4 * Builtin top command: Display a continuously updated profile of
5 * any workload, CPU or specific PID.
7 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
9 * Improvements and fixes by:
11 * Arjan van de Ven <arjan@linux.intel.com>
12 * Yanmin Zhang <yanmin.zhang@intel.com>
13 * Wu Fengguang <fengguang.wu@intel.com>
14 * Mike Galbraith <efault@gmx.de>
15 * Paul Mackerras <paulus@samba.org>
17 * Released under the GPL v2. (and only v2, not any later version)
23 #include "util/color.h"
24 #include "util/evsel.h"
25 #include "util/session.h"
26 #include "util/symbol.h"
27 #include "util/thread.h"
28 #include "util/util.h"
29 #include <linux/rbtree.h>
30 #include "util/parse-options.h"
31 #include "util/parse-events.h"
32 #include "util/cpumap.h"
33 #include "util/xyarray.h"
35 #include "util/debug.h"
49 #include <sys/syscall.h>
50 #include <sys/ioctl.h>
52 #include <sys/prctl.h>
57 #include <linux/unistd.h>
58 #include <linux/types.h>
60 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
62 static bool system_wide = false;
64 static int default_interval = 0;
66 static int count_filter = 5;
67 static int print_entries;
69 static int target_pid = -1;
70 static int target_tid = -1;
71 static struct thread_map *threads;
72 static bool inherit = false;
73 static struct cpu_map *cpus;
74 static int realtime_prio = 0;
75 static bool group = false;
76 static unsigned int page_size;
77 static unsigned int mmap_pages = 16;
78 static int freq = 1000; /* 1 KHz */
80 static int delay_secs = 2;
81 static bool zero = false;
82 static bool dump_symtab = false;
84 static bool hide_kernel_symbols = false;
85 static bool hide_user_symbols = false;
86 static struct winsize winsize;
94 unsigned long count[MAX_COUNTERS];
96 struct source_line *next;
99 static const char *sym_filter = NULL;
100 struct sym_entry *sym_filter_entry = NULL;
101 struct sym_entry *sym_filter_entry_sched = NULL;
102 static int sym_pcnt_filter = 5;
103 static int sym_counter = 0;
104 static struct perf_evsel *sym_evsel = NULL;
105 static int display_weighted = -1;
106 static const char *cpu_list;
112 struct sym_entry_source {
113 struct source_line *source;
114 struct source_line *lines;
115 struct source_line **lines_tail;
116 pthread_mutex_t lock;
120 struct rb_node rb_node;
121 struct list_head node;
122 unsigned long snap_count;
128 struct sym_entry_source *src;
129 unsigned long count[0];
136 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
138 return ((void *)self) + symbol_conf.priv_size;
141 void get_term_dimensions(struct winsize *ws)
143 char *s = getenv("LINES");
146 ws->ws_row = atoi(s);
147 s = getenv("COLUMNS");
149 ws->ws_col = atoi(s);
150 if (ws->ws_row && ws->ws_col)
155 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
156 ws->ws_row && ws->ws_col)
163 static void update_print_entries(struct winsize *ws)
165 print_entries = ws->ws_row;
167 if (print_entries > 9)
171 static void sig_winch_handler(int sig __used)
173 get_term_dimensions(&winsize);
174 update_print_entries(&winsize);
177 static int parse_source(struct sym_entry *syme)
180 struct sym_entry_source *source;
183 char command[PATH_MAX*2];
190 sym = sym_entry__symbol(syme);
194 * We can't annotate with just /proc/kallsyms
196 if (map->dso->origin == DSO__ORIG_KERNEL)
199 if (syme->src == NULL) {
200 syme->src = zalloc(sizeof(*source));
201 if (syme->src == NULL)
203 pthread_mutex_init(&syme->src->lock, NULL);
209 pthread_mutex_lock(&source->lock);
212 path = map->dso->long_name;
214 len = sym->end - sym->start;
217 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
218 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
219 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
221 file = popen(command, "r");
225 pthread_mutex_lock(&source->lock);
226 source->lines_tail = &source->lines;
