1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin top command: Display a continuously updated profile of
6 * any workload, CPU or specific PID.
8 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
11 * Improvements and fixes by:
13 * Arjan van de Ven <arjan@linux.intel.com>
14 * Yanmin Zhang <yanmin.zhang@intel.com>
15 * Wu Fengguang <fengguang.wu@intel.com>
16 * Mike Galbraith <efault@gmx.de>
17 * Paul Mackerras <paulus@samba.org>
23 #include "util/annotate.h"
24 #include "util/bpf-event.h"
25 #include "util/config.h"
26 #include "util/color.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/event.h"
31 #include "util/machine.h"
33 #include "util/session.h"
34 #include "util/symbol.h"
36 #include "util/util.h"
37 #include <linux/rbtree.h>
38 #include <subcmd/parse-options.h>
39 #include "util/parse-events.h"
40 #include "util/callchain.h"
41 #include "util/cpumap.h"
42 #include "util/sort.h"
43 #include "util/string2.h"
44 #include "util/term.h"
45 #include "util/intlist.h"
46 #include "util/parse-branch-options.h"
47 #include "arch/common.h"
50 #include "util/debug.h"
51 #include "util/ordered-events.h"
67 #include <sys/syscall.h>
68 #include <sys/ioctl.h>
70 #include <sys/prctl.h>
73 #include <sys/utsname.h>
76 #include <linux/stringify.h>
77 #include <linux/time64.h>
78 #include <linux/types.h>
80 #include <linux/ctype.h>
82 static volatile int done;
83 static volatile int resize;
85 #define HEADER_LINE_NR 5
87 static void perf_top__update_print_entries(struct perf_top *top)
89 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
92 static void winch_sig(int sig __maybe_unused)
97 static void perf_top__resize(struct perf_top *top)
99 get_term_dimensions(&top->winsize);
100 perf_top__update_print_entries(top);
103 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
107 struct annotation *notes;
111 if (!he || !he->ms.sym)
114 evsel = hists_to_evsel(he->hists);
120 * We can't annotate with just /proc/kallsyms
122 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
123 !dso__is_kcore(map->dso)) {
124 pr_err("Can't annotate %s: No vmlinux file was found in the "
125 "path\n", sym->name);
130 notes = symbol__annotation(sym);
131 pthread_mutex_lock(¬es->lock);
133 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
134 pthread_mutex_unlock(¬es->lock);
135 pr_err("Not enough memory for annotating '%s' symbol!\n",
141 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
143 top->sym_filter_entry = he;
146 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
147 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
150 pthread_mutex_unlock(¬es->lock);
154 static void __zero_source_counters(struct hist_entry *he)
156 struct symbol *sym = he->ms.sym;
157 symbol__annotate_zero_histograms(sym);
160 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
163 int err = uname(&uts);
165 ui__warning("Out of bounds address found:\n\n"
166 "Addr: %" PRIx64 "\n"
168 "Map: %" PRIx64 "-%" PRIx64 "\n"
169 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
173 "Not all samples will be on the annotation output.\n\n"
174 "Please report to linux-kernel@vger.kernel.org\n",
175 ip, map->dso->long_name, dso__symtab_origin(map->dso),
176 map->start, map->end, sym->start, sym->end,
177 sym->binding == STB_GLOBAL ? 'g' :
178 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
179 err ? "[unknown]" : uts.machine,
180 err ? "[unknown]" : uts.release, perf_version_string);
181 if (use_browser <= 0)
184 map->erange_warned = true;
187 static void perf_top__record_precise_ip(struct perf_top *top,
188 struct hist_entry *he,
189 struct perf_sample *sample,
190 struct evsel *evsel, u64 ip)
192 struct annotation *notes;
193 struct symbol *sym = he->ms.sym;
196 if (sym == NULL || (use_browser == 0 &&
197 (top->sym_filter_entry == NULL ||
198 top->sym_filter_entry->ms.sym != sym)))
201 notes = symbol__annotation(sym);
203 if (pthread_mutex_trylock(¬es->lock))
206 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
208 pthread_mutex_unlock(¬es->lock);
212 * This function is now called with he->hists->lock held.
213 * Release it before going to sleep.
