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 bool system_wide = false;
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 target_tid = -1;
69 static pid_t *all_tids = NULL;
70 static int thread_num = 0;
71 static bool inherit = false;
72 static int profile_cpu = -1;
73 static int nr_cpus = 0;
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 int display_weighted = -1;
105 static const char *cpu_list;
111 struct sym_entry_source {
112 struct source_line *source;
113 struct source_line *lines;
114 struct source_line **lines_tail;
115 pthread_mutex_t lock;
119 struct rb_node rb_node;
120 struct list_head node;
121 unsigned long snap_count;
127 struct sym_entry_source *src;
128 unsigned long count[0];
135 static inline struct symbol *sym_entry__symbol(struct sym_entry *self)
137 return ((void *)self) + symbol_conf.priv_size;
140 void get_term_dimensions(struct winsize *ws)
142 char *s = getenv("LINES");
145 ws->ws_row = atoi(s);
146 s = getenv("COLUMNS");
148 ws->ws_col = atoi(s);
149 if (ws->ws_row && ws->ws_col)
154 if (ioctl(1, TIOCGWINSZ, ws) == 0 &&
155 ws->ws_row && ws->ws_col)
162 static void update_print_entries(struct winsize *ws)
164 print_entries = ws->ws_row;
166 if (print_entries > 9)
170 static void sig_winch_handler(int sig __used)
172 get_term_dimensions(&winsize);
173 update_print_entries(&winsize);
176 static int parse_source(struct sym_entry *syme)
179 struct sym_entry_source *source;
182 char command[PATH_MAX*2];
189 sym = sym_entry__symbol(syme);
193 * We can't annotate with just /proc/kallsyms
195 if (map->dso->origin == DSO__ORIG_KERNEL)
198 if (syme->src == NULL) {
199 syme->src = zalloc(sizeof(*source));
200 if (syme->src == NULL)
202 pthread_mutex_init(&syme->src->lock, NULL);
208 pthread_mutex_lock(&source->lock);
211 path = map->dso->long_name;
213 len = sym->end - sym->start;
216 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
217 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->start),
218 BITS_PER_LONG / 4, map__rip_2objdump(map, sym->end), path);
220 file = popen(command, "r");
224 pthread_mutex_lock(&source->lock);
225 source->lines_tail = &source->lines;
226 while (!feof(file)) {
227 struct source_line *src;
231 src = malloc(sizeof(struct source_line));
233 memset(src, 0, sizeof(struct source_line));
235 if (getline(&src->line, &dummy, file) < 0)
240 c = strchr(src->line, '\n');
245 *source->lines_tail = src;
246 source->lines_tail = &src->next;
248 src->eip = strtoull(src->line, &sep, 16);
250 src->eip = map__objdump_2ip(map, src->eip);
251 else /* this line has no ip info (e.g. source line) */
256 sym_filter_entry = syme;
257 pthread_mutex_unlock(&source->lock);
261 static void __zero_source_counters(struct sym_entry *syme)
264 struct source_line *line;
266 line = syme->src->lines;
268 for (i = 0; i < nr_counters; i++)
274 static void record_precise_ip(struct sym_entry *syme, int counter, u64 ip)
276 struct source_line *line;
278 if (syme != sym_filter_entry)
281 if (pthread_mutex_trylock(&syme->src->lock))
284 if (syme->src == NULL || syme->src->source == NULL)
287 for (line = syme->src->lines; line; line = line->next) {
288 /* skip lines without IP info */
291 if (line->eip == ip) {
292 line->count[counter]++;
299 pthread_mutex_unlock(&syme->src->lock);
302 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
304 static void lookup_sym_source(struct sym_entry *syme)
306 struct symbol *symbol = sym_entry__symbol(syme);
307 struct source_line *line;
308 char pattern[PATTERN_LEN + 1];
310 sprintf(pattern, "%0*Lx <", BITS_PER_LONG / 4,
311 map__rip_2objdump(syme->map, symbol->start));
313 pthread_mutex_lock(&syme->src->lock);
314 for (line = syme->src->lines; line; line = line->next) {
