d923127b41b610336e6675d47b1133c4212239a3
[linux-2.6.git] / tools / perf / builtin-top.c
1 /*
2  * builtin-top.c
3  *
4  * Builtin top command: Display a continuously updated profile of
5  * any workload, CPU or specific PID.
6  *
7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8  *
9  * Improvements and fixes by:
10  *
11  *   Arjan van de Ven <arjan@linux.intel.com>
12  *   Yanmin Zhang <yanmin.zhang@intel.com>
13  *   Wu Fengguang <fengguang.wu@intel.com>
14  *   Mike Galbraith <efault@gmx.de>
15  *   Paul Mackerras <paulus@samba.org>
16  *
17  * Released under the GPL v2. (and only v2, not any later version)
18  */
19 #include "builtin.h"
20
21 #include "perf.h"
22
23 #include "util/color.h"
24 #include "util/evlist.h"
25 #include "util/evsel.h"
26 #include "util/session.h"
27 #include "util/symbol.h"
28 #include "util/thread.h"
29 #include "util/thread_map.h"
30 #include "util/util.h"
31 #include <linux/rbtree.h>
32 #include "util/parse-options.h"
33 #include "util/parse-events.h"
34 #include "util/cpumap.h"
35 #include "util/xyarray.h"
36
37 #include "util/debug.h"
38
39 #include <assert.h>
40 #include <fcntl.h>
41
42 #include <stdio.h>
43 #include <termios.h>
44 #include <unistd.h>
45 #include <inttypes.h>
46
47 #include <errno.h>
48 #include <time.h>
49 #include <sched.h>
50 #include <pthread.h>
51
52 #include <sys/syscall.h>
53 #include <sys/ioctl.h>
54 #include <sys/poll.h>
55 #include <sys/prctl.h>
56 #include <sys/wait.h>
57 #include <sys/uio.h>
58 #include <sys/mman.h>
59
60 #include <linux/unistd.h>
61 #include <linux/types.h>
62
63 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
64
65 struct perf_evlist              *evsel_list;
66
67 static bool                     system_wide                     =  false;
68
69 static int                      default_interval                =      0;
70
71 static int                      count_filter                    =      5;
72 static int                      print_entries;
73
74 static int                      target_pid                      =     -1;
75 static int                      target_tid                      =     -1;
76 static struct thread_map        *threads;
77 static bool                     inherit                         =  false;
78 static struct cpu_map           *cpus;
79 static int                      realtime_prio                   =      0;
80 static bool                     group                           =  false;
81 static unsigned int             page_size;
82 static unsigned int             mmap_pages                      =    128;
83 static int                      freq                            =   1000; /* 1 KHz */
84
85 static int                      delay_secs                      =      2;
86 static bool                     zero                            =  false;
87 static bool                     dump_symtab                     =  false;
88
89 static bool                     hide_kernel_symbols             =  false;
90 static bool                     hide_user_symbols               =  false;
91 static struct winsize           winsize;
92
93 /*
94  * Source
95  */
96
97 struct source_line {
98         u64                     eip;
99         unsigned long           count[MAX_COUNTERS];
100         char                    *line;
101         struct source_line      *next;
102 };
103
104 static const char               *sym_filter                     =   NULL;
105 struct sym_entry                *sym_filter_entry               =   NULL;
106 struct sym_entry                *sym_filter_entry_sched         =   NULL;
107 static int                      sym_pcnt_filter                 =      5;
108 static int                      sym_counter                     =      0;
109 static struct perf_evsel        *sym_evsel                      =   NULL;
110 static int                      display_weighted                =     -1;
111 static const char               *cpu_list;
112
113 /*
114  * Symbols
115  */
116
117 struct sym_entry_source {
118         struct source_line      *source;
119         struct source_line      *lines;
120         struct source_line      **lines_tail;
121         pthread_mutex_t         lock;
122 };
123
124 struct sym_entry {
125         struct rb_node          rb_node;
126         struct list_head        node;
127         unsigned long           snap_count;
128         double                  weight;
129         int                     skip;
130         u16                     name_len;
131         u8                      origin;
132         struct map              *map;
133         struct sym_entry_source *src;
134         unsigned long           count[0];
135 };
136
137 /*
138  * Source functions
139  */
140
141 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
142 {
143        return ((void *)self) + symbol_conf.priv_size;
144 }
145
146 void get_term_dimensions(struct winsize *ws)
147 {
148         char *s = getenv("LINES");
149
150         if (s != NULL) {
151                 ws->ws_row = atoi(s);
152                 s = getenv("COLUMNS");
153                 if (s != NULL) {
154                         ws->ws_col = atoi(s);
155                         if (ws->ws_row && ws->ws_col)
156                                 return;
157                 }
158         }
159 #ifdef TIOCGWINSZ
160         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
161             ws->ws_row && ws->ws_col)
162                 return;
163 #endif
164         ws->ws_row = 25;
165         ws->ws_col = 80;
166 }
167
168 static void update_print_entries(struct winsize *ws)
169 {
170         print_entries = ws->ws_row;
171
172         if (print_entries > 9)
173                 print_entries -= 9;
174 }
175
176 static void sig_winch_handler(int sig __used)
177 {
178         get_term_dimensions(&winsize);
179         update_print_entries(&winsize);
180 }
181
182 static int parse_source(struct sym_entry *syme)
183 {
184         struct symbol *sym;
185         struct sym_entry_source *source;
186         struct map *map;
187         FILE *file;
188         char command[PATH_MAX*2];
189         const char *path;
190         u64 len;
191
192         if (!syme)
193                 return -1;
194
195         sym = sym_entry__symbol(syme);
196         map = syme->map;
197
198         /*
199          * We can't annotate with just /proc/kallsyms
200          */
201         if (map->dso->origin == DSO__ORIG_KERNEL)
202                 return -1;
203
204         if (syme->src == NULL) {
205                 syme->src = zalloc(sizeof(*source));
206                 if (syme->src == NULL)
207                         return -1;
208                 pthread_mutex_init(&syme->src->lock, NULL);
209         }
210
