Loading...
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/symbol.h"
37#include "util/synthetic-events.h"
38#include "util/top.h"
39#include "util/util.h"
40#include <linux/rbtree.h>
41#include <subcmd/parse-options.h>
42#include "util/parse-events.h"
43#include "util/callchain.h"
44#include "util/cpumap.h"
45#include "util/sort.h"
46#include "util/string2.h"
47#include "util/term.h"
48#include "util/intlist.h"
49#include "util/parse-branch-options.h"
50#include "arch/common.h"
51#include "ui/ui.h"
52
53#include "util/debug.h"
54#include "util/ordered-events.h"
55
56#include <assert.h>
57#include <elf.h>
58#include <fcntl.h>
59
60#include <stdio.h>
61#include <termios.h>
62#include <unistd.h>
63#include <inttypes.h>
64
65#include <errno.h>
66#include <time.h>
67#include <sched.h>
68#include <signal.h>
69
70#include <sys/syscall.h>
71#include <sys/ioctl.h>
72#include <poll.h>
73#include <sys/prctl.h>
74#include <sys/wait.h>
75#include <sys/uio.h>
76#include <sys/utsname.h>
77#include <sys/mman.h>
78
79#include <linux/stringify.h>
80#include <linux/time64.h>
81#include <linux/types.h>
82#include <linux/err.h>
83
84#include <linux/ctype.h>
85
86static volatile int done;
87static volatile int resize;
88
89#define HEADER_LINE_NR 5
90
91static void perf_top__update_print_entries(struct perf_top *top)
92{
93 top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
94}
95
96static void winch_sig(int sig __maybe_unused)
97{
98 resize = 1;
99}
100
101static void perf_top__resize(struct perf_top *top)
102{
103 get_term_dimensions(&top->winsize);
104 perf_top__update_print_entries(top);
105}
106
107static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
108{
109 struct evsel *evsel;
110 struct symbol *sym;
111 struct annotation *notes;
112 struct map *map;
113 int err = -1;
114
115 if (!he || !he->ms.sym)
116 return -1;
117
118 evsel = hists_to_evsel(he->hists);
119
120 sym = he->ms.sym;
121 map = he->ms.map;
122
123 /*
124 * We can't annotate with just /proc/kallsyms
125 */
126 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
127 !dso__is_kcore(map->dso)) {
128 pr_err("Can't annotate %s: No vmlinux file was found in the "
129 "path\n", sym->name);
130 sleep(1);
131 return -1;
132 }
133
134 notes = symbol__annotation(sym);
135 pthread_mutex_lock(¬es->lock);
136
137 if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
138 pthread_mutex_unlock(¬es->lock);
139 pr_err("Not enough memory for annotating '%s' symbol!\n",
140 sym->name);
141 sleep(1);
142 return err;
143 }
144
145 err = symbol__annotate(sym, map, evsel, 0, &top->annotation_opts, NULL);
146 if (err == 0) {
147 top->sym_filter_entry = he;
148 } else {
149 char msg[BUFSIZ];
150 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
151 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
152 }
153
154 pthread_mutex_unlock(¬es->lock);
155 return err;
156}
157
158static void __zero_source_counters(struct hist_entry *he)
159{
160 struct symbol *sym = he->ms.sym;
161 symbol__annotate_zero_histograms(sym);
162}
163
164static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
165{
166 struct utsname uts;
167 int err = uname(&uts);
168
169 ui__warning("Out of bounds address found:\n\n"
170 "Addr: %" PRIx64 "\n"
171 "DSO: %s %c\n"
172 "Map: %" PRIx64 "-%" PRIx64 "\n"
173 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
174 "Arch: %s\n"
175 "Kernel: %s\n"
176 "Tools: %s\n\n"
177 "Not all samples will be on the annotation output.\n\n"
178 "Please report to linux-kernel@vger.kernel.org\n",
179 ip, map->dso->long_name, dso__symtab_origin(map->dso),
180 map->start, map->end, sym->start, sym->end,
181 sym->binding == STB_GLOBAL ? 'g' :
182 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
183 err ? "[unknown]" : uts.machine,
184 err ? "[unknown]" : uts.release, perf_version_string);
185 if (use_browser <= 0)
186 sleep(5);
187
188 map->erange_warned = true;
189}
190
191static void perf_top__record_precise_ip(struct perf_top *top,
192 struct hist_entry *he,
193 struct perf_sample *sample,
194 struct evsel *evsel, u64 ip)
195{
196 struct annotation *notes;
197 struct symbol *sym = he->ms.sym;
198 int err = 0;
199
200 if (sym == NULL || (use_browser == 0 &&
201 (top->sym_filter_entry == NULL ||
202 top->sym_filter_entry->ms.sym != sym)))
203 return;
204
205 notes = symbol__annotation(sym);
206
207 if (pthread_mutex_trylock(¬es->lock))
208 return;
209
210 err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
211
212 pthread_mutex_unlock(¬es->lock);
213
214 if (unlikely(err)) {
215 /*
216 * This function is now called with he->hists->lock held.
217 * Release it before going to sleep.
218 */
219 pthread_mutex_unlock(&he->hists->lock);
220
221 if (err == -ERANGE && !he->ms.map->erange_warned)
222 ui__warn_map_erange(he->ms.map, sym, ip);
223 else if (err == -ENOMEM) {
224 pr_err("Not enough memory for annotating '%s' symbol!\n",
225 sym->name);
226 sleep(1);
227 }
228
229 pthread_mutex_lock(&he->hists->lock);
230 }
231}
232
233static void perf_top__show_details(struct perf_top *top)
234{
235 struct hist_entry *he = top->sym_filter_entry;
236 struct evsel *evsel;
237 struct annotation *notes;
238 struct symbol *symbol;
239 int more;
240
241 if (!he)
242 return;
243
244 evsel = hists_to_evsel(he->hists);
245
246 symbol = he->ms.sym;
247 notes = symbol__annotation(symbol);
248
249 pthread_mutex_lock(¬es->lock);
250
251 symbol__calc_percent(symbol, evsel);
252
253 if (notes->src == NULL)
254 goto out_unlock;
255
256 printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
257 printf(" Events Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
258
259 more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel, &top->annotation_opts);
260
261 if (top->evlist->enabled) {
262 if (top->zero)
263 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
264 else
265 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
266 }
267 if (more != 0)
268 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
269out_unlock:
270 pthread_mutex_unlock(¬es->lock);
271}
272
273static void perf_top__resort_hists(struct perf_top *t)
274{
275 struct evlist *evlist = t->evlist;
276 struct evsel *pos;
277
278 evlist__for_each_entry(evlist, pos) {
279 struct hists *hists = evsel__hists(pos);
280
281 /*
282 * unlink existing entries so that they can be linked
283 * in a correct order in hists__match() below.
