perf tools: Fix permission checks
[linux-2.6.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 static u64                      min_ip;
100 static u64                      max_ip = -1ll;
101
102 struct sym_entry {
103         struct rb_node          rb_node;
104         struct list_head        node;
105         unsigned long           count[MAX_COUNTERS];
106         unsigned long           snap_count;
107         double                  weight;
108         int                     skip;
109         struct source_line      *source;
110         struct source_line      *lines;
111         struct source_line      **lines_tail;
112         pthread_mutex_t         source_lock;
113 };
114
115 /*
116  * Source functions
117  */
118
119 static void parse_source(struct sym_entry *syme)
120 {
121         struct symbol *sym;
122         struct module *module;
123         struct section *section = NULL;
124         FILE *file;
125         char command[PATH_MAX*2];
126         const char *path = vmlinux_name;
127         u64 start, end, len;
128
129         if (!syme)
130                 return;
131
132         if (syme->lines) {
133                 pthread_mutex_lock(&syme->source_lock);
134                 goto out_assign;
135         }
136
137         sym = (struct symbol *)(syme + 1);
138         module = sym->module;
139
140         if (module)
141                 path = module->path;
142         if (!path)
143                 return;
144
145         start = sym->obj_start;
146         if (!start)
147                 start = sym->start;
148
149         if (module) {
150                 section = module->sections->find_section(module->sections, ".text");
151                 if (section)
152                         start -= section->vma;
153         }
154
155         end = start + sym->end - sym->start + 1;
156         len = sym->end - sym->start;
157
158         sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s", start, end, path);
159
160         file = popen(command, "r");
161         if (!file)
162                 return;
163
164         pthread_mutex_lock(&syme->source_lock);
165         syme->lines_tail = &syme->lines;
166         while (!feof(file)) {
167                 struct source_line *src;
168                 size_t dummy = 0;
169                 char *c;
170
171                 src = malloc(sizeof(struct source_line));
172                 assert(src != NULL);
173                 memset(src, 0, sizeof(struct source_line));
174
175                 if (getline(&src->line, &dummy, file) < 0)
176                         break;
177                 if (!src->line)
178                         break;
179
180                 c = strchr(src->line, '\n');
181                 if (c)
182                         *c = 0;
183
184                 src->next = NULL;
185                 *syme->lines_tail = src;
186                 syme->lines_tail = &src->next;
187
188                 if (strlen(src->line)>8 && src->line[8] == ':') {
189                         src->eip = strtoull(src->line, NULL, 16);
190                         if (section)
191                                 src->eip += section->vma;
192                 }
193                 if (strlen(src->line)>8 && src->line[16] == ':') {
194                         src->eip = strtoull(src->line, NULL, 16);
195                         if (section)
196                                 src->eip += section->vma;
197                 }
198         }
199         pclose(file);
200 out_assign:
201         sym_filter_entry = syme;
202         pthread_mutex_unlock(&syme->source_lock);
203 }
204
205 static void __zero_source_counters(struct sym_entry *syme)
206 {
207         int i;
208         struct source_line *line;
209
210         line = syme->lines;
211         while (line) {
212                 for (i = 0; i < nr_counters; i++)
213                         line->count[i] = 0;
214                 line = line->next;
215         }
216 }
217
218 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
219 {
220         struct source_line *line;
221
222         if (syme != sym_filter_entry)
223                 return;
224
225         if (pthread_mutex_trylock(&syme->source_lock))
226                 return;
227
228         if (!syme->source)
229                 goto out_unlock;
230
231         for (line = syme->lines; line; line = line->next) {
232                 if (line->eip == ip) {
233                         line->count[counter]++;
234                         break;
235                 }
236                 if (line->eip > ip)
237                         break;
238         }
239 out_unlock:
240         pthread_mutex_unlock(&syme->source_lock);
241 }
242
243 static void lookup_sym_source(struct sym_entry *syme)
244 {
245         struct symbol *symbol = (struct symbol *)(syme + 1);
246         struct source_line *line;
247         char pattern[PATH_MAX];
248         char *idx;
249
250         sprintf(pattern, "<%s>:", symbol->name);
251
252         if (symbol->module) {
253                 idx = strstr(pattern, "\t");
254                 if (idx)
255                         *idx = 0;
256         }
257
258         pthread_mutex_lock(&syme->source_lock);
259         for (line = syme->lines; line; line = line->next) {
260                 if (strstr(line->line, pattern)) {
261                         syme->source = line;
262                         break;
263                 }
264         }
265         pthread_mutex_unlock(&syme->source_lock);
266 }
267
268 static void show_lines(struct source_line *queue, int count, int total)
269 {
270         int i;
271         struct source_line *line;
272
273         line = queue;
274         for (i = 0; i < count; i++) {
275                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
276
277                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
278                 line = line->next;
279         }
280 }
281
282 #define TRACE_COUNT     3
283