227 while (!feof(file)) {
228 struct source_line *src;
232 src = malloc(sizeof(struct source_line));
234 memset(src, 0, sizeof(struct source_line));
236 if (getline(&src->line, &dummy, file) < 0)
241 c = strchr(src->line, '\n');
246 *source->lines_tail = src;
247 source->lines_tail = &src->next;
249 src->eip = strtoull(src->line, &sep, 16);
251 src->eip = map__objdump_2ip(map, src->eip);
252 else /* this line has no ip info (e.g. source line) */
257 sym_filter_entry = syme;
258 pthread_mutex_unlock(&source->lock);
262 static void __zero_source_counters(struct sym_entry *syme)
265 struct source_line *line;
267 line = syme->src->lines;
269 for (i = 0; i < nr_counters; i++)
275 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
277 struct source_line *line;
279 if (syme != sym_filter_entry)
282 if (pthread_mutex_trylock(&syme->src->lock))
285 if (syme->src == NULL || syme->src->source == NULL)
288 for (line = syme->src->lines; line; line = line->next) {
289 /* skip lines without IP info */
292 if (line->eip == ip) {
293 line->count[counter]++;
300 pthread_mutex_unlock(&syme->src->lock);
303 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
305 static void lookup_sym_source(struct sym_entry *syme)
307 struct symbol *symbol = sym_entry__symbol(syme);
308 struct source_line *line;
309 char pattern[PATTERN_LEN + 1];
311 sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
312 map__rip_2objdump(syme->map, symbol->start));
314 pthread_mutex_lock(&syme->src->lock);
315 for (line = syme->src->lines; line; line = line->next) {
316 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
317 syme->src->source = line;
321 pthread_mutex_unlock(&syme->src->lock);
324 static void show_lines(struct source_line *queue, int count, int total)
327 struct source_line *line;
330 for (i = 0; i < count; i++) {
331 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
333 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
338 #define TRACE_COUNT 3
340 static void show_details(struct sym_entry *syme)
342 struct symbol *symbol;
343 struct source_line *line;
344 struct source_line *line_queue = NULL;
346 int line_queue_count = 0, total = 0, more = 0;
351 if (!syme->src->source)
352 lookup_sym_source(syme);
354 if (!syme->src->source)
357 symbol = sym_entry__symbol(syme);
358 printf("Showing %s for %s\n", event_name(sym_evsel), symbol->name);
359 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
361 pthread_mutex_lock(&syme->src->lock);
362 line = syme->src->source;
364 total += line->count[sym_counter];
368 line = syme->src->source;
372 if (!line_queue_count)
376 if (line->count[sym_counter])
377 pcnt = 100.0 * line->count[sym_counter] / (float)total;
378 if (pcnt >= (float)sym_pcnt_filter) {
379 if (displayed <= print_entries)
380 show_lines(line_queue, line_queue_count, total);
382 displayed += line_queue_count;
383 line_queue_count = 0;
385 } else if (line_queue_count > TRACE_COUNT) {
386 line_queue = line_queue->next;
390 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
393 pthread_mutex_unlock(&syme->src->lock);
395 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
399 * Symbols will be added here in event__process_sample and will get out
402 static LIST_HEAD(active_symbols);
403 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
406 * Ordering weight: count-1 * count-2 * ... / count-n
408 static double sym_weight(const struct sym_entry *sym)
410 double weight = sym->snap_count;
413 if (!display_weighted)
416 for (counter = 1; counter < nr_counters-1; counter++)
417 weight *= sym->count[counter];
419 weight /= (sym->count[counter] + 1);
425 static long kernel_samples, us_samples;
426 static long exact_samples;
427 static long guest_us_samples, guest_kernel_samples;
428 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
430 static void __list_insert_active_sym(struct sym_entry *syme)
