4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10 * Improvements and fixes by:
12 * Arjan van de Ven <arjan@linux.intel.com>
13 * Yanmin Zhang <yanmin.zhang@intel.com>
14 * Wu Fengguang <fengguang.wu@intel.com>
15 * Mike Galbraith <efault@gmx.de>
16 * Paul Mackerras <paulus@samba.org>
18 * Released under the GPL v2. (and only v2, not any later version)
24 #include "util/annotate.h"
25 #include "util/config.h"
26 #include "util/color.h"
27 #include "util/drv_configs.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/event.h"
31 #include "util/machine.h"
32 #include "util/session.h"
33 #include "util/symbol.h"
34 #include "util/thread.h"
35 #include "util/thread_map.h"
37 #include <linux/rbtree.h>
38 #include <subcmd/parse-options.h>
39 #include "util/parse-events.h"
40 #include "util/cpumap.h"
41 #include "util/xyarray.h"
42 #include "util/sort.h"
43 #include "util/term.h"
44 #include "util/intlist.h"
45 #include "util/parse-branch-options.h"
46 #include "arch/common.h"
48 #include "util/debug.h"
64 #include <sys/syscall.h>
65 #include <sys/ioctl.h>
67 #include <sys/prctl.h>
70 #include <sys/utsname.h>
73 #include <linux/stringify.h>
74 #include <linux/time64.h>
75 #include <linux/types.h>
77 #include "sane_ctype.h"
79 static volatile int done;
80 static volatile int resize;
82 #define HEADER_LINE_NR 5
84 static void perf_top__update_print_entries(struct perf_top *top)
86 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
89 static void winch_sig(int sig __maybe_unused)
94 static void perf_top__resize(struct perf_top *top)
96 get_term_dimensions(&top->winsize);
97 perf_top__update_print_entries(top);
100 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
102 struct perf_evsel *evsel = hists_to_evsel(he->hists);
104 struct annotation *notes;
108 if (!he || !he->ms.sym)
115 * We can't annotate with just /proc/kallsyms
117 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
118 !dso__is_kcore(map->dso)) {
119 pr_err("Can't annotate %s: No vmlinux file was found in the "
120 "path\n", sym->name);
125 notes = symbol__annotation(sym);
126 pthread_mutex_lock(¬es->lock);
128 if (!symbol__hists(sym, top->evlist->nr_entries)) {
129 pthread_mutex_unlock(¬es->lock);
130 pr_err("Not enough memory for annotating '%s' symbol!\n",
136 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
138 top->sym_filter_entry = he;
141 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
142 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
145 pthread_mutex_unlock(¬es->lock);
149 static void __zero_source_counters(struct hist_entry *he)
151 struct symbol *sym = he->ms.sym;
152 symbol__annotate_zero_histograms(sym);
155 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
158 int err = uname(&uts);
160 ui__warning("Out of bounds address found:\n\n"
161 "Addr: %" PRIx64 "\n"
163 "Map: %" PRIx64 "-%" PRIx64 "\n"
164 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
168 "Not all samples will be on the annotation output.\n\n"
169 "Please report to linux-kernel@vger.kernel.org\n",
170 ip, map->dso->long_name, dso__symtab_origin(map->dso),
171 map->start, map->end, sym->start, sym->end,
172 sym->binding == STB_GLOBAL ? 'g' :
173 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
174 err ? "[unknown]" : uts.machine,
175 err ? "[unknown]" : uts.release, perf_version_string);
176 if (use_browser <= 0)
179 map->erange_warned = true;
182 static void perf_top__record_precise_ip(struct perf_top *top,
183 struct hist_entry *he,
184 struct perf_sample *sample,
185 struct perf_evsel *evsel, u64 ip)
187 struct annotation *notes;
188 struct symbol *sym = he->ms.sym;
191 if (sym == NULL || (use_browser == 0 &&
192 (top->sym_filter_entry == NULL ||
193 top->sym_filter_entry->ms.sym != sym)))
196 notes = symbol__annotation(sym);
198 if (pthread_mutex_trylock(¬es->lock))
201 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
203 pthread_mutex_unlock(¬es->lock);
207 * This function is now called with he->hists->lock held.
208 * Release it before going to sleep.
