perf_counter tools: Fix/resurrect perf top annotation in a simple interactive form
[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 <assert.h>
31 #include <fcntl.h>
32
33 #include <stdio.h>
34 #include <termios.h>
35 #include <unistd.h>
36
37 #include <errno.h>
38 #include <time.h>
39 #include <sched.h>
40 #include <pthread.h>
41
42 #include <sys/syscall.h>
43 #include <sys/ioctl.h>
44 #include <sys/poll.h>
45 #include <sys/prctl.h>
46 #include <sys/wait.h>
47 #include <sys/uio.h>
48 #include <sys/mman.h>
49
50 #include <linux/unistd.h>
51 #include <linux/types.h>
52
53 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
54
55 static int                      system_wide                     =  0;
56
57 static int                      default_interval                = 100000;
58
59 static int                      count_filter                    =  5;
60 static int                      print_entries                   = 15;
61
62 static int                      target_pid                      = -1;
63 static int                      inherit                         =  0;
64 static int                      profile_cpu                     = -1;
65 static int                      nr_cpus                         =  0;
66 static unsigned int             realtime_prio                   =  0;
67 static int                      group                           =  0;
68 static unsigned int             page_size;
69 static unsigned int             mmap_pages                      = 16;
70 static int                      freq                            =  0;
71 static int                      verbose                         =  0;
72 static char                     *vmlinux                        =  NULL;
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
94 /*
95  * Symbols
96  */
97
98 static u64                      min_ip;
99 static u64                      max_ip = -1ll;
100
101 struct sym_entry {
102         struct rb_node          rb_node;
103         struct list_head        node;
104         unsigned long           count[MAX_COUNTERS];
105         unsigned long           snap_count;
106         double                  weight;
107         int                     skip;
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 module *module;
122         struct section *section = NULL;
123         FILE *file;
124         char command[PATH_MAX*2], *path = vmlinux;
125         u64 start, end, 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         module = sym->module;
137
138         if (module)
139                 path = module->path;
140         if (!path)
141                 return;
142
143         start = sym->obj_start;
144         if (!start)
145                 start = sym->start;
146
147         if (module) {
148                 section = module->sections->find_section(module->sections, ".text");
149                 if (section)
150                         start -= section->vma;
151         }
152
153         end = start + sym->end - sym->start + 1;
154         len = sym->end - sym->start;
155
156         sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s", start, end, path);
157
158         file = popen(command, "r");
159         if (!file)
160                 return;
161
162         pthread_mutex_lock(&syme->source_lock);
163         syme->lines_tail = &syme->lines;
164         while (!feof(file)) {
165                 struct source_line *src;
166                 size_t dummy = 0;
167                 char *c;
168
169                 src = malloc(sizeof(struct source_line));
170                 assert(src != NULL);
171                 memset(src, 0, sizeof(struct source_line));
172
173                 if (getline(&src->line, &dummy, file) < 0)
174                         break;
175                 if (!src->line)
176                         break;
177
178                 c = strchr(src->line, '\n');
179                 if (c)
180                         *c = 0;
181
182                 src->next = NULL;
183                 *syme->lines_tail = src;
184                 syme->lines_tail = &src->next;
185
186                 if (strlen(src->line)>8 && src->line[8] == ':') {
187                         src->eip = strtoull(src->line, NULL, 16);
188                         if (section)
189                                 src->eip += section->vma;
190                 }
191                 if (strlen(src->line)>8 && src->line[16] == ':') {
192                         src->eip = strtoull(src->line, NULL, 16);
193                         if (section)
194                                 src->eip += section->vma;
195                 }
196         }
197         pclose(file);
198 out_assign:
199         sym_filter_entry = syme;
200         pthread_mutex_unlock(&syme->source_lock);
201 }
202
203 static void __zero_source_counters(struct sym_entry *syme)
204 {
205         int i;
206         struct source_line *line;
207
208         line = syme->lines;
209         while (line) {
210                 for (i = 0; i < nr_counters; i++)
211                         line->count[i] = 0;
212                 line = line->next;
213         }
214 }
215
216 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
217 {
218         struct source_line *line;
219
220         if (syme != sym_filter_entry)
221                 return;
222
223         if (pthread_mutex_trylock(&syme->source_lock))
224                 return;
225
226         if (!syme->source)
227                 goto out_unlock;
228
229         for (line = syme->lines; line; line = line->next) {
230                 if (line->eip == ip) {
231                         line->count[counter]++;
