ee30372f771335bf9bdfd2ece698db895912fffe
[releases.git] / builtin-top.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * builtin-top.c
4  *
5  * Builtin top command: Display a continuously updated profile of
6  * any workload, CPU or specific PID.
7  *
8  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
10  *
11  * Improvements and fixes by:
12  *
13  *   Arjan van de Ven <arjan@linux.intel.com>
14  *   Yanmin Zhang <yanmin.zhang@intel.com>
15  *   Wu Fengguang <fengguang.wu@intel.com>
16  *   Mike Galbraith <efault@gmx.de>
17  *   Paul Mackerras <paulus@samba.org>
18  */
19 #include "builtin.h"
20
21 #include "perf.h"
22
23 #include "util/annotate.h"
24 #include "util/bpf-event.h"
25 #include "util/config.h"
26 #include "util/color.h"
27 #include "util/dso.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/evsel_config.h"
31 #include "util/event.h"
32 #include "util/machine.h"
33 #include "util/map.h"
34 #include "util/mmap.h"
35 #include "util/session.h"
36 #include "util/thread.h"
37 #include "util/symbol.h"
38 #include "util/synthetic-events.h"
39 #include "util/top.h"
40 #include "util/util.h"
41 #include <linux/rbtree.h>
42 #include <subcmd/parse-options.h>
43 #include "util/parse-events.h"
44 #include "util/callchain.h"
45 #include "util/cpumap.h"
46 #include "util/sort.h"
47 #include "util/string2.h"
48 #include "util/term.h"
49 #include "util/intlist.h"
50 #include "util/parse-branch-options.h"
51 #include "arch/common.h"
52 #include "ui/ui.h"
53
54 #include "util/debug.h"
55 #include "util/ordered-events.h"
56 #include "util/pfm.h"
57
58 #include <assert.h>
59 #include <elf.h>
60 #include <fcntl.h>
61
62 #include <stdio.h>
63 #include <termios.h>
64 #include <unistd.h>
65 #include <inttypes.h>
66
67 #include <errno.h>
68 #include <time.h>
69 #include <sched.h>
70 #include <signal.h>
71
72 #include <sys/syscall.h>
73 #include <sys/ioctl.h>
74 #include <poll.h>
75 #include <sys/prctl.h>
76 #include <sys/wait.h>
77 #include <sys/uio.h>
78 #include <sys/utsname.h>
79 #include <sys/mman.h>
80
81 #include <linux/stringify.h>
82 #include <linux/time64.h>
83 #include <linux/types.h>
84 #include <linux/err.h>
85
86 #include <linux/ctype.h>
87 #include <perf/mmap.h>
88
89 static volatile int done;
90 static volatile int resize;
91
92 #define HEADER_LINE_NR  5
93
94 static void perf_top__update_print_entries(struct perf_top *top)
95 {
96         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
97 }
98
99 static void winch_sig(int sig __maybe_unused)
100 {
101         resize = 1;
102 }
103
104 static void perf_top__resize(struct perf_top *top)
105 {
106         get_term_dimensions(&top->winsize);
107         perf_top__update_print_entries(top);
108 }
109
110 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
111 {
112         struct evsel *evsel;
113         struct symbol *sym;
114         struct annotation *notes;
115         struct map *map;
116         int err = -1;
117
118         if (!he || !he->ms.sym)
119                 return -1;
120
121         evsel = hists_to_evsel(he->hists);
122
123         sym = he->ms.sym;
124         map = he->ms.map;
125
126         /*
127          * We can't annotate with just /proc/kallsyms
128          */
129         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
130             !dso__is_kcore(map->dso)) {
131                 pr_err("Can't annotate %s: No vmlinux file was found in the "
132                        "path\n", sym->name);
133                 sleep(1);
134                 return -1;
135         }
136
137         notes = symbol__annotation(sym);
138         pthread_mutex_lock(&notes->lock);
139
140         if (!symbol__hists(sym, top->evlist->core.nr_entries)) {
141                 pthread_mutex_unlock(&notes->lock);
142                 pr_err("Not enough memory for annotating '%s' symbol!\n",
143                        sym->name);
144                 sleep(1);
145                 return err;
146         }
147
148         err = symbol__annotate(&he->ms, evsel, &top->annotation_opts, NULL);
149         if (err == 0) {
150                 top->sym_filter_entry = he;
151         } else {
152                 char msg[BUFSIZ];
153                 symbol__strerror_disassemble(&he->ms, err, msg, sizeof(msg));
154                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
155         }
156
157         pthread_mutex_unlock(&notes->lock);
158         return err;
159 }
160
161 static void __zero_source_counters(struct hist_entry *he)
162 {
163         struct symbol *sym = he->ms.sym;
164         symbol__annotate_zero_histograms(sym);
165 }
166
167 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
168 {
169         struct utsname uts;
170         int err = uname(&uts);
171
172         ui__warning("Out of bounds address found:\n\n"
173                     "Addr:   %" PRIx64 "\n"
174                     "DSO:    %s %c\n"
175                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
176                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
177                     "Arch:   %s\n"
178                     "Kernel: %s\n"
179                     "Tools:  %s\n\n"
180                     "Not all samples will be on the annotation output.\n\n"
181                     "Please report to linux-kernel@vger.kernel.org\n",
182                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
183                     map->start, map->end, sym->start, sym->end,
184                     sym->binding == STB_GLOBAL ? 'g' :
185                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
186                     err ? "[unknown]" : uts.machine,
187                     err ? "[unknown]" : uts.release, perf_version_string);
188         if (use_browser <= 0)
189                 sleep(5);
190
191         map->erange_warned = true;
192 }
193
194 static void perf_top__record_precise_ip(struct perf_top *top,
195                                         struct hist_entry *he,
196                                         struct perf_sample *sample,
197                                         struct evsel *evsel, u64 ip)
198 {
199         struct annotation *notes;
200         struct symbol *sym = he->ms.sym;
201         int err = 0;
202
203         if (sym == NULL || (use_browser == 0 &&
204                             (top->sym_filter_entry == NULL ||
205                              top->sym_filter_entry->ms.sym != sym)))
206                 return;
207
208         notes = symbol__annotation(sym);
209
210         if (pthread_mutex_trylock(&notes->lock))
211                 return;
212
213         err = hist_entry__inc_addr_samples(he, sample, evsel, ip);
214
215         pthread_mutex_unlock(&notes->lock);
216
217         if (unlikely(err)) {
218                 /*
219                  * This function is now called with he->hists->lock held.
220                  * Release it before going to sleep.
