projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf top: Remove dead {min,max}_ip unused variables
[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/symbol.h"
24 #include "util/color.h"
25 #include "util/util.h"
26 #include <linux/rbtree.h>
27 #include "util/parse-options.h"
28 #include "util/parse-events.h"
29
30 #include "util/debug.h"
31
32 #include <assert.h>
33 #include <fcntl.h>
34
35 #include <stdio.h>
36 #include <termios.h>
37 #include <unistd.h>
38
39 #include <errno.h>
40 #include <time.h>
41 #include <sched.h>
42 #include <pthread.h>
43
44 #include <sys/syscall.h>
45 #include <sys/ioctl.h>
46 #include <sys/poll.h>
47 #include <sys/prctl.h>
48 #include <sys/wait.h>
49 #include <sys/uio.h>
50 #include <sys/mman.h>
51
52 #include <linux/unistd.h>
53 #include <linux/types.h>
54
55 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
56
57 static int system_wide = 0;
58
59 static int default_interval = 100000;
60
61 static int count_filter = 5;
62 static int print_entries = 15;
63
64 static int target_pid = -1;
65 static int inherit = 0;
66 static int profile_cpu = -1;
67 static int nr_cpus = 0;
68 static unsigned int realtime_prio = 0;
69 static int group = 0;
70 static unsigned int page_size;
71 static unsigned int mmap_pages = 16;
72 static int freq = 0;
73
74 static int delay_secs = 2;
75 static int zero;
76 static int dump_symtab;
77
78 /*
79 * Source
80 */
81
82 struct source_line {
83 u64 eip;
84 unsigned long count[MAX_COUNTERS];
85 char *line;
86 struct source_line *next;
87 };
88
89 static char *sym_filter = NULL;
90 struct sym_entry *sym_filter_entry = NULL;
91 static int sym_pcnt_filter = 5;
92 static int sym_counter = 0;
93 static int display_weighted = -1;
94
95 /*
96 * Symbols
97 */
98
99 struct sym_entry {
100 struct rb_node rb_node;
101 struct list_head node;
102 unsigned long count[MAX_COUNTERS];
103 unsigned long snap_count;
104 double weight;
105 int skip;
106 struct source_line *source;
107 struct source_line *lines;
108 struct source_line **lines_tail;
109 pthread_mutex_t source_lock;
110 };
111
112 /*
113 * Source functions
114 */
115
116 static void parse_source(struct sym_entry *syme)
117 {
118 struct symbol *sym;
119 struct module *module;
120 struct section *section = NULL;
121 FILE *file;
122 char command[PATH_MAX*2];
123 const char *path = vmlinux_name;
124 u64 start, end, len;
125
126 if (!syme)
127 return;
128
129 if (syme->lines) {
130 pthread_mutex_lock(&syme->source_lock);
131 goto out_assign;
132 }
133
134 sym = (struct symbol *)(syme + 1);
135 module = sym->module;
136
137 if (module)
138 path = module->path;
139 if (!path)
140 return;
141
142 start = sym->obj_start;
143 if (!start)
144 start = sym->start;
145
146 if (module) {
147 section = module->sections->find_section(module->sections, ".text");
148 if (section)
149 start -= section->vma;
150 }
151
152 end = start + sym->end - sym->start + 1;
153 len = sym->end - sym->start;
154
155 sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s", start, end, path);
156
157 file = popen(command, "r");
158 if (!file)
159 return;
160
161 pthread_mutex_lock(&syme->source_lock);
162 syme->lines_tail = &syme->lines;
163 while (!feof(file)) {
164 struct source_line *src;
165 size_t dummy = 0;
166 char *c;
167
168 src = malloc(sizeof(struct source_line));
169 assert(src != NULL);
170 memset(src, 0, sizeof(struct source_line));
171
172 if (getline(&src->line, &dummy, file) < 0)
173 break;
174 if (!src->line)
175 break;
176
177 c = strchr(src->line, '\n');
178 if (c)
179 *c = 0;
180
181 src->next = NULL;
182 *syme->lines_tail = src;
183 syme->lines_tail = &src->next;
184
185 if (strlen(src->line)>8 && src->line[8] == ':') {