432 list_add(&syme->node, &active_symbols);
435 static void list_remove_active_sym(struct sym_entry *syme)
437 pthread_mutex_lock(&active_symbols_lock);
438 list_del_init(&syme->node);
439 pthread_mutex_unlock(&active_symbols_lock);
442 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
444 struct rb_node **p = &tree->rb_node;
445 struct rb_node *parent = NULL;
446 struct sym_entry *iter;
450 iter = rb_entry(parent, struct sym_entry, rb_node);
452 if (se->weight > iter->weight)
458 rb_link_node(&se->rb_node, parent, p);
459 rb_insert_color(&se->rb_node, tree);
462 static void print_sym_table(void)
465 struct perf_evsel *counter;
466 int snap = !display_weighted ? sym_counter : 0;
467 float samples_per_sec = samples/delay_secs;
468 float ksamples_per_sec = kernel_samples/delay_secs;
469 float us_samples_per_sec = (us_samples)/delay_secs;
470 float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs;
471 float guest_us_samples_per_sec = (guest_us_samples)/delay_secs;
472 float esamples_percent = (100.0*exact_samples)/samples;
473 float sum_ksamples = 0.0;
474 struct sym_entry *syme, *n;
475 struct rb_root tmp = RB_ROOT;
477 int sym_width = 0, dso_width = 0, dso_short_width = 0;
478 const int win_width = winsize.ws_col - 1;
480 samples = us_samples = kernel_samples = exact_samples = 0;
481 guest_kernel_samples = guest_us_samples = 0;
483 /* Sort the active symbols */
484 pthread_mutex_lock(&active_symbols_lock);
485 syme = list_entry(active_symbols.next, struct sym_entry, node);
486 pthread_mutex_unlock(&active_symbols_lock);
488 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
489 syme->snap_count = syme->count[snap];
490 if (syme->snap_count != 0) {
492 if ((hide_user_symbols &&
493 syme->origin == PERF_RECORD_MISC_USER) ||
494 (hide_kernel_symbols &&
495 syme->origin == PERF_RECORD_MISC_KERNEL)) {
496 list_remove_active_sym(syme);
499 syme->weight = sym_weight(syme);
500 rb_insert_active_sym(&tmp, syme);
501 sum_ksamples += syme->snap_count;
503 for (j = 0; j < nr_counters; j++)
504 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
506 list_remove_active_sym(syme);
511 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
513 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%"
516 100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) /
520 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%"
521 " guest kernel:%4.1f%% guest us:%4.1f%%"
524 100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) /
526 100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) /
528 100.0 - (100.0 * ((samples_per_sec -
529 guest_kernel_samples_per_sec) /
531 100.0 - (100.0 * ((samples_per_sec -
532 guest_us_samples_per_sec) /
537 if (nr_counters == 1 || !display_weighted) {
538 struct perf_evsel *first;
539 first = list_entry(evsel_list.next, struct perf_evsel, node);
540 printf("%Ld", first->attr.sample_period);
547 if (!display_weighted)
548 printf("%s", event_name(sym_evsel));
549 else list_for_each_entry(counter, &evsel_list, node) {
553 printf("%s", event_name(counter));
558 if (target_pid != -1)
559 printf(" (target_pid: %d", target_pid);
560 else if (target_tid != -1)
561 printf(" (target_tid: %d", target_tid);
566 printf(", CPU%s: %s)\n", cpus->nr > 1 ? "s" : "", cpu_list);
568 if (target_tid != -1)
571 printf(", %d CPU%s)\n", cpus->nr, cpus->nr > 1 ? "s" : "");
574 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
576 if (sym_filter_entry) {
577 show_details(sym_filter_entry);
582 * Find the longest symbol name that will be displayed
584 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
585 syme = rb_entry(nd, struct sym_entry, rb_node);
586 if (++printed > print_entries ||
587 (int)syme->snap_count < count_filter)
590 if (syme->map->dso->long_name_len > dso_width)
591 dso_width = syme->map->dso->long_name_len;
