ddc584b64871019f0eb08dcd447ff511ce639983
[safe/jmp/linux-2.6] / 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/color.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26 #include "util/thread.h"
27 #include "util/util.h"
28 #include <linux/rbtree.h>
29 #include "util/parse-options.h"
30 #include "util/parse-events.h"
31
32 #include "util/debug.h"
33
34 #include <assert.h>
35 #include <fcntl.h>
36
37 #include <stdio.h>
38 #include <termios.h>
39 #include <unistd.h>
40
41 #include <errno.h>
42 #include <time.h>
43 #include <sched.h>
44 #include <pthread.h>
45
46 #include <sys/syscall.h>
47 #include <sys/ioctl.h>
48 #include <sys/poll.h>
49 #include <sys/prctl.h>
50 #include <sys/wait.h>
51 #include <sys/uio.h>
52 #include <sys/mman.h>
53
54 #include <linux/unistd.h>
55 #include <linux/types.h>
56
57 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
58
59 static int                      system_wide                     =      0;
60
61 static int                      default_interval                =      0;
62
63 static int                      count_filter                    =      5;
64 static int                      print_entries;
65
66 static int                      target_pid                      =     -1;
67 static int                      inherit                         =      0;
68 static int                      profile_cpu                     =     -1;
69 static int                      nr_cpus                         =      0;
70 static unsigned int             realtime_prio                   =      0;
71 static int                      group                           =      0;
72 static unsigned int             page_size;
73 static unsigned int             mmap_pages                      =     16;
74 static int                      freq                            =   1000; /* 1 KHz */
75
76 static int                      delay_secs                      =      2;
77 static int                      zero                            =      0;
78 static int                      dump_symtab                     =      0;
79
80 static bool                     hide_kernel_symbols             =  false;
81 static bool                     hide_user_symbols               =  false;
82 static struct winsize           winsize;
83
84 /*
85  * Source
86  */
87
88 struct source_line {
89         u64                     eip;
90         unsigned long           count[MAX_COUNTERS];
91         char                    *line;
92         struct source_line      *next;
93 };
94
95 static char                     *sym_filter                     =   NULL;
96 struct sym_entry                *sym_filter_entry               =   NULL;
97 static int                      sym_pcnt_filter                 =      5;
98 static int                      sym_counter                     =      0;
99 static int                      display_weighted                =     -1;
100
101 /*
102  * Symbols
103  */
104
105 struct sym_entry_source {
106         struct source_line      *source;
107         struct source_line      *lines;
108         struct source_line      **lines_tail;
109         pthread_mutex_t         lock;
110 };
111
112 struct sym_entry {
113         struct rb_node          rb_node;
114         struct list_head        node;
115         unsigned long           snap_count;
116         double                  weight;
117         int                     skip;
118         u16                     name_len;
119         u8                      origin;
120         struct map              *map;
121         struct sym_entry_source *src;
122         unsigned long           count[0];
123 };
124
125 /*
126  * Source functions
127  */
128
129 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
130 {
131        return ((void *)self) + symbol_conf.priv_size;
132 }
133
134 static void get_term_dimensions(struct winsize *ws)
135 {
136         char *s = getenv("LINES");
137
138         if (s != NULL) {
139                 ws->ws_row = atoi(s);
140                 s = getenv("COLUMNS");
141                 if (s != NULL) {
142                         ws->ws_col = atoi(s);
143                         if (ws->ws_row && ws->ws_col)
144                                 return;
145                 }
146         }
147 #ifdef TIOCGWINSZ
148         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
149             ws->ws_row && ws->ws_col)
150                 return;
151 #endif
152         ws->ws_row = 25;
153         ws->ws_col = 80;
154 }
155
156 static void update_print_entries(struct winsize *ws)
157 {
158         print_entries = ws->ws_row;
159
160         if (print_entries > 9)
161                 print_entries -= 9;
162 }
163
164 static void sig_winch_handler(int sig __used)
165 {
166         get_term_dimensions(&winsize);
167         update_print_entries(&winsize);
168 }
169
170 static void parse_source(struct sym_entry *syme)
171 {
172         struct symbol *sym;
173         struct sym_entry_source *source;
174         struct map *map;
175         FILE *file;
176         char command[PATH_MAX*2];
177         const char *path;
178         u64 len;
179
180         if (!syme)
181                 return;
182
183         if (syme->src == NULL) {
184                 syme->src = zalloc(sizeof(*source));
185                 if (syme->src == NULL)
186                         return;
187                 pthread_mutex_init(&syme->src->lock, NULL);
188         }
189
190         source = syme->src;
191
192         if (source->lines) {
193                 pthread_mutex_lock(&source->lock);
194                 goto out_assign;
195         }
196
197         sym = sym_entry__symbol(syme);
198         map = syme->map;
199         path = map->dso->long_name;
200
201         len = sym->end - sym->start;
202
203         sprintf(command,
204                 "objdump --start-address=0x%016Lx "
205                          "--stop-address=0x%016Lx -dS %s",
206                 map->unmap_ip(map, sym->start),
207                 map->unmap_ip(map, sym->end), path);
