4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10 * Improvements and fixes by:
12 * Arjan van de Ven <arjan@linux.intel.com>
13 * Yanmin Zhang <yanmin.zhang@intel.com>
14 * Wu Fengguang <fengguang.wu@intel.com>
15 * Mike Galbraith <efault@gmx.de>
16 * Paul Mackerras <paulus@samba.org>
18 * Released under the GPL v2. (and only v2, not any later version)
24 #include "util/annotate.h"
25 #include "util/config.h"
26 #include "util/color.h"
27 #include "util/drv_configs.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/event.h"
31 #include "util/machine.h"
32 #include "util/session.h"
33 #include "util/symbol.h"
34 #include "util/thread.h"
35 #include "util/thread_map.h"
37 #include <linux/rbtree.h>
38 #include <subcmd/parse-options.h>
39 #include "util/parse-events.h"
40 #include "util/cpumap.h"
41 #include "util/xyarray.h"
42 #include "util/sort.h"
43 #include "util/term.h"
44 #include "util/intlist.h"
45 #include "util/parse-branch-options.h"
46 #include "arch/common.h"
48 #include "util/debug.h"
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 "sane_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)
103 struct perf_evsel *evsel = hists_to_evsel(he->hists);
105 struct annotation *notes;
109 if (!he || !he->ms.sym)
116 * We can't annotate with just /proc/kallsyms
118 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
119 !dso__is_kcore(map->dso)) {
120 pr_err("Can't annotate %s: No vmlinux file was found in the "
121 "path\n", sym->name);
126 notes = symbol__annotation(sym);
127 pthread_mutex_lock(¬es->lock);
129 if (!symbol__hists(sym, top->evlist->nr_entries)) {
130 pthread_mutex_unlock(¬es->lock);
131 pr_err("Not enough memory for annotating '%s' symbol!\n",
137 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
139 top->sym_filter_entry = he;
142 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
143 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
146 pthread_mutex_unlock(¬es->lock);
150 static void __zero_source_counters(struct hist_entry *he)
152 struct symbol *sym = he->ms.sym;
153 symbol__annotate_zero_histograms(sym);
156 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
159 int err = uname(&uts);
161 ui__warning("Out of bounds address found:\n\n"
162 "Addr: %" PRIx64 "\n"
164 "Map: %" PRIx64 "-%" PRIx64 "\n"
165 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
169 "Not all samples will be on the annotation output.\n\n"
170 "Please report to linux-kernel@vger.kernel.org\n",
171 ip, map->dso->long_name, dso__symtab_origin(map->dso),
172 map->start, map->end, sym->start, sym->end,
173 sym->binding == STB_GLOBAL ? 'g' :
174 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
175 err ? "[unknown]" : uts.machine,
176 err ? "[unknown]" : uts.release, perf_version_string);
177 if (use_browser <= 0)
180 map->erange_warned = true;
183 static void perf_top__record_precise_ip(struct perf_top *top,
184 struct hist_entry *he,
185 struct perf_sample *sample,
186 struct perf_evsel *evsel, u64 ip)
188 struct annotation *notes;
189 struct symbol *sym = he->ms.sym;
192 if (sym == NULL || (use_browser == 0 &&
193 (top->sym_filter_entry == NULL ||
194 top->sym_filter_entry->ms.sym != sym)))
197 notes = symbol__annotation(sym);
199 if (pthread_mutex_trylock(¬es->lock))
202 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
204 pthread_mutex_unlock(¬es->lock);
208 * This function is now called with he->hists->lock held.
209 * Release it before going to sleep.
