c9ff3950cd4be55a487ff6cbda76b4f55c47ad5d
[~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/machine.h"
30 #include "util/session.h"
31 #include "util/symbol.h"
32 #include "util/thread.h"
33 #include "util/thread_map.h"
34 #include "util/top.h"
35 #include "util/util.h"
36 #include <linux/rbtree.h>
37 #include "util/parse-options.h"
38 #include "util/parse-events.h"
39 #include "util/cpumap.h"
40 #include "util/xyarray.h"
41 #include "util/sort.h"
42 #include "util/intlist.h"
43
44 #include "util/debug.h"
45
46 #include <assert.h>
47 #include <elf.h>
48 #include <fcntl.h>
49
50 #include <stdio.h>
51 #include <termios.h>
52 #include <unistd.h>
53 #include <inttypes.h>
54
55 #include <errno.h>
56 #include <time.h>
57 #include <sched.h>
58
59 #include <sys/syscall.h>
60 #include <sys/ioctl.h>
61 #include <sys/poll.h>
62 #include <sys/prctl.h>
63 #include <sys/wait.h>
64 #include <sys/uio.h>
65 #include <sys/utsname.h>
66 #include <sys/mman.h>
67
68 #include <linux/unistd.h>
69 #include <linux/types.h>
70
71 void get_term_dimensions(struct winsize *ws)
72 {
73         char *s = getenv("LINES");
74
75         if (s != NULL) {
76                 ws->ws_row = atoi(s);
77                 s = getenv("COLUMNS");
78                 if (s != NULL) {
79                         ws->ws_col = atoi(s);
80                         if (ws->ws_row && ws->ws_col)
81                                 return;
82                 }
83         }
84 #ifdef TIOCGWINSZ
85         if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
86             ws->ws_row && ws->ws_col)
87                 return;
88 #endif
89         ws->ws_row = 25;
90         ws->ws_col = 80;
91 }
92
93 static void perf_top__update_print_entries(struct perf_top *top)
94 {
95         if (top->print_entries > 9)
96                 top->print_entries -= 9;
97 }
98
99 static void perf_top__sig_winch(int sig __maybe_unused,
100                                 siginfo_t *info __maybe_unused, void *arg)
101 {
102         struct perf_top *top = arg;
103
104         get_term_dimensions(&top->winsize);
105         if (!top->print_entries
106             || (top->print_entries+4) > top->winsize.ws_row) {
107                 top->print_entries = top->winsize.ws_row;
108         } else {
109                 top->print_entries += 4;
110                 top->winsize.ws_row = top->print_entries;
111         }
112         perf_top__update_print_entries(top);
113 }
114
115 static int perf_top__parse_source(struct perf_top *top, struct hist_entry *he)
116 {
117         struct symbol *sym;
118         struct annotation *notes;
119         struct map *map;
120         int err = -1;
121
122         if (!he || !he->ms.sym)
123                 return -1;
124
125         sym = he->ms.sym;
126         map = he->ms.map;
127
128         /*
129          * We can't annotate with just /proc/kallsyms
130          */
131         if (map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS) {
132                 pr_err("Can't annotate %s: No vmlinux file was found in the "
133                        "path\n", sym->name);
134                 sleep(1);
135                 return -1;
136         }
137
138         notes = symbol__annotation(sym);
139         if (notes->src != NULL) {
140                 pthread_mutex_lock(&notes->lock);
141                 goto out_assign;
142         }
143
144         pthread_mutex_lock(&notes->lock);
145
146         if (symbol__alloc_hist(sym) < 0) {
147                 pthread_mutex_unlock(&notes->lock);
148                 pr_err("Not enough memory for annotating '%s' symbol!\n",
149                        sym->name);
150                 sleep(1);
151                 return err;
152         }
153
154         err = symbol__annotate(sym, map, 0);
155         if (err == 0) {
156 out_assign:
157                 top->sym_filter_entry = he;
158         }
159
160         pthread_mutex_unlock(&notes->lock);
161         return err;
162 }
163
164 static void __zero_source_counters(struct hist_entry *he)
165 {
166         struct symbol *sym = he->ms.sym;
167         symbol__annotate_zero_histograms(sym);
168 }
169
170 static void ui__warn_map_erange(struct map *map, struct symbol *sym, u64 ip)
171 {
172         struct utsname uts;
173         int err = uname(&uts);
174
175         ui__warning("Out of bounds address found:\n\n"
176                     "Addr:   %" PRIx64 "\n"
177                     "DSO:    %s %c\n"
178                     "Map:    %" PRIx64 "-%" PRIx64 "\n"
179                     "Symbol: %" PRIx64 "-%" PRIx64 " %c %s\n"
180                     "Arch:   %s\n"
181                     "Kernel: %s\n"
182                     "Tools:  %s\n\n"
183                     "Not all samples will be on the annotation output.\n\n"
184                     "Please report to linux-kernel@vger.kernel.org\n",
185                     ip, map->dso->long_name, dso__symtab_origin(map->dso),
186                     map->start, map->end, sym->start, sym->end,
187                     sym->binding == STB_GLOBAL ? 'g' :
188                     sym->binding == STB_LOCAL  ? 'l' : 'w', sym->name,
189                     err ? "[unknown]" : uts.machine,
190                     err ? "[unknown]" : uts.release, perf_version_string);
191         if (use_browser <= 0)
192                 sleep(5);
193         
194         map->erange_warned = true;
195 }
196
197 static void perf_top__record_precise_ip(struct perf_top *top,
198                                         struct hist_entry *he,
199                                         int counter, u64 ip)
200 {
201         struct annotation *notes;
202         struct symbol *sym;
203         int err;
204
205         if (he == NULL || he->ms.sym == NULL ||
206             ((top->sym_filter_entry == NULL ||
207               top->sym_filter_entry->ms.sym != he->ms.sym) && use_browser != 1))
208                 return;
209
210         sym = he->ms.sym;
211         notes = symbol__annotation(sym);
212
