cfg80211: handle failed skb allocation
[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/stringify.h>
71 #include <linux/types.h>
72
73 static volatile int done;
74
75 #define HEADER_LINE_NR  5
76
77 static void perf_top__update_print_entries(struct perf_top *top)
78 {
79         top->print_entries = top->winsize.ws_row - HEADER_LINE_NR;
80 }
81
82 static void perf_top__sig_winch(int sig __maybe_unused,
83                                 siginfo_t *info __maybe_unused, void *arg)
84 {
85         struct perf_top *top = arg;
86
87         get_term_dimensions(&top->winsize);
88         perf_top__update_print_entries(top);
89 }
90
91 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
92 {
93         struct symbol *sym;
94         struct annotation *notes;
95         struct map *map;
96         int err = -1;
97
98         if (!he || !he->ms.sym)
99                 return -1;
100
101         sym = he->ms.sym;
102         map = he->ms.map;
103
104         /*
105          * We can't annotate with just /proc/kallsyms
106          */
107         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
108             !dso__is_kcore(map->dso)) {
109                 pr_err("Can't annotate %s: No vmlinux file was found in the "
110                        "path\n", sym->name);
111                 sleep(1);
112                 return -1;
113         }
114
115         notes = symbol__annotation(sym);
116         if (notes->src != NULL) {
117                 pthread_mutex_lock(&notes->lock);
118                 goto out_assign;
119         }
120
121         pthread_mutex_lock(&notes->lock);
122
123         if (symbol__alloc_hist(sym) < 0) {
124                 pthread_mutex_unlock(&notes->lock);
125                 pr_err("Not enough memory for annotating '%s' symbol!\n",
126                        sym->name);
127                 sleep(1);
128                 return err;
129         }
130
131         err = symbol__annotate(sym, map, 0);
132         if (err == 0) {
133 out_assign:
134                 top->sym_filter_entry = he;
135         }
136
137         pthread_mutex_unlock(&notes->lock);
138         return err;
139 }
140
141 static void __zero_source_counters(struct hist_entry *he)
142 {
143         struct symbol *sym = he->ms.sym;
144         symbol__annotate_zero_histograms(sym);
145 }
146
147 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
148 {
149         struct utsname uts;
150         int err = uname(&uts);
151
152         ui__warning("Out of bounds address found:\n\n"
153                     "Addr:   %" PRIx64 "\n"
154                     "DSO:    %s %c\n"
155                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
156                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
157                     "Arch:   %s\n"
158                     "Kernel: %s\n"
159                     "Tools:  %s\n\n"
160                     "Not all samples will be on the annotation output.\n\n"
161                     "Please report to linux-kernel@vger.kernel.org\n",
162                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
163                     map->start, map->end, sym->start, sym->end,
164                     sym->binding == STB_GLOBAL ? 'g' :
165                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
166                     err ? "[unknown]" : uts.machine,
167                     err ? "[unknown]" : uts.release, perf_version_string);
168         if (use_browser <= 0)
169                 sleep(5);
170
171         map->erange_warned = true;
172 }
173
174 static void perf_top__record_precise_ip(struct perf_top *top,
175                                         struct hist_entry *he,
176                                         int counter, u64 ip)
177 {
178         struct annotation *notes;
179         struct symbol *sym = he->ms.sym;
180         int err = 0;
181
182         if (sym == NULL || (use_browser == 0 &&
183                             (top->sym_filter_entry == NULL ||
184                              top->sym_filter_entry->ms.sym != sym)))
185                 return;
186
187         notes = symbol__annotation(sym);
188
189         if (pthread_mutex_trylock(&notes->lock))
190                 return;
191
192         err = hist_entry__inc_addr_samples(he, counter, ip);
193
194         pthread_mutex_unlock(&notes->lock);
195
196         if (unlikely(err)) {
197                 /*
198                  * This function is now called with he->hists->lock held.
199                  * Release it before going to sleep.