315 if (memcmp(line->line, pattern, PATTERN_LEN) == 0) {
316 syme->src->source = line;
320 pthread_mutex_unlock(&syme->src->lock);
323 static void show_lines(struct source_line *queue, int count, int total)
326 struct source_line *line;
329 for (i = 0; i < count; i++) {
330 float pcnt = 100.0*(float)line->count[sym_counter]/(float)total;
332 printf("%8li %4.1f%%\t%s\n", line->count[sym_counter], pcnt, line->line);
337 #define TRACE_COUNT 3
339 static void show_details(struct sym_entry *syme)
341 struct symbol *symbol;
342 struct source_line *line;
343 struct source_line *line_queue = NULL;
345 int line_queue_count = 0, total = 0, more = 0;
350 if (!syme->src->source)
351 lookup_sym_source(syme);
353 if (!syme->src->source)
356 symbol = sym_entry__symbol(syme);
357 printf("Showing %s for %s\n", event_name(sym_counter), symbol->name);
358 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter);
360 pthread_mutex_lock(&syme->src->lock);
361 line = syme->src->source;
363 total += line->count[sym_counter];
367 line = syme->src->source;
371 if (!line_queue_count)
375 if (line->count[sym_counter])
376 pcnt = 100.0 * line->count[sym_counter] / (float)total;
377 if (pcnt >= (float)sym_pcnt_filter) {
378 if (displayed <= print_entries)
379 show_lines(line_queue, line_queue_count, total);
381 displayed += line_queue_count;
382 line_queue_count = 0;
384 } else if (line_queue_count > TRACE_COUNT) {
385 line_queue = line_queue->next;
389 line->count[sym_counter] = zero ? 0 : line->count[sym_counter] * 7 / 8;
392 pthread_mutex_unlock(&syme->src->lock);
394 printf("%d lines not displayed, maybe increase display entries [e]\n", more);
398 * Symbols will be added here in event__process_sample and will get out
401 static LIST_HEAD(active_symbols);
402 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
405 * Ordering weight: count-1 * count-2 * ... / count-n
407 static double sym_weight(const struct sym_entry *sym)
409 double weight = sym->snap_count;
412 if (!display_weighted)
415 for (counter = 1; counter < nr_counters-1; counter++)
416 weight *= sym->count[counter];
418 weight /= (sym->count[counter] + 1);
424 static long kernel_samples, us_samples;
425 static long exact_samples;
426 static long guest_us_samples, guest_kernel_samples;
427 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
429 static void __list_insert_active_sym(struct sym_entry *syme)
431 list_add(&syme->node, &active_symbols);
434 static void list_remove_active_sym(struct sym_entry *syme)
436 pthread_mutex_lock(&active_symbols_lock);
437 list_del_init(&syme->node);
438 pthread_mutex_unlock(&active_symbols_lock);
441 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
443 struct rb_node **p = &tree->rb_node;
444 struct rb_node *parent = NULL;
445 struct sym_entry *iter;
449 iter = rb_entry(parent, struct sym_entry, rb_node);
451 if (se->weight > iter->weight)
457 rb_link_node(&se->rb_node, parent, p);
458 rb_insert_color(&se->rb_node, tree);
461 static void print_sym_table(void)
464 int counter, snap = !display_weighted ? sym_counter : 0;
465 float samples_per_sec = samples/delay_secs;
466 float ksamples_per_sec = kernel_samples/delay_secs;
467 float us_samples_per_sec = (us_samples)/delay_secs;
468 float guest_kernel_samples_per_sec = (guest_kernel_samples)/delay_secs;
469 float guest_us_samples_per_sec = (guest_us_samples)/delay_secs;
470 float esamples_percent = (100.0*exact_samples)/samples;
471 float sum_ksamples = 0.0;
472 struct sym_entry *syme, *n;
473 struct rb_root tmp = RB_ROOT;
475 int sym_width = 0, dso_width = 0, dso_short_width = 0;
476 const int win_width = winsize.ws_col - 1;
478 samples = us_samples = kernel_samples = exact_samples = 0;
479 guest_kernel_samples = guest_us_samples = 0;