284 static void show_details(struct sym_entry *syme)
285 {
286         struct symbol *symbol;
287         struct source_line *line;
288         struct source_line *line_queue = NULL;
289         int displayed = 0;
290         int line_queue_count = 0, total = 0, more = 0;
291
292         if (!syme)
293                 return;
294
295         if (!syme->source)
296                 lookup_sym_source(syme);
297
298         if (!syme->source)
299                 return;
300
301         symbol = (struct symbol *)(syme + 1);
302         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
303         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
304
305         pthread_mutex_lock(&syme->source_lock);
306         line = syme->source;
307         while (line) {
308                 total += line->count[sym_counter];
309                 line = line->next;
310         }
311
312         line = syme->source;
313         while (line) {
314                 float pcnt = 0.0;
315
316                 if (!line_queue_count)
317                         line_queue = line;
318                 line_queue_count++;
319
320                 if (line->count[sym_counter])
321                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
322                 if (pcnt >= (float)sym_pcnt_filter) {
323                         if (displayed <= print_entries)
324                                 show_lines(line_queue, line_queue_count, total);
325                         else more++;
326                         displayed += line_queue_count;
327                         line_queue_count = 0;
328                         line_queue = NULL;
329                 } else if (line_queue_count > TRACE_COUNT) {
330                         line_queue = line_queue->next;
331                         line_queue_count--;
332                 }
333
334                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
335                 line = line->next;
336         }
337         pthread_mutex_unlock(&syme->source_lock);
338         if (more)
339                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
340 }
341
342 /*
343  * Symbols will be added here in record_ip and will get out
344  * after decayed.
345  */
346 static LIST_HEAD(active_symbols);
347 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
348
349 /*
350  * Ordering weight: count-1 * count-2 * ... / count-n
351  */
352 static double sym_weight(const struct sym_entry *sym)
353 {
354         double weight = sym->snap_count;
355         int counter;
356
357         if (!display_weighted)
358                 return weight;
359
360         for (counter = 1; counter < nr_counters-1; counter++)
361                 weight *= sym->count[counter];
362
363         weight /= (sym->count[counter] + 1);
364
365         return weight;
366 }
367
368 static long                     samples;
369 static long                     userspace_samples;
370 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
371
372 static void __list_insert_active_sym(struct sym_entry *syme)
373 {
374         list_add(&syme->node, &active_symbols);
375 }
376
377 static void list_remove_active_sym(struct sym_entry *syme)
378 {
379         pthread_mutex_lock(&active_symbols_lock);
380         list_del_init(&syme->node);
381         pthread_mutex_unlock(&active_symbols_lock);
382 }
383
384 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
385 {
386         struct rb_node **p = &tree->rb_node;
387         struct rb_node *parent = NULL;
388         struct sym_entry *iter;
389
390         while (*p != NULL) {
391                 parent = *p;
392                 iter = rb_entry(parent, struct sym_entry, rb_node);
393
394                 if (se->weight > iter->weight)
395                         p = &(*p)->rb_left;
396                 else
397                         p = &(*p)->rb_right;
398         }
399
400         rb_link_node(&se->rb_node, parent, p);
401         rb_insert_color(&se->rb_node, tree);
402 }
403
404 static void print_sym_table(void)
405 {
406         int printed = 0, j;
407         int counter, snap = !display_weighted ? sym_counter : 0;
408         float samples_per_sec = samples/delay_secs;
409         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
410         float sum_ksamples = 0.0;
411         struct sym_entry *syme, *n;
412         struct rb_root tmp = RB_ROOT;
413         struct rb_node *nd;
414
415         samples = userspace_samples = 0;
416
417         /* Sort the active symbols */
418         pthread_mutex_lock(&active_symbols_lock);
419         syme = list_entry(active_symbols.next, struct sym_entry, node);
420         pthread_mutex_unlock(&active_symbols_lock);
421
422         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
423                 syme->snap_count = syme->count[snap];
424                 if (syme->snap_count != 0) {
425                         syme->weight = sym_weight(syme);
426                         rb_insert_active_sym(&tmp, syme);
427                         sum_ksamples += syme->snap_count;
428
429                         for (j = 0; j < nr_counters; j++)
430                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
431                 } else
432                         list_remove_active_sym(syme);
433         }
434
435         puts(CONSOLE_CLEAR);
436
437         printf(
438 "------------------------------------------------------------------------------\n");
439         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
440                 samples_per_sec,
441                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
442
443         if (nr_counters == 1 || !display_weighted) {
444                 printf("%Ld", (u64)attrs[0].sample_period);
445                 if (freq)
446                         printf("Hz ");
447                 else
448                         printf(" ");
449         }
450
451         if (!display_weighted)
452                 printf("%s", event_name(sym_counter));
