]> bbs.cooldavid.org Git - net-next-2.6.git/blob - tools/perf/builtin-top.c
perf symbols: Look for vmlinux in more places
[net-next-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;
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 static bool                     hide_kernel_symbols             =  false;
80 static bool                     hide_user_symbols               =  false;
81 static struct winsize           winsize;
82 const char                      *vmlinux_name;
83 static const char               *graph_line                     =
84         "_____________________________________________________________________"
85         "_____________________________________________________________________";
86 static const char               *graph_dotted_line                      =
87         "---------------------------------------------------------------------"
88         "---------------------------------------------------------------------"
89         "---------------------------------------------------------------------";
90
91 /*
92  * Source
93  */
94
95 struct source_line {
96         u64                     eip;
97         unsigned long           count[MAX_COUNTERS];
98         char                    *line;
99         struct source_line      *next;
100 };
101
102 static char                     *sym_filter                     =   NULL;
103 struct sym_entry                *sym_filter_entry               =   NULL;
104 static int                      sym_pcnt_filter                 =      5;
105 static int                      sym_counter                     =      0;
106 static int                      display_weighted                =     -1;
107
108 /*
109  * Symbols
110  */
111
112 struct sym_entry_source {
113         struct source_line      *source;
114         struct source_line      *lines;
115         struct source_line      **lines_tail;
116         pthread_mutex_t         lock;
117 };
118
119 struct sym_entry {
120         struct rb_node          rb_node;
121         struct list_head        node;
122         unsigned long           snap_count;
123         double                  weight;
124         int                     skip;
125         u16                     name_len;
126         u8                      origin;
127         struct map              *map;
128         struct sym_entry_source *src;
129         unsigned long           count[0];
130 };
131
132 /*
133  * Source functions
134  */
135
136 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
137 {
138        return ((void *)self) + symbol__priv_size;
139 }
140
141 static void get_term_dimensions(struct winsize *ws)
142 {
143         char *s = getenv("LINES");
144
145         if (s != NULL) {
146                 ws->ws_row = atoi(s);
147                 s = getenv("COLUMNS");
148                 if (s != NULL) {
149                         ws->ws_col = atoi(s);
150                         if (ws->ws_row && ws->ws_col)
151                                 return;
152                 }
153         }
154 #ifdef TIOCGWINSZ
155         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
156             ws->ws_row && ws->ws_col)
157                 return;
158 #endif
159         ws->ws_row = 25;
160         ws->ws_col = 80;
161 }
162
163 static void update_print_entries(struct winsize *ws)
164 {
165         print_entries = ws->ws_row;
166
167         if (print_entries > 9)
168                 print_entries -= 9;
169 }
170
171 static void sig_winch_handler(int sig __used)
172 {
173         get_term_dimensions(&winsize);
174         update_print_entries(&winsize);
175 }
176
177 static void parse_source(struct sym_entry *syme)
178 {
179         struct symbol *sym;
180         struct sym_entry_source *source;
181         struct map *map;
182         FILE *file;
183         char command[PATH_MAX*2];
184         const char *path;
185         u64 len;
186
187         if (!syme)
188                 return;
189
190         if (syme->src == NULL) {
191                 syme->src = calloc(1, sizeof(*source));
192                 if (syme->src == NULL)
193                         return;
194                 pthread_mutex_init(&syme->src->lock, NULL);
195         }
196
197         source = syme->src;
198
199         if (source->lines) {
200                 pthread_mutex_lock(&source->lock);
201                 goto out_assign;
202         }
203
204         sym = sym_entry__symbol(syme);
205         map = syme->map;
206         path = map->dso->long_name;
207
208         len = sym->end - sym->start;
209
210         sprintf(command,
211                 "objdump --start-address=0x%016Lx "
212                          "--stop-address=0x%016Lx -dS %s",
213                 map->unmap_ip(map, sym->start),
214                 map->unmap_ip(map, sym->end), path);
215
216         file = popen(command, "r");
217         if (!file)
218                 return;
219
220         pthread_mutex_lock(&source->lock);
221         source->lines_tail = &source->lines;
222         while (!feof(file)) {
223                 struct source_line *src;
224                 size_t dummy = 0;
225                 char *c;
226
227                 src = malloc(sizeof(struct source_line));
228                 assert(src != NULL);
229                 memset(src, 0, sizeof(struct source_line));
230
231                 if (getline(&src->line, &dummy, file) < 0)
232                         break;
