]> git.karo-electronics.de Git - karo-tx-linux.git/blob - tools/perf/builtin-top.c
Merge branch 'upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/floppy...
[karo-tx-linux.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 <elf.h>
46 #include <fcntl.h>
47
48 #include <stdio.h>
49 #include <termios.h>
50 #include <unistd.h>
51 #include <inttypes.h>
52
53 #include <errno.h>
54 #include <time.h>
55 #include <sched.h>
56
57 #include <sys/syscall.h>
58 #include <sys/ioctl.h>
59 #include <sys/poll.h>
60 #include <sys/prctl.h>
61 #include <sys/wait.h>
62 #include <sys/uio.h>
63 #include <sys/utsname.h>
64 #include <sys/mman.h>
65
66 #include <linux/unistd.h>
67 #include <linux/types.h>
68
69 void get_term_dimensions(struct winsize *ws)
70 {
71         char *s = getenv("LINES");
72
73         if (s != NULL) {
74                 ws->ws_row = atoi(s);
75                 s = getenv("COLUMNS");
76                 if (s != NULL) {
77                         ws->ws_col = atoi(s);
78                         if (ws->ws_row && ws->ws_col)
79                                 return;
80                 }
81         }
82 #ifdef TIOCGWINSZ
83         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
84             ws->ws_row && ws->ws_col)
85                 return;
86 #endif
87         ws->ws_row = 25;
88         ws->ws_col = 80;
89 }
90
91 static void perf_top__update_print_entries(struct perf_top *top)
92 {
93         if (top->print_entries > 9)
94                 top->print_entries -= 9;
95 }
96
97 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
98 {
99         struct perf_top *top = arg;
100
101         get_term_dimensions(&top->winsize);
102         if (!top->print_entries
103             || (top->print_entries+4) > top->winsize.ws_row) {
104                 top->print_entries = top->winsize.ws_row;
105         } else {
106                 top->print_entries += 4;
107                 top->winsize.ws_row = top->print_entries;
108         }
109         perf_top__update_print_entries(top);
110 }
111
112 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
113 {
114         struct symbol *sym;
115         struct annotation *notes;
116         struct map *map;
117         int err = -1;
118
119         if (!he || !he->ms.sym)
120                 return -1;
121
122         sym = he->ms.sym;
123         map = he->ms.map;
124
125         /*
126          * We can't annotate with just /proc/kallsyms
127          */
128         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
129                 pr_err("Can't annotate %s: No vmlinux file was found in the "
130                        "path\n", sym->name);
131                 sleep(1);
132                 return -1;
133         }
134
135         notes = symbol__annotation(sym);
136         if (notes->src != NULL) {
137                 pthread_mutex_lock(&notes->lock);
138                 goto out_assign;
139         }
140
141         pthread_mutex_lock(&notes->lock);
142
143         if (symbol__alloc_hist(sym) < 0) {
144                 pthread_mutex_unlock(&notes->lock);
145                 pr_err("Not enough memory for annotating '%s' symbol!\n",
146                        sym->name);
147                 sleep(1);
148                 return err;
149         }
150
151         err = symbol__annotate(sym, map, 0);
152         if (err == 0) {
153 out_assign:
154                 top->sym_filter_entry = he;
155         }
156
157         pthread_mutex_unlock(&notes->lock);
158         return err;
159 }
160
161 static void __zero_source_counters(struct hist_entry *he)
162 {
163         struct symbol *sym = he->ms.sym;
164         symbol__annotate_zero_histograms(sym);
165 }
166
167 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
168 {
169         struct utsname uts;
170         int err = uname(&uts);
171
172         ui__warning("Out of bounds address found:\n\n"
173                     "Addr:   %" PRIx64 "\n"
174                     "DSO:    %s %c\n"
175                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
176                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
177                     "Arch:   %s\n"
178                     "Kernel: %s\n"
179                     "Tools:  %s\n\n"
180                     "Not all samples will be on the annotation output.\n\n"
181                     "Please report to linux-kernel@vger.kernel.org\n",
182                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
183                     map->start, map->end, sym->start, sym->end,
184                     sym->binding == STB_GLOBAL ? 'g' :
185                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
186                     err ? "[unknown]" : uts.machine,
187                     err ? "[unknown]" : uts.release, perf_version_string);
188         if (use_browser <= 0)
189                 sleep(5);
190         
191         map->erange_warned = true;
192 }
193
194 static void perf_top__record_precise_ip(struct perf_top *top,
195                                         struct hist_entry *he,
196                                         int counter, u64 ip)
197 {
198         struct annotation *notes;
199         struct symbol *sym;
200         int err;
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 && symbol__alloc_hist(sym) < 0) {
214                 pthread_mutex_unlock(&notes->lock);
215                 pr_err("Not enough memory for annotating '%s' symbol!\n",
216                        sym->name);
217                 sleep(1);
218                 return;
219         }
220
221         ip = he->ms.map->map_ip(he->ms.map, ip);
222         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
223
