Merge tag 'nfs-for-3.6-4' of git://git.linux-nfs.org/projects/trondmy/linux-nfs
[platform/adaptation/renesas_rcar/renesas_kernel.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 #include "util/intlist.h"
42
43 #include "util/debug.h"
44
45 #include <assert.h>
46 #include <elf.h>
47 #include <fcntl.h>
48
49 #include <stdio.h>
50 #include <termios.h>
51 #include <unistd.h>
52 #include <inttypes.h>
53
54 #include <errno.h>
55 #include <time.h>
56 #include <sched.h>
57
58 #include <sys/syscall.h>
59 #include <sys/ioctl.h>
60 #include <sys/poll.h>
61 #include <sys/prctl.h>
62 #include <sys/wait.h>
63 #include <sys/uio.h>
64 #include <sys/utsname.h>
65 #include <sys/mman.h>
66
67 #include <linux/unistd.h>
68 #include <linux/types.h>
69
70 void get_term_dimensions(struct winsize *ws)
71 {
72         char *s = getenv("LINES");
73
74         if (s != NULL) {
75                 ws->ws_row = atoi(s);
76                 s = getenv("COLUMNS");
77                 if (s != NULL) {
78                         ws->ws_col = atoi(s);
79                         if (ws->ws_row && ws->ws_col)
80                                 return;
81                 }
82         }
83 #ifdef TIOCGWINSZ
84         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
85             ws->ws_row && ws->ws_col)
86                 return;
87 #endif
88         ws->ws_row = 25;
89         ws->ws_col = 80;
90 }
91
92 static void perf_top__update_print_entries(struct perf_top *top)
93 {
94         if (top->print_entries > 9)
95                 top->print_entries -= 9;
96 }
97
98 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
99 {
100         struct perf_top *top = arg;
101
102         get_term_dimensions(&top->winsize);
103         if (!top->print_entries
104             || (top->print_entries+4) > top->winsize.ws_row) {
105                 top->print_entries = top->winsize.ws_row;
106         } else {
107                 top->print_entries += 4;
108                 top->winsize.ws_row = top->print_entries;
109         }
110         perf_top__update_print_entries(top);
111 }
112
113 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
114 {
115         struct symbol *sym;
116         struct annotation *notes;
117         struct map *map;
118         int err = -1;
119
120         if (!he || !he->ms.sym)
121                 return -1;
122
123         sym = he->ms.sym;
124         map = he->ms.map;
125
126         /*
127          * We can't annotate with just /proc/kallsyms
128          */
129         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
130                 pr_err("Can't annotate %s: No vmlinux file was found in the "
131                        "path\n", sym->name);
132                 sleep(1);
133                 return -1;
134         }
135
136         notes = symbol__annotation(sym);
137         if (notes->src != NULL) {
138                 pthread_mutex_lock(&notes->lock);
139                 goto out_assign;
140         }
141
142         pthread_mutex_lock(&notes->lock);
143
144         if (symbol__alloc_hist(sym) < 0) {
145                 pthread_mutex_unlock(&notes->lock);
146                 pr_err("Not enough memory for annotating '%s' symbol!\n",
147                        sym->name);
148                 sleep(1);
149                 return err;
150         }
151
152         err = symbol__annotate(sym, map, 0);
153         if (err == 0) {
154 out_assign:
155                 top->sym_filter_entry = he;
156         }
157
158         pthread_mutex_unlock(&notes->lock);
159         return err;
160 }
161
162 static void __zero_source_counters(struct hist_entry *he)
163 {
164         struct symbol *sym = he->ms.sym;
165         symbol__annotate_zero_histograms(sym);
166 }
167
168 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
169 {
170         struct utsname uts;
171         int err = uname(&uts);
172
173         ui__warning("Out of bounds address found:\n\n"
174                     "Addr:   %" PRIx64 "\n"
175                     "DSO:    %s %c\n"
176                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
177                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
178                     "Arch:   %s\n"
179                     "Kernel: %s\n"
180                     "Tools:  %s\n\n"
181                     "Not all samples will be on the annotation output.\n\n"
182                     "Please report to linux-kernel@vger.kernel.org\n",
183                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
184                     map->start, map->end, sym->start, sym->end,
185                     sym->binding == STB_GLOBAL ? 'g' :
186                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
187                     err ? "[unknown]" : uts.machine,
188                     err ? "[unknown]" : uts.release, perf_version_string);
189         if (use_browser <= 0)
190                 sleep(5);
191         
192         map->erange_warned = true;
193 }
194
195 static void perf_top__record_precise_ip(struct perf_top *top,
196                                         struct hist_entry *he,
197                                         int counter, u64 ip)
198 {
199         struct annotation *notes;
200         struct symbol *sym;
201         int err;
202
203         if (he == NULL || he->ms.sym == NULL ||
204             ((top->sym_filter_entry == NULL ||
205               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
206                 return;
207
208         sym = he->ms.sym;
209         notes = symbol__annotation(sym);
210
211         if (pthread_mutex_trylock(&notes->lock))
212                 return;
213
214         if (notes->src == NULL && symbol__alloc_hist(sym) < 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         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
224
225         pthread_mutex_unlock(&notes->lock);
226
227         if (err == -ERANGE && !he->ms.map->erange_warned)
228                 ui__warn_map_erange(he->ms.map, sym, ip);
229 }
230
231 static void perf_top__show_details(struct perf_top *top)
232 {
233         struct hist_entry *he = top->sym_filter_entry;
234         struct annotation *notes;
235         struct symbol *symbol;
236         int more;
237
238         if (!he)
239                 return;
240
241         symbol = he->ms.sym;
242         notes = symbol__annotation(symbol);
243
244         pthread_mutex_lock(&notes->lock);
245
246         if (notes->src == NULL)
247                 goto out_unlock;
248
249         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
