projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'perf/urgent' into perf/core
[deliverable/linux.git] / tools / perf / builtin-top.c
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/color.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26 #include "util/thread.h"
27 #include "util/util.h"
28 #include <linux/rbtree.h>
29 #include "util/parse-options.h"
30 #include "util/parse-events.h"
31 #include "util/cpumap.h"
32
33 #include "util/debug.h"
34
35 #include <assert.h>
36 #include <fcntl.h>
37
38 #include <stdio.h>
39 #include <termios.h>
40 #include <unistd.h>
41
42 #include <errno.h>
43 #include <time.h>
44 #include <sched.h>
45 #include <pthread.h>
46
47 #include <sys/syscall.h>
48 #include <sys/ioctl.h>
49 #include <sys/poll.h>
50 #include <sys/prctl.h>
51 #include <sys/wait.h>
52 #include <sys/uio.h>
53 #include <sys/mman.h>
54
55 #include <linux/unistd.h>
56 #include <linux/types.h>
57
58 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
59
60 static int system_wide = 0;
61
62 static int default_interval = 0;
63
64 static int count_filter = 5;
65 static int print_entries;
66
67 static int target_pid = -1;
68 static int inherit = 0;
69 static int profile_cpu = -1;
70 static int nr_cpus = 0;
71 static unsigned int realtime_prio = 0;
72 static int group = 0;
73 static unsigned int page_size;
74 static unsigned int mmap_pages = 16;
75 static int freq = 1000; /* 1 KHz */
76
77 static int delay_secs = 2;
78 static int zero = 0;
79 static int dump_symtab = 0;
80
81 static bool hide_kernel_symbols = false;
82 static bool hide_user_symbols = false;
83 static struct winsize winsize;
84
85 /*
86 * Source
87 */
88
89 struct source_line {
90 u64 eip;
91 unsigned long count[MAX_COUNTERS];
92 char *line;
93 struct source_line *next;
94 };
95
96 static char *sym_filter = NULL;
97 struct sym_entry *sym_filter_entry = NULL;
98 struct sym_entry *sym_filter_entry_sched = NULL;
99 static int sym_pcnt_filter = 5;
100 static int sym_counter = 0;
101 static int display_weighted = -1;
102
103 /*
104 * Symbols
105 */
106
107 struct sym_entry_source {
108 struct source_line *source;
109 struct source_line *lines;
110 struct source_line **lines_tail;
111 pthread_mutex_t lock;
112 };
113
114 struct sym_entry {
115 struct rb_node rb_node;
116 struct list_head node;
117 unsigned long snap_count;
118 double weight;
119 int skip;
120 u16 name_len;
121 u8 origin;
122 struct map *map;
123 struct sym_entry_source *src;
124 unsigned long count[0];
125 };
126
127 /*
128 * Source functions
129 */
130
131 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
132 {
133 return ((void *)self) + symbol_conf.priv_size;
134 }
135
136 void get_term_dimensions(struct winsize *ws)
137 {
138 char *s = getenv("LINES");
139
140 if (s != NULL) {
141 ws->ws_row = atoi(s);
142 s = getenv("COLUMNS");
143 if (s != NULL) {
144 ws->ws_col = atoi(s);
145 if (ws->ws_row && ws->ws_col)
146 return;
147 }
148 }
149 #ifdef TIOCGWINSZ
150 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
151 ws->ws_row && ws->ws_col)
152 return;
153 #endif
154 ws->ws_row = 25;
155 ws->ws_col = 80;
156 }
157
158 static void update_print_entries(struct winsize *ws)
159 {
160 print_entries = ws->ws_row;
161
162 if (print_entries > 9)
163 print_entries -= 9;
164 }
165
166 static void sig_winch_handler(int sig __used)
167 {
168 get_term_dimensions(&winsize);
169 update_print_entries(&winsize);
170 }
171
172 static int parse_source(struct sym_entry *syme)
173 {
174 struct symbol *sym;
175 struct sym_entry_source *source;
176 struct map *map;
177 FILE *file;
178 char command[PATH_MAX*2];
179 const char *path;
180 u64 len;
181
182 if (!syme)
183 return -1;
184
185 sym = sym_entry__symbol(syme);
186 map = syme->map;
187
188 /*
189 * We can't annotate with just /proc/kallsyms
190 */
191 if (map->dso->origin == DSO__ORIG_KERNEL)
192 return -1;
193
194 if (syme->src == NULL) {
195 syme->src = zalloc(sizeof(*source));
196 if (syme->src == NULL)
197 return -1;
198 pthread_mutex_init(&syme->src->lock, NULL);
199 }
200
201 source = syme->src;
202
203 if (source->lines) {
204 pthread_mutex_lock(&source->lock);
205 goto out_assign;
206 }
207 path = map->dso->long_name;
208
209 len = sym->end - sym->start;
210
211 sprintf(command,
212 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
213 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
214 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
215
