GNU Linux-libre 4.4.290-gnu1
[releases.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 "util/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 static volatile int resize;
74
75 #define HEADER_LINE_NR  5
76
77 static void perf_top__update_print_entries(struct perf_top *top)
78 {
79         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
80 }
81
82 static void perf_top__sig_winch(int sig __maybe_unused,
83                                 siginfo_t *info __maybe_unused, void *arg __maybe_unused)
84 {
85         resize = 1;
86 }
87
88 static void perf_top__resize(struct perf_top *top)
89 {
90         get_term_dimensions(&top->winsize);
91         perf_top__update_print_entries(top);
92 }
93
94 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
95 {
96         struct symbol *sym;
97         struct annotation *notes;
98         struct map *map;
99         int err = -1;
100
101         if (!he || !he->ms.sym)
102                 return -1;
103
104         sym = he->ms.sym;
105         map = he->ms.map;
106
107         /*
108          * We can't annotate with just /proc/kallsyms
109          */
110         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
111             !dso__is_kcore(map->dso)) {
112                 pr_err("Can't annotate %s: No vmlinux file was found in the "
113                        "path\n", sym->name);
114                 sleep(1);
115                 return -1;
116         }
117
118         notes = symbol__annotation(sym);
119         if (notes->src != NULL) {
120                 pthread_mutex_lock(&notes->lock);
121                 goto out_assign;
122         }
123
124         pthread_mutex_lock(&notes->lock);
125
126         if (symbol__alloc_hist(sym) < 0) {
127                 pthread_mutex_unlock(&notes->lock);
128                 pr_err("Not enough memory for annotating '%s' symbol!\n",
129                        sym->name);
130                 sleep(1);
131                 return err;
132         }
133
134         err = symbol__annotate(sym, map, 0);
135         if (err == 0) {
136 out_assign:
137                 top->sym_filter_entry = he;
138         }
139
140         pthread_mutex_unlock(&notes->lock);
141         return err;
142 }
143
144 static void __zero_source_counters(struct hist_entry *he)
145 {
146         struct symbol *sym = he->ms.sym;
147         symbol__annotate_zero_histograms(sym);
148 }
149
150 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
151 {
152         struct utsname uts;
153         int err = uname(&uts);
154
155         ui__warning("Out of bounds address found:\n\n"
156                     "Addr:   %" PRIx64 "\n"
157                     "DSO:    %s %c\n"
158                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
159                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
160                     "Arch:   %s\n"
161                     "Kernel: %s\n"
162                     "Tools:  %s\n\n"
163                     "Not all samples will be on the annotation output.\n\n"
164                     "Please report to linux-kernel@vger.kernel.org\n",
165                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
166                     map->start, map->end, sym->start, sym->end,
167                     sym->binding == STB_GLOBAL ? 'g' :
168                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
169                     err ? "[unknown]" : uts.machine,
170                     err ? "[unknown]" : uts.release, perf_version_string);
171         if (use_browser <= 0)
172                 sleep(5);
173
174         map->erange_warned = true;
175 }
176
177 static void perf_top__record_precise_ip(struct perf_top *top,
178                                         struct hist_entry *he,
179                                         int counter, u64 ip)
180 {
181         struct annotation *notes;
182         struct symbol *sym;
183         int err = 0;
184
185         if (he == NULL || he->ms.sym == NULL ||
186             ((top->sym_filter_entry == NULL ||
187               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
188                 return;
189
190         sym = he->ms.sym;
191         notes = symbol__annotation(sym);
192
193         if (pthread_mutex_trylock(&notes->lock))
194                 return;
195
196         ip = he->ms.map->map_ip(he->ms.map, ip);
197
198         if (ui__has_annotation())
199                 err = hist_entry__inc_addr_samples(he, counter, ip);
200
201         pthread_mutex_unlock(&notes->lock);
202
203         /*
204          * This function is now called with he->hists->lock held.
205          * Release it before going to sleep.