232                         break;
233                 }
234                 if (line->eip > ip)
235                         break;
236         }
237 out_unlock:
238         pthread_mutex_unlock(&syme->source_lock);
239 }
240
241 static void lookup_sym_source(struct sym_entry *syme)
242 {
243         struct symbol *symbol = (struct symbol *)(syme + 1);
244         struct source_line *line;
245         char pattern[PATH_MAX];
246         char *idx;
247
248         sprintf(pattern, "<%s>:", symbol->name);
249
250         if (symbol->module) {
251                 idx = strstr(pattern, "\t");
252                 if (idx)
253                         *idx = 0;
254         }
255
256         pthread_mutex_lock(&syme->source_lock);
257         for (line = syme->lines; line; line = line->next) {
258                 if (strstr(line->line, pattern)) {
259                         syme->source = line;
260                         break;
261                 }
262         }
263         pthread_mutex_unlock(&syme->source_lock);
264 }
265
266 static void show_lines(struct source_line *queue, int count, int total)
267 {
268         int i;
269         struct source_line *line;
270
271         line = queue;
272         for (i = 0; i < count; i++) {
273                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
274
275                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
276                 line = line->next;
277         }
278 }
279
280 #define TRACE_COUNT     3
281
282 static void show_details(struct sym_entry *syme)
283 {
284         struct symbol *symbol;
285         struct source_line *line;
286         struct source_line *line_queue = NULL;
287         int displayed = 0;
288         int line_queue_count = 0, total = 0, more = 0;
289
290         if (!syme)
291                 return;
292
293         if (!syme->source)
294                 lookup_sym_source(syme);
295
296         if (!syme->source)
297                 return;
298
299         symbol = (struct symbol *)(syme + 1);
300         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
301         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
302
303         pthread_mutex_lock(&syme->source_lock);
304         line = syme->source;
305         while (line) {
306                 total += line->count[sym_counter];
307                 line = line->next;
308         }
309
310         line = syme->source;
311         while (line) {
312                 float pcnt = 0.0;
313
314                 if (!line_queue_count)
315                         line_queue = line;
316                 line_queue_count++;
317
318                 if (line->count[sym_counter])
319                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
320                 if (pcnt >= (float)sym_pcnt_filter) {
321                         if (displayed <= print_entries)
322                                 show_lines(line_queue, line_queue_count, total);
323                         else more++;
324                         displayed += line_queue_count;
325                         line_queue_count = 0;
326                         line_queue = NULL;
327                 } else if (line_queue_count > TRACE_COUNT) {
328                         line_queue = line_queue->next;
329                         line_queue_count--;
330                 }
331
332                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
333                 line = line->next;
334         }
335         pthread_mutex_unlock(&syme->source_lock);
336         if (more)
337                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
338 }
339
340 struct dso                      *kernel_dso;
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         for (counter = 1; counter < nr_counters-1; counter++)
358                 weight *= sym->count[counter];
359
360         weight /= (sym->count[counter] + 1);
361
362         return weight;
363 }
364
365 static long                     samples;
366 static long                     userspace_samples;
367 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
368
369 static void __list_insert_active_sym(struct sym_entry *syme)
370 {
371         list_add(&syme->node, &active_symbols);
372 }
373
374 static void list_remove_active_sym(struct sym_entry *syme)
375 {
376         pthread_mutex_lock(&active_symbols_lock);
377         list_del_init(&syme->node);
378         pthread_mutex_unlock(&active_symbols_lock);
379 }
380
381 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
382 {
383         struct rb_node **p = &tree->rb_node;
384         struct rb_node *parent = NULL;
385         struct sym_entry *iter;
386
387         while (*p != NULL) {
388                 parent = *p;
389                 iter = rb_entry(parent, struct sym_entry, rb_node);
390
391                 if (se->weight > iter->weight)
392                         p = &(*p)->rb_left;
393                 else
394                         p = &(*p)->rb_right;
395         }
396
397         rb_link_node(&se->rb_node, parent, p);
398         rb_insert_color(&se->rb_node, tree);
399 }
400
401 static void print_sym_table(void)
402 {
403         int printed = 0, j;
404         int counter;
405         float samples_per_sec = samples/delay_secs;
406         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
407         float sum_ksamples = 0.0;
408         struct sym_entry *syme, *n;
409         struct rb_root tmp = RB_ROOT;
410         struct rb_node *nd;
411
412         samples = userspace_samples = 0;
413
414         /* Sort the active symbols */
415         pthread_mutex_lock(&active_symbols_lock);
