1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin top command: Display a continuously updated profile of
6 * any workload, CPU or specific PID.
8 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
11 * Improvements and fixes by:
13 * Arjan van de Ven <arjan@linux.intel.com>
14 * Yanmin Zhang <yanmin.zhang@intel.com>
15 * Wu Fengguang <fengguang.wu@intel.com>
16 * Mike Galbraith <efault@gmx.de>
17 * Paul Mackerras <paulus@samba.org>
23 #include "util/annotate.h"
24 #include "util/bpf-event.h"
25 #include "util/config.h"
26 #include "util/color.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/event.h"
31 #include "util/machine.h"
33 #include "util/session.h"
34 #include "util/symbol.h"
35 #include "util/synthetic-events.h"
37 #include "util/util.h"
38 #include <linux/rbtree.h>
39 #include <subcmd/parse-options.h>
40 #include "util/parse-events.h"
41 #include "util/callchain.h"
42 #include "util/cpumap.h"
43 #include "util/sort.h"
44 #include "util/string2.h"
45 #include "util/term.h"
46 #include "util/intlist.h"
47 #include "util/parse-branch-options.h"
48 #include "arch/common.h"
51 #include "util/debug.h"
52 #include "util/ordered-events.h"
68 #include <sys/syscall.h>
69 #include <sys/ioctl.h>
71 #include <sys/prctl.h>
74 #include <sys/utsname.h>
77 #include <linux/stringify.h>
78 #include <linux/time64.h>
79 #include <linux/types.h>
80 #include <linux/err.h>
82 #include <linux/ctype.h>
84 static volatile int done;
85 static volatile int resize;
87 #define HEADER_LINE_NR 5
89 static void perf_top__update_print_entries(struct perf_top *top)
91 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
94 static void winch_sig(int sig __maybe_unused)
99 static void perf_top__resize(struct perf_top *top)
101 get_term_dimensions(&top->winsize);
102 perf_top__update_print_entries(top);
105 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
109 struct annotation *notes;
113 if (!he || !he->ms.sym)
116 evsel = hists_to_evsel(he->hists);
122 * We can't annotate with just /proc/kallsyms
124 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
125 !dso__is_kcore(map->dso)) {
126 pr_err("Can't annotate %s: No vmlinux file was found in the "
127 "path\n", sym->name);
132 notes = symbol__annotation(sym);
133 pthread_mutex_lock(¬es->lock);
135 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
136 pthread_mutex_unlock(¬es->lock);
137 pr_err("Not enough memory for annotating '%s' symbol!\n",
143 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
145 top->sym_filter_entry = he;
148 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
149 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
152 pthread_mutex_unlock(¬es->lock);
156 static void __zero_source_counters(struct hist_entry *he)
158 struct symbol *sym = he->ms.sym;
159 symbol__annotate_zero_histograms(sym);
162 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
165 int err = uname(&uts);
167 ui__warning("Out of bounds address found:\n\n"
168 "Addr: %" PRIx64 "\n"
170 "Map: %" PRIx64 "-%" PRIx64 "\n"
171 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
175 "Not all samples will be on the annotation output.\n\n"
176 "Please report to linux-kernel@vger.kernel.org\n",
177 ip, map->dso->long_name, dso__symtab_origin(map->dso),
178 map->start, map->end, sym->start, sym->end,
179 sym->binding == STB_GLOBAL ? 'g' :
180 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
181 err ? "[unknown]" : uts.machine,
182 err ? "[unknown]" : uts.release, perf_version_string);
183 if (use_browser <= 0)
186 map->erange_warned = true;
189 static void perf_top__record_precise_ip(struct perf_top *top,
190 struct hist_entry *he,
191 struct perf_sample *sample,
192 struct evsel *evsel, u64 ip)
194 struct annotation *notes;
195 struct symbol *sym = he->ms.sym;
198 if (sym == NULL || (use_browser == 0 &&
199 (top->sym_filter_entry == NULL ||
200 top->sym_filter_entry->ms.sym != sym)))
203 notes = symbol__annotation(sym);
205 if (pthread_mutex_trylock(¬es->lock))
208 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
210 pthread_mutex_unlock(¬es->lock);
214 * This function is now called with he->hists->lock held.
215 * Release it before going to sleep.