453         else for (counter = 0; counter < nr_counters; counter++) {
454                 if (counter)
455                         printf("/");
456
457                 printf("%s", event_name(counter));
458         }
459
460         printf( "], ");
461
462         if (target_pid != -1)
463                 printf(" (target_pid: %d", target_pid);
464         else
465                 printf(" (all");
466
467         if (profile_cpu != -1)
468                 printf(", cpu: %d)\n", profile_cpu);
469         else {
470                 if (target_pid != -1)
471                         printf(")\n");
472                 else
473                         printf(", %d CPUs)\n", nr_cpus);
474         }
475
476         printf("------------------------------------------------------------------------------\n\n");
477
478         if (sym_filter_entry) {
479                 show_details(sym_filter_entry);
480                 return;
481         }
482
483         if (nr_counters == 1)
484                 printf("             samples    pcnt");
485         else
486                 printf("   weight    samples    pcnt");
487
488         if (verbose)
489                 printf("         RIP       ");
490         printf("   kernel function\n");
491         printf("   %s    _______   _____",
492                nr_counters == 1 ? "      " : "______");
493         if (verbose)
494                 printf("   ________________");
495         printf("   _______________\n\n");
496
497         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
498                 struct symbol *sym;
499                 double pcnt;
500
501                 syme = rb_entry(nd, struct sym_entry, rb_node);
502                 sym = (struct symbol *)(syme + 1);
503
504                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
505                         continue;
506
507                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
508                                          sum_ksamples));
509
510                 if (nr_counters == 1 || !display_weighted)
511                         printf("%20.2f - ", syme->weight);
512                 else
513                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
514
515                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
516                 if (verbose)
517                         printf(" - %016llx", sym->start);
518                 printf(" : %s", sym->name);
519                 if (sym->module)
520                         printf("\t[%s]", sym->module->name);
521                 printf("\n");
522         }
523 }
524
525 static void prompt_integer(int *target, const char *msg)
526 {
527         char *buf = malloc(0), *p;
528         size_t dummy = 0;
529         int tmp;
530
531         fprintf(stdout, "\n%s: ", msg);
532         if (getline(&buf, &dummy, stdin) < 0)
533                 return;
534
535         p = strchr(buf, '\n');
536         if (p)
537                 *p = 0;
538
539         p = buf;
540         while(*p) {
541                 if (!isdigit(*p))
542                         goto out_free;
543                 p++;
544         }
545         tmp = strtoul(buf, NULL, 10);
546         *target = tmp;
547 out_free:
548         free(buf);
549 }
550
551 static void prompt_percent(int *target, const char *msg)
552 {
553         int tmp = 0;
554
555         prompt_integer(&tmp, msg);
556         if (tmp >= 0 && tmp <= 100)
557                 *target = tmp;
558 }
559
560 static void prompt_symbol(struct sym_entry **target, const char *msg)
561 {
562         char *buf = malloc(0), *p;
563         struct sym_entry *syme = *target, *n, *found = NULL;
564         size_t dummy = 0;
565
566         /* zero counters of active symbol */
567         if (syme) {
568                 pthread_mutex_lock(&syme->source_lock);
569                 __zero_source_counters(syme);
570                 *target = NULL;
571                 pthread_mutex_unlock(&syme->source_lock);
572         }
573
574         fprintf(stdout, "\n%s: ", msg);
575         if (getline(&buf, &dummy, stdin) < 0)
576                 goto out_free;
577
578         p = strchr(buf, '\n');
579         if (p)
580                 *p = 0;
581
582         pthread_mutex_lock(&active_symbols_lock);
583         syme = list_entry(active_symbols.next, struct sym_entry, node);
584         pthread_mutex_unlock(&active_symbols_lock);
585
586         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
587                 struct symbol *sym = (struct symbol *)(syme + 1);
588
589                 if (!strcmp(buf, sym->name)) {
590                         found = syme;
591                         break;
592                 }
593         }
594
595         if (!found) {
596                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
597                 sleep(1);
598                 return;
599         } else
600                 parse_source(found);
601
602 out_free:
603         free(buf);
604 }
605
606 static void print_mapped_keys(void)
607 {
608         char *name = NULL;
609
610         if (sym_filter_entry) {
611                 struct symbol *sym = (struct symbol *)(sym_filter_entry+1);
612                 name = sym->name;
613         }
614
615         fprintf(stdout, "\nMapped keys:\n");
616         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
617         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
618
619         if (nr_counters > 1)
620                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
621
622         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
623
624         if (vmlinux_name) {
625                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
626                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
627                 fprintf(stdout, "\t[S]     stop annotation.\n");
628         }
629
630         if (nr_counters > 1)
