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/util.h"
25 #include "util/rbtree.h"
26 #include "util/parse-options.h"
27 #include "util/parse-events.h"
39 #include <sys/syscall.h>
40 #include <sys/ioctl.h>
42 #include <sys/prctl.h>
47 #include <linux/unistd.h>
48 #include <linux/types.h>
50 static int system_wide
= 0;
52 static __u64 default_event_id
[MAX_COUNTERS
] = {
53 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_TASK_CLOCK
),
54 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_CONTEXT_SWITCHES
),
55 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_CPU_MIGRATIONS
),
56 EID(PERF_TYPE_SOFTWARE
, PERF_COUNT_PAGE_FAULTS
),
58 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_CPU_CYCLES
),
59 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_INSTRUCTIONS
),
60 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_CACHE_REFERENCES
),
61 EID(PERF_TYPE_HARDWARE
, PERF_COUNT_CACHE_MISSES
),
63 static int default_interval
= 100000;
64 static int event_count
[MAX_COUNTERS
];
65 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
67 static __u64 count_filter
= 5;
68 static int print_entries
= 15;
70 static int target_pid
= -1;
71 static int profile_cpu
= -1;
72 static int nr_cpus
= 0;
73 static unsigned int realtime_prio
= 0;
75 static unsigned int page_size
;
76 static unsigned int mmap_pages
= 16;
77 static int use_mmap
= 0;
78 static int use_munmap
= 0;
81 static char *sym_filter
;
82 static unsigned long filter_start
;
83 static unsigned long filter_end
;
85 static int delay_secs
= 2;
87 static int dump_symtab
;
89 static const unsigned int default_count
[] = {
102 static uint64_t min_ip
;
103 static uint64_t max_ip
= -1ll;
106 struct rb_node rb_node
;
107 struct list_head node
;
108 unsigned long count
[MAX_COUNTERS
];
109 unsigned long snap_count
;
114 struct sym_entry
*sym_filter_entry
;
116 struct dso
*kernel_dso
;
119 * Symbols will be added here in record_ip and will get out
122 static LIST_HEAD(active_symbols
);
123 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
126 * Ordering weight: count-1 * count-2 * ... / count-n
128 static double sym_weight(const struct sym_entry
*sym
)
130 double weight
= sym
->snap_count
;
133 for (counter
= 1; counter
< nr_counters
-1; counter
++)
134 weight
*= sym
->count
[counter
];
136 weight
/= (sym
->count
[counter
] + 1);
142 static long userspace_events
;
143 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
145 static void __list_insert_active_sym(struct sym_entry
*syme
)
147 list_add(&syme
->node
, &active_symbols
);
150 static void list_remove_active_sym(struct sym_entry
*syme
)
152 pthread_mutex_lock(&active_symbols_lock
);
153 list_del_init(&syme
->node
);
154 pthread_mutex_unlock(&active_symbols_lock
);
157 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
159 struct rb_node
**p
= &tree
->rb_node
;
160 struct rb_node
*parent
= NULL
;
161 struct sym_entry
*iter
;
165 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
167 if (se
->weight
> iter
->weight
)
173 rb_link_node(&se
->rb_node
, parent
, p
);
174 rb_insert_color(&se
->rb_node
, tree
);
177 static void print_sym_table(void)
181 float events_per_sec
= events
/delay_secs
;
182 float kevents_per_sec
= (events
-userspace_events
)/delay_secs
;
183 float sum_kevents
= 0.0;
184 struct sym_entry
*syme
, *n
;
185 struct rb_root tmp
= RB_ROOT
;
188 events
= userspace_events
= 0;
190 /* Sort the active symbols */
191 pthread_mutex_lock(&active_symbols_lock
);
192 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
193 pthread_mutex_unlock(&active_symbols_lock
);
195 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
196 syme
->snap_count
= syme