186 src->eip = strtoull(src->line, NULL, 16);
187 if (section)
188 src->eip += section->vma;
189 }
190 if (strlen(src->line)>8 && src->line[16] == ':') {
191 src->eip = strtoull(src->line, NULL, 16);
192 if (section)
193 src->eip += section->vma;
194 }
195 }
196 pclose(file);
197 out_assign:
198 sym_filter_entry = syme;
199 pthread_mutex_unlock(&syme->source_lock);
200 }
201
202 static void __zero_source_counters(struct sym_entry *syme)
203 {
204 int i;
205 struct source_line *line;
206
207 line = syme->lines;
208 while (line) {
209 for (i = 0; i < nr_counters; i++)
210 line->count[i] = 0;
211 line = line->next;
212 }
213 }
214
215 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
216 {
217 struct source_line *line;
218
219 if (syme != sym_filter_entry)
220 return;
221
222 if (pthread_mutex_trylock(&syme->source_lock))
223 return;
224
225 if (!syme->source)
226 goto out_unlock;
227
228 for (line = syme->lines; line; line = line->next) {
229 if (line->eip == ip) {
230 line->count[counter]++;
231 break;
232 }
233 if (line->eip > ip)
234 break;
235 }
236 out_unlock:
237 pthread_mutex_unlock(&syme->source_lock);
238 }
239
240 static void lookup_sym_source(struct sym_entry *syme)
241 {
242 struct symbol *symbol = (struct symbol *)(syme + 1);
243 struct source_line *line;
244 char pattern[PATH_MAX];
245 char *idx;
246
247 sprintf(pattern, "<%s>:", symbol->name);
248
249 if (symbol->module) {
250 idx = strstr(pattern, "\t");
251 if (idx)
252 *idx = 0;
253 }
254
255 pthread_mutex_lock(&syme->source_lock);
256 for (line = syme->lines; line; line = line->next) {
257 if (strstr(line->line, pattern)) {
258 syme->source = line;
259 break;
260 }
261 }
262 pthread_mutex_unlock(&syme->source_lock);
263 }
264
265 static void show_lines(struct source_line *queue, int count, int total)
266 {
267 int i;
268 struct source_line *line;
269
270 line = queue;
271 for (i = 0; i < count; i++) {
272 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
273
274 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
275 line = line->next;
276 }
277 }
278
279 #define TRACE_COUNT 3
280
281 static void show_details(struct sym_entry *syme)
282 {
283 struct symbol *symbol;
284 struct source_line *line;
285 struct source_line *line_queue = NULL;
286 int displayed = 0;
287 int line_queue_count = 0, total = 0, more = 0;
288
289 if (!syme)
290 return;
291
292 if (!syme->source)
293 lookup_sym_source(syme);
294
295 if (!syme->source)
296 return;
297
298 symbol = (struct symbol *)(syme + 1);
299 printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
300 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
301
302 pthread_mutex_lock(&syme->source_lock);
303 line = syme->source;
304 while (line) {
305 total += line->count[sym_counter];
306 line = line->next;
307 }
308
309 line = syme->source;
310 while (line) {
311 float pcnt = 0.0;
312
313 if (!line_queue_count)
314 line_queue = line;
315 line_queue_count++;
316
317 if (line->count[sym_counter])
318 pcnt = 100.0 * line->count[sym_counter] / (float)total;
319 if (pcnt >= (float)sym_pcnt_filter) {
320 if (displayed <= print_entries)
321 show_lines(line_queue, line_queue_count, total);
322 else more++;
323 displayed += line_queue_count;
324 line_queue_count = 0;
325 line_queue = NULL;
326 } else if (line_queue_count > TRACE_COUNT) {
327 line_queue = line_queue->next;
328 line_queue_count--;
329 }
330
331 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
332 line = line->next;
333 }
334 pthread_mutex_unlock(&syme->source_lock);
335 if (more)
336 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
337 }
338
339 /*
340 * Symbols will be added here in record_ip and will get out
341 * after decayed.