593 if (syme->map->dso->short_name_len > dso_short_width)
594 dso_short_width = syme->map->dso->short_name_len;
596 if (syme->name_len > sym_width)
597 sym_width = syme->name_len;
602 if (sym_width + dso_width > winsize.ws_col - 29) {
603 dso_width = dso_short_width;
604 if (sym_width + dso_width > winsize.ws_col - 29)
605 sym_width = winsize.ws_col - dso_width - 29;
608 if (nr_counters == 1)
609 printf(" samples pcnt");
611 printf(" weight samples pcnt");
615 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
616 printf(" %s _______ _____",
617 nr_counters == 1 ? " " : "______");
619 printf(" ________________");
620 printf(" %-*.*s", sym_width, sym_width, graph_line);
621 printf(" %-*.*s", dso_width, dso_width, graph_line);
624 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
628 syme = rb_entry(nd, struct sym_entry, rb_node);
629 sym = sym_entry__symbol(syme);
630 if (++printed > print_entries || (int)syme->snap_count < count_filter)
633 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
636 if (nr_counters == 1 || !display_weighted)
637 printf("%20.2f ", syme->weight);
639 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
641 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
643 printf(" %016llx", sym->start);
644 printf(" %-*.*s", sym_width, sym_width, sym->name);
645 printf(" %-*.*s\n", dso_width, dso_width,
646 dso_width >= syme->map->dso->long_name_len ?
647 syme->map->dso->long_name :
648 syme->map->dso->short_name);
652 static void prompt_integer(int *target, const char *msg)
654 char *buf = malloc(0), *p;
658 fprintf(stdout, "\n%s: ", msg);
659 if (getline(&buf, &dummy, stdin) < 0)
662 p = strchr(buf, '\n');
672 tmp = strtoul(buf, NULL, 10);
678 static void prompt_percent(int *target, const char *msg)
682 prompt_integer(&tmp, msg);
683 if (tmp >= 0 && tmp <= 100)
687 static void prompt_symbol(struct sym_entry **target, const char *msg)
689 char *buf = malloc(0), *p;
690 struct sym_entry *syme = *target, *n, *found = NULL;
693 /* zero counters of active symbol */
695 pthread_mutex_lock(&syme->src->lock);
696 __zero_source_counters(syme);
698 pthread_mutex_unlock(&syme->src->lock);
701 fprintf(stdout, "\n%s: ", msg);
702 if (getline(&buf, &dummy, stdin) < 0)
705 p = strchr(buf, '\n');
709 pthread_mutex_lock(&active_symbols_lock);
710 syme = list_entry(active_symbols.next, struct sym_entry, node);
711 pthread_mutex_unlock(&active_symbols_lock);
713 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
714 struct symbol *sym = sym_entry__symbol(syme);
716 if (!strcmp(buf, sym->name)) {
723 fprintf(stderr, "Sorry, %s is not active.\n", buf);
733 static void print_mapped_keys(void)
737 if (sym_filter_entry) {
738 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
742 fprintf(stdout, "\nMapped keys:\n");
743 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
744 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
747 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel));
749 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
751 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
752 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
753 fprintf(stdout, "\t[S] stop annotation.\n");
756 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
759 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
760 hide_kernel_symbols ? "yes" : "no");
762 "\t[U] hide user symbols. \t(%s)\n",
763 hide_user_symbols ? "yes" : "no");
764 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
765 fprintf(stdout, "\t[qQ] quit.\n");
768 static int key_mapped(int c)
785 return nr_counters > 1 ? 1 : 0;
793 static void handle_keypress(struct perf_session *session, int c)
795 if (!key_mapped(c)) {
796 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
797 struct termios tc, save;
800 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