211 pthread_mutex_unlock(&he->hists->lock);
213 if (err == -ERANGE && !he->ms.map->erange_warned)
214 ui__warn_map_erange(he->ms.map, sym, ip);
215 else if (err == -ENOMEM) {
216 pr_err("Not enough memory for annotating '%s' symbol!\n",
221 pthread_mutex_lock(&he->hists->lock);
225 static void perf_top__show_details(struct perf_top *top)
227 struct hist_entry *he = top->sym_filter_entry;
228 struct perf_evsel *evsel = hists_to_evsel(he->hists);
229 struct annotation *notes;
230 struct symbol *symbol;
237 notes = symbol__annotation(symbol);
239 pthread_mutex_lock(¬es->lock);
241 symbol__calc_percent(symbol, evsel);
243 if (notes->src == NULL)
246 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
247 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
249 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
251 if (top->evlist->enabled) {
253 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
255 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
258 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
260 pthread_mutex_unlock(¬es->lock);
263 static void perf_top__print_sym_table(struct perf_top *top)
267 const int win_width = top->winsize.ws_col - 1;
268 struct perf_evsel *evsel = top->sym_evsel;
269 struct hists *hists = evsel__hists(evsel);
273 perf_top__header_snprintf(top, bf, sizeof(bf));
276 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
278 if (!top->record_opts.overwrite &&
279 (hists->stats.nr_lost_warned !=
280 hists->stats.nr_events[PERF_RECORD_LOST])) {
281 hists->stats.nr_lost_warned =
282 hists->stats.nr_events[PERF_RECORD_LOST];
283 color_fprintf(stdout, PERF_COLOR_RED,
284 "WARNING: LOST %d chunks, Check IO/CPU overload",
285 hists->stats.nr_lost_warned);
289 if (top->sym_filter_entry) {
290 perf_top__show_details(top);
294 if (top->evlist->enabled) {
296 hists__delete_entries(hists);
298 hists__decay_entries(hists, top->hide_user_symbols,
299 top->hide_kernel_symbols);
303 hists__collapse_resort(hists, NULL);
304 perf_evsel__output_resort(evsel, NULL);
306 hists__output_recalc_col_len(hists, top->print_entries - printed);
308 hists__fprintf(hists, false, top->print_entries - printed, win_width,
309 top->min_percent, stdout, !symbol_conf.use_callchain);
312 static void prompt_integer(int *target, const char *msg)
314 char *buf = malloc(0), *p;
318 fprintf(stdout, "\n%s: ", msg);
319 if (getline(&buf, &dummy, stdin) < 0)
322 p = strchr(buf, '\n');
332 tmp = strtoul(buf, NULL, 10);
338 static void prompt_percent(int *target, const char *msg)
342 prompt_integer(&tmp, msg);
343 if (tmp >= 0 && tmp <= 100)
347 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
349 char *buf = malloc(0), *p;
350 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
351 struct hists *hists = evsel__hists(top->sym_evsel);
352 struct rb_node *next;
355 /* zero counters of active symbol */
357 __zero_source_counters(syme);
358 top->sym_filter_entry = NULL;
361 fprintf(stdout, "\n%s: ", msg);
362 if (getline(&buf, &dummy, stdin) < 0)
365 p = strchr(buf, '\n');
369 next = rb_first(&hists->entries);
371 n = rb_entry(next, struct hist_entry, rb_node);
372 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
376 next = rb_next(&n->rb_node);
380 fprintf(stderr, "Sorry, %s is not active.\n", buf);
383 perf_top__parse_source(top, found);
389 static void perf_top__print_mapped_keys(struct perf_top *top)
393 if (top->sym_filter_entry) {
394 struct symbol *sym = top->sym_filter_entry->ms.sym;
398 fprintf(stdout, "\nMapped keys:\n");
399 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
400 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
402 if (top->evlist->nr_entries > 1)
403 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
405 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
407 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
408 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
409 fprintf(stdout, "\t[S] stop annotation.\n");
412 "\t[K] hide kernel symbols. \t(%s)\n",
413 top->hide_kernel_symbols ? "yes" : "no");
415 "\t[U] hide user symbols. \t(%s)\n",
416 top->hide_user_symbols ? "yes" : "no");
417 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
418 fprintf(stdout, "\t[qQ] quit.\n");
421 static int perf_top__key_mapped(struct perf_top *top, int c)
437 return top->evlist->nr_entries > 1 ? 1 : 0;
445 static bool perf_top__handle_keypress(struct perf_top *top, int c)
449 if (!perf_top__key_mapped(top, c)) {
450 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
453 perf_top__print_mapped_keys(top);
454 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
457 set_term_quiet_input(&save);
459 poll(&stdin_poll, 1, -1);
462 tcsetattr(0, TCSAFLUSH, &save);
463 if (!perf_top__key_mapped(top, c))
469 prompt_integer(&top->delay_secs, "Enter display delay");
470 if (top->delay_secs < 1)
474 prompt_integer(&top->print_entries, "Enter display entries (lines)");
475 if (top->print_entries == 0) {
476 perf_top__resize(top);
477 signal(SIGWINCH, winch_sig);
479 signal(SIGWINCH, SIG_DFL);
483 if (top->evlist->nr_entries > 1) {
484 /* Select 0 as the default event: */
487 fprintf(stderr, "\nAvailable events:");
489 evlist__for_each_entry(top->evlist, top->sym_evsel)
490 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
