Linux-libre 4.9.135-gnu
[librecmc/linux-libre.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/config.h"
26 #include "util/color.h"
27 #include "util/drv_configs.h"
28 #include "util/evlist.h"
29 #include "util/evsel.h"
30 #include "util/machine.h"
31 #include "util/session.h"
32 #include "util/symbol.h"
33 #include "util/thread.h"
34 #include "util/thread_map.h"
35 #include "util/top.h"
36 #include "util/util.h"
37 #include <linux/rbtree.h>
38 #include <subcmd/parse-options.h>
39 #include "util/parse-events.h"
40 #include "util/cpumap.h"
41 #include "util/xyarray.h"
42 #include "util/sort.h"
43 #include "util/intlist.h"
44 #include "util/parse-branch-options.h"
45 #include "arch/common.h"
46
47 #include "util/debug.h"
48
49 #include <assert.h>
50 #include <elf.h>
51 #include <fcntl.h>
52
53 #include <stdio.h>
54 #include <termios.h>
55 #include <unistd.h>
56 #include <inttypes.h>
57
58 #include <errno.h>
59 #include <time.h>
60 #include <sched.h>
61
62 #include <sys/syscall.h>
63 #include <sys/ioctl.h>
64 #include <poll.h>
65 #include <sys/prctl.h>
66 #include <sys/wait.h>
67 #include <sys/uio.h>
68 #include <sys/utsname.h>
69 #include <sys/mman.h>
70
71 #include <linux/stringify.h>
72 #include <linux/time64.h>
73 #include <linux/types.h>
74
75 static volatile int done;
76 static volatile int resize;
77
78 #define HEADER_LINE_NR  5
79
80 static void perf_top__update_print_entries(struct perf_top *top)
81 {
82         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
83 }
84
85 static void perf_top__sig_winch(int sig __maybe_unused,
86                                 siginfo_t *info __maybe_unused, void *arg __maybe_unused)
87 {
88         resize = 1;
89 }
90
91 static void perf_top__resize(struct perf_top *top)
92 {
93         get_term_dimensions(&top->winsize);
94         perf_top__update_print_entries(top);
95 }
96
97 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
98 {
99         struct symbol *sym;
100         struct annotation *notes;
101         struct map *map;
102         int err = -1;
103
104         if (!he || !he->ms.sym)
105                 return -1;
106
107         sym = he->ms.sym;
108         map = he->ms.map;
109
110         /*
111          * We can't annotate with just /proc/kallsyms
112          */
113         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
114             !dso__is_kcore(map->dso)) {
115                 pr_err("Can't annotate %s: No vmlinux file was found in the "
116                        "path\n", sym->name);
117                 sleep(1);
118                 return -1;
119         }
120
121         notes = symbol__annotation(sym);
122         if (notes->src != NULL) {
123                 pthread_mutex_lock(&notes->lock);
124                 goto out_assign;
125         }
126
127         pthread_mutex_lock(&notes->lock);
128
129         if (symbol__alloc_hist(sym) < 0) {
130                 pthread_mutex_unlock(&notes->lock);
131                 pr_err("Not enough memory for annotating '%s' symbol!\n",
132                        sym->name);
133                 sleep(1);
134                 return err;
135         }
136
137         err = symbol__disassemble(sym, map, 0);
138         if (err == 0) {
139 out_assign:
140                 top->sym_filter_entry = he;
141         } else {
142                 char msg[BUFSIZ];
143                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
144                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
145         }
146
147         pthread_mutex_unlock(&notes->lock);
148         return err;
149 }
150
151 static void __zero_source_counters(struct hist_entry *he)
152 {
153         struct symbol *sym = he->ms.sym;
154         symbol__annotate_zero_histograms(sym);
155 }
156
157 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
158 {
159         struct utsname uts;
160         int err = uname(&uts);
161
162         ui__warning("Out of bounds address found:\n\n"
163                     "Addr:   %" PRIx64 "\n"
164                     "DSO:    %s %c\n"
165                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
166                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
167                     "Arch:   %s\n"
168                     "Kernel: %s\n"
169                     "Tools:  %s\n\n"
170                     "Not all samples will be on the annotation output.\n\n"
171                     "Please report to linux-kernel@vger.kernel.org\n",
172                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
173                     map->start, map->end, sym->start, sym->end,
174                     sym->binding == STB_GLOBAL ? 'g' :
175                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
176                     err ? "[unknown]" : uts.machine,
177                     err ? "[unknown]" : uts.release, perf_version_string);
178         if (use_browser <= 0)
179                 sleep(5);
180
181         map->erange_warned = true;
182 }
183
184 static void perf_top__record_precise_ip(struct perf_top *top,
185                                         struct hist_entry *he,
186                                         int counter, u64 ip)
187 {
188         struct annotation *notes;
189         struct symbol *sym = he->ms.sym;
190         int err = 0;
191
192         if (sym == NULL || (use_browser == 0 &&
193                             (top->sym_filter_entry == NULL ||
194                              top->sym_filter_entry->ms.sym != sym)))
195                 return;
196
197         notes = symbol__annotation(sym);
198
199         if (pthread_mutex_trylock(&notes->lock))
200                 return;
201
202         err = hist_entry__inc_addr_samples(he, counter, ip);
203
204         pthread_mutex_unlock(&notes->lock);
205
206         if (unlikely(err)) {
207                 /*
208                  * This function is now called with he->hists->lock held.
