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 static 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 void 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;
184
185 if (syme->src == NULL) {
186 syme->src = zalloc(sizeof(*source));
187 if (syme->src == NULL)
188 return;
189 pthread_mutex_init(&syme->src->lock, NULL);
190 }
191
192 source = syme->src;
193
194 if (source->lines) {
195 pthread_mutex_lock(&source->lock);
196 goto out_assign;
197 }
198
199 sym = sym_entry__symbol(syme);
200 map = syme->map;
201 path = map->dso->long_name;
202
203 len = sym->end - sym->start;
204
205 sprintf(command,
206 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
207 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
208 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
209
210 file = popen(command, "r");
211 if (!file)
212 return;
213
214 pthread_mutex_lock(&source->lock);
215 source->lines_tail = &source->lines;
216 while (!feof(file)) {
217 struct source_line *src;
218 size_t dummy = 0;
219 char *c, *sep;
220
221 src = malloc(sizeof(struct source_line));
222 assert(src != NULL);
223 memset(src, 0, sizeof(struct source_line));
224
225 if (getline(&src->line, &dummy, file) < 0)
226 break;
227 if (!src->line)
228 break;
229
230 c = strchr(src->line, '\n');
231 if (c)
232 *c = 0;
233
234 src->next = NULL;
235 *source->lines_tail = src;
236 source->lines_tail = &src->next;
237
238 src->eip = strtoull(src->line, &sep, 16);
239 if (*sep == ':')
240 src->eip = map__objdump_2ip(map, src->eip);
241 else /* this line has no ip info (e.g. source line) */
242 src->eip = 0;
243 }
244 pclose(file);
245 out_assign:
246 sym_filter_entry = syme;
247 pthread_mutex_unlock(&source->lock);
248 }
249
250 static void __zero_source_counters(struct sym_entry *syme)
251 {
252 int i;
253 struct source_line *line;
254
255 line = syme->src->lines;
256 while (line) {
257 for (i = 0; i < nr_counters; i++)
258 line->count[i] = 0;
259 line = line->next;
260 }
261 }
262
263 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
264 {
265 struct source_line *line;
266
267 if (syme != sym_filter_entry)
268 return;
269
270 if (pthread_mutex_trylock(&syme->src->lock))
271 return;
272
273 if (syme->src == NULL || syme->src->source == NULL)
274 goto out_unlock;
275
276 for (line = syme->src->lines; line; line = line->next) {
277 /* skip lines without IP info */
278 if (line->eip == 0)
279 continue;
280 if (line->eip == ip) {
281 line->count[counter]++;
282 break;
283 }
284 if (line->eip > ip)
285 break;
286 }
287 out_unlock:
288 pthread_mutex_unlock(&syme->src->lock);
289 }
290
291 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
292
293 static void lookup_sym_source(struct sym_entry *syme)
294 {
295 struct symbol *symbol = sym_entry__symbol(syme);
296 struct source_line *line;
297 char pattern[PATTERN_LEN + 1];
298
299 sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
300 map__rip_2objdump(syme->map, symbol->start));
301
302 pthread_mutex_lock(&syme->src->lock);
303 for (line = syme->src->lines; line; line = line->next) {
304 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
305 syme->src->source = line;
306 break;
307 }
308 }
309 pthread_mutex_unlock(&syme->src->lock);
310 }
311
312 static void show_lines(struct source_line *queue, int count, int total)
313 {
314 int i;
315 struct source_line *line;
316
317 line = queue;
318 for (i = 0; i < count; i++) {
319 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
320
321 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
322 line = line->next;
323 }
324 }
325
326 #define TRACE_COUNT 3
327
328 static void show_details(struct sym_entry *syme)
329 {
330 struct symbol *symbol;
331 struct source_line *line;
332 struct source_line *line_queue = NULL;
333 int displayed = 0;
334 int line_queue_count = 0, total = 0, more = 0;
335
336 if (!syme)
337 return;
338
339 if (!syme->src->source)
340 lookup_sym_source(syme);
341
342 if (!syme->src->source)
343 return;
344
345 symbol = sym_entry__symbol(syme);
346 printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
347 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
348
349 pthread_mutex_lock(&syme->src->lock);
350 line = syme->src->source;
351 while (line) {
352 total += line->count[sym_counter];
353 line = line->next;
354 }
355
356 line = syme->src->source;
357 while (line) {
358 float pcnt = 0.0;
359
360 if (!line_queue_count)
361 line_queue = line;
362 line_queue_count++;
363
364 if (line->count[sym_counter])
365 pcnt = 100.0 * line->count[sym_counter] / (float)total;
366 if (pcnt >= (float)sym_pcnt_filter) {
367 if (displayed <= print_entries)
368 show_lines(line_queue, line_queue_count, total);
369 else more++;
370 displayed += line_queue_count;
371 line_queue_count = 0;
372 line_queue = NULL;
373 } else if (line_queue_count > TRACE_COUNT) {
374 line_queue = line_queue->next;
375 line_queue_count--;
376 }
377
378 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
379 line = line->next;
380 }
381 pthread_mutex_unlock(&syme->src->lock);
382 if (more)
383 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
384 }
385
386 /*
387 * Symbols will be added here in event__process_sample and will get out
388 * after decayed.
