Merge commit 'v2.6.33' into perf/core
[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/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 struct sym_entry                *sym_filter_entry_sched         =   NULL;
98 static int                      sym_pcnt_filter                 =      5;
99 static int                      sym_counter                     =      0;
100 static int                      display_weighted                =     -1;
101
102 /*
103  * Symbols
104  */
105
106 struct sym_entry_source {
107         struct source_line      *source;
108         struct source_line      *lines;
109         struct source_line      **lines_tail;
110         pthread_mutex_t         lock;
111 };
112
113 struct sym_entry {
114         struct rb_node          rb_node;
115         struct list_head        node;
116         unsigned long           snap_count;
117         double                  weight;
118         int                     skip;
119         u16                     name_len;
120         u8                      origin;
121         struct map              *map;
122         struct sym_entry_source *src;
123         unsigned long           count[0];
124 };
125
126 /*
127  * Source functions
128  */
129
130 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
131 {
132        return ((void *)self) + symbol_conf.priv_size;
133 }
134
135 static void get_term_dimensions(struct winsize *ws)
136 {
137         char *s = getenv("LINES");
138
139         if (s != NULL) {
140                 ws->ws_row = atoi(s);
141                 s = getenv("COLUMNS");
142                 if (s != NULL) {
143                         ws->ws_col = atoi(s);
144                         if (ws->ws_row && ws->ws_col)
145                                 return;
146                 }
147         }
148 #ifdef TIOCGWINSZ
149         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
150             ws->ws_row && ws->ws_col)
151                 return;
152 #endif
153         ws->ws_row = 25;
154         ws->ws_col = 80;
155 }
156
157 static void update_print_entries(struct winsize *ws)
158 {
159         print_entries = ws->ws_row;
160
161         if (print_entries > 9)
162                 print_entries -= 9;
163 }
164
165 static void sig_winch_handler(int sig __used)
166 {
167         get_term_dimensions(&winsize);
168         update_print_entries(&winsize);
169 }
170
171 static void parse_source(struct sym_entry *syme)
172 {
173         struct symbol *sym;
174         struct sym_entry_source *source;
175         struct map *map;
176         FILE *file;
177         char command[PATH_MAX*2];
178         const char *path;
179         u64 len;
180
181         if (!syme)
182                 return;
183
184         if (syme->src == NULL) {
185                 syme->src = zalloc(sizeof(*source));
186                 if (syme->src == NULL)
187                         return;
188                 pthread_mutex_init(&syme->src->lock, NULL);
189         }
190
191         source = syme->src;
192
193         if (source->lines) {
194                 pthread_mutex_lock(&source->lock);
195                 goto out_assign;
196         }
197
198         sym = sym_entry__symbol(syme);
199         map = syme->map;
200         path = map->dso->long_name;
201
202         len = sym->end - sym->start;
203
204         sprintf(command,
205                 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
206                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
207                 BITS_PER_LONG / 4, map__rip_2objdump(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, *sep;
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                 src->eip = strtoull(src->line, &sep, 16);
238                 if (*sep == ':')
239                         src->eip = map__objdump_2ip(map, src->eip);
240                 else /* this line has no ip info (e.g. source line) */
241                         src->eip = 0;
242         }
243         pclose(file);
244 out_assign:
245         sym_filter_entry = syme;
246         pthread_mutex_unlock(&source->lock);
247 }
248
249 static void __zero_source_counters(struct sym_entry *syme)
250 {
251         int i;
252         struct source_line *line;
253
254         line = syme->src->lines;
255         while (line) {
256                 for (i = 0; i < nr_counters; i++)
257                         line->count[i] = 0;
258                 line = line->next;
259         }
260 }
261
262 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
263 {
264         struct source_line *line;
265
266         if (syme != sym_filter_entry)
267                 return;
268
269         if (pthread_mutex_trylock(&syme->src->lock))
270                 return;
271
272         if (syme->src == NULL || syme->src->source == NULL)
273                 goto out_unlock;
274
275         for (line = syme->src->lines; line; line = line->next) {
276                 /* skip lines without IP info */
277                 if (line->eip == 0)
278                         continue;
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 #define PATTERN_LEN             (BITS_PER_LONG / 4 + 2)
291
292 static void lookup_sym_source(struct sym_entry *syme)
293 {
294         struct symbol *symbol = sym_entry__symbol(syme);
295         struct source_line *line;
296         char pattern[PATTERN_LEN + 1];
297
298         sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
