1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin top command: Display a continuously updated profile of
6 * any workload, CPU or specific PID.
8 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
11 * Improvements and fixes by:
13 * Arjan van de Ven <arjan@linux.intel.com>
14 * Yanmin Zhang <yanmin.zhang@intel.com>
15 * Wu Fengguang <fengguang.wu@intel.com>
16 * Mike Galbraith <efault@gmx.de>
17 * Paul Mackerras <paulus@samba.org>
23 #include "util/annotate.h"
24 #include "util/bpf-event.h"
25 #include "util/config.h"
26 #include "util/color.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/event.h"
31 #include "util/machine.h"
33 #include "util/mmap.h"
34 #include "util/session.h"
35 #include "util/symbol.h"
36 #include "util/synthetic-events.h"
38 #include "util/util.h"
39 #include <linux/rbtree.h>
40 #include <subcmd/parse-options.h>
41 #include "util/parse-events.h"
42 #include "util/callchain.h"
43 #include "util/cpumap.h"
44 #include "util/sort.h"
45 #include "util/string2.h"
46 #include "util/term.h"
47 #include "util/intlist.h"
48 #include "util/parse-branch-options.h"
49 #include "arch/common.h"
52 #include "util/debug.h"
53 #include "util/ordered-events.h"
69 #include <sys/syscall.h>
70 #include <sys/ioctl.h>
72 #include <sys/prctl.h>
75 #include <sys/utsname.h>
78 #include <linux/stringify.h>
79 #include <linux/time64.h>
80 #include <linux/types.h>
81 #include <linux/err.h>
83 #include <linux/ctype.h>
85 static volatile int done;
86 static volatile int resize;
88 #define HEADER_LINE_NR 5
90 static void perf_top__update_print_entries(struct perf_top *top)
92 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
95 static void winch_sig(int sig __maybe_unused)
100 static void perf_top__resize(struct perf_top *top)
102 get_term_dimensions(&top->winsize);
103 perf_top__update_print_entries(top);
106 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
110 struct annotation *notes;
114 if (!he || !he->ms.sym)
117 evsel = hists_to_evsel(he->hists);
123 * We can't annotate with just /proc/kallsyms
125 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
126 !dso__is_kcore(map->dso)) {
127 pr_err("Can't annotate %s: No vmlinux file was found in the "
128 "path\n", sym->name);
133 notes = symbol__annotation(sym);
134 pthread_mutex_lock(¬es->lock);
136 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
137 pthread_mutex_unlock(¬es->lock);
138 pr_err("Not enough memory for annotating '%s' symbol!\n",
144 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
146 top->sym_filter_entry = he;
149 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
150 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
153 pthread_mutex_unlock(¬es->lock);
157 static void __zero_source_counters(struct hist_entry *he)
159 struct symbol *sym = he->ms.sym;
160 symbol__annotate_zero_histograms(sym);
163 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
166 int err = uname(&uts);
168 ui__warning("Out of bounds address found:\n\n"
169 "Addr: %" PRIx64 "\n"
171 "Map: %" PRIx64 "-%" PRIx64 "\n"
172 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
176 "Not all samples will be on the annotation output.\n\n"
177 "Please report to linux-kernel@vger.kernel.org\n",
178 ip, map->dso->long_name, dso__symtab_origin(map->dso),
179 map->start, map->end, sym->start, sym->end,
180 sym->binding == STB_GLOBAL ? 'g' :
181 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
182 err ? "[unknown]" : uts.machine,
183 err ? "[unknown]" : uts.release, perf_version_string);
184 if (use_browser <= 0)
187 map->erange_warned = true;
190 static void perf_top__record_precise_ip(struct perf_top *top,
191 struct hist_entry *he,
192 struct perf_sample *sample,
193 struct evsel *evsel, u64 ip)
195 struct annotation *notes;
196 struct symbol *sym = he->ms.sym;
199 if (sym == NULL || (use_browser == 0 &&
200 (top->sym_filter_entry == NULL ||
201 top->sym_filter_entry->ms.sym != sym)))
204 notes = symbol__annotation(sym);
206 if (pthread_mutex_trylock(¬es->lock))
209 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
211 pthread_mutex_unlock(¬es->lock);
215 * This function is now called with he->hists->lock held.
216 * Release it before going to sleep.
