perf session: Move the global threads list to perf_session
[linux-2.6.git] / tools / perf / builtin-top.c
1 /*
2  * builtin-top.c
3  *
4  * Builtin top command: Display a continuously updated profile of
5  * any workload, CPU or specific PID.
6  *
7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8  *
9  * Improvements and fixes by:
10  *
11  *   Arjan van de Ven <arjan@linux.intel.com>
12  *   Yanmin Zhang <yanmin.zhang@intel.com>
13  *   Wu Fengguang <fengguang.wu@intel.com>
14  *   Mike Galbraith <efault@gmx.de>
15  *   Paul Mackerras <paulus@samba.org>
16  *
17  * Released under the GPL v2. (and only v2, not any later version)
18  */
19 #include "builtin.h"
20
21 #include "perf.h"
22
23 #include "util/color.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26 #include "util/thread.h"
27 #include "util/util.h"
28 #include <linux/rbtree.h>
29 #include "util/parse-options.h"
30 #include "util/parse-events.h"
31
32 #include "util/debug.h"
33
34 #include <assert.h>
35 #include <fcntl.h>
36
37 #include <stdio.h>
38 #include <termios.h>
39 #include <unistd.h>
40
41 #include <errno.h>
42 #include <time.h>
43 #include <sched.h>
44 #include <pthread.h>
45
46 #include <sys/syscall.h>
47 #include <sys/ioctl.h>
48 #include <sys/poll.h>
49 #include <sys/prctl.h>
50 #include <sys/wait.h>
51 #include <sys/uio.h>
52 #include <sys/mman.h>
53
54 #include <linux/unistd.h>
55 #include <linux/types.h>
56
57 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
58
59 static int                      system_wide                     =      0;
60
61 static int                      default_interval                =      0;
62
63 static int                      count_filter                    =      5;
64 static int                      print_entries;
65
66 static int                      target_pid                      =     -1;
67 static int                      inherit                         =      0;
68 static int                      profile_cpu                     =     -1;
69 static int                      nr_cpus                         =      0;
70 static unsigned int             realtime_prio                   =      0;
71 static int                      group                           =      0;
72 static unsigned int             page_size;
73 static unsigned int             mmap_pages                      =     16;
74 static int                      freq                            =   1000; /* 1 KHz */
75
76 static int                      delay_secs                      =      2;
77 static int                      zero                            =      0;
78 static int                      dump_symtab                     =      0;
79
80 static bool                     hide_kernel_symbols             =  false;
81 static bool                     hide_user_symbols               =  false;
82 static struct winsize           winsize;
83 struct symbol_conf              symbol_conf;
84
85 /*
86  * Source
87  */
88
89 struct source_line {
90         u64                     eip;
91         unsigned long           count[MAX_COUNTERS];
92         char                    *line;
93         struct source_line      *next;
94 };
95
96 static char                     *sym_filter                     =   NULL;
97 struct sym_entry                *sym_filter_entry               =   NULL;
98 static int                      sym_pcnt_filter                 =      5;
99 static int                      sym_counter                     =      0;
100 static int                      display_weighted                =     -1;
101
102 /*
103  * Symbols
104  */
105
106 struct sym_entry_source {
107         struct source_line      *source;
108         struct source_line      *lines;
109         struct source_line      **lines_tail;
110         pthread_mutex_t         lock;
111 };
112
113 struct sym_entry {
114         struct rb_node          rb_node;
115         struct list_head        node;
116         unsigned long           snap_count;
117         double                  weight;
118         int                     skip;
119         u16                     name_len;
120         u8                      origin;
121         struct map              *map;
122         struct sym_entry_source *src;
123         unsigned long           count[0];
124 };
125
126 /*
127  * Source functions
128  */
129
130 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
131 {
132        return ((void *)self) + symbol_conf.priv_size;
133 }
134
135 static void get_term_dimensions(struct winsize *ws)
136 {
137         char *s = getenv("LINES");
138
139         if (s != NULL) {
140                 ws->ws_row = atoi(s);
141                 s = getenv("COLUMNS");
142                 if (s != NULL) {
143                         ws->ws_col = atoi(s);
144                         if (ws->ws_row && ws->ws_col)
145                                 return;
146                 }
147         }
148 #ifdef TIOCGWINSZ
149         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
150             ws->ws_row && ws->ws_col)
151                 return;
152 #endif
153         ws->ws_row = 25;
154         ws->ws_col = 80;
155 }
156
157 static void update_print_entries(struct winsize *ws)
158 {
159         print_entries = ws->ws_row;
160
161         if (print_entries > 9)
162                 print_entries -= 9;
163 }
164
165 static void sig_winch_handler(int sig __used)
166 {
167         get_term_dimensions(&winsize);
168         update_print_entries(&winsize);
169 }
170
171 static void parse_source(struct sym_entry *syme)
172 {
173         struct symbol *sym;
174         struct sym_entry_source *source;
175         struct map *map;
176         FILE *file;
177         char command[PATH_MAX*2];
178         const char *path;
179         u64 len;
180
181         if (!syme)
182                 return;
183
184         if (syme->src == NULL) {
185                 syme->src = zalloc(sizeof(*source));
186                 if (syme->src == NULL)
187                         return;
188                 pthread_mutex_init(&syme->src->lock, NULL);
189         }
190
191         source = syme->src;
192
193         if (source->lines) {
194                 pthread_mutex_lock(&source->lock);
195                 goto out_assign;
196         }
197
198         sym = sym_entry__symbol(syme);
199         map = syme->map;