208
209         file = popen(command, "r");
210         if (!file)
211                 return;
212
213         pthread_mutex_lock(&source->lock);
214         source->lines_tail = &source->lines;
215         while (!feof(file)) {
216                 struct source_line *src;
217                 size_t dummy = 0;
218                 char *c;
219
220                 src = malloc(sizeof(struct source_line));
221                 assert(src != NULL);
222                 memset(src, 0, sizeof(struct source_line));
223
224                 if (getline(&src->line, &dummy, file) < 0)
225                         break;
226                 if (!src->line)
227                         break;
228
229                 c = strchr(src->line, '\n');
230                 if (c)
231                         *c = 0;
232
233                 src->next = NULL;
234                 *source->lines_tail = src;
235                 source->lines_tail = &src->next;
236
237                 if (strlen(src->line)>8 && src->line[8] == ':') {
238                         src->eip = strtoull(src->line, NULL, 16);
239                         src->eip = map->unmap_ip(map, src->eip);
240                 }
241                 if (strlen(src->line)>8 && src->line[16] == ':') {
242                         src->eip = strtoull(src->line, NULL, 16);
243                         src->eip = map->unmap_ip(map, src->eip);
244                 }
245         }
246         pclose(file);
247 out_assign:
248         sym_filter_entry = syme;
249         pthread_mutex_unlock(&source->lock);
250 }
251
252 static void __zero_source_counters(struct sym_entry *syme)
253 {
254         int i;
255         struct source_line *line;
256
257         line = syme->src->lines;
258         while (line) {
259                 for (i = 0; i < nr_counters; i++)
260                         line->count[i] = 0;
261                 line = line->next;
262         }
263 }
264
265 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
266 {
267         struct source_line *line;
268
269         if (syme != sym_filter_entry)
270                 return;
271
272         if (pthread_mutex_trylock(&syme->src->lock))
273                 return;
274
275         if (syme->src == NULL || syme->src->source == NULL)
276                 goto out_unlock;
277
278         for (line = syme->src->lines; line; line = line->next) {
279                 if (line->eip == ip) {
280                         line->count[counter]++;
281                         break;
282                 }
283                 if (line->eip > ip)
284                         break;
285         }
286 out_unlock:
287         pthread_mutex_unlock(&syme->src->lock);
288 }
289
290 static void lookup_sym_source(struct sym_entry *syme)
291 {
292         struct symbol *symbol = sym_entry__symbol(syme);
293         struct source_line *line;
294         char pattern[PATH_MAX];
295
296         sprintf(pattern, "<%s>:", symbol->name);
297
298         pthread_mutex_lock(&syme->src->lock);
299         for (line = syme->src->lines; line; line = line->next) {
300                 if (strstr(line->line, pattern)) {
301                         syme->src->source = line;
302                         break;
303                 }
304         }
305         pthread_mutex_unlock(&syme->src->lock);
306 }
307
308 static void show_lines(struct source_line *queue, int count, int total)
309 {
310         int i;
311         struct source_line *line;
312
313         line = queue;
314         for (i = 0; i < count; i++) {
315                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
316
317                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
318                 line = line->next;
319         }
320 }
321
322 #define TRACE_COUNT     3
323
324 static void show_details(struct sym_entry *syme)
325 {
326         struct symbol *symbol;
327         struct source_line *line;
328         struct source_line *line_queue = NULL;
329         int displayed = 0;
330         int line_queue_count = 0, total = 0, more = 0;
331
332         if (!syme)
333                 return;
334
335         if (!syme->src->source)
336                 lookup_sym_source(syme);
337
338         if (!syme->src->source)
339                 return;
340
341         symbol = sym_entry__symbol(syme);
342         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
343         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
344
345         pthread_mutex_lock(&syme->src->lock);
346         line = syme->src->source;
347         while (line) {
348                 total += line->count[sym_counter];
349                 line = line->next;
350         }
351
352         line = syme->src->source;
353         while (line) {
354                 float pcnt = 0.0;
355
356                 if (!line_queue_count)
357                         line_queue = line;
358                 line_queue_count++;
359
360                 if (line->count[sym_counter])
361                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
362                 if (pcnt >= (float)sym_pcnt_filter) {
363                         if (displayed <= print_entries)
364                                 show_lines(line_queue, line_queue_count, total);
365                         else more++;
366                         displayed += line_queue_count;
367                         line_queue_count = 0;
368                         line_queue = NULL;
369                 } else if (line_queue_count > TRACE_COUNT) {
370                         line_queue = line_queue->next;
371                         line_queue_count--;
372                 }
373
374                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
375                 line = line->next;
376         }
377         pthread_mutex_unlock(&syme->src->lock);
378         if (more)
379                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
380 }
381
382 /*
383  * Symbols will be added here in event__process_sample and will get out
384  * after decayed.