213         if (pthread_mutex_trylock(&notes->lock))
214                 return;
215
216         if (notes->src == NULL && symbol__alloc_hist(sym) < 0) {
217                 pthread_mutex_unlock(&notes->lock);
218                 pr_err("Not enough memory for annotating '%s' symbol!\n",
219                        sym->name);
220                 sleep(1);
221                 return;
222         }
223
224         ip = he->ms.map->map_ip(he->ms.map, ip);
225         err = symbol__inc_addr_samples(sym, he->ms.map, counter, ip);
226
227         pthread_mutex_unlock(&notes->lock);
228
229         if (err == -ERANGE && !he->ms.map->erange_warned)
230                 ui__warn_map_erange(he->ms.map, sym, ip);
231 }
232
233 static void perf_top__show_details(struct perf_top *top)
234 {
235         struct hist_entry *he = top->sym_filter_entry;
236         struct annotation *notes;
237         struct symbol *symbol;
238         int more;
239
240         if (!he)
241                 return;
242
243         symbol = he->ms.sym;
244         notes = symbol__annotation(symbol);
245
246         pthread_mutex_lock(&notes->lock);
247
248         if (notes->src == NULL)
249                 goto out_unlock;
250
251         printf("Showing %s for %s\n", perf_evsel__name(top->sym_evsel), symbol->name);
252         printf("  Events  Pcnt (>=%d%%)\n", top->sym_pcnt_filter);
253
254         more = symbol__annotate_printf(symbol, he->ms.map, top->sym_evsel->idx,
255                                        0, top->sym_pcnt_filter, top->print_entries, 4);
256         if (top->zero)
257                 symbol__annotate_zero_histogram(symbol, top->sym_evsel->idx);
258         else
259                 symbol__annotate_decay_histogram(symbol, top->sym_evsel->idx);
260         if (more != 0)
261                 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
262 out_unlock:
263         pthread_mutex_unlock(&notes->lock);
264 }
265
266 static const char               CONSOLE_CLEAR[] = "\e[H\e[2J";
267
268 static struct hist_entry *perf_evsel__add_hist_entry(struct perf_evsel *evsel,
269                                                      struct addr_location *al,
270                                                      struct perf_sample *sample)
271 {
272         struct hist_entry *he;
273
274         he = __hists__add_entry(&evsel->hists, al, NULL, sample->period);
275         if (he == NULL)
276                 return NULL;
277
278         hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
279         return he;
280 }
281
282 static void perf_top__print_sym_table(struct perf_top *top)
283 {
284         char bf[160];
285         int printed = 0;
286         const int win_width = top->winsize.ws_col - 1;
287
288         puts(CONSOLE_CLEAR);
289
290         perf_top__header_snprintf(top, bf, sizeof(bf));
291         printf("%s\n", bf);
292
293         perf_top__reset_sample_counters(top);
294
295         printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
296
297         if (top->sym_evsel->hists.stats.nr_lost_warned !=
298             top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST]) {
299                 top->sym_evsel->hists.stats.nr_lost_warned =
300                         top->sym_evsel->hists.stats.nr_events[PERF_RECORD_LOST];
301                 color_fprintf(stdout, PERF_COLOR_RED,
302                               "WARNING: LOST %d chunks, Check IO/CPU overload",
303                               top->sym_evsel->hists.stats.nr_lost_warned);
304                 ++printed;
305         }
306
307         if (top->sym_filter_entry) {
308                 perf_top__show_details(top);
309                 return;
310         }
311
312         hists__collapse_resort_threaded(&top->sym_evsel->hists);
313         hists__output_resort_threaded(&top->sym_evsel->hists);
314         hists__decay_entries_threaded(&top->sym_evsel->hists,
315                                       top->hide_user_symbols,
316                                       top->hide_kernel_symbols);
317         hists__output_recalc_col_len(&top->sym_evsel->hists,
318                                      top->winsize.ws_row - 3);
319         putchar('\n');
320         hists__fprintf(&top->sym_evsel->hists, false,
321                        top->winsize.ws_row - 4 - printed, win_width, stdout);
322 }
323
324 static void prompt_integer(int *target, const char *msg)
325 {
326         char *buf = malloc(0), *p;
327         size_t dummy = 0;
328         int tmp;
329
330         fprintf(stdout, "\n%s: ", msg);
331         if (getline(&buf, &dummy, stdin) < 0)
332                 return;
333
334         p = strchr(buf, '\n');
335         if (p)
336                 *p = 0;
337
338         p = buf;
339         while(*p) {
340                 if (!isdigit(*p))
341                         goto out_free;
342                 p++;
343         }
344         tmp = strtoul(buf, NULL, 10);
345         *target = tmp;
346 out_free:
347         free(buf);
348 }
349
350 static void prompt_percent(int *target, const char *msg)
351 {
352         int tmp = 0;
353
354         prompt_integer(&tmp, msg);
355         if (tmp >= 0 && tmp <= 100)
356                 *target = tmp;
357 }
358
359 static void perf_top__prompt_symbol(struct perf_top *top, const char *msg)
360 {
361         char *buf = malloc(0), *p;
362         struct hist_entry *syme = top->sym_filter_entry, *n, *found = NULL;
363         struct rb_node *next;
364         size_t dummy = 0;
365
366         /* zero counters of active symbol */
367         if (syme) {
368                 __zero_source_counters(syme);
369                 top->sym_filter_entry = NULL;
370         }
371
372         fprintf(stdout, "\n%s: ", msg);
373         if (getline(&buf, &dummy, stdin) < 0)
374                 goto out_free;
375
376         p = strchr(buf, '\n');
377         if (p)
378                 *p = 0;
379
380         next = rb_first(&top->sym_evsel->hists.entries);
381         while (next) {
