1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * builtin-top.c
4 *
5 * Builtin top command: Display a continuously updated profile of
6 * any workload, CPU or specific PID.
7 *
8 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10 *
11 * Improvements and fixes by:
12 *
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>
18 */
19 #include "builtin.h"
20
21 #include "perf.h"
22
23 #include "util/annotate.h"
24 #include "util/bpf-event.h"
25 #include "util/config.h"
26 #include "util/color.h"
27 #include "util/dso.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/evsel_config.h"
31 #include "util/event.h"
32 #include "util/machine.h"
33 #include "util/map.h"
34 #include "util/mmap.h"
35 #include "util/session.h"
36 #include "util/thread.h"
37 #include "util/symbol.h"
38 #include "util/synthetic-events.h"
39 #include "util/top.h"
40 #include "util/util.h"
41 #include <linux/rbtree.h>
42 #include <subcmd/parse-options.h>
43 #include "util/parse-events.h"
44 #include "util/callchain.h"
45 #include "util/cpumap.h"
46 #include "util/sort.h"
47 #include "util/string2.h"
48 #include "util/term.h"
49 #include "util/intlist.h"
50 #include "util/parse-branch-options.h"
51 #include "arch/common.h"
52 #include "ui/ui.h"
53
54 #include "util/debug.h"
55 #include "util/ordered-events.h"
56 #include "util/pfm.h"
57
58 #include <assert.h>
59 #include <elf.h>
60 #include <fcntl.h>
61
62 #include <stdio.h>
63 #include <termios.h>
64 #include <unistd.h>
65 #include <inttypes.h>
66
67 #include <errno.h>
68 #include <time.h>
69 #include <sched.h>
70 #include <signal.h>
71
72 #include <sys/syscall.h>
73 #include <sys/ioctl.h>
74 #include <poll.h>
75 #include <sys/prctl.h>
76 #include <sys/wait.h>
77 #include <sys/uio.h>
78 #include <sys/utsname.h>
79 #include <sys/mman.h>
80
81 #include <linux/stringify.h>
82 #include <linux/time64.h>
83 #include <linux/types.h>
84 #include <linux/err.h>
85
86 #include <linux/ctype.h>
87 #include <perf/mmap.h>
88
89 static volatile int done;
90 static volatile int resize;
91
92 #define HEADER_LINE_NR 5
93
perf_top__update_print_entries(struct perf_top * top)94 static void perf_top__update_print_entries(struct perf_top *top)
95 {
96 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
97 }
98
winch_sig(int sig __maybe_unused)99 static void winch_sig(int sig __maybe_unused)
100 {
101 resize = 1;
102 }
103
perf_top__resize(struct perf_top * top)104 static void perf_top__resize(struct perf_top *top)
105 {
106 get_term_dimensions(&top->winsize);
107 perf_top__update_print_entries(top);
108 }
109
perf_top__parse_source(struct perf_top * top,struct hist_entry * he)110 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
111 {
112 struct evsel *evsel;
113 struct symbol *sym;
114 struct annotation *notes;
115 struct map *map;
116 int err = -1;
117
118 if (!he || !he->ms.sym)
119 return -1;
120
121 evsel = hists_to_evsel(he->hists);
122
123 sym = he->ms.sym;
124 map = he->ms.map;
125
126 /*
127 * We can't annotate with just /proc/kallsyms
128 */
129 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
130 !dso__is_kcore(map->dso)) {
131 pr_err("Can't annotate %s: No vmlinux file was found in the "
132 "path\n", sym->name);
133 sleep(1);
134 return -1;
135 }
136
137 notes = symbol__annotation(sym);
138 pthread_mutex_lock(¬es->lock);
139
140 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
141 pthread_mutex_unlock(¬es->lock);
142 pr_err("Not enough memory for annotating '%s' symbol!\n",
143 sym->name);
144 sleep(1);
145 return err;
146 }
147
148 err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
149 if (err == 0) {
150 top->sym_filter_entry = he;
151 } else {
152 char msg[BUFSIZ];
153 symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
154 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
155 }
156
157 pthread_mutex_unlock(¬es->lock);
158 return err;
159 }
160
__zero_source_counters(struct hist_entry * he)161 static void __zero_source_counters(struct hist_entry *he)
162 {
163 struct symbol *sym = he->ms.sym;
164 symbol__annotate_zero_histograms(sym);
165 }
166
ui__warn_map_erange(struct map * map,struct symbol * sym,u64 ip)167 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
168 {
169 struct utsname uts;
170 int err = uname(&uts);
171
172 ui__warning("Out of bounds address found:\n\n"
173 "Addr: %" PRIx64 "\n"
174 "DSO: %s %c\n"
175 "Map: %" PRIx64 "-%" PRIx64 "\n"
176 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
177 "Arch: %s\n"
178 "Kernel: %s\n"
179 "Tools: %s\n\n"
180 "Not all samples will be on the annotation output.\n\n"
181 "Please report to linux-kernel@vger.kernel.org\n",
182 ip, map->dso->long_name, dso__symtab_origin(map->dso),
183 map->start, map->end, sym->start, sym->end,
184 sym->binding == STB_GLOBAL ? 'g' :
185 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
186 err ? "[unknown]" : uts.machine,
187 err ? "[unknown]" : uts.release, perf_version_string);
188 if (use_browser <= 0)
189 sleep(5);
190
191 map->erange_warned = true;
192 }
193
perf_top__record_precise_ip(struct perf_top * top,struct hist_entry * he,struct perf_sample * sample,struct evsel * evsel,u64 ip)194 static void perf_top__record_precise_ip(struct perf_top *top,
195 struct hist_entry *he,
196 struct perf_sample *sample,
197 struct evsel *evsel, u64 ip)
198 {
199 struct annotation *notes;
200 struct symbol *sym = he->ms.sym;
201 int err = 0;
202
203 if (sym == NULL || (use_browser == 0 &&
204 (top->sym_filter_entry == NULL ||
205 top->sym_filter_entry->ms.sym != sym)))
206 return;
207
208 notes = symbol__annotation(sym);
209
210 if (pthread_mutex_trylock(¬es->lock))
211 return;
212
213 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
214
215 pthread_mutex_unlock(¬es->lock);
216
217 if (unlikely(err)) {
218 /*
219 * This function is now called with he->hists->lock held.
220 * Release it before going to sleep.