224         pthread_mutex_unlock(&notes->lock);
225
226         if (err == -ERANGE && !he->ms.map->erange_warned)
227                 ui__warn_map_erange(he->ms.map, sym, ip);
228 }
229
230 static void perf_top__show_details(struct perf_top *top)
231 {
232         struct hist_entry *he = top->sym_filter_entry;
233         struct annotation *notes;
234         struct symbol *symbol;
235         int more;
236
237         if (!he)
238                 return;
239
240         symbol = he->ms.sym;
241         notes = symbol__annotation(symbol);
242
243         pthread_mutex_lock(&notes->lock);
244
245         if (notes->src == NULL)
246                 goto out_unlock;
247
248         printf("Showing %s for %s\n", event_name(top->sym_evsel), symbol->name);
249         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
250
251         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
252                                        0, top->sym_pcnt_filter, top->print_entries, 4);
253         if (top->zero)
254                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
255         else
256                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
257         if (more != 0)
258                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
259 out_unlock:
260         pthread_mutex_unlock(&notes->lock);
261 }
262
263 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
264
265 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
266                                                      struct addr_location *al,
267                                                      struct perf_sample *sample)
268 {
269         struct hist_entry *he;
270
271         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
272         if (he == NULL)
273                 return NULL;
274
275         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
276         return he;
277 }
278
279 static void perf_top__print_sym_table(struct perf_top *top)
280 {
281         char bf[160];
282         int printed = 0;
283         const int win_width = top->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                 perf_top__show_details(top);
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,
315                                      top->winsize.ws_row - 3);
316         putchar('\n');
317         hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
318                        top->winsize.ws_row - 4 - printed, win_width, stdout);
319 }
320
321 static void prompt_integer(int *target, const char *msg)
322 {
323         char *buf = malloc(0), *p;
324         size_t dummy = 0;
325         int tmp;
326
327         fprintf(stdout, "\n%s: ", msg);
328         if (getline(&buf, &dummy, stdin) < 0)
329                 return;
330
331         p = strchr(buf, '\n');
332         if (p)
333                 *p = 0;
334
335         p = buf;
336         while(*p) {
337                 if (!isdigit(*p))
338                         goto out_free;
339                 p++;
340         }
341         tmp = strtoul(buf, NULL, 10);
342         *target = tmp;
343 out_free:
344         free(buf);
345 }
346
347 static void prompt_percent(int *target, const char *msg)
348 {
349         int tmp = 0;
350
351         prompt_integer(&tmp, msg);
352         if (tmp >= 0 && tmp <= 100)
353                 *target = tmp;
354 }
355
356 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
357 {
358         char *buf = malloc(0), *p;
359         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
360         struct rb_node *next;
361         size_t dummy = 0;
362
363         /* zero counters of active symbol */
364         if (syme) {
365                 __zero_source_counters(syme);
366                 top->sym_filter_entry = NULL;
367         }
368
369         fprintf(stdout, "\n%s: ", msg);
370         if (getline(&buf, &dummy, stdin) < 0)
371                 goto out_free;
372
373         p = strchr(buf, '\n');
374         if (p)
375                 *p = 0;
376
377         next = rb_first(&top->sym_evsel->hists.entries);
378         while (next) {
379                 n = rb_entry(next, struct hist_entry, rb_node);
380                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
381                         found = n;
382                         break;
383                 }
384                 next = rb_next(&n->rb_node);
385         }
386
387         if (!found) {
388                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
389                 sleep(1);
390         } else
391                 perf_top__parse_source(top, found);
392
393 out_free:
394         free(buf);
395 }
396
397 static void perf_top__print_mapped_keys(struct perf_top *top)
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", top->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 perf_top__key_mapped(struct perf_top *top, 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 perf_top__handle_keypress(struct perf_top *top, int c)
454 {
455         if (!perf_top__key_mapped(top, c)) {
456                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
457                 struct termios tc, save;
458
459                 perf_top__print_mapped_keys(top);
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 (!perf_top__key_mapped(top, 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                                 struct sigaction act = {
488                                         .sa_sigaction = perf_top__sig_winch,
489                                         .sa_flags     = SA_SIGINFO,
490                                 };
491                                 perf_top__sig_winch(SIGWINCH, NULL, top);
