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"
35 #include "util/synthetic-events.h"
37 #include "util/util.h"
38 #include <linux/rbtree.h>
39 #include <subcmd/parse-options.h>
40 #include "util/parse-events.h"
41 #include "util/callchain.h"
42 #include "util/cpumap.h"
43 #include "util/sort.h"
44 #include "util/string2.h"
45 #include "util/term.h"
46 #include "util/intlist.h"
47 #include "util/parse-branch-options.h"
48 #include "arch/common.h"
51 #include "util/debug.h"
52 #include "util/ordered-events.h"
68 #include <sys/syscall.h>
69 #include <sys/ioctl.h>
71 #include <sys/prctl.h>
74 #include <sys/utsname.h>
77 #include <linux/stringify.h>
78 #include <linux/time64.h>
79 #include <linux/types.h>
81 #include <linux/ctype.h>
83 static volatile int done;
84 static volatile int resize;
86 #define HEADER_LINE_NR 5
88 static void perf_top__update_print_entries(struct perf_top *top)
90 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
93 static void winch_sig(int sig __maybe_unused)
98 static void perf_top__resize(struct perf_top *top)
100 get_term_dimensions(&top->winsize);
101 perf_top__update_print_entries(top);
104 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
108 struct annotation *notes;
112 if (!he || !he->ms.sym)
115 evsel = hists_to_evsel(he->hists);
121 * We can't annotate with just /proc/kallsyms
123 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
124 !dso__is_kcore(map->dso)) {
125 pr_err("Can't annotate %s: No vmlinux file was found in the "
126 "path\n", sym->name);
131 notes = symbol__annotation(sym);
132 pthread_mutex_lock(¬es->lock);
134 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
135 pthread_mutex_unlock(¬es->lock);
136 pr_err("Not enough memory for annotating '%s' symbol!\n",
142 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
144 top->sym_filter_entry = he;
147 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
148 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
151 pthread_mutex_unlock(¬es->lock);
155 static void __zero_source_counters(struct hist_entry *he)
157 struct symbol *sym = he->ms.sym;
158 symbol__annotate_zero_histograms(sym);
161 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
164 int err = uname(&uts);
166 ui__warning("Out of bounds address found:\n\n"
167 "Addr: %" PRIx64 "\n"
169 "Map: %" PRIx64 "-%" PRIx64 "\n"
170 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
174 "Not all samples will be on the annotation output.\n\n"
175 "Please report to linux-kernel@vger.kernel.org\n",
176 ip, map->dso->long_name, dso__symtab_origin(map->dso),
177 map->start, map->end, sym->start, sym->end,
178 sym->binding == STB_GLOBAL ? 'g' :
179 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
180 err ? "[unknown]" : uts.machine,
181 err ? "[unknown]" : uts.release, perf_version_string);
182 if (use_browser <= 0)
185 map->erange_warned = true;
188 static void perf_top__record_precise_ip(struct perf_top *top,
189 struct hist_entry *he,
190 struct perf_sample *sample,
191 struct evsel *evsel, u64 ip)
193 struct annotation *notes;
194 struct symbol *sym = he->ms.sym;
197 if (sym == NULL || (use_browser == 0 &&
198 (top->sym_filter_entry == NULL ||
199 top->sym_filter_entry->ms.sym != sym)))
202 notes = symbol__annotation(sym);
204 if (pthread_mutex_trylock(¬es->lock))
207 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
209 pthread_mutex_unlock(¬es->lock);
213 * This function is now called with he->hists->lock held.
214 * Release it before going to sleep.
216 pthread_mutex_unlock(&he->hists->lock);
218 if (err == -ERANGE && !he->ms.map->erange_warned)
219 ui__warn_map_erange(he->ms.map, sym, ip);
220 else if (err == -ENOMEM) {
221 pr_err("Not enough memory for annotating '%s' symbol!\n",
226 pthread_mutex_lock(&he->hists->lock);
230 static void perf_top__show_details(struct perf_top *top)
232 struct hist_entry *he = top->sym_filter_entry;
234 struct annotation *notes;
235 struct symbol *symbol;
241 evsel = hists_to_evsel(he->hists);
244 notes = symbol__annotation(symbol);
246 pthread_mutex_lock(¬es->lock);
248 symbol__calc_percent(symbol, evsel);
250 if (notes->src == NULL)
253 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
254 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
256 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
258 if (top->evlist->enabled) {
260 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
262 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
265 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
267 pthread_mutex_unlock(¬es->lock);
270 static void perf_top__resort_hists(struct perf_top *t)
272 struct evlist *evlist = t->evlist;
275 evlist__for_each_entry(evlist, pos) {
276 struct hists *hists = evsel__hists(pos);
279 * unlink existing entries so that they can be linked
280 * in a correct order in hists__match() below.
