Merge branch 'tty-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh...
[pandora-kernel.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 #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 bool                     system_wide                     =  false;
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 bool                     inherit                         =  false;
72 static int                      nr_cpus                         =      0;
73 static int                      realtime_prio                   =      0;
74 static bool                     group                           =  false;
75 static unsigned int             page_size;
76 static unsigned int             mmap_pages                      =     16;
77 static int                      freq                            =   1000; /* 1 KHz */
78
79 static int                      delay_secs                      =      2;
80 static bool                     zero                            =  false;
81 static bool                     dump_symtab                     =  false;
82
83 static bool                     hide_kernel_symbols             =  false;
84 static bool                     hide_user_symbols               =  false;
85 static struct winsize           winsize;
86
87 /*
88  * Source
89  */
90
91 struct source_line {
92         u64                     eip;
93         unsigned long           count[MAX_COUNTERS];
94         char                    *line;
95         struct source_line      *next;
96 };
97
98 static const char               *sym_filter                     =   NULL;
99 struct sym_entry                *sym_filter_entry               =   NULL;
100 struct sym_entry                *sym_filter_entry_sched         =   NULL;
101 static int                      sym_pcnt_filter                 =      5;
102 static int                      sym_counter                     =      0;
103 static int                      display_weighted                =     -1;
104 static const char               *cpu_list;
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                     kernel_samples, us_samples;
424 static long                     exact_samples;
425 static long                     guest_us_samples, guest_kernel_samples;
426 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
427
428 static void __list_insert_active_sym(struct sym_entry *syme)
429 {
430         list_add(&syme->node, &active_symbols);
431 }
432
433 static void list_remove_active_sym(struct sym_entry *syme)
434 {
435         pthread_mutex_lock(&active_symbols_lock);
436         list_del_init(&syme->node);
437         pthread_mutex_unlock(&active_symbols_lock);
438 }
439
440 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
441 {
442         struct rb_node **p = &tree->rb_node;
443         struct rb_node *parent = NULL;
444         struct sym_entry *iter;
445
446         while (*p != NULL) {
447                 parent = *p;
448                 iter = rb_entry(parent, struct sym_entry, rb_node);
449
450                 if (se->weight > iter->weight)
451                         p = &(*p)->rb_left;
452                 else
453                         p = &(*p)->rb_right;
454         }
455
456         rb_link_node(&se->rb_node, parent, p);
457         rb_insert_color(&se->rb_node, tree);
458 }
459
460 static void print_sym_table(void)
461 {
462         int printed = 0, j;
463         int counter, snap = !display_weighted ? sym_counter : 0;
464         float samples_per_sec = samples/delay_secs;
465         float ksamples_per_sec = kernel_samples/delay_secs;
466         float us_samples_per_sec = (us_samples)/delay_secs;
467         float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs;
468         float guest_us_samples_per_sec = (guest_us_samples)/delay_secs;
469         float esamples_percent = (100.0*exact_samples)/samples;
470         float sum_ksamples = 0.0;
471         struct sym_entry *syme, *n;
472         struct rb_root tmp = RB_ROOT;
473         struct rb_node *nd;
474         int sym_width = 0, dso_width = 0, dso_short_width = 0;
475         const int win_width = winsize.ws_col - 1;
476
477         samples = us_samples = kernel_samples = exact_samples = 0;
478         guest_kernel_samples = guest_us_samples = 0;
479
480         /* Sort the active symbols */
481         pthread_mutex_lock(&active_symbols_lock);
482         syme = list_entry(active_symbols.next, struct sym_entry, node);
483         pthread_mutex_unlock(&active_symbols_lock);
484
485         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
486                 syme->snap_count = syme->count[snap];
487                 if (syme->snap_count != 0) {
488
489                         if ((hide_user_symbols &&
490                              syme->origin == PERF_RECORD_MISC_USER) ||
491                             (hide_kernel_symbols &&
492                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
493                                 list_remove_active_sym(syme);
494                                 continue;
495                         }
496                         syme->weight = sym_weight(syme);
497                         rb_insert_active_sym(&tmp, syme);
498                         sum_ksamples += syme->snap_count;
499
500                         for (j = 0; j < nr_counters; j++)
501                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
502                 } else
503                         list_remove_active_sym(syme);
504         }
505
506         puts(CONSOLE_CLEAR);
507
508         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
509         if (!perf_guest) {
510                 printf("   PerfTop:%8.0f irqs/sec  kernel:%4.1f%%"
