Loading...
1/*
2 * builtin-top.c
3 *
4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
6 *
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 *
9 * Improvements and fixes by:
10 *
11 * Arjan van de Ven <arjan@linux.intel.com>
12 * Yanmin Zhang <yanmin.zhang@intel.com>
13 * Wu Fengguang <fengguang.wu@intel.com>
14 * Mike Galbraith <efault@gmx.de>
15 * Paul Mackerras <paulus@samba.org>
16 *
17 * Released under the GPL v2. (and only v2, not any later version)
18 */
19#include "builtin.h"
20
21#include "perf.h"
22
23#include "util/annotate.h"
24#include "util/cache.h"
25#include "util/color.h"
26#include "util/evlist.h"
27#include "util/evsel.h"
28#include "util/session.h"
29#include "util/symbol.h"
30#include "util/thread.h"
31#include "util/thread_map.h"
32#include "util/top.h"
33#include "util/util.h"
34#include <linux/rbtree.h>
35#include "util/parse-options.h"
36#include "util/parse-events.h"
37#include "util/cpumap.h"
38#include "util/xyarray.h"
39
40#include "util/debug.h"
41
42#include <assert.h>
43#include <fcntl.h>
44
45#include <stdio.h>
46#include <termios.h>
47#include <unistd.h>
48#include <inttypes.h>
49
50#include <errno.h>
51#include <time.h>
52#include <sched.h>
53
54#include <sys/syscall.h>
55#include <sys/ioctl.h>
56#include <sys/poll.h>
57#include <sys/prctl.h>
58#include <sys/wait.h>
59#include <sys/uio.h>
60#include <sys/mman.h>
61
62#include <linux/unistd.h>
63#include <linux/types.h>
64
65static struct perf_top top = {
66 .count_filter = 5,
67 .delay_secs = 2,
68 .display_weighted = -1,
69 .target_pid = -1,
70 .target_tid = -1,
71 .active_symbols = LIST_HEAD_INIT(top.active_symbols),
72 .active_symbols_lock = PTHREAD_MUTEX_INITIALIZER,
73 .active_symbols_cond = PTHREAD_COND_INITIALIZER,
74 .freq = 1000, /* 1 KHz */
75};
76
77static bool system_wide = false;
78
79static bool use_tui, use_stdio;
80
81static int default_interval = 0;
82
83static bool kptr_restrict_warned;
84static bool vmlinux_warned;
85static bool inherit = false;
86static int realtime_prio = 0;
87static bool group = false;
88static unsigned int page_size;
89static unsigned int mmap_pages = 128;
90
91static bool dump_symtab = false;
92
93static struct winsize winsize;
94
95static const char *sym_filter = NULL;
96struct sym_entry *sym_filter_entry_sched = NULL;
97static int sym_pcnt_filter = 5;
98
99/*
100 * Source functions
101 */
102
103void get_term_dimensions(struct winsize *ws)
104{
105 char *s = getenv("LINES");
106
107 if (s != NULL) {
108 ws->ws_row = atoi(s);
109 s = getenv("COLUMNS");
110 if (s != NULL) {
111 ws->ws_col = atoi(s);
112 if (ws->ws_row && ws->ws_col)
113 return;
114 }
115 }
116#ifdef TIOCGWINSZ
117 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
118 ws->ws_row && ws->ws_col)
119 return;
120#endif
121 ws->ws_row = 25;
122 ws->ws_col = 80;
123}
124
125static void update_print_entries(struct winsize *ws)
126{
127 top.print_entries = ws->ws_row;
128
129 if (top.print_entries > 9)
130 top.print_entries -= 9;
131}
132
133static void sig_winch_handler(int sig __used)
134{
135 get_term_dimensions(&winsize);
136 update_print_entries(&winsize);
137}
138
139static int parse_source(struct sym_entry *syme)
140{
141 struct symbol *sym;
142 struct annotation *notes;
143 struct map *map;
144 int err = -1;
145
146 if (!syme)
147 return -1;
148
149 sym = sym_entry__symbol(syme);
150 map = syme->map;
151
152 /*
153 * We can't annotate with just /proc/kallsyms
154 */
155 if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
156 pr_err("Can't annotate %s: No vmlinux file was found in the "
157 "path\n", sym->name);
158 sleep(1);
159 return -1;
160 }
161
162 notes = symbol__annotation(sym);
163 if (notes->src != NULL) {
164 pthread_mutex_lock(¬es->lock);
165 goto out_assign;
166 }
167
168 pthread_mutex_lock(¬es->lock);
169
170 if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
171 pthread_mutex_unlock(¬es->lock);
172 pr_err("Not enough memory for annotating '%s' symbol!\n",
173 sym->name);
174 sleep(1);
175 return err;
176 }
177
178 err = symbol__annotate(sym, syme->map, 0);
179 if (err == 0) {
180out_assign:
181 top.sym_filter_entry = syme;
182 }
183
184 pthread_mutex_unlock(¬es->lock);
185 return err;
186}
187
188static void __zero_source_counters(struct sym_entry *syme)
189{
190 struct symbol *sym = sym_entry__symbol(syme);
191 symbol__annotate_zero_histograms(sym);
192}
193
194static void record_precise_ip(struct sym_entry *syme, struct map *map,
195 int counter, u64 ip)
196{
197 struct annotation *notes;
198 struct symbol *sym;
199
200 if (syme != top.sym_filter_entry)
201 return;
202
203 sym = sym_entry__symbol(syme);
204 notes = symbol__annotation(sym);
205
206 if (pthread_mutex_trylock(¬es->lock))
207 return;
208
209 ip = map->map_ip(map, ip);
210 symbol__inc_addr_samples(sym, map, counter, ip);
211
212 pthread_mutex_unlock(¬es->lock);
213}
214
215static void show_details(struct sym_entry *syme)
216{
217 struct annotation *notes;
218 struct symbol *symbol;
219 int more;
220
221 if (!syme)
222 return;
223
224 symbol = sym_entry__symbol(syme);
225 notes = symbol__annotation(symbol);
226
227 pthread_mutex_lock(¬es->lock);
228
229 if (notes->src == NULL)
230 goto out_unlock;
231
232 printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
233 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
234
235 more = symbol__annotate_printf(symbol, syme->map, top.sym_evsel->idx,
236 0, sym_pcnt_filter, top.print_entries, 4);
237 if (top.zero)
238 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
239 else
240 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
241 if (more != 0)
242 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
243out_unlock:
244 pthread_mutex_unlock(¬es->lock);
245}
246
247static const char CONSOLE_CLEAR[] = "[H[2J";
248
249static void __list_insert_active_sym(struct sym_entry *syme)
250{
251 list_add(&syme->node, &top.active_symbols);
252}
253
254static void print_sym_table(struct perf_session *session)
255{
256 char bf[160];
257 int printed = 0;
258 struct rb_node *nd;
259 struct sym_entry *syme;
260 struct rb_root tmp = RB_ROOT;
261 const int win_width = winsize.ws_col - 1;
262 int sym_width, dso_width, dso_short_width;
263 float sum_ksamples = perf_top__decay_samples(&top, &tmp);
264
265 puts(CONSOLE_CLEAR);
266
267 perf_top__header_snprintf(&top, bf, sizeof(bf));
268 printf("%s\n", bf);
269
270 perf_top__reset_sample_counters(&top);
271
272 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
273
274 if (session->hists.stats.total_lost != 0) {
275 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
276 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
277 session->hists.stats.total_lost);
278 }
279
280 if (top.sym_filter_entry) {
281 show_details(top.sym_filter_entry);
282 return;
283 }
284
285 perf_top__find_widths(&top, &tmp, &dso_width, &dso_short_width,
286 &sym_width);
287
288 if (sym_width + dso_width > winsize.ws_col - 29) {
289 dso_width = dso_short_width;
290 if (sym_width + dso_width > winsize.ws_col - 29)
291 sym_width = winsize.ws_col - dso_width - 29;
292 }
293 putchar('\n');
294 if (top.evlist->nr_entries == 1)
295 printf(" samples pcnt");
296 else
297 printf(" weight samples pcnt");
298
299 if (verbose)
300 printf(" RIP ");
301 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
302 printf(" %s _______ _____",
303 top.evlist->nr_entries == 1 ? " " : "______");
304 if (verbose)
305 printf(" ________________");
306 printf(" %-*.*s", sym_width, sym_width, graph_line);
307 printf(" %-*.*s", dso_width, dso_width, graph_line);
308 puts("\n");
309
310 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
311 struct symbol *sym;
312 double pcnt;
313
314 syme = rb_entry(nd, struct sym_entry, rb_node);
315 sym = sym_entry__symbol(syme);
316 if (++printed > top.print_entries ||
317 (int)syme->snap_count < top.count_filter)
318 continue;
319
320 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
321 sum_ksamples));
322
323 if (top.evlist->nr_entries == 1 || !top.display_weighted)
324 printf("%20.2f ", syme->weight);
325 else
326 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
327
328 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
329 if (verbose)
330 printf(" %016" PRIx64, sym->start);
331 printf(" %-*.*s", sym_width, sym_width, sym->name);
332 printf(" %-*.*s\n", dso_width, dso_width,
333 dso_width >= syme->map->dso->long_name_len ?