282 hists__unlink(hists);
284 if (evlist->enabled) {
286 hists__delete_entries(hists);
288 hists__decay_entries(hists, t->hide_user_symbols,
289 t->hide_kernel_symbols);
293 hists__collapse_resort(hists, NULL);
295 /* Non-group events are considered as leader */
296 if (symbol_conf.event_group &&
297 !perf_evsel__is_group_leader(pos)) {
298 struct hists *leader_hists = evsel__hists(pos->leader);
300 hists__match(leader_hists, hists);
301 hists__link(leader_hists, hists);
305 evlist__for_each_entry(evlist, pos) {
306 perf_evsel__output_resort(pos, NULL);
310 static void perf_top__print_sym_table(struct perf_top *top)
314 const int win_width = top->winsize.ws_col - 1;
315 struct evsel *evsel = top->sym_evsel;
316 struct hists *hists = evsel__hists(evsel);
320 perf_top__header_snprintf(top, bf, sizeof(bf));
323 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
325 if (!top->record_opts.overwrite &&
326 (hists->stats.nr_lost_warned !=
327 hists->stats.nr_events[PERF_RECORD_LOST])) {
328 hists->stats.nr_lost_warned =
329 hists->stats.nr_events[PERF_RECORD_LOST];
330 color_fprintf(stdout, PERF_COLOR_RED,
331 "WARNING: LOST %d chunks, Check IO/CPU overload",
332 hists->stats.nr_lost_warned);
336 if (top->sym_filter_entry) {
337 perf_top__show_details(top);
341 perf_top__resort_hists(top);
343 hists__output_recalc_col_len(hists, top->print_entries - printed);
345 hists__fprintf(hists, false, top->print_entries - printed, win_width,
346 top->min_percent, stdout, !symbol_conf.use_callchain);
349 static void prompt_integer(int *target, const char *msg)
351 char *buf = malloc(0), *p;
355 fprintf(stdout, "\n%s: ", msg);
356 if (getline(&buf, &dummy, stdin) < 0)
359 p = strchr(buf, '\n');
369 tmp = strtoul(buf, NULL, 10);
375 static void prompt_percent(int *target, const char *msg)
379 prompt_integer(&tmp, msg);
380 if (tmp >= 0 && tmp <= 100)
384 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
386 char *buf = malloc(0), *p;
387 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
388 struct hists *hists = evsel__hists(top->sym_evsel);
389 struct rb_node *next;
392 /* zero counters of active symbol */
394 __zero_source_counters(syme);
395 top->sym_filter_entry = NULL;
398 fprintf(stdout, "\n%s: ", msg);
399 if (getline(&buf, &dummy, stdin) < 0)
402 p = strchr(buf, '\n');
406 next = rb_first_cached(&hists->entries);
408 n = rb_entry(next, struct hist_entry, rb_node);
409 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
413 next = rb_next(&n->rb_node);
417 fprintf(stderr, "Sorry, %s is not active.\n", buf);
420 perf_top__parse_source(top, found);
426 static void perf_top__print_mapped_keys(struct perf_top *top)
430 if (top->sym_filter_entry) {
431 struct symbol *sym = top->sym_filter_entry->ms.sym;
435 fprintf(stdout, "\nMapped keys:\n");
436 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
437 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
439 if (top->evlist->core.nr_entries > 1)
440 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
442 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
444 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
445 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
446 fprintf(stdout, "\t[S] stop annotation.\n");
449 "\t[K] hide kernel symbols. \t(%s)\n",
450 top->hide_kernel_symbols ? "yes" : "no");
452 "\t[U] hide user symbols. \t(%s)\n",
453 top->hide_user_symbols ? "yes" : "no");
454 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
455 fprintf(stdout, "\t[qQ] quit.\n");
458 static int perf_top__key_mapped(struct perf_top *top, int c)
474 return top->evlist->core.nr_entries > 1 ? 1 : 0;
482 static bool perf_top__handle_keypress(struct perf_top *top, int c)
486 if (!perf_top__key_mapped(top, c)) {
487 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
490 perf_top__print_mapped_keys(top);
491 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
494 set_term_quiet_input(&save);
496 poll(&stdin_poll, 1, -1);
499 tcsetattr(0, TCSAFLUSH, &save);
500 if (!perf_top__key_mapped(top, c))
506 prompt_integer(&top->delay_secs, "Enter display delay");
507 if (top->delay_secs < 1)
511 prompt_integer(&top->print_entries, "Enter display entries (lines)");
512 if (top->print_entries == 0) {
513 perf_top__resize(top);
514 signal(SIGWINCH, winch_sig);
516 signal(SIGWINCH, SIG_DFL);
520 if (top->evlist->core.nr_entries > 1) {
521 /* Select 0 as the default event: */
524 fprintf(stderr, "\nAvailable events:");
526 evlist__for_each_entry(top->evlist, top->sym_evsel)
527 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
529 prompt_integer(&counter, "Enter details event counter");
531 if (counter >= top->evlist->core.nr_entries) {
532 top->sym_evsel = perf_evlist__first(top->evlist);
533 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
537 evlist__for_each_entry(top->evlist, top->sym_evsel)