389 */
390 static LIST_HEAD(active_symbols);
391 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
392
393 /*
394 * Ordering weight: count-1 * count-2 * ... / count-n
395 */
396 static double sym_weight(const struct sym_entry *sym)
397 {
398 double weight = sym->snap_count;
399 int counter;
400
401 if (!display_weighted)
402 return weight;
403
404 for (counter = 1; counter < nr_counters-1; counter++)
405 weight *= sym->count[counter];
406
407 weight /= (sym->count[counter] + 1);
408
409 return weight;
410 }
411
412 static long samples;
413 static long userspace_samples;
414 static long exact_samples;
415 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
416
417 static void __list_insert_active_sym(struct sym_entry *syme)
418 {
419 list_add(&syme->node, &active_symbols);
420 }
421
422 static void list_remove_active_sym(struct sym_entry *syme)
423 {
424 pthread_mutex_lock(&active_symbols_lock);
425 list_del_init(&syme->node);
426 pthread_mutex_unlock(&active_symbols_lock);
427 }
428
429 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
430 {
431 struct rb_node **p = &tree->rb_node;
432 struct rb_node *parent = NULL;
433 struct sym_entry *iter;
434
435 while (*p != NULL) {
436 parent = *p;
437 iter = rb_entry(parent, struct sym_entry, rb_node);
438
439 if (se->weight > iter->weight)
440 p = &(*p)->rb_left;
441 else
442 p = &(*p)->rb_right;
443 }
444
445 rb_link_node(&se->rb_node, parent, p);
446 rb_insert_color(&se->rb_node, tree);
447 }
448
449 static void print_sym_table(void)
450 {
451 int printed = 0, j;
452 int counter, snap = !display_weighted ? sym_counter : 0;
453 float samples_per_sec = samples/delay_secs;
454 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
455 float esamples_percent = (100.0*exact_samples)/samples;
456 float sum_ksamples = 0.0;
457 struct sym_entry *syme, *n;
458 struct rb_root tmp = RB_ROOT;
459 struct rb_node *nd;
460 int sym_width = 0, dso_width = 0, max_dso_width;
461 const int win_width = winsize.ws_col - 1;
462
463 samples = userspace_samples = exact_samples = 0;
464
465 /* Sort the active symbols */
466 pthread_mutex_lock(&active_symbols_lock);
467 syme = list_entry(active_symbols.next, struct sym_entry, node);
468 pthread_mutex_unlock(&active_symbols_lock);
469
470 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
471 syme->snap_count = syme->count[snap];
472 if (syme->snap_count != 0) {
473
474 if ((hide_user_symbols &&
475 syme->origin == PERF_RECORD_MISC_USER) ||
476 (hide_kernel_symbols &&
477 syme->origin == PERF_RECORD_MISC_KERNEL)) {
478 list_remove_active_sym(syme);
479 continue;
480 }
481 syme->weight = sym_weight(syme);
482 rb_insert_active_sym(&tmp, syme);
483 sum_ksamples += syme->snap_count;
484
485 for (j = 0; j < nr_counters; j++)
486 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
487 } else
488 list_remove_active_sym(syme);
489 }
490
491 puts(CONSOLE_CLEAR);
492