492 prompt_integer(&counter, "Enter details event counter");
494 if (counter >= top->evlist->nr_entries) {
495 top->sym_evsel = perf_evlist__first(top->evlist);
496 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
500 evlist__for_each_entry(top->evlist, top->sym_evsel)
501 if (top->sym_evsel->idx == counter)
504 top->sym_evsel = perf_evlist__first(top->evlist);
507 prompt_integer(&top->count_filter, "Enter display event count filter");
510 prompt_percent(&top->annotation_opts.min_pcnt,
511 "Enter details display event filter (percent)");
514 top->hide_kernel_symbols = !top->hide_kernel_symbols;
518 printf("exiting.\n");
519 if (top->dump_symtab)
520 perf_session__fprintf_dsos(top->session, stderr);
524 perf_top__prompt_symbol(top, "Enter details symbol");
527 if (!top->sym_filter_entry)
530 struct hist_entry *syme = top->sym_filter_entry;
532 top->sym_filter_entry = NULL;
533 __zero_source_counters(syme);
537 top->hide_user_symbols = !top->hide_user_symbols;
540 top->zero = !top->zero;
549 static void perf_top__sort_new_samples(void *arg)
551 struct perf_top *t = arg;
552 struct perf_evsel *evsel = t->sym_evsel;
555 if (t->evlist->selected != NULL)
556 t->sym_evsel = t->evlist->selected;
558 hists = evsel__hists(evsel);
560 if (t->evlist->enabled) {
562 hists__delete_entries(hists);
564 hists__decay_entries(hists, t->hide_user_symbols,
565 t->hide_kernel_symbols);
569 hists__collapse_resort(hists, NULL);
570 perf_evsel__output_resort(evsel, NULL);
572 if (t->lost || t->drop)
573 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
576 static void stop_top(void)
582 static void *display_thread_tui(void *arg)
584 struct perf_evsel *pos;
585 struct perf_top *top = arg;
586 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
587 struct hist_browser_timer hbt = {
588 .timer = perf_top__sort_new_samples,
590 .refresh = top->delay_secs,
593 /* In order to read symbols from other namespaces perf to needs to call
594 * setns(2). This isn't permitted if the struct_fs has multiple users.
595 * unshare(2) the fs so that we may continue to setns into namespaces
596 * that we're observing.
600 perf_top__sort_new_samples(top);
603 * Initialize the uid_filter_str, in the future the TUI will allow
604 * Zooming in/out UIDs. For now just use whatever the user passed
607 evlist__for_each_entry(top->evlist, pos) {
608 struct hists *hists = evsel__hists(pos);
609 hists->uid_filter_str = top->record_opts.target.uid_str;
612 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
614 &top->session->header.env,
615 !top->record_opts.overwrite,
616 &top->annotation_opts);
622 static void display_sig(int sig __maybe_unused)
627 static void display_setup_sig(void)
629 signal(SIGSEGV, sighandler_dump_stack);
630 signal(SIGFPE, sighandler_dump_stack);
631 signal(SIGINT, display_sig);
632 signal(SIGQUIT, display_sig);
633 signal(SIGTERM, display_sig);
636 static void *display_thread(void *arg)
638 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
640 struct perf_top *top = arg;
643 /* In order to read symbols from other namespaces perf to needs to call
644 * setns(2). This isn't permitted if the struct_fs has multiple users.
645 * unshare(2) the fs so that we may continue to setns into namespaces
646 * that we're observing.
651 pthread__unblock_sigwinch();
653 delay_msecs = top->delay_secs * MSEC_PER_SEC;
654 set_term_quiet_input(&save);
659 perf_top__print_sym_table(top);
661 * Either timeout expired or we got an EINTR due to SIGWINCH,
662 * refresh screen in both cases.
664 switch (poll(&stdin_poll, 1, delay_msecs)) {
673 tcsetattr(0, TCSAFLUSH, &save);
675 if (perf_top__handle_keypress(top, c))
681 tcsetattr(0, TCSAFLUSH, &save);
685 static int hist_iter__top_callback(struct hist_entry_iter *iter,
686 struct addr_location *al, bool single,
689 struct perf_top *top = arg;
690 struct hist_entry *he = iter->he;
691 struct perf_evsel *evsel = iter->evsel;
693 if (perf_hpp_list.sym && single)
694 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
696 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
697 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
701 static void perf_event__process_sample(struct perf_tool *tool,
702 const union perf_event *event,
703 struct perf_evsel *evsel,
704 struct perf_sample *sample,
705 struct machine *machine)
707 struct perf_top *top = container_of(tool, struct perf_top, tool);
708 struct addr_location al;
711 if (!machine && perf_guest) {
712 static struct intlist *seen;
715 seen = intlist__new(NULL);
717 if (!intlist__has_entry(seen, sample->pid)) {
718 pr_err("Can't find guest [%d]'s kernel information\n",
720 intlist__add(seen, sample->pid);
726 pr_err("%u unprocessable samples recorded.\r",
727 top->session->evlist->stats.nr_unprocessable_samples++);
731 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
732 top->exact_samples++;
734 if (machine__resolve(machine, &al, sample) < 0)
737 if (!machine->kptr_restrict_warned &&
738 symbol_conf.kptr_restrict &&
739 al.cpumode == PERF_RECORD_MISC_KERNEL) {
740 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
742 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
743 "Check /proc/sys/kernel/kptr_restrict.\n\n"
744 "Kernel%s samples will not be resolved.\n",
745 al.map && map__has_symbols(al.map) ?
