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_events
;
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 events_per_sec
= events
/delay_secs
;
181 float kevents_per_sec
= (events
-userspace_events
)/delay_secs
;
182 float sum_kevents
= 0.0;
183 struct sym_entry
*syme
, *n
;
184 struct rb_root tmp
= RB_ROOT
;
187 events
= userspace_events
= 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_kevents
+= 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
);
207 write(1, CONSOLE_CLEAR
, strlen(CONSOLE_CLEAR
));
210 "------------------------------------------------------------------------------\n");
211 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
213 100.0 - (100.0*((events_per_sec
-kevents_per_sec
)/events_per_sec
)));
215 if (nr_counters
== 1)
216 printf("%d ", event_count
[0]);
218 for (counter
= 0; counter
< nr_counters
; counter
++) {
222 printf("%s", event_name(counter
));
227 if (target_pid
!= -1)
228 printf(" (target_pid: %d", target_pid
);
232 if (profile_cpu
!= -1)
233 printf(", cpu: %d)\n", profile_cpu
);
235 if (target_pid
!= -1)
238 printf(", %d CPUs)\n", nr_cpus
);
241 printf("------------------------------------------------------------------------------\n\n");
243 if (nr_counters
== 1)
244 printf(" events pcnt");
246 printf(" weight events pcnt");
248 printf(" RIP kernel function\n"
249 " ______ ______ _____ ________________ _______________\n\n"
252 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
253 struct sym_entry
*syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
254 struct symbol
*sym
= (struct symbol
*)(syme
+ 1);
255 char *color
= PERF_COLOR_NORMAL
;
258 if (++printed
> print_entries
|| syme
->snap_count
< count_filter
)
261 pcnt
= 100.0 - (100.0 * ((sum_kevents
- syme
->snap_count
) /
265 * We color high-overhead entries in red, low-overhead
266 * entries in green - and keep the middle ground normal:
269 color
= PERF_COLOR_RED
;
271 color
= PERF_COLOR_GREEN
;
273 if (nr_counters
== 1)
274 printf("%19.2f - ", syme
->weight
);
276 printf("%8.1f %10ld - ", syme
->weight
, syme
->snap_count
);
278 color_fprintf(stdout
, color
, "%4.1f%%", pcnt
);
279 printf(" - %016llx : %s\n", sym
->start
, sym
->name
);
283 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
285 if (poll(&stdin_poll
, 1, 0) == 1) {
286 printf("key pressed - exiting.\n");
292 static void *display_thread(void *arg
)
294 printf("PerfTop refresh period: %d seconds\n", delay_secs
);
296 while (!sleep(delay_secs
))
302 static int symbol_filter(struct dso
*self
, struct symbol
*sym
)
304 static int filter_match
;
305 struct sym_entry
*syme
;
306 const char *name
= sym
->name
;
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 /* Tag events to be skipped. */
319 if (!strcmp("default_idle", name
) ||
320 !strcmp("cpu_idle", name
) ||
321 !strcmp("enter_idle", name
) ||
322 !strcmp("exit_idle", name
) ||
323 !strcmp("mwait_idle", name
))
326 if (filter_match
== 1) {
327 filter_end
= sym
->start
;
329 if (filter_end
- filter_start
> 10000) {
331 "hm, too large filter symbol <%s> - skipping.\n",
333 fprintf(stderr
, "symbol filter start: %016lx\n",
335 fprintf(stderr
, " end: %016lx\n",
337 filter_end
= filter_start
= 0;
343 if (filter_match
== 0 && sym_filter
&& !strcmp(name
, sym_filter
)) {
345 filter_start
= sym
->start
;
352 static int parse_symbols(void)
354 struct rb_node
*node
;
357 kernel_dso
= dso__new("[kernel]", sizeof(struct sym_entry
));
358 if (kernel_dso
== NULL
)
361 if (dso__load_kernel(kernel_dso
, NULL
, symbol_filter
, 1) != 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(uint64_t 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(uint64_t ip
, int counter
)
412 if (ip
< min_ip
|| ip
> max_ip
) {
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(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 screw up the events 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
;
480 __u32 pid
, target_pid
;
483 struct perf_event_header header
;
484 __u32 pid
, target_pid
;
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
.misc
& PERF_EVENT_MISC_OVERFLOW
) {
526 if (event
->header
.type
& PERF_SAMPLE_IP
)
527 process_event(event
->ip
.ip
, md
->counter
);
534 static struct pollfd event_array
[MAX_NR_CPUS
* MAX_COUNTERS
];