511                         "  exact: %4.1f%% [",
512                         samples_per_sec,
513                         100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) /
514                                          samples_per_sec)),
515                         esamples_percent);
516         } else {
517                 printf("   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% us:%4.1f%%"
518                         " guest kernel:%4.1f%% guest us:%4.1f%%"
519                         " exact: %4.1f%% [",
520                         samples_per_sec,
521                         100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) /
522                                           samples_per_sec)),
523                         100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) /
524                                           samples_per_sec)),
525                         100.0 - (100.0 * ((samples_per_sec -
526                                                 guest_kernel_samples_per_sec) /
527                                           samples_per_sec)),
528                         100.0 - (100.0 * ((samples_per_sec -
529                                            guest_us_samples_per_sec) /
530                                           samples_per_sec)),
531                         esamples_percent);
532         }
533
534         if (nr_counters == 1 || !display_weighted) {
535                 printf("%Ld", (u64)attrs[0].sample_period);
536                 if (freq)
537                         printf("Hz ");
538                 else
539                         printf(" ");
540         }
541
542         if (!display_weighted)
543                 printf("%s", event_name(sym_counter));
544         else for (counter = 0; counter < nr_counters; counter++) {
545                 if (counter)
546                         printf("/");
547
548                 printf("%s", event_name(counter));
549         }
550
551         printf( "], ");
552
553         if (target_pid != -1)
554                 printf(" (target_pid: %d", target_pid);
555         else if (target_tid != -1)
556                 printf(" (target_tid: %d", target_tid);
557         else
558                 printf(" (all");
559
560         if (cpu_list)
561                 printf(", CPU%s: %s)\n", nr_cpus > 1 ? "s" : "", cpu_list);
562         else {
563                 if (target_tid != -1)
564                         printf(")\n");
565                 else
566                         printf(", %d CPU%s)\n", nr_cpus, nr_cpus > 1 ? "s" : "");
567         }
568
569         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
570
571         if (sym_filter_entry) {
572                 show_details(sym_filter_entry);
573                 return;
574         }
575
576         /*
577          * Find the longest symbol name that will be displayed
578          */
579         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
580                 syme = rb_entry(nd, struct sym_entry, rb_node);
581                 if (++printed > print_entries ||
582                     (int)syme->snap_count < count_filter)
583                         continue;
584
585                 if (syme->map->dso->long_name_len > dso_width)
586                         dso_width = syme->map->dso->long_name_len;
587
588                 if (syme->map->dso->short_name_len > dso_short_width)
589                         dso_short_width = syme->map->dso->short_name_len;
590
591                 if (syme->name_len > sym_width)
592                         sym_width = syme->name_len;
593         }
594
595         printed = 0;
596
597         if (sym_width + dso_width > winsize.ws_col - 29) {
598                 dso_width = dso_short_width;
599                 if (sym_width + dso_width > winsize.ws_col - 29)
600                         sym_width = winsize.ws_col - dso_width - 29;
601         }
602         putchar('\n');
603         if (nr_counters == 1)
604                 printf("             samples  pcnt");
605         else
606                 printf("   weight    samples  pcnt");
607
608         if (verbose)
609                 printf("         RIP       ");
610         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
611         printf("   %s    _______ _____",
612                nr_counters == 1 ? "      " : "______");
613         if (verbose)
614                 printf(" ________________");
615         printf(" %-*.*s", sym_width, sym_width, graph_line);
616         printf(" %-*.*s", dso_width, dso_width, graph_line);
617         puts("\n");
618
619         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
620                 struct symbol *sym;
621                 double pcnt;
622
623                 syme = rb_entry(nd, struct sym_entry, rb_node);
624                 sym = sym_entry__symbol(syme);
625                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
626                         continue;
627
628                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
629                                          sum_ksamples));
630
631                 if (nr_counters == 1 || !display_weighted)
632                         printf("%20.2f ", syme->weight);
633                 else
634                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
635
636                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
637                 if (verbose)
638                         printf(" %016llx", sym->start);
639                 printf(" %-*.*s", sym_width, sym_width, sym->name);
640                 printf(" %-*.*s\n", dso_width, dso_width,
641                        dso_width >= syme->map->dso->long_name_len ?