416         syme = list_entry(active_symbols.next, struct sym_entry, node);
417         pthread_mutex_unlock(&active_symbols_lock);
418
419         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
420                 syme->snap_count = syme->count[0];
421                 if (syme->snap_count != 0) {
422                         syme->weight = sym_weight(syme);
423                         rb_insert_active_sym(&tmp, syme);
424                         sum_ksamples += syme->snap_count;
425
426                         for (j = 0; j < nr_counters; j++)
427                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
428                 } else
429                         list_remove_active_sym(syme);
430         }
431
432         puts(CONSOLE_CLEAR);
433
434         printf(
435 "------------------------------------------------------------------------------\n");
436         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
437                 samples_per_sec,
438                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
439
440         if (nr_counters == 1) {
441                 printf("%Ld", (u64)attrs[0].sample_period);
442                 if (freq)
443                         printf("Hz ");
444                 else
445                         printf(" ");
446         }
447
448         for (counter = 0; counter < nr_counters; counter++) {
449                 if (counter)
450                         printf("/");
451
452                 printf("%s", event_name(counter));
453         }
454
455         printf( "], ");
456
457         if (target_pid != -1)
458                 printf(" (target_pid: %d", target_pid);
459         else
460                 printf(" (all");
461
462         if (profile_cpu != -1)
463                 printf(", cpu: %d)\n", profile_cpu);
464         else {
465                 if (target_pid != -1)
466                         printf(")\n");
467                 else
468                         printf(", %d CPUs)\n", nr_cpus);
469         }
470
471         printf("------------------------------------------------------------------------------\n\n");
472
473         if (sym_filter_entry) {
474                 show_details(sym_filter_entry);
475                 return;
476         }
477
478         if (nr_counters == 1)
479                 printf("             samples    pcnt");
480         else
481                 printf("  weight     samples    pcnt");
482
483         printf("         RIP          kernel function\n"
484                        "  ______     _______   _____   ________________   _______________\n\n"
485         );
486
487         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
488                 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
489                 struct symbol *sym = (struct symbol *)(syme + 1);
490                 double pcnt;
491
492                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
493                         continue;
494
495                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
496                                          sum_ksamples));
497
498                 if (nr_counters == 1)
499                         printf("%20.2f - ", syme->weight);
500                 else
501                         printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
502
503                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
504                 printf(" - %016llx : %s", sym->start, sym->name);
505                 if (sym->module)
506                         printf("\t[%s]", sym->module->name);
507                 printf("\n");
508         }
509 }
510
511 static void prompt_integer(int *target, const char *msg)
512 {
513         char *buf = malloc(0), *p;
514         size_t dummy = 0;
515         int tmp;
516
517         fprintf(stdout, "\n%s: ", msg);
518         if (getline(&buf, &dummy, stdin) < 0)
519                 return;
520
521         p = strchr(buf, '\n');
522         if (p)
523                 *p = 0;
524
525         p = buf;
526         while(*p) {
527                 if (!isdigit(*p))
528                         goto out_free;
529                 p++;
530         }
531         tmp = strtoul(buf, NULL, 10);
532         *target = tmp;
533 out_free:
534         free(buf);
535 }
536
537 static void prompt_percent(int *target, const char *msg)
538 {
539         int tmp = 0;
540
541         prompt_integer(&tmp, msg);
542         if (tmp >= 0 && tmp <= 100)
543                 *target = tmp;
544 }
545
546 static void prompt_symbol(struct sym_entry **target, const char *msg)
547 {
548         char *buf = malloc(0), *p;
549         struct sym_entry *syme = *target, *n, *found = NULL;
550         size_t dummy = 0;
551
552         /* zero counters of active symbol */
553         if (syme) {
554                 pthread_mutex_lock(&syme->source_lock);
555                 __zero_source_counters(syme);
556                 *target = NULL;
557                 pthread_mutex_unlock(&syme->source_lock);
558         }
559
560         fprintf(stdout, "\n%s: ", msg);
561         if (getline(&buf, &dummy, stdin) < 0)
562                 goto out_free;
563
564         p = strchr(buf, '\n');
565         if (p)
566                 *p = 0;
567
568         pthread_mutex_lock(&active_symbols_lock);
569         syme = list_entry(active_symbols.next, struct sym_entry, node);
570         pthread_mutex_unlock(&active_symbols_lock);
571
572         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
573                 struct symbol *sym = (struct symbol *)(syme + 1);
574
575                 if (!strcmp(buf, sym->name)) {
576                         found = syme;
