Merge branch 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[~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         evsel->hists.stats.total_period += sample->period;
239         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
240         return he;
241 }
242
243 static void perf_top__print_sym_table(struct perf_top *top)
244 {
245         char bf[160];
246         int printed = 0;
247         const int win_width = top->winsize.ws_col - 1;
248
249         puts(CONSOLE_CLEAR);
250
251         perf_top__header_snprintf(top, bf, sizeof(bf));
252         printf("%s\n", bf);
253
254         perf_top__reset_sample_counters(top);
255
256         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
257
258         if (top->sym_evsel->hists.stats.nr_lost_warned !=
259             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
260                 top->sym_evsel->hists.stats.nr_lost_warned =
261                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
262                 color_fprintf(stdout, PERF_COLOR_RED,
263                               "WARNING: LOST %d chunks, Check IO/CPU overload",
264                               top->sym_evsel->hists.stats.nr_lost_warned);
265                 ++printed;
266         }
267
268         if (top->sym_filter_entry) {
269                 perf_top__show_details(top);
270                 return;
271         }
272
273         hists__collapse_resort_threaded(&top->sym_evsel->hists);
274         hists__output_resort_threaded(&top->sym_evsel->hists);
275         hists__decay_entries_threaded(&top->sym_evsel->hists,
276                                       top->hide_user_symbols,
277                                       top->hide_kernel_symbols);
278         hists__output_recalc_col_len(&top->sym_evsel->hists,
279                                      top->winsize.ws_row - 3);
280         putchar('\n');
281         hists__fprintf(&top->sym_evsel->hists, NULL, false, false,
282                        top->winsize.ws_row - 4 - printed, win_width, stdout);
283 }
284
285 static void prompt_integer(int *target, const char *msg)
286 {
287         char *buf = malloc(0), *p;
288         size_t dummy = 0;
289         int tmp;
290
291         fprintf(stdout, "\n%s: ", msg);
292         if (getline(&buf, &dummy, stdin) < 0)
293                 return;
294
295         p = strchr(buf, '\n');
296         if (p)
297                 *p = 0;
298
299         p = buf;
300         while(*p) {
301                 if (!isdigit(*p))
302                         goto out_free;
303                 p++;
304         }
305         tmp = strtoul(buf, NULL, 10);
306         *target = tmp;
307 out_free:
308         free(buf);
309 }
310
311 static void prompt_percent(int *target, const char *msg)
312 {
313         int tmp = 0;
314
315         prompt_integer(&tmp, msg);
316         if (tmp >= 0 && tmp <= 100)
317                 *target = tmp;
318 }
319
320 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
321 {
322         char *buf = malloc(0), *p;
323         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
324         struct rb_node *next;
325         size_t dummy = 0;
326
327         /* zero counters of active symbol */
328         if (syme) {
329                 __zero_source_counters(syme);
330                 top->sym_filter_entry = NULL;
331         }
332
333         fprintf(stdout, "\n%s: ", msg);
334         if (getline(&buf, &dummy, stdin) < 0)
335                 goto out_free;
336
337         p = strchr(buf, '\n');
338         if (p)
339                 *p = 0;
340
341         next = rb_first(&top->sym_evsel->hists.entries);
342         while (next) {
343                 n = rb_entry(next, struct hist_entry, rb_node);
344                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
345                         found = n;
346                         break;
347                 }
348                 next = rb_next(&n->rb_node);
349         }
350
351         if (!found) {
352                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
353                 sleep(1);
354         } else
355                 perf_top__parse_source(top, found);
356
357 out_free:
358         free(buf);
359 }
360
361 static void perf_top__print_mapped_keys(struct perf_top *top)
362 {
363         char *name = NULL;
364
365         if (top->sym_filter_entry) {
366                 struct symbol *sym = top->sym_filter_entry->ms.sym;
367                 name = sym->name;
368         }
369
370         fprintf(stdout, "\nMapped keys:\n");
371         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
372         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
373
374         if (top->evlist->nr_entries > 1)
375                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", event_name(top->sym_evsel));
376
