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"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/event.h"
30 #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/sort.h"
42 #include "util/string2.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"
49 #include "util/ordered-events.h"
65 #include <sys/syscall.h>
66 #include <sys/ioctl.h>
68 #include <sys/prctl.h>
71 #include <sys/utsname.h>
74 #include <linux/stringify.h>
75 #include <linux/time64.h>
76 #include <linux/types.h>
78 #include <linux/ctype.h>
80 static volatile int done;
81 static volatile int resize;
83 #define HEADER_LINE_NR 5
85 static void perf_top__update_print_entries(struct perf_top *top)
87 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
90 static void winch_sig(int sig __maybe_unused)
95 static void perf_top__resize(struct perf_top *top)
97 get_term_dimensions(&top->winsize);
98 perf_top__update_print_entries(top);
101 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
105 struct annotation *notes;
109 if (!he || !he->ms.sym)
112 evsel = hists_to_evsel(he->hists);
118 * We can't annotate with just /proc/kallsyms
120 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
121 !dso__is_kcore(map->dso)) {
122 pr_err("Can't annotate %s: No vmlinux file was found in the "
123 "path\n", sym->name);
128 notes = symbol__annotation(sym);
129 pthread_mutex_lock(¬es->lock);
131 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
132 pthread_mutex_unlock(¬es->lock);
133 pr_err("Not enough memory for annotating '%s' symbol!\n",
139 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
141 top->sym_filter_entry = he;
144 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
145 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
148 pthread_mutex_unlock(¬es->lock);
152 static void __zero_source_counters(struct hist_entry *he)
154 struct symbol *sym = he->ms.sym;
155 symbol__annotate_zero_histograms(sym);
158 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
161 int err = uname(&uts);
163 ui__warning("Out of bounds address found:\n\n"
164 "Addr: %" PRIx64 "\n"
166 "Map: %" PRIx64 "-%" PRIx64 "\n"
167 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
171 "Not all samples will be on the annotation output.\n\n"
172 "Please report to linux-kernel@vger.kernel.org\n",
173 ip, map->dso->long_name, dso__symtab_origin(map->dso),
174 map->start, map->end, sym->start, sym->end,
175 sym->binding == STB_GLOBAL ? 'g' :
176 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
177 err ? "[unknown]" : uts.machine,
178 err ? "[unknown]" : uts.release, perf_version_string);
179 if (use_browser <= 0)
182 map->erange_warned = true;
185 static void perf_top__record_precise_ip(struct perf_top *top,
186 struct hist_entry *he,
187 struct perf_sample *sample,
188 struct evsel *evsel, u64 ip)
190 struct annotation *notes;
191 struct symbol *sym = he->ms.sym;
194 if (sym == NULL || (use_browser == 0 &&
195 (top->sym_filter_entry == NULL ||
196 top->sym_filter_entry->ms.sym != sym)))
199 notes = symbol__annotation(sym);
201 if (pthread_mutex_trylock(¬es->lock))
204 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
206 pthread_mutex_unlock(¬es->lock);
210 * This function is now called with he->hists->lock held.
211 * Release it before going to sleep.