206          */
207         pthread_mutex_unlock(&he->hists->lock);
208
209         if (err == -ERANGE && !he->ms.map->erange_warned)
210                 ui__warn_map_erange(he->ms.map, sym, ip);
211         else if (err == -ENOMEM) {
212                 pr_err("Not enough memory for annotating '%s' symbol!\n",
213                        sym->name);
214                 sleep(1);
215         }
216
217         pthread_mutex_lock(&he->hists->lock);
218 }
219
220 static void perf_top__show_details(struct perf_top *top)
221 {
222         struct hist_entry *he = top->sym_filter_entry;
223         struct annotation *notes;
224         struct symbol *symbol;
225         int more;
226
227         if (!he)
228                 return;
229
230         symbol = he->ms.sym;
231         notes = symbol__annotation(symbol);
232
233         pthread_mutex_lock(&notes->lock);
234
235         if (notes->src == NULL)
236                 goto out_unlock;
237
238         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
239         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
240
241         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
242                                        0, top->sym_pcnt_filter, top->print_entries, 4);
243
244         if (top->evlist->enabled) {
245                 if (top->zero)
246                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
247                 else
248                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
249         }
250         if (more != 0)
251                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
252 out_unlock:
253         pthread_mutex_unlock(&notes->lock);
254 }
255
256 static void perf_top__print_sym_table(struct perf_top *top)
257 {
258         char bf[160];
259         int printed = 0;
260         const int win_width = top->winsize.ws_col - 1;
261         struct hists *hists = evsel__hists(top->sym_evsel);
262
263         puts(CONSOLE_CLEAR);
264
265         perf_top__header_snprintf(top, bf, sizeof(bf));
266         printf("%s\n", bf);
267
268         perf_top__reset_sample_counters(top);
269
270         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
271
272         if (hists->stats.nr_lost_warned !=
273             hists->stats.nr_events[PERF_RECORD_LOST]) {
274                 hists->stats.nr_lost_warned =
275                               hists->stats.nr_events[PERF_RECORD_LOST];
276                 color_fprintf(stdout, PERF_COLOR_RED,
277                               "WARNING: LOST %d chunks, Check IO/CPU overload",
278                               hists->stats.nr_lost_warned);
279                 ++printed;
280         }
281
282         if (top->sym_filter_entry) {
283                 perf_top__show_details(top);
284                 return;
285         }
286
287         if (top->evlist->enabled) {
288                 if (top->zero) {
289                         hists__delete_entries(hists);
290                 } else {
291                         hists__decay_entries(hists, top->hide_user_symbols,
292                                              top->hide_kernel_symbols);
293                 }
294         }
295
296         hists__collapse_resort(hists, NULL);
297         hists__output_resort(hists, NULL);
298
299         hists__output_recalc_col_len(hists, top->print_entries - printed);
300         putchar('\n');
301         hists__fprintf(hists, false, top->print_entries - printed, win_width,
302                        top->min_percent, stdout);
303 }
304
305 static void prompt_integer(int *target, const char *msg)
306 {
307         char *buf = malloc(0), *p;
308         size_t dummy = 0;
309         int tmp;
310
311         fprintf(stdout, "\n%s: ", msg);
312         if (getline(&buf, &dummy, stdin) < 0)
313                 return;
314
315         p = strchr(buf, '\n');
316         if (p)
317                 *p = 0;
318
319         p = buf;
320         while(*p) {
321                 if (!isdigit(*p))
322                         goto out_free;
323                 p++;
324         }
325         tmp = strtoul(buf, NULL, 10);
326         *target = tmp;
327 out_free:
328         free(buf);
329 }
330
331 static void prompt_percent(int *target, const char *msg)
332 {
333         int tmp = 0;
334
335         prompt_integer(&tmp, msg);
336         if (tmp >= 0 && tmp <= 100)
337                 *target = tmp;
338 }
339
340 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
341 {
342         char *buf = malloc(0), *p;
343         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
344         struct hists *hists = evsel__hists(top->sym_evsel);
345         struct rb_node *next;
346         size_t dummy = 0;
347
348         /* zero counters of active symbol */
349         if (syme) {
350                 __zero_source_counters(syme);
351                 top->sym_filter_entry = NULL;
352         }
353
354         fprintf(stdout, "\n%s: ", msg);
355         if (getline(&buf, &dummy, stdin) < 0)
356                 goto out_free;
357
358         p = strchr(buf, '\n');
359         if (p)
360                 *p = 0;
361
362         next = rb_first(&hists->entries);
363         while (next) {
364                 n = rb_entry(next, struct hist_entry, rb_node);
365                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
366                         found = n;
367                         break;
368                 }
369                 next = rb_next(&n->rb_node);
370         }
371
372         if (!found) {
373                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
374                 sleep(1);
375         } else
376                 perf_top__parse_source(top, found);
377
378 out_free:
379         free(buf);
380 }
381
382 static void perf_top__print_mapped_keys(struct perf_top *top)
383 {
384         char *name = NULL;
385
386         if (top->sym_filter_entry) {
387                 struct symbol *sym = top->sym_filter_entry->ms.sym;
388                 name = sym->name;
389         }
390
391         fprintf(stdout, "\nMapped keys:\n");
392         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
393         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
394
395         if (top->evlist->nr_entries > 1)
396                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
397
398         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
399
400         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
401         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
402         fprintf(stdout, "\t[S]     stop annotation.\n");
403
404         fprintf(stdout,
405                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
406                 top->hide_kernel_symbols ? "yes" : "no");
407         fprintf(stdout,
408                 "\t[U]     hide user symbols.               \t(%s)\n",
409                 top->hide_user_symbols ? "yes" : "no");
410         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
411         fprintf(stdout, "\t[qQ]    quit.\n");
412 }
