94af190f6843d2c92e68b2dfef8dc51017c42079
[cascardo/linux.git] / tools / perf / builtin-top.c
1 /*
2  * builtin-top.c
3  *
4  * Builtin top command: Display a continuously updated profile of
5  * any workload, CPU or specific PID.
6  *
7  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
8  *               2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
9  *
10  * Improvements and fixes by:
11  *
12  *   Arjan van de Ven <arjan@linux.intel.com>
13  *   Yanmin Zhang <yanmin.zhang@intel.com>
14  *   Wu Fengguang <fengguang.wu@intel.com>
15  *   Mike Galbraith <efault@gmx.de>
16  *   Paul Mackerras <paulus@samba.org>
17  *
18  * Released under the GPL v2. (and only v2, not any later version)
19  */
20 #include "builtin.h"
21
22 #include "perf.h"
23
24 #include "util/annotate.h"
25 #include "util/cache.h"
26 #include "util/color.h"
27 #include "util/evlist.h"
28 #include "util/evsel.h"
29 #include "util/machine.h"
30 #include "util/session.h"
31 #include "util/symbol.h"
32 #include "util/thread.h"
33 #include "util/thread_map.h"
34 #include "util/top.h"
35 #include "util/util.h"
36 #include <linux/rbtree.h>
37 #include <subcmd/parse-options.h>
38 #include "util/parse-events.h"
39 #include "util/cpumap.h"
40 #include "util/xyarray.h"
41 #include "util/sort.h"
42 #include "util/intlist.h"
43 #include "util/parse-branch-options.h"
44 #include "arch/common.h"
45
46 #include "util/debug.h"
47
48 #include <assert.h>
49 #include <elf.h>
50 #include <fcntl.h>
51
52 #include <stdio.h>
53 #include <termios.h>
54 #include <unistd.h>
55 #include <inttypes.h>
56
57 #include <errno.h>
58 #include <time.h>
59 #include <sched.h>
60
61 #include <sys/syscall.h>
62 #include <sys/ioctl.h>
63 #include <poll.h>
64 #include <sys/prctl.h>
65 #include <sys/wait.h>
66 #include <sys/uio.h>
67 #include <sys/utsname.h>
68 #include <sys/mman.h>
69
70 #include <linux/types.h>
71
72 static volatile int done;
73
74 #define HEADER_LINE_NR  5
75
76 static void perf_top__update_print_entries(struct perf_top *top)
77 {
78         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
79 }
80
81 static void perf_top__sig_winch(int sig __maybe_unused,
82                                 siginfo_t *info __maybe_unused, void *arg)
83 {
84         struct perf_top *top = arg;
85
86         get_term_dimensions(&top->winsize);
87         perf_top__update_print_entries(top);
88 }
89
90 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
91 {
92         struct symbol *sym;
93         struct annotation *notes;
94         struct map *map;
95         int err = -1;
96
97         if (!he || !he->ms.sym)
98                 return -1;
99
100         sym = he->ms.sym;
101         map = he->ms.map;
102
103         /*
104          * We can't annotate with just /proc/kallsyms
105          */
106         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
107             !dso__is_kcore(map->dso)) {
108                 pr_err("Can't annotate %s: No vmlinux file was found in the "
109                        "path\n", sym->name);
110                 sleep(1);
111                 return -1;
112         }
113
114         notes = symbol__annotation(sym);
115         if (notes->src != NULL) {
116                 pthread_mutex_lock(&notes->lock);
117                 goto out_assign;
118         }
119
120         pthread_mutex_lock(&notes->lock);
121
122         if (symbol__alloc_hist(sym) < 0) {
123                 pthread_mutex_unlock(&notes->lock);
124                 pr_err("Not enough memory for annotating '%s' symbol!\n",
125                        sym->name);
126                 sleep(1);
127                 return err;
128         }
129
130         err = symbol__annotate(sym, map, 0);
131         if (err == 0) {
132 out_assign:
133                 top->sym_filter_entry = he;
134         }
135
136         pthread_mutex_unlock(&notes->lock);
137         return err;
138 }
139
140 static void __zero_source_counters(struct hist_entry *he)
141 {
142         struct symbol *sym = he->ms.sym;
143         symbol__annotate_zero_histograms(sym);
144 }
145
146 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
147 {
148         struct utsname uts;
149         int err = uname(&uts);
150
151         ui__warning("Out of bounds address found:\n\n"
152                     "Addr:   %" PRIx64 "\n"
153                     "DSO:    %s %c\n"
154                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
155                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
156                     "Arch:   %s\n"
157                     "Kernel: %s\n"
158                     "Tools:  %s\n\n"
159                     "Not all samples will be on the annotation output.\n\n"
160                     "Please report to linux-kernel@vger.kernel.org\n",
161                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
162                     map->start, map->end, sym->start, sym->end,
163                     sym->binding == STB_GLOBAL ? 'g' :
164                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
165                     err ? "[unknown]" : uts.machine,
166                     err ? "[unknown]" : uts.release, perf_version_string);
167         if (use_browser <= 0)
168                 sleep(5);
169
170         map->erange_warned = true;
171 }
172
173 static void perf_top__record_precise_ip(struct perf_top *top,
174                                         struct hist_entry *he,
175                                         int counter, u64 ip)
176 {
177         struct annotation *notes;
178         struct symbol *sym = he->ms.sym;
179         int err = 0;
180
181         if (sym == NULL || (use_browser == 0 &&
182                             (top->sym_filter_entry == NULL ||
183                              top->sym_filter_entry->ms.sym != sym)))
184                 return;
185
186         notes = symbol__annotation(sym);
187
188         if (pthread_mutex_trylock(&notes->lock))
189                 return;
190
191         err = hist_entry__inc_addr_samples(he, counter, ip);
192
193         pthread_mutex_unlock(&notes->lock);
194
195         if (unlikely(err)) {
196                 /*
197                  * This function is now called with he->hists->lock held.
