projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf threads: Move thread_map to separate file
[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/evlist.h"
25 #include "util/evsel.h"
26 #include "util/session.h"
27 #include "util/symbol.h"
28 #include "util/thread.h"
29 #include "util/thread_map.h"
30 #include "util/util.h"
31 #include <linux/rbtree.h>
32 #include "util/parse-options.h"
33 #include "util/parse-events.h"
34 #include "util/cpumap.h"
35 #include "util/xyarray.h"
36
37 #include "util/debug.h"
38
39 #include <assert.h>
40 #include <fcntl.h>
41
42 #include <stdio.h>
43 #include <termios.h>
44 #include <unistd.h>
45 #include <inttypes.h>
46
47 #include <errno.h>
48 #include <time.h>
49 #include <sched.h>
50 #include <pthread.h>
51
52 #include <sys/syscall.h>
53 #include <sys/ioctl.h>
54 #include <sys/poll.h>
55 #include <sys/prctl.h>
56 #include <sys/wait.h>
57 #include <sys/uio.h>
58 #include <sys/mman.h>
59
60 #include <linux/unistd.h>
61 #include <linux/types.h>
62
63 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
64
65 struct perf_evlist *evsel_list;
66
67 static bool system_wide = false;
68
69 static int default_interval = 0;
70
71 static int count_filter = 5;
72 static int print_entries;
73
74 static int target_pid = -1;
75 static int target_tid = -1;
76 static struct thread_map *threads;
77 static bool inherit = false;
78 static struct cpu_map *cpus;
79 static int realtime_prio = 0;
80 static bool group = false;
81 static unsigned int page_size;
82 static unsigned int mmap_pages = 128;
83 static int freq = 1000; /* 1 KHz */
84
85 static int delay_secs = 2;
86 static bool zero = false;
87 static bool dump_symtab = false;
88
89 static bool hide_kernel_symbols = false;
90 static bool hide_user_symbols = false;
91 static struct winsize winsize;
92
93 /*
94 * Source
95 */
96
97 struct source_line {
98 u64 eip;
99 unsigned long count[MAX_COUNTERS];
100 char *line;
101 struct source_line *next;
102 };
103
104 static const char *sym_filter = NULL;
105 struct sym_entry *sym_filter_entry = NULL;
106 struct sym_entry *sym_filter_entry_sched = NULL;
107 static int sym_pcnt_filter = 5;
108 static int sym_counter = 0;
109 static struct perf_evsel *sym_evsel = NULL;
110 static int display_weighted = -1;
111 static const char *cpu_list;
112
113 /*
114 * Symbols
115 */
116
117 struct sym_entry_source {
118 struct source_line *source;
119 struct source_line *lines;
120 struct source_line **lines_tail;
121 pthread_mutex_t lock;
122 };
123
124 struct sym_entry {
125 struct rb_node rb_node;
126 struct list_head node;
127 unsigned long snap_count;
128 double weight;
129 int skip;
130 u16 name_len;
131 u8 origin;
132 struct map *map;
133 struct sym_entry_source *src;
134 unsigned long count[0];
135 };
136
137 /*
138 * Source functions
139 */
140
141 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
142 {
143 return ((void *)self) + symbol_conf.priv_size;
144 }
145
146 void get_term_dimensions(struct winsize *ws)
147 {
148 char *s = getenv("LINES");
149
150 if (s != NULL) {
151 ws->ws_row = atoi(s);
152 s = getenv("COLUMNS");
153 if (s != NULL) {
154 ws->ws_col = atoi(s);
155 if (ws->ws_row && ws->ws_col)
156 return;
157 }
158 }
159 #ifdef TIOCGWINSZ
160 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
161 ws->ws_row && ws->ws_col)
162 return;
163 #endif
164 ws->ws_row = 25;
165 ws->ws_col = 80;
166 }
167
168 static void update_print_entries(struct winsize *ws)
169 {
170 print_entries = ws->ws_row;
171
172 if (print_entries > 9)
173 print_entries -= 9;
174 }
175
176 static void sig_winch_handler(int sig __used)
177 {
178 get_term_dimensions(&winsize);
179 update_print_entries(&winsize);
180 }
181
182 static int parse_source(struct sym_entry *syme)
183 {
184 struct symbol *sym;
185 struct sym_entry_source *source;
186 struct map *map;
187 FILE *file;
188 char command[PATH_MAX*2];
189 const char *path;
190 u64 len;
191
192 if (!syme)
193 return -1;
194
195 sym = sym_entry__symbol(syme);
196 map = syme->map;
197
198 /*
199 * We can't annotate with just /proc/kallsyms
200 */
201 if (map->dso->origin == DSO__ORIG_KERNEL)
202 return -1;
203
204 if (syme->src == NULL) {
205 syme->src = zalloc(sizeof(*source));
206 if (syme->src == NULL)
207 return -1;
208 pthread_mutex_init(&syme->src->lock, NULL);
209 }
210
211 source = syme->src;
212
213 if (source->lines) {
214 pthread_mutex_lock(&source->lock);
215 goto out_assign;
216 }
217 path = map->dso->long_name;
218
219 len = sym->end - sym->start;
220
221 sprintf(command,
222 "objdump --start-address=%#0*" PRIx64 " --stop-address=%#0*" PRIx64 " -dS %s",
