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/color.h"
24 #include "util/evlist.h"
25 #include "util/evsel.h"
26 #include "util/session.h"
27 #include "util/symbol.h"
28 #include "util/thread.h"
29 #include "util/thread_map.h"
30 #include "util/util.h"
31 #include <linux/rbtree.h>
32 #include "util/parse-options.h"
33 #include "util/parse-events.h"
34 #include "util/cpumap.h"
35 #include "util/xyarray.h"
37 #include "util/debug.h"
52 #include <sys/syscall.h>
53 #include <sys/ioctl.h>
55 #include <sys/prctl.h>
60 #include <linux/unistd.h>
61 #include <linux/types.h>
63 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
65 struct perf_evlist
*evsel_list
;
67 static bool system_wide
= false;
69 static int default_interval
= 0;
71 static int count_filter
= 5;
72 static int print_entries
;
74 static int target_pid
= -1;
75 static int target_tid
= -1;
76 static bool inherit
= false;
77 static int realtime_prio
= 0;
78 static bool group
= false;
79 static unsigned int page_size
;
80 static unsigned int mmap_pages
= 128;
81 static int freq
= 1000; /* 1 KHz */
83 static int delay_secs
= 2;
84 static bool zero
= false;
85 static bool dump_symtab
= false;
87 static bool hide_kernel_symbols
= false;
88 static bool hide_user_symbols
= false;
89 static struct winsize winsize
;
97 unsigned long count
[MAX_COUNTERS
];
99 struct source_line
*next
;
102 static const char *sym_filter
= NULL
;
103 struct sym_entry
*sym_filter_entry
= NULL
;
104 struct sym_entry
*sym_filter_entry_sched
= NULL
;
105 static int sym_pcnt_filter
= 5;
106 static int sym_counter
= 0;
107 static struct perf_evsel
*sym_evsel
= NULL
;
108 static int display_weighted
= -1;
109 static const char *cpu_list
;
115 struct sym_entry_source
{
116 struct source_line
*source
;
117 struct source_line
*lines
;
118 struct source_line
**lines_tail
;
119 pthread_mutex_t lock
;
123 struct rb_node rb_node
;
124 struct list_head node
;
125 unsigned long snap_count
;
131 struct sym_entry_source
*src
;
132 unsigned long count
[0];
139 static inline struct symbol
*sym_entry__symbol(struct sym_entry
*self
)
141 return ((void *)self
) + symbol_conf
.priv_size
;
144 void get_term_dimensions(struct winsize
*ws
)
146 char *s
= getenv("LINES");
149 ws
->ws_row
= atoi(s
);
150 s
= getenv("COLUMNS");
152 ws
->ws_col
= atoi(s
);
153 if (ws
->ws_row
&& ws
->ws_col
)
158 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
159 ws
->ws_row
&& ws
->ws_col
)
166 static void update_print_entries(struct winsize
*ws
)
168 print_entries
= ws
->ws_row
;
170 if (print_entries
> 9)
174 static void sig_winch_handler(int sig __used
)
176 get_term_dimensions(&winsize
);
177 update_print_entries(&winsize
);
180 static int parse_source(struct sym_entry
*syme
)
183 struct sym_entry_source
*source
;
186 char command
[PATH_MAX
*2];
193 sym
= sym_entry__symbol(syme
);
197 * We can't annotate with just /proc/kallsyms
199 if (map
->dso
->origin
== DSO__ORIG_KERNEL
)
202 if (syme
->src
== NULL
) {
203 syme
->src
= zalloc(sizeof(*source
));
204 if (syme
->src
== NULL
)
206 pthread_mutex_init(&syme
->src
->lock
, NULL
);
212 pthread_mutex_lock(&source
->lock
);
215 path
= map
->dso
->long_name
;
217 len
= sym
->end
- sym
->start
;
220 "objdump --start-address=%#0*" PRIx64
" --stop-address=%#0*" PRIx64
" -dS %s",
221 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->start