535 static struct mmap_data mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
537 static int __cmd_top(void)
539 struct perf_counter_attr attr
;
541 int i
, counter
, group_fd
, nr_poll
= 0;
545 for (i
= 0; i
< nr_cpus
; i
++) {
547 for (counter
= 0; counter
< nr_counters
; counter
++) {
550 if (target_pid
== -1 && profile_cpu
== -1)
553 memset(&attr
, 0, sizeof(attr
));
554 attr
.config
= event_id
[counter
];
555 attr
.sample_period
= event_count
[counter
];
556 attr
.sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
559 fd
[i
][counter
] = sys_perf_counter_open(&attr
, target_pid
, cpu
, group_fd
, 0);
560 if (fd
[i
][counter
] < 0) {
563 error("syscall returned with %d (%s)\n",
564 fd
[i
][counter
], strerror(err
));
566 printf("Are you root?\n");
569 assert(fd
[i
][counter
] >= 0);
570 fcntl(fd
[i
][counter
], F_SETFL
, O_NONBLOCK
);
573 * First counter acts as the group leader:
575 if (group
&& group_fd
== -1)
576 group_fd
= fd
[i
][counter
];
578 event_array
[nr_poll
].fd
= fd
[i
][counter
];
579 event_array
[nr_poll
].events
= POLLIN
;
582 mmap_array
[i
][counter
].counter
= counter
;
583 mmap_array
[i
][counter
].prev
= 0;
584 mmap_array
[i
][counter
].mask
= mmap_pages
*page_size
- 1;
585 mmap_array
[i
][counter
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
586 PROT_READ
, MAP_SHARED
, fd
[i
][counter
], 0);
587 if (mmap_array
[i
][counter
].base
== MAP_FAILED
)
588 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
592 if (pthread_create(&thread
, NULL
, display_thread
, NULL
)) {
593 printf("Could not create display thread.\n");
598 struct sched_param param
;
600 param
.sched_priority
= realtime_prio
;
601 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
602 printf("Could not set realtime priority.\n");
610 for (i
= 0; i
< nr_cpus
; i
++) {
611 for (counter
= 0; counter
< nr_counters
; counter
++)
612 mmap_read(&mmap_array
[i
][counter
]);
616 ret
= poll(event_array
, nr_poll
, 100);
622 static const char * const top_usage
[] = {
623 "perf top [<options>]",
627 static char events_help_msg
[EVENTS_HELP_MAX
];
629 static const struct option options
[] = {
630 OPT_CALLBACK('e', "event", NULL
, "event",
631 events_help_msg
, parse_events
),
632 OPT_INTEGER('c', "count", &default_interval
,
633 "event period to sample"),
634 OPT_INTEGER('p', "pid", &target_pid
,
635 "profile events on existing pid"),
636 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
637 "system-wide collection from all CPUs"),
638 OPT_INTEGER('C', "CPU", &profile_cpu
,
639 "CPU to profile on"),
640 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
641 "number of mmap data pages"),
642 OPT_INTEGER('r', "realtime", &realtime_prio
,
643 "collect data with this RT SCHED_FIFO priority"),
644 OPT_INTEGER('d', "delay", &delay_secs
,
645 "number of seconds to delay between refreshes"),
646 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
647 "dump the symbol table used for profiling"),
648 OPT_INTEGER('f', "count-filter", &count_filter
,
649 "only display functions with more events than this"),
650 OPT_BOOLEAN('g', "group", &group
,
651 "put the counters into a counter group"),
652 OPT_STRING('s', "sym-filter", &sym_filter
, "pattern",
653 "only display symbols matchig this pattern"),
654 OPT_BOOLEAN('z', "zero", &group
,
655 "zero history across updates"),
656 OPT_INTEGER('F', "freq", &freq
,
657 "profile at this frequency"),
658 OPT_INTEGER('E', "entries", &print_entries
,
659 "display this many functions"),
663 int cmd_top(int argc
, const char **argv
, const char *prefix
)
667 page_size
= sysconf(_SC_PAGE_SIZE
);
669 create_events_help(events_help_msg
);
670 memcpy(event_id
, default_event_id
, sizeof(default_event_id
));
672 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
674 usage_with_options(top_usage
, options
);
677 default_interval
= freq
;
681 /* CPU and PID are mutually exclusive */
682 if (target_pid
!= -1 && profile_cpu
!= -1) {
683 printf("WARNING: PID switch overriding CPU\n");
693 for (counter
= 0; counter
< nr_counters
; counter
++) {
694 if (event_count
[counter
])
697 event_count
[counter
] = default_interval
;
700 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
701 assert(nr_cpus
<= MAX_NR_CPUS
);
702 assert(nr_cpus
>= 0);
704 if (target_pid
!= -1 || profile_cpu
!= -1)