223 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
224 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
225
226 file = popen(command, "r");
227 if (!file)
228 return -1;
229
230 pthread_mutex_lock(&source->lock);
231 source->lines_tail = &source->lines;
232 while (!feof(file)) {
233 struct source_line *src;
234 size_t dummy = 0;
235 char *c, *sep;
236
237 src = malloc(sizeof(struct source_line));
238 assert(src != NULL);
239 memset(src, 0, sizeof(struct source_line));
240
241 if (getline(&src->line, &dummy, file) < 0)
242 break;
243 if (!src->line)
244 break;
245
246 c = strchr(src->line, '\n');
247 if (c)
248 *c = 0;
249
250 src->next = NULL;
251 *source->lines_tail = src;
252 source->lines_tail = &src->next;
253
254 src->eip = strtoull(src->line, &sep, 16);
255 if (*sep == ':')
256 src->eip = map__objdump_2ip(map, src->eip);
257 else /* this line has no ip info (e.g. source line) */
258 src->eip = 0;
259 }
260 pclose(file);
261 out_assign:
262 sym_filter_entry = syme;
263 pthread_mutex_unlock(&source->lock);
264 return 0;
265 }
266
267 static void __zero_source_counters(struct sym_entry *syme)
268 {
269 int i;
270 struct source_line *line;
271
272 line = syme->src->lines;
273 while (line) {
274 for (i = 0; i < evsel_list->nr_entries; i++)
275 line->count[i] = 0;
276 line = line->next;
277 }
278 }
279
280 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
281 {
282 struct source_line *line;
283
284 if (syme != sym_filter_entry)
285 return;
286
287 if (pthread_mutex_trylock(&syme->src->lock))
288 return;
289
290 if (syme->src == NULL || syme->src->source == NULL)
291 goto out_unlock;
292
293 for (line = syme->src->lines; line; line = line->next) {
294 /* skip lines without IP info */
295 if (line->eip == 0)
296 continue;
297 if (line->eip == ip) {
298 line->count[counter]++;
299 break;
300 }
301 if (line->eip > ip)
302 break;
303 }
304 out_unlock:
305 pthread_mutex_unlock(&syme->src->lock);
306 }
307
308 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
309
310 static void lookup_sym_source(struct sym_entry *syme)
311 {
312 struct symbol *symbol = sym_entry__symbol(syme);
313 struct source_line *line;
314 char pattern[PATTERN_LEN + 1];
315
316 sprintf(pattern, "%0*" PRIx64 " <", BITS_PER_LONG / 4,
317 map__rip_2objdump(syme->map, symbol->start));
318
319 pthread_mutex_lock(&syme->src->lock);
320 for (line = syme->src->lines; line; line = line->next) {
321 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
322 syme->src->source = line;
323 break;
324 }
325 }
326 pthread_mutex_unlock(&syme->src->lock);
327 }
328
329 static void show_lines(struct source_line *queue, int count, int total)
330 {
331 int i;
332 struct source_line *line;
333
334 line = queue;
335 for (i = 0; i < count; i++) {
336 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
337
338 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
339 line = line->next;
340 }
341 }
342
343 #define TRACE_COUNT 3
344
345 static void show_details(struct sym_entry *syme)
346 {
347 struct symbol *symbol;
348 struct source_line *line;
349 struct source_line *line_queue = NULL;
350 int displayed = 0;
351 int line_queue_count = 0, total = 0, more = 0;
352
353 if (!syme)
354 return;
355
356 if (!syme->src->source)
357 lookup_sym_source(syme);
358
359 if (!syme->src->source)
360 return;
361
362 symbol = sym_entry__symbol(syme);
363 printf("Showing %s for %s\n", event_name(sym_evsel), symbol->name);
364 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
365
366 pthread_mutex_lock(&syme->src->lock);
367 line = syme->src->source;
368 while (line) {
369 total += line->count[sym_counter];
370 line = line->next;
371 }
372
373 line = syme->src->source;
374 while (line) {
375 float pcnt = 0.0;
376
377 if (!line_queue_count)
378 line_queue = line;
379 line_queue_count++;
380
381 if (line->count[sym_counter])
382 pcnt = 100.0 * line->count[sym_counter] / (float)total;
383 if (pcnt >= (float)sym_pcnt_filter) {
384 if (displayed <= print_entries)
385 show_lines(line_queue, line_queue_count, total);
386 else more++;
387 displayed += line_queue_count;
388 line_queue_count = 0;
389 line_queue = NULL;
390 } else if (line_queue_count > TRACE_COUNT) {
391 line_queue = line_queue->next;
392 line_queue_count--;
393 }
394
395 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
396 line = line->next;
397 }
398 pthread_mutex_unlock(&syme->src->lock);
399 if (more)
400 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
401 }
402
403 /*
404 * Symbols will be added here in event__process_sample and will get out
405 * after decayed.