493 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
494 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% exact: %4.1f%% [",
495 samples_per_sec,
496 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)),
497 esamples_percent);
498
499 if (nr_counters == 1 || !display_weighted) {
500 printf("%Ld", (u64)attrs[0].sample_period);
501 if (freq)
502 printf("Hz ");
503 else
504 printf(" ");
505 }
506
507 if (!display_weighted)
508 printf("%s", event_name(sym_counter));
509 else for (counter = 0; counter < nr_counters; counter++) {
510 if (counter)
511 printf("/");
512
513 printf("%s", event_name(counter));
514 }
515
516 printf( "], ");
517
518 if (target_pid != -1)
519 printf(" (target_pid: %d", target_pid);
520 else
521 printf(" (all");
522
523 if (profile_cpu != -1)
524 printf(", cpu: %d)\n", profile_cpu);
525 else {
526 if (target_pid != -1)
527 printf(")\n");
528 else
529 printf(", %d CPUs)\n", nr_cpus);
530 }
531
532 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
533
534 if (sym_filter_entry) {
535 show_details(sym_filter_entry);
536 return;
537 }
538
539 /*
540 * Find the longest symbol name that will be displayed
541 */
542 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
543 syme = rb_entry(nd, struct sym_entry, rb_node);
544 if (++printed > print_entries ||
545 (int)syme->snap_count < count_filter)
546 continue;
547
548 if (syme->map->dso->long_name_len > dso_width)
549 dso_width = syme->map->dso->long_name_len;
550
551 if (syme->name_len > sym_width)
552 sym_width = syme->name_len;
553 }
554
555 printed = 0;
556
557 max_dso_width = winsize.ws_col - sym_width - 29;
558 if (dso_width > max_dso_width)
559 dso_width = max_dso_width;
560 putchar('\n');
561 if (nr_counters == 1)
562 printf(" samples pcnt");
563 else
564 printf(" weight samples pcnt");
565
566 if (verbose)
567 printf(" RIP ");
568 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
569 printf(" %s _______ _____",
570 nr_counters == 1 ? " " : "______");
571 if (verbose)
572 printf(" ________________");
573 printf(" %-*.*s", sym_width, sym_width, graph_line);
574 printf(" %-*.*s", dso_width, dso_width, graph_line);
575 puts("\n");
576
577 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
578 struct symbol *sym;
579 double pcnt;
580
581 syme = rb_entry(nd, struct sym_entry, rb_node);
582 sym = sym_entry__symbol(syme);
583
584 if (++printed > print_entries || (int)syme->snap_count < count_filter)
585 continue;
586
587 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
588 sum_ksamples));
589
590 if (nr_counters == 1 || !display_weighted)
591 printf("%20.2f ", syme->weight);
592 else
593 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
594
595 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
596 if (verbose)
597 printf(" %016llx", sym->start);
598 printf(" %-*.*s", sym_width, sym_width, sym->name);
599 printf(" %-*.*s\n", dso_width, dso_width,
600 dso_width >= syme->map->dso->long_name_len ?