492                                 sigaction(SIGWINCH, &act, NULL);
493                         } else {
494                                 perf_top__sig_winch(SIGWINCH, NULL, top);
495                                 signal(SIGWINCH, SIG_DFL);
496                         }
497                         break;
498                 case 'E':
499                         if (top->evlist->nr_entries > 1) {
500                                 /* Select 0 as the default event: */
501                                 int counter = 0;
502
503                                 fprintf(stderr, "\nAvailable events:");
504
505                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
506                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
507
508                                 prompt_integer(&counter, "Enter details event counter");
509
510                                 if (counter >= top->evlist->nr_entries) {
511                                         top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
512                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
513                                         sleep(1);
514                                         break;
515                                 }
516                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
517                                         if (top->sym_evsel->idx == counter)
518                                                 break;
519                         } else
520                                 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
521                         break;
522                 case 'f':
523                         prompt_integer(&top->count_filter, "Enter display event count filter");
524                         break;
525                 case 'F':
526                         prompt_percent(&top->sym_pcnt_filter,
527                                        "Enter details display event filter (percent)");
528                         break;
529                 case 'K':
530                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
531                         break;
532                 case 'q':
533                 case 'Q':
534                         printf("exiting.\n");
535                         if (top->dump_symtab)
536                                 perf_session__fprintf_dsos(top->session, stderr);
537                         exit(0);
538                 case 's':
539                         perf_top__prompt_symbol(top, "Enter details symbol");
540                         break;
541                 case 'S':
542                         if (!top->sym_filter_entry)
543                                 break;
544                         else {
545                                 struct hist_entry *syme = top->sym_filter_entry;
546
547                                 top->sym_filter_entry = NULL;
548                                 __zero_source_counters(syme);
549                         }
550                         break;
551                 case 'U':
552                         top->hide_user_symbols = !top->hide_user_symbols;
553                         break;
554                 case 'z':
555                         top->zero = !top->zero;
556                         break;
557                 default:
558                         break;
559         }
560 }
561
562 static void perf_top__sort_new_samples(void *arg)
563 {
564         struct perf_top *t = arg;
565         perf_top__reset_sample_counters(t);
566
567         if (t->evlist->selected != NULL)
568                 t->sym_evsel = t->evlist->selected;
569
570         hists__collapse_resort_threaded(&t->sym_evsel->hists);
571         hists__output_resort_threaded(&t->sym_evsel->hists);
572         hists__decay_entries_threaded(&t->sym_evsel->hists,
573                                       t->hide_user_symbols,
574                                       t->hide_kernel_symbols);
575 }
576
577 static void *display_thread_tui(void *arg)
578 {
579         struct perf_evsel *pos;
580         struct perf_top *top = arg;
581         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
582
583         perf_top__sort_new_samples(top);
584
585         /*
586          * Initialize the uid_filter_str, in the future the TUI will allow
587          * Zooming in/out UIDs. For now juse use whatever the user passed
588          * via --uid.
589          */
590         list_for_each_entry(pos, &top->evlist->entries, node)
591                 pos->hists.uid_filter_str = top->target.uid_str;
592
593         perf_evlist__tui_browse_hists(top->evlist, help,
594                                       perf_top__sort_new_samples,
595                                       top, top->delay_secs);
596
597         exit_browser(0);
598         exit(0);
599         return NULL;
600 }
601
602 static void *display_thread(void *arg)
603 {
604         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
605         struct termios tc, save;
606         struct perf_top *top = arg;
607         int delay_msecs, c;
608
609         tcgetattr(0, &save);
610         tc = save;
611         tc.c_lflag &= ~(ICANON | ECHO);
612         tc.c_cc[VMIN] = 0;
613         tc.c_cc[VTIME] = 0;
614
615         pthread__unblock_sigwinch();
616 repeat:
617         delay_msecs = top->delay_secs * 1000;
618         tcsetattr(0, TCSANOW, &tc);
619         /* trash return*/
620         getc(stdin);
621
622         while (1) {
623                 perf_top__print_sym_table(top);
624                 /*
625                  * Either timeout expired or we got an EINTR due to SIGWINCH,
626                  * refresh screen in both cases.