299                 map__rip_2objdump(syme->map, symbol->start));
300
301         pthread_mutex_lock(&syme->src->lock);
302         for (line = syme->src->lines; line; line = line->next) {
303                 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
304                         syme->src->source = line;
305                         break;
306                 }
307         }
308         pthread_mutex_unlock(&syme->src->lock);
309 }
310
311 static void show_lines(struct source_line *queue, int count, int total)
312 {
313         int i;
314         struct source_line *line;
315
316         line = queue;
317         for (i = 0; i < count; i++) {
318                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
319
320                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
321                 line = line->next;
322         }
323 }
324
325 #define TRACE_COUNT     3
326
327 static void show_details(struct sym_entry *syme)
328 {
329         struct symbol *symbol;
330         struct source_line *line;
331         struct source_line *line_queue = NULL;
332         int displayed = 0;
333         int line_queue_count = 0, total = 0, more = 0;
334
335         if (!syme)
336                 return;
337
338         if (!syme->src->source)
339                 lookup_sym_source(syme);
340
341         if (!syme->src->source)
342                 return;
343
344         symbol = sym_entry__symbol(syme);
345         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
346         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
347
348         pthread_mutex_lock(&syme->src->lock);
349         line = syme->src->source;
350         while (line) {
351                 total += line->count[sym_counter];
352                 line = line->next;
353         }
354
355         line = syme->src->source;
356         while (line) {
357                 float pcnt = 0.0;
358
359                 if (!line_queue_count)
360                         line_queue = line;
361                 line_queue_count++;
362
363                 if (line->count[sym_counter])
364                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
365                 if (pcnt >= (float)sym_pcnt_filter) {
366                         if (displayed <= print_entries)
367                                 show_lines(line_queue, line_queue_count, total);
368                         else more++;
369                         displayed += line_queue_count;
370                         line_queue_count = 0;
371                         line_queue = NULL;
372                 } else if (line_queue_count > TRACE_COUNT) {
373                         line_queue = line_queue->next;
374                         line_queue_count--;
375                 }
376
377                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
378                 line = line->next;
379         }
380         pthread_mutex_unlock(&syme->src->lock);
381         if (more)
382                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
383 }
384
385 /*
386  * Symbols will be added here in event__process_sample and will get out
387  * after decayed.
388  */
389 static LIST_HEAD(active_symbols);
390 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
391
392 /*
393  * Ordering weight: count-1 * count-2 * ... / count-n
394  */
395 static double sym_weight(const struct sym_entry *sym)
396 {
397         double weight = sym->snap_count;
398         int counter;
399
400         if (!display_weighted)
401                 return weight;
402
403         for (counter = 1; counter < nr_counters-1; counter++)
404                 weight *= sym->count[counter];
405
406         weight /= (sym->count[counter] + 1);
407
408         return weight;
409 }
410
411 static long                     samples;
412 static long                     userspace_samples;
413 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
414
415 static void __list_insert_active_sym(struct sym_entry *syme)
416 {
417         list_add(&syme->node, &active_symbols);
418 }
419
420 static void list_remove_active_sym(struct sym_entry *syme)
421 {
422         pthread_mutex_lock(&active_symbols_lock);
423         list_del_init(&syme->node);
424         pthread_mutex_unlock(&active_symbols_lock);
425 }
426
427 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
428 {
429         struct rb_node **p = &tree->rb_node;
430         struct rb_node *parent = NULL;
431         struct sym_entry *iter;
432
433         while (*p != NULL) {
434                 parent = *p;
435                 iter = rb_entry(parent, struct sym_entry, rb_node);
436
437                 if (se->weight > iter->weight)
438                         p = &(*p)->rb_left;
439                 else
440                         p = &(*p)->rb_right;
441         }
442
443         rb_link_node(&se->rb_node, parent, p);
444         rb_insert_color(&se->rb_node, tree);
445 }
446
447 static void print_sym_table(void)
448 {
449         int printed = 0, j;
450         int counter, snap = !display_weighted ? sym_counter : 0;
451         float samples_per_sec = samples/delay_secs;
452         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
453         float sum_ksamples = 0.0;
454         struct sym_entry *syme, *n;
455         struct rb_root tmp = RB_ROOT;
456         struct rb_node *nd;
457         int sym_width = 0, dso_width = 0, max_dso_width;
458         const int win_width = winsize.ws_col - 1;
459
460         samples = userspace_samples = 0;