377         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
378
379         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
380         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
381         fprintf(stdout, "\t[S]     stop annotation.\n");
382
383         fprintf(stdout,
384                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
385                 top->hide_kernel_symbols ? "yes" : "no");
386         fprintf(stdout,
387                 "\t[U]     hide user symbols.               \t(%s)\n",
388                 top->hide_user_symbols ? "yes" : "no");
389         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
390         fprintf(stdout, "\t[qQ]    quit.\n");
391 }
392
393 static int perf_top__key_mapped(struct perf_top *top, int c)
394 {
395         switch (c) {
396                 case 'd':
397                 case 'e':
398                 case 'f':
399                 case 'z':
400                 case 'q':
401                 case 'Q':
402                 case 'K':
403                 case 'U':
404                 case 'F':
405                 case 's':
406                 case 'S':
407                         return 1;
408                 case 'E':
409                         return top->evlist->nr_entries > 1 ? 1 : 0;
410                 default:
411                         break;
412         }
413
414         return 0;
415 }
416
417 static void perf_top__handle_keypress(struct perf_top *top, int c)
418 {
419         if (!perf_top__key_mapped(top, c)) {
420                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
421                 struct termios tc, save;
422
423                 perf_top__print_mapped_keys(top);
424                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
425                 fflush(stdout);
426
427                 tcgetattr(0, &save);
428                 tc = save;
429                 tc.c_lflag &= ~(ICANON | ECHO);
430                 tc.c_cc[VMIN] = 0;
431                 tc.c_cc[VTIME] = 0;
432                 tcsetattr(0, TCSANOW, &tc);
433
434                 poll(&stdin_poll, 1, -1);
435                 c = getc(stdin);
436
437                 tcsetattr(0, TCSAFLUSH, &save);
438                 if (!perf_top__key_mapped(top, c))
439                         return;
440         }
441
442         switch (c) {
443                 case 'd':
444                         prompt_integer(&top->delay_secs, "Enter display delay");
445                         if (top->delay_secs < 1)
446                                 top->delay_secs = 1;
447                         break;
448                 case 'e':
449                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
450                         if (top->print_entries == 0) {
451                                 struct sigaction act = {
452                                         .sa_sigaction = perf_top__sig_winch,
453                                         .sa_flags     = SA_SIGINFO,
454                                 };
455                                 perf_top__sig_winch(SIGWINCH, NULL, top);
456                                 sigaction(SIGWINCH, &act, NULL);
457                         } else
458                                 signal(SIGWINCH, SIG_DFL);
459                         break;
460                 case 'E':
461                         if (top->evlist->nr_entries > 1) {
462                                 /* Select 0 as the default event: */
463                                 int counter = 0;
464
465                                 fprintf(stderr, "\nAvailable events:");
466
467                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
468                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, event_name(top->sym_evsel));
469
470                                 prompt_integer(&counter, "Enter details event counter");
471
472                                 if (counter >= top->evlist->nr_entries) {
473                                         top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
474                                         fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(top->sym_evsel));
475                                         sleep(1);
476                                         break;
477                                 }
478                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
479                                         if (top->sym_evsel->idx == counter)
480                                                 break;
481                         } else
482                                 top->sym_evsel = list_entry(top->evlist->entries.next, struct perf_evsel, node);
483                         break;
484                 case 'f':
485                         prompt_integer(&top->count_filter, "Enter display event count filter");
486                         break;
487                 case 'F':