284 */
285 hists__unlink(hists);
286
287 if (evlist->enabled) {
288 if (t->zero) {
289 hists__delete_entries(hists);
290 } else {
291 hists__decay_entries(hists, t->hide_user_symbols,
292 t->hide_kernel_symbols);
293 }
294 }
295
296 hists__collapse_resort(hists, NULL);
297
298 /* Non-group events are considered as leader */
299 if (symbol_conf.event_group &&
300 !perf_evsel__is_group_leader(pos)) {
301 struct hists *leader_hists = evsel__hists(pos->leader);
302
303 hists__match(leader_hists, hists);
304 hists__link(leader_hists, hists);
305 }
306 }
307
308 evlist__for_each_entry(evlist, pos) {
309 perf_evsel__output_resort(pos, NULL);
310 }
311}
312
313static void perf_top__print_sym_table(struct perf_top *top)
314{
315 char bf[160];
316 int printed = 0;
317 const int win_width = top->winsize.ws_col - 1;
318 struct evsel *evsel = top->sym_evsel;
319 struct hists *hists = evsel__hists(evsel);
320
321 puts(CONSOLE_CLEAR);
322
323 perf_top__header_snprintf(top, bf, sizeof(bf));
324 printf("%s\n", bf);
325
326 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
327
328 if (!top->record_opts.overwrite &&
329 (hists->stats.nr_lost_warned !=
330 hists->stats.nr_events[PERF_RECORD_LOST])) {
331 hists->stats.nr_lost_warned =
332 hists->stats.nr_events[PERF_RECORD_LOST];
333 color_fprintf(stdout, PERF_COLOR_RED,
334 "WARNING: LOST %d chunks, Check IO/CPU overload",
335 hists->stats.nr_lost_warned);
336 ++printed;
337 }
338
339 if (top->sym_filter_entry) {
340 perf_top__show_details(top);
341 return;
342 }
343
344 perf_top__resort_hists(top);
345
346 hists__output_recalc_col_len(hists, top->print_entries - printed);
347 putchar('\n');
348 hists__fprintf(hists, false, top->print_entries - printed, win_width,
349 top->min_percent, stdout, !symbol_conf.use_callchain);
350}
351
352static void prompt_integer(int *target, const char *msg)
353{
354 char *buf = malloc(0), *p;
355 size_t dummy = 0;
356 int tmp;
357
358 fprintf(stdout, "\n%s: ", msg);
359 if (getline(&buf, &dummy, stdin) < 0)
360 return;
361
362 p = strchr(buf, '\n');
363 if (p)
364 *p = 0;
365
366 p = buf;
367 while(*p) {
368 if (!isdigit(*p))
369 goto out_free;
370 p++;
371 }
372 tmp = strtoul(buf, NULL, 10);
373 *target = tmp;
374out_free:
375 free(buf);
376}
377
378static void prompt_percent(int *target, const char *msg)
379{
380 int tmp = 0;
381
382 prompt_integer(&tmp, msg);
383 if (tmp >= 0 && tmp <= 100)
384 *target = tmp;
385}
386
387static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
388{
389 char *buf = malloc(0), *p;
390 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
391 struct hists *hists = evsel__hists(top->sym_evsel);
392 struct rb_node *next;
393 size_t dummy = 0;
394
395 /* zero counters of active symbol */
396 if (syme) {
397 __zero_source_counters(syme);
398 top->sym_filter_entry = NULL;
399 }
400
401 fprintf(stdout, "\n%s: ", msg);
402 if (getline(&buf, &dummy, stdin) < 0)
403 goto out_free;
404
405 p = strchr(buf, '\n');
406 if (p)
407 *p = 0;
408
409 next = rb_first_cached(&hists->entries);
410 while (next) {
411 n = rb_entry(next, struct hist_entry, rb_node);
412 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
413 found = n;
414 break;
415 }
416 next = rb_next(&n->rb_node);
417 }
418
419 if (!found) {
420 fprintf(stderr, "Sorry, %s is not active.\n", buf);
421 sleep(1);
422 } else
423 perf_top__parse_source(top, found);
424
425out_free:
426 free(buf);
427}
428
429static void perf_top__print_mapped_keys(struct perf_top *top)
430{
431 char *name = NULL;
432
433 if (top->sym_filter_entry) {
434 struct symbol *sym = top->sym_filter_entry->ms.sym;
435 name = sym->name;
436 }
437
438 fprintf(stdout, "\nMapped keys:\n");
439 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
440 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
441
442 if (top->evlist->core.nr_entries > 1)
443 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
444
445 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
446
447 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
448 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
449 fprintf(stdout, "\t[S] stop annotation.\n");
450
451 fprintf(stdout,
452 "\t[K] hide kernel symbols. \t(%s)\n",
453 top->hide_kernel_symbols ? "yes" : "no");
454 fprintf(stdout,
455 "\t[U] hide user symbols. \t(%s)\n",
456 top->hide_user_symbols ? "yes" : "no");
457 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
458 fprintf(stdout, "\t[qQ] quit.\n");
459}
460
461static int perf_top__key_mapped(struct perf_top *top, int c)
462{
463 switch (c) {
464 case 'd':
465 case 'e':
466 case 'f':
467 case 'z':
468 case 'q':
469 case 'Q':
470 case 'K':
471 case 'U':
472 case 'F':
473 case 's':
474 case 'S':
475 return 1;
476 case 'E':
477 return top->evlist->core.nr_entries > 1 ? 1 : 0;
478 default:
479 break;
480 }
481
482 return 0;
483}
484
485static bool perf_top__handle_keypress(struct perf_top *top, int c)
486{
487 bool ret = true;
488
489 if (!perf_top__key_mapped(top, c)) {
490 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
491 struct termios save;
492
493 perf_top__print_mapped_keys(top);
494 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