461
462         /* Sort the active symbols */
463         pthread_mutex_lock(&active_symbols_lock);
464         syme = list_entry(active_symbols.next, struct sym_entry, node);
465         pthread_mutex_unlock(&active_symbols_lock);
466
467         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
468                 syme->snap_count = syme->count[snap];
469                 if (syme->snap_count != 0) {
470
471                         if ((hide_user_symbols &&
472                              syme->origin == PERF_RECORD_MISC_USER) ||
473                             (hide_kernel_symbols &&
474                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
475                                 list_remove_active_sym(syme);
476                                 continue;
477                         }
478                         syme->weight = sym_weight(syme);
479                         rb_insert_active_sym(&tmp, syme);
480                         sum_ksamples += syme->snap_count;
481
482                         for (j = 0; j < nr_counters; j++)
483                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
484                 } else
485                         list_remove_active_sym(syme);
486         }
487
488         puts(CONSOLE_CLEAR);
489
490         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
491         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
492                 samples_per_sec,
493                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
494
495         if (nr_counters == 1 || !display_weighted) {
496                 printf("%Ld", (u64)attrs[0].sample_period);
497                 if (freq)
498                         printf("Hz ");
499                 else
500                         printf(" ");
501         }
502
503         if (!display_weighted)
504                 printf("%s", event_name(sym_counter));
505         else for (counter = 0; counter < nr_counters; counter++) {
506                 if (counter)
507                         printf("/");
508
509                 printf("%s", event_name(counter));
510         }
511
512         printf( "], ");
513
514         if (target_pid != -1)
515                 printf(" (target_pid: %d", target_pid);
516         else
517                 printf(" (all");
518
519         if (profile_cpu != -1)
520                 printf(", cpu: %d)\n", profile_cpu);
521         else {
522                 if (target_pid != -1)
523                         printf(")\n");
524                 else
525                         printf(", %d CPUs)\n", nr_cpus);
526         }
527
528         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
529
530         if (sym_filter_entry) {
531                 show_details(sym_filter_entry);
532                 return;
533         }
534
535         /*
536          * Find the longest symbol name that will be displayed
537          */
538         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
539                 syme = rb_entry(nd, struct sym_entry, rb_node);
540                 if (++printed > print_entries ||
541                     (int)syme->snap_count < count_filter)
542                         continue;
543
544                 if (syme->map->dso->long_name_len > dso_width)
545                         dso_width = syme->map->dso->long_name_len;
546
547                 if (syme->name_len > sym_width)
548                         sym_width = syme->name_len;
549         }
550
551         printed = 0;
552
553         max_dso_width = winsize.ws_col - sym_width - 29;
554         if (dso_width > max_dso_width)
555                 dso_width = max_dso_width;
556         putchar('\n');
557         if (nr_counters == 1)
558                 printf("             samples  pcnt");
559         else
560                 printf("   weight    samples  pcnt");
561
562         if (verbose)
563                 printf("         RIP       ");
564         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
565         printf("   %s    _______ _____",
566                nr_counters == 1 ? "      " : "______");
567         if (verbose)
568                 printf(" ________________");
569         printf(" %-*.*s", sym_width, sym_width, graph_line);
570         printf(" %-*.*s", dso_width, dso_width, graph_line);
571         puts("\n");
572
573         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
574                 struct symbol *sym;
575                 double pcnt;
576
577                 syme = rb_entry(nd, struct sym_entry, rb_node);
578                 sym = sym_entry__symbol(syme);
579
580                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
581                         continue;
582
583                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
584                                          sum_ksamples));
585
586                 if (nr_counters == 1 || !display_weighted)
587                         printf("%20.2f ", syme->weight);
588                 else
589                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
590
591                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
592                 if (verbose)
593                         printf(" %016llx", sym->start);
594                 printf(" %-*.*s", sym_width, sym_width, sym->name);
595                 printf(" %-*.*s\n", dso_width, dso_width,
596                        dso_width >= syme->map->dso->long_name_len ?
