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/session.h"
25 #include "util/symbol.h"
26 #include "util/thread.h"
27 #include "util/util.h"
28 #include <linux/rbtree.h>
29 #include "util/parse-options.h"
30 #include "util/parse-events.h"
31 #include "util/cpumap.h"
33 #include "util/debug.h"
47 #include <sys/syscall.h>
48 #include <sys/ioctl.h>
50 #include <sys/prctl.h>
55 #include <linux/unistd.h>
56 #include <linux/types.h>
58 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
60 static int system_wide = 0;
62 static int default_interval = 0;
64 static int count_filter = 5;
65 static int print_entries;
67 static int target_pid = -1;
68 static int inherit = 0;
69 static int profile_cpu = -1;
70 static int nr_cpus = 0;
71 static unsigned int realtime_prio = 0;
73 static unsigned int page_size;
74 static unsigned int mmap_pages = 16;
75 static int freq = 1000; /* 1 KHz */
77 static int delay_secs = 2;
79 static int dump_symtab = 0;
81 static bool hide_kernel_symbols = false;
82 static bool hide_user_symbols = false;
83 static struct winsize winsize;
91 unsigned long count[MAX_COUNTERS];
93 struct source_line *next;
96 static char *sym_filter = NULL;
97 struct sym_entry *sym_filter_entry = NULL;
98 struct sym_entry *sym_filter_entry_sched = NULL;
99 static int sym_pcnt_filter = 5;
100 static int sym_counter = 0;
101 static int display_weighted = -1;
107 struct sym_entry_source {
108 struct source_line *source;
109 struct source_line *lines;
110 struct source_line **lines_tail;
111 pthread_mutex_t lock;
115 struct rb_node rb_node;
116 struct list_head node;
117 unsigned long snap_count;
123 struct sym_entry_source *src;
124 unsigned long count[0];
131 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
133 return ((void *)self) + symbol_conf.priv_size;
136 void get_term_dimensions(struct winsize *ws)
138 char *s = getenv("LINES");
141 ws->ws_row = atoi(s);
142 s = getenv("COLUMNS");
144 ws->ws_col = atoi(s);
145 if (ws->ws_row && ws->ws_col)
150 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
151 ws->ws_row && ws->ws_col)
158 static void update_print_entries(struct winsize *ws)
160 print_entries = ws->ws_row;
162 if (print_entries > 9)
166 static void sig_winch_handler(int sig __used)
168 get_term_dimensions(&winsize);
169 update_print_entries(&winsize);
172 static void parse_source(struct sym_entry *syme)
175 struct sym_entry_source *source;
178 char command[PATH_MAX*2];
185 if (syme->src == NULL) {
186 syme->src = zalloc(sizeof(*source));
187 if (syme->src == NULL)
189 pthread_mutex_init(&syme->src->lock, NULL);
195 pthread_mutex_lock(&source->lock);
199 sym = sym_entry__symbol(syme);
201 path = map->dso->long_name;
203 len = sym->end - sym->start;
206 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
207 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
208 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
210 file = popen(command, "r");
214 pthread_mutex_lock(&source->lock);
215 source->lines_tail = &source->lines;
216 while (!feof(file)) {
217 struct source_line *src;
221 src = malloc(sizeof(struct source_line));
223 memset(src, 0, sizeof(struct source_line));
225 if (getline(&src->line, &dummy, file) < 0)
230 c = strchr(src->line, '\n');
235 *source->lines_tail = src;
236 source->lines_tail = &src->next;
238 src->eip = strtoull(src->line, &sep, 16);
240 src->eip = map__objdump_2ip(map, src->eip);
241 else /* this line has no ip info (e.g. source line) */
246 sym_filter_entry = syme;
247 pthread_mutex_unlock(&source->lock);
250 static void __zero_source_counters(struct sym_entry *syme)
253 struct source_line *line;
255 line = syme->src->lines;
257 for (i = 0; i < nr_counters; i++)
263 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
265 struct source_line *line;
267 if (syme != sym_filter_entry)
270 if (pthread_mutex_trylock(&syme->src->lock))
273 if (syme->src == NULL || syme->src->source == NULL)
276 for (line = syme->src->lines; line; line = line->next) {
277 /* skip lines without IP info */
280 if (line->eip == ip) {
281 line->count[counter]++;
288 pthread_mutex_unlock(&syme->src->lock);
291 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
293 static void lookup_sym_source(struct sym_entry *syme)