601 syme->map->dso->long_name :
602 syme->map->dso->short_name);
603 }
604 }
605
606 static void prompt_integer(int *target, const char *msg)
607 {
608 char *buf = malloc(0), *p;
609 size_t dummy = 0;
610 int tmp;
611
612 fprintf(stdout, "\n%s: ", msg);
613 if (getline(&buf, &dummy, stdin) < 0)
614 return;
615
616 p = strchr(buf, '\n');
617 if (p)
618 *p = 0;
619
620 p = buf;
621 while(*p) {
622 if (!isdigit(*p))
623 goto out_free;
624 p++;
625 }
626 tmp = strtoul(buf, NULL, 10);
627 *target = tmp;
628 out_free:
629 free(buf);
630 }
631
632 static void prompt_percent(int *target, const char *msg)
633 {
634 int tmp = 0;
635
636 prompt_integer(&tmp, msg);
637 if (tmp >= 0 && tmp <= 100)
638 *target = tmp;
639 }
640
641 static void prompt_symbol(struct sym_entry **target, const char *msg)
642 {
643 char *buf = malloc(0), *p;
644 struct sym_entry *syme = *target, *n, *found = NULL;
645 size_t dummy = 0;
646
647 /* zero counters of active symbol */
648 if (syme) {
649 pthread_mutex_lock(&syme->src->lock);
650 __zero_source_counters(syme);
651 *target = NULL;
652 pthread_mutex_unlock(&syme->src->lock);
653 }
654
655 fprintf(stdout, "\n%s: ", msg);
656 if (getline(&buf, &dummy, stdin) < 0)
657 goto out_free;
658
659 p = strchr(buf, '\n');
660 if (p)
661 *p = 0;
662
663 pthread_mutex_lock(&active_symbols_lock);
664 syme = list_entry(active_symbols.next, struct sym_entry, node);
665 pthread_mutex_unlock(&active_symbols_lock);
666
667 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
668 struct symbol *sym = sym_entry__symbol(syme);
669
670 if (!strcmp(buf, sym->name)) {
671 found = syme;
672 break;
673 }
674 }
675
676 if (!found) {
677 fprintf(stderr, "Sorry, %s is not active.\n", buf);
678 sleep(1);
679 return;
680 } else
681 parse_source(found);
682
683 out_free:
684 free(buf);
685 }
686
687 static void print_mapped_keys(void)
688 {
689 char *name = NULL;
690
691 if (sym_filter_entry) {
692 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
693 name = sym->name;
694 }
695
696 fprintf(stdout, "\nMapped keys:\n");
697 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
698 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
699
700 if (nr_counters > 1)
701 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter));
702
703 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
704
705 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
706 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
707 fprintf(stdout, "\t[S] stop annotation.\n");
708
709 if (nr_counters > 1)
710 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
711
712 fprintf(stdout,
713 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
714 hide_kernel_symbols ? "yes" : "no");
715 fprintf(stdout,
716 "\t[U] hide user symbols. \t(%s)\n",
717 hide_user_symbols ? "yes" : "no");
718 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
719 fprintf(stdout, "\t[qQ] quit.\n");
720 }
721
722 static int key_mapped(int c)
723 {
724 switch (c) {
725 case 'd':
726 case 'e':
727 case 'f':
728 case 'z':
729 case 'q':
730 case 'Q':
731 case 'K':
732 case 'U':
733 case 'F':
734 case 's':
735 case 'S':
736 return 1;
737 case 'E':
738 case 'w':
739 return nr_counters > 1 ? 1 : 0;
740 default:
741 break;
742 }
743
744 return 0;
745 }
746
747 static void handle_keypress(int c)
748 {
749 if (!key_mapped(c)) {
750 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
751 struct termios tc, save;
752
753 print_mapped_keys();
754 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
755 fflush(stdout);
756
757 tcgetattr(0, &save);
758 tc = save;
759 tc.c_lflag &= ~(ICANON | ECHO);
760 tc.c_cc[VMIN] = 0;
761 tc.c_cc[VTIME] = 0;
762 tcsetattr(0, TCSANOW, &tc);
763
764 poll(&stdin_poll, 1, -1);
765 c = getc(stdin);
766
767 tcsetattr(0, TCSAFLUSH, &save);
768 if (!key_mapped(c))
769 return;
770 }
771
772 switch (c) {
773 case 'd':
774 prompt_integer(&delay_secs, "Enter display delay");
775 if (delay_secs < 1)
776 delay_secs = 1;
777 break;
778 case 'e':
779 prompt_integer(&print_entries, "Enter display entries (lines)");
780 if (print_entries == 0) {
781 sig_winch_handler(SIGWINCH);
782 signal(SIGWINCH, sig_winch_handler);
783 } else
784 signal(SIGWINCH, SIG_DFL);
785 break;
786 case 'E':
787 if (nr_counters > 1) {
788 int i;
789
790 fprintf(stderr, "\nAvailable events:");
