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 <linux/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 fd[MAX_NR_CPUS][MAX_COUNTERS];
53 static int system_wide = 0;
55 static int default_interval = 100000;
57 static u64 count_filter = 5;
58 static int print_entries = 15;
60 static int target_pid = -1;
61 static int inherit = 0;
62 static int profile_cpu = -1;
63 static int nr_cpus = 0;
64 static unsigned int realtime_prio = 0;
66 static unsigned int page_size;
67 static unsigned int mmap_pages = 16;
69 static int verbose = 0;
70 static char *vmlinux = NULL;
72 static char *sym_filter;
73 static unsigned long filter_start;
74 static unsigned long filter_end;
76 static int delay_secs = 2;
78 static int dump_symtab;
85 static u64 max_ip = -1ll;
88 struct rb_node rb_node;
89 struct list_head node;
90 unsigned long count[MAX_COUNTERS];
91 unsigned long snap_count;
96 struct sym_entry *sym_filter_entry;
98 struct dso *kernel_dso;
101 * Symbols will be added here in record_ip and will get out
104 static LIST_HEAD(active_symbols);
105 static pthread_mutex_t active_symbols_lock = PTHREAD_MUTEX_INITIALIZER;
108 * Ordering weight: count-1 * count-2 * ... / count-n
110 static double sym_weight(const struct sym_entry *sym)
112 double weight = sym->snap_count;
115 for (counter = 1; counter < nr_counters-1; counter++)
116 weight *= sym->count[counter];
118 weight /= (sym->count[counter] + 1);
124 static long userspace_samples;
125 static const char CONSOLE_CLEAR[] = "
\e[H
\e[2J";
127 static void __list_insert_active_sym(struct sym_entry *syme)
129 list_add(&syme->node, &active_symbols);
132 static void list_remove_active_sym(struct sym_entry *syme)
134 pthread_mutex_lock(&active_symbols_lock);
135 list_del_init(&syme->node);
136 pthread_mutex_unlock(&active_symbols_lock);
139 static void rb_insert_active_sym(struct rb_root *tree, struct sym_entry *se)
141 struct rb_node **p = &tree->rb_node;
142 struct rb_node *parent = NULL;
143 struct sym_entry *iter;
147 iter = rb_entry(parent, struct sym_entry, rb_node);
149 if (se->weight > iter->weight)
155 rb_link_node(&se->rb_node, parent, p);
156 rb_insert_color(&se->rb_node, tree);
159 static void print_sym_table(void)
163 float samples_per_sec = samples/delay_secs;
164 float ksamples_per_sec = (samples-userspace_samples)/delay_secs;
165 float sum_ksamples = 0.0;
166 struct sym_entry *syme, *n;
167 struct rb_root tmp = RB_ROOT;
170 samples = userspace_samples = 0;
172 /* Sort the active symbols */
173 pthread_mutex_lock(&active_symbols_lock);
174 syme = list_entry(active_symbols.next, struct sym_entry, node);
175 pthread_mutex_unlock(&active_symbols_lock);
177 list_for_each_entry_safe_from(syme, n, &active_symbols, node) {
178 syme->snap_count = syme->count[0];
179 if (syme->snap_count != 0) {
180 syme->weight = sym_weight(syme);
181 rb_insert_active_sym(&tmp, syme);
182 sum_ksamples += syme->snap_count;
184 for (j = 0; j < nr_counters; j++)
185 syme->count[j] = zero ? 0 : syme->count[j] * 7 / 8;
187 list_remove_active_sym(syme);
193 "------------------------------------------------------------------------------\n");
194 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
196 100.0 - (100.0*((samples_per_sec-ksamples_per_sec)/samples_per_sec)));
198 if (nr_counters == 1) {
199 printf("%Ld", (u64)attrs[0].sample_period);
206 for (counter = 0; counter < nr_counters; counter++) {
210 printf("%s", event_name(counter));
215 if (target_pid != -1)
216 printf(" (target_pid: %d", target_pid);
220 if (profile_cpu != -1)
221 printf(", cpu: %d)\n", profile_cpu);
223 if (target_pid != -1)
226 printf(", %d CPUs)\n", nr_cpus);
229 printf("------------------------------------------------------------------------------\n\n");
231 if (nr_counters == 1)
232 printf(" samples pcnt");
234 printf(" weight samples pcnt");
236 printf(" RIP kernel function\n"
237 " ______ _______ _____ ________________ _______________\n\n"
240 for (nd = rb_first(&tmp); nd; nd = rb_next(nd)) {
241 struct sym_entry *syme = rb_entry(nd, struct sym_entry, rb_node);
242 struct symbol *sym = (struct symbol *)(syme + 1);
245 if (++printed > print_entries || syme->snap_count < count_filter)
248 pcnt = 100.0 - (100.0 * ((sum_ksamples - syme->snap_count) /
251 if (nr_counters == 1)
252 printf("%20.2f - ", syme->weight);
254 printf("%9.1f %10ld - ", syme->weight, syme->snap_count);
256 percent_color_fprintf(stdout, "%4.1f%%", pcnt);
257 printf(" - %016llx : %s", sym->start, sym->name);
259 printf("\t[%s]", sym->module->name);
264 static void *display_thread(void *arg __used)
266 struct pollfd stdin_poll = { .fd = 0, .events = POLLIN };
267 int delay_msecs = delay_secs * 1000;
269 printf("PerfTop refresh period: %d seconds\n", delay_secs);
273 } while (!poll(&stdin_poll, 1, delay_msecs) == 1);
275 printf("key pressed - exiting.\n");
281 /* Tag samples to be skipped. */
282 static const char *skip_symbols[] = {
288 "mwait_idle_with_hints",
289 "ppc64_runlatch_off",
290 "pseries_dedicated_idle_sleep",
294 static int symbol_filter(struct dso *self, struct symbol *sym)
296 static int filter_match;
297 struct sym_entry *syme;
298 const char *name = sym->name;
302 * ppc64 uses function descriptors and appends a '.' to the
303 * start of every instruction address. Remove it.