385  */
386 static LIST_HEAD(active_symbols);
387 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
388
389 /*
390  * Ordering weight: count-1 * count-2 * ... / count-n
391  */
392 static double sym_weight(const struct sym_entry *sym)
393 {
394         double weight = sym->snap_count;
395         int counter;
396
397         if (!display_weighted)
398                 return weight;
399
400         for (counter = 1; counter < nr_counters-1; counter++)
401                 weight *= sym->count[counter];
402
403         weight /= (sym->count[counter] + 1);
404
405         return weight;
406 }
407
408 static long                     samples;
409 static long                     userspace_samples;
410 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
411
412 static void __list_insert_active_sym(struct sym_entry *syme)
413 {
414         list_add(&syme->node, &active_symbols);
415 }
416
417 static void list_remove_active_sym(struct sym_entry *syme)
418 {
419         pthread_mutex_lock(&active_symbols_lock);
420         list_del_init(&syme->node);
421         pthread_mutex_unlock(&active_symbols_lock);
422 }
423
424 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
425 {
426         struct rb_node **p = &tree->rb_node;
427         struct rb_node *parent = NULL;
428         struct sym_entry *iter;
429
430         while (*p != NULL) {
431                 parent = *p;
432                 iter = rb_entry(parent, struct sym_entry, rb_node);
433
434                 if (se->weight > iter->weight)
435                         p = &(*p)->rb_left;
436                 else
437                         p = &(*p)->rb_right;
438         }
439
440         rb_link_node(&se->rb_node, parent, p);
441         rb_insert_color(&se->rb_node, tree);
442 }
443
444 static void print_sym_table(void)
445 {
446         int printed = 0, j;
447         int counter, snap = !display_weighted ? sym_counter : 0;
448         float samples_per_sec = samples/delay_secs;
449         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
450         float sum_ksamples = 0.0;
451         struct sym_entry *syme, *n;
452         struct rb_root tmp = RB_ROOT;
453         struct rb_node *nd;
454         int sym_width = 0, dso_width = 0, max_dso_width;
455         const int win_width = winsize.ws_col - 1;
456
457         samples = userspace_samples = 0;
458
459         /* Sort the active symbols */
460         pthread_mutex_lock(&active_symbols_lock);
461         syme = list_entry(active_symbols.next, struct sym_entry, node);
462         pthread_mutex_unlock(&active_symbols_lock);
463
464         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
465                 syme->snap_count = syme->count[snap];
466                 if (syme->snap_count != 0) {
467
468                         if ((hide_user_symbols &&
469                              syme->origin == PERF_RECORD_MISC_USER) ||
470                             (hide_kernel_symbols &&
471                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
472                                 list_remove_active_sym(syme);
473                                 continue;
474                         }
475                         syme->weight = sym_weight(syme);
476                         rb_insert_active_sym(&tmp, syme);
477                         sum_ksamples += syme->snap_count;
478
479                         for (j = 0; j < nr_counters; j++)
480                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
481                 } else
482                         list_remove_active_sym(syme);
483         }
484
485         puts(CONSOLE_CLEAR);
486
487         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
488         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
489                 samples_per_sec,
490                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
491
492         if (nr_counters == 1 || !display_weighted) {
493                 printf("%Ld", (u64)attrs[0].sample_period);
494                 if (freq)
495                         printf("Hz ");
496                 else
497                         printf(" ");
498         }
499
500         if (!display_weighted)
501                 printf("%s", event_name(sym_counter));
502         else for (counter = 0; counter < nr_counters; counter++) {
503                 if (counter)
504                         printf("/");
505
506                 printf("%s", event_name(counter));
507         }
508
509         printf( "], ");
510
511         if (target_pid != -1)
512                 printf(" (target_pid: %d", target_pid);
513         else
514                 printf(" (all");
515
516         if (profile_cpu != -1)
517                 printf(", cpu: %d)\n", profile_cpu);
518         else {
519                 if (target_pid != -1)
520                         printf(")\n");
521                 else
522                         printf(", %d CPUs)\n", nr_cpus);
523         }
524
525         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
526
527         if (sym_filter_entry) {
528                 show_details(sym_filter_entry);
529                 return;
530         }
531
532         /*
533          * Find the longest symbol name that will be displayed
534          */
535         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
536                 syme = rb_entry(nd, struct sym_entry, rb_node);
537                 if (++printed > print_entries ||
538                     (int)syme->snap_count < count_filter)
539                         continue;
540
541                 if (syme->map->dso->long_name_len > dso_width)
542                         dso_width = syme->map->dso->long_name_len;
543
544                 if (syme->name_len > sym_width)
545                         sym_width = syme->name_len;
546         }
547
548         printed = 0;
549
550         max_dso_width = winsize.ws_col - sym_width - 29;
551         if (dso_width > max_dso_width)
552                 dso_width = max_dso_width;
553         putchar('\n');
554         if (nr_counters == 1)
555                 printf("             samples  pcnt");
