?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf tools: Introduce struct hist_browser_timer
[linux-2.6.git] / tools / perf / builtin-top.c
blob102b43c9905d952c254f4604afa1d38d6088e446
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/cache.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/machine.h"
30 #include "util/session.h"
31 #include "util/symbol.h"
32 #include "util/thread.h"
33 #include "util/thread_map.h"
34 #include "util/top.h"
35 #include "util/util.h"
36 #include <linux/rbtree.h>
37 #include "util/parse-options.h"
38 #include "util/parse-events.h"
39 #include "util/cpumap.h"
40 #include "util/xyarray.h"
41 #include "util/sort.h"
42 #include "util/intlist.h"
43
44 #include "util/debug.h"
45
46 #include <assert.h>
47 #include <elf.h>
48 #include <fcntl.h>
49
50 #include <stdio.h>
51 #include <termios.h>
52 #include <unistd.h>
53 #include <inttypes.h>
54
55 #include <errno.h>
56 #include <time.h>
57 #include <sched.h>
58
59 #include <sys/syscall.h>
60 #include <sys/ioctl.h>
61 #include <sys/poll.h>
62 #include <sys/prctl.h>
63 #include <sys/wait.h>
64 #include <sys/uio.h>
65 #include <sys/utsname.h>
66 #include <sys/mman.h>
67
68 #include <linux/unistd.h>
69 #include <linux/types.h>
70
71 void get_term_dimensions(struct winsize *ws)
72 {
73 char *s = getenv("LINES");
74
75 if (s != NULL) {
76 ws->ws_row = atoi(s);
77 s = getenv("COLUMNS");
78 if (s != NULL) {
79 ws->ws_col = atoi(s);
80 if (ws->ws_row && ws->ws_col)
81 return;
82 }
83 }
84 #ifdef TIOCGWINSZ
85 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
86 ws->ws_row && ws->ws_col)
87 return;
88 #endif
89 ws->ws_row = 25;
90 ws->ws_col = 80;
91 }
92
93 static void perf_top__update_print_entries(struct perf_top *top)
94 {
95 if (top->print_entries > 9)
96 top->print_entries -= 9;
97 }
98
99 static void perf_top__sig_winch(int sig __maybe_unused,
100 siginfo_t *info __maybe_unused, void *arg)
101 {
102 struct perf_top *top = arg;
103
104 get_term_dimensions(&top->winsize);
105 if (!top->print_entries
106 || (top->print_entries+4) > top->winsize.ws_row) {
107 top->print_entries = top->winsize.ws_row;
108 } else {
109 top->print_entries += 4;
110 top->winsize.ws_row = top->print_entries;
111 }
112 perf_top__update_print_entries(top);
113 }
114
115 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
116 {
117 struct symbol *sym;
118 struct annotation *notes;
119 struct map *map;
120 int err = -1;
121
122 if (!he || !he->ms.sym)
123 return -1;
124
125 sym = he->ms.sym;
126 map = he->ms.map;
127
128 /*
129 * We can't annotate with just /proc/kallsyms
130 */
131 if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
132 pr_err("Can't annotate %s: No vmlinux file was found in the "
133 "path\n", sym->name);
134 sleep(1);
135 return -1;
136 }
137
138 notes = symbol__annotation(sym);
139 if (notes->src != NULL) {
140 pthread_mutex_lock(&notes->lock);
141 goto out_assign;
142 }
143
144 pthread_mutex_lock(&notes->lock);
145
146 if (symbol__alloc_hist(sym) < 0) {
147 pthread_mutex_unlock(&notes->lock);
148 pr_err("Not enough memory for annotating '%s' symbol!\n",
149 sym->name);
150 sleep(1);
151 return err;
152 }
153
154 err = symbol__annotate(sym, map, 0);
155 if (err == 0) {
156 out_assign:
157 top->sym_filter_entry = he;
158 }
159
160 pthread_mutex_unlock(&notes->lock);
161 return err;
162 }
163
164 static void __zero_source_counters(struct hist_entry *he)
165 {
166 struct symbol *sym = he->ms.sym;
167 symbol__annotate_zero_histograms(sym);
168 }
169
170 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
171 {
172 struct utsname uts;
173 int err = uname(&uts);
174
175 ui__warning("Out of bounds address found:\n\n"
176 "Addr: %" PRIx64 "\n"
177 "DSO: %s %c\n"
178 "Map: %" PRIx64 "-%" PRIx64 "\n"
179 "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
180 "Arch: %s\n"
181 "Kernel: %s\n"
182 "Tools: %s\n\n"
183 "Not all samples will be on the annotation output.\n\n"
184 "Please report to linux-kernel@vger.kernel.org\n",
185 ip, map->dso->long_name, dso__symtab_origin(map->dso),
186 map->start, map->end, sym->start, sym->end,
187 sym->binding == STB_GLOBAL ? 'g' :
188 sym->binding == STB_LOCAL ? 'l' : 'w', sym->name,
189 err ? "[unknown]" : uts.machine,
190 err ? "[unknown]" : uts.release, perf_version_string);
191 if (use_browser <= 0)
192 sleep(5);
193
194 map->erange_warned = true;
195 }
196
197 static void perf_top__record_precise_ip(struct perf_top *top,
198 struct hist_entry *he,
199 int counter, u64 ip)
200 {
201 struct annotation *notes;
202 struct symbol *sym;
203 int err;
204
205 if (he == NULL || he->ms.sym == NULL ||
206 ((top->sym_filter_entry == NULL ||
207 top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
208 return;
209
210 sym = he->ms.sym;
211 notes = symbol__annotation(sym);
212
213 if (pthread_mutex_trylock(&notes->lock))
214 return;
215
216 if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
217 pthread_mutex_unlock(&notes->lock);