217 pthread_mutex_unlock(&he->hists->lock);
219 if (err == -ERANGE && !he->ms.map->erange_warned)
220 ui__warn_map_erange(he->ms.map, sym, ip);
221 else if (err == -ENOMEM) {
222 pr_err("Not enough memory for annotating '%s' symbol!\n",
227 pthread_mutex_lock(&he->hists->lock);
231 static void perf_top__show_details(struct perf_top *top)
233 struct hist_entry *he = top->sym_filter_entry;
235 struct annotation *notes;
236 struct symbol *symbol;
242 evsel = hists_to_evsel(he->hists);
245 notes = symbol__annotation(symbol);
247 pthread_mutex_lock(¬es->lock);
249 symbol__calc_percent(symbol, evsel);
251 if (notes->src == NULL)
254 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
255 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
257 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
259 if (top->evlist->enabled) {
261 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
263 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
266 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
268 pthread_mutex_unlock(¬es->lock);
271 static void perf_top__resort_hists(struct perf_top *t)
273 struct evlist *evlist = t->evlist;
276 evlist__for_each_entry(evlist, pos) {
277 struct hists *hists = evsel__hists(pos);
280 * unlink existing entries so that they can be linked
281 * in a correct order in hists__match() below.
283 hists__unlink(hists);
285 if (evlist->enabled) {
287 hists__delete_entries(hists);
289 hists__decay_entries(hists, t->hide_user_symbols,
290 t->hide_kernel_symbols);
294 hists__collapse_resort(hists, NULL);
296 /* Non-group events are considered as leader */
297 if (symbol_conf.event_group &&
298 !perf_evsel__is_group_leader(pos)) {
299 struct hists *leader_hists = evsel__hists(pos->leader);
301 hists__match(leader_hists, hists);
302 hists__link(leader_hists, hists);
306 evlist__for_each_entry(evlist, pos) {
307 perf_evsel__output_resort(pos, NULL);
311 static void perf_top__print_sym_table(struct perf_top *top)
315 const int win_width = top->winsize.ws_col - 1;
316 struct evsel *evsel = top->sym_evsel;
317 struct hists *hists = evsel__hists(evsel);
321 perf_top__header_snprintf(top, bf, sizeof(bf));
324 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
326 if (!top->record_opts.overwrite &&
327 (hists->stats.nr_lost_warned !=
328 hists->stats.nr_events[PERF_RECORD_LOST])) {
329 hists->stats.nr_lost_warned =
330 hists->stats.nr_events[PERF_RECORD_LOST];
331 color_fprintf(stdout, PERF_COLOR_RED,
332 "WARNING: LOST %d chunks, Check IO/CPU overload",
333 hists->stats.nr_lost_warned);
337 if (top->sym_filter_entry) {
338 perf_top__show_details(top);
342 perf_top__resort_hists(top);
344 hists__output_recalc_col_len(hists, top->print_entries - printed);
346 hists__fprintf(hists, false, top->print_entries - printed, win_width,
347 top->min_percent, stdout, !symbol_conf.use_callchain);
350 static void prompt_integer(int *target, const char *msg)
352 char *buf = malloc(0), *p;
356 fprintf(stdout, "\n%s: ", msg);
357 if (getline(&buf, &dummy, stdin) < 0)
360 p = strchr(buf, '\n');
370 tmp = strtoul(buf, NULL, 10);
376 static void prompt_percent(int *target, const char *msg)
380 prompt_integer(&tmp, msg);
381 if (tmp >= 0 && tmp <= 100)
385 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
387 char *buf = malloc(0), *p;
388 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
389 struct hists *hists = evsel__hists(top->sym_evsel);
390 struct rb_node *next;
393 /* zero counters of active symbol */
395 __zero_source_counters(syme);
396 top->sym_filter_entry = NULL;
399 fprintf(stdout, "\n%s: ", msg);
400 if (getline(&buf, &dummy, stdin) < 0)
403 p = strchr(buf, '\n');
407 next = rb_first_cached(&hists->entries);
409 n = rb_entry(next, struct hist_entry, rb_node);
410 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
414 next = rb_next(&n->rb_node);
418 fprintf(stderr, "Sorry, %s is not active.\n", buf);
421 perf_top__parse_source(top, found);
427 static void perf_top__print_mapped_keys(struct perf_top *top)
431 if (top->sym_filter_entry) {
432 struct symbol *sym = top->sym_filter_entry->ms.sym;
436 fprintf(stdout, "\nMapped keys:\n");
437 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
438 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
440 if (top->evlist->core.nr_entries > 1)
441 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
443 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
445 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
446 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
447 fprintf(stdout, "\t[S] stop annotation.\n");
450 "\t[K] hide kernel symbols. \t(%s)\n",
