perf tools: Move the prototypes in util/string.h to util.h
[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 #include "util/cpumap.h"
32
33 #include "util/debug.h"
34
35 #include <assert.h>
36 #include <fcntl.h>
37
38 #include <stdio.h>
39 #include <termios.h>
40 #include <unistd.h>
41
42 #include <errno.h>
43 #include <time.h>
44 #include <sched.h>
45 #include <pthread.h>
46
47 #include <sys/syscall.h>
48 #include <sys/ioctl.h>
49 #include <sys/poll.h>
50 #include <sys/prctl.h>
51 #include <sys/wait.h>
52 #include <sys/uio.h>
53 #include <sys/mman.h>
54
55 #include <linux/unistd.h>
56 #include <linux/types.h>
57
58 static int                      *fd[MAX_NR_CPUS][MAX_COUNTERS];
59
60 static int                      system_wide                     =      0;
61
62 static int                      default_interval                =      0;
63
64 static int                      count_filter                    =      5;
65 static int                      print_entries;
66
67 static int                      target_pid                      =     -1;
68 static int                      target_tid                      =     -1;
69 static pid_t                    *all_tids                       =      NULL;
70 static int                      thread_num                      =      0;
71 static int                      inherit                         =      0;
72 static int                      profile_cpu                     =     -1;
73 static int                      nr_cpus                         =      0;
74 static unsigned int             realtime_prio                   =      0;
75 static int                      group                           =      0;
76 static unsigned int             page_size;
77 static unsigned int             mmap_pages                      =     16;
78 static int                      freq                            =   1000; /* 1 KHz */
79
80 static int                      delay_secs                      =      2;
81 static int                      zero                            =      0;
82 static int                      dump_symtab                     =      0;
83
84 static bool                     hide_kernel_symbols             =  false;
85 static bool                     hide_user_symbols               =  false;
86 static struct winsize           winsize;
87
88 /*
89  * Source
90  */
91
92 struct source_line {
93         u64                     eip;
94         unsigned long           count[MAX_COUNTERS];
95         char                    *line;
96         struct source_line      *next;
97 };
98
99 static char                     *sym_filter                     =   NULL;
100 struct sym_entry                *sym_filter_entry               =   NULL;
101 struct sym_entry                *sym_filter_entry_sched         =   NULL;
102 static int                      sym_pcnt_filter                 =      5;
103 static int                      sym_counter                     =      0;
104 static int                      display_weighted                =     -1;
105
106 /*
107  * Symbols
108  */
109
110 struct sym_entry_source {
111         struct source_line      *source;
112         struct source_line      *lines;
113         struct source_line      **lines_tail;
114         pthread_mutex_t         lock;
115 };
116
117 struct sym_entry {
118         struct rb_node          rb_node;
119         struct list_head        node;
120         unsigned long           snap_count;
121         double                  weight;
122         int                     skip;
123         u16                     name_len;
124         u8                      origin;
125         struct map              *map;
126         struct sym_entry_source *src;
127         unsigned long           count[0];
128 };
129
130 /*
131  * Source functions
132  */
133
134 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
135 {
136        return ((void *)self) + symbol_conf.priv_size;
137 }
138
139 void get_term_dimensions(struct winsize *ws)
140 {
141         char *s = getenv("LINES");
142
143         if (s != NULL) {
144                 ws->ws_row = atoi(s);
145                 s = getenv("COLUMNS");
146                 if (s != NULL) {
147                         ws->ws_col = atoi(s);
148                         if (ws->ws_row && ws->ws_col)
149                                 return;
150                 }
151         }
152 #ifdef TIOCGWINSZ
153         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
154             ws->ws_row && ws->ws_col)
155                 return;
156 #endif
157         ws->ws_row = 25;
158         ws->ws_col = 80;
159 }
160
161 static void update_print_entries(struct winsize *ws)
162 {
163         print_entries = ws->ws_row;
164
165         if (print_entries > 9)
166                 print_entries -= 9;
167 }
168
169 static void sig_winch_handler(int sig __used)
170 {
171         get_term_dimensions(&winsize);
172         update_print_entries(&winsize);
173 }
174
175 static int parse_source(struct sym_entry *syme)
176 {
177         struct symbol *sym;
178         struct sym_entry_source *source;
179         struct map *map;
180         FILE *file;
181         char command[PATH_MAX*2];
182         const char *path;
183         u64 len;
184
185         if (!syme)
186                 return -1;
187
188         sym = sym_entry__symbol(syme);
189         map = syme->map;
190
191         /*
192          * We can't annotate with just /proc/kallsyms
193          */
194         if (map->dso->origin == DSO__ORIG_KERNEL)
195                 return -1;
196
197         if (syme->src == NULL) {
198                 syme->src = zalloc(sizeof(*source));
199                 if (syme->src == NULL)
200                         return -1;
201                 pthread_mutex_init(&syme->src->lock, NULL);
202         }
203
204         source = syme->src;
205
206         if (source->lines) {
207                 pthread_mutex_lock(&source->lock);
208                 goto out_assign;
209         }
210         path = map->dso->long_name;
211
212         len = sym->end - sym->start;
213
214         sprintf(command,
215                 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
216                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
