Merge branch 'perf-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[~shefty/rdma-dev.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/session.h"
30 #include "util/symbol.h"
31 #include "util/thread.h"
32 #include "util/thread_map.h"
33 #include "util/top.h"
34 #include "util/util.h"
35 #include <linux/rbtree.h>
36 #include "util/parse-options.h"
37 #include "util/parse-events.h"
38 #include "util/cpumap.h"
39 #include "util/xyarray.h"
40 #include "util/sort.h"
41
42 #include "util/debug.h"
43
44 #include <assert.h>
45 #include <fcntl.h>
46
47 #include <stdio.h>
48 #include <termios.h>
49 #include <unistd.h>
50 #include <inttypes.h>
51
52 #include <errno.h>
53 #include <time.h>
54 #include <sched.h>
55
56 #include <sys/syscall.h>
57 #include <sys/ioctl.h>
58 #include <sys/poll.h>
59 #include <sys/prctl.h>
60 #include <sys/wait.h>
61 #include <sys/uio.h>
62 #include <sys/mman.h>
63
64 #include <linux/unistd.h>
65 #include <linux/types.h>
66
67
68 void get_term_dimensions(struct winsize *ws)
69 {
70         char *s = getenv("LINES");
71
72         if (s != NULL) {
73                 ws->ws_row = atoi(s);
74                 s = getenv("COLUMNS");
75                 if (s != NULL) {
76                         ws->ws_col = atoi(s);
77                         if (ws->ws_row && ws->ws_col)
78                                 return;
79                 }
80         }
81 #ifdef TIOCGWINSZ
82         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
83             ws->ws_row && ws->ws_col)
84                 return;
85 #endif
86         ws->ws_row = 25;
87         ws->ws_col = 80;
88 }
89
90 static void perf_top__update_print_entries(struct perf_top *top)
91 {
92         top->print_entries = top->winsize.ws_row;
93
94         if (top->print_entries > 9)
95                 top->print_entries -= 9;
96 }
97
98 static void perf_top__sig_winch(int sig __used, siginfo_t *info __used, void *arg)
99 {
100         struct perf_top *top = arg;
101
102         get_term_dimensions(&top->winsize);
103         perf_top__update_print_entries(top);
104 }
105
106 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
107 {
108         struct symbol *sym;
109         struct annotation *notes;
110         struct map *map;
111         int err = -1;
112
113         if (!he || !he->ms.sym)
114                 return -1;
115
116         sym = he->ms.sym;
117         map = he->ms.map;
118
119         /*
120          * We can't annotate with just /proc/kallsyms
121          */
122         if (map->dso->symtab_type == SYMTAB__KALLSYMS) {
123                 pr_err("Can't annotate %s: No vmlinux file was found in the "
124                        "path\n", sym->name);
125                 sleep(1);
126                 return -1;
127         }
128
129         notes = symbol__annotation(sym);
130         if (notes->src != NULL) {
131                 pthread_mutex_lock(&notes->lock);
132                 goto out_assign;
133         }
134
135         pthread_mutex_lock(&notes->lock);
136
137         if (symbol__alloc_hist(sym) < 0) {
138                 pthread_mutex_unlock(&notes->lock);
139                 pr_err("Not enough memory for annotating '%s' symbol!\n",
140                        sym->name);
141                 sleep(1);
142                 return err;
143         }
144
145         err = symbol__annotate(sym, map, 0);
146         if (err == 0) {
147 out_assign:
148                 top->sym_filter_entry = he;
149         }
150
151         pthread_mutex_unlock(&notes->lock);
152         return err;
153 }
154
155 static void __zero_source_counters(struct hist_entry *he)
156 {
157         struct symbol *sym = he->ms.sym;
158         symbol__annotate_zero_histograms(sym);
159 }
160
161 static void perf_top__record_precise_ip(struct perf_top *top,
162                                         struct hist_entry *he,
163                                         int counter, u64 ip)
164 {
165         struct annotation *notes;
166         struct symbol *sym;
167
168         if (he == NULL || he->ms.sym == NULL ||
169             ((top->sym_filter_entry == NULL ||
170               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
171                 return;
172
173         sym = he->ms.sym;
174         notes = symbol__annotation(sym);
175
176         if (pthread_mutex_trylock(&notes->lock))
177                 return;
178
179         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
180                 pthread_mutex_unlock(&notes->lock);
181                 pr_err("Not enough memory for annotating '%s' symbol!\n",
182                        sym->name);
183                 sleep(1);
184                 return;
185         }
186
187         ip = he->ms.map->map_ip(he->ms.map, ip);
188         symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
189
190         pthread_mutex_unlock(&notes->lock);
191 }
192
193 static void perf_top__show_details(struct perf_top *top)
194 {
195         struct hist_entry *he = top->sym_filter_entry;
196         struct annotation *notes;
197         struct symbol *symbol;
198         int more;
199
200         if (!he)
201                 return;
202
203         symbol = he->ms.sym;
204         notes = symbol__annotation(symbol);
205
206         pthread_mutex_lock(&notes->lock);
207
208         if (notes->src == NULL)
209                 goto out_unlock;
210
211         printf("Showing %s for %s\n", event_name(top->sym_evsel), symbol->name);
212         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
213
214         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
215                                        0, top->sym_pcnt_filter, top->print_entries, 4);
216         if (top->zero)
217                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
218         else
219                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
220         if (more != 0)