627                  */
628                 switch (poll(&stdin_poll, 1, delay_msecs)) {
629                 case 0:
630                         continue;
631                 case -1:
632                         if (errno == EINTR)
633                                 continue;
634                         /* Fall trhu */
635                 default:
636                         goto process_hotkey;
637                 }
638         }
639 process_hotkey:
640         c = getc(stdin);
641         tcsetattr(0, TCSAFLUSH, &save);
642
643         perf_top__handle_keypress(top, c);
644         goto repeat;
645
646         return NULL;
647 }
648
649 /* Tag samples to be skipped. */
650 static const char *skip_symbols[] = {
651         "intel_idle",
652         "default_idle",
653         "native_safe_halt",
654         "cpu_idle",
655         "enter_idle",
656         "exit_idle",
657         "mwait_idle",
658         "mwait_idle_with_hints",
659         "poll_idle",
660         "ppc64_runlatch_off",
661         "pseries_dedicated_idle_sleep",
662         NULL
663 };
664
665 static int symbol_filter(struct map *map __used, struct symbol *sym)
666 {
667         const char *name = sym->name;
668         int i;
669
670         /*
671          * ppc64 uses function descriptors and appends a '.' to the
672          * start of every instruction address. Remove it.
673          */
674         if (name[0] == '.')
675                 name++;
676
677         if (!strcmp(name, "_text") ||
678             !strcmp(name, "_etext") ||
679             !strcmp(name, "_sinittext") ||
680             !strncmp("init_module", name, 11) ||
681             !strncmp("cleanup_module", name, 14) ||
682             strstr(name, "_text_start") ||
683             strstr(name, "_text_end"))
684                 return 1;
685
686         for (i = 0; skip_symbols[i]; i++) {
687                 if (!strcmp(skip_symbols[i], name)) {
688                         sym->ignore = true;
689                         break;
690                 }
691         }
692
693         return 0;
694 }
695
696 static void perf_event__process_sample(struct perf_tool *tool,
697                                        const union perf_event *event,
698                                        struct perf_evsel *evsel,
699                                        struct perf_sample *sample,
700                                        struct machine *machine)
701 {
702         struct perf_top *top = container_of(tool, struct perf_top, tool);
703         struct symbol *parent = NULL;
704         u64 ip = event->ip.ip;
705         struct addr_location al;
706         int err;
707
708         if (!machine && perf_guest) {
709                 pr_err("Can't find guest [%d]'s kernel information\n",
710                         event->ip.pid);
711                 return;
712         }
713
714         if (!machine) {
715                 pr_err("%u unprocessable samples recorded.",
716                        top->session->hists.stats.nr_unprocessable_samples++);
717                 return;
718         }
719
720         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
721                 top->exact_samples++;
722
723         if (perf_event__preprocess_sample(event, machine, &al, sample,
724                                           symbol_filter) < 0 ||
725             al.filtered)
726                 return;
727
728         if (!top->kptr_restrict_warned &&
729             symbol_conf.kptr_restrict &&
730             al.cpumode == PERF_RECORD_MISC_KERNEL) {
731                 ui__warning(
732 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
733 "Check /proc/sys/kernel/kptr_restrict.\n\n"
734 "Kernel%s samples will not be resolved.\n",
735                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
736                           " modules" : "");
737                 if (use_browser <= 0)
738                         sleep(5);
739                 top->kptr_restrict_warned = true;
740         }
741
742         if (al.sym == NULL) {
743                 const char *msg = "Kernel samples will not be resolved.\n";
744                 /*
745                  * As we do lazy loading of symtabs we only will know if the
746                  * specified vmlinux file is invalid when we actually have a
747                  * hit in kernel space and then try to load it. So if we get
748                  * here and there are _no_ symbols in the DSO backing the
749                  * kernel map, bail out.