198                  * Release it before going to sleep.
199                  */
200                 pthread_mutex_unlock(&he->hists->lock);
201
202                 if (err == -ERANGE && !he->ms.map->erange_warned)
203                         ui__warn_map_erange(he->ms.map, sym, ip);
204                 else if (err == -ENOMEM) {
205                         pr_err("Not enough memory for annotating '%s' symbol!\n",
206                                sym->name);
207                         sleep(1);
208                 }
209
210                 pthread_mutex_lock(&he->hists->lock);
211         }
212 }
213
214 static void perf_top__show_details(struct perf_top *top)
215 {
216         struct hist_entry *he = top->sym_filter_entry;
217         struct annotation *notes;
218         struct symbol *symbol;
219         int more;
220
221         if (!he)
222                 return;
223
224         symbol = he->ms.sym;
225         notes = symbol__annotation(symbol);
226
227         pthread_mutex_lock(&notes->lock);
228
229         if (notes->src == NULL)
230                 goto out_unlock;
231
232         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
233         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
234
235         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
236                                        0, top->sym_pcnt_filter, top->print_entries, 4);
237
238         if (top->evlist->enabled) {
239                 if (top->zero)
240                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
241                 else
242                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
243         }
244         if (more != 0)
245                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
246 out_unlock:
247         pthread_mutex_unlock(&notes->lock);
248 }
249
250 static void perf_top__print_sym_table(struct perf_top *top)
251 {
252         char bf[160];
253         int printed = 0;
254         const int win_width = top->winsize.ws_col - 1;
255         struct perf_evsel *evsel = top->sym_evsel;
256         struct hists *hists = evsel__hists(evsel);
257
258         puts(CONSOLE_CLEAR);
259
260         perf_top__header_snprintf(top, bf, sizeof(bf));
261         printf("%s\n", bf);
262
263         perf_top__reset_sample_counters(top);
264
265         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
266
267         if (hists->stats.nr_lost_warned !=
268             hists->stats.nr_events[PERF_RECORD_LOST]) {
269                 hists->stats.nr_lost_warned =
270                               hists->stats.nr_events[PERF_RECORD_LOST];
271                 color_fprintf(stdout, PERF_COLOR_RED,
272                               "WARNING: LOST %d chunks, Check IO/CPU overload",
273                               hists->stats.nr_lost_warned);
274                 ++printed;
275         }
276
277         if (top->sym_filter_entry) {
278                 perf_top__show_details(top);
279                 return;
280         }
281
282         if (top->evlist->enabled) {
283                 if (top->zero) {
284                         hists__delete_entries(hists);
285                 } else {
286                         hists__decay_entries(hists, top->hide_user_symbols,
287                                              top->hide_kernel_symbols);
288                 }
289         }
290
291         hists__collapse_resort(hists, NULL);
292         perf_evsel__output_resort(evsel, NULL);
293
294         hists__output_recalc_col_len(hists, top->print_entries - printed);
295         putchar('\n');
296         hists__fprintf(hists, false, top->print_entries - printed, win_width,
297                        top->min_percent, stdout);
298 }
299
300 static void prompt_integer(int *target, const char *msg)
301 {
302         char *buf = malloc(0), *p;
303         size_t dummy = 0;
304         int tmp;
305
306         fprintf(stdout, "\n%s: ", msg);
307         if (getline(&buf, &dummy, stdin) < 0)
308                 return;
309
310         p = strchr(buf, '\n');
311         if (p)
312                 *p = 0;
313
314         p = buf;
315         while(*p) {
316                 if (!isdigit(*p))
317                         goto out_free;
318                 p++;
319         }
320         tmp = strtoul(buf, NULL, 10);
321         *target = tmp;
322 out_free:
323         free(buf);
324 }
325
326 static void prompt_percent(int *target, const char *msg)
327 {
328         int tmp = 0;
329
330         prompt_integer(&tmp, msg);
331         if (tmp >= 0 && tmp <= 100)
332                 *target = tmp;
333 }
334
335 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
336 {
337         char *buf = malloc(0), *p;
338         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
339         struct hists *hists = evsel__hists(top->sym_evsel);
340         struct rb_node *next;
341         size_t dummy = 0;
342
343         /* zero counters of active symbol */
344         if (syme) {
345                 __zero_source_counters(syme);
346                 top->sym_filter_entry = NULL;
347         }
348
349         fprintf(stdout, "\n%s: ", msg);
350         if (getline(&buf, &dummy, stdin) < 0)
351                 goto out_free;
352
353         p = strchr(buf, '\n');
354         if (p)
355                 *p = 0;
356
357         next = rb_first(&hists->entries);
358         while (next) {
359                 n = rb_entry(next, struct hist_entry, rb_node);
360                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
361                         found = n;
362                         break;
363                 }
364                 next = rb_next(&n->rb_node);