221                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
222 out_unlock:
223         pthread_mutex_unlock(&notes->lock);
224 }
225
226 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
227
228 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
229                                                      struct addr_location *al,
230                                                      struct perf_sample *sample)
231 {
232         struct hist_entry *he;
233
234         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
235         if (he == NULL)
236                 return NULL;
237
238         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
239         return he;
240 }
241
242 static void perf_top__print_sym_table(struct perf_top *top)
243 {
244         char bf[160];
245         int printed = 0;
246         const int win_width = top->winsize.ws_col - 1;
247
248         puts(CONSOLE_CLEAR);
249
250         perf_top__header_snprintf(top, bf, sizeof(bf));
251         printf("%s\n", bf);
252
253         perf_top__reset_sample_counters(top);
254
255         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
256
257         if (top->sym_evsel->hists.stats.nr_lost_warned !=
258             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
259                 top->sym_evsel->hists.stats.nr_lost_warned =
260                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
261                 color_fprintf(stdout, PERF_COLOR_RED,
262                               "WARNING: LOST %d chunks, Check IO/CPU overload",
263                               top->sym_evsel->hists.stats.nr_lost_warned);
264                 ++printed;
265         }
266
267         if (top->sym_filter_entry) {
268                 perf_top__show_details(top);
269                 return;
270         }
271
272         hists__collapse_resort_threaded(&top->sym_evsel->hists);
273         hists__output_resort_threaded(&top->sym_evsel->hists);
274         hists__decay_entries_threaded(&top->sym_evsel->hists,
275                                       top->hide_user_symbols,
276                                       top->hide_kernel_symbols);
277         hists__output_recalc_col_len(&top->sym_evsel->hists,
278                                      top->winsize.ws_row - 3);
279         putchar('\n');
280         hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
281                        top->winsize.ws_row - 4 - printed, win_width, stdout);
282 }
283
284 static void prompt_integer(int *target, const char *msg)
285 {
286         char *buf = malloc(0), *p;
287         size_t dummy = 0;
288         int tmp;
289
290         fprintf(stdout, "\n%s: ", msg);
291         if (getline(&buf, &dummy, stdin) < 0)
292                 return;
293
294         p = strchr(buf, '\n');
295         if (p)
296                 *p = 0;
297
298         p = buf;
299         while(*p) {
300                 if (!isdigit(*p))
301                         goto out_free;
302                 p++;
303         }
304         tmp = strtoul(buf, NULL, 10);
305         *target = tmp;
306 out_free:
307         free(buf);
308 }
309
310 static void prompt_percent(int *target, const char *msg)
311 {
312         int tmp = 0;
313
314         prompt_integer(&tmp, msg);
315         if (tmp >= 0 && tmp <= 100)
316                 *target = tmp;
317 }
318
319 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
320 {
321         char *buf = malloc(0), *p;
322         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
323         struct rb_node *next;
324         size_t dummy = 0;
325
326         /* zero counters of active symbol */
327         if (syme) {
328                 __zero_source_counters(syme);
329                 top->sym_filter_entry = NULL;
330         }
331
332         fprintf(stdout, "\n%s: ", msg);
333         if (getline(&buf, &dummy, stdin) < 0)
334                 goto out_free;
335
336         p = strchr(buf, '\n');
337         if (p)
338                 *p = 0;
339
340         next = rb_first(&top->sym_evsel->hists.entries);
341         while (next) {
342                 n = rb_entry(next, struct hist_entry, rb_node);
343                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
344                         found = n;
345                         break;
346                 }
347                 next = rb_next(&n->rb_node);
348         }
349
350         if (!found) {
351                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
352                 sleep(1);
353         } else
354                 perf_top__parse_source(top, found);
355
356 out_free:
357         free(buf);
358 }
359
360 static void perf_top__print_mapped_keys(struct perf_top *top)
361 {
362         char *name = NULL;
363
364         if (top->sym_filter_entry) {
365                 struct symbol *sym = top->sym_filter_entry->ms.sym;
366                 name = sym->name;
367         }
368
369         fprintf(stdout, "\nMapped keys:\n");
370         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
371         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
372
373         if (top->evlist->nr_entries > 1)
374                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top->sym_evsel));
375
376         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
377
378         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
379         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
380         fprintf(stdout, "\t[S]     stop annotation.\n");
381
382         fprintf(stdout,
383                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
384                 top->hide_kernel_symbols ? "yes" : "no");
385         fprintf(stdout,
386                 "\t[U]     hide user symbols.               \t(%s)\n",
387                 top->hide_user_symbols ? "yes" : "no");
388         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
389         fprintf(stdout, "\t[qQ]    quit.\n");