221                  */
222                 pthread_mutex_unlock(&he->hists->lock);
223
224                 if (err == -ERANGE && !he->ms.map->erange_warned)
225                         ui__warn_map_erange(he->ms.map, sym, ip);
226                 else if (err == -ENOMEM) {
227                         pr_err("Not enough memory for annotating '%s' symbol!\n",
228                                sym->name);
229                         sleep(1);
230                 }
231
232                 pthread_mutex_lock(&he->hists->lock);
233         }
234 }
235
236 static void perf_top__show_details(struct perf_top *top)
237 {
238         struct hist_entry *he = top->sym_filter_entry;
239         struct evsel *evsel;
240         struct annotation *notes;
241         struct symbol *symbol;
242         int more;
243
244         if (!he)
245                 return;
246
247         evsel = hists_to_evsel(he->hists);
248
249         symbol = he->ms.sym;
250         notes = symbol__annotation(symbol);
251
252         pthread_mutex_lock(&notes->lock);
253
254         symbol__calc_percent(symbol, evsel);
255
256         if (notes->src == NULL)
257                 goto out_unlock;
258
259         printf("Showing %s for %s\n", evsel__name(top->sym_evsel), symbol->name);
260         printf("  Events  Pcnt (>=%d%%)\n", top->annotation_opts.min_pcnt);
261
262         more = symbol__annotate_printf(&he->ms, top->sym_evsel, &top->annotation_opts);
263
264         if (top->evlist->enabled) {
265                 if (top->zero)
266                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
267                 else
268                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
269         }
270         if (more != 0)
271                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
272 out_unlock:
273         pthread_mutex_unlock(&notes->lock);
274 }
275
276 static void perf_top__resort_hists(struct perf_top *t)
277 {
278         struct evlist *evlist = t->evlist;
279         struct evsel *pos;
280
281         evlist__for_each_entry(evlist, pos) {
282                 struct hists *hists = evsel__hists(pos);
283
284                 /*
285                  * unlink existing entries so that they can be linked
286                  * in a correct order in hists__match() below.
287                  */
288                 hists__unlink(hists);
289
290                 if (evlist->enabled) {
291                         if (t->zero) {
292                                 hists__delete_entries(hists);
293                         } else {
294                                 hists__decay_entries(hists, t->hide_user_symbols,
295                                                      t->hide_kernel_symbols);
296                         }
297                 }
298
299                 hists__collapse_resort(hists, NULL);
300
301                 /* Non-group events are considered as leader */
302                 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
303                         struct hists *leader_hists = evsel__hists(pos->leader);
304
305                         hists__match(leader_hists, hists);
306                         hists__link(leader_hists, hists);
307                 }
308         }
309
310         evlist__for_each_entry(evlist, pos) {
311                 evsel__output_resort(pos, NULL);
312         }
313 }
314
315 static void perf_top__print_sym_table(struct perf_top *top)
316 {
317         char bf[160];
318         int printed = 0;
319         const int win_width = top->winsize.ws_col - 1;
320         struct evsel *evsel = top->sym_evsel;
321         struct hists *hists = evsel__hists(evsel);
322
323         puts(CONSOLE_CLEAR);
324
325         perf_top__header_snprintf(top, bf, sizeof(bf));
326         printf("%s\n", bf);
327
328         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
329
330         if (!top->record_opts.overwrite &&
331             (hists->stats.nr_lost_warned !=
332             hists->stats.nr_events[PERF_RECORD_LOST])) {
333                 hists->stats.nr_lost_warned =
334                               hists->stats.nr_events[PERF_RECORD_LOST];
335                 color_fprintf(stdout, PERF_COLOR_RED,
336                               "WARNING: LOST %d chunks, Check IO/CPU overload",
337                               hists->stats.nr_lost_warned);
338                 ++printed;
339         }
340
341         if (top->sym_filter_entry) {
342                 perf_top__show_details(top);
343                 return;
344         }
345
346         perf_top__resort_hists(top);
347
348         hists__output_recalc_col_len(hists, top->print_entries - printed);
349         putchar('\n');
350         hists__fprintf(hists, false, top->print_entries - printed, win_width,
351                        top->min_percent, stdout, !symbol_conf.use_callchain);
352 }
353
354 static void prompt_integer(int *target, const char *msg)
355 {
356         char *buf = malloc(0), *p;
357         size_t dummy = 0;
358         int tmp;
359
360         fprintf(stdout, "\n%s: ", msg);
361         if (getline(&buf, &dummy, stdin) < 0)
362                 return;
363
364         p = strchr(buf, '\n');
365         if (p)
366                 *p = 0;
367
368         p = buf;
369         while(*p) {
370                 if (!isdigit(*p))
371                         goto out_free;
372                 p++;
373         }
374         tmp = strtoul(buf, NULL, 10);
375         *target = tmp;
376 out_free:
377         free(buf);
378 }
379
380 static void prompt_percent(int *target, const char *msg)
381 {
382         int tmp = 0;
383
384         prompt_integer(&tmp, msg);
385         if (tmp >= 0 && tmp <= 100)
386                 *target = tmp;
387 }
388
389 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
390 {
391         char *buf = malloc(0), *p;
392         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
393         struct hists *hists = evsel__hists(top->sym_evsel);
394         struct rb_node *next;
395         size_t dummy = 0;
396
397         /* zero counters of active symbol */
398         if (syme) {
399                 __zero_source_counters(syme);
400                 top->sym_filter_entry = NULL;
401         }
402
403         fprintf(stdout, "\n%s: ", msg);
404         if (getline(&buf, &dummy, stdin) < 0)
405                 goto out_free;
406
407         p = strchr(buf, '\n');
408         if (p)
409                 *p = 0;
410
411         next = rb_first_cached(&hists->entries);
412         while (next) {
413                 n = rb_entry(next, struct hist_entry, rb_node);
414                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
415                         found = n;
416                         break;
417                 }
418                 next = rb_next(&n->rb_node);
419         }
420
421         if (!found) {
422                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
423                 sleep(1);
424         } else
425                 perf_top__parse_source(top, found);
426
427 out_free:
428         free(buf);
429 }
430
431 static void perf_top__print_mapped_keys(struct perf_top *top)
432 {
433         char *name = NULL;
434
435         if (top->sym_filter_entry) {
436                 struct symbol *sym = top->sym_filter_entry->ms.sym;
437                 name = sym->name;
438         }
439
440         fprintf(stdout, "\nMapped keys:\n");
441         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
442         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
443
444         if (top->evlist->core.nr_entries > 1)
445                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", evsel__name(top->sym_evsel));
446
447         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
448
449         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->annotation_opts.min_pcnt);
450         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
451         fprintf(stdout, "\t[S]     stop annotation.\n");
452
453         fprintf(stdout,
454                 "\t[K]     hide kernel symbols.             \t(%s)\n",
455                 top->hide_kernel_symbols ? "yes" : "no");
456         fprintf(stdout,
457                 "\t[U]     hide user symbols.               \t(%s)\n",
458                 top->hide_user_symbols ? "yes" : "no");
459         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
460         fprintf(stdout, "\t[qQ]    quit.\n");
461 }
462
463 static int perf_top__key_mapped(struct perf_top *top, int c)
464 {
465         switch (c) {
466                 case 'd':
467                 case 'e':
468                 case 'f':
469                 case 'z':
470                 case 'q':
471                 case 'Q':
472                 case 'K':
473                 case 'U':
474                 case 'F':
475                 case 's':
476                 case 'S':
