]> Pileus Git - ~andy/linux/blob - tools/perf/builtin-top.c
Merge branch 'for-3.7' of git://git.infradead.org/users/dedekind/l2-mtd
[~andy/linux] / 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 __maybe_unused,
99                                 siginfo_t *info __maybe_unused, void *arg)
100 {
101         struct perf_top *top = arg;
102
103         get_term_dimensions(&top->winsize);
104         if (!top->print_entries
105             || (top->print_entries+4) > top->winsize.ws_row) {
106                 top->print_entries = top->winsize.ws_row;
107         } else {
108                 top->print_entries += 4;
109                 top->winsize.ws_row = top->print_entries;
110         }
111         perf_top__update_print_entries(top);
112 }
113
114 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
115 {
116         struct symbol *sym;
117         struct annotation *notes;
118         struct map *map;
119         int err = -1;
120
121         if (!he || !he->ms.sym)
122                 return -1;
123
124         sym = he->ms.sym;
125         map = he->ms.map;
126
127         /*
128          * We can't annotate with just /proc/kallsyms
129          */
130         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
131                 pr_err("Can't annotate %s: No vmlinux file was found in the "
132                        "path\n", sym->name);
133                 sleep(1);
134                 return -1;
135         }
136
137         notes = symbol__annotation(sym);
138         if (notes->src != NULL) {
139                 pthread_mutex_lock(&notes->lock);
140                 goto out_assign;
141         }
142
143         pthread_mutex_lock(&notes->lock);
144
145         if (symbol__alloc_hist(sym) < 0) {
146                 pthread_mutex_unlock(&notes->lock);
147                 pr_err("Not enough memory for annotating '%s' symbol!\n",
148                        sym->name);
149                 sleep(1);
150                 return err;
151         }
152
153         err = symbol__annotate(sym, map, 0);
154         if (err == 0) {
155 out_assign:
156                 top->sym_filter_entry = he;
157         }
158
159         pthread_mutex_unlock(&notes->lock);
160         return err;
161 }
162
163 static void __zero_source_counters(struct hist_entry *he)
164 {
165         struct symbol *sym = he->ms.sym;
166         symbol__annotate_zero_histograms(sym);
167 }
168
169 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
170 {
171         struct utsname uts;
172         int err = uname(&uts);
173
174         ui__warning("Out of bounds address found:\n\n"
175                     "Addr:   %" PRIx64 "\n"
176                     "DSO:    %s %c\n"
177                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
178                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
179                     "Arch:   %s\n"
180                     "Kernel: %s\n"
181                     "Tools:  %s\n\n"
182                     "Not all samples will be on the annotation output.\n\n"
183                     "Please report to linux-kernel@vger.kernel.org\n",
184                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
185                     map->start, map->end, sym->start, sym->end,
186                     sym->binding == STB_GLOBAL ? 'g' :
187                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
188                     err ? "[unknown]" : uts.machine,
189                     err ? "[unknown]" : uts.release, perf_version_string);
190         if (use_browser <= 0)
191                 sleep(5);
192         
193         map->erange_warned = true;
194 }
195
196 static void perf_top__record_precise_ip(struct perf_top *top,
197                                         struct hist_entry *he,
198                                         int counter, u64 ip)
199 {
200         struct annotation *notes;
201         struct symbol *sym;
202         int err;
203
204         if (he == NULL || he->ms.sym == NULL ||
205             ((top->sym_filter_entry == NULL ||
206               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
207                 return;
208
209         sym = he->ms.sym;
210         notes = symbol__annotation(sym);
211
212         if (pthread_mutex_trylock(&notes->lock))
213                 return;
214
215         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
216                 pthread_mutex_unlock(&notes->lock);
217                 pr_err("Not enough memory for annotating '%s' symbol!\n",
218                        sym->name);
219                 sleep(1);
220                 return;
221         }
222
223         ip = he->ms.map->map_ip(he->ms.map, ip);
224         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
225
226         pthread_mutex_unlock(&notes->lock);
227
228         if (err == -ERANGE && !he->ms.map->erange_warned)
229                 ui__warn_map_erange(he->ms.map, sym, ip);
230 }
231
232 static void perf_top__show_details(struct perf_top *top)
233 {
234         struct hist_entry *he = top->sym_filter_entry;
235         struct annotation *notes;
236         struct symbol *symbol;
237         int more;
238
239         if (!he)
240                 return;
241
242         symbol = he->ms.sym;
243         notes = symbol__annotation(symbol);
244
245         pthread_mutex_lock(&notes->lock);
246
247         if (notes->src == NULL)
248                 goto out_unlock;
249
250         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
251         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
252
253         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
254                                        0, top->sym_pcnt_filter, top->print_entries, 4);
255         if (top->zero)
256                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
257         else
258                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
259         if (more != 0)
260                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
261 out_unlock:
262         pthread_mutex_unlock(&notes->lock);
263 }
264
265 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
266
267 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
