2 * kerneltop.c: show top kernel functions - performance counters showcase
6 make -C Documentation/perf_counter/
10 ------------------------------------------------------------------------------
11 KernelTop: 2669 irqs/sec [cache-misses/cache-refs], (all, cpu: 2)
12 ------------------------------------------------------------------------------
14 weight RIP kernel function
15 ______ ________________ _______________
17 35.20 - ffffffff804ce74b : skb_copy_and_csum_dev
18 33.00 - ffffffff804cb740 : sock_alloc_send_skb
19 31.26 - ffffffff804ce808 : skb_push
20 22.43 - ffffffff80510004 : tcp_established_options
21 19.00 - ffffffff8027d250 : find_get_page
22 15.76 - ffffffff804e4fc9 : eth_type_trans
23 15.20 - ffffffff804d8baa : dst_release
24 14.86 - ffffffff804cf5d8 : skb_release_head_state
25 14.00 - ffffffff802217d5 : read_hpet
26 12.00 - ffffffff804ffb7f : __ip_local_out
27 11.97 - ffffffff804fc0c8 : ip_local_deliver_finish
28 8.54 - ffffffff805001a3 : ip_queue_xmit
32 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
34 * Improvements and fixes by:
36 * Arjan van de Ven <arjan@linux.intel.com>
37 * Yanmin Zhang <yanmin.zhang@intel.com>
38 * Wu Fengguang <fengguang.wu@intel.com>
39 * Mike Galbraith <efault@gmx.de>
40 * Paul Mackerras <paulus@samba.org>
42 * Released under the GPL v2. (and only v2, not any later version)
47 #include "util/symbol.h"
48 #include "util/util.h"
49 #include "util/rbtree.h"
50 #include "util/parse-options.h"
51 #include "util/parse-events.h"
63 #include <sys/syscall.h>
64 #include <sys/ioctl.h>
66 #include <sys/prctl.h>
71 #include <linux/unistd.h>
72 #include <linux/types.h>
74 static int system_wide = 0;
76 static __u64 default_event_id[MAX_COUNTERS] = {
77 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK),
78 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES),
79 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS),
80 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS),
82 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES),
83 EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS),
84 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES),
85 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES),
87 static int default_interval = 100000;
88 static int event_count[MAX_COUNTERS];
89 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
91 static __u64 count_filter = 100;
93 static int target_pid = -1;
94 static int profile_cpu = -1;
95 static int nr_cpus = 0;
96 static unsigned int realtime_prio = 0;
98 static unsigned int page_size;
99 static unsigned int mmap_pages = 16;
100 static int use_mmap = 0;
101 static int use_munmap = 0;
104 static char *sym_filter;
105 static unsigned long filter_start;
106 static unsigned long filter_end;
108 static int delay_secs = 2;
110 static int dump_symtab;
112 static const unsigned int default_count[] = {
125 static uint64_t min_ip;
126 static uint64_t max_ip = -1ll;
129 struct rb_node rb_node;
130 struct list_head node;
131 unsigned long count[MAX_COUNTERS];
135 struct sym_entry *sym_filter_entry;
137 struct dso *kernel_dso;
140 * Symbols will be added here in record_ip and will get out
143 static LIST_HEAD(active_symbols);
146 * Ordering weight: count-1 * count-2 * ... / count-n
148 static double sym_weight(const struct sym_entry *sym)
153 weight = sym->count[0];
155 for (counter = 1; counter < nr_counters-1; counter++)
156 weight *= sym->count[counter];
158 weight /= (sym->count[counter] + 1);
164 static long userspace_events;
165 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
167 static void list_insert_active_sym(struct sym_entry *syme)
169 list_add(&syme->node, &active_symbols);
172 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
