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 struct thread_map
*threads
;
77 static bool inherit
= false;
78 static struct cpu_map
*cpus
;
79 static int realtime_prio
= 0;
80 static bool group
= false;
81 static unsigned int page_size
;
82 static unsigned int mmap_pages
= 128;
83 static int freq
= 1000; /* 1 KHz */
85 static int delay_secs
= 2;
86 static bool zero
= false;
87 static bool dump_symtab
= false;
89 static bool hide_kernel_symbols
= false;
90 static bool hide_user_symbols
= false;
91 static struct winsize winsize
;
99 unsigned long count
[MAX_COUNTERS
];
101 struct source_line
*next
;
104 static const char *sym_filter
= NULL
;
105 struct sym_entry
*sym_filter_entry
= NULL
;
106 struct sym_entry
*sym_filter_entry_sched
= NULL
;
107 static int sym_pcnt_filter
= 5;
108 static int sym_counter
= 0;
109 static struct perf_evsel
*sym_evsel
= NULL
;
110 static int display_weighted
= -1;
111 static const char *cpu_list
;
117 struct sym_entry_source
{
118 struct source_line
*source
;
119 struct source_line
*lines
;
120 struct source_line
**lines_tail
;
121 pthread_mutex_t lock
;
125 struct rb_node rb_node
;
126 struct list_head node
;
127 unsigned long snap_count
;
133 struct sym_entry_source
*src
;
134 unsigned long count
[0];
141 static inline struct symbol
*sym_entry__symbol(struct sym_entry
*self
)
143 return ((void *)self
) + symbol_conf
.priv_size
;
146 void get_term_dimensions(struct winsize
*ws
)
148 char *s
= getenv("LINES");
151 ws
->ws_row
= atoi(s
);
152 s
= getenv("COLUMNS");
154 ws
->ws_col
= atoi(s
);
155 if (ws
->ws_row
&& ws
->ws_col
)
160 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
161 ws
->ws_row
&& ws
->ws_col
)
168 static void update_print_entries(struct winsize
*ws
)
170 print_entries
= ws
->ws_row
;
172 if (print_entries
> 9)
176 static void sig_winch_handler(int sig __used
)
178 get_term_dimensions(&winsize
);
179 update_print_entries(&winsize
);
182 static int parse_source(struct sym_entry
*syme
)
185 struct sym_entry_source
*source
;
188 char command
[PATH_MAX
*2];
195 sym
= sym_entry__symbol(syme
);
199 * We can't annotate with just /proc/kallsyms
201 if (map
->dso
->origin
== DSO__ORIG_KERNEL
)
204 if (syme
->src
== NULL
) {
205 syme
->src
= zalloc(sizeof(*source
));
206 if (syme
->src
== NULL
)
208 pthread_mutex_init(&syme
->src
->lock
, NULL
);
214 pthread_mutex_lock(&source
->lock
);
217 path
= map
->dso
->long_name
;
219 len
= sym
->end
- sym
->start
;
222 "objdump --start-address=%#0*" PRIx64
" --stop-address=%#0*" PRIx64
" -dS %s",
223 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->start
),
224 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->end
), path
);
226 file
= popen(command
, "r");
230 pthread_mutex_lock(&source
->lock
);
231 source
->lines_tail
= &source
->lines
;
232 while (!feof(file
)) {
233 struct source_line
*src
;
237 src
= malloc(sizeof(struct source_line
));
239 memset(src
, 0, sizeof(struct source_line
));
241 if (getline(&src
->line
, &dummy
, file
) < 0)
246 c
= strchr(src
->line
, '\n');
251 *source
->lines_tail
= src
;
252 source
->lines_tail
= &src
->next
;
254 src
->eip
= strtoull(src
->line
, &sep
, 16);
256 src
->eip
= map__objdump_2ip(map
, src
->eip
);
257 else /* this line has no ip info (e.g. source line) */
262 sym_filter_entry
= syme
;
263 pthread_mutex_unlock(&source
->lock
);
267 static void __zero_source_counters(struct sym_entry
*syme
)
270 struct source_line
*line
;
272 line
= syme
->src
->lines