556         else
557                 printf("   weight    samples  pcnt");
558
559         if (verbose)
560                 printf("         RIP       ");
561         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
562         printf("   %s    _______ _____",
563                nr_counters == 1 ? "      " : "______");
564         if (verbose)
565                 printf(" ________________");
566         printf(" %-*.*s", sym_width, sym_width, graph_line);
567         printf(" %-*.*s", dso_width, dso_width, graph_line);
568         puts("\n");
569
570         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
571                 struct symbol *sym;
572                 double pcnt;
573
574                 syme = rb_entry(nd, struct sym_entry, rb_node);
575                 sym = sym_entry__symbol(syme);
576
577                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
578                         continue;
579
580                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
581                                          sum_ksamples));
582
583                 if (nr_counters == 1 || !display_weighted)
584                         printf("%20.2f ", syme->weight);
585                 else
586                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
587
588                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
589                 if (verbose)
590                         printf(" %016llx", sym->start);
591                 printf(" %-*.*s", sym_width, sym_width, sym->name);
592                 printf(" %-*.*s\n", dso_width, dso_width,
593                        dso_width >= syme->map->dso->long_name_len ?
594                                         syme->map->dso->long_name :
595                                         syme->map->dso->short_name);
596         }
597 }
598
599 static void prompt_integer(int *target, const char *msg)
600 {
601         char *buf = malloc(0), *p;
602         size_t dummy = 0;
603         int tmp;
604
605         fprintf(stdout, "\n%s: ", msg);
606         if (getline(&buf, &dummy, stdin) < 0)
607                 return;
608
609         p = strchr(buf, '\n');
610         if (p)
611                 *p = 0;
612
613         p = buf;
614         while(*p) {
615                 if (!isdigit(*p))
616                         goto out_free;
617                 p++;
618         }
619         tmp = strtoul(buf, NULL, 10);
620         *target = tmp;
621 out_free:
622         free(buf);
623 }
624
625 static void prompt_percent(int *target, const char *msg)
626 {
627         int tmp = 0;
628
629         prompt_integer(&tmp, msg);
630         if (tmp >= 0 && tmp <= 100)
631                 *target = tmp;
632 }
633
634 static void prompt_symbol(struct sym_entry **target, const char *msg)
635 {
636         char *buf = malloc(0), *p;
637         struct sym_entry *syme = *target, *n, *found = NULL;
638         size_t dummy = 0;
639
640         /* zero counters of active symbol */
641         if (syme) {
642                 pthread_mutex_lock(&syme->src->lock);
643                 __zero_source_counters(syme);
644                 *target = NULL;
645                 pthread_mutex_unlock(&syme->src->lock);
646         }
647
648         fprintf(stdout, "\n%s: ", msg);
649         if (getline(&buf, &dummy, stdin) < 0)
650                 goto out_free;
651
652         p = strchr(buf, '\n');
653         if (p)
654                 *p = 0;
655
656         pthread_mutex_lock(&active_symbols_lock);
657         syme = list_entry(active_symbols.next, struct sym_entry, node);
658         pthread_mutex_unlock(&active_symbols_lock);
659
660         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
661                 struct symbol *sym = sym_entry__symbol(syme);
662
663                 if (!strcmp(buf, sym->name)) {
664                         found = syme;
665                         break;
666                 }
667         }
668
669         if (!found) {
670                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
671                 sleep(1);
672                 return;
673         } else
674                 parse_source(found);
675
676 out_free:
677         free(buf);
678 }
679
680 static void print_mapped_keys(void)
681 {
682         char *name = NULL;
683
684         if (sym_filter_entry) {
685                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
686                 name = sym->name;
687         }
688
689         fprintf(stdout, "\nMapped keys:\n");
690         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
691         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
692
693         if (nr_counters > 1)
694                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
695
696         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
697
698         if (symbol_conf.vmlinux_name) {
699                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
700                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
701                 fprintf(stdout, "\t[S]     stop annotation.\n");
702         }
703
704         if (nr_counters > 1)
705                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
706
707         fprintf(stdout,
708                 "\t[K]     hide kernel_symbols symbols.             \t(%s)\n",
709                 hide_kernel_symbols ? "yes" : "no");
710         fprintf(stdout,
711                 "\t[U]     hide user symbols.               \t(%s)\n",
712                 hide_user_symbols ? "yes" : "no");
713         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
714         fprintf(stdout, "\t[qQ]    quit.\n");
715 }
716
717 static int key_mapped(int c)
718 {
719         switch (c) {
720                 case 'd':
721                 case 'e':
722                 case 'f':
723                 case 'z':
724                 case 'q':
725                 case 'Q':
726                 case 'K':
727                 case 'U':
728                         return 1;
729                 case 'E':
730                 case 'w':
