Merge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[~shefty/rdma-dev.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  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
9  *
10  * Improvements and fixes by:
11  *
12  *   Arjan van de Ven <arjan@linux.intel.com>
13  *   Yanmin Zhang <yanmin.zhang@intel.com>
14  *   Wu Fengguang <fengguang.wu@intel.com>
15  *   Mike Galbraith <efault@gmx.de>
16  *   Paul Mackerras <paulus@samba.org>
17  *
18  * Released under the GPL v2. (and only v2, not any later version)
19  */
20 #include "builtin.h"
21
22 #include "perf.h"
23
24 #include "util/annotate.h"
25 #include "util/cache.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/session.h"
30 #include "util/symbol.h"
31 #include "util/thread.h"
32 #include "util/thread_map.h"
33 #include "util/top.h"
34 #include "util/util.h"
35 #include <linux/rbtree.h>
36 #include "util/parse-options.h"
37 #include "util/parse-events.h"
38 #include "util/cpumap.h"
39 #include "util/xyarray.h"
40 #include "util/sort.h"
41
42 #include "util/debug.h"
43
44 #include <assert.h>
45 #include <fcntl.h>
46
47 #include <stdio.h>
48 #include <termios.h>
49 #include <unistd.h>
50 #include <inttypes.h>
51
52 #include <errno.h>
53 #include <time.h>
54 #include <sched.h>
55
56 #include <sys/syscall.h>
57 #include <sys/ioctl.h>
58 #include <sys/poll.h>
59 #include <sys/prctl.h>
60 #include <sys/wait.h>
61 #include <sys/uio.h>
62 #include <sys/mman.h>
63
64 #include <linux/unistd.h>
65 #include <linux/types.h>
66
67 static struct perf_top top = {
68         .count_filter           = 5,
69         .delay_secs             = 2,
70         .target_pid             = -1,
71         .target_tid             = -1,
72         .freq                   = 1000, /* 1 KHz */
73 };
74
75 static bool                     system_wide                     =  false;
76
77 static bool                     use_tui, use_stdio;
78
79 static bool                     sort_has_symbols;
80
81 static bool                     dont_use_callchains;
82 static char                     callchain_default_opt[]         = "fractal,0.5,callee";
83
84
85 static int                      default_interval                =      0;
86
87 static bool                     kptr_restrict_warned;
88 static bool                     vmlinux_warned;
89 static bool                     inherit                         =  false;
90 static int                      realtime_prio                   =      0;
91 static bool                     group                           =  false;
92 static bool                     sample_id_all_avail             =   true;
93 static unsigned int             mmap_pages                      =    128;
94
95 static bool                     dump_symtab                     =  false;
96
97 static struct winsize           winsize;
98
99 static const char               *sym_filter                     =   NULL;
100 static int                      sym_pcnt_filter                 =      5;
101
102 /*
103  * Source functions
104  */
105
106 void get_term_dimensions(struct winsize *ws)
107 {
108         char *s = getenv("LINES");
109
110         if (s != NULL) {
111                 ws->ws_row = atoi(s);
112                 s = getenv("COLUMNS");
113                 if (s != NULL) {
114                         ws->ws_col = atoi(s);
115                         if (ws->ws_row && ws->ws_col)
116                                 return;
117                 }
118         }
119 #ifdef TIOCGWINSZ
120         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
121             ws->ws_row && ws->ws_col)
122                 return;
123 #endif
124         ws->ws_row = 25;
125         ws->ws_col = 80;
126 }
127
128 static void update_print_entries(struct winsize *ws)
129 {
130         top.print_entries = ws->ws_row;
131
132         if (top.print_entries > 9)
133                 top.print_entries -= 9;
134 }
135
136 static void sig_winch_handler(int sig __used)
137 {
138         get_term_dimensions(&winsize);
139         update_print_entries(&winsize);
140 }
141
142 static int parse_source(struct hist_entry *he)
143 {
144         struct symbol *sym;
145         struct annotation *notes;
146         struct map *map;
147         int err = -1;
148
149         if (!he || !he->ms.sym)
150                 return -1;
151
152         sym = he->ms.sym;
153         map = he->ms.map;
154
155         /*
156          * We can't annotate with just /proc/kallsyms
157          */
158         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
159                 pr_err("Can't annotate %s: No vmlinux file was found in the "
160                        "path\n", sym->name);
161                 sleep(1);
162                 return -1;
163         }
164
165         notes = symbol__annotation(sym);
166         if (notes->src != NULL) {
167                 pthread_mutex_lock(&notes->lock);
168                 goto out_assign;
169         }
170
171         pthread_mutex_lock(&notes->lock);
172
173         if (symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
174                 pthread_mutex_unlock(&notes->lock);
175                 pr_err("Not enough memory for annotating '%s' symbol!\n",
176                        sym->name);
177                 sleep(1);
178                 return err;
179         }
180
181         err = symbol__annotate(sym, map, 0);
182         if (err == 0) {
183 out_assign:
184                 top.sym_filter_entry = he;
185         }
186