334 syme->map->dso->long_name :
335 syme->map->dso->short_name);
336 }
337}
338
339static void prompt_integer(int *target, const char *msg)
340{
341 char *buf = malloc(0), *p;
342 size_t dummy = 0;
343 int tmp;
344
345 fprintf(stdout, "\n%s: ", msg);
346 if (getline(&buf, &dummy, stdin) < 0)
347 return;
348
349 p = strchr(buf, '\n');
350 if (p)
351 *p = 0;
352
353 p = buf;
354 while(*p) {
355 if (!isdigit(*p))
356 goto out_free;
357 p++;
358 }
359 tmp = strtoul(buf, NULL, 10);
360 *target = tmp;
361out_free:
362 free(buf);
363}
364
365static void prompt_percent(int *target, const char *msg)
366{
367 int tmp = 0;
368
369 prompt_integer(&tmp, msg);
370 if (tmp >= 0 && tmp <= 100)
371 *target = tmp;
372}
373
374static void prompt_symbol(struct sym_entry **target, const char *msg)
375{
376 char *buf = malloc(0), *p;
377 struct sym_entry *syme = *target, *n, *found = NULL;
378 size_t dummy = 0;
379
380 /* zero counters of active symbol */
381 if (syme) {
382 __zero_source_counters(syme);
383 *target = NULL;
384 }
385
386 fprintf(stdout, "\n%s: ", msg);
387 if (getline(&buf, &dummy, stdin) < 0)
388 goto out_free;
389
390 p = strchr(buf, '\n');
391 if (p)
392 *p = 0;
393
394 pthread_mutex_lock(&top.active_symbols_lock);
395 syme = list_entry(top.active_symbols.next, struct sym_entry, node);
396 pthread_mutex_unlock(&top.active_symbols_lock);
397
398 list_for_each_entry_safe_from(syme, n, &top.active_symbols, node) {
399 struct symbol *sym = sym_entry__symbol(syme);
400
401 if (!strcmp(buf, sym->name)) {
402 found = syme;
403 break;
404 }
405 }
406
407 if (!found) {
408 fprintf(stderr, "Sorry, %s is not active.\n", buf);
409 sleep(1);
410 return;
411 } else
412 parse_source(found);
413
414out_free:
415 free(buf);
416}
417
418static void print_mapped_keys(void)
419{
420 char *name = NULL;
421
422 if (top.sym_filter_entry) {
423 struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
424 name = sym->name;
425 }
426
427 fprintf(stdout, "\nMapped keys:\n");
428 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top.delay_secs);
429 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top.print_entries);
430
431 if (top.evlist->nr_entries > 1)
432 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(top.sym_evsel));
433
434 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top.count_filter);
435
436 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
437 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
438 fprintf(stdout, "\t[S] stop annotation.\n");
439
440 if (top.evlist->nr_entries > 1)
441 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", top.display_weighted ? 1 : 0);
442
443 fprintf(stdout,
444 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
445 top.hide_kernel_symbols ? "yes" : "no");
446 fprintf(stdout,
447 "\t[U] hide user symbols. \t(%s)\n",
448 top.hide_user_symbols ? "yes" : "no");
449 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top.zero ? 1 : 0);
450 fprintf(stdout, "\t[qQ] quit.\n");
451}
452
453static int key_mapped(int c)
454{
455 switch (c) {
456 case 'd':
457 case 'e':
458 case 'f':
459 case 'z':
460 case 'q':
461 case 'Q':
462 case 'K':
463 case 'U':
464 case 'F':
465 case 's':
466 case 'S':
467 return 1;
468 case 'E':
469 case 'w':
470 return top.evlist->nr_entries > 1 ? 1 : 0;
471 default:
472 break;
473 }
474
475 return 0;
476}
477
478static void handle_keypress(struct perf_session *session, int c)
479{
480 if (!key_mapped(c)) {
481 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
482 struct termios tc, save;
483
484 print_mapped_keys();
485 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
486 fflush(stdout);
487
488 tcgetattr(0, &save);
489 tc = save;
490 tc.c_lflag &= ~(ICANON | ECHO);
491 tc.c_cc[VMIN] = 0;
492 tc.c_cc[VTIME] = 0;
493 tcsetattr(0, TCSANOW, &tc);
494
495 poll(&stdin_poll, 1, -1);
496 c = getc(stdin);
497
498 tcsetattr(0, TCSAFLUSH, &save);
499 if (!key_mapped(c))
500 return;
501 }
502
503 switch (c) {
504 case 'd':
505 prompt_integer(&top.delay_secs, "Enter display delay");
506 if (top.delay_secs < 1)
507 top.delay_secs = 1;
508 break;
509 case 'e':
510 prompt_integer(&top.print_entries, "Enter display entries (lines)");
511 if (top.print_entries == 0) {
512 sig_winch_handler(SIGWINCH);
513 signal(SIGWINCH, sig_winch_handler);
514 } else
515 signal(SIGWINCH, SIG_DFL);
516 break;
517 case 'E':
518 if (top.evlist->nr_entries > 1) {
519 /* Select 0 as the default event: */
520 int counter = 0;
521
522 fprintf(stderr, "\nAvailable events:");
523
524 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
525 fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
526
527 prompt_integer(&counter, "Enter details event counter");
528
529 if (counter >= top.evlist->nr_entries) {
530 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
531 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
532 sleep(1);
533 break;
534 }
535 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
536 if (top.sym_evsel->idx == counter)
537 break;
538 } else
539 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
540 break;
541 case 'f':
542 prompt_integer(&top.count_filter, "Enter display event count filter");
543 break;
544 case 'F':
545 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
546 break;
547 case 'K':
548 top.hide_kernel_symbols = !top.hide_kernel_symbols;
549 break;
550 case 'q':
551 case 'Q':
552 printf("exiting.\n");
553 if (dump_symtab)
554 perf_session__fprintf_dsos(session, stderr);
555 exit(0);
556 case 's':
557 prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
558 break;
559 case 'S':
560 if (!top.sym_filter_entry)
561 break;
562 else {
563 struct sym_entry *syme = top.sym_filter_entry;
564
565 top.sym_filter_entry = NULL;
566 __zero_source_counters(syme);
567 }
568 break;
569 case 'U':
570 top.hide_user_symbols = !top.hide_user_symbols;
571 break;
572 case 'w':
573 top.display_weighted = ~top.display_weighted;
574 break;
575 case 'z':
576 top.zero = !top.zero;
577 break;
578 default:
579 break;
580 }
581}
582
583static void *display_thread_tui(void *arg __used)
584{
585 int err = 0;
586 pthread_mutex_lock(&top.active_symbols_lock);
587 while (list_empty(&top.active_symbols)) {
588 err = pthread_cond_wait(&top.active_symbols_cond,
589 &top.active_symbols_lock);
590 if (err)
591 break;
592 }
593 pthread_mutex_unlock(&top.active_symbols_lock);
594 if (!err)
595 perf_top__tui_browser(&top);
596 exit_browser(0);
597 exit(0);
598 return NULL;
599}
600
601static void *display_thread(void *arg __used)
602{
603 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
604 struct termios tc, save;
605 int delay_msecs, c;
606 struct perf_session *session = (struct perf_session *) arg;
607
608 tcgetattr(0, &save);
609 tc = save;
610 tc.c_lflag &= ~(ICANON | ECHO);
611 tc.c_cc[VMIN] = 0;
612 tc.c_cc[VTIME] = 0;
613
614repeat:
615 delay_msecs = top.delay_secs * 1000;
616 tcsetattr(0, TCSANOW, &tc);
617 /* trash return*/
618 getc(stdin);
619
620 do {
621 print_sym_table(session);
622 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
623
624 c = getc(stdin);
625 tcsetattr(0, TCSAFLUSH, &save);
626
627 handle_keypress(session, c);
628 goto repeat;
629
630 return NULL;
631}
632
633/* Tag samples to be skipped. */
634static const char *skip_symbols[] = {
635 "default_idle",
636 "native_safe_halt",
637 "cpu_idle",
638 "enter_idle",
639 "exit_idle",
640 "mwait_idle",
641 "mwait_idle_with_hints",
642 "poll_idle",
643 "ppc64_runlatch_off",
644 "pseries_dedicated_idle_sleep",
645 NULL
646};
647
648static int symbol_filter(struct map *map, struct symbol *sym)
649{
650 struct sym_entry *syme;
651 const char *name = sym->name;
652 int i;
653
654 /*
655 * ppc64 uses function descriptors and appends a '.' to the
656 * start of every instruction address. Remove it.
657 */
658 if (name[0] == '.')
659 name++;
660
661 if (!strcmp(name, "_text") ||
662 !strcmp(name, "_etext") ||
663 !strcmp(name, "_sinittext") ||
664 !strncmp("init_module", name, 11) ||
665 !strncmp("cleanup_module", name, 14) ||
666 strstr(name, "_text_start") ||
667 strstr(name, "_text_end"))
668 return 1;
669
670 syme = symbol__priv(sym);
671 syme->map = map;
672 symbol__annotate_init(map, sym);
673
674 if (!top.sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
675 /* schedule initial sym_filter_entry setup */
676 sym_filter_entry_sched = syme;
677 sym_filter = NULL;
678 }
679
680 for (i = 0; skip_symbols[i]; i++) {
681 if (!strcmp(skip_symbols[i], name)) {
682 sym->ignore = true;
683 break;
684 }
685 }
686
687 return 0;
688}
689
690static void perf_event__process_sample(const union perf_event *event,
691 struct perf_sample *sample,
692 struct perf_session *session)
693{
694 u64 ip = event->ip.ip;
695 struct sym_entry *syme;
696 struct addr_location al;
697 struct machine *machine;
698 u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
699
700 ++top.samples;
701
702 switch (origin) {
703 case PERF_RECORD_MISC_USER:
704 ++top.us_samples;
705 if (top.hide_user_symbols)
706 return;
707 machine = perf_session__find_host_machine(session);
708 break;
709 case PERF_RECORD_MISC_KERNEL:
710 ++top.kernel_samples;
711 if (top.hide_kernel_symbols)
712 return;
713 machine = perf_session__find_host_machine(session);
714 break;
715 case PERF_RECORD_MISC_GUEST_KERNEL:
716 ++top.guest_kernel_samples;
717 machine = perf_session__find_machine(session, event->ip.pid);
718 break;
719 case PERF_RECORD_MISC_GUEST_USER:
720 ++top.guest_us_samples;
721 /*
722 * TODO: we don't process guest user from host side
723 * except simple counting.