;
274 for (i
= 0; i
< evsel_list
->nr_entries
; i
++)
280 static void record_precise_ip(struct sym_entry
*syme
, int counter
, u64 ip
)
282 struct source_line
*line
;
284 if (syme
!= sym_filter_entry
)
287 if (pthread_mutex_trylock(&syme
->src
->lock
))
290 if (syme
->src
== NULL
|| syme
->src
->source
== NULL
)
293 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
294 /* skip lines without IP info */
297 if (line
->eip
== ip
) {
298 line
->count
[counter
]++;
305 pthread_mutex_unlock(&syme
->src
->lock
);
308 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
310 static void lookup_sym_source(struct sym_entry
*syme
)
312 struct symbol
*symbol
= sym_entry__symbol(syme
);
313 struct source_line
*line
;
314 char pattern
[PATTERN_LEN
+ 1];
316 sprintf(pattern
, "%0*" PRIx64
" <", BITS_PER_LONG
/ 4,
317 map__rip_2objdump(syme
->map
, symbol
->start
));
319 pthread_mutex_lock(&syme
->src
->lock
);
320 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
321 if (memcmp(line
->line
, pattern
, PATTERN_LEN
) == 0) {
322 syme
->src
->source
= line
;
326 pthread_mutex_unlock(&syme
->src
->lock
);
329 static void show_lines(struct source_line
*queue
, int count
, int total
)
332 struct source_line
*line
;
335 for (i
= 0; i
< count
; i
++) {
336 float pcnt
= 100.0*(float)line
->count
[sym_counter
]/(float)total
;
338 printf("%8li %4.1f%%\t%s\n", line
->count
[sym_counter
], pcnt
, line
->line
);
343 #define TRACE_COUNT 3
345 static void show_details(struct sym_entry
*syme
)
347 struct symbol
*symbol
;
348 struct source_line
*line
;
349 struct source_line
*line_queue
= NULL
;
351 int line_queue_count
= 0, total
= 0, more
= 0;
356 if (!syme
->src
->source
)
357 lookup_sym_source(syme
);
359 if (!syme
->src
->source
)
362 symbol
= sym_entry__symbol(syme
);
363 printf("Showing %s for %s\n", event_name(sym_evsel
), symbol
->name
);
364 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter
);
366 pthread_mutex_lock(&syme
->src
->lock
);
367 line
= syme
->src
->source
;
369 total
+= line
->count
[sym_counter
];
373 line
= syme
->src
->source
;
377 if (!line_queue_count
)
381 if (line
->count
[sym_counter
])
382 pcnt
= 100.0 * line
->count
[sym_counter
] / (float)total
;
383 if (pcnt
>= (float)sym_pcnt_filter
) {
384 if (displayed
<= print_entries
)
385 show_lines(line_queue
, line_queue_count
, total
);
387 displayed
+= line_queue_count
;
388 line_queue_count
= 0;
390 } else if (line_queue_count
> TRACE_COUNT
) {
391 line_queue
= line_queue
->next
;
395 line
->count
[sym_counter
] = zero
? 0 : line
->count
[sym_counter
] * 7 / 8;
398 pthread_mutex_unlock(&syme
->src
->lock
);
400 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
404 * Symbols will be added here in event__process_sample and will get out
407 static LIST_HEAD(active_symbols
);
408 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
411 * Ordering weight: count-1 * count-2 * ... / count-n
413 static double sym_weight(const struct sym_entry
*sym
)
415 double weight
= sym
->snap_count
;
418 if (!display_weighted
)
421 for (counter
= 1; counter
< evsel_list
->nr_entries
- 1; counter
++)
422 weight
*= sym
->count
[counter
];
424 weight
/= (sym
->count
[counter
] + 1);
430 static long kernel_samples
, us_samples
;
431 static long exact_samples
;
432 static long guest_us_samples
, guest_kernel_samples
;
433 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
435 static void __list_insert_active_sym(struct sym_entry
*syme
)
437 list_add(&syme