216 file = popen(command, "r");
217 if (!file)
218 return -1;
219
220 pthread_mutex_lock(&source->lock);
221 source->lines_tail = &source->lines;
222 while (!feof(file)) {
223 struct source_line *src;
224 size_t dummy = 0;
225 char *c, *sep;
226
227 src = malloc(sizeof(struct source_line));
228 assert(src != NULL);
229 memset(src, 0, sizeof(struct source_line));
230
231 if (getline(&src->line, &dummy, file) < 0)
232 break;
233 if (!src->line)
234 break;
235
236 c = strchr(src->line, '\n');
237 if (c)
238 *c = 0;
239
240 src->next = NULL;
241 *source->lines_tail = src;
242 source->lines_tail = &src->next;
243
244 src->eip = strtoull(src->line, &sep, 16);
245 if (*sep == ':')
246 src->eip = map__objdump_2ip(map, src->eip);
247 else /* this line has no ip info (e.g. source line) */
248 src->eip = 0;
249 }
250 pclose(file);
251 out_assign:
252 sym_filter_entry = syme;
253 pthread_mutex_unlock(&source->lock);
254 return 0;
255 }
256
257 static void __zero_source_counters(struct sym_entry *syme)
258 {
259 int i;
260 struct source_line *line;
261
262 line = syme->src->lines;
263 while (line) {
264 for (i = 0; i < nr_counters; i++)
265 line->count[i] = 0;
266 line = line->next;
267 }
268 }
269
270 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
271 {
272 struct source_line *line;
273
274 if (syme != sym_filter_entry)
275 return;
276
277 if (pthread_mutex_trylock(&syme->src->lock))
278 return;
279
280 if (syme->src == NULL || syme->src->source == NULL)
281 goto out_unlock;
282
283 for (line = syme->src->lines; line; line = line->next) {
284 /* skip lines without IP info */
285 if (line->eip == 0)
286 continue;
287 if (line->eip == ip) {
288 line->count[counter]++;
289 break;
290 }
291 if (line->eip > ip)
292 break;
293 }
294 out_unlock:
295 pthread_mutex_unlock(&syme->src->lock);
296 }
297
298 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
299
300 static void lookup_sym_source(struct sym_entry *syme)
301 {
302 struct symbol *symbol = sym_entry__symbol(syme);
303 struct source_line *line;
304 char pattern[PATTERN_LEN + 1];
305
306 sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
307 map__rip_2objdump(syme->map, symbol->start));
308
309 pthread_mutex_lock(&syme->src->lock);
310 for (line = syme->src->lines; line; line = line->next) {
311 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
312 syme->src->source = line;
313 break;
314 }
315 }
316 pthread_mutex_unlock(&syme->src->lock);
317 }
318
319 static void show_lines(struct source_line *queue, int count, int total)
320 {
321 int i;
322 struct source_line *line;
323
324 line = queue;
325 for (i = 0; i < count; i++) {
326 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
327
328 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
329 line = line->next;
330 }
331 }
332
333 #define TRACE_COUNT 3
334
335 static void show_details(struct sym_entry *syme)
336 {
337 struct symbol *symbol;
338 struct source_line *line;
339 struct source_line *line_queue = NULL;
340 int displayed = 0;
341 int line_queue_count = 0, total = 0, more = 0;
342
343 if (!syme)
344 return;
345
346 if (!syme->src->source)
347 lookup_sym_source(syme);
348
349 if (!syme->src->source)
350 return;
351
352 symbol = sym_entry__symbol(syme);
353 printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
354 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
355
356 pthread_mutex_lock(&syme->src->lock);
357 line = syme->src->source;
358 while (line) {
359 total += line->count[sym_counter];
360 line = line->next;
361 }
362
363 line = syme->src->source;
364 while (line) {
365 float pcnt = 0.0;
366
367 if (!line_queue_count)
368 line_queue = line;
369 line_queue_count++;
370
371 if (line->count[sym_counter])
372 pcnt = 100.0 * line->count[sym_counter] / (float)total;
373 if (pcnt >= (float)sym_pcnt_filter) {
374 if (displayed <= print_entries)
375 show_lines(line_queue, line_queue_count, total);
376 else more++;
377 displayed += line_queue_count;
378 line_queue_count = 0;
379 line_queue = NULL;
380 } else if (line_queue_count > TRACE_COUNT) {
381 line_queue = line_queue->next;
382 line_queue_count--;
383 }
384
385 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
386 line = line->next;
387 }
388 pthread_mutex_unlock(&syme->src->lock);
389 if (more)
390 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
391 }
392
393 /*
394 * Symbols will be added here in event__process_sample and will get out
395 * after decayed.