218 pr_err("Not enough memory for annotating '%s' symbol!\n",
219 sym->name);
220 sleep(1);
221 return;
222 }
223
224 ip = he->ms.map->map_ip(he->ms.map, ip);
225 err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
226
227 pthread_mutex_unlock(&notes->lock);
228
229 if (err == -ERANGE && !he->ms.map->erange_warned)
230 ui__warn_map_erange(he->ms.map, sym, ip);
231 }
232
233 static void perf_top__show_details(struct perf_top *top)
234 {
235 struct hist_entry *he = top->sym_filter_entry;
236 struct annotation *notes;
237 struct symbol *symbol;
238 int more;
239
240 if (!he)
241 return;
242
243 symbol = he->ms.sym;
244 notes = symbol__annotation(symbol);
245
246 pthread_mutex_lock(&notes->lock);
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->idx,
255 0, top->sym_pcnt_filter, top->print_entries, 4);
256 if (top->zero)
257 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
258 else
259 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
260 if (more != 0)
261 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
262 out_unlock:
263 pthread_mutex_unlock(&notes->lock);
264 }
265
266 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
267
268 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
269 struct addr_location *al,
270 struct perf_sample *sample)
271 {
272 struct hist_entry *he;
273
274 he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
275 if (he == NULL)
276 return NULL;
277
278 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
279 return he;
280 }
281
282 static void perf_top__print_sym_table(struct perf_top *top)
283 {
284 char bf[160];
285 int printed = 0;
286 const int win_width = top->winsize.ws_col - 1;
287
288 puts(CONSOLE_CLEAR);
289
290 perf_top__header_snprintf(top, bf, sizeof(bf));
291 printf("%s\n", bf);
292
293 perf_top__reset_sample_counters(top);
294
295 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
296
297 if (top->sym_evsel->hists.stats.nr_lost_warned !=
298 top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
299 top->sym_evsel->hists.stats.nr_lost_warned =
300 top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
301 color_fprintf(stdout, PERF_COLOR_RED,
302 "WARNING: LOST %d chunks, Check IO/CPU overload",
303 top->sym_evsel->hists.stats.nr_lost_warned);
304 ++printed;
305 }
306
307 if (top->sym_filter_entry) {
308 perf_top__show_details(top);
309 return;
310 }
311
312 hists__collapse_resort_threaded(&top->sym_evsel->hists);
313 hists__output_resort_threaded(&top->sym_evsel->hists);
314 hists__decay_entries_threaded(&top->sym_evsel->hists,
315 top->hide_user_symbols,
316 top->hide_kernel_symbols);
317 hists__output_recalc_col_len(&top->sym_evsel->hists,
318 top->winsize.ws_row - 3);
319 putchar('\n');
320 hists__fprintf(&top->sym_evsel->hists, false,
321 top->winsize.ws_row - 4 - printed, win_width, stdout);
322 }
323
324 static void prompt_integer(int *target, const char *msg)
325 {
326 char *buf = malloc(0), *p;
327 size_t dummy = 0;
328 int tmp;
329
330 fprintf(stdout, "\n%s: ", msg);
331 if (getline(&buf, &dummy, stdin) < 0)
332 return;
333
334 p = strchr(buf, '\n');
335 if (p)
336 *p = 0;
337
338 p = buf;
339 while(*p) {
340 if (!isdigit(*p))
341 goto out_free;
342 p++;
343 }
344 tmp = strtoul(buf, NULL, 10);
345 *target = tmp;
346 out_free:
347 free(buf);
348 }
349
350 static void prompt_percent(int *target, const char *msg)
351 {
352 int tmp = 0;
353
354 prompt_integer(&tmp, msg);
355 if (tmp >= 0 && tmp <= 100)
356 *target = tmp;
357 }
358
359 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
360 {
361 char *buf = malloc(0), *p;
362 struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
363 struct rb_node *next;
364 size_t dummy = 0;
365
366 /* zero counters of active symbol */
367 if (syme) {
368 __zero_source_counters(syme);
369 top->sym_filter_entry = NULL;
370 }
371
372 fprintf(stdout, "\n%s: ", msg);
373 if (getline(&buf, &dummy, stdin) < 0)
374 goto out_free;
375
376 p = strchr(buf, '\n');
377 if (p)
378 *p = 0;
379
380 next = rb_first(&top->sym_evsel->hists.entries);
381 while (next) {
382 n = rb_entry(next, struct hist_entry, rb_node);
383 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
384 found = n;
385 break;
386 }
387 next = rb_next(&n->rb_node);
388 }
389
390 if (!found) {
391 fprintf(stderr, "Sorry, %s is not active.\n", buf);
392 sleep(1);
393 } else
394 perf_top__parse_source(top, found);
395
396 out_free:
397 free(buf);
398 }
399
400 static void perf_top__print_mapped_keys(struct perf_top *top)
401 {
402 char *name = NULL;
403
404 if (top->sym_filter_entry) {
405 struct symbol *sym = top->sym_filter_entry->ms.sym;
406 name = sym->name;
407 }
408
409 fprintf(stdout, "\nMapped keys:\n");
410 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", top->delay_secs);
411 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", top->print_entries);
412
413 if (top->evlist->nr_entries > 1)