538 if (top->sym_evsel->idx == counter)
541 top->sym_evsel = perf_evlist__first(top->evlist);
544 prompt_integer(&top->count_filter, "Enter display event count filter");
547 prompt_percent(&top->annotation_opts.min_pcnt,
548 "Enter details display event filter (percent)");
551 top->hide_kernel_symbols = !top->hide_kernel_symbols;
555 printf("exiting.\n");
556 if (top->dump_symtab)
557 perf_session__fprintf_dsos(top->session, stderr);
561 perf_top__prompt_symbol(top, "Enter details symbol");
564 if (!top->sym_filter_entry)
567 struct hist_entry *syme = top->sym_filter_entry;
569 top->sym_filter_entry = NULL;
570 __zero_source_counters(syme);
574 top->hide_user_symbols = !top->hide_user_symbols;
577 top->zero = !top->zero;
586 static void perf_top__sort_new_samples(void *arg)
588 struct perf_top *t = arg;
590 if (t->evlist->selected != NULL)
591 t->sym_evsel = t->evlist->selected;
593 perf_top__resort_hists(t);
595 if (t->lost || t->drop)
596 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
599 static void stop_top(void)
605 static void *display_thread_tui(void *arg)
608 struct perf_top *top = arg;
609 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
610 struct hist_browser_timer hbt = {
611 .timer = perf_top__sort_new_samples,
613 .refresh = top->delay_secs,
616 /* In order to read symbols from other namespaces perf to needs to call
617 * setns(2). This isn't permitted if the struct_fs has multiple users.
618 * unshare(2) the fs so that we may continue to setns into namespaces
619 * that we're observing.
623 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
625 perf_top__sort_new_samples(top);
628 * Initialize the uid_filter_str, in the future the TUI will allow
629 * Zooming in/out UIDs. For now just use whatever the user passed
632 evlist__for_each_entry(top->evlist, pos) {
633 struct hists *hists = evsel__hists(pos);
634 hists->uid_filter_str = top->record_opts.target.uid_str;
637 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
639 &top->session->header.env,
640 !top->record_opts.overwrite,
641 &top->annotation_opts);
647 static void display_sig(int sig __maybe_unused)
652 static void display_setup_sig(void)
654 signal(SIGSEGV, sighandler_dump_stack);
655 signal(SIGFPE, sighandler_dump_stack);
656 signal(SIGINT, display_sig);
657 signal(SIGQUIT, display_sig);
658 signal(SIGTERM, display_sig);
661 static void *display_thread(void *arg)
663 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
665 struct perf_top *top = arg;
668 /* In order to read symbols from other namespaces perf to needs to call
669 * setns(2). This isn't permitted if the struct_fs has multiple users.
670 * unshare(2) the fs so that we may continue to setns into namespaces
671 * that we're observing.
675 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
678 pthread__unblock_sigwinch();
680 delay_msecs = top->delay_secs * MSEC_PER_SEC;
681 set_term_quiet_input(&save);
686 perf_top__print_sym_table(top);
688 * Either timeout expired or we got an EINTR due to SIGWINCH,
689 * refresh screen in both cases.
691 switch (poll(&stdin_poll, 1, delay_msecs)) {
700 tcsetattr(0, TCSAFLUSH, &save);
702 if (perf_top__handle_keypress(top, c))
708 tcsetattr(0, TCSAFLUSH, &save);
712 static int hist_iter__top_callback(struct hist_entry_iter *iter,
713 struct addr_location *al, bool single,
716 struct perf_top *top = arg;
717 struct hist_entry *he = iter->he;
718 struct evsel *evsel = iter->evsel;
720 if (perf_hpp_list.sym && single)
721 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
723 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
724 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
728 static void perf_event__process_sample(struct perf_tool *tool,
729 const union perf_event *event,
731 struct perf_sample *sample,
732 struct machine *machine)
734 struct perf_top *top = container_of(tool, struct perf_top, tool);
735 struct addr_location al;
738 if (!machine && perf_guest) {
739 static struct intlist *seen;
742 seen = intlist__new(NULL);
744 if (!intlist__has_entry(seen, sample->pid)) {
745 pr_err("Can't find guest [%d]'s kernel information\n",
747 intlist__add(seen, sample->pid);
753 pr_err("%u unprocessable samples recorded.\r",
754 top->session->evlist->stats.nr_unprocessable_samples++);
758 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
759 top->exact_samples++;
761 if (machine__resolve(machine, &al, sample) < 0)
764 if (!machine->kptr_restrict_warned &&
765 symbol_conf.kptr_restrict &&
766 al.cpumode == PERF_RECORD_MISC_KERNEL) {
767 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
769 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
770 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
771 "Kernel%s samples will not be resolved.\n",
772 al.map && map__has_symbols(al.map) ?