218 pthread_mutex_unlock(&he->hists->lock);
220 if (err == -ERANGE && !he->ms.map->erange_warned)
221 ui__warn_map_erange(he->ms.map, sym, ip);
222 else if (err == -ENOMEM) {
223 pr_err("Not enough memory for annotating '%s' symbol!\n",
228 pthread_mutex_lock(&he->hists->lock);
232 static void perf_top__show_details(struct perf_top *top)
234 struct hist_entry *he = top->sym_filter_entry;
236 struct annotation *notes;
237 struct symbol *symbol;
243 evsel = hists_to_evsel(he->hists);
246 notes = symbol__annotation(symbol);
248 pthread_mutex_lock(¬es->lock);
250 symbol__calc_percent(symbol, evsel);
252 if (notes->src == NULL)
255 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
256 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
258 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
260 if (top->evlist->enabled) {
262 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
264 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
267 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
269 pthread_mutex_unlock(¬es->lock);
272 static void perf_top__resort_hists(struct perf_top *t)
274 struct evlist *evlist = t->evlist;
277 evlist__for_each_entry(evlist, pos) {
278 struct hists *hists = evsel__hists(pos);
281 * unlink existing entries so that they can be linked
282 * in a correct order in hists__match() below.
284 hists__unlink(hists);
286 if (evlist->enabled) {
288 hists__delete_entries(hists);
290 hists__decay_entries(hists, t->hide_user_symbols,
291 t->hide_kernel_symbols);
295 hists__collapse_resort(hists, NULL);
297 /* Non-group events are considered as leader */
298 if (symbol_conf.event_group &&
299 !perf_evsel__is_group_leader(pos)) {
300 struct hists *leader_hists = evsel__hists(pos->leader);
302 hists__match(leader_hists, hists);
303 hists__link(leader_hists, hists);
307 evlist__for_each_entry(evlist, pos) {
308 perf_evsel__output_resort(pos, NULL);
312 static void perf_top__print_sym_table(struct perf_top *top)
316 const int win_width = top->winsize.ws_col - 1;
317 struct evsel *evsel = top->sym_evsel;
318 struct hists *hists = evsel__hists(evsel);
322 perf_top__header_snprintf(top, bf, sizeof(bf));
325 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
327 if (!top->record_opts.overwrite &&
328 (hists->stats.nr_lost_warned !=
329 hists->stats.nr_events[PERF_RECORD_LOST])) {
330 hists->stats.nr_lost_warned =
331 hists->stats.nr_events[PERF_RECORD_LOST];
332 color_fprintf(stdout, PERF_COLOR_RED,
333 "WARNING: LOST %d chunks, Check IO/CPU overload",
334 hists->stats.nr_lost_warned);
338 if (top->sym_filter_entry) {
339 perf_top__show_details(top);
343 perf_top__resort_hists(top);
345 hists__output_recalc_col_len(hists, top->print_entries - printed);
347 hists__fprintf(hists, false, top->print_entries - printed, win_width,
348 top->min_percent, stdout, !symbol_conf.use_callchain);
351 static void prompt_integer(int *target, const char *msg)
353 char *buf = malloc(0), *p;
357 fprintf(stdout, "\n%s: ", msg);
358 if (getline(&buf, &dummy, stdin) < 0)
361 p = strchr(buf, '\n');
371 tmp = strtoul(buf, NULL, 10);
377 static void prompt_percent(int *target, const char *msg)
381 prompt_integer(&tmp, msg);
382 if (tmp >= 0 && tmp <= 100)
386 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
388 char *buf = malloc(0), *p;
389 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
390 struct hists *hists = evsel__hists(top->sym_evsel);
391 struct rb_node *next;
394 /* zero counters of active symbol */
396 __zero_source_counters(syme);
397 top->sym_filter_entry = NULL;
400 fprintf(stdout, "\n%s: ", msg);
401 if (getline(&buf, &dummy, stdin) < 0)
404 p = strchr(buf, '\n');
408 next = rb_first_cached(&hists->entries);
410 n = rb_entry(next, struct hist_entry, rb_node);
411 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
415 next = rb_next(&n->rb_node);
419 fprintf(stderr, "Sorry, %s is not active.\n", buf);
422 perf_top__parse_source(top, found);
428 static void perf_top__print_mapped_keys(struct perf_top *top)
432 if (top->sym_filter_entry) {
433 struct symbol *sym = top->sym_filter_entry->ms.sym;
437 fprintf(stdout, "\nMapped keys:\n");
438 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
439 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
441 if (top->evlist->core.nr_entries > 1)
442 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
444 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
446 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
447 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
448 fprintf(stdout, "\t[S] stop annotation.\n");
451 "\t[K] hide kernel symbols. \t(%s)\n",