414 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", perf_evsel__name(top->sym_evsel));
415
416 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", top->count_filter);
417
418 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
419 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
420 fprintf(stdout, "\t[S] stop annotation.\n");
421
422 fprintf(stdout,
423 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
424 top->hide_kernel_symbols ? "yes" : "no");
425 fprintf(stdout,
426 "\t[U] hide user symbols. \t(%s)\n",
427 top->hide_user_symbols ? "yes" : "no");
428 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", top->zero ? 1 : 0);
429 fprintf(stdout, "\t[qQ] quit.\n");
430 }
431
432 static int perf_top__key_mapped(struct perf_top *top, int c)
433 {
434 switch (c) {
435 case 'd':
436 case 'e':
437 case 'f':
438 case 'z':
439 case 'q':
440 case 'Q':
441 case 'K':
442 case 'U':
443 case 'F':
444 case 's':
445 case 'S':
446 return 1;
447 case 'E':
448 return top->evlist->nr_entries > 1 ? 1 : 0;
449 default:
450 break;
451 }
452
453 return 0;
454 }
455
456 static void perf_top__handle_keypress(struct perf_top *top, int c)
457 {
458 if (!perf_top__key_mapped(top, c)) {
459 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
460 struct termios tc, save;
461
462 perf_top__print_mapped_keys(top);
463 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
464 fflush(stdout);
465
466 tcgetattr(0, &save);
467 tc = save;
468 tc.c_lflag &= ~(ICANON | ECHO);
469 tc.c_cc[VMIN] = 0;
470 tc.c_cc[VTIME] = 0;
471 tcsetattr(0, TCSANOW, &tc);
472
473 poll(&stdin_poll, 1, -1);
474 c = getc(stdin);
475
476 tcsetattr(0, TCSAFLUSH, &save);
477 if (!perf_top__key_mapped(top, c))
478 return;
479 }
480
481 switch (c) {
482 case 'd':
483 prompt_integer(&top->delay_secs, "Enter display delay");
484 if (top->delay_secs < 1)
485 top->delay_secs = 1;
486 break;
487 case 'e':
488 prompt_integer(&top->print_entries, "Enter display entries (lines)");
489 if (top->print_entries == 0) {
490 struct sigaction act = {
491 .sa_sigaction = perf_top__sig_winch,
492 .sa_flags = SA_SIGINFO,
493 };
494 perf_top__sig_winch(SIGWINCH, NULL, top);
495 sigaction(SIGWINCH, &act, NULL);
496 } else {
497 perf_top__sig_winch(SIGWINCH, NULL, top);
498 signal(SIGWINCH, SIG_DFL);
499 }
500 break;
501 case 'E':
502 if (top->evlist->nr_entries > 1) {
503 /* Select 0 as the default event: */
504 int counter = 0;
505
506 fprintf(stderr, "\nAvailable events:");
507
508 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
509 fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
510
511 prompt_integer(&counter, "Enter details event counter");
512
513 if (counter >= top->evlist->nr_entries) {
514 top->sym_evsel = perf_evlist__first(top->evlist);
515 fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
516 sleep(1);
517 break;
518 }
519 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
520 if (top->sym_evsel->idx == counter)
521 break;
522 } else
523 top->sym_evsel = perf_evlist__first(top->evlist);
524 break;
525 case 'f':
526 prompt_integer(&top->count_filter, "Enter display event count filter");
527 break;
528 case 'F':
529 prompt_percent(&top->sym_pcnt_filter,
530 "Enter details display event filter (percent)");
531 break;
532 case 'K':
533 top->hide_kernel_symbols = !top->hide_kernel_symbols;
534 break;
535 case 'q':
536 case 'Q':
537 printf("exiting.\n");
538 if (top->dump_symtab)
539 perf_session__fprintf_dsos(top->session, stderr);
540 exit(0);
541 case 's':
542 perf_top__prompt_symbol(top, "Enter details symbol");
543 break;
544 case 'S':
545 if (!top->sym_filter_entry)
546 break;
547 else {
548 struct hist_entry *syme = top->sym_filter_entry;
549
550 top->sym_filter_entry = NULL;
551 __zero_source_counters(syme);
552 }
553 break;
554 case 'U':
555 top->hide_user_symbols = !top->hide_user_symbols;
556 break;
557 case 'z':
558 top->zero = !top->zero;
559 break;
560 default:
561 break;
562 }
563 }
564
565 static void perf_top__sort_new_samples(void *arg)
566 {
567 struct perf_top *t = arg;
568 perf_top__reset_sample_counters(t);
569
570 if (t->evlist->selected != NULL)
571 t->sym_evsel = t->evlist->selected;
572
573 hists__collapse_resort_threaded(&t->sym_evsel->hists);
574 hists__output_resort_threaded(&t->sym_evsel->hists);
575 hists__decay_entries_threaded(&t->sym_evsel->hists,
576 t->hide_user_symbols,
577 t->hide_kernel_symbols);
578 }
579
580 static void *display_thread_tui(void *arg)
581 {
582 struct perf_evsel *pos;
583 struct perf_top *top = arg;
584 const char *help = "For a higher level overview, try: perf top --sort comm,dso";
585 struct hist_browser_timer hbt = {
586 .timer = perf_top__sort_new_samples,
587 .arg = top,
588 .refresh = top->delay_secs,
589 };
590
591 perf_top__sort_new_samples(top);
592
593 /*
594 * Initialize the uid_filter_str, in the future the TUI will allow
595 * Zooming in/out UIDs. For now juse use whatever the user passed
596 * via --uid.