477                         return 1;
478                 case 'E':
479                         return top->evlist->core.nr_entries > 1 ? 1 : 0;
480                 default:
481                         break;
482         }
483
484         return 0;
485 }
486
487 static bool perf_top__handle_keypress(struct perf_top *top, int c)
488 {
489         bool ret = true;
490
491         if (!perf_top__key_mapped(top, c)) {
492                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
493                 struct termios save;
494
495                 perf_top__print_mapped_keys(top);
496                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
497                 fflush(stdout);
498
499                 set_term_quiet_input(&save);
500
501                 poll(&stdin_poll, 1, -1);
502                 c = getc(stdin);
503
504                 tcsetattr(0, TCSAFLUSH, &save);
505                 if (!perf_top__key_mapped(top, c))
506                         return ret;
507         }
508
509         switch (c) {
510                 case 'd':
511                         prompt_integer(&top->delay_secs, "Enter display delay");
512                         if (top->delay_secs < 1)
513                                 top->delay_secs = 1;
514                         break;
515                 case 'e':
516                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
517                         if (top->print_entries == 0) {
518                                 perf_top__resize(top);
519                                 signal(SIGWINCH, winch_sig);
520                         } else {
521                                 signal(SIGWINCH, SIG_DFL);
522                         }
523                         break;
524                 case 'E':
525                         if (top->evlist->core.nr_entries > 1) {
526                                 /* Select 0 as the default event: */
527                                 int counter = 0;
528
529                                 fprintf(stderr, "\nAvailable events:");
530
531                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
532                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, evsel__name(top->sym_evsel));
533
534                                 prompt_integer(&counter, "Enter details event counter");
535
536                                 if (counter >= top->evlist->core.nr_entries) {
537                                         top->sym_evsel = evlist__first(top->evlist);
538                                         fprintf(stderr, "Sorry, no such event, using %s.\n", evsel__name(top->sym_evsel));
539                                         sleep(1);
540                                         break;
541                                 }
542                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
543                                         if (top->sym_evsel->idx == counter)
544                                                 break;
545                         } else
546                                 top->sym_evsel = evlist__first(top->evlist);
547                         break;
548                 case 'f':
549                         prompt_integer(&top->count_filter, "Enter display event count filter");
550                         break;
551                 case 'F':
552                         prompt_percent(&top->annotation_opts.min_pcnt,
553                                        "Enter details display event filter (percent)");
554                         break;
555                 case 'K':
556                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
557                         break;
558                 case 'q':
559                 case 'Q':
560                         printf("exiting.\n");
561                         if (top->dump_symtab)
562                                 perf_session__fprintf_dsos(top->session, stderr);
563                         ret = false;
564                         break;
565                 case 's':
566                         perf_top__prompt_symbol(top, "Enter details symbol");
567                         break;
568                 case 'S':
569                         if (!top->sym_filter_entry)
570                                 break;
571                         else {
572                                 struct hist_entry *syme = top->sym_filter_entry;
573
574                                 top->sym_filter_entry = NULL;
575                                 __zero_source_counters(syme);
576                         }
577                         break;
578                 case 'U':
579                         top->hide_user_symbols = !top->hide_user_symbols;
580                         break;
581                 case 'z':
582                         top->zero = !top->zero;
583                         break;
584                 default:
585                         break;
586         }
587
588         return ret;
589 }
590
591 static void perf_top__sort_new_samples(void *arg)
592 {
593         struct perf_top *t = arg;
594
595         if (t->evlist->selected != NULL)
596                 t->sym_evsel = t->evlist->selected;
597
598         perf_top__resort_hists(t);
599
600         if (t->lost || t->drop)
601                 pr_warning("Too slow to read ring buffer (change period (-c/-F) or limit CPUs (-C)\n");
602 }
603
604 static void stop_top(void)
605 {
606         session_done = 1;
607         done = 1;
608 }
609
610 static void *display_thread_tui(void *arg)
611 {
612         struct evsel *pos;
613         struct perf_top *top = arg;
614         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
615         struct hist_browser_timer hbt = {
616                 .timer          = perf_top__sort_new_samples,
617                 .arg            = top,
618                 .refresh        = top->delay_secs,
619         };
620         int ret;
621
622         /* In order to read symbols from other namespaces perf to  needs to call
623          * setns(2).  This isn't permitted if the struct_fs has multiple users.
624          * unshare(2) the fs so that we may continue to setns into namespaces
625          * that we're observing.
626          */
627         unshare(CLONE_FS);
628
629         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
630
631 repeat:
632         perf_top__sort_new_samples(top);
633
634         /*
635          * Initialize the uid_filter_str, in the future the TUI will allow
636          * Zooming in/out UIDs. For now just use whatever the user passed
637          * via --uid.
638          */
639         evlist__for_each_entry(top->evlist, pos) {
640                 struct hists *hists = evsel__hists(pos);
641                 hists->uid_filter_str = top->record_opts.target.uid_str;
642         }
643
644         ret = perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
645                                       top->min_percent,
646                                       &top->session->header.env,
647                                       !top->record_opts.overwrite,
648                                       &top->annotation_opts);
649
650         if (ret == K_RELOAD) {
651                 top->zero = true;
652                 goto repeat;
653         } else
654                 stop_top();
655
656         return NULL;
657 }
658
659 static void display_sig(int sig __maybe_unused)
660 {
661         stop_top();
662 }
663
664 static void display_setup_sig(void)
665 {
666         signal(SIGSEGV, sighandler_dump_stack);
667         signal(SIGFPE, sighandler_dump_stack);
668         signal(SIGINT,  display_sig);
669         signal(SIGQUIT, display_sig);
670         signal(SIGTERM, display_sig);
671 }
672
673 static void *display_thread(void *arg)
674 {
675         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
676         struct termios save;
677         struct perf_top *top = arg;
678         int delay_msecs, c;
679
680         /* In order to read symbols from other namespaces perf to  needs to call
681          * setns(2).  This isn't permitted if the struct_fs has multiple users.
682          * unshare(2) the fs so that we may continue to setns into namespaces
683          * that we're observing.
684          */
685         unshare(CLONE_FS);
686
687         prctl(PR_SET_NAME, "perf-top-UI", 0, 0, 0);
688
689         display_setup_sig();
690         pthread__unblock_sigwinch();
691 repeat:
692         delay_msecs = top->delay_secs * MSEC_PER_SEC;
693         set_term_quiet_input(&save);
694         /* trash return*/
695         clearerr(stdin);
696         if (poll(&stdin_poll, 1, 0) > 0)
697                 getc(stdin);
698
699         while (!done) {
700                 perf_top__print_sym_table(top);
701                 /*
702                  * Either timeout expired or we got an EINTR due to SIGWINCH,
703                  * refresh screen in both cases.