805 tc.c_lflag &= ~(ICANON | ECHO);
808 tcsetattr(0, TCSANOW, &tc);
810 poll(&stdin_poll, 1, -1);
813 tcsetattr(0, TCSAFLUSH, &save);
820 prompt_integer(&delay_secs, "Enter display delay");
825 prompt_integer(&print_entries, "Enter display entries (lines)");
826 if (print_entries == 0) {
827 sig_winch_handler(SIGWINCH);
828 signal(SIGWINCH, sig_winch_handler);
830 signal(SIGWINCH, SIG_DFL);
833 if (nr_counters > 1) {
834 fprintf(stderr, "\nAvailable events:");
836 list_for_each_entry(sym_evsel, &evsel_list, node)
837 fprintf(stderr, "\n\t%d %s", sym_evsel->idx, event_name(sym_evsel));
839 prompt_integer(&sym_counter, "Enter details event counter");
841 if (sym_counter >= nr_counters) {
842 sym_evsel = list_entry(evsel_list.next, struct perf_evsel, node);
844 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(sym_evsel));
848 list_for_each_entry(sym_evsel, &evsel_list, node)
849 if (sym_evsel->idx == sym_counter)
851 } else sym_counter = 0;
854 prompt_integer(&count_filter, "Enter display event count filter");
857 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
860 hide_kernel_symbols = !hide_kernel_symbols;
864 printf("exiting.\n");
866 perf_session__fprintf_dsos(session, stderr);
869 prompt_symbol(&sym_filter_entry, "Enter details symbol");
872 if (!sym_filter_entry)
875 struct sym_entry *syme = sym_filter_entry;
877 pthread_mutex_lock(&syme->src->lock);
878 sym_filter_entry = NULL;
879 __zero_source_counters(syme);
880 pthread_mutex_unlock(&syme->src->lock);
884 hide_user_symbols = !hide_user_symbols;
887 display_weighted = ~display_weighted;
897 static void *display_thread(void *arg __used)
899 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
900 struct termios tc, save;
902 struct perf_session *session = (struct perf_session *) arg;
906 tc.c_lflag &= ~(ICANON | ECHO);
911 delay_msecs = delay_secs * 1000;
912 tcsetattr(0, TCSANOW, &tc);
918 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
921 tcsetattr(0, TCSAFLUSH, &save);
923 handle_keypress(session, c);
929 /* Tag samples to be skipped. */
930 static const char *skip_symbols[] = {
936 "mwait_idle_with_hints",
938 "ppc64_runlatch_off",
939 "pseries_dedicated_idle_sleep",
943 static int symbol_filter(struct map *map, struct symbol *sym)
945 struct sym_entry *syme;
946 const char *name = sym->name;
950 * ppc64 uses function descriptors and appends a '.' to the
951 * start of every instruction address. Remove it.
956 if (!strcmp(name, "_text") ||
957 !strcmp(name, "_etext") ||
958 !strcmp(name, "_sinittext") ||
959 !strncmp("init_module", name, 11) ||
960 !strncmp("cleanup_module", name, 14) ||
961 strstr(name, "_text_start") ||
962 strstr(name, "_text_end"))
965 syme = symbol__priv(sym);
969 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
970 /* schedule initial sym_filter_entry setup */
971 sym_filter_entry_sched = syme;
975 for (i = 0; skip_symbols[i]; i++) {
976 if (!strcmp(skip_symbols[i], name)) {
983 syme->name_len = strlen(sym->name);
988 static void event__process_sample(const event_t *self,
989 struct sample_data *sample,
990 struct perf_session *session,
991 struct perf_evsel *evsel)
993 u64 ip = self->ip.ip;
994 struct sym_entry *syme;
995 struct addr_location al;
996 struct machine *machine;
997 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1002 case PERF_RECORD_MISC_USER:
1004 if (hide_user_symbols)
1006 machine = perf_session__find_host_machine(session);
1008 case PERF_RECORD_MISC_KERNEL:
1010 if (hide_kernel_symbols)
1012 machine = perf_session__find_host_machine(session);
1014 case PERF_RECORD_MISC_GUEST_KERNEL:
1015 ++guest_kernel_samples;
1016 machine = perf_session__find_machine(session, self->ip.pid);
1018 case PERF_RECORD_MISC_GUEST_USER:
1021 * TODO: we don't process guest user from host side
1022 * except simple counting.
