Merge tag 'edac_for_4.5' of git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp
[cascardo/linux.git] / tools / perf / builtin-top.c
1 /*
2  * builtin-top.c
3  *
4  * Builtin top command: Display a continuously updated profile of
5  * any workload, CPU or specific PID.
6  *
7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
9  *
10  * Improvements and fixes by:
11  *
12  *   Arjan van de Ven <arjan@linux.intel.com>
13  *   Yanmin Zhang <yanmin.zhang@intel.com>
14  *   Wu Fengguang <fengguang.wu@intel.com>
15  *   Mike Galbraith <efault@gmx.de>
16  *   Paul Mackerras <paulus@samba.org>
17  *
18  * Released under the GPL v2. (and only v2, not any later version)
19  */
20 #include "builtin.h"
21
22 #include "perf.h"
23
24 #include "util/annotate.h"
25 #include "util/cache.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/machine.h"
30 #include "util/session.h"
31 #include "util/symbol.h"
32 #include "util/thread.h"
33 #include "util/thread_map.h"
34 #include "util/top.h"
35 #include "util/util.h"
36 #include <linux/rbtree.h>
37 #include <subcmd/parse-options.h>
38 #include "util/parse-events.h"
39 #include "util/cpumap.h"
40 #include "util/xyarray.h"
41 #include "util/sort.h"
42 #include "util/intlist.h"
43 #include "util/parse-branch-options.h"
44 #include "arch/common.h"
45
46 #include "util/debug.h"
47
48 #include <assert.h>
49 #include <elf.h>
50 #include <fcntl.h>
51
52 #include <stdio.h>
53 #include <termios.h>
54 #include <unistd.h>
55 #include <inttypes.h>
56
57 #include <errno.h>
58 #include <time.h>
59 #include <sched.h>
60
61 #include <sys/syscall.h>
62 #include <sys/ioctl.h>
63 #include <poll.h>
64 #include <sys/prctl.h>
65 #include <sys/wait.h>
66 #include <sys/uio.h>
67 #include <sys/utsname.h>
68 #include <sys/mman.h>
69
70 #include <linux/types.h>
71
72 static volatile int done;
73
74 #define HEADER_LINE_NR  5
75
76 static void perf_top__update_print_entries(struct perf_top *top)
77 {
78         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
79 }
80
81 static void perf_top__sig_winch(int sig __maybe_unused,
82                                 siginfo_t *info __maybe_unused, void *arg)
83 {
84         struct perf_top *top = arg;
85
86         get_term_dimensions(&top->winsize);
87         perf_top__update_print_entries(top);
88 }
89
90 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
91 {
92         struct symbol *sym;
93         struct annotation *notes;
94         struct map *map;
95         int err = -1;
96
97         if (!he || !he->ms.sym)
98                 return -1;
99
100         sym = he->ms.sym;
101         map = he->ms.map;
102
103         /*
104          * We can't annotate with just /proc/kallsyms
105          */
106         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
107             !dso__is_kcore(map->dso)) {
108                 pr_err("Can't annotate %s: No vmlinux file was found in the "
109                        "path\n", sym->name);
110                 sleep(1);
111                 return -1;
112         }
113
114         notes = symbol__annotation(sym);
115         if (notes->src != NULL) {
116                 pthread_mutex_lock(&notes->lock);
117                 goto out_assign;
118         }
119
120         pthread_mutex_lock(&notes->lock);
121
122         if (symbol__alloc_hist(sym) < 0) {
123                 pthread_mutex_unlock(&notes->lock);
124                 pr_err("Not enough memory for annotating '%s' symbol!\n",
125                        sym->name);
126                 sleep(1);
127                 return err;
128         }
129
130         err = symbol__annotate(sym, map, 0);
131         if (err == 0) {
132 out_assign:
133                 top->sym_filter_entry = he;
134         }
135
136         pthread_mutex_unlock(&notes->lock);
137         return err;
138 }
139
140 static void __zero_source_counters(struct hist_entry *he)
141 {
142         struct symbol *sym = he->ms.sym;
143         symbol__annotate_zero_histograms(sym);
144 }
145
146 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
147 {
148         struct utsname uts;
149         int err = uname(&uts);
150
151         ui__warning("Out of bounds address found:\n\n"
152                     "Addr:   %" PRIx64 "\n"
153                     "DSO:    %s %c\n"
154                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
155                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
156                     "Arch:   %s\n"
157                     "Kernel: %s\n"
158                     "Tools:  %s\n\n"
159                     "Not all samples will be on the annotation output.\n\n"
160                     "Please report to linux-kernel@vger.kernel.org\n",
161                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
162                     map->start, map->end, sym->start, sym->end,
163                     sym->binding == STB_GLOBAL ? 'g' :
164                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
165                     err ? "[unknown]" : uts.machine,
166                     err ? "[unknown]" : uts.release, perf_version_string);
167         if (use_browser <= 0)
168                 sleep(5);
169
170         map->erange_warned = true;
171 }
172
173 static void perf_top__record_precise_ip(struct perf_top *top,
174                                         struct hist_entry *he,
175                                         int counter, u64 ip)
176 {
177         struct annotation *notes;
178         struct symbol *sym = he->ms.sym;
179         int err = 0;
180
181         if (sym == NULL || (use_browser == 0 &&
182                             (top->sym_filter_entry == NULL ||
183                              top->sym_filter_entry->ms.sym != sym)))
184                 return;
185
186         notes = symbol__annotation(sym);
187
188         if (pthread_mutex_trylock(&notes->lock))
189                 return;
190
191         err = hist_entry__inc_addr_samples(he, counter, ip);
192
193         pthread_mutex_unlock(&notes->lock);
194
195         if (unlikely(err)) {
196                 /*
197                  * This function is now called with he->hists->lock held.