413
414 static int perf_top__key_mapped(struct perf_top *top, int c)
415 {
416         switch (c) {
417                 case 'd':
418                 case 'e':
419                 case 'f':
420                 case 'z':
421                 case 'q':
422                 case 'Q':
423                 case 'K':
424                 case 'U':
425                 case 'F':
426                 case 's':
427                 case 'S':
428                         return 1;
429                 case 'E':
430                         return top->evlist->nr_entries > 1 ? 1 : 0;
431                 default:
432                         break;
433         }
434
435         return 0;
436 }
437
438 static bool perf_top__handle_keypress(struct perf_top *top, int c)
439 {
440         bool ret = true;
441
442         if (!perf_top__key_mapped(top, c)) {
443                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
444                 struct termios save;
445
446                 perf_top__print_mapped_keys(top);
447                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
448                 fflush(stdout);
449
450                 set_term_quiet_input(&save);
451
452                 poll(&stdin_poll, 1, -1);
453                 c = getc(stdin);
454
455                 tcsetattr(0, TCSAFLUSH, &save);
456                 if (!perf_top__key_mapped(top, c))
457                         return ret;
458         }
459
460         switch (c) {
461                 case 'd':
462                         prompt_integer(&top->delay_secs, "Enter display delay");
463                         if (top->delay_secs < 1)
464                                 top->delay_secs = 1;
465                         break;
466                 case 'e':
467                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
468                         if (top->print_entries == 0) {
469                                 struct sigaction act = {
470                                         .sa_sigaction = perf_top__sig_winch,
471                                         .sa_flags     = SA_SIGINFO,
472                                 };
473                                 perf_top__resize(top);
474                                 sigaction(SIGWINCH, &act, NULL);
475                         } else {
476                                 signal(SIGWINCH, SIG_DFL);
477                         }
478                         break;
479                 case 'E':
480                         if (top->evlist->nr_entries > 1) {
481                                 /* Select 0 as the default event: */
482                                 int counter = 0;
483
484                                 fprintf(stderr, "\nAvailable events:");
485
486                                 evlist__for_each(top->evlist, top->sym_evsel)
487                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
488
489                                 prompt_integer(&counter, "Enter details event counter");
490
491                                 if (counter >= top->evlist->nr_entries) {
492                                         top->sym_evsel = perf_evlist__first(top->evlist);
493                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
494                                         sleep(1);
495                                         break;
496                                 }
497                                 evlist__for_each(top->evlist, top->sym_evsel)
498                                         if (top->sym_evsel->idx == counter)
499                                                 break;
500                         } else
501                                 top->sym_evsel = perf_evlist__first(top->evlist);
502                         break;
503                 case 'f':
504                         prompt_integer(&top->count_filter, "Enter display event count filter");
505                         break;
506                 case 'F':
507                         prompt_percent(&top->sym_pcnt_filter,
508                                        "Enter details display event filter (percent)");
509                         break;
510                 case 'K':
511                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
512                         break;
513                 case 'q':
514                 case 'Q':
515                         printf("exiting.\n");
516                         if (top->dump_symtab)
517                                 perf_session__fprintf_dsos(top->session, stderr);
518                         ret = false;
519                         break;
520                 case 's':
521                         perf_top__prompt_symbol(top, "Enter details symbol");
522                         break;
523                 case 'S':
524                         if (!top->sym_filter_entry)
525                                 break;
526                         else {
527                                 struct hist_entry *syme = top->sym_filter_entry;
528
529                                 top->sym_filter_entry = NULL;
530                                 __zero_source_counters(syme);
531                         }
532                         break;
533                 case 'U':
534                         top->hide_user_symbols = !top->hide_user_symbols;
535                         break;
536                 case 'z':
537                         top->zero = !top->zero;
538                         break;
539                 default:
540                         break;
541         }
542
543         return ret;
544 }
545
546 static void perf_top__sort_new_samples(void *arg)
547 {
548         struct perf_top *t = arg;
549         struct hists *hists;
550
551         perf_top__reset_sample_counters(t);
552
553         if (t->evlist->selected != NULL)
554                 t->sym_evsel = t->evlist->selected;
555
556         hists = evsel__hists(t->sym_evsel);
557
558         if (t->evlist->enabled) {
559                 if (t->zero) {
560                         hists__delete_entries(hists);
561                 } else {
562                         hists__decay_entries(hists, t->hide_user_symbols,
563                                              t->hide_kernel_symbols);
564                 }
565         }
566
567         hists__collapse_resort(hists, NULL);
568         hists__output_resort(hists, NULL);
569 }
570
571 static void *display_thread_tui(void *arg)
572 {
573         struct perf_evsel *pos;
574         struct perf_top *top = arg;
575         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
576         struct hist_browser_timer hbt = {
577                 .timer          = perf_top__sort_new_samples,
578                 .arg            = top,
579                 .refresh        = top->delay_secs,
580         };
581
582         perf_top__sort_new_samples(top);
583
584         /*
585          * Initialize the uid_filter_str, in the future the TUI will allow
586          * Zooming in/out UIDs. For now juse use whatever the user passed
587          * via --uid.