488                         prompt_percent(&top->sym_pcnt_filter,
489                                        "Enter details display event filter (percent)");
490                         break;
491                 case 'K':
492                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
493                         break;
494                 case 'q':
495                 case 'Q':
496                         printf("exiting.\n");
497                         if (top->dump_symtab)
498                                 perf_session__fprintf_dsos(top->session, stderr);
499                         exit(0);
500                 case 's':
501                         perf_top__prompt_symbol(top, "Enter details symbol");
502                         break;
503                 case 'S':
504                         if (!top->sym_filter_entry)
505                                 break;
506                         else {
507                                 struct hist_entry *syme = top->sym_filter_entry;
508
509                                 top->sym_filter_entry = NULL;
510                                 __zero_source_counters(syme);
511                         }
512                         break;
513                 case 'U':
514                         top->hide_user_symbols = !top->hide_user_symbols;
515                         break;
516                 case 'z':
517                         top->zero = !top->zero;
518                         break;
519                 default:
520                         break;
521         }
522 }
523
524 static void perf_top__sort_new_samples(void *arg)
525 {
526         struct perf_top *t = arg;
527         perf_top__reset_sample_counters(t);
528
529         if (t->evlist->selected != NULL)
530                 t->sym_evsel = t->evlist->selected;
531
532         hists__collapse_resort_threaded(&t->sym_evsel->hists);
533         hists__output_resort_threaded(&t->sym_evsel->hists);
534         hists__decay_entries_threaded(&t->sym_evsel->hists,
535                                       t->hide_user_symbols,
536                                       t->hide_kernel_symbols);
537 }
538
539 static void *display_thread_tui(void *arg)
540 {
541         struct perf_top *top = arg;
542         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
543
544         perf_top__sort_new_samples(top);
545         perf_evlist__tui_browse_hists(top->evlist, help,
546                                       perf_top__sort_new_samples,
547                                       top, top->delay_secs);
548
549         exit_browser(0);
550         exit(0);
551         return NULL;
552 }
553
554 static void *display_thread(void *arg)
555 {
556         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
557         struct termios tc, save;
558         struct perf_top *top = arg;
559         int delay_msecs, c;
560
561         tcgetattr(0, &save);
562         tc = save;
563         tc.c_lflag &= ~(ICANON | ECHO);
564         tc.c_cc[VMIN] = 0;
565         tc.c_cc[VTIME] = 0;
566
567         pthread__unblock_sigwinch();
568 repeat:
569         delay_msecs = top->delay_secs * 1000;
570         tcsetattr(0, TCSANOW, &tc);
571         /* trash return*/
572         getc(stdin);
573
574         while (1) {
575                 perf_top__print_sym_table(top);
576                 /*
577                  * Either timeout expired or we got an EINTR due to SIGWINCH,
578                  * refresh screen in both cases.
579                  */
580                 switch (poll(&stdin_poll, 1, delay_msecs)) {
581                 case 0:
582                         continue;
583                 case -1:
584                         if (errno == EINTR)
585                                 continue;
586                         /* Fall trhu */
587                 default:
588                         goto process_hotkey;
589                 }
590         }
591 process_hotkey:
592         c = getc(stdin);
593         tcsetattr(0, TCSAFLUSH, &save);
594
595         perf_top__handle_keypress(top, c);
596         goto repeat;
597
598         return NULL;
599 }
600
601 /* Tag samples to be skipped. */
602 static const char *skip_symbols[] = {
603         "default_idle",
604         "native_safe_halt",
605         "cpu_idle",
606         "enter_idle",
607         "exit_idle",
608         "mwait_idle",
609         "mwait_idle_with_hints",
610         "poll_idle",
611         "ppc64_runlatch_off",
612         "pseries_dedicated_idle_sleep",
613         NULL
614 };
615
616 static int symbol_filter(struct map *map __used, struct symbol *sym)
617 {
618         const char *name = sym->name;
619         int i;
620
621         /*
622          * ppc64 uses function descriptors and appends a '.' to the
623          * start of every instruction address. Remove it.
624          */
625         if (name[0] == '.')