250         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
251
252         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
253                                        0, top->sym_pcnt_filter, top->print_entries, 4);
254         if (top->zero)
255                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
256         else
257                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
258         if (more != 0)
259                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
260 out_unlock:
261         pthread_mutex_unlock(&notes->lock);
262 }
263
264 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
265
266 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
267                                                      struct addr_location *al,
268                                                      struct perf_sample *sample)
269 {
270         struct hist_entry *he;
271
272         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
273         if (he == NULL)
274                 return NULL;
275
276         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
277         return he;
278 }
279
280 static void perf_top__print_sym_table(struct perf_top *top)
281 {
282         char bf[160];
283         int printed = 0;
284         const int win_width = top->winsize.ws_col - 1;
285
286         puts(CONSOLE_CLEAR);
287
288         perf_top__header_snprintf(top, bf, sizeof(bf));
289         printf("%s\n", bf);
290
291         perf_top__reset_sample_counters(top);
292
293         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
294
295         if (top->sym_evsel->hists.stats.nr_lost_warned !=
296             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
297                 top->sym_evsel->hists.stats.nr_lost_warned =
298                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
299                 color_fprintf(stdout, PERF_COLOR_RED,
300                               "WARNING: LOST %d chunks, Check IO/CPU overload",
301                               top->sym_evsel->hists.stats.nr_lost_warned);
302                 ++printed;
303         }
304
305         if (top->sym_filter_entry) {
306                 perf_top__show_details(top);
307                 return;
308         }
309
310         hists__collapse_resort_threaded(&top->sym_evsel->hists);
311         hists__output_resort_threaded(&top->sym_evsel->hists);
312         hists__decay_entries_threaded(&top->sym_evsel->hists,
313                                       top->hide_user_symbols,
314                                       top->hide_kernel_symbols);
315         hists__output_recalc_col_len(&top->sym_evsel->hists,
316                                      top->winsize.ws_row - 3);
317         putchar('\n');
318         hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
319                        top->winsize.ws_row - 4 - printed, win_width, stdout);
320 }
321
322 static void prompt_integer(int *target, const char *msg)
323 {
324         char *buf = malloc(0), *p;
325         size_t dummy = 0;
326         int tmp;
327
328         fprintf(stdout, "\n%s: ", msg);
329         if (getline(&buf, &dummy, stdin) < 0)
330                 return;
331
332         p = strchr(buf, '\n');
333         if (p)
334                 *p = 0;
335
336         p = buf;
337         while(*p) {
338                 if (!isdigit(*p))
339                         goto out_free;
340                 p++;
341         }
342         tmp = strtoul(buf, NULL, 10);
343         *target = tmp;
344 out_free:
345         free(buf);
346 }
347
348 static void prompt_percent(int *target, const char *msg)
349 {
350         int tmp = 0;
351
352         prompt_integer(&tmp, msg);
353         if (tmp >= 0 && tmp <= 100)
354                 *target = tmp;
355 }
356
357 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
358 {
359         char *buf = malloc(0), *p;
360         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
361         struct rb_node *next;
362         size_t dummy = 0;
363
364         /* zero counters of active symbol */
365         if (syme) {
366                 __zero_source_counters(syme);
367                 top->sym_filter_entry = NULL;
368         }
369
370         fprintf(stdout, "\n%s: ", msg);
371         if (getline(&buf, &dummy, stdin) < 0)
372                 goto out_free;
373
374         p = strchr(buf, '\n');
375         if (p)
376                 *p = 0;
377
378         next = rb_first(&top->sym_evsel->hists.entries);
379         while (next) {
380                 n = rb_entry(next, struct hist_entry, rb_node);
381                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
382                         found = n;
383                         break;
384                 }
385                 next = rb_next(&n->rb_node);
386         }
387
388         if (!found) {
389                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
390                 sleep(1);
391         } else
392                 perf_top__parse_source(top, found);
393
394 out_free:
395         free(buf);
396 }
397
398 static void perf_top__print_mapped_keys(struct perf_top *top)
399 {
400         char *name = NULL;
401
402         if (top->sym_filter_entry) {
403                 struct symbol *sym = top->sym_filter_entry->ms.sym;
404                 name = sym->name;
405         }
406
407         fprintf(stdout, "\nMapped keys:\n");
408         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
409         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
410
411         if (top->evlist->nr_entries > 1)
412                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
413
414         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
415
416         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
417         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
418         fprintf(stdout, "\t[S]     stop annotation.\n");
419
420         fprintf(stdout,
421                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
422                 top->hide_kernel_symbols ? "yes" : "no");
423         fprintf(stdout,
424                 "\t[U]     hide user symbols.               \t(%s)\n",
425                 top->hide_user_symbols ? "yes" : "no");
426         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
427         fprintf(stdout, "\t[qQ]    quit.\n");
428 }
429
430 static int perf_top__key_mapped(struct perf_top *top, int c)
431 {
432         switch (c) {
433                 case 'd':
434                 case 'e':
435                 case 'f':
436                 case 'z':