495 fflush(stdout);
496
497 set_term_quiet_input(&save);
498
499 poll(&stdin_poll, 1, -1);
500 c = getc(stdin);
501
502 tcsetattr(0, TCSAFLUSH, &save);
503 if (!perf_top__key_mapped(top, c))
504 return ret;
505 }
506
507 switch (c) {
508 case 'd':
509 prompt_integer(&top->delay_secs, "Enter display delay");
510 if (top->delay_secs < 1)
511 top->delay_secs = 1;
512 break;
513 case 'e':
514 prompt_integer(&top->print_entries, "Enter display entries (lines)");
515 if (top->print_entries == 0) {
516 perf_top__resize(top);
517 signal(SIGWINCH, winch_sig);
518 } else {
519 signal(SIGWINCH, SIG_DFL);
520 }
521 break;
522 case 'E':
523 if (top->evlist->core.nr_entries > 1) {
524 /* Select 0 as the default event: */
525 int counter = 0;
526
527 fprintf(stderr, "\nAvailable events:");
528
529 evlist__for_each_entry(top->evlist, top->sym_evsel)
530 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
531
532 prompt_integer(&counter, "Enter details event counter");
533
534 if (counter >= top->evlist->core.nr_entries) {
535 top->sym_evsel = evlist__first(top->evlist);
536 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
537 sleep(1);
538 break;
539 }
540 evlist__for_each_entry(top->evlist, top->sym_evsel)
541 if (top->sym_evsel->idx == counter)
542 break;
543 } else
544 top->sym_evsel = evlist__first(top->evlist);
545 break;
546 case 'f':
547 prompt_integer(&top->count_filter, "Enter display event count filter");
548 break;
549 case 'F':
550 prompt_percent(&top->annotation_opts.min_pcnt,
551 "Enter details display event filter (percent)");
552 break;
553 case 'K':
554 top->hide_kernel_symbols = !top->hide_kernel_symbols;
555 break;
556 case 'q':
557 case 'Q':
558 printf("exiting.\n");
559 if (top->dump_symtab)
560 perf_session__fprintf_dsos(top->session, stderr);
561 ret = false;
562 break;
563 case 's':
564 perf_top__prompt_symbol(top, "Enter details symbol");
565 break;
566 case 'S':
567 if (!top->sym_filter_entry)
568 break;
569 else {
570 struct hist_entry *syme = top->sym_filter_entry;
571
572 top->sym_filter_entry = NULL;
573 __zero_source_counters(syme);
574 }
575 break;
576 case 'U':
577 top->hide_user_symbols = !top->hide_user_symbols;
578 break;
579 case 'z':
580 top->zero = !top->zero;
581 break;
582 default:
583 break;
584 }
585
586 return ret;
587}
588
589static void perf_top__sort_new_samples(void *arg)
590{
591 struct perf_top *t = arg;
592
593 if (t->evlist->selected != NULL)
594 t->sym_evsel = t->evlist->selected;
595
596 perf_top__resort_hists(t);
597
598 if (t->lost || t->drop)
599 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
600}
601
602static void stop_top(void)
603{
604 session_done = 1;
605 done = 1;
606}
607
608static void *display_thread_tui(void *arg)
609{
610 struct evsel *pos;
611 struct perf_top *top = arg;
612 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
613 struct hist_browser_timer hbt = {
614 .timer = perf_top__sort_new_samples,
615 .arg = top,
616 .refresh = top->delay_secs,
617 };
618
619 /* In order to read symbols from other namespaces perf to needs to call
620 * setns(2). This isn't permitted if the struct_fs has multiple users.
621 * unshare(2) the fs so that we may continue to setns into namespaces
622 * that we're observing.
623 */
624 unshare(CLONE_FS);
625
626 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
627
628 perf_top__sort_new_samples(top);
629
630 /*
631 * Initialize the uid_filter_str, in the future the TUI will allow
632 * Zooming in/out UIDs. For now just use whatever the user passed
633 * via --uid.
634 */
635 evlist__for_each_entry(top->evlist, pos) {
636 struct hists *hists = evsel__hists(pos);
637 hists->uid_filter_str = top->record_opts.target.uid_str;
638 }
639
640 perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
641 top->min_percent,
642 &top->session->header.env,
643 !top->record_opts.overwrite,
644 &top->annotation_opts);
645
646 stop_top();
647 return NULL;
648}
649
650static void display_sig(int sig __maybe_unused)
651{
652 stop_top();
653}
654
655static void display_setup_sig(void)
656{
657 signal(SIGSEGV, sighandler_dump_stack);
658 signal(SIGFPE, sighandler_dump_stack);
659 signal(SIGINT, display_sig);
660 signal(SIGQUIT, display_sig);
661 signal(SIGTERM, display_sig);
662}
663
664static void *display_thread(void *arg)
665{
666 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
667 struct termios save;
668 struct perf_top *top = arg;
669 int delay_msecs, c;
670
671 /* In order to read symbols from other namespaces perf to needs to call
672 * setns(2). This isn't permitted if the struct_fs has multiple users.
673 * unshare(2) the fs so that we may continue to setns into namespaces
674 * that we're observing.
675 */
676 unshare(CLONE_FS);
677
678 prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
679
680 display_setup_sig();
681 pthread__unblock_sigwinch();
682repeat:
683 delay_msecs = top->delay_secs * MSEC_PER_SEC;
684 set_term_quiet_input(&save);
685 /* trash return*/
686 getc(stdin);
687
688 while (!done) {
689 perf_top__print_sym_table(top);
690 /*
691 * Either timeout expired or we got an EINTR due to SIGWINCH,
692 * refresh screen in both cases.