626                 name++;
627
628         if (!strcmp(name, "_text") ||
629             !strcmp(name, "_etext") ||
630             !strcmp(name, "_sinittext") ||
631             !strncmp("init_module", name, 11) ||
632             !strncmp("cleanup_module", name, 14) ||
633             strstr(name, "_text_start") ||
634             strstr(name, "_text_end"))
635                 return 1;
636
637         for (i = 0; skip_symbols[i]; i++) {
638                 if (!strcmp(skip_symbols[i], name)) {
639                         sym->ignore = true;
640                         break;
641                 }
642         }
643
644         return 0;
645 }
646
647 static void perf_event__process_sample(struct perf_tool *tool,
648                                        const union perf_event *event,
649                                        struct perf_evsel *evsel,
650                                        struct perf_sample *sample,
651                                        struct machine *machine)
652 {
653         struct perf_top *top = container_of(tool, struct perf_top, tool);
654         struct symbol *parent = NULL;
655         u64 ip = event->ip.ip;
656         struct addr_location al;
657         int err;
658
659         if (!machine && perf_guest) {
660                 pr_err("Can't find guest [%d]'s kernel information\n",
661                         event->ip.pid);
662                 return;
663         }
664
665         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
666                 top->exact_samples++;
667
668         if (perf_event__preprocess_sample(event, machine, &al, sample,
669                                           symbol_filter) < 0 ||
670             al.filtered)
671                 return;
672
673         if (!top->kptr_restrict_warned &&
674             symbol_conf.kptr_restrict &&
675             al.cpumode == PERF_RECORD_MISC_KERNEL) {
676                 ui__warning(
677 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
678 "Check /proc/sys/kernel/kptr_restrict.\n\n"
679 "Kernel%s samples will not be resolved.\n",
680                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
681                           " modules" : "");
682                 if (use_browser <= 0)
683                         sleep(5);
684                 top->kptr_restrict_warned = true;
685         }
686
687         if (al.sym == NULL) {
688                 const char *msg = "Kernel samples will not be resolved.\n";
689                 /*
690                  * As we do lazy loading of symtabs we only will know if the
691                  * specified vmlinux file is invalid when we actually have a
692                  * hit in kernel space and then try to load it. So if we get
693                  * here and there are _no_ symbols in the DSO backing the
694                  * kernel map, bail out.
695                  *
696                  * We may never get here, for instance, if we use -K/
697                  * --hide-kernel-symbols, even if the user specifies an
698                  * invalid --vmlinux ;-)
699                  */
700                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
701                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
702                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
703                         if (symbol_conf.vmlinux_name) {
704                                 ui__warning("The %s file can't be used.\n%s",
705                                             symbol_conf.vmlinux_name, msg);
706                         } else {
707                                 ui__warning("A vmlinux file was not found.\n%s",
708                                             msg);
709                         }
710
711                         if (use_browser <= 0)
712                                 sleep(5);
713                         top->vmlinux_warned = true;
714                 }
715         }
716
717         if (al.sym == NULL || !al.sym->ignore) {
718                 struct hist_entry *he;
719
720                 if ((sort__has_parent || symbol_conf.use_callchain) &&
721                     sample->callchain) {
722                         err = machine__resolve_callchain(machine, evsel, al.thread,
723                                                          sample->callchain, &parent);
724                         if (err)
725                                 return;
726                 }
727
728                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
729                 if (he == NULL) {
730                         pr_err("Problem incrementing symbol period, skipping event\n");
731                         return;
732                 }
733
734                 if (symbol_conf.use_callchain) {
735                         err = callchain_append(he->callchain, &evsel->hists.callchain_cursor,
736                                                sample->period);
737                         if (err)
738                                 return;
739                 }
740
741                 if (top->sort_has_symbols)
742                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
743         }
744
745         return;
746 }
747
748 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
749 {
750         struct perf_sample sample;
751         struct perf_evsel *evsel;
752         struct perf_session *session = top->session;
753         union perf_event *event;
754         struct machine *machine;
755         u8 origin;
756         int ret;
757
758         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
759                 ret = perf_session__parse_sample(session, event, &sample);
760                 if (ret) {
761                         pr_err("Can't parse sample, err = %d\n", ret);
762                         continue;
763                 }
764
765                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
766                 assert(evsel != NULL);
767
768                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
769
770                 if (event->header.type == PERF_RECORD_SAMPLE)
771                         ++top->samples;
772
773                 switch (origin) {
774                 case PERF_RECORD_MISC_USER:
775                         ++top->us_samples;
776                         if (top->hide_user_symbols)
777                                 continue;
778                         machine = perf_session__find_host_machine(session);
779                         break;
780                 case PERF_RECORD_MISC_KERNEL:
781                         ++top->kernel_samples;
782                         if (top->hide_kernel_symbols)
783                                 continue;
784                         machine = perf_session__find_host_machine(session);
785                         break;
786                 case PERF_RECORD_MISC_GUEST_KERNEL:
787                         ++top->guest_kernel_samples;
788                         machine = perf_session__find_machine(session, event->ip.pid);
789                         break;
790                 case PERF_RECORD_MISC_GUEST_USER:
791                         ++top->guest_us_samples;
792                         /*
793                          * TODO: we don't process guest user from host side
794                          * except simple counting.
