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