209                  * Release it before going to sleep.
210                  */
211                 pthread_mutex_unlock(&he->hists->lock);
212
213                 if (err == -ERANGE && !he->ms.map->erange_warned)
214                         ui__warn_map_erange(he->ms.map, sym, ip);
215                 else if (err == -ENOMEM) {
216                         pr_err("Not enough memory for annotating '%s' symbol!\n",
217                                sym->name);
218                         sleep(1);
219                 }
220
221                 pthread_mutex_lock(&he->hists->lock);
222         }
223 }
224
225 static void perf_top__show_details(struct perf_top *top)
226 {
227         struct hist_entry *he = top->sym_filter_entry;
228         struct annotation *notes;
229         struct symbol *symbol;
230         int more;
231
232         if (!he)
233                 return;
234
235         symbol = he->ms.sym;
236         notes = symbol__annotation(symbol);
237
238         pthread_mutex_lock(&notes->lock);
239
240         if (notes->src == NULL)
241                 goto out_unlock;
242
243         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
244         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
245
246         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
247                                        0, top->sym_pcnt_filter, top->print_entries, 4);
248
249         if (top->evlist->enabled) {
250                 if (top->zero)
251                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
252                 else
253                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
254         }
255         if (more != 0)
256                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
257 out_unlock:
258         pthread_mutex_unlock(&notes->lock);
259 }
260
261 static void perf_top__print_sym_table(struct perf_top *top)
262 {
263         char bf[160];
264         int printed = 0;
265         const int win_width = top->winsize.ws_col - 1;
266         struct perf_evsel *evsel = top->sym_evsel;
267         struct hists *hists = evsel__hists(evsel);
268
269         puts(CONSOLE_CLEAR);
270
271         perf_top__header_snprintf(top, bf, sizeof(bf));
272         printf("%s\n", bf);
273
274         perf_top__reset_sample_counters(top);
275
276         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
277
278         if (hists->stats.nr_lost_warned !=
279             hists->stats.nr_events[PERF_RECORD_LOST]) {
280                 hists->stats.nr_lost_warned =
281                               hists->stats.nr_events[PERF_RECORD_LOST];
282                 color_fprintf(stdout, PERF_COLOR_RED,
283                               "WARNING: LOST %d chunks, Check IO/CPU overload",
284                               hists->stats.nr_lost_warned);
285                 ++printed;
286         }
287
288         if (top->sym_filter_entry) {
289                 perf_top__show_details(top);
290                 return;
291         }
292
293         if (top->evlist->enabled) {
294                 if (top->zero) {
295                         hists__delete_entries(hists);
296                 } else {
297                         hists__decay_entries(hists, top->hide_user_symbols,
298                                              top->hide_kernel_symbols);
299                 }
300         }
301
302         hists__collapse_resort(hists, NULL);
303         perf_evsel__output_resort(evsel, NULL);
304
305         hists__output_recalc_col_len(hists, top->print_entries - printed);
306         putchar('\n');
307         hists__fprintf(hists, false, top->print_entries - printed, win_width,
308                        top->min_percent, stdout, symbol_conf.use_callchain);
309 }
310
311 static void prompt_integer(int *target, const char *msg)
312 {
313         char *buf = malloc(0), *p;
314         size_t dummy = 0;
315         int tmp;
316
317         fprintf(stdout, "\n%s: ", msg);
318         if (getline(&buf, &dummy, stdin) < 0)
319                 return;
320
321         p = strchr(buf, '\n');
322         if (p)
323                 *p = 0;
324
325         p = buf;
326         while(*p) {
327                 if (!isdigit(*p))
328                         goto out_free;
329                 p++;
330         }
331         tmp = strtoul(buf, NULL, 10);
332         *target = tmp;
333 out_free:
334         free(buf);
335 }
336
337 static void prompt_percent(int *target, const char *msg)
338 {
339         int tmp = 0;
340
341         prompt_integer(&tmp, msg);
342         if (tmp >= 0 && tmp <= 100)
343                 *target = tmp;
344 }
345
346 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
347 {
348         char *buf = malloc(0), *p;
349         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
350         struct hists *hists = evsel__hists(top->sym_evsel);
351         struct rb_node *next;
352         size_t dummy = 0;
353
354         /* zero counters of active symbol */
355         if (syme) {
356                 __zero_source_counters(syme);
357                 top->sym_filter_entry = NULL;
358         }
359
360         fprintf(stdout, "\n%s: ", msg);
361         if (getline(&buf, &dummy, stdin) < 0)
362                 goto out_free;
363
364         p = strchr(buf, '\n');
365         if (p)
366                 *p = 0;
367
368         next = rb_first(&hists->entries);
369         while (next) {
370                 n = rb_entry(next, struct hist_entry, rb_node);
371                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
372                         found = n;
373                         break;
374                 }
375                 next = rb_next(&n->rb_node);
376         }
377
378         if (!found) {
379                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
380                 sleep(1);
381         } else
382                 perf_top__parse_source(top, found);
383
384 out_free:
385         free(buf);
386 }
387
388 static void perf_top__print_mapped_keys(struct perf_top *top)
389 {
390         char *name = NULL;
391
392         if (top->sym_filter_entry) {
393                 struct symbol *sym = top->sym_filter_entry->ms.sym;
394                 name = sym->name;
395         }
396
397         fprintf(stdout, "\nMapped keys:\n");
398         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
399         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
400
401         if (top->evlist->nr_entries > 1)
402                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
403
404         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