382                 n = rb_entry(next, struct hist_entry, rb_node);
383                 if (n->ms.sym && !strcmp(buf, n->ms.sym->name)) {
384                         found = n;
385                         break;
386                 }
387                 next = rb_next(&n->rb_node);
388         }
389
390         if (!found) {
391                 fprintf(stderr, "Sorry, %s is not active.\n", buf);
392                 sleep(1);
393         } else
394                 perf_top__parse_source(top, found);
395
396 out_free:
397         free(buf);
398 }
399
400 static void perf_top__print_mapped_keys(struct perf_top *top)
401 {
402         char *name = NULL;
403
404         if (top->sym_filter_entry) {
405                 struct symbol *sym = top->sym_filter_entry->ms.sym;
406                 name = sym->name;
407         }
408
409         fprintf(stdout, "\nMapped keys:\n");
410         fprintf(stdout, "\t[d]     display refresh delay.             \t(%d)\n", top->delay_secs);
411         fprintf(stdout, "\t[e]     display entries (lines).           \t(%d)\n", top->print_entries);
412
413         if (top->evlist->nr_entries > 1)
414                 fprintf(stdout, "\t[E]     active event counter.              \t(%s)\n", perf_evsel__name(top->sym_evsel));
415
416         fprintf(stdout, "\t[f]     profile display filter (count).    \t(%d)\n", top->count_filter);
417
418         fprintf(stdout, "\t[F]     annotate display filter (percent). \t(%d%%)\n", top->sym_pcnt_filter);
419         fprintf(stdout, "\t[s]     annotate symbol.                   \t(%s)\n", name?: "NULL");
420         fprintf(stdout, "\t[S]     stop annotation.\n");
421
422         fprintf(stdout,
423                 "\t[K]     hide kernel_symbols symbols.     \t(%s)\n",
424                 top->hide_kernel_symbols ? "yes" : "no");
425         fprintf(stdout,
426                 "\t[U]     hide user symbols.               \t(%s)\n",
427                 top->hide_user_symbols ? "yes" : "no");
428         fprintf(stdout, "\t[z]     toggle sample zeroing.             \t(%d)\n", top->zero ? 1 : 0);
429         fprintf(stdout, "\t[qQ]    quit.\n");
430 }
431
432 static int perf_top__key_mapped(struct perf_top *top, int c)
433 {
434         switch (c) {
435                 case 'd':
436                 case 'e':
437                 case 'f':
438                 case 'z':
439                 case 'q':
440                 case 'Q':
441                 case 'K':
442                 case 'U':
443                 case 'F':
444                 case 's':
445                 case 'S':
446                         return 1;
447                 case 'E':
448                         return top->evlist->nr_entries > 1 ? 1 : 0;
449                 default:
450                         break;
451         }
452
453         return 0;
454 }
455
456 static void perf_top__handle_keypress(struct perf_top *top, int c)
457 {
458         if (!perf_top__key_mapped(top, c)) {
459                 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
460                 struct termios tc, save;
461
462                 perf_top__print_mapped_keys(top);
463                 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
464                 fflush(stdout);
465
466                 tcgetattr(0, &save);
467                 tc = save;
468                 tc.c_lflag &= ~(ICANON | ECHO);
469                 tc.c_cc[VMIN] = 0;
470                 tc.c_cc[VTIME] = 0;
471                 tcsetattr(0, TCSANOW, &tc);
472
473                 poll(&stdin_poll, 1, -1);
474                 c = getc(stdin);
475
476                 tcsetattr(0, TCSAFLUSH, &save);
477                 if (!perf_top__key_mapped(top, c))
478                         return;
479         }
480
481         switch (c) {
482                 case 'd':
483                         prompt_integer(&top->delay_secs, "Enter display delay");
484                         if (top->delay_secs < 1)
485                                 top->delay_secs = 1;
486                         break;
487                 case 'e':
488                         prompt_integer(&top->print_entries, "Enter display entries (lines)");
489                         if (top->print_entries == 0) {
490                                 struct sigaction act = {
491                                         .sa_sigaction = perf_top__sig_winch,
492                                         .sa_flags     = SA_SIGINFO,
493                                 };
494                                 perf_top__sig_winch(SIGWINCH, NULL, top);
495                                 sigaction(SIGWINCH, &act, NULL);
496                         } else {
497                                 perf_top__sig_winch(SIGWINCH, NULL, top);
498                                 signal(SIGWINCH, SIG_DFL);
499                         }
500                         break;
501                 case 'E':
502                         if (top->evlist->nr_entries > 1) {
503                                 /* Select 0 as the default event: */
504                                 int counter = 0;
505
506                                 fprintf(stderr, "\nAvailable events:");
507
508                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
509                                         fprintf(stderr, "\n\t%d %s", top->sym_evsel->idx, perf_evsel__name(top->sym_evsel));
510
511                                 prompt_integer(&counter, "Enter details event counter");
512
513                                 if (counter >= top->evlist->nr_entries) {
514                                         top->sym_evsel = perf_evlist__first(top->evlist);
515                                         fprintf(stderr, "Sorry, no such event, using %s.\n", perf_evsel__name(top->sym_evsel));
516                                         sleep(1);
517                                         break;
518                                 }
519                                 list_for_each_entry(top->sym_evsel, &top->evlist->entries, node)