597 */
598 list_for_each_entry(pos, &top->evlist->entries, node)
599 pos->hists.uid_filter_str = top->target.uid_str;
600
601 perf_evlist__tui_browse_hists(top->evlist, help, &hbt);
602
603 exit_browser(0);
604 exit(0);
605 return NULL;
606 }
607
608 static void *display_thread(void *arg)
609 {
610 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
611 struct termios tc, save;
612 struct perf_top *top = arg;
613 int delay_msecs, c;
614
615 tcgetattr(0, &save);
616 tc = save;
617 tc.c_lflag &= ~(ICANON | ECHO);
618 tc.c_cc[VMIN] = 0;
619 tc.c_cc[VTIME] = 0;
620
621 pthread__unblock_sigwinch();
622 repeat:
623 delay_msecs = top->delay_secs * 1000;
624 tcsetattr(0, TCSANOW, &tc);
625 /* trash return*/
626 getc(stdin);
627
628 while (1) {
629 perf_top__print_sym_table(top);
630 /*
631 * Either timeout expired or we got an EINTR due to SIGWINCH,
632 * refresh screen in both cases.
633 */
634 switch (poll(&stdin_poll, 1, delay_msecs)) {
635 case 0:
636 continue;
637 case -1:
638 if (errno == EINTR)
639 continue;
640 /* Fall trhu */
641 default:
642 goto process_hotkey;
643 }
644 }
645 process_hotkey:
646 c = getc(stdin);
647 tcsetattr(0, TCSAFLUSH, &save);
648
649 perf_top__handle_keypress(top, c);
650 goto repeat;
651
652 return NULL;
653 }
654
655 /* Tag samples to be skipped. */
656 static const char *skip_symbols[] = {
657 "intel_idle",
658 "default_idle",
659 "native_safe_halt",
660 "cpu_idle",
661 "enter_idle",
662 "exit_idle",
663 "mwait_idle",
664 "mwait_idle_with_hints",
665 "poll_idle",
666 "ppc64_runlatch_off",
667 "pseries_dedicated_idle_sleep",
668 NULL
669 };
670
671 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
672 {
673 const char *name = sym->name;
674 int i;
675
676 /*
677 * ppc64 uses function descriptors and appends a '.' to the
678 * start of every instruction address. Remove it.
679 */
680 if (name[0] == '.')