750                  *
751                  * We may never get here, for instance, if we use -K/
752                  * --hide-kernel-symbols, even if the user specifies an
753                  * invalid --vmlinux ;-)
754                  */
755                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
756                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
757                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
758                         if (symbol_conf.vmlinux_name) {
759                                 ui__warning("The %s file can't be used.\n%s",
760                                             symbol_conf.vmlinux_name, msg);
761                         } else {
762                                 ui__warning("A vmlinux file was not found.\n%s",
763                                             msg);
764                         }
765
766                         if (use_browser <= 0)
767                                 sleep(5);
768                         top->vmlinux_warned = true;
769                 }
770         }
771
772         if (al.sym == NULL || !al.sym->ignore) {
773                 struct hist_entry *he;
774
775                 if ((sort__has_parent || symbol_conf.use_callchain) &&
776                     sample->callchain) {
777                         err = machine__resolve_callchain(machine, evsel, al.thread,
778                                                          sample->callchain, &parent);
779                         if (err)
780                                 return;
781                 }
782
783                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
784                 if (he == NULL) {
785                         pr_err("Problem incrementing symbol period, skipping event\n");
786                         return;
787                 }
788
789                 if (symbol_conf.use_callchain) {
790                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
791                                                sample->period);
792                         if (err)
793                                 return;
794                 }
795
796                 if (top->sort_has_symbols)
797                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
798         }
799
800         return;
801 }
802
803 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
804 {
805         struct perf_sample sample;
806         struct perf_evsel *evsel;
807         struct perf_session *session = top->session;
808         union perf_event *event;
809         struct machine *machine;
810         u8 origin;
811         int ret;
812
813         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
814                 ret = perf_session__parse_sample(session, event, &sample);
815                 if (ret) {
816                         pr_err("Can't parse sample, err = %d\n", ret);
817                         continue;
818                 }
819
820                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
821                 assert(evsel != NULL);
822
823                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
824
825                 if (event->header.type == PERF_RECORD_SAMPLE)
826                         ++top->samples;
827
828                 switch (origin) {
829                 case PERF_RECORD_MISC_USER:
830                         ++top->us_samples;
831                         if (top->hide_user_symbols)
832                                 continue;
833                         machine = perf_session__find_host_machine(session);
834                         break;
835                 case PERF_RECORD_MISC_KERNEL:
836                         ++top->kernel_samples;
837                         if (top->hide_kernel_symbols)
838                                 continue;
839                         machine = perf_session__find_host_machine(session);
840                         break;
841                 case PERF_RECORD_MISC_GUEST_KERNEL:
842                         ++top->guest_kernel_samples;
843                         machine = perf_session__find_machine(session, event->ip.pid);
844                         break;
845                 case PERF_RECORD_MISC_GUEST_USER:
846                         ++top->guest_us_samples;
847                         /*
848                          * TODO: we don't process guest user from host side
849                          * except simple counting.
850                          */
851                         /* Fall thru */
852                 default:
853                         continue;
854                 }
855
856
857                 if (event->header.type == PERF_RECORD_SAMPLE) {
858                         perf_event__process_sample(&top->tool, event, evsel,
859                                                    &sample, machine);
860                 } else if (event->header.type < PERF_RECORD_MAX) {
861                         hists__inc_nr_events(&evsel->hists, event->header.type);
862                         perf_event__process(&top->tool, event, &sample, machine);
863                 } else
864                         ++session->hists.stats.nr_unknown_events;
865         }
866 }
867
868 static void perf_top__mmap_read(struct perf_top *top)
869 {
870         int i;
871
872         for (i = 0; i < top->evlist->nr_mmaps; i++)
873                 perf_top__mmap_read_idx(top, i);
874 }
875
876 static void perf_top__start_counters(struct perf_top *top)
877 {
878         struct perf_evsel *counter, *first;
879         struct perf_evlist *evlist = top->evlist;
880
881         first = list_entry(evlist->entries.next, struct perf_evsel, node);
882
883         list_for_each_entry(counter, &evlist->entries, node) {
884                 struct perf_event_attr *attr = &counter->attr;
885                 struct xyarray *group_fd = NULL;
886
887                 if (top->group && counter != first)
888                         group_fd = first->fd;
889
890                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
891
892                 if (top->freq) {
893                         attr->sample_type |= PERF_SAMPLE_PERIOD;
894                         attr->freq        = 1;
895                         attr->sample_freq = top->freq;
896                 }
897
898                 if (evlist->nr_entries > 1) {
899                         attr->sample_type |= PERF_SAMPLE_ID;
900                         attr->read_format |= PERF_FORMAT_ID;