795                          */
796                         /* Fall thru */
797                 default:
798                         continue;
799                 }
800
801
802                 if (event->header.type == PERF_RECORD_SAMPLE) {
803                         perf_event__process_sample(&top->tool, event, evsel,
804                                                    &sample, machine);
805                 } else if (event->header.type < PERF_RECORD_MAX) {
806                         hists__inc_nr_events(&evsel->hists, event->header.type);
807                         perf_event__process(&top->tool, event, &sample, machine);
808                 } else
809                         ++session->hists.stats.nr_unknown_events;
810         }
811 }
812
813 static void perf_top__mmap_read(struct perf_top *top)
814 {
815         int i;
816
817         for (i = 0; i < top->evlist->nr_mmaps; i++)
818                 perf_top__mmap_read_idx(top, i);
819 }
820
821 static void perf_top__start_counters(struct perf_top *top)
822 {
823         struct perf_evsel *counter, *first;
824         struct perf_evlist *evlist = top->evlist;
825
826         first = list_entry(evlist->entries.next, struct perf_evsel, node);
827
828         list_for_each_entry(counter, &evlist->entries, node) {
829                 struct perf_event_attr *attr = &counter->attr;
830                 struct xyarray *group_fd = NULL;
831
832                 if (top->group && counter != first)
833                         group_fd = first->fd;
834
835                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
836
837                 if (top->freq) {
838                         attr->sample_type |= PERF_SAMPLE_PERIOD;
839                         attr->freq        = 1;
840                         attr->sample_freq = top->freq;
841                 }
842
843                 if (evlist->nr_entries > 1) {
844                         attr->sample_type |= PERF_SAMPLE_ID;
845                         attr->read_format |= PERF_FORMAT_ID;
846                 }
847
848                 if (symbol_conf.use_callchain)
849                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
850
851                 attr->mmap = 1;
852                 attr->comm = 1;
853                 attr->inherit = top->inherit;
854 retry_sample_id:
855                 attr->sample_id_all = top->sample_id_all_avail ? 1 : 0;
856 try_again:
857                 if (perf_evsel__open(counter, top->evlist->cpus,
858                                      top->evlist->threads, top->group,
859                                      group_fd) < 0) {
860                         int err = errno;
861
862                         if (err == EPERM || err == EACCES) {
863                                 ui__error_paranoid();
864                                 goto out_err;
865                         } else if (err == EINVAL && top->sample_id_all_avail) {
866                                 /*
867                                  * Old kernel, no attr->sample_id_type_all field
868                                  */
869                                 top->sample_id_all_avail = false;
870                                 goto retry_sample_id;
871                         }
872                         /*
873                          * If it's cycles then fall back to hrtimer
874                          * based cpu-clock-tick sw counter, which
875                          * is always available even if no PMU support:
876                          */
877                         if (attr->type == PERF_TYPE_HARDWARE &&
878                             attr->config == PERF_COUNT_HW_CPU_CYCLES) {
879                                 if (verbose)
880                                         ui__warning("Cycles event not supported,\n"
881                                                     "trying to fall back to cpu-clock-ticks\n");
882
883                                 attr->type = PERF_TYPE_SOFTWARE;
884                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
885                                 goto try_again;
886                         }
887
888                         if (err == ENOENT) {
889                                 ui__warning("The %s event is not supported.\n",
890                                             event_name(counter));
891                                 goto out_err;
892                         }
893
894                         ui__warning("The sys_perf_event_open() syscall "
895                                     "returned with %d (%s).  /bin/dmesg "
896                                     "may provide additional information.\n"
897                                     "No CONFIG_PERF_EVENTS=y kernel support "
898                                     "configured?\n", err, strerror(err));
899                         goto out_err;
900                 }