390 }
391
392 static int perf_top__key_mapped(struct perf_top *top, int c)
393 {
394         switch (c) {
395                 case 'd':
396                 case 'e':
397                 case 'f':
398                 case 'z':
399                 case 'q':
400                 case 'Q':
401                 case 'K':
402                 case 'U':
403                 case 'F':
404                 case 's':
405                 case 'S':
406                         return 1;
407                 case 'E':
408                         return top->evlist->nr_entries > 1 ? 1 : 0;
409                 default:
410                         break;
411         }
412
413         return 0;
414 }
415
416 static void perf_top__handle_keypress(struct perf_top *top, int c)
417 {
418         if (!perf_top__key_mapped(top, c)) {
419                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
420                 struct termios tc, save;
421
422                 perf_top__print_mapped_keys(top);
423                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
424                 fflush(stdout);
425
426                 tcgetattr(0, &save);
427                 tc = save;
428                 tc.c_lflag &= ~(ICANON | ECHO);
429                 tc.c_cc[VMIN] = 0;
430                 tc.c_cc[VTIME] = 0;
431                 tcsetattr(0, TCSANOW, &tc);
432
433                 poll(&stdin_poll, 1, -1);
434                 c = getc(stdin);
435
436                 tcsetattr(0, TCSAFLUSH, &save);
437                 if (!perf_top__key_mapped(top, c))
438                         return;
439         }
440
441         switch (c) {
442                 case 'd':
443                         prompt_integer(&top->delay_secs, "Enter display delay");
444                         if (top->delay_secs < 1)
445                                 top->delay_secs = 1;
446                         break;
447                 case 'e':
448                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
449                         if (top->print_entries == 0) {
450                                 struct sigaction act = {
451                                         .sa_sigaction = perf_top__sig_winch,
452                                         .sa_flags     = SA_SIGINFO,
453                                 };
454                                 perf_top__sig_winch(SIGWINCH, NULL, top);
455                                 sigaction(SIGWINCH, &act, NULL);
456                         } else
457                                 signal(SIGWINCH, SIG_DFL);
458                         break;
459                 case 'E':
460                         if (top->evlist->nr_entries > 1) {
461                                 /* Select 0 as the default event: */
462                                 int counter = 0;
463
464                                 fprintf(stderr, "\nAvailable events:");
465
466                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
467                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
468
469                                 prompt_integer(&counter, "Enter details event counter");
470
471                                 if (counter >= top->evlist->nr_entries) {
472                                         top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
473                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
474                                         sleep(1);
475                                         break;
476                                 }
477                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
478                                         if (top->sym_evsel->idx == counter)
479                                                 break;
480                         } else
481                                 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
482                         break;
483                 case 'f':
484                         prompt_integer(&top->count_filter, "Enter display event count filter");
485                         break;
486                 case 'F':
487                         prompt_percent(&top->sym_pcnt_filter,
488                                        "Enter details display event filter (percent)");
489                         break;
490                 case 'K':
491                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
492                         break;
493                 case 'q':
494                 case 'Q':
495                         printf("exiting.\n");
496                         if (top->dump_symtab)
497                                 perf_session__fprintf_dsos(top->session, stderr);
498                         exit(0);
499                 case 's':
500                         perf_top__prompt_symbol(top, "Enter details symbol");
501                         break;
502                 case 'S':
503                         if (!top->sym_filter_entry)
504                                 break;
505                         else {
506                                 struct hist_entry *syme = top->sym_filter_entry;
507
508                                 top->sym_filter_entry = NULL;
509                                 __zero_source_counters(syme);
510                         }
511                         break;
512                 case 'U':
513                         top->hide_user_symbols = !top->hide_user_symbols;
514                         break;
515                 case 'z':
516                         top->zero = !top->zero;
517                         break;
518                 default:
519                         break;
520         }
521 }
522
523 static void perf_top__sort_new_samples(void *arg)
524 {
525         struct perf_top *t = arg;
526         perf_top__reset_sample_counters(t);
527
528         if (t->evlist->selected != NULL)
529                 t->sym_evsel = t->evlist->selected;