->node
, &active_symbols
);
440 static void list_remove_active_sym(struct sym_entry
*syme
)
442 pthread_mutex_lock(&active_symbols_lock
);
443 list_del_init(&syme
->node
);
444 pthread_mutex_unlock(&active_symbols_lock
);
447 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
449 struct rb_node
**p
= &tree
->rb_node
;
450 struct rb_node
*parent
= NULL
;
451 struct sym_entry
*iter
;
455 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
457 if (se
->weight
> iter
->weight
)
463 rb_link_node(&se
->rb_node
, parent
, p
);
464 rb_insert_color(&se
->rb_node
, tree
);
467 static void print_sym_table(void)
470 struct perf_evsel
*counter
;
471 int snap
= !display_weighted
? sym_counter
: 0;
472 float samples_per_sec
= samples
/delay_secs
;
473 float ksamples_per_sec
= kernel_samples
/delay_secs
;
474 float us_samples_per_sec
= (us_samples
)/delay_secs
;
475 float guest_kernel_samples_per_sec
= (guest_kernel_samples
)/delay_secs
;
476 float guest_us_samples_per_sec
= (guest_us_samples
)/delay_secs
;
477 float esamples_percent
= (100.0*exact_samples
)/samples
;
478 float sum_ksamples
= 0.0;
479 struct sym_entry
*syme
, *n
;
480 struct rb_root tmp
= RB_ROOT
;
482 int sym_width
= 0, dso_width
= 0, dso_short_width
= 0;
483 const int win_width
= winsize
.ws_col
- 1;
485 samples
= us_samples
= kernel_samples
= exact_samples
= 0;
486 guest_kernel_samples
= guest_us_samples
= 0;
488 /* Sort the active symbols */
489 pthread_mutex_lock(&active_symbols_lock
);
490 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
491 pthread_mutex_unlock(&active_symbols_lock
);
493 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
494 syme
->snap_count
= syme
->count
[snap
];
495 if (syme
->snap_count
!= 0) {
497 if ((hide_user_symbols
&&
498 syme
->origin
== PERF_RECORD_MISC_USER
) ||
499 (hide_kernel_symbols
&&
500 syme
->origin
== PERF_RECORD_MISC_KERNEL
)) {
501 list_remove_active_sym(syme
);
504 syme
->weight
= sym_weight(syme
);
505 rb_insert_active_sym(&tmp
, syme
);
506 sum_ksamples
+= syme
->snap_count
;
508 for (j
= 0; j
< evsel_list
->nr_entries
; j
++)
509 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
511 list_remove_active_sym(syme
);
516 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
518 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%"
521 100.0 - (100.0 * ((samples_per_sec
- ksamples_per_sec
) /
525 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%"
526 " guest kernel:%4.1f%% guest us:%4.1f%%"
529 100.0 - (100.0 * ((samples_per_sec
-ksamples_per_sec
) /
531 100.0 - (100.0 * ((samples_per_sec
-us_samples_per_sec
) /
533 100.0 - (100.0 * ((samples_per_sec
-
534 guest_kernel_samples_per_sec
) /
536 100.0 - (100.0 * ((samples_per_sec
-
537 guest_us_samples_per_sec
) /
542 if (evsel_list
->nr_entries
== 1 || !display_weighted
) {
543 struct perf_evsel
*first
;
544 first
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
545 printf("%" PRIu64
, (uint64_t)first
->attr
.sample_period
);
552 if (!display_weighted
)
553 printf("%s", event_name(sym_evsel
));
554 else list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
558 printf("%s", event_name(counter
));
563 if (target_pid
!= -1)
564 printf(" (target_pid: %d", target_pid
);
565 else if (target_tid
!= -1)
566 printf(" (target_tid: %d", target_tid
);
571 printf(", CPU%s: %s)\n", cpus
->nr
> 1 ? "s" : "", cpu_list
);
573 if (target_tid
!= -1)
576 printf(", %d CPU%s)\n", cpus
->nr
, cpus
->nr
> 1 ? "s" : "");