693 */
694 switch (poll(&stdin_poll, 1, delay_msecs)) {
695 case 0:
696 continue;
697 case -1:
698 if (errno == EINTR)
699 continue;
700 __fallthrough;
701 default:
702 c = getc(stdin);
703 tcsetattr(0, TCSAFLUSH, &save);
704
705 if (perf_top__handle_keypress(top, c))
706 goto repeat;
707 stop_top();
708 }
709 }
710
711 tcsetattr(0, TCSAFLUSH, &save);
712 return NULL;
713}
714
715static int hist_iter__top_callback(struct hist_entry_iter *iter,
716 struct addr_location *al, bool single,
717 void *arg)
718{
719 struct perf_top *top = arg;
720 struct hist_entry *he = iter->he;
721 struct evsel *evsel = iter->evsel;
722
723 if (perf_hpp_list.sym && single)
724 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
725
726 hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
727 !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
728 return 0;
729}
730
731static void perf_event__process_sample(struct perf_tool *tool,
732 const union perf_event *event,
733 struct evsel *evsel,
734 struct perf_sample *sample,
735 struct machine *machine)
736{
737 struct perf_top *top = container_of(tool, struct perf_top, tool);
738 struct addr_location al;
739 int err;
740
741 if (!machine && perf_guest) {
742 static struct intlist *seen;
743
744 if (!seen)
745 seen = intlist__new(NULL);
746
747 if (!intlist__has_entry(seen, sample->pid)) {
748 pr_err("Can't find guest [%d]'s kernel information\n",
749 sample->pid);
750 intlist__add(seen, sample->pid);
751 }
752 return;
753 }
754
755 if (!machine) {
756 pr_err("%u unprocessable samples recorded.\r",
757 top->session->evlist->stats.nr_unprocessable_samples++);
758 return;
759 }
760
761 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
762 top->exact_samples++;
763
764 if (machine__resolve(machine, &al, sample) < 0)
765 return;
766
767 if (!machine->kptr_restrict_warned &&
768 symbol_conf.kptr_restrict &&
769 al.cpumode == PERF_RECORD_MISC_KERNEL) {
770 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
771 ui__warning(
772"Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
773"Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
774"Kernel%s samples will not be resolved.\n",
775 al.map && map__has_symbols(al.map) ?
776 " modules" : "");
777 if (use_browser <= 0)
778 sleep(5);
779 }
780 machine->kptr_restrict_warned = true;
781 }
782
783 if (al.sym == NULL && al.map != NULL) {
784 const char *msg = "Kernel samples will not be resolved.\n";
785 /*
786 * As we do lazy loading of symtabs we only will know if the
787 * specified vmlinux file is invalid when we actually have a
788 * hit in kernel space and then try to load it. So if we get
789 * here and there are _no_ symbols in the DSO backing the
790 * kernel map, bail out.
791 *
792 * We may never get here, for instance, if we use -K/
793 * --hide-kernel-symbols, even if the user specifies an
794 * invalid --vmlinux ;-)
795 */
796 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
797 __map__is_kernel(al.map) && map__has_symbols(al.map)) {
798 if (symbol_conf.vmlinux_name) {
799 char serr[256];
800 dso__strerror_load(al.map->dso, serr, sizeof(serr));
801 ui__warning("The %s file can't be used: %s\n%s",
802 symbol_conf.vmlinux_name, serr, msg);
803 } else {
804 ui__warning("A vmlinux file was not found.\n%s",
805 msg);
806 }
807
808 if (use_browser <= 0)
809 sleep(5);
810 top->vmlinux_warned = true;
811 }
812 }
813
814 if (al.sym == NULL || !al.sym->idle) {
815 struct hists *hists = evsel__hists(evsel);
816 struct hist_entry_iter iter = {
817 .evsel = evsel,
818 .sample = sample,
819 .add_entry_cb = hist_iter__top_callback,
820 };
821
822 if (symbol_conf.cumulate_callchain)
823 iter.ops = &hist_iter_cumulative;
824 else
825 iter.ops = &hist_iter_normal;
826
827 pthread_mutex_lock(&hists->lock);
828
829 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
830 if (err < 0)
831 pr_err("Problem incrementing symbol period, skipping event\n");
832
833 pthread_mutex_unlock(&hists->lock);
834 }
835
836 addr_location__put(&al);
837}
838
839static void
840perf_top__process_lost(struct perf_top *top, union perf_event *event,
841 struct evsel *evsel)
842{
843 struct hists *hists = evsel__hists(evsel);
844
845 top->lost += event->lost.lost;
846 top->lost_total += event->lost.lost;
847 hists->stats.total_lost += event->lost.lost;
848}
849
850static void
851perf_top__process_lost_samples(struct perf_top *top,
852 union perf_event *event,
853 struct evsel *evsel)
854{
855 struct hists *hists = evsel__hists(evsel);
856
857 top->lost += event->lost_samples.lost;
858 top->lost_total += event->lost_samples.lost;
859 hists->stats.total_lost_samples += event->lost_samples.lost;
860}
861
862static u64 last_timestamp;
863
864static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
865{
866 struct record_opts *opts = &top->record_opts;
867 struct evlist *evlist = top->evlist;
868 struct mmap *md;
869 union perf_event *event;
870
871 md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
872 if (perf_mmap__read_init(md) < 0)
873 return;
874
875 while ((event = perf_mmap__read_event(md)) != NULL) {
876 int ret;
877
878 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
879 if (ret && ret != -1)
880 break;
881
882 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
883 if (ret)
884 break;
885
886 perf_mmap__consume(md);
887
888 if (top->qe.rotate) {
889 pthread_mutex_lock(&top->qe.mutex);
890 top->qe.rotate = false;
891 pthread_cond_signal(&top->qe.cond);
892 pthread_mutex_unlock(&top->qe.mutex);
893 }
894 }
895
896 perf_mmap__read_done(md);
897}
898
899static void perf_top__mmap_read(struct perf_top *top)
900{
901 bool overwrite = top->record_opts.overwrite;
902 struct evlist *evlist = top->evlist;
903 int i;
904
905 if (overwrite)
906 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
907
908 for (i = 0; i < top->evlist->core.nr_mmaps; i++)
909 perf_top__mmap_read_idx(top, i);
910
911 if (overwrite) {
912 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
913 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
914 }
915}
916
917/*
918 * Check per-event overwrite term.