481 /* Sort the active symbols */
482 pthread_mutex_lock(&active_symbols_lock);
483 syme = list_entry(active_symbols.next, struct sym_entry, node);
484 pthread_mutex_unlock(&active_symbols_lock);
486 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
487 syme->snap_count = syme->count[snap];
488 if (syme->snap_count != 0) {
490 if ((hide_user_symbols &&
491 syme->origin == PERF_RECORD_MISC_USER) ||
492 (hide_kernel_symbols &&
493 syme->origin == PERF_RECORD_MISC_KERNEL)) {
494 list_remove_active_sym(syme);
497 syme->weight = sym_weight(syme);
498 rb_insert_active_sym(&tmp, syme);
499 sum_ksamples += syme->snap_count;
501 for (j = 0; j < nr_counters; j++)
502 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
504 list_remove_active_sym(syme);
509 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
511 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%"
514 100.0 - (100.0 * ((samples_per_sec - ksamples_per_sec) /
518 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%"
519 " guest kernel:%4.1f%% guest us:%4.1f%%"
522 100.0 - (100.0 * ((samples_per_sec-ksamples_per_sec) /
524 100.0 - (100.0 * ((samples_per_sec-us_samples_per_sec) /
526 100.0 - (100.0 * ((samples_per_sec -
527 guest_kernel_samples_per_sec) /
529 100.0 - (100.0 * ((samples_per_sec -
530 guest_us_samples_per_sec) /
535 if (nr_counters == 1 || !display_weighted) {
536 printf("%Ld", (u64)attrs[0].sample_period);
543 if (!display_weighted)
544 printf("%s", event_name(sym_counter));
545 else for (counter = 0; counter < nr_counters; counter++) {
549 printf("%s", event_name(counter));
554 if (target_pid != -1)
555 printf(" (target_pid: %d", target_pid);
556 else if (target_tid != -1)
557 printf(" (target_tid: %d", target_tid);
561 if (profile_cpu != -1)
562 printf(", cpu: %d)\n", profile_cpu);
564 if (target_tid != -1)
567 printf(", %d CPUs)\n", nr_cpus);
570 printf("%-*.*s\n", win_width, win_width, graph_dotted_line);
572 if (sym_filter_entry) {
573 show_details(sym_filter_entry);
578 * Find the longest symbol name that will be displayed
580 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
581 syme = rb_entry(nd, struct sym_entry, rb_node);
582 if (++printed > print_entries ||
583 (int)syme->snap_count < count_filter)
586 if (syme->map->dso->long_name_len > dso_width)
587 dso_width = syme->map->dso->long_name_len;
589 if (syme->map->dso->short_name_len > dso_short_width)
590 dso_short_width = syme->map->dso->short_name_len;
592 if (syme->name_len > sym_width)
593 sym_width = syme->name_len;
598 if (sym_width + dso_width > winsize.ws_col - 29) {
599 dso_width = dso_short_width;
600 if (sym_width + dso_width > winsize.ws_col - 29)
601 sym_width = winsize.ws_col - dso_width - 29;
604 if (nr_counters == 1)
605 printf(" samples pcnt");
607 printf(" weight samples pcnt");
611 printf(" %-*.*s DSO\n", sym_width, sym_width, "function");
612 printf(" %s _______ _____",
613 nr_counters == 1 ? " " : "______");
615 printf(" ________________");
616 printf(" %-*.*s", sym_width, sym_width, graph_line);
617 printf(" %-*.*s", dso_width, dso_width, graph_line);
620 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
624 syme = rb_entry(nd, struct sym_entry, rb_node);
625 sym = sym_entry__symbol(syme);
626 if (++printed > print_entries || (int)syme->snap_count < count_filter)
629 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
632 if (nr_counters == 1 || !display_weighted)
633 printf("%20.2f ", syme->weight);
635 printf("%9.1f %10ld ", syme->weight, syme->snap_count);
637 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
639 printf(" %016llx", sym->start);
640 printf(" %-*.*s", sym_width, sym_width, sym->name);
641 printf(" %-*.*s\n", dso_width, dso_width,
642 dso_width >= syme->map->dso->long_name_len ?