210 pthread_mutex_unlock(&he->hists->lock);
212 if (err == -ERANGE && !he->ms.map->erange_warned)
213 ui__warn_map_erange(he->ms.map, sym, ip);
214 else if (err == -ENOMEM) {
215 pr_err("Not enough memory for annotating '%s' symbol!\n",
220 pthread_mutex_lock(&he->hists->lock);
224 static void perf_top__show_details(struct perf_top *top)
226 struct hist_entry *he = top->sym_filter_entry;
227 struct perf_evsel *evsel = hists_to_evsel(he->hists);
228 struct annotation *notes;
229 struct symbol *symbol;
236 notes = symbol__annotation(symbol);
238 pthread_mutex_lock(¬es->lock);
240 symbol__calc_percent(symbol, evsel);
242 if (notes->src == NULL)
245 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
246 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
248 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
250 if (top->evlist->enabled) {
252 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
254 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
257 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
259 pthread_mutex_unlock(¬es->lock);
262 static void perf_top__print_sym_table(struct perf_top *top)
266 const int win_width = top->winsize.ws_col - 1;
267 struct perf_evsel *evsel = top->sym_evsel;
268 struct hists *hists = evsel__hists(evsel);
272 perf_top__header_snprintf(top, bf, sizeof(bf));
275 perf_top__reset_sample_counters(top);
277 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
279 if (!top->record_opts.overwrite &&
280 (hists->stats.nr_lost_warned !=
281 hists->stats.nr_events[PERF_RECORD_LOST])) {
282 hists->stats.nr_lost_warned =
283 hists->stats.nr_events[PERF_RECORD_LOST];
284 color_fprintf(stdout, PERF_COLOR_RED,
285 "WARNING: LOST %d chunks, Check IO/CPU overload",
286 hists->stats.nr_lost_warned);
290 if (top->sym_filter_entry) {
291 perf_top__show_details(top);
295 if (top->evlist->enabled) {
297 hists__delete_entries(hists);
299 hists__decay_entries(hists, top->hide_user_symbols,
300 top->hide_kernel_symbols);
304 hists__collapse_resort(hists, NULL);
305 perf_evsel__output_resort(evsel, NULL);
307 hists__output_recalc_col_len(hists, top->print_entries - printed);
309 hists__fprintf(hists, false, top->print_entries - printed, win_width,
310 top->min_percent, stdout, !symbol_conf.use_callchain);
313 static void prompt_integer(int *target, const char *msg)
315 char *buf = malloc(0), *p;
319 fprintf(stdout, "\n%s: ", msg);
320 if (getline(&buf, &dummy, stdin) < 0)
323 p = strchr(buf, '\n');
333 tmp = strtoul(buf, NULL, 10);
339 static void prompt_percent(int *target, const char *msg)
343 prompt_integer(&tmp, msg);
344 if (tmp >= 0 && tmp <= 100)
348 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
350 char *buf = malloc(0), *p;
351 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
352 struct hists *hists = evsel__hists(top->sym_evsel);
353 struct rb_node *next;
356 /* zero counters of active symbol */
358 __zero_source_counters(syme);
359 top->sym_filter_entry = NULL;
362 fprintf(stdout, "\n%s: ", msg);
363 if (getline(&buf, &dummy, stdin) < 0)
366 p = strchr(buf, '\n');
370 next = rb_first(&hists->entries);
372 n = rb_entry(next, struct hist_entry, rb_node);
373 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
377 next = rb_next(&n->rb_node);
381 fprintf(stderr, "Sorry, %s is not active.\n", buf);
384 perf_top__parse_source(top, found);
390 static void perf_top__print_mapped_keys(struct perf_top *top)
394 if (top->sym_filter_entry) {
395 struct symbol *sym = top->sym_filter_entry->ms.sym;
399 fprintf(stdout, "\nMapped keys:\n");
400 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
401 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
403 if (top->evlist->nr_entries > 1)
404 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
406 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
408 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
409 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
410 fprintf(stdout, "\t[S] stop annotation.\n");
413 "\t[K] hide kernel symbols. \t(%s)\n",
414 top->hide_kernel_symbols ? "yes" : "no");
416 "\t[U] hide user symbols. \t(%s)\n",
417 top->hide_user_symbols ? "yes" : "no");
418 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