342 */
343 static LIST_HEAD(active_symbols);
344 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
345
346 /*
347 * Ordering weight: count-1 * count-2 * ... / count-n
348 */
349 static double sym_weight(const struct sym_entry *sym)
350 {
351 double weight = sym->snap_count;
352 int counter;
353
354 if (!display_weighted)
355 return weight;
356
357 for (counter = 1; counter < nr_counters-1; counter++)
358 weight *= sym->count[counter];
359
360 weight /= (sym->count[counter] + 1);
361
362 return weight;
363 }
364
365 static long samples;
366 static long userspace_samples;
367 static const char CONSOLE_CLEAR[] = "\e[H\e[2J";
368
369 static void __list_insert_active_sym(struct sym_entry *syme)
370 {
371 list_add(&syme->node, &active_symbols);
372 }
373
374 static void list_remove_active_sym(struct sym_entry *syme)
375 {
376 pthread_mutex_lock(&active_symbols_lock);
377 list_del_init(&syme->node);
378 pthread_mutex_unlock(&active_symbols_lock);
379 }
380
381 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
382 {
383 struct rb_node **p = &tree->rb_node;
384 struct rb_node *parent = NULL;
385 struct sym_entry *iter;
386
387 while (*p != NULL) {
388 parent = *p;
389 iter = rb_entry(parent, struct sym_entry, rb_node);
390
391 if (se->weight > iter->weight)
392 p = &(*p)->rb_left;
393 else
394 p = &(*p)->rb_right;
395 }
396
397 rb_link_node(&se->rb_node, parent, p);
398 rb_insert_color(&se->rb_node, tree);
399 }
400
401 static void print_sym_table(void)
402 {
403 int printed = 0, j;
404 int counter, snap = !display_weighted ? sym_counter : 0;
405 float samples_per_sec = samples/delay_secs;
406 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
407 float sum_ksamples = 0.0;
408 struct sym_entry *syme, *n;
409 struct rb_root tmp = RB_ROOT;
410 struct rb_node *nd;
411
412 samples = userspace_samples = 0;
413
414 /* Sort the active symbols */
415 pthread_mutex_lock(&active_symbols_lock);
416 syme = list_entry(active_symbols.next, struct sym_entry, node);
417 pthread_mutex_unlock(&active_symbols_lock);
418
419 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
420 syme->snap_count = syme->count[snap];
421 if (syme->snap_count != 0) {
422 syme->weight = sym_weight(syme);
423 rb_insert_active_sym(&tmp, syme);
424 sum_ksamples += syme->snap_count;
425
426 for (j = 0; j < nr_counters; j++)
427 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
428 } else
429 list_remove_active_sym(syme);
430 }
431
432 puts(CONSOLE_CLEAR);
433
434 printf(
435 "------------------------------------------------------------------------------\n");
436 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
437 samples_per_sec,
438 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
439
440 if (nr_counters == 1 || !display_weighted) {
441 printf("%Ld", (u64)attrs[0].sample_period);
442 if (freq)
443 printf("Hz ");
444 else
445 printf(" ");
446 }
447
448 if (!display_weighted)
449 printf("%s", event_name(sym_counter));
450 else for (counter = 0; counter < nr_counters; counter++) {
451 if (counter)
452 printf("/");
453
454 printf("%s", event_name(counter));
455 }
456
457 printf( "], ");
458
459 if (target_pid != -1)
460 printf(" (target_pid: %d", target_pid);
461 else
462 printf(" (all");
463
464 if (profile_cpu != -1)
465 printf(", cpu: %d)\n", profile_cpu);
466 else {
467 if (target_pid != -1)
468 printf(")\n");
469 else
470 printf(", %d CPUs)\n", nr_cpus);
471 }
472
473 printf("------------------------------------------------------------------------------\n\n");
474
475 if (sym_filter_entry) {
476 show_details(sym_filter_entry);
477 return;
478 }
479
480 if (nr_counters == 1)