901         }
902
903         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
904                 ui__warning("Failed to mmap with %d (%s)\n",
905                             errno, strerror(errno));
906                 goto out_err;
907         }
908
909         return;
910
911 out_err:
912         exit_browser(0);
913         exit(0);
914 }
915
916 static int perf_top__setup_sample_type(struct perf_top *top)
917 {
918         if (!top->sort_has_symbols) {
919                 if (symbol_conf.use_callchain) {
920                         ui__warning("Selected -g but \"sym\" not present in --sort/-s.");
921                         return -EINVAL;
922                 }
923         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
924                 if (callchain_register_param(&callchain_param) < 0) {
925                         ui__warning("Can't register callchain params.\n");
926                         return -EINVAL;
927                 }
928         }
929
930         return 0;
931 }
932
933 static int __cmd_top(struct perf_top *top)
934 {
935         pthread_t thread;
936         int ret;
937         /*
938          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
939          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
940          */
941         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
942         if (top->session == NULL)
943                 return -ENOMEM;
944
945         ret = perf_top__setup_sample_type(top);
946         if (ret)
947                 goto out_delete;
948
949         if (top->target_tid != -1)
950                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
951                                                   perf_event__process,
952                                                   &top->session->host_machine);
953         else
954                 perf_event__synthesize_threads(&top->tool, perf_event__process,
955                                                &top->session->host_machine);
956         perf_top__start_counters(top);
957         top->session->evlist = top->evlist;
958         perf_session__update_sample_type(top->session);
959
960         /* Wait for a minimal set of events before starting the snapshot */
961         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
962
963         perf_top__mmap_read(top);
964
965         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
966                                                             display_thread), top)) {
967                 printf("Could not create display thread.\n");
968                 exit(-1);
969         }
970
971         if (top->realtime_prio) {
972                 struct sched_param param;
973
974                 param.sched_priority = top->realtime_prio;
975                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
976                         printf("Could not set realtime priority.\n");
977                         exit(-1);
978                 }
979         }
980
981         while (1) {
982                 u64 hits = top->samples;
983
984                 perf_top__mmap_read(top);
985
986                 if (hits == top->samples)
987                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
988         }
989
990 out_delete:
991         perf_session__delete(top->session);
992         top->session = NULL;
993
994         return 0;
995 }
996
997 static int
998 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
999 {
1000         struct perf_top *top = (struct perf_top *)opt->value;
1001         char *tok, *tok2;
1002         char *endptr;
1003
1004         /*
1005          * --no-call-graph
1006          */
1007         if (unset) {
1008                 top->dont_use_callchains = true;
1009                 return 0;
1010         }
1011
1012         symbol_conf.use_callchain = true;
1013
1014         if (!arg)
1015                 return 0;
1016
1017         tok = strtok((char *)arg, ",");
1018         if (!tok)
1019                 return -1;
1020
1021         /* get the output mode */
1022         if (!strncmp(tok, "graph", strlen(arg)))
1023                 callchain_param.mode = CHAIN_GRAPH_ABS;
1024
1025         else if (!strncmp(tok, "flat", strlen(arg)))
1026                 callchain_param.mode = CHAIN_FLAT;
1027
1028         else if (!strncmp(tok, "fractal", strlen(arg)))
1029                 callchain_param.mode = CHAIN_GRAPH_REL;
1030
1031         else if (!strncmp(tok, "none", strlen(arg))) {
1032                 callchain_param.mode = CHAIN_NONE;
1033                 symbol_conf.use_callchain = false;
1034
1035                 return 0;
1036         } else
1037                 return -1;
1038
1039         /* get the min percentage */
1040         tok = strtok(NULL, ",");
1041         if (!tok)
1042                 goto setup;
1043
1044         callchain_param.min_percent = strtod(tok, &endptr);