365         }
366
367         if (!found) {
368                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
369                 sleep(1);
370         } else
371                 perf_top__parse_source(top, found);
372
373 out_free:
374         free(buf);
375 }
376
377 static void perf_top__print_mapped_keys(struct perf_top *top)
378 {
379         char *name = NULL;
380
381         if (top->sym_filter_entry) {
382                 struct symbol *sym = top->sym_filter_entry->ms.sym;
383                 name = sym->name;
384         }
385
386         fprintf(stdout, "\nMapped keys:\n");
387         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
388         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
389
390         if (top->evlist->nr_entries > 1)
391                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
392
393         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
394
395         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
396         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
397         fprintf(stdout, "\t[S]     stop annotation.\n");
398
399         fprintf(stdout,
400                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
401                 top->hide_kernel_symbols ? "yes" : "no");
402         fprintf(stdout,
403                 "\t[U]     hide user symbols.               \t(%s)\n",
404                 top->hide_user_symbols ? "yes" : "no");
405         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
406         fprintf(stdout, "\t[qQ]    quit.\n");
407 }
408
409 static int perf_top__key_mapped(struct perf_top *top, int c)
410 {
411         switch (c) {
412                 case 'd':
413                 case 'e':
414                 case 'f':
415                 case 'z':
416                 case 'q':
417                 case 'Q':
418                 case 'K':
419                 case 'U':
420                 case 'F':
421                 case 's':
422                 case 'S':
423                         return 1;
424                 case 'E':
425                         return top->evlist->nr_entries > 1 ? 1 : 0;
426                 default:
427                         break;
428         }
429
430         return 0;
431 }
432
433 static bool perf_top__handle_keypress(struct perf_top *top, int c)
434 {
435         bool ret = true;
436
437         if (!perf_top__key_mapped(top, c)) {
438                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
439                 struct termios save;
440
441                 perf_top__print_mapped_keys(top);
442                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
443                 fflush(stdout);
444
445                 set_term_quiet_input(&save);
446
447                 poll(&stdin_poll, 1, -1);
448                 c = getc(stdin);
449
450                 tcsetattr(0, TCSAFLUSH, &save);
451                 if (!perf_top__key_mapped(top, c))
452                         return ret;
453         }
454
455         switch (c) {
456                 case 'd':
457                         prompt_integer(&top->delay_secs, "Enter display delay");
458                         if (top->delay_secs < 1)
459                                 top->delay_secs = 1;
460                         break;
461                 case 'e':
462                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
463                         if (top->print_entries == 0) {
464                                 struct sigaction act = {
465                                         .sa_sigaction = perf_top__sig_winch,
466                                         .sa_flags     = SA_SIGINFO,
467                                 };
468                                 perf_top__sig_winch(SIGWINCH, NULL, top);
469                                 sigaction(SIGWINCH, &act, NULL);
470                         } else {
471                                 signal(SIGWINCH, SIG_DFL);
472                         }
473                         break;
474                 case 'E':
475                         if (top->evlist->nr_entries > 1) {
476                                 /* Select 0 as the default event: */
477                                 int counter = 0;
478
479                                 fprintf(stderr, "\nAvailable events:");
480
481                                 evlist__for_each(top->evlist, top->sym_evsel)
482                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
483
484                                 prompt_integer(&counter, "Enter details event counter");
485
486                                 if (counter >= top->evlist->nr_entries) {
487                                         top->sym_evsel = perf_evlist__first(top->evlist);
488                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
489                                         sleep(1);
490                                         break;
491                                 }
492                                 evlist__for_each(top->evlist, top->sym_evsel)
493                                         if (top->sym_evsel->idx == counter)
494                                                 break;
495                         } else
496                                 top->sym_evsel = perf_evlist__first(top->evlist);
497                         break;
498                 case 'f':
499                         prompt_integer(&top->count_filter, "Enter display event count filter");
500                         break;
501                 case 'F':
502                         prompt_percent(&top->sym_pcnt_filter,
503                                        "Enter details display event filter (percent)");