405
406         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
407         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
408         fprintf(stdout, "\t[S]     stop annotation.\n");
409
410         fprintf(stdout,
411                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
412                 top->hide_kernel_symbols ? "yes" : "no");
413         fprintf(stdout,
414                 "\t[U]     hide user symbols.               \t(%s)\n",
415                 top->hide_user_symbols ? "yes" : "no");
416         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
417         fprintf(stdout, "\t[qQ]    quit.\n");
418 }
419
420 static int perf_top__key_mapped(struct perf_top *top, int c)
421 {
422         switch (c) {
423                 case 'd':
424                 case 'e':
425                 case 'f':
426                 case 'z':
427                 case 'q':
428                 case 'Q':
429                 case 'K':
430                 case 'U':
431                 case 'F':
432                 case 's':
433                 case 'S':
434                         return 1;
435                 case 'E':
436                         return top->evlist->nr_entries > 1 ? 1 : 0;
437                 default:
438                         break;
439         }
440
441         return 0;
442 }
443
444 static bool perf_top__handle_keypress(struct perf_top *top, int c)
445 {
446         bool ret = true;
447
448         if (!perf_top__key_mapped(top, c)) {
449                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
450                 struct termios save;
451
452                 perf_top__print_mapped_keys(top);
453                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
454                 fflush(stdout);
455
456                 set_term_quiet_input(&save);
457
458                 poll(&stdin_poll, 1, -1);
459                 c = getc(stdin);
460
461                 tcsetattr(0, TCSAFLUSH, &save);
462                 if (!perf_top__key_mapped(top, c))
463                         return ret;
464         }
465
466         switch (c) {
467                 case 'd':
468                         prompt_integer(&top->delay_secs, "Enter display delay");
469                         if (top->delay_secs < 1)
470                                 top->delay_secs = 1;
471                         break;
472                 case 'e':
473                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
474                         if (top->print_entries == 0) {
475                                 struct sigaction act = {
476                                         .sa_sigaction = perf_top__sig_winch,
477                                         .sa_flags     = SA_SIGINFO,
478                                 };
479                                 perf_top__resize(top);
480                                 sigaction(SIGWINCH, &act, NULL);
481                         } else {
482                                 signal(SIGWINCH, SIG_DFL);
483                         }
484                         break;
485                 case 'E':
486                         if (top->evlist->nr_entries > 1) {
487                                 /* Select 0 as the default event: */
488                                 int counter = 0;
489
490                                 fprintf(stderr, "\nAvailable events:");
491
492                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
493                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
494
495                                 prompt_integer(&counter, "Enter details event counter");
496
497                                 if (counter >= top->evlist->nr_entries) {
498                                         top->sym_evsel = perf_evlist__first(top->evlist);
499                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
500                                         sleep(1);
501                                         break;
502                                 }
503                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
504                                         if (top->sym_evsel->idx == counter)
505                                                 break;
506                         } else
507                                 top->sym_evsel = perf_evlist__first(top->evlist);
508                         break;
509                 case 'f':
510                         prompt_integer(&top->count_filter, "Enter display event count filter");
511                         break;
512                 case 'F':
513                         prompt_percent(&top->sym_pcnt_filter,
514                                        "Enter details display event filter (percent)");
515                         break;
516                 case 'K':
517                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
518                         break;
519                 case 'q':
520                 case 'Q':
521                         printf("exiting.\n");
522                         if (top->dump_symtab)
523                                 perf_session__fprintf_dsos(top->session, stderr);
524                         ret = false;
525                         break;
526                 case 's':
527                         perf_top__prompt_symbol(top, "Enter details symbol");
528                         break;
529                 case 'S':
530                         if (!top->sym_filter_entry)
531                                 break;
532                         else {
533                                 struct hist_entry *syme = top->sym_filter_entry;
534
535                                 top->sym_filter_entry = NULL;
536                                 __zero_source_counters(syme);
537                         }
538                         break;
539                 case 'U':
540                         top->hide_user_symbols = !top->hide_user_symbols;
541                         break;
542                 case 'z':
543                         top->zero = !top->zero;
544                         break;
545                 default:
546                         break;
547         }
548
549         return ret;
550 }
551
552 static void perf_top__sort_new_samples(void *arg)
553 {
554         struct perf_top *t = arg;
555         struct perf_evsel *evsel = t->sym_evsel;
556         struct hists *hists;
557
558         perf_top__reset_sample_counters(t);
559
560         if (t->evlist->selected != NULL)
561                 t->sym_evsel = t->evlist->selected;
562
563         hists = evsel__hists(evsel);
564