215 pthread_mutex_unlock(&he->hists->lock);
217 if (err == -ERANGE && !he->ms.map->erange_warned)
218 ui__warn_map_erange(he->ms.map, sym, ip);
219 else if (err == -ENOMEM) {
220 pr_err("Not enough memory for annotating '%s' symbol!\n",
225 pthread_mutex_lock(&he->hists->lock);
229 static void perf_top__show_details(struct perf_top *top)
231 struct hist_entry *he = top->sym_filter_entry;
233 struct annotation *notes;
234 struct symbol *symbol;
240 evsel = hists_to_evsel(he->hists);
243 notes = symbol__annotation(symbol);
245 pthread_mutex_lock(¬es->lock);
247 symbol__calc_percent(symbol, evsel);
249 if (notes->src == NULL)
252 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
253 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
255 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
257 if (top->evlist->enabled) {
259 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
261 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
264 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
266 pthread_mutex_unlock(¬es->lock);
269 static void perf_top__resort_hists(struct perf_top *t)
271 struct evlist *evlist = t->evlist;
274 evlist__for_each_entry(evlist, pos) {
275 struct hists *hists = evsel__hists(pos);
278 * unlink existing entries so that they can be linked
279 * in a correct order in hists__match() below.
281 hists__unlink(hists);
283 if (evlist->enabled) {
285 hists__delete_entries(hists);
287 hists__decay_entries(hists, t->hide_user_symbols,
288 t->hide_kernel_symbols);
292 hists__collapse_resort(hists, NULL);
294 /* Non-group events are considered as leader */
295 if (symbol_conf.event_group &&
296 !perf_evsel__is_group_leader(pos)) {
297 struct hists *leader_hists = evsel__hists(pos->leader);
299 hists__match(leader_hists, hists);
300 hists__link(leader_hists, hists);
304 evlist__for_each_entry(evlist, pos) {
305 perf_evsel__output_resort(pos, NULL);
309 static void perf_top__print_sym_table(struct perf_top *top)
313 const int win_width = top->winsize.ws_col - 1;
314 struct evsel *evsel = top->sym_evsel;
315 struct hists *hists = evsel__hists(evsel);
319 perf_top__header_snprintf(top, bf, sizeof(bf));
322 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
324 if (!top->record_opts.overwrite &&
325 (hists->stats.nr_lost_warned !=
326 hists->stats.nr_events[PERF_RECORD_LOST])) {
327 hists->stats.nr_lost_warned =
328 hists->stats.nr_events[PERF_RECORD_LOST];
329 color_fprintf(stdout, PERF_COLOR_RED,
330 "WARNING: LOST %d chunks, Check IO/CPU overload",
331 hists->stats.nr_lost_warned);
335 if (top->sym_filter_entry) {
336 perf_top__show_details(top);
340 perf_top__resort_hists(top);
342 hists__output_recalc_col_len(hists, top->print_entries - printed);
344 hists__fprintf(hists, false, top->print_entries - printed, win_width,
345 top->min_percent, stdout, !symbol_conf.use_callchain);
348 static void prompt_integer(int *target, const char *msg)
350 char *buf = malloc(0), *p;
354 fprintf(stdout, "\n%s: ", msg);
355 if (getline(&buf, &dummy, stdin) < 0)
358 p = strchr(buf, '\n');
368 tmp = strtoul(buf, NULL, 10);
374 static void prompt_percent(int *target, const char *msg)
378 prompt_integer(&tmp, msg);
379 if (tmp >= 0 && tmp <= 100)
383 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
385 char *buf = malloc(0), *p;
386 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
387 struct hists *hists = evsel__hists(top->sym_evsel);
388 struct rb_node *next;
391 /* zero counters of active symbol */
393 __zero_source_counters(syme);
394 top->sym_filter_entry = NULL;
397 fprintf(stdout, "\n%s: ", msg);
398 if (getline(&buf, &dummy, stdin) < 0)
401 p = strchr(buf, '\n');
405 next = rb_first_cached(&hists->entries);
407 n = rb_entry(next, struct hist_entry, rb_node);
408 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
412 next = rb_next(&n->rb_node);
416 fprintf(stderr, "Sorry, %s is not active.\n", buf);
419 perf_top__parse_source(top, found);
425 static void perf_top__print_mapped_keys(struct perf_top *top)
429 if (top->sym_filter_entry) {
430 struct symbol *sym = top->sym_filter_entry->ms.sym;
434 fprintf(stdout, "\nMapped keys:\n");
435 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
436 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
438 if (top->evlist->core.nr_entries > 1)
439 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
441 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
443 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
444 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
445 fprintf(stdout, "\t[S] stop annotation.\n");
448 "\t[K] hide kernel symbols. \t(%s)\n",