504                         break;
505                 case 'K':
506                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
507                         break;
508                 case 'q':
509                 case 'Q':
510                         printf("exiting.\n");
511                         if (top->dump_symtab)
512                                 perf_session__fprintf_dsos(top->session, stderr);
513                         ret = false;
514                         break;
515                 case 's':
516                         perf_top__prompt_symbol(top, "Enter details symbol");
517                         break;
518                 case 'S':
519                         if (!top->sym_filter_entry)
520                                 break;
521                         else {
522                                 struct hist_entry *syme = top->sym_filter_entry;
523
524                                 top->sym_filter_entry = NULL;
525                                 __zero_source_counters(syme);
526                         }
527                         break;
528                 case 'U':
529                         top->hide_user_symbols = !top->hide_user_symbols;
530                         break;
531                 case 'z':
532                         top->zero = !top->zero;
533                         break;
534                 default:
535                         break;
536         }
537
538         return ret;
539 }
540
541 static void perf_top__sort_new_samples(void *arg)
542 {
543         struct perf_top *t = arg;
544         struct perf_evsel *evsel = t->sym_evsel;
545         struct hists *hists;
546
547         perf_top__reset_sample_counters(t);
548
549         if (t->evlist->selected != NULL)
550                 t->sym_evsel = t->evlist->selected;
551
552         hists = evsel__hists(evsel);
553
554         if (t->evlist->enabled) {
555                 if (t->zero) {
556                         hists__delete_entries(hists);
557                 } else {
558                         hists__decay_entries(hists, t->hide_user_symbols,
559                                              t->hide_kernel_symbols);
560                 }
561         }
562
563         hists__collapse_resort(hists, NULL);
564         perf_evsel__output_resort(evsel, NULL);
565 }
566
567 static void *display_thread_tui(void *arg)
568 {
569         struct perf_evsel *pos;
570         struct perf_top *top = arg;
571         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
572         struct hist_browser_timer hbt = {
573                 .timer          = perf_top__sort_new_samples,
574                 .arg            = top,
575                 .refresh        = top->delay_secs,
576         };
577
578         perf_top__sort_new_samples(top);
579
580         /*
581          * Initialize the uid_filter_str, in the future the TUI will allow
582          * Zooming in/out UIDs. For now juse use whatever the user passed
583          * via --uid.
584          */
585         evlist__for_each(top->evlist, pos) {
586                 struct hists *hists = evsel__hists(pos);
587                 hists->uid_filter_str = top->record_opts.target.uid_str;
588         }
589
590         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
591                                       top->min_percent,
592                                       &top->session->header.env);
593
594         done = 1;
595         return NULL;
596 }
597
598 static void display_sig(int sig __maybe_unused)
599 {
600         done = 1;
601 }
602
603 static void display_setup_sig(void)
604 {
605         signal(SIGSEGV, sighandler_dump_stack);
606         signal(SIGFPE, sighandler_dump_stack);
607         signal(SIGINT,  display_sig);
608         signal(SIGQUIT, display_sig);
609         signal(SIGTERM, display_sig);
610 }
611
612 static void *display_thread(void *arg)
613 {
614         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
615         struct termios save;
616         struct perf_top *top = arg;
617         int delay_msecs, c;
618
619         display_setup_sig();
620         pthread__unblock_sigwinch();
621 repeat:
622         delay_msecs = top->delay_secs * 1000;
623         set_term_quiet_input(&save);
624         /* trash return*/
625         getc(stdin);
626
627         while (!done) {
628                 perf_top__print_sym_table(top);
629                 /*
630                  * Either timeout expired or we got an EINTR due to SIGWINCH,
631                  * refresh screen in both cases.
632                  */
633                 switch (poll(&stdin_poll, 1, delay_msecs)) {
634                 case 0:
635                         continue;
636                 case -1:
637                         if (errno == EINTR)
638                                 continue;
639                         /* Fall trhu */
640                 default:
641                         c = getc(stdin);
642                         tcsetattr(0, TCSAFLUSH, &save);
643
644                         if (perf_top__handle_keypress(top, c))
645                                 goto repeat;
646                         done = 1;
647                 }
648         }
649
650         tcsetattr(0, TCSAFLUSH, &save);
651         return NULL;
652 }
653
654 static int symbol_filter(struct map *map, struct symbol *sym)
655 {
656         const char *name = sym->name;
657
658         if (!__map__is_kernel(map))
659                 return 0;
660         /*
661          * ppc64 uses function descriptors and appends a '.' to the
662          * start of every instruction address. Remove it.
663          */
664         if (name[0] == '.')