437                 case 'q':
438                 case 'Q':
439                 case 'K':
440                 case 'U':
441                 case 'F':
442                 case 's':
443                 case 'S':
444                         return 1;
445                 case 'E':
446                         return top->evlist->nr_entries > 1 ? 1 : 0;
447                 default:
448                         break;
449         }
450
451         return 0;
452 }
453
454 static void perf_top__handle_keypress(struct perf_top *top, int c)
455 {
456         if (!perf_top__key_mapped(top, c)) {
457                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
458                 struct termios tc, save;
459
460                 perf_top__print_mapped_keys(top);
461                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
462                 fflush(stdout);
463
464                 tcgetattr(0, &save);
465                 tc = save;
466                 tc.c_lflag &= ~(ICANON | ECHO);
467                 tc.c_cc[VMIN] = 0;
468                 tc.c_cc[VTIME] = 0;
469                 tcsetattr(0, TCSANOW, &tc);
470
471                 poll(&stdin_poll, 1, -1);
472                 c = getc(stdin);
473
474                 tcsetattr(0, TCSAFLUSH, &save);
475                 if (!perf_top__key_mapped(top, c))
476                         return;
477         }
478
479         switch (c) {
480                 case 'd':
481                         prompt_integer(&top->delay_secs, "Enter display delay");
482                         if (top->delay_secs < 1)
483                                 top->delay_secs = 1;
484                         break;
485                 case 'e':
486                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
487                         if (top->print_entries == 0) {
488                                 struct sigaction act = {
489                                         .sa_sigaction = perf_top__sig_winch,
490                                         .sa_flags     = SA_SIGINFO,
491                                 };
492                                 perf_top__sig_winch(SIGWINCH, NULL, top);
493                                 sigaction(SIGWINCH, &act, NULL);
494                         } else {
495                                 perf_top__sig_winch(SIGWINCH, NULL, top);
496                                 signal(SIGWINCH, SIG_DFL);
497                         }
498                         break;
499                 case 'E':
500                         if (top->evlist->nr_entries > 1) {
501                                 /* Select 0 as the default event: */
502                                 int counter = 0;
503
504                                 fprintf(stderr, "\nAvailable events:");
505
506                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
507                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
508
509                                 prompt_integer(&counter, "Enter details event counter");
510
511                                 if (counter >= top->evlist->nr_entries) {
512                                         top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
513                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
514                                         sleep(1);
515                                         break;
516                                 }
517                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
518                                         if (top->sym_evsel->idx == counter)
519                                                 break;
520                         } else
521                                 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
522                         break;
523                 case 'f':
524                         prompt_integer(&top->count_filter, "Enter display event count filter");
525                         break;
526                 case 'F':
527                         prompt_percent(&top->sym_pcnt_filter,
528                                        "Enter details display event filter (percent)");
529                         break;
530                 case 'K':
531                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
532                         break;
533                 case 'q':
534                 case 'Q':
535                         printf("exiting.\n");
536                         if (top->dump_symtab)
537                                 perf_session__fprintf_dsos(top->session, stderr);
538                         exit(0);
539                 case 's':
540                         perf_top__prompt_symbol(top, "Enter details symbol");
541                         break;
542                 case 'S':
543                         if (!top->sym_filter_entry)
544                                 break;
545                         else {
546                                 struct hist_entry *syme = top->sym_filter_entry;
547
548                                 top->sym_filter_entry = NULL;
549                                 __zero_source_counters(syme);
550                         }
551                         break;
552                 case 'U':
553                         top->hide_user_symbols = !top->hide_user_symbols;
554                         break;
555                 case 'z':
556                         top->zero = !top->zero;
557                         break;
558                 default:
559                         break;
560         }
561 }
562
563 static void perf_top__sort_new_samples(void *arg)
564 {
565         struct perf_top *t = arg;
566         perf_top__reset_sample_counters(t);
567
568         if (t->evlist->selected != NULL)
569                 t->sym_evsel = t->evlist->selected;
570
571         hists__collapse_resort_threaded(&t->sym_evsel->hists);
572         hists__output_resort_threaded(&t->sym_evsel->hists);
573         hists__decay_entries_threaded(&t->sym_evsel->hists,
574                                       t->hide_user_symbols,
575                                       t->hide_kernel_symbols);
576 }
577
578 static void *display_thread_tui(void *arg)
579 {
580         struct perf_evsel *pos;
581         struct perf_top *top = arg;
582         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
583
584         perf_top__sort_new_samples(top);
585
586         /*
587          * Initialize the uid_filter_str, in the future the TUI will allow
588          * Zooming in/out UIDs. For now juse use whatever the user passed
589          * via --uid.