200         path = map->dso->long_name;
201
202         len = sym->end - sym->start;
203
204         sprintf(command,
205                 "objdump --start-address=0x%016Lx "
206                          "--stop-address=0x%016Lx -dS %s",
207                 map->unmap_ip(map, sym->start),
208                 map->unmap_ip(map, sym->end), path);
209
210         file = popen(command, "r");
211         if (!file)
212                 return;
213
214         pthread_mutex_lock(&source->lock);
215         source->lines_tail = &source->lines;
216         while (!feof(file)) {
217                 struct source_line *src;
218                 size_t dummy = 0;
219                 char *c;
220
221                 src = malloc(sizeof(struct source_line));
222                 assert(src != NULL);
223                 memset(src, 0, sizeof(struct source_line));
224
225                 if (getline(&src->line, &dummy, file) < 0)
226                         break;
227                 if (!src->line)
228                         break;
229
230                 c = strchr(src->line, '\n');
231                 if (c)
232                         *c = 0;
233
234                 src->next = NULL;
235                 *source->lines_tail = src;
236                 source->lines_tail = &src->next;
237
238                 if (strlen(src->line)>8 && src->line[8] == ':') {
239                         src->eip = strtoull(src->line, NULL, 16);
240                         src->eip = map->unmap_ip(map, src->eip);
241                 }
242                 if (strlen(src->line)>8 && src->line[16] == ':') {
243                         src->eip = strtoull(src->line, NULL, 16);
244                         src->eip = map->unmap_ip(map, src->eip);
245                 }
246         }
247         pclose(file);
248 out_assign:
249         sym_filter_entry = syme;
250         pthread_mutex_unlock(&source->lock);
251 }
252
253 static void __zero_source_counters(struct sym_entry *syme)
254 {
255         int i;
256         struct source_line *line;
257
258         line = syme->src->lines;
259         while (line) {
260                 for (i = 0; i < nr_counters; i++)
261                         line->count[i] = 0;
262                 line = line->next;
263         }
264 }
265
266 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
267 {
268         struct source_line *line;
269
270         if (syme != sym_filter_entry)
271                 return;
272
273         if (pthread_mutex_trylock(&syme->src->lock))
274                 return;
275
276         if (syme->src == NULL || syme->src->source == NULL)
277                 goto out_unlock;
278
279         for (line = syme->src->lines; line; line = line->next) {
280                 if (line->eip == ip) {
281                         line->count[counter]++;
282                         break;
283                 }
284                 if (line->eip > ip)
285                         break;
286         }
287 out_unlock:
288         pthread_mutex_unlock(&syme->src->lock);
289 }
290
291 static void lookup_sym_source(struct sym_entry *syme)
292 {
293         struct symbol *symbol = sym_entry__symbol(syme);
294         struct source_line *line;
295         char pattern[PATH_MAX];
296
297         sprintf(pattern, "<%s>:", symbol->name);
298
299         pthread_mutex_lock(&syme->src->lock);
300         for (line = syme->src->lines; line; line = line->next) {
301                 if (strstr(line->line, pattern)) {
302                         syme->src->source = line;
303                         break;
304                 }
305         }
306         pthread_mutex_unlock(&syme->src->lock);
307 }
308
309 static void show_lines(struct source_line *queue, int count, int total)
310 {
311         int i;
312         struct source_line *line;
313
314         line = queue;
315         for (i = 0; i < count; i++) {
316                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
317
318                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
319                 line = line->next;
320         }
321 }
322
323 #define TRACE_COUNT     3
324
325 static void show_details(struct sym_entry *syme)
326 {
327         struct symbol *symbol;
328         struct source_line *line;
329         struct source_line *line_queue = NULL;
330         int displayed = 0;
331         int line_queue_count = 0, total = 0, more = 0;
332
333         if (!syme)
334                 return;
335
336         if (!syme->src->source)
337                 lookup_sym_source(syme);
338
339         if (!syme->src->source)
340                 return;
341
342         symbol = sym_entry__symbol(syme);
343         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
344         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
345
346         pthread_mutex_lock(&syme->src->lock);
347         line = syme->src->source;
348         while (line) {
349                 total += line->count[sym_counter];
350                 line = line->next;
351         }
352
353         line = syme->src->source;
354         while (line) {
355                 float pcnt = 0.0;
356
357                 if (!line_queue_count)
358                         line_queue = line;
359                 line_queue_count++;
360
361                 if (line->count[sym_counter])
362                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
363                 if (pcnt >= (float)sym_pcnt_filter) {
364                         if (displayed <= print_entries)
365                                 show_lines(line_queue, line_queue_count, total);
366                         else more++;
367                         displayed += line_queue_count;
368                         line_queue_count = 0;
369                         line_queue = NULL;
370                 } else if (line_queue_count > TRACE_COUNT) {
371                         line_queue = line_queue->next;
372                         line_queue_count--;
373                 }
374
375                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
376                 line = line->next;
377         }
378         pthread_mutex_unlock(&syme->src->lock);
379         if (more)
380                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
381 }
382
383 /*
384  * Symbols will be added here in event__process_sample and will get out
385  * after decayed.