396 */
397 static LIST_HEAD(active_symbols);
398 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
399
400 /*
401 * Ordering weight: count-1 * count-2 * ... / count-n
402 */
403 static double sym_weight(const struct sym_entry *sym)
404 {
405 double weight = sym->snap_count;
406 int counter;
407
408 if (!display_weighted)
409 return weight;
410
411 for (counter = 1; counter < nr_counters-1; counter++)
412 weight *= sym->count[counter];
413
414 weight /= (sym->count[counter] + 1);
415
416 return weight;
417 }
418
419 static long samples;
420 static long userspace_samples;
421 static long exact_samples;
422 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
423
424 static void __list_insert_active_sym(struct sym_entry *syme)
425 {
426 list_add(&syme->node, &active_symbols);
427 }
428
429 static void list_remove_active_sym(struct sym_entry *syme)
430 {
431 pthread_mutex_lock(&active_symbols_lock);
432 list_del_init(&syme->node);
433 pthread_mutex_unlock(&active_symbols_lock);
434 }
435
436 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
437 {
438 struct rb_node **p = &tree->rb_node;
439 struct rb_node *parent = NULL;
440 struct sym_entry *iter;
441
442 while (*p != NULL) {
443 parent = *p;
444 iter = rb_entry(parent, struct sym_entry, rb_node);
445
446 if (se->weight > iter->weight)
447 p = &(*p)->rb_left;
448 else
449 p = &(*p)->rb_right;
450 }
451
452 rb_link_node(&se->rb_node, parent, p);
453 rb_insert_color(&se->rb_node, tree);
454 }
455
456 static void print_sym_table(void)
457 {
458 int printed = 0, j;
459 int counter, snap = !display_weighted ? sym_counter : 0;
460 float samples_per_sec = samples/delay_secs;
461 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
462 float esamples_percent = (100.0*exact_samples)/samples;
463 float sum_ksamples = 0.0;
464 struct sym_entry *syme, *n;
465 struct rb_root tmp = RB_ROOT;
466 struct rb_node *nd;
467 int sym_width = 0, dso_width = 0, dso_short_width = 0;
468 const int win_width = winsize.ws_col - 1;
469
470 samples = userspace_samples = exact_samples = 0;
471
472 /* Sort the active symbols */
473 pthread_mutex_lock(&active_symbols_lock);
474 syme = list_entry(active_symbols.next, struct sym_entry, node);
475 pthread_mutex_unlock(&active_symbols_lock);
476
477 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
478 syme->snap_count = syme->count[snap];
479 if (syme->snap_count != 0) {
480
481 if ((hide_user_symbols &&
482 syme->origin == PERF_RECORD_MISC_USER) ||
483 (hide_kernel_symbols &&
484 syme->origin == PERF_RECORD_MISC_KERNEL)) {
485 list_remove_active_sym(syme);
486 continue;
487 }
488 syme->weight = sym_weight(syme);
489 rb_insert_active_sym(&tmp, syme);
490 sum_ksamples += syme->snap_count;
491
492 for (j = 0; j < nr_counters; j++)
493 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
494 } else
495 list_remove_active_sym(syme);
496 }
497
498 puts(CONSOLE_CLEAR);
499
500 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
501 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% exact: %4.1f%% [",
502 samples_per_sec,
503 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)),
504 esamples_percent);
505
506 if (nr_counters == 1 || !display_weighted) {
507 printf("%Ld", (u64)attrs[0].sample_period);
508 if (freq)
509 printf("Hz ");
510 else
511 printf(" ");
512 }
513
514 if (!display_weighted)
515 printf("%s", event_name(sym_counter));
516 else for (counter = 0; counter < nr_counters; counter++) {
517 if (counter)
518 printf("/");
519
520 printf("%s", event_name(counter));
521 }
522
523 printf( "], ");
524
525 if (target_pid != -1)
526 printf(" (target_pid: %d", target_pid);
527 else
528 printf(" (all");
529
530 if (profile_cpu != -1)
531 printf(", cpu: %d)\n", profile_cpu);
532 else {
533 if (target_pid != -1)
534 printf(")\n");
535 else
536 printf(", %d CPUs)\n", nr_cpus);
537 }
538
539 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
540
541 if (sym_filter_entry) {
542 show_details(sym_filter_entry);
543 return;
544 }
545
546 /*
547 * Find the longest symbol name that will be displayed
548 */
549 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
550 syme = rb_entry(nd, struct sym_entry, rb_node);
551 if (++printed > print_entries ||
552 (int)syme->snap_count < count_filter)
553 continue;
554