),
222 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->end
), path
);
224 file
= popen(command
, "r");
228 pthread_mutex_lock(&source
->lock
);
229 source
->lines_tail
= &source
->lines
;
230 while (!feof(file
)) {
231 struct source_line
*src
;
235 src
= malloc(sizeof(struct source_line
));
237 memset(src
, 0, sizeof(struct source_line
));
239 if (getline(&src
->line
, &dummy
, file
) < 0)
244 c
= strchr(src
->line
, '\n');
249 *source
->lines_tail
= src
;
250 source
->lines_tail
= &src
->next
;
252 src
->eip
= strtoull(src
->line
, &sep
, 16);
254 src
->eip
= map__objdump_2ip(map
, src
->eip
);
255 else /* this line has no ip info (e.g. source line) */
260 sym_filter_entry
= syme
;
261 pthread_mutex_unlock(&source
->lock
);
265 static void __zero_source_counters(struct sym_entry
*syme
)
268 struct source_line
*line
;
270 line
= syme
->src
->lines
;
272 for (i
= 0; i
< evsel_list
->nr_entries
; i
++)
278 static void record_precise_ip(struct sym_entry
*syme
, int counter
, u64 ip
)
280 struct source_line
*line
;
282 if (syme
!= sym_filter_entry
)
285 if (pthread_mutex_trylock(&syme
->src
->lock
))
288 if (syme
->src
== NULL
|| syme
->src
->source
== NULL
)
291 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
292 /* skip lines without IP info */
295 if (line
->eip
== ip
) {
296 line
->count
[counter
]++;
303 pthread_mutex_unlock(&syme
->src
->lock
);
306 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
308 static void lookup_sym_source(struct sym_entry
*syme
)
310 struct symbol
*symbol
= sym_entry__symbol(syme
);
311 struct source_line
*line
;
312 char pattern
[PATTERN_LEN
+ 1];
314 sprintf(pattern
, "%0*" PRIx64
" <", BITS_PER_LONG
/ 4,
315 map__rip_2objdump(syme
->map
, symbol
->start
));
317 pthread_mutex_lock(&syme
->src
->lock
);
318 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
319 if (memcmp(line
->line
, pattern
, PATTERN_LEN
) == 0) {
320 syme
->src
->source
= line
;
324 pthread_mutex_unlock(&syme
->src
->lock
);
327 static void show_lines(struct source_line
*queue
, int count
, int total
)
330 struct source_line
*line
;
333 for (i
= 0; i
< count
; i
++) {
334 float pcnt
= 100.0*(float)line
->count
[sym_counter
]/(float)total
;
336 printf("%8li %4.1f%%\t%s\n", line
->count
[sym_counter
], pcnt
, line
->line
);
341 #define TRACE_COUNT 3
343 static void show_details(struct sym_entry
*syme
)
345 struct symbol
*symbol
;
346 struct source_line
*line
;
347 struct source_line
*line_queue
= NULL
;
349 int line_queue_count
= 0, total
= 0, more
= 0;
354 if (!syme
->src
->source
)
355 lookup_sym_source(syme
);
357 if (!syme
->src
->source
)
360 symbol
= sym_entry__symbol(syme
);
361 printf("Showing %s for %s\n", event_name(sym_evsel
), symbol
->name
);
362 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter
);
364 pthread_mutex_lock(&syme
->src
->lock
);
365 line
= syme
->src
->source
;
367 total
+= line
->count
[sym_counter
];
371 line
= syme
->src
->source
;
375 if (!line_queue_count
)
379 if (line
->count
[sym_counter
])
380 pcnt
= 100.0 * line
->count
[sym_counter
] / (float)total
;
381 if (pcnt
>= (float)sym_pcnt_filter
) {
382 if (displayed
<= print_entries
)
383 show_lines(line_queue
, line_queue_count
, total
);
385 displayed
+= line_queue_count
;
386 line_queue_count
= 0;
388 } else if (line_queue_count
> TRACE_COUNT
) {