406 */
407 static LIST_HEAD(active_symbols);
408 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
409
410 /*
411 * Ordering weight: count-1 * count-2 * ... / count-n
412 */
413 static double sym_weight(const struct sym_entry *sym)
414 {
415 double weight = sym->snap_count;
416 int counter;
417
418 if (!display_weighted)
419 return weight;
420
421 for (counter = 1; counter < evsel_list->nr_entries - 1; counter++)
422 weight *= sym->count[counter];
423
424 weight /= (sym->count[counter] + 1);
425
426 return weight;
427 }
428
429 static long samples;
430 static long kernel_samples, us_samples;
431 static long exact_samples;
432 static long guest_us_samples, guest_kernel_samples;
433 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
434
435 static void __list_insert_active_sym(struct sym_entry *syme)
436 {
437 list_add(&syme->node, &active_symbols);
438 }
439
440 static void list_remove_active_sym(struct sym_entry *syme)
441 {
442 pthread_mutex_lock(&active_symbols_lock);
443 list_del_init(&syme->node);
444 pthread_mutex_unlock(&active_symbols_lock);
445 }
446
447 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
448 {
449 struct rb_node **p = &tree->rb_node;
450 struct rb_node *parent = NULL;
451 struct sym_entry *iter;
452
453 while (*p != NULL) {
454 parent = *p;
455 iter = rb_entry(parent, struct sym_entry, rb_node);
456
457 if (se->weight > iter->weight)
458 p = &(*p)->rb_left;
459 else
460 p = &(*p)->rb_right;
461 }
462
463 rb_link_node(&se->rb_node, parent, p);
464 rb_insert_color(&se->rb_node, tree);
465 }
466
467 static void print_sym_table(void)
468 {
469 int printed = 0, j;
470 struct perf_evsel *counter;
471 int snap = !display_weighted ? sym_counter : 0;
472 float samples_per_sec = samples/delay_secs;
473 float ksamples_per_sec = kernel_samples/delay_secs;
474 float us_samples_per_sec = (us_samples)/delay_secs;
475 float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs;
476 float guest_us_samples_per_sec = (guest_us_samples)/delay_secs;
477 float esamples_percent = (100.0*exact_samples)/samples;
478 float sum_ksamples = 0.0;
479 struct sym_entry *syme, *n;
480 struct rb_root tmp = RB_ROOT;
481 struct rb_node *nd;
482 int sym_width = 0, dso_width = 0, dso_short_width = 0;
483 const int win_width = winsize.ws_col - 1;
484
485 samples = us_samples = kernel_samples = exact_samples = 0;
486 guest_kernel_samples = guest_us_samples = 0;
487
488 /* Sort the active symbols */
489 pthread_mutex_lock(&active_symbols_lock);
490 syme = list_entry(active_symbols.next, struct sym_entry, node);
491 pthread_mutex_unlock(&active_symbols_lock);
492
493 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
494 syme->snap_count = syme->count[snap];
495 if (syme->snap_count != 0) {
496
497 if ((hide_user_symbols &&
498 syme->origin == PERF_RECORD_MISC_USER) ||
499 (hide_kernel_symbols &&
500 syme->origin == PERF_RECORD_MISC_KERNEL)) {
501 list_remove_active_sym(syme);
502 continue;
503 }
504 syme->weight = sym_weight(syme);
505 rb_insert_active_sym(&tmp, syme);
506 sum_ksamples += syme->snap_count;
507
508 for (j = 0; j < evsel_list->nr_entries; j++)
509 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
510 } else
511 list_remove_active_sym(syme);
512 }
513
514 puts(CONSOLE_CLEAR);
515
516 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
517 if (!perf_guest) {
518 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%"
519 " exact: %4.1f%% [",
520 samples_per_sec,
521 100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) /
522 samples_per_sec)),
523 esamples_percent);
524 } else {
525 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%"
526 " guest kernel:%4.1f%% guest us:%4.1f%%"
527 " exact: %4.1f%% [",
528 samples_per_sec,
529 100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) /
530 samples_per_sec)),
531 100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) /
532 samples_per_sec)),
533 100.0 - (100.0 * ((samples_per_sec -
534 guest_kernel_samples_per_sec) /
535 samples_per_sec)),
536 100.0 - (100.0 * ((samples_per_sec -
537 guest_us_samples_per_sec) /
538 samples_per_sec)),
539 esamples_percent);
540 }
541
542 if (evsel_list->nr_entries == 1 || !display_weighted) {
543 struct perf_evsel *first;
544 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
545 printf("%" PRIu64, (uint64_t)first->attr.sample_period);
546 if (freq)
547 printf("Hz ");
548 else
549 printf(" ");
550 }
551
552 if (!display_weighted)
553 printf("%s", event_name(sym_evsel));
554 else list_for_each_entry(counter, &evsel_list->entries, node) {
555 if (counter->idx)
556 printf("/");
557
558 printf("%s", event_name(counter));
559 }
560
561 printf( "], ");
562
563 if (target_pid != -1)
564 printf(" (target_pid: %d", target_pid);
565 else if (target_tid != -1)
566 printf(" (target_tid: %d", target_tid);
567 else
568 printf(" (all");
569
570 if (cpu_list)
