perf top: Handle kptr_restrict
[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                     kptr_restrict_warned;
84 static bool                     inherit                         =  false;
85 static int                      realtime_prio                   =      0;
86 static bool                     group                           =  false;
87 static unsigned int             page_size;
88 static unsigned int             mmap_pages                      =    128;
89
90 static bool                     dump_symtab                     =  false;
91
92 static struct winsize           winsize;
93
94 static const char               *sym_filter                     =   NULL;
95 struct sym_entry                *sym_filter_entry_sched         =   NULL;
96 static int                      sym_pcnt_filter                 =      5;
97
98 /*
99  * Source functions
100  */
101
102 void get_term_dimensions(struct winsize *ws)
103 {
104         char *s = getenv("LINES");
105
106         if (s != NULL) {
107                 ws->ws_row = atoi(s);
108                 s = getenv("COLUMNS");
109                 if (s != NULL) {
110                         ws->ws_col = atoi(s);
111                         if (ws->ws_row && ws->ws_col)
112                                 return;
113                 }
114         }
115 #ifdef TIOCGWINSZ
116         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
117             ws->ws_row && ws->ws_col)
118                 return;
119 #endif
120         ws->ws_row = 25;
121         ws->ws_col = 80;
122 }
123
124 static void update_print_entries(struct winsize *ws)
125 {
126         top.print_entries = ws->ws_row;
127
128         if (top.print_entries > 9)
129                 top.print_entries -= 9;
130 }
131
132 static void sig_winch_handler(int sig __used)
133 {
134         get_term_dimensions(&winsize);
135         update_print_entries(&winsize);
136 }
137
138 static int parse_source(struct sym_entry *syme)
139 {
140         struct symbol *sym;
141         struct annotation *notes;
142         struct map *map;
143         int err = -1;
144
145         if (!syme)
146                 return -1;
147
148         sym = sym_entry__symbol(syme);
149         map = syme->map;
150
151         /*
152          * We can't annotate with just /proc/kallsyms
153          */
154         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
155                 pr_err("Can't annotate %s: No vmlinux file was found in the "
156                        "path\n", sym->name);
157                 sleep(1);
158                 return -1;
159         }
160
161         notes = symbol__annotation(sym);
162         if (notes->src != NULL) {
163                 pthread_mutex_lock(&notes->lock);
164                 goto out_assign;
165         }
166
167         pthread_mutex_lock(&notes->lock);
168
169         if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
170                 pthread_mutex_unlock(&notes->lock);
171                 pr_err("Not enough memory for annotating '%s' symbol!\n",
172                        sym->name);
173                 sleep(1);
174                 return err;
175         }
176
177         err = symbol__annotate(sym, syme->map, 0);
178         if (err == 0) {
179 out_assign:
180                 top.sym_filter_entry = syme;
181         }
182
183         pthread_mutex_unlock(&notes->lock);
184         return err;
185 }
186
187 static void __zero_source_counters(struct sym_entry *syme)
188 {
189         struct symbol *sym = sym_entry__symbol(syme);
190         symbol__annotate_zero_histograms(sym);
191 }
192
193 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
194 {
195         struct annotation *notes;
196         struct symbol *sym;
197
198         if (syme != top.sym_filter_entry)
199                 return;
200
201         sym = sym_entry__symbol(syme);
202         notes = symbol__annotation(sym);
203
204         if (pthread_mutex_trylock(&notes->lock))
205                 return;
206
207         ip = syme->map->map_ip(syme->map, ip);
208         symbol__inc_addr_samples(sym, syme->map, counter, ip);
209
210         pthread_mutex_unlock(&notes->lock);
211 }
212
213 static void show_details(struct sym_entry *syme)
214 {
215         struct annotation *notes;
216         struct symbol *symbol;
217         int more;
218
219         if (!syme)
220                 return;
221
222         symbol = sym_entry__symbol(syme);