681 name++;
682
683 if (!strcmp(name, "_text") ||
684 !strcmp(name, "_etext") ||
685 !strcmp(name, "_sinittext") ||
686 !strncmp("init_module", name, 11) ||
687 !strncmp("cleanup_module", name, 14) ||
688 strstr(name, "_text_start") ||
689 strstr(name, "_text_end"))
690 return 1;
691
692 for (i = 0; skip_symbols[i]; i++) {
693 if (!strcmp(skip_symbols[i], name)) {
694 sym->ignore = true;
695 break;
696 }
697 }
698
699 return 0;
700 }
701
702 static void perf_event__process_sample(struct perf_tool *tool,
703 const union perf_event *event,
704 struct perf_evsel *evsel,
705 struct perf_sample *sample,
706 struct machine *machine)
707 {
708 struct perf_top *top = container_of(tool, struct perf_top, tool);
709 struct symbol *parent = NULL;
710 u64 ip = event->ip.ip;
711 struct addr_location al;
712 int err;
713
714 if (!machine && perf_guest) {
715 static struct intlist *seen;
716
717 if (!seen)
718 seen = intlist__new();
719
720 if (!intlist__has_entry(seen, event->ip.pid)) {
721 pr_err("Can't find guest [%d]'s kernel information\n",
722 event->ip.pid);
723 intlist__add(seen, event->ip.pid);
724 }
725 return;
726 }
727
728 if (!machine) {
729 pr_err("%u unprocessable samples recorded.",
730 top->session->hists.stats.nr_unprocessable_samples++);
731 return;
732 }
733
734 if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
735 top->exact_samples++;
736
737 if (perf_event__preprocess_sample(event, machine, &al, sample,
738 symbol_filter) < 0 ||
739 al.filtered)
740 return;
741
742 if (!top->kptr_restrict_warned &&
743 symbol_conf.kptr_restrict &&
744 al.cpumode == PERF_RECORD_MISC_KERNEL) {
745 ui__warning(
746 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
747 "Check /proc/sys/kernel/kptr_restrict.\n\n"
748 "Kernel%s samples will not be resolved.\n",
749 !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
750 " modules" : "");
751 if (use_browser <= 0)
752 sleep(5);
753 top->kptr_restrict_warned = true;
754 }
755
756 if (al.sym == NULL) {
757 const char *msg = "Kernel samples will not be resolved.\n";
758 /*
759 * As we do lazy loading of symtabs we only will know if the
760 * specified vmlinux file is invalid when we actually have a
761 * hit in kernel space and then try to load it. So if we get
762 * here and there are _no_ symbols in the DSO backing the
763 * kernel map, bail out.
764 *
765 * We may never get here, for instance, if we use -K/
766 * --hide-kernel-symbols, even if the user specifies an
767 * invalid --vmlinux ;-)
768 */
769 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
770 al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
771 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
772 if (symbol_conf.vmlinux_name) {
773 ui__warning("The %s file can't be used.\n%s",
774 symbol_conf.vmlinux_name, msg);
775 } else {
776 ui__warning("A vmlinux file was not found.\n%s",
777 msg);
778 }
779
780 if (use_browser <= 0)
781 sleep(5);
782 top->vmlinux_warned = true;
783 }
784 }
785
786 if (al.sym == NULL || !al.sym->ignore) {
787 struct hist_entry *he;
788
789 if ((sort__has_parent || symbol_conf.use_callchain) &&
790 sample->callchain) {
791 err = machine__resolve_callchain(machine, evsel,
792 al.thread, sample,
793 &parent);
794
795 if (err)
796 return;
797 }
798
799 he = perf_evsel__add_hist_entry(evsel, &al, sample);
800 if (he == NULL) {
801 pr_err("Problem incrementing symbol period, skipping event\n");
802 return;
803 }
804
805 if (symbol_conf.use_callchain) {
806 err = callchain_append(he->callchain, &callchain_cursor,
807 sample->period);
808 if (err)
809 return;
810 }
811
812 if (top->sort_has_symbols)
813 perf_top__record_precise_ip(top, he, evsel->idx, ip);
814 }
815
816 return;
817 }
818
819 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
820 {
821 struct perf_sample sample;
822 struct perf_evsel *evsel;
823 struct perf_session *session = top->session;
824 union perf_event *event;
825 struct machine *machine;
826 u8 origin;
827 int ret;
828
829 while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
830 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
831 if (ret) {
832 pr_err("Can't parse sample, err = %d\n", ret);
833 continue;
834 }
835
836 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
837 assert(evsel != NULL);
838
839 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
840
841 if (event->header.type == PERF_RECORD_SAMPLE)
842 ++top->samples;
843
844 switch (origin) {
845 case PERF_RECORD_MISC_USER:
846 ++top->us_samples;
847 if (top->hide_user_symbols)
848 continue;
849 machine = perf_session__find_host_machine(session);
850 break;
851 case PERF_RECORD_MISC_KERNEL:
852 ++top->kernel_samples;
853 if (top->hide_kernel_symbols)
854 continue;
855 machine = perf_session__find_host_machine(session);
856 break;
857 case PERF_RECORD_MISC_GUEST_KERNEL:
858 ++top->guest_kernel_samples;
859 machine = perf_session__find_machine(session, event->ip.pid);
860 break;
861 case PERF_RECORD_MISC_GUEST_USER:
862 ++top->guest_us_samples;
863 /*
864 * TODO: we don't process guest user from host side
865 * except simple counting.
