]> git.karo-electronics.de Git - karo-tx-linux.git/blob - tools/perf/builtin-top.c
47984a838b7335dae05cf75f52ee54e4e624f68e
[karo-tx-linux.git] / tools / perf / builtin-top.c
1 /*
2  * builtin-top.c
3  *
4  * Builtin top command: Display a continuously updated profile of
5  * any workload, CPU or specific PID.
6  *
7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
9  *
10  * Improvements and fixes by:
11  *
12  *   Arjan van de Ven <arjan@linux.intel.com>
13  *   Yanmin Zhang <yanmin.zhang@intel.com>
14  *   Wu Fengguang <fengguang.wu@intel.com>
15  *   Mike Galbraith <efault@gmx.de>
16  *   Paul Mackerras <paulus@samba.org>
17  *
18  * Released under the GPL v2. (and only v2, not any later version)
19  */
20 #include "builtin.h"
21
22 #include "perf.h"
23
24 #include "util/annotate.h"
25 #include "util/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/term.h"
44 #include "util/intlist.h"
45 #include "util/parse-branch-options.h"
46 #include "arch/common.h"
47
48 #include "util/debug.h"
49
50 #include <assert.h>
51 #include <elf.h>
52 #include <fcntl.h>
53
54 #include <stdio.h>
55 #include <termios.h>
56 #include <unistd.h>
57 #include <inttypes.h>
58
59 #include <errno.h>
60 #include <time.h>
61 #include <sched.h>
62
63 #include <sys/syscall.h>
64 #include <sys/ioctl.h>
65 #include <poll.h>
66 #include <sys/prctl.h>
67 #include <sys/wait.h>
68 #include <sys/uio.h>
69 #include <sys/utsname.h>
70 #include <sys/mman.h>
71
72 #include <linux/stringify.h>
73 #include <linux/time64.h>
74 #include <linux/types.h>
75
76 #include "sane_ctype.h"
77
78 static volatile int done;
79
80 #define HEADER_LINE_NR  5
81
82 static void perf_top__update_print_entries(struct perf_top *top)
83 {
84         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
85 }
86
87 static void perf_top__sig_winch(int sig __maybe_unused,
88                                 siginfo_t *info __maybe_unused, void *arg)
89 {
90         struct perf_top *top = arg;
91
92         get_term_dimensions(&top->winsize);
93         perf_top__update_print_entries(top);
94 }
95
96 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
97 {
98         struct symbol *sym;
99         struct annotation *notes;
100         struct map *map;
101         int err = -1;
102
103         if (!he || !he->ms.sym)
104                 return -1;
105
106         sym = he->ms.sym;
107         map = he->ms.map;
108
109         /*
110          * We can't annotate with just /proc/kallsyms
111          */
112         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
113             !dso__is_kcore(map->dso)) {
114                 pr_err("Can't annotate %s: No vmlinux file was found in the "
115                        "path\n", sym->name);
116                 sleep(1);
117                 return -1;
118         }
119
120         notes = symbol__annotation(sym);
121         if (notes->src != NULL) {
122                 pthread_mutex_lock(&notes->lock);
123                 goto out_assign;
124         }
125
126         pthread_mutex_lock(&notes->lock);
127
128         if (symbol__alloc_hist(sym) < 0) {
129                 pthread_mutex_unlock(&notes->lock);
130                 pr_err("Not enough memory for annotating '%s' symbol!\n",
131                        sym->name);
132                 sleep(1);
133                 return err;
134         }
135
136         err = symbol__disassemble(sym, map, NULL, 0);
137         if (err == 0) {
138 out_assign:
139                 top->sym_filter_entry = he;
140         } else {
141                 char msg[BUFSIZ];
142                 symbol__strerror_disassemble(sym, map, err, msg, sizeof(msg));
143                 pr_err("Couldn't annotate %s: %s\n", sym->name, msg);
144         }
145
146         pthread_mutex_unlock(&notes->lock);
147         return err;
148 }
149
150 static void __zero_source_counters(struct hist_entry *he)
151 {
152         struct symbol *sym = he->ms.sym;
153         symbol__annotate_zero_histograms(sym);
154 }
155
156 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
157 {
158         struct utsname uts;
159         int err = uname(&uts);
160
161         ui__warning("Out of bounds address found:\n\n"
162                     "Addr:   %" PRIx64 "\n"
163                     "DSO:    %s %c\n"
164                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
165                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
166                     "Arch:   %s\n"
167                     "Kernel: %s\n"
168                     "Tools:  %s\n\n"
169                     "Not all samples will be on the annotation output.\n\n"
170                     "Please report to linux-kernel@vger.kernel.org\n",
171                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
172                     map->start, map->end, sym->start, sym->end,
173                     sym->binding == STB_GLOBAL ? 'g' :
174                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
175                     err ? "[unknown]" : uts.machine,
176                     err ? "[unknown]" : uts.release, perf_version_string);
177         if (use_browser <= 0)
178                 sleep(5);
179
180         map->erange_warned = true;
181 }
182
183 static void perf_top__record_precise_ip(struct perf_top *top,
184                                         struct hist_entry *he,
185                                         int counter, u64 ip)
186 {
187         struct annotation *notes;
188         struct symbol *sym = he->ms.sym;
189         int err = 0;
190
191         if (sym == NULL || (use_browser == 0 &&
192                             (top->sym_filter_entry == NULL ||
193                              top->sym_filter_entry->ms.sym != sym)))
194                 return;
195
196         notes = symbol__annotation(sym);
197
198         if (pthread_mutex_trylock(&notes->lock))
199                 return;
200
201         err = hist_entry__inc_addr_samples(he, counter, ip);
202
203         pthread_mutex_unlock(&notes->lock);
204
205         if (unlikely(err)) {
206                 /*
207                  * This function is now called with he->hists->lock held.