642                                         syme->map->dso->long_name :
643                                         syme->map->dso->short_name);
644         }
645 }
646
647 static void prompt_integer(int *target, const char *msg)
648 {
649         char *buf = malloc(0), *p;
650         size_t dummy = 0;
651         int tmp;
652
653         fprintf(stdout, "\n%s: ", msg);
654         if (getline(&buf, &dummy, stdin) < 0)
655                 return;
656
657         p = strchr(buf, '\n');
658         if (p)
659                 *p = 0;
660
661         p = buf;
662         while(*p) {
663                 if (!isdigit(*p))
664                         goto out_free;
665                 p++;
666         }
667         tmp = strtoul(buf, NULL, 10);
668         *target = tmp;
669 out_free:
670         free(buf);
671 }
672
673 static void prompt_percent(int *target, const char *msg)
674 {
675         int tmp = 0;
676
677         prompt_integer(&tmp, msg);
678         if (tmp >= 0 && tmp <= 100)
679                 *target = tmp;
680 }
681
682 static void prompt_symbol(struct sym_entry **target, const char *msg)
683 {
684         char *buf = malloc(0), *p;
685         struct sym_entry *syme = *target, *n, *found = NULL;
686         size_t dummy = 0;
687
688         /* zero counters of active symbol */
689         if (syme) {
690                 pthread_mutex_lock(&syme->src->lock);
691                 __zero_source_counters(syme);
692                 *target = NULL;
693                 pthread_mutex_unlock(&syme->src->lock);
694         }
695
696         fprintf(stdout, "\n%s: ", msg);
697         if (getline(&buf, &dummy, stdin) < 0)
698                 goto out_free;
699
700         p = strchr(buf, '\n');
701         if (p)
702                 *p = 0;
703
704         pthread_mutex_lock(&active_symbols_lock);
705         syme = list_entry(active_symbols.next, struct sym_entry, node);
706         pthread_mutex_unlock(&active_symbols_lock);
707
708         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
709                 struct symbol *sym = sym_entry__symbol(syme);
710
711                 if (!strcmp(buf, sym->name)) {
712                         found = syme;
713                         break;
714                 }
715         }
716
717         if (!found) {
718                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
719                 sleep(1);
720                 return;
721         } else
722                 parse_source(found);
723
724 out_free:
725         free(buf);
726 }
727
728 static void print_mapped_keys(void)
729 {
730         char *name = NULL;
731
732         if (sym_filter_entry) {
733                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
734                 name = sym->name;
735         }
736
737         fprintf(stdout, "\nMapped keys:\n");
738         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
739         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
740
741         if (nr_counters > 1)
742                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
743
744         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
745
746         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
747         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
748         fprintf(stdout, "\t[S]     stop annotation.\n");
749
750         if (nr_counters > 1)
751                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
752
753         fprintf(stdout,
754                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
755                 hide_kernel_symbols ? "yes" : "no");
756         fprintf(stdout,
757                 "\t[U]     hide user symbols.               \t(%s)\n",
758                 hide_user_symbols ? "yes" : "no");
759         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
760         fprintf(stdout, "\t[qQ]    quit.\n");
761 }
762
763 static int key_mapped(int c)
764 {
765         switch (c) {
766                 case 'd':
767                 case 'e':
768                 case 'f':
769                 case 'z':
770                 case 'q':
771                 case 'Q':
772                 case 'K':
773                 case 'U':
774                 case 'F':
775                 case 's':
776                 case 'S':
777                         return 1;
778                 case 'E':
779                 case 'w':
780                         return nr_counters > 1 ? 1 : 0;
781                 default:
782                         break;
783         }
784
785         return 0;
786 }
787
788 static void handle_keypress(struct perf_session *session, int c)
789 {
790         if (!key_mapped(c)) {
791                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
792                 struct termios tc, save;
793
794                 print_mapped_keys();
795                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
796                 fflush(stdout);
797
798                 tcgetattr(0, &save);
799                 tc = save;
800                 tc.c_lflag &= ~(ICANON | ECHO);
801                 tc.c_cc[VMIN] = 0;
802                 tc.c_cc[VTIME] = 0;
803                 tcsetattr(0, TCSANOW, &tc);
804
805                 poll(&stdin_poll, 1, -1);