419 fprintf(stdout, "\t[qQ] quit.\n");
422 static int perf_top__key_mapped(struct perf_top *top, int c)
438 return top->evlist->nr_entries > 1 ? 1 : 0;
446 static bool perf_top__handle_keypress(struct perf_top *top, int c)
450 if (!perf_top__key_mapped(top, c)) {
451 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
454 perf_top__print_mapped_keys(top);
455 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
458 set_term_quiet_input(&save);
460 poll(&stdin_poll, 1, -1);
463 tcsetattr(0, TCSAFLUSH, &save);
464 if (!perf_top__key_mapped(top, c))
470 prompt_integer(&top->delay_secs, "Enter display delay");
471 if (top->delay_secs < 1)
475 prompt_integer(&top->print_entries, "Enter display entries (lines)");
476 if (top->print_entries == 0) {
477 perf_top__resize(top);
478 signal(SIGWINCH, winch_sig);
480 signal(SIGWINCH, SIG_DFL);
484 if (top->evlist->nr_entries > 1) {
485 /* Select 0 as the default event: */
488 fprintf(stderr, "\nAvailable events:");
490 evlist__for_each_entry(top->evlist, top->sym_evsel)
491 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
493 prompt_integer(&counter, "Enter details event counter");
495 if (counter >= top->evlist->nr_entries) {
496 top->sym_evsel = perf_evlist__first(top->evlist);
497 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
501 evlist__for_each_entry(top->evlist, top->sym_evsel)
502 if (top->sym_evsel->idx == counter)
505 top->sym_evsel = perf_evlist__first(top->evlist);
508 prompt_integer(&top->count_filter, "Enter display event count filter");
511 prompt_percent(&top->annotation_opts.min_pcnt,
512 "Enter details display event filter (percent)");
515 top->hide_kernel_symbols = !top->hide_kernel_symbols;
519 printf("exiting.\n");
520 if (top->dump_symtab)
521 perf_session__fprintf_dsos(top->session, stderr);
525 perf_top__prompt_symbol(top, "Enter details symbol");
528 if (!top->sym_filter_entry)
531 struct hist_entry *syme = top->sym_filter_entry;
533 top->sym_filter_entry = NULL;
534 __zero_source_counters(syme);
538 top->hide_user_symbols = !top->hide_user_symbols;
541 top->zero = !top->zero;
550 static void perf_top__sort_new_samples(void *arg)
552 struct perf_top *t = arg;
553 struct perf_evsel *evsel = t->sym_evsel;
556 perf_top__reset_sample_counters(t);
558 if (t->evlist->selected != NULL)
559 t->sym_evsel = t->evlist->selected;
561 hists = evsel__hists(evsel);
563 if (t->evlist->enabled) {
565 hists__delete_entries(hists);
567 hists__decay_entries(hists, t->hide_user_symbols,
568 t->hide_kernel_symbols);
572 hists__collapse_resort(hists, NULL);
573 perf_evsel__output_resort(evsel, NULL);
576 static void *display_thread_tui(void *arg)
578 struct perf_evsel *pos;
579 struct perf_top *top = arg;
580 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
581 struct hist_browser_timer hbt = {
582 .timer = perf_top__sort_new_samples,
584 .refresh = top->delay_secs,
587 /* In order to read symbols from other namespaces perf to needs to call
588 * setns(2). This isn't permitted if the struct_fs has multiple users.
589 * unshare(2) the fs so that we may continue to setns into namespaces
590 * that we're observing.
594 perf_top__sort_new_samples(top);
597 * Initialize the uid_filter_str, in the future the TUI will allow
598 * Zooming in/out UIDs. For now juse use whatever the user passed
601 evlist__for_each_entry(top->evlist, pos) {
602 struct hists *hists = evsel__hists(pos);
603 hists->uid_filter_str = top->record_opts.target.uid_str;
606 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
608 &top->session->header.env,
609 !top->record_opts.overwrite,
610 &top->annotation_opts);
616 static void display_sig(int sig __maybe_unused)
621 static void display_setup_sig(void)
623 signal(SIGSEGV, sighandler_dump_stack);
624 signal(SIGFPE, sighandler_dump_stack);
625 signal(SIGINT, display_sig);
626 signal(SIGQUIT, display_sig);
627 signal(SIGTERM, display_sig);
630 static void *display_thread(void *arg)
632 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
634 struct perf_top *top = arg;
637 /* In order to read symbols from other namespaces perf to needs to call
638 * setns(2). This isn't permitted if the struct_fs has multiple users.