->count
[0];
197 if (syme
->snap_count
!= 0) {
198 syme
->weight
= sym_weight(syme
);
199 rb_insert_active_sym(&tmp
, syme
);
200 sum_kevents
+= syme
->snap_count
;
202 for (j
= 0; j
< nr_counters
; j
++)
203 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
205 list_remove_active_sym(syme
);
208 write(1, CONSOLE_CLEAR
, strlen(CONSOLE_CLEAR
));
211 "------------------------------------------------------------------------------\n");
212 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
214 100.0 - (100.0*((events_per_sec
-kevents_per_sec
)/events_per_sec
)));
216 if (nr_counters
== 1)
217 printf("%d ", event_count
[0]);
219 for (counter
= 0; counter
< nr_counters
; counter
++) {
223 printf("%s", event_name(counter
));
228 if (target_pid
!= -1)
229 printf(" (target_pid: %d", target_pid
);
233 if (profile_cpu
!= -1)
234 printf(", cpu: %d)\n", profile_cpu
);
236 if (target_pid
!= -1)
239 printf(", %d CPUs)\n", nr_cpus
);
242 printf("------------------------------------------------------------------------------\n\n");
244 if (nr_counters
== 1)
245 printf(" events pcnt");
247 printf(" weight events pcnt");
249 printf(" RIP kernel function\n"
250 " ______ ______ _____ ________________ _______________\n\n"
253 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
254 struct sym_entry
*syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
255 struct symbol
*sym
= (struct symbol
*)(syme
+ 1);
258 if (++printed
> print_entries
|| syme
->snap_count
< count_filter
)
261 pcnt
= 100.0 - (100.0 * ((sum_kevents
- syme
->snap_count
) /
264 if (nr_counters
== 1)
265 printf("%19.2f - %4.1f%% - %016llx : %s\n",
266 syme
->weight
, pcnt
, sym
->start
, sym
->name
);
268 printf("%8.1f %10ld - %4.1f%% - %016llx : %s\n",
269 syme
->weight
, syme
->snap_count
,
270 pcnt
, sym
->start
, sym
->name
);
274 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
276 if (poll(&stdin_poll
, 1, 0) == 1) {
277 printf("key pressed - exiting.\n");
283 static void *display_thread(void *arg
)
285 printf("PerfTop refresh period: %d seconds\n", delay_secs
);
287 while (!sleep(delay_secs
))
293 static int symbol_filter(struct dso
*self
, struct symbol
*sym
)
295 static int filter_match
;
296 struct sym_entry
*syme
;
297 const char *name
= sym
->name
;
299 if (!strcmp(name
, "_text") ||
300 !strcmp(name
, "_etext") ||
301 !strcmp(name
, "_sinittext") ||
302 !strncmp("init_module", name
, 11) ||
303 !strncmp("cleanup_module", name
, 14) ||
304 strstr(name
, "_text_start") ||
305 strstr(name
, "_text_end"))
308 syme
= dso__sym_priv(self
, sym
);
309 /* Tag events to be skipped. */
310 if (!strcmp("default_idle", name
) ||
311 !strcmp("cpu_idle", name
) ||
312 !strcmp("enter_idle", name
) ||
313 !strcmp("exit_idle", name
) ||
314 !strcmp("mwait_idle", name
))
317 if (filter_match
== 1) {
318 filter_end
= sym
->start
;
320 if (filter_end
- filter_start
> 10000) {
322 "hm, too large filter symbol <%s> - skipping.\n",
324 fprintf(stderr
, "symbol filter start: %016lx\n",
326 fprintf(stderr
, " end: %016lx\n",
328 filter_end
= filter_start
= 0;
334 if (filter_match
== 0 && sym_filter
&& !strcmp(name
, sym_filter
)) {
336 filter_start
= sym
->start
;
343 static int parse_symbols(void)
345 struct rb_node
*node
;
348 kernel_dso
= dso__new("[kernel]", sizeof(struct sym_entry
));
349 if (kernel_dso
== NULL
)
352 if (dso__load_kernel(kernel_dso
, NULL
, symbol_filter
, 1) != 0)
355 node
= rb_first(&kernel_dso
->syms
);
356 sym
= rb_entry(node
, struct symbol
, rb_node
);
359 node
= rb_last(&kernel_dso
->syms
);
360 sym
= rb_entry(node
, struct symbol
, rb_node
);
364 dso__fprintf(kernel_dso
, stderr
);
369 dso__delete(kernel_dso
);