631                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
632
633         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
634         fprintf(stdout, "\t[qQ]    quit.\n");
635 }
636
637 static int key_mapped(int c)
638 {
639         switch (c) {
640                 case 'd':
641                 case 'e':
642                 case 'f':
643                 case 'z':
644                 case 'q':
645                 case 'Q':
646                         return 1;
647                 case 'E':
648                 case 'w':
649                         return nr_counters > 1 ? 1 : 0;
650                 case 'F':
651                 case 's':
652                 case 'S':
653                         return vmlinux_name ? 1 : 0;
654                 default:
655                         break;
656         }
657
658         return 0;
659 }
660
661 static void handle_keypress(int c)
662 {
663         if (!key_mapped(c)) {
664                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
665                 struct termios tc, save;
666
667                 print_mapped_keys();
668                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
669                 fflush(stdout);
670
671                 tcgetattr(0, &save);
672                 tc = save;
673                 tc.c_lflag &= ~(ICANON | ECHO);
674                 tc.c_cc[VMIN] = 0;
675                 tc.c_cc[VTIME] = 0;
676                 tcsetattr(0, TCSANOW, &tc);
677
678                 poll(&stdin_poll, 1, -1);
679                 c = getc(stdin);
680
681                 tcsetattr(0, TCSAFLUSH, &save);
682                 if (!key_mapped(c))
683                         return;
684         }
685
686         switch (c) {
687                 case 'd':
688                         prompt_integer(&delay_secs, "Enter display delay");
689                         if (delay_secs < 1)
690                                 delay_secs = 1;
691                         break;
692                 case 'e':
693                         prompt_integer(&print_entries, "Enter display entries (lines)");
694                         break;
695                 case 'E':
696                         if (nr_counters > 1) {
697                                 int i;
698
699                                 fprintf(stderr, "\nAvailable events:");
700                                 for (i = 0; i < nr_counters; i++)
701                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
702
703                                 prompt_integer(&sym_counter, "Enter details event counter");
704
705                                 if (sym_counter >= nr_counters) {
706                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
707                                         sym_counter = 0;
708                                         sleep(1);
709                                 }
710                         } else sym_counter = 0;
711                         break;
712                 case 'f':
713                         prompt_integer(&count_filter, "Enter display event count filter");
714                         break;
715                 case 'F':
716                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
717                         break;
718                 case 'q':
719                 case 'Q':
720                         printf("exiting.\n");
721                         exit(0);
722                 case 's':
723                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
724                         break;
725                 case 'S':
726                         if (!sym_filter_entry)
727                                 break;
728                         else {
729                                 struct sym_entry *syme = sym_filter_entry;
730
731                                 pthread_mutex_lock(&syme->source_lock);
732                                 sym_filter_entry = NULL;
733                                 __zero_source_counters(syme);
734                                 pthread_mutex_unlock(&syme->source_lock);
735                         }
736                         break;
737                 case 'w':
738                         display_weighted = ~display_weighted;
739                         break;
740                 case 'z':
741                         zero = ~zero;
742                         break;
743                 default:
744                         break;
745         }
746 }
747
748 static void *display_thread(void *arg __used)
749 {
750         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
751         struct termios tc, save;
752         int delay_msecs, c;
753
754         tcgetattr(0, &save);
755         tc = save;
756         tc.c_lflag &= ~(ICANON | ECHO);
757         tc.c_cc[VMIN] = 0;
758         tc.c_cc[VTIME] = 0;
759
760 repeat:
761         delay_msecs = delay_secs * 1000;
762         tcsetattr(0, TCSANOW, &tc);
763         /* trash return*/
764         getc(stdin);
765
766         do {
767                 print_sym_table();
768         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
769
770         c = getc(stdin);
771         tcsetattr(0, TCSAFLUSH, &save);
772
773         handle_keypress(c);
774         goto repeat;
775
776         return NULL;
777 }
778
779 /* Tag samples to be skipped. */
780 static const char *skip_symbols[] = {
781         "default_idle",
782         "cpu_idle",
783         "enter_idle",
784         "exit_idle",
785         "mwait_idle",
786         "mwait_idle_with_hints",
787         "poll_idle",
788         "ppc64_runlatch_off",
789         "pseries_dedicated_idle_sleep",
790         NULL
791 };
792
793 static int symbol_filter(struct dso *self, struct symbol *sym)
794 {
795         struct sym_entry *syme;
796         const char *name = sym->name;
797         int i;
798
799         /*
800          * ppc64 uses function descriptors and appends a '.' to the
801          * start of every instruction address. Remove it.