901                 }
902
903                 if (perf_target__has_cpu(&top->target))
904                         attr->sample_type |= PERF_SAMPLE_CPU;
905
906                 if (symbol_conf.use_callchain)
907                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
908
909                 attr->mmap = 1;
910                 attr->comm = 1;
911                 attr->inherit = top->inherit;
912 fallback_missing_features:
913                 if (top->exclude_guest_missing)
914                         attr->exclude_guest = attr->exclude_host = 0;
915 retry_sample_id:
916                 attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
917 try_again:
918                 if (perf_evsel__open(counter, top->evlist->cpus,
919                                      top->evlist->threads, top->group,
920                                      group_fd) < 0) {
921                         int err = errno;
922
923                         if (err == EPERM || err == EACCES) {
924                                 ui__error_paranoid();
925                                 goto out_err;
926                         } else if (err == EINVAL) {
927                                 if (!top->exclude_guest_missing &&
928                                     (attr->exclude_guest || attr->exclude_host)) {
929                                         pr_debug("Old kernel, cannot exclude "
930                                                  "guest or host samples.\n");
931                                         top->exclude_guest_missing = true;
932                                         goto fallback_missing_features;
933                                 } else if (!top->sample_id_all_missing) {
934                                         /*
935                                          * Old kernel, no attr->sample_id_type_all field
936                                          */
937                                         top->sample_id_all_missing = true;
938                                         goto retry_sample_id;
939                                 }
940                         }
941                         /*
942                          * If it's cycles then fall back to hrtimer
943                          * based cpu-clock-tick sw counter, which
944                          * is always available even if no PMU support:
945                          */
946                         if (attr->type == PERF_TYPE_HARDWARE &&
947                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
948                                 if (verbose)
949                                         ui__warning("Cycles event not supported,\n"
950                                                     "trying to fall back to cpu-clock-ticks\n");
951
952                                 attr->type = PERF_TYPE_SOFTWARE;
953                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
954                                 if (counter->name) {
955                                         free(counter->name);
956                                         counter->name = NULL;
957                                 }
958                                 goto try_again;
959                         }
960
961                         if (err == ENOENT) {
962                                 ui__error("The %s event is not supported.\n",
963                                             event_name(counter));
964                                 goto out_err;
965                         } else if (err == EMFILE) {
966                                 ui__error("Too many events are opened.\n"
967                                             "Try again after reducing the number of events\n");
968                                 goto out_err;
969                         }
970
971                         ui__error("The sys_perf_event_open() syscall "
972                                     "returned with %d (%s).  /bin/dmesg "
973                                     "may provide additional information.\n"
974                                     "No CONFIG_PERF_EVENTS=y kernel support "
975                                     "configured?\n", err, strerror(err));
976                         goto out_err;
977                 }
978         }
979
980         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
981                 ui__error("Failed to mmap with %d (%s)\n",
982                             errno, strerror(errno));
983                 goto out_err;
984         }
985
986         return;
987
988 out_err:
989         exit_browser(0);
990         exit(0);
991 }
992
993 static int perf_top__setup_sample_type(struct perf_top *top)
994 {
995         if (!top->sort_has_symbols) {
996                 if (symbol_conf.use_callchain) {
997                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
998                         return -EINVAL;
999                 }
1000         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
1001                 if (callchain_register_param(&callchain_param) < 0) {
1002                         ui__error("Can't register callchain params.\n");
1003                         return -EINVAL;
1004                 }
1005         }
1006
1007         return 0;
1008 }
1009
1010 static int __cmd_top(struct perf_top *top)
1011 {
1012         pthread_t thread;
1013         int ret;
1014         /*
1015          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1016          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1017          */
1018         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1019         if (top->session == NULL)
1020                 return -ENOMEM;
1021
1022         ret = perf_top__setup_sample_type(top);
1023         if (ret)
1024                 goto out_delete;
1025
1026         if (perf_target__has_task(&top->target))
1027                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
1028                                                   perf_event__process,
1029                                                   &top->session->host_machine);
1030         else
1031                 perf_event__synthesize_threads(&top->tool, perf_event__process,
1032                                                &top->session->host_machine);
1033         perf_top__start_counters(top);
1034         top->session->evlist = top->evlist;
1035         perf_session__update_sample_type(top->session);