520                                         if (top->sym_evsel->idx == counter)
521                                                 break;
522                         } else
523                                 top->sym_evsel = perf_evlist__first(top->evlist);
524                         break;
525                 case 'f':
526                         prompt_integer(&top->count_filter, "Enter display event count filter");
527                         break;
528                 case 'F':
529                         prompt_percent(&top->sym_pcnt_filter,
530                                        "Enter details display event filter (percent)");
531                         break;
532                 case 'K':
533                         top->hide_kernel_symbols = !top->hide_kernel_symbols;
534                         break;
535                 case 'q':
536                 case 'Q':
537                         printf("exiting.\n");
538                         if (top->dump_symtab)
539                                 perf_session__fprintf_dsos(top->session, stderr);
540                         exit(0);
541                 case 's':
542                         perf_top__prompt_symbol(top, "Enter details symbol");
543                         break;
544                 case 'S':
545                         if (!top->sym_filter_entry)
546                                 break;
547                         else {
548                                 struct hist_entry *syme = top->sym_filter_entry;
549
550                                 top->sym_filter_entry = NULL;
551                                 __zero_source_counters(syme);
552                         }
553                         break;
554                 case 'U':
555                         top->hide_user_symbols = !top->hide_user_symbols;
556                         break;
557                 case 'z':
558                         top->zero = !top->zero;
559                         break;
560                 default:
561                         break;
562         }
563 }
564
565 static void perf_top__sort_new_samples(void *arg)
566 {
567         struct perf_top *t = arg;
568         perf_top__reset_sample_counters(t);
569
570         if (t->evlist->selected != NULL)
571                 t->sym_evsel = t->evlist->selected;
572
573         hists__collapse_resort_threaded(&t->sym_evsel->hists);
574         hists__output_resort_threaded(&t->sym_evsel->hists);
575         hists__decay_entries_threaded(&t->sym_evsel->hists,
576                                       t->hide_user_symbols,
577                                       t->hide_kernel_symbols);
578 }
579
580 static void *display_thread_tui(void *arg)
581 {
582         struct perf_evsel *pos;
583         struct perf_top *top = arg;
584         const char *help = "For a higher level overview, try: perf top --sort comm,dso";
585         struct hist_browser_timer hbt = {
586                 .timer          = perf_top__sort_new_samples,
587                 .arg            = top,
588                 .refresh        = top->delay_secs,
589         };
590
591         perf_top__sort_new_samples(top);
592
593         /*
594          * Initialize the uid_filter_str, in the future the TUI will allow
595          * Zooming in/out UIDs. For now juse use whatever the user passed
596          * via --uid.
597          */
598         list_for_each_entry(pos, &top->evlist->entries, node)
599                 pos->hists.uid_filter_str = top->target.uid_str;
600
601         perf_evlist__tui_browse_hists(top->evlist, help, &hbt,
602                                       &top->session->header.env);
603
604         exit_browser(0);
605         exit(0);
606         return NULL;
607 }
608
609 static void *display_thread(void *arg)
610 {
611         struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
612         struct termios tc, save;
613         struct perf_top *top = arg;
614         int delay_msecs, c;
615
616         tcgetattr(0, &save);
617         tc = save;
618         tc.c_lflag &= ~(ICANON | ECHO);
619         tc.c_cc[VMIN] = 0;
620         tc.c_cc[VTIME] = 0;
621
622         pthread__unblock_sigwinch();
623 repeat:
624         delay_msecs = top->delay_secs * 1000;
625         tcsetattr(0, TCSANOW, &tc);
626         /* trash return*/
627         getc(stdin);
628
629         while (1) {
630                 perf_top__print_sym_table(top);
631                 /*
632                  * Either timeout expired or we got an EINTR due to SIGWINCH,
633                  * refresh screen in both cases.
634                  */
635                 switch (poll(&stdin_poll, 1, delay_msecs)) {
636                 case 0:
637                         continue;
638                 case -1:
639                         if (errno == EINTR)
640                                 continue;
641                         /* Fall trhu */
642                 default:
643                         goto process_hotkey;
644                 }
645         }
646 process_hotkey:
647         c = getc(stdin);
648         tcsetattr(0, TCSAFLUSH, &save);
649
650         perf_top__handle_keypress(top, c);
651         goto repeat;
652
653         return NULL;
654 }
655
656 /* Tag samples to be skipped. */
657 static const char *skip_symbols[] = {
658         "intel_idle",
659         "default_idle",
660         "native_safe_halt",
661         "cpu_idle",
662         "enter_idle",
663         "exit_idle",
664         "mwait_idle",
665         "mwait_idle_with_hints",
666         "poll_idle",
667         "ppc64_runlatch_off",
668         "pseries_dedicated_idle_sleep",
669         NULL
670 };
671
672 static int symbol_filter(struct map *map __maybe_unused, struct symbol *sym)
673 {
674         const char *name = sym->name;
675         int i;
676
677         /*
678          * ppc64 uses function descriptors and appends a '.' to the
679          * start of every instruction address. Remove it.
680          */
681         if (name[0] == '.')