386  */
387 static LIST_HEAD(active_symbols);
388 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
389
390 /*
391  * Ordering weight: count-1 * count-2 * ... / count-n
392  */
393 static double sym_weight(const struct sym_entry *sym)
394 {
395         double weight = sym->snap_count;
396         int counter;
397
398         if (!display_weighted)
399                 return weight;
400
401         for (counter = 1; counter < nr_counters-1; counter++)
402                 weight *= sym->count[counter];
403
404         weight /= (sym->count[counter] + 1);
405
406         return weight;
407 }
408
409 static long                     samples;
410 static long                     userspace_samples;
411 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
412
413 static void __list_insert_active_sym(struct sym_entry *syme)
414 {
415         list_add(&syme->node, &active_symbols);
416 }
417
418 static void list_remove_active_sym(struct sym_entry *syme)
419 {
420         pthread_mutex_lock(&active_symbols_lock);
421         list_del_init(&syme->node);
422         pthread_mutex_unlock(&active_symbols_lock);
423 }
424
425 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
426 {
427         struct rb_node **p = &tree->rb_node;
428         struct rb_node *parent = NULL;
429         struct sym_entry *iter;
430
431         while (*p != NULL) {
432                 parent = *p;
433                 iter = rb_entry(parent, struct sym_entry, rb_node);
434
435                 if (se->weight > iter->weight)
436                         p = &(*p)->rb_left;
437                 else
438                         p = &(*p)->rb_right;
439         }
440
441         rb_link_node(&se->rb_node, parent, p);
442         rb_insert_color(&se->rb_node, tree);
443 }
444
445 static void print_sym_table(void)
446 {
447         int printed = 0, j;
448         int counter, snap = !display_weighted ? sym_counter : 0;
449         float samples_per_sec = samples/delay_secs;
450         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
451         float sum_ksamples = 0.0;
452         struct sym_entry *syme, *n;
453         struct rb_root tmp = RB_ROOT;
454         struct rb_node *nd;
455         int sym_width = 0, dso_width = 0, max_dso_width;
456         const int win_width = winsize.ws_col - 1;
457
458         samples = userspace_samples = 0;
459
460         /* Sort the active symbols */
461         pthread_mutex_lock(&active_symbols_lock);
462         syme = list_entry(active_symbols.next, struct sym_entry, node);
463         pthread_mutex_unlock(&active_symbols_lock);
464
465         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
466                 syme->snap_count = syme->count[snap];
467                 if (syme->snap_count != 0) {
468
469                         if ((hide_user_symbols &&
470                              syme->origin == PERF_RECORD_MISC_USER) ||
471                             (hide_kernel_symbols &&
472                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
473                                 list_remove_active_sym(syme);
474                                 continue;
475                         }
476                         syme->weight = sym_weight(syme);
477                         rb_insert_active_sym(&tmp, syme);
478                         sum_ksamples += syme->snap_count;
479
480                         for (j = 0; j < nr_counters; j++)
481                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
482                 } else
483                         list_remove_active_sym(syme);
484         }
485
486         puts(CONSOLE_CLEAR);
487
488         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
489         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
490                 samples_per_sec,
491                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
492
493         if (nr_counters == 1 || !display_weighted) {
494                 printf("%Ld", (u64)attrs[0].sample_period);
495                 if (freq)
496                         printf("Hz ");
497                 else
498                         printf(" ");
499         }
500
501         if (!display_weighted)
502                 printf("%s", event_name(sym_counter));
503         else for (counter = 0; counter < nr_counters; counter++) {
504                 if (counter)
505                         printf("/");
506
507                 printf("%s", event_name(counter));
508         }
509
510         printf( "], ");
511
512         if (target_pid != -1)
513                 printf(" (target_pid: %d", target_pid);
514         else
515                 printf(" (all");
516
517         if (profile_cpu != -1)
518                 printf(", cpu: %d)\n", profile_cpu);
519         else {
520                 if (target_pid != -1)
521                         printf(")\n");
522                 else
523                         printf(", %d CPUs)\n", nr_cpus);
524         }
525
526         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
527
528         if (sym_filter_entry) {
529                 show_details(sym_filter_entry);
530                 return;
531         }
532
533         /*
534          * Find the longest symbol name that will be displayed
535          */
536         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
537                 syme = rb_entry(nd, struct sym_entry, rb_node);
538                 if (++printed > print_entries ||
539                     (int)syme->snap_count < count_filter)
540                         continue;
541
542                 if (syme->map->dso->long_name_len > dso_width)
543                         dso_width = syme->map->dso->long_name_len;
544
545                 if (syme->name_len > sym_width)
546                         sym_width = syme->name_len;
547         }
548
549         printed = 0;
550
551         max_dso_width = winsize.ws_col - sym_width - 29;
552         if (dso_width > max_dso_width)
553                 dso_width = max_dso_width;
554         putchar('\n');
555         if (nr_counters == 1)
556                 printf("             samples  pcnt");