806                 c = getc(stdin);
807
808                 tcsetattr(0, TCSAFLUSH, &save);
809                 if (!key_mapped(c))
810                         return;
811         }
812
813         switch (c) {
814                 case 'd':
815                         prompt_integer(&delay_secs, "Enter display delay");
816                         if (delay_secs < 1)
817                                 delay_secs = 1;
818                         break;
819                 case 'e':
820                         prompt_integer(&print_entries, "Enter display entries (lines)");
821                         if (print_entries == 0) {
822                                 sig_winch_handler(SIGWINCH);
823                                 signal(SIGWINCH, sig_winch_handler);
824                         } else
825                                 signal(SIGWINCH, SIG_DFL);
826                         break;
827                 case 'E':
828                         if (nr_counters > 1) {
829                                 int i;
830
831                                 fprintf(stderr, "\nAvailable events:");
832                                 for (i = 0; i < nr_counters; i++)
833                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
834
835                                 prompt_integer(&sym_counter, "Enter details event counter");
836
837                                 if (sym_counter >= nr_counters) {
838                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
839                                         sym_counter = 0;
840                                         sleep(1);
841                                 }
842                         } else sym_counter = 0;
843                         break;
844                 case 'f':
845                         prompt_integer(&count_filter, "Enter display event count filter");
846                         break;
847                 case 'F':
848                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
849                         break;
850                 case 'K':
851                         hide_kernel_symbols = !hide_kernel_symbols;
852                         break;
853                 case 'q':
854                 case 'Q':
855                         printf("exiting.\n");
856                         if (dump_symtab)
857                                 perf_session__fprintf_dsos(session, stderr);
858                         exit(0);
859                 case 's':
860                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
861                         break;
862                 case 'S':
863                         if (!sym_filter_entry)
864                                 break;
865                         else {
866                                 struct sym_entry *syme = sym_filter_entry;
867
868                                 pthread_mutex_lock(&syme->src->lock);
869                                 sym_filter_entry = NULL;
870                                 __zero_source_counters(syme);
871                                 pthread_mutex_unlock(&syme->src->lock);
872                         }
873                         break;
874                 case 'U':
875                         hide_user_symbols = !hide_user_symbols;
876                         break;
877                 case 'w':
878                         display_weighted = ~display_weighted;
879                         break;
880                 case 'z':
881                         zero = !zero;
882                         break;
883                 default:
884                         break;
885         }
886 }
887
888 static void *display_thread(void *arg __used)
889 {
890         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
891         struct termios tc, save;
892         int delay_msecs, c;
893         struct perf_session *session = (struct perf_session *) arg;
894
895         tcgetattr(0, &save);
896         tc = save;
897         tc.c_lflag &= ~(ICANON | ECHO);
898         tc.c_cc[VMIN] = 0;
899         tc.c_cc[VTIME] = 0;
900
901 repeat:
902         delay_msecs = delay_secs * 1000;
903         tcsetattr(0, TCSANOW, &tc);
904         /* trash return*/
905         getc(stdin);
906
907         do {
908                 print_sym_table();
909         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
910
911         c = getc(stdin);
912         tcsetattr(0, TCSAFLUSH, &save);
913
914         handle_keypress(session, c);
915         goto repeat;
916
917         return NULL;
918 }
919
920 /* Tag samples to be skipped. */
921 static const char *skip_symbols[] = {
922         "default_idle",
923         "cpu_idle",
924         "enter_idle",
925         "exit_idle",
926         "mwait_idle",
927         "mwait_idle_with_hints",
928         "poll_idle",
929         "ppc64_runlatch_off",
930         "pseries_dedicated_idle_sleep",
931         NULL
932 };
933
934 static int symbol_filter(struct map *map, struct symbol *sym)
935 {
936         struct sym_entry *syme;
937         const char *name = sym->name;
938         int i;
939
940         /*
941          * ppc64 uses function descriptors and appends a '.' to the
942          * start of every instruction address. Remove it.
943          */
944         if (name[0] == '.')