565         if (t->evlist->enabled) {
566                 if (t->zero) {
567                         hists__delete_entries(hists);
568                 } else {
569                         hists__decay_entries(hists, t->hide_user_symbols,
570                                              t->hide_kernel_symbols);
571                 }
572         }
573
574         hists__collapse_resort(hists, NULL);
575         perf_evsel__output_resort(evsel, NULL);
576 }
577
578 static void *display_thread_tui(void *arg)
579 {
580         struct perf_evsel *pos;
581         struct perf_top *top = arg;
582         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
583         struct hist_browser_timer hbt = {
584                 .timer          = perf_top__sort_new_samples,
585                 .arg            = top,
586                 .refresh        = top->delay_secs,
587         };
588
589         perf_top__sort_new_samples(top);
590
591         /*
592          * Initialize the uid_filter_str, in the future the TUI will allow
593          * Zooming in/out UIDs. For now juse use whatever the user passed
594          * via --uid.
595          */
596         evlist__for_each_entry(top->evlist, pos) {
597                 struct hists *hists = evsel__hists(pos);
598                 hists->uid_filter_str = top->record_opts.target.uid_str;
599         }
600
601         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
602                                       top->min_percent,
603                                       &top->session->header.env);
604
605         done = 1;
606         return NULL;
607 }
608
609 static void display_sig(int sig __maybe_unused)
610 {
611         done = 1;
612 }
613
614 static void display_setup_sig(void)
615 {
616         signal(SIGSEGV, sighandler_dump_stack);
617         signal(SIGFPE, sighandler_dump_stack);
618         signal(SIGINT,  display_sig);
619         signal(SIGQUIT, display_sig);
620         signal(SIGTERM, display_sig);
621 }
622
623 static void *display_thread(void *arg)
624 {
625         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
626         struct termios save;
627         struct perf_top *top = arg;
628         int delay_msecs, c;
629
630         display_setup_sig();
631         pthread__unblock_sigwinch();
632 repeat:
633         delay_msecs = top->delay_secs * MSEC_PER_SEC;
634         set_term_quiet_input(&save);
635         /* trash return*/
636         getc(stdin);
637
638         while (!done) {
639                 perf_top__print_sym_table(top);
640                 /*
641                  * Either timeout expired or we got an EINTR due to SIGWINCH,
642                  * refresh screen in both cases.
643                  */
644                 switch (poll(&stdin_poll, 1, delay_msecs)) {
645                 case 0:
646                         continue;
647                 case -1:
648                         if (errno == EINTR)
649                                 continue;
650                         __fallthrough;
651                 default:
652                         c = getc(stdin);
653                         tcsetattr(0, TCSAFLUSH, &save);
654
655                         if (perf_top__handle_keypress(top, c))
656                                 goto repeat;
657                         done = 1;
658                 }
659         }
660
661         tcsetattr(0, TCSAFLUSH, &save);
662         return NULL;
663 }
664
665 static int hist_iter__top_callback(struct hist_entry_iter *iter,
666                                    struct addr_location *al, bool single,
667                                    void *arg)
668 {
669         struct perf_top *top = arg;
670         struct hist_entry *he = iter->he;
671         struct perf_evsel *evsel = iter->evsel;
672
673         if (perf_hpp_list.sym && single)
674                 perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
675
676         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
677                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
678         return 0;
679 }
680
681 static void perf_event__process_sample(struct perf_tool *tool,
682                                        const union perf_event *event,
683                                        struct perf_evsel *evsel,
684                                        struct perf_sample *sample,
685                                        struct machine *machine)
686 {
687         struct perf_top *top = container_of(tool, struct perf_top, tool);
688         struct addr_location al;
689         int err;
690
691         if (!machine && perf_guest) {
692                 static struct intlist *seen;
693
694                 if (!seen)
695                         seen = intlist__new(NULL);
696
697                 if (!intlist__has_entry(seen, sample->pid)) {
698                         pr_err("Can't find guest [%d]'s kernel information\n",
699                                 sample->pid);
700                         intlist__add(seen, sample->pid);
701                 }
702                 return;
703         }
704
705         if (!machine) {
706                 pr_err("%u unprocessable samples recorded.\r",
707                        top->session->evlist->stats.nr_unprocessable_samples++);
708                 return;
709         }
710
711         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
712                 top->exact_samples++;
713
714         if (machine__resolve(machine, &al, sample) < 0)
715                 return;
716
717         if (!machine->kptr_restrict_warned &&
718             symbol_conf.kptr_restrict &&
719             al.cpumode == PERF_RECORD_MISC_KERNEL) {
720                 ui__warning(
721 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
722 "Check /proc/sys/kernel/kptr_restrict.\n\n"
723 "Kernel%s samples will not be resolved.\n",
724                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
725                           " modules" : "");
726                 if (use_browser <= 0)
727                         sleep(5);
728                 machine->kptr_restrict_warned = true;
729         }
730
731         if (al.sym == NULL) {
732                 const char *msg = "Kernel samples will not be resolved.\n";
733                 /*
734                  * As we do lazy loading of symtabs we only will know if the
735                  * specified vmlinux file is invalid when we actually have a
736                  * hit in kernel space and then try to load it. So if we get
737                  * here and there are _no_ symbols in the DSO backing the
738                  * kernel map, bail out.