724 */
725 return;
726 default:
727 return;
728 }
729
730 if (!machine && perf_guest) {
731 pr_err("Can't find guest [%d]'s kernel information\n",
732 event->ip.pid);
733 return;
734 }
735
736 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
737 top.exact_samples++;
738
739 if (perf_event__preprocess_sample(event, session, &al, sample,
740 symbol_filter) < 0 ||
741 al.filtered)
742 return;
743
744 if (!kptr_restrict_warned &&
745 symbol_conf.kptr_restrict &&
746 al.cpumode == PERF_RECORD_MISC_KERNEL) {
747 ui__warning(
748"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
749"Check /proc/sys/kernel/kptr_restrict.\n\n"
750"Kernel%s samples will not be resolved.\n",
751 !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
752 " modules" : "");
753 if (use_browser <= 0)
754 sleep(5);
755 kptr_restrict_warned = true;
756 }
757
758 if (al.sym == NULL) {
759 const char *msg = "Kernel samples will not be resolved.\n";
760 /*
761 * As we do lazy loading of symtabs we only will know if the
762 * specified vmlinux file is invalid when we actually have a
763 * hit in kernel space and then try to load it. So if we get
764 * here and there are _no_ symbols in the DSO backing the
765 * kernel map, bail out.
766 *
767 * We may never get here, for instance, if we use -K/
768 * --hide-kernel-symbols, even if the user specifies an
769 * invalid --vmlinux ;-)
770 */
771 if (!kptr_restrict_warned && !vmlinux_warned &&
772 al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
773 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
774 if (symbol_conf.vmlinux_name) {
775 ui__warning("The %s file can't be used.\n%s",
776 symbol_conf.vmlinux_name, msg);
777 } else {
778 ui__warning("A vmlinux file was not found.\n%s",
779 msg);
780 }
781
782 if (use_browser <= 0)
783 sleep(5);
784 vmlinux_warned = true;
785 }
786
787 return;
788 }
789
790 /* let's see, whether we need to install initial sym_filter_entry */
791 if (sym_filter_entry_sched) {
792 top.sym_filter_entry = sym_filter_entry_sched;
793 sym_filter_entry_sched = NULL;
794 if (parse_source(top.sym_filter_entry) < 0) {
795 struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
796
797 pr_err("Can't annotate %s", sym->name);
798 if (top.sym_filter_entry->map->dso->symtab_type == SYMTAB__KALLSYMS) {
799 pr_err(": No vmlinux file was found in the path:\n");
800 machine__fprintf_vmlinux_path(machine, stderr);
801 } else
802 pr_err(".\n");
803 exit(1);
804 }
805 }
806
807 syme = symbol__priv(al.sym);
808 if (!al.sym->ignore) {
809 struct perf_evsel *evsel;
810
811 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
812 assert(evsel != NULL);
813 syme->count[evsel->idx]++;
814 record_precise_ip(syme, al.map, evsel->idx, ip);
815 pthread_mutex_lock(&top.active_symbols_lock);
816 if (list_empty(&syme->node) || !syme->node.next) {
817 static bool first = true;
818 __list_insert_active_sym(syme);
819 if (first) {
820 pthread_cond_broadcast(&top.active_symbols_cond);
821 first = false;
822 }
823 }
824 pthread_mutex_unlock(&top.active_symbols_lock);
825 }
826}
827
828static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
829{
830 struct perf_sample sample;
831 union perf_event *event;
832 int ret;
833
834 while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
835 ret = perf_session__parse_sample(self, event, &sample);
836 if (ret) {
837 pr_err("Can't parse sample, err = %d\n", ret);
838 continue;
839 }
840
841 if (event->header.type == PERF_RECORD_SAMPLE)
842 perf_event__process_sample(event, &sample, self);
843 else
844 perf_event__process(event, &sample, self);
845 }
846}
847
848static void perf_session__mmap_read(struct perf_session *self)
849{
850 int i;
851
852 for (i = 0; i < top.evlist->nr_mmaps; i++)
853 perf_session__mmap_read_idx(self, i);
854}
855
856static void start_counters(struct perf_evlist *evlist)
857{
858 struct perf_evsel *counter;
859
860 list_for_each_entry(counter, &evlist->entries, node) {
861 struct perf_event_attr *attr = &counter->attr;
862
863 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
864
865 if (top.freq) {
866 attr->sample_type |= PERF_SAMPLE_PERIOD;
867 attr->freq = 1;
868 attr->sample_freq = top.freq;
869 }
870
871 if (evlist->nr_entries > 1) {
872 attr->sample_type |= PERF_SAMPLE_ID;
873 attr->read_format |= PERF_FORMAT_ID;
874 }
875
876 attr->mmap = 1;
877 attr->inherit = inherit;
878try_again:
879 if (perf_evsel__open(counter, top.evlist->cpus,
880 top.evlist->threads, group) < 0) {
881 int err = errno;
882
883 if (err == EPERM || err == EACCES) {
884 ui__warning_paranoid();
885 goto out_err;
886 }
887 /*
888 * If it's cycles then fall back to hrtimer
889 * based cpu-clock-tick sw counter, which
890 * is always available even if no PMU support:
891 */
892 if (attr->type == PERF_TYPE_HARDWARE &&
893 attr->config == PERF_COUNT_HW_CPU_CYCLES) {
894 if (verbose)
895 ui__warning("Cycles event not supported,\n"
896 "trying to fall back to cpu-clock-ticks\n");
897
898 attr->type = PERF_TYPE_SOFTWARE;
899 attr->config = PERF_COUNT_SW_CPU_CLOCK;
900 goto try_again;
901 }
902
903 if (err == ENOENT) {
904 ui__warning("The %s event is not supported.\n",
905 event_name(counter));
906 goto out_err;
907 }
908
909 ui__warning("The sys_perf_event_open() syscall "
910 "returned with %d (%s). /bin/dmesg "
911 "may provide additional information.\n"
912 "No CONFIG_PERF_EVENTS=y kernel support "
913 "configured?\n", err, strerror(err));
914 goto out_err;
915 }
916 }
917
918 if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
919 ui__warning("Failed to mmap with %d (%s)\n",
920 errno, strerror(errno));
921 goto out_err;
922 }
923
924 return;
925
926out_err:
927 exit_browser(0);
928 exit(0);
929}
930
931static int __cmd_top(void)
932{
933 pthread_t thread;
934 int ret __used;
935 /*
936 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
937 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
938 */
939 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
940 if (session == NULL)
941 return -ENOMEM;
942
943 if (top.target_tid != -1)
944 perf_event__synthesize_thread_map(top.evlist->threads,
945 perf_event__process, session);
946 else
947 perf_event__synthesize_threads(perf_event__process, session);
948
949 start_counters(top.evlist);
950 session->evlist = top.evlist;
951 perf_session__update_sample_type(session);
952
953 /* Wait for a minimal set of events before starting the snapshot */
954 poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
955
956 perf_session__mmap_read(session);
957
958 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
959 display_thread), session)) {
960 printf("Could not create display thread.\n");
961 exit(-1);
962 }
963
964 if (realtime_prio) {
965 struct sched_param param;
966
967 param.sched_priority = realtime_prio;
968 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
969 printf("Could not set realtime priority.\n");
970 exit(-1);
971 }
972 }
973
974 while (1) {
975 u64 hits = top.samples;
976
977 perf_session__mmap_read(session);
978
979 if (hits == top.samples)
980 ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
981 }
982
983 return 0;
984}
985
986static const char * const top_usage[] = {
987 "perf top [<options>]",
988 NULL
989};
990
991static const struct option options[] = {
992 OPT_CALLBACK('e', "event", &top.evlist, "event",
993 "event selector. use 'perf list' to list available events",
994 parse_events_option),
995 OPT_INTEGER('c', "count", &default_interval,
996 "event period to sample"),
997 OPT_INTEGER('p', "pid", &top.target_pid,
998 "profile events on existing process id"),
999 OPT_INTEGER('t', "tid", &top.target_tid,
1000 "profile events on existing thread id"),
1001 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1002 "system-wide collection from all CPUs"),
1003 OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1004 "list of cpus to monitor"),
1005 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1006 "file", "vmlinux pathname"),
1007 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1008 "hide kernel symbols"),
1009 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1010 OPT_INTEGER('r', "realtime", &realtime_prio,
1011 "collect data with this RT SCHED_FIFO priority"),
1012 OPT_INTEGER('d', "delay", &top.delay_secs,
1013 "number of seconds to delay between refreshes"),
1014 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1015 "dump the symbol table used for profiling"),
1016 OPT_INTEGER('f', "count-filter", &top.count_filter,
1017 "only display functions with more events than this"),
1018 OPT_BOOLEAN('g', "group", &group,
1019 "put the counters into a counter group"),
1020 OPT_BOOLEAN('i', "inherit", &inherit,
1021 "child tasks inherit counters"),
1022 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1023 "symbol to annotate"),
1024 OPT_BOOLEAN('z', "zero", &top.zero,
1025 "zero history across updates"),
1026 OPT_INTEGER('F', "freq", &top.freq,
1027 "profile at this frequency"),
1028 OPT_INTEGER('E', "entries", &top.print_entries,
1029 "display this many functions"),
1030 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1031 "hide user symbols"),
1032 OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1033 OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1034 OPT_INCR('v', "verbose", &verbose,
1035 "be more verbose (show counter open errors, etc)"),
1036 OPT_END()
1037};
1038
1039int cmd_top(int argc, const char **argv, const char *prefix __used)
1040{
1041 struct perf_evsel *pos;
1042 int status = -ENOMEM;
1043
1044 top.evlist = perf_evlist__new(NULL, NULL);
1045 if (top.evlist == NULL)
1046 return -ENOMEM;
1047
1048 page_size = sysconf(_SC_PAGE_SIZE);
1049
1050 argc = parse_options(argc, argv, options, top_usage, 0);
1051 if (argc)
1052 usage_with_options(top_usage, options);
1053
1054 /*
1055 * XXX For now start disabled, only using TUI if explicitely asked for.
