375ed160d9362f4d6755addac29708678d476203
[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/annotate.h"
24 #include "util/cache.h"
25 #include "util/color.h"
26 #include "util/evlist.h"
27 #include "util/evsel.h"
28 #include "util/session.h"
29 #include "util/symbol.h"
30 #include "util/thread.h"
31 #include "util/thread_map.h"
32 #include "util/top.h"
33 #include "util/util.h"
34 #include <linux/rbtree.h>
35 #include "util/parse-options.h"
36 #include "util/parse-events.h"
37 #include "util/cpumap.h"
38 #include "util/xyarray.h"
39
40 #include "util/debug.h"
41
42 #include <assert.h>
43 #include <fcntl.h>
44
45 #include <stdio.h>
46 #include <termios.h>
47 #include <unistd.h>
48 #include <inttypes.h>
49
50 #include <errno.h>
51 #include <time.h>
52 #include <sched.h>
53
54 #include <sys/syscall.h>
55 #include <sys/ioctl.h>
56 #include <sys/poll.h>
57 #include <sys/prctl.h>
58 #include <sys/wait.h>
59 #include <sys/uio.h>
60 #include <sys/mman.h>
61
62 #include <linux/unistd.h>
63 #include <linux/types.h>
64
65 static struct perf_top top = {
66         .count_filter           = 5,
67         .delay_secs             = 2,
68         .display_weighted       = -1,
69         .target_pid             = -1,
70         .target_tid             = -1,
71         .active_symbols         = LIST_HEAD_INIT(top.active_symbols),
72         .active_symbols_lock    = PTHREAD_MUTEX_INITIALIZER,
73         .active_symbols_cond    = PTHREAD_COND_INITIALIZER,
74         .freq                   = 1000, /* 1 KHz */
75 };
76
77 static bool                     system_wide                     =  false;
78
79 static bool                     use_tui, use_stdio;
80
81 static int                      default_interval                =      0;
82
83 static bool                     inherit                         =  false;
84 static int                      realtime_prio                   =      0;
85 static bool                     group                           =  false;
86 static unsigned int             page_size;
87 static unsigned int             mmap_pages                      =    128;
88
89 static bool                     dump_symtab                     =  false;
90
91 static struct winsize           winsize;
92
93 static const char               *sym_filter                     =   NULL;
94 struct sym_entry                *sym_filter_entry_sched         =   NULL;
95 static int                      sym_pcnt_filter                 =      5;
96
97 /*
98  * Source functions
99  */
100
101 void get_term_dimensions(struct winsize *ws)
102 {
103         char *s = getenv("LINES");
104
105         if (s != NULL) {
106                 ws->ws_row = atoi(s);
107                 s = getenv("COLUMNS");
108                 if (s != NULL) {
109                         ws->ws_col = atoi(s);
110                         if (ws->ws_row && ws->ws_col)
111                                 return;
112                 }
113         }
114 #ifdef TIOCGWINSZ
115         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
116             ws->ws_row && ws->ws_col)
117                 return;
118 #endif
119         ws->ws_row = 25;
120         ws->ws_col = 80;
121 }
122
123 static void update_print_entries(struct winsize *ws)
124 {
125         top.print_entries = ws->ws_row;
126
127         if (top.print_entries > 9)
128                 top.print_entries -= 9;
129 }
130
131 static void sig_winch_handler(int sig __used)
132 {
133         get_term_dimensions(&winsize);
134         update_print_entries(&winsize);
135 }
136
137 static int parse_source(struct sym_entry *syme)
138 {
139         struct symbol *sym;
140         struct annotation *notes;
141         struct map *map;
142         int err = -1;
143
144         if (!syme)
145                 return -1;
146
147         sym = sym_entry__symbol(syme);
148         map = syme->map;
149
150         /*
151          * We can't annotate with just /proc/kallsyms
152          */
153         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
154                 pr_err("Can't annotate %s: No vmlinux file was found in the "
155                        "path\n", sym->name);
156                 sleep(1);
157                 return -1;
158         }
159
160         notes = symbol__annotation(sym);
161         if (notes->src != NULL) {
162                 pthread_mutex_lock(&notes->lock);
163                 goto out_assign;
164         }
165
166         pthread_mutex_lock(&notes->lock);
167
168         if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
169                 pthread_mutex_unlock(&notes->lock);
170                 pr_err("Not enough memory for annotating '%s' symbol!\n",
171                        sym->name);
172                 sleep(1);
173                 return err;
174         }
175
176         err = symbol__annotate(sym, syme->map, 0);
177         if (err == 0) {
178 out_assign:
179                 top.sym_filter_entry = syme;
180         }
181
182         pthread_mutex_unlock(&notes->lock);
183         return err;
184 }
185
186 static void __zero_source_counters(struct sym_entry *syme)
187 {
188         struct symbol *sym = sym_entry__symbol(syme);
189         symbol__annotate_zero_histograms(sym);
190 }
191