198                  * Release it before going to sleep.
199                  */
200                 pthread_mutex_unlock(&he->hists->lock);
201
202                 if (err == -ERANGE && !he->ms.map->erange_warned)
203                         ui__warn_map_erange(he->ms.map, sym, ip);
204                 else if (err == -ENOMEM) {
205                         pr_err("Not enough memory for annotating '%s' symbol!\n",
206                                sym->name);
207                         sleep(1);
208                 }
209
210                 pthread_mutex_lock(&he->hists->lock);
211         }
212 }
213
214 static void perf_top__show_details(struct perf_top *top)
215 {
216         struct hist_entry *he = top->sym_filter_entry;
217         struct annotation *notes;
218         struct symbol *symbol;
219         int more;
220
221         if (!he)
222                 return;
223
224         symbol = he->ms.sym;
225         notes = symbol__annotation(symbol);
226
227         pthread_mutex_lock(&notes->lock);
228
229         if (notes->src == NULL)
230                 goto out_unlock;
231
232         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
233         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
234
235         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
236                                        0, top->sym_pcnt_filter, top->print_entries, 4);
237
238         if (top->evlist->enabled) {
239                 if (top->zero)
240                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
241                 else
242                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
243         }
244         if (more != 0)
245                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
246 out_unlock:
247         pthread_mutex_unlock(&notes->lock);
248 }
249
250 static void perf_top__print_sym_table(struct perf_top *top)
251 {
252         char bf[160];
253         int printed = 0;
254         const int win_width = top->winsize.ws_col - 1;
255         struct hists *hists = evsel__hists(top->sym_evsel);
256
257         puts(CONSOLE_CLEAR);
258
259         perf_top__header_snprintf(top, bf, sizeof(bf));
260         printf("%s\n", bf);
261
262         perf_top__reset_sample_counters(top);
263
264         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
265
266         if (hists->stats.nr_lost_warned !=
267             hists->stats.nr_events[PERF_RECORD_LOST]) {
268                 hists->stats.nr_lost_warned =
269                               hists->stats.nr_events[PERF_RECORD_LOST];
270                 color_fprintf(stdout, PERF_COLOR_RED,
271                               "WARNING: LOST %d chunks, Check IO/CPU overload",
272                               hists->stats.nr_lost_warned);
273                 ++printed;
274         }
275
276         if (top->sym_filter_entry) {
277                 perf_top__show_details(top);
278                 return;
279         }
280
281         if (top->evlist->enabled) {
282                 if (top->zero) {
283                         hists__delete_entries(hists);
284                 } else {
285                         hists__decay_entries(hists, top->hide_user_symbols,
286                                              top->hide_kernel_symbols);
287                 }
288         }
289
290         hists__collapse_resort(hists, NULL);
291         hists__output_resort(hists, NULL);
292
293         hists__output_recalc_col_len(hists, top->print_entries - printed);
294         putchar('\n');
295         hists__fprintf(hists, false, top->print_entries - printed, win_width,
296                        top->min_percent, stdout);
297 }
298
299 static void prompt_integer(int *target, const char *msg)
300 {
301         char *buf = malloc(0), *p;
302         size_t dummy = 0;
303         int tmp;
304
305         fprintf(stdout, "\n%s: ", msg);
306         if (getline(&buf, &dummy, stdin) < 0)
307                 return;
308
309         p = strchr(buf, '\n');
310         if (p)
311                 *p = 0;
312
313         p = buf;
314         while(*p) {
315                 if (!isdigit(*p))
316                         goto out_free;
317                 p++;
318         }
319         tmp = strtoul(buf, NULL, 10);
320         *target = tmp;
321 out_free:
322         free(buf);
323 }
324
325 static void prompt_percent(int *target, const char *msg)
326 {
327         int tmp = 0;
328
329         prompt_integer(&tmp, msg);
330         if (tmp >= 0 && tmp <= 100)
331                 *target = tmp;
332 }
333
334 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
335 {
336         char *buf = malloc(0), *p;
337         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
338         struct hists *hists = evsel__hists(top->sym_evsel);
339         struct rb_node *next;
340         size_t dummy = 0;
341
342         /* zero counters of active symbol */
343         if (syme) {
344                 __zero_source_counters(syme);
345                 top->sym_filter_entry = NULL;
346         }
347
348         fprintf(stdout, "\n%s: ", msg);
349         if (getline(&buf, &dummy, stdin) < 0)
350                 goto out_free;
351
352         p = strchr(buf, '\n');
353         if (p)
354                 *p = 0;
355
356         next = rb_first(&hists->entries);
357         while (next) {
358                 n = rb_entry(next, struct hist_entry, rb_node);
359                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
360                         found = n;
361                         break;
362                 }
363                 next = rb_next(&n->rb_node);
364         }
365