208                  * Release it before going to sleep.
209                  */
210                 pthread_mutex_unlock(&he->hists->lock);
211
212                 if (err == -ERANGE && !he->ms.map->erange_warned)
213                         ui__warn_map_erange(he->ms.map, sym, ip);
214                 else if (err == -ENOMEM) {
215                         pr_err("Not enough memory for annotating '%s' symbol!\n",
216                                sym->name);
217                         sleep(1);
218                 }
219
220                 pthread_mutex_lock(&he->hists->lock);
221         }
222 }
223
224 static void perf_top__show_details(struct perf_top *top)
225 {
226         struct hist_entry *he = top->sym_filter_entry;
227         struct annotation *notes;
228         struct symbol *symbol;
229         int more;
230
231         if (!he)
232                 return;
233
234         symbol = he->ms.sym;
235         notes = symbol__annotation(symbol);
236
237         pthread_mutex_lock(&notes->lock);
238
239         if (notes->src == NULL)
240                 goto out_unlock;
241
242         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
243         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
244
245         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
246                                        0, top->sym_pcnt_filter, top->print_entries, 4);
247
248         if (top->evlist->enabled) {
249                 if (top->zero)
250                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
251                 else
252                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
253         }
254         if (more != 0)
255                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
256 out_unlock:
257         pthread_mutex_unlock(&notes->lock);
258 }
259
260 static void perf_top__print_sym_table(struct perf_top *top)
261 {
262         char bf[160];
263         int printed = 0;
264         const int win_width = top->winsize.ws_col - 1;
265         struct perf_evsel *evsel = top->sym_evsel;
266         struct hists *hists = evsel__hists(evsel);
267
268         puts(CONSOLE_CLEAR);
269
270         perf_top__header_snprintf(top, bf, sizeof(bf));
271         printf("%s\n", bf);
272
273         perf_top__reset_sample_counters(top);
274
275         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
276
277         if (hists->stats.nr_lost_warned !=
278             hists->stats.nr_events[PERF_RECORD_LOST]) {
279                 hists->stats.nr_lost_warned =
280                               hists->stats.nr_events[PERF_RECORD_LOST];
281                 color_fprintf(stdout, PERF_COLOR_RED,
282                               "WARNING: LOST %d chunks, Check IO/CPU overload",
283                               hists->stats.nr_lost_warned);
284                 ++printed;
285         }
286
287         if (top->sym_filter_entry) {
288                 perf_top__show_details(top);
289                 return;
290         }
291
292         if (top->evlist->enabled) {
293                 if (top->zero) {
294                         hists__delete_entries(hists);
295                 } else {
296                         hists__decay_entries(hists, top->hide_user_symbols,
297                                              top->hide_kernel_symbols);
298                 }
299         }
300
301         hists__collapse_resort(hists, NULL);
302         perf_evsel__output_resort(evsel, NULL);
303
304         hists__output_recalc_col_len(hists, top->print_entries - printed);
305         putchar('\n');
306         hists__fprintf(hists, false, top->print_entries - printed, win_width,
307                        top->min_percent, stdout, symbol_conf.use_callchain);
308 }
309
310 static void prompt_integer(int *target, const char *msg)
311 {
312         char *buf = malloc(0), *p;
313         size_t dummy = 0;
314         int tmp;
315
316         fprintf(stdout, "\n%s: ", msg);
317         if (getline(&buf, &dummy, stdin) < 0)
318                 return;
319
320         p = strchr(buf, '\n');
321         if (p)
322                 *p = 0;
323
324         p = buf;
325         while(*p) {
326                 if (!isdigit(*p))
327                         goto out_free;
328                 p++;
329         }
330         tmp = strtoul(buf, NULL, 10);
331         *target = tmp;
332 out_free:
333         free(buf);
334 }
335
336 static void prompt_percent(int *target, const char *msg)
337 {
338         int tmp = 0;
339
340         prompt_integer(&tmp, msg);
341         if (tmp >= 0 && tmp <= 100)
342                 *target = tmp;
343 }
344
345 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
346 {
347         char *buf = malloc(0), *p;
348         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
349         struct hists *hists = evsel__hists(top->sym_evsel);
350         struct rb_node *next;
351         size_t dummy = 0;
352
353         /* zero counters of active symbol */
354         if (syme) {
355                 __zero_source_counters(syme);
356                 top->sym_filter_entry = NULL;
357         }
358
359         fprintf(stdout, "\n%s: ", msg);
360         if (getline(&buf, &dummy, stdin) < 0)
361                 goto out_free;
362
363         p = strchr(buf, '\n');
364         if (p)
365                 *p = 0;
366
367         next = rb_first(&hists->entries);
368         while (next) {
369                 n = rb_entry(next, struct hist_entry, rb_node);
370                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
371                         found = n;
372                         break;
373                 }
374                 next = rb_next(&n->rb_node);
375         }
376
377         if (!found) {
378                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
379                 sleep(1);
380         } else
381                 perf_top__parse_source(top, found);
382
383 out_free:
384         free(buf);
385 }
386
387 static void perf_top__print_mapped_keys(struct perf_top *top)
388 {
389         char *name = NULL;
390
391         if (top->sym_filter_entry) {
392                 struct symbol *sym = top->sym_filter_entry->ms.sym;
393                 name = sym->name;
394         }
395
396         fprintf(stdout, "\nMapped keys:\n");
397         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
398         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
399
400         if (top->evlist->nr_entries > 1)
401                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
402
403         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
404