192 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
193 {
194         struct annotation *notes;
195         struct symbol *sym;
196
197         if (syme != top.sym_filter_entry)
198                 return;
199
200         sym = sym_entry__symbol(syme);
201         notes = symbol__annotation(sym);
202
203         if (pthread_mutex_trylock(&notes->lock))
204                 return;
205
206         ip = syme->map->map_ip(syme->map, ip);
207         symbol__inc_addr_samples(sym, syme->map, counter, ip);
208
209         pthread_mutex_unlock(&notes->lock);
210 }
211
212 static void show_details(struct sym_entry *syme)
213 {
214         struct annotation *notes;
215         struct symbol *symbol;
216         int more;
217
218         if (!syme)
219                 return;
220
221         symbol = sym_entry__symbol(syme);
222         notes = symbol__annotation(symbol);
223
224         pthread_mutex_lock(&notes->lock);
225
226         if (notes->src == NULL)
227                 goto out_unlock;
228
229         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
230         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
231
232         more = symbol__annotate_printf(symbol, syme->map, top.sym_evsel->idx,
233                                        0, sym_pcnt_filter, top.print_entries, 4);
234         if (top.zero)
235                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
236         else
237                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
238         if (more != 0)
239                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
240 out_unlock:
241         pthread_mutex_unlock(&notes->lock);
242 }
243
244 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
245
246 static void __list_insert_active_sym(struct sym_entry *syme)
247 {
248         list_add(&syme->node, &top.active_symbols);
249 }
250
251 static void print_sym_table(struct perf_session *session)
252 {
253         char bf[160];
254         int printed = 0;
255         struct rb_node *nd;
256         struct sym_entry *syme;
257         struct rb_root tmp = RB_ROOT;
258         const int win_width = winsize.ws_col - 1;
259         int sym_width, dso_width, dso_short_width;
260         float sum_ksamples = perf_top__decay_samples(&top, &tmp);
261
262         puts(CONSOLE_CLEAR);
263
264         perf_top__header_snprintf(&top, bf, sizeof(bf));
265         printf("%s\n", bf);
266
267         perf_top__reset_sample_counters(&top);
268
269         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
270
271         if (session->hists.stats.total_lost != 0) {
272                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
273                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
274                        session->hists.stats.total_lost);
275         }
276
277         if (top.sym_filter_entry) {
278                 show_details(top.sym_filter_entry);
279                 return;
280         }
281
282         perf_top__find_widths(&top, &tmp, &dso_width, &dso_short_width,
283                               &sym_width);
284
285         if (sym_width + dso_width > winsize.ws_col - 29) {
286                 dso_width = dso_short_width;
287                 if (sym_width + dso_width > winsize.ws_col - 29)
288                         sym_width = winsize.ws_col - dso_width - 29;
289         }
290         putchar('\n');
291         if (top.evlist->nr_entries == 1)
292                 printf("             samples  pcnt");
293         else
294                 printf("   weight    samples  pcnt");
295
296         if (verbose)
297                 printf("         RIP       ");
298         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
299         printf("   %s    _______ _____",
300                top.evlist->nr_entries == 1 ? "      " : "______");
301         if (verbose)
302                 printf(" ________________");
303         printf(" %-*.*s", sym_width, sym_width, graph_line);
304         printf(" %-*.*s", dso_width, dso_width, graph_line);
305         puts("\n");
306
307         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
308                 struct symbol *sym;
309                 double pcnt;
310
311                 syme = rb_entry(nd, struct sym_entry, rb_node);
312                 sym = sym_entry__symbol(syme);
313                 if (++printed > top.print_entries ||
314                     (int)syme->snap_count < top.count_filter)
315                         continue;
316
317                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
318                                          sum_ksamples));
319
320                 if (top.evlist->nr_entries == 1 || !top.display_weighted)
321                         printf("%20.2f ", syme->weight);
322                 else
323                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
324
325                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
326                 if (verbose)
327                         printf(" %016" PRIx64, sym->start);
328                 printf(" %-*.*s", sym_width, sym_width, sym->name);
329                 printf(" %-*.*s\n", dso_width, dso_width,
330                        dso_width >= syme->map->dso->long_name_len ?
