,
571 t->hide_kernel_symbols);
575 hists__collapse_resort(hists, NULL);
576 perf_evsel__output_resort(evsel, NULL);
579 static void *display_thread_tui(void *arg)
581 struct perf_evsel *pos;
582 struct perf_top *top = arg;
583 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
584 struct hist_browser_timer hbt = {
585 .timer = perf_top__sort_new_samples,
587 .refresh = top->delay_secs,
590 /* In order to read symbols from other namespaces perf to needs to call
591 * setns(2). This isn't permitted if the struct_fs has multiple users.
592 * unshare(2) the fs so that we may continue to setns into namespaces
593 * that we're observing.
597 perf_top__sort_new_samples(top);
600 * Initialize the uid_filter_str, in the future the TUI will allow
601 * Zooming in/out UIDs. For now juse use whatever the user passed
604 evlist__for_each_entry(top->evlist, pos) {
605 struct hists *hists = evsel__hists(pos);
606 hists->uid_filter_str = top->record_opts.target.uid_str;
609 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
611 &top->session->header.env);
617 static void display_sig(int sig __maybe_unused)
622 static void display_setup_sig(void)
624 signal(SIGSEGV, sighandler_dump_stack);
625 signal(SIGFPE, sighandler_dump_stack);
626 signal(SIGINT, display_sig);
627 signal(SIGQUIT, display_sig);
628 signal(SIGTERM, display_sig);
631 static void *display_thread(void *arg)
633 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
635 struct perf_top *top = arg;
638 /* In order to read symbols from other namespaces perf to needs to call
639 * setns(2). This isn't permitted if the struct_fs has multiple users.
640 * unshare(2) the fs so that we may continue to setns into namespaces
641 * that we're observing.
646 pthread__unblock_sigwinch();
648 delay_msecs = top->delay_secs * MSEC_PER_SEC;
649 set_term_quiet_input(&save);
654 perf_top__print_sym_table(top);
656 * Either timeout expired or we got an EINTR due to SIGWINCH,
657 * refresh screen in both cases.
659 switch (poll(&stdin_poll, 1, delay_msecs)) {
668 tcsetattr(0, TCSAFLUSH, &save);
670 if (perf_top__handle_keypress(top, c))
676 tcsetattr(0, TCSAFLUSH, &save);
680 static int hist_iter__top_callback(struct hist_entry_iter *iter,
681 struct addr_location *al, bool single,
684 struct perf_top *top = arg;
685 struct hist_entry *he = iter->he;
686 struct perf_evsel *evsel = iter->evsel;
688 if (perf_hpp_list.sym && single)
689 perf_top__record_precise_ip(top, he, iter->sample, evsel->idx, al->addr);
691 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
692 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
696 static void perf_event__process_sample(struct perf_tool *tool,
697 const union perf_event *event,
698 struct perf_evsel *evsel,
699 struct perf_sample *sample,
700 struct machine *machine)
702 struct perf_top *top = container_of(tool, struct perf_top, tool);
703 struct addr_location al;
706 if (!machine && perf_guest) {
707 static struct intlist *seen;
710 seen = intlist__new(NULL);
712 if (!intlist__has_entry(seen, sample->pid)) {
713 pr_err("Can't find guest [%d]'s kernel information\n",
715 intlist__add(seen, sample->pid);
721 pr_err("%u unprocessable samples recorded.\r",
722 top->session->evlist->stats.nr_unprocessable_samples++);
726 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
727 top->exact_samples++;
729 if (machine__resolve(machine, &al, sample) < 0)
732 if (!machine->kptr_restrict_warned &&
733 symbol_conf.kptr_restrict &&
734 al.cpumode == PERF_RECORD_MISC_KERNEL) {
736 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
737 "Check /proc/sys/kernel/kptr_restrict.\n\n"
738 "Kernel%s samples will not be resolved.\n",
739 al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
741 if (use_browser <= 0)
743 machine->kptr_restrict_warned = true;
746 if (al.sym == NULL) {
747 const char *msg = "Kernel samples will not be resolved.\n";
749 * As we do lazy loading of symtabs we only will know if the
750 * specified vmlinux file is invalid when we actually have a
751 * hit in kernel space and then try to load it. So if we get
752 * here and there are _no_ symbols in the DSO backing the
753 * kernel map, bail out.