213 pthread_mutex_unlock(&he->hists->lock);
215 if (err == -ERANGE && !he->ms.map->erange_warned)
216 ui__warn_map_erange(he->ms.map, sym, ip);
217 else if (err == -ENOMEM) {
218 pr_err("Not enough memory for annotating '%s' symbol!\n",
223 pthread_mutex_lock(&he->hists->lock);
227 static void perf_top__show_details(struct perf_top *top)
229 struct hist_entry *he = top->sym_filter_entry;
231 struct annotation *notes;
232 struct symbol *symbol;
238 evsel = hists_to_evsel(he->hists);
241 notes = symbol__annotation(symbol);
243 pthread_mutex_lock(¬es->lock);
245 symbol__calc_percent(symbol, evsel);
247 if (notes->src == NULL)
250 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
251 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
253 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
255 if (top->evlist->enabled) {
257 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
259 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
262 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
264 pthread_mutex_unlock(¬es->lock);
267 static void perf_top__print_sym_table(struct perf_top *top)
271 const int win_width = top->winsize.ws_col - 1;
272 struct evsel *evsel = top->sym_evsel;
273 struct hists *hists = evsel__hists(evsel);
277 perf_top__header_snprintf(top, bf, sizeof(bf));
280 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
282 if (!top->record_opts.overwrite &&
283 (hists->stats.nr_lost_warned !=
284 hists->stats.nr_events[PERF_RECORD_LOST])) {
285 hists->stats.nr_lost_warned =
286 hists->stats.nr_events[PERF_RECORD_LOST];
287 color_fprintf(stdout, PERF_COLOR_RED,
288 "WARNING: LOST %d chunks, Check IO/CPU overload",
289 hists->stats.nr_lost_warned);
293 if (top->sym_filter_entry) {
294 perf_top__show_details(top);
298 if (top->evlist->enabled) {
300 hists__delete_entries(hists);
302 hists__decay_entries(hists, top->hide_user_symbols,
303 top->hide_kernel_symbols);
307 hists__collapse_resort(hists, NULL);
308 perf_evsel__output_resort(evsel, NULL);
310 hists__output_recalc_col_len(hists, top->print_entries - printed);
312 hists__fprintf(hists, false, top->print_entries - printed, win_width,
313 top->min_percent, stdout, !symbol_conf.use_callchain);
316 static void prompt_integer(int *target, const char *msg)
318 char *buf = malloc(0), *p;
322 fprintf(stdout, "\n%s: ", msg);
323 if (getline(&buf, &dummy, stdin) < 0)
326 p = strchr(buf, '\n');
336 tmp = strtoul(buf, NULL, 10);
342 static void prompt_percent(int *target, const char *msg)
346 prompt_integer(&tmp, msg);
347 if (tmp >= 0 && tmp <= 100)
351 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
353 char *buf = malloc(0), *p;
354 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
355 struct hists *hists = evsel__hists(top->sym_evsel);
356 struct rb_node *next;
359 /* zero counters of active symbol */
361 __zero_source_counters(syme);
362 top->sym_filter_entry = NULL;
365 fprintf(stdout, "\n%s: ", msg);
366 if (getline(&buf, &dummy, stdin) < 0)
369 p = strchr(buf, '\n');
373 next = rb_first_cached(&hists->entries);
375 n = rb_entry(next, struct hist_entry, rb_node);
376 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
380 next = rb_next(&n->rb_node);
384 fprintf(stderr, "Sorry, %s is not active.\n", buf);
387 perf_top__parse_source(top, found);
393 static void perf_top__print_mapped_keys(struct perf_top *top)
397 if (top->sym_filter_entry) {
398 struct symbol *sym = top->sym_filter_entry->ms.sym;
402 fprintf(stdout, "\nMapped keys:\n");
403 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
404 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
406 if (top->evlist->core.nr_entries > 1)
407 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
409 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
411 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
412 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
413 fprintf(stdout, "\t[S] stop annotation.\n");
416 "\t[K] hide kernel symbols. \t(%s)\n",
417 top->hide_kernel_symbols ? "yes" : "no");
419 "\t[U] hide user symbols. \t(%s)\n",
420 top->hide_user_symbols ? "yes" : "no");
421 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
422 fprintf(stdout, "\t[qQ] quit.\n");
425 static int perf_top__key_mapped(struct perf_top *top, int c)
441 return top->evlist->core.nr_entries > 1 ? 1 : 0;
449 static bool perf_top__handle_keypress(struct perf_top *top, int c)
453 if (!perf_top__key_mapped(top, c)) {
454 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
457 perf_top__print_mapped_keys(top);
458 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
461 set_term_quiet_input(&save);
463 poll(&stdin_poll, 1, -1);
466 tcsetattr(0, TCSAFLUSH, &save);
467 if (!perf_top__key_mapped(top, c))
473 prompt_integer(&top->delay_secs, "Enter display delay");
474 if (top->delay_secs < 1)
478 prompt_integer(&top->print_entries, "Enter display entries (lines)");
479 if (top->print_entries == 0) {
480 perf_top__resize(top);
481 signal(SIGWINCH, winch_sig);
483 signal(SIGWINCH, SIG_DFL);
487 if (top->evlist->core.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->core.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->annotation_opts.min_pcnt,
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 evsel *evsel = t->sym_evsel;
559 if (t->evlist->selected != NULL)
560 t->sym_evsel = t->evlist->selected;
562 hists = evsel__hists(evsel);
564 if (t->evlist->enabled) {
566 hists__delete_entries(hists);
568 hists__decay_entries(hists, t->hide_user_symbols,
569 t->hide_kernel_symbols);
573 hists__collapse_resort(hists, NULL);
574 perf_evsel__output_resort(evsel, NULL);
576 if (t->lost || t->drop)
577 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
580 static void stop_top(void)
586 static void *display_thread_tui(void *arg)
589 struct perf_top *top = arg;
590 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
591 struct hist_browser_timer hbt = {
592 .timer = perf_top__sort_new_samples,
594 .refresh = top->delay_secs,
597 /* In order to read symbols from other namespaces perf to needs to call
598 * setns(2). This isn't permitted if the struct_fs has multiple users.