389 line_queue
= line_queue
->next
;
393 line
->count
[sym_counter
] = zero
? 0 : line
->count
[sym_counter
] * 7 / 8;
396 pthread_mutex_unlock(&syme
->src
->lock
);
398 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
402 * Symbols will be added here in perf_event__process_sample and will get out
405 static LIST_HEAD(active_symbols
);
406 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
409 * Ordering weight: count-1 * count-2 * ... / count-n
411 static double sym_weight(const struct sym_entry
*sym
)
413 double weight
= sym
->snap_count
;
416 if (!display_weighted
)
419 for (counter
= 1; counter
< evsel_list
->nr_entries
- 1; counter
++)
420 weight
*= sym
->count
[counter
];
422 weight
/= (sym
->count
[counter
] + 1);
428 static long kernel_samples
, us_samples
;
429 static long exact_samples
;
430 static long guest_us_samples
, guest_kernel_samples
;
431 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
433 static void __list_insert_active_sym(struct sym_entry
*syme
)
435 list_add(&syme
->node
, &active_symbols
);
438 static void list_remove_active_sym(struct sym_entry
*syme
)
440 pthread_mutex_lock(&active_symbols_lock
);
441 list_del_init(&syme
->node
);
442 pthread_mutex_unlock(&active_symbols_lock
);
445 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
447 struct rb_node
**p
= &tree
->rb_node
;
448 struct rb_node
*parent
= NULL
;
449 struct sym_entry
*iter
;
453 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
455 if (se
->weight
> iter
->weight
)
461 rb_link_node(&se
->rb_node
, parent
, p
);
462 rb_insert_color(&se
->rb_node
, tree
);
465 static void print_sym_table(struct perf_session
*session
)
468 struct perf_evsel
*counter
;
469 int snap
= !display_weighted
? sym_counter
: 0;
470 float samples_per_sec
= samples
/delay_secs
;
471 float ksamples_per_sec
= kernel_samples
/delay_secs
;
472 float us_samples_per_sec
= (us_samples
)/delay_secs
;
473 float guest_kernel_samples_per_sec
= (guest_kernel_samples
)/delay_secs
;
474 float guest_us_samples_per_sec
= (guest_us_samples
)/delay_secs
;
475 float esamples_percent
= (100.0*exact_samples
)/samples
;
476 float sum_ksamples
= 0.0;
477 struct sym_entry
*syme
, *n
;
478 struct rb_root tmp
= RB_ROOT
;
480 int sym_width
= 0, dso_width
= 0, dso_short_width
= 0;
481 const int win_width
= winsize
.ws_col
- 1;
483 samples
= us_samples
= kernel_samples
= exact_samples
= 0;
484 guest_kernel_samples
= guest_us_samples
= 0;
486 /* Sort the active symbols */
487 pthread_mutex_lock(&active_symbols_lock
);
488 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
489 pthread_mutex_unlock(&active_symbols_lock
);
491 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
492 syme
->snap_count
= syme
->count
[snap
];
493 if (syme
->snap_count
!= 0) {
495 if ((hide_user_symbols
&&
496 syme
->origin
== PERF_RECORD_MISC_USER
) ||
497 (hide_kernel_symbols
&&
498 syme
->origin
== PERF_RECORD_MISC_KERNEL
)) {
499 list_remove_active_sym(syme
);
502 syme
->weight
= sym_weight(syme
);
503 rb_insert_active_sym(&tmp
, syme
);
504 sum_ksamples
+= syme
->snap_count
;
506 for (j
= 0; j
< evsel_list
->nr_entries
; j
++)
507 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
509 list_remove_active_sym(syme
);
515 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%"
518 100.0 - (100.0 * ((samples_per_sec
- ksamples_per_sec
) /