747 if (use_browser <= 0)
750 machine->kptr_restrict_warned = true;
753 if (al.sym == NULL && al.map != NULL) {
754 const char *msg = "Kernel samples will not be resolved.\n";
756 * As we do lazy loading of symtabs we only will know if the
757 * specified vmlinux file is invalid when we actually have a
758 * hit in kernel space and then try to load it. So if we get
759 * here and there are _no_ symbols in the DSO backing the
760 * kernel map, bail out.
762 * We may never get here, for instance, if we use -K/
763 * --hide-kernel-symbols, even if the user specifies an
764 * invalid --vmlinux ;-)
766 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
767 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
768 if (symbol_conf.vmlinux_name) {
770 dso__strerror_load(al.map->dso, serr, sizeof(serr));
771 ui__warning("The %s file can't be used: %s\n%s",
772 symbol_conf.vmlinux_name, serr, msg);
774 ui__warning("A vmlinux file was not found.\n%s",
778 if (use_browser <= 0)
780 top->vmlinux_warned = true;
784 if (al.sym == NULL || !al.sym->idle) {
785 struct hists *hists = evsel__hists(evsel);
786 struct hist_entry_iter iter = {
789 .add_entry_cb = hist_iter__top_callback,
792 if (symbol_conf.cumulate_callchain)
793 iter.ops = &hist_iter_cumulative;
795 iter.ops = &hist_iter_normal;
797 pthread_mutex_lock(&hists->lock);
799 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
801 pr_err("Problem incrementing symbol period, skipping event\n");
803 pthread_mutex_unlock(&hists->lock);
806 addr_location__put(&al);
810 perf_top__process_lost(struct perf_top *top, union perf_event *event,
811 struct perf_evsel *evsel)
813 struct hists *hists = evsel__hists(evsel);
815 top->lost += event->lost.lost;
816 top->lost_total += event->lost.lost;
817 hists->stats.total_lost += event->lost.lost;
821 perf_top__process_lost_samples(struct perf_top *top,
822 union perf_event *event,
823 struct perf_evsel *evsel)
825 struct hists *hists = evsel__hists(evsel);
827 top->lost += event->lost_samples.lost;
828 top->lost_total += event->lost_samples.lost;
829 hists->stats.total_lost_samples += event->lost_samples.lost;
832 static u64 last_timestamp;
834 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
836 struct record_opts *opts = &top->record_opts;
837 struct perf_evlist *evlist = top->evlist;
838 struct perf_mmap *md;
839 union perf_event *event;
841 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
842 if (perf_mmap__read_init(md) < 0)
845 while ((event = perf_mmap__read_event(md)) != NULL) {
848 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
849 if (ret && ret != -1)
852 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
856 perf_mmap__consume(md);
858 if (top->qe.rotate) {
859 pthread_mutex_lock(&top->qe.mutex);
860 top->qe.rotate = false;
861 pthread_cond_signal(&top->qe.cond);
862 pthread_mutex_unlock(&top->qe.mutex);
866 perf_mmap__read_done(md);
869 static void perf_top__mmap_read(struct perf_top *top)
871 bool overwrite = top->record_opts.overwrite;
872 struct perf_evlist *evlist = top->evlist;
876 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
878 for (i = 0; i < top->evlist->nr_mmaps; i++)
879 perf_top__mmap_read_idx(top, i);
882 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
883 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
888 * Check per-event overwrite term.
889 * perf top should support consistent term for all events.
890 * - All events don't have per-event term
891 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
892 * Nothing change, return 0.
893 * - All events have same per-event term
894 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
895 * Using the per-event setting to replace the opts->overwrite if
896 * they are different, then return 0.
897 * - Events have different per-event term
898 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
900 * - Some of the event set per-event term, but some not.