590          */
591         list_for_each_entry(pos, &top->evlist->entries, node)
592                 pos->hists.uid_filter_str = top->target.uid_str;
593
594         perf_evlist__tui_browse_hists(top->evlist, help,
595                                       perf_top__sort_new_samples,
596                                       top, top->delay_secs);
597
598         exit_browser(0);
599         exit(0);
600         return NULL;
601 }
602
603 static void *display_thread(void *arg)
604 {
605         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
606         struct termios tc, save;
607         struct perf_top *top = arg;
608         int delay_msecs, c;
609
610         tcgetattr(0, &save);
611         tc = save;
612         tc.c_lflag &= ~(ICANON | ECHO);
613         tc.c_cc[VMIN] = 0;
614         tc.c_cc[VTIME] = 0;
615
616         pthread__unblock_sigwinch();
617 repeat:
618         delay_msecs = top->delay_secs * 1000;
619         tcsetattr(0, TCSANOW, &tc);
620         /* trash return*/
621         getc(stdin);
622
623         while (1) {
624                 perf_top__print_sym_table(top);
625                 /*
626                  * Either timeout expired or we got an EINTR due to SIGWINCH,
627                  * refresh screen in both cases.
628                  */
629                 switch (poll(&stdin_poll, 1, delay_msecs)) {
630                 case 0:
631                         continue;
632                 case -1:
633                         if (errno == EINTR)
634                                 continue;
635                         /* Fall trhu */
636                 default:
637                         goto process_hotkey;
638                 }
639         }
640 process_hotkey:
641         c = getc(stdin);
642         tcsetattr(0, TCSAFLUSH, &save);
643
644         perf_top__handle_keypress(top, c);
645         goto repeat;
646
647         return NULL;
648 }
649
650 /* Tag samples to be skipped. */
651 static const char *skip_symbols[] = {
652         "intel_idle",
653         "default_idle",
654         "native_safe_halt",
655         "cpu_idle",
656         "enter_idle",
657         "exit_idle",
658         "mwait_idle",
659         "mwait_idle_with_hints",
660         "poll_idle",
661         "ppc64_runlatch_off",
662         "pseries_dedicated_idle_sleep",
663         NULL
664 };
665
666 static int symbol_filter(struct map *map __used, struct symbol *sym)
667 {
668         const char *name = sym->name;
669         int i;
670
671         /*
672          * ppc64 uses function descriptors and appends a '.' to the
673          * start of every instruction address. Remove it.
674          */
675         if (name[0] == '.')
676                 name++;
677
678         if (!strcmp(name, "_text") ||
679             !strcmp(name, "_etext") ||
680             !strcmp(name, "_sinittext") ||
681             !strncmp("init_module", name, 11) ||
682             !strncmp("cleanup_module", name, 14) ||
683             strstr(name, "_text_start") ||
684             strstr(name, "_text_end"))
685                 return 1;
686
687         for (i = 0; skip_symbols[i]; i++) {
688                 if (!strcmp(skip_symbols[i], name)) {
689                         sym->ignore = true;
690                         break;
691                 }
692         }
693
694         return 0;
695 }
696
697 static void perf_event__process_sample(struct perf_tool *tool,
698                                        const union perf_event *event,
699                                        struct perf_evsel *evsel,
700                                        struct perf_sample *sample,
701                                        struct machine *machine)
702 {
703         struct perf_top *top = container_of(tool, struct perf_top, tool);
704         struct symbol *parent = NULL;
705         u64 ip = event->ip.ip;
706         struct addr_location al;
707         int err;
708
709         if (!machine && perf_guest) {
710                 static struct intlist *seen;
711
712                 if (!seen)
713                         seen = intlist__new();
714
715                 if (!intlist__has_entry(seen, event->ip.pid)) {
716                         pr_err("Can't find guest [%d]'s kernel information\n",
717                                 event->ip.pid);
718                         intlist__add(seen, event->ip.pid);
719                 }
720                 return;
721         }
722
723         if (!machine) {
724                 pr_err("%u unprocessable samples recorded.",
725                        top->session->hists.stats.nr_unprocessable_samples++);
726                 return;
727         }
728
729         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
730                 top->exact_samples++;
731
732         if (perf_event__preprocess_sample(event, machine, &al, sample,
733                                           symbol_filter) < 0 ||
734             al.filtered)
735                 return;
736
737         if (!top->kptr_restrict_warned &&
738             symbol_conf.kptr_restrict &&
739             al.cpumode == PERF_RECORD_MISC_KERNEL) {
740                 ui__warning(
741 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
742 "Check /proc/sys/kernel/kptr_restrict.\n\n"
743 "Kernel%s samples will not be resolved.\n",
744                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
745                           " modules" : "");
746                 if (use_browser <= 0)
747                         sleep(5);
748                 top->kptr_restrict_warned = true;
749         }
750
751         if (al.sym == NULL) {
752                 const char *msg = "Kernel samples will not be resolved.\n";
753                 /*
754                  * As we do lazy loading of symtabs we only will know if the
755                  * specified vmlinux file is invalid when we actually have a
756                  * hit in kernel space and then try to load it. So if we get
757                  * here and there are _no_ symbols in the DSO backing the
758                  * kernel map, bail out.