755 * We may never get here, for instance, if we use -K/
756 * --hide-kernel-symbols, even if the user specifies an
757 * invalid --vmlinux ;-)
759 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
760 al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
761 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
762 if (symbol_conf.vmlinux_name) {
764 dso__strerror_load(al.map->dso, serr, sizeof(serr));
765 ui__warning("The %s file can't be used: %s\n%s",
766 symbol_conf.vmlinux_name, serr, msg);
768 ui__warning("A vmlinux file was not found.\n%s",
772 if (use_browser <= 0)
774 top->vmlinux_warned = true;
778 if (al.sym == NULL || !al.sym->idle) {
779 struct hists *hists = evsel__hists(evsel);
780 struct hist_entry_iter iter = {
783 .add_entry_cb = hist_iter__top_callback,
786 if (symbol_conf.cumulate_callchain)
787 iter.ops = &hist_iter_cumulative;
789 iter.ops = &hist_iter_normal;
791 pthread_mutex_lock(&hists->lock);
793 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
795 pr_err("Problem incrementing symbol period, skipping event\n");
797 pthread_mutex_unlock(&hists->lock);
800 addr_location__put(&al);
803 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
805 struct perf_sample sample;
806 struct perf_evsel *evsel;
807 struct perf_session *session = top->session;
808 union perf_event *event;
809 struct machine *machine;
812 while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
813 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
815 pr_err("Can't parse sample, err = %d\n", ret);
819 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
820 assert(evsel != NULL);
822 if (event->header.type == PERF_RECORD_SAMPLE)
825 switch (sample.cpumode) {
826 case PERF_RECORD_MISC_USER:
828 if (top->hide_user_symbols)
830 machine = &session->machines.host;
832 case PERF_RECORD_MISC_KERNEL:
833 ++top->kernel_samples;
834 if (top->hide_kernel_symbols)
836 machine = &session->machines.host;
838 case PERF_RECORD_MISC_GUEST_KERNEL:
839 ++top->guest_kernel_samples;
840 machine = perf_session__find_machine(session,
843 case PERF_RECORD_MISC_GUEST_USER:
844 ++top->guest_us_samples;
846 * TODO: we don't process guest user from host side
847 * except simple counting.
851 if (event->header.type == PERF_RECORD_SAMPLE)
853 machine = &session->machines.host;
858 if (event->header.type == PERF_RECORD_SAMPLE) {
859 perf_event__process_sample(&top->tool, event, evsel,
861 } else if (event->header.type < PERF_RECORD_MAX) {
862 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
863 machine__process_event(machine, event, &sample);
865 ++session->evlist->stats.nr_unknown_events;
867 perf_evlist__mmap_consume(top->evlist, idx);
871 static void perf_top__mmap_read(struct perf_top *top)
875 for (i = 0; i < top->evlist->nr_mmaps; i++)
876 perf_top__mmap_read_idx(top, i);
879 static int perf_top__start_counters(struct perf_top *top)
882 struct perf_evsel *counter;
883 struct perf_evlist *evlist = top->evlist;
884 struct record_opts *opts = &top->record_opts;
886 perf_evlist__config(evlist, opts, &callchain_param);
888 evlist__for_each_entry(evlist, counter) {
890 if (perf_evsel__open(counter, top->evlist->cpus,
891 top->evlist->threads) < 0) {
892 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
894 ui__warning("%s\n", msg);
898 perf_evsel__open_strerror(counter, &opts->target,
899 errno, msg, sizeof(msg));
900 ui__error("%s\n", msg);
905 if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
906 ui__error("Failed to mmap with %d (%s)\n",
907 errno, str_error_r(errno, msg, sizeof(msg)));
917 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
919 if (!perf_hpp_list.sym) {
920 if (callchain->enabled) {
921 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
924 } else if (callchain->mode != CHAIN_NONE) {
925 if (callchain_register_param(callchain) < 0) {
926 ui__error("Can't register callchain params.\n");
934 static int __cmd_top(struct perf_top *top)
937 struct perf_evsel *pos;
938 struct perf_evsel_config_term *err_term;
939 struct perf_evlist *evlist = top->evlist;
940 struct record_opts *opts = &top->record_opts;
944 top->session = perf_session__new(NULL, false, NULL);
945 if (top->session == NULL)
949 ret = perf_env__lookup_objdump(&top->session->header.env);
954 ret = callchain_param__setup_sample_type(&callchain_param);
958 if (perf_session__register_idle_thread(top->session) < 0)
961 if (top->nr_threads_synthesize > 1)
962 perf_set_multithreaded();
964 machine__synthesize_threads(&top->session->machines.host, &opts->target,
965 top->evlist->threads, false,
966 opts->proc_map_timeout,
967 top->nr_threads_synthesize);
969 if (top->nr_threads_synthesize > 1)
970 perf_set_singlethreaded();
972 if (perf_hpp_list.socket) {
973 ret = perf_env__read_cpu_topology_map(&perf_env);
975 goto out_err_cpu_topo;
978 ret = perf_top__start_counters(top);
982 ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
984 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
985 err_term->val.drv_cfg, perf_evsel__name(pos), errno,
986 str_error_r(errno, msg, sizeof(msg)));
990 top->session->evlist = top->evlist;
991 perf_session__set_id_hdr_size(top->session);
994 * When perf is starting the traced process, all the events (apart from
995 * group members) have enable_on_exec=1 set, so don't spoil it by
996 * prematurely enabling them.