731                         return nr_counters > 1 ? 1 : 0;
732                 case 'F':
733                 case 's':
734                 case 'S':
735                         return symbol_conf.vmlinux_name ? 1 : 0;
736                 default:
737                         break;
738         }
739
740         return 0;
741 }
742
743 static void handle_keypress(int c)
744 {
745         if (!key_mapped(c)) {
746                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
747                 struct termios tc, save;
748
749                 print_mapped_keys();
750                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
751                 fflush(stdout);
752
753                 tcgetattr(0, &save);
754                 tc = save;
755                 tc.c_lflag &= ~(ICANON | ECHO);
756                 tc.c_cc[VMIN] = 0;
757                 tc.c_cc[VTIME] = 0;
758                 tcsetattr(0, TCSANOW, &tc);
759
760                 poll(&stdin_poll, 1, -1);
761                 c = getc(stdin);
762
763                 tcsetattr(0, TCSAFLUSH, &save);
764                 if (!key_mapped(c))
765                         return;
766         }
767
768         switch (c) {
769                 case 'd':
770                         prompt_integer(&delay_secs, "Enter display delay");
771                         if (delay_secs < 1)
772                                 delay_secs = 1;
773                         break;
774                 case 'e':
775                         prompt_integer(&print_entries, "Enter display entries (lines)");
776                         if (print_entries == 0) {
777                                 sig_winch_handler(SIGWINCH);
778                                 signal(SIGWINCH, sig_winch_handler);
779                         } else
780                                 signal(SIGWINCH, SIG_DFL);
781                         break;
782                 case 'E':
783                         if (nr_counters > 1) {
784                                 int i;
785
786                                 fprintf(stderr, "\nAvailable events:");
787                                 for (i = 0; i < nr_counters; i++)
788                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
789
790                                 prompt_integer(&sym_counter, "Enter details event counter");
791
792                                 if (sym_counter >= nr_counters) {
793                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
794                                         sym_counter = 0;
795                                         sleep(1);
796                                 }
797                         } else sym_counter = 0;
798                         break;
799                 case 'f':
800                         prompt_integer(&count_filter, "Enter display event count filter");
801                         break;
802                 case 'F':
803                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
804                         break;
805                 case 'K':
806                         hide_kernel_symbols = !hide_kernel_symbols;
807                         break;
808                 case 'q':
809                 case 'Q':
810                         printf("exiting.\n");
811                         if (dump_symtab)
812                                 dsos__fprintf(stderr);
813                         exit(0);
814                 case 's':
815                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
816                         break;
817                 case 'S':
818                         if (!sym_filter_entry)
819                                 break;
820                         else {
821                                 struct sym_entry *syme = sym_filter_entry;
822
823                                 pthread_mutex_lock(&syme->src->lock);
824                                 sym_filter_entry = NULL;
825                                 __zero_source_counters(syme);
826                                 pthread_mutex_unlock(&syme->src->lock);
827                         }
828                         break;
829                 case 'U':
830                         hide_user_symbols = !hide_user_symbols;
831                         break;
832                 case 'w':
833                         display_weighted = ~display_weighted;
834                         break;
835                 case 'z':
836                         zero = ~zero;
837                         break;
838                 default:
839                         break;
840         }
841 }
842
843 static void *display_thread(void *arg __used)
844 {
845         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
846         struct termios tc, save;
847         int delay_msecs, c;
848
849         tcgetattr(0, &save);
850         tc = save;
851         tc.c_lflag &= ~(ICANON | ECHO);
852         tc.c_cc[VMIN] = 0;
853         tc.c_cc[VTIME] = 0;
854
855 repeat:
856         delay_msecs = delay_secs * 1000;
857         tcsetattr(0, TCSANOW, &tc);
858         /* trash return*/
859         getc(stdin);
860
861         do {
862                 print_sym_table();
863         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
864
865         c = getc(stdin);
866         tcsetattr(0, TCSAFLUSH, &save);
867
868         handle_keypress(c);
869         goto repeat;
870
871         return NULL;
872 }
873
874 /* Tag samples to be skipped. */
875 static const char *skip_symbols[] = {
876         "default_idle",
877         "cpu_idle",
878         "enter_idle",
879         "exit_idle",
880         "mwait_idle",
881         "mwait_idle_with_hints",
882         "poll_idle",
883         "ppc64_runlatch_off",
884         "pseries_dedicated_idle_sleep",
885         NULL
886 };
887
888 static int symbol_filter(struct map *map, struct symbol *sym)
889 {
890         struct sym_entry *syme;
891         const char *name = sym->name;
892         int i;
893
894         /*
895          * ppc64 uses function descriptors and appends a '.' to the
896          * start of every instruction address. Remove it.