555 if (syme->map->dso->long_name_len > dso_width)
556 dso_width = syme->map->dso->long_name_len;
557
558 if (syme->map->dso->short_name_len > dso_short_width)
559 dso_short_width = syme->map->dso->short_name_len;
560
561 if (syme->name_len > sym_width)
562 sym_width = syme->name_len;
563 }
564
565 printed = 0;
566
567 if (sym_width + dso_width > winsize.ws_col - 29) {
568 dso_width = dso_short_width;
569 if (sym_width + dso_width > winsize.ws_col - 29)
570 sym_width = winsize.ws_col - dso_width - 29;
571 }
572 putchar('\n');
573 if (nr_counters == 1)
574 printf(" samples pcnt");
575 else
576 printf(" weight samples pcnt");
577
578 if (verbose)
579 printf(" RIP ");
580 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
581 printf(" %s _______ _____",
582 nr_counters == 1 ? " " : "______");
583 if (verbose)
584 printf(" ________________");
585 printf(" %-*.*s", sym_width, sym_width, graph_line);
586 printf(" %-*.*s", dso_width, dso_width, graph_line);
587 puts("\n");
588
589 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
590 struct symbol *sym;
591 double pcnt;
592
593 syme = rb_entry(nd, struct sym_entry, rb_node);
594 sym = sym_entry__symbol(syme);
595
596 if (++printed > print_entries || (int)syme->snap_count < count_filter)
597 continue;
598
599 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
600 sum_ksamples));
601
602 if (nr_counters == 1 || !display_weighted)
603 printf("%20.2f ", syme->weight);
604 else
605 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
606
607 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
608 if (verbose)
609 printf(" %016llx", sym->start);
610 printf(" %-*.*s", sym_width, sym_width, sym->name);
611 printf(" %-*.*s\n", dso_width, dso_width,
612 dso_width >= syme->map->dso->long_name_len ?
613 syme->map->dso->long_name :
614 syme->map->dso->short_name);
615 }
616 }
617
618 static void prompt_integer(int *target, const char *msg)
619 {
620 char *buf = malloc(0), *p;
621 size_t dummy = 0;
622 int tmp;
623
624 fprintf(stdout, "\n%s: ", msg);
625 if (getline(&buf, &dummy, stdin) < 0)
626 return;
627
628 p = strchr(buf, '\n');
629 if (p)
630 *p = 0;
631
632 p = buf;
633 while(*p) {
634 if (!isdigit(*p))
635 goto out_free;
636 p++;
637 }
638 tmp = strtoul(buf, NULL, 10);
639 *target = tmp;
640 out_free:
641 free(buf);
642 }
643
644 static void prompt_percent(int *target, const char *msg)
645 {
646 int tmp = 0;
647
648 prompt_integer(&tmp, msg);
649 if (tmp >= 0 && tmp <= 100)
650 *target = tmp;
651 }
652
653 static void prompt_symbol(struct sym_entry **target, const char *msg)
654 {
655 char *buf = malloc(0), *p;
656 struct sym_entry *syme = *target, *n, *found = NULL;
657 size_t dummy = 0;
658
659 /* zero counters of active symbol */
660 if (syme) {
661 pthread_mutex_lock(&syme->src->lock);
662 __zero_source_counters(syme);
663 *target = NULL;
664 pthread_mutex_unlock(&syme->src->lock);
665 }
666
667 fprintf(stdout, "\n%s: ", msg);
668 if (getline(&buf, &dummy, stdin) < 0)
669 goto out_free;
670
671 p = strchr(buf, '\n');
672 if (p)
673 *p = 0;
674
675 pthread_mutex_lock(&active_symbols_lock);
676 syme = list_entry(active_symbols.next, struct sym_entry, node);
677 pthread_mutex_unlock(&active_symbols_lock);
678
679 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
680 struct symbol *sym = sym_entry__symbol(syme);
681
682 if (!strcmp(buf, sym->name)) {
683 found = syme;
684 break;
685 }
686 }
687
688 if (!found) {
689 fprintf(stderr, "Sorry, %s is not active.\n", buf);
690 sleep(1);
691 return;
692 } else
693 parse_source(found);
694
695 out_free:
696 free(buf);
697 }
698
699 static void print_mapped_keys(void)
700 {
701 char *name = NULL;
702
703 if (sym_filter_entry) {
704 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
705 name = sym->name;
706 }
707
708 fprintf(stdout, "\nMapped keys:\n");
709 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
710 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
711
712 if (nr_counters > 1)
713 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter));
714
715 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
716
717 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
718 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
719 fprintf(stdout, "\t[S] stop annotation.\n");
720
721 if (nr_counters > 1)