221 */
222 pthread_mutex_unlock(&he->hists->lock);
223
224 if (err == -ERANGE && !he->ms.map->erange_warned)
225 ui__warn_map_erange(he->ms.map, sym, ip);
226 else if (err == -ENOMEM) {
227 pr_err("Not enough memory for annotating '%s' symbol!\n",
228 sym->name);
229 sleep(1);
230 }
231
232 pthread_mutex_lock(&he->hists->lock);
233 }
234 }
235
perf_top__show_details(struct perf_top * top)236 static void perf_top__show_details(struct perf_top *top)
237 {
238 struct hist_entry *he = top->sym_filter_entry;
239 struct evsel *evsel;
240 struct annotation *notes;
241 struct symbol *symbol;
242 int more;
243
244 if (!he)
245 return;
246
247 evsel = hists_to_evsel(he->hists);
248
249 symbol = he->ms.sym;
250 notes = symbol__annotation(symbol);
251
252 pthread_mutex_lock(¬es->lock);
253
254 symbol__calc_percent(symbol, evsel);
255
256 if (notes->src == NULL)
257 goto out_unlock;
258
259 printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
260 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
261
262 more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
263
264 if (top->evlist->enabled) {
265 if (top->zero)
266 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
267 else
268 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
269 }
270 if (more != 0)
271 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
272 out_unlock:
273 pthread_mutex_unlock(¬es->lock);
274 }
275
perf_top__resort_hists(struct perf_top * t)276 static void perf_top__resort_hists(struct perf_top *t)
277 {
278 struct evlist *evlist = t->evlist;
279 struct evsel *pos;
280
281 evlist__for_each_entry(evlist, pos) {
282 struct hists *hists = evsel__hists(pos);
283
284 /*
285 * unlink existing entries so that they can be linked
286 * in a correct order in hists__match() below.
287 */
288 hists__unlink(hists);
289
290 if (evlist->enabled) {
291 if (t->zero) {
292 hists__delete_entries(hists);
293 } else {
294 hists__decay_entries(hists, t->hide_user_symbols,
295 t->hide_kernel_symbols);
296 }
297 }
298
299 hists__collapse_resort(hists, NULL);
300
301 /* Non-group events are considered as leader */
302 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
303 struct hists *leader_hists = evsel__hists(pos->leader);
304
305 hists__match(leader_hists, hists);
306 hists__link(leader_hists, hists);
307 }
308 }
309
310 evlist__for_each_entry(evlist, pos) {
311 evsel__output_resort(pos, NULL);
312 }
313 }
314
perf_top__print_sym_table(struct perf_top * top)315 static void perf_top__print_sym_table(struct perf_top *top)
316 {
317 char bf[160];
318 int printed = 0;
319 const int win_width = top->winsize.ws_col - 1;
320 struct evsel *evsel = top->sym_evsel;
321 struct hists *hists = evsel__hists(evsel);
322
323 puts(CONSOLE_CLEAR);
324
325 perf_top__header_snprintf(top, bf, sizeof(bf));
326 printf("%s\n", bf);
327
328 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
329
330 if (!top->record_opts.overwrite &&
331 (top->evlist->stats.nr_lost_warned !=
332 top->evlist->stats.nr_events[PERF_RECORD_LOST])) {
333 top->evlist->stats.nr_lost_warned =
334 top->evlist->stats.nr_events[PERF_RECORD_LOST];
335 color_fprintf(stdout, PERF_COLOR_RED,
336 "WARNING: LOST %d chunks, Check IO/CPU overload",
337 top->evlist->stats.nr_lost_warned);
338 ++printed;
339 }
340
341 if (top->sym_filter_entry) {
342 perf_top__show_details(top);
343 return;
344 }
345
346 perf_top__resort_hists(top);
347
348 hists__output_recalc_col_len(hists, top->print_entries - printed);
349 putchar('\n');
350 hists__fprintf(hists, false, top->print_entries - printed, win_width,
351 top->min_percent, stdout, !symbol_conf.use_callchain);
352 }
353
prompt_integer(int * target,const char * msg)354 static void prompt_integer(int *target, const char *msg)
355 {
356 char *buf = malloc(0), *p;
357 size_t dummy = 0;
358 int tmp;
359
360 fprintf(stdout, "\n%s: ", msg);
361 if (getline(&buf, &dummy, stdin) < 0)
362 return;
363
364 p = strchr(buf, '\n');
365 if (p)
366 *p = 0;
367
368 p = buf;
369 while(*p) {
370 if (!isdigit(*p))
371 goto out_free;
372 p++;
373 }
374 tmp = strtoul(buf, NULL, 10);
375 *target = tmp;
376 out_free:
377 free(buf);
378 }
379
prompt_percent(int * target,const char * msg)380 static void prompt_percent(int *target, const char *msg)
381 {
382 int tmp = 0;
383
384 prompt_integer(&tmp, msg);
385 if (tmp >= 0 && tmp <= 100)
386 *target = tmp;
387 }
388
perf_top__prompt_symbol(struct perf_top * top,const char * msg)389 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
390 {
391 char *buf = malloc(0), *p;
392 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
393 struct hists *hists = evsel__hists(top->sym_evsel);
394 struct rb_node *next;
395 size_t dummy = 0;
396
397 /* zero counters of active symbol */
398 if (syme) {
399 __zero_source_counters(syme);
400 top->sym_filter_entry = NULL;
401 }
402
403 fprintf(stdout, "\n%s: ", msg);
404 if (getline(&buf, &dummy, stdin) < 0)
405 goto out_free;
406
407 p = strchr(buf, '\n');
408 if (p)
409 *p = 0;
410
411 next = rb_first_cached(&hists->entries);
412 while (next) {
413 n = rb_entry(next, struct hist_entry, rb_node);
414 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
415 found = n;
416 break;
417 }
418 next = rb_next(&n->rb_node);
419 }
420
421 if (!found) {
422 fprintf(stderr, "Sorry, %s is not active.\n", buf);
423 sleep(1);
424 } else
425 perf_top__parse_source(top, found);
426
427 out_free:
428 free(buf);
429 }
430
perf_top__print_mapped_keys(struct perf_top * top)431 static void perf_top__print_mapped_keys(struct perf_top *top)
432 {
433 char *name = NULL;
434
435 if (top->sym_filter_entry) {
436 struct symbol *sym = top->sym_filter_entry->ms.sym;
437 name = sym->name;
438 }
439
440 fprintf(stdout, "\nMapped keys:\n");
441 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
442 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
443
444 if (top->evlist->core.nr_entries > 1)
445 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", evsel__name(top->sym_evsel));
446
447 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
448
449 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
450 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
451 fprintf(stdout, "\t[S] stop annotation.\n");
452
453 fprintf(stdout,
454 "\t[K] hide kernel symbols. \t(%s)\n",
455 top->hide_kernel_symbols ? "yes" : "no");
456 fprintf(stdout,
457 "\t[U] hide user symbols. \t(%s)\n",
458 top->hide_user_symbols ? "yes" : "no");
