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 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 profile_cpu
= -1;
62 static int nr_cpus
= 0;
63 static unsigned int realtime_prio
= 0;
65 static unsigned int page_size
;
66 static unsigned int mmap_pages
= 16;
68 static int verbose
= 0;
70 static char *sym_filter
;
71 static unsigned long filter_start
;
72 static unsigned long filter_end
;
74 static int delay_secs
= 2;
76 static int dump_symtab
;
82 static uint64_t min_ip
;
83 static uint64_t max_ip
= -1ll;
86 struct rb_node rb_node
;
87 struct list_head node
;
88 unsigned long count
[MAX_COUNTERS
];
89 unsigned long snap_count
;
94 struct sym_entry
*sym_filter_entry
;
96 struct dso
*kernel_dso
;
99 * Symbols will be added here in record_ip and will get out
102 static LIST_HEAD(active_symbols
);
103 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
106 * Ordering weight: count-1 * count-2 * ... / count-n
108 static double sym_weight(const struct sym_entry
*sym
)
110 double weight
= sym
->snap_count
;
113 for (counter
= 1; counter
< nr_counters
-1; counter
++)
114 weight
*= sym
->count
[counter
];
116 weight
/= (sym
->count
[counter
] + 1);
122 static long userspace_samples
;
123 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
125 static void __list_insert_active_sym(struct sym_entry
*syme
)
127 list_add(&syme
->node
, &active_symbols
);
130 static void list_remove_active_sym(struct sym_entry
*syme
)
132 pthread_mutex_lock(&active_symbols_lock
);
133 list_del_init(&syme
->node
);
134 pthread_mutex_unlock(&active_symbols_lock
);
137 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
139 struct rb_node
**p
= &tree
->rb_node
;
140 struct rb_node
*parent
= NULL
;
141 struct sym_entry
*iter
;
145 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
147 if (se
->weight
> iter
->weight
)
153 rb_link_node(&se
->rb_node
, parent
, p
);
154 rb_insert_color(&se
->rb_node
, tree
);
157 static void print_sym_table(void)
161 float samples_per_sec
= samples
/delay_secs
;
162 float ksamples_per_sec
= (samples
-userspace_samples
)/delay_secs
;
163 float sum_ksamples
= 0.0;
164 struct sym_entry
*syme
, *n
;
165 struct rb_root tmp
= RB_ROOT
;
168 samples
= userspace_samples
= 0;
170 /* Sort the active symbols */
171 pthread_mutex_lock(&active_symbols_lock
);
172 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
173 pthread_mutex_unlock(&active_symbols_lock
);
175 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
176 syme
->snap_count
= syme
->count
[0];
177 if (syme
->snap_count
!= 0) {
178 syme
->weight
= sym_weight(syme
);
179 rb_insert_active_sym(&tmp
, syme
);
180 sum_ksamples
+= syme
->snap_count
;
182 for (j
= 0; j
< nr_counters
; j
++)
183 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
185 list_remove_active_sym(syme
);
191 "------------------------------------------------------------------------------\n");
192 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
194 100.0 - (100.0*((samples_per_sec
-ksamples_per_sec
)/samples_per_sec
)));
196 if (nr_counters
== 1) {
197 printf("%Ld", attrs
[0].sample_period
);
204 for (counter
= 0; counter
< nr_counters
; counter
++) {
208 printf("%s", event_name(counter
));
213 if (target_pid
!= -1)
214 printf(" (target_pid: %d", target_pid
);
218 if (profile_cpu
!= -1)
219 printf(", cpu: %d)\n", profile_cpu
);
221 if (target_pid
!= -1)
224 printf(", %d CPUs)\n", nr_cpus
);
227 printf("------------------------------------------------------------------------------\n\n");
229 if (nr_counters
== 1)
230 printf(" samples pcnt");
232 printf(" weight samples pcnt");
234 printf(" RIP kernel function\n"
235 " ______ _______ _____ ________________ _______________\n\n"
238 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
239 struct sym_entry
*syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
240 struct symbol
*sym
= (struct symbol
*)(syme
+ 1);
241 char *color
= PERF_COLOR_NORMAL
;
244 if (++printed
> print_entries
|| syme
->snap_count
< count_filter
)
247 pcnt
= 100.0 - (100.0 * ((sum_ksamples
- syme
->snap_count
) /
251 * We color high-overhead entries in red, low-overhead
252 * entries in green - and keep the middle ground normal:
255 color
= PERF_COLOR_RED
;
257 color
= PERF_COLOR_GREEN