223         notes = symbol__annotation(symbol);
224
225         pthread_mutex_lock(&notes->lock);
226
227         if (notes->src == NULL)
228                 goto out_unlock;
229
230         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
231         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
232
233         more = symbol__annotate_printf(symbol, syme->map, top.sym_evsel->idx,
234                                        0, sym_pcnt_filter, top.print_entries, 4);
235         if (top.zero)
236                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
237         else
238                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
239         if (more != 0)
240                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
241 out_unlock:
242         pthread_mutex_unlock(&notes->lock);
243 }
244
245 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
246
247 static void __list_insert_active_sym(struct sym_entry *syme)
248 {
249         list_add(&syme->node, &top.active_symbols);
250 }
251
252 static void print_sym_table(struct perf_session *session)
253 {
254         char bf[160];
255         int printed = 0;
256         struct rb_node *nd;
257         struct sym_entry *syme;
258         struct rb_root tmp = RB_ROOT;
259         const int win_width = winsize.ws_col - 1;
260         int sym_width, dso_width, dso_short_width;
261         float sum_ksamples = perf_top__decay_samples(&top, &tmp);
262
263         puts(CONSOLE_CLEAR);
264
265         perf_top__header_snprintf(&top, bf, sizeof(bf));
266         printf("%s\n", bf);
267
268         perf_top__reset_sample_counters(&top);
269
270         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
271
272         if (session->hists.stats.total_lost != 0) {
273                 color_fprintf(stdout, PERF_COLOR_RED, "WARNING:");
274                 printf(" LOST %" PRIu64 " events, Check IO/CPU overload\n",
275                        session->hists.stats.total_lost);
276         }
277
278         if (top.sym_filter_entry) {
279                 show_details(top.sym_filter_entry);
280                 return;
281         }
282
283         perf_top__find_widths(&top, &tmp, &dso_width, &dso_short_width,
284                               &sym_width);
285
286         if (sym_width + dso_width > winsize.ws_col - 29) {
287                 dso_width = dso_short_width;
288                 if (sym_width + dso_width > winsize.ws_col - 29)
289                         sym_width = winsize.ws_col - dso_width - 29;
290         }
291         putchar('\n');
292         if (top.evlist->nr_entries == 1)
293                 printf("             samples  pcnt");
294         else
295                 printf("   weight    samples  pcnt");
296
297         if (verbose)
298                 printf("         RIP       ");
299         printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
300         printf("   %s    _______ _____",
301                top.evlist->nr_entries == 1 ? "      " : "______");
302         if (verbose)
303                 printf(" ________________");
304         printf(" %-*.*s", sym_width, sym_width, graph_line);
305         printf(" %-*.*s", dso_width, dso_width, graph_line);
306         puts("\n");
307
308         for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
309                 struct symbol *sym;
310                 double pcnt;
311
312                 syme = rb_entry(nd, struct sym_entry, rb_node);
313                 sym = sym_entry__symbol(syme);
314                 if (++printed > top.print_entries ||
315                     (int)syme->snap_count < top.count_filter)
316                         continue;
317
318                 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
319                                          sum_ksamples));
320
321                 if (top.evlist->nr_entries == 1 || !top.display_weighted)
322                         printf("%20.2f ", syme->weight);
323                 else
324                         printf("%9.1f %10ld ", syme->weight, syme->snap_count);
325
326                 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
327                 if (verbose)
328                         printf(" %016" PRIx64, sym->start);
329                 printf(" %-*.*s", sym_width, sym_width, sym->name);
330                 printf(" %-*.*s\n", dso_width, dso_width,
331                        dso_width >= syme->map->dso->long_name_len ?