200                  */
201                 pthread_mutex_unlock(&he->hists->lock);
202
203                 if (err == -ERANGE && !he->ms.map->erange_warned)
204                         ui__warn_map_erange(he->ms.map, sym, ip);
205                 else if (err == -ENOMEM) {
206                         pr_err("Not enough memory for annotating '%s' symbol!\n",
207                                sym->name);
208                         sleep(1);
209                 }
210
211                 pthread_mutex_lock(&he->hists->lock);
212         }
213 }
214
215 static void perf_top__show_details(struct perf_top *top)
216 {
217         struct hist_entry *he = top->sym_filter_entry;
218         struct annotation *notes;
219         struct symbol *symbol;
220         int more;
221
222         if (!he)
223                 return;
224
225         symbol = he->ms.sym;
226         notes = symbol__annotation(symbol);
227
228         pthread_mutex_lock(&notes->lock);
229
230         if (notes->src == NULL)
231                 goto out_unlock;
232
233         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
234         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
235
236         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel,
237                                        0, top->sym_pcnt_filter, top->print_entries, 4);
238
239         if (top->evlist->enabled) {
240                 if (top->zero)
241                         symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
242                 else
243                         symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
244         }
245         if (more != 0)
246                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
247 out_unlock:
248         pthread_mutex_unlock(&notes->lock);
249 }
250
251 static void perf_top__print_sym_table(struct perf_top *top)
252 {
253         char bf[160];
254         int printed = 0;
255         const int win_width = top->winsize.ws_col - 1;
256         struct perf_evsel *evsel = top->sym_evsel;
257         struct hists *hists = evsel__hists(evsel);
258
259         puts(CONSOLE_CLEAR);
260
261         perf_top__header_snprintf(top, bf, sizeof(bf));
262         printf("%s\n", bf);
263
264         perf_top__reset_sample_counters(top);
265
266         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
267
268         if (hists->stats.nr_lost_warned !=
269             hists->stats.nr_events[PERF_RECORD_LOST]) {
270                 hists->stats.nr_lost_warned =
271                               hists->stats.nr_events[PERF_RECORD_LOST];
272                 color_fprintf(stdout, PERF_COLOR_RED,
273                               "WARNING: LOST %d chunks, Check IO/CPU overload",
274                               hists->stats.nr_lost_warned);
275                 ++printed;
276         }
277
278         if (top->sym_filter_entry) {
279                 perf_top__show_details(top);
280                 return;
281         }
282
283         if (top->evlist->enabled) {
284                 if (top->zero) {
285                         hists__delete_entries(hists);
286                 } else {
287                         hists__decay_entries(hists, top->hide_user_symbols,
288                                              top->hide_kernel_symbols);
289                 }
290         }
291
292         hists__collapse_resort(hists, NULL);
293         perf_evsel__output_resort(evsel, NULL);
294
295         hists__output_recalc_col_len(hists, top->print_entries - printed);
296         putchar('\n');
297         hists__fprintf(hists, false, top->print_entries - printed, win_width,
298                        top->min_percent, stdout);
299 }
300
301 static void prompt_integer(int *target, const char *msg)
302 {
303         char *buf = malloc(0), *p;
304         size_t dummy = 0;
305         int tmp;
306
307         fprintf(stdout, "\n%s: ", msg);
308         if (getline(&buf, &dummy, stdin) < 0)
309                 return;
310
311         p = strchr(buf, '\n');
312         if (p)
313                 *p = 0;
314
315         p = buf;
316         while(*p) {
317                 if (!isdigit(*p))
318                         goto out_free;
319                 p++;
320         }
321         tmp = strtoul(buf, NULL, 10);
322         *target = tmp;
323 out_free:
324         free(buf);
325 }
326
327 static void prompt_percent(int *target, const char *msg)
328 {
329         int tmp = 0;
330
331         prompt_integer(&tmp, msg);
332         if (tmp >= 0 && tmp <= 100)
333                 *target = tmp;
334 }
335
336 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
337 {
338         char *buf = malloc(0), *p;
339         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
340         struct hists *hists = evsel__hists(top->sym_evsel);
341         struct rb_node *next;
342         size_t dummy = 0;
343
344         /* zero counters of active symbol */
345         if (syme) {
346                 __zero_source_counters(syme);
347                 top->sym_filter_entry = NULL;
348         }
349
350         fprintf(stdout, "\n%s: ", msg);
351         if (getline(&buf, &dummy, stdin) < 0)
352                 goto out_free;
353
354         p = strchr(buf, '\n');
355         if (p)
356                 *p = 0;
357
358         next = rb_first(&hists->entries);
359         while (next) {
360                 n = rb_entry(next, struct hist_entry, rb_node);
361                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
362                         found = n;
363                         break;
364                 }
365                 next = rb_next(&n->rb_node);
366         }
367
368         if (!found) {
369                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
370                 sleep(1);
371         } else
372                 perf_top__parse_source(top, found);
373
374 out_free:
375         free(buf);
376 }
377
378 static void perf_top__print_mapped_keys(struct perf_top *top)
379 {
380         char *name = NULL;
381
382         if (top->sym_filter_entry) {
383                 struct symbol *sym = top->sym_filter_entry->ms.sym;
384                 name = sym->name;
385         }
386
387         fprintf(stdout, "\nMapped keys:\n");
388         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
389         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
390
391         if (top->evlist->nr_entries > 1)
392                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
393
394         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
395
396         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
397         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
398         fprintf(stdout, "\t[S]     stop annotation.\n");
399
400         fprintf(stdout,
401                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
402                 top->hide_kernel_symbols ? "yes" : "no");
403         fprintf(stdout,
404                 "\t[U]     hide user symbols.               \t(%s)\n",
405                 top->hide_user_symbols ? "yes" : "no");
406         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