557         else
558                 printf("   weight    samples  pcnt");
559
560         if (verbose)
561                 printf("         RIP       ");
562         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
563         printf("   %s    _______ _____",
564                nr_counters == 1 ? "      " : "______");
565         if (verbose)
566                 printf(" ________________");
567         printf(" %-*.*s", sym_width, sym_width, graph_line);
568         printf(" %-*.*s", dso_width, dso_width, graph_line);
569         puts("\n");
570
571         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
572                 struct symbol *sym;
573                 double pcnt;
574
575                 syme = rb_entry(nd, struct sym_entry, rb_node);
576                 sym = sym_entry__symbol(syme);
577
578                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
579                         continue;
580
581                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
582                                          sum_ksamples));
583
584                 if (nr_counters == 1 || !display_weighted)
585                         printf("%20.2f ", syme->weight);
586                 else
587                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
588
589                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
590                 if (verbose)
591                         printf(" %016llx", sym->start);
592                 printf(" %-*.*s", sym_width, sym_width, sym->name);
593                 printf(" %-*.*s\n", dso_width, dso_width,
594                        dso_width >= syme->map->dso->long_name_len ?
595                                         syme->map->dso->long_name :
596                                         syme->map->dso->short_name);
597         }
598 }
599
600 static void prompt_integer(int *target, const char *msg)
601 {
602         char *buf = malloc(0), *p;
603         size_t dummy = 0;
604         int tmp;
605
606         fprintf(stdout, "\n%s: ", msg);
607         if (getline(&buf, &dummy, stdin) < 0)
608                 return;
609
610         p = strchr(buf, '\n');
611         if (p)
612                 *p = 0;
613
614         p = buf;
615         while(*p) {
616                 if (!isdigit(*p))
617                         goto out_free;
618                 p++;
619         }
620         tmp = strtoul(buf, NULL, 10);
621         *target = tmp;
622 out_free:
623         free(buf);
624 }
625
626 static void prompt_percent(int *target, const char *msg)
627 {
628         int tmp = 0;
629
630         prompt_integer(&tmp, msg);
631         if (tmp >= 0 && tmp <= 100)
632                 *target = tmp;
633 }
634
635 static void prompt_symbol(struct sym_entry **target, const char *msg)
636 {
637         char *buf = malloc(0), *p;
638         struct sym_entry *syme = *target, *n, *found = NULL;
639         size_t dummy = 0;
640
641         /* zero counters of active symbol */
642         if (syme) {
643                 pthread_mutex_lock(&syme->src->lock);
644                 __zero_source_counters(syme);
645                 *target = NULL;
646                 pthread_mutex_unlock(&syme->src->lock);
647         }
648
649         fprintf(stdout, "\n%s: ", msg);
650         if (getline(&buf, &dummy, stdin) < 0)
651                 goto out_free;
652
653         p = strchr(buf, '\n');
654         if (p)
655                 *p = 0;
656
657         pthread_mutex_lock(&active_symbols_lock);
658         syme = list_entry(active_symbols.next, struct sym_entry, node);
659         pthread_mutex_unlock(&active_symbols_lock);
660
661         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
662                 struct symbol *sym = sym_entry__symbol(syme);
663
664                 if (!strcmp(buf, sym->name)) {
665                         found = syme;
666                         break;
667                 }
668         }
669
670         if (!found) {
671                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
672                 sleep(1);
673                 return;
674         } else
675                 parse_source(found);
676
677 out_free:
678         free(buf);
679 }
680
681 static void print_mapped_keys(void)
682 {
683         char *name = NULL;
684
685         if (sym_filter_entry) {
686                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
687                 name = sym->name;
688         }
689
690         fprintf(stdout, "\nMapped keys:\n");
691         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
692         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
693
694         if (nr_counters > 1)
695                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
696
697         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
698
699         if (symbol_conf.vmlinux_name) {
700                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
701                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
702                 fprintf(stdout, "\t[S]     stop annotation.\n");
703         }
704
705         if (nr_counters > 1)
706                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
707
708         fprintf(stdout,
709                 "\t[K]     hide kernel_symbols symbols.             \t(%s)\n",
710                 hide_kernel_symbols ? "yes" : "no");
711         fprintf(stdout,
712                 "\t[U]     hide user symbols.               \t(%s)\n",
713                 hide_user_symbols ? "yes" : "no");
714         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
715         fprintf(stdout, "\t[qQ]    quit.\n");
716 }
717
718 static int key_mapped(int c)
719 {
720         switch (c) {
721                 case 'd':
722                 case 'e':
723                 case 'f':
724                 case 'z':
725                 case 'q':
726                 case 'Q':
727                 case 'K':
728                 case 'U':
729                         return 1;
730                 case 'E':
731                 case 'w':