187         pthread_mutex_unlock(&notes->lock);
188         return err;
189 }
190
191 static void __zero_source_counters(struct hist_entry *he)
192 {
193         struct symbol *sym = he->ms.sym;
194         symbol__annotate_zero_histograms(sym);
195 }
196
197 static void record_precise_ip(struct hist_entry *he, int counter, u64 ip)
198 {
199         struct annotation *notes;
200         struct symbol *sym;
201
202         if (he == NULL || he->ms.sym == NULL ||
203             ((top.sym_filter_entry == NULL ||
204               top.sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
205                 return;
206
207         sym = he->ms.sym;
208         notes = symbol__annotation(sym);
209
210         if (pthread_mutex_trylock(&notes->lock))
211                 return;
212
213         if (notes->src == NULL &&
214             symbol__alloc_hist(sym, top.evlist->nr_entries) < 0) {
215                 pthread_mutex_unlock(&notes->lock);
216                 pr_err("Not enough memory for annotating '%s' symbol!\n",
217                        sym->name);
218                 sleep(1);
219                 return;
220         }
221
222         ip = he->ms.map->map_ip(he->ms.map, ip);
223         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
224
225         pthread_mutex_unlock(&notes->lock);
226 }
227
228 static void show_details(struct hist_entry *he)
229 {
230         struct annotation *notes;
231         struct symbol *symbol;
232         int more;
233
234         if (!he)
235                 return;
236
237         symbol = he->ms.sym;
238         notes = symbol__annotation(symbol);
239
240         pthread_mutex_lock(&notes->lock);
241
242         if (notes->src == NULL)
243                 goto out_unlock;
244
245         printf("Showing %s for %s\n", event_name(top.sym_evsel), symbol->name);
246         printf("  Events  Pcnt (>=%d%%)\n", sym_pcnt_filter);
247
248         more = symbol__annotate_printf(symbol, he->ms.map, top.sym_evsel->idx,
249                                        0, sym_pcnt_filter, top.print_entries, 4);
250         if (top.zero)
251                 symbol__annotate_zero_histogram(symbol, top.sym_evsel->idx);
252         else
253                 symbol__annotate_decay_histogram(symbol, top.sym_evsel->idx);
254         if (more != 0)
255                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
256 out_unlock:
257         pthread_mutex_unlock(&notes->lock);
258 }
259
260 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
261
262 static struct hist_entry *
263         perf_session__add_hist_entry(struct perf_session *session,
264                                      struct addr_location *al,
265                                      struct perf_sample *sample,
266                                      struct perf_evsel *evsel)
267 {
268         struct hist_entry *he;
269
270         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
271         if (he == NULL)
272                 return NULL;
273
274         session->hists.stats.total_period += sample->period;
275         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
276         return he;
277 }
278
279 static void print_sym_table(void)
280 {
281         char bf[160];
282         int printed = 0;
283         const int win_width = winsize.ws_col - 1;
284
285         puts(CONSOLE_CLEAR);
286
287         perf_top__header_snprintf(&top, bf, sizeof(bf));
288         printf("%s\n", bf);
289
290         perf_top__reset_sample_counters(&top);
291
292         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
293
294         if (top.sym_evsel->hists.stats.nr_lost_warned !=
295             top.sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
296                 top.sym_evsel->hists.stats.nr_lost_warned =
297                         top.sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
298                 color_fprintf(stdout, PERF_COLOR_RED,
299                               "WARNING: LOST %d chunks, Check IO/CPU overload",
300                               top.sym_evsel->hists.stats.nr_lost_warned);
301                 ++printed;
302         }
303
304         if (top.sym_filter_entry) {
305                 show_details(top.sym_filter_entry);
306                 return;
307         }
308
309         hists__collapse_resort_threaded(&top.sym_evsel->hists);
310         hists__output_resort_threaded(&top.sym_evsel->hists);
311         hists__decay_entries_threaded(&top.sym_evsel->hists,
312                                       top.hide_user_symbols,
313                                       top.hide_kernel_symbols);
314         hists__output_recalc_col_len(&top.sym_evsel->hists, winsize.ws_row - 3);
315         putchar('\n');
316         hists__fprintf(&top.sym_evsel->hists, NULL, false, false,
317                        winsize.ws_row - 4 - printed, win_width, stdout);
318 }
319
320 static void prompt_integer(int *target, const char *msg)
321 {
322         char *buf = malloc(0), *p;
323         size_t dummy = 0;
324         int tmp;
325
326         fprintf(stdout, "\n%s: ", msg);
327         if (getline(&buf, &dummy, stdin) < 0)
328                 return;
329
330         p = strchr(buf, '\n');
331         if (p)
332                 *p = 0;
333
334         p = buf;
335         while(*p) {
336                 if (!isdigit(*p))
337                         goto out_free;
338                 p++;
339         }
340         tmp = strtoul(buf, NULL, 10);
341         *target = tmp;
342 out_free:
343         free(buf);
344 }
345
346 static void prompt_percent(int *target, const char *msg)
347 {
348         int tmp = 0;
349
350         prompt_integer(&tmp, msg);
351         if (tmp >= 0 && tmp <= 100)
352                 *target = tmp;
353 }
354
355 static void prompt_symbol(struct hist_entry **target, const char *msg)
356 {
357         char *buf = malloc(0), *p;
358         struct hist_entry *syme = *target, *n, *found = NULL;
359         struct rb_node *next;
360         size_t dummy = 0;
361
362         /* zero counters of active symbol */
363         if (syme) {
364                 __zero_source_counters(syme);
365                 *target = NULL;
366         }
367
368         fprintf(stdout, "\n%s: ", msg);
369         if (getline(&buf, &dummy, stdin) < 0)
370                 goto out_free;
371
372         p = strchr(buf, '\n');
373         if (p)
374                 *p = 0;
375
376         next = rb_first(&top.sym_evsel->hists.entries);
377         while (next) {
378                 n = rb_entry(next, struct hist_entry, rb_node);
379                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
380                         found = n;
381                         break;
382                 }
383                 next = rb_next(&n->rb_node);
384         }
385
386         if (!found) {
387                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
388                 sleep(1);
389                 return;
390         } else
391                 parse_source(found);
392
393 out_free:
394         free(buf);
395 }
396
397 static void print_mapped_keys(void)
398 {
399         char *name = NULL;
400
401         if (top.sym_filter_entry) {
402                 struct symbol *sym = top.sym_filter_entry->ms.sym;
403                 name = sym->name;
404         }
405
406         fprintf(stdout, "\nMapped keys:\n");
407         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top.delay_secs);
408         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top.print_entries);
409
410         if (top.evlist->nr_entries > 1)
411                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top.sym_evsel));
412
413         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top.count_filter);
414
415         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
416         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
417         fprintf(stdout, "\t[S]     stop annotation.\n");
418
419         fprintf(stdout,
420                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
421                 top.hide_kernel_symbols ? "yes" : "no");
422         fprintf(stdout,
423                 "\t[U]     hide user symbols.               \t(%s)\n",
424                 top.hide_user_symbols ? "yes" : "no");
425         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top.zero ? 1 : 0);
426         fprintf(stdout, "\t[qQ]    quit.\n");
427 }
428
429 static int key_mapped(int c)
430 {
431         switch (c) {
432                 case 'd':
433                 case 'e':
434                 case 'f':
435                 case 'z':
436                 case 'q':
437                 case 'Q':
438                 case 'K':
439                 case 'U':
440                 case 'F':
441                 case 's':
442                 case 'S':
443                         return 1;
444                 case 'E':
445                         return top.evlist->nr_entries > 1 ? 1 : 0;
446                 default:
447                         break;
448         }
449
450         return 0;
451 }
452
453 static void handle_keypress(int c)
454 {
455         if (!key_mapped(c)) {
456                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
457                 struct termios tc, save;
458
459                 print_mapped_keys();
460                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
461                 fflush(stdout);
462
463                 tcgetattr(0, &save);
464                 tc = save;
465                 tc.c_lflag &= ~(ICANON | ECHO);
466                 tc.c_cc[VMIN] = 0;
467                 tc.c_cc[VTIME] = 0;
468                 tcsetattr(0, TCSANOW, &tc);
469
470                 poll(&stdin_poll, 1, -1);
471                 c = getc(stdin);
472
473                 tcsetattr(0, TCSAFLUSH, &save);
474                 if (!key_mapped(c))
475                         return;
476         }
477
478         switch (c) {
479                 case 'd':
480                         prompt_integer(&top.delay_secs, "Enter display delay");
481                         if (top.delay_secs < 1)
482                                 top.delay_secs = 1;
483                         break;
484                 case 'e':
485                         prompt_integer(&top.print_entries, "Enter display entries (lines)");
486                         if (top.print_entries == 0) {
487                                 sig_winch_handler(SIGWINCH);
488                                 signal(SIGWINCH, sig_winch_handler);
489                         } else
490                                 signal(SIGWINCH, SIG_DFL);
491                         break;
492                 case 'E':
493                         if (top.evlist->nr_entries > 1) {
494                                 /* Select 0 as the default event: */
495                                 int counter = 0;
496
497                                 fprintf(stderr, "\nAvailable events:");
498
499                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