704                  */
705                 switch (poll(&stdin_poll, 1, delay_msecs)) {
706                 case 0:
707                         continue;
708                 case -1:
709                         if (errno == EINTR)
710                                 continue;
711                         __fallthrough;
712                 default:
713                         c = getc(stdin);
714                         tcsetattr(0, TCSAFLUSH, &save);
715
716                         if (perf_top__handle_keypress(top, c))
717                                 goto repeat;
718                         stop_top();
719                 }
720         }
721
722         tcsetattr(0, TCSAFLUSH, &save);
723         return NULL;
724 }
725
726 static int hist_iter__top_callback(struct hist_entry_iter *iter,
727                                    struct addr_location *al, bool single,
728                                    void *arg)
729 {
730         struct perf_top *top = arg;
731         struct hist_entry *he = iter->he;
732         struct evsel *evsel = iter->evsel;
733
734         if (perf_hpp_list.sym && single)
735                 perf_top__record_precise_ip(top, he, iter->sample, evsel, al->addr);
736
737         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
738                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY),
739                      NULL);
740         return 0;
741 }
742
743 static void perf_event__process_sample(struct perf_tool *tool,
744                                        const union perf_event *event,
745                                        struct evsel *evsel,
746                                        struct perf_sample *sample,
747                                        struct machine *machine)
748 {
749         struct perf_top *top = container_of(tool, struct perf_top, tool);
750         struct addr_location al;
751         int err;
752
753         if (!machine && perf_guest) {
754                 static struct intlist *seen;
755
756                 if (!seen)
757                         seen = intlist__new(NULL);
758
759                 if (!intlist__has_entry(seen, sample->pid)) {
760                         pr_err("Can't find guest [%d]'s kernel information\n",
761                                 sample->pid);
762                         intlist__add(seen, sample->pid);
763                 }
764                 return;
765         }
766
767         if (!machine) {
768                 pr_err("%u unprocessable samples recorded.\r",
769                        top->session->evlist->stats.nr_unprocessable_samples++);
770                 return;
771         }
772
773         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
774                 top->exact_samples++;
775
776         if (machine__resolve(machine, &al, sample) < 0)
777                 return;
778
779         if (top->stitch_lbr)
780                 al.thread->lbr_stitch_enable = true;
781
782         if (!machine->kptr_restrict_warned &&
783             symbol_conf.kptr_restrict &&
784             al.cpumode == PERF_RECORD_MISC_KERNEL) {
785                 if (!perf_evlist__exclude_kernel(top->session->evlist)) {
786                         ui__warning(
787 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
788 "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
789 "Kernel%s samples will not be resolved.\n",
790                           al.map && map__has_symbols(al.map) ?
791                           " modules" : "");
792                         if (use_browser <= 0)
793                                 sleep(5);
794                 }
795                 machine->kptr_restrict_warned = true;
796         }
797
798         if (al.sym == NULL && al.map != NULL) {
799                 const char *msg = "Kernel samples will not be resolved.\n";
800                 /*
801                  * As we do lazy loading of symtabs we only will know if the
802                  * specified vmlinux file is invalid when we actually have a
803                  * hit in kernel space and then try to load it. So if we get
804                  * here and there are _no_ symbols in the DSO backing the
805                  * kernel map, bail out.
806                  *
807                  * We may never get here, for instance, if we use -K/
808                  * --hide-kernel-symbols, even if the user specifies an
809                  * invalid --vmlinux ;-)
810                  */
811                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
812                     __map__is_kernel(al.map) && map__has_symbols(al.map)) {
813                         if (symbol_conf.vmlinux_name) {
814                                 char serr[256];
815                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
816                                 ui__warning("The %s file can't be used: %s\n%s",
817                                             symbol_conf.vmlinux_name, serr, msg);
818                         } else {
819                                 ui__warning("A vmlinux file was not found.\n%s",
820                                             msg);
821                         }
822
823                         if (use_browser <= 0)
824                                 sleep(5);
825                         top->vmlinux_warned = true;
826                 }
827         }
828
829         if (al.sym == NULL || !al.sym->idle) {
830                 struct hists *hists = evsel__hists(evsel);
831                 struct hist_entry_iter iter = {
832                         .evsel          = evsel,
833                         .sample         = sample,
834                         .add_entry_cb   = hist_iter__top_callback,
835                 };
836
837                 if (symbol_conf.cumulate_callchain)
838                         iter.ops = &hist_iter_cumulative;
839                 else
840                         iter.ops = &hist_iter_normal;
841
842                 pthread_mutex_lock(&hists->lock);
843
844                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
845                 if (err < 0)
846                         pr_err("Problem incrementing symbol period, skipping event\n");
847
848                 pthread_mutex_unlock(&hists->lock);
849         }
850
851         addr_location__put(&al);
852 }
853
854 static void
855 perf_top__process_lost(struct perf_top *top, union perf_event *event,
856                        struct evsel *evsel)
857 {
858         struct hists *hists = evsel__hists(evsel);
859
860         top->lost += event->lost.lost;
861         top->lost_total += event->lost.lost;
862         hists->stats.total_lost += event->lost.lost;
863 }
864
865 static void
866 perf_top__process_lost_samples(struct perf_top *top,
867                                union perf_event *event,
868                                struct evsel *evsel)
869 {
870         struct hists *hists = evsel__hists(evsel);
871
872         top->lost += event->lost_samples.lost;
873         top->lost_total += event->lost_samples.lost;
874         hists->stats.total_lost_samples += event->lost_samples.lost;
875 }
876
877 static u64 last_timestamp;
878
879 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
880 {
881         struct record_opts *opts = &top->record_opts;
882         struct evlist *evlist = top->evlist;
883         struct mmap *md;
884         union perf_event *event;
885
886         md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
887         if (perf_mmap__read_init(&md->core) < 0)
888                 return;
889
890         while ((event = perf_mmap__read_event(&md->core)) != NULL) {
891                 int ret;
892
893                 ret = perf_evlist__parse_sample_timestamp(evlist, event, &last_timestamp);
894                 if (ret && ret != -1)
895                         break;
896
897                 ret = ordered_events__queue(top->qe.in, event, last_timestamp, 0);
898                 if (ret)
899                         break;
900
901                 perf_mmap__consume(&md->core);
902
903                 if (top->qe.rotate) {
904                         pthread_mutex_lock(&top->qe.mutex);
905                         top->qe.rotate = false;
906                         pthread_cond_signal(&top->qe.cond);
907                         pthread_mutex_unlock(&top->qe.mutex);
908                 }
909         }
910
911         perf_mmap__read_done(&md->core);
912 }
913
914 static void perf_top__mmap_read(struct perf_top *top)
915 {
916         bool overwrite = top->record_opts.overwrite;
917         struct evlist *evlist = top->evlist;
918         int i;
919
920         if (overwrite)
921                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_DATA_PENDING);
922
923         for (i = 0; i < top->evlist->core.nr_mmaps; i++)
924                 perf_top__mmap_read_idx(top, i);
925
926         if (overwrite) {
927                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
928                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
929         }
930 }
931
932 /*
933  * Check per-event overwrite term.
934  * perf top should support consistent term for all events.
935  * - All events don't have per-event term
936  *   E.g. "cpu/cpu-cycles/,cpu/instructions/"
937  *   Nothing change, return 0.
938  * - All events have same per-event term
939  *   E.g. "cpu/cpu-cycles,no-overwrite/,cpu/instructions,no-overwrite/
940  *   Using the per-event setting to replace the opts->overwrite if
941  *   they are different, then return 0.
942  * - Events have different per-event term
943  *   E.g. "cpu/cpu-cycles,overwrite/,cpu/instructions,no-overwrite/"
944  *   Return -1
945  * - Some of the event set per-event term, but some not.