481 printf(" samples pcnt");
482 else
483 printf(" weight samples pcnt");
484
485 if (verbose)
486 printf(" RIP ");
487 printf(" kernel function\n");
488 printf(" %s _______ _____",
489 nr_counters == 1 ? " " : "______");
490 if (verbose)
491 printf(" ________________");
492 printf(" _______________\n\n");
493
494 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
495 struct symbol *sym;
496 double pcnt;
497
498 syme = rb_entry(nd, struct sym_entry, rb_node);
499 sym = (struct symbol *)(syme + 1);
500
501 if (++printed > print_entries || (int)syme->snap_count < count_filter)
502 continue;
503
504 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
505 sum_ksamples));
506
507 if (nr_counters == 1 || !display_weighted)
508 printf("%20.2f - ", syme->weight);
509 else
510 printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
511
512 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
513 if (verbose)
514 printf(" - %016llx", sym->start);
515 printf(" : %s", sym->name);
516 if (sym->module)
517 printf("\t[%s]", sym->module->name);
518 printf("\n");
519 }
520 }
521
522 static void prompt_integer(int *target, const char *msg)
523 {
524 char *buf = malloc(0), *p;
525 size_t dummy = 0;
526 int tmp;
527
528 fprintf(stdout, "\n%s: ", msg);
529 if (getline(&buf, &dummy, stdin) < 0)
530 return;
531
532 p = strchr(buf, '\n');
533 if (p)
534 *p = 0;
535
536 p = buf;
537 while(*p) {
538 if (!isdigit(*p))
539 goto out_free;
540 p++;
541 }
542 tmp = strtoul(buf, NULL, 10);
543 *target = tmp;
544 out_free:
545 free(buf);
546 }
547
548 static void prompt_percent(int *target, const char *msg)
549 {
550 int tmp = 0;
551
552 prompt_integer(&tmp, msg);
553 if (tmp >= 0 && tmp <= 100)
554 *target = tmp;
555 }
556
557 static void prompt_symbol(struct sym_entry **target, const char *msg)
558 {
559 char *buf = malloc(0), *p;
560 struct sym_entry *syme = *target, *n, *found = NULL;
561 size_t dummy = 0;
562
563 /* zero counters of active symbol */
564 if (syme) {
565 pthread_mutex_lock(&syme->source_lock);
566 __zero_source_counters(syme);
567 *target = NULL;
568 pthread_mutex_unlock(&syme->source_lock);
569 }
570
571 fprintf(stdout, "\n%s: ", msg);
572 if (getline(&buf, &dummy, stdin) < 0)
573 goto out_free;
574
575 p = strchr(buf, '\n');
576 if (p)
577 *p = 0;
578
579 pthread_mutex_lock(&active_symbols_lock);
580 syme = list_entry(active_symbols.next, struct sym_entry, node);
581 pthread_mutex_unlock(&active_symbols_lock);
582
583 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
584 struct symbol *sym = (struct symbol *)(syme + 1);
585
586 if (!strcmp(buf, sym->name)) {
587 found = syme;
588 break;
589 }
590 }
591
592 if (!found) {
593 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
594 sleep(1);
595 return;
596 } else
597 parse_source(found);
598
599 out_free:
600 free(buf);
601 }
602
603 static void print_mapped_keys(void)
604 {
605 char *name = NULL;
606
607 if (sym_filter_entry) {
608 struct symbol *sym = (struct symbol *)(sym_filter_entry+1);
609 name = sym->name;
610 }
611
612 fprintf(stdout, "\nMapped keys:\n");
613 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
614 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
615
616 if (nr_counters > 1)
617 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter));
618
619 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
620
621 if (vmlinux_name) {
622 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
623 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
624 fprintf(stdout, "\t[S] stop annotation.\n");
625 }
626
627 if (nr_counters > 1)
628 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