500                                         fprintf(stderr, "\n\t%d %s", top.sym_evsel->idx, event_name(top.sym_evsel));
501
502                                 prompt_integer(&counter, "Enter details event counter");
503
504                                 if (counter >= top.evlist->nr_entries) {
505                                         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
506                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top.sym_evsel));
507                                         sleep(1);
508                                         break;
509                                 }
510                                 list_for_each_entry(top.sym_evsel, &top.evlist->entries, node)
511                                         if (top.sym_evsel->idx == counter)
512                                                 break;
513                         } else
514                                 top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
515                         break;
516                 case 'f':
517                         prompt_integer(&top.count_filter, "Enter display event count filter");
518                         break;
519                 case 'F':
520                         prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
521                         break;
522                 case 'K':
523                         top.hide_kernel_symbols = !top.hide_kernel_symbols;
524                         break;
525                 case 'q':
526                 case 'Q':
527                         printf("exiting.\n");
528                         if (dump_symtab)
529                                 perf_session__fprintf_dsos(top.session, stderr);
530                         exit(0);
531                 case 's':
532                         prompt_symbol(&top.sym_filter_entry, "Enter details symbol");
533                         break;
534                 case 'S':
535                         if (!top.sym_filter_entry)
536                                 break;
537                         else {
538                                 struct hist_entry *syme = top.sym_filter_entry;
539
540                                 top.sym_filter_entry = NULL;
541                                 __zero_source_counters(syme);
542                         }
543                         break;
544                 case 'U':
545                         top.hide_user_symbols = !top.hide_user_symbols;
546                         break;
547                 case 'z':
548                         top.zero = !top.zero;
549                         break;
550                 default:
551                         break;
552         }
553 }
554
555 static void perf_top__sort_new_samples(void *arg)
556 {
557         struct perf_top *t = arg;
558         perf_top__reset_sample_counters(t);
559
560         if (t->evlist->selected != NULL)
561                 t->sym_evsel = t->evlist->selected;
562
563         hists__collapse_resort_threaded(&t->sym_evsel->hists);
564         hists__output_resort_threaded(&t->sym_evsel->hists);
565         hists__decay_entries_threaded(&t->sym_evsel->hists,
566                                       top.hide_user_symbols,
567                                       top.hide_kernel_symbols);
568 }
569
570 static void *display_thread_tui(void *arg __used)
571 {
572         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
573
574         perf_top__sort_new_samples(&top);
575         perf_evlist__tui_browse_hists(top.evlist, help,
576                                       perf_top__sort_new_samples,
577                                       &top, top.delay_secs);
578
579         exit_browser(0);
580         exit(0);
581         return NULL;
582 }
583
584 static void *display_thread(void *arg __used)
585 {
586         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
587         struct termios tc, save;
588         int delay_msecs, c;
589
590         tcgetattr(0, &save);
591         tc = save;
592         tc.c_lflag &= ~(ICANON | ECHO);
593         tc.c_cc[VMIN] = 0;
594         tc.c_cc[VTIME] = 0;
595
596         pthread__unblock_sigwinch();
597 repeat:
598         delay_msecs = top.delay_secs * 1000;
599         tcsetattr(0, TCSANOW, &tc);
600         /* trash return*/
601         getc(stdin);
602
603         while (1) {
604                 print_sym_table();
605                 /*
606                  * Either timeout expired or we got an EINTR due to SIGWINCH,
607                  * refresh screen in both cases.
608                  */
609                 switch (poll(&stdin_poll, 1, delay_msecs)) {
610                 case 0:
611                         continue;
612                 case -1:
613                         if (errno == EINTR)
614                                 continue;
615                         /* Fall trhu */
616                 default:
617                         goto process_hotkey;
618                 }
619         }
620 process_hotkey:
621         c = getc(stdin);
622         tcsetattr(0, TCSAFLUSH, &save);
623
624         handle_keypress(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 __used, struct symbol *sym)
646 {
647         const char *name = sym->name;
648         int i;
649
650         /*
651          * ppc64 uses function descriptors and appends a '.' to the
652          * start of every instruction address. Remove it.
653          */
654         if (name[0] == '.')