579 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
581 if (sym_filter_entry
) {
582 show_details(sym_filter_entry
);
587 * Find the longest symbol name that will be displayed
589 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
590 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
591 if (++printed
> print_entries
||
592 (int)syme
->snap_count
< count_filter
)
595 if (syme
->map
->dso
->long_name_len
> dso_width
)
596 dso_width
= syme
->map
->dso
->long_name_len
;
598 if (syme
->map
->dso
->short_name_len
> dso_short_width
)
599 dso_short_width
= syme
->map
->dso
->short_name_len
;
601 if (syme
->name_len
> sym_width
)
602 sym_width
= syme
->name_len
;
607 if (sym_width
+ dso_width
> winsize
.ws_col
- 29) {
608 dso_width
= dso_short_width
;
609 if (sym_width
+ dso_width
> winsize
.ws_col
- 29)
610 sym_width
= winsize
.ws_col
- dso_width
- 29;
613 if (evsel_list
->nr_entries
== 1)
614 printf(" samples pcnt");
616 printf(" weight samples pcnt");
620 printf(" %-*.*s DSO\n", sym_width
, sym_width
, "function");
621 printf(" %s _______ _____",
622 evsel_list
->nr_entries
== 1 ? " " : "______");
624 printf(" ________________");
625 printf(" %-*.*s", sym_width
, sym_width
, graph_line
);
626 printf(" %-*.*s", dso_width
, dso_width
, graph_line
);
629 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
633 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
634 sym
= sym_entry__symbol(syme
);
635 if (++printed
> print_entries
|| (int)syme
->snap_count
< count_filter
)
638 pcnt
= 100.0 - (100.0 * ((sum_ksamples
- syme
->snap_count
) /
641 if (evsel_list
->nr_entries
== 1 || !display_weighted
)
642 printf("%20.2f ", syme
->weight
);
644 printf("%9.1f %10ld ", syme
->weight
, syme
->snap_count
);
646 percent_color_fprintf(stdout
, "%4.1f%%", pcnt
);
648 printf(" %016" PRIx64
, sym
->start
);
649 printf(" %-*.*s", sym_width
, sym_width
, sym
->name
);
650 printf(" %-*.*s\n", dso_width
, dso_width
,
651 dso_width
>= syme
->map
->dso
->long_name_len
?
652 syme
->map
->dso
->long_name
:
653 syme
->map
->dso
->short_name
);
657 static void prompt_integer(int *target
, const char *msg
)
659 char *buf
= malloc(0), *p
;
663 fprintf(stdout
, "\n%s: ", msg
);
664 if (getline(&buf
, &dummy
, stdin
) < 0)
667 p
= strchr(buf
, '\n');
677 tmp
= strtoul(buf
, NULL
, 10);
683 static void prompt_percent(int *target
, const char *msg
)
687 prompt_integer(&tmp
, msg
);
688 if (tmp
>= 0 && tmp
<= 100)
692 static void prompt_symbol(struct sym_entry
**target
, const char *msg
)
694 char *buf
= malloc(0), *p
;
695 struct sym_entry
*syme
= *target
, *n
, *found
= NULL
;
698 /* zero counters of active symbol */
700 pthread_mutex_lock(&syme
->src
->lock
);
701 __zero_source_counters(syme
);
703 pthread_mutex_unlock(&syme
->src
->lock
);
706 fprintf(stdout
, "\n%s: ", msg
);
707 if (getline(&buf
, &dummy
, stdin
) < 0)
710 p
= strchr(buf
, '\n');
714 pthread_mutex_lock(&active_symbols_lock
);
715 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
716 pthread_mutex_unlock(&active_symbols_lock
);
718 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
719 struct symbol
*sym
= sym_entry__symbol(syme
);
721 if (!strcmp(buf
, sym
->name
)) {
728 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
738 static void print_mapped_keys(void)
742 if (sym_filter_entry
) {
743 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
747 fprintf(stdout
, "\nMapped keys:\n");
748 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", delay_secs
);