233                 if (!src->line)
234                         break;
235
236                 c = strchr(src->line, '\n');
237                 if (c)
238                         *c = 0;
239
240                 src->next = NULL;
241                 *source->lines_tail = src;
242                 source->lines_tail = &src->next;
243
244                 if (strlen(src->line)>8 && src->line[8] == ':') {
245                         src->eip = strtoull(src->line, NULL, 16);
246                         src->eip = map->unmap_ip(map, src->eip);
247                 }
248                 if (strlen(src->line)>8 && src->line[16] == ':') {
249                         src->eip = strtoull(src->line, NULL, 16);
250                         src->eip = map->unmap_ip(map, src->eip);
251                 }
252         }
253         pclose(file);
254 out_assign:
255         sym_filter_entry = syme;
256         pthread_mutex_unlock(&source->lock);
257 }
258
259 static void __zero_source_counters(struct sym_entry *syme)
260 {
261         int i;
262         struct source_line *line;
263
264         line = syme->src->lines;
265         while (line) {
266                 for (i = 0; i < nr_counters; i++)
267                         line->count[i] = 0;
268                 line = line->next;
269         }
270 }
271
272 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
273 {
274         struct source_line *line;
275
276         if (syme != sym_filter_entry)
277                 return;
278
279         if (pthread_mutex_trylock(&syme->src->lock))
280                 return;
281
282         if (syme->src == NULL || syme->src->source == NULL)
283                 goto out_unlock;
284
285         for (line = syme->src->lines; line; line = line->next) {
286                 if (line->eip == ip) {
287                         line->count[counter]++;
288                         break;
289                 }
290                 if (line->eip > ip)
291                         break;
292         }
293 out_unlock:
294         pthread_mutex_unlock(&syme->src->lock);
295 }
296
297 static void lookup_sym_source(struct sym_entry *syme)
298 {
299         struct symbol *symbol = sym_entry__symbol(syme);
300         struct source_line *line;
301         char pattern[PATH_MAX];
302
303         sprintf(pattern, "<%s>:", symbol->name);
304
305         pthread_mutex_lock(&syme->src->lock);
306         for (line = syme->src->lines; line; line = line->next) {
307                 if (strstr(line->line, pattern)) {
308                         syme->src->source = line;
309                         break;
310                 }
311         }
312         pthread_mutex_unlock(&syme->src->lock);
313 }
314
315 static void show_lines(struct source_line *queue, int count, int total)
316 {
317         int i;
318         struct source_line *line;
319
320         line = queue;
321         for (i = 0; i < count; i++) {
322                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
323
324                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
325                 line = line->next;
326         }
327 }
328
329 #define TRACE_COUNT     3
330
331 static void show_details(struct sym_entry *syme)
332 {
333         struct symbol *symbol;
334         struct source_line *line;
335         struct source_line *line_queue = NULL;
336         int displayed = 0;
337         int line_queue_count = 0, total = 0, more = 0;
338
339         if (!syme)
340                 return;
341
342         if (!syme->src->source)
343                 lookup_sym_source(syme);
344
345         if (!syme->src->source)
346                 return;
347
348         symbol = sym_entry__symbol(syme);
349         printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
350         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
351
352         pthread_mutex_lock(&syme->src->lock);
353         line = syme->src->source;
354         while (line) {
355                 total += line->count[sym_counter];
356                 line = line->next;
357         }
358
359         line = syme->src->source;
360         while (line) {
361                 float pcnt = 0.0;
362
363                 if (!line_queue_count)
364                         line_queue = line;
365                 line_queue_count++;
366
367                 if (line->count[sym_counter])
368                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
369                 if (pcnt >= (float)sym_pcnt_filter) {
370                         if (displayed <= print_entries)
371                                 show_lines(line_queue, line_queue_count, total);
372                         else more++;
373                         displayed += line_queue_count;
374                         line_queue_count = 0;
375                         line_queue = NULL;
376                 } else if (line_queue_count > TRACE_COUNT) {
377                         line_queue = line_queue->next;
378                         line_queue_count--;
379                 }
380
381                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
382                 line = line->next;
383         }
384         pthread_mutex_unlock(&syme->src->lock);
385         if (more)
386                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
387 }
388
389 /*
390  * Symbols will be added here in event__process_sample and will get out
391  * after decayed.