217                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
218
219         file = popen(command, "r");
220         if (!file)
221                 return -1;
222
223         pthread_mutex_lock(&source->lock);
224         source->lines_tail = &source->lines;
225         while (!feof(file)) {
226                 struct source_line *src;
227                 size_t dummy = 0;
228                 char *c, *sep;
229
230                 src = malloc(sizeof(struct source_line));
231                 assert(src != NULL);
232                 memset(src, 0, sizeof(struct source_line));
233
234                 if (getline(&src->line, &dummy, file) < 0)
235                         break;
236                 if (!src->line)
237                         break;
238
239                 c = strchr(src->line, '\n');
240                 if (c)
241                         *c = 0;
242
243                 src->next = NULL;
244                 *source->lines_tail = src;
245                 source->lines_tail = &src->next;
246
247                 src->eip = strtoull(src->line, &sep, 16);
248                 if (*sep == ':')
249                         src->eip = map__objdump_2ip(map, src->eip);
250                 else /* this line has no ip info (e.g. source line) */
251                         src->eip = 0;
252         }
253         pclose(file);
254 out_assign:
255         sym_filter_entry = syme;
256         pthread_mutex_unlock(&source->lock);
257         return 0;
258 }
259
260 static void __zero_source_counters(struct sym_entry *syme)
261 {
262         int i;
263         struct source_line *line;
264
265         line = syme->src->lines;
266         while (line) {
267                 for (i = 0; i < nr_counters; i++)
268                         line->count[i] = 0;
269                 line = line->next;
270         }
271 }
272
273 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
274 {
275         struct source_line *line;
276
277         if (syme != sym_filter_entry)
278                 return;
279
280         if (pthread_mutex_trylock(&syme->src->lock))
281                 return;
282
283         if (syme->src == NULL || syme->src->source == NULL)
284                 goto out_unlock;
285
286         for (line = syme->src->lines; line; line = line->next) {
287                 /* skip lines without IP info */
288                 if (line->eip == 0)
289                         continue;
290                 if (line->eip == ip) {
291                         line->count[counter]++;
292                         break;
293                 }
294                 if (line->eip > ip)
295                         break;
296         }
297 out_unlock:
298         pthread_mutex_unlock(&syme->src->lock);
299 }
300
301 #define PATTERN_LEN             (BITS_PER_LONG / 4 + 2)
302
303 static void lookup_sym_source(struct sym_entry *syme)
304 {
305         struct symbol *symbol = sym_entry__symbol(syme);
306         struct source_line *line;
307         char pattern[PATTERN_LEN + 1];
308
309         sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
310                 map__rip_2objdump(syme->map, symbol->start));
311
312         pthread_mutex_lock(&syme->src->lock);
313         for (line = syme->src->lines; line; line = line->next) {
314                 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
315                         syme->src->source = line;
316                         break;
317                 }
318         }
319         pthread_mutex_unlock(&syme->src->lock);
320 }
321
322 static void show_lines(struct source_line *queue, int count, int total)
323 {
324         int i;
325         struct source_line *line;
326
327         line = queue;
328         for (i = 0; i < count; i++) {
329                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
330
331                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
332                 line = line->next;
333         }
334 }
335
336 #define TRACE_COUNT     3
337
338 static void show_details(struct sym_entry *syme)
339 {
340         struct symbol *symbol;
341         struct source_line *line;
342         struct source_line *line_queue = NULL;
343         int displayed = 0;
344         int line_queue_count = 0, total = 0, more = 0;
345
346         if (!syme)
347                 return;
348
349         if (!syme->src->source)
350                 lookup_sym_source(syme);
351
352         if (!syme->src->source)
353                 return;
354
355         symbol = sym_entry__symbol(syme);
356         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
357         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
358
359         pthread_mutex_lock(&syme->src->lock);
360         line = syme->src->source;
361         while (line) {
362                 total += line->count[sym_counter];
363                 line = line->next;
364         }
365
366         line = syme->src->source;
367         while (line) {
368                 float pcnt = 0.0;
369
370                 if (!line_queue_count)
371                         line_queue = line;
372                 line_queue_count++;
373
374                 if (line->count[sym_counter])
375                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
376                 if (pcnt >= (float)sym_pcnt_filter) {
377                         if (displayed <= print_entries)
378                                 show_lines(line_queue, line_queue_count, total);
379                         else more++;
380                         displayed += line_queue_count;
381                         line_queue_count = 0;
382                         line_queue = NULL;
383                 } else if (line_queue_count > TRACE_COUNT) {
384                         line_queue = line_queue->next;
385                         line_queue_count--;
386                 }
387
388                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
389                 line = line->next;
390         }
391         pthread_mutex_unlock(&syme->src->lock);
392         if (more)
393                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
394 }
395
396 /*
397  * Symbols will be added here in event__process_sample and will get out
398  * after decayed.