174 struct rb_node **p = &tree->rb_node;
175 struct rb_node *parent = NULL;
176 struct sym_entry *iter;
180 iter = rb_entry(parent, struct sym_entry, rb_node);
182 if (sym_weight(se) > sym_weight(iter))
188 rb_link_node(&se->rb_node, parent, p);
189 rb_insert_color(&se->rb_node, tree);
192 static void print_sym_table(void)
196 float events_per_sec = events/delay_secs;
197 float kevents_per_sec = (events-userspace_events)/delay_secs;
198 float sum_kevents = 0.0;
199 struct sym_entry *syme, *n;
200 struct rb_root tmp = RB_ROOT;
203 events = userspace_events = 0;
205 /* Sort the active symbols */
206 list_for_each_entry_safe(syme, n, &active_symbols, node) {
207 if (syme->count[0] != 0) {
208 rb_insert_active_sym(&tmp, syme);
209 sum_kevents += syme->count[0];
211 for (j = 0; j < nr_counters; j++)
212 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
214 list_del_init(&syme->node);
217 write(1, CONSOLE_CLEAR, strlen(CONSOLE_CLEAR));
220 "------------------------------------------------------------------------------\n");
221 printf( " KernelTop:%8.0f irqs/sec kernel:%4.1f%% [",
223 100.0 - (100.0*((events_per_sec-kevents_per_sec)/events_per_sec)));
225 if (nr_counters == 1)
226 printf("%d ", event_count[0]);
228 for (counter = 0; counter < nr_counters; counter++) {
232 printf("%s", event_name(counter));
237 if (target_pid != -1)
238 printf(" (target_pid: %d", target_pid);
242 if (profile_cpu != -1)
243 printf(", cpu: %d)\n", profile_cpu);
245 if (target_pid != -1)
248 printf(", %d CPUs)\n", nr_cpus);
251 printf("------------------------------------------------------------------------------\n\n");
253 if (nr_counters == 1)
254 printf(" events pcnt");
256 printf(" weight events pcnt");
258 printf(" RIP kernel function\n"
259 " ______ ______ _____ ________________ _______________\n\n"
262 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
263 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
264 struct symbol *sym = (struct symbol *)(syme + 1);
267 if (++printed > 18 || syme->count[0] < count_filter)
270 pcnt = 100.0 - (100.0 * ((sum_kevents - syme->count[0]) /
273 if (nr_counters == 1)
274 printf("%19.2f - %4.1f%% - %016llx : %s\n",
276 pcnt, sym->start, sym->name);
278 printf("%8.1f %10ld - %4.1f%% - %016llx : %s\n",
279 sym_weight(syme), syme->count[0],
280 pcnt, sym->start, sym->name);
284 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
286 if (poll(&stdin_poll, 1, 0) == 1) {
287 printf("key pressed - exiting.\n");
293 static void *display_thread(void *arg)
295 printf("KernelTop refresh period: %d seconds\n", delay_secs);
297 while (!sleep(delay_secs))
303 static int symbol_filter(struct dso *self, struct symbol *sym)
305 static int filter_match;
306 struct sym_entry *syme;
307 const char *name = sym->name;
309 if (!strcmp(name, "_text") ||
310 !strcmp(name, "_etext") ||
311 !strcmp(name, "_sinittext") ||
312 !strncmp("init_module", name, 11) ||
313 !strncmp("cleanup_module", name, 14) ||
314 strstr(name, "_text_start") ||
315 strstr(name, "_text_end"))
318 syme = dso__sym_priv(self, sym);
319 /* Tag events to be skipped. */
320 if (!strcmp("default_idle", name) ||
321 !strcmp("cpu_idle", name) ||
322 !strcmp("enter_idle", name) ||
323 !strcmp("exit_idle", name) ||
324 !strcmp("mwait_idle", name))
327 if (filter_match == 1) {
328 filter_end = sym->start;
330 if (filter_end - filter_start > 10000) {
332 "hm, too large filter symbol <%s> - skipping.\n",
334 fprintf(stderr, "symbol filter start: %016lx\n",
336 fprintf(stderr, " end: %016lx\n",
338 filter_end = filter_start = 0;
344 if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
346 filter_start = sym->start;
353 static int parse_symbols(void)