452 top->hide_kernel_symbols ? "yes" : "no");
454 "\t[U] hide user symbols. \t(%s)\n",
455 top->hide_user_symbols ? "yes" : "no");
456 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
457 fprintf(stdout, "\t[qQ] quit.\n");
460 static int perf_top__key_mapped(struct perf_top *top, int c)
476 return top->evlist->core.nr_entries > 1 ? 1 : 0;
484 static bool perf_top__handle_keypress(struct perf_top *top, int c)
488 if (!perf_top__key_mapped(top, c)) {
489 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
492 perf_top__print_mapped_keys(top);
493 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
496 set_term_quiet_input(&save);
498 poll(&stdin_poll, 1, -1);
501 tcsetattr(0, TCSAFLUSH, &save);
502 if (!perf_top__key_mapped(top, c))
508 prompt_integer(&top->delay_secs, "Enter display delay");
509 if (top->delay_secs < 1)
513 prompt_integer(&top->print_entries, "Enter display entries (lines)");
514 if (top->print_entries == 0) {
515 perf_top__resize(top);
516 signal(SIGWINCH, winch_sig);
518 signal(SIGWINCH, SIG_DFL);
522 if (top->evlist->core.nr_entries > 1) {
523 /* Select 0 as the default event: */
526 fprintf(stderr, "\nAvailable events:");
528 evlist__for_each_entry(top->evlist, top->sym_evsel)
529 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
531 prompt_integer(&counter, "Enter details event counter");
533 if (counter >= top->evlist->core.nr_entries) {
534 top->sym_evsel = perf_evlist__first(top->evlist);
535 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
539 evlist__for_each_entry(top->evlist, top->sym_evsel)
540 if (top->sym_evsel->idx == counter)
543 top->sym_evsel = perf_evlist__first(top->evlist);
546 prompt_integer(&top->count_filter, "Enter display event count filter");
549 prompt_percent(&top->annotation_opts.min_pcnt,
550 "Enter details display event filter (percent)");
553 top->hide_kernel_symbols = !top->hide_kernel_symbols;
557 printf("exiting.\n");
558 if (top->dump_symtab)
559 perf_session__fprintf_dsos(top->session, stderr);
563 perf_top__prompt_symbol(top, "Enter details symbol");
566 if (!top->sym_filter_entry)
569 struct hist_entry *syme = top->sym_filter_entry;
571 top->sym_filter_entry = NULL;
572 __zero_source_counters(syme);
576 top->hide_user_symbols = !top->hide_user_symbols;
579 top->zero = !top->zero;
588 static void perf_top__sort_new_samples(void *arg)
590 struct perf_top *t = arg;
592 if (t->evlist->selected != NULL)
593 t->sym_evsel = t->evlist->selected;
595 perf_top__resort_hists(t);
597 if (t->lost || t->drop)
598 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
601 static void stop_top(void)
607 static void *display_thread_tui(void *arg)
610 struct perf_top *top = arg;
611 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
612 struct hist_browser_timer hbt = {
613 .timer = perf_top__sort_new_samples,
615 .refresh = top->delay_secs,
618 /* In order to read symbols from other namespaces perf to needs to call
619 * setns(2). This isn't permitted if the struct_fs has multiple users.
620 * unshare(2) the fs so that we may continue to setns into namespaces
621 * that we're observing.
625 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
627 perf_top__sort_new_samples(top);
630 * Initialize the uid_filter_str, in the future the TUI will allow
631 * Zooming in/out UIDs. For now just use whatever the user passed
634 evlist__for_each_entry(top->evlist, pos) {
635 struct hists *hists = evsel__hists(pos);
636 hists->uid_filter_str = top->record_opts.target.uid_str;
639 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
641 &top->session->header.env,
642 !top->record_opts.overwrite,
643 &top->annotation_opts);
649 static void display_sig(int sig __maybe_unused)
654 static void display_setup_sig(void)
656 signal(SIGSEGV, sighandler_dump_stack);
657 signal(SIGFPE, sighandler_dump_stack);
658 signal(SIGINT, display_sig);
659 signal(SIGQUIT, display_sig);
660 signal(SIGTERM, display_sig);
663 static void *display_thread(void *arg)
665 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
667 struct perf_top *top = arg;
670 /* In order to read symbols from other namespaces perf to needs to call
671 * setns(2). This isn't permitted if the struct_fs has multiple users.
672 * unshare(2) the fs so that we may continue to setns into namespaces
673 * that we're observing.
677 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
680 pthread__unblock_sigwinch();
682 delay_msecs = top->delay_secs * MSEC_PER_SEC;
683 set_term_quiet_input(&save);
688 perf_top__print_sym_table(top);
690 * Either timeout expired or we got an EINTR due to SIGWINCH,
691 * refresh screen in both cases.