407         fprintf(stdout, "\t[qQ]    quit.\n");
408 }
409
410 static int perf_top__key_mapped(struct perf_top *top, int c)
411 {
412         switch (c) {
413                 case 'd':
414                 case 'e':
415                 case 'f':
416                 case 'z':
417                 case 'q':
418                 case 'Q':
419                 case 'K':
420                 case 'U':
421                 case 'F':
422                 case 's':
423                 case 'S':
424                         return 1;
425                 case 'E':
426                         return top->evlist->nr_entries > 1 ? 1 : 0;
427                 default:
428                         break;
429         }
430
431         return 0;
432 }
433
434 static bool perf_top__handle_keypress(struct perf_top *top, int c)
435 {
436         bool ret = true;
437
438         if (!perf_top__key_mapped(top, c)) {
439                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
440                 struct termios save;
441
442                 perf_top__print_mapped_keys(top);
443                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
444                 fflush(stdout);
445
446                 set_term_quiet_input(&save);
447
448                 poll(&stdin_poll, 1, -1);
449                 c = getc(stdin);
450
451                 tcsetattr(0, TCSAFLUSH, &save);
452                 if (!perf_top__key_mapped(top, c))
453                         return ret;
454         }
455
456         switch (c) {
457                 case 'd':
458                         prompt_integer(&top->delay_secs, "Enter display delay");
459                         if (top->delay_secs < 1)
460                                 top->delay_secs = 1;
461                         break;
462                 case 'e':
463                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
464                         if (top->print_entries == 0) {
465                                 struct sigaction act = {
466                                         .sa_sigaction = perf_top__sig_winch,
467                                         .sa_flags     = SA_SIGINFO,
468                                 };
469                                 perf_top__sig_winch(SIGWINCH, NULL, top);
470                                 sigaction(SIGWINCH, &act, NULL);
471                         } else {
472                                 signal(SIGWINCH, SIG_DFL);
473                         }
474                         break;
475                 case 'E':
476                         if (top->evlist->nr_entries > 1) {
477                                 /* Select 0 as the default event: */
478                                 int counter = 0;
479
480                                 fprintf(stderr, "\nAvailable events:");
481
482                                 evlist__for_each(top->evlist, top->sym_evsel)
483                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
484
485                                 prompt_integer(&counter, "Enter details event counter");
486
487                                 if (counter >= top->evlist->nr_entries) {
488                                         top->sym_evsel = perf_evlist__first(top->evlist);
489                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
490                                         sleep(1);
491                                         break;
492                                 }
493                                 evlist__for_each(top->evlist, top->sym_evsel)
494                                         if (top->sym_evsel->idx == counter)
495                                                 break;
496                         } else
497                                 top->sym_evsel = perf_evlist__first(top->evlist);
498                         break;
499                 case 'f':
500                         prompt_integer(&top->count_filter, "Enter display event count filter");
501                         break;
502                 case 'F':
503                         prompt_percent(&top->sym_pcnt_filter,
504                                        "Enter details display event filter (percent)");
505                         break;
506                 case 'K':
507                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
508                         break;
509                 case 'q':
510                 case 'Q':
511                         printf("exiting.\n");
512                         if (top->dump_symtab)
513                                 perf_session__fprintf_dsos(top->session, stderr);
514                         ret = false;
515                         break;
516                 case 's':
517                         perf_top__prompt_symbol(top, "Enter details symbol");
518                         break;
519                 case 'S':
520                         if (!top->sym_filter_entry)
521                                 break;
522                         else {
523                                 struct hist_entry *syme = top->sym_filter_entry;
524
525                                 top->sym_filter_entry = NULL;
526                                 __zero_source_counters(syme);
527                         }
528                         break;
529                 case 'U':
530                         top->hide_user_symbols = !top->hide_user_symbols;
531                         break;
532                 case 'z':
533                         top->zero = !top->zero;
534                         break;
535                 default:
536                         break;
537         }
538
539         return ret;
540 }
541
542 static void perf_top__sort_new_samples(void *arg)
543 {
544         struct perf_top *t = arg;
545         struct perf_evsel *evsel = t->sym_evsel;
546         struct hists *hists;
547
548         perf_top__reset_sample_counters(t);
549
550         if (t->evlist->selected != NULL)
551                 t->sym_evsel = t->evlist->selected;
552
553         hists = evsel__hists(evsel);
554
555         if (t->evlist->enabled) {
556                 if (t->zero) {
557                         hists__delete_entries(hists);
558                 } else {
559                         hists__decay_entries(hists, t->hide_user_symbols,
560                                              t->hide_kernel_symbols);
561                 }
562         }
563
564         hists__collapse_resort(hists, NULL);
565         perf_evsel__output_resort(evsel, NULL);
566 }
567
568 static void *display_thread_tui(void *arg)
569 {
570         struct perf_evsel *pos;
571         struct perf_top *top = arg;
572         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
573         struct hist_browser_timer hbt = {
574                 .timer          = perf_top__sort_new_samples,
575                 .arg            = top,
576                 .refresh        = top->delay_secs,
577         };
578
579         perf_top__sort_new_samples(top);
580
581         /*
582          * Initialize the uid_filter_str, in the future the TUI will allow
583          * Zooming in/out UIDs. For now juse use whatever the user passed
584          * via --uid.