639 * unshare(2) the fs so that we may continue to setns into namespaces
640 * that we're observing.
645 pthread__unblock_sigwinch();
647 delay_msecs = top->delay_secs * MSEC_PER_SEC;
648 set_term_quiet_input(&save);
653 perf_top__print_sym_table(top);
655 * Either timeout expired or we got an EINTR due to SIGWINCH,
656 * refresh screen in both cases.
658 switch (poll(&stdin_poll, 1, delay_msecs)) {
667 tcsetattr(0, TCSAFLUSH, &save);
669 if (perf_top__handle_keypress(top, c))
675 tcsetattr(0, TCSAFLUSH, &save);
679 static int hist_iter__top_callback(struct hist_entry_iter *iter,
680 struct addr_location *al, bool single,
683 struct perf_top *top = arg;
684 struct hist_entry *he = iter->he;
685 struct perf_evsel *evsel = iter->evsel;
687 if (perf_hpp_list.sym && single)
688 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
690 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
691 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
695 static void perf_event__process_sample(struct perf_tool *tool,
696 const union perf_event *event,
697 struct perf_evsel *evsel,
698 struct perf_sample *sample,
699 struct machine *machine)
701 struct perf_top *top = container_of(tool, struct perf_top, tool);
702 struct addr_location al;
705 if (!machine && perf_guest) {
706 static struct intlist *seen;
709 seen = intlist__new(NULL);
711 if (!intlist__has_entry(seen, sample->pid)) {
712 pr_err("Can't find guest [%d]'s kernel information\n",
714 intlist__add(seen, sample->pid);
720 pr_err("%u unprocessable samples recorded.\r",
721 top->session->evlist->stats.nr_unprocessable_samples++);
725 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
726 top->exact_samples++;
728 if (machine__resolve(machine, &al, sample) < 0)
731 if (!machine->kptr_restrict_warned &&
732 symbol_conf.kptr_restrict &&
733 al.cpumode == PERF_RECORD_MISC_KERNEL) {
734 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
736 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
737 "Check /proc/sys/kernel/kptr_restrict.\n\n"
738 "Kernel%s samples will not be resolved.\n",
739 al.map && map__has_symbols(al.map) ?
741 if (use_browser <= 0)
744 machine->kptr_restrict_warned = true;
747 if (al.sym == NULL && al.map != NULL) {
748 const char *msg = "Kernel samples will not be resolved.\n";
750 * As we do lazy loading of symtabs we only will know if the
751 * specified vmlinux file is invalid when we actually have a
752 * hit in kernel space and then try to load it. So if we get
753 * here and there are _no_ symbols in the DSO backing the
754 * kernel map, bail out.
756 * We may never get here, for instance, if we use -K/
757 * --hide-kernel-symbols, even if the user specifies an
758 * invalid --vmlinux ;-)
760 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
761 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
762 if (symbol_conf.vmlinux_name) {
764 dso__strerror_load(al.map->dso, serr, sizeof(serr));
765 ui__warning("The %s file can't be used: %s\n%s",
766 symbol_conf.vmlinux_name, serr, msg);
768 ui__warning("A vmlinux file was not found.\n%s",
772 if (use_browser <= 0)
774 top->vmlinux_warned = true;
778 if (al.sym == NULL || !al.sym->idle) {
779 struct hists *hists = evsel__hists(evsel);
780 struct hist_entry_iter iter = {
783 .add_entry_cb = hist_iter__top_callback,
786 if (symbol_conf.cumulate_callchain)
787 iter.ops = &hist_iter_cumulative;
789 iter.ops = &hist_iter_normal;
791 pthread_mutex_lock(&hists->lock);
793 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
795 pr_err("Problem incrementing symbol period, skipping event\n");
797 pthread_mutex_unlock(&hists->lock);
800 addr_location__put(&al);
803 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
805 struct record_opts *opts = &top->record_opts;
806 struct perf_evlist *evlist = top->evlist;
807 struct perf_sample sample;
808 struct perf_evsel *evsel;
809 struct perf_mmap *md;
810 struct perf_session *session = top->session;
811 union perf_event *event;
812 struct machine *machine;
815 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
816 if (perf_mmap__read_init(md) < 0)
819 while ((event = perf_mmap__read_event(md)) != NULL) {
820 ret = perf_evlist__parse_sample(evlist, event, &sample);
822 pr_err("Can't parse sample, err = %d\n", ret);
826 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
827 assert(evsel != NULL);
829 if (event->header.type == PERF_RECORD_SAMPLE)
832 switch (sample.cpumode) {
833 case PERF_RECORD_MISC_USER:
835 if (top->hide_user_symbols)
837 machine = &session->machines.host;
839 case PERF_RECORD_MISC_KERNEL:
840 ++top->kernel_samples;
841 if (top->hide_kernel_symbols)
843 machine = &session->machines.host;
845 case PERF_RECORD_MISC_GUEST_KERNEL:
846 ++top->guest_kernel_samples;
847 machine = perf_session__find_machine(session,
850 case PERF_RECORD_MISC_GUEST_USER:
851 ++top->guest_us_samples;
853 * TODO: we don't process guest user from host side
854 * except simple counting.