597                                         syme->map->dso->long_name :
598                                         syme->map->dso->short_name);
599         }
600 }
601
602 static void prompt_integer(int *target, const char *msg)
603 {
604         char *buf = malloc(0), *p;
605         size_t dummy = 0;
606         int tmp;
607
608         fprintf(stdout, "\n%s: ", msg);
609         if (getline(&buf, &dummy, stdin) < 0)
610                 return;
611
612         p = strchr(buf, '\n');
613         if (p)
614                 *p = 0;
615
616         p = buf;
617         while(*p) {
618                 if (!isdigit(*p))
619                         goto out_free;
620                 p++;
621         }
622         tmp = strtoul(buf, NULL, 10);
623         *target = tmp;
624 out_free:
625         free(buf);
626 }
627
628 static void prompt_percent(int *target, const char *msg)
629 {
630         int tmp = 0;
631
632         prompt_integer(&tmp, msg);
633         if (tmp >= 0 && tmp <= 100)
634                 *target = tmp;
635 }
636
637 static void prompt_symbol(struct sym_entry **target, const char *msg)
638 {
639         char *buf = malloc(0), *p;
640         struct sym_entry *syme = *target, *n, *found = NULL;
641         size_t dummy = 0;
642
643         /* zero counters of active symbol */
644         if (syme) {
645                 pthread_mutex_lock(&syme->src->lock);
646                 __zero_source_counters(syme);
647                 *target = NULL;
648                 pthread_mutex_unlock(&syme->src->lock);
649         }
650
651         fprintf(stdout, "\n%s: ", msg);
652         if (getline(&buf, &dummy, stdin) < 0)
653                 goto out_free;
654
655         p = strchr(buf, '\n');
656         if (p)
657                 *p = 0;
658
659         pthread_mutex_lock(&active_symbols_lock);
660         syme = list_entry(active_symbols.next, struct sym_entry, node);
661         pthread_mutex_unlock(&active_symbols_lock);
662
663         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
664                 struct symbol *sym = sym_entry__symbol(syme);
665
666                 if (!strcmp(buf, sym->name)) {
667                         found = syme;
668                         break;
669                 }
670         }
671
672         if (!found) {
673                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
674                 sleep(1);
675                 return;
676         } else
677                 parse_source(found);
678
679 out_free:
680         free(buf);
681 }
682
683 static void print_mapped_keys(void)
684 {
685         char *name = NULL;
686
687         if (sym_filter_entry) {
688                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
689                 name = sym->name;
690         }
691
692         fprintf(stdout, "\nMapped keys:\n");
693         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
694         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
695
696         if (nr_counters > 1)
697                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
698
699         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
700
701         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
702         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
703         fprintf(stdout, "\t[S]     stop annotation.\n");
704
705         if (nr_counters > 1)
706                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
707
708         fprintf(stdout,
709                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
710                 hide_kernel_symbols ? "yes" : "no");
711         fprintf(stdout,
712                 "\t[U]     hide user symbols.               \t(%s)\n",
713                 hide_user_symbols ? "yes" : "no");
714         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
715         fprintf(stdout, "\t[qQ]    quit.\n");
716 }
717
718 static int key_mapped(int c)
719 {
720         switch (c) {
721                 case 'd':
722                 case 'e':
723                 case 'f':
724                 case 'z':
725                 case 'q':
726                 case 'Q':
727                 case 'K':
728                 case 'U':
729                 case 'F':
730                 case 's':
731                 case 'S':
732                         return 1;
733                 case 'E':
734                 case 'w':
735                         return nr_counters > 1 ? 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
914         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
915                 /* schedule initial sym_filter_entry setup */
916                 sym_filter_entry_sched = syme;
917                 sym_filter = NULL;
918         }
919
920         for (i = 0; skip_symbols[i]; i++) {
921                 if (!strcmp(skip_symbols[i], name)) {
922                         syme->skip = 1;
923                         break;
924                 }
925         }
926
927         if (!syme->skip)
928                 syme->name_len = strlen(sym->name);
929
930         return 0;
931 }
932
933 static void event__process_sample(const event_t *self,
934                                  struct perf_session *session, int counter)
935 {
936         u64 ip = self->ip.ip;
937         struct sym_entry *syme;
938         struct addr_location al;
939         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
940
941         ++samples;
942
943         switch (origin) {
944         case PERF_RECORD_MISC_USER:
945                 ++userspace_samples;
946                 if (hide_user_symbols)
947                         return;
948                 break;
949         case PERF_RECORD_MISC_KERNEL:
950                 if (hide_kernel_symbols)
951                         return;
952                 break;
953         default:
954                 return;
955         }
956
957         if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
958             al.filtered)
959                 return;
960
961         if (al.sym == NULL) {
962                 /*
963                  * As we do lazy loading of symtabs we only will know if the
964                  * specified vmlinux file is invalid when we actually have a
965                  * hit in kernel space and then try to load it. So if we get
966                  * here and there are _no_ symbols in the DSO backing the
967                  * kernel map, bail out.