405         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
406         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
407         fprintf(stdout, "\t[S]     stop annotation.\n");
408
409         fprintf(stdout,
410                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
411                 top->hide_kernel_symbols ? "yes" : "no");
412         fprintf(stdout,
413                 "\t[U]     hide user symbols.               \t(%s)\n",
414                 top->hide_user_symbols ? "yes" : "no");
415         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
416         fprintf(stdout, "\t[qQ]    quit.\n");
417 }
418
419 static int perf_top__key_mapped(struct perf_top *top, int c)
420 {
421         switch (c) {
422                 case 'd':
423                 case 'e':
424                 case 'f':
425                 case 'z':
426                 case 'q':
427                 case 'Q':
428                 case 'K':
429                 case 'U':
430                 case 'F':
431                 case 's':
432                 case 'S':
433                         return 1;
434                 case 'E':
435                         return top->evlist->nr_entries > 1 ? 1 : 0;
436                 default:
437                         break;
438         }
439
440         return 0;
441 }
442
443 static bool perf_top__handle_keypress(struct perf_top *top, int c)
444 {
445         bool ret = true;
446
447         if (!perf_top__key_mapped(top, c)) {
448                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
449                 struct termios save;
450
451                 perf_top__print_mapped_keys(top);
452                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
453                 fflush(stdout);
454
455                 set_term_quiet_input(&save);
456
457                 poll(&stdin_poll, 1, -1);
458                 c = getc(stdin);
459
460                 tcsetattr(0, TCSAFLUSH, &save);
461                 if (!perf_top__key_mapped(top, c))
462                         return ret;
463         }
464
465         switch (c) {
466                 case 'd':
467                         prompt_integer(&top->delay_secs, "Enter display delay");
468                         if (top->delay_secs < 1)
469                                 top->delay_secs = 1;
470                         break;
471                 case 'e':
472                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
473                         if (top->print_entries == 0) {
474                                 struct sigaction act = {
475                                         .sa_sigaction = perf_top__sig_winch,
476                                         .sa_flags     = SA_SIGINFO,
477                                 };
478                                 perf_top__sig_winch(SIGWINCH, NULL, top);
479                                 sigaction(SIGWINCH, &act, NULL);
480                         } else {
481                                 signal(SIGWINCH, SIG_DFL);
482                         }
483                         break;
484                 case 'E':
485                         if (top->evlist->nr_entries > 1) {
486                                 /* Select 0 as the default event: */
487                                 int counter = 0;
488
489                                 fprintf(stderr, "\nAvailable events:");
490
491                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
492                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
493
494                                 prompt_integer(&counter, "Enter details event counter");
495
496                                 if (counter >= top->evlist->nr_entries) {
497                                         top->sym_evsel = perf_evlist__first(top->evlist);
498                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
499                                         sleep(1);
500                                         break;
501                                 }
502                                 evlist__for_each_entry(top->evlist, top->sym_evsel)
503                                         if (top->sym_evsel->idx == counter)
504                                                 break;
505                         } else
506                                 top->sym_evsel = perf_evlist__first(top->evlist);
507                         break;
508                 case 'f':
509                         prompt_integer(&top->count_filter, "Enter display event count filter");
510                         break;
511                 case 'F':
512                         prompt_percent(&top->sym_pcnt_filter,
513                                        "Enter details display event filter (percent)");
514                         break;
515                 case 'K':
516                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
517                         break;
518                 case 'q':
519                 case 'Q':
520                         printf("exiting.\n");
521                         if (top->dump_symtab)
522                                 perf_session__fprintf_dsos(top->session, stderr);
523                         ret = false;
524                         break;
525                 case 's':
526                         perf_top__prompt_symbol(top, "Enter details symbol");
527                         break;
528                 case 'S':
529                         if (!top->sym_filter_entry)
530                                 break;
531                         else {
532                                 struct hist_entry *syme = top->sym_filter_entry;
533
534                                 top->sym_filter_entry = NULL;
535                                 __zero_source_counters(syme);
536                         }
537                         break;
538                 case 'U':
539                         top->hide_user_symbols = !top->hide_user_symbols;
540                         break;
541                 case 'z':
542                         top->zero = !top->zero;
543                         break;
544                 default:
545                         break;
546         }
547
548         return ret;
549 }
550
551 static void perf_top__sort_new_samples(void *arg)
552 {
553         struct perf_top *t = arg;
554         struct perf_evsel *evsel = t->sym_evsel;
555         struct hists *hists;
556
557         perf_top__reset_sample_counters(t);
558
559         if (t->evlist->selected != NULL)
560                 t->sym_evsel = t->evlist->selected;
561
562         hists = evsel__hists(evsel);
563