682                 name++;
683
684         if (!strcmp(name, "_text") ||
685             !strcmp(name, "_etext") ||
686             !strcmp(name, "_sinittext") ||
687             !strncmp("init_module", name, 11) ||
688             !strncmp("cleanup_module", name, 14) ||
689             strstr(name, "_text_start") ||
690             strstr(name, "_text_end"))
691                 return 1;
692
693         for (i = 0; skip_symbols[i]; i++) {
694                 if (!strcmp(skip_symbols[i], name)) {
695                         sym->ignore = true;
696                         break;
697                 }
698         }
699
700         return 0;
701 }
702
703 static void perf_event__process_sample(struct perf_tool *tool,
704                                        const union perf_event *event,
705                                        struct perf_evsel *evsel,
706                                        struct perf_sample *sample,
707                                        struct machine *machine)
708 {
709         struct perf_top *top = container_of(tool, struct perf_top, tool);
710         struct symbol *parent = NULL;
711         u64 ip = event->ip.ip;
712         struct addr_location al;
713         int err;
714
715         if (!machine && perf_guest) {
716                 static struct intlist *seen;
717
718                 if (!seen)
719                         seen = intlist__new();
720
721                 if (!intlist__has_entry(seen, event->ip.pid)) {
722                         pr_err("Can't find guest [%d]'s kernel information\n",
723                                 event->ip.pid);
724                         intlist__add(seen, event->ip.pid);
725                 }
726                 return;
727         }
728
729         if (!machine) {
730                 pr_err("%u unprocessable samples recorded.",
731                        top->session->hists.stats.nr_unprocessable_samples++);
732                 return;
733         }
734
735         if (event->header.misc & PERF_RECORD_MISC_EXACT_IP)
736                 top->exact_samples++;
737
738         if (perf_event__preprocess_sample(event, machine, &al, sample,
739                                           symbol_filter) < 0 ||
740             al.filtered)
741                 return;
742
743         if (!top->kptr_restrict_warned &&
744             symbol_conf.kptr_restrict &&
745             al.cpumode == PERF_RECORD_MISC_KERNEL) {
746                 ui__warning(
747 "Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
748 "Check /proc/sys/kernel/kptr_restrict.\n\n"
749 "Kernel%s samples will not be resolved.\n",
750                           !RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION]) ?
751                           " modules" : "");
752                 if (use_browser <= 0)
753                         sleep(5);
754                 top->kptr_restrict_warned = true;
755         }
756
757         if (al.sym == NULL) {
758                 const char *msg = "Kernel samples will not be resolved.\n";
759                 /*
760                  * As we do lazy loading of symtabs we only will know if the
761                  * specified vmlinux file is invalid when we actually have a
762                  * hit in kernel space and then try to load it. So if we get
763                  * here and there are _no_ symbols in the DSO backing the
764                  * kernel map, bail out.
765                  *
766                  * We may never get here, for instance, if we use -K/
767                  * --hide-kernel-symbols, even if the user specifies an
768                  * invalid --vmlinux ;-)
769                  */
770                 if (!top->kptr_restrict_warned && !top->vmlinux_warned &&
771                     al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
772                     RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
773                         if (symbol_conf.vmlinux_name) {
774                                 ui__warning("The %s file can't be used.\n%s",
775                                             symbol_conf.vmlinux_name, msg);
776                         } else {
777                                 ui__warning("A vmlinux file was not found.\n%s",
778                                             msg);
779                         }
780
781                         if (use_browser <= 0)
782                                 sleep(5);
783                         top->vmlinux_warned = true;
784                 }
785         }
786
787         if (al.sym == NULL || !al.sym->ignore) {
788                 struct hist_entry *he;
789
790                 if ((sort__has_parent || symbol_conf.use_callchain) &&
791                     sample->callchain) {
792                         err = machine__resolve_callchain(machine, evsel,
793                                                          al.thread, sample,
794                                                          &parent);
795
796                         if (err)
797                                 return;
798                 }
799
800                 he = perf_evsel__add_hist_entry(evsel, &al, sample);
801                 if (he == NULL) {
802                         pr_err("Problem incrementing symbol period, skipping event\n");
803                         return;
804                 }
805
806                 if (symbol_conf.use_callchain) {
807                         err = callchain_append(he->callchain, &callchain_cursor,
808                                                sample->period);
809                         if (err)
810                                 return;
811                 }
812
813                 if (top->sort_has_symbols)
814                         perf_top__record_precise_ip(top, he, evsel->idx, ip);
815         }
816
817         return;
818 }
819
820 static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
821 {
822         struct perf_sample sample;
823         struct perf_evsel *evsel;
824         struct perf_session *session = top->session;
825         union perf_event *event;
826         struct machine *machine;
827         u8 origin;
828         int ret;
829
830         while ((event = perf_evlist__mmap_read(top->evlist, idx)) != NULL) {
831                 ret = perf_evlist__parse_sample(top->evlist, event, &sample);
832                 if (ret) {
833                         pr_err("Can't parse sample, err = %d\n", ret);
834                         continue;
835                 }
836
837                 evsel = perf_evlist__id2evsel(session->evlist, sample.id);
838                 assert(evsel != NULL);
839
840                 origin = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
841
842                 if (event->header.type == PERF_RECORD_SAMPLE)
843                         ++top->samples;
844
845                 switch (origin) {
846                 case PERF_RECORD_MISC_USER:
847                         ++top->us_samples;
848                         if (top->hide_user_symbols)
849                                 continue;
850                         machine = perf_session__find_host_machine(session);
851                         break;
852                 case PERF_RECORD_MISC_KERNEL:
853                         ++top->kernel_samples;
854                         if (top->hide_kernel_symbols)
855                                 continue;
856                         machine = perf_session__find_host_machine(session);
857                         break;
858                 case PERF_RECORD_MISC_GUEST_KERNEL:
859                         ++top->guest_kernel_samples;
860                         machine = perf_session__find_machine(session, event->ip.pid);
861                         break;
862                 case PERF_RECORD_MISC_GUEST_USER:
863                         ++top->guest_us_samples;
864                         /*
865                          * TODO: we don't process guest user from host side
866                          * except simple counting.