655                 name++;
656
657         if (!strcmp(name, "_text") ||
658             !strcmp(name, "_etext") ||
659             !strcmp(name, "_sinittext") ||
660             !strncmp("init_module", name, 11) ||
661             !strncmp("cleanup_module", name, 14) ||
662             strstr(name, "_text_start") ||
663             strstr(name, "_text_end"))
664                 return 1;
665
666         for (i = 0; skip_symbols[i]; i++) {
667                 if (!strcmp(skip_symbols[i], name)) {
668                         sym->ignore = true;
669                         break;
670                 }
671         }
672
673         return 0;
674 }
675
676 static void perf_event__process_sample(const union perf_event *event,
677                                        struct perf_evsel *evsel,
678                                        struct perf_sample *sample,
679                                        struct perf_session *session)
680 {
681         struct symbol *parent = NULL;
682         u64 ip = event->ip.ip;
683         struct addr_location al;
684         struct machine *machine;
685         int err;
686         u8 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
687
688         ++top.samples;
689
690         switch (origin) {
691         case PERF_RECORD_MISC_USER:
692                 ++top.us_samples;
693                 if (top.hide_user_symbols)
694                         return;
695                 machine = perf_session__find_host_machine(session);
696                 break;
697         case PERF_RECORD_MISC_KERNEL:
698                 ++top.kernel_samples;
699                 if (top.hide_kernel_symbols)
700                         return;
701                 machine = perf_session__find_host_machine(session);
702                 break;
703         case PERF_RECORD_MISC_GUEST_KERNEL:
704                 ++top.guest_kernel_samples;
705                 machine = perf_session__find_machine(session, event->ip.pid);
706                 break;
707         case PERF_RECORD_MISC_GUEST_USER:
708                 ++top.guest_us_samples;
709                 /*
710                  * TODO: we don't process guest user from host side
711                  * except simple counting.
712                  */
713                 return;
714         default:
715                 return;
716         }
717
718         if (!machine && perf_guest) {
719                 pr_err("Can't find guest [%d]'s kernel information\n",
720                         event->ip.pid);
721                 return;
722         }
723
724         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
725                 top.exact_samples++;
726
727         if (perf_event__preprocess_sample(event, session, &al, sample,
728                                           symbol_filter) < 0 ||
729             al.filtered)
730                 return;
731
732         if (!kptr_restrict_warned &&
733             symbol_conf.kptr_restrict &&
734             al.cpumode == PERF_RECORD_MISC_KERNEL) {
735                 ui__warning(
736 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
737 "Check /proc/sys/kernel/kptr_restrict.\n\n"
738 "Kernel%s samples will not be resolved.\n",
739                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
740                           " modules" : "");
741                 if (use_browser <= 0)
742                         sleep(5);
743                 kptr_restrict_warned = true;
744         }
745
746         if (al.sym == NULL) {
747                 const char *msg = "Kernel samples will not be resolved.\n";
748                 /*
749                  * As we do lazy loading of symtabs we only will know if the
750                  * specified vmlinux file is invalid when we actually have a
751                  * hit in kernel space and then try to load it. So if we get
752                  * here and there are _no_ symbols in the DSO backing the
753                  * kernel map, bail out.
754                  *
755                  * We may never get here, for instance, if we use -K/
756                  * --hide-kernel-symbols, even if the user specifies an
757                  * invalid --vmlinux ;-)
758                  */
759                 if (!kptr_restrict_warned && !vmlinux_warned &&
760                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
761                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
762                         if (symbol_conf.vmlinux_name) {
763                                 ui__warning("The %s file can't be used.\n%s",
764                                             symbol_conf.vmlinux_name, msg);
765                         } else {
766                                 ui__warning("A vmlinux file was not found.\n%s",
767                                             msg);
768                         }
769
770                         if (use_browser <= 0)
771                                 sleep(5);
772                         vmlinux_warned = true;
773                 }
774         }
775
776         if (al.sym == NULL || !al.sym->ignore) {
777                 struct hist_entry *he;
778
779                 if ((sort__has_parent || symbol_conf.use_callchain) &&
780                     sample->callchain) {
781                         err = perf_session__resolve_callchain(session, al.thread,
782                                                               sample->callchain, &parent);
783                         if (err)
784                                 return;
785                 }
786
787                 he = perf_session__add_hist_entry(session, &al, sample, evsel);
788                 if (he == NULL) {
789                         pr_err("Problem incrementing symbol period, skipping event\n");
790                         return;
791                 }
792
793                 if (symbol_conf.use_callchain) {
794                         err = callchain_append(he->callchain, &session->callchain_cursor,
795                                                sample->period);
796                         if (err)
797                                 return;
798                 }
799
800                 if (sort_has_symbols)
801                         record_precise_ip(he, evsel->idx, ip);
802         }
803
804         return;
805 }
806
807 static void perf_session__mmap_read_idx(struct perf_session *self, int idx)
808 {