564         if (t->evlist->enabled) {
565                 if (t->zero) {
566                         hists__delete_entries(hists);
567                 } else {
568                         hists__decay_entries(hists, t->hide_user_symbols,
569                                              t->hide_kernel_symbols);
570                 }
571         }
572
573         hists__collapse_resort(hists, NULL);
574         perf_evsel__output_resort(evsel, NULL);
575 }
576
577 static void *display_thread_tui(void *arg)
578 {
579         struct perf_evsel *pos;
580         struct perf_top *top = arg;
581         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
582         struct hist_browser_timer hbt = {
583                 .timer          = perf_top__sort_new_samples,
584                 .arg            = top,
585                 .refresh        = top->delay_secs,
586         };
587
588         perf_top__sort_new_samples(top);
589
590         /*
591          * Initialize the uid_filter_str, in the future the TUI will allow
592          * Zooming in/out UIDs. For now juse use whatever the user passed
593          * via --uid.
594          */
595         evlist__for_each_entry(top->evlist, pos) {
596                 struct hists *hists = evsel__hists(pos);
597                 hists->uid_filter_str = top->record_opts.target.uid_str;
598         }
599
600         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
601                                       top->min_percent,
602                                       &top->session->header.env);
603
604         done = 1;
605         return NULL;
606 }
607
608 static void display_sig(int sig __maybe_unused)
609 {
610         done = 1;
611 }
612
613 static void display_setup_sig(void)
614 {
615         signal(SIGSEGV, sighandler_dump_stack);
616         signal(SIGFPE, sighandler_dump_stack);
617         signal(SIGINT,  display_sig);
618         signal(SIGQUIT, display_sig);
619         signal(SIGTERM, display_sig);
620 }
621
622 static void *display_thread(void *arg)
623 {
624         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
625         struct termios save;
626         struct perf_top *top = arg;
627         int delay_msecs, c;
628
629         display_setup_sig();
630         pthread__unblock_sigwinch();
631 repeat:
632         delay_msecs = top->delay_secs * MSEC_PER_SEC;
633         set_term_quiet_input(&save);
634         /* trash return*/
635         getc(stdin);
636
637         while (!done) {
638                 perf_top__print_sym_table(top);
639                 /*
640                  * Either timeout expired or we got an EINTR due to SIGWINCH,
641                  * refresh screen in both cases.
642                  */
643                 switch (poll(&stdin_poll, 1, delay_msecs)) {
644                 case 0:
645                         continue;
646                 case -1:
647                         if (errno == EINTR)
648                                 continue;
649                         __fallthrough;
650                 default:
651                         c = getc(stdin);
652                         tcsetattr(0, TCSAFLUSH, &save);
653
654                         if (perf_top__handle_keypress(top, c))
655                                 goto repeat;
656                         done = 1;
657                 }
658         }
659
660         tcsetattr(0, TCSAFLUSH, &save);
661         return NULL;
662 }
663
664 static int hist_iter__top_callback(struct hist_entry_iter *iter,
665                                    struct addr_location *al, bool single,
666                                    void *arg)
667 {
668         struct perf_top *top = arg;
669         struct hist_entry *he = iter->he;
670         struct perf_evsel *evsel = iter->evsel;
671
672         if (perf_hpp_list.sym && single)
673                 perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
674
675         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
676                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
677         return 0;
678 }
679
680 static void perf_event__process_sample(struct perf_tool *tool,
681                                        const union perf_event *event,
682                                        struct perf_evsel *evsel,
683                                        struct perf_sample *sample,
684                                        struct machine *machine)
685 {
686         struct perf_top *top = container_of(tool, struct perf_top, tool);
687         struct addr_location al;
688         int err;
689
690         if (!machine && perf_guest) {
691                 static struct intlist *seen;
692
693                 if (!seen)
694                         seen = intlist__new(NULL);
695
696                 if (!intlist__has_entry(seen, sample->pid)) {
697                         pr_err("Can't find guest [%d]'s kernel information\n",
698                                 sample->pid);
699                         intlist__add(seen, sample->pid);
700                 }
701                 return;
702         }
703
704         if (!machine) {
705                 pr_err("%u unprocessable samples recorded.\r",
706                        top->session->evlist->stats.nr_unprocessable_samples++);
707                 return;
708         }
709
710         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
711                 top->exact_samples++;
712
713         if (machine__resolve(machine, &al, sample) < 0)
714                 return;
715
716         if (!machine->kptr_restrict_warned &&
717             symbol_conf.kptr_restrict &&
718             al.cpumode == PERF_RECORD_MISC_KERNEL) {
719                 ui__warning(
720 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
721 "Check /proc/sys/kernel/kptr_restrict.\n\n"
722 "Kernel%s samples will not be resolved.\n",
723                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
724                           " modules" : "");
725                 if (use_browser <= 0)
726                         sleep(5);
727                 machine->kptr_restrict_warned = true;
728         }
729
730         if (al.sym == NULL) {
731                 const char *msg = "Kernel samples will not be resolved.\n";
732                 /*
733                  * As we do lazy loading of symtabs we only will know if the
734                  * specified vmlinux file is invalid when we actually have a
735                  * hit in kernel space and then try to load it. So if we get
736                  * here and there are _no_ symbols in the DSO backing the
737                  * kernel map, bail out.