451 top->hide_kernel_symbols ? "yes" : "no");
453 "\t[U] hide user symbols. \t(%s)\n",
454 top->hide_user_symbols ? "yes" : "no");
455 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
456 fprintf(stdout, "\t[qQ] quit.\n");
459 static int perf_top__key_mapped(struct perf_top *top, int c)
475 return top->evlist->core.nr_entries > 1 ? 1 : 0;
483 static bool perf_top__handle_keypress(struct perf_top *top, int c)
487 if (!perf_top__key_mapped(top, c)) {
488 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
491 perf_top__print_mapped_keys(top);
492 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
495 set_term_quiet_input(&save);
497 poll(&stdin_poll, 1, -1);
500 tcsetattr(0, TCSAFLUSH, &save);
501 if (!perf_top__key_mapped(top, c))
507 prompt_integer(&top->delay_secs, "Enter display delay");
508 if (top->delay_secs < 1)
512 prompt_integer(&top->print_entries, "Enter display entries (lines)");
513 if (top->print_entries == 0) {
514 perf_top__resize(top);
515 signal(SIGWINCH, winch_sig);
517 signal(SIGWINCH, SIG_DFL);
521 if (top->evlist->core.nr_entries > 1) {
522 /* Select 0 as the default event: */
525 fprintf(stderr, "\nAvailable events:");
527 evlist__for_each_entry(top->evlist, top->sym_evsel)
528 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
530 prompt_integer(&counter, "Enter details event counter");
532 if (counter >= top->evlist->core.nr_entries) {
533 top->sym_evsel = perf_evlist__first(top->evlist);
534 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
538 evlist__for_each_entry(top->evlist, top->sym_evsel)
539 if (top->sym_evsel->idx == counter)
542 top->sym_evsel = perf_evlist__first(top->evlist);
545 prompt_integer(&top->count_filter, "Enter display event count filter");
548 prompt_percent(&top->annotation_opts.min_pcnt,
549 "Enter details display event filter (percent)");
552 top->hide_kernel_symbols = !top->hide_kernel_symbols;
556 printf("exiting.\n");
557 if (top->dump_symtab)
558 perf_session__fprintf_dsos(top->session, stderr);
562 perf_top__prompt_symbol(top, "Enter details symbol");
565 if (!top->sym_filter_entry)
568 struct hist_entry *syme = top->sym_filter_entry;
570 top->sym_filter_entry = NULL;
571 __zero_source_counters(syme);
575 top->hide_user_symbols = !top->hide_user_symbols;
578 top->zero = !top->zero;
587 static void perf_top__sort_new_samples(void *arg)
589 struct perf_top *t = arg;
591 if (t->evlist->selected != NULL)
592 t->sym_evsel = t->evlist->selected;
594 perf_top__resort_hists(t);
596 if (t->lost || t->drop)
597 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
600 static void stop_top(void)
606 static void *display_thread_tui(void *arg)
609 struct perf_top *top = arg;
610 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
611 struct hist_browser_timer hbt = {
612 .timer = perf_top__sort_new_samples,
614 .refresh = top->delay_secs,
617 /* In order to read symbols from other namespaces perf to needs to call
618 * setns(2). This isn't permitted if the struct_fs has multiple users.
619 * unshare(2) the fs so that we may continue to setns into namespaces
620 * that we're observing.
624 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
626 perf_top__sort_new_samples(top);
629 * Initialize the uid_filter_str, in the future the TUI will allow
630 * Zooming in/out UIDs. For now just use whatever the user passed
633 evlist__for_each_entry(top->evlist, pos) {
634 struct hists *hists = evsel__hists(pos);
635 hists->uid_filter_str = top->record_opts.target.uid_str;
638 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
640 &top->session->header.env,
641 !top->record_opts.overwrite,
642 &top->annotation_opts);
648 static void display_sig(int sig __maybe_unused)
653 static void display_setup_sig(void)
655 signal(SIGSEGV, sighandler_dump_stack);
656 signal(SIGFPE, sighandler_dump_stack);
657 signal(SIGINT, display_sig);
658 signal(SIGQUIT, display_sig);
659 signal(SIGTERM, display_sig);
662 static void *display_thread(void *arg)
664 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
666 struct perf_top *top = arg;
669 /* In order to read symbols from other namespaces perf to needs to call
670 * setns(2). This isn't permitted if the struct_fs has multiple users.
671 * unshare(2) the fs so that we may continue to setns into namespaces
672 * that we're observing.
676 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
679 pthread__unblock_sigwinch();
681 delay_msecs = top->delay_secs * MSEC_PER_SEC;
682 set_term_quiet_input(&save);
687 perf_top__print_sym_table(top);
689 * Either timeout expired or we got an EINTR due to SIGWINCH,
690 * refresh screen in both cases.