1029 if (!machine && perf_guest) {
1030 pr_err("Can't find guest [%d]'s kernel information\n",
1035 if (self->header.misc & PERF_RECORD_MISC_EXACT_IP)
1038 if (event__preprocess_sample(self, session, &al, sample,
1039 symbol_filter) < 0 ||
1043 if (al.sym == NULL) {
1045 * As we do lazy loading of symtabs we only will know if the
1046 * specified vmlinux file is invalid when we actually have a
1047 * hit in kernel space and then try to load it. So if we get
1048 * here and there are _no_ symbols in the DSO backing the
1049 * kernel map, bail out.
1051 * We may never get here, for instance, if we use -K/
1052 * --hide-kernel-symbols, even if the user specifies an
1053 * invalid --vmlinux ;-)
1055 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
1056 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
1057 pr_err("The %s file can't be used\n",
1058 symbol_conf.vmlinux_name);
1065 /* let's see, whether we need to install initial sym_filter_entry */
1066 if (sym_filter_entry_sched) {
1067 sym_filter_entry = sym_filter_entry_sched;
1068 sym_filter_entry_sched = NULL;
1069 if (parse_source(sym_filter_entry) < 0) {
1070 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1072 pr_err("Can't annotate %s", sym->name);
1073 if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1074 pr_err(": No vmlinux file was found in the path:\n");
1075 machine__fprintf_vmlinux_path(machine, stderr);
1082 syme = symbol__priv(al.sym);
1084 syme->count[evsel->idx]++;
1085 syme->origin = origin;
1086 record_precise_ip(syme, evsel->idx, ip);
1087 pthread_mutex_lock(&active_symbols_lock);
1088 if (list_empty(&syme->node) || !syme->node.next)
1089 __list_insert_active_sym(syme);
1090 pthread_mutex_unlock(&active_symbols_lock);
1100 static int perf_evsel__alloc_mmap_per_thread(struct perf_evsel *evsel,
1101 int ncpus, int nthreads)
1103 evsel->priv = xyarray__new(ncpus, nthreads, sizeof(struct mmap_data));
1104 return evsel->priv != NULL ? 0 : -ENOMEM;
1107 static void perf_evsel__free_mmap(struct perf_evsel *evsel)
1109 xyarray__delete(evsel->priv);
1113 static unsigned int mmap_read_head(struct mmap_data *md)
1115 struct perf_event_mmap_page *pc = md->base;
1118 head = pc->data_head;
1124 static void perf_session__mmap_read_counter(struct perf_session *self,
1125 struct perf_evsel *evsel,
1126 int cpu, int thread_idx)
1128 struct xyarray *mmap_array = evsel->priv;
1129 struct mmap_data *md = xyarray__entry(mmap_array, cpu, thread_idx);
1130 unsigned int head = mmap_read_head(md);
1131 unsigned int old = md->prev;
1132 unsigned char *data = md->base + page_size;
1133 struct sample_data sample;
1137 * If we're further behind than half the buffer, there's a chance
1138 * the writer will bite our tail and mess up the samples under us.
1140 * If we somehow ended up ahead of the head, we got messed up.
1142 * In either case, truncate and restart at head.
1145 if (diff > md->mask / 2 || diff < 0) {
1146 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1149 * head points to a known good entry, start there.
1154 for (; old != head;) {
1155 event_t *event = (event_t *)&data[old & md->mask];
1159 size_t size = event->header.size;
1162 * Event straddles the mmap boundary -- header should always
1163 * be inside due to u64 alignment of output.