295 struct symbol *symbol = sym_entry__symbol(syme);
296 struct source_line *line;
297 char pattern[PATTERN_LEN + 1];
299 sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
300 map__rip_2objdump(syme->map, symbol->start));
302 pthread_mutex_lock(&syme->src->lock);
303 for (line = syme->src->lines; line; line = line->next) {
304 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
305 syme->src->source = line;
309 pthread_mutex_unlock(&syme->src->lock);
312 static void show_lines(struct source_line *queue, int count, int total)
315 struct source_line *line;
318 for (i = 0; i < count; i++) {
319 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
321 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
326 #define TRACE_COUNT 3
328 static void show_details(struct sym_entry *syme)
330 struct symbol *symbol;
331 struct source_line *line;
332 struct source_line *line_queue = NULL;
334 int line_queue_count = 0, total = 0, more = 0;
339 if (!syme->src->source)
340 lookup_sym_source(syme);
342 if (!syme->src->source)
345 symbol = sym_entry__symbol(syme);
346 printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
347 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
349 pthread_mutex_lock(&syme->src->lock);
350 line = syme->src->source;
352 total += line->count[sym_counter];
356 line = syme->src->source;
360 if (!line_queue_count)
364 if (line->count[sym_counter])
365 pcnt = 100.0 * line->count[sym_counter] / (float)total;
366 if (pcnt >= (float)sym_pcnt_filter) {
367 if (displayed <= print_entries)
368 show_lines(line_queue, line_queue_count, total);
370 displayed += line_queue_count;
371 line_queue_count = 0;
373 } else if (line_queue_count > TRACE_COUNT) {
374 line_queue = line_queue->next;
378 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
381 pthread_mutex_unlock(&syme->src->lock);
383 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
387 * Symbols will be added here in event__process_sample and will get out
390 static LIST_HEAD(active_symbols);
391 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
394 * Ordering weight: count-1 * count-2 * ... / count-n
396 static double sym_weight(const struct sym_entry *sym)
398 double weight = sym->snap_count;
401 if (!display_weighted)
404 for (counter = 1; counter < nr_counters-1; counter++)
405 weight *= sym->count[counter];
407 weight /= (sym->count[counter] + 1);
413 static long userspace_samples;
414 static long exact_samples;
415 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
417 static void __list_insert_active_sym(struct sym_entry *syme)
419 list_add(&syme->node, &active_symbols);
422 static void list_remove_active_sym(struct sym_entry *syme)
424 pthread_mutex_lock(&active_symbols_lock);
425 list_del_init(&syme->node);
426 pthread_mutex_unlock(&active_symbols_lock);
429 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
431 struct rb_node **p = &tree->rb_node;
432 struct rb_node *parent = NULL;
433 struct sym_entry *iter;
437 iter = rb_entry(parent, struct sym_entry, rb_node);
439 if (se->weight > iter->weight)
445 rb_link_node(&se->rb_node, parent, p);
446 rb_insert_color(&se->rb_node, tree);
449 static void print_sym_table(void)
452 int counter, snap = !display_weighted ? sym_counter : 0;
453 float samples_per_sec = samples/delay_secs;
454 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
455 float esamples_percent = (100.0*exact_samples)/samples;
456 float sum_ksamples = 0.0;
457 struct sym_entry *syme, *n;
458 struct rb_root tmp = RB_ROOT;
460 int sym_width = 0, dso_width = 0, max_dso_width;
461 const int win_width = winsize.ws_col - 1;
463 samples = userspace_samples = exact_samples = 0;
465 /* Sort the active symbols */
466 pthread_mutex_lock(&active_symbols_lock);
467 syme = list_entry(active_symbols.next, struct sym_entry, node);
468 pthread_mutex_unlock(&active_symbols_lock);
470 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
471 syme->snap_count = syme->count[snap];
472 if (syme->snap_count != 0) {
474 if ((hide_user_symbols &&
475 syme->origin == PERF_RECORD_MISC_USER) ||
476 (hide_kernel_symbols &&
477 syme->origin == PERF_RECORD_MISC_KERNEL)) {
478 list_remove_active_sym(syme);
481 syme->weight = sym_weight(syme);