588          */
589         evlist__for_each(top->evlist, pos) {
590                 struct hists *hists = evsel__hists(pos);
591                 hists->uid_filter_str = top->record_opts.target.uid_str;
592         }
593
594         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
595                                       top->min_percent,
596                                       &top->session->header.env);
597
598         done = 1;
599         return NULL;
600 }
601
602 static void display_sig(int sig __maybe_unused)
603 {
604         done = 1;
605 }
606
607 static void display_setup_sig(void)
608 {
609         signal(SIGSEGV, sighandler_dump_stack);
610         signal(SIGFPE, sighandler_dump_stack);
611         signal(SIGINT,  display_sig);
612         signal(SIGQUIT, display_sig);
613         signal(SIGTERM, display_sig);
614 }
615
616 static void *display_thread(void *arg)
617 {
618         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
619         struct termios save;
620         struct perf_top *top = arg;
621         int delay_msecs, c;
622
623         display_setup_sig();
624         pthread__unblock_sigwinch();
625 repeat:
626         delay_msecs = top->delay_secs * 1000;
627         set_term_quiet_input(&save);
628         /* trash return*/
629         clearerr(stdin);
630         if (poll(&stdin_poll, 1, 0) > 0)
631                 getc(stdin);
632
633         while (!done) {
634                 perf_top__print_sym_table(top);
635                 /*
636                  * Either timeout expired or we got an EINTR due to SIGWINCH,
637                  * refresh screen in both cases.
638                  */
639                 switch (poll(&stdin_poll, 1, delay_msecs)) {
640                 case 0:
641                         continue;
642                 case -1:
643                         if (errno == EINTR)
644                                 continue;
645                         __fallthrough;
646                 default:
647                         c = getc(stdin);
648                         tcsetattr(0, TCSAFLUSH, &save);
649
650                         if (perf_top__handle_keypress(top, c))
651                                 goto repeat;
652                         done = 1;
653                 }
654         }
655
656         tcsetattr(0, TCSAFLUSH, &save);
657         return NULL;
658 }
659
660 static int symbol_filter(struct map *map, struct symbol *sym)
661 {
662         const char *name = sym->name;
663
664         if (!__map__is_kernel(map))
665                 return 0;
666         /*
667          * ppc64 uses function descriptors and appends a '.' to the
668          * start of every instruction address. Remove it.
669          */
670         if (name[0] == '.')
671                 name++;
672
673         if (!strcmp(name, "_text") ||
674             !strcmp(name, "_etext") ||
675             !strcmp(name, "_sinittext") ||
676             !strncmp("init_module", name, 11) ||
677             !strncmp("cleanup_module", name, 14) ||
678             strstr(name, "_text_start") ||
679             strstr(name, "_text_end"))
680                 return 1;
681
682         if (symbol__is_idle(sym))
683                 sym->ignore = true;
684
685         return 0;
686 }
687
688 static int hist_iter__top_callback(struct hist_entry_iter *iter,
689                                    struct addr_location *al, bool single,
690                                    void *arg)
691 {
692         struct perf_top *top = arg;
693         struct hist_entry *he = iter->he;
694         struct perf_evsel *evsel = iter->evsel;
695
696         if (sort__has_sym && single) {
697                 u64 ip = al->addr;
698
699                 if (al->map)
700                         ip = al->map->unmap_ip(al->map, ip);
701
702                 perf_top__record_precise_ip(top, he, evsel->idx, ip);
703         }
704
705         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
706                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
707         return 0;
708 }
709
710 static void perf_event__process_sample(struct perf_tool *tool,
711                                        const union perf_event *event,
712                                        struct perf_evsel *evsel,
713                                        struct perf_sample *sample,
714                                        struct machine *machine)
715 {
716         struct perf_top *top = container_of(tool, struct perf_top, tool);
717         struct addr_location al;
718         int err;
719
720         if (!machine && perf_guest) {
721                 static struct intlist *seen;
722
723                 if (!seen)
724                         seen = intlist__new(NULL);
725
726                 if (!intlist__has_entry(seen, sample->pid)) {
727                         pr_err("Can't find guest [%d]'s kernel information\n",
728                                 sample->pid);
729                         intlist__add(seen, sample->pid);
730                 }
731                 return;
732         }
733
734         if (!machine) {
735                 pr_err("%u unprocessable samples recorded.\r",
736                        top->session->evlist->stats.nr_unprocessable_samples++);
737                 return;
738         }
739
740         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
741                 top->exact_samples++;
742
743         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
744                 return;
745
746         if (!top->kptr_restrict_warned &&
747             symbol_conf.kptr_restrict &&
748             al.cpumode == PERF_RECORD_MISC_KERNEL) {
749                 ui__warning(
750 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
751 "Check /proc/sys/kernel/kptr_restrict.\n\n"
752 "Kernel%s samples will not be resolved.\n",
753                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
754                           " modules" : "");
755                 if (use_browser <= 0)
756                         sleep(5);
757                 top->kptr_restrict_warned = true;
758         }
759
760         if (al.sym == NULL) {
761                 const char *msg = "Kernel samples will not be resolved.\n";
762                 /*
763                  * As we do lazy loading of symtabs we only will know if the
764                  * specified vmlinux file is invalid when we actually have a
765                  * hit in kernel space and then try to load it. So if we get
766                  * here and there are _no_ symbols in the DSO backing the
767                  * kernel map, bail out.