585          */
586         evlist__for_each(top->evlist, pos) {
587                 struct hists *hists = evsel__hists(pos);
588                 hists->uid_filter_str = top->record_opts.target.uid_str;
589         }
590
591         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
592                                       top->min_percent,
593                                       &top->session->header.env);
594
595         done = 1;
596         return NULL;
597 }
598
599 static void display_sig(int sig __maybe_unused)
600 {
601         done = 1;
602 }
603
604 static void display_setup_sig(void)
605 {
606         signal(SIGSEGV, sighandler_dump_stack);
607         signal(SIGFPE, sighandler_dump_stack);
608         signal(SIGINT,  display_sig);
609         signal(SIGQUIT, display_sig);
610         signal(SIGTERM, display_sig);
611 }
612
613 static void *display_thread(void *arg)
614 {
615         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
616         struct termios save;
617         struct perf_top *top = arg;
618         int delay_msecs, c;
619
620         display_setup_sig();
621         pthread__unblock_sigwinch();
622 repeat:
623         delay_msecs = top->delay_secs * 1000;
624         set_term_quiet_input(&save);
625         /* trash return*/
626         getc(stdin);
627
628         while (!done) {
629                 perf_top__print_sym_table(top);
630                 /*
631                  * Either timeout expired or we got an EINTR due to SIGWINCH,
632                  * refresh screen in both cases.
633                  */
634                 switch (poll(&stdin_poll, 1, delay_msecs)) {
635                 case 0:
636                         continue;
637                 case -1:
638                         if (errno == EINTR)
639                                 continue;
640                         /* Fall trhu */
641                 default:
642                         c = getc(stdin);
643                         tcsetattr(0, TCSAFLUSH, &save);
644
645                         if (perf_top__handle_keypress(top, c))
646                                 goto repeat;
647                         done = 1;
648                 }
649         }
650
651         tcsetattr(0, TCSAFLUSH, &save);
652         return NULL;
653 }
654
655 static int symbol_filter(struct map *map, struct symbol *sym)
656 {
657         const char *name = sym->name;
658
659         if (!__map__is_kernel(map))
660                 return 0;
661         /*
662          * ppc64 uses function descriptors and appends a '.' to the
663          * start of every instruction address. Remove it.
664          */
665         if (name[0] == '.')
666                 name++;
667
668         if (!strcmp(name, "_text") ||
669             !strcmp(name, "_etext") ||
670             !strcmp(name, "_sinittext") ||
671             !strncmp("init_module", name, 11) ||
672             !strncmp("cleanup_module", name, 14) ||
673             strstr(name, "_text_start") ||
674             strstr(name, "_text_end"))
675                 return 1;
676
677         if (symbol__is_idle(sym))
678                 sym->ignore = true;
679
680         return 0;
681 }
682
683 static int hist_iter__top_callback(struct hist_entry_iter *iter,
684                                    struct addr_location *al, bool single,
685                                    void *arg)
686 {
687         struct perf_top *top = arg;
688         struct hist_entry *he = iter->he;
689         struct perf_evsel *evsel = iter->evsel;
690
691         if (perf_hpp_list.sym && single)
692                 perf_top__record_precise_ip(top, he, evsel->idx, al->addr);
693
694         hist__account_cycles(iter->sample->branch_stack, al, iter->sample,
695                      !(top->record_opts.branch_stack & PERF_SAMPLE_BRANCH_ANY));
696         return 0;
697 }
698
699 static void perf_event__process_sample(struct perf_tool *tool,
700                                        const union perf_event *event,
701                                        struct perf_evsel *evsel,
702                                        struct perf_sample *sample,
703                                        struct machine *machine)
704 {
705         struct perf_top *top = container_of(tool, struct perf_top, tool);
706         struct addr_location al;
707         int err;
708
709         if (!machine && perf_guest) {
710                 static struct intlist *seen;
711
712                 if (!seen)
713                         seen = intlist__new(NULL);
714
715                 if (!intlist__has_entry(seen, sample->pid)) {
716                         pr_err("Can't find guest [%d]'s kernel information\n",
717                                 sample->pid);
718                         intlist__add(seen, sample->pid);
719                 }
720                 return;
721         }
722
723         if (!machine) {
724                 pr_err("%u unprocessable samples recorded.\r",
725                        top->session->evlist->stats.nr_unprocessable_samples++);
726                 return;
727         }
728
729         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
730                 top->exact_samples++;
731
732         if (machine__resolve(machine, &al, sample) < 0)
733                 return;
734
735         if (!top->kptr_restrict_warned &&
736             symbol_conf.kptr_restrict &&
737             al.cpumode == PERF_RECORD_MISC_KERNEL) {
738                 ui__warning(
739 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
740 "Check /proc/sys/kernel/kptr_restrict.\n\n"
741 "Kernel%s samples will not be resolved.\n",
742                           al.map && !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
743                           " modules" : "");
744                 if (use_browser <= 0)
745                         sleep(5);
746                 top->kptr_restrict_warned = true;
747         }
748
749         if (al.sym == NULL) {
750                 const char *msg = "Kernel samples will not be resolved.\n";
751                 /*
752                  * As we do lazy loading of symtabs we only will know if the
753                  * specified vmlinux file is invalid when we actually have a
754                  * hit in kernel space and then try to load it. So if we get
755                  * here and there are _no_ symbols in the DSO backing the
756                  * kernel map, bail out.