665                 name++;
666
667         if (!strcmp(name, "_text") ||
668             !strcmp(name, "_etext") ||
669             !strcmp(name, "_sinittext") ||
670             !strncmp("init_module", name, 11) ||
671             !strncmp("cleanup_module", name, 14) ||
672             strstr(name, "_text_start") ||
673             strstr(name, "_text_end"))
674                 return 1;
675
676         if (symbol__is_idle(sym))
677                 sym->ignore = true;
678
679         return 0;
680 }
681
682 static int hist_iter__top_callback(struct hist_entry_iter *iter,
683                                    struct addr_location *al, bool single,
684                                    void *arg)
685 {
686         struct perf_top *top = arg;
687         struct hist_entry *he = iter->he;
688         struct perf_evsel *evsel = iter->evsel;
689
690         if (sort__has_sym && single)
691                 perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
692
693         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
694                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
695         return 0;
696 }
697
698 static void perf_event__process_sample(struct perf_tool *tool,
699                                        const union perf_event *event,
700                                        struct perf_evsel *evsel,
701                                        struct perf_sample *sample,
702                                        struct machine *machine)
703 {
704         struct perf_top *top = container_of(tool, struct perf_top, tool);
705         struct addr_location al;
706         int err;
707
708         if (!machine && perf_guest) {
709                 static struct intlist *seen;
710
711                 if (!seen)
712                         seen = intlist__new(NULL);
713
714                 if (!intlist__has_entry(seen, sample->pid)) {
715                         pr_err("Can't find guest [%d]'s kernel information\n",
716                                 sample->pid);
717                         intlist__add(seen, sample->pid);
718                 }
719                 return;
720         }
721
722         if (!machine) {
723                 pr_err("%u unprocessable samples recorded.\r",
724                        top->session->evlist->stats.nr_unprocessable_samples++);
725                 return;
726         }
727
728         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
729                 top->exact_samples++;
730
731         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0)
732                 return;
733
734         if (!top->kptr_restrict_warned &&
735             symbol_conf.kptr_restrict &&
736             al.cpumode == PERF_RECORD_MISC_KERNEL) {
737                 ui__warning(
738 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
739 "Check /proc/sys/kernel/kptr_restrict.\n\n"
740 "Kernel%s samples will not be resolved.\n",
741                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
742                           " modules" : "");
743                 if (use_browser <= 0)
744                         sleep(5);
745                 top->kptr_restrict_warned = true;
746         }
747
748         if (al.sym == NULL) {
749                 const char *msg = "Kernel samples will not be resolved.\n";
750                 /*
751                  * As we do lazy loading of symtabs we only will know if the
752                  * specified vmlinux file is invalid when we actually have a
753                  * hit in kernel space and then try to load it. So if we get
754                  * here and there are _no_ symbols in the DSO backing the
755                  * kernel map, bail out.
756                  *
757                  * We may never get here, for instance, if we use -K/
758                  * --hide-kernel-symbols, even if the user specifies an
759                  * invalid --vmlinux ;-)
760                  */
761                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
762                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
763                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
764                         if (symbol_conf.vmlinux_name) {
765                                 char serr[256];
766                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
767                                 ui__warning("The %s file can't be used: %s\n%s",
768                                             symbol_conf.vmlinux_name, serr, msg);
769                         } else {
770                                 ui__warning("A vmlinux file was not found.\n%s",
771                                             msg);
772                         }
773
774                         if (use_browser <= 0)
775                                 sleep(5);
776                         top->vmlinux_warned = true;
777                 }
778         }
779
780         if (al.sym == NULL || !al.sym->ignore) {
781                 struct hists *hists = evsel__hists(evsel);
782                 struct hist_entry_iter iter = {
783                         .evsel          = evsel,
784                         .sample         = sample,
785                         .add_entry_cb   = hist_iter__top_callback,
786                 };
787
788                 if (symbol_conf.cumulate_callchain)
789                         iter.ops = &hist_iter_cumulative;
790                 else
791                         iter.ops = &hist_iter_normal;
792
793                 pthread_mutex_lock(&hists->lock);
794
795                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
796                 if (err < 0)
797                         pr_err("Problem incrementing symbol period, skipping event\n");
798
799                 pthread_mutex_unlock(&hists->lock);
800         }
801
802         addr_location__put(&al);
803 }
804
805 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
806 {
807         struct perf_sample sample;
808         struct perf_evsel *evsel;
809         struct perf_session *session = top->session;
810         union perf_event *event;
811         struct machine *machine;
812         u8 origin;
813         int ret;
814
815         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
816                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
817                 if (ret) {
818                         pr_err("Can't parse sample, err = %d\n", ret);
819                         goto next_event;
820                 }
821
822                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
823                 assert(evsel != NULL);
824
825                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
826
827                 if (event->header.type == PERF_RECORD_SAMPLE)
828                         ++top->samples;
829
830                 switch (origin) {
831                 case PERF_RECORD_MISC_USER:
832                         ++top->us_samples;
833                         if (top->hide_user_symbols)
834                                 goto next_event;
835                         machine = &session->machines.host;
836                         break;
837                 case PERF_RECORD_MISC_KERNEL:
838                         ++top->kernel_samples;
839                         if (top->hide_kernel_symbols)
840                                 goto next_event;
841                         machine = &session->machines.host;
842                         break;
843                 case PERF_RECORD_MISC_GUEST_KERNEL:
844                         ++top->guest_kernel_samples;
845                         machine = perf_session__find_machine(session,
846                                                              sample.pid);
847                         break;
848                 case PERF_RECORD_MISC_GUEST_USER:
849                         ++top->guest_us_samples;
850                         /*
851                          * TODO: we don't process guest user from host side
852                          * except simple counting.