759                  *
760                  * We may never get here, for instance, if we use -K/
761                  * --hide-kernel-symbols, even if the user specifies an
762                  * invalid --vmlinux ;-)
763                  */
764                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
765                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
766                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
767                         if (symbol_conf.vmlinux_name) {
768                                 ui__warning("The %s file can't be used.\n%s",
769                                             symbol_conf.vmlinux_name, msg);
770                         } else {
771                                 ui__warning("A vmlinux file was not found.\n%s",
772                                             msg);
773                         }
774
775                         if (use_browser <= 0)
776                                 sleep(5);
777                         top->vmlinux_warned = true;
778                 }
779         }
780
781         if (al.sym == NULL || !al.sym->ignore) {
782                 struct hist_entry *he;
783
784                 if ((sort__has_parent || symbol_conf.use_callchain) &&
785                     sample->callchain) {
786                         err = machine__resolve_callchain(machine, al.thread,
787                                                          sample->callchain, &parent);
788                         if (err)
789                                 return;
790                 }
791
792                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
793                 if (he == NULL) {
794                         pr_err("Problem incrementing symbol period, skipping event\n");
795                         return;
796                 }
797
798                 if (symbol_conf.use_callchain) {
799                         err = callchain_append(he->callchain, &callchain_cursor,
800                                                sample->period);
801                         if (err)
802                                 return;
803                 }
804
805                 if (top->sort_has_symbols)
806                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
807         }
808
809         return;
810 }
811
812 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
813 {
814         struct perf_sample sample;
815         struct perf_evsel *evsel;
816         struct perf_session *session = top->session;
817         union perf_event *event;
818         struct machine *machine;
819         u8 origin;
820         int ret;
821
822         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
823                 ret = perf_evlist__parse_sample(top->evlist, event, &sample, false);
824                 if (ret) {
825                         pr_err("Can't parse sample, err = %d\n", ret);
826                         continue;
827                 }
828
829                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
830                 assert(evsel != NULL);
831
832                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
833
834                 if (event->header.type == PERF_RECORD_SAMPLE)
835                         ++top->samples;
836
837                 switch (origin) {
838                 case PERF_RECORD_MISC_USER:
839                         ++top->us_samples;
840                         if (top->hide_user_symbols)
841                                 continue;
842                         machine = perf_session__find_host_machine(session);
843                         break;
844                 case PERF_RECORD_MISC_KERNEL:
845                         ++top->kernel_samples;
846                         if (top->hide_kernel_symbols)
847                                 continue;
848                         machine = perf_session__find_host_machine(session);
849                         break;
850                 case PERF_RECORD_MISC_GUEST_KERNEL:
851                         ++top->guest_kernel_samples;
852                         machine = perf_session__find_machine(session, event->ip.pid);
853                         break;
854                 case PERF_RECORD_MISC_GUEST_USER:
855                         ++top->guest_us_samples;
856                         /*
857                          * TODO: we don't process guest user from host side
858                          * except simple counting.