809         struct perf_sample sample;
810         struct perf_evsel *evsel;
811         union perf_event *event;
812         int ret;
813
814         while ((event = perf_evlist__mmap_read(top.evlist, idx)) != NULL) {
815                 ret = perf_session__parse_sample(self, event, &sample);
816                 if (ret) {
817                         pr_err("Can't parse sample, err = %d\n", ret);
818                         continue;
819                 }
820
821                 evsel = perf_evlist__id2evsel(self->evlist, sample.id);
822                 assert(evsel != NULL);
823
824                 if (event->header.type == PERF_RECORD_SAMPLE)
825                         perf_event__process_sample(event, evsel, &sample, self);
826                 else if (event->header.type < PERF_RECORD_MAX) {
827                         hists__inc_nr_events(&evsel->hists, event->header.type);
828                         perf_event__process(event, &sample, self);
829                 } else
830                         ++self->hists.stats.nr_unknown_events;
831         }
832 }
833
834 static void perf_session__mmap_read(struct perf_session *self)
835 {
836         int i;
837
838         for (i = 0; i < top.evlist->nr_mmaps; i++)
839                 perf_session__mmap_read_idx(self, i);
840 }
841
842 static void start_counters(struct perf_evlist *evlist)
843 {
844         struct perf_evsel *counter, *first;
845
846         first = list_entry(evlist->entries.next, struct perf_evsel, node);
847
848         list_for_each_entry(counter, &evlist->entries, node) {
849                 struct perf_event_attr *attr = &counter->attr;
850                 struct xyarray *group_fd = NULL;
851
852                 if (group && counter != first)
853                         group_fd = first->fd;
854
855                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
856
857                 if (top.freq) {
858                         attr->sample_type |= PERF_SAMPLE_PERIOD;
859                         attr->freq        = 1;
860                         attr->sample_freq = top.freq;
861                 }
862
863                 if (evlist->nr_entries > 1) {
864                         attr->sample_type |= PERF_SAMPLE_ID;
865                         attr->read_format |= PERF_FORMAT_ID;
866                 }
867
868                 if (symbol_conf.use_callchain)
869                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
870
871                 attr->mmap = 1;
872                 attr->comm = 1;
873                 attr->inherit = inherit;
874 retry_sample_id:
875                 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
876 try_again:
877                 if (perf_evsel__open(counter, top.evlist->cpus,
878                                      top.evlist->threads, group,
879                                      group_fd) < 0) {
880                         int err = errno;
881
882                         if (err == EPERM || err == EACCES) {
883                                 ui__error_paranoid();
884                                 goto out_err;
885                         } else if (err == EINVAL && sample_id_all_avail) {
886                                 /*
887                                  * Old kernel, no attr->sample_id_type_all field
888                                  */
889                                 sample_id_all_avail = false;
890                                 goto retry_sample_id;
891                         }
892                         /*
893                          * If it's cycles then fall back to hrtimer
894                          * based cpu-clock-tick sw counter, which
895                          * is always available even if no PMU support:
896                          */
897                         if (attr->type == PERF_TYPE_HARDWARE &&
898                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
899                                 if (verbose)
900                                         ui__warning("Cycles event not supported,\n"
901                                                     "trying to fall back to cpu-clock-ticks\n");
902
903                                 attr->type = PERF_TYPE_SOFTWARE;
904                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
905                                 goto try_again;
906                         }
907
908                         if (err == ENOENT) {
909                                 ui__warning("The %s event is not supported.\n",
910                                             event_name(counter));
911                                 goto out_err;
912                         }
913
914                         ui__warning("The sys_perf_event_open() syscall "
915                                     "returned with %d (%s).  /bin/dmesg "
916                                     "may provide additional information.\n"
917                                     "No CONFIG_PERF_EVENTS=y kernel support "
918                                     "configured?\n", err, strerror(err));
919                         goto out_err;
920                 }
921         }
922
923         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0) {
924                 ui__warning("Failed to mmap with %d (%s)\n",
925                             errno, strerror(errno));
926                 goto out_err;
927         }
928
929         return;
930
931 out_err:
932         exit_browser(0);
933         exit(0);
934 }
935
936 static int setup_sample_type(void)
937 {
938         if (!sort_has_symbols) {
939                 if (symbol_conf.use_callchain) {
940                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
941                         return -EINVAL;
942                 }
943         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
944                 if (callchain_register_param(&callchain_param) < 0) {
945                         ui__warning("Can't register callchain params.\n");
946                         return -EINVAL;
947                 }
948         }
949
950         return 0;
951 }
952
953 static int __cmd_top(void)
954 {
955         pthread_t thread;
956         int ret;
957         /*
958          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
959          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
960          */
961         top.session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