366         if (!found) {
367                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
368                 sleep(1);
369         } else
370                 perf_top__parse_source(top, found);
371
372 out_free:
373         free(buf);
374 }
375
376 static void perf_top__print_mapped_keys(struct perf_top *top)
377 {
378         char *name = NULL;
379
380         if (top->sym_filter_entry) {
381                 struct symbol *sym = top->sym_filter_entry->ms.sym;
382                 name = sym->name;
383         }
384
385         fprintf(stdout, "\nMapped keys:\n");
386         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
387         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
388
389         if (top->evlist->nr_entries > 1)
390                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
391
392         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
393
394         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
395         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
396         fprintf(stdout, "\t[S]     stop annotation.\n");
397
398         fprintf(stdout,
399                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
400                 top->hide_kernel_symbols ? "yes" : "no");
401         fprintf(stdout,
402                 "\t[U]     hide user symbols.               \t(%s)\n",
403                 top->hide_user_symbols ? "yes" : "no");
404         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
405         fprintf(stdout, "\t[qQ]    quit.\n");
406 }
407
408 static int perf_top__key_mapped(struct perf_top *top, int c)
409 {
410         switch (c) {
411                 case 'd':
412                 case 'e':
413                 case 'f':
414                 case 'z':
415                 case 'q':
416                 case 'Q':
417                 case 'K':
418                 case 'U':
419                 case 'F':
420                 case 's':
421                 case 'S':
422                         return 1;
423                 case 'E':
424                         return top->evlist->nr_entries > 1 ? 1 : 0;
425                 default:
426                         break;
427         }
428
429         return 0;
430 }
431
432 static bool perf_top__handle_keypress(struct perf_top *top, int c)
433 {
434         bool ret = true;
435
436         if (!perf_top__key_mapped(top, c)) {
437                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
438                 struct termios save;
439
440                 perf_top__print_mapped_keys(top);
441                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
442                 fflush(stdout);
443
444                 set_term_quiet_input(&save);
445
446                 poll(&stdin_poll, 1, -1);
447                 c = getc(stdin);
448
449                 tcsetattr(0, TCSAFLUSH, &save);
450                 if (!perf_top__key_mapped(top, c))
451                         return ret;
452         }
453
454         switch (c) {
455                 case 'd':
456                         prompt_integer(&top->delay_secs, "Enter display delay");
457                         if (top->delay_secs < 1)
458                                 top->delay_secs = 1;
459                         break;
460                 case 'e':
461                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
462                         if (top->print_entries == 0) {
463                                 struct sigaction act = {
464                                         .sa_sigaction = perf_top__sig_winch,
465                                         .sa_flags     = SA_SIGINFO,
466                                 };
467                                 perf_top__sig_winch(SIGWINCH, NULL, top);
468                                 sigaction(SIGWINCH, &act, NULL);
469                         } else {
470                                 signal(SIGWINCH, SIG_DFL);
471                         }
472                         break;
473                 case 'E':
474                         if (top->evlist->nr_entries > 1) {
475                                 /* Select 0 as the default event: */
476                                 int counter = 0;
477
478                                 fprintf(stderr, "\nAvailable events:");
479
480                                 evlist__for_each(top->evlist, top->sym_evsel)
481                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
482
483                                 prompt_integer(&counter, "Enter details event counter");
484
485                                 if (counter >= top->evlist->nr_entries) {
486                                         top->sym_evsel = perf_evlist__first(top->evlist);
487                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
488                                         sleep(1);
489                                         break;
490                                 }
491                                 evlist__for_each(top->evlist, top->sym_evsel)
492                                         if (top->sym_evsel->idx == counter)
493                                                 break;
494                         } else
495                                 top->sym_evsel = perf_evlist__first(top->evlist);
496                         break;
497                 case 'f':
498                         prompt_integer(&top->count_filter, "Enter display event count filter");
499                         break;
500                 case 'F':
501                         prompt_percent(&top->sym_pcnt_filter,
502                                        "Enter details display event filter (percent)");