268                                                      struct addr_location *al,
269                                                      struct perf_sample *sample)
270 {
271         struct hist_entry *he;
272
273         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
274         if (he == NULL)
275                 return NULL;
276
277         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
278         return he;
279 }
280
281 static void perf_top__print_sym_table(struct perf_top *top)
282 {
283         char bf[160];
284         int printed = 0;
285         const int win_width = top->winsize.ws_col - 1;
286
287         puts(CONSOLE_CLEAR);
288
289         perf_top__header_snprintf(top, bf, sizeof(bf));
290         printf("%s\n", bf);
291
292         perf_top__reset_sample_counters(top);
293
294         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
295
296         if (top->sym_evsel->hists.stats.nr_lost_warned !=
297             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
298                 top->sym_evsel->hists.stats.nr_lost_warned =
299                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
300                 color_fprintf(stdout, PERF_COLOR_RED,
301                               "WARNING: LOST %d chunks, Check IO/CPU overload",
302                               top->sym_evsel->hists.stats.nr_lost_warned);
303                 ++printed;
304         }
305
306         if (top->sym_filter_entry) {
307                 perf_top__show_details(top);
308                 return;
309         }
310
311         hists__collapse_resort_threaded(&top->sym_evsel->hists);
312         hists__output_resort_threaded(&top->sym_evsel->hists);
313         hists__decay_entries_threaded(&top->sym_evsel->hists,
314                                       top->hide_user_symbols,
315                                       top->hide_kernel_symbols);
316         hists__output_recalc_col_len(&top->sym_evsel->hists,
317                                      top->winsize.ws_row - 3);
318         putchar('\n');
319         hists__fprintf(&top->sym_evsel->hists, false,
320                        top->winsize.ws_row - 4 - printed, win_width, stdout);
321 }
322
323 static void prompt_integer(int *target, const char *msg)
324 {
325         char *buf = malloc(0), *p;
326         size_t dummy = 0;
327         int tmp;
328
329         fprintf(stdout, "\n%s: ", msg);
330         if (getline(&buf, &dummy, stdin) < 0)
331                 return;
332
333         p = strchr(buf, '\n');
334         if (p)
335                 *p = 0;
336
337         p = buf;
338         while(*p) {
339                 if (!isdigit(*p))
340                         goto out_free;
341                 p++;
342         }
343         tmp = strtoul(buf, NULL, 10);
344         *target = tmp;
345 out_free:
346         free(buf);
347 }
348
349 static void prompt_percent(int *target, const char *msg)
350 {
351         int tmp = 0;
352
353         prompt_integer(&tmp, msg);
354         if (tmp >= 0 && tmp <= 100)
355                 *target = tmp;
356 }
357
358 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
359 {
360         char *buf = malloc(0), *p;
361         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
362         struct rb_node *next;
363         size_t dummy = 0;
364
365         /* zero counters of active symbol */
366         if (syme) {
367                 __zero_source_counters(syme);
368                 top->sym_filter_entry = NULL;
369         }
370
371         fprintf(stdout, "\n%s: ", msg);
372         if (getline(&buf, &dummy, stdin) < 0)
373                 goto out_free;
374
375         p = strchr(buf, '\n');
376         if (p)
377                 *p = 0;
378
379         next = rb_first(&top->sym_evsel->hists.entries);
380         while (next) {
381                 n = rb_entry(next, struct hist_entry, rb_node);
382                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
383                         found = n;
384                         break;
385                 }
386                 next = rb_next(&n->rb_node);
387         }
388
389         if (!found) {
390                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
391                 sleep(1);
392         } else
393                 perf_top__parse_source(top, found);
394
395 out_free:
396         free(buf);
397 }
398
399 static void perf_top__print_mapped_keys(struct perf_top *top)
400 {
401         char *name = NULL;
402
403         if (top->sym_filter_entry) {
404                 struct symbol *sym = top->sym_filter_entry->ms.sym;
405                 name = sym->name;
406         }
407
408         fprintf(stdout, "\nMapped keys:\n");
409         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
410         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
411
412         if (top->evlist->nr_entries > 1)
413                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
414
415         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
416
417         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
418         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
419         fprintf(stdout, "\t[S]     stop annotation.\n");
420
421         fprintf(stdout,
422                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
423                 top->hide_kernel_symbols ? "yes" : "no");
424         fprintf(stdout,
425                 "\t[U]     hide user symbols.               \t(%s)\n",
426                 top->hide_user_symbols ? "yes" : "no");
427         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
428         fprintf(stdout, "\t[qQ]    quit.\n");
429 }
430
431 static int perf_top__key_mapped(struct perf_top *top, int c)
432 {
433         switch (c) {
434                 case 'd':
435                 case 'e':
436                 case 'f':
437                 case 'z':
438                 case 'q':
439                 case 'Q':
440                 case 'K':
441                 case 'U':
442                 case 'F':
443                 case 's':
444                 case 'S':
445                         return 1;
446                 case 'E':
447                         return top->evlist->nr_entries > 1 ? 1 : 0;