692 switch (poll(&stdin_poll, 1, delay_msecs)) {
701 tcsetattr(0, TCSAFLUSH, &save);
703 if (perf_top__handle_keypress(top, c))
709 tcsetattr(0, TCSAFLUSH, &save);
713 static int hist_iter__top_callback(struct hist_entry_iter *iter,
714 struct addr_location *al, bool single,
717 struct perf_top *top = arg;
718 struct hist_entry *he = iter->he;
719 struct evsel *evsel = iter->evsel;
721 if (perf_hpp_list.sym && single)
722 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
724 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
725 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
729 static void perf_event__process_sample(struct perf_tool *tool,
730 const union perf_event *event,
732 struct perf_sample *sample,
733 struct machine *machine)
735 struct perf_top *top = container_of(tool, struct perf_top, tool);
736 struct addr_location al;
739 if (!machine && perf_guest) {
740 static struct intlist *seen;
743 seen = intlist__new(NULL);
745 if (!intlist__has_entry(seen, sample->pid)) {
746 pr_err("Can't find guest [%d]'s kernel information\n",
748 intlist__add(seen, sample->pid);
754 pr_err("%u unprocessable samples recorded.\r",
755 top->session->evlist->stats.nr_unprocessable_samples++);
759 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
760 top->exact_samples++;
762 if (machine__resolve(machine, &al, sample) < 0)
765 if (!machine->kptr_restrict_warned &&
766 symbol_conf.kptr_restrict &&
767 al.cpumode == PERF_RECORD_MISC_KERNEL) {
768 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
770 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
771 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
772 "Kernel%s samples will not be resolved.\n",
773 al.map && map__has_symbols(al.map) ?
775 if (use_browser <= 0)
778 machine->kptr_restrict_warned = true;
781 if (al.sym == NULL && al.map != NULL) {
782 const char *msg = "Kernel samples will not be resolved.\n";
784 * As we do lazy loading of symtabs we only will know if the
785 * specified vmlinux file is invalid when we actually have a
786 * hit in kernel space and then try to load it. So if we get
787 * here and there are _no_ symbols in the DSO backing the
788 * kernel map, bail out.
790 * We may never get here, for instance, if we use -K/
791 * --hide-kernel-symbols, even if the user specifies an
792 * invalid --vmlinux ;-)
794 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
795 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
796 if (symbol_conf.vmlinux_name) {
798 dso__strerror_load(al.map->dso, serr, sizeof(serr));
799 ui__warning("The %s file can't be used: %s\n%s",
800 symbol_conf.vmlinux_name, serr, msg);
802 ui__warning("A vmlinux file was not found.\n%s",
806 if (use_browser <= 0)
808 top->vmlinux_warned = true;
812 if (al.sym == NULL || !al.sym->idle) {
813 struct hists *hists = evsel__hists(evsel);
814 struct hist_entry_iter iter = {
817 .add_entry_cb = hist_iter__top_callback,
820 if (symbol_conf.cumulate_callchain)
821 iter.ops = &hist_iter_cumulative;
823 iter.ops = &hist_iter_normal;
825 pthread_mutex_lock(&hists->lock);
827 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
829 pr_err("Problem incrementing symbol period, skipping event\n");
831 pthread_mutex_unlock(&hists->lock);
834 addr_location__put(&al);
838 perf_top__process_lost(struct perf_top *top, union perf_event *event,
841 struct hists *hists = evsel__hists(evsel);
843 top->lost += event->lost.lost;
844 top->lost_total += event->lost.lost;
845 hists->stats.total_lost += event->lost.lost;
849 perf_top__process_lost_samples(struct perf_top *top,
850 union perf_event *event,
853 struct hists *hists = evsel__hists(evsel);
855 top->lost += event->lost_samples.lost;
856 top->lost_total += event->lost_samples.lost;
857 hists->stats.total_lost_samples += event->lost_samples.lost;
860 static u64 last_timestamp;
862 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
864 struct record_opts *opts = &top->record_opts;
865 struct evlist *evlist = top->evlist;
867 union perf_event *event;
869 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
870 if (perf_mmap__read_init(md) < 0)
873 while ((event = perf_mmap__read_event(md)) != NULL) {
876 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
877 if (ret && ret != -1)
880 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
884 perf_mmap__consume(md);
886 if (top->qe.rotate) {
887 pthread_mutex_lock(&top->qe.mutex);
888 top->qe.rotate = false;
889 pthread_cond_signal(&top->qe.cond);
890 pthread_mutex_unlock(&top->qe.mutex);
894 perf_mmap__read_done(md);
897 static void perf_top__mmap_read(struct perf_top *top)
899 bool overwrite = top->record_opts.overwrite;
900 struct evlist *evlist = top->evlist;
904 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
906 for (i = 0; i < top->evlist->nr_mmaps; i++)
907 perf_top__mmap_read_idx(top, i);
910 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
911 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
916 * Check per-event overwrite term.
917 * perf top should support consistent term for all events.
918 * - All events don't have per-event term
919 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
920 * Nothing change, return 0.
921 * - All events have same per-event term
922 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
923 * Using the per-event setting to replace the opts->overwrite if
924 * they are different, then return 0.