693 switch (poll(&stdin_poll, 1, delay_msecs)) {
702 tcsetattr(0, TCSAFLUSH, &save);
704 if (perf_top__handle_keypress(top, c))
710 tcsetattr(0, TCSAFLUSH, &save);
714 static int hist_iter__top_callback(struct hist_entry_iter *iter,
715 struct addr_location *al, bool single,
718 struct perf_top *top = arg;
719 struct hist_entry *he = iter->he;
720 struct evsel *evsel = iter->evsel;
722 if (perf_hpp_list.sym && single)
723 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
725 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
726 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
730 static void perf_event__process_sample(struct perf_tool *tool,
731 const union perf_event *event,
733 struct perf_sample *sample,
734 struct machine *machine)
736 struct perf_top *top = container_of(tool, struct perf_top, tool);
737 struct addr_location al;
740 if (!machine && perf_guest) {
741 static struct intlist *seen;
744 seen = intlist__new(NULL);
746 if (!intlist__has_entry(seen, sample->pid)) {
747 pr_err("Can't find guest [%d]'s kernel information\n",
749 intlist__add(seen, sample->pid);
755 pr_err("%u unprocessable samples recorded.\r",
756 top->session->evlist->stats.nr_unprocessable_samples++);
760 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
761 top->exact_samples++;
763 if (machine__resolve(machine, &al, sample) < 0)
766 if (!machine->kptr_restrict_warned &&
767 symbol_conf.kptr_restrict &&
768 al.cpumode == PERF_RECORD_MISC_KERNEL) {
769 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
771 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
772 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
773 "Kernel%s samples will not be resolved.\n",
774 al.map && map__has_symbols(al.map) ?
776 if (use_browser <= 0)
779 machine->kptr_restrict_warned = true;
782 if (al.sym == NULL && al.map != NULL) {
783 const char *msg = "Kernel samples will not be resolved.\n";
785 * As we do lazy loading of symtabs we only will know if the
786 * specified vmlinux file is invalid when we actually have a
787 * hit in kernel space and then try to load it. So if we get
788 * here and there are _no_ symbols in the DSO backing the
789 * kernel map, bail out.
791 * We may never get here, for instance, if we use -K/
792 * --hide-kernel-symbols, even if the user specifies an
793 * invalid --vmlinux ;-)
795 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
796 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
797 if (symbol_conf.vmlinux_name) {
799 dso__strerror_load(al.map->dso, serr, sizeof(serr));
800 ui__warning("The %s file can't be used: %s\n%s",
801 symbol_conf.vmlinux_name, serr, msg);
803 ui__warning("A vmlinux file was not found.\n%s",
807 if (use_browser <= 0)
809 top->vmlinux_warned = true;
813 if (al.sym == NULL || !al.sym->idle) {
814 struct hists *hists = evsel__hists(evsel);
815 struct hist_entry_iter iter = {
818 .add_entry_cb = hist_iter__top_callback,
821 if (symbol_conf.cumulate_callchain)
822 iter.ops = &hist_iter_cumulative;
824 iter.ops = &hist_iter_normal;
826 pthread_mutex_lock(&hists->lock);
828 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
830 pr_err("Problem incrementing symbol period, skipping event\n");
832 pthread_mutex_unlock(&hists->lock);
835 addr_location__put(&al);
839 perf_top__process_lost(struct perf_top *top, union perf_event *event,
842 struct hists *hists = evsel__hists(evsel);
844 top->lost += event->lost.lost;
845 top->lost_total += event->lost.lost;
846 hists->stats.total_lost += event->lost.lost;
850 perf_top__process_lost_samples(struct perf_top *top,
851 union perf_event *event,
854 struct hists *hists = evsel__hists(evsel);
856 top->lost += event->lost_samples.lost;
857 top->lost_total += event->lost_samples.lost;
858 hists->stats.total_lost_samples += event->lost_samples.lost;
861 static u64 last_timestamp;
863 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
865 struct record_opts *opts = &top->record_opts;
866 struct evlist *evlist = top->evlist;
868 union perf_event *event;
870 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
871 if (perf_mmap__read_init(md) < 0)
874 while ((event = perf_mmap__read_event(md)) != NULL) {
877 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
878 if (ret && ret != -1)
881 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
885 perf_mmap__consume(md);
887 if (top->qe.rotate) {
888 pthread_mutex_lock(&top->qe.mutex);
889 top->qe.rotate = false;
890 pthread_cond_signal(&top->qe.cond);
891 pthread_mutex_unlock(&top->qe.mutex);
895 perf_mmap__read_done(md);
898 static void perf_top__mmap_read(struct perf_top *top)
900 bool overwrite = top->record_opts.overwrite;
901 struct evlist *evlist = top->evlist;
905 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
907 for (i = 0; i < top->evlist->nr_mmaps; i++)
908 perf_top__mmap_read_idx(top, i);
911 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
912 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
917 * Check per-event overwrite term.
918 * perf top should support consistent term for all events.
919 * - All events don't have per-event term
920 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
921 * Nothing change, return 0.
922 * - All events have same per-event term
923 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
924 * Using the per-event setting to replace the opts->overwrite if
925 * they are different, then return 0.