530
531         hists__collapse_resort_threaded(&t->sym_evsel->hists);
532         hists__output_resort_threaded(&t->sym_evsel->hists);
533         hists__decay_entries_threaded(&t->sym_evsel->hists,
534                                       t->hide_user_symbols,
535                                       t->hide_kernel_symbols);
536 }
537
538 static void *display_thread_tui(void *arg)
539 {
540         struct perf_top *top = arg;
541         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
542
543         perf_top__sort_new_samples(top);
544         perf_evlist__tui_browse_hists(top->evlist, help,
545                                       perf_top__sort_new_samples,
546                                       top, top->delay_secs);
547
548         exit_browser(0);
549         exit(0);
550         return NULL;
551 }
552
553 static void *display_thread(void *arg)
554 {
555         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
556         struct termios tc, save;
557         struct perf_top *top = arg;
558         int delay_msecs, c;
559
560         tcgetattr(0, &save);
561         tc = save;
562         tc.c_lflag &= ~(ICANON | ECHO);
563         tc.c_cc[VMIN] = 0;
564         tc.c_cc[VTIME] = 0;
565
566         pthread__unblock_sigwinch();
567 repeat:
568         delay_msecs = top->delay_secs * 1000;
569         tcsetattr(0, TCSANOW, &tc);
570         /* trash return*/
571         getc(stdin);
572
573         while (1) {
574                 perf_top__print_sym_table(top);
575                 /*
576                  * Either timeout expired or we got an EINTR due to SIGWINCH,
577                  * refresh screen in both cases.
578                  */
579                 switch (poll(&stdin_poll, 1, delay_msecs)) {
580                 case 0:
581                         continue;
582                 case -1:
583                         if (errno == EINTR)
584                                 continue;
585                         /* Fall trhu */
586                 default:
587                         goto process_hotkey;
588                 }
589         }
590 process_hotkey:
591         c = getc(stdin);
592         tcsetattr(0, TCSAFLUSH, &save);
593
594         perf_top__handle_keypress(top, c);
595         goto repeat;
596
597         return NULL;
598 }
599
600 /* Tag samples to be skipped. */
601 static const char *skip_symbols[] = {
602         "default_idle",
603         "native_safe_halt",
604         "cpu_idle",
605         "enter_idle",
606         "exit_idle",
607         "mwait_idle",
608         "mwait_idle_with_hints",
609         "poll_idle",
610         "ppc64_runlatch_off",
611         "pseries_dedicated_idle_sleep",
612         NULL
613 };
614
615 static int symbol_filter(struct map *map __used, struct symbol *sym)
616 {
617         const char *name = sym->name;
618         int i;
619
620         /*
621          * ppc64 uses function descriptors and appends a '.' to the
622          * start of every instruction address. Remove it.
623          */
624         if (name[0] == '.')
625                 name++;
626
627         if (!strcmp(name, "_text") ||
628             !strcmp(name, "_etext") ||
629             !strcmp(name, "_sinittext") ||
630             !strncmp("init_module", name, 11) ||
631             !strncmp("cleanup_module", name, 14) ||
632             strstr(name, "_text_start") ||
633             strstr(name, "_text_end"))
634                 return 1;
635
636         for (i = 0; skip_symbols[i]; i++) {
637                 if (!strcmp(skip_symbols[i], name)) {
638                         sym->ignore = true;
639                         break;
640                 }
641         }
642
643         return 0;
644 }
645
646 static void perf_event__process_sample(struct perf_tool *tool,
647                                        const union perf_event *event,
648                                        struct perf_evsel *evsel,
649                                        struct perf_sample *sample,
650                                        struct machine *machine)
651 {
652         struct perf_top *top = container_of(tool, struct perf_top, tool);
653         struct symbol *parent = NULL;
654         u64 ip = event->ip.ip;
655         struct addr_location al;
656         int err;
657
658         if (!machine && perf_guest) {
659                 pr_err("Can't find guest [%d]'s kernel information\n",
660                         event->ip.pid);
661                 return;
662         }
663
664         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
665                 top->exact_samples++;
666
667         if (perf_event__preprocess_sample(event, machine, &al, sample,
668                                           symbol_filter) < 0 ||
669             al.filtered)
670                 return;
671
672         if (!top->kptr_restrict_warned &&
673             symbol_conf.kptr_restrict &&
674             al.cpumode == PERF_RECORD_MISC_KERNEL) {
675                 ui__warning(
676 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
677 "Check /proc/sys/kernel/kptr_restrict.\n\n"
678 "Kernel%s samples will not be resolved.\n",
679                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
680                           " modules" : "");
681                 if (use_browser <= 0)
682                         sleep(5);
683                 top->kptr_restrict_warned = true;
684         }
685
686         if (al.sym == NULL) {
687                 const char *msg = "Kernel samples will not be resolved.\n";
688                 /*
689                  * As we do lazy loading of symtabs we only will know if the
690                  * specified vmlinux file is invalid when we actually have a
691                  * hit in kernel space and then try to load it. So if we get
692                  * here and there are _no_ symbols in the DSO backing the
693                  * kernel map, bail out.