399  */
400 static LIST_HEAD(active_symbols);
401 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
402
403 /*
404  * Ordering weight: count-1 * count-2 * ... / count-n
405  */
406 static double sym_weight(const struct sym_entry *sym)
407 {
408         double weight = sym->snap_count;
409         int counter;
410
411         if (!display_weighted)
412                 return weight;
413
414         for (counter = 1; counter < nr_counters-1; counter++)
415                 weight *= sym->count[counter];
416
417         weight /= (sym->count[counter] + 1);
418
419         return weight;
420 }
421
422 static long                     samples;
423 static long                     userspace_samples;
424 static long                     exact_samples;
425 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
426
427 static void __list_insert_active_sym(struct sym_entry *syme)
428 {
429         list_add(&syme->node, &active_symbols);
430 }
431
432 static void list_remove_active_sym(struct sym_entry *syme)
433 {
434         pthread_mutex_lock(&active_symbols_lock);
435         list_del_init(&syme->node);
436         pthread_mutex_unlock(&active_symbols_lock);
437 }
438
439 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
440 {
441         struct rb_node **p = &tree->rb_node;
442         struct rb_node *parent = NULL;
443         struct sym_entry *iter;
444
445         while (*p != NULL) {
446                 parent = *p;
447                 iter = rb_entry(parent, struct sym_entry, rb_node);
448
449                 if (se->weight > iter->weight)
450                         p = &(*p)->rb_left;
451                 else
452                         p = &(*p)->rb_right;
453         }
454
455         rb_link_node(&se->rb_node, parent, p);
456         rb_insert_color(&se->rb_node, tree);
457 }
458
459 static void print_sym_table(void)
460 {
461         int printed = 0, j;
462         int counter, snap = !display_weighted ? sym_counter : 0;
463         float samples_per_sec = samples/delay_secs;
464         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
465         float esamples_percent = (100.0*exact_samples)/samples;
466         float sum_ksamples = 0.0;
467         struct sym_entry *syme, *n;
468         struct rb_root tmp = RB_ROOT;
469         struct rb_node *nd;
470         int sym_width = 0, dso_width = 0, dso_short_width = 0;
471         const int win_width = winsize.ws_col - 1;
472
473         samples = userspace_samples = exact_samples = 0;
474
475         /* Sort the active symbols */
476         pthread_mutex_lock(&active_symbols_lock);
477         syme = list_entry(active_symbols.next, struct sym_entry, node);
478         pthread_mutex_unlock(&active_symbols_lock);
479
480         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
481                 syme->snap_count = syme->count[snap];
482                 if (syme->snap_count != 0) {
483
484                         if ((hide_user_symbols &&
485                              syme->origin == PERF_RECORD_MISC_USER) ||
486                             (hide_kernel_symbols &&
487                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
488                                 list_remove_active_sym(syme);
489                                 continue;
490                         }
491                         syme->weight = sym_weight(syme);
492                         rb_insert_active_sym(&tmp, syme);
493                         sum_ksamples += syme->snap_count;
494
495                         for (j = 0; j < nr_counters; j++)
496                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
497                 } else
498                         list_remove_active_sym(syme);
499         }
500
501         puts(CONSOLE_CLEAR);
502
503         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
504         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%%  exact: %4.1f%% [",
505                 samples_per_sec,
506                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)),
507                 esamples_percent);
508
509         if (nr_counters == 1 || !display_weighted) {
510                 printf("%Ld", (u64)attrs[0].sample_period);
511                 if (freq)
512                         printf("Hz ");
513                 else
514                         printf(" ");
515         }
516
517         if (!display_weighted)
518                 printf("%s", event_name(sym_counter));
519         else for (counter = 0; counter < nr_counters; counter++) {
520                 if (counter)
521                         printf("/");
522
523                 printf("%s", event_name(counter));
524         }
525
526         printf( "], ");
527
528         if (target_pid != -1)
529                 printf(" (target_pid: %d", target_pid);
530         else if (target_tid != -1)
531                 printf(" (target_tid: %d", target_tid);
532         else
533                 printf(" (all");
534
535         if (profile_cpu != -1)
536                 printf(", cpu: %d)\n", profile_cpu);
537         else {
538                 if (target_tid != -1)
539                         printf(")\n");
540                 else
541                         printf(", %d CPUs)\n", nr_cpus);
542         }
543
544         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
545
546         if (sym_filter_entry) {
547                 show_details(sym_filter_entry);
548                 return;
549         }
550
551         /*
552          * Find the longest symbol name that will be displayed
553          */
554         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
555                 syme = rb_entry(nd, struct sym_entry, rb_node);
556                 if (++printed > print_entries ||
557                     (int)syme->snap_count < count_filter)
558                         continue;
559
560                 if (syme->map->dso->long_name_len > dso_width)
561                         dso_width = syme->map->dso->long_name_len;
562
563                 if (syme->map->dso->short_name_len > dso_short_width)
564                         dso_short_width = syme->map->dso->short_name_len;
565
566                 if (syme->name_len > sym_width)
567                         sym_width = syme->name_len;
568         }
569
570         printed = 0;
571
572         if (sym_width + dso_width > winsize.ws_col - 29) {
573                 dso_width = dso_short_width;
574                 if (sym_width + dso_width > winsize.ws_col - 29)
575                         sym_width = winsize.ws_col - dso_width - 29;
576         }
577         putchar('\n');
578         if (nr_counters == 1)
579                 printf("             samples  pcnt");
580         else
581                 printf("   weight    samples  pcnt");
582
583         if (verbose)
584                 printf("         RIP       ");
585         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
586         printf("   %s    _______ _____",
587                nr_counters == 1 ? "      " : "______");
588         if (verbose)
589                 printf(" ________________");
590         printf(" %-*.*s", sym_width, sym_width, graph_line);
591         printf(" %-*.*s", dso_width, dso_width, graph_line);
592         puts("\n");
593
594         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
595                 struct symbol *sym;
596                 double pcnt;
597
598                 syme = rb_entry(nd, struct sym_entry, rb_node);
599                 sym = sym_entry__symbol(syme);
600
601                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
602                         continue;
603
604                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
605                                          sum_ksamples));
606
607                 if (nr_counters == 1 || !display_weighted)
608                         printf("%20.2f ", syme->weight);
609                 else
610                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
611
612                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
613                 if (verbose)
614                         printf(" %016llx", sym->start);
615                 printf(" %-*.*s", sym_width, sym_width, sym->name);
616                 printf(" %-*.*s\n", dso_width, dso_width,
617                        dso_width >= syme->map->dso->long_name_len ?