459 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
460 fprintf(stdout, "\t[qQ] quit.\n");
461 }
462
perf_top__key_mapped(struct perf_top * top,int c)463 static int perf_top__key_mapped(struct perf_top *top, int c)
464 {
465 switch (c) {
466 case 'd':
467 case 'e':
468 case 'f':
469 case 'z':
470 case 'q':
471 case 'Q':
472 case 'K':
473 case 'U':
474 case 'F':
475 case 's':
476 case 'S':
477 return 1;
478 case 'E':
479 return top->evlist->core.nr_entries > 1 ? 1 : 0;
480 default:
481 break;
482 }
483
484 return 0;
485 }
486
perf_top__handle_keypress(struct perf_top * top,int c)487 static bool perf_top__handle_keypress(struct perf_top *top, int c)
488 {
489 bool ret = true;
490
491 if (!perf_top__key_mapped(top, c)) {
492 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
493 struct termios save;
494
495 perf_top__print_mapped_keys(top);
496 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
497 fflush(stdout);
498
499 set_term_quiet_input(&save);
500
501 poll(&stdin_poll, 1, -1);
502 c = getc(stdin);
503
504 tcsetattr(0, TCSAFLUSH, &save);
505 if (!perf_top__key_mapped(top, c))
506 return ret;
507 }
508
509 switch (c) {
510 case 'd':
511 prompt_integer(&top->delay_secs, "Enter display delay");
512 if (top->delay_secs < 1)
513 top->delay_secs = 1;
514 break;
515 case 'e':
516 prompt_integer(&top->print_entries, "Enter display entries (lines)");
517 if (top->print_entries == 0) {
518 perf_top__resize(top);
519 signal(SIGWINCH, winch_sig);
520 } else {
521 signal(SIGWINCH, SIG_DFL);
522 }
523 break;
524 case 'E':
525 if (top->evlist->core.nr_entries > 1) {
526 /* Select 0 as the default event: */
527 int counter = 0;
528
529 fprintf(stderr, "\nAvailable events:");
530
531 evlist__for_each_entry(top->evlist, top->sym_evsel)
532 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, evsel__name(top->sym_evsel));
533
534 prompt_integer(&counter, "Enter details event counter");
535
536 if (counter >= top->evlist->core.nr_entries) {
537 top->sym_evsel = evlist__first(top->evlist);
538 fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
539 sleep(1);
540 break;
541 }
542 evlist__for_each_entry(top->evlist, top->sym_evsel)
543 if (top->sym_evsel->idx == counter)
544 break;
545 } else
546 top->sym_evsel = evlist__first(top->evlist);
547 break;
548 case 'f':
549 prompt_integer(&top->count_filter, "Enter display event count filter");
550 break;
551 case 'F':
552 prompt_percent(&top->annotation_opts.min_pcnt,
553 "Enter details display event filter (percent)");
554 break;
555 case 'K':
556 top->hide_kernel_symbols = !top->hide_kernel_symbols;
557 break;
558 case 'q':
559 case 'Q':
560 printf("exiting.\n");
561 if (top->dump_symtab)
562 perf_session__fprintf_dsos(top->session, stderr);
563 ret = false;
564 break;
565 case 's':
566 perf_top__prompt_symbol(top, "Enter details symbol");
567 break;
568 case 'S':
569 if (!top->sym_filter_entry)
570 break;
571 else {
572 struct hist_entry *syme = top->sym_filter_entry;
573
574 top->sym_filter_entry = NULL;
575 __zero_source_counters(syme);
576 }
577 break;
578 case 'U':
579 top->hide_user_symbols = !top->hide_user_symbols;
580 break;
581 case 'z':
582 top->zero = !top->zero;
583 break;
584 default:
585 break;
586 }
587
588 return ret;
589 }
590
perf_top__sort_new_samples(void * arg)591 static void perf_top__sort_new_samples(void *arg)
592 {
593 struct perf_top *t = arg;
594
595 if (t->evlist->selected != NULL)
596 t->sym_evsel = t->evlist->selected;
597
598 perf_top__resort_hists(t);
599
600 if (t->lost || t->drop)
601 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
602 }
603
stop_top(void)604 static void stop_top(void)
605 {
606 session_done = 1;
607 done = 1;
608 }
609
display_thread_tui(void * arg)610 static void *display_thread_tui(void *arg)
611 {
612 struct evsel *pos;
613 struct perf_top *top = arg;
614 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
615 struct hist_browser_timer hbt = {
616 .timer = perf_top__sort_new_samples,
617 .arg = top,
618 .refresh = top->delay_secs,
619 };
620 int ret;
621
622 /* In order to read symbols from other namespaces perf to needs to call
623 * setns(2). This isn't permitted if the struct_fs has multiple users.
624 * unshare(2) the fs so that we may continue to setns into namespaces
625 * that we're observing.
626 */
627 unshare(CLONE_FS);
628
629 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
630
631 repeat:
632 perf_top__sort_new_samples(top);
633
634 /*
635 * Initialize the uid_filter_str, in the future the TUI will allow
636 * Zooming in/out UIDs. For now just use whatever the user passed
637 * via --uid.
638 */
639 evlist__for_each_entry(top->evlist, pos) {
640 struct hists *hists = evsel__hists(pos);
641 hists->uid_filter_str = top->record_opts.target.uid_str;
642 }
643
644 ret = evlist__tui_browse_hists(top->evlist, help, &hbt, top->min_percent,
645 &top->session->header.env, !top->record_opts.overwrite,
646 &top->annotation_opts);
647 if (ret == K_RELOAD) {
648 top->zero = true;
649 goto repeat;
650 } else
651 stop_top();
652
653 return NULL;
654 }
655
display_sig(int sig __maybe_unused)656 static void display_sig(int sig __maybe_unused)
657 {
658 stop_top();
659 }
660
display_setup_sig(void)661 static void display_setup_sig(void)
662 {
663 signal(SIGSEGV, sighandler_dump_stack);
664 signal(SIGFPE, sighandler_dump_stack);
665 signal(SIGINT, display_sig);
666 signal(SIGQUIT, display_sig);
667 signal(SIGTERM, display_sig);
668 }
669
display_thread(void * arg)670 static void *display_thread(void *arg)
671 {
672 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
673 struct termios save;
674 struct perf_top *top = arg;
675 int delay_msecs, c;
676
677 /* In order to read symbols from other namespaces perf to needs to call
678 * setns(2). This isn't permitted if the struct_fs has multiple users.
679 * unshare(2) the fs so that we may continue to setns into namespaces
680 * that we're observing.
681 */
682 unshare(CLONE_FS);
683
684 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
685
686 display_setup_sig();
687 pthread__unblock_sigwinch();
688 repeat:
689 delay_msecs = top->delay_secs * MSEC_PER_SEC;
690 set_term_quiet_input(&save);
691 /* trash return*/
692 clearerr(stdin);
693 if (poll(&stdin_poll, 1, 0) > 0)
694 getc(stdin);
695
696 while (!done) {
697 perf_top__print_sym_table(top);
698 /*
699 * Either timeout expired or we got an EINTR due to SIGWINCH,
700 * refresh screen in both cases.