331                                         syme->map->dso->long_name :
332                                         syme->map->dso->short_name);
333         }
334 }
335
336 static void prompt_integer(int *target, const char *msg)
337 {
338         char *buf = malloc(0), *p;
339         size_t dummy = 0;
340         int tmp;
341
342         fprintf(stdout, "\n%s: ", msg);
343         if (getline(&buf, &dummy, stdin) < 0)
344                 return;
345
346         p = strchr(buf, '\n');
347         if (p)
348                 *p = 0;
349
350         p = buf;
351         while(*p) {
352                 if (!isdigit(*p))
353                         goto out_free;
354                 p++;
355         }
356         tmp = strtoul(buf, NULL, 10);
357         *target = tmp;
358 out_free:
359         free(buf);
360 }
361
362 static void prompt_percent(int *target, const char *msg)
363 {
364         int tmp = 0;
365
366         prompt_integer(&tmp, msg);
367         if (tmp >= 0 && tmp <= 100)
368                 *target = tmp;
369 }
370
371 static void prompt_symbol(struct sym_entry **target, const char *msg)
372 {
373         char *buf = malloc(0), *p;
374         struct sym_entry *syme = *target, *n, *found = NULL;
375         size_t dummy = 0;
376
377         /* zero counters of active symbol */
378         if (syme) {
379                 __zero_source_counters(syme);
380                 *target = NULL;
381         }
382
383         fprintf(stdout, "\n%s: ", msg);
384         if (getline(&buf, &dummy, stdin) < 0)
385                 goto out_free;
386
387         p = strchr(buf, '\n');
388         if (p)
389                 *p = 0;
390
391         pthread_mutex_lock(&top.active_symbols_lock);
392         syme = list_entry(top.active_symbols.next, struct sym_entry, node);
393         pthread_mutex_unlock(&top.active_symbols_lock);
394
395         list_for_each_entry_safe_from(syme, n, &top.active_symbols, node) {
396                 struct symbol *sym = sym_entry__symbol(syme);
397
398                 if (!strcmp(buf, sym->name)) {
399                         found = syme;
400                         break;
401                 }
402         }
403
404         if (!found) {
405                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
406                 sleep(1);
407                 return;
408         } else
409                 parse_source(found);
410
411 out_free:
412         free(buf);
413 }
414
415 static void print_mapped_keys(void)
416 {
417         char *name = NULL;
418
419         if (top.sym_filter_entry) {
420                 struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
421                 name = sym->name;
422         }
423
424         fprintf(stdout, "\nMapped keys:\n");
425         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
426         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
427
428         if (top.evlist->nr_entries > 1)
429                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
430
431         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
432
433         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
434         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
435         fprintf(stdout, "\t[S]     stop annotation.\n");
436
437         if (top.evlist->nr_entries > 1)
438                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", top.display_weighted ? 1 : 0);
439
440         fprintf(stdout,
441                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
442                 top.hide_kernel_symbols ? "yes" : "no");
443         fprintf(stdout,
444                 "\t[U]     hide user symbols.               \t(%s)\n",
445                 top.hide_user_symbols ? "yes" : "no");
446         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
447         fprintf(stdout, "\t[qQ]    quit.\n");
448 }
449
450 static int key_mapped(int c)
451 {
452         switch (c) {
453                 case 'd':
454                 case 'e':
455                 case 'f':
456                 case 'z':
457                 case 'q':
458                 case 'Q':
459                 case 'K':
460                 case 'U':
461                 case 'F':
462                 case 's':
463                 case 'S':
464                         return 1;
465                 case 'E':
466                 case 'w':
467                         return top.evlist->nr_entries > 1 ? 1 : 0;
468                 default:
469                         break;
470         }
471
472         return 0;
473 }
474
475 static void handle_keypress(struct perf_session *session, int c)
476 {
477         if (!key_mapped(c)) {