449 top->hide_kernel_symbols ? "yes" : "no");
451 "\t[U] hide user symbols. \t(%s)\n",
452 top->hide_user_symbols ? "yes" : "no");
453 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
454 fprintf(stdout, "\t[qQ] quit.\n");
457 static int perf_top__key_mapped(struct perf_top *top, int c)
473 return top->evlist->core.nr_entries > 1 ? 1 : 0;
481 static bool perf_top__handle_keypress(struct perf_top *top, int c)
485 if (!perf_top__key_mapped(top, c)) {
486 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
489 perf_top__print_mapped_keys(top);
490 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
493 set_term_quiet_input(&save);
495 poll(&stdin_poll, 1, -1);
498 tcsetattr(0, TCSAFLUSH, &save);
499 if (!perf_top__key_mapped(top, c))
505 prompt_integer(&top->delay_secs, "Enter display delay");
506 if (top->delay_secs < 1)
510 prompt_integer(&top->print_entries, "Enter display entries (lines)");
511 if (top->print_entries == 0) {
512 perf_top__resize(top);
513 signal(SIGWINCH, winch_sig);
515 signal(SIGWINCH, SIG_DFL);
519 if (top->evlist->core.nr_entries > 1) {
520 /* Select 0 as the default event: */
523 fprintf(stderr, "\nAvailable events:");
525 evlist__for_each_entry(top->evlist, top->sym_evsel)
526 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
528 prompt_integer(&counter, "Enter details event counter");
530 if (counter >= top->evlist->core.nr_entries) {
531 top->sym_evsel = perf_evlist__first(top->evlist);
532 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
536 evlist__for_each_entry(top->evlist, top->sym_evsel)
537 if (top->sym_evsel->idx == counter)
540 top->sym_evsel = perf_evlist__first(top->evlist);
543 prompt_integer(&top->count_filter, "Enter display event count filter");
546 prompt_percent(&top->annotation_opts.min_pcnt,
547 "Enter details display event filter (percent)");
550 top->hide_kernel_symbols = !top->hide_kernel_symbols;
554 printf("exiting.\n");
555 if (top->dump_symtab)
556 perf_session__fprintf_dsos(top->session, stderr);
560 perf_top__prompt_symbol(top, "Enter details symbol");
563 if (!top->sym_filter_entry)
566 struct hist_entry *syme = top->sym_filter_entry;
568 top->sym_filter_entry = NULL;
569 __zero_source_counters(syme);
573 top->hide_user_symbols = !top->hide_user_symbols;
576 top->zero = !top->zero;
585 static void perf_top__sort_new_samples(void *arg)
587 struct perf_top *t = arg;
589 if (t->evlist->selected != NULL)
590 t->sym_evsel = t->evlist->selected;
592 perf_top__resort_hists(t);
594 if (t->lost || t->drop)
595 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
598 static void stop_top(void)
604 static void *display_thread_tui(void *arg)
607 struct perf_top *top = arg;
608 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
609 struct hist_browser_timer hbt = {
610 .timer = perf_top__sort_new_samples,
612 .refresh = top->delay_secs,
615 /* In order to read symbols from other namespaces perf to needs to call
616 * setns(2). This isn't permitted if the struct_fs has multiple users.
617 * unshare(2) the fs so that we may continue to setns into namespaces
618 * that we're observing.
622 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
624 perf_top__sort_new_samples(top);
627 * Initialize the uid_filter_str, in the future the TUI will allow
628 * Zooming in/out UIDs. For now just use whatever the user passed
631 evlist__for_each_entry(top->evlist, pos) {
632 struct hists *hists = evsel__hists(pos);
633 hists->uid_filter_str = top->record_opts.target.uid_str;
636 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
638 &top->session->header.env,
639 !top->record_opts.overwrite,
640 &top->annotation_opts);
646 static void display_sig(int sig __maybe_unused)
651 static void display_setup_sig(void)
653 signal(SIGSEGV, sighandler_dump_stack);
654 signal(SIGFPE, sighandler_dump_stack);
655 signal(SIGINT, display_sig);
656 signal(SIGQUIT, display_sig);
657 signal(SIGTERM, display_sig);
660 static void *display_thread(void *arg)
662 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
664 struct perf_top *top = arg;
667 /* In order to read symbols from other namespaces perf to needs to call
668 * setns(2). This isn't permitted if the struct_fs has multiple users.
669 * unshare(2) the fs so that we may continue to setns into namespaces
670 * that we're observing.
674 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
677 pthread__unblock_sigwinch();
679 delay_msecs = top->delay_secs * MSEC_PER_SEC;
680 set_term_quiet_input(&save);
685 perf_top__print_sym_table(top);
687 * Either timeout expired or we got an EINTR due to SIGWINCH,
688 * refresh screen in both cases.