858 if (event->header.type == PERF_RECORD_SAMPLE)
860 machine = &session->machines.host;
865 if (event->header.type == PERF_RECORD_SAMPLE) {
866 perf_event__process_sample(&top->tool, event, evsel,
868 } else if (event->header.type < PERF_RECORD_MAX) {
869 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
870 machine__process_event(machine, event, &sample);
872 ++session->evlist->stats.nr_unknown_events;
874 perf_mmap__consume(md);
877 perf_mmap__read_done(md);
880 static void perf_top__mmap_read(struct perf_top *top)
882 bool overwrite = top->record_opts.overwrite;
883 struct perf_evlist *evlist = top->evlist;
884 unsigned long long start, end;
889 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
891 for (i = 0; i < top->evlist->nr_mmaps; i++)
892 perf_top__mmap_read_idx(top, i);
895 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
896 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
900 if ((end - start) > (unsigned long long)top->delay_secs * NSEC_PER_SEC)
901 ui__warning("Too slow to read ring buffer.\n"
902 "Please try increasing the period (-c) or\n"
903 "decreasing the freq (-F) or\n"
904 "limiting the number of CPUs (-C)\n");
908 * Check per-event overwrite term.
909 * perf top should support consistent term for all events.
910 * - All events don't have per-event term
911 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
912 * Nothing change, return 0.
913 * - All events have same per-event term
914 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
915 * Using the per-event setting to replace the opts->overwrite if
916 * they are different, then return 0.
917 * - Events have different per-event term
918 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
920 * - Some of the event set per-event term, but some not.
921 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
924 static int perf_top__overwrite_check(struct perf_top *top)
926 struct record_opts *opts = &top->record_opts;
927 struct perf_evlist *evlist = top->evlist;
928 struct perf_evsel_config_term *term;
929 struct list_head *config_terms;
930 struct perf_evsel *evsel;
931 int set, overwrite = -1;
933 evlist__for_each_entry(evlist, evsel) {
935 config_terms = &evsel->config_terms;
936 list_for_each_entry(term, config_terms, list) {
937 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
938 set = term->val.overwrite ? 1 : 0;
941 /* no term for current and previous event (likely) */
942 if ((overwrite < 0) && (set < 0))
945 /* has term for both current and previous event, compare */
946 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
949 /* no term for current event but has term for previous one */
950 if ((overwrite >= 0) && (set < 0))
953 /* has term for current event */
954 if ((overwrite < 0) && (set >= 0)) {
955 /* if it's first event, set overwrite */
956 if (evsel == perf_evlist__first(evlist))
963 if ((overwrite >= 0) && (opts->overwrite != overwrite))
964 opts->overwrite = overwrite;
969 static int perf_top_overwrite_fallback(struct perf_top *top,
970 struct perf_evsel *evsel)
972 struct record_opts *opts = &top->record_opts;
973 struct perf_evlist *evlist = top->evlist;
974 struct perf_evsel *counter;
976 if (!opts->overwrite)
979 /* only fall back when first event fails */
980 if (evsel != perf_evlist__first(evlist))
983 evlist__for_each_entry(evlist, counter)
984 counter->attr.write_backward = false;
985 opts->overwrite = false;
986 pr_debug2("fall back to non-overwrite mode\n");
990 static int perf_top__start_counters(struct perf_top *top)
993 struct perf_evsel *counter;
994 struct perf_evlist *evlist = top->evlist;
995 struct record_opts *opts = &top->record_opts;
997 if (perf_top__overwrite_check(top)) {
998 ui__error("perf top only support consistent per-event "
999 "overwrite setting for all events\n");
1003 perf_evlist__config(evlist, opts, &callchain_param);
1005 evlist__for_each_entry(evlist, counter) {
1007 if (perf_evsel__open(counter, top->evlist->cpus,
1008 top->evlist->threads) < 0) {
1011 * Specially handle overwrite fall back.