478                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
479                 struct termios tc, save;
480
481                 print_mapped_keys();
482                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
483                 fflush(stdout);
484
485                 tcgetattr(0, &save);
486                 tc = save;
487                 tc.c_lflag &= ~(ICANON | ECHO);
488                 tc.c_cc[VMIN] = 0;
489                 tc.c_cc[VTIME] = 0;
490                 tcsetattr(0, TCSANOW, &tc);
491
492                 poll(&stdin_poll, 1, -1);
493                 c = getc(stdin);
494
495                 tcsetattr(0, TCSAFLUSH, &save);
496                 if (!key_mapped(c))
497                         return;
498         }
499
500         switch (c) {
501                 case 'd':
502                         prompt_integer(&top.delay_secs, "Enter display delay");
503                         if (top.delay_secs < 1)
504                                 top.delay_secs = 1;
505                         break;
506                 case 'e':
507                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
508                         if (top.print_entries == 0) {
509                                 sig_winch_handler(SIGWINCH);
510                                 signal(SIGWINCH, sig_winch_handler);
511                         } else
512                                 signal(SIGWINCH, SIG_DFL);
513                         break;
514                 case 'E':
515                         if (top.evlist->nr_entries > 1) {
516                                 /* Select 0 as the default event: */
517                                 int counter = 0;
518
519                                 fprintf(stderr, "\nAvailable events:");
520
521                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
522                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
523
524                                 prompt_integer(&counter, "Enter details event counter");
525
526                                 if (counter >= top.evlist->nr_entries) {
527                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
528                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
529                                         sleep(1);
530                                         break;
531                                 }
532                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
533                                         if (top.sym_evsel->idx == counter)
534                                                 break;
535                         } else
536                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
537                         break;
538                 case 'f':
539                         prompt_integer(&top.count_filter, "Enter display event count filter");
540                         break;
541                 case 'F':
542                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
543                         break;
544                 case 'K':
545                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
546                         break;
547                 case 'q':
548                 case 'Q':
549                         printf("exiting.\n");
550                         if (dump_symtab)
551                                 perf_session__fprintf_dsos(session, stderr);
552                         exit(0);
553                 case 's':
554                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
555                         break;
556                 case 'S':
557                         if (!top.sym_filter_entry)
558                                 break;
559                         else {
560                                 struct sym_entry *syme = top.sym_filter_entry;
561
562                                 top.sym_filter_entry = NULL;
563                                 __zero_source_counters(syme);
564                         }
565                         break;
566                 case 'U':
567                         top.hide_user_symbols = !top.hide_user_symbols;
568                         break;
569                 case 'w':
570                         top.display_weighted = ~top.display_weighted;
571                         break;
572                 case 'z':
573                         top.zero = !top.zero;
574                         break;
575                 default:
576                         break;
577         }
578 }
579
580 static void *display_thread_tui(void *arg __used)
581 {
582         int err = 0;
583         pthread_mutex_lock(&top.active_symbols_lock);