768                  *
769                  * We may never get here, for instance, if we use -K/
770                  * --hide-kernel-symbols, even if the user specifies an
771                  * invalid --vmlinux ;-)
772                  */
773                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
774                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
775                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
776                         if (symbol_conf.vmlinux_name) {
777                                 char serr[256];
778                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
779                                 ui__warning("The %s file can't be used: %s\n%s",
780                                             symbol_conf.vmlinux_name, serr, msg);
781                         } else {
782                                 ui__warning("A vmlinux file was not found.\n%s",
783                                             msg);
784                         }
785
786                         if (use_browser <= 0)
787                                 sleep(5);
788                         top->vmlinux_warned = true;
789                 }
790         }
791
792         if (al.sym == NULL || !al.sym->ignore) {
793                 struct hists *hists = evsel__hists(evsel);
794                 struct hist_entry_iter iter = {
795                         .evsel          = evsel,
796                         .sample         = sample,
797                         .add_entry_cb   = hist_iter__top_callback,
798                 };
799
800                 if (symbol_conf.cumulate_callchain)
801                         iter.ops = &hist_iter_cumulative;
802                 else
803                         iter.ops = &hist_iter_normal;
804
805                 pthread_mutex_lock(&hists->lock);
806
807                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
808                 if (err < 0)
809                         pr_err("Problem incrementing symbol period, skipping event\n");
810
811                 pthread_mutex_unlock(&hists->lock);
812         }
813
814         addr_location__put(&al);
815 }
816
817 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
818 {
819         struct perf_sample sample;
820         struct perf_evsel *evsel;
821         struct perf_session *session = top->session;
822         union perf_event *event;
823         struct machine *machine;
824         u8 origin;
825         int ret;
826
827         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
828                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
829                 if (ret) {
830                         pr_err("Can't parse sample, err = %d\n", ret);
831                         goto next_event;
832                 }
833
834                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
835                 assert(evsel != NULL);
836
837                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
838
839                 if (event->header.type == PERF_RECORD_SAMPLE)
840                         ++top->samples;
841
842                 switch (origin) {
843                 case PERF_RECORD_MISC_USER:
844                         ++top->us_samples;
845                         if (top->hide_user_symbols)
846                                 goto next_event;
847                         machine = &session->machines.host;
848                         break;
849                 case PERF_RECORD_MISC_KERNEL:
850                         ++top->kernel_samples;
851                         if (top->hide_kernel_symbols)
852                                 goto next_event;
853                         machine = &session->machines.host;
854                         break;
855                 case PERF_RECORD_MISC_GUEST_KERNEL:
856                         ++top->guest_kernel_samples;
857                         machine = perf_session__find_machine(session,
858                                                              sample.pid);
859                         break;
860                 case PERF_RECORD_MISC_GUEST_USER:
861                         ++top->guest_us_samples;
862                         /*
863                          * TODO: we don't process guest user from host side
864                          * except simple counting.