1012 * Because perf top is the only tool which has
1013 * overwrite mode by default, support
1014 * both overwrite and non-overwrite mode, and
1015 * require consistent mode for all events.
1017 * May move it to generic code with more tools
1018 * have similar attribute.
1020 if (perf_missing_features.write_backward &&
1021 perf_top_overwrite_fallback(top, counter))
1024 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1026 ui__warning("%s\n", msg);
1030 perf_evsel__open_strerror(counter, &opts->target,
1031 errno, msg, sizeof(msg));
1032 ui__error("%s\n", msg);
1037 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1038 ui__error("Failed to mmap with %d (%s)\n",
1039 errno, str_error_r(errno, msg, sizeof(msg)));
1049 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1051 if (!perf_hpp_list.sym) {
1052 if (callchain->enabled) {
1053 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1056 } else if (callchain->mode != CHAIN_NONE) {
1057 if (callchain_register_param(callchain) < 0) {
1058 ui__error("Can't register callchain params.\n");
1066 static int __cmd_top(struct perf_top *top)
1069 struct perf_evsel *pos;
1070 struct perf_evsel_config_term *err_term;
1071 struct perf_evlist *evlist = top->evlist;
1072 struct record_opts *opts = &top->record_opts;
1076 top->session = perf_session__new(NULL, false, NULL);
1077 if (top->session == NULL)
1080 if (!top->annotation_opts.objdump_path) {
1081 ret = perf_env__lookup_objdump(&top->session->header.env,
1082 &top->annotation_opts.objdump_path);
1087 ret = callchain_param__setup_sample_type(&callchain_param);
1091 if (perf_session__register_idle_thread(top->session) < 0)
1094 if (top->nr_threads_synthesize > 1)
1095 perf_set_multithreaded();
1097 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1098 top->evlist->threads, false,
1099 opts->proc_map_timeout,
1100 top->nr_threads_synthesize);
1102 if (top->nr_threads_synthesize > 1)
1103 perf_set_singlethreaded();
1105 if (perf_hpp_list.socket) {
1106 ret = perf_env__read_cpu_topology_map(&perf_env);
1108 goto out_err_cpu_topo;
1111 ret = perf_top__start_counters(top);
1115 ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1117 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1118 err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1119 str_error_r(errno, msg, sizeof(msg)));
1123 top->session->evlist = top->evlist;
1124 perf_session__set_id_hdr_size(top->session);
1127 * When perf is starting the traced process, all the events (apart from
1128 * group members) have enable_on_exec=1 set, so don't spoil it by
1129 * prematurely enabling them.
1131 * XXX 'top' still doesn't start workloads like record, trace, but should,
1132 * so leave the check here.