925 * - Events have different per-event term
926 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
928 * - Some of the event set per-event term, but some not.
929 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
932 static int perf_top__overwrite_check(struct perf_top *top)
934 struct record_opts *opts = &top->record_opts;
935 struct evlist *evlist = top->evlist;
936 struct perf_evsel_config_term *term;
937 struct list_head *config_terms;
939 int set, overwrite = -1;
941 evlist__for_each_entry(evlist, evsel) {
943 config_terms = &evsel->config_terms;
944 list_for_each_entry(term, config_terms, list) {
945 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
946 set = term->val.overwrite ? 1 : 0;
949 /* no term for current and previous event (likely) */
950 if ((overwrite < 0) && (set < 0))
953 /* has term for both current and previous event, compare */
954 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
957 /* no term for current event but has term for previous one */
958 if ((overwrite >= 0) && (set < 0))
961 /* has term for current event */
962 if ((overwrite < 0) && (set >= 0)) {
963 /* if it's first event, set overwrite */
964 if (evsel == perf_evlist__first(evlist))
971 if ((overwrite >= 0) && (opts->overwrite != overwrite))
972 opts->overwrite = overwrite;
977 static int perf_top_overwrite_fallback(struct perf_top *top,
980 struct record_opts *opts = &top->record_opts;
981 struct evlist *evlist = top->evlist;
982 struct evsel *counter;
984 if (!opts->overwrite)
987 /* only fall back when first event fails */
988 if (evsel != perf_evlist__first(evlist))
991 evlist__for_each_entry(evlist, counter)
992 counter->core.attr.write_backward = false;
993 opts->overwrite = false;
994 pr_debug2("fall back to non-overwrite mode\n");
998 static int perf_top__start_counters(struct perf_top *top)
1001 struct evsel *counter;
1002 struct evlist *evlist = top->evlist;
1003 struct record_opts *opts = &top->record_opts;
1005 if (perf_top__overwrite_check(top)) {
1006 ui__error("perf top only support consistent per-event "
1007 "overwrite setting for all events\n");
1011 perf_evlist__config(evlist, opts, &callchain_param);
1013 evlist__for_each_entry(evlist, counter) {
1015 if (evsel__open(counter, top->evlist->core.cpus,
1016 top->evlist->core.threads) < 0) {
1019 * Specially handle overwrite fall back.
1020 * Because perf top is the only tool which has
1021 * overwrite mode by default, support
1022 * both overwrite and non-overwrite mode, and
1023 * require consistent mode for all events.
1025 * May move it to generic code with more tools
1026 * have similar attribute.
1028 if (perf_missing_features.write_backward &&
1029 perf_top_overwrite_fallback(top, counter))
1032 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1034 ui__warning("%s\n", msg);
1038 perf_evsel__open_strerror(counter, &opts->target,
1039 errno, msg, sizeof(msg));
1040 ui__error("%s\n", msg);
1045 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1046 ui__error("Failed to mmap with %d (%s)\n",
1047 errno, str_error_r(errno, msg, sizeof(msg)));
1057 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1059 if (callchain->mode != CHAIN_NONE) {
1060 if (callchain_register_param(callchain) < 0) {
1061 ui__error("Can't register callchain params.\n");
1069 static struct ordered_events *rotate_queues(struct perf_top *top)
1071 struct ordered_events *in = top->qe.in;
1073 if (top->qe.in == &top->qe.data[1])
1074 top->qe.in = &top->qe.data[0];
1076 top->qe.in = &top->qe.data[1];
1081 static void *process_thread(void *arg)
1083 struct perf_top *top = arg;
1086 struct ordered_events *out, *in = top->qe.in;
1088 if (!in->nr_events) {
1093 out = rotate_queues(top);
1095 pthread_mutex_lock(&top->qe.mutex);
1096 top->qe.rotate = true;
1097 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1098 pthread_mutex_unlock(&top->qe.mutex);
1100 if (ordered_events__flush(out, OE_FLUSH__TOP))
1101 pr_err("failed to process events\n");
1108 * Allow only 'top->delay_secs' seconds behind samples.