690 switch (poll(&stdin_poll, 1, delay_msecs)) {
699 tcsetattr(0, TCSAFLUSH, &save);
701 if (perf_top__handle_keypress(top, c))
707 tcsetattr(0, TCSAFLUSH, &save);
711 static int hist_iter__top_callback(struct hist_entry_iter *iter,
712 struct addr_location *al, bool single,
715 struct perf_top *top = arg;
716 struct hist_entry *he = iter->he;
717 struct evsel *evsel = iter->evsel;
719 if (perf_hpp_list.sym && single)
720 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
722 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
723 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
727 static void perf_event__process_sample(struct perf_tool *tool,
728 const union perf_event *event,
730 struct perf_sample *sample,
731 struct machine *machine)
733 struct perf_top *top = container_of(tool, struct perf_top, tool);
734 struct addr_location al;
737 if (!machine && perf_guest) {
738 static struct intlist *seen;
741 seen = intlist__new(NULL);
743 if (!intlist__has_entry(seen, sample->pid)) {
744 pr_err("Can't find guest [%d]'s kernel information\n",
746 intlist__add(seen, sample->pid);
752 pr_err("%u unprocessable samples recorded.\r",
753 top->session->evlist->stats.nr_unprocessable_samples++);
757 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
758 top->exact_samples++;
760 if (machine__resolve(machine, &al, sample) < 0)
763 if (!machine->kptr_restrict_warned &&
764 symbol_conf.kptr_restrict &&
765 al.cpumode == PERF_RECORD_MISC_KERNEL) {
766 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
768 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
769 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
770 "Kernel%s samples will not be resolved.\n",
771 al.map && map__has_symbols(al.map) ?
773 if (use_browser <= 0)
776 machine->kptr_restrict_warned = true;
779 if (al.sym == NULL && al.map != NULL) {
780 const char *msg = "Kernel samples will not be resolved.\n";
782 * As we do lazy loading of symtabs we only will know if the
783 * specified vmlinux file is invalid when we actually have a
784 * hit in kernel space and then try to load it. So if we get
785 * here and there are _no_ symbols in the DSO backing the
786 * kernel map, bail out.
788 * We may never get here, for instance, if we use -K/
789 * --hide-kernel-symbols, even if the user specifies an
790 * invalid --vmlinux ;-)
792 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
793 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
794 if (symbol_conf.vmlinux_name) {
796 dso__strerror_load(al.map->dso, serr, sizeof(serr));
797 ui__warning("The %s file can't be used: %s\n%s",
798 symbol_conf.vmlinux_name, serr, msg);
800 ui__warning("A vmlinux file was not found.\n%s",
804 if (use_browser <= 0)
806 top->vmlinux_warned = true;
810 if (al.sym == NULL || !al.sym->idle) {
811 struct hists *hists = evsel__hists(evsel);
812 struct hist_entry_iter iter = {
815 .add_entry_cb = hist_iter__top_callback,
818 if (symbol_conf.cumulate_callchain)
819 iter.ops = &hist_iter_cumulative;
821 iter.ops = &hist_iter_normal;
823 pthread_mutex_lock(&hists->lock);
825 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
827 pr_err("Problem incrementing symbol period, skipping event\n");
829 pthread_mutex_unlock(&hists->lock);
832 addr_location__put(&al);
836 perf_top__process_lost(struct perf_top *top, union perf_event *event,
839 struct hists *hists = evsel__hists(evsel);
841 top->lost += event->lost.lost;
842 top->lost_total += event->lost.lost;
843 hists->stats.total_lost += event->lost.lost;
847 perf_top__process_lost_samples(struct perf_top *top,
848 union perf_event *event,
851 struct hists *hists = evsel__hists(evsel);
853 top->lost += event->lost_samples.lost;
854 top->lost_total += event->lost_samples.lost;
855 hists->stats.total_lost_samples += event->lost_samples.lost;
858 static u64 last_timestamp;
860 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
862 struct record_opts *opts = &top->record_opts;
863 struct evlist *evlist = top->evlist;
864 struct perf_mmap *md;
865 union perf_event *event;
867 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
868 if (perf_mmap__read_init(md) < 0)
871 while ((event = perf_mmap__read_event(md)) != NULL) {
874 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
875 if (ret && ret != -1)
878 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
882 perf_mmap__consume(md);
884 if (top->qe.rotate) {
885 pthread_mutex_lock(&top->qe.mutex);
886 top->qe.rotate = false;
887 pthread_cond_signal(&top->qe.cond);
888 pthread_mutex_unlock(&top->qe.mutex);
892 perf_mmap__read_done(md);
895 static void perf_top__mmap_read(struct perf_top *top)
897 bool overwrite = top->record_opts.overwrite;
898 struct evlist *evlist = top->evlist;
902 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
904 for (i = 0; i < top->evlist->nr_mmaps; i++)
905 perf_top__mmap_read_idx(top, i);
908 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
909 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
914 * Check per-event overwrite term.
915 * perf top should support consistent term for all events.
916 * - All events don't have per-event term
917 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
918 * Nothing change, return 0.
919 * - All events have same per-event term
920 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
921 * Using the per-event setting to replace the opts->overwrite if
922 * they are different, then return 0.