739                  *
740                  * We may never get here, for instance, if we use -K/
741                  * --hide-kernel-symbols, even if the user specifies an
742                  * invalid --vmlinux ;-)
743                  */
744                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
745                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
746                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
747                         if (symbol_conf.vmlinux_name) {
748                                 char serr[256];
749                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
750                                 ui__warning("The %s file can't be used: %s\n%s",
751                                             symbol_conf.vmlinux_name, serr, msg);
752                         } else {
753                                 ui__warning("A vmlinux file was not found.\n%s",
754                                             msg);
755                         }
756
757                         if (use_browser <= 0)
758                                 sleep(5);
759                         top->vmlinux_warned = true;
760                 }
761         }
762
763         if (al.sym == NULL || !al.sym->idle) {
764                 struct hists *hists = evsel__hists(evsel);
765                 struct hist_entry_iter iter = {
766                         .evsel          = evsel,
767                         .sample         = sample,
768                         .add_entry_cb   = hist_iter__top_callback,
769                 };
770
771                 if (symbol_conf.cumulate_callchain)
772                         iter.ops = &hist_iter_cumulative;
773                 else
774                         iter.ops = &hist_iter_normal;
775
776                 pthread_mutex_lock(&hists->lock);
777
778                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
779                 if (err < 0)
780                         pr_err("Problem incrementing symbol period, skipping event\n");
781
782                 pthread_mutex_unlock(&hists->lock);
783         }
784
785         addr_location__put(&al);
786 }
787
788 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
789 {
790         struct perf_sample sample;
791         struct perf_evsel *evsel;
792         struct perf_session *session = top->session;
793         union perf_event *event;
794         struct machine *machine;
795         int ret;
796
797         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
798                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
799                 if (ret) {
800                         pr_err("Can't parse sample, err = %d\n", ret);
801                         goto next_event;
802                 }
803
804                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
805                 assert(evsel != NULL);
806
807                 if (event->header.type == PERF_RECORD_SAMPLE)
808                         ++top->samples;
809
810                 switch (sample.cpumode) {
811                 case PERF_RECORD_MISC_USER:
812                         ++top->us_samples;
813                         if (top->hide_user_symbols)
814                                 goto next_event;
815                         machine = &session->machines.host;
816                         break;
817                 case PERF_RECORD_MISC_KERNEL:
818                         ++top->kernel_samples;
819                         if (top->hide_kernel_symbols)
820                                 goto next_event;
821                         machine = &session->machines.host;
822                         break;
823                 case PERF_RECORD_MISC_GUEST_KERNEL:
824                         ++top->guest_kernel_samples;
825                         machine = perf_session__find_machine(session,
826                                                              sample.pid);
827                         break;
828                 case PERF_RECORD_MISC_GUEST_USER:
829                         ++top->guest_us_samples;
830                         /*
831                          * TODO: we don't process guest user from host side
832                          * except simple counting.