522 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%"
523 " guest kernel:%4.1f%% guest us:%4.1f%%"
526 100.0 - (100.0 * ((samples_per_sec
-ksamples_per_sec
) /
528 100.0 - (100.0 * ((samples_per_sec
-us_samples_per_sec
) /
530 100.0 - (100.0 * ((samples_per_sec
-
531 guest_kernel_samples_per_sec
) /
533 100.0 - (100.0 * ((samples_per_sec
-
534 guest_us_samples_per_sec
) /
539 if (evsel_list
->nr_entries
== 1 || !display_weighted
) {
540 struct perf_evsel
*first
;
541 first
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
542 printf("%" PRIu64
, (uint64_t)first
->attr
.sample_period
);
549 if (!display_weighted
)
550 printf("%s", event_name(sym_evsel
));
551 else list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
555 printf("%s", event_name(counter
));
560 if (target_pid
!= -1)
561 printf(" (target_pid: %d", target_pid
);
562 else if (target_tid
!= -1)
563 printf(" (target_tid: %d", target_tid
);
568 printf(", CPU%s: %s)\n", evsel_list
->cpus
->nr
> 1 ? "s" : "", cpu_list
);
570 if (target_tid
!= -1)
573 printf(", %d CPU%s)\n", evsel_list
->cpus
->nr
,
574 evsel_list
->cpus
->nr
> 1 ? "s" : "");
577 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
579 if (session
->hists
.stats
.total_lost
!= 0) {
580 color_fprintf(stdout
, PERF_COLOR_RED
, "WARNING:");
581 printf(" LOST %" PRIu64
" events, Check IO/CPU overload\n",
582 session
->hists
.stats
.total_lost
);
585 if (sym_filter_entry
) {
586 show_details(sym_filter_entry
);
591 * Find the longest symbol name that will be displayed
593 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
594 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
595 if (++printed
> print_entries
||
596 (int)syme
->snap_count
< count_filter
)
599 if (syme
->map
->dso
->long_name_len
> dso_width
)
600 dso_width
= syme
->map
->dso
->long_name_len
;
602 if (syme
->map
->dso
->short_name_len
> dso_short_width
)
603 dso_short_width
= syme
->map
->dso
->short_name_len
;
605 if (syme
->name_len
> sym_width
)
606 sym_width
= syme
->name_len
;
611 if (sym_width
+ dso_width
> winsize
.ws_col
- 29) {
612 dso_width
= dso_short_width
;
613 if (sym_width
+ dso_width
> winsize
.ws_col
- 29)
614 sym_width
= winsize
.ws_col
- dso_width
- 29;
617 if (evsel_list
->nr_entries
== 1)
618 printf(" samples pcnt");
620 printf(" weight samples pcnt");
624 printf(" %-*.*s DSO\n", sym_width
, sym_width
, "function");
625 printf(" %s _______ _____",
626 evsel_list
->nr_entries
== 1 ? " " : "______");
628 printf(" ________________");
629 printf(" %-*.*s", sym_width
, sym_width
, graph_line
);
630 printf(" %-*.*s", dso_width
, dso_width
, graph_line
);
633 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
637 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
638 sym
= sym_entry__symbol(syme
);
639 if (++printed
> print_entries
|| (int)syme
->snap_count
< count_filter
)
642 pcnt
= 100.0 - (100.0 * ((sum_ksamples
- syme
->snap_count
) /
645 if (evsel_list
->nr_entries
== 1 || !display_weighted
)
646 printf("%20.2f ", syme
->weight
);
648 printf("%9.1f %10ld ", syme
->weight
, syme
->snap_count
);
650 percent_color_fprintf(stdout
, "%4.1f%%", pcnt
);
652 printf(" %016" PRIx64
, sym
->start
);
653 printf(" %-*.*s", sym_width
, sym_width
, sym
->name
);
654 printf(" %-*.*s\n", dso_width
, dso_width
,
655 dso_width
>= syme
->map
->dso
->long_name_len
?