643 syme->map->dso->long_name :
644 syme->map->dso->short_name);
648 static void prompt_integer(int *target, const char *msg)
650 char *buf = malloc(0), *p;
654 fprintf(stdout, "\n%s: ", msg);
655 if (getline(&buf, &dummy, stdin) < 0)
658 p = strchr(buf, '\n');
668 tmp = strtoul(buf, NULL, 10);
674 static void prompt_percent(int *target, const char *msg)
678 prompt_integer(&tmp, msg);
679 if (tmp >= 0 && tmp <= 100)
683 static void prompt_symbol(struct sym_entry **target, const char *msg)
685 char *buf = malloc(0), *p;
686 struct sym_entry *syme = *target, *n, *found = NULL;
689 /* zero counters of active symbol */
691 pthread_mutex_lock(&syme->src->lock);
692 __zero_source_counters(syme);
694 pthread_mutex_unlock(&syme->src->lock);
697 fprintf(stdout, "\n%s: ", msg);
698 if (getline(&buf, &dummy, stdin) < 0)
701 p = strchr(buf, '\n');
705 pthread_mutex_lock(&active_symbols_lock);
706 syme = list_entry(active_symbols.next, struct sym_entry, node);
707 pthread_mutex_unlock(&active_symbols_lock);
709 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
710 struct symbol *sym = sym_entry__symbol(syme);
712 if (!strcmp(buf, sym->name)) {
719 fprintf(stderr, "Sorry, %s is not active.\n", buf);
729 static void print_mapped_keys(void)
733 if (sym_filter_entry) {
734 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
738 fprintf(stdout, "\nMapped keys:\n");
739 fprintf(stdout, "\t[d] display refresh delay. \t(%d)\n", delay_secs);
740 fprintf(stdout, "\t[e] display entries (lines). \t(%d)\n", print_entries);
743 fprintf(stdout, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter));
745 fprintf(stdout, "\t[f] profile display filter (count). \t(%d)\n", count_filter);
747 fprintf(stdout, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter);
748 fprintf(stdout, "\t[s] annotate symbol. \t(%s)\n", name?: "NULL");
749 fprintf(stdout, "\t[S] stop annotation.\n");
752 fprintf(stdout, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted ? 1 : 0);
755 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
756 hide_kernel_symbols ? "yes" : "no");
758 "\t[U] hide user symbols. \t(%s)\n",
759 hide_user_symbols ? "yes" : "no");
760 fprintf(stdout, "\t[z] toggle sample zeroing. \t(%d)\n", zero ? 1 : 0);
761 fprintf(stdout, "\t[qQ] quit.\n");
764 static int key_mapped(int c)
781 return nr_counters > 1 ? 1 : 0;
789 static void handle_keypress(struct perf_session *session, int c)
791 if (!key_mapped(c)) {
792 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
793 struct termios tc, save;
796 fprintf(stdout, "\nEnter selection, or unmapped key to continue: ");
801 tc.c_lflag &= ~(ICANON | ECHO);
804 tcsetattr(0, TCSANOW, &tc);
806 poll(&stdin_poll, 1, -1);
809 tcsetattr(0, TCSAFLUSH, &save);
816 prompt_integer(&delay_secs, "Enter display delay");
821 prompt_integer(&print_entries, "Enter display entries (lines)");
822 if (print_entries == 0) {
823 sig_winch_handler(SIGWINCH);
824 signal(SIGWINCH, sig_winch_handler);
826 signal(SIGWINCH, SIG_DFL);
829 if (nr_counters > 1) {
832 fprintf(stderr, "\nAvailable events:");
833 for (i = 0; i < nr_counters; i++)