332                                         syme->map->dso->long_name :
333                                         syme->map->dso->short_name);
334         }
335 }
336
337 static void prompt_integer(int *target, const char *msg)
338 {
339         char *buf = malloc(0), *p;
340         size_t dummy = 0;
341         int tmp;
342
343         fprintf(stdout, "\n%s: ", msg);
344         if (getline(&buf, &dummy, stdin) < 0)
345                 return;
346
347         p = strchr(buf, '\n');
348         if (p)
349                 *p = 0;
350
351         p = buf;
352         while(*p) {
353                 if (!isdigit(*p))
354                         goto out_free;
355                 p++;
356         }
357         tmp = strtoul(buf, NULL, 10);
358         *target = tmp;
359 out_free:
360         free(buf);
361 }
362
363 static void prompt_percent(int *target, const char *msg)
364 {
365         int tmp = 0;
366
367         prompt_integer(&tmp, msg);
368         if (tmp >= 0 && tmp <= 100)
369                 *target = tmp;
370 }
371
372 static void prompt_symbol(struct sym_entry **target, const char *msg)
373 {
374         char *buf = malloc(0), *p;
375         struct sym_entry *syme = *target, *n, *found = NULL;
376         size_t dummy = 0;
377
378         /* zero counters of active symbol */
379         if (syme) {
380                 __zero_source_counters(syme);
381                 *target = NULL;
382         }
383
384         fprintf(stdout, "\n%s: ", msg);
385         if (getline(&buf, &dummy, stdin) < 0)
386                 goto out_free;
387
388         p = strchr(buf, '\n');
389         if (p)
390                 *p = 0;
391
392         pthread_mutex_lock(&top.active_symbols_lock);
393         syme = list_entry(top.active_symbols.next, struct sym_entry, node);
394         pthread_mutex_unlock(&top.active_symbols_lock);
395
396         list_for_each_entry_safe_from(syme, n, &top.active_symbols, node) {
397                 struct symbol *sym = sym_entry__symbol(syme);
398
399                 if (!strcmp(buf, sym->name)) {
400                         found = syme;
401                         break;
402                 }
403         }
404
405         if (!found) {
406                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
407                 sleep(1);
408                 return;
409         } else
410                 parse_source(found);
411
412 out_free:
413         free(buf);
414 }
415
416 static void print_mapped_keys(void)
417 {
418         char *name = NULL;
419
420         if (top.sym_filter_entry) {
421                 struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
422                 name = sym->name;
423         }
424
425         fprintf(stdout, "\nMapped keys:\n");
426         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
427         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
428
429         if (top.evlist->nr_entries > 1)
430                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
431
432         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
433
434         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
435         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
436         fprintf(stdout, "\t[S]     stop annotation.\n");
437
438         if (top.evlist->nr_entries > 1)
439                 fprintf(stdout, "\t[w]     toggle display weighted/count[E]r. \t(%d)\n", top.display_weighted ? 1 : 0);
440
441         fprintf(stdout,
442                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
443                 top.hide_kernel_symbols ? "yes" : "no");
444         fprintf(stdout,
445                 "\t[U]     hide user symbols.               \t(%s)\n",
446                 top.hide_user_symbols ? "yes" : "no");
447         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
448         fprintf(stdout, "\t[qQ]    quit.\n");
449 }
450
451 static int key_mapped(int c)
452 {
453         switch (c) {
454                 case 'd':
455                 case 'e':
456                 case 'f':
457                 case 'z':
458                 case 'q':
459                 case 'Q':
460                 case 'K':
461                 case 'U':
462                 case 'F':
463                 case 's':
464                 case 'S':
465                         return 1;
466                 case 'E':
467                 case 'w':
468                         return top.evlist->nr_entries > 1 ? 1 : 0;
469                 default:
470                         break;
471         }
472
473         return 0;
474 }
475
476 static void handle_keypress(struct perf_session *session, int c)
477 {
478         if (!key_mapped(c)) {
479                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
480                 struct termios tc, save;
481
482                 print_mapped_keys();
483                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
484                 fflush(stdout);
485
486                 tcgetattr(0, &save);
487                 tc = save;
488                 tc.c_lflag &= ~(ICANON | ECHO);
489                 tc.c_cc[VMIN] = 0;
490                 tc.c_cc[VTIME] = 0;
491                 tcsetattr(0, TCSANOW, &tc);
492
493                 poll(&stdin_poll, 1, -1);
494                 c = getc(stdin);
495
496                 tcsetattr(0, TCSAFLUSH, &save);
497                 if (!key_mapped(c))
498                         return;
499         }
500
501         switch (c) {
502                 case 'd':
503                         prompt_integer(&top.delay_secs, "Enter display delay");
504                         if (top.delay_secs < 1)
505                                 top.delay_secs = 1;
506                         break;