1036
1037         /* Wait for a minimal set of events before starting the snapshot */
1038         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1039
1040         perf_top__mmap_read(top);
1041
1042         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1043                                                             display_thread), top)) {
1044                 ui__error("Could not create display thread.\n");
1045                 exit(-1);
1046         }
1047
1048         if (top->realtime_prio) {
1049                 struct sched_param param;
1050
1051                 param.sched_priority = top->realtime_prio;
1052                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1053                         ui__error("Could not set realtime priority.\n");
1054                         exit(-1);
1055                 }
1056         }
1057
1058         while (1) {
1059                 u64 hits = top->samples;
1060
1061                 perf_top__mmap_read(top);
1062
1063                 if (hits == top->samples)
1064                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1065         }
1066
1067 out_delete:
1068         perf_session__delete(top->session);
1069         top->session = NULL;
1070
1071         return 0;
1072 }
1073
1074 static int
1075 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1076 {
1077         struct perf_top *top = (struct perf_top *)opt->value;
1078         char *tok, *tok2;
1079         char *endptr;
1080
1081         /*
1082          * --no-call-graph
1083          */
1084         if (unset) {
1085                 top->dont_use_callchains = true;
1086                 return 0;
1087         }
1088
1089         symbol_conf.use_callchain = true;
1090
1091         if (!arg)
1092                 return 0;
1093
1094         tok = strtok((char *)arg, ",");
1095         if (!tok)
1096                 return -1;
1097
1098         /* get the output mode */
1099         if (!strncmp(tok, "graph", strlen(arg)))
1100                 callchain_param.mode = CHAIN_GRAPH_ABS;
1101
1102         else if (!strncmp(tok, "flat", strlen(arg)))
1103                 callchain_param.mode = CHAIN_FLAT;
1104
1105         else if (!strncmp(tok, "fractal", strlen(arg)))
1106                 callchain_param.mode = CHAIN_GRAPH_REL;
1107
1108         else if (!strncmp(tok, "none", strlen(arg))) {
1109                 callchain_param.mode = CHAIN_NONE;
1110                 symbol_conf.use_callchain = false;
1111
1112                 return 0;
1113         } else
1114                 return -1;
1115
1116         /* get the min percentage */
1117         tok = strtok(NULL, ",");
1118         if (!tok)
1119                 goto setup;
1120
1121         callchain_param.min_percent = strtod(tok, &endptr);
1122         if (tok == endptr)
1123                 return -1;
1124
1125         /* get the print limit */
1126         tok2 = strtok(NULL, ",");
1127         if (!tok2)
1128                 goto setup;
1129
1130         if (tok2[0] != 'c') {
1131                 callchain_param.print_limit = strtod(tok2, &endptr);
1132                 tok2 = strtok(NULL, ",");
1133                 if (!tok2)
1134                         goto setup;
1135         }
1136
1137         /* get the call chain order */
1138         if (!strcmp(tok2, "caller"))
1139                 callchain_param.order = ORDER_CALLER;
1140         else if (!strcmp(tok2, "callee"))
1141                 callchain_param.order = ORDER_CALLEE;
1142         else
1143                 return -1;
1144 setup:
1145         if (callchain_register_param(&callchain_param) < 0) {
1146                 fprintf(stderr, "Can't register callchain params\n");
1147                 return -1;
1148         }
1149         return 0;
1150 }
1151
1152 static const char * const top_usage[] = {
1153         "perf top [<options>]",
1154         NULL
1155 };
1156
1157 int cmd_top(int argc, const char **argv, const char *prefix __used)
1158 {
1159         struct perf_evsel *pos;
1160         int status;
1161         char errbuf[BUFSIZ];
1162         struct perf_top top = {
1163                 .count_filter        = 5,
1164                 .delay_secs          = 2,
1165                 .freq                = 4000, /* 4 KHz */
1166                 .mmap_pages          = 128,
1167                 .sym_pcnt_filter     = 5,
1168                 .target              = {
1169                         .uses_mmap   = true,
1170                 },
1171         };
1172         char callchain_default_opt[] = "fractal,0.5,callee";
1173         const struct option options[] = {
1174         OPT_CALLBACK('e', "event", &top.evlist, "event",
1175                      "event selector. use 'perf list' to list available events",
1176                      parse_events_option),
1177         OPT_INTEGER('c', "count", &top.default_interval,
1178                     "event period to sample"),
1179         OPT_STRING('p', "pid", &top.target.pid, "pid",
1180                     "profile events on existing process id"),
1181         OPT_STRING('t', "tid", &top.target.tid, "tid",
1182                     "profile events on existing thread id"),
1183         OPT_BOOLEAN('a', "all-cpus", &top.target.system_wide,
1184                             "system-wide collection from all CPUs"),
1185         OPT_STRING('C', "cpu", &top.target.cpu_list, "cpu",
1186                     "list of cpus to monitor"),
1187         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1188                    "file", "vmlinux pathname"),
1189         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1190                     "hide kernel symbols"),
1191         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1192         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1193                     "collect data with this RT SCHED_FIFO priority"),
1194         OPT_INTEGER('d', "delay", &top.delay_secs,
1195                     "number of seconds to delay between refreshes"),
1196         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1197                             "dump the symbol table used for profiling"),
1198         OPT_INTEGER('f', "count-filter", &top.count_filter,
1199                     "only display functions with more events than this"),
1200         OPT_BOOLEAN('g', "group", &top.group,
1201                             "put the counters into a counter group"),
1202         OPT_BOOLEAN('i', "inherit", &top.inherit,