308 if (!strcmp(name, "_text") ||
309 !strcmp(name, "_etext") ||
310 !strcmp(name, "_sinittext") ||
311 !strncmp("init_module", name, 11) ||
312 !strncmp("cleanup_module", name, 14) ||
313 strstr(name, "_text_start") ||
314 strstr(name, "_text_end"))
317 syme = dso__sym_priv(self, sym);
318 for (i = 0; skip_symbols[i]; i++) {
319 if (!strcmp(skip_symbols[i], name)) {
325 if (filter_match == 1) {
326 filter_end = sym->start;
328 if (filter_end - filter_start > 10000) {
330 "hm, too large filter symbol <%s> - skipping.\n",
332 fprintf(stderr, "symbol filter start: %016lx\n",
334 fprintf(stderr, " end: %016lx\n",
336 filter_end = filter_start = 0;
342 if (filter_match == 0 && sym_filter && !strcmp(name, sym_filter)) {
344 filter_start = sym->start;
351 static int parse_symbols(void)
353 struct rb_node *node;
355 int modules = vmlinux ? 1 : 0;
357 kernel_dso = dso__new("[kernel]", sizeof(struct sym_entry));
358 if (kernel_dso == NULL)
361 if (dso__load_kernel(kernel_dso, vmlinux, symbol_filter, verbose, modules) <= 0)
364 node = rb_first(&kernel_dso->syms);
365 sym = rb_entry(node, struct symbol, rb_node);
368 node = rb_last(&kernel_dso->syms);
369 sym = rb_entry(node, struct symbol, rb_node);
373 dso__fprintf(kernel_dso, stderr);
378 dso__delete(kernel_dso);
383 #define TRACE_COUNT 3
386 * Binary search in the histogram table and record the hit:
388 static void record_ip(u64 ip, int counter)
390 struct symbol *sym = dso__find_symbol(kernel_dso, ip);
393 struct sym_entry *syme = dso__sym_priv(kernel_dso, sym);
396 syme->count[counter]++;
397 pthread_mutex_lock(&active_symbols_lock);
398 if (list_empty(&syme->node) || !syme->node.next)
399 __list_insert_active_sym(syme);
400 pthread_mutex_unlock(&active_symbols_lock);
408 static void process_event(u64 ip, int counter, int user)
417 record_ip(ip, counter);
427 static unsigned int mmap_read_head(struct mmap_data *md)
429 struct perf_counter_mmap_page *pc = md->base;
432 head = pc->data_head;
438 struct timeval last_read, this_read;
440 static void mmap_read_counter(struct mmap_data *md)
442 unsigned int head = mmap_read_head(md);
443 unsigned int old = md->prev;
444 unsigned char *data = md->base + page_size;
447 gettimeofday(&this_read, NULL);
450 * If we're further behind than half the buffer, there's a chance
451 * the writer will bite our tail and mess up the samples under us.
453 * If we somehow ended up ahead of the head, we got messed up.
455 * In either case, truncate and restart at head.
458 if (diff > md->mask / 2 || diff < 0) {
462 timersub(&this_read, &last_read, &iv);
463 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
465 fprintf(stderr, "WARNING: failed to keep up with mmap data."
466 " Last read %lu msecs ago.\n", msecs);
469 * head points to a known good entry, start there.
474 last_read = this_read;
476 for (; old != head;) {
478 struct perf_event_header header;
483 struct perf_event_header header;
488 char filename[PATH_MAX];
491 typedef union event_union {
492 struct perf_event_header header;
494 struct mmap_event mmap;
497 event_t *event = (event_t *)&data[old & md->mask];
501 size_t size = event->header.size;
504 * Event straddles the mmap boundary -- header should always
505 * be inside due to u64 alignment of output.