694                  *
695                  * We may never get here, for instance, if we use -K/
696                  * --hide-kernel-symbols, even if the user specifies an
697                  * invalid --vmlinux ;-)
698                  */
699                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
700                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
701                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
702                         if (symbol_conf.vmlinux_name) {
703                                 ui__warning("The %s file can't be used.\n%s",
704                                             symbol_conf.vmlinux_name, msg);
705                         } else {
706                                 ui__warning("A vmlinux file was not found.\n%s",
707                                             msg);
708                         }
709
710                         if (use_browser <= 0)
711                                 sleep(5);
712                         top->vmlinux_warned = true;
713                 }
714         }
715
716         if (al.sym == NULL || !al.sym->ignore) {
717                 struct hist_entry *he;
718
719                 if ((sort__has_parent || symbol_conf.use_callchain) &&
720                     sample->callchain) {
721                         err = machine__resolve_callchain(machine, evsel, al.thread,
722                                                          sample->callchain, &parent);
723                         if (err)
724                                 return;
725                 }
726
727                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
728                 if (he == NULL) {
729                         pr_err("Problem incrementing symbol period, skipping event\n");
730                         return;
731                 }
732
733                 if (symbol_conf.use_callchain) {
734                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
735                                                sample->period);
736                         if (err)
737                                 return;
738                 }
739
740                 if (top->sort_has_symbols)
741                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
742         }
743
744         return;
745 }
746
747 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
748 {
749         struct perf_sample sample;
750         struct perf_evsel *evsel;
751         struct perf_session *session = top->session;
752         union perf_event *event;
753         struct machine *machine;
754         u8 origin;
755         int ret;
756
757         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
758                 ret = perf_session__parse_sample(session, event, &sample);
759                 if (ret) {
760                         pr_err("Can't parse sample, err = %d\n", ret);
761                         continue;
762                 }
763
764                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
765                 assert(evsel != NULL);
766
767                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
768
769                 if (event->header.type == PERF_RECORD_SAMPLE)
770                         ++top->samples;
771
772                 switch (origin) {
773                 case PERF_RECORD_MISC_USER:
774                         ++top->us_samples;
775                         if (top->hide_user_symbols)
776                                 continue;
777                         machine = perf_session__find_host_machine(session);
778                         break;
779                 case PERF_RECORD_MISC_KERNEL:
780                         ++top->kernel_samples;
781                         if (top->hide_kernel_symbols)
782                                 continue;
783                         machine = perf_session__find_host_machine(session);
784                         break;
785                 case PERF_RECORD_MISC_GUEST_KERNEL:
786                         ++top->guest_kernel_samples;
787                         machine = perf_session__find_machine(session, event->ip.pid);
788                         break;
789                 case PERF_RECORD_MISC_GUEST_USER:
790                         ++top->guest_us_samples;
791                         /*
792                          * TODO: we don't process guest user from host side
793                          * except simple counting.