926 * - Events have different per-event term
927 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
929 * - Some of the event set per-event term, but some not.
930 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
933 static int perf_top__overwrite_check(struct perf_top *top)
935 struct record_opts *opts = &top->record_opts;
936 struct evlist *evlist = top->evlist;
937 struct perf_evsel_config_term *term;
938 struct list_head *config_terms;
940 int set, overwrite = -1;
942 evlist__for_each_entry(evlist, evsel) {
944 config_terms = &evsel->config_terms;
945 list_for_each_entry(term, config_terms, list) {
946 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
947 set = term->val.overwrite ? 1 : 0;
950 /* no term for current and previous event (likely) */
951 if ((overwrite < 0) && (set < 0))
954 /* has term for both current and previous event, compare */
955 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
958 /* no term for current event but has term for previous one */
959 if ((overwrite >= 0) && (set < 0))
962 /* has term for current event */
963 if ((overwrite < 0) && (set >= 0)) {
964 /* if it's first event, set overwrite */
965 if (evsel == perf_evlist__first(evlist))
972 if ((overwrite >= 0) && (opts->overwrite != overwrite))
973 opts->overwrite = overwrite;
978 static int perf_top_overwrite_fallback(struct perf_top *top,
981 struct record_opts *opts = &top->record_opts;
982 struct evlist *evlist = top->evlist;
983 struct evsel *counter;
985 if (!opts->overwrite)
988 /* only fall back when first event fails */
989 if (evsel != perf_evlist__first(evlist))
992 evlist__for_each_entry(evlist, counter)
993 counter->core.attr.write_backward = false;
994 opts->overwrite = false;
995 pr_debug2("fall back to non-overwrite mode\n");
999 static int perf_top__start_counters(struct perf_top *top)
1002 struct evsel *counter;
1003 struct evlist *evlist = top->evlist;
1004 struct record_opts *opts = &top->record_opts;
1006 if (perf_top__overwrite_check(top)) {
1007 ui__error("perf top only support consistent per-event "
1008 "overwrite setting for all events\n");
1012 perf_evlist__config(evlist, opts, &callchain_param);
1014 evlist__for_each_entry(evlist, counter) {
1016 if (evsel__open(counter, top->evlist->core.cpus,
1017 top->evlist->core.threads) < 0) {
1020 * Specially handle overwrite fall back.
1021 * Because perf top is the only tool which has
1022 * overwrite mode by default, support
1023 * both overwrite and non-overwrite mode, and
1024 * require consistent mode for all events.
1026 * May move it to generic code with more tools
1027 * have similar attribute.
1029 if (perf_missing_features.write_backward &&
1030 perf_top_overwrite_fallback(top, counter))
1033 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1035 ui__warning("%s\n", msg);
1039 perf_evsel__open_strerror(counter, &opts->target,
1040 errno, msg, sizeof(msg));
1041 ui__error("%s\n", msg);
1046 if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1047 ui__error("Failed to mmap with %d (%s)\n",
1048 errno, str_error_r(errno, msg, sizeof(msg)));
1058 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1060 if (callchain->mode != CHAIN_NONE) {
1061 if (callchain_register_param(callchain) < 0) {
1062 ui__error("Can't register callchain params.\n");
1070 static struct ordered_events *rotate_queues(struct perf_top *top)
1072 struct ordered_events *in = top->qe.in;
1074 if (top->qe.in == &top->qe.data[1])
1075 top->qe.in = &top->qe.data[0];
1077 top->qe.in = &top->qe.data[1];
1082 static void *process_thread(void *arg)
1084 struct perf_top *top = arg;
1087 struct ordered_events *out, *in = top->qe.in;
1089 if (!in->nr_events) {
1094 out = rotate_queues(top);
1096 pthread_mutex_lock(&top->qe.mutex);
1097 top->qe.rotate = true;
1098 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1099 pthread_mutex_unlock(&top->qe.mutex);
1101 if (ordered_events__flush(out, OE_FLUSH__TOP))
1102 pr_err("failed to process events\n");
1109 * Allow only 'top->delay_secs' seconds behind samples.