392  */
393 static LIST_HEAD(active_symbols);
394 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
395
396 /*
397  * Ordering weight: count-1 * count-2 * ... / count-n
398  */
399 static double sym_weight(const struct sym_entry *sym)
400 {
401         double weight = sym->snap_count;
402         int counter;
403
404         if (!display_weighted)
405                 return weight;
406
407         for (counter = 1; counter < nr_counters-1; counter++)
408                 weight *= sym->count[counter];
409
410         weight /= (sym->count[counter] + 1);
411
412         return weight;
413 }
414
415 static long                     samples;
416 static long                     userspace_samples;
417 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
418
419 static void __list_insert_active_sym(struct sym_entry *syme)
420 {
421         list_add(&syme->node, &active_symbols);
422 }
423
424 static void list_remove_active_sym(struct sym_entry *syme)
425 {
426         pthread_mutex_lock(&active_symbols_lock);
427         list_del_init(&syme->node);
428         pthread_mutex_unlock(&active_symbols_lock);
429 }
430
431 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
432 {
433         struct rb_node **p = &tree->rb_node;
434         struct rb_node *parent = NULL;
435         struct sym_entry *iter;
436
437         while (*p != NULL) {
438                 parent = *p;
439                 iter = rb_entry(parent, struct sym_entry, rb_node);
440
441                 if (se->weight > iter->weight)
442                         p = &(*p)->rb_left;
443                 else
444                         p = &(*p)->rb_right;
445         }
446
447         rb_link_node(&se->rb_node, parent, p);
448         rb_insert_color(&se->rb_node, tree);
449 }
450
451 static void print_sym_table(void)
452 {
453         int printed = 0, j;
454         int counter, snap = !display_weighted ? sym_counter : 0;
455         float samples_per_sec = samples/delay_secs;
456         float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
457         float sum_ksamples = 0.0;
458         struct sym_entry *syme, *n;
459         struct rb_root tmp = RB_ROOT;
460         struct rb_node *nd;
461         int sym_width = 0, dso_width = 0;
462         const int win_width = winsize.ws_col - 1;
463         struct dso *unique_dso = NULL, *first_dso = NULL;
464
465         samples = userspace_samples = 0;
466
467         /* Sort the active symbols */
468         pthread_mutex_lock(&active_symbols_lock);
469         syme = list_entry(active_symbols.next, struct sym_entry, node);
470         pthread_mutex_unlock(&active_symbols_lock);
471
472         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
473                 syme->snap_count = syme->count[snap];
474                 if (syme->snap_count != 0) {
475
476                         if ((hide_user_symbols &&
477                              syme->origin == PERF_RECORD_MISC_USER) ||
478                             (hide_kernel_symbols &&
479                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
480                                 list_remove_active_sym(syme);
481                                 continue;
482                         }
483                         syme->weight = sym_weight(syme);
484                         rb_insert_active_sym(&tmp, syme);
485                         sum_ksamples += syme->snap_count;
486
487                         for (j = 0; j < nr_counters; j++)
488                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
489                 } else
490                         list_remove_active_sym(syme);
491         }
492
493         puts(CONSOLE_CLEAR);
494
495         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
496         printf( "   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% [",
497                 samples_per_sec,
498                 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
499
500         if (nr_counters == 1 || !display_weighted) {
501                 printf("%Ld", (u64)attrs[0].sample_period);
502                 if (freq)
503                         printf("Hz ");
504                 else
505                         printf(" ");
506         }
507
508         if (!display_weighted)
509                 printf("%s", event_name(sym_counter));
510         else for (counter = 0; counter < nr_counters; counter++) {
511                 if (counter)
512                         printf("/");
513
514                 printf("%s", event_name(counter));
515         }
516
517         printf( "], ");
518
519         if (target_pid != -1)
520                 printf(" (target_pid: %d", target_pid);
521         else
522                 printf(" (all");
523
524         if (profile_cpu != -1)
525                 printf(", cpu: %d)\n", profile_cpu);
526         else {
527                 if (target_pid != -1)
528                         printf(")\n");
529                 else
530                         printf(", %d CPUs)\n", nr_cpus);
531         }
532
533         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
534
535         if (sym_filter_entry) {
536                 show_details(sym_filter_entry);
537                 return;
538         }
539
540         /*
541          * Find the longest symbol name that will be displayed
542          */
543         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
544                 syme = rb_entry(nd, struct sym_entry, rb_node);
545                 if (++printed > print_entries ||
546                     (int)syme->snap_count < count_filter)
547                         continue;
548
549                 if (first_dso == NULL)
550                         unique_dso = first_dso = syme->map->dso;
551                 else if (syme->map->dso != first_dso)
552                         unique_dso = NULL;
553
554                 if (syme->map->dso->long_name_len > dso_width)
555                         dso_width = syme->map->dso->long_name_len;
556
557                 if (syme->name_len > sym_width)
558                         sym_width = syme->name_len;
559         }
560
561         printed = 0;
562
563         if (unique_dso)
564                 printf("DSO: %s\n", unique_dso->long_name);
565         else {
566                 int max_dso_width = winsize.ws_col - sym_width - 29;
567                 if (dso_width > max_dso_width)
568                         dso_width = max_dso_width;
569                 putchar('\n');
570         }
571         if (nr_counters == 1)
572                 printf("             samples  pcnt");
573         else
574                 printf("   weight    samples  pcnt");
575
576         if (verbose)
577                 printf("         RIP       ");
578         printf(" %-*.*s", sym_width, sym_width, "function");
579         if (!unique_dso)
580                 printf(" DSO");
581         putchar('\n');
582         printf("   %s    _______ _____",
583                nr_counters == 1 ? "      " : "______");
584         if (verbose)
585                 printf(" ________________");
586         printf(" %-*.*s", sym_width, sym_width, graph_line);
587         if (!unique_dso)
588                 printf(" %-*.*s", dso_width, dso_width, graph_line);
589         puts("\n");
590
591         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
592                 struct symbol *sym;
593                 double pcnt;
594
595                 syme = rb_entry(nd, struct sym_entry, rb_node);
596                 sym = sym_entry__symbol(syme);
597
598                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
599                         continue;
600
601                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
602                                          sum_ksamples));
603
604                 if (nr_counters == 1 || !display_weighted)
605                         printf("%20.2f ", syme->weight);
606                 else
607                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
608
609                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
610                 if (verbose)
611                         printf(" %016llx", sym->start);
612                 printf(" %-*.*s", sym_width, sym_width, sym->name);
613                 if (!unique_dso)
614                         printf(" %-*.*s", dso_width, dso_width,
615                                dso_width >= syme->map->dso->long_name_len ?