211         source = syme->src;
212
213         if (source->lines) {
214                 pthread_mutex_lock(&source->lock);
215                 goto out_assign;
216         }
217         path = map->dso->long_name;
218
219         len = sym->end - sym->start;
220
221         sprintf(command,
222                 "objdump --start-address=%#0*" PRIx64 " --stop-address=%#0*" PRIx64 " -dS %s",
223                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
224                 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
225
226         file = popen(command, "r");
227         if (!file)
228                 return -1;
229
230         pthread_mutex_lock(&source->lock);
231         source->lines_tail = &source->lines;
232         while (!feof(file)) {
233                 struct source_line *src;
234                 size_t dummy = 0;
235                 char *c, *sep;
236
237                 src = malloc(sizeof(struct source_line));
238                 assert(src != NULL);
239                 memset(src, 0, sizeof(struct source_line));
240
241                 if (getline(&src->line, &dummy, file) < 0)
242                         break;
243                 if (!src->line)
244                         break;
245
246                 c = strchr(src->line, '\n');
247                 if (c)
248                         *c = 0;
249
250                 src->next = NULL;
251                 *source->lines_tail = src;
252                 source->lines_tail = &src->next;
253
254                 src->eip = strtoull(src->line, &sep, 16);
255                 if (*sep == ':')
256                         src->eip = map__objdump_2ip(map, src->eip);
257                 else /* this line has no ip info (e.g. source line) */
258                         src->eip = 0;
259         }
260         pclose(file);
261 out_assign:
262         sym_filter_entry = syme;
263         pthread_mutex_unlock(&source->lock);
264         return 0;
265 }
266
267 static void __zero_source_counters(struct sym_entry *syme)
268 {
269         int i;
270         struct source_line *line;
271
272         line = syme->src->lines;
273         while (line) {
274                 for (i = 0; i < evsel_list->nr_entries; i++)
275                         line->count[i] = 0;
276                 line = line->next;
277         }
278 }
279
280 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
281 {
282         struct source_line *line;
283
284         if (syme != sym_filter_entry)
285                 return;
286
287         if (pthread_mutex_trylock(&syme->src->lock))
288                 return;
289
290         if (syme->src == NULL || syme->src->source == NULL)
291                 goto out_unlock;
292
293         for (line = syme->src->lines; line; line = line->next) {
294                 /* skip lines without IP info */
295                 if (line->eip == 0)
296                         continue;
297                 if (line->eip == ip) {
298                         line->count[counter]++;
299                         break;
300                 }
301                 if (line->eip > ip)
302                         break;
303         }
304 out_unlock:
305         pthread_mutex_unlock(&syme->src->lock);
306 }
307
308 #define PATTERN_LEN             (BITS_PER_LONG / 4 + 2)
309
310 static void lookup_sym_source(struct sym_entry *syme)
311 {
312         struct symbol *symbol = sym_entry__symbol(syme);
313         struct source_line *line;
314         char pattern[PATTERN_LEN + 1];
315
316         sprintf(pattern, "%0*" PRIx64 " <", BITS_PER_LONG / 4,
317                 map__rip_2objdump(syme->map, symbol->start));
318
319         pthread_mutex_lock(&syme->src->lock);
320         for (line = syme->src->lines; line; line = line->next) {
321                 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
322                         syme->src->source = line;
323                         break;
324                 }
325         }
326         pthread_mutex_unlock(&syme->src->lock);
327 }
328
329 static void show_lines(struct source_line *queue, int count, int total)
330 {
331         int i;
332         struct source_line *line;
333
334         line = queue;
335         for (i = 0; i < count; i++) {
336                 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
337
338                 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
339                 line = line->next;
340         }
341 }
342
343 #define TRACE_COUNT     3
344
345 static void show_details(struct sym_entry *syme)
346 {
347         struct symbol *symbol;
348         struct source_line *line;
349         struct source_line *line_queue = NULL;
350         int displayed = 0;
351         int line_queue_count = 0, total = 0, more = 0;
352
353         if (!syme)
354                 return;
355
356         if (!syme->src->source)
357                 lookup_sym_source(syme);
358
359         if (!syme->src->source)
360                 return;
361
362         symbol = sym_entry__symbol(syme);
363         printf("Showing %s for %s\n", event_name(sym_evsel), symbol->name);
364         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
365
366         pthread_mutex_lock(&syme->src->lock);
367         line = syme->src->source;
368         while (line) {
369                 total += line->count[sym_counter];
370                 line = line->next;
371         }
372
373         line = syme->src->source;
374         while (line) {
375                 float pcnt = 0.0;
376
377                 if (!line_queue_count)
378                         line_queue = line;
379                 line_queue_count++;
380
381                 if (line->count[sym_counter])
382                         pcnt = 100.0 * line->count[sym_counter] / (float)total;
383                 if (pcnt >= (float)sym_pcnt_filter) {
384                         if (displayed <= print_entries)
385                                 show_lines(line_queue, line_queue_count, total);
386                         else more++;
387                         displayed += line_queue_count;
388                         line_queue_count = 0;
389                         line_queue = NULL;
390                 } else if (line_queue_count > TRACE_COUNT) {
391                         line_queue = line_queue->next;
392                         line_queue_count--;
393                 }
394
395                 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
396                 line = line->next;
397         }
398         pthread_mutex_unlock(&syme->src->lock);
399         if (more)
400                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
401 }
402
403 /*
404  * Symbols will be added here in event__process_sample and will get out
405  * after decayed.