901 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
904 static int perf_top__overwrite_check(struct perf_top *top)
906 struct record_opts *opts = &top->record_opts;
907 struct perf_evlist *evlist = top->evlist;
908 struct perf_evsel_config_term *term;
909 struct list_head *config_terms;
910 struct perf_evsel *evsel;
911 int set, overwrite = -1;
913 evlist__for_each_entry(evlist, evsel) {
915 config_terms = &evsel->config_terms;
916 list_for_each_entry(term, config_terms, list) {
917 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
918 set = term->val.overwrite ? 1 : 0;
921 /* no term for current and previous event (likely) */
922 if ((overwrite < 0) && (set < 0))
925 /* has term for both current and previous event, compare */
926 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
929 /* no term for current event but has term for previous one */
930 if ((overwrite >= 0) && (set < 0))
933 /* has term for current event */
934 if ((overwrite < 0) && (set >= 0)) {
935 /* if it's first event, set overwrite */
936 if (evsel == perf_evlist__first(evlist))
943 if ((overwrite >= 0) && (opts->overwrite != overwrite))
944 opts->overwrite = overwrite;
949 static int perf_top_overwrite_fallback(struct perf_top *top,
950 struct perf_evsel *evsel)
952 struct record_opts *opts = &top->record_opts;
953 struct perf_evlist *evlist = top->evlist;
954 struct perf_evsel *counter;
956 if (!opts->overwrite)
959 /* only fall back when first event fails */
960 if (evsel != perf_evlist__first(evlist))
963 evlist__for_each_entry(evlist, counter)
964 counter->attr.write_backward = false;
965 opts->overwrite = false;
966 pr_debug2("fall back to non-overwrite mode\n");
970 static int perf_top__start_counters(struct perf_top *top)
973 struct perf_evsel *counter;
974 struct perf_evlist *evlist = top->evlist;
975 struct record_opts *opts = &top->record_opts;
977 if (perf_top__overwrite_check(top)) {
978 ui__error("perf top only support consistent per-event "
979 "overwrite setting for all events\n");
983 perf_evlist__config(evlist, opts, &callchain_param);
985 evlist__for_each_entry(evlist, counter) {
987 if (perf_evsel__open(counter, top->evlist->cpus,
988 top->evlist->threads) < 0) {
991 * Specially handle overwrite fall back.
992 * Because perf top is the only tool which has
993 * overwrite mode by default, support
994 * both overwrite and non-overwrite mode, and
995 * require consistent mode for all events.
997 * May move it to generic code with more tools
998 * have similar attribute.
1000 if (perf_missing_features.write_backward &&
1001 perf_top_overwrite_fallback(top, counter))
1004 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1006 ui__warning("%s\n", msg);
1010 perf_evsel__open_strerror(counter, &opts->target,
1011 errno, msg, sizeof(msg));
1012 ui__error("%s\n", msg);
1017 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1018 ui__error("Failed to mmap with %d (%s)\n",
1019 errno, str_error_r(errno, msg, sizeof(msg)));
1029 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1031 if (!perf_hpp_list.sym) {
1032 if (callchain->enabled) {
1033 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1036 } else if (callchain->mode != CHAIN_NONE) {
1037 if (callchain_register_param(callchain) < 0) {
1038 ui__error("Can't register callchain params.\n");
1046 static struct ordered_events *rotate_queues(struct perf_top *top)
1048 struct ordered_events *in = top->qe.in;
1050 if (top->qe.in == &top->qe.data[1])
1051 top->qe.in = &top->qe.data[0];
1053 top->qe.in = &top->qe.data[1];
1058 static void *process_thread(void *arg)
1060 struct perf_top *top = arg;
1063 struct ordered_events *out, *in = top->qe.in;
1065 if (!in->nr_events) {
1070 out = rotate_queues(top);
1072 pthread_mutex_lock(&top->qe.mutex);
1073 top->qe.rotate = true;
1074 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1075 pthread_mutex_unlock(&top->qe.mutex);
1077 if (ordered_events__flush(out, OE_FLUSH__TOP))
1078 pr_err("failed to process events\n");
1085 * Allow only 'top->delay_secs' seconds behind samples.