;
259 if (nr_counters
== 1)
260 printf("%20.2f - ", syme
->weight
);
262 printf("%9.1f %10ld - ", syme
->weight
, syme
->snap_count
);
264 color_fprintf(stdout
, color
, "%4.1f%%", pcnt
);
265 printf(" - %016llx : %s\n", sym
->start
, sym
->name
);
269 static void *display_thread(void *arg
)
271 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
272 int delay_msecs
= delay_secs
* 1000;
274 printf("PerfTop refresh period: %d seconds\n", delay_secs
);
278 } while (!poll(&stdin_poll
, 1, delay_msecs
) == 1);
280 printf("key pressed - exiting.\n");
286 static int symbol_filter(struct dso
*self
, struct symbol
*sym
)
288 static int filter_match
;
289 struct sym_entry
*syme
;
290 const char *name
= sym
->name
;
292 if (!strcmp(name
, "_text") ||
293 !strcmp(name
, "_etext") ||
294 !strcmp(name
, "_sinittext") ||
295 !strncmp("init_module", name
, 11) ||
296 !strncmp("cleanup_module", name
, 14) ||
297 strstr(name
, "_text_start") ||
298 strstr(name
, "_text_end"))
301 syme
= dso__sym_priv(self
, sym
);
302 /* Tag samples to be skipped. */
303 if (!strcmp("default_idle", name
) ||
304 !strcmp("cpu_idle", name
) ||
305 !strcmp("enter_idle", name
) ||
306 !strcmp("exit_idle", name
) ||
307 !strcmp("mwait_idle", name
))
310 if (filter_match
== 1) {
311 filter_end
= sym
->start
;
313 if (filter_end
- filter_start
> 10000) {
315 "hm, too large filter symbol <%s> - skipping.\n",
317 fprintf(stderr
, "symbol filter start: %016lx\n",
319 fprintf(stderr
, " end: %016lx\n",
321 filter_end
= filter_start
= 0;
327 if (filter_match
== 0 && sym_filter
&& !strcmp(name
, sym_filter
)) {
329 filter_start
= sym
->start
;
336 static int parse_symbols(void)
338 struct rb_node
*node
;
341 kernel_dso
= dso__new("[kernel]", sizeof(struct sym_entry
));
342 if (kernel_dso
== NULL
)
345 if (dso__load_kernel(kernel_dso
, NULL
, symbol_filter
, 1) != 0)
348 node
= rb_first(&kernel_dso
->syms
);
349 sym
= rb_entry(node
, struct symbol
, rb_node
);
352 node
= rb_last(&kernel_dso
->syms
);
353 sym
= rb_entry(node
, struct symbol
, rb_node
);
357 dso__fprintf(kernel_dso
, stderr
);
362 dso__delete(kernel_dso
);
367 #define TRACE_COUNT 3
370 * Binary search in the histogram table and record the hit:
372 static void record_ip(uint64_t ip
, int counter
)
374 struct symbol
*sym
= dso__find_symbol(kernel_dso
, ip
);
377 struct sym_entry
*syme
= dso__sym_priv(kernel_dso
, sym
);
380 syme
->count
[counter
]++;
381 pthread_mutex_lock(&active_symbols_lock
);
382 if (list_empty(&syme
->node
) || !syme
->node
.next
)
383 __list_insert_active_sym(syme
);
384 pthread_mutex_unlock(&active_symbols_lock
);
392 static void process_event(uint64_t ip
, int counter
)
396 if (ip
< min_ip
|| ip
> max_ip
) {
401 record_ip(ip
, counter
);
411 static unsigned int mmap_read_head(struct mmap_data
*md
)
413 struct perf_counter_mmap_page
*pc
= md
->base
;
416 head
= pc
->data_head
;
422 struct timeval last_read
, this_read
;
424 static void mmap_read_counter(struct mmap_data
*md
)
426 unsigned int head
= mmap_read_head(md
);
427 unsigned int old
= md
->prev
;
428 unsigned char *data
= md
->base
+ page_size
;
431 gettimeofday(&this_read
, NULL
);
434 * If we're further behind than half the buffer, there's a chance
435 * the writer will bite our tail and mess up the samples under us.
437 * If we somehow ended up ahead of the head, we got messed up.
439 * In either case, truncate and restart at head.
442 if (diff
> md
->mask
/ 2 || diff
< 0) {
446 timersub(&this_read
, &last_read
, &iv
);
447 msecs
= iv
.tv_sec
*1000 + iv
.tv_usec
/1000;
449 fprintf(stderr
, "WARNING: failed to keep up with mmap data."
450 " Last read %lu msecs ago.\n", msecs
);
453 * head points to a known good entry, start there.
458 last_read
= this_read
;
460 for (; old
!= head
;) {
462 struct perf_event_header header
;
464 __u32 pid
, target_pid
;
467 struct perf_event_header header
;
468 __u32 pid
, target_pid
;
472 char filename
[PATH_MAX
];
475 typedef union event_union
{
476 struct perf_event_header header
;
478 struct mmap_event mmap
;
481 event_t
*event
= (event_t
*)&data
[old
& md
->mask
];
485 size_t size
= event
->header
.size
;
488 * Event straddles the mmap boundary -- header should always
489 * be inside due to u64 alignment of output.