1056 * Change that when handle_keys equivalent gets written, live annotation
1057 * done, etc.
1058 */
1059 use_browser = 0;
1060
1061 if (use_stdio)
1062 use_browser = 0;
1063 else if (use_tui)
1064 use_browser = 1;
1065
1066 setup_browser(false);
1067
1068 /* CPU and PID are mutually exclusive */
1069 if (top.target_tid > 0 && top.cpu_list) {
1070 printf("WARNING: PID switch overriding CPU\n");
1071 sleep(1);
1072 top.cpu_list = NULL;
1073 }
1074
1075 if (top.target_pid != -1)
1076 top.target_tid = top.target_pid;
1077
1078 if (perf_evlist__create_maps(top.evlist, top.target_pid,
1079 top.target_tid, top.cpu_list) < 0)
1080 usage_with_options(top_usage, options);
1081
1082 if (!top.evlist->nr_entries &&
1083 perf_evlist__add_default(top.evlist) < 0) {
1084 pr_err("Not enough memory for event selector list\n");
1085 return -ENOMEM;
1086 }
1087
1088 if (top.delay_secs < 1)
1089 top.delay_secs = 1;
1090
1091 /*
1092 * User specified count overrides default frequency.
1093 */
1094 if (default_interval)
1095 top.freq = 0;
1096 else if (top.freq) {
1097 default_interval = top.freq;
1098 } else {
1099 fprintf(stderr, "frequency and count are zero, aborting\n");
1100 exit(EXIT_FAILURE);
1101 }
1102
1103 list_for_each_entry(pos, &top.evlist->entries, node) {
1104 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1105 top.evlist->threads->nr) < 0)
1106 goto out_free_fd;
1107 /*
1108 * Fill in the ones not specifically initialized via -c:
1109 */
1110 if (pos->attr.sample_period)
1111 continue;
1112
1113 pos->attr.sample_period = default_interval;
1114 }
1115
1116 if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1117 perf_evlist__alloc_mmap(top.evlist) < 0)
1118 goto out_free_fd;
1119
1120 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1121
1122 symbol_conf.priv_size = (sizeof(struct sym_entry) + sizeof(struct annotation) +
1123 (top.evlist->nr_entries + 1) * sizeof(unsigned long));
1124
1125 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1126 if (symbol__init() < 0)
1127 return -1;
1128
1129 get_term_dimensions(&winsize);
1130 if (top.print_entries == 0) {
1131 update_print_entries(&winsize);
1132 signal(SIGWINCH, sig_winch_handler);
1133 }
1134
1135 status = __cmd_top();
1136out_free_fd:
1137 perf_evlist__delete(top.evlist);
1138
1139 return status;
1140}
1/*
2 * builtin-top.c
3 *
4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
6 *
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8 * 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
9 *
10 * Improvements and fixes by:
11 *
12 * Arjan van de Ven <arjan@linux.intel.com>
13 * Yanmin Zhang <yanmin.zhang@intel.com>
14 * Wu Fengguang <fengguang.wu@intel.com>
15 * Mike Galbraith <efault@gmx.de>
16 * Paul Mackerras <paulus@samba.org>
17 *
18 * Released under the GPL v2. (and only v2, not any later version)
19 */
20#include "builtin.h"
21
22#include "perf.h"
23
24#include "util/annotate.h"
25#include "util/config.h"
26#include "util/color.h"
27#include "util/drv_configs.h"
28#include "util/evlist.h"
29#include "util/evsel.h"
30#include "util/event.h"
31#include "util/machine.h"
32#include "util/session.h"
33#include "util/symbol.h"
34#include "util/thread.h"
35#include "util/thread_map.h"
36#include "util/top.h"
37#include <linux/rbtree.h>
38#include <subcmd/parse-options.h>
39#include "util/parse-events.h"
40#include "util/cpumap.h"
41#include "util/xyarray.h"
42#include "util/sort.h"
43#include "util/term.h"
44#include "util/intlist.h"
45#include "util/parse-branch-options.h"
46#include "arch/common.h"
47
48#include "util/debug.h"
49
50#include <assert.h>
51#include <elf.h>
52#include <fcntl.h>
53
54#include <stdio.h>
55#include <termios.h>
56#include <unistd.h>
57#include <inttypes.h>
58
59#include <errno.h>
60#include <time.h>
61#include <sched.h>
62#include <signal.h>
63
64#include <sys/syscall.h>
65#include <sys/ioctl.h>
66#include <poll.h>
67#include <sys/prctl.h>
68#include <sys/wait.h>
69#include <sys/uio.h>
70#include <sys/utsname.h>
71#include <sys/mman.h>
72
73#include <linux/stringify.h>
74#include <linux/time64.h>
75#include <linux/types.h>
76
77#include "sane_ctype.h"
78
79static volatile int done;
80static volatile int resize;
81
82#define HEADER_LINE_NR 5
83
84static void perf_top__update_print_entries(struct perf_top *top)
85{
86 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
87}
88
89static void winch_sig(int sig __maybe_unused)
90{
91 resize = 1;
92}
93
94static void perf_top__resize(struct perf_top *top)
95{
96 get_term_dimensions(&top->winsize);
97 perf_top__update_print_entries(top);
98}
99
100static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
101{
102 struct perf_evsel *evsel = hists_to_evsel(he->hists);
103 struct symbol *sym;
104 struct annotation *notes;
105 struct map *map;
106 int err = -1;
107
108 if (!he || !he->ms.sym)
109 return -1;
110
111 sym = he->ms.sym;
112 map = he->ms.map;
113
114 /*
115 * We can't annotate with just /proc/kallsyms
116 */
117 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
118 !dso__is_kcore(map->dso)) {
119 pr_err("Can't annotate %s: No vmlinux file was found in the "
120 "path\n", sym->name);
121 sleep(1);
122 return -1;
123 }
124
125 notes = symbol__annotation(sym);
126 if (notes->src != NULL) {
127 pthread_mutex_lock(¬es->lock);
128 goto out_assign;
129 }
130
131 pthread_mutex_lock(¬es->lock);
132
133 if (symbol__alloc_hist(sym) < 0) {
134 pthread_mutex_unlock(¬es->lock);
135 pr_err("Not enough memory for annotating '%s' symbol!\n",
136 sym->name);
137 sleep(1);
138 return err;
139 }
140
141 err = symbol__annotate(sym, map, evsel, 0, NULL);
142 if (err == 0) {
143out_assign:
144 top->sym_filter_entry = he;
145 } else {
146 char msg[BUFSIZ];
147 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
148 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
149 }
150
151 pthread_mutex_unlock(¬es->lock);
152 return err;
153}
154
155static void __zero_source_counters(struct hist_entry *he)
156{
157 struct symbol *sym = he->ms.sym;
158 symbol__annotate_zero_histograms(sym);
159}
160
161static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
162{
163 struct utsname uts;
164 int err = uname(&uts);
165
166 ui__warning("Out of bounds address found:\n\n"
167 "Addr: %" PRIx64 "\n"
168 "DSO: %s %c\n"
169 "Map: %" PRIx64 "-%" PRIx64 "\n"
170 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
171 "Arch: %s\n"
172 "Kernel: %s\n"
173 "Tools: %s\n\n"
174 "Not all samples will be on the annotation output.\n\n"
175 "Please report to linux-kernel@vger.kernel.org\n",
176 ip, map->dso->long_name, dso__symtab_origin(map->dso),
177 map->start, map->end, sym->start, sym->end,
178 sym->binding == STB_GLOBAL ? 'g' :
179 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
180 err ? "[unknown]" : uts.machine,
181 err ? "[unknown]" : uts.release, perf_version_string);
182 if (use_browser <= 0)
183 sleep(5);
184
185 map->erange_warned = true;
186}
187
188static void perf_top__record_precise_ip(struct perf_top *top,
189 struct hist_entry *he,
190 struct perf_sample *sample,
191 int counter, u64 ip)
192{
193 struct annotation *notes;
194 struct symbol *sym = he->ms.sym;
195 int err = 0;
196
197 if (sym == NULL || (use_browser == 0 &&
198 (top->sym_filter_entry == NULL ||
199 top->sym_filter_entry->ms.sym != sym)))
200 return;
201
202 notes = symbol__annotation(sym);
203
204 if (pthread_mutex_trylock(¬es->lock))
205 return;
206
207 err = hist_entry__inc_addr_samples(he, sample, counter, ip);
208
209 pthread_mutex_unlock(¬es->lock);
210
211 if (unlikely(err)) {
212 /*
213 * This function is now called with he->hists->lock held.