794                          */
795                         /* Fall thru */
796                 default:
797                         continue;
798                 }
799
800
801                 if (event->header.type == PERF_RECORD_SAMPLE) {
802                         perf_event__process_sample(&top->tool, event, evsel,
803                                                    &sample, machine);
804                 } else if (event->header.type < PERF_RECORD_MAX) {
805                         hists__inc_nr_events(&evsel->hists, event->header.type);
806                         perf_event__process(&top->tool, event, &sample, machine);
807                 } else
808                         ++session->hists.stats.nr_unknown_events;
809         }
810 }
811
812 static void perf_top__mmap_read(struct perf_top *top)
813 {
814         int i;
815
816         for (i = 0; i < top->evlist->nr_mmaps; i++)
817                 perf_top__mmap_read_idx(top, i);
818 }
819
820 static void perf_top__start_counters(struct perf_top *top)
821 {
822         struct perf_evsel *counter, *first;
823         struct perf_evlist *evlist = top->evlist;
824
825         first = list_entry(evlist->entries.next, struct perf_evsel, node);
826
827         list_for_each_entry(counter, &evlist->entries, node) {
828                 struct perf_event_attr *attr = &counter->attr;
829                 struct xyarray *group_fd = NULL;
830
831                 if (top->group && counter != first)
832                         group_fd = first->fd;
833
834                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
835
836                 if (top->freq) {
837                         attr->sample_type |= PERF_SAMPLE_PERIOD;
838                         attr->freq        = 1;
839                         attr->sample_freq = top->freq;
840                 }
841
842                 if (evlist->nr_entries > 1) {
843                         attr->sample_type |= PERF_SAMPLE_ID;
844                         attr->read_format |= PERF_FORMAT_ID;
845                 }
846
847                 if (symbol_conf.use_callchain)
848                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
849
850                 attr->mmap = 1;
851                 attr->comm = 1;
852                 attr->inherit = top->inherit;
853 retry_sample_id:
854                 attr->sample_id_all = top->sample_id_all_avail ? 1 : 0;
855 try_again:
856                 if (perf_evsel__open(counter, top->evlist->cpus,
857                                      top->evlist->threads, top->group,
858                                      group_fd) < 0) {
859                         int err = errno;
860
861                         if (err == EPERM || err == EACCES) {
862                                 ui__error_paranoid();
863                                 goto out_err;
864                         } else if (err == EINVAL && top->sample_id_all_avail) {
865                                 /*
866                                  * Old kernel, no attr->sample_id_type_all field
867                                  */
868                                 top->sample_id_all_avail = false;
869                                 goto retry_sample_id;
870                         }
871                         /*
872                          * If it's cycles then fall back to hrtimer
873                          * based cpu-clock-tick sw counter, which
874                          * is always available even if no PMU support:
875                          */
876                         if (attr->type == PERF_TYPE_HARDWARE &&
877                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
878                                 if (verbose)
879                                         ui__warning("Cycles event not supported,\n"
880                                                     "trying to fall back to cpu-clock-ticks\n");
881
882                                 attr->type = PERF_TYPE_SOFTWARE;
883                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
884                                 goto try_again;
885                         }
886
887                         if (err == ENOENT) {
888                                 ui__warning("The %s event is not supported.\n",
889                                             event_name(counter));
890                                 goto out_err;
891                         } else if (err == EMFILE) {
892                                 ui__warning("Too many events are opened.\n"
893                                             "Try again after reducing the number of events\n");
894                                 goto out_err;
895                         }
896
897                         ui__warning("The sys_perf_event_open() syscall "
898                                     "returned with %d (%s).  /bin/dmesg "
899                                     "may provide additional information.\n"
900                                     "No CONFIG_PERF_EVENTS=y kernel support "
901                                     "configured?\n", err, strerror(err));
902                         goto out_err;
903                 }
904         }
905
906         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
907                 ui__warning("Failed to mmap with %d (%s)\n",
908                             errno, strerror(errno));
909                 goto out_err;
910         }
911
912         return;
913
914 out_err:
915         exit_browser(0);
916         exit(0);
917 }
918
919 static int perf_top__setup_sample_type(struct perf_top *top)
920 {
921         if (!top->sort_has_symbols) {
922                 if (symbol_conf.use_callchain) {
923                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
924                         return -EINVAL;
925                 }
926         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
927                 if (callchain_register_param(&callchain_param) < 0) {
928                         ui__warning("Can't register callchain params.\n");
929                         return -EINVAL;
930                 }
931         }
932
933         return 0;
934 }
935
936 static int __cmd_top(struct perf_top *top)
937 {
938         pthread_t thread;
939         int ret;
940         /*
941          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
942          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
943          */
944         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
945         if (top->session == NULL)
946                 return -ENOMEM;
947