833                          */
834                         goto next_event;
835                 default:
836                         if (event->header.type == PERF_RECORD_SAMPLE)
837                                 goto next_event;
838                         machine = &session->machines.host;
839                         break;
840                 }
841
842
843                 if (event->header.type == PERF_RECORD_SAMPLE) {
844                         perf_event__process_sample(&top->tool, event, evsel,
845                                                    &sample, machine);
846                 } else if (event->header.type < PERF_RECORD_MAX) {
847                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
848                         machine__process_event(machine, event, &sample);
849                 } else
850                         ++session->evlist->stats.nr_unknown_events;
851 next_event:
852                 perf_evlist__mmap_consume(top->evlist, idx);
853         }
854 }
855
856 static void perf_top__mmap_read(struct perf_top *top)
857 {
858         int i;
859
860         for (i = 0; i < top->evlist->nr_mmaps; i++)
861                 perf_top__mmap_read_idx(top, i);
862 }
863
864 static int perf_top__start_counters(struct perf_top *top)
865 {
866         char msg[512];
867         struct perf_evsel *counter;
868         struct perf_evlist *evlist = top->evlist;
869         struct record_opts *opts = &top->record_opts;
870
871         perf_evlist__config(evlist, opts, &callchain_param);
872
873         evlist__for_each_entry(evlist, counter) {
874 try_again:
875                 if (perf_evsel__open(counter, top->evlist->cpus,
876                                      top->evlist->threads) < 0) {
877                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
878                                 if (verbose)
879                                         ui__warning("%s\n", msg);
880                                 goto try_again;
881                         }
882
883                         perf_evsel__open_strerror(counter, &opts->target,
884                                                   errno, msg, sizeof(msg));
885                         ui__error("%s\n", msg);
886                         goto out_err;
887                 }
888         }
889
890         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
891                 ui__error("Failed to mmap with %d (%s)\n",
892                             errno, str_error_r(errno, msg, sizeof(msg)));
893                 goto out_err;
894         }
895
896         return 0;
897
898 out_err:
899         return -1;
900 }
901
902 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
903 {
904         if (!perf_hpp_list.sym) {
905                 if (callchain->enabled) {
906                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
907                         return -EINVAL;
908                 }
909         } else if (callchain->mode != CHAIN_NONE) {
910                 if (callchain_register_param(callchain) < 0) {
911                         ui__error("Can't register callchain params.\n");
912                         return -EINVAL;
913                 }
914         }
915
916         return 0;
917 }
918
919 static int __cmd_top(struct perf_top *top)
920 {
921         char msg[512];
922         struct perf_evsel *pos;
923         struct perf_evsel_config_term *err_term;
924         struct perf_evlist *evlist = top->evlist;
925         struct record_opts *opts = &top->record_opts;
926         pthread_t thread;
927         int ret;
928
929         top->session = perf_session__new(NULL, false, NULL);
930         if (top->session == NULL)
931                 return -1;
932
933         if (!objdump_path) {
934                 ret = perf_env__lookup_objdump(&top->session->header.env);
935                 if (ret)
936                         goto out_delete;
937         }
938
939         ret = callchain_param__setup_sample_type(&callchain_param);
940         if (ret)
941                 goto out_delete;
942
943         if (perf_session__register_idle_thread(top->session) < 0)
944                 goto out_delete;
945
946         machine__synthesize_threads(&top->session->machines.host, &opts->target,
947                                     top->evlist->threads, false, opts->proc_map_timeout);
948
949         if (perf_hpp_list.socket) {
950                 ret = perf_env__read_cpu_topology_map(&perf_env);
951                 if (ret < 0)
952                         goto out_err_cpu_topo;
953         }
954
955         ret = perf_top__start_counters(top);
956         if (ret)
957                 goto out_delete;
958
959         ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
960         if (ret) {
961                 error("failed to set config \"%s\" on event %s with %d (%s)\n",
962                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
963                         str_error_r(errno, msg, sizeof(msg)));
964                 goto out_delete;
965         }
966
967         top->session->evlist = top->evlist;
968         perf_session__set_id_hdr_size(top->session);
969
970         /*
971          * When perf is starting the traced process, all the events (apart from
972          * group members) have enable_on_exec=1 set, so don't spoil it by
973          * prematurely enabling them.
974          *
975          * XXX 'top' still doesn't start workloads like record, trace, but should,
976          * so leave the check here.