919 * perf top should support consistent term for all events.
920 * - All events don't have per-event term
921 * E.g. "cpu/cpu-cycles/,cpu/instructions/"
922 * Nothing change, return 0.
923 * - All events have same per-event term
924 * E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
925 * Using the per-event setting to replace the opts->overwrite if
926 * they are different, then return 0.
927 * - Events have different per-event term
928 * E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
929 * Return -1
930 * - Some of the event set per-event term, but some not.
931 * E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
932 * Return -1
933 */
934static int perf_top__overwrite_check(struct perf_top *top)
935{
936 struct record_opts *opts = &top->record_opts;
937 struct evlist *evlist = top->evlist;
938 struct perf_evsel_config_term *term;
939 struct list_head *config_terms;
940 struct evsel *evsel;
941 int set, overwrite = -1;
942
943 evlist__for_each_entry(evlist, evsel) {
944 set = -1;
945 config_terms = &evsel->config_terms;
946 list_for_each_entry(term, config_terms, list) {
947 if (term->type == PERF_EVSEL__CONFIG_TERM_OVERWRITE)
948 set = term->val.overwrite ? 1 : 0;
949 }
950
951 /* no term for current and previous event (likely) */
952 if ((overwrite < 0) && (set < 0))
953 continue;
954
955 /* has term for both current and previous event, compare */
956 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
957 return -1;
958
959 /* no term for current event but has term for previous one */
960 if ((overwrite >= 0) && (set < 0))
961 return -1;
962
963 /* has term for current event */
964 if ((overwrite < 0) && (set >= 0)) {
965 /* if it's first event, set overwrite */
966 if (evsel == evlist__first(evlist))
967 overwrite = set;
968 else
969 return -1;
970 }
971 }
972
973 if ((overwrite >= 0) && (opts->overwrite != overwrite))
974 opts->overwrite = overwrite;
975
976 return 0;
977}
978
979static int perf_top_overwrite_fallback(struct perf_top *top,
980 struct evsel *evsel)
981{
982 struct record_opts *opts = &top->record_opts;
983 struct evlist *evlist = top->evlist;
984 struct evsel *counter;
985
986 if (!opts->overwrite)
987 return 0;
988
989 /* only fall back when first event fails */
990 if (evsel != evlist__first(evlist))
991 return 0;
992
993 evlist__for_each_entry(evlist, counter)
994 counter->core.attr.write_backward = false;
995 opts->overwrite = false;
996 pr_debug2("fall back to non-overwrite mode\n");
997 return 1;
998}
999
1000static int perf_top__start_counters(struct perf_top *top)
1001{
1002 char msg[BUFSIZ];
1003 struct evsel *counter;
1004 struct evlist *evlist = top->evlist;
1005 struct record_opts *opts = &top->record_opts;
1006
1007 if (perf_top__overwrite_check(top)) {
1008 ui__error("perf top only support consistent per-event "
1009 "overwrite setting for all events\n");
1010 goto out_err;
1011 }
1012
1013 perf_evlist__config(evlist, opts, &callchain_param);
1014
1015 evlist__for_each_entry(evlist, counter) {
1016try_again:
1017 if (evsel__open(counter, top->evlist->core.cpus,
1018 top->evlist->core.threads) < 0) {
1019
1020 /*
1021 * Specially handle overwrite fall back.
1022 * Because perf top is the only tool which has
1023 * overwrite mode by default, support
1024 * both overwrite and non-overwrite mode, and
1025 * require consistent mode for all events.
1026 *
1027 * May move it to generic code with more tools
1028 * have similar attribute.
1029 */
1030 if (perf_missing_features.write_backward &&
1031 perf_top_overwrite_fallback(top, counter))
1032 goto try_again;
1033
1034 if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
1035 if (verbose > 0)
1036 ui__warning("%s\n", msg);
1037 goto try_again;
1038 }
1039
1040 perf_evsel__open_strerror(counter, &opts->target,
1041 errno, msg, sizeof(msg));
1042 ui__error("%s\n", msg);
1043 goto out_err;
1044 }
1045 }
1046
1047 if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1048 ui__error("Failed to mmap with %d (%s)\n",
1049 errno, str_error_r(errno, msg, sizeof(msg)));
1050 goto out_err;
1051 }
1052
1053 return 0;
1054
1055out_err:
1056 return -1;
1057}
1058
1059static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1060{
1061 if (callchain->mode != CHAIN_NONE) {
1062 if (callchain_register_param(callchain) < 0) {
1063 ui__error("Can't register callchain params.\n");
1064 return -EINVAL;
1065 }
1066 }
1067
1068 return 0;
1069}
1070
1071static struct ordered_events *rotate_queues(struct perf_top *top)
1072{
1073 struct ordered_events *in = top->qe.in;
1074
1075 if (top->qe.in == &top->qe.data[1])
1076 top->qe.in = &top->qe.data[0];
1077 else
1078 top->qe.in = &top->qe.data[1];
1079
1080 return in;
1081}
1082
1083static void *process_thread(void *arg)
1084{
1085 struct perf_top *top = arg;
1086
1087 while (!done) {
1088 struct ordered_events *out, *in = top->qe.in;
1089
1090 if (!in->nr_events) {
1091 usleep(100);
1092 continue;
1093 }
1094
1095 out = rotate_queues(top);
1096
1097 pthread_mutex_lock(&top->qe.mutex);
1098 top->qe.rotate = true;
1099 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1100 pthread_mutex_unlock(&top->qe.mutex);
1101
1102 if (ordered_events__flush(out, OE_FLUSH__TOP))
1103 pr_err("failed to process events\n");
1104 }
1105
1106 return NULL;
1107}
1108
1109/*
1110 * Allow only 'top->delay_secs' seconds behind samples.