865                          */
866                         goto next_event;
867                 default:
868                         if (event->header.type == PERF_RECORD_SAMPLE)
869                                 goto next_event;
870                         machine = &session->machines.host;
871                         break;
872                 }
873
874
875                 if (event->header.type == PERF_RECORD_SAMPLE) {
876                         perf_event__process_sample(&top->tool, event, evsel,
877                                                    &sample, machine);
878                 } else if (event->header.type < PERF_RECORD_MAX) {
879                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
880                         machine__process_event(machine, event, &sample);
881                 } else
882                         ++session->evlist->stats.nr_unknown_events;
883 next_event:
884                 perf_evlist__mmap_consume(top->evlist, idx);
885         }
886 }
887
888 static void perf_top__mmap_read(struct perf_top *top)
889 {
890         int i;
891
892         for (i = 0; i < top->evlist->nr_mmaps; i++)
893                 perf_top__mmap_read_idx(top, i);
894 }
895
896 static int perf_top__start_counters(struct perf_top *top)
897 {
898         char msg[512];
899         struct perf_evsel *counter;
900         struct perf_evlist *evlist = top->evlist;
901         struct record_opts *opts = &top->record_opts;
902
903         perf_evlist__config(evlist, opts);
904
905         evlist__for_each(evlist, counter) {
906 try_again:
907                 if (perf_evsel__open(counter, top->evlist->cpus,
908                                      top->evlist->threads) < 0) {
909                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
910                                 if (verbose)
911                                         ui__warning("%s\n", msg);
912                                 goto try_again;
913                         }
914
915                         perf_evsel__open_strerror(counter, &opts->target,
916                                                   errno, msg, sizeof(msg));
917                         ui__error("%s\n", msg);
918                         goto out_err;
919                 }
920         }
921
922         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
923                 ui__error("Failed to mmap with %d (%s)\n",
924                             errno, strerror_r(errno, msg, sizeof(msg)));
925                 goto out_err;
926         }
927
928         return 0;
929
930 out_err:
931         return -1;
932 }
933
934 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
935 {
936         if (!sort__has_sym) {
937                 if (symbol_conf.use_callchain) {
938                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
939                         return -EINVAL;
940                 }
941         } else if (callchain_param.mode != CHAIN_NONE) {
942                 if (callchain_register_param(&callchain_param) < 0) {
943                         ui__error("Can't register callchain params.\n");
944                         return -EINVAL;
945                 }
946         }
947
948         return 0;
949 }
950
951 static int __cmd_top(struct perf_top *top)
952 {
953         struct record_opts *opts = &top->record_opts;
954         pthread_t thread;
955         int ret;
956
957         top->session = perf_session__new(NULL, false, NULL);
958         if (top->session == NULL)
959                 return -1;
960
961         machines__set_symbol_filter(&top->session->machines, symbol_filter);
962
963         if (!objdump_path) {
964                 ret = perf_env__lookup_objdump(&top->session->header.env);
965                 if (ret)
966                         goto out_delete;
967         }
968
969         ret = perf_top__setup_sample_type(top);
970         if (ret)
971                 goto out_delete;
972
973         if (perf_session__register_idle_thread(top->session) == NULL)
974                 goto out_delete;
975
976         machine__synthesize_threads(&top->session->machines.host, &opts->target,
977                                     top->evlist->threads, false, opts->proc_map_timeout);
978
979         if (sort__has_socket) {
980                 ret = perf_env__read_cpu_topology_map(&perf_env);
981                 if (ret < 0)
982                         goto out_err_cpu_topo;
983         }
984
985         ret = perf_top__start_counters(top);
986         if (ret)
987                 goto out_delete;
988
989         top->session->evlist = top->evlist;
990         perf_session__set_id_hdr_size(top->session);
991
992         /*
993          * When perf is starting the traced process, all the events (apart from
994          * group members) have enable_on_exec=1 set, so don't spoil it by
995          * prematurely enabling them.
996          *
997          * XXX 'top' still doesn't start workloads like record, trace, but should,
998          * so leave the check here.
999          */
1000         if (!target__none(&opts->target))
1001                 perf_evlist__enable(top->evlist);
1002
1003         /* Wait for a minimal set of events before starting the snapshot */
1004         perf_evlist__poll(top->evlist, 100);
1005
1006         perf_top__mmap_read(top);
1007
1008         ret = -1;
1009         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1010                                                             display_thread), top)) {
1011                 ui__error("Could not create display thread.\n");
1012                 goto out_delete;
1013         }
1014
1015         if (top->realtime_prio) {
1016                 struct sched_param param;
1017
1018                 param.sched_priority = top->realtime_prio;
1019                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1020                         ui__error("Could not set realtime priority.\n");
1021                         goto out_join;
1022                 }
1023         }
1024
1025         while (!done) {
1026                 u64 hits = top->samples;
1027
1028                 perf_top__mmap_read(top);
1029
1030                 if (hits == top->samples)
1031                         ret = perf_evlist__poll(top->evlist, 100);
1032
1033                 if (resize) {
1034                         perf_top__resize(top);
1035                         resize = 0;
1036                 }
1037         }
1038
1039         ret = 0;
1040 out_join:
1041         pthread_join(thread, NULL);
1042 out_delete:
1043         perf_session__delete(top->session);
1044         top->session = NULL;
1045
1046         return ret;
1047
1048 out_err_cpu_topo: {
1049         char errbuf[BUFSIZ];
1050         const char *err = strerror_r(-ret, errbuf, sizeof(errbuf));