866 */
867 /* Fall thru */
868 default:
869 continue;
870 }
871
872
873 if (event->header.type == PERF_RECORD_SAMPLE) {
874 perf_event__process_sample(&top->tool, event, evsel,
875 &sample, machine);
876 } else if (event->header.type < PERF_RECORD_MAX) {
877 hists__inc_nr_events(&evsel->hists, event->header.type);
878 machine__process_event(machine, event);
879 } else
880 ++session->hists.stats.nr_unknown_events;
881 }
882 }
883
884 static void perf_top__mmap_read(struct perf_top *top)
885 {
886 int i;
887
888 for (i = 0; i < top->evlist->nr_mmaps; i++)
889 perf_top__mmap_read_idx(top, i);
890 }
891
892 static void perf_top__start_counters(struct perf_top *top)
893 {
894 struct perf_evsel *counter;
895 struct perf_evlist *evlist = top->evlist;
896
897 if (top->group)
898 perf_evlist__set_leader(evlist);
899
900 list_for_each_entry(counter, &evlist->entries, node) {
901 struct perf_event_attr *attr = &counter->attr;
902
903 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
904
905 if (top->freq) {
906 attr->sample_type |= PERF_SAMPLE_PERIOD;
907 attr->freq = 1;
908 attr->sample_freq = top->freq;
909 }
910
911 if (evlist->nr_entries > 1) {
912 attr->sample_type |= PERF_SAMPLE_ID;
913 attr->read_format |= PERF_FORMAT_ID;
914 }
915
916 if (perf_target__has_cpu(&top->target))
917 attr->sample_type |= PERF_SAMPLE_CPU;
918
919 if (symbol_conf.use_callchain)
920 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
921
922 attr->mmap = 1;
923 attr->comm = 1;
924 attr->inherit = top->inherit;
925 fallback_missing_features:
926 if (top->exclude_guest_missing)
927 attr->exclude_guest = attr->exclude_host = 0;
928 retry_sample_id:
929 attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
930 try_again:
931 if (perf_evsel__open(counter, top->evlist->cpus,
932 top->evlist->threads) < 0) {
933 int err = errno;
934
935 if (err == EPERM || err == EACCES) {
936 ui__error_paranoid();
937 goto out_err;
938 } else if (err == EINVAL) {
939 if (!top->exclude_guest_missing &&
940 (attr->exclude_guest || attr->exclude_host)) {
941 pr_debug("Old kernel, cannot exclude "
942 "guest or host samples.\n");
943 top->exclude_guest_missing = true;
944 goto fallback_missing_features;
945 } else if (!top->sample_id_all_missing) {
946 /*
947 * Old kernel, no attr->sample_id_type_all field
948 */
949 top->sample_id_all_missing = true;
950 goto retry_sample_id;
951 }
952 }
953 /*
954 * If it's cycles then fall back to hrtimer
955 * based cpu-clock-tick sw counter, which
956 * is always available even if no PMU support:
957 */
958 if ((err == ENOENT || err == ENXIO) &&
959 (attr->type == PERF_TYPE_HARDWARE) &&
960 (attr->config == PERF_COUNT_HW_CPU_CYCLES)) {
961
962 if (verbose)
963 ui__warning("Cycles event not supported,\n"
964 "trying to fall back to cpu-clock-ticks\n");
965
966 attr->type = PERF_TYPE_SOFTWARE;
967 attr->config = PERF_COUNT_SW_CPU_CLOCK;
968 if (counter->name) {
969 free(counter->name);
970 counter->name = NULL;
971 }
972 goto try_again;
973 }
974
975 if (err == ENOENT) {
976 ui__error("The %s event is not supported.\n",
977 perf_evsel__name(counter));
978 goto out_err;
979 } else if (err == EMFILE) {
980 ui__error("Too many events are opened.\n"
981 "Try again after reducing the number of events\n");
982 goto out_err;
983 } else if ((err == EOPNOTSUPP) && (attr->precise_ip)) {
984 ui__error("\'precise\' request may not be supported. "
985 "Try removing 'p' modifier\n");
986 goto out_err;
987 }
988
989 ui__error("The sys_perf_event_open() syscall "
990 "returned with %d (%s). /bin/dmesg "
991 "may provide additional information.\n"
992 "No CONFIG_PERF_EVENTS=y kernel support "
993 "configured?\n", err, strerror(err));
994 goto out_err;
995 }
996 }
997
998 if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
999 ui__error("Failed to mmap with %d (%s)\n",
1000 errno, strerror(errno));
1001 goto out_err;
1002 }
1003
1004 return;
1005
1006 out_err:
1007 exit_browser(0);
1008 exit(0);
1009 }
1010
1011 static int perf_top__setup_sample_type(struct perf_top *top)
1012 {
1013 if (!top->sort_has_symbols) {
1014 if (symbol_conf.use_callchain) {
1015 ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1016 return -EINVAL;
1017 }
1018 } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
1019 if (callchain_register_param(&callchain_param) < 0) {
1020 ui__error("Can't register callchain params.\n");
1021 return -EINVAL;
1022 }
1023 }
1024
1025 return 0;
1026 }
1027
1028 static int __cmd_top(struct perf_top *top)
1029 {
1030 pthread_t thread;
1031 int ret;
1032 /*
1033 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1034 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1035 */
1036 top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1037 if (top->session == NULL)
1038 return -ENOMEM;
1039
1040 ret = perf_top__setup_sample_type(top);
1041 if (ret)
1042 goto out_delete;
1043
1044 if (perf_target__has_task(&top->target))
1045 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
1046 perf_event__process,