732                         return nr_counters > 1 ? 1 : 0;
733                 case 'F':
734                 case 's':
735                 case 'S':
736                         return symbol_conf.vmlinux_name ? 1 : 0;
737                 default:
738                         break;
739         }
740
741         return 0;
742 }
743
744 static void handle_keypress(int c)
745 {
746         if (!key_mapped(c)) {
747                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
748                 struct termios tc, save;
749
750                 print_mapped_keys();
751                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
752                 fflush(stdout);
753
754                 tcgetattr(0, &save);
755                 tc = save;
756                 tc.c_lflag &= ~(ICANON | ECHO);
757                 tc.c_cc[VMIN] = 0;
758                 tc.c_cc[VTIME] = 0;
759                 tcsetattr(0, TCSANOW, &tc);
760
761                 poll(&stdin_poll, 1, -1);
762                 c = getc(stdin);
763
764                 tcsetattr(0, TCSAFLUSH, &save);
765                 if (!key_mapped(c))
766                         return;
767         }
768
769         switch (c) {
770                 case 'd':
771                         prompt_integer(&delay_secs, "Enter display delay");
772                         if (delay_secs < 1)
773                                 delay_secs = 1;
774                         break;
775                 case 'e':
776                         prompt_integer(&print_entries, "Enter display entries (lines)");
777                         if (print_entries == 0) {
778                                 sig_winch_handler(SIGWINCH);
779                                 signal(SIGWINCH, sig_winch_handler);
780                         } else
781                                 signal(SIGWINCH, SIG_DFL);
782                         break;
783                 case 'E':
784                         if (nr_counters > 1) {
785                                 int i;
786
787                                 fprintf(stderr, "\nAvailable events:");
788                                 for (i = 0; i < nr_counters; i++)
789                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
790
791                                 prompt_integer(&sym_counter, "Enter details event counter");
792
793                                 if (sym_counter >= nr_counters) {
794                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
795                                         sym_counter = 0;
796                                         sleep(1);
797                                 }
798                         } else sym_counter = 0;
799                         break;
800                 case 'f':
801                         prompt_integer(&count_filter, "Enter display event count filter");
802                         break;
803                 case 'F':
804                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
805                         break;
806                 case 'K':
807                         hide_kernel_symbols = !hide_kernel_symbols;
808                         break;
809                 case 'q':
810                 case 'Q':
811                         printf("exiting.\n");
812                         if (dump_symtab)
813                                 dsos__fprintf(stderr);
814                         exit(0);
815                 case 's':
816                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
817                         break;
818                 case 'S':
819                         if (!sym_filter_entry)
820                                 break;
821                         else {
822                                 struct sym_entry *syme = sym_filter_entry;
823
824                                 pthread_mutex_lock(&syme->src->lock);
825                                 sym_filter_entry = NULL;
826                                 __zero_source_counters(syme);
827                                 pthread_mutex_unlock(&syme->src->lock);
828                         }
829                         break;
830                 case 'U':
831                         hide_user_symbols = !hide_user_symbols;
832                         break;
833                 case 'w':
834                         display_weighted = ~display_weighted;
835                         break;
836                 case 'z':
837                         zero = ~zero;
838                         break;
839                 default:
840                         break;
841         }
842 }
843
844 static void *display_thread(void *arg __used)
845 {
846         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
847         struct termios tc, save;
848         int delay_msecs, c;
849
850         tcgetattr(0, &save);
851         tc = save;
852         tc.c_lflag &= ~(ICANON | ECHO);
853         tc.c_cc[VMIN] = 0;
854         tc.c_cc[VTIME] = 0;
855
856 repeat:
857         delay_msecs = delay_secs * 1000;
858         tcsetattr(0, TCSANOW, &tc);
859         /* trash return*/
860         getc(stdin);
861
862         do {
863                 print_sym_table();
864         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
865
866         c = getc(stdin);
867         tcsetattr(0, TCSAFLUSH, &save);
868
869         handle_keypress(c);
870         goto repeat;
871
872         return NULL;
873 }
874
875 /* Tag samples to be skipped. */
876 static const char *skip_symbols[] = {
877         "default_idle",
878         "cpu_idle",
879         "enter_idle",
880         "exit_idle",
881         "mwait_idle",
882         "mwait_idle_with_hints",
883         "poll_idle",
884         "ppc64_runlatch_off",
885         "pseries_dedicated_idle_sleep",
886         NULL
887 };
888
889 static int symbol_filter(struct map *map, struct symbol *sym)
890 {
891         struct sym_entry *syme;
892         const char *name = sym->name;
893         int i;
894
895         /*
896          * ppc64 uses function descriptors and appends a '.' to the
897          * start of every instruction address. Remove it.