482 rb_insert_active_sym(&tmp, syme);
483 sum_ksamples += syme->snap_count;
485 for (j = 0; j < nr_counters; j++)
486 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
488 list_remove_active_sym(syme);
493 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
494 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% exact: %4.1f%% [",
496 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)),
499 if (nr_counters == 1 || !display_weighted) {
500 printf("%Ld", (u64)attrs[0].sample_period);
507 if (!display_weighted)
508 printf("%s", event_name(sym_counter));
509 else for (counter = 0; counter < nr_counters; counter++) {
513 printf("%s", event_name(counter));
518 if (target_pid != -1)
519 printf(" (target_pid: %d", target_pid);
523 if (profile_cpu != -1)
524 printf(", cpu: %d)\n", profile_cpu);
526 if (target_pid != -1)
529 printf(", %d CPUs)\n", nr_cpus);
532 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
534 if (sym_filter_entry) {
535 show_details(sym_filter_entry);
540 * Find the longest symbol name that will be displayed
542 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
543 syme = rb_entry(nd, struct sym_entry, rb_node);
544 if (++printed > print_entries ||
545 (int)syme->snap_count < count_filter)
548 if (syme->map->dso->long_name_len > dso_width)
549 dso_width = syme->map->dso->long_name_len;
551 if (syme->name_len > sym_width)
552 sym_width = syme->name_len;
557 max_dso_width = winsize.ws_col - sym_width - 29;
558 if (dso_width > max_dso_width)
559 dso_width = max_dso_width;
561 if (nr_counters == 1)
562 printf(" samples pcnt");
564 printf(" weight samples pcnt");
568 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
569 printf(" %s _______ _____",
570 nr_counters == 1 ? " " : "______");
572 printf(" ________________");
573 printf(" %-*.*s", sym_width, sym_width, graph_line);
574 printf(" %-*.*s", dso_width, dso_width, graph_line);
577 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
581 syme = rb_entry(nd, struct sym_entry, rb_node);
582 sym = sym_entry__symbol(syme);
584 if (++printed > print_entries || (int)syme->snap_count < count_filter)
587 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
590 if (nr_counters == 1 || !display_weighted)
591 printf("%20.2f ", syme->weight);
593 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
595 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
597 printf(" %016llx", sym->start);
598 printf(" %-*.*s", sym_width, sym_width, sym->name);
599 printf(" %-*.*s\n", dso_width, dso_width,
600 dso_width >= syme->map->dso->long_name_len ?
601 syme->map->dso->long_name :
602 syme->map->dso->short_name);
606 static void prompt_integer(int *target, const char *msg)
608 char *buf = malloc(0), *p;
612 fprintf(stdout, "\n%s: ", msg);
613 if (getline(&buf, &dummy, stdin) < 0)
616 p = strchr(buf, '\n');
626 tmp = strtoul(buf, NULL, 10);
632 static void prompt_percent(int *target, const char *msg)
636 prompt_integer(&tmp, msg);
637 if (tmp >= 0 && tmp <= 100)
641 static void prompt_symbol(struct sym_entry **target, const char *msg)
643 char *buf = malloc(0), *p;
644 struct sym_entry *syme = *target, *n, *found = NULL;
647 /* zero counters of active symbol */
649 pthread_mutex_lock(&syme->src->lock);
650 __zero_source_counters(syme);
652 pthread_mutex_unlock(&syme->src->lock);
655 fprintf(stdout, "\n%s: ", msg);
656 if (getline(&buf, &dummy, stdin) < 0)
659 p = strchr(buf, '\n');
663 pthread_mutex_lock(&active_symbols_lock);
664 syme = list_entry(active_symbols.next, struct sym_entry, node);
665 pthread_mutex_unlock(&active_symbols_lock);
667 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
668 struct symbol *sym = sym_entry__symbol(syme);
670 if (!strcmp(buf, sym->name)) {
677 fprintf(stderr, "Sorry, %s is not active.\n", buf);
687 static void print_mapped_keys(void)
691 if (sym_filter_entry) {
692 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
696 fprintf(stdout, "\nMapped keys:\n");
697 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
698 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
701 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter));
703 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