757                  *
758                  * We may never get here, for instance, if we use -K/
759                  * --hide-kernel-symbols, even if the user specifies an
760                  * invalid --vmlinux ;-)
761                  */
762                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
763                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
764                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
765                         if (symbol_conf.vmlinux_name) {
766                                 char serr[256];
767                                 dso__strerror_load(al.map->dso, serr, sizeof(serr));
768                                 ui__warning("The %s file can't be used: %s\n%s",
769                                             symbol_conf.vmlinux_name, serr, msg);
770                         } else {
771                                 ui__warning("A vmlinux file was not found.\n%s",
772                                             msg);
773                         }
774
775                         if (use_browser <= 0)
776                                 sleep(5);
777                         top->vmlinux_warned = true;
778                 }
779         }
780
781         if (al.sym == NULL || !al.sym->ignore) {
782                 struct hists *hists = evsel__hists(evsel);
783                 struct hist_entry_iter iter = {
784                         .evsel          = evsel,
785                         .sample         = sample,
786                         .add_entry_cb   = hist_iter__top_callback,
787                 };
788
789                 if (symbol_conf.cumulate_callchain)
790                         iter.ops = &hist_iter_cumulative;
791                 else
792                         iter.ops = &hist_iter_normal;
793
794                 pthread_mutex_lock(&hists->lock);
795
796                 err = hist_entry_iter__add(&iter, &al, top->max_stack, top);
797                 if (err < 0)
798                         pr_err("Problem incrementing symbol period, skipping event\n");
799
800                 pthread_mutex_unlock(&hists->lock);
801         }
802
803         addr_location__put(&al);
804 }
805
806 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
807 {
808         struct perf_sample sample;
809         struct perf_evsel *evsel;
810         struct perf_session *session = top->session;
811         union perf_event *event;
812         struct machine *machine;
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                 if (event->header.type == PERF_RECORD_SAMPLE)
826                         ++top->samples;
827
828                 switch (sample.cpumode) {
829                 case PERF_RECORD_MISC_USER:
830                         ++top->us_samples;
831                         if (top->hide_user_symbols)
832                                 goto next_event;
833                         machine = &session->machines.host;
834                         break;
835                 case PERF_RECORD_MISC_KERNEL:
836                         ++top->kernel_samples;
837                         if (top->hide_kernel_symbols)
838                                 goto next_event;
839                         machine = &session->machines.host;
840                         break;
841                 case PERF_RECORD_MISC_GUEST_KERNEL:
842                         ++top->guest_kernel_samples;
843                         machine = perf_session__find_machine(session,
844                                                              sample.pid);
845                         break;
846                 case PERF_RECORD_MISC_GUEST_USER:
847                         ++top->guest_us_samples;
848                         /*
849                          * TODO: we don't process guest user from host side
850                          * except simple counting.
851                          */
852                         goto next_event;
853                 default:
854                         if (event->header.type == PERF_RECORD_SAMPLE)
855                                 goto next_event;
856                         machine = &session->machines.host;
857                         break;
858                 }
859
860
861                 if (event->header.type == PERF_RECORD_SAMPLE) {
862                         perf_event__process_sample(&top->tool, event, evsel,
863                                                    &sample, machine);
864                 } else if (event->header.type < PERF_RECORD_MAX) {
865                         hists__inc_nr_events(evsel__hists(evsel), event->header.type);
866                         machine__process_event(machine, event, &sample);
867                 } else
868                         ++session->evlist->stats.nr_unknown_events;
869 next_event:
870                 perf_evlist__mmap_consume(top->evlist, idx);
871         }
872 }
873
874 static void perf_top__mmap_read(struct perf_top *top)
875 {
876         int i;
877
878         for (i = 0; i < top->evlist->nr_mmaps; i++)
879                 perf_top__mmap_read_idx(top, i);
880 }
881
882 static int perf_top__start_counters(struct perf_top *top)
883 {
884         char msg[512];
885         struct perf_evsel *counter;
886         struct perf_evlist *evlist = top->evlist;
887         struct record_opts *opts = &top->record_opts;
888
889         perf_evlist__config(evlist, opts, &callchain_param);
890
891         evlist__for_each(evlist, counter) {
892 try_again:
893                 if (perf_evsel__open(counter, top->evlist->cpus,
894                                      top->evlist->threads) < 0) {
895                         if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
896                                 if (verbose)
897                                         ui__warning("%s\n", msg);
898                                 goto try_again;
899                         }
900
901                         perf_evsel__open_strerror(counter, &opts->target,
902                                                   errno, msg, sizeof(msg));
903                         ui__error("%s\n", msg);
904                         goto out_err;
905                 }
906         }
907
908         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
909                 ui__error("Failed to mmap with %d (%s)\n",