618                                         syme->map->dso->long_name :
619                                         syme->map->dso->short_name);
620         }
621 }
622
623 static void prompt_integer(int *target, const char *msg)
624 {
625         char *buf = malloc(0), *p;
626         size_t dummy = 0;
627         int tmp;
628
629         fprintf(stdout, "\n%s: ", msg);
630         if (getline(&buf, &dummy, stdin) < 0)
631                 return;
632
633         p = strchr(buf, '\n');
634         if (p)
635                 *p = 0;
636
637         p = buf;
638         while(*p) {
639                 if (!isdigit(*p))
640                         goto out_free;
641                 p++;
642         }
643         tmp = strtoul(buf, NULL, 10);
644         *target = tmp;
645 out_free:
646         free(buf);
647 }
648
649 static void prompt_percent(int *target, const char *msg)
650 {
651         int tmp = 0;
652
653         prompt_integer(&tmp, msg);
654         if (tmp >= 0 && tmp <= 100)
655                 *target = tmp;
656 }
657
658 static void prompt_symbol(struct sym_entry **target, const char *msg)
659 {
660         char *buf = malloc(0), *p;
661         struct sym_entry *syme = *target, *n, *found = NULL;
662         size_t dummy = 0;
663
664         /* zero counters of active symbol */
665         if (syme) {
666                 pthread_mutex_lock(&syme->src->lock);
667                 __zero_source_counters(syme);
668                 *target = NULL;
669                 pthread_mutex_unlock(&syme->src->lock);
670         }
671
672         fprintf(stdout, "\n%s: ", msg);
673         if (getline(&buf, &dummy, stdin) < 0)
674                 goto out_free;
675
676         p = strchr(buf, '\n');
677         if (p)
678                 *p = 0;
679
680         pthread_mutex_lock(&active_symbols_lock);
681         syme = list_entry(active_symbols.next, struct sym_entry, node);
682         pthread_mutex_unlock(&active_symbols_lock);
683
684         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
685                 struct symbol *sym = sym_entry__symbol(syme);
686
687                 if (!strcmp(buf, sym->name)) {
688                         found = syme;
689                         break;
690                 }
691         }
692
693         if (!found) {
694                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
695                 sleep(1);
696                 return;
697         } else
698                 parse_source(found);
699
700 out_free:
701         free(buf);
702 }
703
704 static void print_mapped_keys(void)
705 {
706         char *name = NULL;
707
708         if (sym_filter_entry) {
709                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
710                 name = sym->name;
711         }
712
713         fprintf(stdout, "\nMapped keys:\n");
714         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
715         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
716
717         if (nr_counters > 1)
718                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
719
720         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
721
722         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
723         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
724         fprintf(stdout, "\t[S]     stop annotation.\n");
725
726         if (nr_counters > 1)
727                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
728
729         fprintf(stdout,
730                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
731                 hide_kernel_symbols ? "yes" : "no");
732         fprintf(stdout,
733                 "\t[U]     hide user symbols.               \t(%s)\n",
734                 hide_user_symbols ? "yes" : "no");
735         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
736         fprintf(stdout, "\t[qQ]    quit.\n");
737 }
738
739 static int key_mapped(int c)
740 {
741         switch (c) {
742                 case 'd':
743                 case 'e':
744                 case 'f':
745                 case 'z':
746                 case 'q':
747                 case 'Q':
748                 case 'K':
749                 case 'U':
750                 case 'F':
751                 case 's':
752                 case 'S':
753                         return 1;
754                 case 'E':
755                 case 'w':
756                         return nr_counters > 1 ? 1 : 0;
757                 default:
758                         break;
759         }
760
761         return 0;
762 }
763
764 static void handle_keypress(int c)
765 {
766         if (!key_mapped(c)) {
767                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
768                 struct termios tc, save;
769
770                 print_mapped_keys();
771                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
772                 fflush(stdout);
773
774                 tcgetattr(0, &save);
775                 tc = save;
776                 tc.c_lflag &= ~(ICANON | ECHO);
777                 tc.c_cc[VMIN] = 0;
778                 tc.c_cc[VTIME] = 0;
779                 tcsetattr(0, TCSANOW, &tc);
780
781                 poll(&stdin_poll, 1, -1);
782                 c = getc(stdin);
783
784                 tcsetattr(0, TCSAFLUSH, &save);
785                 if (!key_mapped(c))
786                         return;
787         }
788
789         switch (c) {
790                 case 'd':
791                         prompt_integer(&delay_secs, "Enter display delay");
792                         if (delay_secs < 1)
793                                 delay_secs = 1;
794                         break;
795                 case 'e':
796                         prompt_integer(&print_entries, "Enter display entries (lines)");
797                         if (print_entries == 0) {
798                                 sig_winch_handler(SIGWINCH);
799                                 signal(SIGWINCH, sig_winch_handler);
800                         } else
801                                 signal(SIGWINCH, SIG_DFL);
802                         break;
803                 case 'E':
804                         if (nr_counters > 1) {
805                                 int i;
806
807                                 fprintf(stderr, "\nAvailable events:");
808                                 for (i = 0; i < nr_counters; i++)
809                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