1087 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1089 union perf_event *event = qevent->event;
1090 u64 delay_timestamp;
1092 if (event->header.type != PERF_RECORD_SAMPLE)
1095 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1096 return delay_timestamp < last_timestamp;
1099 static int deliver_event(struct ordered_events *qe,
1100 struct ordered_event *qevent)
1102 struct perf_top *top = qe->data;
1103 struct perf_evlist *evlist = top->evlist;
1104 struct perf_session *session = top->session;
1105 union perf_event *event = qevent->event;
1106 struct perf_sample sample;
1107 struct perf_evsel *evsel;
1108 struct machine *machine;
1111 if (should_drop(qevent, top)) {
1117 ret = perf_evlist__parse_sample(evlist, event, &sample);
1119 pr_err("Can't parse sample, err = %d\n", ret);
1123 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1124 assert(evsel != NULL);
1126 if (event->header.type == PERF_RECORD_SAMPLE)
1129 switch (sample.cpumode) {
1130 case PERF_RECORD_MISC_USER:
1132 if (top->hide_user_symbols)
1134 machine = &session->machines.host;
1136 case PERF_RECORD_MISC_KERNEL:
1137 ++top->kernel_samples;
1138 if (top->hide_kernel_symbols)
1140 machine = &session->machines.host;
1142 case PERF_RECORD_MISC_GUEST_KERNEL:
1143 ++top->guest_kernel_samples;
1144 machine = perf_session__find_machine(session,
1147 case PERF_RECORD_MISC_GUEST_USER:
1148 ++top->guest_us_samples;
1150 * TODO: we don't process guest user from host side
1151 * except simple counting.
1155 if (event->header.type == PERF_RECORD_SAMPLE)
1157 machine = &session->machines.host;
1161 if (event->header.type == PERF_RECORD_SAMPLE) {
1162 perf_event__process_sample(&top->tool, event, evsel,
1164 } else if (event->header.type == PERF_RECORD_LOST) {
1165 perf_top__process_lost(top, event, evsel);
1166 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1167 perf_top__process_lost_samples(top, event, evsel);
1168 } else if (event->header.type < PERF_RECORD_MAX) {
1169 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1170 machine__process_event(machine, event, &sample);
1172 ++session->evlist->stats.nr_unknown_events;
1179 static void init_process_thread(struct perf_top *top)
1181 ordered_events__init(&top->qe.data[0], deliver_event, top);
1182 ordered_events__init(&top->qe.data[1], deliver_event, top);
1183 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1184 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1185 top->qe.in = &top->qe.data[0];
1186 pthread_mutex_init(&top->qe.mutex, NULL);
1187 pthread_cond_init(&top->qe.cond, NULL);
1190 static int __cmd_top(struct perf_top *top)
1193 struct perf_evsel *pos;
1194 struct perf_evsel_config_term *err_term;
1195 struct perf_evlist *evlist = top->evlist;
1196 struct record_opts *opts = &top->record_opts;
1197 pthread_t thread, thread_process;
1200 top->session = perf_session__new(NULL, false, NULL);
1201 if (top->session == NULL)
1204 if (!top->annotation_opts.objdump_path) {
1205 ret = perf_env__lookup_objdump(&top->session->header.env,
1206 &top->annotation_opts.objdump_path);
1211 ret = callchain_param__setup_sample_type(&callchain_param);
1215 if (perf_session__register_idle_thread(top->session) < 0)
1218 if (top->nr_threads_synthesize > 1)
1219 perf_set_multithreaded();
1221 init_process_thread(top);
1223 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1224 top->evlist->threads, false,
1225 top->nr_threads_synthesize);
1227 if (top->nr_threads_synthesize > 1)
1228 perf_set_singlethreaded();
1230 if (perf_hpp_list.socket) {
1231 ret = perf_env__read_cpu_topology_map(&perf_env);
1233 goto out_err_cpu_topo;
1236 ret = perf_top__start_counters(top);
1240 ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1242 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1243 err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1244 str_error_r(errno, msg, sizeof(msg)));
1248 top->session->evlist = top->evlist;
1249 perf_session__set_id_hdr_size(top->session);
1252 * When perf is starting the traced process, all the events (apart from
1253 * group members) have enable_on_exec=1 set, so don't spoil it by
1254 * prematurely enabling them.
1256 * XXX 'top' still doesn't start workloads like record, trace, but should,
1257 * so leave the check here.