962         if (top.session == NULL)
963                 return -ENOMEM;
964
965         ret = setup_sample_type();
966         if (ret)
967                 goto out_delete;
968
969         if (top.target_tid != -1)
970                 perf_event__synthesize_thread_map(top.evlist->threads,
971                                                   perf_event__process, top.session);
972         else
973                 perf_event__synthesize_threads(perf_event__process, top.session);
974
975         start_counters(top.evlist);
976         top.session->evlist = top.evlist;
977         perf_session__update_sample_type(top.session);
978
979         /* Wait for a minimal set of events before starting the snapshot */
980         poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
981
982         perf_session__mmap_read(top.session);
983
984         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
985                                                              display_thread), NULL)) {
986                 printf("Could not create display thread.\n");
987                 exit(-1);
988         }
989
990         if (realtime_prio) {
991                 struct sched_param param;
992
993                 param.sched_priority = realtime_prio;
994                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
995                         printf("Could not set realtime priority.\n");
996                         exit(-1);
997                 }
998         }
999
1000         while (1) {
1001                 u64 hits = top.samples;
1002
1003                 perf_session__mmap_read(top.session);
1004
1005                 if (hits == top.samples)
1006                         ret = poll(top.evlist->pollfd, top.evlist->nr_fds, 100);
1007         }
1008
1009 out_delete:
1010         perf_session__delete(top.session);
1011         top.session = NULL;
1012
1013         return 0;
1014 }
1015
1016 static int
1017 parse_callchain_opt(const struct option *opt __used, const char *arg,
1018                     int unset)
1019 {
1020         char *tok, *tok2;
1021         char *endptr;
1022
1023         /*
1024          * --no-call-graph
1025          */
1026         if (unset) {
1027                 dont_use_callchains = true;
1028                 return 0;
1029         }
1030
1031         symbol_conf.use_callchain = true;
1032
1033         if (!arg)
1034                 return 0;
1035
1036         tok = strtok((char *)arg, ",");
1037         if (!tok)
1038                 return -1;
1039
1040         /* get the output mode */
1041         if (!strncmp(tok, "graph", strlen(arg)))
1042                 callchain_param.mode = CHAIN_GRAPH_ABS;
1043
1044         else if (!strncmp(tok, "flat", strlen(arg)))
1045                 callchain_param.mode = CHAIN_FLAT;
1046
1047         else if (!strncmp(tok, "fractal", strlen(arg)))
1048                 callchain_param.mode = CHAIN_GRAPH_REL;
1049
1050         else if (!strncmp(tok, "none", strlen(arg))) {
1051                 callchain_param.mode = CHAIN_NONE;
1052                 symbol_conf.use_callchain = false;
1053
1054                 return 0;
1055         }
1056
1057         else
1058                 return -1;
1059
1060         /* get the min percentage */
1061         tok = strtok(NULL, ",");
1062         if (!tok)
1063                 goto setup;
1064
1065         callchain_param.min_percent = strtod(tok, &endptr);
1066         if (tok == endptr)
1067                 return -1;
1068
1069         /* get the print limit */
1070         tok2 = strtok(NULL, ",");
1071         if (!tok2)
1072                 goto setup;
1073
1074         if (tok2[0] != 'c') {
1075                 callchain_param.print_limit = strtod(tok2, &endptr);
1076                 tok2 = strtok(NULL, ",");
1077                 if (!tok2)
1078                         goto setup;
1079         }
1080
1081         /* get the call chain order */
1082         if (!strcmp(tok2, "caller"))
1083                 callchain_param.order = ORDER_CALLER;
1084         else if (!strcmp(tok2, "callee"))
1085                 callchain_param.order = ORDER_CALLEE;
1086         else
1087                 return -1;
1088 setup:
1089         if (callchain_register_param(&callchain_param) < 0) {
1090                 fprintf(stderr, "Can't register callchain params\n");
1091                 return -1;
1092         }
1093         return 0;
1094 }
1095
1096 static const char * const top_usage[] = {
1097         "perf top [<options>]",
1098         NULL
1099 };
1100
1101 static const struct option options[] = {
1102         OPT_CALLBACK('e', "event", &top.evlist, "event",
1103                      "event selector. use 'perf list' to list available events",
1104                      parse_events_option),
1105         OPT_INTEGER('c', "count", &default_interval,
1106                     "event period to sample"),
1107         OPT_INTEGER('p', "pid", &top.target_pid,
1108                     "profile events on existing process id"),
1109         OPT_INTEGER('t', "tid", &top.target_tid,
1110                     "profile events on existing thread id"),
1111         OPT_BOOLEAN('a', "all-cpus", &system_wide,
1112                             "system-wide collection from all CPUs"),
1113         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1114                     "list of cpus to monitor"),
1115         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1116                    "file", "vmlinux pathname"),
1117         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1118                     "hide kernel symbols"),
1119         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1120         OPT_INTEGER('r', "realtime", &realtime_prio,
1121                     "collect data with this RT SCHED_FIFO priority"),
1122         OPT_INTEGER('d', "delay", &top.delay_secs,
1123                     "number of seconds to delay between refreshes"),
1124         OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1125                             "dump the symbol table used for profiling"),
1126         OPT_INTEGER('f', "count-filter", &top.count_filter,
1127                     "only display functions with more events than this"),
1128         OPT_BOOLEAN('g', "group", &group,
1129                             "put the counters into a counter group"),
1130         OPT_BOOLEAN('i', "inherit", &inherit,