701 */
702 switch (poll(&stdin_poll, 1, delay_msecs)) {
703 case 0:
704 continue;
705 case -1:
706 if (errno == EINTR)
707 continue;
708 __fallthrough;
709 default:
710 c = getc(stdin);
711 tcsetattr(0, TCSAFLUSH, &save);
712
713 if (perf_top__handle_keypress(top, c))
714 goto repeat;
715 stop_top();
716 }
717 }
718
719 tcsetattr(0, TCSAFLUSH, &save);
720 return NULL;
721 }
722
hist_iter__top_callback(struct hist_entry_iter * iter,struct addr_location * al,bool single,void * arg)723 static int hist_iter__top_callback(struct hist_entry_iter *iter,
724 struct addr_location *al, bool single,
725 void *arg)
726 {
727 struct perf_top *top = arg;
728 struct hist_entry *he = iter->he;
729 struct evsel *evsel = iter->evsel;
730
731 if (perf_hpp_list.sym && single)
732 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
733
734 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
735 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
736 NULL);
737 return 0;
738 }
739
perf_event__process_sample(struct perf_tool * tool,const union perf_event * event,struct evsel * evsel,struct perf_sample * sample,struct machine * machine)740 static void perf_event__process_sample(struct perf_tool *tool,
741 const union perf_event *event,
742 struct evsel *evsel,
743 struct perf_sample *sample,
744 struct machine *machine)
745 {
746 struct perf_top *top = container_of(tool, struct perf_top, tool);
747 struct addr_location al;
748 int err;
749
750 if (!machine && perf_guest) {
751 static struct intlist *seen;
752
753 if (!seen)
754 seen = intlist__new(NULL);
755
756 if (!intlist__has_entry(seen, sample->pid)) {
757 pr_err("Can't find guest [%d]'s kernel information\n",
758 sample->pid);
759 intlist__add(seen, sample->pid);
760 }
761 return;
762 }
763
764 if (!machine) {
765 pr_err("%u unprocessable samples recorded.\r",
766 top->session->evlist->stats.nr_unprocessable_samples++);
767 return;
768 }
769
770 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
771 top->exact_samples++;
772
773 if (machine__resolve(machine, &al, sample) < 0)
774 return;
775
776 if (top->stitch_lbr)
777 al.thread->lbr_stitch_enable = true;
778
779 if (!machine->kptr_restrict_warned &&
780 symbol_conf.kptr_restrict &&
781 al.cpumode == PERF_RECORD_MISC_KERNEL) {
782 if (!evlist__exclude_kernel(top->session->evlist)) {
783 ui__warning(
784 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
785 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
786 "Kernel%s samples will not be resolved.\n",
787 al.map && map__has_symbols(al.map) ?
788 " modules" : "");
789 if (use_browser <= 0)
790 sleep(5);
791 }
792 machine->kptr_restrict_warned = true;
793 }
794
795 if (al.sym == NULL && al.map != NULL) {
796 const char *msg = "Kernel samples will not be resolved.\n";
797 /*
798 * As we do lazy loading of symtabs we only will know if the
799 * specified vmlinux file is invalid when we actually have a
800 * hit in kernel space and then try to load it. So if we get
801 * here and there are _no_ symbols in the DSO backing the
802 * kernel map, bail out.
803 *
804 * We may never get here, for instance, if we use -K/
805 * --hide-kernel-symbols, even if the user specifies an
806 * invalid --vmlinux ;-)
807 */
808 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
809 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
810 if (symbol_conf.vmlinux_name) {
811 char serr[256];
812 dso__strerror_load(al.map->dso, serr, sizeof(serr));
813 ui__warning("The %s file can't be used: %s\n%s",
814 symbol_conf.vmlinux_name, serr, msg);
815 } else {
816 ui__warning("A vmlinux file was not found.\n%s",
817 msg);
818 }
819
820 if (use_browser <= 0)
821 sleep(5);
822 top->vmlinux_warned = true;
823 }
824 }
825
826 if (al.sym == NULL || !al.sym->idle) {
827 struct hists *hists = evsel__hists(evsel);
828 struct hist_entry_iter iter = {
829 .evsel = evsel,
830 .sample = sample,
831 .add_entry_cb = hist_iter__top_callback,
832 };
833
834 if (symbol_conf.cumulate_callchain)
835 iter.ops = &hist_iter_cumulative;
836 else
837 iter.ops = &hist_iter_normal;
838
839 pthread_mutex_lock(&hists->lock);
840
841 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
842 if (err < 0)
843 pr_err("Problem incrementing symbol period, skipping event\n");
844
845 pthread_mutex_unlock(&hists->lock);
846 }
847
848 addr_location__put(&al);
849 }
850
851 static void
perf_top__process_lost(struct perf_top * top,union perf_event * event,struct evsel * evsel)852 perf_top__process_lost(struct perf_top *top, union perf_event *event,
853 struct evsel *evsel)
854 {
855 top->lost += event->lost.lost;
856 top->lost_total += event->lost.lost;
857 evsel->evlist->stats.total_lost += event->lost.lost;
858 }
859
860 static void
perf_top__process_lost_samples(struct perf_top * top,union perf_event * event,struct evsel * evsel)861 perf_top__process_lost_samples(struct perf_top *top,
862 union perf_event *event,
863 struct evsel *evsel)
864 {
865 top->lost += event->lost_samples.lost;
866 top->lost_total += event->lost_samples.lost;
867 evsel->evlist->stats.total_lost_samples += event->lost_samples.lost;
868 }
869
870 static u64 last_timestamp;
871
perf_top__mmap_read_idx(struct perf_top * top,int idx)872 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
873 {
874 struct record_opts *opts = &top->record_opts;
875 struct evlist *evlist = top->evlist;
876 struct mmap *md;
877 union perf_event *event;
878
879 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
880 if (perf_mmap__read_init(&md->core) < 0)
881 return;
882
883 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
884 int ret;
885
886 ret = evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
887 if (ret && ret != -1)
888 break;
889
890 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
891 if (ret)
892 break;
893
894 perf_mmap__consume(&md->core);
895
896 if (top->qe.rotate) {
897 pthread_mutex_lock(&top->qe.mutex);
898 top->qe.rotate = false;
899 pthread_cond_signal(&top->qe.cond);
900 pthread_mutex_unlock(&top->qe.mutex);
901 }
902 }
903
904 perf_mmap__read_done(&md->core);
905 }
906
perf_top__mmap_read(struct perf_top * top)907 static void perf_top__mmap_read(struct perf_top *top)
908 {
909 bool overwrite = top->record_opts.overwrite;
910 struct evlist *evlist = top->evlist;
911 int i;
912
913 if (overwrite)
914 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
915
916 for (i = 0; i < top->evlist->core.nr_mmaps; i++)
917 perf_top__mmap_read_idx(top, i);
918
919 if (overwrite) {
920 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
921 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
922 }
923 }
924
925 /*
926 * Check per-event overwrite term.