656 syme
->map
->dso
->long_name
:
657 syme
->map
->dso
->short_name
);
661 static void prompt_integer(int *target
, const char *msg
)
663 char *buf
= malloc(0), *p
;
667 fprintf(stdout
, "\n%s: ", msg
);
668 if (getline(&buf
, &dummy
, stdin
) < 0)
671 p
= strchr(buf
, '\n');
681 tmp
= strtoul(buf
, NULL
, 10);
687 static void prompt_percent(int *target
, const char *msg
)
691 prompt_integer(&tmp
, msg
);
692 if (tmp
>= 0 && tmp
<= 100)
696 static void prompt_symbol(struct sym_entry
**target
, const char *msg
)
698 char *buf
= malloc(0), *p
;
699 struct sym_entry
*syme
= *target
, *n
, *found
= NULL
;
702 /* zero counters of active symbol */
704 pthread_mutex_lock(&syme
->src
->lock
);
705 __zero_source_counters(syme
);
707 pthread_mutex_unlock(&syme
->src
->lock
);
710 fprintf(stdout
, "\n%s: ", msg
);
711 if (getline(&buf
, &dummy
, stdin
) < 0)
714 p
= strchr(buf
, '\n');
718 pthread_mutex_lock(&active_symbols_lock
);
719 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
720 pthread_mutex_unlock(&active_symbols_lock
);
722 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
723 struct symbol
*sym
= sym_entry__symbol(syme
);
725 if (!strcmp(buf
, sym
->name
)) {
732 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
742 static void print_mapped_keys(void)
746 if (sym_filter_entry
) {
747 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
751 fprintf(stdout
, "\nMapped keys:\n");
752 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", delay_secs
);
753 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", print_entries
);
755 if (evsel_list
->nr_entries
> 1)
756 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel
));
758 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", count_filter
);
760 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter
);
761 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
762 fprintf(stdout
, "\t[S] stop annotation.\n");
764 if (evsel_list
->nr_entries
> 1)
765 fprintf(stdout
, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted
? 1 : 0);
768 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
769 hide_kernel_symbols
? "yes" : "no");
771 "\t[U] hide user symbols. \t(%s)\n",
772 hide_user_symbols
? "yes" : "no");
773 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", zero
? 1 : 0);
774 fprintf(stdout
, "\t[qQ] quit.\n");
777 static int key_mapped(int c
)
794 return evsel_list
->nr_entries
> 1 ? 1 : 0;
802 static void handle_keypress(struct perf_session
*session
, int c
)
804 if (!key_mapped(c
)) {
805 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
806 struct termios tc
, save
;
809 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
814 tc
.c_lflag
&= ~(ICANON
| ECHO
);
817 tcsetattr(0, TCSANOW
, &tc
);
819 poll(&stdin_poll
, 1, -1);
822 tcsetattr(0, TCSAFLUSH
, &save
);
829 prompt_integer(&delay_secs
, "Enter display delay");
834 prompt_integer(&print_entries
, "Enter display entries (lines)");
835 if (print_entries
== 0) {
836 sig_winch_handler(SIGWINCH
);
837 signal(SIGWINCH
, sig_winch_handler
);
839 signal(SIGWINCH
, SIG_DFL
);
842 if (evsel_list
->nr_entries
> 1) {
843 fprintf(stderr
, "\nAvailable events:");
845 list_for_each_entry(sym_evsel
, &evsel_list
->entries
, node
)
846 fprintf(stderr
, "\n\t%d %s", sym_evsel
->idx
, event_name(sym_evsel
));
848 prompt_integer(&sym_counter
, "Enter details event counter");
850 if (sym_counter
>= evsel_list
->nr_entries
) {
851 sym_evsel
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
853 fprintf(stderr
, "Sorry, no such event, using %s.\n", event_name(sym_evsel
));
857 list_for_each_entry(sym_evsel
, &evsel_list
->entries
, node
)
858 if (sym_evsel
->idx
== sym_counter
)
860 } else sym_counter
= 0;
863 prompt_integer(&count_filter
, "Enter display event count filter");
866 prompt_percent(&sym_pcnt_filter
, "Enter details display event filter (percent)");
869 hide_kernel_symbols
= !hide_kernel_symbols
;
873 printf("exiting.\n");
875 perf_session__fprintf_dsos(session
, stderr
);
878 prompt_symbol(&sym_filter_entry
, "Enter details symbol");
881 if (!sym_filter_entry
)
884 struct sym_entry
*syme
= sym_filter_entry
;
886 pthread_mutex_lock(&syme
->src
->lock
);
887 sym_filter_entry
= NULL
;
888 __zero_source_counters(syme
);
889 pthread_mutex_unlock(&syme
->src
->lock
);
893 hide_user_symbols
= !hide_user_symbols
;
896 display_weighted
= ~display_weighted
;
906 static void *display_thread(void *arg __used
)
908 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
909 struct termios tc
, save
;
911 struct perf_session
*session
= (struct perf_session
*) arg
;
915 tc
.c_lflag
&= ~(ICANON
| ECHO
);
920 delay_msecs
= delay_secs
* 1000;
921 tcsetattr(0, TCSANOW
, &tc
);
926 print_sym_table(session
);
927 } while (!poll(&stdin_poll
, 1, delay_msecs
) == 1);
930 tcsetattr(0, TCSAFLUSH
, &save
);
932 handle_keypress(session
, c
);
938 /* Tag samples to be skipped. */
939 static const char *skip_symbols
[] = {
946 "mwait_idle_with_hints",
948 "ppc64_runlatch_off",
949 "pseries_dedicated_idle_sleep",
953 static int symbol_filter(struct map
*map
, struct symbol
*sym
)
955 struct sym_entry
*syme
;
956 const char *name
= sym
->name
;
960 * ppc64 uses function descriptors and appends a '.' to the
961 * start of every instruction address. Remove it.