1111 */
1112static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1113{
1114 union perf_event *event = qevent->event;
1115 u64 delay_timestamp;
1116
1117 if (event->header.type != PERF_RECORD_SAMPLE)
1118 return false;
1119
1120 delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1121 return delay_timestamp < last_timestamp;
1122}
1123
1124static int deliver_event(struct ordered_events *qe,
1125 struct ordered_event *qevent)
1126{
1127 struct perf_top *top = qe->data;
1128 struct evlist *evlist = top->evlist;
1129 struct perf_session *session = top->session;
1130 union perf_event *event = qevent->event;
1131 struct perf_sample sample;
1132 struct evsel *evsel;
1133 struct machine *machine;
1134 int ret = -1;
1135
1136 if (should_drop(qevent, top)) {
1137 top->drop++;
1138 top->drop_total++;
1139 return 0;
1140 }
1141
1142 ret = perf_evlist__parse_sample(evlist, event, &sample);
1143 if (ret) {
1144 pr_err("Can't parse sample, err = %d\n", ret);
1145 goto next_event;
1146 }
1147
1148 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1149 assert(evsel != NULL);
1150
1151 if (event->header.type == PERF_RECORD_SAMPLE) {
1152 if (evswitch__discard(&top->evswitch, evsel))
1153 return 0;
1154 ++top->samples;
1155 }
1156
1157 switch (sample.cpumode) {
1158 case PERF_RECORD_MISC_USER:
1159 ++top->us_samples;
1160 if (top->hide_user_symbols)
1161 goto next_event;
1162 machine = &session->machines.host;
1163 break;
1164 case PERF_RECORD_MISC_KERNEL:
1165 ++top->kernel_samples;
1166 if (top->hide_kernel_symbols)
1167 goto next_event;
1168 machine = &session->machines.host;
1169 break;
1170 case PERF_RECORD_MISC_GUEST_KERNEL:
1171 ++top->guest_kernel_samples;
1172 machine = perf_session__find_machine(session,
1173 sample.pid);
1174 break;
1175 case PERF_RECORD_MISC_GUEST_USER:
1176 ++top->guest_us_samples;
1177 /*
1178 * TODO: we don't process guest user from host side
1179 * except simple counting.
1180 */
1181 goto next_event;
1182 default:
1183 if (event->header.type == PERF_RECORD_SAMPLE)
1184 goto next_event;
1185 machine = &session->machines.host;
1186 break;
1187 }
1188
1189 if (event->header.type == PERF_RECORD_SAMPLE) {
1190 perf_event__process_sample(&top->tool, event, evsel,
1191 &sample, machine);
1192 } else if (event->header.type == PERF_RECORD_LOST) {
1193 perf_top__process_lost(top, event, evsel);
1194 } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1195 perf_top__process_lost_samples(top, event, evsel);
1196 } else if (event->header.type < PERF_RECORD_MAX) {
1197 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1198 machine__process_event(machine, event, &sample);
1199 } else
1200 ++session->evlist->stats.nr_unknown_events;
1201
1202 ret = 0;
1203next_event:
1204 return ret;
1205}
1206
1207static void init_process_thread(struct perf_top *top)
1208{
1209 ordered_events__init(&top->qe.data[0], deliver_event, top);
1210 ordered_events__init(&top->qe.data[1], deliver_event, top);
1211 ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1212 ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1213 top->qe.in = &top->qe.data[0];
1214 pthread_mutex_init(&top->qe.mutex, NULL);
1215 pthread_cond_init(&top->qe.cond, NULL);
1216}
1217
1218static int __cmd_top(struct perf_top *top)
1219{
1220 struct record_opts *opts = &top->record_opts;
1221 pthread_t thread, thread_process;
1222 int ret;
1223
1224 if (!top->annotation_opts.objdump_path) {
1225 ret = perf_env__lookup_objdump(&top->session->header.env,
1226 &top->annotation_opts.objdump_path);
1227 if (ret)
1228 return ret;
1229 }
1230
1231 ret = callchain_param__setup_sample_type(&callchain_param);
1232 if (ret)
1233 return ret;
1234
1235 if (perf_session__register_idle_thread(top->session) < 0)
1236 return ret;
1237
1238 if (top->nr_threads_synthesize > 1)
1239 perf_set_multithreaded();
1240
1241 init_process_thread(top);
1242
1243 if (opts->record_namespaces)
1244 top->tool.namespace_events = true;
1245
1246 ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1247 &top->session->machines.host,
1248 &top->record_opts);
1249 if (ret < 0)
1250 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1251
1252 machine__synthesize_threads(&top->session->machines.host, &opts->target,
1253 top->evlist->core.threads, false,
1254 top->nr_threads_synthesize);
1255
1256 if (top->nr_threads_synthesize > 1)
1257 perf_set_singlethreaded();
1258
1259 if (perf_hpp_list.socket) {
1260 ret = perf_env__read_cpu_topology_map(&perf_env);
1261 if (ret < 0) {
1262 char errbuf[BUFSIZ];
1263 const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1264
1265 ui__error("Could not read the CPU topology map: %s\n", err);
1266 return ret;
1267 }
1268 }
1269
1270 ret = perf_top__start_counters(top);
1271 if (ret)
1272 return ret;
1273
1274 top->session->evlist = top->evlist;
1275 perf_session__set_id_hdr_size(top->session);
1276
1277 /*
1278 * When perf is starting the traced process, all the events (apart from
1279 * group members) have enable_on_exec=1 set, so don't spoil it by
1280 * prematurely enabling them.
1281 *
1282 * XXX 'top' still doesn't start workloads like record, trace, but should,
1283 * so leave the check here.