503                         break;
504                 case 'K':
505                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
506                         break;
507                 case 'q':
508                 case 'Q':
509                         printf("exiting.\n");
510                         if (top->dump_symtab)
511                                 perf_session__fprintf_dsos(top->session, stderr);
512                         ret = false;
513                         break;
514                 case 's':
515                         perf_top__prompt_symbol(top, "Enter details symbol");
516                         break;
517                 case 'S':
518                         if (!top->sym_filter_entry)
519                                 break;
520                         else {
521                                 struct hist_entry *syme = top->sym_filter_entry;
522
523                                 top->sym_filter_entry = NULL;
524                                 __zero_source_counters(syme);
525                         }
526                         break;
527                 case 'U':
528                         top->hide_user_symbols = !top->hide_user_symbols;
529                         break;
530                 case 'z':
531                         top->zero = !top->zero;
532                         break;
533                 default:
534                         break;
535         }
536
537         return ret;
538 }
539
540 static void perf_top__sort_new_samples(void *arg)
541 {
542         struct perf_top *t = arg;
543         struct hists *hists;
544
545         perf_top__reset_sample_counters(t);
546
547         if (t->evlist->selected != NULL)
548                 t->sym_evsel = t->evlist->selected;
549
550         hists = evsel__hists(t->sym_evsel);
551
552         if (t->evlist->enabled) {
553                 if (t->zero) {
554                         hists__delete_entries(hists);
555                 } else {
556                         hists__decay_entries(hists, t->hide_user_symbols,
557                                              t->hide_kernel_symbols);
558                 }
559         }
560
561         hists__collapse_resort(hists, NULL);
562         hists__output_resort(hists, NULL);
563 }
564
565 static void *display_thread_tui(void *arg)
566 {
567         struct perf_evsel *pos;
568         struct perf_top *top = arg;
569         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
570         struct hist_browser_timer hbt = {
571                 .timer          = perf_top__sort_new_samples,
572                 .arg            = top,
573                 .refresh        = top->delay_secs,
574         };
575
576         perf_top__sort_new_samples(top);
577
578         /*
579          * Initialize the uid_filter_str, in the future the TUI will allow
580          * Zooming in/out UIDs. For now juse use whatever the user passed
581          * via --uid.
582          */
583         evlist__for_each(top->evlist, pos) {
584                 struct hists *hists = evsel__hists(pos);
585                 hists->uid_filter_str = top->record_opts.target.uid_str;
586         }
587
588         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
589                                       top->min_percent,
590                                       &top->session->header.env);
591
592         done = 1;
593         return NULL;
594 }
595
596 static void display_sig(int sig __maybe_unused)
597 {
598         done = 1;
599 }
600
601 static void display_setup_sig(void)
602 {
603         signal(SIGSEGV, sighandler_dump_stack);
604         signal(SIGFPE, sighandler_dump_stack);
605         signal(SIGINT,  display_sig);
606         signal(SIGQUIT, display_sig);
607         signal(SIGTERM, display_sig);
608 }
609
610 static void *display_thread(void *arg)
611 {
612         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
613         struct termios save;
614         struct perf_top *top = arg;
615         int delay_msecs, c;
616
617         display_setup_sig();
618         pthread__unblock_sigwinch();
619 repeat:
620         delay_msecs = top->delay_secs * 1000;
621         set_term_quiet_input(&save);
622         /* trash return*/
623         getc(stdin);
624
625         while (!done) {
626                 perf_top__print_sym_table(top);
627                 /*
628                  * Either timeout expired or we got an EINTR due to SIGWINCH,
629                  * refresh screen in both cases.
630                  */
631                 switch (poll(&stdin_poll, 1, delay_msecs)) {
632                 case 0:
633                         continue;
634                 case -1:
635                         if (errno == EINTR)
636                                 continue;
637                         /* Fall trhu */
638                 default:
639                         c = getc(stdin);
640                         tcsetattr(0, TCSAFLUSH, &save);
641
642                         if (perf_top__handle_keypress(top, c))
643                                 goto repeat;
644                         done = 1;
645                 }
646         }
647
648         tcsetattr(0, TCSAFLUSH, &save);
649         return NULL;
650 }
651
652 static int symbol_filter(struct map *map, struct symbol *sym)
653 {
654         const char *name = sym->name;
655
656         if (!__map__is_kernel(map))
657                 return 0;
658         /*
659          * ppc64 uses function descriptors and appends a '.' to the
660          * start of every instruction address. Remove it.
661          */
662         if (name[0] == '.')