948         ret = perf_top__setup_sample_type(top);
949         if (ret)
950                 goto out_delete;
951
952         if (top->target_tid != -1)
953                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
954                                                   perf_event__process,
955                                                   &top->session->host_machine);
956         else
957                 perf_event__synthesize_threads(&top->tool, perf_event__process,
958                                                &top->session->host_machine);
959         perf_top__start_counters(top);
960         top->session->evlist = top->evlist;
961         perf_session__update_sample_type(top->session);
962
963         /* Wait for a minimal set of events before starting the snapshot */
964         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
965
966         perf_top__mmap_read(top);
967
968         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
969                                                             display_thread), top)) {
970                 printf("Could not create display thread.\n");
971                 exit(-1);
972         }
973
974         if (top->realtime_prio) {
975                 struct sched_param param;
976
977                 param.sched_priority = top->realtime_prio;
978                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
979                         printf("Could not set realtime priority.\n");
980                         exit(-1);
981                 }
982         }
983
984         while (1) {
985                 u64 hits = top->samples;
986
987                 perf_top__mmap_read(top);
988
989                 if (hits == top->samples)
990                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
991         }
992
993 out_delete:
994         perf_session__delete(top->session);
995         top->session = NULL;
996
997         return 0;
998 }
999
1000 static int
1001 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1002 {
1003         struct perf_top *top = (struct perf_top *)opt->value;
1004         char *tok, *tok2;
1005         char *endptr;
1006
1007         /*
1008          * --no-call-graph
1009          */
1010         if (unset) {
1011                 top->dont_use_callchains = true;
1012                 return 0;
1013         }
1014
1015         symbol_conf.use_callchain = true;
1016
1017         if (!arg)
1018                 return 0;
1019
1020         tok = strtok((char *)arg, ",");
1021         if (!tok)
1022                 return -1;
1023
1024         /* get the output mode */
1025         if (!strncmp(tok, "graph", strlen(arg)))
1026                 callchain_param.mode = CHAIN_GRAPH_ABS;
1027
1028         else if (!strncmp(tok, "flat", strlen(arg)))
1029                 callchain_param.mode = CHAIN_FLAT;
1030
1031         else if (!strncmp(tok, "fractal", strlen(arg)))
1032                 callchain_param.mode = CHAIN_GRAPH_REL;
1033
1034         else if (!strncmp(tok, "none", strlen(arg))) {
1035                 callchain_param.mode = CHAIN_NONE;
1036                 symbol_conf.use_callchain = false;
1037
1038                 return 0;
1039         } else
1040                 return -1;
1041
1042         /* get the min percentage */
1043         tok = strtok(NULL, ",");
1044         if (!tok)
1045                 goto setup;
1046
1047         callchain_param.min_percent = strtod(tok, &endptr);
1048         if (tok == endptr)
1049                 return -1;
1050
1051         /* get the print limit */
1052         tok2 = strtok(NULL, ",");
1053         if (!tok2)
1054                 goto setup;
1055
1056         if (tok2[0] != 'c') {
1057                 callchain_param.print_limit = strtod(tok2, &endptr);
1058                 tok2 = strtok(NULL, ",");
1059                 if (!tok2)
1060                         goto setup;
1061         }
1062
1063         /* get the call chain order */
1064         if (!strcmp(tok2, "caller"))
1065                 callchain_param.order = ORDER_CALLER;
1066         else if (!strcmp(tok2, "callee"))
1067                 callchain_param.order = ORDER_CALLEE;
1068         else
1069                 return -1;
1070 setup:
1071         if (callchain_register_param(&callchain_param) < 0) {
1072                 fprintf(stderr, "Can't register callchain params\n");
1073                 return -1;
1074         }
1075         return 0;
1076 }
1077
1078 static const char * const top_usage[] = {
1079         "perf top [<options>]",
1080         NULL
1081 };
1082
1083 int cmd_top(int argc, const char **argv, const char *prefix __used)
1084 {
1085         struct perf_evsel *pos;
1086         int status = -ENOMEM;
1087         struct perf_top top = {
1088                 .count_filter        = 5,
1089                 .delay_secs          = 2,
1090                 .target_pid          = -1,
1091                 .target_tid          = -1,
1092                 .freq                = 1000, /* 1 KHz */
1093                 .sample_id_all_avail = true,
1094                 .mmap_pages          = 128,
1095                 .sym_pcnt_filter     = 5,
1096         };
1097         char callchain_default_opt[] = "fractal,0.5,callee";
1098         const struct option options[] = {
1099         OPT_CALLBACK('e', "event", &top.evlist, "event",
1100                      "event selector. use 'perf list' to list available events",
1101                      parse_events_option),
1102         OPT_INTEGER('c', "count", &top.default_interval,
1103                     "event period to sample"),
1104         OPT_INTEGER('p', "pid", &top.target_pid,
1105                     "profile events on existing process id"),
1106         OPT_INTEGER('t', "tid", &top.target_tid,
1107                     "profile events on existing thread id"),
1108         OPT_BOOLEAN('a', "all-cpus", &top.system_wide,
1109                             "system-wide collection from all CPUs"),
1110         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1111                     "list of cpus to monitor"),
1112         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1113                    "file", "vmlinux pathname"),
1114         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1115                     "hide kernel symbols"),
1116         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1117         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1118                     "collect data with this RT SCHED_FIFO priority"),
1119         OPT_INTEGER('d', "delay", &top.delay_secs,