1259 if (!target__none(&opts->target))
1260 perf_evlist__enable(top->evlist);
1263 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1264 ui__error("Could not create process thread.\n");
1268 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1269 display_thread), top)) {
1270 ui__error("Could not create display thread.\n");
1271 goto out_join_thread;
1274 if (top->realtime_prio) {
1275 struct sched_param param;
1277 param.sched_priority = top->realtime_prio;
1278 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1279 ui__error("Could not set realtime priority.\n");
1284 /* Wait for a minimal set of events before starting the snapshot */
1285 perf_evlist__poll(top->evlist, 100);
1287 perf_top__mmap_read(top);
1290 u64 hits = top->samples;
1292 perf_top__mmap_read(top);
1294 if (opts->overwrite || (hits == top->samples))
1295 ret = perf_evlist__poll(top->evlist, 100);
1298 perf_top__resize(top);
1305 pthread_join(thread, NULL);
1307 pthread_cond_signal(&top->qe.cond);
1308 pthread_join(thread_process, NULL);
1310 perf_session__delete(top->session);
1311 top->session = NULL;
1316 char errbuf[BUFSIZ];
1317 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1319 ui__error("Could not read the CPU topology map: %s\n", err);
1325 callchain_opt(const struct option *opt, const char *arg, int unset)
1327 symbol_conf.use_callchain = true;
1328 return record_callchain_opt(opt, arg, unset);
1332 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1334 struct callchain_param *callchain = opt->value;
1336 callchain->enabled = !unset;
1337 callchain->record_mode = CALLCHAIN_FP;
1343 symbol_conf.use_callchain = false;
1344 callchain->record_mode = CALLCHAIN_NONE;
1348 return parse_callchain_top_opt(arg);
1351 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1353 if (!strcmp(var, "top.call-graph")) {
1354 var = "call-graph.record-mode";
1355 return perf_default_config(var, value, cb);
1357 if (!strcmp(var, "top.children")) {
1358 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1366 parse_percent_limit(const struct option *opt, const char *arg,
1367 int unset __maybe_unused)
1369 struct perf_top *top = opt->value;
1371 top->min_percent = strtof(arg, NULL);
1375 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1376 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1378 int cmd_top(int argc, const char **argv)
1380 char errbuf[BUFSIZ];
1381 struct perf_top top = {
1385 .mmap_pages = UINT_MAX,
1386 .user_freq = UINT_MAX,
1387 .user_interval = ULLONG_MAX,
1388 .freq = 4000, /* 4 KHz */
1393 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1394 * when we pause, fix that and reenable. Probably using a
1395 * separate evlist with a dummy event, i.e. a non-overwrite
1396 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1397 * stays in overwrite mode. -acme
1400 .sample_time = true,
1402 .max_stack = sysctl__max_stack(),
1403 .annotation_opts = annotation__default_options,
1404 .nr_threads_synthesize = UINT_MAX,
1406 struct record_opts *opts = &top.record_opts;
1407 struct target *target = &opts->target;
1408 const struct option options[] = {
1409 OPT_CALLBACK('e', "event", &top.evlist, "event",
1410 "event selector. use 'perf list' to list available events",
1411 parse_events_option),
1412 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1413 OPT_STRING('p', "pid", &target->pid, "pid",
1414 "profile events on existing process id"),
1415 OPT_STRING('t', "tid", &target->tid, "tid",
1416 "profile events on existing thread id"),
1417 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1418 "system-wide collection from all CPUs"),
1419 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1420 "list of cpus to monitor"),
1421 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1422 "file", "vmlinux pathname"),
1423 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1424 "don't load vmlinux even if found"),
1425 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1426 "file", "kallsyms pathname"),
1427 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1428 "hide kernel symbols"),
1429 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1430 "number of mmap data pages",
1431 perf_evlist__parse_mmap_pages),
1432 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1433 "collect data with this RT SCHED_FIFO priority"),
1434 OPT_INTEGER('d', "delay", &top.delay_secs,
1435 "number of seconds to delay between refreshes"),
1436 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1437 "dump the symbol table used for profiling"),
1438 OPT_INTEGER('f', "count-filter", &top.count_filter,
1439 "only display functions with more events than this"),
1440 OPT_BOOLEAN(0, "group", &opts->group,
1441 "put the counters into a counter group"),
1442 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1443 "child tasks do not inherit counters"),
1444 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1445 "symbol to annotate"),
1446 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1447 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1448 "profile at this frequency",
1449 record__parse_freq),
1450 OPT_INTEGER('E', "entries", &top.print_entries,
1451 "display this many functions"),
1452 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1453 "hide user symbols"),
1454 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1455 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1456 OPT_INCR('v', "verbose", &verbose,
1457 "be more verbose (show counter open errors, etc)"),
1458 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1459 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1460 " Please refer the man page for the complete list."),
1461 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1462 "output field(s): overhead, period, sample plus all of sort keys"),
1463 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1464 "Show a column with the number of samples"),