214 * Release it before going to sleep.
215 */
216 pthread_mutex_unlock(&he->hists->lock);
217
218 if (err == -ERANGE && !he->ms.map->erange_warned)
219 ui__warn_map_erange(he->ms.map, sym, ip);
220 else if (err == -ENOMEM) {
221 pr_err("Not enough memory for annotating '%s' symbol!\n",
222 sym->name);
223 sleep(1);
224 }
225
226 pthread_mutex_lock(&he->hists->lock);
227 }
228}
229
230static void perf_top__show_details(struct perf_top *top)
231{
232 struct hist_entry *he = top->sym_filter_entry;
233 struct perf_evsel *evsel = hists_to_evsel(he->hists);
234 struct annotation *notes;
235 struct symbol *symbol;
236 int more;
237
238 if (!he)
239 return;
240
241 symbol = he->ms.sym;
242 notes = symbol__annotation(symbol);
243
244 pthread_mutex_lock(¬es->lock);
245
246 symbol__calc_percent(symbol, evsel);
247
248 if (notes->src == NULL)
249 goto out_unlock;
250
251 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
252 printf(" Events Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
253
254 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
255 0, top->sym_pcnt_filter, top->print_entries, 4);
256
257 if (top->evlist->enabled) {
258 if (top->zero)
259 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
260 else
261 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
262 }
263 if (more != 0)
264 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
265out_unlock:
266 pthread_mutex_unlock(¬es->lock);
267}
268
269static void perf_top__print_sym_table(struct perf_top *top)
270{
271 char bf[160];
272 int printed = 0;
273 const int win_width = top->winsize.ws_col - 1;
274 struct perf_evsel *evsel = top->sym_evsel;
275 struct hists *hists = evsel__hists(evsel);
276
277 puts(CONSOLE_CLEAR);
278
279 perf_top__header_snprintf(top, bf, sizeof(bf));
280 printf("%s\n", bf);
281
282 perf_top__reset_sample_counters(top);
283
284 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
285
286 if (!top->record_opts.overwrite &&
287 (hists->stats.nr_lost_warned !=
288 hists->stats.nr_events[PERF_RECORD_LOST])) {
289 hists->stats.nr_lost_warned =
290 hists->stats.nr_events[PERF_RECORD_LOST];
291 color_fprintf(stdout, PERF_COLOR_RED,
292 "WARNING: LOST %d chunks, Check IO/CPU overload",
293 hists->stats.nr_lost_warned);
294 ++printed;
295 }
296
297 if (top->sym_filter_entry) {
298 perf_top__show_details(top);
299 return;
300 }
301
302 if (top->evlist->enabled) {
303 if (top->zero) {
304 hists__delete_entries(hists);
305 } else {
306 hists__decay_entries(hists, top->hide_user_symbols,
307 top->hide_kernel_symbols);
308 }
309 }
310
311 hists__collapse_resort(hists, NULL);
312 perf_evsel__output_resort(evsel, NULL);
313
314 hists__output_recalc_col_len(hists, top->print_entries - printed);
315 putchar('\n');
316 hists__fprintf(hists, false, top->print_entries - printed, win_width,
317 top->min_percent, stdout, symbol_conf.use_callchain);
318}
319
320static void prompt_integer(int *target, const char *msg)
321{
322 char *buf = malloc(0), *p;
323 size_t dummy = 0;
324 int tmp;
325
326 fprintf(stdout, "\n%s: ", msg);
327 if (getline(&buf, &dummy, stdin) < 0)
328 return;
329
330 p = strchr(buf, '\n');
331 if (p)
332 *p = 0;
333
334 p = buf;
335 while(*p) {
336 if (!isdigit(*p))
337 goto out_free;
338 p++;
339 }
340 tmp = strtoul(buf, NULL, 10);
341 *target = tmp;
342out_free:
343 free(buf);
344}
345
346static void prompt_percent(int *target, const char *msg)
347{
348 int tmp = 0;
349
350 prompt_integer(&tmp, msg);
351 if (tmp >= 0 && tmp <= 100)
352 *target = tmp;
353}
354
355static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
356{
357 char *buf = malloc(0), *p;
358 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
359 struct hists *hists = evsel__hists(top->sym_evsel);
360 struct rb_node *next;
361 size_t dummy = 0;
362
363 /* zero counters of active symbol */
364 if (syme) {
365 __zero_source_counters(syme);
366 top->sym_filter_entry = NULL;
367 }
368
369 fprintf(stdout, "\n%s: ", msg);
370 if (getline(&buf, &dummy, stdin) < 0)
371 goto out_free;
372
373 p = strchr(buf, '\n');
374 if (p)
375 *p = 0;
376
377 next = rb_first(&hists->entries);
378 while (next) {
379 n = rb_entry(next, struct hist_entry, rb_node);
380 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
381 found = n;
382 break;
383 }
384 next = rb_next(&n->rb_node);
385 }
386
387 if (!found) {
388 fprintf(stderr, "Sorry, %s is not active.\n", buf);
389 sleep(1);
390 } else
391 perf_top__parse_source(top, found);
392
393out_free:
394 free(buf);
395}
396
397static void perf_top__print_mapped_keys(struct perf_top *top)
398{
399 char *name = NULL;
400
401 if (top->sym_filter_entry) {
402 struct symbol *sym = top->sym_filter_entry->ms.sym;
403 name = sym->name;
404 }
405
406 fprintf(stdout, "\nMapped keys:\n");
407 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
408 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
409
410 if (top->evlist->nr_entries > 1)
411 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
412
413 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
414
415 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
416 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
417 fprintf(stdout, "\t[S] stop annotation.\n");
418
419 fprintf(stdout,
420 "\t[K] hide kernel symbols. \t(%s)\n",
421 top->hide_kernel_symbols ? "yes" : "no");
422 fprintf(stdout,
423 "\t[U] hide user symbols. \t(%s)\n",
424 top->hide_user_symbols ? "yes" : "no");
425 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
426 fprintf(stdout, "\t[qQ] quit.\n");
427}
428
429static int perf_top__key_mapped(struct perf_top *top, int c)
430{
431 switch (c) {
432 case 'd':
433 case 'e':
434 case 'f':
435 case 'z':
436 case 'q':
437 case 'Q':
438 case 'K':
439 case 'U':
440 case 'F':
441 case 's':
442 case 'S':
443 return 1;
444 case 'E':
445 return top->evlist->nr_entries > 1 ? 1 : 0;
446 default:
447 break;
448 }
449
450 return 0;
451}
452
453static bool perf_top__handle_keypress(struct perf_top *top, int c)
454{
455 bool ret = true;
456
457 if (!perf_top__key_mapped(top, c)) {
458 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
459 struct termios save;
460
461 perf_top__print_mapped_keys(top);
462 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
463 fflush(stdout);
464
465 set_term_quiet_input(&save);
466
467 poll(&stdin_poll, 1, -1);
468 c = getc(stdin);
469
470 tcsetattr(0, TCSAFLUSH, &save);
471 if (!perf_top__key_mapped(top, c))
472 return ret;
473 }
474
475 switch (c) {
476 case 'd':
477 prompt_integer(&top->delay_secs, "Enter display delay");
478 if (top->delay_secs < 1)
479 top->delay_secs = 1;
480 break;
481 case 'e':
482 prompt_integer(&top->print_entries, "Enter display entries (lines)");
483 if (top->print_entries == 0) {
484 perf_top__resize(top);
485 signal(SIGWINCH, winch_sig);
486 } else {
487 signal(SIGWINCH, SIG_DFL);
488 }
489 break;
490 case 'E':
491 if (top->evlist->nr_entries > 1) {
492 /* Select 0 as the default event: */
493 int counter = 0;
494
495 fprintf(stderr, "\nAvailable events:");
496
497 evlist__for_each_entry(top->evlist, top->sym_evsel)
498 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
499
500 prompt_integer(&counter, "Enter details event counter");
501
502 if (counter >= top->evlist->nr_entries) {
503 top->sym_evsel = perf_evlist__first(top->evlist);
504 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
505 sleep(1);
506 break;
507 }
508 evlist__for_each_entry(top->evlist, top->sym_evsel)
509 if (top->sym_evsel->idx == counter)
510 break;
511 } else
512 top->sym_evsel = perf_evlist__first(top->evlist);
513 break;
514 case 'f':
515 prompt_integer(&top->count_filter, "Enter display event count filter");
516 break;
517 case 'F':
518 prompt_percent(&top->sym_pcnt_filter,
519 "Enter details display event filter (percent)");
520 break;
521 case 'K':
522 top->hide_kernel_symbols = !top->hide_kernel_symbols;
523 break;
524 case 'q':
525 case 'Q':
526 printf("exiting.\n");
527 if (top->dump_symtab)
528 perf_session__fprintf_dsos(top->session, stderr);
529 ret = false;
530 break;
531 case 's':
532 perf_top__prompt_symbol(top, "Enter details symbol");
533 break;
534 case 'S':
535 if (!top->sym_filter_entry)
536 break;
537 else {
538 struct hist_entry *syme = top->sym_filter_entry;
539
540 top->sym_filter_entry = NULL;
541 __zero_source_counters(syme);
542 }
543 break;
544 case 'U':
545 top->hide_user_symbols = !top->hide_user_symbols;
546 break;
547 case 'z':
548 top->zero = !top->zero;
549 break;
550 default:
551 break;
552 }
553
554 return ret;
555}
556
557static void perf_top__sort_new_samples(void *arg)
558{
559 struct perf_top *t = arg;
560 struct perf_evsel *evsel = t->sym_evsel;
561 struct hists *hists;
562
563 perf_top__reset_sample_counters(t);
564
565 if (t->evlist->selected != NULL)
566 t->sym_evsel = t->evlist->selected;
567
568 hists = evsel__hists(evsel);
569
570 if (t->evlist->enabled) {
571 if (t->zero) {
572 hists__delete_entries(hists);
573 } else {
574 hists__decay_entries(hists, t->hide_user_symbols,
575 t->hide_kernel_symbols);
576 }
577 }
578
579 hists__collapse_resort(hists, NULL);
580 perf_evsel__output_resort(evsel, NULL);
581}
582
583static void *display_thread_tui(void *arg)
584{
585 struct perf_evsel *pos;
586 struct perf_top *top = arg;
587 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
588 struct hist_browser_timer hbt = {
589 .timer = perf_top__sort_new_samples,
590 .arg = top,
591 .refresh = top->delay_secs,
592 };
593
594 /* In order to read symbols from other namespaces perf to needs to call
595 * setns(2). This isn't permitted if the struct_fs has multiple users.