867                          */
868                         /* Fall thru */
869                 default:
870                         continue;
871                 }
872
873
874                 if (event->header.type == PERF_RECORD_SAMPLE) {
875                         perf_event__process_sample(&top->tool, event, evsel,
876                                                    &sample, machine);
877                 } else if (event->header.type < PERF_RECORD_MAX) {
878                         hists__inc_nr_events(&evsel->hists, event->header.type);
879                         machine__process_event(machine, event);
880                 } else
881                         ++session->hists.stats.nr_unknown_events;
882         }
883 }
884
885 static void perf_top__mmap_read(struct perf_top *top)
886 {
887         int i;
888
889         for (i = 0; i < top->evlist->nr_mmaps; i++)
890                 perf_top__mmap_read_idx(top, i);
891 }
892
893 static void perf_top__start_counters(struct perf_top *top)
894 {
895         struct perf_evsel *counter;
896         struct perf_evlist *evlist = top->evlist;
897
898         if (top->group)
899                 perf_evlist__set_leader(evlist);
900
901         list_for_each_entry(counter, &evlist->entries, node) {
902                 struct perf_event_attr *attr = &counter->attr;
903
904                 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
905
906                 if (top->freq) {
907                         attr->sample_type |= PERF_SAMPLE_PERIOD;
908                         attr->freq        = 1;
909                         attr->sample_freq = top->freq;
910                 }
911
912                 if (evlist->nr_entries > 1) {
913                         attr->sample_type |= PERF_SAMPLE_ID;
914                         attr->read_format |= PERF_FORMAT_ID;
915                 }
916
917                 if (perf_target__has_cpu(&top->target))
918                         attr->sample_type |= PERF_SAMPLE_CPU;
919
920                 if (symbol_conf.use_callchain)
921                         attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
922
923                 attr->mmap = 1;
924                 attr->comm = 1;
925                 attr->inherit = top->inherit;
926 fallback_missing_features:
927                 if (top->exclude_guest_missing)
928                         attr->exclude_guest = attr->exclude_host = 0;
929 retry_sample_id:
930                 attr->sample_id_all = top->sample_id_all_missing ? 0 : 1;
931 try_again:
932                 if (perf_evsel__open(counter, top->evlist->cpus,
933                                      top->evlist->threads) < 0) {
934                         int err = errno;
935
936                         if (err == EPERM || err == EACCES) {
937                                 ui__error_paranoid();
938                                 goto out_err;
939                         } else if (err == EINVAL) {
940                                 if (!top->exclude_guest_missing &&
941                                     (attr->exclude_guest || attr->exclude_host)) {
942                                         pr_debug("Old kernel, cannot exclude "
943                                                  "guest or host samples.\n");
944                                         top->exclude_guest_missing = true;
945                                         goto fallback_missing_features;
946                                 } else if (!top->sample_id_all_missing) {
947                                         /*
948                                          * Old kernel, no attr->sample_id_type_all field
949                                          */
950                                         top->sample_id_all_missing = true;
951                                         goto retry_sample_id;
952                                 }
953                         }
954                         /*
955                          * If it's cycles then fall back to hrtimer
956                          * based cpu-clock-tick sw counter, which
957                          * is always available even if no PMU support:
958                          */
959                         if ((err == ENOENT || err == ENXIO) &&
960                             (attr->type == PERF_TYPE_HARDWARE) &&
961                             (attr->config == PERF_COUNT_HW_CPU_CYCLES)) {
962
963                                 if (verbose)
964                                         ui__warning("Cycles event not supported,\n"
965                                                     "trying to fall back to cpu-clock-ticks\n");
966
967                                 attr->type = PERF_TYPE_SOFTWARE;
968                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
969                                 if (counter->name) {
970                                         free(counter->name);
971                                         counter->name = NULL;
972                                 }
973                                 goto try_again;
974                         }
975
976                         if (err == ENOENT) {
977                                 ui__error("The %s event is not supported.\n",
978                                           perf_evsel__name(counter));
979                                 goto out_err;
980                         } else if (err == EMFILE) {
981                                 ui__error("Too many events are opened.\n"
982                                             "Try again after reducing the number of events\n");
983                                 goto out_err;
984                         } else if ((err == EOPNOTSUPP) && (attr->precise_ip)) {
985                                 ui__error("\'precise\' request may not be supported. "
986                                           "Try removing 'p' modifier\n");
987                                 goto out_err;
988                         }
989
990                         ui__error("The sys_perf_event_open() syscall "
991                                     "returned with %d (%s).  /bin/dmesg "
992                                     "may provide additional information.\n"
993                                     "No CONFIG_PERF_EVENTS=y kernel support "
994                                     "configured?\n", err, strerror(err));
995                         goto out_err;
996                 }
997         }
998
999         if (perf_evlist__mmap(evlist, top->mmap_pages, false) < 0) {
1000                 ui__error("Failed to mmap with %d (%s)\n",
1001                             errno, strerror(errno));
1002                 goto out_err;
1003         }
1004
1005         return;
1006
1007 out_err:
1008         exit_browser(0);
1009         exit(0);
1010 }
1011
1012 static int perf_top__setup_sample_type(struct perf_top *top)
1013 {
1014         if (!top->sort_has_symbols) {
1015                 if (symbol_conf.use_callchain) {
1016                         ui__error("Selected -g but \"sym\" not present in --sort/-s.");
1017                         return -EINVAL;
1018                 }
1019         } else if (!top->dont_use_callchains && callchain_param.mode != CHAIN_NONE) {
1020                 if (callchain_register_param(&callchain_param) < 0) {