507                 case 'e':
508                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
509                         if (top.print_entries == 0) {
510                                 sig_winch_handler(SIGWINCH);
511                                 signal(SIGWINCH, sig_winch_handler);
512                         } else
513                                 signal(SIGWINCH, SIG_DFL);
514                         break;
515                 case 'E':
516                         if (top.evlist->nr_entries > 1) {
517                                 /* Select 0 as the default event: */
518                                 int counter = 0;
519
520                                 fprintf(stderr, "\nAvailable events:");
521
522                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
523                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
524
525                                 prompt_integer(&counter, "Enter details event counter");
526
527                                 if (counter >= top.evlist->nr_entries) {
528                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
529                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
530                                         sleep(1);
531                                         break;
532                                 }
533                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
534                                         if (top.sym_evsel->idx == counter)
535                                                 break;
536                         } else
537                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
538                         break;
539                 case 'f':
540                         prompt_integer(&top.count_filter, "Enter display event count filter");
541                         break;
542                 case 'F':
543                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
544                         break;
545                 case 'K':
546                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
547                         break;
548                 case 'q':
549                 case 'Q':
550                         printf("exiting.\n");
551                         if (dump_symtab)
552                                 perf_session__fprintf_dsos(session, stderr);
553                         exit(0);
554                 case 's':
555                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
556                         break;
557                 case 'S':
558                         if (!top.sym_filter_entry)
559                                 break;
560                         else {
561                                 struct sym_entry *syme = top.sym_filter_entry;
562
563                                 top.sym_filter_entry = NULL;
564                                 __zero_source_counters(syme);
565                         }
566                         break;
567                 case 'U':
568                         top.hide_user_symbols = !top.hide_user_symbols;
569                         break;
570                 case 'w':
571                         top.display_weighted = ~top.display_weighted;
572                         break;
573                 case 'z':
574                         top.zero = !top.zero;
575                         break;
576                 default:
577                         break;
578         }
579 }
580
581 static void *display_thread_tui(void *arg __used)
582 {
583         int err = 0;
584         pthread_mutex_lock(&top.active_symbols_lock);
585         while (list_empty(&top.active_symbols)) {
586                 err = pthread_cond_wait(&top.active_symbols_cond,
587                                         &top.active_symbols_lock);
588                 if (err)
589                         break;
590         }
591         pthread_mutex_unlock(&top.active_symbols_lock);
592         if (!err)
593                 perf_top__tui_browser(&top);
594         exit_browser(0);
595         exit(0);
596         return NULL;
597 }
598
599 static void *display_thread(void *arg __used)
600 {
601         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
602         struct termios tc, save;
603         int delay_msecs, c;
604         struct perf_session *session = (struct perf_session *) arg;
605
606         tcgetattr(0, &save);
607         tc = save;
608         tc.c_lflag &= ~(ICANON | ECHO);
609         tc.c_cc[VMIN] = 0;
610         tc.c_cc[VTIME] = 0;
611
612 repeat:
613         delay_msecs = top.delay_secs * 1000;
614         tcsetattr(0, TCSANOW, &tc);
615         /* trash return*/
616         getc(stdin);
617
618         do {
619                 print_sym_table(session);
620         } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
621
622         c = getc(stdin);
623         tcsetattr(0, TCSAFLUSH, &save);
624
625         handle_keypress(session, c);
626         goto repeat;
627
628         return NULL;
629 }
630
631 /* Tag samples to be skipped. */
632 static const char *skip_symbols[] = {
633         "default_idle",
634         "native_safe_halt",
635         "cpu_idle",
636         "enter_idle",
637         "exit_idle",
638         "mwait_idle",
639         "mwait_idle_with_hints",
640         "poll_idle",
641         "ppc64_runlatch_off",
642         "pseries_dedicated_idle_sleep",
643         NULL
644 };
645
646 static int symbol_filter(struct map *map, struct symbol *sym)
647 {
648         struct sym_entry *syme;
649         const char *name = sym->name;
650         int i;
651
652         /*
653          * ppc64 uses function descriptors and appends a '.' to the
654          * start of every instruction address. Remove it.