998 * XXX 'top' still doesn't start workloads like record, trace, but should,
999 * so leave the check here.
1001 if (!target__none(&opts->target))
1002 perf_evlist__enable(top->evlist);
1004 /* Wait for a minimal set of events before starting the snapshot */
1005 perf_evlist__poll(top->evlist, 100);
1007 perf_top__mmap_read(top);
1010 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1011 display_thread), top)) {
1012 ui__error("Could not create display thread.\n");
1016 if (top->realtime_prio) {
1017 struct sched_param param;
1019 param.sched_priority = top->realtime_prio;
1020 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1021 ui__error("Could not set realtime priority.\n");
1027 u64 hits = top->samples;
1029 perf_top__mmap_read(top);
1031 if (hits == top->samples)
1032 ret = perf_evlist__poll(top->evlist, 100);
1037 pthread_join(thread, NULL);
1039 perf_session__delete(top->session);
1040 top->session = NULL;
1045 char errbuf[BUFSIZ];
1046 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1048 ui__error("Could not read the CPU topology map: %s\n", err);
1054 callchain_opt(const struct option *opt, const char *arg, int unset)
1056 symbol_conf.use_callchain = true;
1057 return record_callchain_opt(opt, arg, unset);
1061 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1063 struct callchain_param *callchain = opt->value;
1065 callchain->enabled = !unset;
1066 callchain->record_mode = CALLCHAIN_FP;
1072 symbol_conf.use_callchain = false;
1073 callchain->record_mode = CALLCHAIN_NONE;
1077 return parse_callchain_top_opt(arg);
1080 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1082 if (!strcmp(var, "top.call-graph"))
1083 var = "call-graph.record-mode"; /* fall-through */
1084 if (!strcmp(var, "top.children")) {
1085 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1093 parse_percent_limit(const struct option *opt, const char *arg,
1094 int unset __maybe_unused)
1096 struct perf_top *top = opt->value;
1098 top->min_percent = strtof(arg, NULL);
1102 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1103 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1105 int cmd_top(int argc, const char **argv)
1107 char errbuf[BUFSIZ];
1108 struct perf_top top = {
1112 .mmap_pages = UINT_MAX,
1113 .user_freq = UINT_MAX,
1114 .user_interval = ULLONG_MAX,
1115 .freq = 4000, /* 4 KHz */
1119 .proc_map_timeout = 500,
1121 .max_stack = sysctl_perf_event_max_stack,
1122 .sym_pcnt_filter = 5,
1123 .nr_threads_synthesize = UINT_MAX,
1125 struct record_opts *opts = &top.record_opts;
1126 struct target *target = &opts->target;
1127 const struct option options[] = {
1128 OPT_CALLBACK('e', "event", &top.evlist, "event",
1129 "event selector. use 'perf list' to list available events",
1130 parse_events_option),
1131 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1132 OPT_STRING('p', "pid", &target->pid, "pid",
1133 "profile events on existing process id"),
1134 OPT_STRING('t', "tid", &target->tid, "tid",
1135 "profile events on existing thread id"),
1136 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1137 "system-wide collection from all CPUs"),
1138 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1139 "list of cpus to monitor"),
1140 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1141 "file", "vmlinux pathname"),
1142 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1143 "don't load vmlinux even if found"),
1144 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1145 "hide kernel symbols"),
1146 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1147 "number of mmap data pages",
1148 perf_evlist__parse_mmap_pages),
1149 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1150 "collect data with this RT SCHED_FIFO priority"),
1151 OPT_INTEGER('d', "delay", &top.delay_secs,
1152 "number of seconds to delay between refreshes"),
1153 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1154 "dump the symbol table used for profiling"),
1155 OPT_INTEGER('f', "count-filter", &top.count_filter,
1156 "only display functions with more events than this"),
1157 OPT_BOOLEAN(0, "group", &opts->group,
1158 "put the counters into a counter group"),
1159 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1160 "child tasks do not inherit counters"),
1161 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1162 "symbol to annotate"),
1163 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1164 OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1165 OPT_INTEGER('E', "entries", &top.print_entries,
1166 "display this many functions"),
1167 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1168 "hide user symbols"),
1169 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1170 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1171 OPT_INCR('v', "verbose", &verbose,
1172 "be more verbose (show counter open errors, etc)"),
1173 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1174 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1175 " Please refer the man page for the complete list."),
1176 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1177 "output field(s): overhead, period, sample plus all of sort keys"),
1178 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1179 "Show a column with the number of samples"),
1180 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1181 NULL, "enables call-graph recording and display",
1183 OPT_CALLBACK(0, "call-graph", &callchain_param,
1184 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1185 top_callchain_help, &parse_callchain_opt),