923 * - Events have different per-event term
924 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
926 * - Some of the event set per-event term, but some not.
927 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
930 static int perf_top__overwrite_check(struct perf_top *top)
932 struct record_opts *opts = &top->record_opts;
933 struct evlist *evlist = top->evlist;
934 struct perf_evsel_config_term *term;
935 struct list_head *config_terms;
937 int set, overwrite = -1;
939 evlist__for_each_entry(evlist, evsel) {
941 config_terms = &evsel->config_terms;
942 list_for_each_entry(term, config_terms, list) {
943 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
944 set = term->val.overwrite ? 1 : 0;
947 /* no term for current and previous event (likely) */
948 if ((overwrite < 0) && (set < 0))
951 /* has term for both current and previous event, compare */
952 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
955 /* no term for current event but has term for previous one */
956 if ((overwrite >= 0) && (set < 0))
959 /* has term for current event */
960 if ((overwrite < 0) && (set >= 0)) {
961 /* if it's first event, set overwrite */
962 if (evsel == perf_evlist__first(evlist))
969 if ((overwrite >= 0) && (opts->overwrite != overwrite))
970 opts->overwrite = overwrite;
975 static int perf_top_overwrite_fallback(struct perf_top *top,
978 struct record_opts *opts = &top->record_opts;
979 struct evlist *evlist = top->evlist;
980 struct evsel *counter;
982 if (!opts->overwrite)
985 /* only fall back when first event fails */
986 if (evsel != perf_evlist__first(evlist))
989 evlist__for_each_entry(evlist, counter)
990 counter->core.attr.write_backward = false;
991 opts->overwrite = false;
992 pr_debug2("fall back to non-overwrite mode\n");
996 static int perf_top__start_counters(struct perf_top *top)
999 struct evsel *counter;
1000 struct evlist *evlist = top->evlist;
1001 struct record_opts *opts = &top->record_opts;
1003 if (perf_top__overwrite_check(top)) {
1004 ui__error("perf top only support consistent per-event "
1005 "overwrite setting for all events\n");
1009 perf_evlist__config(evlist, opts, &callchain_param);
1011 evlist__for_each_entry(evlist, counter) {
1013 if (evsel__open(counter, top->evlist->core.cpus,
1014 top->evlist->core.threads) < 0) {
1017 * Specially handle overwrite fall back.
1018 * Because perf top is the only tool which has
1019 * overwrite mode by default, support
1020 * both overwrite and non-overwrite mode, and
1021 * require consistent mode for all events.
1023 * May move it to generic code with more tools
1024 * have similar attribute.
1026 if (perf_missing_features.write_backward &&
1027 perf_top_overwrite_fallback(top, counter))
1030 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1032 ui__warning("%s\n", msg);
1036 perf_evsel__open_strerror(counter, &opts->target,
1037 errno, msg, sizeof(msg));
1038 ui__error("%s\n", msg);
1043 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1044 ui__error("Failed to mmap with %d (%s)\n",
1045 errno, str_error_r(errno, msg, sizeof(msg)));
1055 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1057 if (callchain->mode != CHAIN_NONE) {
1058 if (callchain_register_param(callchain) < 0) {
1059 ui__error("Can't register callchain params.\n");
1067 static struct ordered_events *rotate_queues(struct perf_top *top)
1069 struct ordered_events *in = top->qe.in;
1071 if (top->qe.in == &top->qe.data[1])
1072 top->qe.in = &top->qe.data[0];
1074 top->qe.in = &top->qe.data[1];
1079 static void *process_thread(void *arg)
1081 struct perf_top *top = arg;
1084 struct ordered_events *out, *in = top->qe.in;
1086 if (!in->nr_events) {
1091 out = rotate_queues(top);
1093 pthread_mutex_lock(&top->qe.mutex);
1094 top->qe.rotate = true;
1095 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1096 pthread_mutex_unlock(&top->qe.mutex);
1098 if (ordered_events__flush(out, OE_FLUSH__TOP))
1099 pr_err("failed to process events\n");
1106 * Allow only 'top->delay_secs' seconds behind samples.