1051
1052         ui__error("Could not read the CPU topology map: %s\n", err);
1053         goto out_delete;
1054 }
1055 }
1056
1057 static int
1058 callchain_opt(const struct option *opt, const char *arg, int unset)
1059 {
1060         symbol_conf.use_callchain = true;
1061         return record_callchain_opt(opt, arg, unset);
1062 }
1063
1064 static int
1065 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1066 {
1067         struct record_opts *record = (struct record_opts *)opt->value;
1068
1069         record->callgraph_set = true;
1070         callchain_param.enabled = !unset;
1071         callchain_param.record_mode = CALLCHAIN_FP;
1072
1073         /*
1074          * --no-call-graph
1075          */
1076         if (unset) {
1077                 symbol_conf.use_callchain = false;
1078                 callchain_param.record_mode = CALLCHAIN_NONE;
1079                 return 0;
1080         }
1081
1082         return parse_callchain_top_opt(arg);
1083 }
1084
1085 static int perf_top_config(const char *var, const char *value, void *cb)
1086 {
1087         if (!strcmp(var, "top.call-graph"))
1088                 var = "call-graph.record-mode"; /* fall-through */
1089         if (!strcmp(var, "top.children")) {
1090                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1091                 return 0;
1092         }
1093
1094         return perf_default_config(var, value, cb);
1095 }
1096
1097 static int
1098 parse_percent_limit(const struct option *opt, const char *arg,
1099                     int unset __maybe_unused)
1100 {
1101         struct perf_top *top = opt->value;
1102
1103         top->min_percent = strtof(arg, NULL);
1104         return 0;
1105 }
1106
1107 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1108         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1109
1110 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1111 {
1112         char errbuf[BUFSIZ];
1113         struct perf_top top = {
1114                 .count_filter        = 5,
1115                 .delay_secs          = 2,
1116                 .record_opts = {
1117                         .mmap_pages     = UINT_MAX,
1118                         .user_freq      = UINT_MAX,
1119                         .user_interval  = ULLONG_MAX,
1120                         .freq           = 4000, /* 4 KHz */
1121                         .target         = {
1122                                 .uses_mmap   = true,
1123                         },
1124                         .proc_map_timeout    = 500,
1125                 },
1126                 .max_stack           = PERF_MAX_STACK_DEPTH,
1127                 .sym_pcnt_filter     = 5,
1128         };
1129         struct record_opts *opts = &top.record_opts;
1130         struct target *target = &opts->target;
1131         const struct option options[] = {
1132         OPT_CALLBACK('e', "event", &top.evlist, "event",
1133                      "event selector. use 'perf list' to list available events",
1134                      parse_events_option),
1135         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1136         OPT_STRING('p', "pid", &target->pid, "pid",
1137                     "profile events on existing process id"),
1138         OPT_STRING('t', "tid", &target->tid, "tid",
1139                     "profile events on existing thread id"),
1140         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1141                             "system-wide collection from all CPUs"),
1142         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1143                     "list of cpus to monitor"),
1144         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1145                    "file", "vmlinux pathname"),
1146         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1147                     "don't load vmlinux even if found"),
1148         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1149                     "hide kernel symbols"),
1150         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1151                      "number of mmap data pages",
1152                      perf_evlist__parse_mmap_pages),
1153         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1154                     "collect data with this RT SCHED_FIFO priority"),
1155         OPT_INTEGER('d', "delay", &top.delay_secs,
1156                     "number of seconds to delay between refreshes"),
1157         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1158                             "dump the symbol table used for profiling"),
1159         OPT_INTEGER('f', "count-filter", &top.count_filter,
1160                     "only display functions with more events than this"),
1161         OPT_BOOLEAN(0, "group", &opts->group,
1162                             "put the counters into a counter group"),
1163         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1164                     "child tasks do not inherit counters"),
1165         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1166                     "symbol to annotate"),
1167         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1168         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1169         OPT_INTEGER('E', "entries", &top.print_entries,
1170                     "display this many functions"),
1171         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1172                     "hide user symbols"),
1173         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1174         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1175         OPT_INCR('v', "verbose", &verbose,
1176                     "be more verbose (show counter open errors, etc)"),
1177         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1178                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1179                    " Please refer the man page for the complete list."),
1180         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1181                    "output field(s): overhead, period, sample plus all of sort keys"),
1182         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1183                     "Show a column with the number of samples"),
1184         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1185                            NULL, "enables call-graph recording and display",
1186                            &callchain_opt),
1187         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1188                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1189                      top_callchain_help, &parse_callchain_opt),
1190         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1191                     "Accumulate callchains of children and show total overhead as well"),