1284 */
1285 if (!target__none(&opts->target))
1286 evlist__enable(top->evlist);
1287
1288 ret = -1;
1289 if (pthread_create(&thread_process, NULL, process_thread, top)) {
1290 ui__error("Could not create process thread.\n");
1291 return ret;
1292 }
1293
1294 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1295 display_thread), top)) {
1296 ui__error("Could not create display thread.\n");
1297 goto out_join_thread;
1298 }
1299
1300 if (top->realtime_prio) {
1301 struct sched_param param;
1302
1303 param.sched_priority = top->realtime_prio;
1304 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1305 ui__error("Could not set realtime priority.\n");
1306 goto out_join;
1307 }
1308 }
1309
1310 /* Wait for a minimal set of events before starting the snapshot */
1311 evlist__poll(top->evlist, 100);
1312
1313 perf_top__mmap_read(top);
1314
1315 while (!done) {
1316 u64 hits = top->samples;
1317
1318 perf_top__mmap_read(top);
1319
1320 if (opts->overwrite || (hits == top->samples))
1321 ret = evlist__poll(top->evlist, 100);
1322
1323 if (resize) {
1324 perf_top__resize(top);
1325 resize = 0;
1326 }
1327 }
1328
1329 ret = 0;
1330out_join:
1331 pthread_join(thread, NULL);
1332out_join_thread:
1333 pthread_cond_signal(&top->qe.cond);
1334 pthread_join(thread_process, NULL);
1335 return ret;
1336}
1337
1338static int
1339callchain_opt(const struct option *opt, const char *arg, int unset)
1340{
1341 symbol_conf.use_callchain = true;
1342 return record_callchain_opt(opt, arg, unset);
1343}
1344
1345static int
1346parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1347{
1348 struct callchain_param *callchain = opt->value;
1349
1350 callchain->enabled = !unset;
1351 callchain->record_mode = CALLCHAIN_FP;
1352
1353 /*
1354 * --no-call-graph
1355 */
1356 if (unset) {
1357 symbol_conf.use_callchain = false;
1358 callchain->record_mode = CALLCHAIN_NONE;
1359 return 0;
1360 }
1361
1362 return parse_callchain_top_opt(arg);
1363}
1364
1365static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1366{
1367 if (!strcmp(var, "top.call-graph")) {
1368 var = "call-graph.record-mode";
1369 return perf_default_config(var, value, cb);
1370 }
1371 if (!strcmp(var, "top.children")) {
1372 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1373 return 0;
1374 }
1375
1376 return 0;
1377}
1378
1379static int
1380parse_percent_limit(const struct option *opt, const char *arg,
1381 int unset __maybe_unused)
1382{
1383 struct perf_top *top = opt->value;
1384
1385 top->min_percent = strtof(arg, NULL);
1386 return 0;
1387}
1388
1389const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1390 "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1391
1392int cmd_top(int argc, const char **argv)
1393{
1394 char errbuf[BUFSIZ];
1395 struct perf_top top = {
1396 .count_filter = 5,
1397 .delay_secs = 2,
1398 .record_opts = {
1399 .mmap_pages = UINT_MAX,
1400 .user_freq = UINT_MAX,
1401 .user_interval = ULLONG_MAX,
1402 .freq = 4000, /* 4 KHz */
1403 .target = {
1404 .uses_mmap = true,
1405 },
1406 /*
1407 * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1408 * when we pause, fix that and reenable. Probably using a
1409 * separate evlist with a dummy event, i.e. a non-overwrite
1410 * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1411 * stays in overwrite mode. -acme
1412 * */
1413 .overwrite = 0,
1414 .sample_time = true,
1415 .sample_time_set = true,
1416 },
1417 .max_stack = sysctl__max_stack(),
1418 .annotation_opts = annotation__default_options,
1419 .nr_threads_synthesize = UINT_MAX,
1420 };
1421 struct record_opts *opts = &top.record_opts;
1422 struct target *target = &opts->target;
1423 const struct option options[] = {
1424 OPT_CALLBACK('e', "event", &top.evlist, "event",
1425 "event selector. use 'perf list' to list available events",
1426 parse_events_option),
1427 OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1428 OPT_STRING('p', "pid", &target->pid, "pid",
1429 "profile events on existing process id"),
1430 OPT_STRING('t', "tid", &target->tid, "tid",
1431 "profile events on existing thread id"),
1432 OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1433 "system-wide collection from all CPUs"),
1434 OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1435 "list of cpus to monitor"),
1436 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1437 "file", "vmlinux pathname"),
1438 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1439 "don't load vmlinux even if found"),
1440 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1441 "file", "kallsyms pathname"),
1442 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1443 "hide kernel symbols"),
1444 OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1445 "number of mmap data pages",
1446 perf_evlist__parse_mmap_pages),
1447 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1448 "collect data with this RT SCHED_FIFO priority"),
1449 OPT_INTEGER('d', "delay", &top.delay_secs,
1450 "number of seconds to delay between refreshes"),
1451 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1452 "dump the symbol table used for profiling"),
1453 OPT_INTEGER('f', "count-filter", &top.count_filter,
1454 "only display functions with more events than this"),
1455 OPT_BOOLEAN(0, "group", &opts->group,
1456 "put the counters into a counter group"),
1457 OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1458 "child tasks do not inherit counters"),
1459 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1460 "symbol to annotate"),
1461 OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1462 OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1463 "profile at this frequency",
1464 record__parse_freq),
1465 OPT_INTEGER('E', "entries", &top.print_entries,
1466 "display this many functions"),
1467 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1468 "hide user symbols"),
1469 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1470 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1471 OPT_INCR('v', "verbose", &verbose,
1472 "be more verbose (show counter open errors, etc)"),
1473 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1474 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1475 " Please refer the man page for the complete list."),
1476 OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1477 "output field(s): overhead, period, sample plus all of sort keys"),
1478 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1479 "Show a column with the number of samples"),
1480 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1481 NULL, "enables call-graph recording and display",
1482 &callchain_opt),
1483 OPT_CALLBACK(0, "call-graph", &callchain_param,
1484 "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1485 top_callchain_help, &parse_callchain_opt),
1486 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1487 "Accumulate callchains of children and show total overhead as well"),
1488 OPT_INTEGER(0, "max-stack", &top.max_stack,
1489 "Set the maximum stack depth when parsing the callchain. "