1108 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1110 union perf_event *event = qevent->event;
1111 u64 delay_timestamp;
1113 if (event->header.type != PERF_RECORD_SAMPLE)
1116 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1117 return delay_timestamp < last_timestamp;
1120 static int deliver_event(struct ordered_events *qe,
1121 struct ordered_event *qevent)
1123 struct perf_top *top = qe->data;
1124 struct evlist *evlist = top->evlist;
1125 struct perf_session *session = top->session;
1126 union perf_event *event = qevent->event;
1127 struct perf_sample sample;
1128 struct evsel *evsel;
1129 struct machine *machine;
1132 if (should_drop(qevent, top)) {
1138 ret = perf_evlist__parse_sample(evlist, event, &sample);
1140 pr_err("Can't parse sample, err = %d\n", ret);
1144 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1145 assert(evsel != NULL);
1147 if (event->header.type == PERF_RECORD_SAMPLE) {
1148 if (evswitch__discard(&top->evswitch, evsel))
1153 switch (sample.cpumode) {
1154 case PERF_RECORD_MISC_USER:
1156 if (top->hide_user_symbols)
1158 machine = &session->machines.host;
1160 case PERF_RECORD_MISC_KERNEL:
1161 ++top->kernel_samples;
1162 if (top->hide_kernel_symbols)
1164 machine = &session->machines.host;
1166 case PERF_RECORD_MISC_GUEST_KERNEL:
1167 ++top->guest_kernel_samples;
1168 machine = perf_session__find_machine(session,
1171 case PERF_RECORD_MISC_GUEST_USER:
1172 ++top->guest_us_samples;
1174 * TODO: we don't process guest user from host side
1175 * except simple counting.
1179 if (event->header.type == PERF_RECORD_SAMPLE)
1181 machine = &session->machines.host;
1185 if (event->header.type == PERF_RECORD_SAMPLE) {
1186 perf_event__process_sample(&top->tool, event, evsel,
1188 } else if (event->header.type == PERF_RECORD_LOST) {
1189 perf_top__process_lost(top, event, evsel);
1190 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1191 perf_top__process_lost_samples(top, event, evsel);
1192 } else if (event->header.type < PERF_RECORD_MAX) {
1193 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1194 machine__process_event(machine, event, &sample);
1196 ++session->evlist->stats.nr_unknown_events;
1203 static void init_process_thread(struct perf_top *top)
1205 ordered_events__init(&top->qe.data[0], deliver_event, top);
1206 ordered_events__init(&top->qe.data[1], deliver_event, top);
1207 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1208 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1209 top->qe.in = &top->qe.data[0];
1210 pthread_mutex_init(&top->qe.mutex, NULL);
1211 pthread_cond_init(&top->qe.cond, NULL);
1214 static int __cmd_top(struct perf_top *top)
1216 struct record_opts *opts = &top->record_opts;
1217 pthread_t thread, thread_process;
1220 if (!top->annotation_opts.objdump_path) {
1221 ret = perf_env__lookup_objdump(&top->session->header.env,
1222 &top->annotation_opts.objdump_path);
1227 ret = callchain_param__setup_sample_type(&callchain_param);
1231 if (perf_session__register_idle_thread(top->session) < 0)
1234 if (top->nr_threads_synthesize > 1)
1235 perf_set_multithreaded();
1237 init_process_thread(top);
1239 if (opts->record_namespaces)
1240 top->tool.namespace_events = true;
1242 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1243 &top->session->machines.host,
1246 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1248 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1249 top->evlist->core.threads, false,
1250 top->nr_threads_synthesize);
1252 if (top->nr_threads_synthesize > 1)
1253 perf_set_singlethreaded();
1255 if (perf_hpp_list.socket) {
1256 ret = perf_env__read_cpu_topology_map(&perf_env);
1258 char errbuf[BUFSIZ];
1259 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1261 ui__error("Could not read the CPU topology map: %s\n", err);
1266 ret = perf_top__start_counters(top);
1270 top->session->evlist = top->evlist;
1271 perf_session__set_id_hdr_size(top->session);
1274 * When perf is starting the traced process, all the events (apart from
1275 * group members) have enable_on_exec=1 set, so don't spoil it by
1276 * prematurely enabling them.
1278 * XXX 'top' still doesn't start workloads like record, trace, but should,
1279 * so leave the check here.