616                                                 syme->map->dso->long_name :
617                                                 syme->map->dso->short_name);
618                 printf("\n");
619         }
620 }
621
622 static void prompt_integer(int *target, const char *msg)
623 {
624         char *buf = malloc(0), *p;
625         size_t dummy = 0;
626         int tmp;
627
628         fprintf(stdout, "\n%s: ", msg);
629         if (getline(&buf, &dummy, stdin) < 0)
630                 return;
631
632         p = strchr(buf, '\n');
633         if (p)
634                 *p = 0;
635
636         p = buf;
637         while(*p) {
638                 if (!isdigit(*p))
639                         goto out_free;
640                 p++;
641         }
642         tmp = strtoul(buf, NULL, 10);
643         *target = tmp;
644 out_free:
645         free(buf);
646 }
647
648 static void prompt_percent(int *target, const char *msg)
649 {
650         int tmp = 0;
651
652         prompt_integer(&tmp, msg);
653         if (tmp >= 0 && tmp <= 100)
654                 *target = tmp;
655 }
656
657 static void prompt_symbol(struct sym_entry **target, const char *msg)
658 {
659         char *buf = malloc(0), *p;
660         struct sym_entry *syme = *target, *n, *found = NULL;
661         size_t dummy = 0;
662
663         /* zero counters of active symbol */
664         if (syme) {
665                 pthread_mutex_lock(&syme->src->lock);
666                 __zero_source_counters(syme);
667                 *target = NULL;
668                 pthread_mutex_unlock(&syme->src->lock);
669         }
670
671         fprintf(stdout, "\n%s: ", msg);
672         if (getline(&buf, &dummy, stdin) < 0)
673                 goto out_free;
674
675         p = strchr(buf, '\n');
676         if (p)
677                 *p = 0;
678
679         pthread_mutex_lock(&active_symbols_lock);
680         syme = list_entry(active_symbols.next, struct sym_entry, node);
681         pthread_mutex_unlock(&active_symbols_lock);
682
683         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
684                 struct symbol *sym = sym_entry__symbol(syme);
685
686                 if (!strcmp(buf, sym->name)) {
687                         found = syme;
688                         break;
689                 }
690         }
691
692         if (!found) {
693                 fprintf(stderr, "Sorry, %s is not active.\n", sym_filter);
694                 sleep(1);
695                 return;
696         } else
697                 parse_source(found);
698
699 out_free:
700         free(buf);
701 }
702
703 static void print_mapped_keys(void)
704 {
705         char *name = NULL;
706
707         if (sym_filter_entry) {
708                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
709                 name = sym->name;
710         }
711
712         fprintf(stdout, "\nMapped keys:\n");
713         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
714         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
715
716         if (nr_counters > 1)
717                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_counter));
718
719         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
720
721         if (vmlinux_name) {
722                 fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
723                 fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
724                 fprintf(stdout, "\t[S]     stop annotation.\n");
725         }
726
727         if (nr_counters > 1)
728                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
729
730         fprintf(stdout,
731                 "\t[K]     hide kernel_symbols symbols.             \t(%s)\n",
732                 hide_kernel_symbols ? "yes" : "no");
733         fprintf(stdout,
734                 "\t[U]     hide user symbols.               \t(%s)\n",
735                 hide_user_symbols ? "yes" : "no");
736         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
737         fprintf(stdout, "\t[qQ]    quit.\n");
738 }
739
740 static int key_mapped(int c)
741 {
742         switch (c) {
743                 case 'd':
744                 case 'e':
745                 case 'f':
746                 case 'z':
747                 case 'q':
748                 case 'Q':
749                 case 'K':
750                 case 'U':
751                         return 1;
752                 case 'E':
753                 case 'w':
754                         return nr_counters > 1 ? 1 : 0;
755                 case 'F':
756                 case 's':
757                 case 'S':
758                         return vmlinux_name ? 1 : 0;
759                 default:
760                         break;
761         }
762
763         return 0;
764 }
765
766 static void handle_keypress(int c)
767 {
768         if (!key_mapped(c)) {
769                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
770                 struct termios tc, save;
771
772                 print_mapped_keys();
773                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
774                 fflush(stdout);
775
776                 tcgetattr(0, &save);
777                 tc = save;
778                 tc.c_lflag &= ~(ICANON | ECHO);
779                 tc.c_cc[VMIN] = 0;