596 * unshare(2) the fs so that we may continue to setns into namespaces
597 * that we're observing.
598 */
599 unshare(CLONE_FS);
600
601 perf_top__sort_new_samples(top);
602
603 /*
604 * Initialize the uid_filter_str, in the future the TUI will allow
605 * Zooming in/out UIDs. For now juse use whatever the user passed
606 * via --uid.
607 */
608 evlist__for_each_entry(top->evlist, pos) {
609 struct hists *hists = evsel__hists(pos);
610 hists->uid_filter_str = top->record_opts.target.uid_str;
611 }
612
613 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
614 top->min_percent,
615 &top->session->header.env,
616 !top->record_opts.overwrite);
617
618 done = 1;
619 return NULL;
620}
621
622static void display_sig(int sig __maybe_unused)
623{
624 done = 1;
625}
626
627static void display_setup_sig(void)
628{
629 signal(SIGSEGV, sighandler_dump_stack);
630 signal(SIGFPE, sighandler_dump_stack);
631 signal(SIGINT, display_sig);
632 signal(SIGQUIT, display_sig);
633 signal(SIGTERM, display_sig);
634}
635
636static void *display_thread(void *arg)
637{
638 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
639 struct termios save;
640 struct perf_top *top = arg;
641 int delay_msecs, c;
642
643 /* In order to read symbols from other namespaces perf to needs to call
644 * setns(2). This isn't permitted if the struct_fs has multiple users.
645 * unshare(2) the fs so that we may continue to setns into namespaces
646 * that we're observing.
647 */
648 unshare(CLONE_FS);
649
650 display_setup_sig();
651 pthread__unblock_sigwinch();
652repeat:
653 delay_msecs = top->delay_secs * MSEC_PER_SEC;
654 set_term_quiet_input(&save);
655 /* trash return*/
656 getc(stdin);
657
658 while (!done) {
659 perf_top__print_sym_table(top);
660 /*
661 * Either timeout expired or we got an EINTR due to SIGWINCH,
662 * refresh screen in both cases.
663 */
664 switch (poll(&stdin_poll, 1, delay_msecs)) {
665 case 0:
666 continue;
667 case -1:
668 if (errno == EINTR)
669 continue;
670 __fallthrough;
671 default:
672 c = getc(stdin);
673 tcsetattr(0, TCSAFLUSH, &save);
674
675 if (perf_top__handle_keypress(top, c))
676 goto repeat;
677 done = 1;
678 }
679 }
680
681 tcsetattr(0, TCSAFLUSH, &save);
682 return NULL;
683}
684
685static int hist_iter__top_callback(struct hist_entry_iter *iter,
686 struct addr_location *al, bool single,
687 void *arg)
688{
689 struct perf_top *top = arg;
690 struct hist_entry *he = iter->he;
691 struct perf_evsel *evsel = iter->evsel;
692
693 if (perf_hpp_list.sym && single)
694 perf_top__record_precise_ip(top, he, iter->sample, evsel->idx, al->addr);
695
696 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
697 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
698 return 0;
699}
700
701static void perf_event__process_sample(struct perf_tool *tool,
702 const union perf_event *event,
703 struct perf_evsel *evsel,
704 struct perf_sample *sample,
705 struct machine *machine)
706{
707 struct perf_top *top = container_of(tool, struct perf_top, tool);
708 struct addr_location al;
709 int err;
710
711 if (!machine && perf_guest) {
712 static struct intlist *seen;
713
714 if (!seen)
715 seen = intlist__new(NULL);
716
717 if (!intlist__has_entry(seen, sample->pid)) {
718 pr_err("Can't find guest [%d]'s kernel information\n",
719 sample->pid);
720 intlist__add(seen, sample->pid);
721 }
722 return;
723 }
724
725 if (!machine) {
726 pr_err("%u unprocessable samples recorded.\r",
727 top->session->evlist->stats.nr_unprocessable_samples++);
728 return;
729 }
730
731 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
732 top->exact_samples++;
733
734 if (machine__resolve(machine, &al, sample) < 0)
735 return;
736
737 if (!machine->kptr_restrict_warned &&
738 symbol_conf.kptr_restrict &&
739 al.cpumode == PERF_RECORD_MISC_KERNEL) {
740 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
741 ui__warning(
742"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
743"Check /proc/sys/kernel/kptr_restrict.\n\n"
744"Kernel%s samples will not be resolved.\n",
745 al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
746 " modules" : "");
747 if (use_browser <= 0)
748 sleep(5);
749 }
750 machine->kptr_restrict_warned = true;
751 }
752
753 if (al.sym == NULL) {
754 const char *msg = "Kernel samples will not be resolved.\n";
755 /*
756 * As we do lazy loading of symtabs we only will know if the
757 * specified vmlinux file is invalid when we actually have a
758 * hit in kernel space and then try to load it. So if we get
759 * here and there are _no_ symbols in the DSO backing the
760 * kernel map, bail out.
761 *
762 * We may never get here, for instance, if we use -K/
763 * --hide-kernel-symbols, even if the user specifies an
764 * invalid --vmlinux ;-)
765 */
766 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
767 al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
768 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
769 if (symbol_conf.vmlinux_name) {
770 char serr[256];
771 dso__strerror_load(al.map->dso, serr, sizeof(serr));
772 ui__warning("The %s file can't be used: %s\n%s",
773 symbol_conf.vmlinux_name, serr, msg);
774 } else {
775 ui__warning("A vmlinux file was not found.\n%s",
776 msg);
777 }
778
779 if (use_browser <= 0)
780 sleep(5);
781 top->vmlinux_warned = true;
782 }
783 }
784
785 if (al.sym == NULL || !al.sym->idle) {
786 struct hists *hists = evsel__hists(evsel);
787 struct hist_entry_iter iter = {
788 .evsel = evsel,
789 .sample = sample,
790 .add_entry_cb = hist_iter__top_callback,
791 };
792
793 if (symbol_conf.cumulate_callchain)
794 iter.ops = &hist_iter_cumulative;
795 else
796 iter.ops = &hist_iter_normal;
797
798 pthread_mutex_lock(&hists->lock);
799
800 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
801 if (err < 0)
802 pr_err("Problem incrementing symbol period, skipping event\n");
803
804 pthread_mutex_unlock(&hists->lock);
805 }
806
807 addr_location__put(&al);
808}
809
810static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
811{
812 struct record_opts *opts = &top->record_opts;
813 struct perf_evlist *evlist = top->evlist;
814 struct perf_sample sample;
815 struct perf_evsel *evsel;
816 struct perf_mmap *md;
817 struct perf_session *session = top->session;
818 union perf_event *event;
819 struct machine *machine;
820 int ret;
821
822 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
823 if (perf_mmap__read_init(md) < 0)
824 return;
825
826 while ((event = perf_mmap__read_event(md)) != NULL) {
827 ret = perf_evlist__parse_sample(evlist, event, &sample);
828 if (ret) {
829 pr_err("Can't parse sample, err = %d\n", ret);
830 goto next_event;
831 }
832
833 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
834 assert(evsel != NULL);
835
836 if (event->header.type == PERF_RECORD_SAMPLE)
837 ++top->samples;
838
839 switch (sample.cpumode) {
840 case PERF_RECORD_MISC_USER:
841 ++top->us_samples;
842 if (top->hide_user_symbols)
843 goto next_event;
844 machine = &session->machines.host;
845 break;
846 case PERF_RECORD_MISC_KERNEL:
847 ++top->kernel_samples;
848 if (top->hide_kernel_symbols)
849 goto next_event;
850 machine = &session->machines.host;
851 break;
852 case PERF_RECORD_MISC_GUEST_KERNEL:
853 ++top->guest_kernel_samples;
854 machine = perf_session__find_machine(session,
855 sample.pid);
856 break;
857 case PERF_RECORD_MISC_GUEST_USER:
858 ++top->guest_us_samples;
859 /*
860 * TODO: we don't process guest user from host side
861 * except simple counting.