946  *   E.g. "cpu/cpu-cycles/,cpu/instructions,no-overwrite/"
947  *   Return -1
948  */
949 static int perf_top__overwrite_check(struct perf_top *top)
950 {
951         struct record_opts *opts = &top->record_opts;
952         struct evlist *evlist = top->evlist;
953         struct evsel_config_term *term;
954         struct list_head *config_terms;
955         struct evsel *evsel;
956         int set, overwrite = -1;
957
958         evlist__for_each_entry(evlist, evsel) {
959                 set = -1;
960                 config_terms = &evsel->config_terms;
961                 list_for_each_entry(term, config_terms, list) {
962                         if (term->type == EVSEL__CONFIG_TERM_OVERWRITE)
963                                 set = term->val.overwrite ? 1 : 0;
964                 }
965
966                 /* no term for current and previous event (likely) */
967                 if ((overwrite < 0) && (set < 0))
968                         continue;
969
970                 /* has term for both current and previous event, compare */
971                 if ((overwrite >= 0) && (set >= 0) && (overwrite != set))
972                         return -1;
973
974                 /* no term for current event but has term for previous one */
975                 if ((overwrite >= 0) && (set < 0))
976                         return -1;
977
978                 /* has term for current event */
979                 if ((overwrite < 0) && (set >= 0)) {
980                         /* if it's first event, set overwrite */
981                         if (evsel == evlist__first(evlist))
982                                 overwrite = set;
983                         else
984                                 return -1;
985                 }
986         }
987
988         if ((overwrite >= 0) && (opts->overwrite != overwrite))
989                 opts->overwrite = overwrite;
990
991         return 0;
992 }
993
994 static int perf_top_overwrite_fallback(struct perf_top *top,
995                                        struct evsel *evsel)
996 {
997         struct record_opts *opts = &top->record_opts;
998         struct evlist *evlist = top->evlist;
999         struct evsel *counter;
1000
1001         if (!opts->overwrite)
1002                 return 0;
1003
1004         /* only fall back when first event fails */
1005         if (evsel != evlist__first(evlist))
1006                 return 0;
1007
1008         evlist__for_each_entry(evlist, counter)
1009                 counter->core.attr.write_backward = false;
1010         opts->overwrite = false;
1011         pr_debug2("fall back to non-overwrite mode\n");
1012         return 1;
1013 }
1014
1015 static int perf_top__start_counters(struct perf_top *top)
1016 {
1017         char msg[BUFSIZ];
1018         struct evsel *counter;
1019         struct evlist *evlist = top->evlist;
1020         struct record_opts *opts = &top->record_opts;
1021
1022         if (perf_top__overwrite_check(top)) {
1023                 ui__error("perf top only support consistent per-event "
1024                           "overwrite setting for all events\n");
1025                 goto out_err;
1026         }
1027
1028         perf_evlist__config(evlist, opts, &callchain_param);
1029
1030         evlist__for_each_entry(evlist, counter) {
1031 try_again:
1032                 if (evsel__open(counter, top->evlist->core.cpus,
1033                                      top->evlist->core.threads) < 0) {
1034
1035                         /*
1036                          * Specially handle overwrite fall back.
1037                          * Because perf top is the only tool which has
1038                          * overwrite mode by default, support
1039                          * both overwrite and non-overwrite mode, and
1040                          * require consistent mode for all events.
1041                          *
1042                          * May move it to generic code with more tools
1043                          * have similar attribute.
1044                          */
1045                         if (perf_missing_features.write_backward &&
1046                             perf_top_overwrite_fallback(top, counter))
1047                                 goto try_again;
1048
1049                         if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
1050                                 if (verbose > 0)
1051                                         ui__warning("%s\n", msg);
1052                                 goto try_again;
1053                         }
1054
1055                         evsel__open_strerror(counter, &opts->target, errno, msg, sizeof(msg));
1056                         ui__error("%s\n", msg);
1057                         goto out_err;
1058                 }
1059         }
1060
1061         if (evlist__mmap(evlist, opts->mmap_pages) < 0) {
1062                 ui__error("Failed to mmap with %d (%s)\n",
1063                             errno, str_error_r(errno, msg, sizeof(msg)));
1064                 goto out_err;
1065         }
1066
1067         return 0;
1068
1069 out_err:
1070         return -1;
1071 }
1072
1073 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
1074 {
1075         if (callchain->mode != CHAIN_NONE) {
1076                 if (callchain_register_param(callchain) < 0) {
1077                         ui__error("Can't register callchain params.\n");
1078                         return -EINVAL;
1079                 }
1080         }
1081
1082         return 0;
1083 }
1084
1085 static struct ordered_events *rotate_queues(struct perf_top *top)
1086 {
1087         struct ordered_events *in = top->qe.in;
1088
1089         if (top->qe.in == &top->qe.data[1])
1090                 top->qe.in = &top->qe.data[0];
1091         else
1092                 top->qe.in = &top->qe.data[1];
1093
1094         return in;
1095 }
1096
1097 static void *process_thread(void *arg)
1098 {
1099         struct perf_top *top = arg;
1100
1101         while (!done) {
1102                 struct ordered_events *out, *in = top->qe.in;
1103
1104                 if (!in->nr_events) {
1105                         usleep(100);
1106                         continue;
1107                 }
1108
1109                 out = rotate_queues(top);
1110
1111                 pthread_mutex_lock(&top->qe.mutex);
1112                 top->qe.rotate = true;
1113                 pthread_cond_wait(&top->qe.cond, &top->qe.mutex);
1114                 pthread_mutex_unlock(&top->qe.mutex);
1115
1116                 if (ordered_events__flush(out, OE_FLUSH__TOP))
1117                         pr_err("failed to process events\n");
1118         }
1119
1120         return NULL;
1121 }
1122
1123 /*
1124  * Allow only 'top->delay_secs' seconds behind samples.
1125  */
1126 static int should_drop(struct ordered_event *qevent, struct perf_top *top)
1127 {
1128         union perf_event *event = qevent->event;
1129         u64 delay_timestamp;
1130
1131         if (event->header.type != PERF_RECORD_SAMPLE)
1132                 return false;
1133
1134         delay_timestamp = qevent->timestamp + top->delay_secs * NSEC_PER_SEC;
1135         return delay_timestamp < last_timestamp;
1136 }
1137
1138 static int deliver_event(struct ordered_events *qe,
1139                          struct ordered_event *qevent)
1140 {
1141         struct perf_top *top = qe->data;
1142         struct evlist *evlist = top->evlist;
1143         struct perf_session *session = top->session;
1144         union perf_event *event = qevent->event;
1145         struct perf_sample sample;
1146         struct evsel *evsel;
1147         struct machine *machine;
1148         int ret = -1;
1149
1150         if (should_drop(qevent, top)) {
1151                 top->drop++;
1152                 top->drop_total++;
1153                 return 0;
1154         }
1155
1156         ret = perf_evlist__parse_sample(evlist, event, &sample);
1157         if (ret) {
1158                 pr_err("Can't parse sample, err = %d\n", ret);
1159                 goto next_event;
1160         }
1161
1162         evsel = perf_evlist__id2evsel(session->evlist, sample.id);
1163         assert(evsel != NULL);
1164
1165         if (event->header.type == PERF_RECORD_SAMPLE) {
1166                 if (evswitch__discard(&top->evswitch, evsel))
1167                         return 0;
1168                 ++top->samples;
1169         }
1170
1171         switch (sample.cpumode) {
1172         case PERF_RECORD_MISC_USER:
1173                 ++top->us_samples;
1174                 if (top->hide_user_symbols)
1175                         goto next_event;
1176                 machine = &session->machines.host;
1177                 break;
1178         case PERF_RECORD_MISC_KERNEL:
1179                 ++top->kernel_samples;
1180                 if (top->hide_kernel_symbols)
1181                         goto next_event;
1182                 machine = &session->machines.host;
1183                 break;
1184         case PERF_RECORD_MISC_GUEST_KERNEL:
1185                 ++top->guest_kernel_samples;
1186                 machine = perf_session__find_machine(session,
1187                                                      sample.pid);
1188                 break;
1189         case PERF_RECORD_MISC_GUEST_USER:
1190                 ++top->guest_us_samples;
1191                 /*
1192                  * TODO: we don't process guest user from host side
1193                  * except simple counting.