722 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
723
724 fprintf(stdout,
725 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
726 hide_kernel_symbols ? "yes" : "no");
727 fprintf(stdout,
728 "\t[U] hide user symbols. \t(%s)\n",
729 hide_user_symbols ? "yes" : "no");
730 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
731 fprintf(stdout, "\t[qQ] quit.\n");
732 }
733
734 static int key_mapped(int c)
735 {
736 switch (c) {
737 case 'd':
738 case 'e':
739 case 'f':
740 case 'z':
741 case 'q':
742 case 'Q':
743 case 'K':
744 case 'U':
745 case 'F':
746 case 's':
747 case 'S':
748 return 1;
749 case 'E':
750 case 'w':
751 return nr_counters > 1 ? 1 : 0;
752 default:
753 break;
754 }
755
756 return 0;
757 }
758
759 static void handle_keypress(int c)
760 {
761 if (!key_mapped(c)) {
762 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
763 struct termios tc, save;
764
765 print_mapped_keys();
766 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
767 fflush(stdout);
768
769 tcgetattr(0, &save);
770 tc = save;
771 tc.c_lflag &= ~(ICANON | ECHO);
772 tc.c_cc[VMIN] = 0;
773 tc.c_cc[VTIME] = 0;
774 tcsetattr(0, TCSANOW, &tc);
775
776 poll(&stdin_poll, 1, -1);
777 c = getc(stdin);
778
779 tcsetattr(0, TCSAFLUSH, &save);
780 if (!key_mapped(c))
781 return;
782 }
783
784 switch (c) {
785 case 'd':
786 prompt_integer(&delay_secs, "Enter display delay");
787 if (delay_secs < 1)
788 delay_secs = 1;
789 break;
790 case 'e':
791 prompt_integer(&print_entries, "Enter display entries (lines)");
792 if (print_entries == 0) {
793 sig_winch_handler(SIGWINCH);
794 signal(SIGWINCH, sig_winch_handler);
795 } else
796 signal(SIGWINCH, SIG_DFL);
797 break;
798 case 'E':
799 if (nr_counters > 1) {
800 int i;
801
802 fprintf(stderr, "\nAvailable events:");
803 for (i = 0; i < nr_counters; i++)
804 fprintf(stderr, "\n\t%d %s", i, event_name(i));
805
806 prompt_integer(&sym_counter, "Enter details event counter");
807
808 if (sym_counter >= nr_counters) {
809 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
810 sym_counter = 0;
811 sleep(1);
812 }
813 } else sym_counter = 0;
814 break;
815 case 'f':
816 prompt_integer(&count_filter, "Enter display event count filter");
817 break;
818 case 'F':
819 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
820 break;
821 case 'K':
822 hide_kernel_symbols = !hide_kernel_symbols;
823 break;
824 case 'q':
825 case 'Q':
826 printf("exiting.\n");
827 if (dump_symtab)
828 dsos__fprintf(stderr);
829 exit(0);
830 case 's':
831 prompt_symbol(&sym_filter_entry, "Enter details symbol");
832 break;
833 case 'S':
834 if (!sym_filter_entry)
835 break;
836 else {
837 struct sym_entry *syme = sym_filter_entry;
838
839 pthread_mutex_lock(&syme->src->lock);
840 sym_filter_entry = NULL;
841 __zero_source_counters(syme);
842 pthread_mutex_unlock(&syme->src->lock);
843 }
844 break;
845 case 'U':
846 hide_user_symbols = !hide_user_symbols;
847 break;
848 case 'w':
849 display_weighted = ~display_weighted;
850 break;
851 case 'z':
852 zero = ~zero;
853 break;
854 default:
855 break;
856 }
857 }
858
859 static void *display_thread(void *arg __used)
860 {
861 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
862 struct termios tc, save;
863 int delay_msecs, c;
864
865 tcgetattr(0, &save);
866 tc = save;
867 tc.c_lflag &= ~(ICANON | ECHO);
868 tc.c_cc[VMIN] = 0;
869 tc.c_cc[VTIME] = 0;
870
871 repeat:
872 delay_msecs = delay_secs * 1000;
873 tcsetattr(0, TCSANOW, &tc);
874 /* trash return*/
875 getc(stdin);
876
877 do {
878 print_sym_table();
879 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
880
881 c = getc(stdin);
882 tcsetattr(0, TCSAFLUSH, &save);
883
884 handle_keypress(c);
885 goto repeat;
886
887 return NULL;
888 }
889
890 /* Tag samples to be skipped. */
891 static const char *skip_symbols[] = {
892 "default_idle",
893 "cpu_idle",
894 "enter_idle",
895 "exit_idle",
896 "mwait_idle",
897 "mwait_idle_with_hints",
898 "poll_idle",
899 "ppc64_runlatch_off",
900 "pseries_dedicated_idle_sleep",
901 NULL
902 };
903
904 static int symbol_filter(struct map *map, struct symbol *sym)
905 {
906 struct sym_entry *syme;
907 const char *name = sym->name;
908 int i;
909
910 /*
911 * ppc64 uses function descriptors and appends a '.' to the
912 * start of every instruction address. Remove it.