1045         if (tok == endptr)
1046                 return -1;
1047
1048         /* get the print limit */
1049         tok2 = strtok(NULL, ",");
1050         if (!tok2)
1051                 goto setup;
1052
1053         if (tok2[0] != 'c') {
1054                 callchain_param.print_limit = strtod(tok2, &endptr);
1055                 tok2 = strtok(NULL, ",");
1056                 if (!tok2)
1057                         goto setup;
1058         }
1059
1060         /* get the call chain order */
1061         if (!strcmp(tok2, "caller"))
1062                 callchain_param.order = ORDER_CALLER;
1063         else if (!strcmp(tok2, "callee"))
1064                 callchain_param.order = ORDER_CALLEE;
1065         else
1066                 return -1;
1067 setup:
1068         if (callchain_register_param(&callchain_param) < 0) {
1069                 fprintf(stderr, "Can't register callchain params\n");
1070                 return -1;
1071         }
1072         return 0;
1073 }
1074
1075 static const char * const top_usage[] = {
1076         "perf top [<options>]",
1077         NULL
1078 };
1079
1080 int cmd_top(int argc, const char **argv, const char *prefix __used)
1081 {
1082         struct perf_evsel *pos;
1083         int status = -ENOMEM;
1084         struct perf_top top = {
1085                 .count_filter        = 5,
1086                 .delay_secs          = 2,
1087                 .target_pid          = -1,
1088                 .target_tid          = -1,
1089                 .freq                = 1000, /* 1 KHz */
1090                 .sample_id_all_avail = true,
1091                 .mmap_pages          = 128,
1092                 .sym_pcnt_filter     = 5,
1093         };
1094         char callchain_default_opt[] = "fractal,0.5,callee";
1095         const struct option options[] = {
1096         OPT_CALLBACK('e', "event", &top.evlist, "event",
1097                      "event selector. use 'perf list' to list available events",
1098                      parse_events_option),
1099         OPT_INTEGER('c', "count", &top.default_interval,
1100                     "event period to sample"),
1101         OPT_INTEGER('p', "pid", &top.target_pid,
1102                     "profile events on existing process id"),
1103         OPT_INTEGER('t', "tid", &top.target_tid,
1104                     "profile events on existing thread id"),
1105         OPT_BOOLEAN('a', "all-cpus", &top.system_wide,
1106                             "system-wide collection from all CPUs"),
1107         OPT_STRING('C', "cpu", &top.cpu_list, "cpu",
1108                     "list of cpus to monitor"),
1109         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1110                    "file", "vmlinux pathname"),
1111         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1112                     "hide kernel symbols"),
1113         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1114         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1115                     "collect data with this RT SCHED_FIFO priority"),
1116         OPT_INTEGER('d', "delay", &top.delay_secs,
1117                     "number of seconds to delay between refreshes"),
1118         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1119                             "dump the symbol table used for profiling"),
1120         OPT_INTEGER('f', "count-filter", &top.count_filter,
1121                     "only display functions with more events than this"),
1122         OPT_BOOLEAN('g', "group", &top.group,
1123                             "put the counters into a counter group"),
1124         OPT_BOOLEAN('i', "inherit", &top.inherit,
1125                     "child tasks inherit counters"),
1126         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1127                     "symbol to annotate"),
1128         OPT_BOOLEAN('z', "zero", &top.zero,
1129                     "zero history across updates"),
1130         OPT_INTEGER('F', "freq", &top.freq,
1131                     "profile at this frequency"),
1132         OPT_INTEGER('E', "entries", &top.print_entries,
1133                     "display this many functions"),
1134         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1135                     "hide user symbols"),
1136         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1137         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1138         OPT_INCR('v', "verbose", &verbose,
1139                     "be more verbose (show counter open errors, etc)"),
1140         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1141                    "sort by key(s): pid, comm, dso, symbol, parent"),
1142         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1143                     "Show a column with the number of samples"),