774 if (use_browser <= 0)
777 machine->kptr_restrict_warned = true;
780 if (al.sym == NULL && al.map != NULL) {
781 const char *msg = "Kernel samples will not be resolved.\n";
783 * As we do lazy loading of symtabs we only will know if the
784 * specified vmlinux file is invalid when we actually have a
785 * hit in kernel space and then try to load it. So if we get
786 * here and there are _no_ symbols in the DSO backing the
787 * kernel map, bail out.
789 * We may never get here, for instance, if we use -K/
790 * --hide-kernel-symbols, even if the user specifies an
791 * invalid --vmlinux ;-)
793 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
794 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
795 if (symbol_conf.vmlinux_name) {
797 dso__strerror_load(al.map->dso, serr, sizeof(serr));
798 ui__warning("The %s file can't be used: %s\n%s",
799 symbol_conf.vmlinux_name, serr, msg);
801 ui__warning("A vmlinux file was not found.\n%s",
805 if (use_browser <= 0)
807 top->vmlinux_warned = true;
811 if (al.sym == NULL || !al.sym->idle) {
812 struct hists *hists = evsel__hists(evsel);
813 struct hist_entry_iter iter = {
816 .add_entry_cb = hist_iter__top_callback,
819 if (symbol_conf.cumulate_callchain)
820 iter.ops = &hist_iter_cumulative;
822 iter.ops = &hist_iter_normal;
824 pthread_mutex_lock(&hists->lock);
826 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
828 pr_err("Problem incrementing symbol period, skipping event\n");
830 pthread_mutex_unlock(&hists->lock);
833 addr_location__put(&al);
837 perf_top__process_lost(struct perf_top *top, union perf_event *event,
840 struct hists *hists = evsel__hists(evsel);
842 top->lost += event->lost.lost;
843 top->lost_total += event->lost.lost;
844 hists->stats.total_lost += event->lost.lost;
848 perf_top__process_lost_samples(struct perf_top *top,
849 union perf_event *event,
852 struct hists *hists = evsel__hists(evsel);
854 top->lost += event->lost_samples.lost;
855 top->lost_total += event->lost_samples.lost;
856 hists->stats.total_lost_samples += event->lost_samples.lost;
859 static u64 last_timestamp;
861 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
863 struct record_opts *opts = &top->record_opts;
864 struct evlist *evlist = top->evlist;
865 struct perf_mmap *md;
866 union perf_event *event;
868 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
869 if (perf_mmap__read_init(md) < 0)
872 while ((event = perf_mmap__read_event(md)) != NULL) {
875 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
876 if (ret && ret != -1)
879 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
883 perf_mmap__consume(md);
885 if (top->qe.rotate) {
886 pthread_mutex_lock(&top->qe.mutex);
887 top->qe.rotate = false;
888 pthread_cond_signal(&top->qe.cond);
889 pthread_mutex_unlock(&top->qe.mutex);
893 perf_mmap__read_done(md);
896 static void perf_top__mmap_read(struct perf_top *top)
898 bool overwrite = top->record_opts.overwrite;
899 struct evlist *evlist = top->evlist;
903 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
905 for (i = 0; i < top->evlist->nr_mmaps; i++)
906 perf_top__mmap_read_idx(top, i);
909 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
910 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
915 * Check per-event overwrite term.
916 * perf top should support consistent term for all events.
917 * - All events don't have per-event term
918 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
919 * Nothing change, return 0.
920 * - All events have same per-event term
921 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
922 * Using the per-event setting to replace the opts->overwrite if
923 * they are different, then return 0.
924 * - Events have different per-event term
925 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
927 * - Some of the event set per-event term, but some not.