913 */
914 if (name[0] == '.')
915 name++;
916
917 if (!strcmp(name, "_text") ||
918 !strcmp(name, "_etext") ||
919 !strcmp(name, "_sinittext") ||
920 !strncmp("init_module", name, 11) ||
921 !strncmp("cleanup_module", name, 14) ||
922 strstr(name, "_text_start") ||
923 strstr(name, "_text_end"))
924 return 1;
925
926 syme = symbol__priv(sym);
927 syme->map = map;
928 syme->src = NULL;
929
930 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
931 /* schedule initial sym_filter_entry setup */
932 sym_filter_entry_sched = syme;
933 sym_filter = NULL;
934 }
935
936 for (i = 0; skip_symbols[i]; i++) {
937 if (!strcmp(skip_symbols[i], name)) {
938 syme->skip = 1;
939 break;
940 }
941 }
942
943 if (!syme->skip)
944 syme->name_len = strlen(sym->name);
945
946 return 0;
947 }
948
949 static void event__process_sample(const event_t *self,
950 struct perf_session *session, int counter)
951 {
952 u64 ip = self->ip.ip;
953 struct sym_entry *syme;
954 struct addr_location al;
955 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
956
957 ++samples;
958
959 switch (origin) {
960 case PERF_RECORD_MISC_USER:
961 ++userspace_samples;
962 if (hide_user_symbols)
963 return;
964 break;
965 case PERF_RECORD_MISC_KERNEL:
966 if (hide_kernel_symbols)
967 return;
968 break;
969 default:
970 return;
971 }
972
973 if (self->header.misc & PERF_RECORD_MISC_EXACT)
974 exact_samples++;
975
976 if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
977 al.filtered)
978 return;
979
980 if (al.sym == NULL) {
981 /*
982 * As we do lazy loading of symtabs we only will know if the
983 * specified vmlinux file is invalid when we actually have a
984 * hit in kernel space and then try to load it. So if we get
985 * here and there are _no_ symbols in the DSO backing the
986 * kernel map, bail out.
987 *
988 * We may never get here, for instance, if we use -K/
989 * --hide-kernel-symbols, even if the user specifies an
990 * invalid --vmlinux ;-)
991 */
992 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
993 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
994 pr_err("The %s file can't be used\n",
995 symbol_conf.vmlinux_name);
996 exit(1);
997 }
998
999 return;
1000 }
1001
1002 /* let's see, whether we need to install initial sym_filter_entry */
1003 if (sym_filter_entry_sched) {
1004 sym_filter_entry = sym_filter_entry_sched;
1005 sym_filter_entry_sched = NULL;
1006 if (parse_source(sym_filter_entry) < 0) {
1007 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1008
1009 pr_err("Can't annotate %s", sym->name);
1010 if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1011 pr_err(": No vmlinux file was found in the path:\n");
1012 vmlinux_path__fprintf(stderr);
1013 } else
1014 pr_err(".\n");
1015 exit(1);
1016 }
1017 }
1018
1019 syme = symbol__priv(al.sym);
1020 if (!syme->skip) {
1021 syme->count[counter]++;
1022 syme->origin = origin;
1023 record_precise_ip(syme, counter, ip);
1024 pthread_mutex_lock(&active_symbols_lock);
1025 if (list_empty(&syme->node) || !syme->node.next)
1026 __list_insert_active_sym(syme);
1027 pthread_mutex_unlock(&active_symbols_lock);
1028 }
1029 }
1030
1031 static int event__process(event_t *event, struct perf_session *session)
1032 {
1033 switch (event->header.type) {
1034 case PERF_RECORD_COMM:
1035 event__process_comm(event, session);
1036 break;
1037 case PERF_RECORD_MMAP:
1038 event__process_mmap(event, session);
1039 break;
1040 case PERF_RECORD_FORK:
1041 case PERF_RECORD_EXIT:
1042 event__process_task(event, session);
1043 break;
1044 default:
1045 break;
1046 }
1047
1048 return 0;
1049 }
1050
1051 struct mmap_data {
1052 int counter;
1053 void *base;
1054 int mask;
1055 unsigned int prev;
1056 };
1057
1058 static unsigned int mmap_read_head(struct mmap_data *md)
1059 {
1060 struct perf_event_mmap_page *pc = md->base;
1061 int head;
1062
1063 head = pc->data_head;
1064 rmb();
1065
1066 return head;
1067 }
1068
1069 static void perf_session__mmap_read_counter(struct perf_session *self,
1070 struct mmap_data *md)
1071 {
1072 unsigned int head = mmap_read_head(md);
1073 unsigned int old = md->prev;
1074 unsigned char *data = md->base + page_size;
1075 int diff;
1076
1077 /*
1078 * If we're further behind than half the buffer, there's a chance
1079 * the writer will bite our tail and mess up the samples under us.