810
811                                 prompt_integer(&sym_counter, "Enter details event counter");
812
813                                 if (sym_counter >= nr_counters) {
814                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
815                                         sym_counter = 0;
816                                         sleep(1);
817                                 }
818                         } else sym_counter = 0;
819                         break;
820                 case 'f':
821                         prompt_integer(&count_filter, "Enter display event count filter");
822                         break;
823                 case 'F':
824                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
825                         break;
826                 case 'K':
827                         hide_kernel_symbols = !hide_kernel_symbols;
828                         break;
829                 case 'q':
830                 case 'Q':
831                         printf("exiting.\n");
832                         if (dump_symtab)
833                                 dsos__fprintf(stderr);
834                         exit(0);
835                 case 's':
836                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
837                         break;
838                 case 'S':
839                         if (!sym_filter_entry)
840                                 break;
841                         else {
842                                 struct sym_entry *syme = sym_filter_entry;
843
844                                 pthread_mutex_lock(&syme->src->lock);
845                                 sym_filter_entry = NULL;
846                                 __zero_source_counters(syme);
847                                 pthread_mutex_unlock(&syme->src->lock);
848                         }
849                         break;
850                 case 'U':
851                         hide_user_symbols = !hide_user_symbols;
852                         break;
853                 case 'w':
854                         display_weighted = ~display_weighted;
855                         break;
856                 case 'z':
857                         zero = ~zero;
858                         break;
859                 default:
860                         break;
861         }
862 }
863
864 static void *display_thread(void *arg __used)
865 {
866         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
867         struct termios tc, save;
868         int delay_msecs, c;
869
870         tcgetattr(0, &save);
871         tc = save;
872         tc.c_lflag &= ~(ICANON | ECHO);
873         tc.c_cc[VMIN] = 0;
874         tc.c_cc[VTIME] = 0;
875
876 repeat:
877         delay_msecs = delay_secs * 1000;
878         tcsetattr(0, TCSANOW, &tc);
879         /* trash return*/
880         getc(stdin);
881
882         do {
883                 print_sym_table();
884         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
885
886         c = getc(stdin);
887         tcsetattr(0, TCSAFLUSH, &save);
888
889         handle_keypress(c);
890         goto repeat;
891
892         return NULL;
893 }
894
895 /* Tag samples to be skipped. */
896 static const char *skip_symbols[] = {
897         "default_idle",
898         "cpu_idle",
899         "enter_idle",
900         "exit_idle",
901         "mwait_idle",
902         "mwait_idle_with_hints",
903         "poll_idle",
904         "ppc64_runlatch_off",
905         "pseries_dedicated_idle_sleep",
906         NULL
907 };
908
909 static int symbol_filter(struct map *map, struct symbol *sym)
910 {
911         struct sym_entry *syme;
912         const char *name = sym->name;
913         int i;
914
915         /*
916          * ppc64 uses function descriptors and appends a '.' to the
917          * start of every instruction address. Remove it.
918          */
919         if (name[0] == '.')
920                 name++;
921
922         if (!strcmp(name, "_text") ||
923             !strcmp(name, "_etext") ||
924             !strcmp(name, "_sinittext") ||
925             !strncmp("init_module", name, 11) ||
926             !strncmp("cleanup_module", name, 14) ||
927             strstr(name, "_text_start") ||
928             strstr(name, "_text_end"))
929                 return 1;
930
931         syme = symbol__priv(sym);
932         syme->map = map;
933         syme->src = NULL;
934
935         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
936                 /* schedule initial sym_filter_entry setup */
937                 sym_filter_entry_sched = syme;
938                 sym_filter = NULL;
939         }
940
941         for (i = 0; skip_symbols[i]; i++) {
942                 if (!strcmp(skip_symbols[i], name)) {
943                         syme->skip = 1;
944                         break;
945                 }
946         }
947
948         if (!syme->skip)
949                 syme->name_len = strlen(sym->name);
950
951         return 0;
952 }
953
954 static void event__process_sample(const event_t *self,
955                                  struct perf_session *session, int counter)
956 {
957         u64 ip = self->ip.ip;
958         struct sym_entry *syme;
959         struct addr_location al;
960         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
961
962         ++samples;
963
964         switch (origin) {
965         case PERF_RECORD_MISC_USER:
966                 ++userspace_samples;
967                 if (hide_user_symbols)
968                         return;
969                 break;
970         case PERF_RECORD_MISC_KERNEL:
971                 if (hide_kernel_symbols)
972                         return;
973                 break;
974         default:
975                 return;
976         }
977
978         if (self->header.misc & PERF_RECORD_MISC_EXACT)
979                 exact_samples++;
980
981         if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
982             al.filtered)
983                 return;
984
985         if (al.sym == NULL) {
986                 /*
987                  * As we do lazy loading of symtabs we only will know if the
988                  * specified vmlinux file is invalid when we actually have a
989                  * hit in kernel space and then try to load it. So if we get
990                  * here and there are _no_ symbols in the DSO backing the
991                  * kernel map, bail out.