834 fprintf(stderr, "\n\t%d %s", i, event_name(i));
836 prompt_integer(&sym_counter, "Enter details event counter");
838 if (sym_counter >= nr_counters) {
839 fprintf(stderr, "Sorry, no such event, using %s.\n", event_name(0));
843 } else sym_counter = 0;
846 prompt_integer(&count_filter, "Enter display event count filter");
849 prompt_percent(&sym_pcnt_filter, "Enter details display event filter (percent)");
852 hide_kernel_symbols = !hide_kernel_symbols;
856 printf("exiting.\n");
858 perf_session__fprintf_dsos(session, stderr);
861 prompt_symbol(&sym_filter_entry, "Enter details symbol");
864 if (!sym_filter_entry)
867 struct sym_entry *syme = sym_filter_entry;
869 pthread_mutex_lock(&syme->src->lock);
870 sym_filter_entry = NULL;
871 __zero_source_counters(syme);
872 pthread_mutex_unlock(&syme->src->lock);
876 hide_user_symbols = !hide_user_symbols;
879 display_weighted = ~display_weighted;
889 static void *display_thread(void *arg __used)
891 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
892 struct termios tc, save;
894 struct perf_session *session = (struct perf_session *) arg;
898 tc.c_lflag &= ~(ICANON | ECHO);
903 delay_msecs = delay_secs * 1000;
904 tcsetattr(0, TCSANOW, &tc);
910 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
913 tcsetattr(0, TCSAFLUSH, &save);
915 handle_keypress(session, c);
921 /* Tag samples to be skipped. */
922 static const char *skip_symbols[] = {
928 "mwait_idle_with_hints",
930 "ppc64_runlatch_off",
931 "pseries_dedicated_idle_sleep",
935 static int symbol_filter(struct map *map, struct symbol *sym)
937 struct sym_entry *syme;
938 const char *name = sym->name;
942 * ppc64 uses function descriptors and appends a '.' to the
943 * start of every instruction address. Remove it.
948 if (!strcmp(name, "_text") ||
949 !strcmp(name, "_etext") ||
950 !strcmp(name, "_sinittext") ||
951 !strncmp("init_module", name, 11) ||
952 !strncmp("cleanup_module", name, 14) ||
953 strstr(name, "_text_start") ||
954 strstr(name, "_text_end"))
957 syme = symbol__priv(sym);
961 if (!sym_filter_entry && sym_filter && !strcmp(name, sym_filter)) {
962 /* schedule initial sym_filter_entry setup */
963 sym_filter_entry_sched = syme;
967 for (i = 0; skip_symbols[i]; i++) {
968 if (!strcmp(skip_symbols[i], name)) {
975 syme->name_len = strlen(sym->name);
980 static void event__process_sample(const event_t *self,
981 struct perf_session *session, int counter)
983 u64 ip = self->ip.ip;
984 struct sym_entry *syme;
985 struct addr_location al;
986 struct sample_data data;
987 struct machine *machine;
988 u8 origin = self->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
993 case PERF_RECORD_MISC_USER:
995 if (hide_user_symbols)
997 machine = perf_session__find_host_machine(session);
999 case PERF_RECORD_MISC_KERNEL:
1001 if (hide_kernel_symbols)
1003 machine = perf_session__find_host_machine(session);
1005 case PERF_RECORD_MISC_GUEST_KERNEL:
1006 ++guest_kernel_samples;
1007 machine = perf_session__find_machine(session, self->ip.pid);
1009 case PERF_RECORD_MISC_GUEST_USER:
1012 * TODO: we don't process guest user from host side
1013 * except simple counting.