853                          */
854                         goto next_event;
855                 default:
856                         if (event->header.type == PERF_RECORD_SAMPLE)
857                                 goto next_event;
858                         machine = &session->machines.host;
859                         break;
860                 }
861
862
863                 if (event->header.type == PERF_RECORD_SAMPLE) {
864                         perf_event__process_sample(&top->tool, event, evsel,
865                                                    &sample, machine);
866                 } else if (event->header.type < PERF_RECORD_MAX) {
867                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
868                         machine__process_event(machine, event, &sample);
869                 } else
870                         ++session->evlist->stats.nr_unknown_events;
871 next_event:
872                 perf_evlist__mmap_consume(top->evlist, idx);
873         }
874 }
875
876 static void perf_top__mmap_read(struct perf_top *top)
877 {
878         int i;
879
880         for (i = 0; i < top->evlist->nr_mmaps; i++)
881                 perf_top__mmap_read_idx(top, i);
882 }
883
884 static int perf_top__start_counters(struct perf_top *top)
885 {
886         char msg[512];
887         struct perf_evsel *counter;
888         struct perf_evlist *evlist = top->evlist;
889         struct record_opts *opts = &top->record_opts;
890
891         perf_evlist__config(evlist, opts);
892
893         evlist__for_each(evlist, counter) {
894 try_again:
895                 if (perf_evsel__open(counter, top->evlist->cpus,
896                                      top->evlist->threads) < 0) {
897                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
898                                 if (verbose)
899                                         ui__warning("%s\n", msg);
900                                 goto try_again;
901                         }
902
903                         perf_evsel__open_strerror(counter, &opts->target,
904                                                   errno, msg, sizeof(msg));
905                         ui__error("%s\n", msg);
906                         goto out_err;
907                 }
908         }
909
910         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
911                 ui__error("Failed to mmap with %d (%s)\n",
912                             errno, strerror_r(errno, msg, sizeof(msg)));
913                 goto out_err;
914         }
915
916         return 0;
917
918 out_err:
919         return -1;
920 }
921
922 static int perf_top__setup_sample_type(struct perf_top *top __maybe_unused)
923 {
924         if (!sort__has_sym) {
925                 if (symbol_conf.use_callchain) {
926                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
927                         return -EINVAL;
928                 }
929         } else if (callchain_param.mode != CHAIN_NONE) {
930                 if (callchain_register_param(&callchain_param) < 0) {
931                         ui__error("Can't register callchain params.\n");
932                         return -EINVAL;
933                 }
934         }
935
936         return 0;
937 }
938
939 static int __cmd_top(struct perf_top *top)
940 {
941         struct record_opts *opts = &top->record_opts;
942         pthread_t thread;
943         int ret;
944
945         top->session = perf_session__new(NULL, false, NULL);
946         if (top->session == NULL)
947                 return -1;
948
949         machines__set_symbol_filter(&top->session->machines, symbol_filter);
950
951         if (!objdump_path) {
952                 ret = perf_env__lookup_objdump(&top->session->header.env);
953                 if (ret)
954                         goto out_delete;
955         }
956
957         ret = perf_top__setup_sample_type(top);
958         if (ret)
959                 goto out_delete;
960
961         if (perf_session__register_idle_thread(top->session) < 0)
962                 goto out_delete;
963
964         machine__synthesize_threads(&top->session->machines.host, &opts->target,
965                                     top->evlist->threads, false, opts->proc_map_timeout);
966
967         if (sort__has_socket) {
968                 ret = perf_env__read_cpu_topology_map(&perf_env);
969                 if (ret < 0)
970                         goto out_err_cpu_topo;
971         }
972
973         ret = perf_top__start_counters(top);
974         if (ret)
975                 goto out_delete;
976
977         top->session->evlist = top->evlist;
978         perf_session__set_id_hdr_size(top->session);
979
980         /*
981          * When perf is starting the traced process, all the events (apart from
982          * group members) have enable_on_exec=1 set, so don't spoil it by
983          * prematurely enabling them.
984          *
985          * XXX 'top' still doesn't start workloads like record, trace, but should,
986          * so leave the check here.
987          */
988         if (!target__none(&opts->target))
989                 perf_evlist__enable(top->evlist);
990
991         /* Wait for a minimal set of events before starting the snapshot */
992         perf_evlist__poll(top->evlist, 100);
993
994         perf_top__mmap_read(top);
995
996         ret = -1;
997         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
998                                                             display_thread), top)) {
999                 ui__error("Could not create display thread.\n");
1000                 goto out_delete;
1001         }
1002
1003         if (top->realtime_prio) {
1004                 struct sched_param param;
1005
1006                 param.sched_priority = top->realtime_prio;
1007                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1008                         ui__error("Could not set realtime priority.\n");
1009                         goto out_join;
1010                 }
1011         }
1012
1013         while (!done) {
1014                 u64 hits = top->samples;
1015
1016                 perf_top__mmap_read(top);
1017
1018                 if (hits == top->samples)
1019                         ret = perf_evlist__poll(top->evlist, 100);
1020         }
1021
1022         ret = 0;
1023 out_join:
1024         pthread_join(thread, NULL);
1025 out_delete:
1026         perf_session__delete(top->session);
1027         top->session = NULL;
1028
1029         return ret;
1030
1031 out_err_cpu_topo: {
1032         char errbuf[BUFSIZ];
1033         const char *err = strerror_r(-ret, errbuf, sizeof(errbuf));
1034
1035         ui__error("Could not read the CPU topology map: %s\n", err);
1036         goto out_delete;
1037 }
1038 }
1039
1040 static int
1041 callchain_opt(const struct option *opt, const char *arg, int unset)
1042 {
1043         symbol_conf.use_callchain = true;
1044         return record_callchain_opt(opt, arg, unset);
1045 }
1046
1047 static int