927 * perf top should support consistent term for all events.
928 * - All events don't have per-event term
929 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
930 * Nothing change, return 0.
931 * - All events have same per-event term
932 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
933 * Using the per-event setting to replace the opts->overwrite if
934 * they are different, then return 0.
935 * - Events have different per-event term
936 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
937 * Return -1
938 * - Some of the event set per-event term, but some not.
939 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
940 * Return -1
941 */
perf_top__overwrite_check(struct perf_top * top)942 static int perf_top__overwrite_check(struct perf_top *top)
943 {
944 struct record_opts *opts = &top->record_opts;
945 struct evlist *evlist = top->evlist;
946 struct evsel_config_term *term;
947 struct list_head *config_terms;
948 struct evsel *evsel;
949 int set, overwrite = -1;
950
951 evlist__for_each_entry(evlist, evsel) {
952 set = -1;
953 config_terms = &evsel->config_terms;
954 list_for_each_entry(term, config_terms, list) {
955 if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
956 set = term->val.overwrite ? 1 : 0;
957 }
958
959 /* no term for current and previous event (likely) */
960 if ((overwrite < 0) && (set < 0))
961 continue;
962
963 /* has term for both current and previous event, compare */
964 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
965 return -1;
966
967 /* no term for current event but has term for previous one */
968 if ((overwrite >= 0) && (set < 0))
969 return -1;
970
971 /* has term for current event */
972 if ((overwrite < 0) && (set >= 0)) {
973 /* if it's first event, set overwrite */
974 if (evsel == evlist__first(evlist))
975 overwrite = set;
976 else
977 return -1;
978 }
979 }
980
981 if ((overwrite >= 0) && (opts->overwrite != overwrite))
982 opts->overwrite = overwrite;
983
984 return 0;
985 }
986
perf_top_overwrite_fallback(struct perf_top * top,struct evsel * evsel)987 static int perf_top_overwrite_fallback(struct perf_top *top,
988 struct evsel *evsel)
989 {
990 struct record_opts *opts = &top->record_opts;
991 struct evlist *evlist = top->evlist;
992 struct evsel *counter;
993
994 if (!opts->overwrite)
995 return 0;
996
997 /* only fall back when first event fails */
998 if (evsel != evlist__first(evlist))
999 return 0;
1000
1001 evlist__for_each_entry(evlist, counter)
1002 counter->core.attr.write_backward = false;
1003 opts->overwrite = false;
1004 pr_debug2("fall back to non-overwrite mode\n");
1005 return 1;
1006 }
1007
perf_top__start_counters(struct perf_top * top)1008 static int perf_top__start_counters(struct perf_top *top)
1009 {
1010 char msg[BUFSIZ];
1011 struct evsel *counter;
1012 struct evlist *evlist = top->evlist;
1013 struct record_opts *opts = &top->record_opts;
1014
1015 if (perf_top__overwrite_check(top)) {
1016 ui__error("perf top only support consistent per-event "
1017 "overwrite setting for all events\n");
1018 goto out_err;
1019 }
1020
1021 evlist__config(evlist, opts, &callchain_param);
1022
1023 evlist__for_each_entry(evlist, counter) {
1024 try_again:
1025 if (evsel__open(counter, top->evlist->core.cpus,
1026 top->evlist->core.threads) < 0) {
1027
1028 /*
1029 * Specially handle overwrite fall back.
1030 * Because perf top is the only tool which has
1031 * overwrite mode by default, support
1032 * both overwrite and non-overwrite mode, and
1033 * require consistent mode for all events.
1034 *
1035 * May move it to generic code with more tools
1036 * have similar attribute.
1037 */
1038 if (perf_missing_features.write_backward &&
1039 perf_top_overwrite_fallback(top, counter))
1040 goto try_again;
1041
1042 if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
1043 if (verbose > 0)
1044 ui__warning("%s\n", msg);
1045 goto try_again;
1046 }
1047
1048 evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1049 ui__error("%s\n", msg);
1050 goto out_err;
1051 }
1052 }
1053
1054 if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1055 ui__error("Failed to mmap with %d (%s)\n",
1056 errno, str_error_r(errno, msg, sizeof(msg)));
1057 goto out_err;
1058 }
1059
1060 return 0;
1061
1062 out_err:
1063 return -1;
1064 }
1065
callchain_param__setup_sample_type(struct callchain_param * callchain)1066 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1067 {
1068 if (callchain->mode != CHAIN_NONE) {
1069 if (callchain_register_param(callchain) < 0) {
1070 ui__error("Can't register callchain params.\n");
1071 return -EINVAL;
1072 }
1073 }
1074
1075 return 0;
1076 }
1077
rotate_queues(struct perf_top * top)1078 static struct ordered_events *rotate_queues(struct perf_top *top)
1079 {
1080 struct ordered_events *in = top->qe.in;
1081
1082 if (top->qe.in == &top->qe.data[1])
1083 top->qe.in = &top->qe.data[0];
1084 else
1085 top->qe.in = &top->qe.data[1];
1086
1087 return in;
1088 }
1089
process_thread(void * arg)1090 static void *process_thread(void *arg)
1091 {
1092 struct perf_top *top = arg;
1093
1094 while (!done) {
1095 struct ordered_events *out, *in = top->qe.in;
1096
1097 if (!in->nr_events) {
1098 usleep(100);
1099 continue;
1100 }
1101
1102 out = rotate_queues(top);
1103
1104 pthread_mutex_lock(&top->qe.mutex);
1105 top->qe.rotate = true;
1106 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1107 pthread_mutex_unlock(&top->qe.mutex);
1108
1109 if (ordered_events__flush(out, OE_FLUSH__TOP))
1110 pr_err("failed to process events\n");
1111 }
1112
1113 return NULL;
1114 }
1115
1116 /*
1117 * Allow only 'top->delay_secs' seconds behind samples.