629
630 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
631 fprintf(stdout, "\t[qQ] quit.\n");
632 }
633
634 static int key_mapped(int c)
635 {
636 switch (c) {
637 case 'd':
638 case 'e':
639 case 'f':
640 case 'z':
641 case 'q':
642 case 'Q':
643 return 1;
644 case 'E':
645 case 'w':
646 return nr_counters > 1 ? 1 : 0;
647 case 'F':
648 case 's':
649 case 'S':
650 return vmlinux_name ? 1 : 0;
651 default:
652 break;
653 }
654
655 return 0;
656 }
657
658 static void handle_keypress(int c)
659 {
660 if (!key_mapped(c)) {
661 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
662 struct termios tc, save;
663
664 print_mapped_keys();
665 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
666 fflush(stdout);
667
668 tcgetattr(0, &save);
669 tc = save;
670 tc.c_lflag &= ~(ICANON | ECHO);
671 tc.c_cc[VMIN] = 0;
672 tc.c_cc[VTIME] = 0;
673 tcsetattr(0, TCSANOW, &tc);
674
675 poll(&stdin_poll, 1, -1);
676 c = getc(stdin);
677
678 tcsetattr(0, TCSAFLUSH, &save);
679 if (!key_mapped(c))
680 return;
681 }
682
683 switch (c) {
684 case 'd':
685 prompt_integer(&delay_secs, "Enter display delay");
686 break;
687 case 'e':
688 prompt_integer(&print_entries, "Enter display entries (lines)");
689 break;
690 case 'E':
691 if (nr_counters > 1) {
692 int i;
693
694 fprintf(stderr, "\nAvailable events:");
695 for (i = 0; i < nr_counters; i++)
696 fprintf(stderr, "\n\t%d %s", i, event_name(i));
697
698 prompt_integer(&sym_counter, "Enter details event counter");
699
700 if (sym_counter >= nr_counters) {
701 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
702 sym_counter = 0;
703 sleep(1);
704 }
705 } else sym_counter = 0;
706 break;
707 case 'f':
708 prompt_integer(&count_filter, "Enter display event count filter");
709 break;
710 case 'F':
711 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
712 break;
713 case 'q':
714 case 'Q':
715 printf("exiting.\n");
716 exit(0);
717 case 's':
718 prompt_symbol(&sym_filter_entry, "Enter details symbol");
719 break;
720 case 'S':
721 if (!sym_filter_entry)
722 break;
723 else {
724 struct sym_entry *syme = sym_filter_entry;
725
726 pthread_mutex_lock(&syme->source_lock);
727 sym_filter_entry = NULL;
728 __zero_source_counters(syme);
729 pthread_mutex_unlock(&syme->source_lock);
730 }
731 break;
732 case 'w':
733 display_weighted = ~display_weighted;
734 break;
735 case 'z':
736 zero = ~zero;
737 break;
738 default:
739 break;
740 }
741 }
742
743 static void *display_thread(void *arg __used)
744 {
745 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
746 struct termios tc, save;
747 int delay_msecs, c;
748
749 tcgetattr(0, &save);
750 tc = save;
751 tc.c_lflag &= ~(ICANON | ECHO);
752 tc.c_cc[VMIN] = 0;
753 tc.c_cc[VTIME] = 0;
754
755 repeat:
756 delay_msecs = delay_secs * 1000;
757 tcsetattr(0, TCSANOW, &tc);
758 /* trash return*/
759 getc(stdin);
760
761 do {
762 print_sym_table();
763 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
764
765 c = getc(stdin);
766 tcsetattr(0, TCSAFLUSH, &save);
767
768 handle_keypress(c);
769 goto repeat;
770
771 return NULL;
772 }
773
774 /* Tag samples to be skipped. */
775 static const char *skip_symbols[] = {
776 "default_idle",
777 "cpu_idle",
778 "enter_idle",
779 "exit_idle",
780 "mwait_idle",
781 "mwait_idle_with_hints",
782 "ppc64_runlatch_off",
783 "pseries_dedicated_idle_sleep",
784 NULL
785 };
786
787 static int symbol_filter(struct dso *self, struct symbol *sym)
788 {
789 struct sym_entry *syme;
790 const char *name = sym->name;
791 int i;
792
793 /*
794 * ppc64 uses function descriptors and appends a '.' to the
795 * start of every instruction address. Remove it.
