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