1134 if (!target__none(&opts->target))
1135 perf_evlist__enable(top->evlist);
1137 /* Wait for a minimal set of events before starting the snapshot */
1138 perf_evlist__poll(top->evlist, 100);
1140 perf_top__mmap_read(top);
1143 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1144 display_thread), top)) {
1145 ui__error("Could not create display thread.\n");
1149 if (top->realtime_prio) {
1150 struct sched_param param;
1152 param.sched_priority = top->realtime_prio;
1153 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1154 ui__error("Could not set realtime priority.\n");
1160 u64 hits = top->samples;
1162 perf_top__mmap_read(top);
1164 if (opts->overwrite || (hits == top->samples))
1165 ret = perf_evlist__poll(top->evlist, 100);
1168 perf_top__resize(top);
1175 pthread_join(thread, NULL);
1177 perf_session__delete(top->session);
1178 top->session = NULL;
1183 char errbuf[BUFSIZ];
1184 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1186 ui__error("Could not read the CPU topology map: %s\n", err);
1192 callchain_opt(const struct option *opt, const char *arg, int unset)
1194 symbol_conf.use_callchain = true;
1195 return record_callchain_opt(opt, arg, unset);
1199 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1201 struct callchain_param *callchain = opt->value;
1203 callchain->enabled = !unset;
1204 callchain->record_mode = CALLCHAIN_FP;
1210 symbol_conf.use_callchain = false;
1211 callchain->record_mode = CALLCHAIN_NONE;
1215 return parse_callchain_top_opt(arg);
1218 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1220 if (!strcmp(var, "top.call-graph")) {
1221 var = "call-graph.record-mode";
1222 return perf_default_config(var, value, cb);
1224 if (!strcmp(var, "top.children")) {
1225 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1233 parse_percent_limit(const struct option *opt, const char *arg,
1234 int unset __maybe_unused)
1236 struct perf_top *top = opt->value;
1238 top->min_percent = strtof(arg, NULL);
1242 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1243 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1245 int cmd_top(int argc, const char **argv)
1247 char errbuf[BUFSIZ];
1248 struct perf_top top = {
1252 .mmap_pages = UINT_MAX,
1253 .user_freq = UINT_MAX,
1254 .user_interval = ULLONG_MAX,
1255 .freq = 4000, /* 4 KHz */
1259 .proc_map_timeout = 500,
1262 .max_stack = sysctl__max_stack(),
1263 .annotation_opts = annotation__default_options,
1264 .nr_threads_synthesize = UINT_MAX,
1266 struct record_opts *opts = &top.record_opts;
1267 struct target *target = &opts->target;
1268 const struct option options[] = {
1269 OPT_CALLBACK('e', "event", &top.evlist, "event",
1270 "event selector. use 'perf list' to list available events",
1271 parse_events_option),
1272 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1273 OPT_STRING('p', "pid", &target->pid, "pid",
1274 "profile events on existing process id"),
1275 OPT_STRING('t', "tid", &target->tid, "tid",
1276 "profile events on existing thread id"),
1277 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1278 "system-wide collection from all CPUs"),
1279 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1280 "list of cpus to monitor"),
1281 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1282 "file", "vmlinux pathname"),
1283 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1284 "don't load vmlinux even if found"),
1285 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1286 "hide kernel symbols"),
1287 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1288 "number of mmap data pages",
1289 perf_evlist__parse_mmap_pages),
1290 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1291 "collect data with this RT SCHED_FIFO priority"),
1292 OPT_INTEGER('d', "delay", &top.delay_secs,
1293 "number of seconds to delay between refreshes"),
1294 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1295 "dump the symbol table used for profiling"),
1296 OPT_INTEGER('f', "count-filter", &top.count_filter,
1297 "only display functions with more events than this"),
1298 OPT_BOOLEAN(0, "group", &opts->group,
1299 "put the counters into a counter group"),
1300 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1301 "child tasks do not inherit counters"),
1302 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1303 "symbol to annotate"),
1304 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1305 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1306 "profile at this frequency",
1307 record__parse_freq),
1308 OPT_INTEGER('E', "entries", &top.print_entries,
1309 "display this many functions"),
1310 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1311 "hide user symbols"),
1312 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1313 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1314 OPT_INCR('v', "verbose", &verbose,
1315 "be more verbose (show counter open errors, etc)"),
1316 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1317 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1318 " Please refer the man page for the complete list."),
1319 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1320 "output field(s): overhead, period, sample plus all of sort keys"),
1321 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1322 "Show a column with the number of samples"),
1323 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1324 NULL, "enables call-graph recording and display",
1326 OPT_CALLBACK(0, "call-graph", &callchain_param,