945                 name++;
946
947         if (!strcmp(name, "_text") ||
948             !strcmp(name, "_etext") ||
949             !strcmp(name, "_sinittext") ||
950             !strncmp("init_module", name, 11) ||
951             !strncmp("cleanup_module", name, 14) ||
952             strstr(name, "_text_start") ||
953             strstr(name, "_text_end"))
954                 return 1;
955
956         syme = symbol__priv(sym);
957         syme->map = map;
958         syme->src = NULL;
959
960         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
961                 /* schedule initial sym_filter_entry setup */
962                 sym_filter_entry_sched = syme;
963                 sym_filter = NULL;
964         }
965
966         for (i = 0; skip_symbols[i]; i++) {
967                 if (!strcmp(skip_symbols[i], name)) {
968                         syme->skip = 1;
969                         break;
970                 }
971         }
972
973         if (!syme->skip)
974                 syme->name_len = strlen(sym->name);
975
976         return 0;
977 }
978
979 static void event__process_sample(const event_t *self,
980                                  struct perf_session *session, int counter)
981 {
982         u64 ip = self->ip.ip;
983         struct sym_entry *syme;
984         struct addr_location al;
985         struct sample_data data;
986         struct machine *machine;
987         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
988
989         ++samples;
990
991         switch (origin) {
992         case PERF_RECORD_MISC_USER:
993                 ++us_samples;
994                 if (hide_user_symbols)
995                         return;
996                 machine = perf_session__find_host_machine(session);
997                 break;
998         case PERF_RECORD_MISC_KERNEL:
999                 ++kernel_samples;
1000                 if (hide_kernel_symbols)
1001                         return;
1002                 machine = perf_session__find_host_machine(session);
1003                 break;
1004         case PERF_RECORD_MISC_GUEST_KERNEL:
1005                 ++guest_kernel_samples;
1006                 machine = perf_session__find_machine(session, self->ip.pid);
1007                 break;
1008         case PERF_RECORD_MISC_GUEST_USER:
1009                 ++guest_us_samples;
1010                 /*
1011                  * TODO: we don't process guest user from host side
1012                  * except simple counting.
1013                  */
1014                 return;
1015         default:
1016                 return;
1017         }
1018
1019         if (!machine && perf_guest) {
1020                 pr_err("Can't find guest [%d]'s kernel information\n",
1021                         self->ip.pid);
1022                 return;
1023         }
1024
1025         if (self->header.misc & PERF_RECORD_MISC_EXACT_IP)
1026                 exact_samples++;
1027
1028         if (event__preprocess_sample(self, session, &al, &data,
1029                                      symbol_filter) < 0 ||
1030             al.filtered)
1031                 return;
1032
1033         if (al.sym == NULL) {
1034                 /*
1035                  * As we do lazy loading of symtabs we only will know if the
1036                  * specified vmlinux file is invalid when we actually have a
1037                  * hit in kernel space and then try to load it. So if we get
1038                  * here and there are _no_ symbols in the DSO backing the
1039                  * kernel map, bail out.
1040                  *
1041                  * We may never get here, for instance, if we use -K/
1042                  * --hide-kernel-symbols, even if the user specifies an
1043                  * invalid --vmlinux ;-)
1044                  */
1045                 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
1046                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
1047                         pr_err("The %s file can't be used\n",
1048                                symbol_conf.vmlinux_name);
1049                         exit(1);
1050                 }
1051
1052                 return;
1053         }
1054
1055         /* let's see, whether we need to install initial sym_filter_entry */
1056         if (sym_filter_entry_sched) {
1057                 sym_filter_entry = sym_filter_entry_sched;
1058                 sym_filter_entry_sched = NULL;
1059                 if (parse_source(sym_filter_entry) < 0) {
1060                         struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1061
1062                         pr_err("Can't annotate %s", sym->name);
1063                         if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1064                                 pr_err(": No vmlinux file was found in the path:\n");
1065                                 machine__fprintf_vmlinux_path(machine, stderr);
1066                         } else
1067                                 pr_err(".\n");
1068                         exit(1);
1069                 }
1070         }
1071
1072         syme = symbol__priv(al.sym);
1073         if (!syme->skip) {
1074                 syme->count[counter]++;
1075                 syme->origin = origin;
1076                 record_precise_ip(syme, counter, ip);
1077                 pthread_mutex_lock(&active_symbols_lock);
1078                 if (list_empty(&syme->node) || !syme->node.next)
1079                         __list_insert_active_sym(syme);
1080                 pthread_mutex_unlock(&active_symbols_lock);
1081         }
1082 }
1083
1084 struct mmap_data {
1085         int                     counter;
1086         void                    *base;
1087         int                     mask;
1088         unsigned int            prev;
1089 };
1090
1091 static unsigned int mmap_read_head(struct mmap_data *md)
1092 {
1093         struct perf_event_mmap_page *pc = md->base;
1094         int head;
1095
1096         head = pc->data_head;
1097         rmb();
1098
1099         return head;
1100 }
1101
1102 static void perf_session__mmap_read_counter(struct perf_session *self,
1103                                             struct mmap_data *md)
1104 {
1105         unsigned int head = mmap_read_head(md);
1106         unsigned int old = md->prev;
1107         unsigned char *data = md->base + page_size;
1108         int diff;
1109
1110         /*
1111          * If we're further behind than half the buffer, there's a chance
1112          * the writer will bite our tail and mess up the samples under us.