599 * unshare(2) the fs so that we may continue to setns into namespaces
600 * that we're observing.
604 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
606 perf_top__sort_new_samples(top);
609 * Initialize the uid_filter_str, in the future the TUI will allow
610 * Zooming in/out UIDs. For now just use whatever the user passed
613 evlist__for_each_entry(top->evlist, pos) {
614 struct hists *hists = evsel__hists(pos);
615 hists->uid_filter_str = top->record_opts.target.uid_str;
618 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
620 &top->session->header.env,
621 !top->record_opts.overwrite,
622 &top->annotation_opts);
628 static void display_sig(int sig __maybe_unused)
633 static void display_setup_sig(void)
635 signal(SIGSEGV, sighandler_dump_stack);
636 signal(SIGFPE, sighandler_dump_stack);
637 signal(SIGINT, display_sig);
638 signal(SIGQUIT, display_sig);
639 signal(SIGTERM, display_sig);
642 static void *display_thread(void *arg)
644 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
646 struct perf_top *top = arg;
649 /* In order to read symbols from other namespaces perf to needs to call
650 * setns(2). This isn't permitted if the struct_fs has multiple users.
651 * unshare(2) the fs so that we may continue to setns into namespaces
652 * that we're observing.
656 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
659 pthread__unblock_sigwinch();
661 delay_msecs = top->delay_secs * MSEC_PER_SEC;
662 set_term_quiet_input(&save);
667 perf_top__print_sym_table(top);
669 * Either timeout expired or we got an EINTR due to SIGWINCH,
670 * refresh screen in both cases.
672 switch (poll(&stdin_poll, 1, delay_msecs)) {
681 tcsetattr(0, TCSAFLUSH, &save);
683 if (perf_top__handle_keypress(top, c))
689 tcsetattr(0, TCSAFLUSH, &save);
693 static int hist_iter__top_callback(struct hist_entry_iter *iter,
694 struct addr_location *al, bool single,
697 struct perf_top *top = arg;
698 struct hist_entry *he = iter->he;
699 struct evsel *evsel = iter->evsel;
701 if (perf_hpp_list.sym && single)
702 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
704 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
705 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
709 static void perf_event__process_sample(struct perf_tool *tool,
710 const union perf_event *event,
712 struct perf_sample *sample,
713 struct machine *machine)
715 struct perf_top *top = container_of(tool, struct perf_top, tool);
716 struct addr_location al;
719 if (!machine && perf_guest) {
720 static struct intlist *seen;
723 seen = intlist__new(NULL);
725 if (!intlist__has_entry(seen, sample->pid)) {
726 pr_err("Can't find guest [%d]'s kernel information\n",
728 intlist__add(seen, sample->pid);
734 pr_err("%u unprocessable samples recorded.\r",
735 top->session->evlist->stats.nr_unprocessable_samples++);
739 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
740 top->exact_samples++;
742 if (machine__resolve(machine, &al, sample) < 0)
745 if (!machine->kptr_restrict_warned &&
746 symbol_conf.kptr_restrict &&
747 al.cpumode == PERF_RECORD_MISC_KERNEL) {
748 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
750 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
751 "Check /proc/sys/kernel/kptr_restrict.\n\n"
752 "Kernel%s samples will not be resolved.\n",
753 al.map && map__has_symbols(al.map) ?
755 if (use_browser <= 0)
758 machine->kptr_restrict_warned = true;
761 if (al.sym == NULL && al.map != NULL) {
762 const char *msg = "Kernel samples will not be resolved.\n";
764 * As we do lazy loading of symtabs we only will know if the
765 * specified vmlinux file is invalid when we actually have a
766 * hit in kernel space and then try to load it. So if we get
767 * here and there are _no_ symbols in the DSO backing the
768 * kernel map, bail out.