1118 */
should_drop(struct ordered_event * qevent,struct perf_top * top)1119 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1120 {
1121 union perf_event *event = qevent->event;
1122 u64 delay_timestamp;
1123
1124 if (event->header.type != PERF_RECORD_SAMPLE)
1125 return false;
1126
1127 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1128 return delay_timestamp < last_timestamp;
1129 }
1130
deliver_event(struct ordered_events * qe,struct ordered_event * qevent)1131 static int deliver_event(struct ordered_events *qe,
1132 struct ordered_event *qevent)
1133 {
1134 struct perf_top *top = qe->data;
1135 struct evlist *evlist = top->evlist;
1136 struct perf_session *session = top->session;
1137 union perf_event *event = qevent->event;
1138 struct perf_sample sample;
1139 struct evsel *evsel;
1140 struct machine *machine;
1141 int ret = -1;
1142
1143 if (should_drop(qevent, top)) {
1144 top->drop++;
1145 top->drop_total++;
1146 return 0;
1147 }
1148
1149 ret = evlist__parse_sample(evlist, event, &sample);
1150 if (ret) {
1151 pr_err("Can't parse sample, err = %d\n", ret);
1152 goto next_event;
1153 }
1154
1155 evsel = evlist__id2evsel(session->evlist, sample.id);
1156 assert(evsel != NULL);
1157
1158 if (event->header.type == PERF_RECORD_SAMPLE) {
1159 if (evswitch__discard(&top->evswitch, evsel))
1160 return 0;
1161 ++top->samples;
1162 }
1163
1164 switch (sample.cpumode) {
1165 case PERF_RECORD_MISC_USER:
1166 ++top->us_samples;
1167 if (top->hide_user_symbols)
1168 goto next_event;
1169 machine = &session->machines.host;
1170 break;
1171 case PERF_RECORD_MISC_KERNEL:
1172 ++top->kernel_samples;
1173 if (top->hide_kernel_symbols)
1174 goto next_event;
1175 machine = &session->machines.host;
1176 break;
1177 case PERF_RECORD_MISC_GUEST_KERNEL:
1178 ++top->guest_kernel_samples;
1179 machine = perf_session__find_machine(session,
1180 sample.pid);
1181 break;
1182 case PERF_RECORD_MISC_GUEST_USER:
1183 ++top->guest_us_samples;
1184 /*
1185 * TODO: we don't process guest user from host side
1186 * except simple counting.
1187 */
1188 goto next_event;
1189 default:
1190 if (event->header.type == PERF_RECORD_SAMPLE)
1191 goto next_event;
1192 machine = &session->machines.host;
1193 break;
1194 }
1195
1196 if (event->header.type == PERF_RECORD_SAMPLE) {
1197 perf_event__process_sample(&top->tool, event, evsel,
1198 &sample, machine);
1199 } else if (event->header.type == PERF_RECORD_LOST) {
1200 perf_top__process_lost(top, event, evsel);
1201 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1202 perf_top__process_lost_samples(top, event, evsel);
1203 } else if (event->header.type < PERF_RECORD_MAX) {
1204 events_stats__inc(&session->evlist->stats, event->header.type);
1205 machine__process_event(machine, event, &sample);
1206 } else
1207 ++session->evlist->stats.nr_unknown_events;
1208
1209 ret = 0;
1210 next_event:
1211 return ret;
1212 }
1213
init_process_thread(struct perf_top * top)1214 static void init_process_thread(struct perf_top *top)
1215 {
1216 ordered_events__init(&top->qe.data[0], deliver_event, top);
1217 ordered_events__init(&top->qe.data[1], deliver_event, top);
1218 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1219 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1220 top->qe.in = &top->qe.data[0];
1221 pthread_mutex_init(&top->qe.mutex, NULL);
1222 pthread_cond_init(&top->qe.cond, NULL);
1223 }
1224
__cmd_top(struct perf_top * top)1225 static int __cmd_top(struct perf_top *top)
1226 {
1227 struct record_opts *opts = &top->record_opts;
1228 pthread_t thread, thread_process;
1229 int ret;
1230
1231 if (!top->annotation_opts.objdump_path) {
1232 ret = perf_env__lookup_objdump(&top->session->header.env,
1233 &top->annotation_opts.objdump_path);
1234 if (ret)
1235 return ret;
1236 }
1237
1238 ret = callchain_param__setup_sample_type(&callchain_param);
1239 if (ret)
1240 return ret;
1241
1242 if (perf_session__register_idle_thread(top->session) < 0)
1243 return ret;
1244
1245 if (top->nr_threads_synthesize > 1)
1246 perf_set_multithreaded();
1247
1248 init_process_thread(top);
1249
1250 if (opts->record_namespaces)
1251 top->tool.namespace_events = true;
1252 if (opts->record_cgroup) {
1253 #ifdef HAVE_FILE_HANDLE
1254 top->tool.cgroup_events = true;
1255 #else
1256 pr_err("cgroup tracking is not supported.\n");
1257 return -1;
1258 #endif
1259 }
1260
1261 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1262 &top->session->machines.host,
1263 &top->record_opts);
1264 if (ret < 0)
1265 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1266
1267 ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1268 &top->session->machines.host);
1269 if (ret < 0)
1270 pr_debug("Couldn't synthesize cgroup events.\n");
1271
1272 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1273 top->evlist->core.threads, false,
1274 top->nr_threads_synthesize);
1275
1276 if (top->nr_threads_synthesize > 1)
1277 perf_set_singlethreaded();
1278
1279 if (perf_hpp_list.socket) {
1280 ret = perf_env__read_cpu_topology_map(&perf_env);
1281 if (ret < 0) {
1282 char errbuf[BUFSIZ];
1283 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1284
1285 ui__error("Could not read the CPU topology map: %s\n", err);
1286 return ret;
1287 }
1288 }
1289
1290 ret = perf_top__start_counters(top);
1291 if (ret)
1292 return ret;
1293
1294 top->session->evlist = top->evlist;
1295 perf_session__set_id_hdr_size(top->session);
1296
1297 /*
1298 * When perf is starting the traced process, all the events (apart from
1299 * group members) have enable_on_exec=1 set, so don't spoil it by
1300 * prematurely enabling them.
1301 *
1302 * XXX 'top' still doesn't start workloads like record, trace, but should,
1303 * so leave the check here.