1131                     "child tasks inherit counters"),
1132         OPT_STRING(0, "sym-annotate", &sym_filter, "symbol name",
1133                     "symbol to annotate"),
1134         OPT_BOOLEAN('z', "zero", &top.zero,
1135                     "zero history across updates"),
1136         OPT_INTEGER('F', "freq", &top.freq,
1137                     "profile at this frequency"),
1138         OPT_INTEGER('E', "entries", &top.print_entries,
1139                     "display this many functions"),
1140         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1141                     "hide user symbols"),
1142         OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
1143         OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
1144         OPT_INCR('v', "verbose", &verbose,
1145                     "be more verbose (show counter open errors, etc)"),
1146         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1147                    "sort by key(s): pid, comm, dso, symbol, parent"),
1148         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1149                     "Show a column with the number of samples"),
1150         OPT_CALLBACK_DEFAULT('G', "call-graph", NULL, "output_type,min_percent, call_order",
1151                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1152                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1153                      callchain_default_opt),
1154         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1155                     "Show a column with the sum of periods"),
1156         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1157                    "only consider symbols in these dsos"),
1158         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1159                    "only consider symbols in these comms"),
1160         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1161                    "only consider these symbols"),
1162         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1163                     "Interleave source code with assembly code (default)"),
1164         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1165                     "Display raw encoding of assembly instructions (default)"),
1166         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1167                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1168         OPT_END()
1169 };
1170
1171 int cmd_top(int argc, const char **argv, const char *prefix __used)
1172 {
1173         struct perf_evsel *pos;
1174         int status = -ENOMEM;
1175
1176         top.evlist = perf_evlist__new(NULL, NULL);
1177         if (top.evlist == NULL)
1178                 return -ENOMEM;
1179
1180         symbol_conf.exclude_other = false;
1181
1182         argc = parse_options(argc, argv, options, top_usage, 0);
1183         if (argc)
1184                 usage_with_options(top_usage, options);
1185
1186         if (sort_order == default_sort_order)
1187                 sort_order = "dso,symbol";
1188
1189         setup_sorting(top_usage, options);
1190
1191         if (use_stdio)
1192                 use_browser = 0;
1193         else if (use_tui)
1194                 use_browser = 1;
1195
1196         setup_browser(false);
1197
1198         /* CPU and PID are mutually exclusive */
1199         if (top.target_tid > 0 && top.cpu_list) {
1200                 printf("WARNING: PID switch overriding CPU\n");
1201                 sleep(1);
1202                 top.cpu_list = NULL;
1203         }
1204
1205         if (top.target_pid != -1)
1206                 top.target_tid = top.target_pid;
1207
1208         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1209                                      top.target_tid, top.cpu_list) < 0)
1210                 usage_with_options(top_usage, options);
1211
1212         if (!top.evlist->nr_entries &&
1213             perf_evlist__add_default(top.evlist) < 0) {
1214                 pr_err("Not enough memory for event selector list\n");
1215                 return -ENOMEM;
1216         }
1217
1218         if (top.delay_secs < 1)
1219                 top.delay_secs = 1;
1220
1221         /*
1222          * User specified count overrides default frequency.
1223          */
1224         if (default_interval)
1225                 top.freq = 0;
1226         else if (top.freq) {
1227                 default_interval = top.freq;
1228         } else {
1229                 fprintf(stderr, "frequency and count are zero, aborting\n");
1230                 exit(EXIT_FAILURE);
1231         }
1232
1233         list_for_each_entry(pos, &top.evlist->entries, node) {
1234                 if (perf_evsel__alloc_fd(pos, top.evlist->cpus->nr,
1235                                          top.evlist->threads->nr) < 0)
1236                         goto out_free_fd;
1237                 /*
1238                  * Fill in the ones not specifically initialized via -c:
1239                  */
1240                 if (pos->attr.sample_period)
1241                         continue;
1242
1243                 pos->attr.sample_period = default_interval;
1244         }
1245
1246         if (perf_evlist__alloc_pollfd(top.evlist) < 0 ||
1247             perf_evlist__alloc_mmap(top.evlist) < 0)
1248                 goto out_free_fd;
1249
1250         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1251
1252         symbol_conf.priv_size = sizeof(struct annotation);
1253
1254         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1255         if (symbol__init() < 0)
1256                 return -1;
1257
1258         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1259         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1260         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1261
1262         /*
1263          * Avoid annotation data structures overhead when symbols aren't on the
1264          * sort list.
1265          */
1266         sort_has_symbols = sort_sym.list.next != NULL;
1267
1268         get_term_dimensions(&winsize);
1269         if (top.print_entries == 0) {
1270                 update_print_entries(&winsize);
1271                 signal(SIGWINCH, sig_winch_handler);
1272         }
1273
1274         status = __cmd_top();
1275 out_free_fd:
1276         perf_evlist__delete(top.evlist);
1277
1278         return status;
1279 }