1120                     "number of seconds to delay between refreshes"),
1121         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1122                             "dump the symbol table used for profiling"),
1123         OPT_INTEGER('f', "count-filter", &top.count_filter,
1124                     "only display functions with more events than this"),
1125         OPT_BOOLEAN('g', "group", &top.group,
1126                             "put the counters into a counter group"),
1127         OPT_BOOLEAN('i', "inherit", &top.inherit,
1128                     "child tasks inherit counters"),
1129         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1130                     "symbol to annotate"),
1131         OPT_BOOLEAN('z', "zero", &top.zero,
1132                     "zero history across updates"),
1133         OPT_INTEGER('F', "freq", &top.freq,
1134                     "profile at this frequency"),
1135         OPT_INTEGER('E', "entries", &top.print_entries,
1136                     "display this many functions"),
1137         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1138                     "hide user symbols"),
1139         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1140         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1141         OPT_INCR('v', "verbose", &verbose,
1142                     "be more verbose (show counter open errors, etc)"),
1143         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1144                    "sort by key(s): pid, comm, dso, symbol, parent"),
1145         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1146                     "Show a column with the number of samples"),
1147         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1148                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1149                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1150                      callchain_default_opt),
1151         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1152                     "Show a column with the sum of periods"),
1153         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1154                    "only consider symbols in these dsos"),
1155         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1156                    "only consider symbols in these comms"),
1157         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1158                    "only consider these symbols"),
1159         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1160                     "Interleave source code with assembly code (default)"),
1161         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1162                     "Display raw encoding of assembly instructions (default)"),
1163         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1164                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1165         OPT_END()
1166         };
1167
1168         top.evlist = perf_evlist__new(NULL, NULL);
1169         if (top.evlist == NULL)
1170                 return -ENOMEM;
1171
1172         symbol_conf.exclude_other = false;
1173
1174         argc = parse_options(argc, argv, options, top_usage, 0);
1175         if (argc)
1176                 usage_with_options(top_usage, options);
1177
1178         if (sort_order == default_sort_order)
1179                 sort_order = "dso,symbol";
1180
1181         setup_sorting(top_usage, options);
1182
1183         if (top.use_stdio)
1184                 use_browser = 0;
1185         else if (top.use_tui)
1186                 use_browser = 1;
1187
1188         setup_browser(false);
1189
1190         /* CPU and PID are mutually exclusive */
1191         if (top.target_tid > 0 && top.cpu_list) {
1192                 printf("WARNING: PID switch overriding CPU\n");
1193                 sleep(1);
1194                 top.cpu_list = NULL;
1195         }
1196
1197         if (top.target_pid != -1)
1198                 top.target_tid = top.target_pid;
1199
1200         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1201                                      top.target_tid, top.cpu_list) < 0)
1202                 usage_with_options(top_usage, options);
1203
1204         if (!top.evlist->nr_entries &&
1205             perf_evlist__add_default(top.evlist) < 0) {
1206                 pr_err("Not enough memory for event selector list\n");
1207                 return -ENOMEM;
1208         }
1209
1210         symbol_conf.nr_events = top.evlist->nr_entries;
1211
1212         if (top.delay_secs < 1)
1213                 top.delay_secs = 1;
1214
1215         /*
1216          * User specified count overrides default frequency.
1217          */
1218         if (top.default_interval)
1219                 top.freq = 0;
1220         else if (top.freq) {
1221                 top.default_interval = top.freq;
1222         } else {
1223                 fprintf(stderr, "frequency and count are zero, aborting\n");
1224                 exit(EXIT_FAILURE);
1225         }
1226
1227         list_for_each_entry(pos, &top.evlist->entries, node) {
1228                 /*
1229                  * Fill in the ones not specifically initialized via -c:
1230                  */
1231                 if (!pos->attr.sample_period)
1232                         pos->attr.sample_period = top.default_interval;
1233         }
1234
1235         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1236
1237         symbol_conf.priv_size = sizeof(struct annotation);
1238
1239         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1240         if (symbol__init() < 0)
1241                 return -1;
1242
1243         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1244         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1245         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1246
1247         /*
1248          * Avoid annotation data structures overhead when symbols aren't on the
1249          * sort list.
1250          */
1251         top.sort_has_symbols = sort_sym.list.next != NULL;
1252
1253         get_term_dimensions(&top.winsize);
1254         if (top.print_entries == 0) {
1255                 struct sigaction act = {
1256                         .sa_sigaction = perf_top__sig_winch,
1257                         .sa_flags     = SA_SIGINFO,
1258                 };
1259                 perf_top__update_print_entries(&top);
1260                 sigaction(SIGWINCH, &act, NULL);
1261         }
1262
1263         status = __cmd_top(&top);
1264
1265         perf_evlist__delete(top.evlist);
1266
1267         return status;
1268 }