1113          *
1114          * If we somehow ended up ahead of the head, we got messed up.
1115          *
1116          * In either case, truncate and restart at head.
1117          */
1118         diff = head - old;
1119         if (diff > md->mask / 2 || diff < 0) {
1120                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1121
1122                 /*
1123                  * head points to a known good entry, start there.
1124                  */
1125                 old = head;
1126         }
1127
1128         for (; old != head;) {
1129                 event_t *event = (event_t *)&data[old & md->mask];
1130
1131                 event_t event_copy;
1132
1133                 size_t size = event->header.size;
1134
1135                 /*
1136                  * Event straddles the mmap boundary -- header should always
1137                  * be inside due to u64 alignment of output.
1138                  */
1139                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1140                         unsigned int offset = old;
1141                         unsigned int len = min(sizeof(*event), size), cpy;
1142                         void *dst = &event_copy;
1143
1144                         do {
1145                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1146                                 memcpy(dst, &data[offset & md->mask], cpy);
1147                                 offset += cpy;
1148                                 dst += cpy;
1149                                 len -= cpy;
1150                         } while (len);
1151
1152                         event = &event_copy;
1153                 }
1154
1155                 if (event->header.type == PERF_RECORD_SAMPLE)
1156                         event__process_sample(event, self, md->counter);
1157                 else
1158                         event__process(event, self);
1159                 old += size;
1160         }
1161
1162         md->prev = old;
1163 }
1164
1165 static struct pollfd *event_array;
1166 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1167
1168 static void perf_session__mmap_read(struct perf_session *self)
1169 {
1170         int i, counter, thread_index;
1171
1172         for (i = 0; i < nr_cpus; i++) {
1173                 for (counter = 0; counter < nr_counters; counter++)
1174                         for (thread_index = 0;
1175                                 thread_index < thread_num;
1176                                 thread_index++) {
1177                                 perf_session__mmap_read_counter(self,
1178                                         &mmap_array[i][counter][thread_index]);
1179                         }
1180         }
1181 }
1182
1183 int nr_poll;
1184 int group_fd;
1185
1186 static void start_counter(int i, int counter)
1187 {
1188         struct perf_event_attr *attr;
1189         int cpu = -1;
1190         int thread_index;
1191
1192         if (target_tid == -1)
1193                 cpu = cpumap[i];
1194
1195         attr = attrs + counter;
1196
1197         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1198
1199         if (freq) {
1200                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1201                 attr->freq              = 1;
1202                 attr->sample_freq       = freq;
1203         }
1204
1205         attr->inherit           = (cpu < 0) && inherit;
1206         attr->mmap              = 1;
1207
1208         for (thread_index = 0; thread_index < thread_num; thread_index++) {
1209 try_again:
1210                 fd[i][counter][thread_index] = sys_perf_event_open(attr,
1211                                 all_tids[thread_index], cpu, group_fd, 0);
1212
1213                 if (fd[i][counter][thread_index] < 0) {
1214                         int err = errno;
1215
1216                         if (err == EPERM || err == EACCES)
1217                                 die("No permission - are you root?\n");
1218                         /*
1219                          * If it's cycles then fall back to hrtimer
1220                          * based cpu-clock-tick sw counter, which
1221                          * is always available even if no PMU support:
1222                          */
1223                         if (attr->type == PERF_TYPE_HARDWARE
1224                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1225
1226                                 if (verbose)
1227                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
1228
1229                                 attr->type = PERF_TYPE_SOFTWARE;
1230                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1231                                 goto try_again;
1232                         }
1233                         printf("\n");
1234                         error("perfcounter syscall returned with %d (%s)\n",