406  */
407 static LIST_HEAD(active_symbols);
408 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
409
410 /*
411  * Ordering weight: count-1 * count-2 * ... / count-n
412  */
413 static double sym_weight(const struct sym_entry *sym)
414 {
415         double weight = sym->snap_count;
416         int counter;
417
418         if (!display_weighted)
419                 return weight;
420
421         for (counter = 1; counter < evsel_list->nr_entries - 1; counter++)
422                 weight *= sym->count[counter];
423
424         weight /= (sym->count[counter] + 1);
425
426         return weight;
427 }
428
429 static long                     samples;
430 static long                     kernel_samples, us_samples;
431 static long                     exact_samples;
432 static long                     guest_us_samples, guest_kernel_samples;
433 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
434
435 static void __list_insert_active_sym(struct sym_entry *syme)
436 {
437         list_add(&syme->node, &active_symbols);
438 }
439
440 static void list_remove_active_sym(struct sym_entry *syme)
441 {
442         pthread_mutex_lock(&active_symbols_lock);
443         list_del_init(&syme->node);
444         pthread_mutex_unlock(&active_symbols_lock);
445 }
446
447 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
448 {
449         struct rb_node **p = &tree->rb_node;
450         struct rb_node *parent = NULL;
451         struct sym_entry *iter;
452
453         while (*p != NULL) {
454                 parent = *p;
455                 iter = rb_entry(parent, struct sym_entry, rb_node);
456
457                 if (se->weight > iter->weight)
458                         p = &(*p)->rb_left;
459                 else
460                         p = &(*p)->rb_right;
461         }
462
463         rb_link_node(&se->rb_node, parent, p);
464         rb_insert_color(&se->rb_node, tree);
465 }
466
467 static void print_sym_table(struct perf_session *session)
468 {
469         int printed = 0, j;
470         struct perf_evsel *counter;
471         int snap = !display_weighted ? sym_counter : 0;
472         float samples_per_sec = samples/delay_secs;
473         float ksamples_per_sec = kernel_samples/delay_secs;
474         float us_samples_per_sec = (us_samples)/delay_secs;
475         float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs;
476         float guest_us_samples_per_sec = (guest_us_samples)/delay_secs;
477         float esamples_percent = (100.0*exact_samples)/samples;
478         float sum_ksamples = 0.0;
479         struct sym_entry *syme, *n;
480         struct rb_root tmp = RB_ROOT;
481         struct rb_node *nd;
482         int sym_width = 0, dso_width = 0, dso_short_width = 0;
483         const int win_width = winsize.ws_col - 1;
484
485         samples = us_samples = kernel_samples = exact_samples = 0;
486         guest_kernel_samples = guest_us_samples = 0;
487
488         /* Sort the active symbols */
489         pthread_mutex_lock(&active_symbols_lock);
490         syme = list_entry(active_symbols.next, struct sym_entry, node);
491         pthread_mutex_unlock(&active_symbols_lock);
492
493         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
494                 syme->snap_count = syme->count[snap];
495                 if (syme->snap_count != 0) {
496
497                         if ((hide_user_symbols &&
498                              syme->origin == PERF_RECORD_MISC_USER) ||
499                             (hide_kernel_symbols &&
500                              syme->origin == PERF_RECORD_MISC_KERNEL)) {
501                                 list_remove_active_sym(syme);
502                                 continue;
503                         }
504                         syme->weight = sym_weight(syme);
505                         rb_insert_active_sym(&tmp, syme);
506                         sum_ksamples += syme->snap_count;
507
508                         for (j = 0; j < evsel_list->nr_entries; j++)
509                                 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
510                 } else
511                         list_remove_active_sym(syme);
512         }
513
514         puts(CONSOLE_CLEAR);
515
516         if (!perf_guest) {
517                 printf("   PerfTop:%8.0f irqs/sec  kernel:%4.1f%%"
518                         "  exact: %4.1f%% [",
519                         samples_per_sec,
520                         100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) /
521                                          samples_per_sec)),
522                         esamples_percent);
523         } else {
524                 printf("   PerfTop:%8.0f irqs/sec  kernel:%4.1f%% us:%4.1f%%"
525                         " guest kernel:%4.1f%% guest us:%4.1f%%"
526                         " exact: %4.1f%% [",
527                         samples_per_sec,
528                         100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) /
529                                           samples_per_sec)),
530                         100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) /
531                                           samples_per_sec)),
532                         100.0 - (100.0 * ((samples_per_sec -
533                                                 guest_kernel_samples_per_sec) /
534                                           samples_per_sec)),
535                         100.0 - (100.0 * ((samples_per_sec -
536                                            guest_us_samples_per_sec) /
537                                           samples_per_sec)),
538                         esamples_percent);
539         }
540
541         if (evsel_list->nr_entries == 1 || !display_weighted) {
542                 struct perf_evsel *first;
543                 first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
544                 printf("%" PRIu64, (uint64_t)first->attr.sample_period);
545                 if (freq)
546                         printf("Hz ");
547                 else
548                         printf(" ");
549         }
550
551         if (!display_weighted)
552                 printf("%s", event_name(sym_evsel));
553         else list_for_each_entry(counter, &evsel_list->entries, node) {
554                 if (counter->idx)