966 if (!strcmp(name
, "_text") ||
967 !strcmp(name
, "_etext") ||
968 !strcmp(name
, "_sinittext") ||
969 !strncmp("init_module", name
, 11) ||
970 !strncmp("cleanup_module", name
, 14) ||
971 strstr(name
, "_text_start") ||
972 strstr(name
, "_text_end"))
975 syme
= symbol__priv(sym
);
979 if (!sym_filter_entry
&& sym_filter
&& !strcmp(name
, sym_filter
)) {
980 /* schedule initial sym_filter_entry setup */
981 sym_filter_entry_sched
= syme
;
985 for (i
= 0; skip_symbols
[i
]; i
++) {
986 if (!strcmp(skip_symbols
[i
], name
)) {
993 syme
->name_len
= strlen(sym
->name
);
998 static void perf_event__process_sample(const union perf_event
*event
,
999 struct perf_sample
*sample
,
1000 struct perf_session
*session
)
1002 u64 ip
= event
->ip
.ip
;
1003 struct sym_entry
*syme
;
1004 struct addr_location al
;
1005 struct machine
*machine
;
1006 u8 origin
= event
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
1011 case PERF_RECORD_MISC_USER
:
1013 if (hide_user_symbols
)
1015 machine
= perf_session__find_host_machine(session
);
1017 case PERF_RECORD_MISC_KERNEL
:
1019 if (hide_kernel_symbols
)
1021 machine
= perf_session__find_host_machine(session
);
1023 case PERF_RECORD_MISC_GUEST_KERNEL
:
1024 ++guest_kernel_samples
;
1025 machine
= perf_session__find_machine(session
, event
->ip
.pid
);
1027 case PERF_RECORD_MISC_GUEST_USER
:
1030 * TODO: we don't process guest user from host side
1031 * except simple counting.
1038 if (!machine
&& perf_guest
) {
1039 pr_err("Can't find guest [%d]'s kernel information\n",
1044 if (event
->header
.misc
& PERF_RECORD_MISC_EXACT_IP
)
1047 if (perf_event__preprocess_sample(event
, session
, &al
, sample
,
1048 symbol_filter
) < 0 ||
1052 if (al
.sym
== NULL
) {
1054 * As we do lazy loading of symtabs we only will know if the
1055 * specified vmlinux file is invalid when we actually have a
1056 * hit in kernel space and then try to load it. So if we get
1057 * here and there are _no_ symbols in the DSO backing the
1058 * kernel map, bail out.