1020 if (!machine && perf_guest) {
1021 pr_err("Can't find guest [%d]'s kernel information\n",
1026 if (self->header.misc & PERF_RECORD_MISC_EXACT_IP)
1029 if (event__preprocess_sample(self, session, &al, &data,
1030 symbol_filter) < 0 ||
1034 if (al.sym == NULL) {
1036 * As we do lazy loading of symtabs we only will know if the
1037 * specified vmlinux file is invalid when we actually have a
1038 * hit in kernel space and then try to load it. So if we get
1039 * here and there are _no_ symbols in the DSO backing the
1040 * kernel map, bail out.
1042 * We may never get here, for instance, if we use -K/
1043 * --hide-kernel-symbols, even if the user specifies an
1044 * invalid --vmlinux ;-)
1046 if (al.map == machine->vmlinux_maps[MAP__FUNCTION] &&
1047 RB_EMPTY_ROOT(&al.map->dso->symbols[MAP__FUNCTION])) {
1048 pr_err("The %s file can't be used\n",
1049 symbol_conf.vmlinux_name);
1056 /* let's see, whether we need to install initial sym_filter_entry */
1057 if (sym_filter_entry_sched) {
1058 sym_filter_entry = sym_filter_entry_sched;
1059 sym_filter_entry_sched = NULL;
1060 if (parse_source(sym_filter_entry) < 0) {
1061 struct symbol *sym = sym_entry__symbol(sym_filter_entry);
1063 pr_err("Can't annotate %s", sym->name);
1064 if (sym_filter_entry->map->dso->origin == DSO__ORIG_KERNEL) {
1065 pr_err(": No vmlinux file was found in the path:\n");
1066 machine__fprintf_vmlinux_path(machine, stderr);
1073 syme = symbol__priv(al.sym);
1075 syme->count[counter]++;
1076 syme->origin = origin;
1077 record_precise_ip(syme, counter, ip);
1078 pthread_mutex_lock(&active_symbols_lock);
1079 if (list_empty(&syme->node) || !syme->node.next)
1080 __list_insert_active_sym(syme);
1081 pthread_mutex_unlock(&active_symbols_lock);
1085 static int event__process(event_t *event, struct perf_session *session)
1087 switch (event->header.type) {
1088 case PERF_RECORD_COMM:
1089 event__process_comm(event, session);
1091 case PERF_RECORD_MMAP:
1092 event__process_mmap(event, session);
1094 case PERF_RECORD_FORK:
1095 case PERF_RECORD_EXIT:
1096 event__process_task(event, session);
1112 static unsigned int mmap_read_head(struct mmap_data *md)
1114 struct perf_event_mmap_page *pc = md->base;
1117 head = pc->data_head;
1123 static void perf_session__mmap_read_counter(struct perf_session *self,
1124 struct mmap_data *md)
1126 unsigned int head = mmap_read_head(md);
1127 unsigned int old = md->prev;
1128 unsigned char *data = md->base + page_size;
1132 * If we're further behind than half the buffer, there's a chance
1133 * the writer will bite our tail and mess up the samples under us.
1135 * If we somehow ended up ahead of the head, we got messed up.
1137 * In either case, truncate and restart at head.
1140 if (diff > md->mask / 2 || diff < 0) {
1141 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
1144 * head points to a known good entry, start there.
1149 for (; old != head;) {
1150 event_t *event = (event_t *)&data[old & md->mask];
1154 size_t size = event->header.size;
1157 * Event straddles the mmap boundary -- header should always
1158 * be inside due to u64 alignment of output.