1490 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1491 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1492 "ignore callees of these functions in call graphs",
1493 report_parse_ignore_callees_opt),
1494 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1495 "Show a column with the sum of periods"),
1496 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1497 "only consider symbols in these dsos"),
1498 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1499 "only consider symbols in these comms"),
1500 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1501 "only consider these symbols"),
1502 OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1503 "Interleave source code with assembly code (default)"),
1504 OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1505 "Display raw encoding of assembly instructions (default)"),
1506 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1507 "Enable kernel symbol demangling"),
1508 OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1509 OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1510 "objdump binary to use for disassembly and annotations"),
1511 OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1512 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1513 OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1514 OPT_CALLBACK(0, "percent-limit", &top, "percent",
1515 "Don't show entries under that percent", parse_percent_limit),
1516 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1517 "How to display percentage of filtered entries", parse_filter_percentage),
1518 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1519 "width[,width...]",
1520 "don't try to adjust column width, use these fixed values"),
1521 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1522 "per thread proc mmap processing timeout in ms"),
1523 OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1524 "branch any", "sample any taken branches",
1525 parse_branch_stack),
1526 OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1527 "branch filter mask", "branch stack filter modes",
1528 parse_branch_stack),
1529 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1530 "Show raw trace event output (do not use print fmt or plugins)"),
1531 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1532 "Show entries in a hierarchy"),
1533 OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1534 "Use a backward ring buffer, default: no"),
1535 OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1536 OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1537 "number of thread to run event synthesize"),
1538 OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1539 "Record namespaces events"),
1540 OPTS_EVSWITCH(&top.evswitch),
1541 OPT_END()
1542 };
1543 struct evlist *sb_evlist = NULL;
1544 const char * const top_usage[] = {
1545 "perf top [<options>]",
1546 NULL
1547 };
1548 int status = hists__init();
1549
1550 if (status < 0)
1551 return status;
1552
1553 top.annotation_opts.min_pcnt = 5;
1554 top.annotation_opts.context = 4;
1555
1556 top.evlist = evlist__new();
1557 if (top.evlist == NULL)
1558 return -ENOMEM;
1559
1560 status = perf_config(perf_top_config, &top);
1561 if (status)
1562 return status;
1563
1564 argc = parse_options(argc, argv, options, top_usage, 0);
1565 if (argc)
1566 usage_with_options(top_usage, options);
1567
1568 if (!top.evlist->core.nr_entries &&
1569 perf_evlist__add_default(top.evlist) < 0) {
1570 pr_err("Not enough memory for event selector list\n");
1571 goto out_delete_evlist;
1572 }
1573
1574 status = evswitch__init(&top.evswitch, top.evlist, stderr);
1575 if (status)
1576 goto out_delete_evlist;
1577
1578 if (symbol_conf.report_hierarchy) {
1579 /* disable incompatible options */
1580 symbol_conf.event_group = false;
1581 symbol_conf.cumulate_callchain = false;
1582
1583 if (field_order) {
1584 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1585 parse_options_usage(top_usage, options, "fields", 0);
1586 parse_options_usage(NULL, options, "hierarchy", 0);
1587 goto out_delete_evlist;
1588 }
1589 }
1590
1591 if (opts->branch_stack && callchain_param.enabled)
1592 symbol_conf.show_branchflag_count = true;
1593
1594 sort__mode = SORT_MODE__TOP;
1595 /* display thread wants entries to be collapsed in a different tree */
1596 perf_hpp_list.need_collapse = 1;
1597
1598 if (top.use_stdio)
1599 use_browser = 0;
1600 else if (top.use_tui)
1601 use_browser = 1;
1602
1603 setup_browser(false);
1604
1605 if (setup_sorting(top.evlist) < 0) {
1606 if (sort_order)
1607 parse_options_usage(top_usage, options, "s", 1);
1608 if (field_order)
1609 parse_options_usage(sort_order ? NULL : top_usage,
1610 options, "fields", 0);
1611 goto out_delete_evlist;
1612 }
1613
1614 status = target__validate(target);
1615 if (status) {
1616 target__strerror(target, status, errbuf, BUFSIZ);
1617 ui__warning("%s\n", errbuf);
1618 }
1619
1620 status = target__parse_uid(target);
1621 if (status) {
1622 int saved_errno = errno;
1623
1624 target__strerror(target, status, errbuf, BUFSIZ);
1625 ui__error("%s\n", errbuf);
1626
1627 status = -saved_errno;
1628 goto out_delete_evlist;
1629 }
1630
1631 if (target__none(target))
1632 target->system_wide = true;
1633
1634 if (perf_evlist__create_maps(top.evlist, target) < 0) {
1635 ui__error("Couldn't create thread/CPU maps: %s\n",
1636 errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1637 goto out_delete_evlist;
1638 }
1639
1640 if (top.delay_secs < 1)
1641 top.delay_secs = 1;
1642
1643 if (record_opts__config(opts)) {
1644 status = -EINVAL;
1645 goto out_delete_evlist;
1646 }
1647
1648 top.sym_evsel = evlist__first(top.evlist);
1649
1650 if (!callchain_param.enabled) {
1651 symbol_conf.cumulate_callchain = false;
1652 perf_hpp__cancel_cumulate();
1653 }
1654
1655 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1656 callchain_param.order = ORDER_CALLER;
1657
1658 status = symbol__annotation_init();
1659 if (status < 0)
1660 goto out_delete_evlist;
1661
1662 annotation_config__init();
1663
1664 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1665 status = symbol__init(NULL);
1666 if (status < 0)
1667 goto out_delete_evlist;
1668
1669 sort__setup_elide(stdout);
1670
1671 get_term_dimensions(&top.winsize);
1672 if (top.print_entries == 0) {
1673 perf_top__update_print_entries(&top);
1674 signal(SIGWINCH, winch_sig);
1675 }
1676
1677 top.session = perf_session__new(NULL, false, NULL);
1678 if (IS_ERR(top.session)) {
1679 status = PTR_ERR(top.session);
1680 goto out_delete_evlist;
1681 }
1682
1683 if (!top.record_opts.no_bpf_event)
1684 bpf_event__add_sb_event(&sb_evlist, &perf_env);
1685
1686 if (perf_evlist__start_sb_thread(sb_evlist, target)) {
1687 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1688 opts->no_bpf_event = true;
1689 }
1690
1691 status = __cmd_top(&top);
1692
1693 if (!opts->no_bpf_event)
1694 perf_evlist__stop_sb_thread(sb_evlist);
1695
1696out_delete_evlist:
1697 evlist__delete(top.evlist);
1698 perf_session__delete(top.session);
1699
1700 return status;
1701}
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}