1111 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1113 union perf_event *event = qevent->event;
1114 u64 delay_timestamp;
1116 if (event->header.type != PERF_RECORD_SAMPLE)
1119 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1120 return delay_timestamp < last_timestamp;
1123 static int deliver_event(struct ordered_events *qe,
1124 struct ordered_event *qevent)
1126 struct perf_top *top = qe->data;
1127 struct evlist *evlist = top->evlist;
1128 struct perf_session *session = top->session;
1129 union perf_event *event = qevent->event;
1130 struct perf_sample sample;
1131 struct evsel *evsel;
1132 struct machine *machine;
1135 if (should_drop(qevent, top)) {
1141 ret = perf_evlist__parse_sample(evlist, event, &sample);
1143 pr_err("Can't parse sample, err = %d\n", ret);
1147 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1148 assert(evsel != NULL);
1150 if (event->header.type == PERF_RECORD_SAMPLE) {
1151 if (evswitch__discard(&top->evswitch, evsel))
1156 switch (sample.cpumode) {
1157 case PERF_RECORD_MISC_USER:
1159 if (top->hide_user_symbols)
1161 machine = &session->machines.host;
1163 case PERF_RECORD_MISC_KERNEL:
1164 ++top->kernel_samples;
1165 if (top->hide_kernel_symbols)
1167 machine = &session->machines.host;
1169 case PERF_RECORD_MISC_GUEST_KERNEL:
1170 ++top->guest_kernel_samples;
1171 machine = perf_session__find_machine(session,
1174 case PERF_RECORD_MISC_GUEST_USER:
1175 ++top->guest_us_samples;
1177 * TODO: we don't process guest user from host side
1178 * except simple counting.
1182 if (event->header.type == PERF_RECORD_SAMPLE)
1184 machine = &session->machines.host;
1188 if (event->header.type == PERF_RECORD_SAMPLE) {
1189 perf_event__process_sample(&top->tool, event, evsel,
1191 } else if (event->header.type == PERF_RECORD_LOST) {
1192 perf_top__process_lost(top, event, evsel);
1193 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1194 perf_top__process_lost_samples(top, event, evsel);
1195 } else if (event->header.type < PERF_RECORD_MAX) {
1196 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1197 machine__process_event(machine, event, &sample);
1199 ++session->evlist->stats.nr_unknown_events;
1206 static void init_process_thread(struct perf_top *top)
1208 ordered_events__init(&top->qe.data[0], deliver_event, top);
1209 ordered_events__init(&top->qe.data[1], deliver_event, top);
1210 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1211 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1212 top->qe.in = &top->qe.data[0];
1213 pthread_mutex_init(&top->qe.mutex, NULL);
1214 pthread_cond_init(&top->qe.cond, NULL);
1217 static int __cmd_top(struct perf_top *top)
1219 struct record_opts *opts = &top->record_opts;
1220 pthread_t thread, thread_process;
1223 if (!top->annotation_opts.objdump_path) {
1224 ret = perf_env__lookup_objdump(&top->session->header.env,
1225 &top->annotation_opts.objdump_path);
1230 ret = callchain_param__setup_sample_type(&callchain_param);
1234 if (perf_session__register_idle_thread(top->session) < 0)
1237 if (top->nr_threads_synthesize > 1)
1238 perf_set_multithreaded();
1240 init_process_thread(top);
1242 if (opts->record_namespaces)
1243 top->tool.namespace_events = true;
1245 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1246 &top->session->machines.host,
1249 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1251 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1252 top->evlist->core.threads, false,
1253 top->nr_threads_synthesize);
1255 if (top->nr_threads_synthesize > 1)
1256 perf_set_singlethreaded();
1258 if (perf_hpp_list.socket) {
1259 ret = perf_env__read_cpu_topology_map(&perf_env);
1261 char errbuf[BUFSIZ];
1262 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1264 ui__error("Could not read the CPU topology map: %s\n", err);
1269 ret = perf_top__start_counters(top);
1273 top->session->evlist = top->evlist;
1274 perf_session__set_id_hdr_size(top->session);
1277 * When perf is starting the traced process, all the events (apart from
1278 * group members) have enable_on_exec=1 set, so don't spoil it by
1279 * prematurely enabling them.
1281 * XXX 'top' still doesn't start workloads like record, trace, but should,
1282 * so leave the check here.