897          */
898         if (name[0] == '.')
899                 name++;
900
901         if (!strcmp(name, "_text") ||
902             !strcmp(name, "_etext") ||
903             !strcmp(name, "_sinittext") ||
904             !strncmp("init_module", name, 11) ||
905             !strncmp("cleanup_module", name, 14) ||
906             strstr(name, "_text_start") ||
907             strstr(name, "_text_end"))
908                 return 1;
909
910         syme = symbol__priv(sym);
911         syme->map = map;
912         syme->src = NULL;
913         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
914                 sym_filter_entry = syme;
915
916         for (i = 0; skip_symbols[i]; i++) {
917                 if (!strcmp(skip_symbols[i], name)) {
918                         syme->skip = 1;
919                         break;
920                 }
921         }
922
923         if (!syme->skip)
924                 syme->name_len = strlen(sym->name);
925
926         return 0;
927 }
928
929 static void event__process_sample(const event_t *self,
930                                  struct perf_session *session, int counter)
931 {
932         u64 ip = self->ip.ip;
933         struct sym_entry *syme;
934         struct addr_location al;
935         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
936
937         switch (origin) {
938         case PERF_RECORD_MISC_USER:
939                 if (hide_user_symbols)
940                         return;
941                 break;
942         case PERF_RECORD_MISC_KERNEL:
943                 if (hide_kernel_symbols)
944                         return;
945                 break;
946         default:
947                 return;
948         }
949
950         if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
951             al.sym == NULL || al.filtered)
952                 return;
953
954         syme = symbol__priv(al.sym);
955         if (!syme->skip) {
956                 syme->count[counter]++;
957                 syme->origin = origin;
958                 record_precise_ip(syme, counter, ip);
959                 pthread_mutex_lock(&active_symbols_lock);
960                 if (list_empty(&syme->node) || !syme->node.next)
961                         __list_insert_active_sym(syme);
962                 pthread_mutex_unlock(&active_symbols_lock);
963                 if (origin == PERF_RECORD_MISC_USER)
964                         ++userspace_samples;
965                 ++samples;
966         }
967 }
968
969 static int event__process(event_t *event, struct perf_session *session)
970 {
971         switch (event->header.type) {
972         case PERF_RECORD_COMM:
973                 event__process_comm(event, session);
974                 break;
975         case PERF_RECORD_MMAP:
976                 event__process_mmap(event, session);
977                 break;
978         default:
979                 break;
980         }
981
982         return 0;
983 }
984
985 struct mmap_data {
986         int                     counter;
987         void                    *base;
988         int                     mask;
989         unsigned int            prev;
990 };
991
992 static unsigned int mmap_read_head(struct mmap_data *md)
993 {
994         struct perf_event_mmap_page *pc = md->base;
995         int head;
996
997         head = pc->data_head;
998         rmb();
999
1000         return head;
1001 }
1002
1003 static void perf_session__mmap_read_counter(struct perf_session *self,
1004                                             struct mmap_data *md)
1005 {
1006         unsigned int head = mmap_read_head(md);
1007         unsigned int old = md->prev;
1008         unsigned char *data = md->base + page_size;
1009         int diff;
1010
1011         /*
1012          * If we're further behind than half the buffer, there's a chance
1013          * the writer will bite our tail and mess up the samples under us.
1014          *
1015          * If we somehow ended up ahead of the head, we got messed up.
1016          *
1017          * In either case, truncate and restart at head.
1018          */
1019         diff = head - old;
1020         if (diff > md->mask / 2 || diff < 0) {
1021                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1022
1023                 /*
1024                  * head points to a known good entry, start there.