977          */
978         if (!target__none(&opts->target))
979                 perf_evlist__enable(top->evlist);
980
981         /* Wait for a minimal set of events before starting the snapshot */
982         perf_evlist__poll(top->evlist, 100);
983
984         perf_top__mmap_read(top);
985
986         ret = -1;
987         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
988                                                             display_thread), top)) {
989                 ui__error("Could not create display thread.\n");
990                 goto out_delete;
991         }
992
993         if (top->realtime_prio) {
994                 struct sched_param param;
995
996                 param.sched_priority = top->realtime_prio;
997                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
998                         ui__error("Could not set realtime priority.\n");
999                         goto out_join;
1000                 }
1001         }
1002
1003         while (!done) {
1004                 u64 hits = top->samples;
1005
1006                 perf_top__mmap_read(top);
1007
1008                 if (hits == top->samples)
1009                         ret = perf_evlist__poll(top->evlist, 100);
1010
1011                 if (resize) {
1012                         perf_top__resize(top);
1013                         resize = 0;
1014                 }
1015         }
1016
1017         ret = 0;
1018 out_join:
1019         pthread_join(thread, NULL);
1020 out_delete:
1021         perf_session__delete(top->session);
1022         top->session = NULL;
1023
1024         return ret;
1025
1026 out_err_cpu_topo: {
1027         char errbuf[BUFSIZ];
1028         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1029
1030         ui__error("Could not read the CPU topology map: %s\n", err);
1031         goto out_delete;
1032 }
1033 }
1034
1035 static int
1036 callchain_opt(const struct option *opt, const char *arg, int unset)
1037 {
1038         symbol_conf.use_callchain = true;
1039         return record_callchain_opt(opt, arg, unset);
1040 }
1041
1042 static int
1043 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1044 {
1045         struct callchain_param *callchain = opt->value;
1046
1047         callchain->enabled = !unset;
1048         callchain->record_mode = CALLCHAIN_FP;
1049
1050         /*
1051          * --no-call-graph
1052          */
1053         if (unset) {
1054                 symbol_conf.use_callchain = false;
1055                 callchain->record_mode = CALLCHAIN_NONE;
1056                 return 0;
1057         }
1058
1059         return parse_callchain_top_opt(arg);
1060 }
1061
1062 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1063 {
1064         if (!strcmp(var, "top.call-graph")) {
1065                 var = "call-graph.record-mode";
1066                 return perf_default_config(var, value, cb);
1067         }
1068         if (!strcmp(var, "top.children")) {
1069                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1070                 return 0;
1071         }
1072
1073         return 0;
1074 }
1075
1076 static int
1077 parse_percent_limit(const struct option *opt, const char *arg,
1078                     int unset __maybe_unused)
1079 {
1080         struct perf_top *top = opt->value;
1081
1082         top->min_percent = strtof(arg, NULL);
1083         return 0;
1084 }
1085
1086 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1087         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1088
1089 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1090 {
1091         char errbuf[BUFSIZ];
1092         struct perf_top top = {
1093                 .count_filter        = 5,
1094                 .delay_secs          = 2,
1095                 .record_opts = {
1096                         .mmap_pages     = UINT_MAX,
1097                         .user_freq      = UINT_MAX,
1098                         .user_interval  = ULLONG_MAX,
1099                         .freq           = 4000, /* 4 KHz */
1100                         .target         = {
1101                                 .uses_mmap   = true,
1102                         },
1103                         .proc_map_timeout    = 500,
1104                 },
1105                 .max_stack           = sysctl_perf_event_max_stack,
1106                 .sym_pcnt_filter     = 5,
1107         };
1108         struct record_opts *opts = &top.record_opts;
1109         struct target *target = &opts->target;
1110         const struct option options[] = {
1111         OPT_CALLBACK('e', "event", &top.evlist, "event",
1112                      "event selector. use 'perf list' to list available events",
1113                      parse_events_option),
1114         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1115         OPT_STRING('p', "pid", &target->pid, "pid",
1116                     "profile events on existing process id"),
1117         OPT_STRING('t', "tid", &target->tid, "tid",
1118                     "profile events on existing thread id"),
1119         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1120                             "system-wide collection from all CPUs"),
1121         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1122                     "list of cpus to monitor"),
1123         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1124                    "file", "vmlinux pathname"),
1125         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1126                     "don't load vmlinux even if found"),
1127         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1128                     "hide kernel symbols"),
1129         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1130                      "number of mmap data pages",
1131                      perf_evlist__parse_mmap_pages),
1132         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1133                     "collect data with this RT SCHED_FIFO priority"),
1134         OPT_INTEGER('d', "delay", &top.delay_secs,
1135                     "number of seconds to delay between refreshes"),
1136         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1137                             "dump the symbol table used for profiling"),
1138         OPT_INTEGER('f', "count-filter", &top.count_filter,
1139                     "only display functions with more events than this"),
1140         OPT_BOOLEAN(0, "group", &opts->group,
1141                             "put the counters into a counter group"),
1142         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1143                     "child tasks do not inherit counters"),
1144         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1145                     "symbol to annotate"),
1146         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1147         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1148         OPT_INTEGER('E', "entries", &top.print_entries,
1149                     "display this many functions"),
1150         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1151                     "hide user symbols"),
1152         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1153         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1154         OPT_INCR('v', "verbose", &verbose,
1155                     "be more verbose (show counter open errors, etc)"),
1156         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1157                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1158                    " Please refer the man page for the complete list."),
1159         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1160                    "output field(s): overhead, period, sample plus all of sort keys"),