555                         printf("/");
556
557                 printf("%s", event_name(counter));
558         }
559
560         printf( "], ");
561
562         if (target_pid != -1)
563                 printf(" (target_pid: %d", target_pid);
564         else if (target_tid != -1)
565                 printf(" (target_tid: %d", target_tid);
566         else
567                 printf(" (all");
568
569         if (cpu_list)
570                 printf(", CPU%s: %s)\n", cpus->nr > 1 ? "s" : "", cpu_list);
571         else {
572                 if (target_tid != -1)
573                         printf(")\n");
574                 else
575                         printf(", %d CPU%s)\n", cpus->nr, cpus->nr > 1 ? "s" : "");
576         }
577
578         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
579
580         if (session->hists.stats.total_lost != 0) {
581                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
582                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
583                        session->hists.stats.total_lost);
584         }
585
586         if (sym_filter_entry) {
587                 show_details(sym_filter_entry);
588                 return;
589         }
590
591         /*
592          * Find the longest symbol name that will be displayed
593          */
594         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
595                 syme = rb_entry(nd, struct sym_entry, rb_node);
596                 if (++printed > print_entries ||
597                     (int)syme->snap_count < count_filter)
598                         continue;
599
600                 if (syme->map->dso->long_name_len > dso_width)
601                         dso_width = syme->map->dso->long_name_len;
602
603                 if (syme->map->dso->short_name_len > dso_short_width)
604                         dso_short_width = syme->map->dso->short_name_len;
605
606                 if (syme->name_len > sym_width)
607                         sym_width = syme->name_len;
608         }
609
610         printed = 0;
611
612         if (sym_width + dso_width > winsize.ws_col - 29) {
613                 dso_width = dso_short_width;
614                 if (sym_width + dso_width > winsize.ws_col - 29)
615                         sym_width = winsize.ws_col - dso_width - 29;
616         }
617         putchar('\n');
618         if (evsel_list->nr_entries == 1)
619                 printf("             samples  pcnt");
620         else
621                 printf("   weight    samples  pcnt");
622
623         if (verbose)
624                 printf("         RIP       ");
625         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
626         printf("   %s    _______ _____",
627                evsel_list->nr_entries == 1 ? "      " : "______");
628         if (verbose)
629                 printf(" ________________");
630         printf(" %-*.*s", sym_width, sym_width, graph_line);
631         printf(" %-*.*s", dso_width, dso_width, graph_line);
632         puts("\n");
633
634         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
635                 struct symbol *sym;
636                 double pcnt;
637
638                 syme = rb_entry(nd, struct sym_entry, rb_node);
639                 sym = sym_entry__symbol(syme);
640                 if (++printed > print_entries || (int)syme->snap_count < count_filter)
641                         continue;
642
643                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
644                                          sum_ksamples));
645
646                 if (evsel_list->nr_entries == 1 || !display_weighted)
647                         printf("%20.2f ", syme->weight);
648                 else
649                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
650
651                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
652                 if (verbose)
653                         printf(" %016" PRIx64, sym->start);
654                 printf(" %-*.*s", sym_width, sym_width, sym->name);
655                 printf(" %-*.*s\n", dso_width, dso_width,
656                        dso_width >= syme->map->dso->long_name_len ?
657                                         syme->map->dso->long_name :
658                                         syme->map->dso->short_name);
659         }
660 }
661
662 static void prompt_integer(int *target, const char *msg)
663 {
664         char *buf = malloc(0), *p;
665         size_t dummy = 0;
666         int tmp;
667
668         fprintf(stdout, "\n%s: ", msg);
669         if (getline(&buf, &dummy, stdin) < 0)
670                 return;
671
672         p = strchr(buf, '\n');
673         if (p)
674                 *p = 0;
675
676         p = buf;
677         while(*p) {
678                 if (!isdigit(*p))
679                         goto out_free;
680                 p++;
681         }
682         tmp = strtoul(buf, NULL, 10);
683         *target = tmp;
684 out_free:
685         free(buf);
686 }
687
688 static void prompt_percent(int *target, const char *msg)
689 {
690         int tmp = 0;
691
692         prompt_integer(&tmp, msg);
693         if (tmp >= 0 && tmp <= 100)
694                 *target = tmp;
695 }
696
697 static void prompt_symbol(struct sym_entry **target, const char *msg)
698 {
699         char *buf = malloc(0), *p;
700         struct sym_entry *syme = *target, *n, *found = NULL;
701         size_t dummy = 0;
702
703         /* zero counters of active symbol */
704         if (syme) {
705                 pthread_mutex_lock(&syme->src->lock);
706                 __zero_source_counters(syme);
707                 *target = NULL;
708                 pthread_mutex_unlock(&syme->src->lock);
709         }
710
711         fprintf(stdout, "\n%s: ", msg);
712         if (getline(&buf, &dummy, stdin) < 0)
713                 goto out_free;
714
715         p = strchr(buf, '\n');
716         if (p)
717                 *p = 0;
718
719         pthread_mutex_lock(&active_symbols_lock);
720         syme = list_entry(active_symbols.next, struct sym_entry, node);
721         pthread_mutex_unlock(&active_symbols_lock);
722
723         list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
724                 struct symbol *sym = sym_entry__symbol(syme);
725