738                  *
739                  * We may never get here, for instance, if we use -K/
740                  * --hide-kernel-symbols, even if the user specifies an
741                  * invalid --vmlinux ;-)
742                  */
743                 if (!machine->kptr_restrict_warned && !top->vmlinux_warned &&
744                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
745                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
746                         if (symbol_conf.vmlinux_name) {
747                                 char serr[256];
748                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
749                                 ui__warning("The %s file can't be used: %s\n%s",
750                                             symbol_conf.vmlinux_name, serr, msg);
751                         } else {
752                                 ui__warning("A vmlinux file was not found.\n%s",
753                                             msg);
754                         }
755
756                         if (use_browser <= 0)
757                                 sleep(5);
758                         top->vmlinux_warned = true;
759                 }
760         }
761
762         if (al.sym == NULL || !al.sym->idle) {
763                 struct hists *hists = evsel__hists(evsel);
764                 struct hist_entry_iter iter = {
765                         .evsel          = evsel,
766                         .sample         = sample,
767                         .add_entry_cb   = hist_iter__top_callback,
768                 };
769
770                 if (symbol_conf.cumulate_callchain)
771                         iter.ops = &hist_iter_cumulative;
772                 else
773                         iter.ops = &hist_iter_normal;
774
775                 pthread_mutex_lock(&hists->lock);
776
777                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
778                 if (err < 0)
779                         pr_err("Problem incrementing symbol period, skipping event\n");
780
781                 pthread_mutex_unlock(&hists->lock);
782         }
783
784         addr_location__put(&al);
785 }
786
787 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
788 {
789         struct perf_sample sample;
790         struct perf_evsel *evsel;
791         struct perf_session *session = top->session;
792         union perf_event *event;
793         struct machine *machine;
794         int ret;
795
796         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
797                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
798                 if (ret) {
799                         pr_err("Can't parse sample, err = %d\n", ret);
800                         goto next_event;
801                 }
802
803                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
804                 assert(evsel != NULL);
805
806                 if (event->header.type == PERF_RECORD_SAMPLE)
807                         ++top->samples;
808
809                 switch (sample.cpumode) {
810                 case PERF_RECORD_MISC_USER:
811                         ++top->us_samples;
812                         if (top->hide_user_symbols)
813                                 goto next_event;
814                         machine = &session->machines.host;
815                         break;
816                 case PERF_RECORD_MISC_KERNEL:
817                         ++top->kernel_samples;
818                         if (top->hide_kernel_symbols)
819                                 goto next_event;
820                         machine = &session->machines.host;
821                         break;
822                 case PERF_RECORD_MISC_GUEST_KERNEL:
823                         ++top->guest_kernel_samples;
824                         machine = perf_session__find_machine(session,
825                                                              sample.pid);
826                         break;
827                 case PERF_RECORD_MISC_GUEST_USER:
828                         ++top->guest_us_samples;
829                         /*
830                          * TODO: we don't process guest user from host side
831                          * except simple counting.