1284 if (!target__none(&opts->target))
1285 evlist__enable(top->evlist);
1288 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1289 ui__error("Could not create process thread.\n");
1293 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1294 display_thread), top)) {
1295 ui__error("Could not create display thread.\n");
1296 goto out_join_thread;
1299 if (top->realtime_prio) {
1300 struct sched_param param;
1302 param.sched_priority = top->realtime_prio;
1303 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1304 ui__error("Could not set realtime priority.\n");
1309 /* Wait for a minimal set of events before starting the snapshot */
1310 perf_evlist__poll(top->evlist, 100);
1312 perf_top__mmap_read(top);
1315 u64 hits = top->samples;
1317 perf_top__mmap_read(top);
1319 if (opts->overwrite || (hits == top->samples))
1320 ret = perf_evlist__poll(top->evlist, 100);
1323 perf_top__resize(top);
1330 pthread_join(thread, NULL);
1332 pthread_cond_signal(&top->qe.cond);
1333 pthread_join(thread_process, NULL);
1338 callchain_opt(const struct option *opt, const char *arg, int unset)
1340 symbol_conf.use_callchain = true;
1341 return record_callchain_opt(opt, arg, unset);
1345 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1347 struct callchain_param *callchain = opt->value;
1349 callchain->enabled = !unset;
1350 callchain->record_mode = CALLCHAIN_FP;
1356 symbol_conf.use_callchain = false;
1357 callchain->record_mode = CALLCHAIN_NONE;
1361 return parse_callchain_top_opt(arg);
1364 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1366 if (!strcmp(var, "top.call-graph")) {
1367 var = "call-graph.record-mode";
1368 return perf_default_config(var, value, cb);
1370 if (!strcmp(var, "top.children")) {
1371 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1379 parse_percent_limit(const struct option *opt, const char *arg,
1380 int unset __maybe_unused)
1382 struct perf_top *top = opt->value;
1384 top->min_percent = strtof(arg, NULL);
1388 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1389 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1391 int cmd_top(int argc, const char **argv)
1393 char errbuf[BUFSIZ];
1394 struct perf_top top = {
1398 .mmap_pages = UINT_MAX,
1399 .user_freq = UINT_MAX,
1400 .user_interval = ULLONG_MAX,
1401 .freq = 4000, /* 4 KHz */
1406 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1407 * when we pause, fix that and reenable. Probably using a
1408 * separate evlist with a dummy event, i.e. a non-overwrite
1409 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1410 * stays in overwrite mode. -acme
1413 .sample_time = true,
1414 .sample_time_set = true,
1416 .max_stack = sysctl__max_stack(),
1417 .annotation_opts = annotation__default_options,
1418 .nr_threads_synthesize = UINT_MAX,
1420 struct record_opts *opts = &top.record_opts;
1421 struct target *target = &opts->target;
1422 const struct option options[] = {
1423 OPT_CALLBACK('e', "event", &top.evlist, "event",
1424 "event selector. use 'perf list' to list available events",
1425 parse_events_option),
1426 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1427 OPT_STRING('p', "pid", &target->pid, "pid",
1428 "profile events on existing process id"),
1429 OPT_STRING('t', "tid", &target->tid, "tid",
1430 "profile events on existing thread id"),
1431 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1432 "system-wide collection from all CPUs"),
1433 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1434 "list of cpus to monitor"),
1435 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1436 "file", "vmlinux pathname"),
1437 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1438 "don't load vmlinux even if found"),
1439 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1440 "file", "kallsyms pathname"),
1441 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1442 "hide kernel symbols"),
1443 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1444 "number of mmap data pages",
1445 perf_evlist__parse_mmap_pages),
1446 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1447 "collect data with this RT SCHED_FIFO priority"),
1448 OPT_INTEGER('d', "delay", &top.delay_secs,
1449 "number of seconds to delay between refreshes"),
1450 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1451 "dump the symbol table used for profiling"),
1452 OPT_INTEGER('f', "count-filter", &top.count_filter,
1453 "only display functions with more events than this"),
1454 OPT_BOOLEAN(0, "group", &opts->group,
1455 "put the counters into a counter group"),
1456 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1457 "child tasks do not inherit counters"),
1458 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1459 "symbol to annotate"),
1460 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1461 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1462 "profile at this frequency",
1463 record__parse_freq),
1464 OPT_INTEGER('E', "entries", &top.print_entries,
1465 "display this many functions"),
1466 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1467 "hide user symbols"),
1468 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1469 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1470 OPT_INCR('v', "verbose", &verbose,
1471 "be more verbose (show counter open errors, etc)"),
1472 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1473 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1474 " Please refer the man page for the complete list."),
1475 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1476 "output field(s): overhead, period, sample plus all of sort keys"),
1477 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1478 "Show a column with the number of samples"),
1479 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1480 NULL, "enables call-graph recording and display",
1482 OPT_CALLBACK(0, "call-graph", &callchain_param,
1483 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1484 top_callchain_help, &parse_callchain_opt),
1485 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1486 "Accumulate callchains of children and show total overhead as well"),
1487 OPT_INTEGER(0, "max-stack", &top.max_stack,
1488 "Set the maximum stack depth when parsing the callchain. "