780                 tc.c_cc[VTIME] = 0;
781                 tcsetattr(0, TCSANOW, &tc);
782
783                 poll(&stdin_poll, 1, -1);
784                 c = getc(stdin);
785
786                 tcsetattr(0, TCSAFLUSH, &save);
787                 if (!key_mapped(c))
788                         return;
789         }
790
791         switch (c) {
792                 case 'd':
793                         prompt_integer(&delay_secs, "Enter display delay");
794                         if (delay_secs < 1)
795                                 delay_secs = 1;
796                         break;
797                 case 'e':
798                         prompt_integer(&print_entries, "Enter display entries (lines)");
799                         if (print_entries == 0) {
800                                 sig_winch_handler(SIGWINCH);
801                                 signal(SIGWINCH, sig_winch_handler);
802                         } else
803                                 signal(SIGWINCH, SIG_DFL);
804                         break;
805                 case 'E':
806                         if (nr_counters > 1) {
807                                 int i;
808
809                                 fprintf(stderr, "\nAvailable events:");
810                                 for (i = 0; i < nr_counters; i++)
811                                         fprintf(stderr, "\n\t%d %s", i, event_name(i));
812
813                                 prompt_integer(&sym_counter, "Enter details event counter");
814
815                                 if (sym_counter >= nr_counters) {
816                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
817                                         sym_counter = 0;
818                                         sleep(1);
819                                 }
820                         } else sym_counter = 0;
821                         break;
822                 case 'f':
823                         prompt_integer(&count_filter, "Enter display event count filter");
824                         break;
825                 case 'F':
826                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
827                         break;
828                 case 'K':
829                         hide_kernel_symbols = !hide_kernel_symbols;
830                         break;
831                 case 'q':
832                 case 'Q':
833                         printf("exiting.\n");
834                         if (dump_symtab)
835                                 dsos__fprintf(stderr);
836                         exit(0);
837                 case 's':
838                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
839                         break;
840                 case 'S':
841                         if (!sym_filter_entry)
842                                 break;
843                         else {
844                                 struct sym_entry *syme = sym_filter_entry;
845
846                                 pthread_mutex_lock(&syme->src->lock);
847                                 sym_filter_entry = NULL;
848                                 __zero_source_counters(syme);
849                                 pthread_mutex_unlock(&syme->src->lock);
850                         }
851                         break;
852                 case 'U':
853                         hide_user_symbols = !hide_user_symbols;
854                         break;
855                 case 'w':
856                         display_weighted = ~display_weighted;
857                         break;
858                 case 'z':
859                         zero = ~zero;
860                         break;
861                 default:
862                         break;
863         }
864 }
865
866 static void *display_thread(void *arg __used)
867 {
868         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
869         struct termios tc, save;
870         int delay_msecs, c;
871
872         tcgetattr(0, &save);
873         tc = save;
874         tc.c_lflag &= ~(ICANON | ECHO);
875         tc.c_cc[VMIN] = 0;
876         tc.c_cc[VTIME] = 0;
877
878 repeat:
879         delay_msecs = delay_secs * 1000;
880         tcsetattr(0, TCSANOW, &tc);
881         /* trash return*/
882         getc(stdin);
883
884         do {
885                 print_sym_table();
886         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
887
888         c = getc(stdin);
889         tcsetattr(0, TCSAFLUSH, &save);
890
891         handle_keypress(c);
892         goto repeat;
893
894         return NULL;
895 }
896
897 /* Tag samples to be skipped. */
898 static const char *skip_symbols[] = {
899         "default_idle",
900         "cpu_idle",
901         "enter_idle",
902         "exit_idle",
903         "mwait_idle",
904         "mwait_idle_with_hints",
905         "poll_idle",
906         "ppc64_runlatch_off",
907         "pseries_dedicated_idle_sleep",
908         NULL
909 };
910
911 static int symbol_filter(struct map *map, struct symbol *sym)
912 {
913         struct sym_entry *syme;
914         const char *name = sym->name;
915         int i;
916
917         /*
918          * ppc64 uses function descriptors and appends a '.' to the
919          * start of every instruction address. Remove it.
920          */
921         if (name[0] == '.')