571 printf(", CPU%s: %s)\n", cpus->nr > 1 ? "s" : "", cpu_list);
572 else {
573 if (target_tid != -1)
574 printf(")\n");
575 else
576 printf(", %d CPU%s)\n", cpus->nr, cpus->nr > 1 ? "s" : "");
577 }
578
579 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
580
581 if (sym_filter_entry) {
582 show_details(sym_filter_entry);
583 return;
584 }
585
586 /*
587 * Find the longest symbol name that will be displayed
588 */
589 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
590 syme = rb_entry(nd, struct sym_entry, rb_node);
591 if (++printed > print_entries ||
592 (int)syme->snap_count < count_filter)
593 continue;
594
595 if (syme->map->dso->long_name_len > dso_width)
596 dso_width = syme->map->dso->long_name_len;
597
598 if (syme->map->dso->short_name_len > dso_short_width)
599 dso_short_width = syme->map->dso->short_name_len;
600
601 if (syme->name_len > sym_width)
602 sym_width = syme->name_len;
603 }
604
605 printed = 0;
606
607 if (sym_width + dso_width > winsize.ws_col - 29) {
608 dso_width = dso_short_width;
609 if (sym_width + dso_width > winsize.ws_col - 29)
610 sym_width = winsize.ws_col - dso_width - 29;
611 }
612 putchar('\n');
613 if (evsel_list->nr_entries == 1)
614 printf(" samples pcnt");
615 else
616 printf(" weight samples pcnt");
617
618 if (verbose)
619 printf(" RIP ");
620 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
621 printf(" %s _______ _____",
622 evsel_list->nr_entries == 1 ? " " : "______");
623 if (verbose)
624 printf(" ________________");
625 printf(" %-*.*s", sym_width, sym_width, graph_line);
626 printf(" %-*.*s", dso_width, dso_width, graph_line);
627 puts("\n");
628
629 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
630 struct symbol *sym;
631 double pcnt;
632
633 syme = rb_entry(nd, struct sym_entry, rb_node);
634 sym = sym_entry__symbol(syme);
635 if (++printed > print_entries || (int)syme->snap_count < count_filter)
636 continue;
637
638 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
639 sum_ksamples));
640
641 if (evsel_list->nr_entries == 1 || !display_weighted)
642 printf("%20.2f ", syme->weight);
643 else
644 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
645
646 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
647 if (verbose)
648 printf(" %016" PRIx64, sym->start);
649 printf(" %-*.*s", sym_width, sym_width, sym->name);
650 printf(" %-*.*s\n", dso_width, dso_width,
651 dso_width >= syme->map->dso->long_name_len ?
652 syme->map->dso->long_name :
653 syme->map->dso->short_name);
654 }
655 }
656
657 static void prompt_integer(int *target, const char *msg)
658 {
659 char *buf = malloc(0), *p;
660 size_t dummy = 0;
661 int tmp;
662
663 fprintf(stdout, "\n%s: ", msg);
664 if (getline(&buf, &dummy, stdin) < 0)
665 return;
666
667 p = strchr(buf, '\n');
668 if (p)
669 *p = 0;
670
671 p = buf;
672 while(*p) {
673 if (!isdigit(*p))
674 goto out_free;
675 p++;
676 }
677 tmp = strtoul(buf, NULL, 10);
678 *target = tmp;
679 out_free:
680 free(buf);
681 }
682
683 static void prompt_percent(int *target, const char *msg)
684 {
685 int tmp = 0;
686
687 prompt_integer(&tmp, msg);
688 if (tmp >= 0 && tmp <= 100)
689 *target = tmp;
690 }
691
692 static void prompt_symbol(struct sym_entry **target, const char *msg)
693 {
694 char *buf = malloc(0), *p;
695 struct sym_entry *syme = *target, *n, *found = NULL;
696 size_t dummy = 0;
697
698 /* zero counters of active symbol */
699 if (syme) {
700 pthread_mutex_lock(&syme->src->lock);
701 __zero_source_counters(syme);
702 *target = NULL;
703 pthread_mutex_unlock(&syme->src->lock);
704 }
705
706 fprintf(stdout, "\n%s: ", msg);
707 if (getline(&buf, &dummy, stdin) < 0)
708 goto out_free;
709
710 p = strchr(buf, '\n');
711 if (p)
712 *p = 0;
713
714 pthread_mutex_lock(&active_symbols_lock);
715 syme = list_entry(active_symbols.next, struct sym_entry, node);
716 pthread_mutex_unlock(&active_symbols_lock);
717
718 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
719 struct symbol *sym = sym_entry__symbol(syme);
720
721 if (!strcmp(buf, sym->name)) {
722 found = syme;
723 break;
724 }
725 }
726
727 if (!found) {
728 fprintf(stderr, "Sorry, %s is not active.\n", buf);
729 sleep(1);
730 return;
731 } else
732 parse_source(found);
733
734 out_free:
735 free(buf);
736 }
737
738 static void print_mapped_keys(void)
739 {
740 char *name = NULL;
741
742 if (sym_filter_entry) {
743 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
744 name = sym->name;
745 }
746
747 fprintf(stdout, "\nMapped keys:\n");
748 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
749 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
750
751 if (evsel_list->nr_entries > 1)
752 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel));
753
754 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
755
756 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