448                 default:
449                         break;
450         }
451
452         return 0;
453 }
454
455 static void perf_top__handle_keypress(struct perf_top *top, int c)
456 {
457         if (!perf_top__key_mapped(top, c)) {
458                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
459                 struct termios tc, save;
460
461                 perf_top__print_mapped_keys(top);
462                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
463                 fflush(stdout);
464
465                 tcgetattr(0, &save);
466                 tc = save;
467                 tc.c_lflag &= ~(ICANON | ECHO);
468                 tc.c_cc[VMIN] = 0;
469                 tc.c_cc[VTIME] = 0;
470                 tcsetattr(0, TCSANOW, &tc);
471
472                 poll(&stdin_poll, 1, -1);
473                 c = getc(stdin);
474
475                 tcsetattr(0, TCSAFLUSH, &save);
476                 if (!perf_top__key_mapped(top, c))
477                         return;
478         }
479
480         switch (c) {
481                 case 'd':
482                         prompt_integer(&top->delay_secs, "Enter display delay");
483                         if (top->delay_secs < 1)
484                                 top->delay_secs = 1;
485                         break;
486                 case 'e':
487                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
488                         if (top->print_entries == 0) {
489                                 struct sigaction act = {
490                                         .sa_sigaction = perf_top__sig_winch,
491                                         .sa_flags     = SA_SIGINFO,
492                                 };
493                                 perf_top__sig_winch(SIGWINCH, NULL, top);
494                                 sigaction(SIGWINCH, &act, NULL);
495                         } else {
496                                 perf_top__sig_winch(SIGWINCH, NULL, top);
497                                 signal(SIGWINCH, SIG_DFL);
498                         }
499                         break;
500                 case 'E':
501                         if (top->evlist->nr_entries > 1) {
502                                 /* Select 0 as the default event: */
503                                 int counter = 0;
504
505                                 fprintf(stderr, "\nAvailable events:");
506
507                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
508                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
509
510                                 prompt_integer(&counter, "Enter details event counter");
511
512                                 if (counter >= top->evlist->nr_entries) {
513                                         top->sym_evsel = perf_evlist__first(top->evlist);
514                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
515                                         sleep(1);
516                                         break;
517                                 }
518                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
519                                         if (top->sym_evsel->idx == counter)
520                                                 break;
521                         } else
522                                 top->sym_evsel = perf_evlist__first(top->evlist);
523                         break;
524                 case 'f':
525                         prompt_integer(&top->count_filter, "Enter display event count filter");
526                         break;
527                 case 'F':
528                         prompt_percent(&top->sym_pcnt_filter,
529                                        "Enter details display event filter (percent)");
530                         break;
531                 case 'K':
532                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
533                         break;
534                 case 'q':
535                 case 'Q':
536                         printf("exiting.\n");
537                         if (top->dump_symtab)
538                                 perf_session__fprintf_dsos(top->session, stderr);
539                         exit(0);
540                 case 's':
541                         perf_top__prompt_symbol(top, "Enter details symbol");
542                         break;
543                 case 'S':
544                         if (!top->sym_filter_entry)
545                                 break;
546                         else {
547                                 struct hist_entry *syme = top->sym_filter_entry;
548
549                                 top->sym_filter_entry = NULL;
550                                 __zero_source_counters(syme);
551                         }
552                         break;
553                 case 'U':
554                         top->hide_user_symbols = !top->hide_user_symbols;
555                         break;
556                 case 'z':
557                         top->zero = !top->zero;
558                         break;
559                 default:
560                         break;
561         }
562 }
563
564 static void perf_top__sort_new_samples(void *arg)
565 {
566         struct perf_top *t = arg;
567         perf_top__reset_sample_counters(t);
568
569         if (t->evlist->selected != NULL)
570                 t->sym_evsel = t->evlist->selected;
571
572         hists__collapse_resort_threaded(&t->sym_evsel->hists);
573         hists__output_resort_threaded(&t->sym_evsel->hists);
574         hists__decay_entries_threaded(&t->sym_evsel->hists,
575                                       t->hide_user_symbols,
576                                       t->hide_kernel_symbols);
577 }
578
579 static void *display_thread_tui(void *arg)
580 {
581         struct perf_evsel *pos;
582         struct perf_top *top = arg;
583         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
584
585         perf_top__sort_new_samples(top);
586
587         /*
588          * Initialize the uid_filter_str, in the future the TUI will allow
589          * Zooming in/out UIDs. For now juse use whatever the user passed
590          * via --uid.