992                  *
993                  * We may never get here, for instance, if we use -K/
994                  * --hide-kernel-symbols, even if the user specifies an
995                  * invalid --vmlinux ;-)
996                  */
997                 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
998                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
999                         pr_err("The %s file can't be used\n",
1000                                symbol_conf.vmlinux_name);
1001                         exit(1);
1002                 }
1003
1004                 return;
1005         }
1006
1007         /* let's see, whether we need to install initial sym_filter_entry */
1008         if (sym_filter_entry_sched) {
1009                 sym_filter_entry = sym_filter_entry_sched;
1010                 sym_filter_entry_sched = NULL;
1011                 if (parse_source(sym_filter_entry) < 0) {
1012                         struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1013
1014                         pr_err("Can't annotate %s", sym->name);
1015                         if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1016                                 pr_err(": No vmlinux file was found in the path:\n");
1017                                 vmlinux_path__fprintf(stderr);
1018                         } else
1019                                 pr_err(".\n");
1020                         exit(1);
1021                 }
1022         }
1023
1024         syme = symbol__priv(al.sym);
1025         if (!syme->skip) {
1026                 syme->count[counter]++;
1027                 syme->origin = origin;
1028                 record_precise_ip(syme, counter, ip);
1029                 pthread_mutex_lock(&active_symbols_lock);
1030                 if (list_empty(&syme->node) || !syme->node.next)
1031                         __list_insert_active_sym(syme);
1032                 pthread_mutex_unlock(&active_symbols_lock);
1033         }
1034 }
1035
1036 static int event__process(event_t *event, struct perf_session *session)
1037 {
1038         switch (event->header.type) {
1039         case PERF_RECORD_COMM:
1040                 event__process_comm(event, session);
1041                 break;
1042         case PERF_RECORD_MMAP:
1043                 event__process_mmap(event, session);
1044                 break;
1045         case PERF_RECORD_FORK:
1046         case PERF_RECORD_EXIT:
1047                 event__process_task(event, session);
1048                 break;
1049         default:
1050                 break;
1051         }
1052
1053         return 0;
1054 }
1055
1056 struct mmap_data {
1057         int                     counter;
1058         void                    *base;
1059         int                     mask;
1060         unsigned int            prev;
1061 };
1062
1063 static unsigned int mmap_read_head(struct mmap_data *md)
1064 {
1065         struct perf_event_mmap_page *pc = md->base;
1066         int head;
1067
1068         head = pc->data_head;
1069         rmb();
1070
1071         return head;
1072 }
1073
1074 static void perf_session__mmap_read_counter(struct perf_session *self,
1075                                             struct mmap_data *md)
1076 {
1077         unsigned int head = mmap_read_head(md);
1078         unsigned int old = md->prev;
1079         unsigned char *data = md->base + page_size;
1080         int diff;
1081
1082         /*
1083          * If we're further behind than half the buffer, there's a chance
1084          * the writer will bite our tail and mess up the samples under us.
1085          *
1086          * If we somehow ended up ahead of the head, we got messed up.
1087          *
1088          * In either case, truncate and restart at head.
1089          */
1090         diff = head - old;
1091         if (diff > md->mask / 2 || diff < 0) {
1092                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1093
1094                 /*
1095                  * head points to a known good entry, start there.
1096                  */
1097                 old = head;
1098         }
1099
1100         for (; old != head;) {
1101                 event_t *event = (event_t *)&data[old & md->mask];
1102
1103                 event_t event_copy;
1104
1105                 size_t size = event->header.size;
1106
1107                 /*
1108                  * Event straddles the mmap boundary -- header should always
1109                  * be inside due to u64 alignment of output.