1281 if (!target__none(&opts->target))
1282 evlist__enable(top->evlist);
1285 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1286 ui__error("Could not create process thread.\n");
1290 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1291 display_thread), top)) {
1292 ui__error("Could not create display thread.\n");
1293 goto out_join_thread;
1296 if (top->realtime_prio) {
1297 struct sched_param param;
1299 param.sched_priority = top->realtime_prio;
1300 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1301 ui__error("Could not set realtime priority.\n");
1306 /* Wait for a minimal set of events before starting the snapshot */
1307 perf_evlist__poll(top->evlist, 100);
1309 perf_top__mmap_read(top);
1312 u64 hits = top->samples;
1314 perf_top__mmap_read(top);
1316 if (opts->overwrite || (hits == top->samples))
1317 ret = perf_evlist__poll(top->evlist, 100);
1320 perf_top__resize(top);
1327 pthread_join(thread, NULL);
1329 pthread_cond_signal(&top->qe.cond);
1330 pthread_join(thread_process, NULL);
1335 callchain_opt(const struct option *opt, const char *arg, int unset)
1337 symbol_conf.use_callchain = true;
1338 return record_callchain_opt(opt, arg, unset);
1342 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1344 struct callchain_param *callchain = opt->value;
1346 callchain->enabled = !unset;
1347 callchain->record_mode = CALLCHAIN_FP;
1353 symbol_conf.use_callchain = false;
1354 callchain->record_mode = CALLCHAIN_NONE;
1358 return parse_callchain_top_opt(arg);
1361 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1363 if (!strcmp(var, "top.call-graph")) {
1364 var = "call-graph.record-mode";
1365 return perf_default_config(var, value, cb);
1367 if (!strcmp(var, "top.children")) {
1368 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1376 parse_percent_limit(const struct option *opt, const char *arg,
1377 int unset __maybe_unused)
1379 struct perf_top *top = opt->value;
1381 top->min_percent = strtof(arg, NULL);
1385 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1386 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1388 int cmd_top(int argc, const char **argv)
1390 char errbuf[BUFSIZ];
1391 struct perf_top top = {
1395 .mmap_pages = UINT_MAX,
1396 .user_freq = UINT_MAX,
1397 .user_interval = ULLONG_MAX,
1398 .freq = 4000, /* 4 KHz */
1403 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1404 * when we pause, fix that and reenable. Probably using a
1405 * separate evlist with a dummy event, i.e. a non-overwrite
1406 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1407 * stays in overwrite mode. -acme
1410 .sample_time = true,
1411 .sample_time_set = true,
1413 .max_stack = sysctl__max_stack(),
1414 .annotation_opts = annotation__default_options,
1415 .nr_threads_synthesize = UINT_MAX,
1417 struct record_opts *opts = &top.record_opts;
1418 struct target *target = &opts->target;
1419 const struct option options[] = {
1420 OPT_CALLBACK('e', "event", &top.evlist, "event",
1421 "event selector. use 'perf list' to list available events",
1422 parse_events_option),
1423 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1424 OPT_STRING('p', "pid", &target->pid, "pid",
1425 "profile events on existing process id"),
1426 OPT_STRING('t', "tid", &target->tid, "tid",
1427 "profile events on existing thread id"),
1428 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1429 "system-wide collection from all CPUs"),
1430 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1431 "list of cpus to monitor"),
1432 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1433 "file", "vmlinux pathname"),
1434 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1435 "don't load vmlinux even if found"),
1436 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1437 "file", "kallsyms pathname"),
1438 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1439 "hide kernel symbols"),
1440 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1441 "number of mmap data pages",
1442 perf_evlist__parse_mmap_pages),
1443 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1444 "collect data with this RT SCHED_FIFO priority"),
1445 OPT_INTEGER('d', "delay", &top.delay_secs,
1446 "number of seconds to delay between refreshes"),
1447 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1448 "dump the symbol table used for profiling"),
1449 OPT_INTEGER('f', "count-filter", &top.count_filter,
1450 "only display functions with more events than this"),
1451 OPT_BOOLEAN(0, "group", &opts->group,
1452 "put the counters into a counter group"),
1453 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1454 "child tasks do not inherit counters"),
1455 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1456 "symbol to annotate"),
1457 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1458 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1459 "profile at this frequency",
1460 record__parse_freq),
1461 OPT_INTEGER('E', "entries", &top.print_entries,
1462 "display this many functions"),
1463 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1464 "hide user symbols"),
1465 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1466 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1467 OPT_INCR('v', "verbose", &verbose,
1468 "be more verbose (show counter open errors, etc)"),
1469 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1470 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1471 " Please refer the man page for the complete list."),
1472 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1473 "output field(s): overhead, period, sample plus all of sort keys"),
1474 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1475 "Show a column with the number of samples"),
1476 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1477 NULL, "enables call-graph recording and display",
1479 OPT_CALLBACK(0, "call-graph", &callchain_param,
1480 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1481 top_callchain_help, &parse_callchain_opt),
1482 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1483 "Accumulate callchains of children and show total overhead as well"),
1484 OPT_INTEGER(0, "max-stack", &top.max_stack,
1485 "Set the maximum stack depth when parsing the callchain. "