757 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
758 fprintf(stdout, "\t[S] stop annotation.\n");
759
760 if (evsel_list->nr_entries > 1)
761 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
762
763 fprintf(stdout,
764 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
765 hide_kernel_symbols ? "yes" : "no");
766 fprintf(stdout,
767 "\t[U] hide user symbols. \t(%s)\n",
768 hide_user_symbols ? "yes" : "no");
769 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
770 fprintf(stdout, "\t[qQ] quit.\n");
771 }
772
773 static int key_mapped(int c)
774 {
775 switch (c) {
776 case 'd':
777 case 'e':
778 case 'f':
779 case 'z':
780 case 'q':
781 case 'Q':
782 case 'K':
783 case 'U':
784 case 'F':
785 case 's':
786 case 'S':
787 return 1;
788 case 'E':
789 case 'w':
790 return evsel_list->nr_entries > 1 ? 1 : 0;
791 default:
792 break;
793 }
794
795 return 0;
796 }
797
798 static void handle_keypress(struct perf_session *session, int c)
799 {
800 if (!key_mapped(c)) {
801 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
802 struct termios tc, save;
803
804 print_mapped_keys();
805 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
806 fflush(stdout);
807
808 tcgetattr(0, &save);
809 tc = save;
810 tc.c_lflag &= ~(ICANON | ECHO);
811 tc.c_cc[VMIN] = 0;
812 tc.c_cc[VTIME] = 0;
813 tcsetattr(0, TCSANOW, &tc);
814
815 poll(&stdin_poll, 1, -1);
816 c = getc(stdin);
817
818 tcsetattr(0, TCSAFLUSH, &save);
819 if (!key_mapped(c))
820 return;
821 }
822
823 switch (c) {
824 case 'd':
825 prompt_integer(&delay_secs, "Enter display delay");
826 if (delay_secs < 1)
827 delay_secs = 1;
828 break;
829 case 'e':
830 prompt_integer(&print_entries, "Enter display entries (lines)");
831 if (print_entries == 0) {
832 sig_winch_handler(SIGWINCH);
833 signal(SIGWINCH, sig_winch_handler);
834 } else
835 signal(SIGWINCH, SIG_DFL);
836 break;
837 case 'E':
838 if (evsel_list->nr_entries > 1) {
839 fprintf(stderr, "\nAvailable events:");
840
841 list_for_each_entry(sym_evsel, &evsel_list->entries, node)
842 fprintf(stderr, "\n\t%d %s", sym_evsel->idx, event_name(sym_evsel));
843
844 prompt_integer(&sym_counter, "Enter details event counter");
845
846 if (sym_counter >= evsel_list->nr_entries) {
847 sym_evsel = list_entry(evsel_list->entries.next, struct perf_evsel, node);
848 sym_counter = 0;
849 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(sym_evsel));
850 sleep(1);
851 break;
852 }
853 list_for_each_entry(sym_evsel, &evsel_list->entries, node)
854 if (sym_evsel->idx == sym_counter)
855 break;
856 } else sym_counter = 0;
857 break;
858 case 'f':
859 prompt_integer(&count_filter, "Enter display event count filter");
860 break;
861 case 'F':
862 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
863 break;
864 case 'K':
865 hide_kernel_symbols = !hide_kernel_symbols;
866 break;
867 case 'q':
868 case 'Q':
869 printf("exiting.\n");
870 if (dump_symtab)
871 perf_session__fprintf_dsos(session, stderr);
872 exit(0);
873 case 's':
874 prompt_symbol(&sym_filter_entry, "Enter details symbol");
875 break;
876 case 'S':
877 if (!sym_filter_entry)
878 break;
879 else {
880 struct sym_entry *syme = sym_filter_entry;
881
882 pthread_mutex_lock(&syme->src->lock);
883 sym_filter_entry = NULL;
884 __zero_source_counters(syme);
885 pthread_mutex_unlock(&syme->src->lock);
886 }
887 break;
888 case 'U':
889 hide_user_symbols = !hide_user_symbols;
890 break;
891 case 'w':
892 display_weighted = ~display_weighted;
893 break;
894 case 'z':
895 zero = !zero;
896 break;
897 default:
898 break;
899 }
900 }
901
902 static void *display_thread(void *arg __used)
903 {
904 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
905 struct termios tc, save;
906 int delay_msecs, c;
907 struct perf_session *session = (struct perf_session *) arg;
908
909 tcgetattr(0, &save);
910 tc = save;
911 tc.c_lflag &= ~(ICANON | ECHO);
912 tc.c_cc[VMIN] = 0;
913 tc.c_cc[VTIME] = 0;
914
915 repeat:
916 delay_msecs = delay_secs * 1000;
917 tcsetattr(0, TCSANOW, &tc);
918 /* trash return*/
919 getc(stdin);
920
921 do {
922 print_sym_table();
923 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
924
925 c = getc(stdin);
926 tcsetattr(0, TCSAFLUSH, &save);
927
928 handle_keypress(session, c);
929 goto repeat;
930
931 return NULL;
932 }
933
934 /* Tag samples to be skipped. */
935 static const char *skip_symbols[] = {
936 "default_idle",
937 "native_safe_halt",
938 "cpu_idle",
939 "enter_idle",
940 "exit_idle",
941 "mwait_idle",
942 "mwait_idle_with_hints",
943 "poll_idle",
944 "ppc64_runlatch_off",
945 "pseries_dedicated_idle_sleep",
946 NULL
947 };
948
949 static int symbol_filter(struct map *map, struct symbol *sym)
950 {
951 struct sym_entry *syme;
952 const char *name = sym->name;
953 int i;
954
955 /*
956 * ppc64 uses function descriptors and appends a '.' to the
957 * start of every instruction address. Remove it.