655          */
656         if (name[0] == '.')
657                 name++;
658
659         if (!strcmp(name, "_text") ||
660             !strcmp(name, "_etext") ||
661             !strcmp(name, "_sinittext") ||
662             !strncmp("init_module", name, 11) ||
663             !strncmp("cleanup_module", name, 14) ||
664             strstr(name, "_text_start") ||
665             strstr(name, "_text_end"))
666                 return 1;
667
668         syme = symbol__priv(sym);
669         syme->map = map;
670         symbol__annotate_init(map, sym);
671
672         if (!top.sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
673                 /* schedule initial sym_filter_entry setup */
674                 sym_filter_entry_sched = syme;
675                 sym_filter = NULL;
676         }
677
678         for (i = 0; skip_symbols[i]; i++) {
679                 if (!strcmp(skip_symbols[i], name)) {
680                         sym->ignore = true;
681                         break;
682                 }
683         }
684
685         return 0;
686 }
687
688 static void perf_event__process_sample(const union perf_event *event,
689                                        struct perf_sample *sample,
690                                        struct perf_session *session)
691 {
692         u64 ip = event->ip.ip;
693         struct sym_entry *syme;
694         struct addr_location al;
695         struct machine *machine;
696         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
697
698         ++top.samples;
699
700         switch (origin) {
701         case PERF_RECORD_MISC_USER:
702                 ++top.us_samples;
703                 if (top.hide_user_symbols)
704                         return;
705                 machine = perf_session__find_host_machine(session);
706                 break;
707         case PERF_RECORD_MISC_KERNEL:
708                 ++top.kernel_samples;
709                 if (top.hide_kernel_symbols)
710                         return;
711                 machine = perf_session__find_host_machine(session);
712                 break;
713         case PERF_RECORD_MISC_GUEST_KERNEL:
714                 ++top.guest_kernel_samples;
715                 machine = perf_session__find_machine(session, event->ip.pid);
716                 break;
717         case PERF_RECORD_MISC_GUEST_USER:
718                 ++top.guest_us_samples;
719                 /*
720                  * TODO: we don't process guest user from host side
721                  * except simple counting.
722                  */
723                 return;
724         default:
725                 return;
726         }
727
728         if (!machine && perf_guest) {
729                 pr_err("Can't find guest [%d]'s kernel information\n",
730                         event->ip.pid);
731                 return;
732         }
733
734         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
735                 top.exact_samples++;
736
737         if (perf_event__preprocess_sample(event, session, &al, sample,
738                                           symbol_filter) < 0 ||
739             al.filtered)
740                 return;
741
742         if (!kptr_restrict_warned &&
743             symbol_conf.kptr_restrict &&
744             al.cpumode == PERF_RECORD_MISC_KERNEL) {
745                 ui__warning(
746 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
747 "Check /proc/sys/kernel/kptr_restrict.\n\n"
748 "Kernel%s samples will not be resolved.\n",
749                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
750                           " modules" : "");
751                 if (use_browser <= 0)
752                         sleep(5);
753                 kptr_restrict_warned = true;
754         }
755
756         if (al.sym == NULL) {
757                 /*
758                  * As we do lazy loading of symtabs we only will know if the
759                  * specified vmlinux file is invalid when we actually have a
760                  * hit in kernel space and then try to load it. So if we get
761                  * here and there are _no_ symbols in the DSO backing the
762                  * kernel map, bail out.