859                          */
860                         /* Fall thru */
861                 default:
862                         continue;
863                 }
864
865
866                 if (event->header.type == PERF_RECORD_SAMPLE) {
867                         perf_event__process_sample(&top->tool, event, evsel,
868                                                    &sample, machine);
869                 } else if (event->header.type < PERF_RECORD_MAX) {
870                         hists__inc_nr_events(&evsel->hists, event->header.type);
871                         perf_event__process(&top->tool, event, &sample, machine);
872                 } else
873                         ++session->hists.stats.nr_unknown_events;
874         }
875 }
876
877 static void perf_top__mmap_read(struct perf_top *top)
878 {
879         int i;
880
881         for (i = 0; i < top->evlist->nr_mmaps; i++)
882                 perf_top__mmap_read_idx(top, i);
883 }
884
885 static void perf_top__start_counters(struct perf_top *top)
886 {
887         struct perf_evsel *counter, *first;
888         struct perf_evlist *evlist = top->evlist;
889
890         first = list_entry(evlist->entries.next, struct perf_evsel, node);
891
892         list_for_each_entry(counter, &evlist->entries, node) {
893                 struct perf_event_attr *attr = &counter->attr;
894                 struct xyarray *group_fd = NULL;
895
896                 if (top->group && counter != first)
897                         group_fd = first->fd;
898
899                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
900
901                 if (top->freq) {
902                         attr->sample_type |= PERF_SAMPLE_PERIOD;
903                         attr->freq        = 1;
904                         attr->sample_freq = top->freq;
905                 }
906
907                 if (evlist->nr_entries > 1) {
908                         attr->sample_type |= PERF_SAMPLE_ID;
909                         attr->read_format |= PERF_FORMAT_ID;
910                 }
911
912                 if (perf_target__has_cpu(&top->target))
913                         attr->sample_type |= PERF_SAMPLE_CPU;
914
915                 if (symbol_conf.use_callchain)
916                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
917
918                 attr->mmap = 1;
919                 attr->comm = 1;
920                 attr->inherit = top->inherit;
921 fallback_missing_features:
922                 if (top->exclude_guest_missing)
923                         attr->exclude_guest = attr->exclude_host = 0;
924 retry_sample_id:
925                 attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
926 try_again:
927                 if (perf_evsel__open(counter, top->evlist->cpus,
928                                      top->evlist->threads, top->group,
929                                      group_fd) < 0) {
930                         int err = errno;
931
932                         if (err == EPERM || err == EACCES) {
933                                 ui__error_paranoid();
934                                 goto out_err;
935                         } else if (err == EINVAL) {
936                                 if (!top->exclude_guest_missing &&
937                                     (attr->exclude_guest || attr->exclude_host)) {
938                                         pr_debug("Old kernel, cannot exclude "
939                                                  "guest or host samples.\n");
940                                         top->exclude_guest_missing = true;
941                                         goto fallback_missing_features;
942                                 } else if (!top->sample_id_all_missing) {
943                                         /*
944                                          * Old kernel, no attr->sample_id_type_all field
945                                          */
946                                         top->sample_id_all_missing = true;
947                                         goto retry_sample_id;
948                                 }
949                         }
950                         /*
951                          * If it's cycles then fall back to hrtimer
952                          * based cpu-clock-tick sw counter, which
953                          * is always available even if no PMU support:
954                          */
955                         if ((err == ENOENT || err == ENXIO) &&
956                             (attr->type == PERF_TYPE_HARDWARE) &&
957                             (attr->config == PERF_COUNT_HW_CPU_CYCLES)) {
958
959                                 if (verbose)
960                                         ui__warning("Cycles event not supported,\n"
961                                                     "trying to fall back to cpu-clock-ticks\n");
962
963                                 attr->type = PERF_TYPE_SOFTWARE;
964                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
965                                 if (counter->name) {
966                                         free(counter->name);
967                                         counter->name = NULL;
968                                 }
969                                 goto try_again;
970                         }
971
972                         if (err == ENOENT) {
973                                 ui__error("The %s event is not supported.\n",
974                                           perf_evsel__name(counter));
975                                 goto out_err;
976                         } else if (err == EMFILE) {
977                                 ui__error("Too many events are opened.\n"
978                                             "Try again after reducing the number of events\n");
979                                 goto out_err;
980                         }
981
982                         ui__error("The sys_perf_event_open() syscall "
983                                     "returned with %d (%s).  /bin/dmesg "
984                                     "may provide additional information.\n"
985                                     "No CONFIG_PERF_EVENTS=y kernel support "
986                                     "configured?\n", err, strerror(err));
987                         goto out_err;
988                 }
989         }
990
991         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
992                 ui__error("Failed to mmap with %d (%s)\n",
993                             errno, strerror(errno));
994                 goto out_err;
995         }
996
997         return;
998
999 out_err:
1000         exit_browser(0);
1001         exit(0);
1002 }
1003
1004 static int perf_top__setup_sample_type(struct perf_top *top)
1005 {
1006         if (!top->sort_has_symbols) {
1007                 if (symbol_conf.use_callchain) {
1008                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1009                         return -EINVAL;
1010                 }
1011         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
1012                 if (callchain_register_param(&callchain_param) < 0) {