791 for (i = 0; i < nr_counters; i++)
792 fprintf(stderr, "\n\t%d %s", i, event_name(i));
793
794 prompt_integer(&sym_counter, "Enter details event counter");
795
796 if (sym_counter >= nr_counters) {
797 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
798 sym_counter = 0;
799 sleep(1);
800 }
801 } else sym_counter = 0;
802 break;
803 case 'f':
804 prompt_integer(&count_filter, "Enter display event count filter");
805 break;
806 case 'F':
807 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
808 break;
809 case 'K':
810 hide_kernel_symbols = !hide_kernel_symbols;
811 break;
812 case 'q':
813 case 'Q':
814 printf("exiting.\n");
815 if (dump_symtab)
816 dsos__fprintf(stderr);
817 exit(0);
818 case 's':
819 prompt_symbol(&sym_filter_entry, "Enter details symbol");
820 break;
821 case 'S':
822 if (!sym_filter_entry)
823 break;
824 else {
825 struct sym_entry *syme = sym_filter_entry;
826
827 pthread_mutex_lock(&syme->src->lock);
828 sym_filter_entry = NULL;
829 __zero_source_counters(syme);
830 pthread_mutex_unlock(&syme->src->lock);
831 }
832 break;
833 case 'U':
834 hide_user_symbols = !hide_user_symbols;
835 break;
836 case 'w':
837 display_weighted = ~display_weighted;
838 break;
839 case 'z':
840 zero = ~zero;
841 break;
842 default:
843 break;
844 }
845 }
846
847 static void *display_thread(void *arg __used)
848 {
849 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
850 struct termios tc, save;
851 int delay_msecs, c;
852
853 tcgetattr(0, &save);
854 tc = save;
855 tc.c_lflag &= ~(ICANON | ECHO);
856 tc.c_cc[VMIN] = 0;
857 tc.c_cc[VTIME] = 0;
858
859 repeat:
860 delay_msecs = delay_secs * 1000;
861 tcsetattr(0, TCSANOW, &tc);
862 /* trash return*/
863 getc(stdin);
864
865 do {
866 print_sym_table();
867 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
868
869 c = getc(stdin);
870 tcsetattr(0, TCSAFLUSH, &save);
871
872 handle_keypress(c);
873 goto repeat;
874
875 return NULL;
876 }
877
878 /* Tag samples to be skipped. */
879 static const char *skip_symbols[] = {
880 "default_idle",
881 "cpu_idle",
882 "enter_idle",
883 "exit_idle",
884 "mwait_idle",
885 "mwait_idle_with_hints",
886 "poll_idle",
887 "ppc64_runlatch_off",
888 "pseries_dedicated_idle_sleep",
889 NULL
890 };
891
892 static int symbol_filter(struct map *map, struct symbol *sym)
893 {
894 struct sym_entry *syme;
895 const char *name = sym->name;
896 int i;
897
898 /*
899 * ppc64 uses function descriptors and appends a '.' to the
900 * start of every instruction address. Remove it.
901 */
902 if (name[0] == '.')
903 name++;
904
905 if (!strcmp(name, "_text") ||
906 !strcmp(name, "_etext") ||
907 !strcmp(name, "_sinittext") ||
908 !strncmp("init_module", name, 11) ||
909 !strncmp("cleanup_module", name, 14) ||
910 strstr(name, "_text_start") ||
911 strstr(name, "_text_end"))
912 return 1;
913
914 syme = symbol__priv(sym);
915 syme->map = map;
916 syme->src = NULL;
917
918 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
919 /* schedule initial sym_filter_entry setup */
920 sym_filter_entry_sched = syme;
921 sym_filter = NULL;
922 }
923
924 for (i = 0; skip_symbols[i]; i++) {
925 if (!strcmp(skip_symbols[i], name)) {
926 syme->skip = 1;
927 break;
928 }
929 }
930
931 if (!syme->skip)
932 syme->name_len = strlen(sym->name);
933
934 return 0;
935 }
936
937 static void event__process_sample(const event_t *self,
938 struct perf_session *session, int counter)
939 {
940 u64 ip = self->ip.ip;
941 struct sym_entry *syme;
942 struct addr_location al;
943 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
944
945 ++samples;
946
947 switch (origin) {
948 case PERF_RECORD_MISC_USER:
949 ++userspace_samples;
950 if (hide_user_symbols)
951 return;
952 break;
953 case PERF_RECORD_MISC_KERNEL:
954 if (hide_kernel_symbols)
955 return;
956 break;
957 default:
958 return;
959 }
960
961 if (self->header.misc & PERF_RECORD_MISC_EXACT)
962 exact_samples++;
963
964 if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
965 al.filtered)
966 return;
967
968 if (al.sym == NULL) {
969 /*
970 * As we do lazy loading of symtabs we only will know if the
971 * specified vmlinux file is invalid when we actually have a
972 * hit in kernel space and then try to load it. So if we get
973 * here and there are _no_ symbols in the DSO backing the
974 * kernel map, bail out.