910                             errno, strerror_r(errno, msg, sizeof(msg)));
911                 goto out_err;
912         }
913
914         return 0;
915
916 out_err:
917         return -1;
918 }
919
920 static int callchain_param__setup_sample_type(struct callchain_param *callchain)
921 {
922         if (!perf_hpp_list.sym) {
923                 if (callchain->enabled) {
924                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
925                         return -EINVAL;
926                 }
927         } else if (callchain->mode != CHAIN_NONE) {
928                 if (callchain_register_param(callchain) < 0) {
929                         ui__error("Can't register callchain params.\n");
930                         return -EINVAL;
931                 }
932         }
933
934         return 0;
935 }
936
937 static int __cmd_top(struct perf_top *top)
938 {
939         struct record_opts *opts = &top->record_opts;
940         pthread_t thread;
941         int ret;
942
943         top->session = perf_session__new(NULL, false, NULL);
944         if (top->session == NULL)
945                 return -1;
946
947         machines__set_symbol_filter(&top->session->machines, symbol_filter);
948
949         if (!objdump_path) {
950                 ret = perf_env__lookup_objdump(&top->session->header.env);
951                 if (ret)
952                         goto out_delete;
953         }
954
955         ret = callchain_param__setup_sample_type(&callchain_param);
956         if (ret)
957                 goto out_delete;
958
959         if (perf_session__register_idle_thread(top->session) < 0)
960                 goto out_delete;
961
962         machine__synthesize_threads(&top->session->machines.host, &opts->target,
963                                     top->evlist->threads, false, opts->proc_map_timeout);
964
965         if (perf_hpp_list.socket) {
966                 ret = perf_env__read_cpu_topology_map(&perf_env);
967                 if (ret < 0)
968                         goto out_err_cpu_topo;
969         }
970
971         ret = perf_top__start_counters(top);
972         if (ret)
973                 goto out_delete;
974
975         top->session->evlist = top->evlist;
976         perf_session__set_id_hdr_size(top->session);
977
978         /*
979          * When perf is starting the traced process, all the events (apart from
980          * group members) have enable_on_exec=1 set, so don't spoil it by
981          * prematurely enabling them.
982          *
983          * XXX 'top' still doesn't start workloads like record, trace, but should,
984          * so leave the check here.
985          */
986         if (!target__none(&opts->target))
987                 perf_evlist__enable(top->evlist);
988
989         /* Wait for a minimal set of events before starting the snapshot */
990         perf_evlist__poll(top->evlist, 100);
991
992         perf_top__mmap_read(top);
993
994         ret = -1;
995         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
996                                                             display_thread), top)) {
997                 ui__error("Could not create display thread.\n");
998                 goto out_delete;
999         }
1000
1001         if (top->realtime_prio) {
1002                 struct sched_param param;
1003
1004                 param.sched_priority = top->realtime_prio;
1005                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1006                         ui__error("Could not set realtime priority.\n");
1007                         goto out_join;
1008                 }
1009         }
1010
1011         while (!done) {
1012                 u64 hits = top->samples;
1013
1014                 perf_top__mmap_read(top);
1015
1016                 if (hits == top->samples)
1017                         ret = perf_evlist__poll(top->evlist, 100);
1018         }
1019
1020         ret = 0;
1021 out_join:
1022         pthread_join(thread, NULL);
1023 out_delete:
1024         perf_session__delete(top->session);
1025         top->session = NULL;
1026
1027         return ret;
1028
1029 out_err_cpu_topo: {
1030         char errbuf[BUFSIZ];
1031         const char *err = strerror_r(-ret, errbuf, sizeof(errbuf));
1032
1033         ui__error("Could not read the CPU topology map: %s\n", err);
1034         goto out_delete;
1035 }
1036 }
1037
1038 static int
1039 callchain_opt(const struct option *opt, const char *arg, int unset)
1040 {
1041         symbol_conf.use_callchain = true;
1042         return record_callchain_opt(opt, arg, unset);
1043 }
1044
1045 static int
1046 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1047 {
1048         struct callchain_param *callchain = opt->value;
1049
1050         callchain->enabled = !unset;
1051         callchain->record_mode = CALLCHAIN_FP;
1052
1053         /*
1054          * --no-call-graph
1055          */
1056         if (unset) {
1057                 symbol_conf.use_callchain = false;
1058                 callchain->record_mode = CALLCHAIN_NONE;
1059                 return 0;
1060         }
1061
1062         return parse_callchain_top_opt(arg);
1063 }
1064
1065 static int perf_top_config(const char *var, const char *value, void *cb __maybe_unused)
1066 {
1067         if (!strcmp(var, "top.call-graph"))
1068                 var = "call-graph.record-mode"; /* fall-through */
1069         if (!strcmp(var, "top.children")) {
1070                 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
1071                 return 0;
1072         }
1073
1074         return 0;
1075 }
1076
1077 static int
1078 parse_percent_limit(const struct option *opt, const char *arg,
1079                     int unset __maybe_unused)
1080 {
1081         struct perf_top *top = opt->value;
1082
1083         top->min_percent = strtof(arg, NULL);
1084         return 0;
1085 }
1086
1087 const char top_callchain_help[] = CALLCHAIN_RECORD_HELP CALLCHAIN_REPORT_HELP
1088         "\n\t\t\t\tDefault: fp,graph,0.5,caller,function";
1089
1090 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1091 {
1092         char errbuf[BUFSIZ];
1093         struct perf_top top = {
1094                 .count_filter        = 5,
1095                 .delay_secs          = 2,
1096                 .record_opts = {