584         while (list_empty(&top.active_symbols)) {
585                 err = pthread_cond_wait(&top.active_symbols_cond,
586                                         &top.active_symbols_lock);
587                 if (err)
588                         break;
589         }
590         pthread_mutex_unlock(&top.active_symbols_lock);
591         if (!err)
592                 perf_top__tui_browser(&top);
593         exit_browser(0);
594         exit(0);
595         return NULL;
596 }
597
598 static void *display_thread(void *arg __used)
599 {
600         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
601         struct termios tc, save;
602         int delay_msecs, c;
603         struct perf_session *session = (struct perf_session *) arg;
604
605         tcgetattr(0, &save);
606         tc = save;
607         tc.c_lflag &= ~(ICANON | ECHO);
608         tc.c_cc[VMIN] = 0;
609         tc.c_cc[VTIME] = 0;
610
611 repeat:
612         delay_msecs = top.delay_secs * 1000;
613         tcsetattr(0, TCSANOW, &tc);
614         /* trash return*/
615         getc(stdin);
616
617         do {
618                 print_sym_table(session);
619         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
620
621         c = getc(stdin);
622         tcsetattr(0, TCSAFLUSH, &save);
623
624         handle_keypress(session, c);
625         goto repeat;
626
627         return NULL;
628 }
629
630 /* Tag samples to be skipped. */
631 static const char *skip_symbols[] = {
632         "default_idle",
633         "native_safe_halt",
634         "cpu_idle",
635         "enter_idle",
636         "exit_idle",
637         "mwait_idle",
638         "mwait_idle_with_hints",
639         "poll_idle",
640         "ppc64_runlatch_off",
641         "pseries_dedicated_idle_sleep",
642         NULL
643 };
644
645 static int symbol_filter(struct map *map, struct symbol *sym)
646 {
647         struct sym_entry *syme;
648         const char *name = sym->name;
649         int i;
650
651         /*
652          * ppc64 uses function descriptors and appends a '.' to the
653          * start of every instruction address. Remove it.
654          */
655         if (name[0] == '.')
656                 name++;
657
658         if (!strcmp(name, "_text") ||
659             !strcmp(name, "_etext") ||
660             !strcmp(name, "_sinittext") ||
661             !strncmp("init_module", name, 11) ||
662             !strncmp("cleanup_module", name, 14) ||
663             strstr(name, "_text_start") ||
664             strstr(name, "_text_end"))
665                 return 1;
666
667         syme = symbol__priv(sym);
668         syme->map = map;
669         symbol__annotate_init(map, sym);
670
671         if (!top.sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
672                 /* schedule initial sym_filter_entry setup */
673                 sym_filter_entry_sched = syme;
674                 sym_filter = NULL;
675         }
676
677         for (i = 0; skip_symbols[i]; i++) {
678                 if (!strcmp(skip_symbols[i], name)) {
679                         sym->ignore = true;
680                         break;
681                 }
682         }
683
684         return 0;
685 }
686
687 static void perf_event__process_sample(const union perf_event *event,
688                                        struct perf_sample *sample,
689                                        struct perf_session *session)
690 {
691         u64 ip = event->ip.ip;
692         struct sym_entry *syme;
693         struct addr_location al;
694         struct machine *machine;
695         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
696
697         ++top.samples;
698
699         switch (origin) {
700         case PERF_RECORD_MISC_USER:
701                 ++top.us_samples;
702                 if (top.hide_user_symbols)
703                         return;
704                 machine = perf_session__find_host_machine(session);
705                 break;
706         case PERF_RECORD_MISC_KERNEL:
707                 ++top.kernel_samples;
708                 if (top.hide_kernel_symbols)
709                         return;
710                 machine = perf_session__find_host_machine(session);
711                 break;
712         case PERF_RECORD_MISC_GUEST_KERNEL:
713                 ++top.guest_kernel_samples;
714                 machine = perf_session__find_machine(session, event->ip.pid);
715                 break;
716         case PERF_RECORD_MISC_GUEST_USER:
717                 ++top.guest_us_samples;
718                 /*
719                  * TODO: we don't process guest user from host side
720                  * except simple counting.