1194                  */
1195                 goto next_event;
1196         default:
1197                 if (event->header.type == PERF_RECORD_SAMPLE)
1198                         goto next_event;
1199                 machine = &session->machines.host;
1200                 break;
1201         }
1202
1203         if (event->header.type == PERF_RECORD_SAMPLE) {
1204                 perf_event__process_sample(&top->tool, event, evsel,
1205                                            &sample, machine);
1206         } else if (event->header.type == PERF_RECORD_LOST) {
1207                 perf_top__process_lost(top, event, evsel);
1208         } else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
1209                 perf_top__process_lost_samples(top, event, evsel);
1210         } else if (event->header.type < PERF_RECORD_MAX) {
1211                 hists__inc_nr_events(evsel__hists(evsel), event->header.type);
1212                 machine__process_event(machine, event, &sample);
1213         } else
1214                 ++session->evlist->stats.nr_unknown_events;
1215
1216         ret = 0;
1217 next_event:
1218         return ret;
1219 }
1220
1221 static void init_process_thread(struct perf_top *top)
1222 {
1223         ordered_events__init(&top->qe.data[0], deliver_event, top);
1224         ordered_events__init(&top->qe.data[1], deliver_event, top);
1225         ordered_events__set_copy_on_queue(&top->qe.data[0], true);
1226         ordered_events__set_copy_on_queue(&top->qe.data[1], true);
1227         top->qe.in = &top->qe.data[0];
1228         pthread_mutex_init(&top->qe.mutex, NULL);
1229         pthread_cond_init(&top->qe.cond, NULL);
1230 }
1231
1232 static int __cmd_top(struct perf_top *top)
1233 {
1234         struct record_opts *opts = &top->record_opts;
1235         pthread_t thread, thread_process;
1236         int ret;
1237
1238         if (!top->annotation_opts.objdump_path) {
1239                 ret = perf_env__lookup_objdump(&top->session->header.env,
1240                                                &top->annotation_opts.objdump_path);
1241                 if (ret)
1242                         return ret;
1243         }
1244
1245         ret = callchain_param__setup_sample_type(&callchain_param);
1246         if (ret)
1247                 return ret;
1248
1249         if (perf_session__register_idle_thread(top->session) < 0)
1250                 return ret;
1251
1252         if (top->nr_threads_synthesize > 1)
1253                 perf_set_multithreaded();
1254
1255         init_process_thread(top);
1256
1257         if (opts->record_namespaces)
1258                 top->tool.namespace_events = true;
1259         if (opts->record_cgroup) {
1260 #ifdef HAVE_FILE_HANDLE
1261                 top->tool.cgroup_events = true;
1262 #else
1263                 pr_err("cgroup tracking is not supported.\n");
1264                 return -1;
1265 #endif
1266         }
1267
1268         ret = perf_event__synthesize_bpf_events(top->session, perf_event__process,
1269                                                 &top->session->machines.host,
1270                                                 &top->record_opts);
1271         if (ret < 0)
1272                 pr_debug("Couldn't synthesize BPF events: Pre-existing BPF programs won't have symbols resolved.\n");
1273
1274         ret = perf_event__synthesize_cgroups(&top->tool, perf_event__process,
1275                                              &top->session->machines.host);
1276         if (ret < 0)
1277                 pr_debug("Couldn't synthesize cgroup events.\n");
1278
1279         machine__synthesize_threads(&top->session->machines.host, &opts->target,
1280                                     top->evlist->core.threads, false,
1281                                     top->nr_threads_synthesize);
1282
1283         if (top->nr_threads_synthesize > 1)
1284                 perf_set_singlethreaded();
1285
1286         if (perf_hpp_list.socket) {
1287                 ret = perf_env__read_cpu_topology_map(&perf_env);
1288                 if (ret < 0) {
1289                         char errbuf[BUFSIZ];
1290                         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1291
1292                         ui__error("Could not read the CPU topology map: %s\n", err);
1293                         return ret;
1294                 }
1295         }
1296
1297         ret = perf_top__start_counters(top);
1298         if (ret)
1299                 return ret;
1300
1301         top->session->evlist = top->evlist;
1302         perf_session__set_id_hdr_size(top->session);
1303
1304         /*
1305          * When perf is starting the traced process, all the events (apart from
1306          * group members) have enable_on_exec=1 set, so don't spoil it by
1307          * prematurely enabling them.
1308          *
1309          * XXX 'top' still doesn't start workloads like record, trace, but should,
1310          * so leave the check here.
1311          */
1312         if (!target__none(&opts->target))
1313                 evlist__enable(top->evlist);
1314
1315         ret = -1;
1316         if (pthread_create(&thread_process, NULL, process_thread, top)) {
1317                 ui__error("Could not create process thread.\n");
1318                 return ret;
1319         }
1320
1321         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1322                                                             display_thread), top)) {
1323                 ui__error("Could not create display thread.\n");
1324                 goto out_join_thread;
1325         }
1326
1327         if (top->realtime_prio) {
1328                 struct sched_param param;
1329
1330                 param.sched_priority = top->realtime_prio;
1331                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1332                         ui__error("Could not set realtime priority.\n");
1333                         goto out_join;
1334                 }
1335         }
1336
1337         /* Wait for a minimal set of events before starting the snapshot */
1338         evlist__poll(top->evlist, 100);
1339
1340         perf_top__mmap_read(top);
1341
1342         while (!done) {
1343                 u64 hits = top->samples;
1344
1345                 perf_top__mmap_read(top);
1346
1347                 if (opts->overwrite || (hits == top->samples))
1348                         ret = evlist__poll(top->evlist, 100);
1349
1350                 if (resize) {
1351                         perf_top__resize(top);
1352                         resize = 0;
1353                 }
1354         }
1355
1356         ret = 0;
1357 out_join:
1358         pthread_join(thread, NULL);
1359 out_join_thread:
1360         pthread_cond_signal(&top->qe.cond);
1361         pthread_join(thread_process, NULL);
1362         return ret;
1363 }
1364
1365 static int
1366 callchain_opt(const struct option *opt, const char *arg, int unset)
1367 {
1368         symbol_conf.use_callchain = true;
1369         return record_callchain_opt(opt, arg, unset);
1370 }
1371
1372 static int
1373 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1374 {
1375         struct callchain_param *callchain = opt->value;
1376
1377         callchain->enabled = !unset;
1378         callchain->record_mode = CALLCHAIN_FP;
1379
1380         /*
1381          * --no-call-graph
1382          */
1383         if (unset) {
1384                 symbol_conf.use_callchain = false;
1385                 callchain->record_mode = CALLCHAIN_NONE;
1386                 return 0;
1387         }
1388
1389         return parse_callchain_top_opt(arg);
1390 }
1391
1392 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1393 {
1394         if (!strcmp(var, "top.call-graph")) {
1395                 var = "call-graph.record-mode";
1396                 return perf_default_config(var, value, cb);
1397         }
1398         if (!strcmp(var, "top.children")) {
1399                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1400                 return 0;
1401         }
1402
1403         return 0;
1404 }
1405
1406 static int
1407 parse_percent_limit(const struct option *opt, const char *arg,
1408                     int unset __maybe_unused)
1409 {
1410         struct perf_top *top = opt->value;
1411
1412         top->min_percent = strtof(arg, NULL);
1413         return 0;
1414 }
1415
1416 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1417         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1418
1419 int cmd_top(int argc, const char **argv)
1420 {
1421         char errbuf[BUFSIZ];
1422         struct perf_top top = {
1423                 .count_filter        = 5,
1424                 .delay_secs          = 2,
1425                 .record_opts = {
1426                         .mmap_pages     = UINT_MAX,
1427                         .user_freq      = UINT_MAX,
1428                         .user_interval  = ULLONG_MAX,
1429                         .freq           = 4000, /* 4 KHz */
1430                         .target         = {
1431                                 .uses_mmap   = true,
1432                         },
1433                         /*
1434                          * FIXME: This will lose PERF_RECORD_MMAP and other metadata
1435                          * when we pause, fix that and reenable. Probably using a
1436                          * separate evlist with a dummy event, i.e. a non-overwrite
1437                          * ring buffer just for metadata events, while PERF_RECORD_SAMPLE
1438                          * stays in overwrite mode. -acme
1439                          * */
1440                         .overwrite      = 0,
1441                         .sample_time    = true,
1442                         .sample_time_set = true,
1443                 },
1444                 .max_stack           = sysctl__max_stack(),
1445                 .annotation_opts     = annotation__default_options,
1446                 .nr_threads_synthesize = UINT_MAX,
1447         };