1097                         .mmap_pages     = UINT_MAX,
1098                         .user_freq      = UINT_MAX,
1099                         .user_interval  = ULLONG_MAX,
1100                         .freq           = 4000, /* 4 KHz */
1101                         .target         = {
1102                                 .uses_mmap   = true,
1103                         },
1104                         .proc_map_timeout    = 500,
1105                 },
1106                 .max_stack           = sysctl_perf_event_max_stack,
1107                 .sym_pcnt_filter     = 5,
1108         };
1109         struct record_opts *opts = &top.record_opts;
1110         struct target *target = &opts->target;
1111         const struct option options[] = {
1112         OPT_CALLBACK('e', "event", &top.evlist, "event",
1113                      "event selector. use 'perf list' to list available events",
1114                      parse_events_option),
1115         OPT_U64('c', "count", &opts->user_interval, "event period to sample"),
1116         OPT_STRING('p', "pid", &target->pid, "pid",
1117                     "profile events on existing process id"),
1118         OPT_STRING('t', "tid", &target->tid, "tid",
1119                     "profile events on existing thread id"),
1120         OPT_BOOLEAN('a', "all-cpus", &target->system_wide,
1121                             "system-wide collection from all CPUs"),
1122         OPT_STRING('C', "cpu", &target->cpu_list, "cpu",
1123                     "list of cpus to monitor"),
1124         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1125                    "file", "vmlinux pathname"),
1126         OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1127                     "don't load vmlinux even if found"),
1128         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1129                     "hide kernel symbols"),
1130         OPT_CALLBACK('m', "mmap-pages", &opts->mmap_pages, "pages",
1131                      "number of mmap data pages",
1132                      perf_evlist__parse_mmap_pages),
1133         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1134                     "collect data with this RT SCHED_FIFO priority"),
1135         OPT_INTEGER('d', "delay", &top.delay_secs,
1136                     "number of seconds to delay between refreshes"),
1137         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1138                             "dump the symbol table used for profiling"),
1139         OPT_INTEGER('f', "count-filter", &top.count_filter,
1140                     "only display functions with more events than this"),
1141         OPT_BOOLEAN(0, "group", &opts->group,
1142                             "put the counters into a counter group"),
1143         OPT_BOOLEAN('i', "no-inherit", &opts->no_inherit,
1144                     "child tasks do not inherit counters"),
1145         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1146                     "symbol to annotate"),
1147         OPT_BOOLEAN('z', "zero", &top.zero, "zero history across updates"),
1148         OPT_UINTEGER('F', "freq", &opts->user_freq, "profile at this frequency"),
1149         OPT_INTEGER('E', "entries", &top.print_entries,
1150                     "display this many functions"),
1151         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1152                     "hide user symbols"),
1153         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1154         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1155         OPT_INCR('v', "verbose", &verbose,
1156                     "be more verbose (show counter open errors, etc)"),
1157         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1158                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1159                    " Please refer the man page for the complete list."),
1160         OPT_STRING(0, "fields", &field_order, "key[,keys...]",
1161                    "output field(s): overhead, period, sample plus all of sort keys"),
1162         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1163                     "Show a column with the number of samples"),
1164         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1165                            NULL, "enables call-graph recording and display",
1166                            &callchain_opt),
1167         OPT_CALLBACK(0, "call-graph", &callchain_param,
1168                      "record_mode[,record_size],print_type,threshold[,print_limit],order,sort_key[,branch]",
1169                      top_callchain_help, &parse_callchain_opt),
1170         OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1171                     "Accumulate callchains of children and show total overhead as well"),
1172         OPT_INTEGER(0, "max-stack", &top.max_stack,
1173                     "Set the maximum stack depth when parsing the callchain. "
1174                     "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1175         OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1176                    "ignore callees of these functions in call graphs",
1177                    report_parse_ignore_callees_opt),
1178         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1179                     "Show a column with the sum of periods"),
1180         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1181                    "only consider symbols in these dsos"),
1182         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1183                    "only consider symbols in these comms"),
1184         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1185                    "only consider these symbols"),
1186         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1187                     "Interleave source code with assembly code (default)"),
1188         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1189                     "Display raw encoding of assembly instructions (default)"),
1190         OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1191                     "Enable kernel symbol demangling"),
1192         OPT_STRING(0, "objdump", &objdump_path, "path",
1193                     "objdump binary to use for disassembly and annotations"),