1192         OPT_INTEGER(0, "max-stack", &top.max_stack,
1193                     "Set the maximum stack depth when parsing the callchain. "
1194                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1195         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1196                    "ignore callees of these functions in call graphs",
1197                    report_parse_ignore_callees_opt),
1198         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1199                     "Show a column with the sum of periods"),
1200         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1201                    "only consider symbols in these dsos"),
1202         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1203                    "only consider symbols in these comms"),
1204         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1205                    "only consider these symbols"),
1206         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1207                     "Interleave source code with assembly code (default)"),
1208         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1209                     "Display raw encoding of assembly instructions (default)"),
1210         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1211                     "Enable kernel symbol demangling"),
1212         OPT_STRING(0, "objdump", &objdump_path, "path",
1213                     "objdump binary to use for disassembly and annotations"),
1214         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1215                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1216         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1217         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1218                      "Don't show entries under that percent", parse_percent_limit),
1219         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1220                      "How to display percentage of filtered entries", parse_filter_percentage),
1221         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1222                    "width[,width...]",
1223                    "don't try to adjust column width, use these fixed values"),
1224         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1225                         "per thread proc mmap processing timeout in ms"),
1226         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1227                      "branch any", "sample any taken branches",
1228                      parse_branch_stack),
1229         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1230                      "branch filter mask", "branch stack filter modes",
1231                      parse_branch_stack),
1232         OPT_END()
1233         };
1234         const char * const top_usage[] = {
1235                 "perf top [<options>]",
1236                 NULL
1237         };
1238         int status = hists__init();
1239
1240         if (status < 0)
1241                 return status;
1242
1243         top.evlist = perf_evlist__new();
1244         if (top.evlist == NULL)
1245                 return -ENOMEM;
1246
1247         perf_config(perf_top_config, &top);
1248
1249         argc = parse_options(argc, argv, options, top_usage, 0);
1250         if (argc)
1251                 usage_with_options(top_usage, options);
1252
1253         sort__mode = SORT_MODE__TOP;
1254         /* display thread wants entries to be collapsed in a different tree */
1255         sort__need_collapse = 1;
1256
1257         if (setup_sorting() < 0) {
1258                 if (sort_order)
1259                         parse_options_usage(top_usage, options, "s", 1);
1260                 if (field_order)
1261                         parse_options_usage(sort_order ? NULL : top_usage,
1262                                             options, "fields", 0);
1263                 goto out_delete_evlist;
1264         }
1265
1266         if (top.use_stdio)
1267                 use_browser = 0;
1268         else if (top.use_tui)
1269                 use_browser = 1;
1270
1271         setup_browser(false);
1272
1273         status = target__validate(target);
1274         if (status) {
1275                 target__strerror(target, status, errbuf, BUFSIZ);
1276                 ui__warning("%s\n", errbuf);
1277         }
1278
1279         status = target__parse_uid(target);
1280         if (status) {
1281                 int saved_errno = errno;
1282
1283                 target__strerror(target, status, errbuf, BUFSIZ);
1284                 ui__error("%s\n", errbuf);
1285
1286                 status = -saved_errno;
1287                 goto out_delete_evlist;
1288         }
1289
1290         if (target__none(target))
1291                 target->system_wide = true;
1292
1293         if (perf_evlist__create_maps(top.evlist, target) < 0)
1294                 usage_with_options(top_usage, options);
1295
1296         if (!top.evlist->nr_entries &&
1297             perf_evlist__add_default(top.evlist) < 0) {
1298                 ui__error("Not enough memory for event selector list\n");
1299                 goto out_delete_evlist;
1300         }
1301
1302         symbol_conf.nr_events = top.evlist->nr_entries;
1303
1304         if (top.delay_secs < 1)
1305                 top.delay_secs = 1;
1306
1307         if (record_opts__config(opts)) {
1308                 status = -EINVAL;
1309                 goto out_delete_evlist;
1310         }
1311
1312         top.sym_evsel = perf_evlist__first(top.evlist);
1313
1314         if (!symbol_conf.use_callchain) {
1315                 symbol_conf.cumulate_callchain = false;
1316                 perf_hpp__cancel_cumulate();
1317         }
1318
1319         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1320                 callchain_param.order = ORDER_CALLER;
1321
1322         symbol_conf.priv_size = sizeof(struct annotation);
1323
1324         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1325         status = symbol__init(NULL);
1326         if (status < 0)
1327                 goto out_delete_evlist;
1328
1329         sort__setup_elide(stdout);
1330
1331         get_term_dimensions(&top.winsize);
1332         if (top.print_entries == 0) {
1333                 struct sigaction act = {
1334                         .sa_sigaction = perf_top__sig_winch,
1335                         .sa_flags     = SA_SIGINFO,
1336                 };
1337                 perf_top__update_print_entries(&top);
1338                 sigaction(SIGWINCH, &act, NULL);
1339         }
1340
1341         status = __cmd_top(&top);
1342
1343 out_delete_evlist:
1344         perf_evlist__delete(top.evlist);
1345
1346         return status;
1347 }