491 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
492 unsigned int offset
= old
;
493 unsigned int len
= min(sizeof(*event
), size
), cpy
;
494 void *dst
= &event_copy
;
497 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
498 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
509 if (event
->header
.misc
& PERF_EVENT_MISC_OVERFLOW
) {
510 if (event
->header
.type
& PERF_SAMPLE_IP
)
511 process_event(event
->ip
.ip
, md
->counter
);
518 static struct pollfd event_array
[MAX_NR_CPUS
* MAX_COUNTERS
];
519 static struct mmap_data mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
521 static void mmap_read(void)
525 for (i
= 0; i
< nr_cpus
; i
++) {
526 for (counter
= 0; counter
< nr_counters
; counter
++)
527 mmap_read_counter(&mmap_array
[i
][counter
]);
534 static void start_counter(int i
, int counter
)
536 struct perf_counter_attr
*attr
;
540 if (target_pid
== -1 && profile_cpu
== -1)
543 attr
= attrs
+ counter
;
545 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
549 fd
[i
][counter
] = sys_perf_counter_open(attr
, target_pid
, cpu
, group_fd
, 0);
551 if (fd
[i
][counter
] < 0) {
555 die("No permission - are you root?\n");
557 * If it's cycles then fall back to hrtimer
558 * based cpu-clock-tick sw counter, which
559 * is always available even if no PMU support:
561 if (attr
->type
== PERF_TYPE_HARDWARE
562 && attr
->config
== PERF_COUNT_CPU_CYCLES
) {
565 warning(" ... trying to fall back to cpu-clock-ticks\n");
567 attr
->type
= PERF_TYPE_SOFTWARE
;
568 attr
->config
= PERF_COUNT_CPU_CLOCK
;
572 error("perfcounter syscall returned with %d (%s)\n",
573 fd
[i
][counter
], strerror(err
));
574 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
577 assert(fd
[i
][counter
] >= 0);
578 fcntl(fd
[i
][counter
], F_SETFL
, O_NONBLOCK
);
581 * First counter acts as the group leader:
583 if (group
&& group_fd
== -1)
584 group_fd
= fd
[i
][counter
];
586 event_array
[nr_poll
].fd
= fd
[i
][counter
];
587 event_array
[nr_poll
].events
= POLLIN
;
590 mmap_array
[i
][counter
].counter
= counter
;
591 mmap_array
[i
][counter
].prev
= 0;
592 mmap_array
[i
][counter
].mask
= mmap_pages
*page_size
- 1;
593 mmap_array
[i
][counter
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
594 PROT_READ
, MAP_SHARED
, fd
[i
][counter
], 0);
595 if (mmap_array
[i
][counter
].base
== MAP_FAILED
)
596 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
599 static int __cmd_top(void)
605 for (i
= 0; i
< nr_cpus
; i
++) {
607 for (counter
= 0; counter
< nr_counters
; counter
++)
608 start_counter(i
, counter
);
611 /* Wait for a minimal set of events before starting the snapshot */
612 poll(event_array
, nr_poll
, 100);
616 if (pthread_create(&thread
, NULL
, display_thread
, NULL
)) {
617 printf("Could not create display thread.\n");
622 struct sched_param param
;
624 param
.sched_priority
= realtime_prio
;
625 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
626 printf("Could not set realtime priority.\n");
637 ret
= poll(event_array
, nr_poll
, 100);
643 static const char * const top_usage
[] = {
644 "perf top [<options>]",
648 static const struct option options
[] = {
649 OPT_CALLBACK('e', "event", NULL
, "event",
650 "event selector. use 'perf list' to list available events",
652 OPT_INTEGER('c', "count", &default_interval
,
653 "event period to sample"),
654 OPT_INTEGER('p', "pid", &target_pid
,
655 "profile events on existing pid"),
656 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
657 "system-wide collection from all CPUs"),
658 OPT_INTEGER('C', "CPU", &profile_cpu
,
659 "CPU to profile on"),
660 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
661 "number of mmap data pages"),
662 OPT_INTEGER('r', "realtime", &realtime_prio
,
663 "collect data with this RT SCHED_FIFO priority"),
664 OPT_INTEGER('d', "delay", &delay_secs
,
665 "number of seconds to delay between refreshes"),
666 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
667 "dump the symbol table used for profiling"),
668 OPT_INTEGER('f', "count-filter", &count_filter
,
669 "only display functions with more events than this"),
670 OPT_BOOLEAN('g', "group", &group
,
671 "put the counters into a counter group"),
672 OPT_STRING('s', "sym-filter", &sym_filter
, "pattern",
673 "only display symbols matchig this pattern"),
674 OPT_BOOLEAN('z', "zero", &group
,
675 "zero history across updates"),
676 OPT_INTEGER('F', "freq", &freq
,
677 "profile at this frequency"),
678 OPT_INTEGER('E', "entries", &print_entries
,
679 "display this many functions"),
680 OPT_BOOLEAN('v', "verbose", &verbose
,
681 "be more verbose (show counter open errors, etc)"),
685 int cmd_top(int argc
, const char **argv
, const char *prefix
)
689 page_size
= sysconf(_SC_PAGE_SIZE
);
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");
716 * Fill in the ones not specifically initialized via -c:
718 for (counter
= 0; counter
< nr_counters
; counter
++) {
719 if (attrs
[counter
].sample_period
)
722 attrs
[counter
].sample_period
= default_interval
;
725 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
726 assert(nr_cpus
<= MAX_NR_CPUS
);
727 assert(nr_cpus
>= 0);
729 if (target_pid
!= -1 || profile_cpu
!= -1)