763                  *
764                  * We may never get here, for instance, if we use -K/
765                  * --hide-kernel-symbols, even if the user specifies an
766                  * invalid --vmlinux ;-)
767                  */
768                 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
769                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
770                         ui__warning("The %s file can't be used\n",
771                                     symbol_conf.vmlinux_name);
772                         exit_browser(0);
773                         exit(1);
774                 }
775
776                 return;
777         }
778
779         /* let's see, whether we need to install initial sym_filter_entry */
780         if (sym_filter_entry_sched) {
781                 top.sym_filter_entry = sym_filter_entry_sched;
782                 sym_filter_entry_sched = NULL;
783                 if (parse_source(top.sym_filter_entry) < 0) {
784                         struct symbol *sym = sym_entry__symbol(top.sym_filter_entry);
785
786                         pr_err("Can't annotate %s", sym->name);
787                         if (top.sym_filter_entry->map->dso->symtab_type == SYMTAB__KALLSYMS) {
788                                 pr_err(": No vmlinux file was found in the path:\n");
789                                 machine__fprintf_vmlinux_path(machine, stderr);
790                         } else
791                                 pr_err(".\n");
792                         exit(1);
793                 }
794         }
795
796         syme = symbol__priv(al.sym);
797         if (!al.sym->ignore) {
798                 struct perf_evsel *evsel;
799
800                 evsel = perf_evlist__id2evsel(top.evlist, sample->id);
801                 assert(evsel != NULL);
802                 syme->count[evsel->idx]++;
803                 record_precise_ip(syme, evsel->idx, ip);
804                 pthread_mutex_lock(&top.active_symbols_lock);
805                 if (list_empty(&syme->node) || !syme->node.next) {
806                         static bool first = true;
807                         __list_insert_active_sym(syme);
808                         if (first) {
809                                 pthread_cond_broadcast(&top.active_symbols_cond);
810                                 first = false;
811                         }
812                 }
813                 pthread_mutex_unlock(&top.active_symbols_lock);
814         }
815 }
816
817 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
818 {
819         struct perf_sample sample;
820         union perf_event *event;
821         int ret;
822
823         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
824                 ret = perf_session__parse_sample(self, event, &sample);
825                 if (ret) {
826                         pr_err("Can't parse sample, err = %d\n", ret);
827                         continue;
828                 }
829
830                 if (event->header.type == PERF_RECORD_SAMPLE)
831                         perf_event__process_sample(event, &sample, self);
832                 else
833                         perf_event__process(event, &sample, self);
834         }
835 }
836
837 static void perf_session__mmap_read(struct perf_session *self)
838 {
839         int i;
840
841         for (i = 0; i < top.evlist->nr_mmaps; i++)
842                 perf_session__mmap_read_idx(self, i);
843 }
844
845 static void start_counters(struct perf_evlist *evlist)
846 {
847         struct perf_evsel *counter;
848
849         list_for_each_entry(counter, &evlist->entries, node) {
850                 struct perf_event_attr *attr = &counter->attr;
851
852                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
853
854                 if (top.freq) {
855                         attr->sample_type |= PERF_SAMPLE_PERIOD;
856                         attr->freq        = 1;
857                         attr->sample_freq = top.freq;
858                 }
859
860                 if (evlist->nr_entries > 1) {
861                         attr->sample_type |= PERF_SAMPLE_ID;
862                         attr->read_format |= PERF_FORMAT_ID;
863                 }
864
865                 attr->mmap = 1;
866                 attr->inherit = inherit;
867 try_again:
868                 if (perf_evsel__open(counter, top.evlist->cpus,
869                                      top.evlist->threads, group) < 0) {
870                         int err = errno;
871
872                         if (err == EPERM || err == EACCES) {
873                                 ui__warning_paranoid();
874                                 goto out_err;
875                         }
876                         /*
877                          * If it's cycles then fall back to hrtimer
878                          * based cpu-clock-tick sw counter, which
879                          * is always available even if no PMU support:
880                          */
881                         if (attr->type == PERF_TYPE_HARDWARE &&