1110                  */
1111                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1112                         unsigned int offset = old;
1113                         unsigned int len = min(sizeof(*event), size), cpy;
1114                         void *dst = &event_copy;
1115
1116                         do {
1117                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1118                                 memcpy(dst, &data[offset & md->mask], cpy);
1119                                 offset += cpy;
1120                                 dst += cpy;
1121                                 len -= cpy;
1122                         } while (len);
1123
1124                         event = &event_copy;
1125                 }
1126
1127                 if (event->header.type == PERF_RECORD_SAMPLE)
1128                         event__process_sample(event, self, md->counter);
1129                 else
1130                         event__process(event, self);
1131                 old += size;
1132         }
1133
1134         md->prev = old;
1135 }
1136
1137 static struct pollfd *event_array;
1138 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1139
1140 static void perf_session__mmap_read(struct perf_session *self)
1141 {
1142         int i, counter, thread_index;
1143
1144         for (i = 0; i < nr_cpus; i++) {
1145                 for (counter = 0; counter < nr_counters; counter++)
1146                         for (thread_index = 0;
1147                                 thread_index < thread_num;
1148                                 thread_index++) {
1149                                 perf_session__mmap_read_counter(self,
1150                                         &mmap_array[i][counter][thread_index]);
1151                         }
1152         }
1153 }
1154
1155 int nr_poll;
1156 int group_fd;
1157
1158 static void start_counter(int i, int counter)
1159 {
1160         struct perf_event_attr *attr;
1161         int cpu;
1162         int thread_index;
1163
1164         cpu = profile_cpu;
1165         if (target_tid == -1 && profile_cpu == -1)
1166                 cpu = cpumap[i];
1167
1168         attr = attrs + counter;
1169
1170         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1171
1172         if (freq) {
1173                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1174                 attr->freq              = 1;
1175                 attr->sample_freq       = freq;
1176         }
1177
1178         attr->inherit           = (cpu < 0) && inherit;
1179         attr->mmap              = 1;
1180
1181         for (thread_index = 0; thread_index < thread_num; thread_index++) {
1182 try_again:
1183                 fd[i][counter][thread_index] = sys_perf_event_open(attr,
1184                                 all_tids[thread_index], cpu, group_fd, 0);
1185
1186                 if (fd[i][counter][thread_index] < 0) {
1187                         int err = errno;
1188
1189                         if (err == EPERM || err == EACCES)
1190                                 die("No permission - are you root?\n");
1191                         /*
1192                          * If it's cycles then fall back to hrtimer
1193                          * based cpu-clock-tick sw counter, which
1194                          * is always available even if no PMU support:
1195                          */
1196                         if (attr->type == PERF_TYPE_HARDWARE
1197                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1198
1199                                 if (verbose)
1200                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
1201
1202                                 attr->type = PERF_TYPE_SOFTWARE;
1203                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1204                                 goto try_again;
1205                         }
1206                         printf("\n");
1207                         error("perfcounter syscall returned with %d (%s)\n",
1208                                         fd[i][counter][thread_index], strerror(err));
1209                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1210                         exit(-1);
1211                 }
1212                 assert(fd[i][counter][thread_index] >= 0);
1213                 fcntl(fd[i][counter][thread_index], F_SETFL, O_NONBLOCK);
1214
1215                 /*
1216                  * First counter acts as the group leader:
1217                  */
1218                 if (group && group_fd == -1)
1219                         group_fd = fd[i][counter][thread_index];
1220
1221                 event_array[nr_poll].fd = fd[i][counter][thread_index];
1222                 event_array[nr_poll].events = POLLIN;
1223                 nr_poll++;
1224
1225                 mmap_array[i][counter][thread_index].counter = counter;
1226                 mmap_array[i][counter][thread_index].prev = 0;
1227                 mmap_array[i][counter][thread_index].mask = mmap_pages*page_size - 1;
1228                 mmap_array[i][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
1229                                 PROT_READ, MAP_SHARED, fd[i][counter][thread_index], 0);
1230                 if (mmap_array[i][counter][thread_index].base == MAP_FAILED)
1231                         die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1232         }
1233 }
1234
1235 static int __cmd_top(void)
1236 {
1237         pthread_t thread;
1238         int i, counter;
1239         int ret;
1240         /*
1241          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1242          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1243          */
1244         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1245         if (session == NULL)
1246                 return -ENOMEM;
1247
1248         if (target_tid != -1)
1249                 event__synthesize_thread(target_tid, event__process, session);
1250         else
1251                 event__synthesize_threads(event__process, session);
1252
1253         for (i = 0; i < nr_cpus; i++) {
1254                 group_fd = -1;
1255                 for (counter = 0; counter < nr_counters; counter++)
1256                         start_counter(i, counter);
1257         }
1258
1259         /* Wait for a minimal set of events before starting the snapshot */
1260         poll(&event_array[0], nr_poll, 100);
1261
1262         perf_session__mmap_read(session);
1263
1264         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1265                 printf("Could not create display thread.\n");
1266                 exit(-1);
1267         }
1268
1269         if (realtime_prio) {
1270                 struct sched_param param;
1271
1272                 param.sched_priority = realtime_prio;