1013                         ui__error("Can't register callchain params.\n");
1014                         return -EINVAL;
1015                 }
1016         }
1017
1018         return 0;
1019 }
1020
1021 static int __cmd_top(struct perf_top *top)
1022 {
1023         pthread_t thread;
1024         int ret;
1025         /*
1026          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1027          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1028          */
1029         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1030         if (top->session == NULL)
1031                 return -ENOMEM;
1032
1033         ret = perf_top__setup_sample_type(top);
1034         if (ret)
1035                 goto out_delete;
1036
1037         if (perf_target__has_task(&top->target))
1038                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
1039                                                   perf_event__process,
1040                                                   &top->session->host_machine);
1041         else
1042                 perf_event__synthesize_threads(&top->tool, perf_event__process,
1043                                                &top->session->host_machine);
1044         perf_top__start_counters(top);
1045         top->session->evlist = top->evlist;
1046         perf_session__set_id_hdr_size(top->session);
1047
1048         /* Wait for a minimal set of events before starting the snapshot */
1049         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1050
1051         perf_top__mmap_read(top);
1052
1053         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1054                                                             display_thread), top)) {
1055                 ui__error("Could not create display thread.\n");
1056                 exit(-1);
1057         }
1058
1059         if (top->realtime_prio) {
1060                 struct sched_param param;
1061
1062                 param.sched_priority = top->realtime_prio;
1063                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1064                         ui__error("Could not set realtime priority.\n");
1065                         exit(-1);
1066                 }
1067         }
1068
1069         while (1) {
1070                 u64 hits = top->samples;
1071
1072                 perf_top__mmap_read(top);
1073
1074                 if (hits == top->samples)
1075                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1076         }
1077
1078 out_delete:
1079         perf_session__delete(top->session);
1080         top->session = NULL;
1081
1082         return 0;
1083 }
1084
1085 static int
1086 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1087 {
1088         struct perf_top *top = (struct perf_top *)opt->value;
1089         char *tok, *tok2;
1090         char *endptr;
1091
1092         /*
1093          * --no-call-graph
1094          */
1095         if (unset) {
1096                 top->dont_use_callchains = true;
1097                 return 0;
1098         }
1099
1100         symbol_conf.use_callchain = true;
1101
1102         if (!arg)
1103                 return 0;
1104
1105         tok = strtok((char *)arg, ",");
1106         if (!tok)
1107                 return -1;
1108
1109         /* get the output mode */
1110         if (!strncmp(tok, "graph", strlen(arg)))
1111                 callchain_param.mode = CHAIN_GRAPH_ABS;
1112
1113         else if (!strncmp(tok, "flat", strlen(arg)))
1114                 callchain_param.mode = CHAIN_FLAT;
1115
1116         else if (!strncmp(tok, "fractal", strlen(arg)))
1117                 callchain_param.mode = CHAIN_GRAPH_REL;
1118
1119         else if (!strncmp(tok, "none", strlen(arg))) {
1120                 callchain_param.mode = CHAIN_NONE;
1121                 symbol_conf.use_callchain = false;
1122
1123                 return 0;
1124         } else
1125                 return -1;
1126
1127         /* get the min percentage */
1128         tok = strtok(NULL, ",");
1129         if (!tok)
1130                 goto setup;
1131
1132         callchain_param.min_percent = strtod(tok, &endptr);
1133         if (tok == endptr)
1134                 return -1;
1135
1136         /* get the print limit */
1137         tok2 = strtok(NULL, ",");
1138         if (!tok2)
1139                 goto setup;
1140
1141         if (tok2[0] != 'c') {
1142                 callchain_param.print_limit = strtod(tok2, &endptr);
1143                 tok2 = strtok(NULL, ",");
1144                 if (!tok2)
1145                         goto setup;
1146         }
1147
1148         /* get the call chain order */
1149         if (!strcmp(tok2, "caller"))
1150                 callchain_param.order = ORDER_CALLER;
1151         else if (!strcmp(tok2, "callee"))
1152                 callchain_param.order = ORDER_CALLEE;
1153         else
1154                 return -1;
1155 setup:
1156         if (callchain_register_param(&callchain_param) < 0) {
1157                 fprintf(stderr, "Can't register callchain params\n");
1158                 return -1;
1159         }
1160         return 0;
1161 }
1162
1163 static const char * const top_usage[] = {
1164         "perf top [<options>]",
1165         NULL
1166 };
1167
1168 int cmd_top(int argc, const char **argv, const char *prefix __used)
1169 {
1170         struct perf_evsel *pos;
1171         int status;
1172         char errbuf[BUFSIZ];
1173         struct perf_top top = {
1174                 .count_filter        = 5,
1175                 .delay_secs          = 2,
1176                 .freq                = 4000, /* 4 KHz */
1177                 .mmap_pages          = 128,
1178                 .sym_pcnt_filter     = 5,
1179                 .target              = {
1180                         .uses_mmap   = true,
1181                 },
1182         };
1183         char callchain_default_opt[] = "fractal,0.5,callee";
1184         const struct option options[] = {
1185         OPT_CALLBACK('e', "event", &top.evlist, "event",
1186                      "event selector. use 'perf list' to list available events",
1187                      parse_events_option),
1188         OPT_INTEGER('c', "count", &top.default_interval,
1189                     "event period to sample"),
1190         OPT_STRING('p', "pid", &top.target.pid, "pid",
1191                     "profile events on existing process id"),
1192         OPT_STRING('t', "tid", &top.target.tid, "tid",
1193                     "profile events on existing thread id"),
1194         OPT_BOOLEAN('a', "all-cpus", &top.target.system_wide,
1195                             "system-wide collection from all CPUs"),
1196         OPT_STRING('C', "cpu", &top.target.cpu_list, "cpu",
1197                     "list of cpus to monitor"),
1198         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1199                    "file", "vmlinux pathname"),
1200         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1201                     "hide kernel symbols"),