882                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
883                                 if (verbose)
884                                         ui__warning("Cycles event not supported,\n"
885                                                     "trying to fall back to cpu-clock-ticks\n");
886
887                                 attr->type = PERF_TYPE_SOFTWARE;
888                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
889                                 goto try_again;
890                         }
891
892                         if (err == ENOENT) {
893                                 ui__warning("The %s event is not supported.\n",
894                                             event_name(counter));
895                                 goto out_err;
896                         }
897
898                         ui__warning("The sys_perf_event_open() syscall "
899                                     "returned with %d (%s).  /bin/dmesg "
900                                     "may provide additional information.\n"
901                                     "No CONFIG_PERF_EVENTS=y kernel support "
902                                     "configured?\n", err, strerror(err));
903                         goto out_err;
904                 }
905         }
906
907         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
908                 ui__warning("Failed to mmap with %d (%s)\n",
909                             errno, strerror(errno));
910                 goto out_err;
911         }
912
913         return;
914
915 out_err:
916         exit_browser(0);
917         exit(0);
918 }
919
920 static int __cmd_top(void)
921 {
922         pthread_t thread;
923         int ret __used;
924         /*
925          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
926          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
927          */
928         struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
929         if (session == NULL)
930                 return -ENOMEM;
931
932         if (top.target_tid != -1)
933                 perf_event__synthesize_thread_map(top.evlist->threads,
934                                                   perf_event__process, session);
935         else
936                 perf_event__synthesize_threads(perf_event__process, session);
937
938         start_counters(top.evlist);
939         session->evlist = top.evlist;
940         perf_session__update_sample_type(session);
941
942         /* Wait for a minimal set of events before starting the snapshot */
943         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
944
945         perf_session__mmap_read(session);
946
947         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
948                                                              display_thread), session)) {
949                 printf("Could not create display thread.\n");
950                 exit(-1);
951         }
952
953         if (realtime_prio) {
954                 struct sched_param param;
955
956                 param.sched_priority = realtime_prio;
957                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
958                         printf("Could not set realtime priority.\n");
959                         exit(-1);
960                 }
961         }
962
963         while (1) {
964                 u64 hits = top.samples;
965
966                 perf_session__mmap_read(session);
967
968                 if (hits == top.samples)
969                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
970         }
971
972         return 0;
973 }
974
975 static const char * const top_usage[] = {
976         "perf top [<options>]",
977         NULL
978 };
979
980 static const struct option options[] = {
981         OPT_CALLBACK('e', "event", &top.evlist, "event",
982                      "event selector. use 'perf list' to list available events",
983                      parse_events),
984         OPT_INTEGER('c', "count", &default_interval,
985                     "event period to sample"),
986         OPT_INTEGER('p', "pid", &top.target_pid,
987                     "profile events on existing process id"),
988         OPT_INTEGER('t', "tid", &top.target_tid,
989                     "profile events on existing thread id"),
990         OPT_BOOLEAN('a', "all-cpus", &system_wide,
991                             "system-wide collection from all CPUs"),
992         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
993                     "list of cpus to monitor"),
994         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
995                    "file", "vmlinux pathname"),
996         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
997                     "hide kernel symbols"),
998         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
999         OPT_INTEGER('r', "realtime", &realtime_prio,
1000                     "collect data with this RT SCHED_FIFO priority"),
1001         OPT_INTEGER('d', "delay", &top.delay_secs,
1002                     "number of seconds to delay between refreshes"),
