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/symbol.h"
24 #include "util/color.h"
25 #include "util/util.h"
26 #include "util/rbtree.h"
27 #include "util/parse-options.h"
28 #include "util/parse-events.h"
40 #include <sys/syscall.h>
41 #include <sys/ioctl.h>
43 #include <sys/prctl.h>
48 #include <linux/unistd.h>
49 #include <linux/types.h>
51 static int system_wide = 0;
53 static __u64 default_event_id[MAX_COUNTERS] = {
54 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_TASK_CLOCK),
55 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CONTEXT_SWITCHES),
56 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_CPU_MIGRATIONS),
57 EID(PERF_TYPE_SOFTWARE, PERF_COUNT_PAGE_FAULTS),
59 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CPU_CYCLES),
60 EID(PERF_TYPE_HARDWARE, PERF_COUNT_INSTRUCTIONS),
61 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_REFERENCES),
62 EID(PERF_TYPE_HARDWARE, PERF_COUNT_CACHE_MISSES),
64 static int default_interval = 100000;
65 static int event_count[MAX_COUNTERS];
66 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
68 static __u64 count_filter = 5;
69 static int print_entries = 15;
71 static int target_pid = -1;
72 static int profile_cpu = -1;
73 static int nr_cpus = 0;
74 static unsigned int realtime_prio = 0;
76 static unsigned int page_size;
77 static unsigned int mmap_pages = 16;
80 static char *sym_filter;
81 static unsigned long filter_start;
82 static unsigned long filter_end;
84 static int delay_secs = 2;
86 static int dump_symtab;
88 static const unsigned int default_count[] = {
101 static uint64_t min_ip;
102 static uint64_t max_ip = -1ll;
105 struct rb_node rb_node;
106 struct list_head node;
107 unsigned long count[MAX_COUNTERS];
108 unsigned long snap_count;
113 struct sym_entry *sym_filter_entry;
115 struct dso *kernel_dso;
118 * Symbols will be added here in record_ip and will get out
121 static LIST_HEAD(active_symbols);
122 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
125 * Ordering weight: count-1 * count-2 * ... / count-n
127 static double sym_weight(const struct sym_entry *sym)
129 double weight = sym->snap_count;
132 for (counter = 1; counter < nr_counters-1; counter++)
133 weight *= sym->count[counter];
135 weight /= (sym->count[counter] + 1);
141 static long userspace_samples;
142 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
144 static void __list_insert_active_sym(struct sym_entry *syme)
146 list_add(&syme->node, &active_symbols);
149 static void list_remove_active_sym(struct sym_entry *syme)
151 pthread_mutex_lock(&active_symbols_lock);
152 list_del_init(&syme->node);
153 pthread_mutex_unlock(&active_symbols_lock);
156 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
158 struct rb_node **p = &tree->rb_node;
159 struct rb_node *parent = NULL;
160 struct sym_entry *iter;
164 iter = rb_entry(parent, struct sym_entry, rb_node);
166 if (se->weight > iter->weight)
172 rb_link_node(&se->rb_node, parent, p);
173 rb_insert_color(&se->rb_node, tree);
176 static void print_sym_table(void)
180 float samples_per_sec = samples/delay_secs;
181 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
182 float sum_ksamples = 0.0;
183 struct sym_entry *syme, *n;
184 struct rb_root tmp = RB_ROOT;
187 samples = userspace_samples = 0;
189 /* Sort the active symbols */
190 pthread_mutex_lock(&active_symbols_lock);
191 syme = list_entry(active_symbols.next, struct sym_entry, node);
192 pthread_mutex_unlock(&active_symbols_lock);
194 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
195 syme->snap_count = syme->count[0];
196 if (syme->snap_count != 0) {
197 syme->weight = sym_weight(syme);
198 rb_insert_active_sym(&tmp, syme);
199 sum_ksamples += syme->snap_count;
201 for (j = 0; j < nr_counters; j++)
202 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
204 list_remove_active_sym(syme);
210 "------------------------------------------------------------------------------\n");
211 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
213 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
215 if (nr_counters == 1) {
216 printf("%d", event_count[0]);
223 for (counter = 0; counter < nr_counters; counter++) {
227 printf("%s", event_name(counter));
232 if (target_pid != -1)
233 printf(" (target_pid: %d", target_pid);
237 if (profile_cpu != -1)
238 printf(", cpu: %d)\n", profile_cpu);
240 if (target_pid != -1)
243 printf(", %d CPUs)\n", nr_cpus);
246 printf("------------------------------------------------------------------------------\n\n");
248 if (nr_counters == 1)
249 printf(" samples pcnt");
251 printf(" weight samples pcnt");
253 printf(" RIP kernel function\n"