898          */
899         if (name[0] == '.')
900                 name++;
901
902         if (!strcmp(name, "_text") ||
903             !strcmp(name, "_etext") ||
904             !strcmp(name, "_sinittext") ||
905             !strncmp("init_module", name, 11) ||
906             !strncmp("cleanup_module", name, 14) ||
907             strstr(name, "_text_start") ||
908             strstr(name, "_text_end"))
909                 return 1;
910
911         syme = symbol__priv(sym);
912         syme->map = map;
913         syme->src = NULL;
914         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
915                 sym_filter_entry = syme;
916
917         for (i = 0; skip_symbols[i]; i++) {
918                 if (!strcmp(skip_symbols[i], name)) {
919                         syme->skip = 1;
920                         break;
921                 }
922         }
923
924         if (!syme->skip)
925                 syme->name_len = strlen(sym->name);
926
927         return 0;
928 }
929
930 static void event__process_sample(const event_t *self,
931                                  struct perf_session *session, int counter)
932 {
933         u64 ip = self->ip.ip;
934         struct sym_entry *syme;
935         struct addr_location al;
936         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
937
938         switch (origin) {
939         case PERF_RECORD_MISC_USER:
940                 if (hide_user_symbols)
941                         return;
942                 break;
943         case PERF_RECORD_MISC_KERNEL:
944                 if (hide_kernel_symbols)
945                         return;
946                 break;
947         default:
948                 return;
949         }
950
951         if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
952             al.sym == NULL)
953                 return;
954
955         syme = symbol__priv(al.sym);
956         if (!syme->skip) {
957                 syme->count[counter]++;
958                 syme->origin = origin;
959                 record_precise_ip(syme, counter, ip);
960                 pthread_mutex_lock(&active_symbols_lock);
961                 if (list_empty(&syme->node) || !syme->node.next)
962                         __list_insert_active_sym(syme);
963                 pthread_mutex_unlock(&active_symbols_lock);
964                 if (origin == PERF_RECORD_MISC_USER)
965                         ++userspace_samples;
966                 ++samples;
967         }
968 }
969
970 static int event__process(event_t *event, struct perf_session *session)
971 {
972         switch (event->header.type) {
973         case PERF_RECORD_COMM:
974                 event__process_comm(event, session);
975                 break;
976         case PERF_RECORD_MMAP:
977                 event__process_mmap(event, session);
978                 break;
979         default:
980                 break;
981         }
982
983         return 0;
984 }
985
986 struct mmap_data {
987         int                     counter;
988         void                    *base;
989         int                     mask;
990         unsigned int            prev;
991 };
992
993 static unsigned int mmap_read_head(struct mmap_data *md)
994 {
995         struct perf_event_mmap_page *pc = md->base;
996         int head;
997
998         head = pc->data_head;
999         rmb();
1000
1001         return head;
1002 }
1003
1004 static void perf_session__mmap_read_counter(struct perf_session *self,
1005                                             struct mmap_data *md)
1006 {
1007         unsigned int head = mmap_read_head(md);
1008         unsigned int old = md->prev;
1009         unsigned char *data = md->base + page_size;
1010         int diff;
1011
1012         /*
1013          * If we're further behind than half the buffer, there's a chance
1014          * the writer will bite our tail and mess up the samples under us.
1015          *
1016          * If we somehow ended up ahead of the head, we got messed up.