663                 name++;
664
665         if (!strcmp(name, "_text") ||
666             !strcmp(name, "_etext") ||
667             !strcmp(name, "_sinittext") ||
668             !strncmp("init_module", name, 11) ||
669             !strncmp("cleanup_module", name, 14) ||
670             strstr(name, "_text_start") ||
671             strstr(name, "_text_end"))
672                 return 1;
673
674         if (symbol__is_idle(sym))
675                 sym->ignore = true;
676
677         return 0;
678 }
679
680 static int hist_iter__top_callback(struct hist_entry_iter *iter,
681                                    struct addr_location *al, bool single,
682                                    void *arg)
683 {
684         struct perf_top *top = arg;
685         struct hist_entry *he = iter->he;
686         struct perf_evsel *evsel = iter->evsel;
687
688         if (sort__has_sym && single)
689                 perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
690
691         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
692                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
693         return 0;
694 }
695
696 static void perf_event__process_sample(struct perf_tool *tool,
697                                        const union perf_event *event,
698                                        struct perf_evsel *evsel,
699                                        struct perf_sample *sample,
700                                        struct machine *machine)
701 {
702         struct perf_top *top = container_of(tool, struct perf_top, tool);
703         struct addr_location al;
704         int err;
705
706         if (!machine && perf_guest) {
707                 static struct intlist *seen;
708
709                 if (!seen)
710                         seen = intlist__new(NULL);
711
712                 if (!intlist__has_entry(seen, sample->pid)) {
713                         pr_err("Can't find guest [%d]'s kernel information\n",
714                                 sample->pid);
715                         intlist__add(seen, sample->pid);
716                 }
717                 return;
718         }
719
720         if (!machine) {
721                 pr_err("%u unprocessable samples recorded.\r",
722                        top->session->evlist->stats.nr_unprocessable_samples++);
723                 return;
724         }
725
726         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
727                 top->exact_samples++;
728
729         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
730                 return;
731
732         if (!top->kptr_restrict_warned &&
733             symbol_conf.kptr_restrict &&
734             al.cpumode == PERF_RECORD_MISC_KERNEL) {
735                 ui__warning(
736 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
737 "Check /proc/sys/kernel/kptr_restrict.\n\n"
738 "Kernel%s samples will not be resolved.\n",
739                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
740                           " modules" : "");
741                 if (use_browser <= 0)
742                         sleep(5);
743                 top->kptr_restrict_warned = true;
744         }
745
746         if (al.sym == NULL) {
747                 const char *msg = "Kernel samples will not be resolved.\n";
748                 /*
749                  * As we do lazy loading of symtabs we only will know if the
750                  * specified vmlinux file is invalid when we actually have a
751                  * hit in kernel space and then try to load it. So if we get
752                  * here and there are _no_ symbols in the DSO backing the
753                  * kernel map, bail out.
754                  *
755                  * We may never get here, for instance, if we use -K/
756                  * --hide-kernel-symbols, even if the user specifies an
757                  * invalid --vmlinux ;-)
758                  */
759                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
760                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
761                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
762                         if (symbol_conf.vmlinux_name) {
763                                 char serr[256];
764                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
765                                 ui__warning("The %s file can't be used: %s\n%s",
766                                             symbol_conf.vmlinux_name, serr, msg);
767                         } else {
768                                 ui__warning("A vmlinux file was not found.\n%s",
769                                             msg);
770                         }
771
772                         if (use_browser <= 0)
773                                 sleep(5);
774                         top->vmlinux_warned = true;
775                 }
776         }
777
778         if (al.sym == NULL || !al.sym->ignore) {
779                 struct hists *hists = evsel__hists(evsel);
780                 struct hist_entry_iter iter = {
781                         .evsel          = evsel,
782                         .sample         = sample,
783                         .add_entry_cb   = hist_iter__top_callback,
784                 };
785
786                 if (symbol_conf.cumulate_callchain)
787                         iter.ops = &hist_iter_cumulative;
788                 else
789                         iter.ops = &hist_iter_normal;
790
791                 pthread_mutex_lock(&hists->lock);
792
793                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
794                 if (err < 0)
795                         pr_err("Problem incrementing symbol period, skipping event\n");
796
797                 pthread_mutex_unlock(&hists->lock);
798         }
799
800         addr_location__put(&al);
801 }
802
803 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
804 {
805         struct perf_sample sample;
806         struct perf_evsel *evsel;
807         struct perf_session *session = top->session;
808         union perf_event *event;
809         struct machine *machine;
810         u8 origin;
811         int ret;
812
813         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
814                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
815                 if (ret) {
816                         pr_err("Can't parse sample, err = %d\n", ret);
817                         goto next_event;
818                 }
819
820                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
821                 assert(evsel != NULL);
822
823                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
824
825                 if (event->header.type == PERF_RECORD_SAMPLE)
826                         ++top->samples;
827
828                 switch (origin) {
829                 case PERF_RECORD_MISC_USER:
830                         ++top->us_samples;
831                         if (top->hide_user_symbols)
832                                 goto next_event;
833                         machine = &session->machines.host;
834                         break;
835                 case PERF_RECORD_MISC_KERNEL:
836                         ++top->kernel_samples;
837                         if (top->hide_kernel_symbols)
838                                 goto next_event;
839                         machine = &session->machines.host;
840                         break;
841                 case PERF_RECORD_MISC_GUEST_KERNEL:
842                         ++top->guest_kernel_samples;
843                         machine = perf_session__find_machine(session,
844                                                              sample.pid);
845                         break;
846                 case PERF_RECORD_MISC_GUEST_USER:
847                         ++top->guest_us_samples;
848                         /*
849                          * TODO: we don't process guest user from host side
850                          * except simple counting.