770 * We may never get here, for instance, if we use -K/
771 * --hide-kernel-symbols, even if the user specifies an
772 * invalid --vmlinux ;-)
774 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
775 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
776 if (symbol_conf.vmlinux_name) {
778 dso__strerror_load(al.map->dso, serr, sizeof(serr));
779 ui__warning("The %s file can't be used: %s\n%s",
780 symbol_conf.vmlinux_name, serr, msg);
782 ui__warning("A vmlinux file was not found.\n%s",
786 if (use_browser <= 0)
788 top->vmlinux_warned = true;
792 if (al.sym == NULL || !al.sym->idle) {
793 struct hists *hists = evsel__hists(evsel);
794 struct hist_entry_iter iter = {
797 .add_entry_cb = hist_iter__top_callback,
800 if (symbol_conf.cumulate_callchain)
801 iter.ops = &hist_iter_cumulative;
803 iter.ops = &hist_iter_normal;
805 pthread_mutex_lock(&hists->lock);
807 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
809 pr_err("Problem incrementing symbol period, skipping event\n");
811 pthread_mutex_unlock(&hists->lock);
814 addr_location__put(&al);
818 perf_top__process_lost(struct perf_top *top, union perf_event *event,
821 struct hists *hists = evsel__hists(evsel);
823 top->lost += event->lost.lost;
824 top->lost_total += event->lost.lost;
825 hists->stats.total_lost += event->lost.lost;
829 perf_top__process_lost_samples(struct perf_top *top,
830 union perf_event *event,
833 struct hists *hists = evsel__hists(evsel);
835 top->lost += event->lost_samples.lost;
836 top->lost_total += event->lost_samples.lost;
837 hists->stats.total_lost_samples += event->lost_samples.lost;
840 static u64 last_timestamp;
842 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
844 struct record_opts *opts = &top->record_opts;
845 struct evlist *evlist = top->evlist;
846 struct perf_mmap *md;
847 union perf_event *event;
849 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
850 if (perf_mmap__read_init(md) < 0)
853 while ((event = perf_mmap__read_event(md)) != NULL) {
856 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
857 if (ret && ret != -1)
860 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
864 perf_mmap__consume(md);
866 if (top->qe.rotate) {
867 pthread_mutex_lock(&top->qe.mutex);
868 top->qe.rotate = false;
869 pthread_cond_signal(&top->qe.cond);
870 pthread_mutex_unlock(&top->qe.mutex);
874 perf_mmap__read_done(md);
877 static void perf_top__mmap_read(struct perf_top *top)
879 bool overwrite = top->record_opts.overwrite;
880 struct evlist *evlist = top->evlist;
884 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
886 for (i = 0; i < top->evlist->nr_mmaps; i++)
887 perf_top__mmap_read_idx(top, i);
890 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
891 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
896 * Check per-event overwrite term.
897 * perf top should support consistent term for all events.
898 * - All events don't have per-event term
899 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
900 * Nothing change, return 0.
901 * - All events have same per-event term
902 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
903 * Using the per-event setting to replace the opts->overwrite if
904 * they are different, then return 0.
905 * - Events have different per-event term
906 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
908 * - Some of the event set per-event term, but some not.
909 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
912 static int perf_top__overwrite_check(struct perf_top *top)
914 struct record_opts *opts = &top->record_opts;
915 struct evlist *evlist = top->evlist;
916 struct perf_evsel_config_term *term;
917 struct list_head *config_terms;
919 int set, overwrite = -1;
921 evlist__for_each_entry(evlist, evsel) {
923 config_terms = &evsel->config_terms;
924 list_for_each_entry(term, config_terms, list) {
925 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
926 set = term->val.overwrite ? 1 : 0;
929 /* no term for current and previous event (likely) */
930 if ((overwrite < 0) && (set < 0))
933 /* has term for both current and previous event, compare */
934 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
937 /* no term for current event but has term for previous one */
938 if ((overwrite >= 0) && (set < 0))
941 /* has term for current event */
942 if ((overwrite < 0) && (set >= 0)) {
943 /* if it's first event, set overwrite */
944 if (evsel == perf_evlist__first(evlist))
951 if ((overwrite >= 0) && (opts->overwrite != overwrite))
952 opts->overwrite = overwrite;
957 static int perf_top_overwrite_fallback(struct perf_top *top,
960 struct record_opts *opts = &top->record_opts;
961 struct evlist *evlist = top->evlist;
962 struct evsel *counter;
964 if (!opts->overwrite)
967 /* only fall back when first event fails */
968 if (evsel != perf_evlist__first(evlist))
971 evlist__for_each_entry(evlist, counter)
972 counter->core.attr.write_backward = false;
973 opts->overwrite = false;
974 pr_debug2("fall back to non-overwrite mode\n");
978 static int perf_top__start_counters(struct perf_top *top)
981 struct evsel *counter;
982 struct evlist *evlist = top->evlist;
983 struct record_opts *opts = &top->record_opts;
985 if (perf_top__overwrite_check(top)) {
986 ui__error("perf top only support consistent per-event "
987 "overwrite setting for all events\n");
991 perf_evlist__config(evlist, opts, &callchain_param);
993 evlist__for_each_entry(evlist, counter) {
995 if (evsel__open(counter, top->evlist->core.cpus,
996 top->evlist->core.threads) < 0) {
999 * Specially handle overwrite fall back.