355 struct rb_node *node;
358 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
359 if (kernel_dso == NULL)
362 if (dso__load_kernel(kernel_dso, NULL, symbol_filter) != 0)
365 node = rb_first(&kernel_dso->syms);
366 sym = rb_entry(node, struct symbol, rb_node);
369 node = rb_last(&kernel_dso->syms);
370 sym = rb_entry(node, struct symbol, rb_node);
374 dso__fprintf(kernel_dso, stdout);
379 dso__delete(kernel_dso);
384 #define TRACE_COUNT 3
387 * Binary search in the histogram table and record the hit:
389 static void record_ip(uint64_t ip, int counter)
391 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
394 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
397 syme->count[counter]++;
398 if (list_empty(&syme->node) || !syme->node.next)
399 list_insert_active_sym(syme);
407 static void process_event(uint64_t ip, int counter)
411 if (ip < min_ip || ip > max_ip) {
416 record_ip(ip, counter);
426 static unsigned int mmap_read_head(struct mmap_data *md)
428 struct perf_counter_mmap_page *pc = md->base;
431 head = pc->data_head;
437 struct timeval last_read, this_read;
439 static void mmap_read(struct mmap_data *md)
441 unsigned int head = mmap_read_head(md);
442 unsigned int old = md->prev;
443 unsigned char *data = md->base + page_size;
446 gettimeofday(&this_read, NULL);
449 * If we're further behind than half the buffer, there's a chance
450 * the writer will bite our tail and screw up the events under us.
452 * If we somehow ended up ahead of the head, we got messed up.
454 * In either case, truncate and restart at head.
457 if (diff > md->mask / 2 || diff < 0) {
461 timersub(&this_read, &last_read, &iv);
462 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
464 fprintf(stderr, "WARNING: failed to keep up with mmap data."
465 " Last read %lu msecs ago.\n", msecs);
468 * head points to a known good entry, start there.
473 last_read = this_read;
475 for (; old != head;) {
477 struct perf_event_header header;
479 __u32 pid, target_pid;
482 struct perf_event_header header;
483 __u32 pid, target_pid;
487 char filename[PATH_MAX];
490 typedef union event_union {
491 struct perf_event_header header;
493 struct mmap_event mmap;
496 event_t *event = (event_t *)&data[old & md->mask];
500 size_t size = event->header.size;
503 * Event straddles the mmap boundary -- header should always
504 * be inside due to u64 alignment of output.
506 if ((old & md->mask) + size != ((old + size) & md->mask)) {
507 unsigned int offset = old;
508 unsigned int len = min(sizeof(*event), size), cpy;
509 void *dst = &event_copy;
512 cpy = min(md->mask + 1 - (offset & md->mask), len);
513 memcpy(dst, &data[offset & md->mask], cpy);
524 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
525 if (event->header.type & PERF_RECORD_IP)
526 process_event(event->ip.ip, md->counter);
528 switch (event->header.type) {
529 case PERF_EVENT_MMAP:
530 case PERF_EVENT_MUNMAP:
531 printf("%s: %Lu %Lu %Lu %s\n",
532 event->header.type == PERF_EVENT_MMAP
537 event->mmap.filename);
546 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
547 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
549 static int __cmd_top(void)
551 struct perf_counter_hw_event hw_event;
553 int i, counter, group_fd, nr_poll = 0;
557 for (i = 0; i < nr_cpus; i++) {
559 for (counter = 0; counter < nr_counters; counter++) {
562 if (target_pid == -1 && profile_cpu == -1)
565 memset(&hw_event, 0, sizeof(hw_event));
566 hw_event.config = event_id[counter];
567 hw_event.irq_period = event_count[counter];
568 hw_event.record_type = PERF_RECORD_IP | PERF_RECORD_TID;
570 hw_event.mmap = use_mmap;
571 hw_event.munmap = use_munmap;
572 hw_event.freq = freq;
574 fd[i][counter] = sys_perf_counter_open(&hw_event, target_pid, cpu, group_fd, 0);