577                         break;
578                 }
579         }
580
581         if (!found) {
582                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
583                 sleep(1);
584                 return;
585         } else
586                 parse_source(found);
587
588 out_free:
589         free(buf);
590 }
591
592 static void print_known_keys(void)
593 {
594         fprintf(stdout, "\nknown keys:\n");
595         fprintf(stdout, "\t[d]     select display delay.\n");
596         fprintf(stdout, "\t[e]     select display entries (lines).\n");
597         fprintf(stdout, "\t[E]     select annotation event counter.\n");
598         fprintf(stdout, "\t[f]     select normal display count filter.\n");
599         fprintf(stdout, "\t[F]     select annotation display count filter (percentage).\n");
600         fprintf(stdout, "\t[qQ]    quit.\n");
601         fprintf(stdout, "\t[s]     select annotation symbol and start annotation.\n");
602         fprintf(stdout, "\t[S]     stop annotation, revert to normal display.\n");
603         fprintf(stdout, "\t[z]     toggle event count zeroing.\n");
604 }
605
606 static void handle_keypress(int c)
607 {
608         int once = 0;
609 repeat:
610         switch (c) {
611                 case 'd':
612                         prompt_integer(&delay_secs, "Enter display delay");
613                         break;
614                 case 'e':
615                         prompt_integer(&print_entries, "Enter display entries (lines)");
616                         break;
617                 case 'E':
618                         if (nr_counters > 1) {
619                                 int i;
620
621                                 fprintf(stderr, "\nAvailable events:");
622                                 for (i = 0; i < nr_counters; i++)
623                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
624
625                                 prompt_integer(&sym_counter, "Enter details event counter");
626
627                                 if (sym_counter >= nr_counters) {
628                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
629                                         sym_counter = 0;
630                                         sleep(1);
631                                 }
632                         } else sym_counter = 0;
633                         break;
634                 case 'f':
635                         prompt_integer(&count_filter, "Enter display event count filter");
636                         break;
637                 case 'F':
638                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
639                         break;
640                 case 'q':
641                 case 'Q':
642                         printf("exiting.\n");
643                         exit(0);
644                 case 's':
645                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
646                         break;
647                 case 'S':
648                         if (!sym_filter_entry)
649                                 break;
650                         else {
651                                 struct sym_entry *syme = sym_filter_entry;
652
653                                 pthread_mutex_lock(&syme->source_lock);
654                                 sym_filter_entry = NULL;
655                                 __zero_source_counters(syme);
656                                 pthread_mutex_unlock(&syme->source_lock);
657                         }
658                         break;
659                 case 'z':
660                         zero = ~zero;
661                         break;
662                 default: {
663                         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
664                         struct termios tc, save;
665
666                         if (!once) {
667                                 print_known_keys();
668                                 once++;
669                         }
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                         goto repeat;
683                 }
684         }
685 }
686
687 static void *display_thread(void *arg __used)
688 {
689         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
690         struct termios tc, save;
691         int delay_msecs, c;
692
693         tcgetattr(0, &save);
694         tc = save;
695         tc.c_lflag &= ~(ICANON | ECHO);
696         tc.c_cc[VMIN] = 0;
697         tc.c_cc[VTIME] = 0;
698 repeat:
699         delay_msecs = delay_secs * 1000;
700         tcsetattr(0, TCSANOW, &tc);
701         /* trash return*/
702         getc(stdin);
703
704         do {
705                 print_sym_table();
706         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
707
708         c = getc(stdin);
709         tcsetattr(0, TCSAFLUSH, &save);
710
711         handle_keypress(c);
712         goto repeat;
713
714         return NULL;
715 }
716
717 /* Tag samples to be skipped. */
718 static const char *skip_symbols[] = {
719         "default_idle",
720         "cpu_idle",
721         "enter_idle",
722         "exit_idle",
723         "mwait_idle",
724         "mwait_idle_with_hints",
725         "ppc64_runlatch_off",
726         "pseries_dedicated_idle_sleep",
727         NULL
728 };
729
730 static int symbol_filter(struct dso *self, struct symbol *sym)
731 {
732         struct sym_entry *syme;
733         const char *name = sym->name;
734         int i;
735
736         /*
737          * ppc64 uses function descriptors and appends a '.' to the
738          * start of every instruction address. Remove it.