832                          */
833                         goto next_event;
834                 default:
835                         if (event->header.type == PERF_RECORD_SAMPLE)
836                                 goto next_event;
837                         machine = &session->machines.host;
838                         break;
839                 }
840
841
842                 if (event->header.type == PERF_RECORD_SAMPLE) {
843                         perf_event__process_sample(&top->tool, event, evsel,
844                                                    &sample, machine);
845                 } else if (event->header.type < PERF_RECORD_MAX) {
846                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
847                         machine__process_event(machine, event, &sample);
848                 } else
849                         ++session->evlist->stats.nr_unknown_events;
850 next_event:
851                 perf_evlist__mmap_consume(top->evlist, idx);
852         }
853 }
854
855 static void perf_top__mmap_read(struct perf_top *top)
856 {
857         int i;
858
859         for (i = 0; i < top->evlist->nr_mmaps; i++)
860                 perf_top__mmap_read_idx(top, i);
861 }
862
863 static int perf_top__start_counters(struct perf_top *top)
864 {
865         char msg[BUFSIZ];
866         struct perf_evsel *counter;
867         struct perf_evlist *evlist = top->evlist;
868         struct record_opts *opts = &top->record_opts;
869
870         perf_evlist__config(evlist, opts, &callchain_param);
871
872         evlist__for_each_entry(evlist, counter) {
873 try_again:
874                 if (perf_evsel__open(counter, top->evlist->cpus,
875                                      top->evlist->threads) < 0) {
876                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
877                                 if (verbose > 0)
878                                         ui__warning("%s\n", msg);
879                                 goto try_again;
880                         }
881
882                         perf_evsel__open_strerror(counter, &opts->target,
883                                                   errno, msg, sizeof(msg));
884                         ui__error("%s\n", msg);
885                         goto out_err;
886                 }
887         }
888
889         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
890                 ui__error("Failed to mmap with %d (%s)\n",
891                             errno, str_error_r(errno, msg, sizeof(msg)));
892                 goto out_err;
893         }
894
895         return 0;
896
897 out_err:
898         return -1;
899 }
900
901 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
902 {
903         if (!perf_hpp_list.sym) {
904                 if (callchain->enabled) {
905                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
906                         return -EINVAL;
907                 }
908         } else if (callchain->mode != CHAIN_NONE) {
909                 if (callchain_register_param(callchain) < 0) {
910                         ui__error("Can't register callchain params.\n");
911                         return -EINVAL;
912                 }
913         }
914
915         return 0;
916 }
917
918 static int __cmd_top(struct perf_top *top)
919 {
920         char msg[512];
921         struct perf_evsel *pos;
922         struct perf_evsel_config_term *err_term;
923         struct perf_evlist *evlist = top->evlist;
924         struct record_opts *opts = &top->record_opts;
925         pthread_t thread;
926         int ret;
927
928         top->session = perf_session__new(NULL, false, NULL);
929         if (top->session == NULL)
930                 return -1;
931
932         if (!objdump_path) {
933                 ret = perf_env__lookup_objdump(&top->session->header.env);
934                 if (ret)
935                         goto out_delete;
936         }
937
938         ret = callchain_param__setup_sample_type(&callchain_param);
939         if (ret)
940                 goto out_delete;
941
942         if (perf_session__register_idle_thread(top->session) < 0)
943                 goto out_delete;
944
945         machine__synthesize_threads(&top->session->machines.host, &opts->target,
946                                     top->evlist->threads, false, opts->proc_map_timeout);
947
948         if (perf_hpp_list.socket) {
949                 ret = perf_env__read_cpu_topology_map(&perf_env);
950                 if (ret < 0)
951                         goto out_err_cpu_topo;
952         }
953
954         ret = perf_top__start_counters(top);
955         if (ret)
956                 goto out_delete;
957
958         ret = perf_evlist__apply_drv_configs(evlist, &pos, &err_term);
959         if (ret) {
960                 error("failed to set config \"%s\" on event %s with %d (%s)\n",
961                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
962                         str_error_r(errno, msg, sizeof(msg)));
963                 goto out_delete;
964         }
965
966         top->session->evlist = top->evlist;
967         perf_session__set_id_hdr_size(top->session);
968
969         /*
970          * When perf is starting the traced process, all the events (apart from
971          * group members) have enable_on_exec=1 set, so don't spoil it by
972          * prematurely enabling them.
973          *
974          * XXX 'top' still doesn't start workloads like record, trace, but should,
975          * so leave the check here.
976          */
977         if (!target__none(&opts->target))
978                 perf_evlist__enable(top->evlist);
979
980         /* Wait for a minimal set of events before starting the snapshot */
981         perf_evlist__poll(top->evlist, 100);
982
983         perf_top__mmap_read(top);
984
985         ret = -1;
986         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
987                                                             display_thread), top)) {
988                 ui__error("Could not create display thread.\n");
989                 goto out_delete;
990         }
991
992         if (top->realtime_prio) {
993                 struct sched_param param;
994
995                 param.sched_priority = top->realtime_prio;
996                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
997                         ui__error("Could not set realtime priority.\n");
998                         goto out_join;
999                 }
1000         }
1001
1002         while (!done) {
1003                 u64 hits = top->samples;
1004
1005                 perf_top__mmap_read(top);
1006
1007                 if (hits == top->samples)
1008                         ret = perf_evlist__poll(top->evlist, 100);
1009         }
1010
1011         ret = 0;
1012 out_join:
1013         pthread_join(thread, NULL);
1014 out_delete:
1015         perf_session__delete(top->session);
1016         top->session = NULL;
1017
1018         return ret;
1019
1020 out_err_cpu_topo: {
1021         char errbuf[BUFSIZ];
1022         const char *err = str_error_r(-ret, errbuf, sizeof(errbuf));
1023
1024         ui__error("Could not read the CPU topology map: %s\n", err);
1025         goto out_delete;