1448         struct record_opts *opts = &top.record_opts;
1449         struct target *target = &opts->target;
1450         const struct option options[] = {
1451         OPT_CALLBACK('e', "event", &top.evlist, "event",
1452                      "event selector. use 'perf list' to list available events",
1453                      parse_events_option),
1454         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1455         OPT_STRING('p', "pid", &target->pid, "pid",
1456                     "profile events on existing process id"),
1457         OPT_STRING('t', "tid", &target->tid, "tid",
1458                     "profile events on existing thread id"),
1459         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1460                             "system-wide collection from all CPUs"),
1461         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1462                     "list of cpus to monitor"),
1463         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1464                    "file", "vmlinux pathname"),
1465         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1466                     "don't load vmlinux even if found"),
1467         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1468                    "file", "kallsyms pathname"),
1469         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1470                     "hide kernel symbols"),
1471         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1472                      "number of mmap data pages",
1473                      perf_evlist__parse_mmap_pages),
1474         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1475                     "collect data with this RT SCHED_FIFO priority"),
1476         OPT_INTEGER('d', "delay", &top.delay_secs,
1477                     "number of seconds to delay between refreshes"),
1478         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1479                             "dump the symbol table used for profiling"),
1480         OPT_INTEGER('f', "count-filter", &top.count_filter,
1481                     "only display functions with more events than this"),
1482         OPT_BOOLEAN(0, "group", &opts->group,
1483                             "put the counters into a counter group"),
1484         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1485                     "child tasks do not inherit counters"),
1486         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1487                     "symbol to annotate"),
1488         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1489         OPT_CALLBACK('F', "freq", &top.record_opts, "freq or 'max'",
1490                      "profile at this frequency",
1491                       record__parse_freq),
1492         OPT_INTEGER('E', "entries", &top.print_entries,
1493                     "display this many functions"),
1494         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1495                     "hide user symbols"),
1496         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1497         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1498         OPT_INCR('v', "verbose", &verbose,
1499                     "be more verbose (show counter open errors, etc)"),
1500         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1501                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1502                    " Please refer the man page for the complete list."),
1503         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1504                    "output field(s): overhead, period, sample plus all of sort keys"),
1505         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1506                     "Show a column with the number of samples"),
1507         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1508                            NULL, "enables call-graph recording and display",
1509                            &callchain_opt),
1510         OPT_CALLBACK(0, "call-graph", &callchain_param,
1511                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1512                      top_callchain_help, &parse_callchain_opt),
1513         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1514                     "Accumulate callchains of children and show total overhead as well"),
1515         OPT_INTEGER(0, "max-stack", &top.max_stack,
1516                     "Set the maximum stack depth when parsing the callchain. "
1517                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1518         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1519                    "ignore callees of these functions in call graphs",
1520                    report_parse_ignore_callees_opt),
1521         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1522                     "Show a column with the sum of periods"),
1523         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1524                    "only consider symbols in these dsos"),
1525         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1526                    "only consider symbols in these comms"),
1527         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1528                    "only consider these symbols"),
1529         OPT_BOOLEAN(0, "source", &top.annotation_opts.annotate_src,
1530                     "Interleave source code with assembly code (default)"),
1531         OPT_BOOLEAN(0, "asm-raw", &top.annotation_opts.show_asm_raw,
1532                     "Display raw encoding of assembly instructions (default)"),
1533         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1534                     "Enable kernel symbol demangling"),
1535         OPT_BOOLEAN(0, "no-bpf-event", &top.record_opts.no_bpf_event, "do not record bpf events"),
1536         OPT_STRING(0, "objdump", &top.annotation_opts.objdump_path, "path",
1537                     "objdump binary to use for disassembly and annotations"),
1538         OPT_STRING('M', "disassembler-style", &top.annotation_opts.disassembler_style, "disassembler style",
1539                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1540         OPT_STRING(0, "prefix", &top.annotation_opts.prefix, "prefix",
1541                     "Add prefix to source file path names in programs (with --prefix-strip)"),
1542         OPT_STRING(0, "prefix-strip", &top.annotation_opts.prefix_strip, "N",
1543                     "Strip first N entries of source file path name in programs (with --prefix)"),
1544         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1545         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1546                      "Don't show entries under that percent", parse_percent_limit),
1547         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1548                      "How to display percentage of filtered entries", parse_filter_percentage),
1549         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1550                    "width[,width...]",
1551                    "don't try to adjust column width, use these fixed values"),
1552         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
1553                         "per thread proc mmap processing timeout in ms"),
1554         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1555                      "branch any", "sample any taken branches",
1556                      parse_branch_stack),
1557         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1558                      "branch filter mask", "branch stack filter modes",
1559                      parse_branch_stack),
1560         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1561                     "Show raw trace event output (do not use print fmt or plugins)"),
1562         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1563                     "Show entries in a hierarchy"),
1564         OPT_BOOLEAN(0, "overwrite", &top.record_opts.overwrite,
1565                     "Use a backward ring buffer, default: no"),
1566         OPT_BOOLEAN(0, "force", &symbol_conf.force, "don't complain, do it"),
1567         OPT_UINTEGER(0, "num-thread-synthesize", &top.nr_threads_synthesize,
1568                         "number of thread to run event synthesize"),
1569         OPT_BOOLEAN(0, "namespaces", &opts->record_namespaces,
1570                     "Record namespaces events"),
1571         OPT_BOOLEAN(0, "all-cgroups", &opts->record_cgroup,
1572                     "Record cgroup events"),
1573         OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1574                     "Sort the output by the event at the index n in group. "
1575                     "If n is invalid, sort by the first event. "
1576                     "WARNING: should be used on grouped events."),
1577         OPT_BOOLEAN(0, "stitch-lbr", &top.stitch_lbr,
1578                     "Enable LBR callgraph stitching approach"),
1579 #ifdef HAVE_LIBPFM
1580         OPT_CALLBACK(0, "pfm-events", &top.evlist, "event",
1581                 "libpfm4 event selector. use 'perf list' to list available events",
1582                 parse_libpfm_events_option),
1583 #endif
1584         OPTS_EVSWITCH(&top.evswitch),
1585         OPT_END()
1586         };
1587         const char * const top_usage[] = {
1588                 "perf top [<options>]",
1589                 NULL
1590         };
1591         int status = hists__init();
1592
1593         if (status < 0)
1594                 return status;
1595
1596         top.annotation_opts.min_pcnt = 5;
1597         top.annotation_opts.context  = 4;
1598
1599         top.evlist = evlist__new();
1600         if (top.evlist == NULL)
1601                 return -ENOMEM;
1602
1603         status = perf_config(perf_top_config, &top);
1604         if (status)
1605                 return status;
1606         /*
1607          * Since the per arch annotation init routine may need the cpuid, read
1608          * it here, since we are not getting this from the perf.data header.