922                 name++;
923
924         if (!strcmp(name, "_text") ||
925             !strcmp(name, "_etext") ||
926             !strcmp(name, "_sinittext") ||
927             !strncmp("init_module", name, 11) ||
928             !strncmp("cleanup_module", name, 14) ||
929             strstr(name, "_text_start") ||
930             strstr(name, "_text_end"))
931                 return 1;
932
933         syme = symbol__priv(sym);
934         syme->map = map;
935         syme->src = NULL;
936         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter))
937                 sym_filter_entry = syme;
938
939         for (i = 0; skip_symbols[i]; i++) {
940                 if (!strcmp(skip_symbols[i], name)) {
941                         syme->skip = 1;
942                         break;
943                 }
944         }
945
946         if (!syme->skip)
947                 syme->name_len = strlen(sym->name);
948
949         return 0;
950 }
951
952 static void event__process_sample(const event_t *self, int counter)
953 {
954         u64 ip = self->ip.ip;
955         struct map *map;
956         struct sym_entry *syme;
957         struct symbol *sym;
958         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
959
960         switch (origin) {
961         case PERF_RECORD_MISC_USER: {
962                 struct thread *thread;
963
964                 if (hide_user_symbols)
965                         return;
966
967                 thread = threads__findnew(self->ip.pid);
968                 if (thread == NULL)
969                         return;
970
971                 map = thread__find_map(thread, ip);
972                 if (map != NULL) {
973                         ip = map->map_ip(map, ip);
974                         sym = map__find_symbol(map, ip, symbol_filter);
975                         if (sym == NULL)
976                                 return;
977                         userspace_samples++;
978                         break;
979                 }
980         }
981                 /*
982                  * If this is outside of all known maps,
983                  * and is a negative address, try to look it
984                  * up in the kernel dso, as it might be a
985                  * vsyscall or vdso (which executes in user-mode).
986                  */
987                 if ((long long)ip >= 0)
988                         return;
989                 /* Fall thru */
990         case PERF_RECORD_MISC_KERNEL:
991                 if (hide_kernel_symbols)
992                         return;
993
994                 sym = kernel_maps__find_symbol(ip, &map, symbol_filter);
995                 if (sym == NULL)
996                         return;
997                 break;
998         default:
999                 return;
1000         }
1001
1002         syme = symbol__priv(sym);
1003
1004         if (!syme->skip) {
1005                 syme->count[counter]++;
1006                 syme->origin = origin;
1007                 record_precise_ip(syme, counter, ip);
1008                 pthread_mutex_lock(&active_symbols_lock);
1009                 if (list_empty(&syme->node) || !syme->node.next)
1010                         __list_insert_active_sym(syme);
1011                 pthread_mutex_unlock(&active_symbols_lock);
1012                 ++samples;
1013                 return;
1014         }
1015 }
1016
1017 static void event__process_mmap(event_t *self)
1018 {
1019         struct thread *thread = threads__findnew(self->mmap.pid);
1020
1021         if (thread != NULL) {
1022                 struct map *map = map__new(&self->mmap, NULL, 0);
1023                 if (map != NULL)
1024                         thread__insert_map(thread, map);
1025         }
1026 }
1027
1028 static void event__process_comm(event_t *self)
1029 {
1030         struct thread *thread = threads__findnew(self->comm.pid);
1031
1032         if (thread != NULL)
1033                 thread__set_comm(thread, self->comm.comm);
1034 }
1035
1036 static int event__process(event_t *event)
1037 {
1038         switch (event->header.type) {
1039         case PERF_RECORD_COMM:
1040                 event__process_comm(event);
1041                 break;
1042         case PERF_RECORD_MMAP:
1043                 event__process_mmap(event);
1044                 break;
1045         default:
1046                 break;
1047         }
1048
1049         return 0;
1050 }
1051
1052 struct mmap_data {
1053         int                     counter;
1054         void                    *base;
1055         int                     mask;
1056         unsigned int            prev;
1057 };
1058
1059 static unsigned int mmap_read_head(struct mmap_data *md)
1060 {
1061         struct perf_event_mmap_page *pc = md->base;
1062         int head;
1063
1064         head = pc->data_head;
1065         rmb();
1066
1067         return head;
1068 }
1069
1070 static void mmap_read_counter(struct mmap_data *md)
1071 {
1072         unsigned int head = mmap_read_head(md);
1073         unsigned int old = md->prev;
1074         unsigned char *data = md->base + page_size;
1075         int diff;
1076
1077         /*
1078          * If we're further behind than half the buffer, there's a chance
1079          * the writer will bite our tail and mess up the samples under us.
1080          *
1081          * If we somehow ended up ahead of the head, we got messed up.
1082          *
1083          * In either case, truncate and restart at head.
