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