928 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
931 static int perf_top__overwrite_check(struct perf_top *top)
933 struct record_opts *opts = &top->record_opts;
934 struct evlist *evlist = top->evlist;
935 struct perf_evsel_config_term *term;
936 struct list_head *config_terms;
938 int set, overwrite = -1;
940 evlist__for_each_entry(evlist, evsel) {
942 config_terms = &evsel->config_terms;
943 list_for_each_entry(term, config_terms, list) {
944 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
945 set = term->val.overwrite ? 1 : 0;
948 /* no term for current and previous event (likely) */
949 if ((overwrite < 0) && (set < 0))
952 /* has term for both current and previous event, compare */
953 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
956 /* no term for current event but has term for previous one */
957 if ((overwrite >= 0) && (set < 0))
960 /* has term for current event */
961 if ((overwrite < 0) && (set >= 0)) {
962 /* if it's first event, set overwrite */
963 if (evsel == perf_evlist__first(evlist))
970 if ((overwrite >= 0) && (opts->overwrite != overwrite))
971 opts->overwrite = overwrite;
976 static int perf_top_overwrite_fallback(struct perf_top *top,
979 struct record_opts *opts = &top->record_opts;
980 struct evlist *evlist = top->evlist;
981 struct evsel *counter;
983 if (!opts->overwrite)
986 /* only fall back when first event fails */
987 if (evsel != perf_evlist__first(evlist))
990 evlist__for_each_entry(evlist, counter)
991 counter->core.attr.write_backward = false;
992 opts->overwrite = false;
993 pr_debug2("fall back to non-overwrite mode\n");
997 static int perf_top__start_counters(struct perf_top *top)
1000 struct evsel *counter;
1001 struct evlist *evlist = top->evlist;
1002 struct record_opts *opts = &top->record_opts;
1004 if (perf_top__overwrite_check(top)) {
1005 ui__error("perf top only support consistent per-event "
1006 "overwrite setting for all events\n");
1010 perf_evlist__config(evlist, opts, &callchain_param);
1012 evlist__for_each_entry(evlist, counter) {
1014 if (evsel__open(counter, top->evlist->core.cpus,
1015 top->evlist->core.threads) < 0) {
1018 * Specially handle overwrite fall back.
1019 * Because perf top is the only tool which has
1020 * overwrite mode by default, support
1021 * both overwrite and non-overwrite mode, and
1022 * require consistent mode for all events.
1024 * May move it to generic code with more tools
1025 * have similar attribute.
1027 if (perf_missing_features.write_backward &&
1028 perf_top_overwrite_fallback(top, counter))
1031 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1033 ui__warning("%s\n", msg);
1037 perf_evsel__open_strerror(counter, &opts->target,
1038 errno, msg, sizeof(msg));
1039 ui__error("%s\n", msg);
1044 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1045 ui__error("Failed to mmap with %d (%s)\n",
1046 errno, str_error_r(errno, msg, sizeof(msg)));
1056 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1058 if (callchain->mode != CHAIN_NONE) {
1059 if (callchain_register_param(callchain) < 0) {
1060 ui__error("Can't register callchain params.\n");
1068 static struct ordered_events *rotate_queues(struct perf_top *top)
1070 struct ordered_events *in = top->qe.in;
1072 if (top->qe.in == &top->qe.data[1])
1073 top->qe.in = &top->qe.data[0];
1075 top->qe.in = &top->qe.data[1];
1080 static void *process_thread(void *arg)
1082 struct perf_top *top = arg;
1085 struct ordered_events *out, *in = top->qe.in;
1087 if (!in->nr_events) {
1092 out = rotate_queues(top);
1094 pthread_mutex_lock(&top->qe.mutex);
1095 top->qe.rotate = true;
1096 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1097 pthread_mutex_unlock(&top->qe.mutex);
1099 if (ordered_events__flush(out, OE_FLUSH__TOP))
1100 pr_err("failed to process events\n");
1107 * Allow only 'top->delay_secs' seconds behind samples.
1109 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1111 union perf_event *event = qevent->event;
1112 u64 delay_timestamp;
1114 if (event->header.type != PERF_RECORD_SAMPLE)
1117 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1118 return delay_timestamp < last_timestamp;
1121 static int deliver_event(struct ordered_events *qe,
1122 struct ordered_event *qevent)
1124 struct perf_top *top = qe->data;
1125 struct evlist *evlist = top->evlist;
1126 struct perf_session *session = top->session;
1127 union perf_event *event = qevent->event;
1128 struct perf_sample sample;
1129 struct evsel *evsel;
1130 struct machine *machine;
1133 if (should_drop(qevent, top)) {
1139 ret = perf_evlist__parse_sample(evlist, event, &sample);
1141 pr_err("Can't parse sample, err = %d\n", ret);
1145 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1146 assert(evsel != NULL);
1148 if (event->header.type == PERF_RECORD_SAMPLE) {
1149 if (evswitch__discard(&top->evswitch, evsel))
1154 switch (sample.cpumode) {
1155 case PERF_RECORD_MISC_USER:
1157 if (top->hide_user_symbols)
1159 machine = &session->machines.host;
1161 case PERF_RECORD_MISC_KERNEL:
1162 ++top->kernel_samples;
1163 if (top->hide_kernel_symbols)
1165 machine = &session->machines.host;
1167 case PERF_RECORD_MISC_GUEST_KERNEL:
1168 ++top->guest_kernel_samples;
1169 machine = perf_session__find_machine(session,
1172 case PERF_RECORD_MISC_GUEST_USER:
1173 ++top->guest_us_samples;
1175 * TODO: we don't process guest user from host side
1176 * except simple counting.