591          */
592         list_for_each_entry(pos, &top->evlist->entries, node)
593                 pos->hists.uid_filter_str = top->target.uid_str;
594
595         perf_evlist__tui_browse_hists(top->evlist, help,
596                                       perf_top__sort_new_samples,
597                                       top, top->delay_secs);
598
599         exit_browser(0);
600         exit(0);
601         return NULL;
602 }
603
604 static void *display_thread(void *arg)
605 {
606         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
607         struct termios tc, save;
608         struct perf_top *top = arg;
609         int delay_msecs, c;
610
611         tcgetattr(0, &save);
612         tc = save;
613         tc.c_lflag &= ~(ICANON | ECHO);
614         tc.c_cc[VMIN] = 0;
615         tc.c_cc[VTIME] = 0;
616
617         pthread__unblock_sigwinch();
618 repeat:
619         delay_msecs = top->delay_secs * 1000;
620         tcsetattr(0, TCSANOW, &tc);
621         /* trash return*/
622         getc(stdin);
623
624         while (1) {
625                 perf_top__print_sym_table(top);
626                 /*
627                  * Either timeout expired or we got an EINTR due to SIGWINCH,
628                  * refresh screen in both cases.
629                  */
630                 switch (poll(&stdin_poll, 1, delay_msecs)) {
631                 case 0:
632                         continue;
633                 case -1:
634                         if (errno == EINTR)
635                                 continue;
636                         /* Fall trhu */
637                 default:
638                         goto process_hotkey;
639                 }
640         }
641 process_hotkey:
642         c = getc(stdin);
643         tcsetattr(0, TCSAFLUSH, &save);
644
645         perf_top__handle_keypress(top, c);
646         goto repeat;
647
648         return NULL;
649 }
650
651 /* Tag samples to be skipped. */
652 static const char *skip_symbols[] = {
653         "intel_idle",
654         "default_idle",
655         "native_safe_halt",
656         "cpu_idle",
657         "enter_idle",
658         "exit_idle",
659         "mwait_idle",
660         "mwait_idle_with_hints",
661         "poll_idle",
662         "ppc64_runlatch_off",
663         "pseries_dedicated_idle_sleep",
664         NULL
665 };
666
667 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
668 {
669         const char *name = sym->name;
670         int i;
671
672         /*
673          * ppc64 uses function descriptors and appends a '.' to the
674          * start of every instruction address. Remove it.
675          */
676         if (name[0] == '.')
677                 name++;
678
679         if (!strcmp(name, "_text") ||
680             !strcmp(name, "_etext") ||
681             !strcmp(name, "_sinittext") ||
682             !strncmp("init_module", name, 11) ||
683             !strncmp("cleanup_module", name, 14) ||
684             strstr(name, "_text_start") ||
685             strstr(name, "_text_end"))
686                 return 1;
687
688         for (i = 0; skip_symbols[i]; i++) {
689                 if (!strcmp(skip_symbols[i], name)) {
690                         sym->ignore = true;
691                         break;
692                 }
693         }
694
695         return 0;
696 }
697
698 static void perf_event__process_sample(struct perf_tool *tool,
699                                        const union perf_event *event,
700                                        struct perf_evsel *evsel,
701                                        struct perf_sample *sample,
702                                        struct machine *machine)
703 {
704         struct perf_top *top = container_of(tool, struct perf_top, tool);
705         struct symbol *parent = NULL;
706         u64 ip = event->ip.ip;
707         struct addr_location al;
708         int err;
709
710         if (!machine && perf_guest) {
711                 static struct intlist *seen;
712
713                 if (!seen)
714                         seen = intlist__new();
715
716                 if (!intlist__has_entry(seen, event->ip.pid)) {
717                         pr_err("Can't find guest [%d]'s kernel information\n",
718                                 event->ip.pid);
719                         intlist__add(seen, event->ip.pid);
720                 }
721                 return;
722         }
723
724         if (!machine) {
725                 pr_err("%u unprocessable samples recorded.",
726                        top->session->hists.stats.nr_unprocessable_samples++);
727                 return;
728         }
729
730         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
731                 top->exact_samples++;
732
733         if (perf_event__preprocess_sample(event, machine, &al, sample,
734                                           symbol_filter) < 0 ||
735             al.filtered)
736                 return;
737
738         if (!top->kptr_restrict_warned &&
739             symbol_conf.kptr_restrict &&
740             al.cpumode == PERF_RECORD_MISC_KERNEL) {
741                 ui__warning(
742 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
743 "Check /proc/sys/kernel/kptr_restrict.\n\n"
744 "Kernel%s samples will not be resolved.\n",
745                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
746                           " modules" : "");
747                 if (use_browser <= 0)
748                         sleep(5);
749                 top->kptr_restrict_warned = true;
750         }
751
752         if (al.sym == NULL) {
753                 const char *msg = "Kernel samples will not be resolved.\n";
754                 /*
755                  * As we do lazy loading of symtabs we only will know if the
756                  * specified vmlinux file is invalid when we actually have a
757                  * hit in kernel space and then try to load it. So if we get
758                  * here and there are _no_ symbols in the DSO backing the
759                  * kernel map, bail out.