1080 *
1081 * If we somehow ended up ahead of the head, we got messed up.
1082 *
1083 * In either case, truncate and restart at head.
1084 */
1085 diff = head - old;
1086 if (diff > md->mask / 2 || diff < 0) {
1087 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1088
1089 /*
1090 * head points to a known good entry, start there.
1091 */
1092 old = head;
1093 }
1094
1095 for (; old != head;) {
1096 event_t *event = (event_t *)&data[old & md->mask];
1097
1098 event_t event_copy;
1099
1100 size_t size = event->header.size;
1101
1102 /*
1103 * Event straddles the mmap boundary -- header should always
1104 * be inside due to u64 alignment of output.
1105 */
1106 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1107 unsigned int offset = old;
1108 unsigned int len = min(sizeof(*event), size), cpy;
1109 void *dst = &event_copy;
1110
1111 do {
1112 cpy = min(md->mask + 1 - (offset & md->mask), len);
1113 memcpy(dst, &data[offset & md->mask], cpy);
1114 offset += cpy;
1115 dst += cpy;
1116 len -= cpy;
1117 } while (len);
1118
1119 event = &event_copy;
1120 }
1121
1122 if (event->header.type == PERF_RECORD_SAMPLE)
1123 event__process_sample(event, self, md->counter);
1124 else
1125 event__process(event, self);
1126 old += size;
1127 }
1128
1129 md->prev = old;
1130 }
1131
1132 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1133 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1134
1135 static void perf_session__mmap_read(struct perf_session *self)
1136 {
1137 int i, counter;
1138
1139 for (i = 0; i < nr_cpus; i++) {
1140 for (counter = 0; counter < nr_counters; counter++)
1141 perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1142 }
1143 }
1144
1145 int nr_poll;
1146 int group_fd;
1147
1148 static void start_counter(int i, int counter)
1149 {
1150 struct perf_event_attr *attr;
1151 int cpu;
1152
1153 cpu = profile_cpu;
1154 if (target_pid == -1 && profile_cpu == -1)
1155 cpu = cpumap[i];
1156
1157 attr = attrs + counter;
1158
1159 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1160
1161 if (freq) {
1162 attr->sample_type |= PERF_SAMPLE_PERIOD;
1163 attr->freq = 1;
1164 attr->sample_freq = freq;
1165 }
1166
1167 attr->inherit = (cpu < 0) && inherit;
1168 attr->mmap = 1;
1169
1170 try_again:
1171 fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1172
1173 if (fd[i][counter] < 0) {
1174 int err = errno;
1175
1176 if (err == EPERM || err == EACCES)
1177 die("No permission - are you root?\n");
1178 /*
1179 * If it's cycles then fall back to hrtimer
1180 * based cpu-clock-tick sw counter, which
1181 * is always available even if no PMU support:
1182 */
1183 if (attr->type == PERF_TYPE_HARDWARE
1184 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1185
1186 if (verbose)
1187 warning(" ... trying to fall back to cpu-clock-ticks\n");
1188
1189 attr->type = PERF_TYPE_SOFTWARE;
1190 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1191 goto try_again;
1192 }
1193 printf("\n");
1194 error("perfcounter syscall returned with %d (%s)\n",
1195 fd[i][counter], strerror(err));
1196 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1197 exit(-1);
1198 }
1199 assert(fd[i][counter] >= 0);
1200 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1201
1202 /*
1203 * First counter acts as the group leader:
1204 */
1205 if (group && group_fd == -1)
1206 group_fd = fd[i][counter];
1207
1208 event_array[nr_poll].fd = fd[i][counter];
1209 event_array[nr_poll].events = POLLIN;
1210 nr_poll++;
1211
1212 mmap_array[i][counter].counter = counter;
1213 mmap_array[i][counter].prev = 0;
1214 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1215 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1216 PROT_READ, MAP_SHARED, fd[i][counter], 0);
1217 if (mmap_array[i][counter].base == MAP_FAILED)
1218 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1219 }
1220
1221 static int __cmd_top(void)
1222 {
1223 pthread_t thread;
1224 int i, counter;
1225 int ret;
1226 /*
1227 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1228 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1229 */
1230 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1231 if (session == NULL)
1232 return -ENOMEM;
1233
1234 if (target_pid != -1)
1235 event__synthesize_thread(target_pid, event__process, session);
1236 else
1237 event__synthesize_threads(event__process, session);
1238
1239 for (i = 0; i < nr_cpus; i++) {
1240 group_fd = -1;
1241 for (counter = 0; counter < nr_counters; counter++)
1242 start_counter(i, counter);
1243 }