1047 &top->session->host_machine);
1048 else
1049 perf_event__synthesize_threads(&top->tool, perf_event__process,
1050 &top->session->host_machine);
1051 perf_top__start_counters(top);
1052 top->session->evlist = top->evlist;
1053 perf_session__set_id_hdr_size(top->session);
1054
1055 /* Wait for a minimal set of events before starting the snapshot */
1056 poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1057
1058 perf_top__mmap_read(top);
1059
1060 if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1061 display_thread), top)) {
1062 ui__error("Could not create display thread.\n");
1063 exit(-1);
1064 }
1065
1066 if (top->realtime_prio) {
1067 struct sched_param param;
1068
1069 param.sched_priority = top->realtime_prio;
1070 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1071 ui__error("Could not set realtime priority.\n");
1072 exit(-1);
1073 }
1074 }
1075
1076 while (1) {
1077 u64 hits = top->samples;
1078
1079 perf_top__mmap_read(top);
1080
1081 if (hits == top->samples)
1082 ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1083 }
1084
1085 out_delete:
1086 perf_session__delete(top->session);
1087 top->session = NULL;
1088
1089 return 0;
1090 }
1091
1092 static int
1093 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1094 {
1095 struct perf_top *top = (struct perf_top *)opt->value;
1096 char *tok, *tok2;
1097 char *endptr;
1098
1099 /*
1100 * --no-call-graph
1101 */
1102 if (unset) {
1103 top->dont_use_callchains = true;
1104 return 0;
1105 }
1106
1107 symbol_conf.use_callchain = true;
1108
1109 if (!arg)
1110 return 0;
1111
1112 tok = strtok((char *)arg, ",");
1113 if (!tok)
1114 return -1;
1115
1116 /* get the output mode */
1117 if (!strncmp(tok, "graph", strlen(arg)))
1118 callchain_param.mode = CHAIN_GRAPH_ABS;
1119
1120 else if (!strncmp(tok, "flat", strlen(arg)))
1121 callchain_param.mode = CHAIN_FLAT;
1122
1123 else if (!strncmp(tok, "fractal", strlen(arg)))
1124 callchain_param.mode = CHAIN_GRAPH_REL;
1125
1126 else if (!strncmp(tok, "none", strlen(arg))) {
1127 callchain_param.mode = CHAIN_NONE;
1128 symbol_conf.use_callchain = false;
1129
1130 return 0;
1131 } else
1132 return -1;
1133
1134 /* get the min percentage */
1135 tok = strtok(NULL, ",");
1136 if (!tok)
1137 goto setup;
1138
1139 callchain_param.min_percent = strtod(tok, &endptr);
1140 if (tok == endptr)
1141 return -1;
1142
1143 /* get the print limit */
1144 tok2 = strtok(NULL, ",");
1145 if (!tok2)
1146 goto setup;
1147
1148 if (tok2[0] != 'c') {
1149 callchain_param.print_limit = strtod(tok2, &endptr);
1150 tok2 = strtok(NULL, ",");
1151 if (!tok2)
1152 goto setup;
1153 }
1154
1155 /* get the call chain order */
1156 if (!strcmp(tok2, "caller"))
1157 callchain_param.order = ORDER_CALLER;
1158 else if (!strcmp(tok2, "callee"))
1159 callchain_param.order = ORDER_CALLEE;
1160 else
1161 return -1;
1162 setup:
1163 if (callchain_register_param(&callchain_param) < 0) {
1164 fprintf(stderr, "Can't register callchain params\n");
1165 return -1;
1166 }
1167 return 0;
1168 }
1169
1170 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1171 {
1172 struct perf_evsel *pos;
1173 int status;
1174 char errbuf[BUFSIZ];
1175 struct perf_top top = {
1176 .count_filter = 5,
1177 .delay_secs = 2,
1178 .freq = 4000, /* 4 KHz */
1179 .mmap_pages = 128,
1180 .sym_pcnt_filter = 5,
1181 .target = {
1182 .uses_mmap = true,
1183 },
1184 };
1185 char callchain_default_opt[] = "fractal,0.5,callee";
1186 const struct option options[] = {
1187 OPT_CALLBACK('e', "event", &top.evlist, "event",
1188 "event selector. use 'perf list' to list available events",
1189 parse_events_option),
1190 OPT_INTEGER('c', "count", &top.default_interval,
1191 "event period to sample"),
1192 OPT_STRING('p', "pid", &top.target.pid, "pid",
1193 "profile events on existing process id"),
1194 OPT_STRING('t', "tid", &top.target.tid, "tid",
1195 "profile events on existing thread id"),
1196 OPT_BOOLEAN('a', "all-cpus", &top.target.system_wide,
1197 "system-wide collection from all CPUs"),
1198 OPT_STRING('C', "cpu", &top.target.cpu_list, "cpu",
1199 "list of cpus to monitor"),
1200 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1201 "file", "vmlinux pathname"),
1202 OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1203 "hide kernel symbols"),
1204 OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1205 OPT_INTEGER('r', "realtime", &top.realtime_prio,
1206 "collect data with this RT SCHED_FIFO priority"),
1207 OPT_INTEGER('d', "delay", &top.delay_secs,
1208 "number of seconds to delay between refreshes"),
1209 OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1210 "dump the symbol table used for profiling"),
1211 OPT_INTEGER('f', "count-filter", &top.count_filter,
1212 "only display functions with more events than this"),
1213 OPT_BOOLEAN('g', "group", &top.group,
1214 "put the counters into a counter group"),
1215 OPT_BOOLEAN('i', "inherit", &top.inherit,
1216 "child tasks inherit counters"),