958 */
959 if (name[0] == '.')
960 name++;
961
962 if (!strcmp(name, "_text") ||
963 !strcmp(name, "_etext") ||
964 !strcmp(name, "_sinittext") ||
965 !strncmp("init_module", name, 11) ||
966 !strncmp("cleanup_module", name, 14) ||
967 strstr(name, "_text_start") ||
968 strstr(name, "_text_end"))
969 return 1;
970
971 syme = symbol__priv(sym);
972 syme->map = map;
973 syme->src = NULL;
974
975 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
976 /* schedule initial sym_filter_entry setup */
977 sym_filter_entry_sched = syme;
978 sym_filter = NULL;
979 }
980
981 for (i = 0; skip_symbols[i]; i++) {
982 if (!strcmp(skip_symbols[i], name)) {
983 syme->skip = 1;
984 break;
985 }
986 }
987
988 if (!syme->skip)
989 syme->name_len = strlen(sym->name);
990
991 return 0;
992 }
993
994 static void event__process_sample(const event_t *self,
995 struct sample_data *sample,
996 struct perf_session *session)
997 {
998 u64 ip = self->ip.ip;
999 struct sym_entry *syme;
1000 struct addr_location al;
1001 struct machine *machine;
1002 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1003
1004 ++samples;
1005
1006 switch (origin) {
1007 case PERF_RECORD_MISC_USER:
1008 ++us_samples;
1009 if (hide_user_symbols)
1010 return;
1011 machine = perf_session__find_host_machine(session);
1012 break;
1013 case PERF_RECORD_MISC_KERNEL:
1014 ++kernel_samples;
1015 if (hide_kernel_symbols)
1016 return;
1017 machine = perf_session__find_host_machine(session);
1018 break;
1019 case PERF_RECORD_MISC_GUEST_KERNEL:
1020 ++guest_kernel_samples;
1021 machine = perf_session__find_machine(session, self->ip.pid);
1022 break;
1023 case PERF_RECORD_MISC_GUEST_USER:
1024 ++guest_us_samples;
1025 /*
1026 * TODO: we don't process guest user from host side
1027 * except simple counting.
1028 */
1029 return;
1030 default:
1031 return;
1032 }
1033
1034 if (!machine && perf_guest) {
1035 pr_err("Can't find guest [%d]'s kernel information\n",
1036 self->ip.pid);
1037 return;
1038 }
1039
1040 if (self->header.misc & PERF_RECORD_MISC_EXACT_IP)
1041 exact_samples++;
1042
1043 if (event__preprocess_sample(self, session, &al, sample,
1044 symbol_filter) < 0 ||
1045 al.filtered)
1046 return;
1047
1048 if (al.sym == NULL) {
1049 /*
1050 * As we do lazy loading of symtabs we only will know if the
1051 * specified vmlinux file is invalid when we actually have a
1052 * hit in kernel space and then try to load it. So if we get
1053 * here and there are _no_ symbols in the DSO backing the
1054 * kernel map, bail out.