374 #define TRACE_COUNT 3
377 * Binary search in the histogram table and record the hit:
379 static void record_ip(uint64_t ip
, int counter
)
381 struct symbol
*sym
= dso__find_symbol(kernel_dso
, ip
);
384 struct sym_entry
*syme
= dso__sym_priv(kernel_dso
, sym
);
387 syme
->count
[counter
]++;
388 pthread_mutex_lock(&active_symbols_lock
);
389 if (list_empty(&syme
->node
) || !syme
->node
.next
)
390 __list_insert_active_sym(syme
);
391 pthread_mutex_unlock(&active_symbols_lock
);
399 static void process_event(uint64_t ip
, int counter
)
403 if (ip
< min_ip
|| ip
> max_ip
) {
408 record_ip(ip
, counter
);
418 static unsigned int mmap_read_head(struct mmap_data
*md
)
420 struct perf_counter_mmap_page
*pc
= md
->base
;
423 head
= pc
->data_head
;
429 struct timeval last_read
, this_read
;
431 static void mmap_read(struct mmap_data
*md
)
433 unsigned int head
= mmap_read_head(md
);
434 unsigned int old
= md
->prev
;
435 unsigned char *data
= md
->base
+ page_size
;
438 gettimeofday(&this_read
, NULL
);
441 * If we're further behind than half the buffer, there's a chance
442 * the writer will bite our tail and screw up the events under us.
444 * If we somehow ended up ahead of the head, we got messed up.
446 * In either case, truncate and restart at head.
449 if (diff
> md
->mask
/ 2 || diff
< 0) {
453 timersub(&this_read
, &last_read
, &iv
);
454 msecs
= iv
.tv_sec
*1000 + iv
.tv_usec
/1000;
456 fprintf(stderr
, "WARNING: failed to keep up with mmap data."
457 " Last read %lu msecs ago.\n", msecs
);
460 * head points to a known good entry, start there.
465 last_read
= this_read
;
467 for (; old
!= head
;) {
469 struct perf_event_header header
;
471 __u32 pid
, target_pid
;
474 struct perf_event_header header
;
475 __u32 pid
, target_pid
;
479 char filename
[PATH_MAX
];
482 typedef union event_union
{
483 struct perf_event_header header
;
485 struct mmap_event mmap
;
488 event_t
*event
= (event_t
*)&data
[old
& md
->mask
];
492 size_t size
= event
->header
.size
;
495 * Event straddles the mmap boundary -- header should always
496 * be inside due to u64 alignment of output.
498 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
499 unsigned int offset
= old
;
500 unsigned int len
= min(sizeof(*event
), size
), cpy
;
501 void *dst
= &event_copy
;
504 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
505 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
516 if (event
->header
.misc
& PERF_EVENT_MISC_OVERFLOW
) {
517 if (event
->header
.type
& PERF_SAMPLE_IP
)
518 process_event(event
->ip
.ip
, md
->counter
);
520 switch (event
->header
.type
) {
521 case PERF_EVENT_MMAP
:
522 case PERF_EVENT_MUNMAP
:
523 printf("%s: %Lu %Lu %Lu %s\n",
524 event
->header
.type
== PERF_EVENT_MMAP
529 event
->mmap
.filename
);
538 static struct pollfd event_array
[MAX_NR_CPUS
* MAX_COUNTERS
];
539 static struct mmap_data mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
541 static int __cmd_top(void)
543 struct perf_counter_attr attr
;
545 int i
, counter
, group_fd
, nr_poll
= 0;
549 for (i
= 0; i
< nr_cpus
; i
++) {
551 for (counter
= 0; counter
< nr_counters
; counter
++) {
554 if (target_pid
== -1 && profile_cpu
== -1)
557 memset(&attr
, 0, sizeof(attr
));
558 attr
.config
= event_id
[counter
];
559 attr
.sample_period
= event_count
[counter
];
560 attr
.sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
561 attr
.mmap
= use_mmap
;
562 attr
.munmap
= use_munmap
;
565 fd
[i
][counter
] = sys_perf_counter_open(&attr
, target_pid
, cpu
, group_fd
, 0);
566 if (fd
[i
][counter
] < 0) {
569 error("syscall returned with %d (%s)\n",
570 fd
[i
][counter
], strerror(err
));
572 printf("Are you root?\n");