1180 if (event->header.type == PERF_RECORD_SAMPLE)
1182 machine = &session->machines.host;
1186 if (event->header.type == PERF_RECORD_SAMPLE) {
1187 perf_event__process_sample(&top->tool, event, evsel,
1189 } else if (event->header.type == PERF_RECORD_LOST) {
1190 perf_top__process_lost(top, event, evsel);
1191 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1192 perf_top__process_lost_samples(top, event, evsel);
1193 } else if (event->header.type < PERF_RECORD_MAX) {
1194 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1195 machine__process_event(machine, event, &sample);
1197 ++session->evlist->stats.nr_unknown_events;
1204 static void init_process_thread(struct perf_top *top)
1206 ordered_events__init(&top->qe.data[0], deliver_event, top);
1207 ordered_events__init(&top->qe.data[1], deliver_event, top);
1208 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1209 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1210 top->qe.in = &top->qe.data[0];
1211 pthread_mutex_init(&top->qe.mutex, NULL);
1212 pthread_cond_init(&top->qe.cond, NULL);
1215 static int __cmd_top(struct perf_top *top)
1217 struct record_opts *opts = &top->record_opts;
1218 pthread_t thread, thread_process;
1221 if (!top->annotation_opts.objdump_path) {
1222 ret = perf_env__lookup_objdump(&top->session->header.env,
1223 &top->annotation_opts.objdump_path);
1228 ret = callchain_param__setup_sample_type(&callchain_param);
1232 if (perf_session__register_idle_thread(top->session) < 0)
1235 if (top->nr_threads_synthesize > 1)
1236 perf_set_multithreaded();
1238 init_process_thread(top);
1240 if (opts->record_namespaces)
1241 top->tool.namespace_events = true;
1243 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1244 &top->session->machines.host,
1247 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1249 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1250 top->evlist->core.threads, false,
1251 top->nr_threads_synthesize);
1253 if (top->nr_threads_synthesize > 1)
1254 perf_set_singlethreaded();
1256 if (perf_hpp_list.socket) {
1257 ret = perf_env__read_cpu_topology_map(&perf_env);
1259 char errbuf[BUFSIZ];
1260 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1262 ui__error("Could not read the CPU topology map: %s\n", err);
1267 ret = perf_top__start_counters(top);
1271 top->session->evlist = top->evlist;
1272 perf_session__set_id_hdr_size(top->session);
1275 * When perf is starting the traced process, all the events (apart from
1276 * group members) have enable_on_exec=1 set, so don't spoil it by
1277 * prematurely enabling them.
1279 * XXX 'top' still doesn't start workloads like record, trace, but should,
1280 * so leave the check here.