968                  *
969                  * We may never get here, for instance, if we use -K/
970                  * --hide-kernel-symbols, even if the user specifies an
971                  * invalid --vmlinux ;-)
972                  */
973                 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
974                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
975                         pr_err("The %s file can't be used\n",
976                                symbol_conf.vmlinux_name);
977                         exit(1);
978                 }
979
980                 return;
981         }
982
983         /* let's see, whether we need to install initial sym_filter_entry */
984         if (sym_filter_entry_sched) {
985                 sym_filter_entry = sym_filter_entry_sched;
986                 sym_filter_entry_sched = NULL;
987                 parse_source(sym_filter_entry);
988         }
989
990         syme = symbol__priv(al.sym);
991         if (!syme->skip) {
992                 syme->count[counter]++;
993                 syme->origin = origin;
994                 record_precise_ip(syme, counter, ip);
995                 pthread_mutex_lock(&active_symbols_lock);
996                 if (list_empty(&syme->node) || !syme->node.next)
997                         __list_insert_active_sym(syme);
998                 pthread_mutex_unlock(&active_symbols_lock);
999         }
1000 }
1001
1002 static int event__process(event_t *event, struct perf_session *session)
1003 {
1004         switch (event->header.type) {
1005         case PERF_RECORD_COMM:
1006                 event__process_comm(event, session);
1007                 break;
1008         case PERF_RECORD_MMAP:
1009                 event__process_mmap(event, session);
1010                 break;
1011         case PERF_RECORD_FORK:
1012         case PERF_RECORD_EXIT:
1013                 event__process_task(event, session);
1014                 break;
1015         default:
1016                 break;
1017         }
1018
1019         return 0;
1020 }
1021
1022 struct mmap_data {
1023         int                     counter;
1024         void                    *base;
1025         int                     mask;
1026         unsigned int            prev;
1027 };
1028
1029 static unsigned int mmap_read_head(struct mmap_data *md)
1030 {
1031         struct perf_event_mmap_page *pc = md->base;
1032         int head;
1033
1034         head = pc->data_head;
1035         rmb();
1036
1037         return head;
1038 }
1039
1040 static void perf_session__mmap_read_counter(struct perf_session *self,
1041                                             struct mmap_data *md)
1042 {
1043         unsigned int head = mmap_read_head(md);
1044         unsigned int old = md->prev;
1045         unsigned char *data = md->base + page_size;
1046         int diff;
1047
1048         /*
1049          * If we're further behind than half the buffer, there's a chance
1050          * the writer will bite our tail and mess up the samples under us.
1051          *
1052          * If we somehow ended up ahead of the head, we got messed up.
1053          *
1054          * In either case, truncate and restart at head.
1055          */
1056         diff = head - old;
1057         if (diff > md->mask / 2 || diff < 0) {
1058                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1059
1060                 /*
1061                  * head points to a known good entry, start there.
1062                  */
1063                 old = head;
1064         }
1065
1066         for (; old != head;) {
1067                 event_t *event = (event_t *)&data[old & md->mask];
1068
1069                 event_t event_copy;
1070
1071                 size_t size = event->header.size;
1072
1073                 /*
1074                  * Event straddles the mmap boundary -- header should always
1075                  * be inside due to u64 alignment of output.