862 */
863 goto next_event;
864 default:
865 if (event->header.type == PERF_RECORD_SAMPLE)
866 goto next_event;
867 machine = &session->machines.host;
868 break;
869 }
870
871
872 if (event->header.type == PERF_RECORD_SAMPLE) {
873 perf_event__process_sample(&top->tool, event, evsel,
874 &sample, machine);
875 } else if (event->header.type < PERF_RECORD_MAX) {
876 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
877 machine__process_event(machine, event, &sample);
878 } else
879 ++session->evlist->stats.nr_unknown_events;
880next_event:
881 perf_mmap__consume(md);
882 }
883
884 perf_mmap__read_done(md);
885}
886
887static void perf_top__mmap_read(struct perf_top *top)
888{
889 bool overwrite = top->record_opts.overwrite;
890 struct perf_evlist *evlist = top->evlist;
891 unsigned long long start, end;
892 int i;
893
894 start = rdclock();
895 if (overwrite)
896 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
897
898 for (i = 0; i < top->evlist->nr_mmaps; i++)
899 perf_top__mmap_read_idx(top, i);
900
901 if (overwrite) {
902 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
903 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
904 }
905 end = rdclock();
906
907 if ((end - start) > (unsigned long long)top->delay_secs * NSEC_PER_SEC)
908 ui__warning("Too slow to read ring buffer.\n"
909 "Please try increasing the period (-c) or\n"
910 "decreasing the freq (-F) or\n"
911 "limiting the number of CPUs (-C)\n");
912}
913
914/*
915 * Check per-event overwrite term.
916 * perf top should support consistent term for all events.
917 * - All events don't have per-event term
918 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
919 * Nothing change, return 0.
920 * - All events have same per-event term
921 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
922 * Using the per-event setting to replace the opts->overwrite if
923 * they are different, then return 0.
924 * - Events have different per-event term
925 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
926 * Return -1
927 * - Some of the event set per-event term, but some not.
928 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
929 * Return -1
930 */
931static int perf_top__overwrite_check(struct perf_top *top)
932{
933 struct record_opts *opts = &top->record_opts;
934 struct perf_evlist *evlist = top->evlist;
935 struct perf_evsel_config_term *term;
936 struct list_head *config_terms;
937 struct perf_evsel *evsel;
938 int set, overwrite = -1;
939
940 evlist__for_each_entry(evlist, evsel) {
941 set = -1;
942 config_terms = &evsel->config_terms;
943 list_for_each_entry(term, config_terms, list) {
944 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
945 set = term->val.overwrite ? 1 : 0;
946 }
947
948 /* no term for current and previous event (likely) */
949 if ((overwrite < 0) && (set < 0))
950 continue;
951
952 /* has term for both current and previous event, compare */
953 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
954 return -1;
955
956 /* no term for current event but has term for previous one */
957 if ((overwrite >= 0) && (set < 0))
958 return -1;
959
960 /* has term for current event */
961 if ((overwrite < 0) && (set >= 0)) {
962 /* if it's first event, set overwrite */
963 if (evsel == perf_evlist__first(evlist))
964 overwrite = set;
965 else
966 return -1;
967 }
968 }
969
970 if ((overwrite >= 0) && (opts->overwrite != overwrite))
971 opts->overwrite = overwrite;
972
973 return 0;
974}
975
976static int perf_top_overwrite_fallback(struct perf_top *top,
977 struct perf_evsel *evsel)
978{
979 struct record_opts *opts = &top->record_opts;
980 struct perf_evlist *evlist = top->evlist;
981 struct perf_evsel *counter;
982
983 if (!opts->overwrite)
984 return 0;
985
986 /* only fall back when first event fails */
987 if (evsel != perf_evlist__first(evlist))
988 return 0;
989
990 evlist__for_each_entry(evlist, counter)
991 counter->attr.write_backward = false;
992 opts->overwrite = false;
993 pr_debug2("fall back to non-overwrite mode\n");
994 return 1;
995}
996
997static int perf_top__start_counters(struct perf_top *top)
998{
999 char msg[BUFSIZ];
1000 struct perf_evsel *counter;
1001 struct perf_evlist *evlist = top->evlist;
1002 struct record_opts *opts = &top->record_opts;
1003
1004 if (perf_top__overwrite_check(top)) {
1005 ui__error("perf top only support consistent per-event "
1006 "overwrite setting for all events\n");
1007 goto out_err;
1008 }
1009
1010 perf_evlist__config(evlist, opts, &callchain_param);
1011
1012 evlist__for_each_entry(evlist, counter) {
1013try_again:
1014 if (perf_evsel__open(counter, top->evlist->cpus,
1015 top->evlist->threads) < 0) {
1016
1017 /*
1018 * Specially handle overwrite fall back.
1019 * Because perf top is the only tool which has
1020 * overwrite mode by default, support
1021 * both overwrite and non-overwrite mode, and
1022 * require consistent mode for all events.
1023 *
1024 * May move it to generic code with more tools
1025 * have similar attribute.
1026 */
1027 if (perf_missing_features.write_backward &&
1028 perf_top_overwrite_fallback(top, counter))
1029 goto try_again;
1030
1031 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1032 if (verbose > 0)
1033 ui__warning("%s\n", msg);
1034 goto try_again;
1035 }
1036
1037 perf_evsel__open_strerror(counter, &opts->target,
1038 errno, msg, sizeof(msg));
1039 ui__error("%s\n", msg);
1040 goto out_err;
1041 }
1042 }
1043
1044 if (perf_evlist__mmap(evlist, opts->mmap_pages) < 0) {
1045 ui__error("Failed to mmap with %d (%s)\n",
1046 errno, str_error_r(errno, msg, sizeof(msg)));
1047 goto out_err;
1048 }
1049
1050 return 0;
1051
1052out_err:
1053 return -1;
1054}
1055
1056static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1057{
1058 if (!perf_hpp_list.sym) {
1059 if (callchain->enabled) {
1060 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1061 return -EINVAL;
1062 }
1063 } else if (callchain->mode != CHAIN_NONE) {
1064 if (callchain_register_param(callchain) < 0) {
1065 ui__error("Can't register callchain params.\n");
1066 return -EINVAL;
1067 }
1068 }
1069
1070 return 0;
1071}
1072
1073static int __cmd_top(struct perf_top *top)
1074{
1075 char msg[512];
1076 struct perf_evsel *pos;
1077 struct perf_evsel_config_term *err_term;
1078 struct perf_evlist *evlist = top->evlist;
1079 struct record_opts *opts = &top->record_opts;
1080 pthread_t thread;
1081 int ret;
1082
1083 top->session = perf_session__new(NULL, false, NULL);
1084 if (top->session == NULL)
1085 return -1;
1086
1087 if (!objdump_path) {
1088 ret = perf_env__lookup_objdump(&top->session->header.env);
1089 if (ret)
1090 goto out_delete;
1091 }
1092
1093 ret = callchain_param__setup_sample_type(&callchain_param);
1094 if (ret)
1095 goto out_delete;
1096
1097 if (perf_session__register_idle_thread(top->session) < 0)
1098 goto out_delete;
1099
1100 if (top->nr_threads_synthesize > 1)
1101 perf_set_multithreaded();
1102
1103 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1104 top->evlist->threads, false,
1105 opts->proc_map_timeout,
1106 top->nr_threads_synthesize);
1107
1108 if (top->nr_threads_synthesize > 1)
1109 perf_set_singlethreaded();
1110
1111 if (perf_hpp_list.socket) {
1112 ret = perf_env__read_cpu_topology_map(&perf_env);
1113 if (ret < 0)
1114 goto out_err_cpu_topo;
1115 }
1116
1117 ret = perf_top__start_counters(top);
1118 if (ret)
1119 goto out_delete;
1120
1121 ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
1122 if (ret) {
1123 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
1124 err_term->val.drv_cfg, perf_evsel__name(pos), errno,
1125 str_error_r(errno, msg, sizeof(msg)));
1126 goto out_delete;
1127 }
1128
1129 top->session->evlist = top->evlist;
1130 perf_session__set_id_hdr_size(top->session);
1131
1132 /*
1133 * When perf is starting the traced process, all the events (apart from
1134 * group members) have enable_on_exec=1 set, so don't spoil it by
1135 * prematurely enabling them.
1136 *
1137 * XXX 'top' still doesn't start workloads like record, trace, but should,
1138 * so leave the check here.