1235                                         fd[i][counter][thread_index], strerror(err));
1236                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1237                         exit(-1);
1238                 }
1239                 assert(fd[i][counter][thread_index] >= 0);
1240                 fcntl(fd[i][counter][thread_index], F_SETFL, O_NONBLOCK);
1241
1242                 /*
1243                  * First counter acts as the group leader:
1244                  */
1245                 if (group && group_fd == -1)
1246                         group_fd = fd[i][counter][thread_index];
1247
1248                 event_array[nr_poll].fd = fd[i][counter][thread_index];
1249                 event_array[nr_poll].events = POLLIN;
1250                 nr_poll++;
1251
1252                 mmap_array[i][counter][thread_index].counter = counter;
1253                 mmap_array[i][counter][thread_index].prev = 0;
1254                 mmap_array[i][counter][thread_index].mask = mmap_pages*page_size - 1;
1255                 mmap_array[i][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
1256                                 PROT_READ, MAP_SHARED, fd[i][counter][thread_index], 0);
1257                 if (mmap_array[i][counter][thread_index].base == MAP_FAILED)
1258                         die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1259         }
1260 }
1261
1262 static int __cmd_top(void)
1263 {
1264         pthread_t thread;
1265         int i, counter;
1266         int ret;
1267         /*
1268          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1269          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1270          */
1271         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false);
1272         if (session == NULL)
1273                 return -ENOMEM;
1274
1275         if (target_tid != -1)
1276                 event__synthesize_thread(target_tid, event__process, session);
1277         else
1278                 event__synthesize_threads(event__process, session);
1279
1280         for (i = 0; i < nr_cpus; i++) {
1281                 group_fd = -1;
1282                 for (counter = 0; counter < nr_counters; counter++)
1283                         start_counter(i, counter);
1284         }
1285
1286         /* Wait for a minimal set of events before starting the snapshot */
1287         poll(&event_array[0], nr_poll, 100);
1288
1289         perf_session__mmap_read(session);
1290
1291         if (pthread_create(&thread, NULL, display_thread, session)) {
1292                 printf("Could not create display thread.\n");
1293                 exit(-1);
1294         }
1295
1296         if (realtime_prio) {
1297                 struct sched_param param;
1298
1299                 param.sched_priority = realtime_prio;
1300                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1301                         printf("Could not set realtime priority.\n");
1302                         exit(-1);
1303                 }
1304         }
1305
1306         while (1) {
1307                 int hits = samples;
1308
1309                 perf_session__mmap_read(session);
1310
1311                 if (hits == samples)
1312                         ret = poll(event_array, nr_poll, 100);
1313         }
1314
1315         return 0;
1316 }
1317
1318 static const char * const top_usage[] = {
1319         "perf top [<options>]",
1320         NULL
1321 };
1322
1323 static const struct option options[] = {
1324         OPT_CALLBACK('e', "event", NULL, "event",
1325                      "event selector. use 'perf list' to list available events",
1326                      parse_events),
1327         OPT_INTEGER('c', "count", &default_interval,
1328                     "event period to sample"),
1329         OPT_INTEGER('p', "pid", &target_pid,
1330                     "profile events on existing process id"),
1331         OPT_INTEGER('t', "tid", &target_tid,
1332                     "profile events on existing thread id"),
1333         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1334                             "system-wide collection from all CPUs"),
1335         OPT_STRING('C', "cpu", &cpu_list, "cpu",
1336                     "list of cpus to monitor"),
1337         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1338                    "file", "vmlinux pathname"),
1339         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1340                     "hide kernel symbols"),
1341         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1342         OPT_INTEGER('r', "realtime", &realtime_prio,
1343                     "collect data with this RT SCHED_FIFO priority"),
1344         OPT_INTEGER('d', "delay", &delay_secs,
1345                     "number of seconds to delay between refreshes"),