796 */
797 if (name[0] == '.')
798 name++;
799
800 if (!strcmp(name, "_text") ||
801 !strcmp(name, "_etext") ||
802 !strcmp(name, "_sinittext") ||
803 !strncmp("init_module", name, 11) ||
804 !strncmp("cleanup_module", name, 14) ||
805 strstr(name, "_text_start") ||
806 strstr(name, "_text_end"))
807 return 1;
808
809 syme = dso__sym_priv(self, sym);
810 pthread_mutex_init(&syme->source_lock, NULL);
811 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
812 sym_filter_entry = syme;
813
814 for (i = 0; skip_symbols[i]; i++) {
815 if (!strcmp(skip_symbols[i], name)) {
816 syme->skip = 1;
817 break;
818 }
819 }
820
821 return 0;
822 }
823
824 static int parse_symbols(void)
825 {
826 int use_modules = vmlinux_name ? 1 : 0;
827
828 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
829 if (kernel_dso == NULL)
830 return -1;
831
832 if (dso__load_kernel(kernel_dso, vmlinux_name, symbol_filter, verbose, use_modules) <= 0)
833 goto out_delete_dso;
834
835 if (dump_symtab)
836 dso__fprintf(kernel_dso, stderr);
837
838 return 0;
839
840 out_delete_dso:
841 dso__delete(kernel_dso);
842 kernel_dso = NULL;
843 return -1;
844 }
845
846 /*
847 * Binary search in the histogram table and record the hit:
848 */
849 static void record_ip(u64 ip, int counter)
850 {
851 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
852
853 if (sym != NULL) {
854 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
855
856 if (!syme->skip) {
857 syme->count[counter]++;
858 record_precise_ip(syme, counter, ip);
859 pthread_mutex_lock(&active_symbols_lock);
860 if (list_empty(&syme->node) || !syme->node.next)
861 __list_insert_active_sym(syme);
862 pthread_mutex_unlock(&active_symbols_lock);
863 return;
864 }
865 }
866
867 samples--;
868 }
869
870 static void process_event(u64 ip, int counter, int user)
871 {
872 samples++;
873
874 if (user) {
875 userspace_samples++;
876 return;
877 }
878
879 record_ip(ip, counter);
880 }
881
882 struct mmap_data {
883 int counter;
884 void *base;
885 int mask;
886 unsigned int prev;
887 };
888
889 static unsigned int mmap_read_head(struct mmap_data *md)
890 {
891 struct perf_event_mmap_page *pc = md->base;
892 int head;
893
894 head = pc->data_head;
895 rmb();
896
897 return head;
898 }
899
900 struct timeval last_read, this_read;
901
902 static void mmap_read_counter(struct mmap_data *md)
903 {
904 unsigned int head = mmap_read_head(md);
905 unsigned int old = md->prev;
906 unsigned char *data = md->base + page_size;
907 int diff;
908
909 gettimeofday(&this_read, NULL);
910
911 /*
912 * If we're further behind than half the buffer, there's a chance
913 * the writer will bite our tail and mess up the samples under us.
914 *
915 * If we somehow ended up ahead of the head, we got messed up.
916 *
917 * In either case, truncate and restart at head.
918 */
919 diff = head - old;
920 if (diff > md->mask / 2 || diff < 0) {
921 struct timeval iv;
922 unsigned long msecs;
923
924 timersub(&this_read, &last_read, &iv);
925 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
926
927 fprintf(stderr, "WARNING: failed to keep up with mmap data."
928 " Last read %lu msecs ago.\n", msecs);
929
930 /*
931 * head points to a known good entry, start there.