1465 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1466 NULL, "enables call-graph recording and display",
1468 OPT_CALLBACK(0, "call-graph", &callchain_param,
1469 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1470 top_callchain_help, &parse_callchain_opt),
1471 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1472 "Accumulate callchains of children and show total overhead as well"),
1473 OPT_INTEGER(0, "max-stack", &top.max_stack,
1474 "Set the maximum stack depth when parsing the callchain. "
1475 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1476 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1477 "ignore callees of these functions in call graphs",
1478 report_parse_ignore_callees_opt),
1479 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1480 "Show a column with the sum of periods"),
1481 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1482 "only consider symbols in these dsos"),
1483 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1484 "only consider symbols in these comms"),
1485 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1486 "only consider these symbols"),
1487 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1488 "Interleave source code with assembly code (default)"),
1489 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1490 "Display raw encoding of assembly instructions (default)"),
1491 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1492 "Enable kernel symbol demangling"),
1493 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1494 "objdump binary to use for disassembly and annotations"),
1495 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1496 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1497 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1498 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1499 "Don't show entries under that percent", parse_percent_limit),
1500 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1501 "How to display percentage of filtered entries", parse_filter_percentage),
1502 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1504 "don't try to adjust column width, use these fixed values"),
1505 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1506 "per thread proc mmap processing timeout in ms"),
1507 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1508 "branch any", "sample any taken branches",
1509 parse_branch_stack),
1510 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1511 "branch filter mask", "branch stack filter modes",
1512 parse_branch_stack),
1513 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1514 "Show raw trace event output (do not use print fmt or plugins)"),
1515 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1516 "Show entries in a hierarchy"),
1517 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1518 "Use a backward ring buffer, default: no"),
1519 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1520 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1521 "number of thread to run event synthesize"),
1524 const char * const top_usage[] = {
1525 "perf top [<options>]",
1528 int status = hists__init();
1533 top.annotation_opts.min_pcnt = 5;
1534 top.annotation_opts.context = 4;
1536 top.evlist = perf_evlist__new();
1537 if (top.evlist == NULL)
1540 status = perf_config(perf_top_config, &top);
1544 argc = parse_options(argc, argv, options, top_usage, 0);
1546 usage_with_options(top_usage, options);
1548 if (!top.evlist->nr_entries &&
1549 perf_evlist__add_default(top.evlist) < 0) {
1550 pr_err("Not enough memory for event selector list\n");
1551 goto out_delete_evlist;
1554 if (symbol_conf.report_hierarchy) {
1555 /* disable incompatible options */
1556 symbol_conf.event_group = false;
1557 symbol_conf.cumulate_callchain = false;
1560 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1561 parse_options_usage(top_usage, options, "fields", 0);
1562 parse_options_usage(NULL, options, "hierarchy", 0);
1563 goto out_delete_evlist;
1567 if (opts->branch_stack && callchain_param.enabled)
1568 symbol_conf.show_branchflag_count = true;
1570 sort__mode = SORT_MODE__TOP;
1571 /* display thread wants entries to be collapsed in a different tree */
1572 perf_hpp_list.need_collapse = 1;
1576 else if (top.use_tui)
1579 setup_browser(false);
1581 if (setup_sorting(top.evlist) < 0) {
1583 parse_options_usage(top_usage, options, "s", 1);
1585 parse_options_usage(sort_order ? NULL : top_usage,
1586 options, "fields", 0);
1587 goto out_delete_evlist;
1590 status = target__validate(target);
1592 target__strerror(target, status, errbuf, BUFSIZ);
1593 ui__warning("%s\n", errbuf);
1596 status = target__parse_uid(target);
1598 int saved_errno = errno;
1600 target__strerror(target, status, errbuf, BUFSIZ);
1601 ui__error("%s\n", errbuf);
1603 status = -saved_errno;
1604 goto out_delete_evlist;
1607 if (target__none(target))
1608 target->system_wide = true;
1610 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1611 ui__error("Couldn't create thread/CPU maps: %s\n",
1612 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1613 goto out_delete_evlist;
1616 if (top.delay_secs < 1)
1619 if (record_opts__config(opts)) {
1621 goto out_delete_evlist;
1624 top.sym_evsel = perf_evlist__first(top.evlist);
1626 if (!callchain_param.enabled) {
1627 symbol_conf.cumulate_callchain = false;
1628 perf_hpp__cancel_cumulate();
1631 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1632 callchain_param.order = ORDER_CALLER;
1634 status = symbol__annotation_init();
1636 goto out_delete_evlist;
1638 annotation_config__init();
1640 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1641 if (symbol__init(NULL) < 0)
1644 sort__setup_elide(stdout);
1646 get_term_dimensions(&top.winsize);
1647 if (top.print_entries == 0) {
1648 perf_top__update_print_entries(&top);
1649 signal(SIGWINCH, winch_sig);
1652 status = __cmd_top(&top);
1655 perf_evlist__delete(top.evlist);