802          */
803         if (name[0] == '.')
804                 name++;
805
806         if (!strcmp(name, "_text") ||
807             !strcmp(name, "_etext") ||
808             !strcmp(name, "_sinittext") ||
809             !strncmp("init_module", name, 11) ||
810             !strncmp("cleanup_module", name, 14) ||
811             strstr(name, "_text_start") ||
812             strstr(name, "_text_end"))
813                 return 1;
814
815         syme = dso__sym_priv(self, sym);
816         pthread_mutex_init(&syme->source_lock, NULL);
817         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
818                 sym_filter_entry = syme;
819
820         for (i = 0; skip_symbols[i]; i++) {
821                 if (!strcmp(skip_symbols[i], name)) {
822                         syme->skip = 1;
823                         break;
824                 }
825         }
826
827         return 0;
828 }
829
830 static int parse_symbols(void)
831 {
832         struct rb_node *node;
833         struct symbol  *sym;
834         int use_modules = vmlinux_name ? 1 : 0;
835
836         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
837         if (kernel_dso == NULL)
838                 return -1;
839
840         if (dso__load_kernel(kernel_dso, vmlinux_name, symbol_filter, verbose, use_modules) <= 0)
841                 goto out_delete_dso;
842
843         node = rb_first(&kernel_dso->syms);
844         sym = rb_entry(node, struct symbol, rb_node);
845         min_ip = sym->start;
846
847         node = rb_last(&kernel_dso->syms);
848         sym = rb_entry(node, struct symbol, rb_node);
849         max_ip = sym->end;
850
851         if (dump_symtab)
852                 dso__fprintf(kernel_dso, stderr);
853
854         return 0;
855
856 out_delete_dso:
857         dso__delete(kernel_dso);
858         kernel_dso = NULL;
859         return -1;
860 }
861
862 /*
863  * Binary search in the histogram table and record the hit:
864  */
865 static void record_ip(u64 ip, int counter)
866 {
867         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
868
869         if (sym != NULL) {
870                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
871
872                 if (!syme->skip) {
873                         syme->count[counter]++;
874                         record_precise_ip(syme, counter, ip);
875                         pthread_mutex_lock(&active_symbols_lock);
876                         if (list_empty(&syme->node) || !syme->node.next)
877                                 __list_insert_active_sym(syme);
878                         pthread_mutex_unlock(&active_symbols_lock);
879                         return;
880                 }
881         }
882
883         samples--;
884 }
885
886 static void process_event(u64 ip, int counter, int user)
887 {
888         samples++;
889
890         if (user) {
891                 userspace_samples++;
892                 return;
893         }
894
895         record_ip(ip, counter);
896 }
897
898 struct mmap_data {
899         int                     counter;
900         void                    *base;
901         int                     mask;
902         unsigned int            prev;
903 };
904
905 static unsigned int mmap_read_head(struct mmap_data *md)
906 {
907         struct perf_event_mmap_page *pc = md->base;
908         int head;
909
910         head = pc->data_head;
911         rmb();
912
913         return head;
914 }
915
916 struct timeval last_read, this_read;
917
918 static void mmap_read_counter(struct mmap_data *md)
919 {
920         unsigned int head = mmap_read_head(md);
921         unsigned int old = md->prev;
922         unsigned char *data = md->base + page_size;
923         int diff;
924
925         gettimeofday(&this_read, NULL);
926
927         /*
928          * If we're further behind than half the buffer, there's a chance
929          * the writer will bite our tail and mess up the samples under us.