1055 *
1056 * We may never get here, for instance, if we use -K/
1057 * --hide-kernel-symbols, even if the user specifies an
1058 * invalid --vmlinux ;-)
1059 */
1060 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
1061 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
1062 pr_err("The %s file can't be used\n",
1063 symbol_conf.vmlinux_name);
1064 exit(1);
1065 }
1066
1067 return;
1068 }
1069
1070 /* let's see, whether we need to install initial sym_filter_entry */
1071 if (sym_filter_entry_sched) {
1072 sym_filter_entry = sym_filter_entry_sched;
1073 sym_filter_entry_sched = NULL;
1074 if (parse_source(sym_filter_entry) < 0) {
1075 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1076
1077 pr_err("Can't annotate %s", sym->name);
1078 if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1079 pr_err(": No vmlinux file was found in the path:\n");
1080 machine__fprintf_vmlinux_path(machine, stderr);
1081 } else
1082 pr_err(".\n");
1083 exit(1);
1084 }
1085 }
1086
1087 syme = symbol__priv(al.sym);
1088 if (!syme->skip) {
1089 struct perf_evsel *evsel;
1090
1091 syme->origin = origin;
1092 evsel = perf_evlist__id2evsel(evsel_list, sample->id);
1093 assert(evsel != NULL);
1094 syme->count[evsel->idx]++;
1095 record_precise_ip(syme, evsel->idx, ip);
1096 pthread_mutex_lock(&active_symbols_lock);
1097 if (list_empty(&syme->node) || !syme->node.next)
1098 __list_insert_active_sym(syme);
1099 pthread_mutex_unlock(&active_symbols_lock);
1100 }
1101 }
1102
1103 static void perf_session__mmap_read_cpu(struct perf_session *self, int cpu)
1104 {
1105 struct sample_data sample;
1106 event_t *event;
1107
1108 while ((event = perf_evlist__read_on_cpu(evsel_list, cpu)) != NULL) {
1109 event__parse_sample(event, self, &sample);
1110
1111 if (event->header.type == PERF_RECORD_SAMPLE)
1112 event__process_sample(event, &sample, self);
1113 else
1114 event__process(event, &sample, self);
1115 }
1116 }
1117
1118 static void perf_session__mmap_read(struct perf_session *self)
1119 {
1120 int i;
1121
1122 for (i = 0; i < cpus->nr; i++)
1123 perf_session__mmap_read_cpu(self, i);
1124 }
1125
1126 static void start_counters(struct perf_evlist *evlist)
1127 {
1128 struct perf_evsel *counter;
1129
1130 list_for_each_entry(counter, &evlist->entries, node) {
1131 struct perf_event_attr *attr = &counter->attr;
1132
1133 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1134
1135 if (freq) {
1136 attr->sample_type |= PERF_SAMPLE_PERIOD;
1137 attr->freq = 1;
1138 attr->sample_freq = freq;
1139 }
1140
1141 if (evlist->nr_entries > 1) {
1142 attr->sample_type |= PERF_SAMPLE_ID;
1143 attr->read_format |= PERF_FORMAT_ID;
1144 }
1145
1146 attr->mmap = 1;
1147 try_again:
1148 if (perf_evsel__open(counter, cpus, threads, group, inherit) < 0) {
1149 int err = errno;
1150
1151 if (err == EPERM || err == EACCES)
1152 die("Permission error - are you root?\n"
1153 "\t Consider tweaking"
1154 " /proc/sys/kernel/perf_event_paranoid.\n");
1155 /*
1156 * If it's cycles then fall back to hrtimer
1157 * based cpu-clock-tick sw counter, which
1158 * is always available even if no PMU support:
1159 */
1160 if (attr->type == PERF_TYPE_HARDWARE &&
1161 attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1162
1163 if (verbose)
1164 warning(" ... trying to fall back to cpu-clock-ticks\n");
1165
1166 attr->type = PERF_TYPE_SOFTWARE;
1167 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1168 goto try_again;
1169 }
1170 printf("\n");
1171 error("sys_perf_event_open() syscall returned with %d "
1172 "(%s). /bin/dmesg may provide additional information.\n",
1173 err, strerror(err));
1174 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1175 exit(-1);
1176 }
1177 }
1178
1179 if (perf_evlist__mmap(evlist, cpus, threads, mmap_pages, true) < 0)
1180 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1181 }
1182
1183 static int __cmd_top(void)
1184 {
1185 pthread_t thread;
1186 struct perf_evsel *first;
1187 int ret;
1188 /*
1189 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1190 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1191 */
1192 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1193 if (session == NULL)
1194 return -ENOMEM;
1195
1196 if (target_tid != -1)
1197 event__synthesize_thread(target_tid, event__process, session);
1198 else
1199 event__synthesize_threads(event__process, session);
1200
1201 start_counters(evsel_list);
1202 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
1203 perf_session__set_sample_type(session, first->attr.sample_type);
1204
1205 /* Wait for a minimal set of events before starting the snapshot */
1206 poll(evsel_list->pollfd, evsel_list->nr_fds, 100);
1207
1208 perf_session__mmap_read(session);
1209
1210 if (pthread_create(&thread, NULL, display_thread, session)) {
1211 printf("Could not create display thread.\n");
1212 exit(-1);
1213 }
1214
1215 if (realtime_prio) {
1216 struct sched_param param;
1217
1218 param.sched_priority = realtime_prio;
1219 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1220 printf("Could not set realtime priority.\n");