851                          */
852                         goto next_event;
853                 default:
854                         if (event->header.type == PERF_RECORD_SAMPLE)
855                                 goto next_event;
856                         machine = &session->machines.host;
857                         break;
858                 }
859
860
861                 if (event->header.type == PERF_RECORD_SAMPLE) {
862                         perf_event__process_sample(&top->tool, event, evsel,
863                                                    &sample, machine);
864                 } else if (event->header.type < PERF_RECORD_MAX) {
865                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
866                         machine__process_event(machine, event, &sample);
867                 } else
868                         ++session->evlist->stats.nr_unknown_events;
869 next_event:
870                 perf_evlist__mmap_consume(top->evlist, idx);
871         }
872 }
873
874 static void perf_top__mmap_read(struct perf_top *top)
875 {
876         int i;
877
878         for (i = 0; i < top->evlist->nr_mmaps; i++)
879                 perf_top__mmap_read_idx(top, i);
880 }
881
882 static int perf_top__start_counters(struct perf_top *top)
883 {
884         char msg[512];
885         struct perf_evsel *counter;
886         struct perf_evlist *evlist = top->evlist;
887         struct record_opts *opts = &top->record_opts;
888
889         perf_evlist__config(evlist, opts);
890
891         evlist__for_each(evlist, counter) {
892 try_again:
893                 if (perf_evsel__open(counter, top->evlist->cpus,
894                                      top->evlist->threads) < 0) {
895                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
896                                 if (verbose)
897                                         ui__warning("%s\n", msg);
898                                 goto try_again;
899                         }
900
901                         perf_evsel__open_strerror(counter, &opts->target,
902                                                   errno, msg, sizeof(msg));
903                         ui__error("%s\n", msg);
904                         goto out_err;
905                 }
906         }
907
908         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
909                 ui__error("Failed to mmap with %d (%s)\n",
910                             errno, strerror_r(errno, msg, sizeof(msg)));
911                 goto out_err;
912         }
913
914         return 0;
915
916 out_err:
917         return -1;
918 }
919
920 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
921 {
922         if (!sort__has_sym) {
923                 if (symbol_conf.use_callchain) {
924                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
925                         return -EINVAL;
926                 }
927         } else if (callchain_param.mode != CHAIN_NONE) {
928                 if (callchain_register_param(&callchain_param) < 0) {
929                         ui__error("Can't register callchain params.\n");
930                         return -EINVAL;
931                 }
932         }
933
934         return 0;
935 }
936
937 static int __cmd_top(struct perf_top *top)
938 {
939         struct record_opts *opts = &top->record_opts;
940         pthread_t thread;
941         int ret;
942
943         top->session = perf_session__new(NULL, false, NULL);
944         if (top->session == NULL)
945                 return -1;
946
947         machines__set_symbol_filter(&top->session->machines, symbol_filter);
948
949         if (!objdump_path) {
950                 ret = perf_env__lookup_objdump(&top->session->header.env);
951                 if (ret)
952                         goto out_delete;
953         }
954
955         ret = perf_top__setup_sample_type(top);
956         if (ret)
957                 goto out_delete;
958
959         if (perf_session__register_idle_thread(top->session) < 0)
960                 goto out_delete;
961
962         machine__synthesize_threads(&top->session->machines.host, &opts->target,
963                                     top->evlist->threads, false, opts->proc_map_timeout);
964
965         if (sort__has_socket) {
966                 ret = perf_env__read_cpu_topology_map(&perf_env);
967                 if (ret < 0)
968                         goto out_err_cpu_topo;
969         }
970
971         ret = perf_top__start_counters(top);
972         if (ret)
973                 goto out_delete;
974
975         top->session->evlist = top->evlist;
976         perf_session__set_id_hdr_size(top->session);
977
978         /*
979          * When perf is starting the traced process, all the events (apart from
980          * group members) have enable_on_exec=1 set, so don't spoil it by
981          * prematurely enabling them.
982          *
983          * XXX 'top' still doesn't start workloads like record, trace, but should,
984          * so leave the check here.