1160 if ((old & md->mask) + size != ((old + size) & md->mask)) {
1161 unsigned int offset = old;
1162 unsigned int len = min(sizeof(*event), size), cpy;
1163 void *dst = &event_copy;
1166 cpy = min(md->mask + 1 - (offset & md->mask), len);
1167 memcpy(dst, &data[offset & md->mask], cpy);
1173 event = &event_copy;
1176 if (event->header.type == PERF_RECORD_SAMPLE)
1177 event__process_sample(event, self, md->counter);
1179 event__process(event, self);
1186 static struct pollfd *event_array;
1187 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
1189 static void perf_session__mmap_read(struct perf_session *self)
1191 int i, counter, thread_index;
1193 for (i = 0; i < nr_cpus; i++) {
1194 for (counter = 0; counter < nr_counters; counter++)
1195 for (thread_index = 0;
1196 thread_index < thread_num;
1198 perf_session__mmap_read_counter(self,
1199 &mmap_array[i][counter][thread_index]);
1207 static void start_counter(int i, int counter)
1209 struct perf_event_attr *attr;
1214 if (target_tid == -1 && profile_cpu == -1)
1217 attr = attrs + counter;
1219 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
1222 attr->sample_type |= PERF_SAMPLE_PERIOD;
1224 attr->sample_freq = freq;
1227 attr->inherit = (cpu < 0) && inherit;
1230 for (thread_index = 0; thread_index < thread_num; thread_index++) {
1232 fd[i][counter][thread_index] = sys_perf_event_open(attr,
1233 all_tids[thread_index], cpu, group_fd, 0);
1235 if (fd[i][counter][thread_index] < 0) {
1238 if (err == EPERM || err == EACCES)
1239 die("No permission - are you root?\n");
1241 * If it's cycles then fall back to hrtimer
1242 * based cpu-clock-tick sw counter, which
1243 * is always available even if no PMU support:
1245 if (attr->type == PERF_TYPE_HARDWARE
1246 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
1249 warning(" ... trying to fall back to cpu-clock-ticks\n");
1251 attr->type = PERF_TYPE_SOFTWARE;
1252 attr->config = PERF_COUNT_SW_CPU_CLOCK;
1256 error("perfcounter syscall returned with %d (%s)\n",
1257 fd[i][counter][thread_index], strerror(err));
1258 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1261 assert(fd[i][counter][thread_index] >= 0);
1262 fcntl(fd[i][counter][thread_index], F_SETFL, O_NONBLOCK);
1265 * First counter acts as the group leader:
1267 if (group && group_fd == -1)
1268 group_fd = fd[i][counter][thread_index];
1270 event_array[nr_poll].fd = fd[i][counter][thread_index];
1271 event_array[nr_poll].events = POLLIN;
1274 mmap_array[i][counter][thread_index].counter = counter;
1275 mmap_array[i][counter][thread_index].prev = 0;
1276 mmap_array[i][counter][thread_index].mask = mmap_pages*page_size - 1;
1277 mmap_array[i][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
1278 PROT_READ, MAP_SHARED, fd[i][counter][thread_index], 0);
1279 if (mmap_array[i][counter][thread_index].base == MAP_FAILED)
1280 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
1284 static int __cmd_top(void)
1290 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1291 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1293 struct perf_session *session = perf_session__new(NULL, O_WRONLY, false, false);
1294 if (session == NULL)
1297 if (target_tid != -1)
1298 event__synthesize_thread(target_tid, event__process, session);
1300 event__synthesize_threads(event__process, session);
1302 for (i = 0; i < nr_cpus; i++) {
1304 for (counter = 0; counter < nr_counters; counter++)
1305 start_counter(i, counter);
1308 /* Wait for a minimal set of events before starting the snapshot */
1309 poll(&event_array[0], nr_poll, 100);
1311 perf_session__mmap_read(session);
1313 if (pthread_create(&thread, NULL, display_thread, session)) {
1314 printf("Could not create display thread.\n");
1318 if (realtime_prio) {
1319 struct sched_param param;
1321 param.sched_priority = realtime_prio;
1322 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1323 printf("Could not set realtime priority.\n");
1331 perf_session__mmap_read(session);