575 if (fd[i][counter] < 0) {
577 printf("kerneltop error: syscall returned with %d (%s)\n",
578 fd[i][counter], strerror(err));
580 printf("Are you root?\n");
583 assert(fd[i][counter] >= 0);
584 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
587 * First counter acts as the group leader:
589 if (group && group_fd == -1)
590 group_fd = fd[i][counter];
592 event_array[nr_poll].fd = fd[i][counter];
593 event_array[nr_poll].events = POLLIN;
596 mmap_array[i][counter].counter = counter;
597 mmap_array[i][counter].prev = 0;
598 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
599 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
600 PROT_READ, MAP_SHARED, fd[i][counter], 0);
601 if (mmap_array[i][counter].base == MAP_FAILED) {
602 printf("kerneltop error: failed to mmap with %d (%s)\n",
603 errno, strerror(errno));
609 if (pthread_create(&thread, NULL, display_thread, NULL)) {
610 printf("Could not create display thread.\n");
615 struct sched_param param;
617 param.sched_priority = realtime_prio;
618 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
619 printf("Could not set realtime priority.\n");
627 for (i = 0; i < nr_cpus; i++) {
628 for (counter = 0; counter < nr_counters; counter++)
629 mmap_read(&mmap_array[i][counter]);
633 ret = poll(event_array, nr_poll, 100);
639 static const char * const top_usage[] = {
640 "perf top [<options>]",
644 static char events_help_msg[EVENTS_HELP_MAX];
646 static const struct option options[] = {
647 OPT_CALLBACK('e', "event", NULL, "event",
648 events_help_msg, parse_events),
649 OPT_INTEGER('c', "count", &default_interval,
650 "event period to sample"),
651 OPT_INTEGER('p', "pid", &target_pid,
652 "profile events on existing pid"),
653 OPT_BOOLEAN('a', "all-cpus", &system_wide,
654 "system-wide collection from all CPUs"),
655 OPT_INTEGER('C', "CPU", &profile_cpu,
656 "CPU to profile on"),
657 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
658 "number of mmap data pages"),
659 OPT_INTEGER('r', "realtime", &realtime_prio,
660 "collect data with this RT SCHED_FIFO priority"),
661 OPT_INTEGER('d', "delay", &delay_secs,
662 "number of seconds to delay between refreshes"),
663 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
664 "dump the symbol table used for profiling"),
665 OPT_INTEGER('f', "--count-filter", &count_filter,
666 "only display functions with more events than this"),
667 OPT_BOOLEAN('g', "group", &group,
668 "put the counters into a counter group"),
669 OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
670 "only display symbols matchig this pattern"),
671 OPT_BOOLEAN('z', "zero", &group,
672 "zero history across updates"),
673 OPT_BOOLEAN('M', "use-mmap", &use_mmap,
674 "track mmap events"),
675 OPT_BOOLEAN('U', "use-munmap", &use_munmap,
676 "track munmap events"),
677 OPT_INTEGER('F', "--freq", &freq,
678 "profile at this frequency"),
682 int cmd_top(int argc, const char **argv, const char *prefix)
686 page_size = sysconf(_SC_PAGE_SIZE);
688 create_events_help(events_help_msg);
689 memcpy(event_id, default_event_id, sizeof(default_event_id));
691 argc = parse_options(argc, argv, options, top_usage, 0);
693 usage_with_options(top_usage, options);
696 default_interval = freq;
700 /* CPU and PID are mutually exclusive */
701 if (target_pid != -1 && profile_cpu != -1) {
702 printf("WARNING: PID switch overriding CPU\n");
712 for (counter = 0; counter < nr_counters; counter++) {
713 if (event_count[counter])
716 event_count[counter] = default_interval;
719 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
720 assert(nr_cpus <= MAX_NR_CPUS);
721 assert(nr_cpus >= 0);
723 if (target_pid != -1 || profile_cpu != -1)