705 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
706 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
707 fprintf(stdout, "\t[S] stop annotation.\n");
710 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
713 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
714 hide_kernel_symbols ? "yes" : "no");
716 "\t[U] hide user symbols. \t(%s)\n",
717 hide_user_symbols ? "yes" : "no");
718 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
719 fprintf(stdout, "\t[qQ] quit.\n");
722 static int key_mapped(int c)
739 return nr_counters > 1 ? 1 : 0;
747 static void handle_keypress(int c)
749 if (!key_mapped(c)) {
750 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
751 struct termios tc, save;
754 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
759 tc.c_lflag &= ~(ICANON | ECHO);
762 tcsetattr(0, TCSANOW, &tc);
764 poll(&stdin_poll, 1, -1);
767 tcsetattr(0, TCSAFLUSH, &save);
774 prompt_integer(&delay_secs, "Enter display delay");
779 prompt_integer(&print_entries, "Enter display entries (lines)");
780 if (print_entries == 0) {
781 sig_winch_handler(SIGWINCH);
782 signal(SIGWINCH, sig_winch_handler);
784 signal(SIGWINCH, SIG_DFL);
787 if (nr_counters > 1) {
790 fprintf(stderr, "\nAvailable events:");
791 for (i = 0; i < nr_counters; i++)
792 fprintf(stderr, "\n\t%d %s", i, event_name(i));
794 prompt_integer(&sym_counter, "Enter details event counter");
796 if (sym_counter >= nr_counters) {
797 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
801 } else sym_counter = 0;
804 prompt_integer(&count_filter, "Enter display event count filter");
807 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
810 hide_kernel_symbols = !hide_kernel_symbols;
814 printf("exiting.\n");
816 dsos__fprintf(stderr);
819 prompt_symbol(&sym_filter_entry, "Enter details symbol");
822 if (!sym_filter_entry)
825 struct sym_entry *syme = sym_filter_entry;
827 pthread_mutex_lock(&syme->src->lock);
828 sym_filter_entry = NULL;
829 __zero_source_counters(syme);
830 pthread_mutex_unlock(&syme->src->lock);
834 hide_user_symbols = !hide_user_symbols;
837 display_weighted = ~display_weighted;
847 static void *display_thread(void *arg __used)
849 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
850 struct termios tc, save;
855 tc.c_lflag &= ~(ICANON | ECHO);
860 delay_msecs = delay_secs * 1000;
861 tcsetattr(0, TCSANOW, &tc);
867 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
870 tcsetattr(0, TCSAFLUSH, &save);
878 /* Tag samples to be skipped. */
879 static const char *skip_symbols[] = {
885 "mwait_idle_with_hints",
887 "ppc64_runlatch_off",
888 "pseries_dedicated_idle_sleep",
892 static int symbol_filter(struct map *map, struct symbol *sym)
894 struct sym_entry *syme;
895 const char *name = sym->name;
899 * ppc64 uses function descriptors and appends a '.' to the
900 * start of every instruction address. Remove it.
905 if (!strcmp(name, "_text") ||
906 !strcmp(name, "_etext") ||
907 !strcmp(name, "_sinittext") ||
908 !strncmp("init_module", name, 11) ||
909 !strncmp("cleanup_module", name, 14) ||
910 strstr(name, "_text_start") ||
911 strstr(name, "_text_end"))
914 syme = symbol__priv(sym);
918 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
919 /* schedule initial sym_filter_entry setup */
920 sym_filter_entry_sched = syme;
924 for (i = 0; skip_symbols[i]; i++) {
925 if (!strcmp(skip_symbols[i], name)) {
932 syme->name_len = strlen(sym->name);
937 static void event__process_sample(const event_t *self,
938 struct perf_session *session, int counter)
940 u64 ip = self->ip.ip;
941 struct sym_entry *syme;
942 struct addr_location al;
943 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
948 case PERF_RECORD_MISC_USER:
950 if (hide_user_symbols)
953 case PERF_RECORD_MISC_KERNEL:
954 if (hide_kernel_symbols)
961 if (self->header.misc & PERF_RECORD_MISC_EXACT)
964 if (event__preprocess_sample(self, session, &al, symbol_filter) < 0 ||
968 if (al.sym == NULL) {
970 * As we do lazy loading of symtabs we only will know if the
971 * specified vmlinux file is invalid when we actually have a
972 * hit in kernel space and then try to load it. So if we get
973 * here and there are _no_ symbols in the DSO backing the
974 * kernel map, bail out.