1186 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1187 "Accumulate callchains of children and show total overhead as well"),
1188 OPT_INTEGER(0, "max-stack", &top.max_stack,
1189 "Set the maximum stack depth when parsing the callchain. "
1190 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1191 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1192 "ignore callees of these functions in call graphs",
1193 report_parse_ignore_callees_opt),
1194 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1195 "Show a column with the sum of periods"),
1196 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1197 "only consider symbols in these dsos"),
1198 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1199 "only consider symbols in these comms"),
1200 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1201 "only consider these symbols"),
1202 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1203 "Interleave source code with assembly code (default)"),
1204 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1205 "Display raw encoding of assembly instructions (default)"),
1206 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1207 "Enable kernel symbol demangling"),
1208 OPT_STRING(0, "objdump", &objdump_path, "path",
1209 "objdump binary to use for disassembly and annotations"),
1210 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1211 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1212 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1213 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1214 "Don't show entries under that percent", parse_percent_limit),
1215 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1216 "How to display percentage of filtered entries", parse_filter_percentage),
1217 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1219 "don't try to adjust column width, use these fixed values"),
1220 OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1221 "per thread proc mmap processing timeout in ms"),
1222 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1223 "branch any", "sample any taken branches",
1224 parse_branch_stack),
1225 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1226 "branch filter mask", "branch stack filter modes",
1227 parse_branch_stack),
1228 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1229 "Show raw trace event output (do not use print fmt or plugins)"),
1230 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1231 "Show entries in a hierarchy"),
1232 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1233 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1234 "number of thread to run event synthesize"),
1237 const char * const top_usage[] = {
1238 "perf top [<options>]",
1241 int status = hists__init();
1246 top.evlist = perf_evlist__new();
1247 if (top.evlist == NULL)
1250 status = perf_config(perf_top_config, &top);
1254 argc = parse_options(argc, argv, options, top_usage, 0);
1256 usage_with_options(top_usage, options);
1258 if (!top.evlist->nr_entries &&
1259 perf_evlist__add_default(top.evlist) < 0) {
1260 pr_err("Not enough memory for event selector list\n");
1261 goto out_delete_evlist;
1264 if (symbol_conf.report_hierarchy) {
1265 /* disable incompatible options */
1266 symbol_conf.event_group = false;
1267 symbol_conf.cumulate_callchain = false;
1270 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1271 parse_options_usage(top_usage, options, "fields", 0);
1272 parse_options_usage(NULL, options, "hierarchy", 0);
1273 goto out_delete_evlist;
1277 sort__mode = SORT_MODE__TOP;
1278 /* display thread wants entries to be collapsed in a different tree */
1279 perf_hpp_list.need_collapse = 1;
1283 else if (top.use_tui)
1286 setup_browser(false);
1288 if (setup_sorting(top.evlist) < 0) {
1290 parse_options_usage(top_usage, options, "s", 1);
1292 parse_options_usage(sort_order ? NULL : top_usage,
1293 options, "fields", 0);
1294 goto out_delete_evlist;
1297 status = target__validate(target);
1299 target__strerror(target, status, errbuf, BUFSIZ);
1300 ui__warning("%s\n", errbuf);
1303 status = target__parse_uid(target);
1305 int saved_errno = errno;
1307 target__strerror(target, status, errbuf, BUFSIZ);
1308 ui__error("%s\n", errbuf);
1310 status = -saved_errno;
1311 goto out_delete_evlist;
1314 if (target__none(target))
1315 target->system_wide = true;
1317 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1318 ui__error("Couldn't create thread/CPU maps: %s\n",
1319 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1320 goto out_delete_evlist;
1323 symbol_conf.nr_events = top.evlist->nr_entries;
1325 if (top.delay_secs < 1)
1328 if (record_opts__config(opts)) {
1330 goto out_delete_evlist;
1333 top.sym_evsel = perf_evlist__first(top.evlist);
1335 if (!callchain_param.enabled) {
1336 symbol_conf.cumulate_callchain = false;
1337 perf_hpp__cancel_cumulate();
1340 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1341 callchain_param.order = ORDER_CALLER;
1343 status = symbol__annotation_init();
1345 goto out_delete_evlist;
1347 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1348 if (symbol__init(NULL) < 0)
1351 sort__setup_elide(stdout);
1353 get_term_dimensions(&top.winsize);
1354 if (top.print_entries == 0) {
1355 struct sigaction act = {
1356 .sa_sigaction = perf_top__sig_winch,
1357 .sa_flags = SA_SIGINFO,
1359 perf_top__update_print_entries(&top);
1360 sigaction(SIGWINCH, &act, NULL);
1363 status = __cmd_top(&top);
1366 perf_evlist__delete(top.evlist);