1076                  */
1077                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1078                         unsigned int offset = old;
1079                         unsigned int len = min(sizeof(*event), size), cpy;
1080                         void *dst = &event_copy;
1081
1082                         do {
1083                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1084                                 memcpy(dst, &data[offset & md->mask], cpy);
1085                                 offset += cpy;
1086                                 dst += cpy;
1087                                 len -= cpy;
1088                         } while (len);
1089
1090                         event = &event_copy;
1091                 }
1092
1093                 if (event->header.type == PERF_RECORD_SAMPLE)
1094                         event__process_sample(event, self, md->counter);
1095                 else
1096                         event__process(event, self);
1097                 old += size;
1098         }
1099
1100         md->prev = old;
1101 }
1102
1103 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1104 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1105
1106 static void perf_session__mmap_read(struct perf_session *self)
1107 {
1108         int i, counter;
1109
1110         for (i = 0; i < nr_cpus; i++) {
1111                 for (counter = 0; counter < nr_counters; counter++)
1112                         perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1113         }
1114 }
1115
1116 int nr_poll;
1117 int group_fd;
1118
1119 static void start_counter(int i, int counter)
1120 {
1121         struct perf_event_attr *attr;
1122         int cpu;
1123
1124         cpu = profile_cpu;
1125         if (target_pid == -1 && profile_cpu == -1)
1126                 cpu = i;
1127
1128         attr = attrs + counter;
1129
1130         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1131
1132         if (freq) {
1133                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1134                 attr->freq              = 1;
1135                 attr->sample_freq       = freq;
1136         }
1137
1138         attr->inherit           = (cpu < 0) && inherit;
1139         attr->mmap              = 1;
1140
1141 try_again:
1142         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1143
1144         if (fd[i][counter] < 0) {
1145                 int err = errno;
1146
1147                 if (err == EPERM || err == EACCES)
1148                         die("No permission - are you root?\n");
1149                 /*
1150                  * If it's cycles then fall back to hrtimer
1151                  * based cpu-clock-tick sw counter, which
1152                  * is always available even if no PMU support:
1153                  */
1154                 if (attr->type == PERF_TYPE_HARDWARE
1155                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1156
1157                         if (verbose)
1158                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1159
1160                         attr->type = PERF_TYPE_SOFTWARE;
1161                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1162                         goto try_again;
1163                 }
1164                 printf("\n");
1165                 error("perfcounter syscall returned with %d (%s)\n",
1166                         fd[i][counter], strerror(err));
1167                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1168                 exit(-1);
1169         }
1170         assert(fd[i][counter] >= 0);
1171         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1172
1173         /*
1174          * First counter acts as the group leader:
1175          */
1176         if (group && group_fd == -1)
1177                 group_fd = fd[i][counter];
1178
1179         event_array[nr_poll].fd = fd[i][counter];
1180         event_array[nr_poll].events = POLLIN;
1181         nr_poll++;
1182
1183         mmap_array[i][counter].counter = counter;
1184         mmap_array[i][counter].prev = 0;
1185         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1186         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1187                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1188         if (mmap_array[i][counter].base == MAP_FAILED)
1189                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1190 }
1191
1192 static int __cmd_top(void)
1193 {
1194         pthread_t thread;
1195         int i, counter;
1196         int ret;
1197         /*
1198          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1199          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1200          */
1201         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1202         if (session == NULL)
1203                 return -ENOMEM;
1204
1205         if (target_pid != -1)
1206                 event__synthesize_thread(target_pid, event__process, session);
1207         else
1208                 event__synthesize_threads(event__process, session);
1209
1210         for (i = 0; i < nr_cpus; i++) {
1211                 group_fd = -1;
1212                 for (counter = 0; counter < nr_counters; counter++)
1213                         start_counter(i, counter);
1214         }
1215
1216         /* Wait for a minimal set of events before starting the snapshot */
1217         poll(event_array, nr_poll, 100);
1218
1219         perf_session__mmap_read(session);
1220
1221         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1222                 printf("Could not create display thread.\n");