976 * We may never get here, for instance, if we use -K/
977 * --hide-kernel-symbols, even if the user specifies an
978 * invalid --vmlinux ;-)
980 if (al.map == session->vmlinux_maps[MAP__FUNCTION] &&
981 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
982 pr_err("The %s file can't be used\n",
983 symbol_conf.vmlinux_name);
990 /* let's see, whether we need to install initial sym_filter_entry */
991 if (sym_filter_entry_sched) {
992 sym_filter_entry = sym_filter_entry_sched;
993 sym_filter_entry_sched = NULL;
994 parse_source(sym_filter_entry);
997 syme = symbol__priv(al.sym);
999 syme->count[counter]++;
1000 syme->origin = origin;
1001 record_precise_ip(syme, counter, ip);
1002 pthread_mutex_lock(&active_symbols_lock);
1003 if (list_empty(&syme->node) || !syme->node.next)
1004 __list_insert_active_sym(syme);
1005 pthread_mutex_unlock(&active_symbols_lock);
1009 static int event__process(event_t *event, struct perf_session *session)
1011 switch (event->header.type) {
1012 case PERF_RECORD_COMM:
1013 event__process_comm(event, session);
1015 case PERF_RECORD_MMAP:
1016 event__process_mmap(event, session);
1018 case PERF_RECORD_FORK:
1019 case PERF_RECORD_EXIT:
1020 event__process_task(event, session);
1036 static unsigned int mmap_read_head(struct mmap_data *md)
1038 struct perf_event_mmap_page *pc = md->base;
1041 head = pc->data_head;
1047 static void perf_session__mmap_read_counter(struct perf_session *self,
1048 struct mmap_data *md)
1050 unsigned int head = mmap_read_head(md);
1051 unsigned int old = md->prev;
1052 unsigned char *data = md->base + page_size;
1056 * If we're further behind than half the buffer, there's a chance
1057 * the writer will bite our tail and mess up the samples under us.
1059 * If we somehow ended up ahead of the head, we got messed up.
1061 * In either case, truncate and restart at head.
1064 if (diff > md->mask / 2 || diff < 0) {
1065 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1068 * head points to a known good entry, start there.
1073 for (; old != head;) {
1074 event_t *event = (event_t *)&data[old & md->mask];
1078 size_t size = event->header.size;
1081 * Event straddles the mmap boundary -- header should always
1082 * be inside due to u64 alignment of output.
1084 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1085 unsigned int offset = old;
1086 unsigned int len = min(sizeof(*event), size), cpy;
1087 void *dst = &event_copy;
1090 cpy = min(md->mask + 1 - (offset & md->mask), len);
1091 memcpy(dst, &data[offset & md->mask], cpy);
1097 event = &event_copy;
1100 if (event->header.type == PERF_RECORD_SAMPLE)
1101 event__process_sample(event, self, md->counter);
1103 event__process(event, self);
1110 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
1111 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1113 static void perf_session__mmap_read(struct perf_session *self)
1117 for (i = 0; i < nr_cpus; i++) {
1118 for (counter = 0; counter < nr_counters; counter++)
1119 perf_session__mmap_read_counter(self, &mmap_array[i][counter]);
1126 static void start_counter(int i, int counter)
1128 struct perf_event_attr *attr;
1132 if (target_pid == -1 && profile_cpu == -1)
1135 attr = attrs + counter;
1137 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1140 attr->sample_type |= PERF_SAMPLE_PERIOD;
1142 attr->sample_freq = freq;
1145 attr->inherit = (cpu < 0) && inherit;
1149 fd[i][counter] = sys_perf_event_open(attr, target_pid, cpu, group_fd, 0);
1151 if (fd[i][counter] < 0) {
1154 if (err == EPERM || err == EACCES)
1155 die("No permission - are you root?\n");
1157 * If it's cycles then fall back to hrtimer
1158 * based cpu-clock-tick sw counter, which
1159 * is always available even if no PMU support:
1161 if (attr->type == PERF_TYPE_HARDWARE
1162 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1165 warning(" ... trying to fall back to cpu-clock-ticks\n");
1167 attr->type = PERF_TYPE_SOFTWARE;
1168 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1172 error("perfcounter syscall returned with %d (%s)\n",
1173 fd[i][counter], strerror(err));
1174 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1177 assert(fd[i][counter] >= 0);
1178 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
1181 * First counter acts as the group leader:
1183 if (group && group_fd == -1)
1184 group_fd = fd[i][counter];
1186 event_array[nr_poll].fd = fd[i][counter];
1187 event_array[nr_poll].events = POLLIN;
1190 mmap_array[i][counter].counter = counter;
1191 mmap_array[i][counter].prev = 0;