1060 * We may never get here, for instance, if we use -K/
1061 * --hide-kernel-symbols, even if the user specifies an
1062 * invalid --vmlinux ;-)
1064 if (al
.map
== machine
->vmlinux_maps
[MAP__FUNCTION
] &&
1065 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
1066 pr_err("The %s file can't be used\n",
1067 symbol_conf
.vmlinux_name
);
1074 /* let's see, whether we need to install initial sym_filter_entry */
1075 if (sym_filter_entry_sched
) {
1076 sym_filter_entry
= sym_filter_entry_sched
;
1077 sym_filter_entry_sched
= NULL
;
1078 if (parse_source(sym_filter_entry
) < 0) {
1079 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
1081 pr_err("Can't annotate %s", sym
->name
);
1082 if (sym_filter_entry
->map
->dso
->origin
== DSO__ORIG_KERNEL
) {
1083 pr_err(": No vmlinux file was found in the path:\n");
1084 machine__fprintf_vmlinux_path(machine
, stderr
);
1091 syme
= symbol__priv(al
.sym
);
1093 struct perf_evsel
*evsel
;
1095 syme
->origin
= origin
;
1096 evsel
= perf_evlist__id2evsel(evsel_list
, sample
->id
);
1097 assert(evsel
!= NULL
);
1098 syme
->count
[evsel
->idx
]++;
1099 record_precise_ip(syme
, evsel
->idx
, ip
);
1100 pthread_mutex_lock(&active_symbols_lock
);
1101 if (list_empty(&syme
->node
) || !syme
->node
.next
)
1102 __list_insert_active_sym(syme
);
1103 pthread_mutex_unlock(&active_symbols_lock
);
1107 static void perf_session__mmap_read_cpu(struct perf_session
*self
, int cpu
)
1109 struct perf_sample sample
;
1110 union perf_event
*event
;
1112 while ((event
= perf_evlist__read_on_cpu(evsel_list
, cpu
)) != NULL
) {
1113 perf_session__parse_sample(self
, event
, &sample
);
1115 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
1116 perf_event__process_sample(event
, &sample
, self
);
1118 perf_event__process(event
, &sample
, self
);
1122 static void perf_session__mmap_read(struct perf_session
*self
)
1126 for (i
= 0; i
< evsel_list
->cpus
->nr
; i
++)
1127 perf_session__mmap_read_cpu(self
, i
);
1130 static void start_counters(struct perf_evlist
*evlist
)
1132 struct perf_evsel
*counter
;
1134 list_for_each_entry(counter
, &evlist
->entries
, node
) {
1135 struct perf_event_attr
*attr
= &counter
->attr
;
1137 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
1140 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
1142 attr
->sample_freq
= freq
;
1145 if (evlist
->nr_entries
> 1) {
1146 attr
->sample_type
|= PERF_SAMPLE_ID
;
1147 attr
->read_format
|= PERF_FORMAT_ID
;
1152 if (perf_evsel__open(counter
, evsel_list
->cpus
,
1153 evsel_list
->threads
, group
, inherit
) < 0) {
1156 if (err
== EPERM
|| err
== EACCES
)
1157 die("Permission error - are you root?\n"
1158 "\t Consider tweaking"
1159 " /proc/sys/kernel/perf_event_paranoid.\n");
1161 * If it's cycles then fall back to hrtimer
1162 * based cpu-clock-tick sw counter, which
1163 * is always available even if no PMU support:
1165 if (attr
->type
== PERF_TYPE_HARDWARE
&&
1166 attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
1169 warning(" ... trying to fall back to cpu-clock-ticks\n");
1171 attr
->type
= PERF_TYPE_SOFTWARE
;
1172 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
1176 error("sys_perf_event_open() syscall returned with %d "
1177 "(%s). /bin/dmesg may provide additional information.\n",
1178 err
, strerror(err
));
1179 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1184 if (perf_evlist__mmap(evlist
, mmap_pages
, false) < 0)
1185 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
1188 static int __cmd_top(void)
1191 struct perf_evsel
*first
;
1194 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1195 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1197 struct perf_session
*session
= perf_session__new(NULL
, O_WRONLY
, false, false, NULL
);
1198 if (session
== NULL
)
1201 if (target_tid
!= -1)
1202 perf_event__synthesize_thread(target_tid
, perf_event__process
,
1205 perf_event__synthesize_threads(perf_event__process
, session
);
1207 start_counters(evsel_list
);
1208 first
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
1209 perf_session__set_sample_type(session
, first
->attr
.sample_type
);
1211 /* Wait for a minimal set of events before starting the snapshot */
1212 poll(evsel_list
->pollfd
, evsel_list
->nr_fds
, 100);
1214 perf_session__mmap_read(session
);
1216 if (pthread_create(&thread
, NULL
, display_thread
, session
)) {