1273                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1274                         printf("Could not set realtime priority.\n");
1275                         exit(-1);
1276                 }
1277         }
1278
1279         while (1) {
1280                 int hits = samples;
1281
1282                 perf_session__mmap_read(session);
1283
1284                 if (hits == samples)
1285                         ret = poll(event_array, nr_poll, 100);
1286         }
1287
1288         return 0;
1289 }
1290
1291 static const char * const top_usage[] = {
1292         "perf top [<options>]",
1293         NULL
1294 };
1295
1296 static const struct option options[] = {
1297         OPT_CALLBACK('e', "event", NULL, "event",
1298                      "event selector. use 'perf list' to list available events",
1299                      parse_events),
1300         OPT_INTEGER('c', "count", &default_interval,
1301                     "event period to sample"),
1302         OPT_INTEGER('p', "pid", &target_pid,
1303                     "profile events on existing process id"),
1304         OPT_INTEGER('t', "tid", &target_tid,
1305                     "profile events on existing thread id"),
1306         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1307                             "system-wide collection from all CPUs"),
1308         OPT_INTEGER('C', "CPU", &profile_cpu,
1309                     "CPU to profile on"),
1310         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1311                    "file", "vmlinux pathname"),
1312         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1313                     "hide kernel symbols"),
1314         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1315                     "number of mmap data pages"),
1316         OPT_INTEGER('r', "realtime", &realtime_prio,
1317                     "collect data with this RT SCHED_FIFO priority"),
1318         OPT_INTEGER('d', "delay", &delay_secs,
1319                     "number of seconds to delay between refreshes"),
1320         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1321                             "dump the symbol table used for profiling"),
1322         OPT_INTEGER('f', "count-filter", &count_filter,
1323                     "only display functions with more events than this"),
1324         OPT_BOOLEAN('g', "group", &group,
1325                             "put the counters into a counter group"),
1326         OPT_BOOLEAN('i', "inherit", &inherit,
1327                     "child tasks inherit counters"),
1328         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1329                     "symbol to annotate"),
1330         OPT_BOOLEAN('z', "zero", &zero,
1331                     "zero history across updates"),
1332         OPT_INTEGER('F', "freq", &freq,
1333                     "profile at this frequency"),
1334         OPT_INTEGER('E', "entries", &print_entries,
1335                     "display this many functions"),
1336         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1337                     "hide user symbols"),
1338         OPT_BOOLEAN('v', "verbose", &verbose,
1339                     "be more verbose (show counter open errors, etc)"),
1340         OPT_END()
1341 };
1342
1343 int cmd_top(int argc, const char **argv, const char *prefix __used)
1344 {
1345         int counter;
1346         int i,j;
1347
1348         page_size = sysconf(_SC_PAGE_SIZE);
1349
1350         argc = parse_options(argc, argv, options, top_usage, 0);
1351         if (argc)
1352                 usage_with_options(top_usage, options);
1353
1354         if (target_pid != -1) {
1355                 target_tid = target_pid;
1356                 thread_num = find_all_tid(target_pid, &all_tids);
1357                 if (thread_num <= 0) {
1358                         fprintf(stderr, "Can't find all threads of pid %d\n",
1359                                 target_pid);
1360                         usage_with_options(top_usage, options);
1361                 }
1362         } else {
1363                 all_tids=malloc(sizeof(pid_t));
1364                 if (!all_tids)
1365                         return -ENOMEM;
1366
1367                 all_tids[0] = target_tid;
1368                 thread_num = 1;
1369         }
1370
1371         for (i = 0; i < MAX_NR_CPUS; i++) {
1372                 for (j = 0; j < MAX_COUNTERS; j++) {
1373                         fd[i][j] = malloc(sizeof(int)*thread_num);
1374                         mmap_array[i][j] = zalloc(
1375                                 sizeof(struct mmap_data)*thread_num);
1376                         if (!fd[i][j] || !mmap_array[i][j])
1377                                 return -ENOMEM;
1378                 }
1379         }
1380         event_array = malloc(
1381                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
1382         if (!event_array)
1383                 return -ENOMEM;
1384
1385         /* CPU and PID are mutually exclusive */
1386         if (target_tid > 0 && profile_cpu != -1) {
1387                 printf("WARNING: PID switch overriding CPU\n");
1388                 sleep(1);
1389                 profile_cpu = -1;
1390         }
1391
1392         if (!nr_counters)
1393                 nr_counters = 1;
1394
1395         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1396                                  (nr_counters + 1) * sizeof(unsigned long));
1397
1398         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1399         if (symbol__init() < 0)
1400                 return -1;
1401
1402         if (delay_secs < 1)
1403                 delay_secs = 1;
1404
1405         /*
1406          * User specified count overrides default frequency.
1407          */
1408         if (default_interval)
1409                 freq = 0;
1410         else if (freq) {
1411                 default_interval = freq;
1412         } else {
1413                 fprintf(stderr, "frequency and count are zero, aborting\n");
1414                 exit(EXIT_FAILURE);
1415         }
1416
1417         /*
1418          * Fill in the ones not specifically initialized via -c:
1419          */
1420         for (counter = 0; counter < nr_counters; counter++) {
1421                 if (attrs[counter].sample_period)
1422                         continue;
1423
1424                 attrs[counter].sample_period = default_interval;
1425         }
1426
1427         if (target_tid != -1 || profile_cpu != -1)
1428                 nr_cpus = 1;
1429         else
1430                 nr_cpus = read_cpu_map();
1431
1432         get_term_dimensions(&winsize);
1433         if (print_entries == 0) {
1434                 update_print_entries(&winsize);
1435                 signal(SIGWINCH, sig_winch_handler);
1436         }
1437
1438         return __cmd_top();
1439 }