721                  */
722                 return;
723         default:
724                 return;
725         }
726
727         if (!machine && perf_guest) {
728                 pr_err("Can't find guest [%d]'s kernel information\n",
729                         event->ip.pid);
730                 return;
731         }
732
733         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
734                 top.exact_samples++;
735
736         if (perf_event__preprocess_sample(event, session, &al, sample,
737                                           symbol_filter) < 0 ||
738             al.filtered)
739                 return;
740
741         if (al.sym == NULL) {
742                 /*
743                  * As we do lazy loading of symtabs we only will know if the
744                  * specified vmlinux file is invalid when we actually have a
745                  * hit in kernel space and then try to load it. So if we get
746                  * here and there are _no_ symbols in the DSO backing the
747                  * kernel map, bail out.
748                  *
749                  * We may never get here, for instance, if we use -K/
750                  * --hide-kernel-symbols, even if the user specifies an
751                  * invalid --vmlinux ;-)
752                  */
753                 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
754                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
755                         ui__warning("The %s file can't be used\n",
756                                     symbol_conf.vmlinux_name);
757                         exit_browser(0);
758                         exit(1);
759                 }
760
761                 return;
762         }
763
764         /* let's see, whether we need to install initial sym_filter_entry */
765         if (sym_filter_entry_sched) {
766                 top.sym_filter_entry = sym_filter_entry_sched;
767                 sym_filter_entry_sched = NULL;
768                 if (parse_source(top.sym_filter_entry) < 0) {
769                         struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
770
771                         pr_err("Can't annotate %s", sym->name);
772                         if (top.sym_filter_entry->map->dso->symtab_type == SYMTAB__KALLSYMS) {
773                                 pr_err(": No vmlinux file was found in the path:\n");
774                                 machine__fprintf_vmlinux_path(machine, stderr);
775                         } else
776                                 pr_err(".\n");
777                         exit(1);
778                 }
779         }
780
781         syme = symbol__priv(al.sym);
782         if (!al.sym->ignore) {
783                 struct perf_evsel *evsel;
784
785                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
786                 assert(evsel != NULL);
787                 syme->count[evsel->idx]++;
788                 record_precise_ip(syme, evsel->idx, ip);
789                 pthread_mutex_lock(&top.active_symbols_lock);
790                 if (list_empty(&syme->node) || !syme->node.next) {
791                         static bool first = true;
792                         __list_insert_active_sym(syme);
793                         if (first) {
794                                 pthread_cond_broadcast(&top.active_symbols_cond);
795                                 first = false;
796                         }
797                 }
798                 pthread_mutex_unlock(&top.active_symbols_lock);
799         }
800 }
801
802 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
803 {
804         struct perf_sample sample;
805         union perf_event *event;
806         int ret;
807
808         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
809                 ret = perf_session__parse_sample(self, event, &sample);
810                 if (ret) {
811                         pr_err("Can't parse sample, err = %d\n", ret);
812                         continue;
813                 }
814
815                 if (event->header.type == PERF_RECORD_SAMPLE)
816                         perf_event__process_sample(event, &sample, self);
817                 else
818                         perf_event__process(event, &sample, self);
819         }
820 }
821
822 static void perf_session__mmap_read(struct perf_session *self)
823 {
824         int i;
825
826         for (i = 0; i < top.evlist->nr_mmaps; i++)
827                 perf_session__mmap_read_idx(self, i);
828 }
829
830 static void start_counters(struct perf_evlist *evlist)
831 {
832         struct perf_evsel *counter;
833
834         list_for_each_entry(counter, &evlist->entries, node) {
835                 struct perf_event_attr *attr = &counter->attr;
836
837                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
838
839                 if (top.freq) {
840                         attr->sample_type |= PERF_SAMPLE_PERIOD;
841                         attr->freq        = 1;
842                         attr->sample_freq = top.freq;
843                 }
844
845                 if (evlist->nr_entries > 1) {
846                         attr->sample_type |= PERF_SAMPLE_ID;
847                         attr->read_format |= PERF_FORMAT_ID;
848                 }
849
850                 attr->mmap = 1;
851                 attr->inherit = inherit;
852 try_again:
853                 if (perf_evsel__open(counter, top.evlist->cpus,