726                 if (!strcmp(buf, sym->name)) {
727                         found = syme;
728                         break;
729                 }
730         }
731
732         if (!found) {
733                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
734                 sleep(1);
735                 return;
736         } else
737                 parse_source(found);
738
739 out_free:
740         free(buf);
741 }
742
743 static void print_mapped_keys(void)
744 {
745         char *name = NULL;
746
747         if (sym_filter_entry) {
748                 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
749                 name = sym->name;
750         }
751
752         fprintf(stdout, "\nMapped keys:\n");
753         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", delay_secs);
754         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", print_entries);
755
756         if (evsel_list->nr_entries > 1)
757                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(sym_evsel));
758
759         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", count_filter);
760
761         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
762         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
763         fprintf(stdout, "\t[S]     stop annotation.\n");
764
765         if (evsel_list->nr_entries > 1)
766                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
767
768         fprintf(stdout,
769                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
770                 hide_kernel_symbols ? "yes" : "no");
771         fprintf(stdout,
772                 "\t[U]     hide user symbols.               \t(%s)\n",
773                 hide_user_symbols ? "yes" : "no");
774         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", zero ? 1 : 0);
775         fprintf(stdout, "\t[qQ]    quit.\n");
776 }
777
778 static int key_mapped(int c)
779 {
780         switch (c) {
781                 case 'd':
782                 case 'e':
783                 case 'f':
784                 case 'z':
785                 case 'q':
786                 case 'Q':
787                 case 'K':
788                 case 'U':
789                 case 'F':
790                 case 's':
791                 case 'S':
792                         return 1;
793                 case 'E':
794                 case 'w':
795                         return evsel_list->nr_entries > 1 ? 1 : 0;
796                 default:
797                         break;
798         }
799
800         return 0;
801 }
802
803 static void handle_keypress(struct perf_session *session, int c)
804 {
805         if (!key_mapped(c)) {
806                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
807                 struct termios tc, save;
808
809                 print_mapped_keys();
810                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
811                 fflush(stdout);
812
813                 tcgetattr(0, &save);
814                 tc = save;
815                 tc.c_lflag &= ~(ICANON | ECHO);
816                 tc.c_cc[VMIN] = 0;
817                 tc.c_cc[VTIME] = 0;
818                 tcsetattr(0, TCSANOW, &tc);
819
820                 poll(&stdin_poll, 1, -1);
821                 c = getc(stdin);
822
823                 tcsetattr(0, TCSAFLUSH, &save);
824                 if (!key_mapped(c))
825                         return;
826         }
827
828         switch (c) {
829                 case 'd':
830                         prompt_integer(&delay_secs, "Enter display delay");
831                         if (delay_secs < 1)
832                                 delay_secs = 1;
833                         break;
834                 case 'e':
835                         prompt_integer(&print_entries, "Enter display entries (lines)");
836                         if (print_entries == 0) {
837                                 sig_winch_handler(SIGWINCH);
838                                 signal(SIGWINCH, sig_winch_handler);
839                         } else
840                                 signal(SIGWINCH, SIG_DFL);
841                         break;
842                 case 'E':
843                         if (evsel_list->nr_entries > 1) {
844                                 fprintf(stderr, "\nAvailable events:");
845
846                                 list_for_each_entry(sym_evsel, &evsel_list->entries, node)
847                                         fprintf(stderr, "\n\t%d %s", sym_evsel->idx, event_name(sym_evsel));
848
849                                 prompt_integer(&sym_counter, "Enter details event counter");
850
851                                 if (sym_counter >= evsel_list->nr_entries) {
852                                         sym_evsel = list_entry(evsel_list->entries.next, struct perf_evsel, node);
853                                         sym_counter = 0;
854                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(sym_evsel));
855                                         sleep(1);
856                                         break;
857                                 }
858                                 list_for_each_entry(sym_evsel, &evsel_list->entries, node)
859                                         if (sym_evsel->idx == sym_counter)
860                                                 break;
861                         } else sym_counter = 0;
862                         break;
863                 case 'f':
864                         prompt_integer(&count_filter, "Enter display event count filter");
865                         break;
866                 case 'F':
867                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
868                         break;
869                 case 'K':
870                         hide_kernel_symbols = !hide_kernel_symbols;
871                         break;
872                 case 'q':
873                 case 'Q':
874                         printf("exiting.\n");
875                         if (dump_symtab)
876                                 perf_session__fprintf_dsos(session, stderr);
877                         exit(0);
878                 case 's':
879                         prompt_symbol(&sym_filter_entry, "Enter details symbol");
880                         break;
