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