254 " ______ _______ _____ ________________ _______________\n\n"
257 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
258 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
259 struct symbol *sym = (struct symbol *)(syme + 1);
260 char *color = PERF_COLOR_NORMAL;
263 if (++printed > print_entries || syme->snap_count < count_filter)
266 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
270 * We color high-overhead entries in red, low-overhead
271 * entries in green - and keep the middle ground normal:
274 color = PERF_COLOR_RED;
276 color = PERF_COLOR_GREEN;
278 if (nr_counters == 1)
279 printf("%20.2f - ", syme->weight);
281 printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
283 color_fprintf(stdout, color, "%4.1f%%", pcnt);
284 printf(" - %016llx : %s\n", sym->start, sym->name);
288 static void *display_thread(void *arg)
290 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
291 int delay_msecs = delay_secs * 1000;
293 printf("PerfTop refresh period: %d seconds\n", delay_secs);
297 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
299 printf("key pressed - exiting.\n");
305 static int symbol_filter(struct dso *self, struct symbol *sym)
307 static int filter_match;
308 struct sym_entry *syme;
309 const char *name = sym->name;
311 if (!strcmp(name, "_text") ||
312 !strcmp(name, "_etext") ||
313 !strcmp(name, "_sinittext") ||
314 !strncmp("init_module", name, 11) ||
315 !strncmp("cleanup_module", name, 14) ||
316 strstr(name, "_text_start") ||
317 strstr(name, "_text_end"))
320 syme = dso__sym_priv(self, sym);
321 /* Tag samples to be skipped. */
322 if (!strcmp("default_idle", name) ||
323 !strcmp("cpu_idle", name) ||
324 !strcmp("enter_idle", name) ||
325 !strcmp("exit_idle", name) ||
326 !strcmp("mwait_idle", name))
329 if (filter_match == 1) {
330 filter_end = sym->start;
332 if (filter_end - filter_start > 10000) {
334 "hm, too large filter symbol <%s> - skipping.\n",
336 fprintf(stderr, "symbol filter start: %016lx\n",
338 fprintf(stderr, " end: %016lx\n",
340 filter_end = filter_start = 0;
346 if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
348 filter_start = sym->start;
355 static int parse_symbols(void)
357 struct rb_node *node;
360 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
361 if (kernel_dso == NULL)
364 if (dso__load_kernel(kernel_dso, NULL, symbol_filter, 1) != 0)
367 node = rb_first(&kernel_dso->syms);
368 sym = rb_entry(node, struct symbol, rb_node);
371 node = rb_last(&kernel_dso->syms);
372 sym = rb_entry(node, struct symbol, rb_node);
376 dso__fprintf(kernel_dso, stderr);
381 dso__delete(kernel_dso);
386 #define TRACE_COUNT 3
389 * Binary search in the histogram table and record the hit:
391 static void record_ip(uint64_t ip, int counter)
393 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
396 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
399 syme->count[counter]++;
400 pthread_mutex_lock(&active_symbols_lock);
401 if (list_empty(&syme->node) || !syme->node.next)
402 __list_insert_active_sym(syme);
403 pthread_mutex_unlock(&active_symbols_lock);
411 static void process_event(uint64_t ip, int counter)
415 if (ip < min_ip || ip > max_ip) {
420 record_ip(ip, counter);
430 static unsigned int mmap_read_head(struct mmap_data *md)
432 struct perf_counter_mmap_page *pc = md->base;
435 head = pc->data_head;
441 struct timeval last_read, this_read;
443 static void mmap_read(struct mmap_data *md)
445 unsigned int head = mmap_read_head(md);
446 unsigned int old = md->prev;
447 unsigned char *data = md->base + page_size;
450 gettimeofday(&this_read, NULL);
453 * If we're further behind than half the buffer, there's a chance
454 * the writer will bite our tail and mess up the samples under us.
456 * If we somehow ended up ahead of the head, we got messed up.
458 * In either case, truncate and restart at head.
461 if (diff > md->mask / 2 || diff < 0) {
465 timersub(&this_read, &last_read, &iv);
466 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
468 fprintf(stderr, "WARNING: failed to keep up with mmap data."
469 " Last read %lu msecs ago.\n", msecs);
472 * head points to a known good entry, start there.
477 last_read = this_read;
479 for (; old != head;) {
481 struct perf_event_header header;
483 __u32 pid, target_pid;
486 struct perf_event_header header;
487 __u32 pid, target_pid;
491 char filename[PATH_MAX];
494 typedef union event_union {
495 struct perf_event_header header;
497 struct mmap_event mmap;
500 event_t *event = (event_t *)&data[old & md->mask];
504 size_t size = event->header.size;
507 * Event straddles the mmap boundary -- header should always
508 * be inside due to u64 alignment of output.