739          */
740         if (name[0] == '.')
741                 name++;
742
743         if (!strcmp(name, "_text") ||
744             !strcmp(name, "_etext") ||
745             !strcmp(name, "_sinittext") ||
746             !strncmp("init_module", name, 11) ||
747             !strncmp("cleanup_module", name, 14) ||
748             strstr(name, "_text_start") ||
749             strstr(name, "_text_end"))
750                 return 1;
751
752         syme = dso__sym_priv(self, sym);
753         pthread_mutex_init(&syme->source_lock, NULL);
754         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
755                 sym_filter_entry = syme;
756
757         for (i = 0; skip_symbols[i]; i++) {
758                 if (!strcmp(skip_symbols[i], name)) {
759                         syme->skip = 1;
760                         break;
761                 }
762         }
763
764         return 0;
765 }
766
767 static int parse_symbols(void)
768 {
769         struct rb_node *node;
770         struct symbol  *sym;
771         int modules = vmlinux ? 1 : 0;
772
773         kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
774         if (kernel_dso == NULL)
775                 return -1;
776
777         if (dso__load_kernel(kernel_dso, vmlinux, symbol_filter, verbose, modules) <= 0)
778                 goto out_delete_dso;
779
780         node = rb_first(&kernel_dso->syms);
781         sym = rb_entry(node, struct symbol, rb_node);
782         min_ip = sym->start;
783
784         node = rb_last(&kernel_dso->syms);
785         sym = rb_entry(node, struct symbol, rb_node);
786         max_ip = sym->end;
787
788         if (dump_symtab)
789                 dso__fprintf(kernel_dso, stderr);
790
791         return 0;
792
793 out_delete_dso:
794         dso__delete(kernel_dso);
795         kernel_dso = NULL;
796         return -1;
797 }
798
799 /*
800  * Binary search in the histogram table and record the hit:
801  */
802 static void record_ip(u64 ip, int counter)
803 {
804         struct symbol *sym = dso__find_symbol(kernel_dso, ip);
805
806         if (sym != NULL) {
807                 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
808
809                 if (!syme->skip) {
810                         syme->count[counter]++;
811                         record_precise_ip(syme, counter, ip);
812                         pthread_mutex_lock(&active_symbols_lock);
813                         if (list_empty(&syme->node) || !syme->node.next)
814                                 __list_insert_active_sym(syme);
815                         pthread_mutex_unlock(&active_symbols_lock);
816                         return;
817                 }
818         }
819
820         samples--;
821 }
822
823 static void process_event(u64 ip, int counter, int user)
824 {
825         samples++;
826
827         if (user) {
828                 userspace_samples++;
829                 return;
830         }
831
832         record_ip(ip, counter);
833 }
834
835 struct mmap_data {
836         int                     counter;
837         void                    *base;
838         int                     mask;
839         unsigned int            prev;
840 };
841
842 static unsigned int mmap_read_head(struct mmap_data *md)
843 {
844         struct perf_counter_mmap_page *pc = md->base;
845         int head;
846
847         head = pc->data_head;
848         rmb();
849
850         return head;
851 }
852
853 struct timeval last_read, this_read;
854
855 static void mmap_read_counter(struct mmap_data *md)
856 {
857         unsigned int head = mmap_read_head(md);
858         unsigned int old = md->prev;
859         unsigned char *data = md->base + page_size;
860         int diff;
861
862         gettimeofday(&this_read, NULL);
863
864         /*
865          * If we're further behind than half the buffer, there's a chance
866          * the writer will bite our tail and mess up the samples under us.
867          *
868          * If we somehow ended up ahead of the head, we got messed up.
869          *
870          * In either case, truncate and restart at head.