881                 case 'S':
882                         if (!sym_filter_entry)
883                                 break;
884                         else {
885                                 struct sym_entry *syme = sym_filter_entry;
886
887                                 pthread_mutex_lock(&syme->src->lock);
888                                 sym_filter_entry = NULL;
889                                 __zero_source_counters(syme);
890                                 pthread_mutex_unlock(&syme->src->lock);
891                         }
892                         break;
893                 case 'U':
894                         hide_user_symbols = !hide_user_symbols;
895                         break;
896                 case 'w':
897                         display_weighted = ~display_weighted;
898                         break;
899                 case 'z':
900                         zero = !zero;
901                         break;
902                 default:
903                         break;
904         }
905 }
906
907 static void *display_thread(void *arg __used)
908 {
909         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
910         struct termios tc, save;
911         int delay_msecs, c;
912         struct perf_session *session = (struct perf_session *) arg;
913
914         tcgetattr(0, &save);
915         tc = save;
916         tc.c_lflag &= ~(ICANON | ECHO);
917         tc.c_cc[VMIN] = 0;
918         tc.c_cc[VTIME] = 0;
919
920 repeat:
921         delay_msecs = delay_secs * 1000;
922         tcsetattr(0, TCSANOW, &tc);
923         /* trash return*/
924         getc(stdin);
925
926         do {
927                 print_sym_table(session);
928         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
929
930         c = getc(stdin);
931         tcsetattr(0, TCSAFLUSH, &save);
932
933         handle_keypress(session, c);
934         goto repeat;
935
936         return NULL;
937 }
938
939 /* Tag samples to be skipped. */
940 static const char *skip_symbols[] = {
941         "default_idle",
942         "native_safe_halt",
943         "cpu_idle",
944         "enter_idle",
945         "exit_idle",
946         "mwait_idle",
947         "mwait_idle_with_hints",
948         "poll_idle",
949         "ppc64_runlatch_off",
950         "pseries_dedicated_idle_sleep",
951         NULL
952 };
953
954 static int symbol_filter(struct map *map, struct symbol *sym)
955 {
956         struct sym_entry *syme;
957         const char *name = sym->name;
958         int i;
959
960         /*
961          * ppc64 uses function descriptors and appends a '.' to the
962          * start of every instruction address. Remove it.
963          */
964         if (name[0] == '.')
965                 name++;
966
967         if (!strcmp(name, "_text") ||
968             !strcmp(name, "_etext") ||
969             !strcmp(name, "_sinittext") ||
970             !strncmp("init_module", name, 11) ||
971             !strncmp("cleanup_module", name, 14) ||
972             strstr(name, "_text_start") ||
973             strstr(name, "_text_end"))
974                 return 1;
975
976         syme = symbol__priv(sym);
977         syme->map = map;
978         syme->src = NULL;
979
980         if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
981                 /* schedule initial sym_filter_entry setup */
982                 sym_filter_entry_sched = syme;
983                 sym_filter = NULL;
984         }
985
986         for (i = 0; skip_symbols[i]; i++) {
987                 if (!strcmp(skip_symbols[i], name)) {
988                         syme->skip = 1;
989                         break;
990                 }
991         }
992
993         if (!syme->skip)
994                 syme->name_len = strlen(sym->name);
995
996         return 0;
997 }
998
999 static void event__process_sample(const event_t *self,
1000                                   struct perf_sample *sample,
1001                                   struct perf_session *session)
1002 {
1003         u64 ip = self->ip.ip;
1004         struct sym_entry *syme;
1005         struct addr_location al;
1006         struct machine *machine;
1007         u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1008
1009         ++samples;
1010
1011         switch (origin) {
1012         case PERF_RECORD_MISC_USER:
1013                 ++us_samples;
1014                 if (hide_user_symbols)
1015                         return;
1016                 machine = perf_session__find_host_machine(session);
1017                 break;
1018         case PERF_RECORD_MISC_KERNEL:
1019                 ++kernel_samples;
1020                 if (hide_kernel_symbols)
1021                         return;
1022                 machine = perf_session__find_host_machine(session);
1023                 break;
1024         case PERF_RECORD_MISC_GUEST_KERNEL:
1025                 ++guest_kernel_samples;
1026                 machine = perf_session__find_machine(session, self->ip.pid);
1027                 break;
1028         case PERF_RECORD_MISC_GUEST_USER:
1029                 ++guest_us_samples;
1030                 /*
1031                  * TODO: we don't process guest user from host side
1032                  * except simple counting.
1033                  */
1034                 return;
1035         default:
1036                 return;
1037         }
1038
1039         if (!machine && perf_guest) {
1040                 pr_err("Can't find guest [%d]'s kernel information\n",
1041                         self->ip.pid);
1042                 return;
1043         }
1044
1045         if (self->header.misc & PERF_RECORD_MISC_EXACT_IP)
1046                 exact_samples++;
1047
1048         if (event__preprocess_sample(self, session, &al, sample,
1049                                      symbol_filter) < 0 ||
1050             al.filtered)
1051                 return;
1052
1053         if (al.sym == NULL) {
1054                 /*
1055                  * As we do lazy loading of symtabs we only will know if the
1056                  * specified vmlinux file is invalid when we actually have a
1057                  * hit in kernel space and then try to load it. So if we get
1058                  * here and there are _no_ symbols in the DSO backing the
1059                  * kernel map, bail out.