760                  *
761                  * We may never get here, for instance, if we use -K/
762                  * --hide-kernel-symbols, even if the user specifies an
763                  * invalid --vmlinux ;-)
764                  */
765                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
766                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
767                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
768                         if (symbol_conf.vmlinux_name) {
769                                 ui__warning("The %s file can't be used.\n%s",
770                                             symbol_conf.vmlinux_name, msg);
771                         } else {
772                                 ui__warning("A vmlinux file was not found.\n%s",
773                                             msg);
774                         }
775
776                         if (use_browser <= 0)
777                                 sleep(5);
778                         top->vmlinux_warned = true;
779                 }
780         }
781
782         if (al.sym == NULL || !al.sym->ignore) {
783                 struct hist_entry *he;
784
785                 if ((sort__has_parent || symbol_conf.use_callchain) &&
786                     sample->callchain) {
787                         err = machine__resolve_callchain(machine, evsel,
788                                                          al.thread, sample,
789                                                          &parent);
790
791                         if (err)
792                                 return;
793                 }
794
795                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
796                 if (he == NULL) {
797                         pr_err("Problem incrementing symbol period, skipping event\n");
798                         return;
799                 }
800
801                 if (symbol_conf.use_callchain) {
802                         err = callchain_append(he->callchain, &callchain_cursor,
803                                                sample->period);
804                         if (err)
805                                 return;
806                 }
807
808                 if (top->sort_has_symbols)
809                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
810         }
811
812         return;
813 }
814
815 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
816 {
817         struct perf_sample sample;
818         struct perf_evsel *evsel;
819         struct perf_session *session = top->session;
820         union perf_event *event;
821         struct machine *machine;
822         u8 origin;
823         int ret;
824
825         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
826                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
827                 if (ret) {
828                         pr_err("Can't parse sample, err = %d\n", ret);
829                         continue;
830                 }
831
832                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
833                 assert(evsel != NULL);
834
835                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
836
837                 if (event->header.type == PERF_RECORD_SAMPLE)
838                         ++top->samples;
839
840                 switch (origin) {
841                 case PERF_RECORD_MISC_USER:
842                         ++top->us_samples;
843                         if (top->hide_user_symbols)
844                                 continue;
845                         machine = perf_session__find_host_machine(session);
846                         break;
847                 case PERF_RECORD_MISC_KERNEL:
848                         ++top->kernel_samples;
849                         if (top->hide_kernel_symbols)
850                                 continue;
851                         machine = perf_session__find_host_machine(session);
852                         break;
853                 case PERF_RECORD_MISC_GUEST_KERNEL:
854                         ++top->guest_kernel_samples;
855                         machine = perf_session__find_machine(session, event->ip.pid);
856                         break;
857                 case PERF_RECORD_MISC_GUEST_USER:
858                         ++top->guest_us_samples;
859                         /*
860                          * TODO: we don't process guest user from host side
861                          * except simple counting.