932 */
933 old = head;
934 }
935
936 last_read = this_read;
937
938 for (; old != head;) {
939 event_t *event = (event_t *)&data[old & md->mask];
940
941 event_t event_copy;
942
943 size_t size = event->header.size;
944
945 /*
946 * Event straddles the mmap boundary -- header should always
947 * be inside due to u64 alignment of output.
948 */
949 if ((old & md->mask) + size != ((old + size) & md->mask)) {
950 unsigned int offset = old;
951 unsigned int len = min(sizeof(*event), size), cpy;
952 void *dst = &event_copy;
953
954 do {
955 cpy = min(md->mask + 1 - (offset & md->mask), len);
956 memcpy(dst, &data[offset & md->mask], cpy);
957 offset += cpy;
958 dst += cpy;
959 len -= cpy;
960 } while (len);
961
962 event = &event_copy;
963 }
964
965 old += size;
966
967 if (event->header.type == PERF_RECORD_SAMPLE) {
968 int user =
969 (event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK) == PERF_RECORD_MISC_USER;
970 process_event(event->ip.ip, md->counter, user);
971 }
972 }
973
974 md->prev = old;
975 }
976
977 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
978 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
979
980 static void mmap_read(void)
981 {
982 int i, counter;
983
984 for (i = 0; i < nr_cpus; i++) {
985 for (counter = 0; counter < nr_counters; counter++)
986 mmap_read_counter(&mmap_array[i][counter]);
987 }
988 }
989
990 int nr_poll;
991 int group_fd;
992
993 static void start_counter(int i, int counter)
994 {
995 struct perf_event_attr *attr;
996 int cpu;
997
998 cpu = profile_cpu;
999 if (target_pid == -1 && profile_cpu == -1)
1000 cpu = i;
1001
1002 attr = attrs + counter;
1003
1004 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1005 attr->freq = freq;
1006 attr->inherit = (cpu < 0) && inherit;
1007
1008 try_again:
1009 fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1010
1011 if (fd[i][counter] < 0) {
1012 int err = errno;
1013
1014 if (err == EPERM)
1015 die("No permission - are you root?\n");
1016 /*
1017 * If it's cycles then fall back to hrtimer
1018 * based cpu-clock-tick sw counter, which
1019 * is always available even if no PMU support:
1020 */
1021 if (attr->type == PERF_TYPE_HARDWARE
1022 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1023
1024 if (verbose)
1025 warning(" ... trying to fall back to cpu-clock-ticks\n");
1026
1027 attr->type = PERF_TYPE_SOFTWARE;
1028 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1029 goto try_again;
1030 }
1031 printf("\n");
1032 error("perfcounter syscall returned with %d (%s)\n",
1033 fd[i][counter], strerror(err));
1034 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1035 exit(-1);
1036 }
1037 assert(fd[i][counter] >= 0);
1038 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1039
1040 /*
1041 * First counter acts as the group leader:
1042 */
1043 if (group && group_fd == -1)
1044 group_fd = fd[i][counter];
1045
1046 event_array[nr_poll].fd = fd[i][counter];
1047 event_array[nr_poll].events = POLLIN;
1048 nr_poll++;
1049
1050 mmap_array[i][counter].counter = counter;
1051 mmap_array[i][counter].prev = 0;
1052 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1053 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1054 PROT_READ, MAP_SHARED, fd[i][counter], 0);
1055 if (mmap_array[i][counter].base == MAP_FAILED)
1056 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1057 }
1058
1059 static int __cmd_top(void)
1060 {
1061 pthread_t thread;
1062 int i, counter;
1063 int ret;
1064
1065 for (i = 0; i < nr_cpus; i++) {
1066 group_fd = -1;
1067 for (counter = 0; counter < nr_counters; counter++)
1068 start_counter(i, counter);
1069 }
1070
1071 /* Wait for a minimal set of events before starting the snapshot */
1072 poll(event_array, nr_poll, 100);
1073