1202         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1203         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1204                     "collect data with this RT SCHED_FIFO priority"),
1205         OPT_INTEGER('d', "delay", &top.delay_secs,
1206                     "number of seconds to delay between refreshes"),
1207         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1208                             "dump the symbol table used for profiling"),
1209         OPT_INTEGER('f', "count-filter", &top.count_filter,
1210                     "only display functions with more events than this"),
1211         OPT_BOOLEAN('g', "group", &top.group,
1212                             "put the counters into a counter group"),
1213         OPT_BOOLEAN('i', "inherit", &top.inherit,
1214                     "child tasks inherit counters"),
1215         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1216                     "symbol to annotate"),
1217         OPT_BOOLEAN('z', "zero", &top.zero,
1218                     "zero history across updates"),
1219         OPT_INTEGER('F', "freq", &top.freq,
1220                     "profile at this frequency"),
1221         OPT_INTEGER('E', "entries", &top.print_entries,
1222                     "display this many functions"),
1223         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1224                     "hide user symbols"),
1225         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1226         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1227         OPT_INCR('v', "verbose", &verbose,
1228                     "be more verbose (show counter open errors, etc)"),
1229         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1230                    "sort by key(s): pid, comm, dso, symbol, parent"),
1231         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1232                     "Show a column with the number of samples"),
1233         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1234                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1235                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1236                      callchain_default_opt),
1237         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1238                     "Show a column with the sum of periods"),
1239         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1240                    "only consider symbols in these dsos"),
1241         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1242                    "only consider symbols in these comms"),
1243         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1244                    "only consider these symbols"),
1245         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1246                     "Interleave source code with assembly code (default)"),
1247         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1248                     "Display raw encoding of assembly instructions (default)"),
1249         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1250                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1251         OPT_STRING('u', "uid", &top.target.uid_str, "user", "user to profile"),
1252         OPT_END()
1253         };
1254
1255         top.evlist = perf_evlist__new(NULL, NULL);
1256         if (top.evlist == NULL)
1257                 return -ENOMEM;
1258
1259         symbol_conf.exclude_other = false;
1260
1261         argc = parse_options(argc, argv, options, top_usage, 0);
1262         if (argc)
1263                 usage_with_options(top_usage, options);
1264
1265         if (sort_order == default_sort_order)
1266                 sort_order = "dso,symbol";
1267
1268         setup_sorting(top_usage, options);
1269
1270         if (top.use_stdio)
1271                 use_browser = 0;
1272         else if (top.use_tui)
1273                 use_browser = 1;
1274
1275         setup_browser(false);
1276
1277         status = perf_target__validate(&top.target);
1278         if (status) {
1279                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1280                 ui__warning("%s", errbuf);
1281         }
1282
1283         status = perf_target__parse_uid(&top.target);
1284         if (status) {
1285                 int saved_errno = errno;
1286
1287                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1288                 ui__error("%s", errbuf);
1289
1290                 status = -saved_errno;
1291                 goto out_delete_evlist;
1292         }
1293
1294         if (perf_target__none(&top.target))
1295                 top.target.system_wide = true;
1296
1297         if (perf_evlist__create_maps(top.evlist, &top.target) < 0)
1298                 usage_with_options(top_usage, options);
1299
1300         if (!top.evlist->nr_entries &&
1301             perf_evlist__add_default(top.evlist) < 0) {
1302                 ui__error("Not enough memory for event selector list\n");
1303                 return -ENOMEM;
1304         }
1305
1306         symbol_conf.nr_events = top.evlist->nr_entries;
1307
1308         if (top.delay_secs < 1)
1309                 top.delay_secs = 1;
1310
1311         /*
1312          * User specified count overrides default frequency.
1313          */
1314         if (top.default_interval)
1315                 top.freq = 0;
1316         else if (top.freq) {
1317                 top.default_interval = top.freq;
1318         } else {
1319                 ui__error("frequency and count are zero, aborting\n");
1320                 exit(EXIT_FAILURE);
1321         }
1322
1323         list_for_each_entry(pos, &top.evlist->entries, node) {
1324                 /*
1325                  * Fill in the ones not specifically initialized via -c:
1326                  */
1327                 if (!pos->attr.sample_period)
1328                         pos->attr.sample_period = top.default_interval;
1329         }
1330
1331         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1332
1333         symbol_conf.priv_size = sizeof(struct annotation);
1334
1335         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1336         if (symbol__init() < 0)
1337                 return -1;
1338
1339         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1340         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1341         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1342
1343         /*
1344          * Avoid annotation data structures overhead when symbols aren't on the
1345          * sort list.
1346          */
1347         top.sort_has_symbols = sort_sym.list.next != NULL;
1348
1349         get_term_dimensions(&top.winsize);
1350         if (top.print_entries == 0) {
1351                 struct sigaction act = {
1352                         .sa_sigaction = perf_top__sig_winch,
1353                         .sa_flags     = SA_SIGINFO,
1354                 };
1355                 perf_top__update_print_entries(&top);
1356                 sigaction(SIGWINCH, &act, NULL);
1357         }
1358
1359         status = __cmd_top(&top);
1360
1361 out_delete_evlist:
1362         perf_evlist__delete(top.evlist);
1363
1364         return status;
1365 }