1194         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1195                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1196         OPT_STRING('u', "uid", &target->uid_str, "user", "user to profile"),
1197         OPT_CALLBACK(0, "percent-limit", &top, "percent",
1198                      "Don't show entries under that percent", parse_percent_limit),
1199         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1200                      "How to display percentage of filtered entries", parse_filter_percentage),
1201         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1202                    "width[,width...]",
1203                    "don't try to adjust column width, use these fixed values"),
1204         OPT_UINTEGER(0, "proc-map-timeout", &opts->proc_map_timeout,
1205                         "per thread proc mmap processing timeout in ms"),
1206         OPT_CALLBACK_NOOPT('b', "branch-any", &opts->branch_stack,
1207                      "branch any", "sample any taken branches",
1208                      parse_branch_stack),
1209         OPT_CALLBACK('j', "branch-filter", &opts->branch_stack,
1210                      "branch filter mask", "branch stack filter modes",
1211                      parse_branch_stack),
1212         OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1213                     "Show raw trace event output (do not use print fmt or plugins)"),
1214         OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1215                     "Show entries in a hierarchy"),
1216         OPT_END()
1217         };
1218         const char * const top_usage[] = {
1219                 "perf top [<options>]",
1220                 NULL
1221         };
1222         int status = hists__init();
1223
1224         if (status < 0)
1225                 return status;
1226
1227         top.evlist = perf_evlist__new();
1228         if (top.evlist == NULL)
1229                 return -ENOMEM;
1230
1231         perf_config(perf_top_config, &top);
1232
1233         argc = parse_options(argc, argv, options, top_usage, 0);
1234         if (argc)
1235                 usage_with_options(top_usage, options);
1236
1237         if (!top.evlist->nr_entries &&
1238             perf_evlist__add_default(top.evlist) < 0) {
1239                 pr_err("Not enough memory for event selector list\n");
1240                 goto out_delete_evlist;
1241         }
1242
1243         if (symbol_conf.report_hierarchy) {
1244                 /* disable incompatible options */
1245                 symbol_conf.event_group = false;
1246                 symbol_conf.cumulate_callchain = false;
1247
1248                 if (field_order) {
1249                         pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1250                         parse_options_usage(top_usage, options, "fields", 0);
1251                         parse_options_usage(NULL, options, "hierarchy", 0);
1252                         goto out_delete_evlist;
1253                 }
1254         }
1255
1256         sort__mode = SORT_MODE__TOP;
1257         /* display thread wants entries to be collapsed in a different tree */
1258         perf_hpp_list.need_collapse = 1;
1259
1260         if (top.use_stdio)
1261                 use_browser = 0;
1262         else if (top.use_tui)
1263                 use_browser = 1;
1264
1265         setup_browser(false);
1266
1267         if (setup_sorting(top.evlist) < 0) {
1268                 if (sort_order)
1269                         parse_options_usage(top_usage, options, "s", 1);
1270                 if (field_order)
1271                         parse_options_usage(sort_order ? NULL : top_usage,
1272                                             options, "fields", 0);
1273                 goto out_delete_evlist;
1274         }
1275
1276         status = target__validate(target);
1277         if (status) {
1278                 target__strerror(target, status, errbuf, BUFSIZ);
1279                 ui__warning("%s\n", errbuf);
1280         }
1281
1282         status = target__parse_uid(target);
1283         if (status) {
1284                 int saved_errno = errno;
1285
1286                 target__strerror(target, status, errbuf, BUFSIZ);
1287                 ui__error("%s\n", errbuf);
1288
1289                 status = -saved_errno;
1290                 goto out_delete_evlist;
1291         }
1292
1293         if (target__none(target))
1294                 target->system_wide = true;
1295
1296         if (perf_evlist__create_maps(top.evlist, target) < 0) {
1297                 ui__error("Couldn't create thread/CPU maps: %s\n",
1298                           errno == ENOENT ? "No such process" : strerror_r(errno, errbuf, sizeof(errbuf)));
1299                 goto out_delete_evlist;
1300         }
1301
1302         symbol_conf.nr_events = top.evlist->nr_entries;
1303
1304         if (top.delay_secs < 1)
1305                 top.delay_secs = 1;
1306
1307         if (record_opts__config(opts)) {
1308                 status = -EINVAL;
1309                 goto out_delete_evlist;
1310         }
1311
1312         top.sym_evsel = perf_evlist__first(top.evlist);
1313
1314         if (!callchain_param.enabled) {
1315                 symbol_conf.cumulate_callchain = false;
1316                 perf_hpp__cancel_cumulate();
1317         }
1318
1319         if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1320                 callchain_param.order = ORDER_CALLER;
1321
1322         symbol_conf.priv_size = sizeof(struct annotation);
1323
1324         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1325         if (symbol__init(NULL) < 0)
1326                 return -1;
1327
1328         sort__setup_elide(stdout);
1329
1330         get_term_dimensions(&top.winsize);
1331         if (top.print_entries == 0) {
1332                 struct sigaction act = {
1333                         .sa_sigaction = perf_top__sig_winch,
1334                         .sa_flags     = SA_SIGINFO,
1335                 };
1336                 perf_top__update_print_entries(&top);
1337                 sigaction(SIGWINCH, &act, NULL);
1338         }
1339
1340         status = __cmd_top(&top);
1341
1342 out_delete_evlist:
1343         perf_evlist__delete(top.evlist);
1344
1345         return status;
1346 }