1110 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1112 union perf_event *event = qevent->event;
1113 u64 delay_timestamp;
1115 if (event->header.type != PERF_RECORD_SAMPLE)
1118 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1119 return delay_timestamp < last_timestamp;
1122 static int deliver_event(struct ordered_events *qe,
1123 struct ordered_event *qevent)
1125 struct perf_top *top = qe->data;
1126 struct evlist *evlist = top->evlist;
1127 struct perf_session *session = top->session;
1128 union perf_event *event = qevent->event;
1129 struct perf_sample sample;
1130 struct evsel *evsel;
1131 struct machine *machine;
1134 if (should_drop(qevent, top)) {
1140 ret = perf_evlist__parse_sample(evlist, event, &sample);
1142 pr_err("Can't parse sample, err = %d\n", ret);
1146 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1147 assert(evsel != NULL);
1149 if (event->header.type == PERF_RECORD_SAMPLE) {
1150 if (evswitch__discard(&top->evswitch, evsel))
1155 switch (sample.cpumode) {
1156 case PERF_RECORD_MISC_USER:
1158 if (top->hide_user_symbols)
1160 machine = &session->machines.host;
1162 case PERF_RECORD_MISC_KERNEL:
1163 ++top->kernel_samples;
1164 if (top->hide_kernel_symbols)
1166 machine = &session->machines.host;
1168 case PERF_RECORD_MISC_GUEST_KERNEL:
1169 ++top->guest_kernel_samples;
1170 machine = perf_session__find_machine(session,
1173 case PERF_RECORD_MISC_GUEST_USER:
1174 ++top->guest_us_samples;
1176 * TODO: we don't process guest user from host side
1177 * except simple counting.
1181 if (event->header.type == PERF_RECORD_SAMPLE)
1183 machine = &session->machines.host;
1187 if (event->header.type == PERF_RECORD_SAMPLE) {
1188 perf_event__process_sample(&top->tool, event, evsel,
1190 } else if (event->header.type == PERF_RECORD_LOST) {
1191 perf_top__process_lost(top, event, evsel);
1192 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1193 perf_top__process_lost_samples(top, event, evsel);
1194 } else if (event->header.type < PERF_RECORD_MAX) {
1195 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1196 machine__process_event(machine, event, &sample);
1198 ++session->evlist->stats.nr_unknown_events;
1205 static void init_process_thread(struct perf_top *top)
1207 ordered_events__init(&top->qe.data[0], deliver_event, top);
1208 ordered_events__init(&top->qe.data[1], deliver_event, top);
1209 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1210 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1211 top->qe.in = &top->qe.data[0];
1212 pthread_mutex_init(&top->qe.mutex, NULL);
1213 pthread_cond_init(&top->qe.cond, NULL);
1216 static int __cmd_top(struct perf_top *top)
1218 struct record_opts *opts = &top->record_opts;
1219 pthread_t thread, thread_process;
1222 if (!top->annotation_opts.objdump_path) {
1223 ret = perf_env__lookup_objdump(&top->session->header.env,
1224 &top->annotation_opts.objdump_path);
1229 ret = callchain_param__setup_sample_type(&callchain_param);
1233 if (perf_session__register_idle_thread(top->session) < 0)
1236 if (top->nr_threads_synthesize > 1)
1237 perf_set_multithreaded();
1239 init_process_thread(top);
1241 if (opts->record_namespaces)
1242 top->tool.namespace_events = true;
1244 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1245 &top->session->machines.host,
1248 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1250 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1251 top->evlist->core.threads, false,
1252 top->nr_threads_synthesize);
1254 if (top->nr_threads_synthesize > 1)
1255 perf_set_singlethreaded();
1257 if (perf_hpp_list.socket) {
1258 ret = perf_env__read_cpu_topology_map(&perf_env);
1260 char errbuf[BUFSIZ];
1261 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1263 ui__error("Could not read the CPU topology map: %s\n", err);
1268 ret = perf_top__start_counters(top);
1272 top->session->evlist = top->evlist;
1273 perf_session__set_id_hdr_size(top->session);
1276 * When perf is starting the traced process, all the events (apart from
1277 * group members) have enable_on_exec=1 set, so don't spoil it by
1278 * prematurely enabling them.
1280 * XXX 'top' still doesn't start workloads like record, trace, but should,
1281 * so leave the check here.