575 assert(fd
[i
][counter
] >= 0);
576 fcntl(fd
[i
][counter
], F_SETFL
, O_NONBLOCK
);
579 * First counter acts as the group leader:
581 if (group
&& group_fd
== -1)
582 group_fd
= fd
[i
][counter
];
584 event_array
[nr_poll
].fd
= fd
[i
][counter
];
585 event_array
[nr_poll
].events
= POLLIN
;
588 mmap_array
[i
][counter
].counter
= counter
;
589 mmap_array
[i
][counter
].prev
= 0;
590 mmap_array
[i
][counter
].mask
= mmap_pages
*page_size
- 1;
591 mmap_array
[i
][counter
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
592 PROT_READ
, MAP_SHARED
, fd
[i
][counter
], 0);
593 if (mmap_array
[i
][counter
].base
== MAP_FAILED
)
594 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
598 if (pthread_create(&thread
, NULL
, display_thread
, NULL
)) {
599 printf("Could not create display thread.\n");
604 struct sched_param param
;
606 param
.sched_priority
= realtime_prio
;
607 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
608 printf("Could not set realtime priority.\n");
616 for (i
= 0; i
< nr_cpus
; i
++) {
617 for (counter
= 0; counter
< nr_counters
; counter
++)
618 mmap_read(&mmap_array
[i
][counter
]);
622 ret
= poll(event_array
, nr_poll
, 100);
628 static const char * const top_usage
[] = {
629 "perf top [<options>]",
633 static char events_help_msg
[EVENTS_HELP_MAX
];
635 static const struct option options
[] = {
636 OPT_CALLBACK('e', "event", NULL
, "event",
637 events_help_msg
, parse_events
),
638 OPT_INTEGER('c', "count", &default_interval
,
639 "event period to sample"),
640 OPT_INTEGER('p', "pid", &target_pid
,
641 "profile events on existing pid"),
642 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
643 "system-wide collection from all CPUs"),
644 OPT_INTEGER('C', "CPU", &profile_cpu
,
645 "CPU to profile on"),
646 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
647 "number of mmap data pages"),
648 OPT_INTEGER('r', "realtime", &realtime_prio
,
649 "collect data with this RT SCHED_FIFO priority"),
650 OPT_INTEGER('d', "delay", &delay_secs
,
651 "number of seconds to delay between refreshes"),
652 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
653 "dump the symbol table used for profiling"),
654 OPT_INTEGER('f', "count-filter", &count_filter
,
655 "only display functions with more events than this"),
656 OPT_BOOLEAN('g', "group", &group
,
657 "put the counters into a counter group"),
658 OPT_STRING('s', "sym-filter", &sym_filter
, "pattern",
659 "only display symbols matchig this pattern"),
660 OPT_BOOLEAN('z', "zero", &group
,
661 "zero history across updates"),
662 OPT_BOOLEAN('M', "use-mmap", &use_mmap
,
663 "track mmap events"),
664 OPT_BOOLEAN('U', "use-munmap", &use_munmap
,
665 "track munmap events"),
666 OPT_INTEGER('F', "freq", &freq
,
667 "profile at this frequency"),
668 OPT_INTEGER('E', "entries", &print_entries
,
669 "display this many functions"),
673 int cmd_top(int argc
, const char **argv
, const char *prefix
)
677 page_size
= sysconf(_SC_PAGE_SIZE
);
679 create_events_help(events_help_msg
);
680 memcpy(event_id
, default_event_id
, sizeof(default_event_id
));
682 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
684 usage_with_options(top_usage
, options
);
687 default_interval
= freq
;
691 /* CPU and PID are mutually exclusive */
692 if (target_pid
!= -1 && profile_cpu
!= -1) {
693 printf("WARNING: PID switch overriding CPU\n");
703 for (counter
= 0; counter
< nr_counters
; counter
++) {
704 if (event_count
[counter
])
707 event_count
[counter
] = default_interval
;
710 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
711 assert(nr_cpus
<= MAX_NR_CPUS
);
712 assert(nr_cpus
>= 0);
714 if (target_pid
!= -1 || profile_cpu
!= -1)