1025                  */
1026                 old = head;
1027         }
1028
1029         for (; old != head;) {
1030                 event_t *event = (event_t *)&data[old & md->mask];
1031
1032                 event_t event_copy;
1033
1034                 size_t size = event->header.size;
1035
1036                 /*
1037                  * Event straddles the mmap boundary -- header should always
1038                  * be inside due to u64 alignment of output.
1039                  */
1040                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1041                         unsigned int offset = old;
1042                         unsigned int len = min(sizeof(*event), size), cpy;
1043                         void *dst = &event_copy;
1044
1045                         do {
1046                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1047                                 memcpy(dst, &data[offset & md->mask], cpy);
1048                                 offset += cpy;
1049                                 dst += cpy;
1050                                 len -= cpy;
1051                         } while (len);
1052
1053                         event = &event_copy;
1054                 }
1055
1056                 if (event->header.type == PERF_RECORD_SAMPLE)
1057                         event__process_sample(event, self, md->counter);
1058                 else
1059                         event__process(event, self);
1060                 old += size;
1061         }
1062
1063         md->prev = old;
1064 }
1065
1066 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1067 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1068
1069 static void perf_session__mmap_read(struct perf_session *self)
1070 {
1071         int i, counter;
1072
1073         for (i = 0; i < nr_cpus; i++) {
1074                 for (counter = 0; counter < nr_counters; counter++)
1075                         perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1076         }
1077 }
1078
1079 int nr_poll;
1080 int group_fd;
1081
1082 static void start_counter(int i, int counter)
1083 {
1084         struct perf_event_attr *attr;
1085         int cpu;
1086
1087         cpu = profile_cpu;
1088         if (target_pid == -1 && profile_cpu == -1)
1089                 cpu = i;
1090
1091         attr = attrs + counter;
1092
1093         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1094
1095         if (freq) {
1096                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1097                 attr->freq              = 1;
1098                 attr->sample_freq       = freq;
1099         }
1100
1101         attr->inherit           = (cpu < 0) && inherit;
1102         attr->mmap              = 1;
1103
1104 try_again:
1105         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1106
1107         if (fd[i][counter] < 0) {
1108                 int err = errno;
1109
1110                 if (err == EPERM || err == EACCES)
1111                         die("No permission - are you root?\n");
1112                 /*
1113                  * If it's cycles then fall back to hrtimer
1114                  * based cpu-clock-tick sw counter, which
1115                  * is always available even if no PMU support:
1116                  */
1117                 if (attr->type == PERF_TYPE_HARDWARE
1118                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1119
1120                         if (verbose)
1121                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1122
1123                         attr->type = PERF_TYPE_SOFTWARE;
1124                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1125                         goto try_again;
1126                 }
1127                 printf("\n");
1128                 error("perfcounter syscall returned with %d (%s)\n",
1129                         fd[i][counter], strerror(err));
1130                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1131                 exit(-1);
1132         }
1133         assert(fd[i][counter] >= 0);
1134         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1135
1136         /*
1137          * First counter acts as the group leader:
1138          */
1139         if (group && group_fd == -1)
1140                 group_fd = fd[i][counter];
1141
1142         event_array[nr_poll].fd = fd[i][counter];
1143         event_array[nr_poll].events = POLLIN;
1144         nr_poll++;
1145
1146         mmap_array[i][counter].counter = counter;
1147         mmap_array[i][counter].prev = 0;
1148         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1149         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1150                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1151         if (mmap_array[i][counter].base == MAP_FAILED)
1152                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1153 }
1154
1155 static int __cmd_top(void)
1156 {
1157         pthread_t thread;
1158         int i, counter;
1159         int ret;
1160         /*
1161          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1162          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1163          */
1164         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1165         if (session == NULL)
1166                 return -ENOMEM;
1167
1168         if (target_pid != -1)
1169                 event__synthesize_thread(target_pid, event__process, session);
1170         else
1171                 event__synthesize_threads(event__process, session);
1172
1173         for (i = 0; i < nr_cpus; i++) {
1174                 group_fd = -1;
1175                 for (counter = 0; counter < nr_counters; counter++)
1176                         start_counter(i, counter);
1177         }
1178
1179         /* Wait for a minimal set of events before starting the snapshot */
1180         poll(event_array, nr_poll, 100);
1181