1084          */
1085         diff = head - old;
1086         if (diff > md->mask / 2 || diff < 0) {
1087                 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1088
1089                 /*
1090                  * head points to a known good entry, start there.
1091                  */
1092                 old = head;
1093         }
1094
1095         for (; old != head;) {
1096                 event_t *event = (event_t *)&data[old & md->mask];
1097
1098                 event_t event_copy;
1099
1100                 size_t size = event->header.size;
1101
1102                 /*
1103                  * Event straddles the mmap boundary -- header should always
1104                  * be inside due to u64 alignment of output.
1105                  */
1106                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1107                         unsigned int offset = old;
1108                         unsigned int len = min(sizeof(*event), size), cpy;
1109                         void *dst = &event_copy;
1110
1111                         do {
1112                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
1113                                 memcpy(dst, &data[offset & md->mask], cpy);
1114                                 offset += cpy;
1115                                 dst += cpy;
1116                                 len -= cpy;
1117                         } while (len);
1118
1119                         event = &event_copy;
1120                 }
1121
1122                 if (event->header.type == PERF_RECORD_SAMPLE)
1123                         event__process_sample(event, md->counter);
1124                 else
1125                         event__process(event);
1126                 old += size;
1127         }
1128
1129         md->prev = old;
1130 }
1131
1132 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1133 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1134
1135 static void mmap_read(void)
1136 {
1137         int i, counter;
1138
1139         for (i = 0; i < nr_cpus; i++) {
1140                 for (counter = 0; counter < nr_counters; counter++)
1141                         mmap_read_counter(&mmap_array[i][counter]);
1142         }
1143 }
1144
1145 int nr_poll;
1146 int group_fd;
1147
1148 static void start_counter(int i, int counter)
1149 {
1150         struct perf_event_attr *attr;
1151         int cpu;
1152
1153         cpu = profile_cpu;
1154         if (target_pid == -1 && profile_cpu == -1)
1155                 cpu = i;
1156
1157         attr = attrs + counter;
1158
1159         attr->sample_type       = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1160
1161         if (freq) {
1162                 attr->sample_type       |= PERF_SAMPLE_PERIOD;
1163                 attr->freq              = 1;
1164                 attr->sample_freq       = freq;
1165         }
1166
1167         attr->inherit           = (cpu < 0) && inherit;
1168         attr->mmap              = 1;
1169
1170 try_again:
1171         fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1172
1173         if (fd[i][counter] < 0) {
1174                 int err = errno;
1175
1176                 if (err == EPERM || err == EACCES)
1177                         die("No permission - are you root?\n");
1178                 /*
1179                  * If it's cycles then fall back to hrtimer
1180                  * based cpu-clock-tick sw counter, which
1181                  * is always available even if no PMU support:
1182                  */
1183                 if (attr->type == PERF_TYPE_HARDWARE
1184                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1185
1186                         if (verbose)
1187                                 warning(" ... trying to fall back to cpu-clock-ticks\n");
1188
1189                         attr->type = PERF_TYPE_SOFTWARE;
1190                         attr->config = PERF_COUNT_SW_CPU_CLOCK;
1191                         goto try_again;
1192                 }
1193                 printf("\n");
1194                 error("perfcounter syscall returned with %d (%s)\n",
1195                         fd[i][counter], strerror(err));
1196                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1197                 exit(-1);
1198         }
1199         assert(fd[i][counter] >= 0);
1200         fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1201
1202         /*
1203          * First counter acts as the group leader:
1204          */
1205         if (group && group_fd == -1)
1206                 group_fd = fd[i][counter];
1207
1208         event_array[nr_poll].fd = fd[i][counter];
1209         event_array[nr_poll].events = POLLIN;
1210         nr_poll++;
1211
1212         mmap_array[i][counter].counter = counter;
1213         mmap_array[i][counter].prev = 0;
1214         mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1215         mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1216                         PROT_READ, MAP_SHARED, fd[i][counter], 0);
1217         if (mmap_array[i][counter].base == MAP_FAILED)
1218                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1219 }
1220
1221 static int __cmd_top(void)
1222 {
1223         pthread_t thread;
1224         int i, counter;
1225         int ret;
1226
1227         if (target_pid != -1)
1228                 event__synthesize_thread(target_pid, event__process);
1229         else
1230                 event__synthesize_threads(event__process);
1231
1232         for (i = 0; i < nr_cpus; i++) {
1233                 group_fd = -1;
1234                 for (counter = 0; counter < nr_counters; counter++)
1235                         start_counter(i, counter);
1236         }
1237
1238         /* Wait for a minimal set of events before starting the snapshot */