1165 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1166 unsigned int offset = old;
1167 unsigned int len = min(sizeof(*event), size), cpy;
1168 void *dst = &event_copy;
1171 cpy = min(md->mask + 1 - (offset & md->mask), len);
1172 memcpy(dst, &data[offset & md->mask], cpy);
1178 event = &event_copy;
1181 event__parse_sample(event, self, &sample);
1182 if (event->header.type == PERF_RECORD_SAMPLE)
1183 event__process_sample(event, &sample, self, evsel);
1185 event__process(event, &sample, self);
1192 static struct pollfd *event_array;
1194 static void perf_session__mmap_read(struct perf_session *self)
1196 struct perf_evsel *counter;
1197 int i, thread_index;
1199 for (i = 0; i < cpus->nr; i++) {
1200 list_for_each_entry(counter, &evsel_list, node) {
1201 for (thread_index = 0;
1202 thread_index < threads->nr;
1204 perf_session__mmap_read_counter(self,
1205 counter, i, thread_index);
1214 static void start_counter(int i, struct perf_evsel *evsel)
1216 struct xyarray *mmap_array = evsel->priv;
1217 struct mmap_data *mm;
1218 struct perf_event_attr *attr;
1222 if (target_tid == -1)
1225 attr = &evsel->attr;
1227 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1230 attr->sample_type |= PERF_SAMPLE_PERIOD;
1232 attr->sample_freq = freq;
1235 attr->inherit = (cpu < 0) && inherit;
1238 for (thread_index = 0; thread_index < threads->nr; thread_index++) {
1240 FD(evsel, i, thread_index) = sys_perf_event_open(attr,
1241 threads->map[thread_index], cpu, group_fd, 0);
1243 if (FD(evsel, i, thread_index) < 0) {
1246 if (err == EPERM || err == EACCES)
1247 die("Permission error - are you root?\n"
1248 "\t Consider tweaking"
1249 " /proc/sys/kernel/perf_event_paranoid.\n");
1251 * If it's cycles then fall back to hrtimer
1252 * based cpu-clock-tick sw counter, which
1253 * is always available even if no PMU support:
1255 if (attr->type == PERF_TYPE_HARDWARE
1256 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1259 warning(" ... trying to fall back to cpu-clock-ticks\n");
1261 attr->type = PERF_TYPE_SOFTWARE;
1262 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1266 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
1267 FD(evsel, i, thread_index), strerror(err));
1268 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1271 assert(FD(evsel, i, thread_index) >= 0);
1272 fcntl(FD(evsel, i, thread_index), F_SETFL, O_NONBLOCK);
1275 * First counter acts as the group leader:
1277 if (group && group_fd == -1)
1278 group_fd = FD(evsel, i, thread_index);
1280 event_array[nr_poll].fd = FD(evsel, i, thread_index);
1281 event_array[nr_poll].events = POLLIN;
1284 mm = xyarray__entry(mmap_array, i, thread_index);
1286 mm->mask = mmap_pages*page_size - 1;
1287 mm->base = mmap(NULL, (mmap_pages+1)*page_size,
1288 PROT_READ, MAP_SHARED, FD(evsel, i, thread_index), 0);
1289 if (mm->base == MAP_FAILED)
1290 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1294 static int __cmd_top(void)
1297 struct perf_evsel *counter;
1300 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1301 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1303 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false, NULL);
1304 if (session == NULL)
1307 if (target_tid != -1)
1308 event__synthesize_thread(target_tid, event__process, session);
1310 event__synthesize_threads(event__process, session);
1312 for (i = 0; i < cpus->nr; i++) {
1314 list_for_each_entry(counter, &evsel_list, node)
1315 start_counter(i, counter);
1318 /* Wait for a minimal set of events before starting the snapshot */
1319 poll(&event_array[0], nr_poll, 100);
1321 perf_session__mmap_read(session);
1323 if (pthread_create(&thread, NULL, display_thread, session)) {
1324 printf("Could not create display thread.\n");
1328 if (realtime_prio) {
1329 struct sched_param param;
1331 param.sched_priority = realtime_prio;
1332 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1333 printf("Could not set realtime priority.\n");
1341 perf_session__mmap_read(session);
1343 if (hits == samples)