975 *
976 * We may never get here, for instance, if we use -K/
977 * --hide-kernel-symbols, even if the user specifies an
978 * invalid --vmlinux ;-)
979 */
980 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
981 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
982 pr_err("The %s file can't be used\n",
983 symbol_conf.vmlinux_name);
984 exit(1);
985 }
986
987 return;
988 }
989
990 /* let's see, whether we need to install initial sym_filter_entry */
991 if (sym_filter_entry_sched) {
992 sym_filter_entry = sym_filter_entry_sched;
993 sym_filter_entry_sched = NULL;
994 parse_source(sym_filter_entry);
995 }
996
997 syme = symbol__priv(al.sym);
998 if (!syme->skip) {
999 syme->count[counter]++;
1000 syme->origin = origin;
1001 record_precise_ip(syme, counter, ip);
1002 pthread_mutex_lock(&active_symbols_lock);
1003 if (list_empty(&syme->node) || !syme->node.next)
1004 __list_insert_active_sym(syme);
1005 pthread_mutex_unlock(&active_symbols_lock);
1006 }
1007 }
1008
1009 static int event__process(event_t *event, struct perf_session *session)
1010 {
1011 switch (event->header.type) {
1012 case PERF_RECORD_COMM:
1013 event__process_comm(event, session);
1014 break;
1015 case PERF_RECORD_MMAP:
1016 event__process_mmap(event, session);
1017 break;
1018 case PERF_RECORD_FORK:
1019 case PERF_RECORD_EXIT:
1020 event__process_task(event, session);
1021 break;
1022 default:
1023 break;
1024 }
1025
1026 return 0;
1027 }
1028
1029 struct mmap_data {
1030 int counter;
1031 void *base;
1032 int mask;
1033 unsigned int prev;
1034 };
1035
1036 static unsigned int mmap_read_head(struct mmap_data *md)
1037 {
1038 struct perf_event_mmap_page *pc = md->base;
1039 int head;
1040
1041 head = pc->data_head;
1042 rmb();
1043
1044 return head;
1045 }
1046
1047 static void perf_session__mmap_read_counter(struct perf_session *self,
1048 struct mmap_data *md)
1049 {
1050 unsigned int head = mmap_read_head(md);
1051 unsigned int old = md->prev;
1052 unsigned char *data = md->base + page_size;
1053 int diff;
1054
1055 /*
1056 * If we're further behind than half the buffer, there's a chance
1057 * the writer will bite our tail and mess up the samples under us.
1058 *
1059 * If we somehow ended up ahead of the head, we got messed up.
1060 *
1061 * In either case, truncate and restart at head.
1062 */
1063 diff = head - old;
1064 if (diff > md->mask / 2 || diff < 0) {
1065 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1066
1067 /*
1068 * head points to a known good entry, start there.
1069 */
1070 old = head;
1071 }
1072
1073 for (; old != head;) {
1074 event_t *event = (event_t *)&data[old & md->mask];
1075
1076 event_t event_copy;
1077
1078 size_t size = event->header.size;
1079
1080 /*
1081 * Event straddles the mmap boundary -- header should always
1082 * be inside due to u64 alignment of output.