1017          *
1018          * In either case, truncate and restart at head.
1019          */
1020         diff = head - old;
1021         if (diff > md->mask / 2 || diff < 0) {
1022                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1023
1024                 /*
1025                  * head points to a known good entry, start there.
1026                  */
1027                 old = head;
1028         }
1029
1030         for (; old != head;) {
1031                 event_t *event = (event_t *)&data[old & md->mask];
1032
1033                 event_t event_copy;
1034
1035                 size_t size = event->header.size;
1036
1037                 /*
1038                  * Event straddles the mmap boundary -- header should always
1039                  * be inside due to u64 alignment of output.
1040                  */
1041                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1042                         unsigned int offset = old;
1043                         unsigned int len = min(sizeof(*event), size), cpy;
1044                         void *dst = &event_copy;
1045
1046                         do {
1047                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1048                                 memcpy(dst, &data[offset & md->mask], cpy);
1049                                 offset += cpy;
1050                                 dst += cpy;
1051                                 len -= cpy;
1052                         } while (len);
1053
1054                         event = &event_copy;
1055                 }
1056
1057                 if (event->header.type == PERF_RECORD_SAMPLE)
1058                         event__process_sample(event, self, md->counter);
1059                 else
1060                         event__process(event, self);
1061                 old += size;
1062         }
1063
1064         md->prev = old;
1065 }
1066
1067 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1068 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1069
1070 static void perf_session__mmap_read(struct perf_session *self)
1071 {
1072         int i, counter;
1073
1074         for (i = 0; i < nr_cpus; i++) {
1075                 for (counter = 0; counter < nr_counters; counter++)
1076                         perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1077         }
1078 }
1079
1080 int nr_poll;
1081 int group_fd;
1082
1083 static void start_counter(int i, int counter)
1084 {
1085         struct perf_event_attr *attr;
1086         int cpu;
1087
1088         cpu = profile_cpu;
1089         if (target_pid == -1 && profile_cpu == -1)
1090                 cpu = i;
1091
1092         attr = attrs + counter;
1093
1094         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1095
1096         if (freq) {
1097                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1098                 attr->freq              = 1;
1099                 attr->sample_freq       = freq;
1100         }
1101
1102         attr->inherit           = (cpu < 0) && inherit;
1103         attr->mmap              = 1;
1104
1105 try_again:
1106         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1107
1108         if (fd[i][counter] < 0) {
1109                 int err = errno;
1110
1111                 if (err == EPERM || err == EACCES)
1112                         die("No permission - are you root?\n");
1113                 /*
1114                  * If it's cycles then fall back to hrtimer
1115                  * based cpu-clock-tick sw counter, which
1116                  * is always available even if no PMU support:
1117                  */
1118                 if (attr->type == PERF_TYPE_HARDWARE
1119                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1120
1121                         if (verbose)
1122                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1123
1124                         attr->type = PERF_TYPE_SOFTWARE;
1125                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1126                         goto try_again;
1127                 }
1128                 printf("\n");
1129                 error("perfcounter syscall returned with %d (%s)\n",
1130                         fd[i][counter], strerror(err));
1131                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1132                 exit(-1);
1133         }
1134         assert(fd[i][counter] >= 0);
1135         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1136
1137         /*
1138          * First counter acts as the group leader:
1139          */
1140         if (group && group_fd == -1)
1141                 group_fd = fd[i][counter];
1142
1143         event_array[nr_poll].fd = fd[i][counter];
1144         event_array[nr_poll].events = POLLIN;
1145         nr_poll++;
1146
1147         mmap_array[i][counter].counter = counter;
1148         mmap_array[i][counter].prev = 0;
1149         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1150         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1151                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1152         if (mmap_array[i][counter].base == MAP_FAILED)
1153                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1154 }
1155
1156 static int __cmd_top(void)
1157 {
1158         pthread_t thread;
1159         int i, counter;
1160         int ret;
1161         /*
1162          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1163          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1164          */
1165         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1166
1167         if (session == NULL)
1168                 return -ENOMEM;
1169
1170         if (target_pid != -1)
1171                 event__synthesize_thread(target_pid, event__process, session);
1172         else
1173                 event__synthesize_threads(event__process, session);
1174
1175         for (i = 0; i < nr_cpus; i++) {
1176                 group_fd = -1;
1177                 for (counter = 0; counter < nr_counters; counter++)
1178                         start_counter(i, counter);