1161         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1162                     "Show a column with the number of samples"),
1163         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1164                            NULL, "enables call-graph recording and display",
1165                            &callchain_opt),
1166         OPT_CALLBACK(0, "call-graph", &callchain_param,
1167                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1168                      top_callchain_help, &parse_callchain_opt),
1169         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1170                     "Accumulate callchains of children and show total overhead as well"),
1171         OPT_INTEGER(0, "max-stack", &top.max_stack,
1172                     "Set the maximum stack depth when parsing the callchain. "
1173                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1174         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1175                    "ignore callees of these functions in call graphs",
1176                    report_parse_ignore_callees_opt),
1177         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1178                     "Show a column with the sum of periods"),
1179         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1180                    "only consider symbols in these dsos"),
1181         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1182                    "only consider symbols in these comms"),
1183         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1184                    "only consider these symbols"),
1185         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1186                     "Interleave source code with assembly code (default)"),
1187         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1188                     "Display raw encoding of assembly instructions (default)"),
1189         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1190                     "Enable kernel symbol demangling"),
1191         OPT_STRING(0, "objdump", &objdump_path, "path",
1192                     "objdump binary to use for disassembly and annotations"),
1193         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1194                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1195         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1196         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1197                      "Don't show entries under that percent", parse_percent_limit),
1198         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1199                      "How to display percentage of filtered entries", parse_filter_percentage),
1200         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1201                    "width[,width...]",
1202                    "don't try to adjust column width, use these fixed values"),
1203         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1204                         "per thread proc mmap processing timeout in ms"),
1205         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1206                      "branch any", "sample any taken branches",
1207                      parse_branch_stack),
1208         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1209                      "branch filter mask", "branch stack filter modes",
1210                      parse_branch_stack),
1211         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1212                     "Show raw trace event output (do not use print fmt or plugins)"),
1213         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1214                     "Show entries in a hierarchy"),
1215         OPT_END()
1216         };
1217         const char * const top_usage[] = {
1218                 "perf top [<options>]",
1219                 NULL
1220         };
1221         int status = hists__init();
1222
1223         if (status < 0)
1224                 return status;
1225
1226         top.evlist = perf_evlist__new();
1227         if (top.evlist == NULL)
1228                 return -ENOMEM;
1229
1230         perf_config(perf_top_config, &top);
1231
1232         argc = parse_options(argc, argv, options, top_usage, 0);
1233         if (argc)
1234                 usage_with_options(top_usage, options);
1235
1236         if (!top.evlist->nr_entries &&
1237             perf_evlist__add_default(top.evlist) < 0) {
1238                 pr_err("Not enough memory for event selector list\n");
1239                 goto out_delete_evlist;
1240         }
1241
1242         if (symbol_conf.report_hierarchy) {
1243                 /* disable incompatible options */
1244                 symbol_conf.event_group = false;
1245                 symbol_conf.cumulate_callchain = false;
1246
1247                 if (field_order) {
1248                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1249                         parse_options_usage(top_usage, options, "fields", 0);
1250                         parse_options_usage(NULL, options, "hierarchy", 0);
1251                         goto out_delete_evlist;
1252                 }
1253         }
1254
1255         sort__mode = SORT_MODE__TOP;
1256         /* display thread wants entries to be collapsed in a different tree */
1257         perf_hpp_list.need_collapse = 1;
1258
1259         if (top.use_stdio)
1260                 use_browser = 0;
1261         else if (top.use_tui)
1262                 use_browser = 1;
1263
1264         setup_browser(false);
1265
1266         if (setup_sorting(top.evlist) < 0) {
1267                 if (sort_order)
1268                         parse_options_usage(top_usage, options, "s", 1);
1269                 if (field_order)
1270                         parse_options_usage(sort_order ? NULL : top_usage,
1271                                             options, "fields", 0);
1272                 goto out_delete_evlist;
1273         }
1274
1275         status = target__validate(target);
1276         if (status) {
1277                 target__strerror(target, status, errbuf, BUFSIZ);
1278                 ui__warning("%s\n", errbuf);
1279         }
1280
1281         status = target__parse_uid(target);
1282         if (status) {
1283                 int saved_errno = errno;
1284
1285                 target__strerror(target, status, errbuf, BUFSIZ);
1286                 ui__error("%s\n", errbuf);
1287
1288                 status = -saved_errno;
1289                 goto out_delete_evlist;
1290         }
1291
1292         if (target__none(target))
1293                 target->system_wide = true;
1294
1295         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1296                 ui__error("Couldn't create thread/CPU maps: %s\n",
1297                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1298                 goto out_delete_evlist;
1299         }
1300
1301         symbol_conf.nr_events = top.evlist->nr_entries;
1302
1303         if (top.delay_secs < 1)
1304                 top.delay_secs = 1;
1305
1306         if (record_opts__config(opts)) {
1307                 status = -EINVAL;
1308                 goto out_delete_evlist;
1309         }
1310
1311         top.sym_evsel = perf_evlist__first(top.evlist);
1312
1313         if (!callchain_param.enabled) {
1314                 symbol_conf.cumulate_callchain = false;
1315                 perf_hpp__cancel_cumulate();
1316         }
1317
1318         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1319                 callchain_param.order = ORDER_CALLER;
1320
1321         status = symbol__annotation_init();
1322         if (status < 0)
1323                 goto out_delete_evlist;
1324
1325         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1326         if (symbol__init(NULL) < 0)
1327                 return -1;
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 }