854                                      top.evlist->threads, group) < 0) {
855                         int err = errno;
856
857                         if (err == EPERM || err == EACCES) {
858                                 ui__warning_paranoid();
859                                 goto out_err;
860                         }
861                         /*
862                          * If it's cycles then fall back to hrtimer
863                          * based cpu-clock-tick sw counter, which
864                          * is always available even if no PMU support:
865                          */
866                         if (attr->type == PERF_TYPE_HARDWARE &&
867                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
868                                 if (verbose)
869                                         ui__warning("Cycles event not supported,\n"
870                                                     "trying to fall back to cpu-clock-ticks\n");
871
872                                 attr->type = PERF_TYPE_SOFTWARE;
873                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
874                                 goto try_again;
875                         }
876
877                         if (err == ENOENT) {
878                                 ui__warning("The %s event is not supported.\n",
879                                             event_name(counter));
880                                 goto out_err;
881                         }
882
883                         ui__warning("The sys_perf_event_open() syscall "
884                                     "returned with %d (%s).  /bin/dmesg "
885                                     "may provide additional information.\n"
886                                     "No CONFIG_PERF_EVENTS=y kernel support "
887                                     "configured?\n", err, strerror(err));
888                         goto out_err;
889                 }
890         }
891
892         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
893                 ui__warning("Failed to mmap with %d (%s)\n",
894                             errno, strerror(errno));
895                 goto out_err;
896         }
897
898         return;
899
900 out_err:
901         exit_browser(0);
902         exit(0);
903 }
904
905 static int __cmd_top(void)
906 {
907         pthread_t thread;
908         int ret __used;
909         /*
910          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
911          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
912          */
913         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
914         if (session == NULL)
915                 return -ENOMEM;
916
917         if (top.target_tid != -1)
918                 perf_event__synthesize_thread_map(top.evlist->threads,
919                                                   perf_event__process, session);
920         else
921                 perf_event__synthesize_threads(perf_event__process, session);
922
923         start_counters(top.evlist);
924         session->evlist = top.evlist;
925         perf_session__update_sample_type(session);
926
927         /* Wait for a minimal set of events before starting the snapshot */
928         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
929
930         perf_session__mmap_read(session);
931
932         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
933                                                              display_thread), session)) {
934                 printf("Could not create display thread.\n");
935                 exit(-1);
936         }
937
938         if (realtime_prio) {
939                 struct sched_param param;
940
941                 param.sched_priority = realtime_prio;
942                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
943                         printf("Could not set realtime priority.\n");
944                         exit(-1);
945                 }
946         }
947
948         while (1) {
949                 u64 hits = top.samples;
950
951                 perf_session__mmap_read(session);
952
953                 if (hits == top.samples)
954                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
955         }
956
957         return 0;
958 }
959
960 static const char * const top_usage[] = {
961         "perf top [<options>]",
962         NULL
963 };
964
965 static const struct option options[] = {
966         OPT_CALLBACK('e', "event", &top.evlist, "event",
967                      "event selector. use 'perf list' to list available events",
968                      parse_events),
969         OPT_INTEGER('c', "count", &default_interval,
970                     "event period to sample"),
971         OPT_INTEGER('p', "pid", &top.target_pid,
972                     "profile events on existing process id"),
973         OPT_INTEGER('t', "tid", &top.target_tid,
974                     "profile events on existing thread id"),
975         OPT_BOOLEAN('a', "all-cpus", &system_wide,
976                             "system-wide collection from all CPUs"),
977         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
978                     "list of cpus to monitor"),
979         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
980                    "file", "vmlinux pathname"),
981         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
982                     "hide kernel symbols"),