1344 ret = poll(event_array, nr_poll, 100);
1350 static const char * const top_usage[] = {
1351 "perf top [<options>]",
1355 static const struct option options[] = {
1356 OPT_CALLBACK('e', "event", NULL, "event",
1357 "event selector. use 'perf list' to list available events",
1359 OPT_INTEGER('c', "count", &default_interval,
1360 "event period to sample"),
1361 OPT_INTEGER('p', "pid", &target_pid,
1362 "profile events on existing process id"),
1363 OPT_INTEGER('t', "tid", &target_tid,
1364 "profile events on existing thread id"),
1365 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1366 "system-wide collection from all CPUs"),
1367 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1368 "list of cpus to monitor"),
1369 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1370 "file", "vmlinux pathname"),
1371 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1372 "hide kernel symbols"),
1373 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1374 OPT_INTEGER('r', "realtime", &realtime_prio,
1375 "collect data with this RT SCHED_FIFO priority"),
1376 OPT_INTEGER('d', "delay", &delay_secs,
1377 "number of seconds to delay between refreshes"),
1378 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1379 "dump the symbol table used for profiling"),
1380 OPT_INTEGER('f', "count-filter", &count_filter,
1381 "only display functions with more events than this"),
1382 OPT_BOOLEAN('g', "group", &group,
1383 "put the counters into a counter group"),
1384 OPT_BOOLEAN('i', "inherit", &inherit,
1385 "child tasks inherit counters"),
1386 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1387 "symbol to annotate"),
1388 OPT_BOOLEAN('z', "zero", &zero,
1389 "zero history across updates"),
1390 OPT_INTEGER('F', "freq", &freq,
1391 "profile at this frequency"),
1392 OPT_INTEGER('E', "entries", &print_entries,
1393 "display this many functions"),
1394 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1395 "hide user symbols"),
1396 OPT_INCR('v', "verbose", &verbose,
1397 "be more verbose (show counter open errors, etc)"),
1401 int cmd_top(int argc, const char **argv, const char *prefix __used)
1403 struct perf_evsel *pos;
1404 int status = -ENOMEM;
1406 page_size = sysconf(_SC_PAGE_SIZE);
1408 argc = parse_options(argc, argv, options, top_usage, 0);
1410 usage_with_options(top_usage, options);
1412 if (target_pid != -1)
1413 target_tid = target_pid;
1415 threads = thread_map__new(target_pid, target_tid);
1416 if (threads == NULL) {
1417 pr_err("Problems finding threads of monitor\n");
1418 usage_with_options(top_usage, options);
1421 event_array = malloc((sizeof(struct pollfd) *
1422 MAX_NR_CPUS * MAX_COUNTERS * threads->nr));
1426 /* CPU and PID are mutually exclusive */
1427 if (target_tid > 0 && cpu_list) {
1428 printf("WARNING: PID switch overriding CPU\n");
1433 if (!nr_counters && perf_evsel_list__create_default() < 0) {
1434 pr_err("Not enough memory for event selector list\n");
1442 * User specified count overrides default frequency.
1444 if (default_interval)
1447 default_interval = freq;
1449 fprintf(stderr, "frequency and count are zero, aborting\n");
1453 if (target_tid != -1)
1454 cpus = cpu_map__dummy_new();
1456 cpus = cpu_map__new(cpu_list);
1459 usage_with_options(top_usage, options);
1461 list_for_each_entry(pos, &evsel_list, node) {
1462 if (perf_evsel__alloc_mmap_per_thread(pos, cpus->nr, threads->nr) < 0 ||
1463 perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
1466 * Fill in the ones not specifically initialized via -c:
1468 if (pos->attr.sample_period)
1471 pos->attr.sample_period = default_interval;
1474 symbol_conf.priv_size = (sizeof(struct sym_entry) +
1475 (nr_counters + 1) * sizeof(unsigned long));
1477 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1478 if (symbol__init() < 0)
1481 get_term_dimensions(&winsize);
1482 if (print_entries == 0) {
1483 update_print_entries(&winsize);
1484 signal(SIGWINCH, sig_winch_handler);
1487 status = __cmd_top();
1489 list_for_each_entry(pos, &evsel_list, node)
1490 perf_evsel__free_mmap(pos);