1026 }
1027 }
1028
1029 static int
1030 callchain_opt(const struct option *opt, const char *arg, int unset)
1031 {
1032         symbol_conf.use_callchain = true;
1033         return record_callchain_opt(opt, arg, unset);
1034 }
1035
1036 static int
1037 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1038 {
1039         struct callchain_param *callchain = opt->value;
1040
1041         callchain->enabled = !unset;
1042         callchain->record_mode = CALLCHAIN_FP;
1043
1044         /*
1045          * --no-call-graph
1046          */
1047         if (unset) {
1048                 symbol_conf.use_callchain = false;
1049                 callchain->record_mode = CALLCHAIN_NONE;
1050                 return 0;
1051         }
1052
1053         return parse_callchain_top_opt(arg);
1054 }
1055
1056 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1057 {
1058         if (!strcmp(var, "top.call-graph"))
1059                 var = "call-graph.record-mode"; /* fall-through */
1060         if (!strcmp(var, "top.children")) {
1061                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1062                 return 0;
1063         }
1064
1065         return 0;
1066 }
1067
1068 static int
1069 parse_percent_limit(const struct option *opt, const char *arg,
1070                     int unset __maybe_unused)
1071 {
1072         struct perf_top *top = opt->value;
1073
1074         top->min_percent = strtof(arg, NULL);
1075         return 0;
1076 }
1077
1078 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1079         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1080
1081 int cmd_top(int argc, const char **argv)
1082 {
1083         char errbuf[BUFSIZ];
1084         struct perf_top top = {
1085                 .count_filter        = 5,
1086                 .delay_secs          = 2,
1087                 .record_opts = {
1088                         .mmap_pages     = UINT_MAX,
1089                         .user_freq      = UINT_MAX,
1090                         .user_interval  = ULLONG_MAX,
1091                         .freq           = 4000, /* 4 KHz */
1092                         .target         = {
1093                                 .uses_mmap   = true,
1094                         },
1095                         .proc_map_timeout    = 500,
1096                 },
1097                 .max_stack           = sysctl_perf_event_max_stack,
1098                 .sym_pcnt_filter     = 5,
1099         };
1100         struct record_opts *opts = &top.record_opts;
1101         struct target *target = &opts->target;
1102         const struct option options[] = {
1103         OPT_CALLBACK('e', "event", &top.evlist, "event",
1104                      "event selector. use 'perf list' to list available events",
1105                      parse_events_option),
1106         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1107         OPT_STRING('p', "pid", &target->pid, "pid",
1108                     "profile events on existing process id"),
1109         OPT_STRING('t', "tid", &target->tid, "tid",
1110                     "profile events on existing thread id"),
1111         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1112                             "system-wide collection from all CPUs"),
1113         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1114                     "list of cpus to monitor"),
1115         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1116                    "file", "vmlinux pathname"),
1117         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1118                     "don't load vmlinux even if found"),
1119         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1120                     "hide kernel symbols"),
1121         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1122                      "number of mmap data pages",
1123                      perf_evlist__parse_mmap_pages),
1124         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1125                     "collect data with this RT SCHED_FIFO priority"),
1126         OPT_INTEGER('d', "delay", &top.delay_secs,
1127                     "number of seconds to delay between refreshes"),
1128         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1129                             "dump the symbol table used for profiling"),
1130         OPT_INTEGER('f', "count-filter", &top.count_filter,
1131                     "only display functions with more events than this"),
1132         OPT_BOOLEAN(0, "group", &opts->group,
1133                             "put the counters into a counter group"),
1134         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1135                     "child tasks do not inherit counters"),
1136         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1137                     "symbol to annotate"),
1138         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1139         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1140         OPT_INTEGER('E', "entries", &top.print_entries,
1141                     "display this many functions"),
1142         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1143                     "hide user symbols"),
1144         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1145         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1146         OPT_INCR('v', "verbose", &verbose,
1147                     "be more verbose (show counter open errors, etc)"),
1148         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1149                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1150                    " Please refer the man page for the complete list."),
1151         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1152                    "output field(s): overhead, period, sample plus all of sort keys"),
1153         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1154                     "Show a column with the number of samples"),
1155         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1156                            NULL, "enables call-graph recording and display",
1157                            &callchain_opt),
1158         OPT_CALLBACK(0, "call-graph", &callchain_param,
1159                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1160                      top_callchain_help, &parse_callchain_opt),
1161         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1162                     "Accumulate callchains of children and show total overhead as well"),
1163         OPT_INTEGER(0, "max-stack", &top.max_stack,
1164                     "Set the maximum stack depth when parsing the callchain. "
1165                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1166         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1167                    "ignore callees of these functions in call graphs",
1168                    report_parse_ignore_callees_opt),
1169         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1170                     "Show a column with the sum of periods"),
1171         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1172                    "only consider symbols in these dsos"),