1060                  *
1061                  * We may never get here, for instance, if we use -K/
1062                  * --hide-kernel-symbols, even if the user specifies an
1063                  * invalid --vmlinux ;-)
1064                  */
1065                 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
1066                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
1067                         pr_err("The %s file can't be used\n",
1068                                symbol_conf.vmlinux_name);
1069                         exit(1);
1070                 }
1071
1072                 return;
1073         }
1074
1075         /* let's see, whether we need to install initial sym_filter_entry */
1076         if (sym_filter_entry_sched) {
1077                 sym_filter_entry = sym_filter_entry_sched;
1078                 sym_filter_entry_sched = NULL;
1079                 if (parse_source(sym_filter_entry) < 0) {
1080                         struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1081
1082                         pr_err("Can't annotate %s", sym->name);
1083                         if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1084                                 pr_err(": No vmlinux file was found in the path:\n");
1085                                 machine__fprintf_vmlinux_path(machine, stderr);
1086                         } else
1087                                 pr_err(".\n");
1088                         exit(1);
1089                 }
1090         }
1091
1092         syme = symbol__priv(al.sym);
1093         if (!syme->skip) {
1094                 struct perf_evsel *evsel;
1095
1096                 syme->origin = origin;
1097                 evsel = perf_evlist__id2evsel(evsel_list, sample->id);
1098                 assert(evsel != NULL);
1099                 syme->count[evsel->idx]++;
1100                 record_precise_ip(syme, evsel->idx, ip);
1101                 pthread_mutex_lock(&active_symbols_lock);
1102                 if (list_empty(&syme->node) || !syme->node.next)
1103                         __list_insert_active_sym(syme);
1104                 pthread_mutex_unlock(&active_symbols_lock);
1105         }
1106 }
1107
1108 static void perf_session__mmap_read_cpu(struct perf_session *self, int cpu)
1109 {
1110         struct perf_sample sample;
1111         event_t *event;
1112
1113         while ((event = perf_evlist__read_on_cpu(evsel_list, cpu)) != NULL) {
1114                 perf_session__parse_sample(self, event, &sample);
1115
1116                 if (event->header.type == PERF_RECORD_SAMPLE)
1117                         event__process_sample(event, &sample, self);
1118                 else
1119                         event__process(event, &sample, self);
1120         }
1121 }
1122
1123 static void perf_session__mmap_read(struct perf_session *self)
1124 {
1125         int i;
1126
1127         for (i = 0; i < cpus->nr; i++)
1128                 perf_session__mmap_read_cpu(self, i);
1129 }
1130
1131 static void start_counters(struct perf_evlist *evlist)
1132 {
1133         struct perf_evsel *counter;
1134
1135         list_for_each_entry(counter, &evlist->entries, node) {
1136                 struct perf_event_attr *attr = &counter->attr;
1137
1138                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1139
1140                 if (freq) {
1141                         attr->sample_type |= PERF_SAMPLE_PERIOD;
1142                         attr->freq        = 1;
1143                         attr->sample_freq = freq;
1144                 }
1145
1146                 if (evlist->nr_entries > 1) {
1147                         attr->sample_type |= PERF_SAMPLE_ID;
1148                         attr->read_format |= PERF_FORMAT_ID;
1149                 }
1150
1151                 attr->mmap = 1;
1152 try_again:
1153                 if (perf_evsel__open(counter, cpus, threads, group, inherit) < 0) {
1154                         int err = errno;
1155
1156                         if (err == EPERM || err == EACCES)
1157                                 die("Permission error - are you root?\n"
1158                                         "\t Consider tweaking"
1159                                         " /proc/sys/kernel/perf_event_paranoid.\n");
1160                         /*
1161                          * If it's cycles then fall back to hrtimer
1162                          * based cpu-clock-tick sw counter, which
1163                          * is always available even if no PMU support:
1164                          */
1165                         if (attr->type == PERF_TYPE_HARDWARE &&
1166                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1167
1168                                 if (verbose)
1169                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
1170
1171                                 attr->type = PERF_TYPE_SOFTWARE;
1172                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1173                                 goto try_again;
1174                         }
1175                         printf("\n");
1176                         error("sys_perf_event_open() syscall returned with %d "
1177                               "(%s).  /bin/dmesg may provide additional information.\n",
1178                               err, strerror(err));
1179                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1180                         exit(-1);
1181                 }
1182         }
1183
1184         if (perf_evlist__mmap(evlist, cpus, threads, mmap_pages, false) < 0)
1185                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1186 }
1187
1188 static int __cmd_top(void)
1189 {
1190         pthread_t thread;
1191         struct perf_evsel *first;
1192         int ret;
1193         /*
1194          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1195          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1196          */
1197         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1198         if (session == NULL)
1199                 return -ENOMEM;
1200
1201         if (target_tid != -1)
1202                 event__synthesize_thread(target_tid, event__process, session);
1203         else
1204                 event__synthesize_threads(event__process, session);
1205
1206         start_counters(evsel_list);
1207         first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
1208         perf_session__set_sample_type(session, first->attr.sample_type);
1209
1210         /* Wait for a minimal set of events before starting the snapshot */
1211         poll(evsel_list->pollfd, evsel_list->nr_fds, 100);
1212
1213         perf_session__mmap_read(session);
1214
1215         if (pthread_create(&thread, NULL, display_thread, session)) {
1216                 printf("Could not create display thread.\n");
1217                 exit(-1);
1218         }
1219