1139 */
1140 if (!target__none(&opts->target))
1141 perf_evlist__enable(top->evlist);
1142
1143 /* Wait for a minimal set of events before starting the snapshot */
1144 perf_evlist__poll(top->evlist, 100);
1145
1146 perf_top__mmap_read(top);
1147
1148 ret = -1;
1149 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1150 display_thread), top)) {
1151 ui__error("Could not create display thread.\n");
1152 goto out_delete;
1153 }
1154
1155 if (top->realtime_prio) {
1156 struct sched_param param;
1157
1158 param.sched_priority = top->realtime_prio;
1159 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1160 ui__error("Could not set realtime priority.\n");
1161 goto out_join;
1162 }
1163 }
1164
1165 while (!done) {
1166 u64 hits = top->samples;
1167
1168 perf_top__mmap_read(top);
1169
1170 if (opts->overwrite || (hits == top->samples))
1171 ret = perf_evlist__poll(top->evlist, 100);
1172
1173 if (resize) {
1174 perf_top__resize(top);
1175 resize = 0;
1176 }
1177 }
1178
1179 ret = 0;
1180out_join:
1181 pthread_join(thread, NULL);
1182out_delete:
1183 perf_session__delete(top->session);
1184 top->session = NULL;
1185
1186 return ret;
1187
1188out_err_cpu_topo: {
1189 char errbuf[BUFSIZ];
1190 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1191
1192 ui__error("Could not read the CPU topology map: %s\n", err);
1193 goto out_delete;
1194}
1195}
1196
1197static int
1198callchain_opt(const struct option *opt, const char *arg, int unset)
1199{
1200 symbol_conf.use_callchain = true;
1201 return record_callchain_opt(opt, arg, unset);
1202}
1203
1204static int
1205parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1206{
1207 struct callchain_param *callchain = opt->value;
1208
1209 callchain->enabled = !unset;
1210 callchain->record_mode = CALLCHAIN_FP;
1211
1212 /*
1213 * --no-call-graph
1214 */
1215 if (unset) {
1216 symbol_conf.use_callchain = false;
1217 callchain->record_mode = CALLCHAIN_NONE;
1218 return 0;
1219 }
1220
1221 return parse_callchain_top_opt(arg);
1222}
1223
1224static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1225{
1226 if (!strcmp(var, "top.call-graph")) {
1227 var = "call-graph.record-mode";
1228 return perf_default_config(var, value, cb);
1229 }
1230 if (!strcmp(var, "top.children")) {
1231 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1232 return 0;
1233 }
1234
1235 return 0;
1236}
1237
1238static int
1239parse_percent_limit(const struct option *opt, const char *arg,
1240 int unset __maybe_unused)
1241{
1242 struct perf_top *top = opt->value;
1243
1244 top->min_percent = strtof(arg, NULL);
1245 return 0;
1246}
1247
1248const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1249 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1250
1251int cmd_top(int argc, const char **argv)
1252{
1253 char errbuf[BUFSIZ];
1254 struct perf_top top = {
1255 .count_filter = 5,
1256 .delay_secs = 2,
1257 .record_opts = {
1258 .mmap_pages = UINT_MAX,
1259 .user_freq = UINT_MAX,
1260 .user_interval = ULLONG_MAX,
1261 .freq = 4000, /* 4 KHz */
1262 .target = {
1263 .uses_mmap = true,
1264 },
1265 .proc_map_timeout = 500,
1266 .overwrite = 1,
1267 },
1268 .max_stack = sysctl_perf_event_max_stack,
1269 .sym_pcnt_filter = 5,
1270 .nr_threads_synthesize = UINT_MAX,
1271 };
1272 struct record_opts *opts = &top.record_opts;
1273 struct target *target = &opts->target;
1274 const struct option options[] = {
1275 OPT_CALLBACK('e', "event", &top.evlist, "event",
1276 "event selector. use 'perf list' to list available events",
1277 parse_events_option),
1278 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1279 OPT_STRING('p', "pid", &target->pid, "pid",
1280 "profile events on existing process id"),
1281 OPT_STRING('t', "tid", &target->tid, "tid",
1282 "profile events on existing thread id"),
1283 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1284 "system-wide collection from all CPUs"),
1285 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1286 "list of cpus to monitor"),
1287 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1288 "file", "vmlinux pathname"),
1289 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1290 "don't load vmlinux even if found"),
1291 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1292 "hide kernel symbols"),
1293 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1294 "number of mmap data pages",
1295 perf_evlist__parse_mmap_pages),
1296 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1297 "collect data with this RT SCHED_FIFO priority"),
1298 OPT_INTEGER('d', "delay", &top.delay_secs,
1299 "number of seconds to delay between refreshes"),
1300 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1301 "dump the symbol table used for profiling"),
1302 OPT_INTEGER('f', "count-filter", &top.count_filter,
1303 "only display functions with more events than this"),
1304 OPT_BOOLEAN(0, "group", &opts->group,
1305 "put the counters into a counter group"),
1306 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1307 "child tasks do not inherit counters"),
1308 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1309 "symbol to annotate"),
1310 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1311 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1312 "profile at this frequency",
1313 record__parse_freq),
1314 OPT_INTEGER('E', "entries", &top.print_entries,
1315 "display this many functions"),
1316 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1317 "hide user symbols"),
1318 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1319 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1320 OPT_INCR('v', "verbose", &verbose,
1321 "be more verbose (show counter open errors, etc)"),
1322 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1323 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1324 " Please refer the man page for the complete list."),
1325 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1326 "output field(s): overhead, period, sample plus all of sort keys"),
1327 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1328 "Show a column with the number of samples"),
1329 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1330 NULL, "enables call-graph recording and display",
1331 &callchain_opt),
1332 OPT_CALLBACK(0, "call-graph", &callchain_param,
1333 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1334 top_callchain_help, &parse_callchain_opt),
1335 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1336 "Accumulate callchains of children and show total overhead as well"),
1337 OPT_INTEGER(0, "max-stack", &top.max_stack,
1338 "Set the maximum stack depth when parsing the callchain. "
1339 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1340 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1341 "ignore callees of these functions in call graphs",
1342 report_parse_ignore_callees_opt),
1343 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1344 "Show a column with the sum of periods"),
1345 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1346 "only consider symbols in these dsos"),
1347 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1348 "only consider symbols in these comms"),
1349 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1350 "only consider these symbols"),
1351 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1352 "Interleave source code with assembly code (default)"),
1353 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1354 "Display raw encoding of assembly instructions (default)"),
1355 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1356 "Enable kernel symbol demangling"),
1357 OPT_STRING(0, "objdump", &objdump_path, "path",
1358 "objdump binary to use for disassembly and annotations"),
1359 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1360 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1361 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1362 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1363 "Don't show entries under that percent", parse_percent_limit),
1364 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1365 "How to display percentage of filtered entries", parse_filter_percentage),
1366 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1367 "width[,width...]",
1368 "don't try to adjust column width, use these fixed values"),
1369 OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1370 "per thread proc mmap processing timeout in ms"),
1371 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1372 "branch any", "sample any taken branches",
1373 parse_branch_stack),
1374 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1375 "branch filter mask", "branch stack filter modes",
1376 parse_branch_stack),
1377 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1378 "Show raw trace event output (do not use print fmt or plugins)"),
1379 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1380 "Show entries in a hierarchy"),
1381 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1382 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1383 "number of thread to run event synthesize"),
1384 OPT_END()
1385 };
1386 const char * const top_usage[] = {
1387 "perf top [<options>]",
1388 NULL
1389 };
1390 int status = hists__init();
1391
1392 if (status < 0)
1393 return status;
1394
1395 top.evlist = perf_evlist__new();
1396 if (top.evlist == NULL)
1397 return -ENOMEM;
1398
1399 status = perf_config(perf_top_config, &top);
1400 if (status)
1401 return status;
1402
1403 argc = parse_options(argc, argv, options, top_usage, 0);
1404 if (argc)
1405 usage_with_options(top_usage, options);
1406
1407 if (!top.evlist->nr_entries &&
1408 perf_evlist__add_default(top.evlist) < 0) {
1409 pr_err("Not enough memory for event selector list\n");
1410 goto out_delete_evlist;
1411 }
1412
1413 if (symbol_conf.report_hierarchy) {
1414 /* disable incompatible options */
1415 symbol_conf.event_group = false;
1416 symbol_conf.cumulate_callchain = false;
1417
1418 if (field_order) {
1419 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1420 parse_options_usage(top_usage, options, "fields", 0);
1421 parse_options_usage(NULL, options, "hierarchy", 0);
1422 goto out_delete_evlist;
1423 }
1424 }
1425
1426 sort__mode = SORT_MODE__TOP;
1427 /* display thread wants entries to be collapsed in a different tree */
1428 perf_hpp_list.need_collapse = 1;
1429
1430 if (top.use_stdio)
1431 use_browser = 0;
1432 else if (top.use_tui)
1433 use_browser = 1;
1434
1435 setup_browser(false);
1436
1437 if (setup_sorting(top.evlist) < 0) {
1438 if (sort_order)
1439 parse_options_usage(top_usage, options, "s", 1);
1440 if (field_order)
1441 parse_options_usage(sort_order ? NULL : top_usage,
1442 options, "fields", 0);
1443 goto out_delete_evlist;
1444 }
1445
1446 status = target__validate(target);
1447 if (status) {
1448 target__strerror(target, status, errbuf, BUFSIZ);
1449 ui__warning("%s\n", errbuf);
1450 }
1451
1452 status = target__parse_uid(target);
1453 if (status) {
1454 int saved_errno = errno;
1455
1456 target__strerror(target, status, errbuf, BUFSIZ);
1457 ui__error("%s\n", errbuf);
1458
1459 status = -saved_errno;
1460 goto out_delete_evlist;
1461 }
1462
1463 if (target__none(target))
1464 target->system_wide = true;
1465
1466 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1467 ui__error("Couldn't create thread/CPU maps: %s\n",
1468 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1469 goto out_delete_evlist;
1470 }
1471
1472 symbol_conf.nr_events = top.evlist->nr_entries;
1473
1474 if (top.delay_secs < 1)
1475 top.delay_secs = 1;
1476
1477 if (record_opts__config(opts)) {
1478 status = -EINVAL;
1479 goto out_delete_evlist;
1480 }
1481
1482 top.sym_evsel = perf_evlist__first(top.evlist);
1483
1484 if (!callchain_param.enabled) {
1485 symbol_conf.cumulate_callchain = false;
1486 perf_hpp__cancel_cumulate();
1487 }
1488
1489 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1490 callchain_param.order = ORDER_CALLER;
1491
1492 status = symbol__annotation_init();
1493 if (status < 0)
1494 goto out_delete_evlist;
1495
1496 annotation_config__init();
1497
1498 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1499 if (symbol__init(NULL) < 0)
1500 return -1;
1501
1502 sort__setup_elide(stdout);
1503
1504 get_term_dimensions(&top.winsize);
1505 if (top.print_entries == 0) {
1506 perf_top__update_print_entries(&top);
1507 signal(SIGWINCH, winch_sig);
1508 }
1509
1510 status = __cmd_top(&top);
1511
1512out_delete_evlist:
1513 perf_evlist__delete(top.evlist);
1514
1515 return status;
1516}