1144         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1145                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1146                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1147                      callchain_default_opt),
1148         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1149                     "Show a column with the sum of periods"),
1150         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1151                    "only consider symbols in these dsos"),
1152         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1153                    "only consider symbols in these comms"),
1154         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1155                    "only consider these symbols"),
1156         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1157                     "Interleave source code with assembly code (default)"),
1158         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1159                     "Display raw encoding of assembly instructions (default)"),
1160         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1161                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1162         OPT_END()
1163         };
1164
1165         top.evlist = perf_evlist__new(NULL, NULL);
1166         if (top.evlist == NULL)
1167                 return -ENOMEM;
1168
1169         symbol_conf.exclude_other = false;
1170
1171         argc = parse_options(argc, argv, options, top_usage, 0);
1172         if (argc)
1173                 usage_with_options(top_usage, options);
1174
1175         if (sort_order == default_sort_order)
1176                 sort_order = "dso,symbol";
1177
1178         setup_sorting(top_usage, options);
1179
1180         if (top.use_stdio)
1181                 use_browser = 0;
1182         else if (top.use_tui)
1183                 use_browser = 1;
1184
1185         setup_browser(false);
1186
1187         /* CPU and PID are mutually exclusive */
1188         if (top.target_tid > 0 && top.cpu_list) {
1189                 printf("WARNING: PID switch overriding CPU\n");
1190                 sleep(1);
1191                 top.cpu_list = NULL;
1192         }
1193
1194         if (top.target_pid != -1)
1195                 top.target_tid = top.target_pid;
1196
1197         if (perf_evlist__create_maps(top.evlist, top.target_pid,
1198                                      top.target_tid, top.cpu_list) < 0)
1199                 usage_with_options(top_usage, options);
1200
1201         if (!top.evlist->nr_entries &&
1202             perf_evlist__add_default(top.evlist) < 0) {
1203                 pr_err("Not enough memory for event selector list\n");
1204                 return -ENOMEM;
1205         }
1206
1207         symbol_conf.nr_events = top.evlist->nr_entries;
1208
1209         if (top.delay_secs < 1)
1210                 top.delay_secs = 1;
1211
1212         /*
1213          * User specified count overrides default frequency.
1214          */
1215         if (top.default_interval)
1216                 top.freq = 0;
1217         else if (top.freq) {
1218                 top.default_interval = top.freq;
1219         } else {
1220                 fprintf(stderr, "frequency and count are zero, aborting\n");
1221                 exit(EXIT_FAILURE);
1222         }
1223
1224         list_for_each_entry(pos, &top.evlist->entries, node) {
1225                 /*
1226                  * Fill in the ones not specifically initialized via -c:
1227                  */
1228                 if (!pos->attr.sample_period)
1229                         pos->attr.sample_period = top.default_interval;
1230         }
1231
1232         top.sym_evsel = list_entry(top.evlist->entries.next, struct perf_evsel, node);
1233
1234         symbol_conf.priv_size = sizeof(struct annotation);
1235
1236         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1237         if (symbol__init() < 0)
1238                 return -1;
1239
1240         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1241         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1242         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1243
1244         /*
1245          * Avoid annotation data structures overhead when symbols aren't on the
1246          * sort list.
1247          */
1248         top.sort_has_symbols = sort_sym.list.next != NULL;
1249
1250         get_term_dimensions(&top.winsize);
1251         if (top.print_entries == 0) {
1252                 struct sigaction act = {
1253                         .sa_sigaction = perf_top__sig_winch,
1254                         .sa_flags     = SA_SIGINFO,
1255                 };
1256                 perf_top__update_print_entries(&top);
1257                 sigaction(SIGWINCH, &act, NULL);
1258         }
1259
1260         status = __cmd_top(&top);
1261
1262         perf_evlist__delete(top.evlist);
1263
1264         return status;
1265 }