1220         if (realtime_prio) {
1221                 struct sched_param param;
1222
1223                 param.sched_priority = realtime_prio;
1224                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1225                         printf("Could not set realtime priority.\n");
1226                         exit(-1);
1227                 }
1228         }
1229
1230         while (1) {
1231                 int hits = samples;
1232
1233                 perf_session__mmap_read(session);
1234
1235                 if (hits == samples)
1236                         ret = poll(evsel_list->pollfd, evsel_list->nr_fds, 100);
1237         }
1238
1239         return 0;
1240 }
1241
1242 static const char * const top_usage[] = {
1243         "perf top [<options>]",
1244         NULL
1245 };
1246
1247 static const struct option options[] = {
1248         OPT_CALLBACK('e', "event", &evsel_list, "event",
1249                      "event selector. use 'perf list' to list available events",
1250                      parse_events),
1251         OPT_INTEGER('c', "count", &default_interval,
1252                     "event period to sample"),
1253         OPT_INTEGER('p', "pid", &target_pid,
1254                     "profile events on existing process id"),
1255         OPT_INTEGER('t', "tid", &target_tid,
1256                     "profile events on existing thread id"),
1257         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1258                             "system-wide collection from all CPUs"),
1259         OPT_STRING('C', "cpu", &cpu_list, "cpu",
1260                     "list of cpus to monitor"),
1261         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1262                    "file", "vmlinux pathname"),
1263         OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1264                     "hide kernel symbols"),
1265         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1266         OPT_INTEGER('r', "realtime", &realtime_prio,
1267                     "collect data with this RT SCHED_FIFO priority"),
1268         OPT_INTEGER('d', "delay", &delay_secs,
1269                     "number of seconds to delay between refreshes"),
1270         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1271                             "dump the symbol table used for profiling"),
1272         OPT_INTEGER('f', "count-filter", &count_filter,
1273                     "only display functions with more events than this"),
1274         OPT_BOOLEAN('g', "group", &group,
1275                             "put the counters into a counter group"),
1276         OPT_BOOLEAN('i', "inherit", &inherit,
1277                     "child tasks inherit counters"),
1278         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1279                     "symbol to annotate"),
1280         OPT_BOOLEAN('z', "zero", &zero,
1281                     "zero history across updates"),
1282         OPT_INTEGER('F', "freq", &freq,
1283                     "profile at this frequency"),
1284         OPT_INTEGER('E', "entries", &print_entries,
1285                     "display this many functions"),
1286         OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1287                     "hide user symbols"),
1288         OPT_INCR('v', "verbose", &verbose,
1289                     "be more verbose (show counter open errors, etc)"),
1290         OPT_END()
1291 };
1292
1293 int cmd_top(int argc, const char **argv, const char *prefix __used)
1294 {
1295         struct perf_evsel *pos;
1296         int status = -ENOMEM;
1297
1298         evsel_list = perf_evlist__new();
1299         if (evsel_list == NULL)
1300                 return -ENOMEM;
1301
1302         page_size = sysconf(_SC_PAGE_SIZE);
1303
1304         argc = parse_options(argc, argv, options, top_usage, 0);
1305         if (argc)
1306                 usage_with_options(top_usage, options);
1307
1308         if (target_pid != -1)
1309                 target_tid = target_pid;
1310
1311         threads = thread_map__new(target_pid, target_tid);
1312         if (threads == NULL) {
1313                 pr_err("Problems finding threads of monitor\n");
1314                 usage_with_options(top_usage, options);
1315         }
1316
1317         /* CPU and PID are mutually exclusive */
1318         if (target_tid > 0 && cpu_list) {
1319                 printf("WARNING: PID switch overriding CPU\n");
1320                 sleep(1);
1321                 cpu_list = NULL;
1322         }
1323
1324         if (!evsel_list->nr_entries &&
1325             perf_evlist__add_default(evsel_list) < 0) {
1326                 pr_err("Not enough memory for event selector list\n");
1327                 return -ENOMEM;
1328         }
1329
1330         if (delay_secs < 1)
1331                 delay_secs = 1;
1332
1333         /*
1334          * User specified count overrides default frequency.
1335          */
1336         if (default_interval)
1337                 freq = 0;
1338         else if (freq) {
1339                 default_interval = freq;
1340         } else {
1341                 fprintf(stderr, "frequency and count are zero, aborting\n");
1342                 exit(EXIT_FAILURE);
1343         }
1344
1345         if (target_tid != -1)
1346                 cpus = cpu_map__dummy_new();
1347         else
1348                 cpus = cpu_map__new(cpu_list);
1349
1350         if (cpus == NULL)
1351                 usage_with_options(top_usage, options);
1352
1353         list_for_each_entry(pos, &evsel_list->entries, node) {
1354                 if (perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
1355                         goto out_free_fd;
1356                 /*
1357                  * Fill in the ones not specifically initialized via -c:
1358                  */
1359                 if (pos->attr.sample_period)
1360                         continue;
1361
1362                 pos->attr.sample_period = default_interval;
1363         }
1364
1365         if (perf_evlist__alloc_pollfd(evsel_list, cpus->nr, threads->nr) < 0 ||
1366             perf_evlist__alloc_mmap(evsel_list, cpus->nr) < 0)
1367                 goto out_free_fd;
1368
1369         sym_evsel = list_entry(evsel_list->entries.next, struct perf_evsel, node);
1370
1371         symbol_conf.priv_size = (sizeof(struct sym_entry) +
1372                                  (evsel_list->nr_entries + 1) * sizeof(unsigned long));
1373
1374         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1375         if (symbol__init() < 0)
1376                 return -1;
1377
1378         get_term_dimensions(&winsize);
1379         if (print_entries == 0) {
1380                 update_print_entries(&winsize);
1381                 signal(SIGWINCH, sig_winch_handler);
1382         }
1383
1384         status = __cmd_top();
1385 out_free_fd:
1386         perf_evlist__delete(evsel_list);
1387
1388         return status;
1389 }