1003         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1004                             "dump the symbol table used for profiling"),
1005         OPT_INTEGER('f', "count-filter", &top.count_filter,
1006                     "only display functions with more events than this"),
1007         OPT_BOOLEAN('g', "group", &group,
1008                             "put the counters into a counter group"),
1009         OPT_BOOLEAN('i', "inherit", &inherit,
1010                     "child tasks inherit counters"),
1011         OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1012                     "symbol to annotate"),
1013         OPT_BOOLEAN('z', "zero", &top.zero,
1014                     "zero history across updates"),
1015         OPT_INTEGER('F', "freq", &top.freq,
1016                     "profile at this frequency"),
1017         OPT_INTEGER('E', "entries", &top.print_entries,
1018                     "display this many functions"),
1019         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1020                     "hide user symbols"),
1021         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1022         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1023         OPT_INCR('v', "verbose", &verbose,
1024                     "be more verbose (show counter open errors, etc)"),
1025         OPT_END()
1026 };
1027
1028 int cmd_top(int argc, const char **argv, const char *prefix __used)
1029 {
1030         struct perf_evsel *pos;
1031         int status = -ENOMEM;
1032
1033         top.evlist = perf_evlist__new(NULL, NULL);
1034         if (top.evlist == NULL)
1035                 return -ENOMEM;
1036
1037         page_size = sysconf(_SC_PAGE_SIZE);
1038
1039         argc = parse_options(argc, argv, options, top_usage, 0);
1040         if (argc)
1041                 usage_with_options(top_usage, options);
1042
1043         /*
1044          * XXX For now start disabled, only using TUI if explicitely asked for.
1045          * Change that when handle_keys equivalent gets written, live annotation
1046          * done, etc.
1047          */
1048         use_browser = 0;
1049
1050         if (use_stdio)
1051                 use_browser = 0;
1052         else if (use_tui)
1053                 use_browser = 1;
1054
1055         setup_browser(false);
1056
1057         /* CPU and PID are mutually exclusive */
1058         if (top.target_tid > 0 && top.cpu_list) {
1059                 printf("WARNING: PID switch overriding CPU\n");
1060                 sleep(1);
1061                 top.cpu_list = NULL;
1062         }
1063
1064         if (top.target_pid != -1)
1065                 top.target_tid = top.target_pid;
1066
1067         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1068                                      top.target_tid, top.cpu_list) < 0)
1069                 usage_with_options(top_usage, options);
1070
1071         if (!top.evlist->nr_entries &&
1072             perf_evlist__add_default(top.evlist) < 0) {
1073                 pr_err("Not enough memory for event selector list\n");
1074                 return -ENOMEM;
1075         }
1076
1077         if (top.delay_secs < 1)
1078                 top.delay_secs = 1;
1079
1080         /*
1081          * User specified count overrides default frequency.
1082          */
1083         if (default_interval)
1084                 top.freq = 0;
1085         else if (top.freq) {
1086                 default_interval = top.freq;
1087         } else {
1088                 fprintf(stderr, "frequency and count are zero, aborting\n");
1089                 exit(EXIT_FAILURE);
1090         }
1091
1092         list_for_each_entry(pos, &top.evlist->entries, node) {
1093                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1094                                          top.evlist->threads->nr) < 0)
1095                         goto out_free_fd;
1096                 /*
1097                  * Fill in the ones not specifically initialized via -c:
1098                  */
1099                 if (pos->attr.sample_period)
1100                         continue;
1101
1102                 pos->attr.sample_period = default_interval;
1103         }
1104
1105         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1106             perf_evlist__alloc_mmap(top.evlist) < 0)
1107                 goto out_free_fd;
1108
1109         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1110
1111         symbol_conf.priv_size = (sizeof(struct sym_entry) + sizeof(struct annotation) +
1112                                  (top.evlist->nr_entries + 1) * sizeof(unsigned long));
1113
1114         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1115         if (symbol__init() < 0)
1116                 return -1;
1117
1118         get_term_dimensions(&winsize);
1119         if (top.print_entries == 0) {
1120                 update_print_entries(&winsize);
1121                 signal(SIGWINCH, sig_winch_handler);
1122         }
1123
1124         status = __cmd_top();
1125 out_free_fd:
1126         perf_evlist__delete(top.evlist);
1127
1128         return status;
1129 }