862                          */
863                         /* Fall thru */
864                 default:
865                         continue;
866                 }
867
868
869                 if (event->header.type == PERF_RECORD_SAMPLE) {
870                         perf_event__process_sample(&top->tool, event, evsel,
871                                                    &sample, machine);
872                 } else if (event->header.type < PERF_RECORD_MAX) {
873                         hists__inc_nr_events(&evsel->hists, event->header.type);
874                         perf_event__process(&top->tool, event, &sample, machine);
875                 } else
876                         ++session->hists.stats.nr_unknown_events;
877         }
878 }
879
880 static void perf_top__mmap_read(struct perf_top *top)
881 {
882         int i;
883
884         for (i = 0; i < top->evlist->nr_mmaps; i++)
885                 perf_top__mmap_read_idx(top, i);
886 }
887
888 static void perf_top__start_counters(struct perf_top *top)
889 {
890         struct perf_evsel *counter;
891         struct perf_evlist *evlist = top->evlist;
892
893         if (top->group)
894                 perf_evlist__set_leader(evlist);
895
896         list_for_each_entry(counter, &evlist->entries, node) {
897                 struct perf_event_attr *attr = &counter->attr;
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) < 0) {
929                         int err = errno;
930
931                         if (err == EPERM || err == EACCES) {
932                                 ui__error_paranoid();
933                                 goto out_err;
934                         } else if (err == EINVAL) {
935                                 if (!top->exclude_guest_missing &&
936                                     (attr->exclude_guest || attr->exclude_host)) {
937                                         pr_debug("Old kernel, cannot exclude "
938                                                  "guest or host samples.\n");
939                                         top->exclude_guest_missing = true;
940                                         goto fallback_missing_features;
941                                 } else if (!top->sample_id_all_missing) {
942                                         /*
943                                          * Old kernel, no attr->sample_id_type_all field
944                                          */
945                                         top->sample_id_all_missing = true;
946                                         goto retry_sample_id;
947                                 }
948                         }
949                         /*
950                          * If it's cycles then fall back to hrtimer
951                          * based cpu-clock-tick sw counter, which
952                          * is always available even if no PMU support:
953                          */
954                         if ((err == ENOENT || err == ENXIO) &&
955                             (attr->type == PERF_TYPE_HARDWARE) &&
956                             (attr->config == PERF_COUNT_HW_CPU_CYCLES)) {
957
958                                 if (verbose)
959                                         ui__warning("Cycles event not supported,\n"
960                                                     "trying to fall back to cpu-clock-ticks\n");
961
962                                 attr->type = PERF_TYPE_SOFTWARE;
963                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
964                                 if (counter->name) {
965                                         free(counter->name);
966                                         counter->name = NULL;
967                                 }
968                                 goto try_again;
969                         }
970
971                         if (err == ENOENT) {
972                                 ui__error("The %s event is not supported.\n",
973                                           perf_evsel__name(counter));
974                                 goto out_err;
975                         } else if (err == EMFILE) {
976                                 ui__error("Too many events are opened.\n"
977                                             "Try again after reducing the number of events\n");
978                                 goto out_err;
979                         }
980
981                         ui__error("The sys_perf_event_open() syscall "
982                                     "returned with %d (%s).  /bin/dmesg "
983                                     "may provide additional information.\n"
984                                     "No CONFIG_PERF_EVENTS=y kernel support "
985                                     "configured?\n", err, strerror(err));
986                         goto out_err;
987                 }
988         }
989
990         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
991                 ui__error("Failed to mmap with %d (%s)\n",
992                             errno, strerror(errno));
993                 goto out_err;
994         }
995
996         return;
997
998 out_err:
999         exit_browser(0);
1000         exit(0);
1001 }
1002
1003 static int perf_top__setup_sample_type(struct perf_top *top)
1004 {
1005         if (!top->sort_has_symbols) {
1006                 if (symbol_conf.use_callchain) {
1007                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1008                         return -EINVAL;
1009                 }
1010         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
1011                 if (callchain_register_param(&callchain_param) < 0) {
1012                         ui__error("Can't register callchain params.\n");
1013                         return -EINVAL;
1014                 }
1015         }
1016
1017         return 0;
1018 }
1019
1020 static int __cmd_top(struct perf_top *top)
1021 {
1022         pthread_t thread;
1023         int ret;
1024         /*
1025          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1026          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1027          */
1028         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1029         if (top->session == NULL)
1030                 return -ENOMEM;
1031
1032         ret = perf_top__setup_sample_type(top);
1033         if (ret)
1034                 goto out_delete;
1035
1036         if (perf_target__has_task(&top->target))
1037                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
1038                                                   perf_event__process,