1179         }
1180
1181         /* Wait for a minimal set of events before starting the snapshot */
1182         poll(event_array, nr_poll, 100);
1183
1184         perf_session__mmap_read(session);
1185
1186         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1187                 printf("Could not create display thread.\n");
1188                 exit(-1);
1189         }
1190
1191         if (realtime_prio) {
1192                 struct sched_param param;
1193
1194                 param.sched_priority = realtime_prio;
1195                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1196                         printf("Could not set realtime priority.\n");
1197                         exit(-1);
1198                 }
1199         }
1200
1201         while (1) {
1202                 int hits = samples;
1203
1204                 perf_session__mmap_read(session);
1205
1206                 if (hits == samples)
1207                         ret = poll(event_array, nr_poll, 100);
1208         }
1209
1210         return 0;
1211 }
1212
1213 static const char * const top_usage[] = {
1214         "perf top [<options>]",
1215         NULL
1216 };
1217
1218 static const struct option options[] = {
1219         OPT_CALLBACK('e', "event", NULL, "event",
1220                      "event selector. use 'perf list' to list available events",
1221                      parse_events),
1222         OPT_INTEGER('c', "count", &default_interval,
1223                     "event period to sample"),
1224         OPT_INTEGER('p', "pid", &target_pid,
1225                     "profile events on existing pid"),
1226         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1227                             "system-wide collection from all CPUs"),
1228         OPT_INTEGER('C', "CPU", &profile_cpu,
1229                     "CPU to profile on"),
1230         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1231                    "file", "vmlinux pathname"),
1232         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1233                     "hide kernel symbols"),
1234         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1235                     "number of mmap data pages"),
1236         OPT_INTEGER('r', "realtime", &realtime_prio,
1237                     "collect data with this RT SCHED_FIFO priority"),
1238         OPT_INTEGER('d', "delay", &delay_secs,
1239                     "number of seconds to delay between refreshes"),
1240         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1241                             "dump the symbol table used for profiling"),
1242         OPT_INTEGER('f', "count-filter", &count_filter,
1243                     "only display functions with more events than this"),
1244         OPT_BOOLEAN('g', "group", &group,
1245                             "put the counters into a counter group"),
1246         OPT_BOOLEAN('i', "inherit", &inherit,
1247                     "child tasks inherit counters"),
1248         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1249                     "symbol to annotate - requires -k option"),
1250         OPT_BOOLEAN('z', "zero", &zero,
1251                     "zero history across updates"),
1252         OPT_INTEGER('F', "freq", &freq,
1253                     "profile at this frequency"),
1254         OPT_INTEGER('E', "entries", &print_entries,
1255                     "display this many functions"),
1256         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1257                     "hide user symbols"),
1258         OPT_BOOLEAN('v', "verbose", &verbose,
1259                     "be more verbose (show counter open errors, etc)"),
1260         OPT_END()
1261 };
1262
1263 int cmd_top(int argc, const char **argv, const char *prefix __used)
1264 {
1265         int counter;
1266
1267         page_size = sysconf(_SC_PAGE_SIZE);
1268
1269         argc = parse_options(argc, argv, options, top_usage, 0);
1270         if (argc)
1271                 usage_with_options(top_usage, options);
1272
1273         /* CPU and PID are mutually exclusive */
1274         if (target_pid != -1 && profile_cpu != -1) {
1275                 printf("WARNING: PID switch overriding CPU\n");
1276                 sleep(1);
1277                 profile_cpu = -1;
1278         }
1279
1280         if (!nr_counters)
1281                 nr_counters = 1;
1282
1283         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1284                                  (nr_counters + 1) * sizeof(unsigned long));
1285         if (symbol_conf.vmlinux_name == NULL)
1286                 symbol_conf.try_vmlinux_path = true;
1287         if (symbol__init(&symbol_conf) < 0)
1288                 return -1;
1289
1290         if (delay_secs < 1)
1291                 delay_secs = 1;
1292
1293         parse_source(sym_filter_entry);
1294
1295         /*
1296          * User specified count overrides default frequency.
1297          */
1298         if (default_interval)
1299                 freq = 0;
1300         else if (freq) {
1301                 default_interval = freq;
1302         } else {
1303                 fprintf(stderr, "frequency and count are zero, aborting\n");
1304                 exit(EXIT_FAILURE);
1305         }
1306
1307         /*
1308          * Fill in the ones not specifically initialized via -c:
1309          */
1310         for (counter = 0; counter < nr_counters; counter++) {
1311                 if (attrs[counter].sample_period)
1312                         continue;
1313
1314                 attrs[counter].sample_period = default_interval;
1315         }
1316
1317         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1318         assert(nr_cpus <= MAX_NR_CPUS);
1319         assert(nr_cpus >= 0);
1320
1321         if (target_pid != -1 || profile_cpu != -1)
1322                 nr_cpus = 1;
1323
1324         get_term_dimensions(&winsize);
1325         if (print_entries == 0) {
1326                 update_print_entries(&winsize);
1327                 signal(SIGWINCH, sig_winch_handler);
1328         }
1329
1330         return __cmd_top();
1331 }