1221 exit(-1);
1222 }
1223 }
1224
1225 while (1) {
1226 int hits = samples;
1227
1228 perf_session__mmap_read(session);
1229
1230 if (hits == samples)
1231 ret = poll(evsel_list->pollfd, evsel_list->nr_fds, 100);
1232 }
1233
1234 return 0;
1235 }
1236
1237 static const char * const top_usage[] = {
1238 "perf top [<options>]",
1239 NULL
1240 };
1241
1242 static const struct option options[] = {
1243 OPT_CALLBACK('e', "event", &evsel_list, "event",
1244 "event selector. use 'perf list' to list available events",
1245 parse_events),
1246 OPT_INTEGER('c', "count", &default_interval,
1247 "event period to sample"),
1248 OPT_INTEGER('p', "pid", &target_pid,
1249 "profile events on existing process id"),
1250 OPT_INTEGER('t', "tid", &target_tid,
1251 "profile events on existing thread id"),
1252 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1253 "system-wide collection from all CPUs"),
1254 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1255 "list of cpus to monitor"),
1256 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1257 "file", "vmlinux pathname"),
1258 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1259 "hide kernel symbols"),
1260 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1261 OPT_INTEGER('r', "realtime", &realtime_prio,
1262 "collect data with this RT SCHED_FIFO priority"),
1263 OPT_INTEGER('d', "delay", &delay_secs,
1264 "number of seconds to delay between refreshes"),
1265 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1266 "dump the symbol table used for profiling"),
1267 OPT_INTEGER('f', "count-filter", &count_filter,
1268 "only display functions with more events than this"),
1269 OPT_BOOLEAN('g', "group", &group,
1270 "put the counters into a counter group"),
1271 OPT_BOOLEAN('i', "inherit", &inherit,
1272 "child tasks inherit counters"),
1273 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1274 "symbol to annotate"),
1275 OPT_BOOLEAN('z', "zero", &zero,
1276 "zero history across updates"),
1277 OPT_INTEGER('F', "freq", &freq,
1278 "profile at this frequency"),
1279 OPT_INTEGER('E', "entries", &print_entries,
1280 "display this many functions"),
1281 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1282 "hide user symbols"),
1283 OPT_INCR('v', "verbose", &verbose,
1284 "be more verbose (show counter open errors, etc)"),
1285 OPT_END()
1286 };
1287
1288 int cmd_top(int argc, const char **argv, const char *prefix __used)
1289 {
1290 struct perf_evsel *pos;
1291 int status = -ENOMEM;
1292
1293 evsel_list = perf_evlist__new();
1294 if (evsel_list == NULL)
1295 return -ENOMEM;
1296
1297 page_size = sysconf(_SC_PAGE_SIZE);
1298
1299 argc = parse_options(argc, argv, options, top_usage, 0);
1300 if (argc)
1301 usage_with_options(top_usage, options);
1302
1303 if (target_pid != -1)
1304 target_tid = target_pid;
1305
1306 threads = thread_map__new(target_pid, target_tid);
1307 if (threads == NULL) {
1308 pr_err("Problems finding threads of monitor\n");
1309 usage_with_options(top_usage, options);
1310 }
1311
1312 /* CPU and PID are mutually exclusive */
1313 if (target_tid > 0 && cpu_list) {
1314 printf("WARNING: PID switch overriding CPU\n");
1315 sleep(1);
1316 cpu_list = NULL;
1317 }
1318
1319 if (!evsel_list->nr_entries &&
1320 perf_evlist__add_default(evsel_list) < 0) {
1321 pr_err("Not enough memory for event selector list\n");
1322 return -ENOMEM;
1323 }
1324
1325 if (delay_secs < 1)
1326 delay_secs = 1;
1327
1328 /*
1329 * User specified count overrides default frequency.
1330 */
1331 if (default_interval)
1332 freq = 0;
1333 else if (freq) {
1334 default_interval = freq;
1335 } else {
1336 fprintf(stderr, "frequency and count are zero, aborting\n");
1337 exit(EXIT_FAILURE);
1338 }
1339
1340 if (target_tid != -1)
1341 cpus = cpu_map__dummy_new();
1342 else
1343 cpus = cpu_map__new(cpu_list);
1344
1345 if (cpus == NULL)
1346 usage_with_options(top_usage, options);
1347
1348 list_for_each_entry(pos, &evsel_list->entries, node) {
1349 if (perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
1350 goto out_free_fd;
1351 /*
1352 * Fill in the ones not specifically initialized via -c:
1353 */
1354 if (pos->attr.sample_period)
1355 continue;
1356
1357 pos->attr.sample_period = default_interval;
1358 }
1359
1360 if (perf_evlist__alloc_pollfd(evsel_list, cpus->nr, threads->nr) < 0 ||
1361 perf_evlist__alloc_mmap(evsel_list, cpus->nr) < 0)
1362 goto out_free_fd;
1363
1364 sym_evsel = list_entry(evsel_list->entries.next, struct perf_evsel, node);
1365
1366 symbol_conf.priv_size = (sizeof(struct sym_entry) +
1367 (evsel_list->nr_entries + 1) * sizeof(unsigned long));
1368
1369 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1370 if (symbol__init() < 0)
1371 return -1;
1372
1373 get_term_dimensions(&winsize);
1374 if (print_entries == 0) {
1375 update_print_entries(&winsize);
1376 signal(SIGWINCH, sig_winch_handler);
1377 }
1378
1379 status = __cmd_top();
1380 out_free_fd:
1381 perf_evlist__delete(evsel_list);
1382
1383 return status;
1384 }
Linux kernel - Deliverables
This page took 0.090281 seconds and 5 git commands to generate.