985          */
986         if (!target__none(&opts->target))
987                 perf_evlist__enable(top->evlist);
988
989         /* Wait for a minimal set of events before starting the snapshot */
990         perf_evlist__poll(top->evlist, 100);
991
992         perf_top__mmap_read(top);
993
994         ret = -1;
995         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
996                                                             display_thread), top)) {
997                 ui__error("Could not create display thread.\n");
998                 goto out_delete;
999         }
1000
1001         if (top->realtime_prio) {
1002                 struct sched_param param;
1003
1004                 param.sched_priority = top->realtime_prio;
1005                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1006                         ui__error("Could not set realtime priority.\n");
1007                         goto out_join;
1008                 }
1009         }
1010
1011         while (!done) {
1012                 u64 hits = top->samples;
1013
1014                 perf_top__mmap_read(top);
1015
1016                 if (hits == top->samples)
1017                         ret = perf_evlist__poll(top->evlist, 100);
1018         }
1019
1020         ret = 0;
1021 out_join:
1022         pthread_join(thread, NULL);
1023 out_delete:
1024         perf_session__delete(top->session);
1025         top->session = NULL;
1026
1027         return ret;
1028
1029 out_err_cpu_topo: {
1030         char errbuf[BUFSIZ];
1031         const char *err = strerror_r(-ret, errbuf, sizeof(errbuf));
1032
1033         ui__error("Could not read the CPU topology map: %s\n", err);
1034         goto out_delete;
1035 }
1036 }
1037
1038 static int
1039 callchain_opt(const struct option *opt, const char *arg, int unset)
1040 {
1041         symbol_conf.use_callchain = true;
1042         return record_callchain_opt(opt, arg, unset);
1043 }
1044
1045 static int
1046 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1047 {
1048         struct record_opts *record = (struct record_opts *)opt->value;
1049
1050         record->callgraph_set = true;
1051         callchain_param.enabled = !unset;
1052         callchain_param.record_mode = CALLCHAIN_FP;
1053
1054         /*
1055          * --no-call-graph
1056          */
1057         if (unset) {
1058                 symbol_conf.use_callchain = false;
1059                 callchain_param.record_mode = CALLCHAIN_NONE;
1060                 return 0;
1061         }
1062
1063         return parse_callchain_top_opt(arg);
1064 }
1065
1066 static int perf_top_config(const char *var, const char *value, void *cb)
1067 {
1068         if (!strcmp(var, "top.call-graph"))
1069                 var = "call-graph.record-mode"; /* fall-through */
1070         if (!strcmp(var, "top.children")) {
1071                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1072                 return 0;
1073         }
1074
1075         return perf_default_config(var, value, cb);
1076 }
1077
1078 static int
1079 parse_percent_limit(const struct option *opt, const char *arg,
1080                     int unset __maybe_unused)
1081 {
1082         struct perf_top *top = opt->value;
1083
1084         top->min_percent = strtof(arg, NULL);
1085         return 0;
1086 }
1087
1088 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1089         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1090
1091 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1092 {
1093         char errbuf[BUFSIZ];
1094         struct perf_top top = {
1095                 .count_filter        = 5,
1096                 .delay_secs          = 2,
1097                 .record_opts = {
1098                         .mmap_pages     = UINT_MAX,
1099                         .user_freq      = UINT_MAX,
1100                         .user_interval  = ULLONG_MAX,
1101                         .freq           = 4000, /* 4 KHz */
1102                         .target         = {
1103                                 .uses_mmap   = true,
1104                         },
1105                         .proc_map_timeout    = 500,
1106                 },
1107                 .max_stack           = PERF_MAX_STACK_DEPTH,
1108                 .sym_pcnt_filter     = 5,
1109         };
1110         struct record_opts *opts = &top.record_opts;
1111         struct target *target = &opts->target;
1112         const struct option options[] = {
1113         OPT_CALLBACK('e', "event", &top.evlist, "event",
1114                      "event selector. use 'perf list' to list available events",
1115                      parse_events_option),
1116         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1117         OPT_STRING('p', "pid", &target->pid, "pid",
1118                     "profile events on existing process id"),
1119         OPT_STRING('t', "tid", &target->tid, "tid",
1120                     "profile events on existing thread id"),
1121         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1122                             "system-wide collection from all CPUs"),
1123         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1124                     "list of cpus to monitor"),
1125         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1126                    "file", "vmlinux pathname"),
1127         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1128                     "don't load vmlinux even if found"),
1129         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1130                     "hide kernel symbols"),
1131         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1132                      "number of mmap data pages",
1133                      perf_evlist__parse_mmap_pages),
1134         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1135                     "collect data with this RT SCHED_FIFO priority"),
1136         OPT_INTEGER('d', "delay", &top.delay_secs,
1137                     "number of seconds to delay between refreshes"),
1138         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1139                             "dump the symbol table used for profiling"),
1140         OPT_INTEGER('f', "count-filter", &top.count_filter,
1141                     "only display functions with more events than this"),
1142         OPT_BOOLEAN(0, "group", &opts->group,
1143                             "put the counters into a counter group"),
1144         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1145                     "child tasks do not inherit counters"),
1146         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1147                     "symbol to annotate"),
1148         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1149         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1150         OPT_INTEGER('E', "entries", &top.print_entries,
1151                     "display this many functions"),
1152         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1153                     "hide user symbols"),
1154         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1155         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1156         OPT_INCR('v', "verbose", &verbose,
1157                     "be more verbose (show counter open errors, etc)"),
1158         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1159                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1160                    " Please refer the man page for the complete list."),
1161         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1162                    "output field(s): overhead, period, sample plus all of sort keys"),