1000 * Because perf top is the only tool which has
1001 * overwrite mode by default, support
1002 * both overwrite and non-overwrite mode, and
1003 * require consistent mode for all events.
1005 * May move it to generic code with more tools
1006 * have similar attribute.
1008 if (perf_missing_features.write_backward &&
1009 perf_top_overwrite_fallback(top, counter))
1012 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1014 ui__warning("%s\n", msg);
1018 perf_evsel__open_strerror(counter, &opts->target,
1019 errno, msg, sizeof(msg));
1020 ui__error("%s\n", msg);
1025 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1026 ui__error("Failed to mmap with %d (%s)\n",
1027 errno, str_error_r(errno, msg, sizeof(msg)));
1037 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1039 if (callchain->mode != CHAIN_NONE) {
1040 if (callchain_register_param(callchain) < 0) {
1041 ui__error("Can't register callchain params.\n");
1049 static struct ordered_events *rotate_queues(struct perf_top *top)
1051 struct ordered_events *in = top->qe.in;
1053 if (top->qe.in == &top->qe.data[1])
1054 top->qe.in = &top->qe.data[0];
1056 top->qe.in = &top->qe.data[1];
1061 static void *process_thread(void *arg)
1063 struct perf_top *top = arg;
1066 struct ordered_events *out, *in = top->qe.in;
1068 if (!in->nr_events) {
1073 out = rotate_queues(top);
1075 pthread_mutex_lock(&top->qe.mutex);
1076 top->qe.rotate = true;
1077 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1078 pthread_mutex_unlock(&top->qe.mutex);
1080 if (ordered_events__flush(out, OE_FLUSH__TOP))
1081 pr_err("failed to process events\n");
1088 * Allow only 'top->delay_secs' seconds behind samples.
1090 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1092 union perf_event *event = qevent->event;
1093 u64 delay_timestamp;
1095 if (event->header.type != PERF_RECORD_SAMPLE)
1098 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1099 return delay_timestamp < last_timestamp;
1102 static int deliver_event(struct ordered_events *qe,
1103 struct ordered_event *qevent)
1105 struct perf_top *top = qe->data;
1106 struct evlist *evlist = top->evlist;
1107 struct perf_session *session = top->session;
1108 union perf_event *event = qevent->event;
1109 struct perf_sample sample;
1110 struct evsel *evsel;
1111 struct machine *machine;
1114 if (should_drop(qevent, top)) {
1120 ret = perf_evlist__parse_sample(evlist, event, &sample);
1122 pr_err("Can't parse sample, err = %d\n", ret);
1126 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1127 assert(evsel != NULL);
1129 if (event->header.type == PERF_RECORD_SAMPLE)
1132 switch (sample.cpumode) {
1133 case PERF_RECORD_MISC_USER:
1135 if (top->hide_user_symbols)
1137 machine = &session->machines.host;
1139 case PERF_RECORD_MISC_KERNEL:
1140 ++top->kernel_samples;
1141 if (top->hide_kernel_symbols)
1143 machine = &session->machines.host;
1145 case PERF_RECORD_MISC_GUEST_KERNEL:
1146 ++top->guest_kernel_samples;
1147 machine = perf_session__find_machine(session,
1150 case PERF_RECORD_MISC_GUEST_USER:
1151 ++top->guest_us_samples;
1153 * TODO: we don't process guest user from host side
1154 * except simple counting.