1333 if (hits == samples)
1334 ret = poll(event_array, nr_poll, 100);
1340 static const char * const top_usage[] = {
1341 "perf top [<options>]",
1345 static const struct option options[] = {
1346 OPT_CALLBACK('e', "event", NULL, "event",
1347 "event selector. use 'perf list' to list available events",
1349 OPT_INTEGER('c', "count", &default_interval,
1350 "event period to sample"),
1351 OPT_INTEGER('p', "pid", &target_pid,
1352 "profile events on existing process id"),
1353 OPT_INTEGER('t', "tid", &target_tid,
1354 "profile events on existing thread id"),
1355 OPT_BOOLEAN('a', "all-cpus", &system_wide,
1356 "system-wide collection from all CPUs"),
1357 OPT_STRING('C', "cpu", &cpu_list, "cpu",
1358 "list of cpus to monitor"),
1359 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1360 "file", "vmlinux pathname"),
1361 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols,
1362 "hide kernel symbols"),
1363 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
1364 OPT_INTEGER('r', "realtime", &realtime_prio,
1365 "collect data with this RT SCHED_FIFO priority"),
1366 OPT_INTEGER('d', "delay", &delay_secs,
1367 "number of seconds to delay between refreshes"),
1368 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
1369 "dump the symbol table used for profiling"),
1370 OPT_INTEGER('f', "count-filter", &count_filter,
1371 "only display functions with more events than this"),
1372 OPT_BOOLEAN('g', "group", &group,
1373 "put the counters into a counter group"),
1374 OPT_BOOLEAN('i', "inherit", &inherit,
1375 "child tasks inherit counters"),
1376 OPT_STRING('s', "sym-annotate", &sym_filter, "symbol name",
1377 "symbol to annotate"),
1378 OPT_BOOLEAN('z', "zero", &zero,
1379 "zero history across updates"),
1380 OPT_INTEGER('F', "freq", &freq,
1381 "profile at this frequency"),
1382 OPT_INTEGER('E', "entries", &print_entries,
1383 "display this many functions"),
1384 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols,
1385 "hide user symbols"),
1386 OPT_INCR('v', "verbose", &verbose,
1387 "be more verbose (show counter open errors, etc)"),
1391 int cmd_top(int argc, const char **argv, const char *prefix __used)
1396 page_size = sysconf(_SC_PAGE_SIZE);
1398 argc = parse_options(argc, argv, options, top_usage, 0);
1400 usage_with_options(top_usage, options);
1402 if (target_pid != -1) {
1403 target_tid = target_pid;
1404 thread_num = find_all_tid(target_pid, &all_tids);
1405 if (thread_num <= 0) {
1406 fprintf(stderr, "Can't find all threads of pid %d\n",
1408 usage_with_options(top_usage, options);
1411 all_tids=malloc(sizeof(pid_t));
1415 all_tids[0] = target_tid;
1419 for (i = 0; i < MAX_NR_CPUS; i++) {
1420 for (j = 0; j < MAX_COUNTERS; j++) {
1421 fd[i][j] = malloc(sizeof(int)*thread_num);
1422 mmap_array[i][j] = zalloc(
1423 sizeof(struct mmap_data)*thread_num);
1424 if (!fd[i][j] || !mmap_array[i][j])
1428 event_array = malloc(
1429 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
1433 /* CPU and PID are mutually exclusive */
1434 if (target_tid > 0 && cpu_list) {
1435 printf("WARNING: PID switch overriding CPU\n");
1443 symbol_conf.priv_size = (sizeof(struct sym_entry) +
1444 (nr_counters + 1) * sizeof(unsigned long));
1446 symbol_conf.try_vmlinux_path = (symbol_conf.vmlinux_name == NULL);
1447 if (symbol__init() < 0)
1454 * User specified count overrides default frequency.
1456 if (default_interval)
1459 default_interval = freq;
1461 fprintf(stderr, "frequency and count are zero, aborting\n");
1466 * Fill in the ones not specifically initialized via -c:
1468 for (counter = 0; counter < nr_counters; counter++) {
1469 if (attrs[counter].sample_period)
1472 attrs[counter].sample_period = default_interval;
1475 if (target_tid != -1)
1478 nr_cpus = read_cpu_map(cpu_list);
1481 usage_with_options(top_usage, options);
1483 get_term_dimensions(&winsize);
1484 if (print_entries == 0) {
1485 update_print_entries(&winsize);
1486 signal(SIGWINCH, sig_winch_handler);