1282 if (!target__none(&opts->target))
1283 evlist__enable(top->evlist);
1286 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1287 ui__error("Could not create process thread.\n");
1291 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1292 display_thread), top)) {
1293 ui__error("Could not create display thread.\n");
1294 goto out_join_thread;
1297 if (top->realtime_prio) {
1298 struct sched_param param;
1300 param.sched_priority = top->realtime_prio;
1301 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1302 ui__error("Could not set realtime priority.\n");
1307 /* Wait for a minimal set of events before starting the snapshot */
1308 perf_evlist__poll(top->evlist, 100);
1310 perf_top__mmap_read(top);
1313 u64 hits = top->samples;
1315 perf_top__mmap_read(top);
1317 if (opts->overwrite || (hits == top->samples))
1318 ret = perf_evlist__poll(top->evlist, 100);
1321 perf_top__resize(top);
1328 pthread_join(thread, NULL);
1330 pthread_cond_signal(&top->qe.cond);
1331 pthread_join(thread_process, NULL);
1336 callchain_opt(const struct option *opt, const char *arg, int unset)
1338 symbol_conf.use_callchain = true;
1339 return record_callchain_opt(opt, arg, unset);
1343 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1345 struct callchain_param *callchain = opt->value;
1347 callchain->enabled = !unset;
1348 callchain->record_mode = CALLCHAIN_FP;
1354 symbol_conf.use_callchain = false;
1355 callchain->record_mode = CALLCHAIN_NONE;
1359 return parse_callchain_top_opt(arg);
1362 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1364 if (!strcmp(var, "top.call-graph")) {
1365 var = "call-graph.record-mode";
1366 return perf_default_config(var, value, cb);
1368 if (!strcmp(var, "top.children")) {
1369 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1377 parse_percent_limit(const struct option *opt, const char *arg,
1378 int unset __maybe_unused)
1380 struct perf_top *top = opt->value;
1382 top->min_percent = strtof(arg, NULL);
1386 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1387 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1389 int cmd_top(int argc, const char **argv)
1391 char errbuf[BUFSIZ];
1392 struct perf_top top = {
1396 .mmap_pages = UINT_MAX,
1397 .user_freq = UINT_MAX,
1398 .user_interval = ULLONG_MAX,
1399 .freq = 4000, /* 4 KHz */
1404 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1405 * when we pause, fix that and reenable. Probably using a
1406 * separate evlist with a dummy event, i.e. a non-overwrite
1407 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1408 * stays in overwrite mode. -acme
1411 .sample_time = true,
1412 .sample_time_set = true,
1414 .max_stack = sysctl__max_stack(),
1415 .annotation_opts = annotation__default_options,
1416 .nr_threads_synthesize = UINT_MAX,
1418 struct record_opts *opts = &top.record_opts;
1419 struct target *target = &opts->target;
1420 const struct option options[] = {
1421 OPT_CALLBACK('e', "event", &top.evlist, "event",
1422 "event selector. use 'perf list' to list available events",
1423 parse_events_option),
1424 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1425 OPT_STRING('p', "pid", &target->pid, "pid",
1426 "profile events on existing process id"),
1427 OPT_STRING('t', "tid", &target->tid, "tid",
1428 "profile events on existing thread id"),
1429 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1430 "system-wide collection from all CPUs"),
1431 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1432 "list of cpus to monitor"),
1433 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1434 "file", "vmlinux pathname"),
1435 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1436 "don't load vmlinux even if found"),
1437 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1438 "file", "kallsyms pathname"),
1439 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1440 "hide kernel symbols"),
1441 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1442 "number of mmap data pages",
1443 perf_evlist__parse_mmap_pages),
1444 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1445 "collect data with this RT SCHED_FIFO priority"),
1446 OPT_INTEGER('d', "delay", &top.delay_secs,
1447 "number of seconds to delay between refreshes"),
1448 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1449 "dump the symbol table used for profiling"),
1450 OPT_INTEGER('f', "count-filter", &top.count_filter,
1451 "only display functions with more events than this"),
1452 OPT_BOOLEAN(0, "group", &opts->group,
1453 "put the counters into a counter group"),
1454 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1455 "child tasks do not inherit counters"),
1456 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1457 "symbol to annotate"),
1458 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1459 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1460 "profile at this frequency",
1461 record__parse_freq),
1462 OPT_INTEGER('E', "entries", &top.print_entries,
1463 "display this many functions"),
1464 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1465 "hide user symbols"),
1466 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1467 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1468 OPT_INCR('v', "verbose", &verbose,
1469 "be more verbose (show counter open errors, etc)"),
1470 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1471 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1472 " Please refer the man page for the complete list."),
1473 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1474 "output field(s): overhead, period, sample plus all of sort keys"),
1475 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1476 "Show a column with the number of samples"),
1477 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1478 NULL, "enables call-graph recording and display",
1480 OPT_CALLBACK(0, "call-graph", &callchain_param,
1481 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1482 top_callchain_help, &parse_callchain_opt),
1483 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1484 "Accumulate callchains of children and show total overhead as well"),
1485 OPT_INTEGER(0, "max-stack", &top.max_stack,
1486 "Set the maximum stack depth when parsing the callchain. "