1244
1245 /* Wait for a minimal set of events before starting the snapshot */
1246 poll(event_array, nr_poll, 100);
1247
1248 perf_session__mmap_read(session);
1249
1250 if (pthread_create(&thread, NULL, display_thread, NULL)) {
1251 printf("Could not create display thread.\n");
1252 exit(-1);
1253 }
1254
1255 if (realtime_prio) {
1256 struct sched_param param;
1257
1258 param.sched_priority = realtime_prio;
1259 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1260 printf("Could not set realtime priority.\n");
1261 exit(-1);
1262 }
1263 }
1264
1265 while (1) {
1266 int hits = samples;
1267
1268 perf_session__mmap_read(session);
1269
1270 if (hits == samples)
1271 ret = poll(event_array, nr_poll, 100);
1272 }
1273
1274 return 0;
1275 }
1276
1277 static const char * const top_usage[] = {
1278 "perf top [<options>]",
1279 NULL
1280 };
1281
1282 static const struct option options[] = {
1283 OPT_CALLBACK('e', "event", NULL, "event",
1284 "event selector. use 'perf list' to list available events",
1285 parse_events),
1286 OPT_INTEGER('c', "count", &default_interval,
1287 "event period to sample"),
1288 OPT_INTEGER('p', "pid", &target_pid,
1289 "profile events on existing pid"),
1290 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1291 "system-wide collection from all CPUs"),
1292 OPT_INTEGER('C', "CPU", &profile_cpu,
1293 "CPU to profile on"),
1294 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1295 "file", "vmlinux pathname"),
1296 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1297 "hide kernel symbols"),
1298 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1299 "number of mmap data pages"),
1300 OPT_INTEGER('r', "realtime", &realtime_prio,
1301 "collect data with this RT SCHED_FIFO priority"),
1302 OPT_INTEGER('d', "delay", &delay_secs,
1303 "number of seconds to delay between refreshes"),
1304 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1305 "dump the symbol table used for profiling"),
1306 OPT_INTEGER('f', "count-filter", &count_filter,
1307 "only display functions with more events than this"),
1308 OPT_BOOLEAN('g', "group", &group,
1309 "put the counters into a counter group"),
1310 OPT_BOOLEAN('i', "inherit", &inherit,
1311 "child tasks inherit counters"),
1312 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1313 "symbol to annotate"),
1314 OPT_BOOLEAN('z', "zero", &zero,
1315 "zero history across updates"),
1316 OPT_INTEGER('F', "freq", &freq,
1317 "profile at this frequency"),
1318 OPT_INTEGER('E', "entries", &print_entries,
1319 "display this many functions"),
1320 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1321 "hide user symbols"),
1322 OPT_BOOLEAN('v', "verbose", &verbose,
1323 "be more verbose (show counter open errors, etc)"),
1324 OPT_END()
1325 };
1326
1327 int cmd_top(int argc, const char **argv, const char *prefix __used)
1328 {
1329 int counter;
1330
1331 page_size = sysconf(_SC_PAGE_SIZE);
1332
1333 argc = parse_options(argc, argv, options, top_usage, 0);
1334 if (argc)
1335 usage_with_options(top_usage, options);
1336
1337 /* CPU and PID are mutually exclusive */
1338 if (target_pid != -1 && profile_cpu != -1) {
1339 printf("WARNING: PID switch overriding CPU\n");
1340 sleep(1);
1341 profile_cpu = -1;
1342 }
1343
1344 if (!nr_counters)
1345 nr_counters = 1;
1346
1347 symbol_conf.priv_size = (sizeof(struct sym_entry) +
1348 (nr_counters + 1) * sizeof(unsigned long));
1349
1350 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1351 if (symbol__init() < 0)
1352 return -1;
1353
1354 if (delay_secs < 1)
1355 delay_secs = 1;
1356
1357 /*
1358 * User specified count overrides default frequency.
1359 */
1360 if (default_interval)
1361 freq = 0;
1362 else if (freq) {
1363 default_interval = freq;
1364 } else {
1365 fprintf(stderr, "frequency and count are zero, aborting\n");
1366 exit(EXIT_FAILURE);
1367 }
1368
1369 /*
1370 * Fill in the ones not specifically initialized via -c:
1371 */
1372 for (counter = 0; counter < nr_counters; counter++) {
1373 if (attrs[counter].sample_period)
1374 continue;
1375
1376 attrs[counter].sample_period = default_interval;
1377 }
1378
1379 if (target_pid != -1 || profile_cpu != -1)
1380 nr_cpus = 1;
1381 else
1382 nr_cpus = read_cpu_map();
1383
1384 get_term_dimensions(&winsize);
1385 if (print_entries == 0) {
1386 update_print_entries(&winsize);
1387 signal(SIGWINCH, sig_winch_handler);
1388 }
1389
1390 return __cmd_top();
1391 }
Linux kernel - Deliverables
This page took 0.060779 seconds and 5 git commands to generate.