1083 */
1084 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1085 unsigned int offset = old;
1086 unsigned int len = min(sizeof(*event), size), cpy;
1087 void *dst = &event_copy;
1088
1089 do {
1090 cpy = min(md->mask + 1 - (offset & md->mask), len);
1091 memcpy(dst, &data[offset & md->mask], cpy);
1092 offset += cpy;
1093 dst += cpy;
1094 len -= cpy;
1095 } while (len);
1096
1097 event = &event_copy;
1098 }
1099
1100 if (event->header.type == PERF_RECORD_SAMPLE)
1101 event__process_sample(event, self, md->counter);
1102 else
1103 event__process(event, self);
1104 old += size;
1105 }
1106
1107 md->prev = old;
1108 }
1109
1110 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1111 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1112
1113 static void perf_session__mmap_read(struct perf_session *self)
1114 {
1115 int i, counter;
1116
1117 for (i = 0; i < nr_cpus; i++) {
1118 for (counter = 0; counter < nr_counters; counter++)
1119 perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1120 }
1121 }
1122
1123 int nr_poll;
1124 int group_fd;
1125
1126 static void start_counter(int i, int counter)
1127 {
1128 struct perf_event_attr *attr;
1129 int cpu;
1130
1131 cpu = profile_cpu;
1132 if (target_pid == -1 && profile_cpu == -1)
1133 cpu = cpumap[i];
1134
1135 attr = attrs + counter;
1136
1137 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1138
1139 if (freq) {
1140 attr->sample_type |= PERF_SAMPLE_PERIOD;
1141 attr->freq = 1;
1142 attr->sample_freq = freq;
1143 }
1144
1145 attr->inherit = (cpu < 0) && inherit;
1146 attr->mmap = 1;
1147
1148 try_again:
1149 fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1150
1151 if (fd[i][counter] < 0) {
1152 int err = errno;
1153
1154 if (err == EPERM || err == EACCES)
1155 die("No permission - are you root?\n");
1156 /*
1157 * If it's cycles then fall back to hrtimer
1158 * based cpu-clock-tick sw counter, which
1159 * is always available even if no PMU support:
1160 */
1161 if (attr->type == PERF_TYPE_HARDWARE
1162 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1163
1164 if (verbose)
1165 warning(" ... trying to fall back to cpu-clock-ticks\n");
1166
1167 attr->type = PERF_TYPE_SOFTWARE;
1168 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1169 goto try_again;
1170 }
1171 printf("\n");
1172 error("perfcounter syscall returned with %d (%s)\n",
1173 fd[i][counter], strerror(err));
1174 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1175 exit(-1);
1176 }
1177 assert(fd[i][counter] >= 0);
1178 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1179
1180 /*
1181 * First counter acts as the group leader:
1182 */
1183 if (group && group_fd == -1)
1184 group_fd = fd[i][counter];
1185
1186 event_array[nr_poll].fd = fd[i][counter];
1187 event_array[nr_poll].events = POLLIN;
1188 nr_poll++;
1189
1190 mmap_array[i][counter].counter = counter;
1191 mmap_array[i][counter].prev = 0;
1192 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1193 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1194 PROT_READ, MAP_SHARED, fd[i][counter], 0);
1195 if (mmap_array[i][counter].base == MAP_FAILED)
1196 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1197 }
1198
1199 static int __cmd_top(void)
1200 {
1201 pthread_t thread;
1202 int i, counter;
1203 int ret;
1204 /*
1205 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1206 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1207 */
1208 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1209 if (session == NULL)
1210 return -ENOMEM;
1211
1212 if (target_pid != -1)
1213 event__synthesize_thread(target_pid, event__process, session);
1214 else
1215 event__synthesize_threads(event__process, session);
1216
1217 for (i = 0; i < nr_cpus; i++) {
1218 group_fd = -1;
1219 for (counter = 0; counter < nr_counters; counter++)
1220 start_counter(i, counter);
1221 }
1222
1223 /* Wait for a minimal set of events before starting the snapshot */
1224 poll(event_array, nr_poll, 100);
1225
1226 perf_session__mmap_read(session);
1227
1228 if (pthread_create(&thread, NULL, display_thread, NULL)) {
1229 printf("Could not create display thread.\n");
1230 exit(-1);
1231 }
1232
1233 if (realtime_prio) {