510 if ((old & md->mask) + size != ((old + size) & md->mask)) {
511 unsigned int offset = old;
512 unsigned int len = min(sizeof(*event), size), cpy;
513 void *dst = &event_copy;
516 cpy = min(md->mask + 1 - (offset & md->mask), len);
517 memcpy(dst, &data[offset & md->mask], cpy);
528 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
529 if (event->header.type & PERF_SAMPLE_IP)
530 process_event(event->ip.ip, md->counter);
537 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
538 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
540 static int __cmd_top(void)
542 struct perf_counter_attr attr;
544 int i, counter, group_fd, nr_poll = 0;
548 for (i = 0; i < nr_cpus; i++) {
550 for (counter = 0; counter < nr_counters; counter++) {
553 if (target_pid == -1 && profile_cpu == -1)
556 memset(&attr, 0, sizeof(attr));
557 attr.config = event_id[counter];
558 attr.sample_period = event_count[counter];
559 attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
562 fd[i][counter] = sys_perf_counter_open(&attr, target_pid, cpu, group_fd, 0);
563 if (fd[i][counter] < 0) {
566 error("syscall returned with %d (%s)\n",
567 fd[i][counter], strerror(err));
569 printf("Are you root?\n");
572 assert(fd[i][counter] >= 0);
573 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
576 * First counter acts as the group leader:
578 if (group && group_fd == -1)
579 group_fd = fd[i][counter];
581 event_array[nr_poll].fd = fd[i][counter];
582 event_array[nr_poll].events = POLLIN;
585 mmap_array[i][counter].counter = counter;
586 mmap_array[i][counter].prev = 0;
587 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
588 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
589 PROT_READ, MAP_SHARED, fd[i][counter], 0);
590 if (mmap_array[i][counter].base == MAP_FAILED)
591 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
595 if (pthread_create(&thread, NULL, display_thread, NULL)) {
596 printf("Could not create display thread.\n");
601 struct sched_param param;
603 param.sched_priority = realtime_prio;
604 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
605 printf("Could not set realtime priority.\n");
613 for (i = 0; i < nr_cpus; i++) {
614 for (counter = 0; counter < nr_counters; counter++)
615 mmap_read(&mmap_array[i][counter]);
619 ret = poll(event_array, nr_poll, 100);
625 static const char * const top_usage[] = {
626 "perf top [<options>]",
630 static char events_help_msg[EVENTS_HELP_MAX];
632 static const struct option options[] = {
633 OPT_CALLBACK('e', "event", NULL, "event",
634 events_help_msg, parse_events),
635 OPT_INTEGER('c', "count", &default_interval,
636 "event period to sample"),
637 OPT_INTEGER('p', "pid", &target_pid,
638 "profile events on existing pid"),
639 OPT_BOOLEAN('a', "all-cpus", &system_wide,
640 "system-wide collection from all CPUs"),
641 OPT_INTEGER('C', "CPU", &profile_cpu,
642 "CPU to profile on"),
643 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
644 "number of mmap data pages"),
645 OPT_INTEGER('r', "realtime", &realtime_prio,
646 "collect data with this RT SCHED_FIFO priority"),
647 OPT_INTEGER('d', "delay", &delay_secs,
648 "number of seconds to delay between refreshes"),
649 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
650 "dump the symbol table used for profiling"),
651 OPT_INTEGER('f', "count-filter", &count_filter,
652 "only display functions with more events than this"),
653 OPT_BOOLEAN('g', "group", &group,
654 "put the counters into a counter group"),
655 OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
656 "only display symbols matchig this pattern"),
657 OPT_BOOLEAN('z', "zero", &group,
658 "zero history across updates"),
659 OPT_INTEGER('F', "freq", &freq,
660 "profile at this frequency"),
661 OPT_INTEGER('E', "entries", &print_entries,
662 "display this many functions"),
666 int cmd_top(int argc, const char **argv, const char *prefix)
670 page_size = sysconf(_SC_PAGE_SIZE);
672 create_events_help(events_help_msg);
673 memcpy(event_id, default_event_id, sizeof(default_event_id));
675 argc = parse_options(argc, argv, options, top_usage, 0);
677 usage_with_options(top_usage, options);
680 default_interval = freq;
684 /* CPU and PID are mutually exclusive */
685 if (target_pid != -1 && profile_cpu != -1) {
686 printf("WARNING: PID switch overriding CPU\n");
696 for (counter = 0; counter < nr_counters; counter++) {
697 if (event_count[counter])
700 event_count[counter] = default_interval;
703 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
704 assert(nr_cpus <= MAX_NR_CPUS);
705 assert(nr_cpus >= 0);
707 if (target_pid != -1 || profile_cpu != -1)