871          */
872         diff = head - old;
873         if (diff > md->mask / 2 || diff < 0) {
874                 struct timeval iv;
875                 unsigned long msecs;
876
877                 timersub(&this_read, &last_read, &iv);
878                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
879
880                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
881                                 "  Last read %lu msecs ago.\n", msecs);
882
883                 /*
884                  * head points to a known good entry, start there.
885                  */
886                 old = head;
887         }
888
889         last_read = this_read;
890
891         for (; old != head;) {
892                 struct ip_event {
893                         struct perf_event_header header;
894                         u64 ip;
895                         u32 pid, target_pid;
896                 };
897                 struct mmap_event {
898                         struct perf_event_header header;
899                         u32 pid, target_pid;
900                         u64 start;
901                         u64 len;
902                         u64 pgoff;
903                         char filename[PATH_MAX];
904                 };
905
906                 typedef union event_union {
907                         struct perf_event_header header;
908                         struct ip_event ip;
909                         struct mmap_event mmap;
910                 } event_t;
911
912                 event_t *event = (event_t *)&data[old & md->mask];
913
914                 event_t event_copy;
915
916                 size_t size = event->header.size;
917
918                 /*
919                  * Event straddles the mmap boundary -- header should always
920                  * be inside due to u64 alignment of output.
921                  */
922                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
923                         unsigned int offset = old;
924                         unsigned int len = min(sizeof(*event), size), cpy;
925                         void *dst = &event_copy;
926
927                         do {
928                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
929                                 memcpy(dst, &data[offset & md->mask], cpy);
930                                 offset += cpy;
931                                 dst += cpy;
932                                 len -= cpy;
933                         } while (len);
934
935                         event = &event_copy;
936                 }
937
938                 old += size;
939
940                 if (event->header.type == PERF_EVENT_SAMPLE) {
941                         int user =
942         (event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
943                         process_event(event->ip.ip, md->counter, user);
944                 }
945         }
946
947         md->prev = old;
948 }
949
950 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
951 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
952
953 static void mmap_read(void)
954 {
955         int i, counter;
956
957         for (i = 0; i < nr_cpus; i++) {
958                 for (counter = 0; counter < nr_counters; counter++)
959                         mmap_read_counter(&mmap_array[i][counter]);
960         }
961 }
962
963 int nr_poll;
964 int group_fd;
965
966 static void start_counter(int i, int counter)
967 {
968         struct perf_counter_attr *attr;
969         int cpu;
970
971         cpu = profile_cpu;
972         if (target_pid == -1 && profile_cpu == -1)
973                 cpu = i;
974
975         attr = attrs + counter;
976
977         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
978         attr->freq              = freq;
979         attr->inherit           = (cpu < 0) && inherit;
980
981 try_again:
982         fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
983
984         if (fd[i][counter] < 0) {
985                 int err = errno;
986
987                 if (err == EPERM)
988                         die("No permission - are you root?\n");
989                 /*
990                  * If it's cycles then fall back to hrtimer
991                  * based cpu-clock-tick sw counter, which
992                  * is always available even if no PMU support:
993                  */
994                 if (attr->type == PERF_TYPE_HARDWARE
995                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
996
997                         if (verbose)
998                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
999
1000                         attr->type = PERF_TYPE_SOFTWARE;
1001                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1002                         goto try_again;
1003                 }
1004                 printf("\n");
1005                 error("perfcounter syscall returned with %d (%s)\n",
1006                         fd[i][counter], strerror(err));
1007                 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
1008                 exit(-1);
1009         }
1010         assert(fd[i][counter] >= 0);
1011         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1012
1013         /*
1014          * First counter acts as the group leader:
1015          */
1016         if (group && group_fd == -1)
1017                 group_fd = fd[i][counter];
1018
1019         event_array[nr_poll].fd = fd[i][counter];
1020         event_array[nr_poll].events = POLLIN;
1021         nr_poll++;
1022
1023         mmap_array[i][counter].counter = counter;
1024         mmap_array[i][counter].prev = 0;