1163         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1164                     "Show a column with the number of samples"),
1165         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1166                            NULL, "enables call-graph recording and display",
1167                            &callchain_opt),
1168         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1169                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1170                      top_callchain_help, &parse_callchain_opt),
1171         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1172                     "Accumulate callchains of children and show total overhead as well"),
1173         OPT_INTEGER(0, "max-stack", &top.max_stack,
1174                     "Set the maximum stack depth when parsing the callchain. "
1175                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1176         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1177                    "ignore callees of these functions in call graphs",
1178                    report_parse_ignore_callees_opt),
1179         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1180                     "Show a column with the sum of periods"),
1181         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1182                    "only consider symbols in these dsos"),
1183         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1184                    "only consider symbols in these comms"),
1185         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1186                    "only consider these symbols"),
1187         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1188                     "Interleave source code with assembly code (default)"),
1189         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1190                     "Display raw encoding of assembly instructions (default)"),
1191         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1192                     "Enable kernel symbol demangling"),
1193         OPT_STRING(0, "objdump", &objdump_path, "path",
1194                     "objdump binary to use for disassembly and annotations"),
1195         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1196                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1197         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1198         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1199                      "Don't show entries under that percent", parse_percent_limit),
1200         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1201                      "How to display percentage of filtered entries", parse_filter_percentage),
1202         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1203                    "width[,width...]",
1204                    "don't try to adjust column width, use these fixed values"),
1205         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1206                         "per thread proc mmap processing timeout in ms"),
1207         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1208                      "branch any", "sample any taken branches",
1209                      parse_branch_stack),
1210         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1211                      "branch filter mask", "branch stack filter modes",
1212                      parse_branch_stack),
1213         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1214                     "Show raw trace event output (do not use print fmt or plugins)"),
1215         OPT_END()
1216         };
1217         const char * const top_usage[] = {
1218                 "perf top [<options>]",
1219                 NULL
1220         };
1221         int status = hists__init();
1222
1223         if (status < 0)
1224                 return status;
1225
1226         top.evlist = perf_evlist__new();
1227         if (top.evlist == NULL)
1228                 return -ENOMEM;
1229
1230         perf_config(perf_top_config, &top);
1231
1232         argc = parse_options(argc, argv, options, top_usage, 0);
1233         if (argc)
1234                 usage_with_options(top_usage, options);
1235
1236         if (!top.evlist->nr_entries &&
1237             perf_evlist__add_default(top.evlist) < 0) {
1238                 pr_err("Not enough memory for event selector list\n");
1239                 goto out_delete_evlist;
1240         }
1241
1242         sort__mode = SORT_MODE__TOP;
1243         /* display thread wants entries to be collapsed in a different tree */
1244         sort__need_collapse = 1;
1245
1246         if (setup_sorting(top.evlist) < 0) {
1247                 if (sort_order)
1248                         parse_options_usage(top_usage, options, "s", 1);
1249                 if (field_order)
1250                         parse_options_usage(sort_order ? NULL : top_usage,
1251                                             options, "fields", 0);
1252                 goto out_delete_evlist;
1253         }
1254
1255         if (top.use_stdio)
1256                 use_browser = 0;
1257         else if (top.use_tui)
1258                 use_browser = 1;
1259
1260         setup_browser(false);
1261
1262         status = target__validate(target);
1263         if (status) {
1264                 target__strerror(target, status, errbuf, BUFSIZ);
1265                 ui__warning("%s\n", errbuf);
1266         }
1267
1268         status = target__parse_uid(target);
1269         if (status) {
1270                 int saved_errno = errno;
1271
1272                 target__strerror(target, status, errbuf, BUFSIZ);
1273                 ui__error("%s\n", errbuf);
1274
1275                 status = -saved_errno;
1276                 goto out_delete_evlist;
1277         }
1278
1279         if (target__none(target))
1280                 target->system_wide = true;
1281
1282         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1283                 ui__error("Couldn't create thread/CPU maps: %s\n",
1284                           errno == ENOENT ? "No such process" : strerror_r(errno, errbuf, sizeof(errbuf)));
1285                 goto out_delete_evlist;
1286         }
1287
1288         symbol_conf.nr_events = top.evlist->nr_entries;
1289
1290         if (top.delay_secs < 1)
1291                 top.delay_secs = 1;
1292
1293         if (record_opts__config(opts)) {
1294                 status = -EINVAL;
1295                 goto out_delete_evlist;
1296         }
1297
1298         top.sym_evsel = perf_evlist__first(top.evlist);
1299
1300         if (!symbol_conf.use_callchain) {
1301                 symbol_conf.cumulate_callchain = false;
1302                 perf_hpp__cancel_cumulate();
1303         }
1304
1305         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1306                 callchain_param.order = ORDER_CALLER;
1307
1308         symbol_conf.priv_size = sizeof(struct annotation);
1309
1310         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1311         if (symbol__init(NULL) < 0)
1312                 return -1;
1313
1314         sort__setup_elide(stdout);
1315
1316         get_term_dimensions(&top.winsize);
1317         if (top.print_entries == 0) {
1318                 struct sigaction act = {
1319                         .sa_sigaction = perf_top__sig_winch,
1320                         .sa_flags     = SA_SIGINFO,
1321                 };
1322                 perf_top__update_print_entries(&top);
1323                 sigaction(SIGWINCH, &act, NULL);
1324         }
1325
1326         status = __cmd_top(&top);
1327
1328 out_delete_evlist:
1329         perf_evlist__delete(top.evlist);
1330
1331         return status;
1332 }