1074 mmap_read();
1075
1076 if (pthread_create(&thread, NULL, display_thread, NULL)) {
1077 printf("Could not create display thread.\n");
1078 exit(-1);
1079 }
1080
1081 if (realtime_prio) {
1082 struct sched_param param;
1083
1084 param.sched_priority = realtime_prio;
1085 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1086 printf("Could not set realtime priority.\n");
1087 exit(-1);
1088 }
1089 }
1090
1091 while (1) {
1092 int hits = samples;
1093
1094 mmap_read();
1095
1096 if (hits == samples)
1097 ret = poll(event_array, nr_poll, 100);
1098 }
1099
1100 return 0;
1101 }
1102
1103 static const char * const top_usage[] = {
1104 "perf top [<options>]",
1105 NULL
1106 };
1107
1108 static const struct option options[] = {
1109 OPT_CALLBACK('e', "event", NULL, "event",
1110 "event selector. use 'perf list' to list available events",
1111 parse_events),
1112 OPT_INTEGER('c', "count", &default_interval,
1113 "event period to sample"),
1114 OPT_INTEGER('p', "pid", &target_pid,
1115 "profile events on existing pid"),
1116 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1117 "system-wide collection from all CPUs"),
1118 OPT_INTEGER('C', "CPU", &profile_cpu,
1119 "CPU to profile on"),
1120 OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1121 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1122 "number of mmap data pages"),
1123 OPT_INTEGER('r', "realtime", &realtime_prio,
1124 "collect data with this RT SCHED_FIFO priority"),
1125 OPT_INTEGER('d', "delay", &delay_secs,
1126 "number of seconds to delay between refreshes"),
1127 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1128 "dump the symbol table used for profiling"),
1129 OPT_INTEGER('f', "count-filter", &count_filter,
1130 "only display functions with more events than this"),
1131 OPT_BOOLEAN('g', "group", &group,
1132 "put the counters into a counter group"),
1133 OPT_BOOLEAN('i', "inherit", &inherit,
1134 "child tasks inherit counters"),
1135 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1136 "symbol to annotate - requires -k option"),
1137 OPT_BOOLEAN('z', "zero", &zero,
1138 "zero history across updates"),
1139 OPT_INTEGER('F', "freq", &freq,
1140 "profile at this frequency"),
1141 OPT_INTEGER('E', "entries", &print_entries,
1142 "display this many functions"),
1143 OPT_BOOLEAN('v', "verbose", &verbose,
1144 "be more verbose (show counter open errors, etc)"),
1145 OPT_END()
1146 };
1147
1148 int cmd_top(int argc, const char **argv, const char *prefix __used)
1149 {
1150 int counter;
1151
1152 symbol__init();
1153
1154 page_size = sysconf(_SC_PAGE_SIZE);
1155
1156 argc = parse_options(argc, argv, options, top_usage, 0);
1157 if (argc)
1158 usage_with_options(top_usage, options);
1159
1160 if (freq) {
1161 default_interval = freq;
1162 freq = 1;
1163 }
1164
1165 /* CPU and PID are mutually exclusive */
1166 if (target_pid != -1 && profile_cpu != -1) {
1167 printf("WARNING: PID switch overriding CPU\n");
1168 sleep(1);
1169 profile_cpu = -1;
1170 }
1171
1172 if (!nr_counters)
1173 nr_counters = 1;
1174
1175 if (delay_secs < 1)
1176 delay_secs = 1;
1177
1178 parse_symbols();
1179 parse_source(sym_filter_entry);
1180
1181 /*
1182 * Fill in the ones not specifically initialized via -c:
1183 */
1184 for (counter = 0; counter < nr_counters; counter++) {
1185 if (attrs[counter].sample_period)
1186 continue;
1187
1188 attrs[counter].sample_period = default_interval;
1189 }
1190
1191 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1192 assert(nr_cpus <= MAX_NR_CPUS);
1193 assert(nr_cpus >= 0);
1194
1195 if (target_pid != -1 || profile_cpu != -1)
1196 nr_cpus = 1;
1197
1198 return __cmd_top();
1199 }
Linux kernel - Deliverables
This page took 0.079505 seconds and 6 git commands to generate.