1234 struct sched_param param;
1235
1236 param.sched_priority = realtime_prio;
1237 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1238 printf("Could not set realtime priority.\n");
1239 exit(-1);
1240 }
1241 }
1242
1243 while (1) {
1244 int hits = samples;
1245
1246 perf_session__mmap_read(session);
1247
1248 if (hits == samples)
1249 ret = poll(event_array, nr_poll, 100);
1250 }
1251
1252 return 0;
1253 }
1254
1255 static const char * const top_usage[] = {
1256 "perf top [<options>]",
1257 NULL
1258 };
1259
1260 static const struct option options[] = {
1261 OPT_CALLBACK('e', "event", NULL, "event",
1262 "event selector. use 'perf list' to list available events",
1263 parse_events),
1264 OPT_INTEGER('c', "count", &default_interval,
1265 "event period to sample"),
1266 OPT_INTEGER('p', "pid", &target_pid,
1267 "profile events on existing pid"),
1268 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1269 "system-wide collection from all CPUs"),
1270 OPT_INTEGER('C', "CPU", &profile_cpu,
1271 "CPU to profile on"),
1272 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1273 "file", "vmlinux pathname"),
1274 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1275 "hide kernel symbols"),
1276 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1277 "number of mmap data pages"),
1278 OPT_INTEGER('r', "realtime", &realtime_prio,
1279 "collect data with this RT SCHED_FIFO priority"),
1280 OPT_INTEGER('d', "delay", &delay_secs,
1281 "number of seconds to delay between refreshes"),
1282 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1283 "dump the symbol table used for profiling"),
1284 OPT_INTEGER('f', "count-filter", &count_filter,
1285 "only display functions with more events than this"),
1286 OPT_BOOLEAN('g', "group", &group,
1287 "put the counters into a counter group"),
1288 OPT_BOOLEAN('i', "inherit", &inherit,
1289 "child tasks inherit counters"),
1290 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1291 "symbol to annotate"),
1292 OPT_BOOLEAN('z', "zero", &zero,
1293 "zero history across updates"),
1294 OPT_INTEGER('F', "freq", &freq,
1295 "profile at this frequency"),
1296 OPT_INTEGER('E', "entries", &print_entries,
1297 "display this many functions"),
1298 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1299 "hide user symbols"),
1300 OPT_BOOLEAN('v', "verbose", &verbose,
1301 "be more verbose (show counter open errors, etc)"),
1302 OPT_END()
1303 };
1304
1305 int cmd_top(int argc, const char **argv, const char *prefix __used)
1306 {
1307 int counter;
1308
1309 page_size = sysconf(_SC_PAGE_SIZE);
1310
1311 argc = parse_options(argc, argv, options, top_usage, 0);
1312 if (argc)
1313 usage_with_options(top_usage, options);
1314
1315 /* CPU and PID are mutually exclusive */
1316 if (target_pid != -1 && profile_cpu != -1) {
1317 printf("WARNING: PID switch overriding CPU\n");
1318 sleep(1);
1319 profile_cpu = -1;
1320 }
1321
1322 if (!nr_counters)
1323 nr_counters = 1;
1324
1325 symbol_conf.priv_size = (sizeof(struct sym_entry) +
1326 (nr_counters + 1) * sizeof(unsigned long));
1327
1328 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1329 if (symbol__init() < 0)
1330 return -1;
1331
1332 if (delay_secs < 1)
1333 delay_secs = 1;
1334
1335 /*
1336 * User specified count overrides default frequency.
1337 */
1338 if (default_interval)
1339 freq = 0;
1340 else if (freq) {
1341 default_interval = freq;
1342 } else {
1343 fprintf(stderr, "frequency and count are zero, aborting\n");
1344 exit(EXIT_FAILURE);
1345 }
1346
1347 /*
1348 * Fill in the ones not specifically initialized via -c:
1349 */
1350 for (counter = 0; counter < nr_counters; counter++) {
1351 if (attrs[counter].sample_period)
1352 continue;
1353
1354 attrs[counter].sample_period = default_interval;
1355 }
1356
1357 if (target_pid != -1 || profile_cpu != -1)
1358 nr_cpus = 1;
1359 else
1360 nr_cpus = read_cpu_map();
1361
1362 get_term_dimensions(&winsize);
1363 if (print_entries == 0) {
1364 update_print_entries(&winsize);
1365 signal(SIGWINCH, sig_winch_handler);
1366 }
1367
1368 return __cmd_top();
1369 }
Linux kernel - Deliverables
This page took 0.064799 seconds and 5 git commands to generate.