1182         perf_session__mmap_read(session);
1183
1184         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1185                 printf("Could not create display thread.\n");
1186                 exit(-1);
1187         }
1188
1189         if (realtime_prio) {
1190                 struct sched_param param;
1191
1192                 param.sched_priority = realtime_prio;
1193                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1194                         printf("Could not set realtime priority.\n");
1195                         exit(-1);
1196                 }
1197         }
1198
1199         while (1) {
1200                 int hits = samples;
1201
1202                 perf_session__mmap_read(session);
1203
1204                 if (hits == samples)
1205                         ret = poll(event_array, nr_poll, 100);
1206         }
1207
1208         return 0;
1209 }
1210
1211 static const char * const top_usage[] = {
1212         "perf top [<options>]",
1213         NULL
1214 };
1215
1216 static const struct option options[] = {
1217         OPT_CALLBACK('e', "event", NULL, "event",
1218                      "event selector. use 'perf list' to list available events",
1219                      parse_events),
1220         OPT_INTEGER('c', "count", &default_interval,
1221                     "event period to sample"),
1222         OPT_INTEGER('p', "pid", &target_pid,
1223                     "profile events on existing pid"),
1224         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1225                             "system-wide collection from all CPUs"),
1226         OPT_INTEGER('C', "CPU", &profile_cpu,
1227                     "CPU to profile on"),
1228         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1229                    "file", "vmlinux pathname"),
1230         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1231                     "hide kernel symbols"),
1232         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1233                     "number of mmap data pages"),
1234         OPT_INTEGER('r', "realtime", &realtime_prio,
1235                     "collect data with this RT SCHED_FIFO priority"),
1236         OPT_INTEGER('d', "delay", &delay_secs,
1237                     "number of seconds to delay between refreshes"),
1238         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1239                             "dump the symbol table used for profiling"),
1240         OPT_INTEGER('f', "count-filter", &count_filter,
1241                     "only display functions with more events than this"),
1242         OPT_BOOLEAN('g', "group", &group,
1243                             "put the counters into a counter group"),
1244         OPT_BOOLEAN('i', "inherit", &inherit,
1245                     "child tasks inherit counters"),
1246         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1247                     "symbol to annotate - requires -k option"),
1248         OPT_BOOLEAN('z', "zero", &zero,
1249                     "zero history across updates"),
1250         OPT_INTEGER('F', "freq", &freq,
1251                     "profile at this frequency"),
1252         OPT_INTEGER('E', "entries", &print_entries,
1253                     "display this many functions"),
1254         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1255                     "hide user symbols"),
1256         OPT_BOOLEAN('v', "verbose", &verbose,
1257                     "be more verbose (show counter open errors, etc)"),
1258         OPT_END()
1259 };
1260
1261 int cmd_top(int argc, const char **argv, const char *prefix __used)
1262 {
1263         int counter;
1264
1265         page_size = sysconf(_SC_PAGE_SIZE);
1266
1267         argc = parse_options(argc, argv, options, top_usage, 0);
1268         if (argc)
1269                 usage_with_options(top_usage, options);
1270
1271         /* CPU and PID are mutually exclusive */
1272         if (target_pid != -1 && profile_cpu != -1) {
1273                 printf("WARNING: PID switch overriding CPU\n");
1274                 sleep(1);
1275                 profile_cpu = -1;
1276         }
1277
1278         if (!nr_counters)
1279                 nr_counters = 1;
1280
1281         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1282                                  (nr_counters + 1) * sizeof(unsigned long));
1283         if (symbol_conf.vmlinux_name == NULL)
1284                 symbol_conf.try_vmlinux_path = true;
1285         if (symbol__init() < 0)
1286                 return -1;
1287
1288         if (delay_secs < 1)
1289                 delay_secs = 1;
1290
1291         parse_source(sym_filter_entry);
1292
1293         /*
1294          * User specified count overrides default frequency.
1295          */
1296         if (default_interval)
1297                 freq = 0;
1298         else if (freq) {
1299                 default_interval = freq;
1300         } else {
1301                 fprintf(stderr, "frequency and count are zero, aborting\n");
1302                 exit(EXIT_FAILURE);
1303         }
1304
1305         /*
1306          * Fill in the ones not specifically initialized via -c:
1307          */
1308         for (counter = 0; counter < nr_counters; counter++) {
1309                 if (attrs[counter].sample_period)
1310                         continue;
1311
1312                 attrs[counter].sample_period = default_interval;
1313         }
1314
1315         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1316         assert(nr_cpus <= MAX_NR_CPUS);
1317         assert(nr_cpus >= 0);
1318
1319         if (target_pid != -1 || profile_cpu != -1)
1320                 nr_cpus = 1;
1321
1322         get_term_dimensions(&winsize);
1323         if (print_entries == 0) {
1324                 update_print_entries(&winsize);
1325                 signal(SIGWINCH, sig_winch_handler);
1326         }
1327
1328         return __cmd_top();
1329 }