1304 */
1305 if (!target__none(&opts->target))
1306 evlist__enable(top->evlist);
1307
1308 ret = -1;
1309 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1310 ui__error("Could not create process thread.\n");
1311 return ret;
1312 }
1313
1314 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1315 display_thread), top)) {
1316 ui__error("Could not create display thread.\n");
1317 goto out_join_thread;
1318 }
1319
1320 if (top->realtime_prio) {
1321 struct sched_param param;
1322
1323 param.sched_priority = top->realtime_prio;
1324 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1325 ui__error("Could not set realtime priority.\n");
1326 goto out_join;
1327 }
1328 }
1329
1330 /* Wait for a minimal set of events before starting the snapshot */
1331 evlist__poll(top->evlist, 100);
1332
1333 perf_top__mmap_read(top);
1334
1335 while (!done) {
1336 u64 hits = top->samples;
1337
1338 perf_top__mmap_read(top);
1339
1340 if (opts->overwrite || (hits == top->samples))
1341 ret = evlist__poll(top->evlist, 100);
1342
1343 if (resize) {
1344 perf_top__resize(top);
1345 resize = 0;
1346 }
1347 }
1348
1349 ret = 0;
1350 out_join:
1351 pthread_join(thread, NULL);
1352 out_join_thread:
1353 pthread_cond_signal(&top->qe.cond);
1354 pthread_join(thread_process, NULL);
1355 return ret;
1356 }
1357
1358 static int
callchain_opt(const struct option * opt,const char * arg,int unset)1359 callchain_opt(const struct option *opt, const char *arg, int unset)
1360 {
1361 symbol_conf.use_callchain = true;
1362 return record_callchain_opt(opt, arg, unset);
1363 }
1364
1365 static int
parse_callchain_opt(const struct option * opt,const char * arg,int unset)1366 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1367 {
1368 struct callchain_param *callchain = opt->value;
1369
1370 callchain->enabled = !unset;
1371 callchain->record_mode = CALLCHAIN_FP;
1372
1373 /*
1374 * --no-call-graph
1375 */
1376 if (unset) {
1377 symbol_conf.use_callchain = false;
1378 callchain->record_mode = CALLCHAIN_NONE;
1379 return 0;
1380 }
1381
1382 return parse_callchain_top_opt(arg);
1383 }
1384
perf_top_config(const char * var,const char * value,void * cb __maybe_unused)1385 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1386 {
1387 if (!strcmp(var, "top.call-graph")) {
1388 var = "call-graph.record-mode";
1389 return perf_default_config(var, value, cb);
1390 }
1391 if (!strcmp(var, "top.children")) {
1392 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1393 return 0;
1394 }
1395
1396 return 0;
1397 }
1398
1399 static int
parse_percent_limit(const struct option * opt,const char * arg,int unset __maybe_unused)1400 parse_percent_limit(const struct option *opt, const char *arg,
1401 int unset __maybe_unused)
1402 {
1403 struct perf_top *top = opt->value;
1404
1405 top->min_percent = strtof(arg, NULL);
1406 return 0;
1407 }
1408
1409 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1410 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1411
cmd_top(int argc,const char ** argv)1412 int cmd_top(int argc, const char **argv)
1413 {
1414 char errbuf[BUFSIZ];
1415 struct perf_top top = {
1416 .count_filter = 5,
1417 .delay_secs = 2,
1418 .record_opts = {
1419 .mmap_pages = UINT_MAX,
1420 .user_freq = UINT_MAX,
1421 .user_interval = ULLONG_MAX,
1422 .freq = 4000, /* 4 KHz */
1423 .target = {
1424 .uses_mmap = true,
1425 },
1426 /*
1427 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1428 * when we pause, fix that and reenable. Probably using a
1429 * separate evlist with a dummy event, i.e. a non-overwrite
1430 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1431 * stays in overwrite mode. -acme
1432 * */
1433 .overwrite = 0,
1434 .sample_time = true,
1435 .sample_time_set = true,
1436 },
1437 .max_stack = sysctl__max_stack(),
1438 .annotation_opts = annotation__default_options,
1439 .nr_threads_synthesize = UINT_MAX,
1440 };
1441 struct record_opts *opts = &top.record_opts;
1442 struct target *target = &opts->target;
1443 const struct option options[] = {
1444 OPT_CALLBACK('e', "event", &top.evlist, "event",
1445 "event selector. use 'perf list' to list available events",
1446 parse_events_option),
1447 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1448 OPT_STRING('p', "pid", &target->pid, "pid",
1449 "profile events on existing process id"),
1450 OPT_STRING('t', "tid", &target->tid, "tid",
1451 "profile events on existing thread id"),
1452 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1453 "system-wide collection from all CPUs"),
1454 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1455 "list of cpus to monitor"),
1456 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1457 "file", "vmlinux pathname"),
1458 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1459 "don't load vmlinux even if found"),
1460 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1461 "file", "kallsyms pathname"),
1462 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1463 "hide kernel symbols"),
1464 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1465 "number of mmap data pages", evlist__parse_mmap_pages),
1466 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1467 "collect data with this RT SCHED_FIFO priority"),
1468 OPT_INTEGER('d', "delay", &top.delay_secs,
1469 "number of seconds to delay between refreshes"),
1470 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1471 "dump the symbol table used for profiling"),
1472 OPT_INTEGER('f', "count-filter", &top.count_filter,
1473 "only display functions with more events than this"),
1474 OPT_BOOLEAN(0, "group", &opts->group,
1475 "put the counters into a counter group"),
1476 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1477 "child tasks do not inherit counters"),
1478 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1479 "symbol to annotate"),
1480 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1481 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1482 "profile at this frequency",
1483 record__parse_freq),
1484 OPT_INTEGER('E', "entries", &top.print_entries,
1485 "display this many functions"),
1486 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1487 "hide user symbols"),
1488 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1489 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1490 OPT_INCR('v', "verbose", &verbose,
1491 "be more verbose (show counter open errors, etc)"),
1492 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1493 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1494 " Please refer the man page for the complete list."),
1495 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1496 "output field(s): overhead, period, sample plus all of sort keys"),
1497 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1498 "Show a column with the number of samples"),
1499 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1500 NULL, "enables call-graph recording and display",
1501 &callchain_opt),
1502 OPT_CALLBACK(0, "call-graph", &callchain_param,
1503 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1504 top_callchain_help, &parse_callchain_opt),
1505 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1506 "Accumulate callchains of children and show total overhead as well"),
1507 OPT_INTEGER(0, "max-stack", &top.max_stack,
1508 "Set the maximum stack depth when parsing the callchain. "
1509 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1510 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1511 "ignore callees of these functions in call graphs",
1512 report_parse_ignore_callees_opt),
1513 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1514 "Show a column with the sum of periods"),
1515 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1516 "only consider symbols in these dsos"),
1517 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1518 "only consider symbols in these comms"),
1519 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1520 "only consider these symbols"),
1521 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1522 "Interleave source code with assembly code (default)"),
1523 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1524 "Display raw encoding of assembly instructions (default)"),
1525 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1526 "Enable kernel symbol demangling"),
1527 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1528 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1529 "objdump binary to use for disassembly and annotations"),
1530 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1531 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1532 OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix",
1533 "Add prefix to source file path names in programs (with --prefix-strip)"),
1534 OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N",
1535 "Strip first N entries of source file path name in programs (with --prefix)"),
1536 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1537 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1538 "Don't show entries under that percent", parse_percent_limit),
1539 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1540 "How to display percentage of filtered entries", parse_filter_percentage),
1541 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1542 "width[,width...]",
1543 "don't try to adjust column width, use these fixed values"),
1544 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1545 "per thread proc mmap processing timeout in ms"),
1546 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1547 "branch any", "sample any taken branches",
1548 parse_branch_stack),
1549 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1550 "branch filter mask", "branch stack filter modes",
1551 parse_branch_stack),
1552 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1553 "Show raw trace event output (do not use print fmt or plugins)"),
1554 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1555 "Show entries in a hierarchy"),
1556 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1557 "Use a backward ring buffer, default: no"),
1558 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1559 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1560 "number of thread to run event synthesize"),
1561 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1562 "Record namespaces events"),
1563 OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1564 "Record cgroup events"),
1565 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1566 "Sort the output by the event at the index n in group. "
1567 "If n is invalid, sort by the first event. "
1568 "WARNING: should be used on grouped events."),
1569 OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1570 "Enable LBR callgraph stitching approach"),
1571 #ifdef HAVE_LIBPFM
1572 OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1573 "libpfm4 event selector. use 'perf list' to list available events",
1574 parse_libpfm_events_option),
1575 #endif
1576 OPTS_EVSWITCH(&top.evswitch),
1577 OPT_END()
1578 };
1579 const char * const top_usage[] = {
1580 "perf top [<options>]",
1581 NULL
1582 };
1583 int status = hists__init();
1584
1585 if (status < 0)
1586 return status;
1587
1588 top.annotation_opts.min_pcnt = 5;
1589 top.annotation_opts.context = 4;
1590
1591 top.evlist = evlist__new();
1592 if (top.evlist == NULL)
1593 return -ENOMEM;
1594
1595 status = perf_config(perf_top_config, &top);
1596 if (status)
1597 return status;
1598 /*
1599 * Since the per arch annotation init routine may need the cpuid, read
1600 * it here, since we are not getting this from the perf.data header.