1239         poll(event_array, nr_poll, 100);
1240
1241         mmap_read();
1242
1243         if (pthread_create(&thread, NULL, display_thread, NULL)) {
1244                 printf("Could not create display thread.\n");
1245                 exit(-1);
1246         }
1247
1248         if (realtime_prio) {
1249                 struct sched_param param;
1250
1251                 param.sched_priority = realtime_prio;
1252                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1253                         printf("Could not set realtime priority.\n");
1254                         exit(-1);
1255                 }
1256         }
1257
1258         while (1) {
1259                 int hits = samples;
1260
1261                 mmap_read();
1262
1263                 if (hits == samples)
1264                         ret = poll(event_array, nr_poll, 100);
1265         }
1266
1267         return 0;
1268 }
1269
1270 static const char * const top_usage[] = {
1271         "perf top [<options>]",
1272         NULL
1273 };
1274
1275 static const struct option options[] = {
1276         OPT_CALLBACK('e', "event", NULL, "event",
1277                      "event selector. use 'perf list' to list available events",
1278                      parse_events),
1279         OPT_INTEGER('c', "count", &default_interval,
1280                     "event period to sample"),
1281         OPT_INTEGER('p', "pid", &target_pid,
1282                     "profile events on existing pid"),
1283         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1284                             "system-wide collection from all CPUs"),
1285         OPT_INTEGER('C', "CPU", &profile_cpu,
1286                     "CPU to profile on"),
1287         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1288         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1289                     "hide kernel symbols"),
1290         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1291                     "number of mmap data pages"),
1292         OPT_INTEGER('r', "realtime", &realtime_prio,
1293                     "collect data with this RT SCHED_FIFO priority"),
1294         OPT_INTEGER('d', "delay", &delay_secs,
1295                     "number of seconds to delay between refreshes"),
1296         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1297                             "dump the symbol table used for profiling"),
1298         OPT_INTEGER('f', "count-filter", &count_filter,
1299                     "only display functions with more events than this"),
1300         OPT_BOOLEAN('g', "group", &group,
1301                             "put the counters into a counter group"),
1302         OPT_BOOLEAN('i', "inherit", &inherit,
1303                     "child tasks inherit counters"),
1304         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1305                     "symbol to annotate - requires -k option"),
1306         OPT_BOOLEAN('z', "zero", &zero,
1307                     "zero history across updates"),
1308         OPT_INTEGER('F', "freq", &freq,
1309                     "profile at this frequency"),
1310         OPT_INTEGER('E', "entries", &print_entries,
1311                     "display this many functions"),
1312         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1313                     "hide user symbols"),
1314         OPT_BOOLEAN('v', "verbose", &verbose,
1315                     "be more verbose (show counter open errors, etc)"),
1316         OPT_END()
1317 };
1318
1319 int cmd_top(int argc, const char **argv, const char *prefix __used)
1320 {
1321         int counter, err;
1322
1323         page_size = sysconf(_SC_PAGE_SIZE);
1324
1325         argc = parse_options(argc, argv, options, top_usage, 0);
1326         if (argc)
1327                 usage_with_options(top_usage, options);
1328
1329         /* CPU and PID are mutually exclusive */
1330         if (target_pid != -1 && profile_cpu != -1) {
1331                 printf("WARNING: PID switch overriding CPU\n");
1332                 sleep(1);
1333                 profile_cpu = -1;
1334         }
1335
1336         if (!nr_counters)
1337                 nr_counters = 1;
1338
1339         symbol__init(sizeof(struct sym_entry) +
1340                      (nr_counters + 1) * sizeof(unsigned long));
1341
1342         if (delay_secs < 1)
1343                 delay_secs = 1;
1344
1345         err = kernel_maps__init(vmlinux_name, !vmlinux_name, true);
1346         if (err < 0)
1347                 return err;
1348         parse_source(sym_filter_entry);
1349
1350         /*
1351          * User specified count overrides default frequency.
1352          */
1353         if (default_interval)
1354                 freq = 0;
1355         else if (freq) {
1356                 default_interval = freq;
1357         } else {
1358                 fprintf(stderr, "frequency and count are zero, aborting\n");
1359                 exit(EXIT_FAILURE);
1360         }
1361
1362         /*
1363          * Fill in the ones not specifically initialized via -c:
1364          */
1365         for (counter = 0; counter < nr_counters; counter++) {
1366                 if (attrs[counter].sample_period)
1367                         continue;
1368
1369                 attrs[counter].sample_period = default_interval;
1370         }
1371
1372         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1373         assert(nr_cpus <= MAX_NR_CPUS);
1374         assert(nr_cpus >= 0);
1375
1376         if (target_pid != -1 || profile_cpu != -1)
1377                 nr_cpus = 1;
1378
1379         get_term_dimensions(&winsize);
1380         if (print_entries == 0) {
1381                 update_print_entries(&winsize);
1382                 signal(SIGWINCH, sig_winch_handler);
1383         }
1384
1385         return __cmd_top();
1386 }