1327 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1328 top_callchain_help, &parse_callchain_opt),
1329 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1330 "Accumulate callchains of children and show total overhead as well"),
1331 OPT_INTEGER(0, "max-stack", &top.max_stack,
1332 "Set the maximum stack depth when parsing the callchain. "
1333 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1334 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1335 "ignore callees of these functions in call graphs",
1336 report_parse_ignore_callees_opt),
1337 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1338 "Show a column with the sum of periods"),
1339 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1340 "only consider symbols in these dsos"),
1341 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1342 "only consider symbols in these comms"),
1343 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1344 "only consider these symbols"),
1345 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1346 "Interleave source code with assembly code (default)"),
1347 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1348 "Display raw encoding of assembly instructions (default)"),
1349 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1350 "Enable kernel symbol demangling"),
1351 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1352 "objdump binary to use for disassembly and annotations"),
1353 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1354 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1355 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1356 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1357 "Don't show entries under that percent", parse_percent_limit),
1358 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1359 "How to display percentage of filtered entries", parse_filter_percentage),
1360 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1362 "don't try to adjust column width, use these fixed values"),
1363 OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1364 "per thread proc mmap processing timeout in ms"),
1365 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1366 "branch any", "sample any taken branches",
1367 parse_branch_stack),
1368 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1369 "branch filter mask", "branch stack filter modes",
1370 parse_branch_stack),
1371 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1372 "Show raw trace event output (do not use print fmt or plugins)"),
1373 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1374 "Show entries in a hierarchy"),
1375 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1376 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1377 "number of thread to run event synthesize"),
1380 const char * const top_usage[] = {
1381 "perf top [<options>]",
1384 int status = hists__init();
1389 top.annotation_opts.min_pcnt = 5;
1390 top.annotation_opts.context = 4;
1392 top.evlist = perf_evlist__new();
1393 if (top.evlist == NULL)
1396 status = perf_config(perf_top_config, &top);
1400 argc = parse_options(argc, argv, options, top_usage, 0);
1402 usage_with_options(top_usage, options);
1404 if (!top.evlist->nr_entries &&
1405 perf_evlist__add_default(top.evlist) < 0) {
1406 pr_err("Not enough memory for event selector list\n");
1407 goto out_delete_evlist;
1410 if (symbol_conf.report_hierarchy) {
1411 /* disable incompatible options */
1412 symbol_conf.event_group = false;
1413 symbol_conf.cumulate_callchain = false;
1416 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1417 parse_options_usage(top_usage, options, "fields", 0);
1418 parse_options_usage(NULL, options, "hierarchy", 0);
1419 goto out_delete_evlist;
1423 sort__mode = SORT_MODE__TOP;
1424 /* display thread wants entries to be collapsed in a different tree */
1425 perf_hpp_list.need_collapse = 1;
1429 else if (top.use_tui)
1432 setup_browser(false);
1434 if (setup_sorting(top.evlist) < 0) {
1436 parse_options_usage(top_usage, options, "s", 1);
1438 parse_options_usage(sort_order ? NULL : top_usage,
1439 options, "fields", 0);
1440 goto out_delete_evlist;
1443 status = target__validate(target);
1445 target__strerror(target, status, errbuf, BUFSIZ);
1446 ui__warning("%s\n", errbuf);
1449 status = target__parse_uid(target);
1451 int saved_errno = errno;
1453 target__strerror(target, status, errbuf, BUFSIZ);
1454 ui__error("%s\n", errbuf);
1456 status = -saved_errno;
1457 goto out_delete_evlist;
1460 if (target__none(target))
1461 target->system_wide = true;
1463 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1464 ui__error("Couldn't create thread/CPU maps: %s\n",
1465 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1466 goto out_delete_evlist;
1469 if (top.delay_secs < 1)
1472 if (record_opts__config(opts)) {
1474 goto out_delete_evlist;
1477 top.sym_evsel = perf_evlist__first(top.evlist);
1479 if (!callchain_param.enabled) {
1480 symbol_conf.cumulate_callchain = false;
1481 perf_hpp__cancel_cumulate();
1484 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1485 callchain_param.order = ORDER_CALLER;
1487 status = symbol__annotation_init();
1489 goto out_delete_evlist;
1491 annotation_config__init();
1493 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1494 if (symbol__init(NULL) < 0)
1497 sort__setup_elide(stdout);
1499 get_term_dimensions(&top.winsize);
1500 if (top.print_entries == 0) {
1501 perf_top__update_print_entries(&top);
1502 signal(SIGWINCH, winch_sig);
1505 status = __cmd_top(&top);
1508 perf_evlist__delete(top.evlist);