1039                                                   &top->session->host_machine);
1040         else
1041                 perf_event__synthesize_threads(&top->tool, perf_event__process,
1042                                                &top->session->host_machine);
1043         perf_top__start_counters(top);
1044         top->session->evlist = top->evlist;
1045         perf_session__set_id_hdr_size(top->session);
1046
1047         /* Wait for a minimal set of events before starting the snapshot */
1048         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1049
1050         perf_top__mmap_read(top);
1051
1052         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1053                                                             display_thread), top)) {
1054                 ui__error("Could not create display thread.\n");
1055                 exit(-1);
1056         }
1057
1058         if (top->realtime_prio) {
1059                 struct sched_param param;
1060
1061                 param.sched_priority = top->realtime_prio;
1062                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1063                         ui__error("Could not set realtime priority.\n");
1064                         exit(-1);
1065                 }
1066         }
1067
1068         while (1) {
1069                 u64 hits = top->samples;
1070
1071                 perf_top__mmap_read(top);
1072
1073                 if (hits == top->samples)
1074                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1075         }
1076
1077 out_delete:
1078         perf_session__delete(top->session);
1079         top->session = NULL;
1080
1081         return 0;
1082 }
1083
1084 static int
1085 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1086 {
1087         struct perf_top *top = (struct perf_top *)opt->value;
1088         char *tok, *tok2;
1089         char *endptr;
1090
1091         /*
1092          * --no-call-graph
1093          */
1094         if (unset) {
1095                 top->dont_use_callchains = true;
1096                 return 0;
1097         }
1098
1099         symbol_conf.use_callchain = true;
1100
1101         if (!arg)
1102                 return 0;
1103
1104         tok = strtok((char *)arg, ",");
1105         if (!tok)
1106                 return -1;
1107
1108         /* get the output mode */
1109         if (!strncmp(tok, "graph", strlen(arg)))
1110                 callchain_param.mode = CHAIN_GRAPH_ABS;
1111
1112         else if (!strncmp(tok, "flat", strlen(arg)))
1113                 callchain_param.mode = CHAIN_FLAT;
1114
1115         else if (!strncmp(tok, "fractal", strlen(arg)))
1116                 callchain_param.mode = CHAIN_GRAPH_REL;
1117
1118         else if (!strncmp(tok, "none", strlen(arg))) {
1119                 callchain_param.mode = CHAIN_NONE;
1120                 symbol_conf.use_callchain = false;
1121
1122                 return 0;
1123         } else
1124                 return -1;
1125
1126         /* get the min percentage */
1127         tok = strtok(NULL, ",");
1128         if (!tok)
1129                 goto setup;
1130
1131         callchain_param.min_percent = strtod(tok, &endptr);
1132         if (tok == endptr)
1133                 return -1;
1134
1135         /* get the print limit */
1136         tok2 = strtok(NULL, ",");
1137         if (!tok2)
1138                 goto setup;
1139
1140         if (tok2[0] != 'c') {
1141                 callchain_param.print_limit = strtod(tok2, &endptr);
1142                 tok2 = strtok(NULL, ",");
1143                 if (!tok2)
1144                         goto setup;
1145         }
1146
1147         /* get the call chain order */
1148         if (!strcmp(tok2, "caller"))
1149                 callchain_param.order = ORDER_CALLER;
1150         else if (!strcmp(tok2, "callee"))
1151                 callchain_param.order = ORDER_CALLEE;
1152         else
1153                 return -1;
1154 setup:
1155         if (callchain_register_param(&callchain_param) < 0) {
1156                 fprintf(stderr, "Can't register callchain params\n");
1157                 return -1;
1158         }
1159         return 0;
1160 }
1161
1162 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1163 {
1164         struct perf_evsel *pos;
1165         int status;
1166         char errbuf[BUFSIZ];
1167         struct perf_top top = {
1168                 .count_filter        = 5,
1169                 .delay_secs          = 2,
1170                 .freq                = 4000, /* 4 KHz */
1171                 .mmap_pages          = 128,
1172                 .sym_pcnt_filter     = 5,
1173                 .target              = {
1174                         .uses_mmap   = true,
1175                 },
1176         };
1177         char callchain_default_opt[] = "fractal,0.5,callee";
1178         const struct option options[] = {
1179         OPT_CALLBACK('e', "event", &top.evlist, "event",
1180                      "event selector. use 'perf list' to list available events",
1181                      parse_events_option),
1182         OPT_INTEGER('c', "count", &top.default_interval,
1183                     "event period to sample"),
1184         OPT_STRING('p', "pid", &top.target.pid, "pid",
1185                     "profile events on existing process id"),
1186         OPT_STRING('t', "tid", &top.target.tid, "tid",
1187                     "profile events on existing thread id"),
1188         OPT_BOOLEAN('a', "all-cpus", &top.target.system_wide,
1189                             "system-wide collection from all CPUs"),
1190         OPT_STRING('C', "cpu", &top.target.cpu_list, "cpu",
1191                     "list of cpus to monitor"),
1192         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1193                    "file", "vmlinux pathname"),
1194         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1195                     "hide kernel symbols"),
1196         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1197         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1198                     "collect data with this RT SCHED_FIFO priority"),
1199         OPT_INTEGER('d', "delay", &top.delay_secs,
1200                     "number of seconds to delay between refreshes"),
1201         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1202                             "dump the symbol table used for profiling"),
1203         OPT_INTEGER('f', "count-filter", &top.count_filter,
1204                     "only display functions with more events than this"),
1205         OPT_BOOLEAN('g', "group", &top.group,
1206                             "put the counters into a counter group"),