930          *
931          * If we somehow ended up ahead of the head, we got messed up.
932          *
933          * In either case, truncate and restart at head.
934          */
935         diff = head - old;
936         if (diff > md->mask / 2 || diff < 0) {
937                 struct timeval iv;
938                 unsigned long msecs;
939
940                 timersub(&this_read, &last_read, &iv);
941                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
942
943                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
944                                 "  Last read %lu msecs ago.\n", msecs);
945
946                 /*
947                  * head points to a known good entry, start there.
948                  */
949                 old = head;
950         }
951
952         last_read = this_read;
953
954         for (; old != head;) {
955                 event_t *event = (event_t *)&data[old & md->mask];
956
957                 event_t event_copy;
958
959                 size_t size = event->header.size;
960
961                 /*
962                  * Event straddles the mmap boundary -- header should always
963                  * be inside due to u64 alignment of output.
964                  */
965                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
966                         unsigned int offset = old;
967                         unsigned int len = min(sizeof(*event), size), cpy;
968                         void *dst = &event_copy;
969
970                         do {
971                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
972                                 memcpy(dst, &data[offset & md->mask], cpy);
973                                 offset += cpy;
974                                 dst += cpy;
975                                 len -= cpy;
976                         } while (len);
977
978                         event = &event_copy;
979                 }
980
981                 old += size;
982
983                 if (event->header.type == PERF_RECORD_SAMPLE) {
984                         int user =
985         (event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK) == PERF_RECORD_MISC_USER;
986                         process_event(event->ip.ip, md->counter, user);
987                 }
988         }
989
990         md->prev = old;
991 }
992
993 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
994 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
995
996 static void mmap_read(void)
997 {
998         int i, counter;
999
1000         for (i = 0; i < nr_cpus; i++) {
1001                 for (counter = 0; counter < nr_counters; counter++)
1002                         mmap_read_counter(&mmap_array[i][counter]);
1003         }
1004 }
1005
1006 int nr_poll;
1007 int group_fd;
1008
1009 static void start_counter(int i, int counter)
1010 {
1011         struct perf_event_attr *attr;
1012         int cpu;
1013
1014         cpu = profile_cpu;
1015         if (target_pid == -1 && profile_cpu == -1)
1016                 cpu = i;
1017
1018         attr = attrs + counter;
1019
1020         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1021         attr->freq              = freq;
1022         attr->inherit           = (cpu < 0) && inherit;
1023
1024 try_again:
1025         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1026
1027         if (fd[i][counter] < 0) {
1028                 int err = errno;
1029
1030                 if (err == EPERM || err == EACCES)
1031                         die("No permission - are you root?\n");
1032                 /*
1033                  * If it's cycles then fall back to hrtimer
1034                  * based cpu-clock-tick sw counter, which
1035                  * is always available even if no PMU support:
1036                  */
1037                 if (attr->type == PERF_TYPE_HARDWARE
1038                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1039
1040                         if (verbose)
1041                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1042
1043                         attr->type = PERF_TYPE_SOFTWARE;
1044                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1045                         goto try_again;
1046                 }
1047                 printf("\n");
1048                 error("perfcounter syscall returned with %d (%s)\n",
1049                         fd[i][counter], strerror(err));
1050                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1051                 exit(-1);
1052         }
1053         assert(fd[i][counter] >= 0);
1054         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1055
1056         /*
1057          * First counter acts as the group leader:
1058          */
1059         if (group && group_fd == -1)
1060                 group_fd = fd[i][counter];
1061
1062         event_array[nr_poll].fd = fd[i][counter];
1063         event_array[nr_poll].events = POLLIN;
1064         nr_poll++;
1065
1066         mmap_array[i][counter].counter = counter;
1067         mmap_array[i][counter].prev = 0;
1068         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1069         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1070                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1071         if (mmap_array[i][counter].base == MAP_FAILED)