1223                 exit(-1);
1224         }
1225
1226         if (realtime_prio) {
1227                 struct sched_param param;
1228
1229                 param.sched_priority = realtime_prio;
1230                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1231                         printf("Could not set realtime priority.\n");
1232                         exit(-1);
1233                 }
1234         }
1235
1236         while (1) {
1237                 int hits = samples;
1238
1239                 perf_session__mmap_read(session);
1240
1241                 if (hits == samples)
1242                         ret = poll(event_array, nr_poll, 100);
1243         }
1244
1245         return 0;
1246 }
1247
1248 static const char * const top_usage[] = {
1249         "perf top [<options>]",
1250         NULL
1251 };
1252
1253 static const struct option options[] = {
1254         OPT_CALLBACK('e', "event", NULL, "event",
1255                      "event selector. use 'perf list' to list available events",
1256                      parse_events),
1257         OPT_INTEGER('c', "count", &default_interval,
1258                     "event period to sample"),
1259         OPT_INTEGER('p', "pid", &target_pid,
1260                     "profile events on existing pid"),
1261         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1262                             "system-wide collection from all CPUs"),
1263         OPT_INTEGER('C', "CPU", &profile_cpu,
1264                     "CPU to profile on"),
1265         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1266                    "file", "vmlinux pathname"),
1267         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1268                     "hide kernel symbols"),
1269         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1270                     "number of mmap data pages"),
1271         OPT_INTEGER('r', "realtime", &realtime_prio,
1272                     "collect data with this RT SCHED_FIFO priority"),
1273         OPT_INTEGER('d', "delay", &delay_secs,
1274                     "number of seconds to delay between refreshes"),
1275         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1276                             "dump the symbol table used for profiling"),
1277         OPT_INTEGER('f', "count-filter", &count_filter,
1278                     "only display functions with more events than this"),
1279         OPT_BOOLEAN('g', "group", &group,
1280                             "put the counters into a counter group"),
1281         OPT_BOOLEAN('i', "inherit", &inherit,
1282                     "child tasks inherit counters"),
1283         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1284                     "symbol to annotate"),
1285         OPT_BOOLEAN('z', "zero", &zero,
1286                     "zero history across updates"),
1287         OPT_INTEGER('F', "freq", &freq,
1288                     "profile at this frequency"),
1289         OPT_INTEGER('E', "entries", &print_entries,
1290                     "display this many functions"),
1291         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1292                     "hide user symbols"),
1293         OPT_BOOLEAN('v', "verbose", &verbose,
1294                     "be more verbose (show counter open errors, etc)"),
1295         OPT_END()
1296 };
1297
1298 int cmd_top(int argc, const char **argv, const char *prefix __used)
1299 {
1300         int counter;
1301
1302         page_size = sysconf(_SC_PAGE_SIZE);
1303
1304         argc = parse_options(argc, argv, options, top_usage, 0);
1305         if (argc)
1306                 usage_with_options(top_usage, options);
1307
1308         /* CPU and PID are mutually exclusive */
1309         if (target_pid != -1 && profile_cpu != -1) {
1310                 printf("WARNING: PID switch overriding CPU\n");
1311                 sleep(1);
1312                 profile_cpu = -1;
1313         }
1314
1315         if (!nr_counters)
1316                 nr_counters = 1;
1317
1318         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1319                                  (nr_counters + 1) * sizeof(unsigned long));
1320
1321         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1322         if (symbol__init() < 0)
1323                 return -1;
1324
1325         if (delay_secs < 1)
1326                 delay_secs = 1;
1327
1328         /*
1329          * User specified count overrides default frequency.
1330          */
1331         if (default_interval)
1332                 freq = 0;
1333         else if (freq) {
1334                 default_interval = freq;
1335         } else {
1336                 fprintf(stderr, "frequency and count are zero, aborting\n");
1337                 exit(EXIT_FAILURE);
1338         }
1339
1340         /*
1341          * Fill in the ones not specifically initialized via -c:
1342          */
1343         for (counter = 0; counter < nr_counters; counter++) {
1344                 if (attrs[counter].sample_period)
1345                         continue;
1346
1347                 attrs[counter].sample_period = default_interval;
1348         }
1349
1350         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1351         assert(nr_cpus <= MAX_NR_CPUS);
1352         assert(nr_cpus >= 0);
1353
1354         if (target_pid != -1 || profile_cpu != -1)
1355                 nr_cpus = 1;
1356
1357         get_term_dimensions(&winsize);
1358         if (print_entries == 0) {
1359                 update_print_entries(&winsize);
1360                 signal(SIGWINCH, sig_winch_handler);
1361         }
1362
1363         return __cmd_top();
1364 }