1487 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1488 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1489 "ignore callees of these functions in call graphs",
1490 report_parse_ignore_callees_opt),
1491 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1492 "Show a column with the sum of periods"),
1493 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1494 "only consider symbols in these dsos"),
1495 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1496 "only consider symbols in these comms"),
1497 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1498 "only consider these symbols"),
1499 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1500 "Interleave source code with assembly code (default)"),
1501 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1502 "Display raw encoding of assembly instructions (default)"),
1503 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1504 "Enable kernel symbol demangling"),
1505 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1506 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1507 "objdump binary to use for disassembly and annotations"),
1508 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1509 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1510 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1511 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1512 "Don't show entries under that percent", parse_percent_limit),
1513 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1514 "How to display percentage of filtered entries", parse_filter_percentage),
1515 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1517 "don't try to adjust column width, use these fixed values"),
1518 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1519 "per thread proc mmap processing timeout in ms"),
1520 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1521 "branch any", "sample any taken branches",
1522 parse_branch_stack),
1523 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1524 "branch filter mask", "branch stack filter modes",
1525 parse_branch_stack),
1526 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1527 "Show raw trace event output (do not use print fmt or plugins)"),
1528 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1529 "Show entries in a hierarchy"),
1530 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1531 "Use a backward ring buffer, default: no"),
1532 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1533 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1534 "number of thread to run event synthesize"),
1535 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1536 "Record namespaces events"),
1537 OPTS_EVSWITCH(&top.evswitch),
1540 struct evlist *sb_evlist = NULL;
1541 const char * const top_usage[] = {
1542 "perf top [<options>]",
1545 int status = hists__init();
1550 top.annotation_opts.min_pcnt = 5;
1551 top.annotation_opts.context = 4;
1553 top.evlist = evlist__new();
1554 if (top.evlist == NULL)
1557 status = perf_config(perf_top_config, &top);
1561 argc = parse_options(argc, argv, options, top_usage, 0);
1563 usage_with_options(top_usage, options);
1565 if (!top.evlist->core.nr_entries &&
1566 perf_evlist__add_default(top.evlist) < 0) {
1567 pr_err("Not enough memory for event selector list\n");
1568 goto out_delete_evlist;
1571 status = evswitch__init(&top.evswitch, top.evlist, stderr);
1573 goto out_delete_evlist;
1575 if (symbol_conf.report_hierarchy) {
1576 /* disable incompatible options */
1577 symbol_conf.event_group = false;
1578 symbol_conf.cumulate_callchain = false;
1581 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1582 parse_options_usage(top_usage, options, "fields", 0);
1583 parse_options_usage(NULL, options, "hierarchy", 0);
1584 goto out_delete_evlist;
1588 if (opts->branch_stack && callchain_param.enabled)
1589 symbol_conf.show_branchflag_count = true;
1591 sort__mode = SORT_MODE__TOP;
1592 /* display thread wants entries to be collapsed in a different tree */
1593 perf_hpp_list.need_collapse = 1;
1597 else if (top.use_tui)
1600 setup_browser(false);
1602 if (setup_sorting(top.evlist) < 0) {
1604 parse_options_usage(top_usage, options, "s", 1);
1606 parse_options_usage(sort_order ? NULL : top_usage,
1607 options, "fields", 0);
1608 goto out_delete_evlist;
1611 status = target__validate(target);
1613 target__strerror(target, status, errbuf, BUFSIZ);
1614 ui__warning("%s\n", errbuf);
1617 status = target__parse_uid(target);
1619 int saved_errno = errno;
1621 target__strerror(target, status, errbuf, BUFSIZ);
1622 ui__error("%s\n", errbuf);
1624 status = -saved_errno;
1625 goto out_delete_evlist;
1628 if (target__none(target))
1629 target->system_wide = true;
1631 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1632 ui__error("Couldn't create thread/CPU maps: %s\n",
1633 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1634 goto out_delete_evlist;
1637 if (top.delay_secs < 1)
1640 if (record_opts__config(opts)) {
1642 goto out_delete_evlist;
1645 top.sym_evsel = perf_evlist__first(top.evlist);
1647 if (!callchain_param.enabled) {
1648 symbol_conf.cumulate_callchain = false;
1649 perf_hpp__cancel_cumulate();
1652 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1653 callchain_param.order = ORDER_CALLER;
1655 status = symbol__annotation_init();
1657 goto out_delete_evlist;
1659 annotation_config__init();
1661 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1662 status = symbol__init(NULL);
1664 goto out_delete_evlist;
1666 sort__setup_elide(stdout);
1668 get_term_dimensions(&top.winsize);
1669 if (top.print_entries == 0) {
1670 perf_top__update_print_entries(&top);
1671 signal(SIGWINCH, winch_sig);
1674 top.session = perf_session__new(NULL, false, NULL);
1675 if (top.session == NULL) {
1677 goto out_delete_evlist;
1680 if (!top.record_opts.no_bpf_event)
1681 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1683 if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1684 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1685 opts->no_bpf_event = true;
1688 status = __cmd_top(&top);
1690 if (!opts->no_bpf_event)
1691 perf_evlist__stop_sb_thread(sb_evlist);
1694 evlist__delete(top.evlist);
1695 perf_session__delete(top.session);