1283 if (!target__none(&opts->target))
1284 evlist__enable(top->evlist);
1287 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1288 ui__error("Could not create process thread.\n");
1292 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1293 display_thread), top)) {
1294 ui__error("Could not create display thread.\n");
1295 goto out_join_thread;
1298 if (top->realtime_prio) {
1299 struct sched_param param;
1301 param.sched_priority = top->realtime_prio;
1302 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1303 ui__error("Could not set realtime priority.\n");
1308 /* Wait for a minimal set of events before starting the snapshot */
1309 perf_evlist__poll(top->evlist, 100);
1311 perf_top__mmap_read(top);
1314 u64 hits = top->samples;
1316 perf_top__mmap_read(top);
1318 if (opts->overwrite || (hits == top->samples))
1319 ret = perf_evlist__poll(top->evlist, 100);
1322 perf_top__resize(top);
1329 pthread_join(thread, NULL);
1331 pthread_cond_signal(&top->qe.cond);
1332 pthread_join(thread_process, NULL);
1337 callchain_opt(const struct option *opt, const char *arg, int unset)
1339 symbol_conf.use_callchain = true;
1340 return record_callchain_opt(opt, arg, unset);
1344 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1346 struct callchain_param *callchain = opt->value;
1348 callchain->enabled = !unset;
1349 callchain->record_mode = CALLCHAIN_FP;
1355 symbol_conf.use_callchain = false;
1356 callchain->record_mode = CALLCHAIN_NONE;
1360 return parse_callchain_top_opt(arg);
1363 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1365 if (!strcmp(var, "top.call-graph")) {
1366 var = "call-graph.record-mode";
1367 return perf_default_config(var, value, cb);
1369 if (!strcmp(var, "top.children")) {
1370 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1378 parse_percent_limit(const struct option *opt, const char *arg,
1379 int unset __maybe_unused)
1381 struct perf_top *top = opt->value;
1383 top->min_percent = strtof(arg, NULL);
1387 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1388 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1390 int cmd_top(int argc, const char **argv)
1392 char errbuf[BUFSIZ];
1393 struct perf_top top = {
1397 .mmap_pages = UINT_MAX,
1398 .user_freq = UINT_MAX,
1399 .user_interval = ULLONG_MAX,
1400 .freq = 4000, /* 4 KHz */
1405 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1406 * when we pause, fix that and reenable. Probably using a
1407 * separate evlist with a dummy event, i.e. a non-overwrite
1408 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1409 * stays in overwrite mode. -acme
1412 .sample_time = true,
1413 .sample_time_set = true,
1415 .max_stack = sysctl__max_stack(),
1416 .annotation_opts = annotation__default_options,
1417 .nr_threads_synthesize = UINT_MAX,
1419 struct record_opts *opts = &top.record_opts;
1420 struct target *target = &opts->target;
1421 const struct option options[] = {
1422 OPT_CALLBACK('e', "event", &top.evlist, "event",
1423 "event selector. use 'perf list' to list available events",
1424 parse_events_option),
1425 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1426 OPT_STRING('p', "pid", &target->pid, "pid",
1427 "profile events on existing process id"),
1428 OPT_STRING('t', "tid", &target->tid, "tid",
1429 "profile events on existing thread id"),
1430 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1431 "system-wide collection from all CPUs"),
1432 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1433 "list of cpus to monitor"),
1434 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1435 "file", "vmlinux pathname"),
1436 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1437 "don't load vmlinux even if found"),
1438 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1439 "file", "kallsyms pathname"),
1440 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1441 "hide kernel symbols"),
1442 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1443 "number of mmap data pages",
1444 perf_evlist__parse_mmap_pages),
1445 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1446 "collect data with this RT SCHED_FIFO priority"),
1447 OPT_INTEGER('d', "delay", &top.delay_secs,
1448 "number of seconds to delay between refreshes"),
1449 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1450 "dump the symbol table used for profiling"),
1451 OPT_INTEGER('f', "count-filter", &top.count_filter,
1452 "only display functions with more events than this"),
1453 OPT_BOOLEAN(0, "group", &opts->group,
1454 "put the counters into a counter group"),
1455 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1456 "child tasks do not inherit counters"),
1457 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1458 "symbol to annotate"),
1459 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1460 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1461 "profile at this frequency",
1462 record__parse_freq),
1463 OPT_INTEGER('E', "entries", &top.print_entries,
1464 "display this many functions"),
1465 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1466 "hide user symbols"),
1467 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1468 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1469 OPT_INCR('v', "verbose", &verbose,
1470 "be more verbose (show counter open errors, etc)"),
1471 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1472 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1473 " Please refer the man page for the complete list."),
1474 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1475 "output field(s): overhead, period, sample plus all of sort keys"),
1476 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1477 "Show a column with the number of samples"),
1478 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1479 NULL, "enables call-graph recording and display",
1481 OPT_CALLBACK(0, "call-graph", &callchain_param,
1482 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1483 top_callchain_help, &parse_callchain_opt),
1484 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1485 "Accumulate callchains of children and show total overhead as well"),
1486 OPT_INTEGER(0, "max-stack", &top.max_stack,
1487 "Set the maximum stack depth when parsing the callchain. "