1048 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1049 {
1050         struct record_opts *record = (struct record_opts *)opt->value;
1051
1052         record->callgraph_set = true;
1053         callchain_param.enabled = !unset;
1054         callchain_param.record_mode = CALLCHAIN_FP;
1055
1056         /*
1057          * --no-call-graph
1058          */
1059         if (unset) {
1060                 symbol_conf.use_callchain = false;
1061                 callchain_param.record_mode = CALLCHAIN_NONE;
1062                 return 0;
1063         }
1064
1065         return parse_callchain_top_opt(arg);
1066 }
1067
1068 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1069 {
1070         if (!strcmp(var, "top.call-graph"))
1071                 var = "call-graph.record-mode"; /* fall-through */
1072         if (!strcmp(var, "top.children")) {
1073                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1074                 return 0;
1075         }
1076
1077         return 0;
1078 }
1079
1080 static int
1081 parse_percent_limit(const struct option *opt, const char *arg,
1082                     int unset __maybe_unused)
1083 {
1084         struct perf_top *top = opt->value;
1085
1086         top->min_percent = strtof(arg, NULL);
1087         return 0;
1088 }
1089
1090 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1091         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1092
1093 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1094 {
1095         char errbuf[BUFSIZ];
1096         struct perf_top top = {
1097                 .count_filter        = 5,
1098                 .delay_secs          = 2,
1099                 .record_opts = {
1100                         .mmap_pages     = UINT_MAX,
1101                         .user_freq      = UINT_MAX,
1102                         .user_interval  = ULLONG_MAX,
1103                         .freq           = 4000, /* 4 KHz */
1104                         .target         = {
1105                                 .uses_mmap   = true,
1106                         },
1107                         .proc_map_timeout    = 500,
1108                 },
1109                 .max_stack           = PERF_MAX_STACK_DEPTH,
1110                 .sym_pcnt_filter     = 5,
1111         };
1112         struct record_opts *opts = &top.record_opts;
1113         struct target *target = &opts->target;
1114         const struct option options[] = {
1115         OPT_CALLBACK('e', "event", &top.evlist, "event",
1116                      "event selector. use 'perf list' to list available events",
1117                      parse_events_option),
1118         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1119         OPT_STRING('p', "pid", &target->pid, "pid",
1120                     "profile events on existing process id"),
1121         OPT_STRING('t', "tid", &target->tid, "tid",
1122                     "profile events on existing thread id"),
1123         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1124                             "system-wide collection from all CPUs"),
1125         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1126                     "list of cpus to monitor"),
1127         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1128                    "file", "vmlinux pathname"),
1129         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1130                     "don't load vmlinux even if found"),
1131         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1132                     "hide kernel symbols"),
1133         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1134                      "number of mmap data pages",
1135                      perf_evlist__parse_mmap_pages),
1136         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1137                     "collect data with this RT SCHED_FIFO priority"),
1138         OPT_INTEGER('d', "delay", &top.delay_secs,
1139                     "number of seconds to delay between refreshes"),
1140         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1141                             "dump the symbol table used for profiling"),
1142         OPT_INTEGER('f', "count-filter", &top.count_filter,
1143                     "only display functions with more events than this"),
1144         OPT_BOOLEAN(0, "group", &opts->group,
1145                             "put the counters into a counter group"),
1146         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1147                     "child tasks do not inherit counters"),
1148         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1149                     "symbol to annotate"),
1150         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1151         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1152         OPT_INTEGER('E', "entries", &top.print_entries,
1153                     "display this many functions"),
1154         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1155                     "hide user symbols"),
1156         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1157         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1158         OPT_INCR('v', "verbose", &verbose,
1159                     "be more verbose (show counter open errors, etc)"),
1160         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1161                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1162                    " Please refer the man page for the complete list."),
1163         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1164                    "output field(s): overhead, period, sample plus all of sort keys"),
1165         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1166                     "Show a column with the number of samples"),
1167         OPT_CALLBACK_NOOPT('g', NULL, &top.record_opts,
1168                            NULL, "enables call-graph recording and display",
1169                            &callchain_opt),
1170         OPT_CALLBACK(0, "call-graph", &top.record_opts,
1171                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1172                      top_callchain_help, &parse_callchain_opt),
1173         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1174                     "Accumulate callchains of children and show total overhead as well"),
1175         OPT_INTEGER(0, "max-stack", &top.max_stack,
1176                     "Set the maximum stack depth when parsing the callchain. "
1177                     "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
1178         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1179                    "ignore callees of these functions in call graphs",
1180                    report_parse_ignore_callees_opt),
1181         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1182                     "Show a column with the sum of periods"),
1183         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1184                    "only consider symbols in these dsos"),