1192 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
1193 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
1194 PROT_READ, MAP_SHARED, fd[i][counter], 0);
1195 if (mmap_array[i][counter].base == MAP_FAILED)
1196 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1199 static int __cmd_top(void)
1205 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1206 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1208 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false);
1209 if (session == NULL)
1212 if (target_pid != -1)
1213 event__synthesize_thread(target_pid, event__process, session);
1215 event__synthesize_threads(event__process, session);
1217 for (i = 0; i < nr_cpus; i++) {
1219 for (counter = 0; counter < nr_counters; counter++)
1220 start_counter(i, counter);
1223 /* Wait for a minimal set of events before starting the snapshot */
1224 poll(event_array, nr_poll, 100);
1226 perf_session__mmap_read(session);
1228 if (pthread_create(&thread, NULL, display_thread, NULL)) {
1229 printf("Could not create display thread.\n");
1233 if (realtime_prio) {
1234 struct sched_param param;
1236 param.sched_priority = realtime_prio;
1237 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1238 printf("Could not set realtime priority.\n");
1246 perf_session__mmap_read(session);
1248 if (hits == samples)
1249 ret = poll(event_array, nr_poll, 100);
1255 static const char * const top_usage[] = {
1256 "perf top [<options>]",
1260 static const struct option options[] = {
1261 OPT_CALLBACK('e', "event", NULL, "event",
1262 "event selector. use 'perf list' to list available events",
1264 OPT_INTEGER('c', "count", &default_interval,
1265 "event period to sample"),
1266 OPT_INTEGER('p', "pid", &target_pid,
1267 "profile events on existing pid"),
1268 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1269 "system-wide collection from all CPUs"),
1270 OPT_INTEGER('C', "CPU", &profile_cpu,
1271 "CPU to profile on"),
1272 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1273 "file", "vmlinux pathname"),
1274 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1275 "hide kernel symbols"),
1276 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
1277 "number of mmap data pages"),
1278 OPT_INTEGER('r', "realtime", &realtime_prio,
1279 "collect data with this RT SCHED_FIFO priority"),
1280 OPT_INTEGER('d', "delay", &delay_secs,
1281 "number of seconds to delay between refreshes"),
1282 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1283 "dump the symbol table used for profiling"),
1284 OPT_INTEGER('f', "count-filter", &count_filter,
1285 "only display functions with more events than this"),
1286 OPT_BOOLEAN('g', "group", &group,
1287 "put the counters into a counter group"),
1288 OPT_BOOLEAN('i', "inherit", &inherit,
1289 "child tasks inherit counters"),
1290 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1291 "symbol to annotate"),
1292 OPT_BOOLEAN('z', "zero", &zero,
1293 "zero history across updates"),
1294 OPT_INTEGER('F', "freq", &freq,
1295 "profile at this frequency"),
1296 OPT_INTEGER('E', "entries", &print_entries,
1297 "display this many functions"),
1298 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1299 "hide user symbols"),
1300 OPT_BOOLEAN('v', "verbose", &verbose,
1301 "be more verbose (show counter open errors, etc)"),
1305 int cmd_top(int argc, const char **argv, const char *prefix __used)
1309 page_size = sysconf(_SC_PAGE_SIZE);
1311 argc = parse_options(argc, argv, options, top_usage, 0);
1313 usage_with_options(top_usage, options);
1315 /* CPU and PID are mutually exclusive */
1316 if (target_pid != -1 && profile_cpu != -1) {
1317 printf("WARNING: PID switch overriding CPU\n");
1325 symbol_conf.priv_size = (sizeof(struct sym_entry) +
1326 (nr_counters + 1) * sizeof(unsigned long));
1328 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1329 if (symbol__init() < 0)
1336 * User specified count overrides default frequency.
1338 if (default_interval)
1341 default_interval = freq;
1343 fprintf(stderr, "frequency and count are zero, aborting\n");
1348 * Fill in the ones not specifically initialized via -c:
1350 for (counter = 0; counter < nr_counters; counter++) {
1351 if (attrs[counter].sample_period)
1354 attrs[counter].sample_period = default_interval;
1357 if (target_pid != -1 || profile_cpu != -1)
1360 nr_cpus = read_cpu_map();
1362 get_term_dimensions(&winsize);
1363 if (print_entries == 0) {
1364 update_print_entries(&winsize);
1365 signal(SIGWINCH, sig_winch_handler);