1021                         ui__error("Can't register callchain params.\n");
1022                         return -EINVAL;
1023                 }
1024         }
1025
1026         return 0;
1027 }
1028
1029 static int __cmd_top(struct perf_top *top)
1030 {
1031         pthread_t thread;
1032         int ret;
1033         /*
1034          * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1035          * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1036          */
1037         top->session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1038         if (top->session == NULL)
1039                 return -ENOMEM;
1040
1041         ret = perf_top__setup_sample_type(top);
1042         if (ret)
1043                 goto out_delete;
1044
1045         if (perf_target__has_task(&top->target))
1046                 perf_event__synthesize_thread_map(&top->tool, top->evlist->threads,
1047                                                   perf_event__process,
1048                                                   &top->session->host_machine);
1049         else
1050                 perf_event__synthesize_threads(&top->tool, perf_event__process,
1051                                                &top->session->host_machine);
1052         perf_top__start_counters(top);
1053         top->session->evlist = top->evlist;
1054         perf_session__set_id_hdr_size(top->session);
1055
1056         /* Wait for a minimal set of events before starting the snapshot */
1057         poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1058
1059         perf_top__mmap_read(top);
1060
1061         if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
1062                                                             display_thread), top)) {
1063                 ui__error("Could not create display thread.\n");
1064                 exit(-1);
1065         }
1066
1067         if (top->realtime_prio) {
1068                 struct sched_param param;
1069
1070                 param.sched_priority = top->realtime_prio;
1071                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
1072                         ui__error("Could not set realtime priority.\n");
1073                         exit(-1);
1074                 }
1075         }
1076
1077         while (1) {
1078                 u64 hits = top->samples;
1079
1080                 perf_top__mmap_read(top);
1081
1082                 if (hits == top->samples)
1083                         ret = poll(top->evlist->pollfd, top->evlist->nr_fds, 100);
1084         }
1085
1086 out_delete:
1087         perf_session__delete(top->session);
1088         top->session = NULL;
1089
1090         return 0;
1091 }
1092
1093 static int
1094 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1095 {
1096         struct perf_top *top = (struct perf_top *)opt->value;
1097         char *tok, *tok2;
1098         char *endptr;
1099
1100         /*
1101          * --no-call-graph
1102          */
1103         if (unset) {
1104                 top->dont_use_callchains = true;
1105                 return 0;
1106         }
1107
1108         symbol_conf.use_callchain = true;
1109
1110         if (!arg)
1111                 return 0;
1112
1113         tok = strtok((char *)arg, ",");
1114         if (!tok)
1115                 return -1;
1116
1117         /* get the output mode */
1118         if (!strncmp(tok, "graph", strlen(arg)))
1119                 callchain_param.mode = CHAIN_GRAPH_ABS;
1120
1121         else if (!strncmp(tok, "flat", strlen(arg)))
1122                 callchain_param.mode = CHAIN_FLAT;
1123
1124         else if (!strncmp(tok, "fractal", strlen(arg)))
1125                 callchain_param.mode = CHAIN_GRAPH_REL;
1126
1127         else if (!strncmp(tok, "none", strlen(arg))) {
1128                 callchain_param.mode = CHAIN_NONE;
1129                 symbol_conf.use_callchain = false;
1130
1131                 return 0;
1132         } else
1133                 return -1;
1134
1135         /* get the min percentage */
1136         tok = strtok(NULL, ",");
1137         if (!tok)
1138                 goto setup;
1139
1140         callchain_param.min_percent = strtod(tok, &endptr);
1141         if (tok == endptr)
1142                 return -1;
1143
1144         /* get the print limit */
1145         tok2 = strtok(NULL, ",");
1146         if (!tok2)
1147                 goto setup;
1148
1149         if (tok2[0] != 'c') {
1150                 callchain_param.print_limit = strtod(tok2, &endptr);
1151                 tok2 = strtok(NULL, ",");
1152                 if (!tok2)
1153                         goto setup;
1154         }
1155
1156         /* get the call chain order */
1157         if (!strcmp(tok2, "caller"))
1158                 callchain_param.order = ORDER_CALLER;
1159         else if (!strcmp(tok2, "callee"))
1160                 callchain_param.order = ORDER_CALLEE;
1161         else
1162                 return -1;
1163 setup:
1164         if (callchain_register_param(&callchain_param) < 0) {
1165                 fprintf(stderr, "Can't register callchain params\n");
1166                 return -1;
1167         }
1168         return 0;
1169 }
1170
1171 int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
1172 {
1173         struct perf_evsel *pos;
1174         int status;
1175         char errbuf[BUFSIZ];
1176         struct perf_top top = {
1177                 .count_filter        = 5,
1178                 .delay_secs          = 2,
1179                 .freq                = 4000, /* 4 KHz */
1180                 .mmap_pages          = 128,
1181                 .sym_pcnt_filter     = 5,
1182                 .target              = {
1183                         .uses_mmap   = true,
1184                 },
1185         };
1186         char callchain_default_opt[] = "fractal,0.5,callee";
1187         const struct option options[] = {
1188         OPT_CALLBACK('e', "event", &top.evlist, "event",
1189                      "event selector. use 'perf list' to list available events",
1190                      parse_events_option),
1191         OPT_INTEGER('c', "count", &top.default_interval,
1192                     "event period to sample"),
1193         OPT_STRING('p', "pid", &top.target.pid, "pid",
1194                     "profile events on existing process id"),
1195         OPT_STRING('t', "tid", &top.target.tid, "tid",
1196                     "profile events on existing thread id"),
1197         OPT_BOOLEAN('a', "all-cpus", &top.target.system_wide,
1198                             "system-wide collection from all CPUs"),
1199         OPT_STRING('C', "cpu", &top.target.cpu_list, "cpu",
1200                     "list of cpus to monitor"),
1201         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1202                    "file", "vmlinux pathname"),
1203         OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
1204                     "hide kernel symbols"),
1205         OPT_UINTEGER('m', "mmap-pages", &top.mmap_pages, "number of mmap data pages"),
1206         OPT_INTEGER('r', "realtime", &top.realtime_prio,
1207                     "collect data with this RT SCHED_FIFO priority"),
1208         OPT_INTEGER('d', "delay", &top.delay_secs,
1209                     "number of seconds to delay between refreshes"),