1185         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1186                    "only consider symbols in these comms"),
1187         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1188                    "only consider these symbols"),
1189         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1190                     "Interleave source code with assembly code (default)"),
1191         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1192                     "Display raw encoding of assembly instructions (default)"),
1193         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1194                     "Enable kernel symbol demangling"),
1195         OPT_STRING(0, "objdump", &objdump_path, "path",
1196                     "objdump binary to use for disassembly and annotations"),
1197         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1198                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1199         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1200         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1201                      "Don't show entries under that percent", parse_percent_limit),
1202         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1203                      "How to display percentage of filtered entries", parse_filter_percentage),
1204         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1205                    "width[,width...]",
1206                    "don't try to adjust column width, use these fixed values"),
1207         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1208                         "per thread proc mmap processing timeout in ms"),
1209         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1210                      "branch any", "sample any taken branches",
1211                      parse_branch_stack),
1212         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1213                      "branch filter mask", "branch stack filter modes",
1214                      parse_branch_stack),
1215         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1216                     "Show raw trace event output (do not use print fmt or plugins)"),
1217         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1218                     "Show entries in a hierarchy"),
1219         OPT_END()
1220         };
1221         const char * const top_usage[] = {
1222                 "perf top [<options>]",
1223                 NULL
1224         };
1225         int status = hists__init();
1226
1227         if (status < 0)
1228                 return status;
1229
1230         top.evlist = perf_evlist__new();
1231         if (top.evlist == NULL)
1232                 return -ENOMEM;
1233
1234         perf_config(perf_top_config, &top);
1235
1236         argc = parse_options(argc, argv, options, top_usage, 0);
1237         if (argc)
1238                 usage_with_options(top_usage, options);
1239
1240         if (!top.evlist->nr_entries &&
1241             perf_evlist__add_default(top.evlist) < 0) {
1242                 pr_err("Not enough memory for event selector list\n");
1243                 goto out_delete_evlist;
1244         }
1245
1246         if (symbol_conf.report_hierarchy) {
1247                 /* disable incompatible options */
1248                 symbol_conf.event_group = false;
1249                 symbol_conf.cumulate_callchain = false;
1250
1251                 if (field_order) {
1252                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1253                         parse_options_usage(top_usage, options, "fields", 0);
1254                         parse_options_usage(NULL, options, "hierarchy", 0);
1255                         goto out_delete_evlist;
1256                 }
1257         }
1258
1259         sort__mode = SORT_MODE__TOP;
1260         /* display thread wants entries to be collapsed in a different tree */
1261         sort__need_collapse = 1;
1262
1263         if (top.use_stdio)
1264                 use_browser = 0;
1265         else if (top.use_tui)
1266                 use_browser = 1;
1267
1268         setup_browser(false);
1269
1270         if (setup_sorting(top.evlist) < 0) {
1271                 if (sort_order)
1272                         parse_options_usage(top_usage, options, "s", 1);
1273                 if (field_order)
1274                         parse_options_usage(sort_order ? NULL : top_usage,
1275                                             options, "fields", 0);
1276                 goto out_delete_evlist;
1277         }
1278
1279         status = target__validate(target);
1280         if (status) {
1281                 target__strerror(target, status, errbuf, BUFSIZ);
1282                 ui__warning("%s\n", errbuf);
1283         }
1284
1285         status = target__parse_uid(target);
1286         if (status) {
1287                 int saved_errno = errno;
1288
1289                 target__strerror(target, status, errbuf, BUFSIZ);
1290                 ui__error("%s\n", errbuf);
1291
1292                 status = -saved_errno;
1293                 goto out_delete_evlist;
1294         }
1295
1296         if (target__none(target))
1297                 target->system_wide = true;
1298
1299         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1300                 ui__error("Couldn't create thread/CPU maps: %s\n",
1301                           errno == ENOENT ? "No such process" : strerror_r(errno, errbuf, sizeof(errbuf)));
1302                 goto out_delete_evlist;
1303         }
1304
1305         symbol_conf.nr_events = top.evlist->nr_entries;
1306
1307         if (top.delay_secs < 1)
1308                 top.delay_secs = 1;
1309
1310         if (record_opts__config(opts)) {
1311                 status = -EINVAL;
1312                 goto out_delete_evlist;
1313         }
1314
1315         top.sym_evsel = perf_evlist__first(top.evlist);
1316
1317         if (!symbol_conf.use_callchain) {
1318                 symbol_conf.cumulate_callchain = false;
1319                 perf_hpp__cancel_cumulate();
1320         }
1321
1322         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1323                 callchain_param.order = ORDER_CALLER;
1324
1325         symbol_conf.priv_size = sizeof(struct annotation);
1326
1327         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1328         if (symbol__init(NULL) < 0)
1329                 return -1;
1330
1331         sort__setup_elide(stdout);
1332
1333         get_term_dimensions(&top.winsize);
1334         if (top.print_entries == 0) {
1335                 struct sigaction act = {
1336                         .sa_sigaction = perf_top__sig_winch,
1337                         .sa_flags     = SA_SIGINFO,
1338                 };
1339                 perf_top__update_print_entries(&top);
1340                 sigaction(SIGWINCH, &act, NULL);
1341         }
1342
1343         status = __cmd_top(&top);
1344
1345 out_delete_evlist:
1346         perf_evlist__delete(top.evlist);
1347
1348         return status;
1349 }