1158 if (event->header.type == PERF_RECORD_SAMPLE)
1160 machine = &session->machines.host;
1164 if (event->header.type == PERF_RECORD_SAMPLE) {
1165 perf_event__process_sample(&top->tool, event, evsel,
1167 } else if (event->header.type == PERF_RECORD_LOST) {
1168 perf_top__process_lost(top, event, evsel);
1169 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1170 perf_top__process_lost_samples(top, event, evsel);
1171 } else if (event->header.type < PERF_RECORD_MAX) {
1172 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1173 machine__process_event(machine, event, &sample);
1175 ++session->evlist->stats.nr_unknown_events;
1182 static void init_process_thread(struct perf_top *top)
1184 ordered_events__init(&top->qe.data[0], deliver_event, top);
1185 ordered_events__init(&top->qe.data[1], deliver_event, top);
1186 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1187 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1188 top->qe.in = &top->qe.data[0];
1189 pthread_mutex_init(&top->qe.mutex, NULL);
1190 pthread_cond_init(&top->qe.cond, NULL);
1193 static int __cmd_top(struct perf_top *top)
1195 struct record_opts *opts = &top->record_opts;
1196 pthread_t thread, thread_process;
1199 if (!top->annotation_opts.objdump_path) {
1200 ret = perf_env__lookup_objdump(&top->session->header.env,
1201 &top->annotation_opts.objdump_path);
1206 ret = callchain_param__setup_sample_type(&callchain_param);
1210 if (perf_session__register_idle_thread(top->session) < 0)
1213 if (top->nr_threads_synthesize > 1)
1214 perf_set_multithreaded();
1216 init_process_thread(top);
1218 if (opts->record_namespaces)
1219 top->tool.namespace_events = true;
1221 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1222 &top->session->machines.host,
1225 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1227 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1228 top->evlist->core.threads, false,
1229 top->nr_threads_synthesize);
1231 if (top->nr_threads_synthesize > 1)
1232 perf_set_singlethreaded();
1234 if (perf_hpp_list.socket) {
1235 ret = perf_env__read_cpu_topology_map(&perf_env);
1237 char errbuf[BUFSIZ];
1238 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1240 ui__error("Could not read the CPU topology map: %s\n", err);
1245 ret = perf_top__start_counters(top);
1249 top->session->evlist = top->evlist;
1250 perf_session__set_id_hdr_size(top->session);
1253 * When perf is starting the traced process, all the events (apart from
1254 * group members) have enable_on_exec=1 set, so don't spoil it by
1255 * prematurely enabling them.
1257 * XXX 'top' still doesn't start workloads like record, trace, but should,
1258 * so leave the check here.
1260 if (!target__none(&opts->target))
1261 evlist__enable(top->evlist);
1264 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1265 ui__error("Could not create process thread.\n");
1269 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1270 display_thread), top)) {
1271 ui__error("Could not create display thread.\n");
1272 goto out_join_thread;
1275 if (top->realtime_prio) {
1276 struct sched_param param;
1278 param.sched_priority = top->realtime_prio;
1279 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1280 ui__error("Could not set realtime priority.\n");
1285 /* Wait for a minimal set of events before starting the snapshot */
1286 perf_evlist__poll(top->evlist, 100);
1288 perf_top__mmap_read(top);
1291 u64 hits = top->samples;
1293 perf_top__mmap_read(top);
1295 if (opts->overwrite || (hits == top->samples))
1296 ret = perf_evlist__poll(top->evlist, 100);
1299 perf_top__resize(top);
1306 pthread_join(thread, NULL);
1308 pthread_cond_signal(&top->qe.cond);
1309 pthread_join(thread_process, NULL);
1314 callchain_opt(const struct option *opt, const char *arg, int unset)
1316 symbol_conf.use_callchain = true;
1317 return record_callchain_opt(opt, arg, unset);
1321 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1323 struct callchain_param *callchain = opt->value;
1325 callchain->enabled = !unset;
1326 callchain->record_mode = CALLCHAIN_FP;
1332 symbol_conf.use_callchain = false;
1333 callchain->record_mode = CALLCHAIN_NONE;
1337 return parse_callchain_top_opt(arg);
1340 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1342 if (!strcmp(var, "top.call-graph")) {
1343 var = "call-graph.record-mode";
1344 return perf_default_config(var, value, cb);
1346 if (!strcmp(var, "top.children")) {
1347 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1355 parse_percent_limit(const struct option *opt, const char *arg,
1356 int unset __maybe_unused)
1358 struct perf_top *top = opt->value;
1360 top->min_percent = strtof(arg, NULL);
1364 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1365 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1367 int cmd_top(int argc, const char **argv)
1369 char errbuf[BUFSIZ];
1370 struct perf_top top = {
1374 .mmap_pages = UINT_MAX,
1375 .user_freq = UINT_MAX,
1376 .user_interval = ULLONG_MAX,
1377 .freq = 4000, /* 4 KHz */