1346         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1347                             "dump the symbol table used for profiling"),
1348         OPT_INTEGER('f', "count-filter", &count_filter,
1349                     "only display functions with more events than this"),
1350         OPT_BOOLEAN('g', "group", &group,
1351                             "put the counters into a counter group"),
1352         OPT_BOOLEAN('i', "inherit", &inherit,
1353                     "child tasks inherit counters"),
1354         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1355                     "symbol to annotate"),
1356         OPT_BOOLEAN('z', "zero", &zero,
1357                     "zero history across updates"),
1358         OPT_INTEGER('F', "freq", &freq,
1359                     "profile at this frequency"),
1360         OPT_INTEGER('E', "entries", &print_entries,
1361                     "display this many functions"),
1362         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1363                     "hide user symbols"),
1364         OPT_INCR('v', "verbose", &verbose,
1365                     "be more verbose (show counter open errors, etc)"),
1366         OPT_END()
1367 };
1368
1369 int cmd_top(int argc, const char **argv, const char *prefix __used)
1370 {
1371         int counter;
1372         int i,j;
1373
1374         page_size = sysconf(_SC_PAGE_SIZE);
1375
1376         argc = parse_options(argc, argv, options, top_usage, 0);
1377         if (argc)
1378                 usage_with_options(top_usage, options);
1379
1380         if (target_pid != -1) {
1381                 target_tid = target_pid;
1382                 thread_num = find_all_tid(target_pid, &all_tids);
1383                 if (thread_num <= 0) {
1384                         fprintf(stderr, "Can't find all threads of pid %d\n",
1385                                 target_pid);
1386                         usage_with_options(top_usage, options);
1387                 }
1388         } else {
1389                 all_tids=malloc(sizeof(pid_t));
1390                 if (!all_tids)
1391                         return -ENOMEM;
1392
1393                 all_tids[0] = target_tid;
1394                 thread_num = 1;
1395         }
1396
1397         for (i = 0; i < MAX_NR_CPUS; i++) {
1398                 for (j = 0; j < MAX_COUNTERS; j++) {
1399                         fd[i][j] = malloc(sizeof(int)*thread_num);
1400                         mmap_array[i][j] = zalloc(
1401                                 sizeof(struct mmap_data)*thread_num);
1402                         if (!fd[i][j] || !mmap_array[i][j])
1403                                 return -ENOMEM;
1404                 }
1405         }
1406         event_array = malloc(
1407                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
1408         if (!event_array)
1409                 return -ENOMEM;
1410
1411         /* CPU and PID are mutually exclusive */
1412         if (target_tid > 0 && cpu_list) {
1413                 printf("WARNING: PID switch overriding CPU\n");
1414                 sleep(1);
1415                 cpu_list = NULL;
1416         }
1417
1418         if (!nr_counters)
1419                 nr_counters = 1;
1420
1421         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1422                                  (nr_counters + 1) * sizeof(unsigned long));
1423
1424         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1425         if (symbol__init() < 0)
1426                 return -1;
1427
1428         if (delay_secs < 1)
1429                 delay_secs = 1;
1430
1431         /*
1432          * User specified count overrides default frequency.
1433          */
1434         if (default_interval)
1435                 freq = 0;
1436         else if (freq) {
1437                 default_interval = freq;
1438         } else {
1439                 fprintf(stderr, "frequency and count are zero, aborting\n");
1440                 exit(EXIT_FAILURE);
1441         }
1442
1443         /*
1444          * Fill in the ones not specifically initialized via -c:
1445          */
1446         for (counter = 0; counter < nr_counters; counter++) {
1447                 if (attrs[counter].sample_period)
1448                         continue;
1449
1450                 attrs[counter].sample_period = default_interval;
1451         }
1452
1453         if (target_tid != -1)
1454                 nr_cpus = 1;
1455         else
1456                 nr_cpus = read_cpu_map(cpu_list);
1457
1458         if (nr_cpus < 1)
1459                 usage_with_options(top_usage, options);
1460
1461         get_term_dimensions(&winsize);
1462         if (print_entries == 0) {
1463                 update_print_entries(&winsize);
1464                 signal(SIGWINCH, sig_winch_handler);
1465         }
1466
1467         return __cmd_top();
1468 }