1601 */
1602 status = perf_env__read_cpuid(&perf_env);
1603 if (status) {
1604 /*
1605 * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1606 * warn the user explicitly.
1607 */
1608 eprintf(status == ENOSYS ? 1 : 0, verbose,
1609 "Couldn't read the cpuid for this machine: %s\n",
1610 str_error_r(errno, errbuf, sizeof(errbuf)));
1611 }
1612 top.evlist->env = &perf_env;
1613
1614 argc = parse_options(argc, argv, options, top_usage, 0);
1615 if (argc)
1616 usage_with_options(top_usage, options);
1617
1618 if (annotate_check_args(&top.annotation_opts) < 0)
1619 goto out_delete_evlist;
1620
1621 if (!top.evlist->core.nr_entries &&
1622 evlist__add_default(top.evlist) < 0) {
1623 pr_err("Not enough memory for event selector list\n");
1624 goto out_delete_evlist;
1625 }
1626
1627 status = evswitch__init(&top.evswitch, top.evlist, stderr);
1628 if (status)
1629 goto out_delete_evlist;
1630
1631 if (symbol_conf.report_hierarchy) {
1632 /* disable incompatible options */
1633 symbol_conf.event_group = false;
1634 symbol_conf.cumulate_callchain = false;
1635
1636 if (field_order) {
1637 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1638 parse_options_usage(top_usage, options, "fields", 0);
1639 parse_options_usage(NULL, options, "hierarchy", 0);
1640 goto out_delete_evlist;
1641 }
1642 }
1643
1644 if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1645 pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1646 goto out_delete_evlist;
1647 }
1648
1649 if (opts->branch_stack && callchain_param.enabled)
1650 symbol_conf.show_branchflag_count = true;
1651
1652 sort__mode = SORT_MODE__TOP;
1653 /* display thread wants entries to be collapsed in a different tree */
1654 perf_hpp_list.need_collapse = 1;
1655
1656 if (top.use_stdio)
1657 use_browser = 0;
1658 else if (top.use_tui)
1659 use_browser = 1;
1660
1661 setup_browser(false);
1662
1663 if (setup_sorting(top.evlist) < 0) {
1664 if (sort_order)
1665 parse_options_usage(top_usage, options, "s", 1);
1666 if (field_order)
1667 parse_options_usage(sort_order ? NULL : top_usage,
1668 options, "fields", 0);
1669 goto out_delete_evlist;
1670 }
1671
1672 status = target__validate(target);
1673 if (status) {
1674 target__strerror(target, status, errbuf, BUFSIZ);
1675 ui__warning("%s\n", errbuf);
1676 }
1677
1678 status = target__parse_uid(target);
1679 if (status) {
1680 int saved_errno = errno;
1681
1682 target__strerror(target, status, errbuf, BUFSIZ);
1683 ui__error("%s\n", errbuf);
1684
1685 status = -saved_errno;
1686 goto out_delete_evlist;
1687 }
1688
1689 if (target__none(target))
1690 target->system_wide = true;
1691
1692 if (evlist__create_maps(top.evlist, target) < 0) {
1693 ui__error("Couldn't create thread/CPU maps: %s\n",
1694 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1695 goto out_delete_evlist;
1696 }
1697
1698 if (top.delay_secs < 1)
1699 top.delay_secs = 1;
1700
1701 if (record_opts__config(opts)) {
1702 status = -EINVAL;
1703 goto out_delete_evlist;
1704 }
1705
1706 top.sym_evsel = evlist__first(top.evlist);
1707
1708 if (!callchain_param.enabled) {
1709 symbol_conf.cumulate_callchain = false;
1710 perf_hpp__cancel_cumulate();
1711 }
1712
1713 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1714 callchain_param.order = ORDER_CALLER;
1715
1716 status = symbol__annotation_init();
1717 if (status < 0)
1718 goto out_delete_evlist;
1719
1720 annotation_config__init(&top.annotation_opts);
1721
1722 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1723 status = symbol__init(NULL);
1724 if (status < 0)
1725 goto out_delete_evlist;
1726
1727 sort__setup_elide(stdout);
1728
1729 get_term_dimensions(&top.winsize);
1730 if (top.print_entries == 0) {
1731 perf_top__update_print_entries(&top);
1732 signal(SIGWINCH, winch_sig);
1733 }
1734
1735 top.session = perf_session__new(NULL, false, NULL);
1736 if (IS_ERR(top.session)) {
1737 status = PTR_ERR(top.session);
1738 goto out_delete_evlist;
1739 }
1740
1741 #ifdef HAVE_LIBBPF_SUPPORT
1742 if (!top.record_opts.no_bpf_event) {
1743 top.sb_evlist = evlist__new();
1744
1745 if (top.sb_evlist == NULL) {
1746 pr_err("Couldn't create side band evlist.\n.");
1747 goto out_delete_evlist;
1748 }
1749
1750 if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1751 pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1752 goto out_delete_evlist;
1753 }
1754 }
1755 #endif
1756
1757 if (evlist__start_sb_thread(top.sb_evlist, target)) {
1758 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1759 opts->no_bpf_event = true;
1760 }
1761
1762 status = __cmd_top(&top);
1763
1764 if (!opts->no_bpf_event)
1765 evlist__stop_sb_thread(top.sb_evlist);
1766
1767 out_delete_evlist:
1768 evlist__delete(top.evlist);
1769 perf_session__delete(top.session);
1770
1771 return status;
1772 }
1773