1488 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1489 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1490 "ignore callees of these functions in call graphs",
1491 report_parse_ignore_callees_opt),
1492 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1493 "Show a column with the sum of periods"),
1494 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1495 "only consider symbols in these dsos"),
1496 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1497 "only consider symbols in these comms"),
1498 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1499 "only consider these symbols"),
1500 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1501 "Interleave source code with assembly code (default)"),
1502 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1503 "Display raw encoding of assembly instructions (default)"),
1504 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1505 "Enable kernel symbol demangling"),
1506 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1507 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1508 "objdump binary to use for disassembly and annotations"),
1509 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1510 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1511 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1512 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1513 "Don't show entries under that percent", parse_percent_limit),
1514 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1515 "How to display percentage of filtered entries", parse_filter_percentage),
1516 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1518 "don't try to adjust column width, use these fixed values"),
1519 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1520 "per thread proc mmap processing timeout in ms"),
1521 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1522 "branch any", "sample any taken branches",
1523 parse_branch_stack),
1524 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1525 "branch filter mask", "branch stack filter modes",
1526 parse_branch_stack),
1527 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1528 "Show raw trace event output (do not use print fmt or plugins)"),
1529 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1530 "Show entries in a hierarchy"),
1531 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1532 "Use a backward ring buffer, default: no"),
1533 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1534 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1535 "number of thread to run event synthesize"),
1536 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1537 "Record namespaces events"),
1538 OPTS_EVSWITCH(&top.evswitch),
1541 struct evlist *sb_evlist = NULL;
1542 const char * const top_usage[] = {
1543 "perf top [<options>]",
1546 int status = hists__init();
1551 top.annotation_opts.min_pcnt = 5;
1552 top.annotation_opts.context = 4;
1554 top.evlist = evlist__new();
1555 if (top.evlist == NULL)
1558 status = perf_config(perf_top_config, &top);
1562 argc = parse_options(argc, argv, options, top_usage, 0);
1564 usage_with_options(top_usage, options);
1566 if (!top.evlist->core.nr_entries &&
1567 perf_evlist__add_default(top.evlist) < 0) {
1568 pr_err("Not enough memory for event selector list\n");
1569 goto out_delete_evlist;
1572 status = evswitch__init(&top.evswitch, top.evlist, stderr);
1574 goto out_delete_evlist;
1576 if (symbol_conf.report_hierarchy) {
1577 /* disable incompatible options */
1578 symbol_conf.event_group = false;
1579 symbol_conf.cumulate_callchain = false;
1582 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1583 parse_options_usage(top_usage, options, "fields", 0);
1584 parse_options_usage(NULL, options, "hierarchy", 0);
1585 goto out_delete_evlist;
1589 if (opts->branch_stack && callchain_param.enabled)
1590 symbol_conf.show_branchflag_count = true;
1592 sort__mode = SORT_MODE__TOP;
1593 /* display thread wants entries to be collapsed in a different tree */
1594 perf_hpp_list.need_collapse = 1;
1598 else if (top.use_tui)
1601 setup_browser(false);
1603 if (setup_sorting(top.evlist) < 0) {
1605 parse_options_usage(top_usage, options, "s", 1);
1607 parse_options_usage(sort_order ? NULL : top_usage,
1608 options, "fields", 0);
1609 goto out_delete_evlist;
1612 status = target__validate(target);
1614 target__strerror(target, status, errbuf, BUFSIZ);
1615 ui__warning("%s\n", errbuf);
1618 status = target__parse_uid(target);
1620 int saved_errno = errno;
1622 target__strerror(target, status, errbuf, BUFSIZ);
1623 ui__error("%s\n", errbuf);
1625 status = -saved_errno;
1626 goto out_delete_evlist;
1629 if (target__none(target))
1630 target->system_wide = true;
1632 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1633 ui__error("Couldn't create thread/CPU maps: %s\n",
1634 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1635 goto out_delete_evlist;
1638 if (top.delay_secs < 1)
1641 if (record_opts__config(opts)) {
1643 goto out_delete_evlist;
1646 top.sym_evsel = perf_evlist__first(top.evlist);
1648 if (!callchain_param.enabled) {
1649 symbol_conf.cumulate_callchain = false;
1650 perf_hpp__cancel_cumulate();
1653 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1654 callchain_param.order = ORDER_CALLER;
1656 status = symbol__annotation_init();
1658 goto out_delete_evlist;
1660 annotation_config__init();
1662 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1663 status = symbol__init(NULL);
1665 goto out_delete_evlist;
1667 sort__setup_elide(stdout);
1669 get_term_dimensions(&top.winsize);
1670 if (top.print_entries == 0) {
1671 perf_top__update_print_entries(&top);
1672 signal(SIGWINCH, winch_sig);
1675 top.session = perf_session__new(NULL, false, NULL);
1676 if (IS_ERR(top.session)) {
1677 status = PTR_ERR(top.session);
1678 goto out_delete_evlist;
1681 if (!top.record_opts.no_bpf_event)
1682 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1684 if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1685 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1686 opts->no_bpf_event = true;
1689 status = __cmd_top(&top);
1691 if (!opts->no_bpf_event)
1692 perf_evlist__stop_sb_thread(sb_evlist);
1695 evlist__delete(top.evlist);
1696 perf_session__delete(top.session);