1207         OPT_BOOLEAN('i', "inherit", &top.inherit,
1208                     "child tasks inherit counters"),
1209         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1210                     "symbol to annotate"),
1211         OPT_BOOLEAN('z', "zero", &top.zero,
1212                     "zero history across updates"),
1213         OPT_INTEGER('F', "freq", &top.freq,
1214                     "profile at this frequency"),
1215         OPT_INTEGER('E', "entries", &top.print_entries,
1216                     "display this many functions"),
1217         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1218                     "hide user symbols"),
1219         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1220         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1221         OPT_INCR('v', "verbose", &verbose,
1222                     "be more verbose (show counter open errors, etc)"),
1223         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1224                    "sort by key(s): pid, comm, dso, symbol, parent"),
1225         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1226                     "Show a column with the number of samples"),
1227         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1228                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1229                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1230                      callchain_default_opt),
1231         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1232                     "Show a column with the sum of periods"),
1233         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1234                    "only consider symbols in these dsos"),
1235         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1236                    "only consider symbols in these comms"),
1237         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1238                    "only consider these symbols"),
1239         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1240                     "Interleave source code with assembly code (default)"),
1241         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1242                     "Display raw encoding of assembly instructions (default)"),
1243         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1244                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1245         OPT_STRING('u', "uid", &top.target.uid_str, "user", "user to profile"),
1246         OPT_END()
1247         };
1248         const char * const top_usage[] = {
1249                 "perf top [<options>]",
1250                 NULL
1251         };
1252
1253         top.evlist = perf_evlist__new(NULL, NULL);
1254         if (top.evlist == NULL)
1255                 return -ENOMEM;
1256
1257         symbol_conf.exclude_other = false;
1258
1259         argc = parse_options(argc, argv, options, top_usage, 0);
1260         if (argc)
1261                 usage_with_options(top_usage, options);
1262
1263         if (sort_order == default_sort_order)
1264                 sort_order = "dso,symbol";
1265
1266         setup_sorting(top_usage, options);
1267
1268         if (top.use_stdio)
1269                 use_browser = 0;
1270         else if (top.use_tui)
1271                 use_browser = 1;
1272
1273         setup_browser(false);
1274
1275         status = perf_target__validate(&top.target);
1276         if (status) {
1277                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1278                 ui__warning("%s", errbuf);
1279         }
1280
1281         status = perf_target__parse_uid(&top.target);
1282         if (status) {
1283                 int saved_errno = errno;
1284
1285                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1286                 ui__error("%s", errbuf);
1287
1288                 status = -saved_errno;
1289                 goto out_delete_evlist;
1290         }
1291
1292         if (perf_target__none(&top.target))
1293                 top.target.system_wide = true;
1294
1295         if (perf_evlist__create_maps(top.evlist, &top.target) < 0)
1296                 usage_with_options(top_usage, options);
1297
1298         if (!top.evlist->nr_entries &&
1299             perf_evlist__add_default(top.evlist) < 0) {
1300                 ui__error("Not enough memory for event selector list\n");
1301                 return -ENOMEM;
1302         }
1303
1304         symbol_conf.nr_events = top.evlist->nr_entries;
1305
1306         if (top.delay_secs < 1)
1307                 top.delay_secs = 1;
1308
1309         /*
1310          * User specified count overrides default frequency.
1311          */
1312         if (top.default_interval)
1313                 top.freq = 0;
1314         else if (top.freq) {
1315                 top.default_interval = top.freq;
1316         } else {
1317                 ui__error("frequency and count are zero, aborting\n");
1318                 exit(EXIT_FAILURE);
1319         }
1320
1321         list_for_each_entry(pos, &top.evlist->entries, node) {
1322                 /*
1323                  * Fill in the ones not specifically initialized via -c:
1324                  */
1325                 if (!pos->attr.sample_period)
1326                         pos->attr.sample_period = top.default_interval;
1327         }
1328
1329         top.sym_evsel = perf_evlist__first(top.evlist);
1330
1331         symbol_conf.priv_size = sizeof(struct annotation);
1332
1333         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1334         if (symbol__init() < 0)
1335                 return -1;
1336
1337         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1338         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1339         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1340
1341         /*
1342          * Avoid annotation data structures overhead when symbols aren't on the
1343          * sort list.
1344          */
1345         top.sort_has_symbols = sort_sym.list.next != NULL;
1346
1347         get_term_dimensions(&top.winsize);
1348         if (top.print_entries == 0) {
1349                 struct sigaction act = {
1350                         .sa_sigaction = perf_top__sig_winch,
1351                         .sa_flags     = SA_SIGINFO,
1352                 };
1353                 perf_top__update_print_entries(&top);
1354                 sigaction(SIGWINCH, &act, NULL);
1355         }
1356
1357         status = __cmd_top(&top);
1358
1359 out_delete_evlist:
1360         perf_evlist__delete(top.evlist);
1361
1362         return status;
1363 }