1210         OPT_BOOLEAN('D', "dump-symtab", &top.dump_symtab,
1211                             "dump the symbol table used for profiling"),
1212         OPT_INTEGER('f', "count-filter", &top.count_filter,
1213                     "only display functions with more events than this"),
1214         OPT_BOOLEAN('g', "group", &top.group,
1215                             "put the counters into a counter group"),
1216         OPT_BOOLEAN('i', "inherit", &top.inherit,
1217                     "child tasks inherit counters"),
1218         OPT_STRING(0, "sym-annotate", &top.sym_filter, "symbol name",
1219                     "symbol to annotate"),
1220         OPT_BOOLEAN('z', "zero", &top.zero,
1221                     "zero history across updates"),
1222         OPT_INTEGER('F', "freq", &top.freq,
1223                     "profile at this frequency"),
1224         OPT_INTEGER('E', "entries", &top.print_entries,
1225                     "display this many functions"),
1226         OPT_BOOLEAN('U', "hide_user_symbols", &top.hide_user_symbols,
1227                     "hide user symbols"),
1228         OPT_BOOLEAN(0, "tui", &top.use_tui, "Use the TUI interface"),
1229         OPT_BOOLEAN(0, "stdio", &top.use_stdio, "Use the stdio interface"),
1230         OPT_INCR('v', "verbose", &verbose,
1231                     "be more verbose (show counter open errors, etc)"),
1232         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1233                    "sort by key(s): pid, comm, dso, symbol, parent"),
1234         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1235                     "Show a column with the number of samples"),
1236         OPT_CALLBACK_DEFAULT('G', "call-graph", &top, "output_type,min_percent, call_order",
1237                      "Display callchains using output_type (graph, flat, fractal, or none), min percent threshold and callchain order. "
1238                      "Default: fractal,0.5,callee", &parse_callchain_opt,
1239                      callchain_default_opt),
1240         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1241                     "Show a column with the sum of periods"),
1242         OPT_STRING(0, "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1243                    "only consider symbols in these dsos"),
1244         OPT_STRING(0, "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1245                    "only consider symbols in these comms"),
1246         OPT_STRING(0, "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1247                    "only consider these symbols"),
1248         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
1249                     "Interleave source code with assembly code (default)"),
1250         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
1251                     "Display raw encoding of assembly instructions (default)"),
1252         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
1253                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
1254         OPT_STRING('u', "uid", &top.target.uid_str, "user", "user to profile"),
1255         OPT_END()
1256         };
1257         const char * const top_usage[] = {
1258                 "perf top [<options>]",
1259                 NULL
1260         };
1261
1262         top.evlist = perf_evlist__new(NULL, NULL);
1263         if (top.evlist == NULL)
1264                 return -ENOMEM;
1265
1266         symbol_conf.exclude_other = false;
1267
1268         argc = parse_options(argc, argv, options, top_usage, 0);
1269         if (argc)
1270                 usage_with_options(top_usage, options);
1271
1272         if (sort_order == default_sort_order)
1273                 sort_order = "dso,symbol";
1274
1275         setup_sorting(top_usage, options);
1276
1277         if (top.use_stdio)
1278                 use_browser = 0;
1279         else if (top.use_tui)
1280                 use_browser = 1;
1281
1282         setup_browser(false);
1283
1284         status = perf_target__validate(&top.target);
1285         if (status) {
1286                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1287                 ui__warning("%s", errbuf);
1288         }
1289
1290         status = perf_target__parse_uid(&top.target);
1291         if (status) {
1292                 int saved_errno = errno;
1293
1294                 perf_target__strerror(&top.target, status, errbuf, BUFSIZ);
1295                 ui__error("%s", errbuf);
1296
1297                 status = -saved_errno;
1298                 goto out_delete_evlist;
1299         }
1300
1301         if (perf_target__none(&top.target))
1302                 top.target.system_wide = true;
1303
1304         if (perf_evlist__create_maps(top.evlist, &top.target) < 0)
1305                 usage_with_options(top_usage, options);
1306
1307         if (!top.evlist->nr_entries &&
1308             perf_evlist__add_default(top.evlist) < 0) {
1309                 ui__error("Not enough memory for event selector list\n");
1310                 return -ENOMEM;
1311         }
1312
1313         symbol_conf.nr_events = top.evlist->nr_entries;
1314
1315         if (top.delay_secs < 1)
1316                 top.delay_secs = 1;
1317
1318         /*
1319          * User specified count overrides default frequency.
1320          */
1321         if (top.default_interval)
1322                 top.freq = 0;
1323         else if (top.freq) {
1324                 top.default_interval = top.freq;
1325         } else {
1326                 ui__error("frequency and count are zero, aborting\n");
1327                 exit(EXIT_FAILURE);
1328         }
1329
1330         list_for_each_entry(pos, &top.evlist->entries, node) {
1331                 /*
1332                  * Fill in the ones not specifically initialized via -c:
1333                  */
1334                 if (!pos->attr.sample_period)
1335                         pos->attr.sample_period = top.default_interval;
1336         }
1337
1338         top.sym_evsel = perf_evlist__first(top.evlist);
1339
1340         symbol_conf.priv_size = sizeof(struct annotation);
1341
1342         symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1343         if (symbol__init() < 0)
1344                 return -1;
1345
1346         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
1347         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
1348         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
1349
1350         /*
1351          * Avoid annotation data structures overhead when symbols aren't on the
1352          * sort list.
1353          */
1354         top.sort_has_symbols = sort_sym.list.next != NULL;
1355
1356         get_term_dimensions(&top.winsize);
1357         if (top.print_entries == 0) {
1358                 struct sigaction act = {
1359                         .sa_sigaction = perf_top__sig_winch,
1360                         .sa_flags     = SA_SIGINFO,
1361                 };
1362                 perf_top__update_print_entries(&top);
1363                 sigaction(SIGWINCH, &act, NULL);
1364         }
1365
1366         status = __cmd_top(&top);
1367
1368 out_delete_evlist:
1369         perf_evlist__delete(top.evlist);
1370
1371         return status;
1372 }