983         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
984         OPT_INTEGER('r', "realtime", &realtime_prio,
985                     "collect data with this RT SCHED_FIFO priority"),
986         OPT_INTEGER('d', "delay", &top.delay_secs,
987                     "number of seconds to delay between refreshes"),
988         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
989                             "dump the symbol table used for profiling"),
990         OPT_INTEGER('f', "count-filter", &top.count_filter,
991                     "only display functions with more events than this"),
992         OPT_BOOLEAN('g', "group", &group,
993                             "put the counters into a counter group"),
994         OPT_BOOLEAN('i', "inherit", &inherit,
995                     "child tasks inherit counters"),
996         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
997                     "symbol to annotate"),
998         OPT_BOOLEAN('z', "zero", &top.zero,
999                     "zero history across updates"),
1000         OPT_INTEGER('F', "freq", &top.freq,
1001                     "profile at this frequency"),
1002         OPT_INTEGER('E', "entries", &top.print_entries,
1003                     "display this many functions"),
1004         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1005                     "hide user symbols"),
1006         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1007         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1008         OPT_INCR('v', "verbose", &verbose,
1009                     "be more verbose (show counter open errors, etc)"),
1010         OPT_END()
1011 };
1012
1013 int cmd_top(int argc, const char **argv, const char *prefix __used)
1014 {
1015         struct perf_evsel *pos;
1016         int status = -ENOMEM;
1017
1018         top.evlist = perf_evlist__new(NULL, NULL);
1019         if (top.evlist == NULL)
1020                 return -ENOMEM;
1021
1022         page_size = sysconf(_SC_PAGE_SIZE);
1023
1024         argc = parse_options(argc, argv, options, top_usage, 0);
1025         if (argc)
1026                 usage_with_options(top_usage, options);
1027
1028         /*
1029          * XXX For now start disabled, only using TUI if explicitely asked for.
1030          * Change that when handle_keys equivalent gets written, live annotation
1031          * done, etc.
1032          */
1033         use_browser = 0;
1034
1035         if (use_stdio)
1036                 use_browser = 0;
1037         else if (use_tui)
1038                 use_browser = 1;
1039
1040         setup_browser(false);
1041
1042         /* CPU and PID are mutually exclusive */
1043         if (top.target_tid > 0 && top.cpu_list) {
1044                 printf("WARNING: PID switch overriding CPU\n");
1045                 sleep(1);
1046                 top.cpu_list = NULL;
1047         }
1048
1049         if (top.target_pid != -1)
1050                 top.target_tid = top.target_pid;
1051
1052         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1053                                      top.target_tid, top.cpu_list) < 0)
1054                 usage_with_options(top_usage, options);
1055
1056         if (!top.evlist->nr_entries &&
1057             perf_evlist__add_default(top.evlist) < 0) {
1058                 pr_err("Not enough memory for event selector list\n");
1059                 return -ENOMEM;
1060         }
1061
1062         if (top.delay_secs < 1)
1063                 top.delay_secs = 1;
1064
1065         /*
1066          * User specified count overrides default frequency.
1067          */
1068         if (default_interval)
1069                 top.freq = 0;
1070         else if (top.freq) {
1071                 default_interval = top.freq;
1072         } else {
1073                 fprintf(stderr, "frequency and count are zero, aborting\n");
1074                 exit(EXIT_FAILURE);
1075         }
1076
1077         list_for_each_entry(pos, &top.evlist->entries, node) {
1078                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1079                                          top.evlist->threads->nr) < 0)
1080                         goto out_free_fd;
1081                 /*
1082                  * Fill in the ones not specifically initialized via -c:
1083                  */
1084                 if (pos->attr.sample_period)
1085                         continue;
1086
1087                 pos->attr.sample_period = default_interval;
1088         }
1089
1090         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1091             perf_evlist__alloc_mmap(top.evlist) < 0)
1092                 goto out_free_fd;
1093
1094         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1095
1096         symbol_conf.priv_size = (sizeof(struct sym_entry) + sizeof(struct annotation) +
1097                                  (top.evlist->nr_entries + 1) * sizeof(unsigned long));
1098
1099         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1100         if (symbol__init() < 0)
1101                 return -1;
1102
1103         get_term_dimensions(&winsize);
1104         if (top.print_entries == 0) {
1105                 update_print_entries(&winsize);
1106                 signal(SIGWINCH, sig_winch_handler);
1107         }
1108
1109         status = __cmd_top();
1110 out_free_fd:
1111         perf_evlist__delete(top.evlist);
1112
1113         return status;
1114 }