1173         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1174                    "only consider symbols in these comms"),
1175         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1176                    "only consider these symbols"),
1177         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1178                     "Interleave source code with assembly code (default)"),
1179         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1180                     "Display raw encoding of assembly instructions (default)"),
1181         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1182                     "Enable kernel symbol demangling"),
1183         OPT_STRING(0, "objdump", &objdump_path, "path",
1184                     "objdump binary to use for disassembly and annotations"),
1185         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1186                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1187         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1188         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1189                      "Don't show entries under that percent", parse_percent_limit),
1190         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1191                      "How to display percentage of filtered entries", parse_filter_percentage),
1192         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1193                    "width[,width...]",
1194                    "don't try to adjust column width, use these fixed values"),
1195         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1196                         "per thread proc mmap processing timeout in ms"),
1197         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1198                      "branch any", "sample any taken branches",
1199                      parse_branch_stack),
1200         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1201                      "branch filter mask", "branch stack filter modes",
1202                      parse_branch_stack),
1203         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1204                     "Show raw trace event output (do not use print fmt or plugins)"),
1205         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1206                     "Show entries in a hierarchy"),
1207         OPT_END()
1208         };
1209         const char * const top_usage[] = {
1210                 "perf top [<options>]",
1211                 NULL
1212         };
1213         int status = hists__init();
1214
1215         if (status < 0)
1216                 return status;
1217
1218         top.evlist = perf_evlist__new();
1219         if (top.evlist == NULL)
1220                 return -ENOMEM;
1221
1222         status = perf_config(perf_top_config, &top);
1223         if (status)
1224                 return status;
1225
1226         argc = parse_options(argc, argv, options, top_usage, 0);
1227         if (argc)
1228                 usage_with_options(top_usage, options);
1229
1230         if (!top.evlist->nr_entries &&
1231             perf_evlist__add_default(top.evlist) < 0) {
1232                 pr_err("Not enough memory for event selector list\n");
1233                 goto out_delete_evlist;
1234         }
1235
1236         if (symbol_conf.report_hierarchy) {
1237                 /* disable incompatible options */
1238                 symbol_conf.event_group = false;
1239                 symbol_conf.cumulate_callchain = false;
1240
1241                 if (field_order) {
1242                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1243                         parse_options_usage(top_usage, options, "fields", 0);
1244                         parse_options_usage(NULL, options, "hierarchy", 0);
1245                         goto out_delete_evlist;
1246                 }
1247         }
1248
1249         sort__mode = SORT_MODE__TOP;
1250         /* display thread wants entries to be collapsed in a different tree */
1251         perf_hpp_list.need_collapse = 1;
1252
1253         if (top.use_stdio)
1254                 use_browser = 0;
1255         else if (top.use_tui)
1256                 use_browser = 1;
1257
1258         setup_browser(false);
1259
1260         if (setup_sorting(top.evlist) < 0) {
1261                 if (sort_order)
1262                         parse_options_usage(top_usage, options, "s", 1);
1263                 if (field_order)
1264                         parse_options_usage(sort_order ? NULL : top_usage,
1265                                             options, "fields", 0);
1266                 goto out_delete_evlist;
1267         }
1268
1269         status = target__validate(target);
1270         if (status) {
1271                 target__strerror(target, status, errbuf, BUFSIZ);
1272                 ui__warning("%s\n", errbuf);
1273         }
1274
1275         status = target__parse_uid(target);
1276         if (status) {
1277                 int saved_errno = errno;
1278
1279                 target__strerror(target, status, errbuf, BUFSIZ);
1280                 ui__error("%s\n", errbuf);
1281
1282                 status = -saved_errno;
1283                 goto out_delete_evlist;
1284         }
1285
1286         if (target__none(target))
1287                 target->system_wide = true;
1288
1289         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1290                 ui__error("Couldn't create thread/CPU maps: %s\n",
1291                           errno == ENOENT ? "No such process" : str_error_r(errno, errbuf, sizeof(errbuf)));
1292                 goto out_delete_evlist;
1293         }
1294
1295         symbol_conf.nr_events = top.evlist->nr_entries;
1296
1297         if (top.delay_secs < 1)
1298                 top.delay_secs = 1;
1299
1300         if (record_opts__config(opts)) {
1301                 status = -EINVAL;
1302                 goto out_delete_evlist;
1303         }
1304
1305         top.sym_evsel = perf_evlist__first(top.evlist);
1306
1307         if (!callchain_param.enabled) {
1308                 symbol_conf.cumulate_callchain = false;
1309                 perf_hpp__cancel_cumulate();
1310         }
1311
1312         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1313                 callchain_param.order = ORDER_CALLER;
1314
1315         status = symbol__annotation_init();
1316         if (status < 0)
1317                 goto out_delete_evlist;
1318
1319         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1320         if (symbol__init(NULL) < 0)
1321                 return -1;
1322
1323         sort__setup_elide(stdout);
1324
1325         get_term_dimensions(&top.winsize);
1326         if (top.print_entries == 0) {
1327                 struct sigaction act = {
1328                         .sa_sigaction = perf_top__sig_winch,
1329                         .sa_flags     = SA_SIGINFO,
1330                 };
1331                 perf_top__update_print_entries(&top);
1332                 sigaction(SIGWINCH, &act, NULL);
1333         }
1334
1335         status = __cmd_top(&top);
1336
1337 out_delete_evlist:
1338         perf_evlist__delete(top.evlist);
1339
1340         return status;
1341 }