1382 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1383 * when we pause, fix that and reenable. Probably using a
1384 * separate evlist with a dummy event, i.e. a non-overwrite
1385 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1386 * stays in overwrite mode. -acme
1389 .sample_time = true,
1390 .sample_time_set = true,
1392 .max_stack = sysctl__max_stack(),
1393 .annotation_opts = annotation__default_options,
1394 .nr_threads_synthesize = UINT_MAX,
1396 struct record_opts *opts = &top.record_opts;
1397 struct target *target = &opts->target;
1398 const struct option options[] = {
1399 OPT_CALLBACK('e', "event", &top.evlist, "event",
1400 "event selector. use 'perf list' to list available events",
1401 parse_events_option),
1402 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1403 OPT_STRING('p', "pid", &target->pid, "pid",
1404 "profile events on existing process id"),
1405 OPT_STRING('t', "tid", &target->tid, "tid",
1406 "profile events on existing thread id"),
1407 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1408 "system-wide collection from all CPUs"),
1409 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1410 "list of cpus to monitor"),
1411 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1412 "file", "vmlinux pathname"),
1413 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1414 "don't load vmlinux even if found"),
1415 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1416 "file", "kallsyms pathname"),
1417 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1418 "hide kernel symbols"),
1419 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1420 "number of mmap data pages",
1421 perf_evlist__parse_mmap_pages),
1422 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1423 "collect data with this RT SCHED_FIFO priority"),
1424 OPT_INTEGER('d', "delay", &top.delay_secs,
1425 "number of seconds to delay between refreshes"),
1426 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1427 "dump the symbol table used for profiling"),
1428 OPT_INTEGER('f', "count-filter", &top.count_filter,
1429 "only display functions with more events than this"),
1430 OPT_BOOLEAN(0, "group", &opts->group,
1431 "put the counters into a counter group"),
1432 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1433 "child tasks do not inherit counters"),
1434 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1435 "symbol to annotate"),
1436 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1437 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1438 "profile at this frequency",
1439 record__parse_freq),
1440 OPT_INTEGER('E', "entries", &top.print_entries,
1441 "display this many functions"),
1442 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1443 "hide user symbols"),
1444 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1445 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1446 OPT_INCR('v', "verbose", &verbose,
1447 "be more verbose (show counter open errors, etc)"),
1448 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1449 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1450 " Please refer the man page for the complete list."),
1451 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1452 "output field(s): overhead, period, sample plus all of sort keys"),
1453 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1454 "Show a column with the number of samples"),
1455 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1456 NULL, "enables call-graph recording and display",
1458 OPT_CALLBACK(0, "call-graph", &callchain_param,
1459 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1460 top_callchain_help, &parse_callchain_opt),
1461 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1462 "Accumulate callchains of children and show total overhead as well"),
1463 OPT_INTEGER(0, "max-stack", &top.max_stack,
1464 "Set the maximum stack depth when parsing the callchain. "
1465 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1466 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1467 "ignore callees of these functions in call graphs",
1468 report_parse_ignore_callees_opt),
1469 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1470 "Show a column with the sum of periods"),
1471 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1472 "only consider symbols in these dsos"),
1473 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1474 "only consider symbols in these comms"),
1475 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1476 "only consider these symbols"),
1477 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1478 "Interleave source code with assembly code (default)"),
1479 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1480 "Display raw encoding of assembly instructions (default)"),
1481 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1482 "Enable kernel symbol demangling"),
1483 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1484 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1485 "objdump binary to use for disassembly and annotations"),