1609          */
1610         status = perf_env__read_cpuid(&perf_env);
1611         if (status) {
1612                 /*
1613                  * Some arches do not provide a get_cpuid(), so just use pr_debug, otherwise
1614                  * warn the user explicitely.
1615                  */
1616                 eprintf(status == ENOSYS ? 1 : 0, verbose,
1617                         "Couldn't read the cpuid for this machine: %s\n",
1618                         str_error_r(errno, errbuf, sizeof(errbuf)));
1619         }
1620         top.evlist->env = &perf_env;
1621
1622         argc = parse_options(argc, argv, options, top_usage, 0);
1623         if (argc)
1624                 usage_with_options(top_usage, options);
1625
1626         if (annotate_check_args(&top.annotation_opts) < 0)
1627                 goto out_delete_evlist;
1628
1629         if (!top.evlist->core.nr_entries &&
1630             evlist__add_default(top.evlist) < 0) {
1631                 pr_err("Not enough memory for event selector list\n");
1632                 goto out_delete_evlist;
1633         }
1634
1635         status = evswitch__init(&top.evswitch, top.evlist, stderr);
1636         if (status)
1637                 goto out_delete_evlist;
1638
1639         if (symbol_conf.report_hierarchy) {
1640                 /* disable incompatible options */
1641                 symbol_conf.event_group = false;
1642                 symbol_conf.cumulate_callchain = false;
1643
1644                 if (field_order) {
1645                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1646                         parse_options_usage(top_usage, options, "fields", 0);
1647                         parse_options_usage(NULL, options, "hierarchy", 0);
1648                         goto out_delete_evlist;
1649                 }
1650         }
1651
1652         if (top.stitch_lbr && !(callchain_param.record_mode == CALLCHAIN_LBR)) {
1653                 pr_err("Error: --stitch-lbr must be used with --call-graph lbr\n");
1654                 goto out_delete_evlist;
1655         }
1656
1657         if (opts->branch_stack && callchain_param.enabled)
1658                 symbol_conf.show_branchflag_count = true;
1659
1660         sort__mode = SORT_MODE__TOP;
1661         /* display thread wants entries to be collapsed in a different tree */
1662         perf_hpp_list.need_collapse = 1;
1663
1664         if (top.use_stdio)
1665                 use_browser = 0;
1666         else if (top.use_tui)
1667                 use_browser = 1;
1668
1669         setup_browser(false);
1670
1671         if (setup_sorting(top.evlist) < 0) {
1672                 if (sort_order)
1673                         parse_options_usage(top_usage, options, "s", 1);
1674                 if (field_order)
1675                         parse_options_usage(sort_order ? NULL : top_usage,
1676                                             options, "fields", 0);
1677                 goto out_delete_evlist;
1678         }
1679
1680         status = target__validate(target);
1681         if (status) {
1682                 target__strerror(target, status, errbuf, BUFSIZ);
1683                 ui__warning("%s\n", errbuf);
1684         }
1685
1686         status = target__parse_uid(target);
1687         if (status) {
1688                 int saved_errno = errno;
1689
1690                 target__strerror(target, status, errbuf, BUFSIZ);
1691                 ui__error("%s\n", errbuf);
1692
1693                 status = -saved_errno;
1694                 goto out_delete_evlist;
1695         }
1696
1697         if (target__none(target))
1698                 target->system_wide = true;
1699
1700         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1701                 ui__error("Couldn't create thread/CPU maps: %s\n",
1702                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1703                 goto out_delete_evlist;
1704         }
1705
1706         if (top.delay_secs < 1)
1707                 top.delay_secs = 1;
1708
1709         if (record_opts__config(opts)) {
1710                 status = -EINVAL;
1711                 goto out_delete_evlist;
1712         }
1713
1714         top.sym_evsel = evlist__first(top.evlist);
1715
1716         if (!callchain_param.enabled) {
1717                 symbol_conf.cumulate_callchain = false;
1718                 perf_hpp__cancel_cumulate();
1719         }
1720
1721         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1722                 callchain_param.order = ORDER_CALLER;
1723
1724         status = symbol__annotation_init();
1725         if (status < 0)
1726                 goto out_delete_evlist;
1727
1728         annotation_config__init(&top.annotation_opts);
1729
1730         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1731         status = symbol__init(NULL);
1732         if (status < 0)
1733                 goto out_delete_evlist;
1734
1735         sort__setup_elide(stdout);
1736
1737         get_term_dimensions(&top.winsize);
1738         if (top.print_entries == 0) {
1739                 perf_top__update_print_entries(&top);
1740                 signal(SIGWINCH, winch_sig);
1741         }
1742
1743         top.session = perf_session__new(NULL, false, NULL);
1744         if (IS_ERR(top.session)) {
1745                 status = PTR_ERR(top.session);
1746                 top.session = NULL;
1747                 goto out_delete_evlist;
1748         }
1749
1750 #ifdef HAVE_LIBBPF_SUPPORT
1751         if (!top.record_opts.no_bpf_event) {
1752                 top.sb_evlist = evlist__new();
1753
1754                 if (top.sb_evlist == NULL) {
1755                         pr_err("Couldn't create side band evlist.\n.");
1756                         goto out_delete_evlist;
1757                 }
1758
1759                 if (evlist__add_bpf_sb_event(top.sb_evlist, &perf_env)) {
1760                         pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1761                         goto out_delete_evlist;
1762                 }
1763         }
1764 #endif
1765
1766         if (perf_evlist__start_sb_thread(top.sb_evlist, target)) {
1767                 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1768                 opts->no_bpf_event = true;
1769         }
1770
1771         status = __cmd_top(&top);
1772
1773         if (!opts->no_bpf_event)
1774                 perf_evlist__stop_sb_thread(top.sb_evlist);
1775
1776 out_delete_evlist:
1777         evlist__delete(top.evlist);
1778         perf_session__delete(top.session);
1779
1780         return status;
1781 }