1025         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1026         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1027                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1028         if (mmap_array[i][counter].base == MAP_FAILED)
1029                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1030 }
1031
1032 static int __cmd_top(void)
1033 {
1034         pthread_t thread;
1035         int i, counter;
1036         int ret;
1037
1038         for (i = 0; i < nr_cpus; i++) {
1039                 group_fd = -1;
1040                 for (counter = 0; counter < nr_counters; counter++)
1041                         start_counter(i, counter);
1042         }
1043
1044         /* Wait for a minimal set of events before starting the snapshot */
1045         poll(event_array, nr_poll, 100);
1046
1047         mmap_read();
1048
1049         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1050                 printf("Could not create display thread.\n");
1051                 exit(-1);
1052         }
1053
1054         if (realtime_prio) {
1055                 struct sched_param param;
1056
1057                 param.sched_priority = realtime_prio;
1058                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1059                         printf("Could not set realtime priority.\n");
1060                         exit(-1);
1061                 }
1062         }
1063
1064         while (1) {
1065                 int hits = samples;
1066
1067                 mmap_read();
1068
1069                 if (hits == samples)
1070                         ret = poll(event_array, nr_poll, 100);
1071         }
1072
1073         return 0;
1074 }
1075
1076 static const char * const top_usage[] = {
1077         "perf top [<options>]",
1078         NULL
1079 };
1080
1081 static const struct option options[] = {
1082         OPT_CALLBACK('e', "event", NULL, "event",
1083                      "event selector. use 'perf list' to list available events",
1084                      parse_events),
1085         OPT_INTEGER('c', "count", &default_interval,
1086                     "event period to sample"),
1087         OPT_INTEGER('p', "pid", &target_pid,
1088                     "profile events on existing pid"),
1089         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1090                             "system-wide collection from all CPUs"),
1091         OPT_INTEGER('C', "CPU", &profile_cpu,
1092                     "CPU to profile on"),
1093         OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1094         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1095                     "number of mmap data pages"),
1096         OPT_INTEGER('r', "realtime", &realtime_prio,
1097                     "collect data with this RT SCHED_FIFO priority"),
1098         OPT_INTEGER('d', "delay", &delay_secs,
1099                     "number of seconds to delay between refreshes"),
1100         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1101                             "dump the symbol table used for profiling"),
1102         OPT_INTEGER('f', "count-filter", &count_filter,
1103                     "only display functions with more events than this"),
1104         OPT_BOOLEAN('g', "group", &group,
1105                             "put the counters into a counter group"),
1106         OPT_BOOLEAN('i', "inherit", &inherit,
1107                     "child tasks inherit counters"),
1108         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1109                     "symbol to annotate - requires -k option"),
1110         OPT_BOOLEAN('z', "zero", &zero,
1111                     "zero history across updates"),
1112         OPT_INTEGER('F', "freq", &freq,
1113                     "profile at this frequency"),
1114         OPT_INTEGER('E', "entries", &print_entries,
1115                     "display this many functions"),
1116         OPT_BOOLEAN('v', "verbose", &verbose,
1117                     "be more verbose (show counter open errors, etc)"),
1118         OPT_END()
1119 };
1120
1121 int cmd_top(int argc, const char **argv, const char *prefix __used)
1122 {
1123         int counter;
1124
1125         symbol__init();
1126
1127         page_size = sysconf(_SC_PAGE_SIZE);
1128
1129         argc = parse_options(argc, argv, options, top_usage, 0);
1130         if (argc)
1131                 usage_with_options(top_usage, options);
1132
1133         if (freq) {
1134                 default_interval = freq;
1135                 freq = 1;
1136         }
1137
1138         /* CPU and PID are mutually exclusive */
1139         if (target_pid != -1 && profile_cpu != -1) {
1140                 printf("WARNING: PID switch overriding CPU\n");
1141                 sleep(1);
1142                 profile_cpu = -1;
1143         }
1144
1145         if (!nr_counters)
1146                 nr_counters = 1;
1147
1148         if (delay_secs < 1)
1149                 delay_secs = 1;
1150
1151         parse_symbols();
1152         parse_source(sym_filter_entry);
1153
1154         /*
1155          * Fill in the ones not specifically initialized via -c:
1156          */
1157         for (counter = 0; counter < nr_counters; counter++) {
1158                 if (attrs[counter].sample_period)
1159                         continue;
1160
1161                 attrs[counter].sample_period = default_interval;
1162         }
1163
1164         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1165         assert(nr_cpus <= MAX_NR_CPUS);
1166         assert(nr_cpus >= 0);
1167
1168         if (target_pid != -1 || profile_cpu != -1)
1169                 nr_cpus = 1;
1170
1171         return __cmd_top();
1172 }