1217 OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1218 "symbol to annotate"),
1219 OPT_BOOLEAN('z', "zero", &top.zero,
1220 "zero history across updates"),
1221 OPT_INTEGER('F', "freq", &top.freq,
1222 "profile at this frequency"),
1223 OPT_INTEGER('E', "entries", &top.print_entries,
1224 "display this many functions"),
1225 OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1226 "hide user symbols"),
1227 OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1228 OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1229 OPT_INCR('v', "verbose", &verbose,
1230 "be more verbose (show counter open errors, etc)"),
1231 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1232 "sort by key(s): pid, comm, dso, symbol, parent"),
1233 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1234 "Show a column with the number of samples"),
1235 OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1236 "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1237 "Default: fractal,0.5,callee", &parse_callchain_opt,
1238 callchain_default_opt),
1239 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1240 "Show a column with the sum of periods"),
1241 OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1242 "only consider symbols in these dsos"),
1243 OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1244 "only consider symbols in these comms"),
1245 OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1246 "only consider these symbols"),
1247 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1248 "Interleave source code with assembly code (default)"),
1249 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1250 "Display raw encoding of assembly instructions (default)"),
1251 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1252 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1253 OPT_STRING('u', "uid", &top.target.uid_str, "user", "user to profile"),
1254 OPT_END()
1255 };
1256 const char * const top_usage[] = {
1257 "perf top [<options>]",
1258 NULL
1259 };
1260
1261 top.evlist = perf_evlist__new(NULL, NULL);
1262 if (top.evlist == NULL)
1263 return -ENOMEM;
1264
1265 symbol_conf.exclude_other = false;
1266
1267 argc = parse_options(argc, argv, options, top_usage, 0);
1268 if (argc)
1269 usage_with_options(top_usage, options);
1270
1271 if (sort_order == default_sort_order)
1272 sort_order = "dso,symbol";
1273
1274 setup_sorting(top_usage, options);
1275
1276 if (top.use_stdio)
1277 use_browser = 0;
1278 else if (top.use_tui)
1279 use_browser = 1;
1280
1281 setup_browser(false);
1282
1283 status = perf_target__validate(&top.target);
1284 if (status) {
1285 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1286 ui__warning("%s", errbuf);
1287 }
1288
1289 status = perf_target__parse_uid(&top.target);
1290 if (status) {
1291 int saved_errno = errno;
1292
1293 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1294 ui__error("%s", errbuf);
1295
1296 status = -saved_errno;
1297 goto out_delete_evlist;
1298 }
1299
1300 if (perf_target__none(&top.target))
1301 top.target.system_wide = true;
1302
1303 if (perf_evlist__create_maps(top.evlist, &top.target) < 0)
1304 usage_with_options(top_usage, options);
1305
1306 if (!top.evlist->nr_entries &&
1307 perf_evlist__add_default(top.evlist) < 0) {
1308 ui__error("Not enough memory for event selector list\n");
1309 return -ENOMEM;
1310 }
1311
1312 symbol_conf.nr_events = top.evlist->nr_entries;
1313
1314 if (top.delay_secs < 1)
1315 top.delay_secs = 1;
1316
1317 /*
1318 * User specified count overrides default frequency.
1319 */
1320 if (top.default_interval)
1321 top.freq = 0;
1322 else if (top.freq) {
1323 top.default_interval = top.freq;
1324 } else {
1325 ui__error("frequency and count are zero, aborting\n");
1326 exit(EXIT_FAILURE);
1327 }
1328
1329 list_for_each_entry(pos, &top.evlist->entries, node) {
1330 /*
1331 * Fill in the ones not specifically initialized via -c:
1332 */
1333 if (!pos->attr.sample_period)
1334 pos->attr.sample_period = top.default_interval;
1335 }
1336
1337 top.sym_evsel = perf_evlist__first(top.evlist);
1338
1339 symbol_conf.priv_size = sizeof(struct annotation);
1340
1341 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1342 if (symbol__init() < 0)
1343 return -1;
1344
1345 sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1346 sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1347 sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1348
1349 /*
1350 * Avoid annotation data structures overhead when symbols aren't on the
1351 * sort list.
1352 */
1353 top.sort_has_symbols = sort_sym.list.next != NULL;
1354
1355 get_term_dimensions(&top.winsize);
1356 if (top.print_entries == 0) {
1357 struct sigaction act = {
1358 .sa_sigaction = perf_top__sig_winch,
1359 .sa_flags = SA_SIGINFO,
1360 };
1361 perf_top__update_print_entries(&top);
1362 sigaction(SIGWINCH, &act, NULL);
1363 }
1364
1365 status = __cmd_top(&top);
1366
1367 out_delete_evlist:
1368 perf_evlist__delete(top.evlist);
1369
1370 return status;
1371 }
Linus' kernel tree