1489 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1490 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1491 "ignore callees of these functions in call graphs",
1492 report_parse_ignore_callees_opt),
1493 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1494 "Show a column with the sum of periods"),
1495 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1496 "only consider symbols in these dsos"),
1497 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1498 "only consider symbols in these comms"),
1499 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1500 "only consider these symbols"),
1501 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1502 "Interleave source code with assembly code (default)"),
1503 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1504 "Display raw encoding of assembly instructions (default)"),
1505 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1506 "Enable kernel symbol demangling"),
1507 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1508 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1509 "objdump binary to use for disassembly and annotations"),
1510 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1511 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1512 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1513 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1514 "Don't show entries under that percent", parse_percent_limit),
1515 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1516 "How to display percentage of filtered entries", parse_filter_percentage),
1517 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1519 "don't try to adjust column width, use these fixed values"),
1520 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1521 "per thread proc mmap processing timeout in ms"),
1522 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1523 "branch any", "sample any taken branches",
1524 parse_branch_stack),
1525 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1526 "branch filter mask", "branch stack filter modes",
1527 parse_branch_stack),
1528 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1529 "Show raw trace event output (do not use print fmt or plugins)"),
1530 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1531 "Show entries in a hierarchy"),
1532 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1533 "Use a backward ring buffer, default: no"),
1534 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1535 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1536 "number of thread to run event synthesize"),
1537 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1538 "Record namespaces events"),
1539 OPTS_EVSWITCH(&top.evswitch),
1542 struct evlist *sb_evlist = NULL;
1543 const char * const top_usage[] = {
1544 "perf top [<options>]",
1547 int status = hists__init();
1552 top.annotation_opts.min_pcnt = 5;
1553 top.annotation_opts.context = 4;
1555 top.evlist = evlist__new();
1556 if (top.evlist == NULL)
1559 status = perf_config(perf_top_config, &top);
1563 argc = parse_options(argc, argv, options, top_usage, 0);
1565 usage_with_options(top_usage, options);
1567 if (!top.evlist->core.nr_entries &&
1568 perf_evlist__add_default(top.evlist) < 0) {
1569 pr_err("Not enough memory for event selector list\n");
1570 goto out_delete_evlist;
1573 status = evswitch__init(&top.evswitch, top.evlist, stderr);
1575 goto out_delete_evlist;
1577 if (symbol_conf.report_hierarchy) {
1578 /* disable incompatible options */
1579 symbol_conf.event_group = false;
1580 symbol_conf.cumulate_callchain = false;
1583 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1584 parse_options_usage(top_usage, options, "fields", 0);
1585 parse_options_usage(NULL, options, "hierarchy", 0);
1586 goto out_delete_evlist;
1590 if (opts->branch_stack && callchain_param.enabled)
1591 symbol_conf.show_branchflag_count = true;
1593 sort__mode = SORT_MODE__TOP;
1594 /* display thread wants entries to be collapsed in a different tree */
1595 perf_hpp_list.need_collapse = 1;
1599 else if (top.use_tui)
1602 setup_browser(false);
1604 if (setup_sorting(top.evlist) < 0) {
1606 parse_options_usage(top_usage, options, "s", 1);
1608 parse_options_usage(sort_order ? NULL : top_usage,
1609 options, "fields", 0);
1610 goto out_delete_evlist;
1613 status = target__validate(target);
1615 target__strerror(target, status, errbuf, BUFSIZ);
1616 ui__warning("%s\n", errbuf);
1619 status = target__parse_uid(target);
1621 int saved_errno = errno;
1623 target__strerror(target, status, errbuf, BUFSIZ);
1624 ui__error("%s\n", errbuf);
1626 status = -saved_errno;
1627 goto out_delete_evlist;
1630 if (target__none(target))
1631 target->system_wide = true;
1633 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1634 ui__error("Couldn't create thread/CPU maps: %s\n",
1635 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1636 goto out_delete_evlist;
1639 if (top.delay_secs < 1)
1642 if (record_opts__config(opts)) {
1644 goto out_delete_evlist;
1647 top.sym_evsel = perf_evlist__first(top.evlist);
1649 if (!callchain_param.enabled) {
1650 symbol_conf.cumulate_callchain = false;
1651 perf_hpp__cancel_cumulate();
1654 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1655 callchain_param.order = ORDER_CALLER;
1657 status = symbol__annotation_init();
1659 goto out_delete_evlist;
1661 annotation_config__init();
1663 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1664 status = symbol__init(NULL);
1666 goto out_delete_evlist;
1668 sort__setup_elide(stdout);
1670 get_term_dimensions(&top.winsize);
1671 if (top.print_entries == 0) {
1672 perf_top__update_print_entries(&top);
1673 signal(SIGWINCH, winch_sig);
1676 top.session = perf_session__new(NULL, false, NULL);
1677 if (IS_ERR(top.session)) {
1678 status = PTR_ERR(top.session);
1679 goto out_delete_evlist;
1682 if (!top.record_opts.no_bpf_event)
1683 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1685 if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1686 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1687 opts->no_bpf_event = true;
1690 status = __cmd_top(&top);
1692 if (!opts->no_bpf_event)
1693 perf_evlist__stop_sb_thread(sb_evlist);
1696 evlist__delete(top.evlist);
1697 perf_session__delete(top.session);