1486 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1487 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1488 "ignore callees of these functions in call graphs",
1489 report_parse_ignore_callees_opt),
1490 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1491 "Show a column with the sum of periods"),
1492 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1493 "only consider symbols in these dsos"),
1494 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1495 "only consider symbols in these comms"),
1496 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1497 "only consider these symbols"),
1498 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1499 "Interleave source code with assembly code (default)"),
1500 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1501 "Display raw encoding of assembly instructions (default)"),
1502 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1503 "Enable kernel symbol demangling"),
1504 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1505 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1506 "objdump binary to use for disassembly and annotations"),
1507 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1508 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1509 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1510 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1511 "Don't show entries under that percent", parse_percent_limit),
1512 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1513 "How to display percentage of filtered entries", parse_filter_percentage),
1514 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1516 "don't try to adjust column width, use these fixed values"),
1517 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1518 "per thread proc mmap processing timeout in ms"),
1519 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1520 "branch any", "sample any taken branches",
1521 parse_branch_stack),
1522 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1523 "branch filter mask", "branch stack filter modes",
1524 parse_branch_stack),
1525 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1526 "Show raw trace event output (do not use print fmt or plugins)"),
1527 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1528 "Show entries in a hierarchy"),
1529 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1530 "Use a backward ring buffer, default: no"),
1531 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1532 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1533 "number of thread to run event synthesize"),
1534 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1535 "Record namespaces events"),
1536 OPTS_EVSWITCH(&top.evswitch),
1539 struct evlist *sb_evlist = NULL;
1540 const char * const top_usage[] = {
1541 "perf top [<options>]",
1544 int status = hists__init();
1549 top.annotation_opts.min_pcnt = 5;
1550 top.annotation_opts.context = 4;
1552 top.evlist = evlist__new();
1553 if (top.evlist == NULL)
1556 status = perf_config(perf_top_config, &top);
1560 argc = parse_options(argc, argv, options, top_usage, 0);
1562 usage_with_options(top_usage, options);
1564 if (!top.evlist->core.nr_entries &&
1565 perf_evlist__add_default(top.evlist) < 0) {
1566 pr_err("Not enough memory for event selector list\n");
1567 goto out_delete_evlist;
1570 status = evswitch__init(&top.evswitch, top.evlist, stderr);
1572 goto out_delete_evlist;
1574 if (symbol_conf.report_hierarchy) {
1575 /* disable incompatible options */
1576 symbol_conf.event_group = false;
1577 symbol_conf.cumulate_callchain = false;
1580 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1581 parse_options_usage(top_usage, options, "fields", 0);
1582 parse_options_usage(NULL, options, "hierarchy", 0);
1583 goto out_delete_evlist;
1587 if (opts->branch_stack && callchain_param.enabled)
1588 symbol_conf.show_branchflag_count = true;
1590 sort__mode = SORT_MODE__TOP;
1591 /* display thread wants entries to be collapsed in a different tree */
1592 perf_hpp_list.need_collapse = 1;
1596 else if (top.use_tui)
1599 setup_browser(false);
1601 if (setup_sorting(top.evlist) < 0) {
1603 parse_options_usage(top_usage, options, "s", 1);
1605 parse_options_usage(sort_order ? NULL : top_usage,
1606 options, "fields", 0);
1607 goto out_delete_evlist;
1610 status = target__validate(target);
1612 target__strerror(target, status, errbuf, BUFSIZ);
1613 ui__warning("%s\n", errbuf);
1616 status = target__parse_uid(target);
1618 int saved_errno = errno;
1620 target__strerror(target, status, errbuf, BUFSIZ);
1621 ui__error("%s\n", errbuf);
1623 status = -saved_errno;
1624 goto out_delete_evlist;
1627 if (target__none(target))
1628 target->system_wide = true;
1630 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1631 ui__error("Couldn't create thread/CPU maps: %s\n",
1632 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1633 goto out_delete_evlist;
1636 if (top.delay_secs < 1)
1639 if (record_opts__config(opts)) {
1641 goto out_delete_evlist;
1644 top.sym_evsel = perf_evlist__first(top.evlist);
1646 if (!callchain_param.enabled) {
1647 symbol_conf.cumulate_callchain = false;
1648 perf_hpp__cancel_cumulate();
1651 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1652 callchain_param.order = ORDER_CALLER;
1654 status = symbol__annotation_init();
1656 goto out_delete_evlist;
1658 annotation_config__init();
1660 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1661 status = symbol__init(NULL);
1663 goto out_delete_evlist;
1665 sort__setup_elide(stdout);
1667 get_term_dimensions(&top.winsize);
1668 if (top.print_entries == 0) {
1669 perf_top__update_print_entries(&top);
1670 signal(SIGWINCH, winch_sig);
1673 top.session = perf_session__new(NULL, false, NULL);
1674 if (top.session == NULL) {
1676 goto out_delete_evlist;
1679 if (!top.record_opts.no_bpf_event)
1680 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1682 if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1683 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1684 opts->no_bpf_event = true;
1687 status = __cmd_top(&top);
1689 if (!opts->no_bpf_event)
1690 perf_evlist__stop_sb_thread(sb_evlist);
1693 evlist__delete(top.evlist);
1694 perf_session__delete(top.session);