1072                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1073 }
1074
1075 static int __cmd_top(void)
1076 {
1077         pthread_t thread;
1078         int i, counter;
1079         int ret;
1080
1081         for (i = 0; i < nr_cpus; i++) {
1082                 group_fd = -1;
1083                 for (counter = 0; counter < nr_counters; counter++)
1084                         start_counter(i, counter);
1085         }
1086
1087         /* Wait for a minimal set of events before starting the snapshot */
1088         poll(event_array, nr_poll, 100);
1089
1090         mmap_read();
1091
1092         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1093                 printf("Could not create display thread.\n");
1094                 exit(-1);
1095         }
1096
1097         if (realtime_prio) {
1098                 struct sched_param param;
1099
1100                 param.sched_priority = realtime_prio;
1101                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1102                         printf("Could not set realtime priority.\n");
1103                         exit(-1);
1104                 }
1105         }
1106
1107         while (1) {
1108                 int hits = samples;
1109
1110                 mmap_read();
1111
1112                 if (hits == samples)
1113                         ret = poll(event_array, nr_poll, 100);
1114         }
1115
1116         return 0;
1117 }
1118
1119 static const char * const top_usage[] = {
1120         "perf top [<options>]",
1121         NULL
1122 };
1123
1124 static const struct option options[] = {
1125         OPT_CALLBACK('e', "event", NULL, "event",
1126                      "event selector. use 'perf list' to list available events",
1127                      parse_events),
1128         OPT_INTEGER('c', "count", &default_interval,
1129                     "event period to sample"),
1130         OPT_INTEGER('p', "pid", &target_pid,
1131                     "profile events on existing pid"),
1132         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1133                             "system-wide collection from all CPUs"),
1134         OPT_INTEGER('C', "CPU", &profile_cpu,
1135                     "CPU to profile on"),
1136         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1137         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1138                     "number of mmap data pages"),
1139         OPT_INTEGER('r', "realtime", &realtime_prio,
1140                     "collect data with this RT SCHED_FIFO priority"),
1141         OPT_INTEGER('d', "delay", &delay_secs,
1142                     "number of seconds to delay between refreshes"),
1143         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1144                             "dump the symbol table used for profiling"),
1145         OPT_INTEGER('f', "count-filter", &count_filter,
1146                     "only display functions with more events than this"),
1147         OPT_BOOLEAN('g', "group", &group,
1148                             "put the counters into a counter group"),
1149         OPT_BOOLEAN('i', "inherit", &inherit,
1150                     "child tasks inherit counters"),
1151         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1152                     "symbol to annotate - requires -k option"),
1153         OPT_BOOLEAN('z', "zero", &zero,
1154                     "zero history across updates"),
1155         OPT_INTEGER('F', "freq", &freq,
1156                     "profile at this frequency"),
1157         OPT_INTEGER('E', "entries", &print_entries,
1158                     "display this many functions"),
1159         OPT_BOOLEAN('v', "verbose", &verbose,
1160                     "be more verbose (show counter open errors, etc)"),
1161         OPT_END()
1162 };
1163
1164 int cmd_top(int argc, const char **argv, const char *prefix __used)
1165 {
1166         int counter;
1167
1168         symbol__init();
1169
1170         page_size = sysconf(_SC_PAGE_SIZE);
1171
1172         argc = parse_options(argc, argv, options, top_usage, 0);
1173         if (argc)
1174                 usage_with_options(top_usage, options);
1175
1176         if (freq) {
1177                 default_interval = freq;
1178                 freq = 1;
1179         }
1180
1181         /* CPU and PID are mutually exclusive */
1182         if (target_pid != -1 && profile_cpu != -1) {
1183                 printf("WARNING: PID switch overriding CPU\n");
1184                 sleep(1);
1185                 profile_cpu = -1;
1186         }
1187
1188         if (!nr_counters)
1189                 nr_counters = 1;
1190
1191         if (delay_secs < 1)
1192                 delay_secs = 1;
1193
1194         parse_symbols();
1195         parse_source(sym_filter_entry);
1196
1197         /*
1198          * Fill in the ones not specifically initialized via -c:
1199          */
1200         for (counter = 0; counter < nr_counters; counter++) {
1201                 if (attrs[counter].sample_period)
1202                         continue;
1203
1204                 attrs[counter].sample_period = default_interval;
1205         }
1206
1207         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1208         assert(nr_cpus <= MAX_NR_CPUS);
1209         assert(nr_cpus >= 0);
1210
1211         if (target_pid != -1 || profile_cpu != -1)
1212                 nr_cpus = 1;
1213
1214         return __cmd_top();
1215 }