1203                     "child tasks inherit counters"),
1204         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1205                     "symbol to annotate"),
1206         OPT_BOOLEAN('z', "zero", &top.zero,
1207                     "zero history across updates"),
1208         OPT_INTEGER('F', "freq", &top.freq,
1209                     "profile at this frequency"),
1210         OPT_INTEGER('E', "entries", &top.print_entries,
1211                     "display this many functions"),
1212         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1213                     "hide user symbols"),
1214         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1215         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1216         OPT_INCR('v', "verbose", &verbose,
1217                     "be more verbose (show counter open errors, etc)"),
1218         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1219                    "sort by key(s): pid, comm, dso, symbol, parent"),
1220         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1221                     "Show a column with the number of samples"),
1222         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1223                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1224                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1225                      callchain_default_opt),
1226         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1227                     "Show a column with the sum of periods"),
1228         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1229                    "only consider symbols in these dsos"),
1230         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1231                    "only consider symbols in these comms"),
1232         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1233                    "only consider these symbols"),
1234         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1235                     "Interleave source code with assembly code (default)"),
1236         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1237                     "Display raw encoding of assembly instructions (default)"),
1238         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1239                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1240         OPT_STRING('u', "uid", &top.target.uid_str, "user", "user to profile"),
1241         OPT_END()
1242         };
1243
1244         top.evlist = perf_evlist__new(NULL, NULL);
1245         if (top.evlist == NULL)
1246                 return -ENOMEM;
1247
1248         symbol_conf.exclude_other = false;
1249
1250         argc = parse_options(argc, argv, options, top_usage, 0);
1251         if (argc)
1252                 usage_with_options(top_usage, options);
1253
1254         if (sort_order == default_sort_order)
1255                 sort_order = "dso,symbol";
1256
1257         setup_sorting(top_usage, options);
1258
1259         if (top.use_stdio)
1260                 use_browser = 0;
1261         else if (top.use_tui)
1262                 use_browser = 1;
1263
1264         setup_browser(false);
1265
1266         status = perf_target__validate(&top.target);
1267         if (status) {
1268                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1269                 ui__warning("%s", errbuf);
1270         }
1271
1272         status = perf_target__parse_uid(&top.target);
1273         if (status) {
1274                 int saved_errno = errno;
1275
1276                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1277                 ui__error("%s", errbuf);
1278
1279                 status = -saved_errno;
1280                 goto out_delete_evlist;
1281         }
1282
1283         if (perf_target__none(&top.target))
1284                 top.target.system_wide = true;
1285
1286         if (perf_evlist__create_maps(top.evlist, &top.target) < 0)
1287                 usage_with_options(top_usage, options);
1288
1289         if (!top.evlist->nr_entries &&
1290             perf_evlist__add_default(top.evlist) < 0) {
1291                 ui__error("Not enough memory for event selector list\n");
1292                 return -ENOMEM;
1293         }
1294
1295         symbol_conf.nr_events = top.evlist->nr_entries;
1296
1297         if (top.delay_secs < 1)
1298                 top.delay_secs = 1;
1299
1300         /*
1301          * User specified count overrides default frequency.
1302          */
1303         if (top.default_interval)
1304                 top.freq = 0;
1305         else if (top.freq) {
1306                 top.default_interval = top.freq;
1307         } else {
1308                 ui__error("frequency and count are zero, aborting\n");
1309                 exit(EXIT_FAILURE);
1310         }
1311
1312         list_for_each_entry(pos, &top.evlist->entries, node) {
1313                 /*
1314                  * Fill in the ones not specifically initialized via -c:
1315                  */
1316                 if (!pos->attr.sample_period)
1317                         pos->attr.sample_period = top.default_interval;
1318         }
1319
1320         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1321
1322         symbol_conf.priv_size = sizeof(struct annotation);
1323
1324         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1325         if (symbol__init() < 0)
1326                 return -1;
1327
1328         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1329         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1330         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1331
1332         /*
1333          * Avoid annotation data structures overhead when symbols aren't on the
1334          * sort list.
1335          */
1336         top.sort_has_symbols = sort_sym.list.next != NULL;
1337
1338         get_term_dimensions(&top.winsize);
1339         if (top.print_entries == 0) {
1340                 struct sigaction act = {
1341                         .sa_sigaction = perf_top__sig_winch,
1342                         .sa_flags     = SA_SIGINFO,
1343                 };
1344                 perf_top__update_print_entries(&top);
1345                 sigaction(SIGWINCH, &act, NULL);
1346         }
1347
1348         status = __cmd_top(&top);
1349
1350 out_delete_evlist:
1351         perf_evlist__delete(top.evlist);
1352
1353         return status;
1354 }