1486 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1487 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1488 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1489 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1490 "Don't show entries under that percent", parse_percent_limit),
1491 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1492 "How to display percentage of filtered entries", parse_filter_percentage),
1493 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1495 "don't try to adjust column width, use these fixed values"),
1496 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1497 "per thread proc mmap processing timeout in ms"),
1498 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1499 "branch any", "sample any taken branches",
1500 parse_branch_stack),
1501 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1502 "branch filter mask", "branch stack filter modes",
1503 parse_branch_stack),
1504 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1505 "Show raw trace event output (do not use print fmt or plugins)"),
1506 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1507 "Show entries in a hierarchy"),
1508 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1509 "Use a backward ring buffer, default: no"),
1510 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1511 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1512 "number of thread to run event synthesize"),
1513 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1514 "Record namespaces events"),
1517 struct evlist *sb_evlist = NULL;
1518 const char * const top_usage[] = {
1519 "perf top [<options>]",
1522 int status = hists__init();
1527 top.annotation_opts.min_pcnt = 5;
1528 top.annotation_opts.context = 4;
1530 top.evlist = evlist__new();
1531 if (top.evlist == NULL)
1534 status = perf_config(perf_top_config, &top);
1538 argc = parse_options(argc, argv, options, top_usage, 0);
1540 usage_with_options(top_usage, options);
1542 if (!top.evlist->core.nr_entries &&
1543 perf_evlist__add_default(top.evlist) < 0) {
1544 pr_err("Not enough memory for event selector list\n");
1545 goto out_delete_evlist;
1548 if (symbol_conf.report_hierarchy) {
1549 /* disable incompatible options */
1550 symbol_conf.event_group = false;
1551 symbol_conf.cumulate_callchain = false;
1554 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1555 parse_options_usage(top_usage, options, "fields", 0);
1556 parse_options_usage(NULL, options, "hierarchy", 0);
1557 goto out_delete_evlist;
1561 if (opts->branch_stack && callchain_param.enabled)
1562 symbol_conf.show_branchflag_count = true;
1564 sort__mode = SORT_MODE__TOP;
1565 /* display thread wants entries to be collapsed in a different tree */
1566 perf_hpp_list.need_collapse = 1;
1570 else if (top.use_tui)
1573 setup_browser(false);
1575 if (setup_sorting(top.evlist) < 0) {
1577 parse_options_usage(top_usage, options, "s", 1);
1579 parse_options_usage(sort_order ? NULL : top_usage,
1580 options, "fields", 0);
1581 goto out_delete_evlist;
1584 status = target__validate(target);
1586 target__strerror(target, status, errbuf, BUFSIZ);
1587 ui__warning("%s\n", errbuf);
1590 status = target__parse_uid(target);
1592 int saved_errno = errno;
1594 target__strerror(target, status, errbuf, BUFSIZ);
1595 ui__error("%s\n", errbuf);
1597 status = -saved_errno;
1598 goto out_delete_evlist;
1601 if (target__none(target))
1602 target->system_wide = true;
1604 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1605 ui__error("Couldn't create thread/CPU maps: %s\n",
1606 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1607 goto out_delete_evlist;
1610 if (top.delay_secs < 1)
1613 if (record_opts__config(opts)) {
1615 goto out_delete_evlist;
1618 top.sym_evsel = perf_evlist__first(top.evlist);
1620 if (!callchain_param.enabled) {
1621 symbol_conf.cumulate_callchain = false;
1622 perf_hpp__cancel_cumulate();
1625 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1626 callchain_param.order = ORDER_CALLER;
1628 status = symbol__annotation_init();
1630 goto out_delete_evlist;
1632 annotation_config__init();
1634 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1635 status = symbol__init(NULL);
1637 goto out_delete_evlist;
1639 sort__setup_elide(stdout);
1641 get_term_dimensions(&top.winsize);
1642 if (top.print_entries == 0) {
1643 perf_top__update_print_entries(&top);
1644 signal(SIGWINCH, winch_sig);
1647 top.session = perf_session__new(NULL, false, NULL);
1648 if (top.session == NULL) {
1650 goto out_delete_evlist;
1653 if (!top.record_opts.no_bpf_event)
1654 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1656 if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1657 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1658 opts->no_bpf_event = true;
1661 status = __cmd_top(&top);
1663 if (!opts->no_bpf_event)
1664 perf_evlist__stop_sb_thread(sb_evlist);
1667 evlist__delete(top.evlist);
1668 perf_session__delete(top.session);