507 if ((old & md->mask) + size != ((old + size) & md->mask)) {
508 unsigned int offset = old;
509 unsigned int len = min(sizeof(*event), size), cpy;
510 void *dst = &event_copy;
513 cpy = min(md->mask + 1 - (offset & md->mask), len);
514 memcpy(dst, &data[offset & md->mask], cpy);
525 if (event->header.type == PERF_EVENT_SAMPLE) {
527 (event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK) == PERF_EVENT_MISC_USER;
528 process_event(event->ip.ip, md->counter, user);
535 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
536 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
538 static void mmap_read(void)
542 for (i = 0; i < nr_cpus; i++) {
543 for (counter = 0; counter < nr_counters; counter++)
544 mmap_read_counter(&mmap_array[i][counter]);
551 static void start_counter(int i, int counter)
553 struct perf_counter_attr *attr;
557 if (target_pid == -1 && profile_cpu == -1)
560 attr = attrs + counter;
562 attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
564 attr->inherit = (cpu < 0) && inherit;
567 fd[i][counter] = sys_perf_counter_open(attr, target_pid, cpu, group_fd, 0);
569 if (fd[i][counter] < 0) {
573 die("No permission - are you root?\n");
575 * If it's cycles then fall back to hrtimer
576 * based cpu-clock-tick sw counter, which
577 * is always available even if no PMU support:
579 if (attr->type == PERF_TYPE_HARDWARE
580 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
583 warning(" ... trying to fall back to cpu-clock-ticks\n");
585 attr->type = PERF_TYPE_SOFTWARE;
586 attr->config = PERF_COUNT_SW_CPU_CLOCK;
590 error("perfcounter syscall returned with %d (%s)\n",
591 fd[i][counter], strerror(err));
592 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
595 assert(fd[i][counter] >= 0);
596 fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
599 * First counter acts as the group leader:
601 if (group && group_fd == -1)
602 group_fd = fd[i][counter];
604 event_array[nr_poll].fd = fd[i][counter];
605 event_array[nr_poll].events = POLLIN;
608 mmap_array[i][counter].counter = counter;
609 mmap_array[i][counter].prev = 0;
610 mmap_array[i][counter].mask = mmap_pages*page_size - 1;
611 mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
612 PROT_READ, MAP_SHARED, fd[i][counter], 0);
613 if (mmap_array[i][counter].base == MAP_FAILED)
614 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
617 static int __cmd_top(void)
623 for (i = 0; i < nr_cpus; i++) {
625 for (counter = 0; counter < nr_counters; counter++)
626 start_counter(i, counter);
629 /* Wait for a minimal set of events before starting the snapshot */
630 poll(event_array, nr_poll, 100);
634 if (pthread_create(&thread, NULL, display_thread, NULL)) {
635 printf("Could not create display thread.\n");
640 struct sched_param param;
642 param.sched_priority = realtime_prio;
643 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
644 printf("Could not set realtime priority.\n");
655 ret = poll(event_array, nr_poll, 100);
661 static const char * const top_usage[] = {
662 "perf top [<options>]",
666 static const struct option options[] = {
667 OPT_CALLBACK('e', "event", NULL, "event",
668 "event selector. use 'perf list' to list available events",
670 OPT_INTEGER('c', "count", &default_interval,
671 "event period to sample"),
672 OPT_INTEGER('p', "pid", &target_pid,
673 "profile events on existing pid"),
674 OPT_BOOLEAN('a', "all-cpus", &system_wide,
675 "system-wide collection from all CPUs"),
676 OPT_INTEGER('C', "CPU", &profile_cpu,
677 "CPU to profile on"),
678 OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
679 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
680 "number of mmap data pages"),
681 OPT_INTEGER('r', "realtime", &realtime_prio,
682 "collect data with this RT SCHED_FIFO priority"),
683 OPT_INTEGER('d', "delay", &delay_secs,
684 "number of seconds to delay between refreshes"),
685 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab,
686 "dump the symbol table used for profiling"),
687 OPT_INTEGER('f', "count-filter", &count_filter,
688 "only display functions with more events than this"),
689 OPT_BOOLEAN('g', "group", &group,
690 "put the counters into a counter group"),
691 OPT_BOOLEAN('i', "inherit", &inherit,
692 "child tasks inherit counters"),
693 OPT_STRING('s', "sym-filter", &sym_filter, "pattern",
694 "only display symbols matchig this pattern"),
695 OPT_BOOLEAN('z', "zero", &zero,
696 "zero history across updates"),
697 OPT_INTEGER('F', "freq", &freq,
698 "profile at this frequency"),
699 OPT_INTEGER('E', "entries", &print_entries,
700 "display this many functions"),
701 OPT_BOOLEAN('v', "verbose", &verbose,
702 "be more verbose (show counter open errors, etc)"),
706 int cmd_top(int argc, const char **argv, const char *prefix __used)
712 page_size = sysconf(_SC_PAGE_SIZE);
714 argc = parse_options(argc, argv, options, top_usage, 0);
716 usage_with_options(top_usage, options);
719 default_interval = freq;
723 /* CPU and PID are mutually exclusive */
724 if (target_pid != -1 && profile_cpu != -1) {
725 printf("WARNING: PID switch overriding CPU\n");
739 * Fill in the ones not specifically initialized via -c:
741 for (counter = 0; counter < nr_counters; counter++) {
742 if (attrs[counter].sample_period)
745 attrs[counter].sample_period = default_interval;
748 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
749 assert(nr_cpus <= MAX_NR_CPUS);
750 assert(nr_cpus >= 0);
752 if (target_pid != -1 || profile_cpu != -1)