1217 printf("Could not create display thread.\n");
1221 if (realtime_prio
) {
1222 struct sched_param param
;
1224 param
.sched_priority
= realtime_prio
;
1225 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1226 printf("Could not set realtime priority.\n");
1234 perf_session__mmap_read(session
);
1236 if (hits
== samples
)
1237 ret
= poll(evsel_list
->pollfd
, evsel_list
->nr_fds
, 100);
1243 static const char * const top_usage
[] = {
1244 "perf top [<options>]",
1248 static const struct option options
[] = {
1249 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1250 "event selector. use 'perf list' to list available events",
1252 OPT_INTEGER('c', "count", &default_interval
,
1253 "event period to sample"),
1254 OPT_INTEGER('p', "pid", &target_pid
,
1255 "profile events on existing process id"),
1256 OPT_INTEGER('t', "tid", &target_tid
,
1257 "profile events on existing thread id"),
1258 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1259 "system-wide collection from all CPUs"),
1260 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
1261 "list of cpus to monitor"),
1262 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1263 "file", "vmlinux pathname"),
1264 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols
,
1265 "hide kernel symbols"),
1266 OPT_UINTEGER('m', "mmap-pages", &mmap_pages
, "number of mmap data pages"),
1267 OPT_INTEGER('r', "realtime", &realtime_prio
,
1268 "collect data with this RT SCHED_FIFO priority"),
1269 OPT_INTEGER('d', "delay", &delay_secs
,
1270 "number of seconds to delay between refreshes"),
1271 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
1272 "dump the symbol table used for profiling"),
1273 OPT_INTEGER('f', "count-filter", &count_filter
,
1274 "only display functions with more events than this"),
1275 OPT_BOOLEAN('g', "group", &group
,
1276 "put the counters into a counter group"),
1277 OPT_BOOLEAN('i', "inherit", &inherit
,
1278 "child tasks inherit counters"),
1279 OPT_STRING('s', "sym-annotate", &sym_filter
, "symbol name",
1280 "symbol to annotate"),
1281 OPT_BOOLEAN('z', "zero", &zero
,
1282 "zero history across updates"),
1283 OPT_INTEGER('F', "freq", &freq
,
1284 "profile at this frequency"),
1285 OPT_INTEGER('E', "entries", &print_entries
,
1286 "display this many functions"),
1287 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols
,
1288 "hide user symbols"),
1289 OPT_INCR('v', "verbose", &verbose
,
1290 "be more verbose (show counter open errors, etc)"),
1294 int cmd_top(int argc
, const char **argv
, const char *prefix __used
)
1296 struct perf_evsel
*pos
;
1297 int status
= -ENOMEM
;
1299 evsel_list
= perf_evlist__new(NULL
, NULL
);
1300 if (evsel_list
== NULL
)
1303 page_size
= sysconf(_SC_PAGE_SIZE
);
1305 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1307 usage_with_options(top_usage
, options
);
1309 /* CPU and PID are mutually exclusive */
1310 if (target_tid
> 0 && cpu_list
) {
1311 printf("WARNING: PID switch overriding CPU\n");
1316 if (target_pid
!= -1)
1317 target_tid
= target_pid
;
1319 if (perf_evlist__create_maps(evsel_list
, target_pid
,
1320 target_tid
, cpu_list
) < 0)
1321 usage_with_options(top_usage
, options
);
1323 if (!evsel_list
->nr_entries
&&
1324 perf_evlist__add_default(evsel_list
) < 0) {
1325 pr_err("Not enough memory for event selector list\n");
1333 * User specified count overrides default frequency.
1335 if (default_interval
)
1338 default_interval
= freq
;
1340 fprintf(stderr
, "frequency and count are zero, aborting\n");
1344 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1345 if (perf_evsel__alloc_fd(pos
, evsel_list
->cpus
->nr
,
1346 evsel_list
->threads
->nr
) < 0)
1349 * Fill in the ones not specifically initialized via -c:
1351 if (pos
->attr
.sample_period
)
1354 pos
->attr
.sample_period
= default_interval
;
1357 if (perf_evlist__alloc_pollfd(evsel_list
) < 0 ||
1358 perf_evlist__alloc_mmap(evsel_list
) < 0)
1361 sym_evsel
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
1363 symbol_conf
.priv_size
= (sizeof(struct sym_entry
) +
1364 (evsel_list
->nr_entries
+ 1) * sizeof(unsigned long));
1366 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1367 if (symbol__init() < 0)
1370 get_term_dimensions(&winsize
);
1371 if (print_entries
== 0) {
1372 update_print_entries(&winsize
);
1373 signal(SIGWINCH
, sig_winch_handler
);
1376 status
= __cmd_top();
1378 perf_evlist__delete(evsel_list
);