749 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", print_entries
);
751 if (evsel_list
->nr_entries
> 1)
752 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel
));
754 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", count_filter
);
756 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter
);
757 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
758 fprintf(stdout
, "\t[S] stop annotation.\n");
760 if (evsel_list
->nr_entries
> 1)
761 fprintf(stdout
, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted
? 1 : 0);
764 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
765 hide_kernel_symbols
? "yes" : "no");
767 "\t[U] hide user symbols. \t(%s)\n",
768 hide_user_symbols
? "yes" : "no");
769 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", zero
? 1 : 0);
770 fprintf(stdout
, "\t[qQ] quit.\n");
773 static int key_mapped(int c
)
790 return evsel_list
->nr_entries
> 1 ? 1 : 0;
798 static void handle_keypress(struct perf_session
*session
, int c
)
800 if (!key_mapped(c
)) {
801 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
802 struct termios tc
, save
;
805 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
810 tc
.c_lflag
&= ~(ICANON
| ECHO
);
813 tcsetattr(0, TCSANOW
, &tc
);
815 poll(&stdin_poll
, 1, -1);
818 tcsetattr(0, TCSAFLUSH
, &save
);
825 prompt_integer(&delay_secs
, "Enter display delay");
830 prompt_integer(&print_entries
, "Enter display entries (lines)");
831 if (print_entries
== 0) {
832 sig_winch_handler(SIGWINCH
);
833 signal(SIGWINCH
, sig_winch_handler
);
835 signal(SIGWINCH
, SIG_DFL
);
838 if (evsel_list
->nr_entries
> 1) {
839 fprintf(stderr
, "\nAvailable events:");
841 list_for_each_entry(sym_evsel
, &evsel_list
->entries
, node
)
842 fprintf(stderr
, "\n\t%d %s", sym_evsel
->idx
, event_name(sym_evsel
));
844 prompt_integer(&sym_counter
, "Enter details event counter");
846 if (sym_counter
>= evsel_list
->nr_entries
) {
847 sym_evsel
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
849 fprintf(stderr
, "Sorry, no such event, using %s.\n", event_name(sym_evsel
));
853 list_for_each_entry(sym_evsel
, &evsel_list
->entries
, node
)
854 if (sym_evsel
->idx
== sym_counter
)
856 } else sym_counter
= 0;
859 prompt_integer(&count_filter
, "Enter display event count filter");
862 prompt_percent(&sym_pcnt_filter
, "Enter details display event filter (percent)");
865 hide_kernel_symbols
= !hide_kernel_symbols
;
869 printf("exiting.\n");
871 perf_session__fprintf_dsos(session
, stderr
);
874 prompt_symbol(&sym_filter_entry
, "Enter details symbol");
877 if (!sym_filter_entry
)
880 struct sym_entry
*syme
= sym_filter_entry
;
882 pthread_mutex_lock(&syme
->src
->lock
);
883 sym_filter_entry
= NULL
;
884 __zero_source_counters(syme
);
885 pthread_mutex_unlock(&syme
->src
->lock
);
889 hide_user_symbols
= !hide_user_symbols
;
892 display_weighted
= ~display_weighted
;
902 static void *display_thread(void *arg __used
)
904 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
905 struct termios tc
, save
;
907 struct perf_session
*session
= (struct perf_session
*) arg
;
911 tc
.c_lflag
&= ~(ICANON
| ECHO
);
916 delay_msecs
= delay_secs
* 1000;
917 tcsetattr(0, TCSANOW
, &tc
);
923 } while (!poll(&stdin_poll
, 1, delay_msecs
) == 1);
926 tcsetattr(0, TCSAFLUSH
, &save
);
928 handle_keypress(session
, c
);
934 /* Tag samples to be skipped. */
935 static const char *skip_symbols
[] = {
942 "mwait_idle_with_hints",
944 "ppc64_runlatch_off",
945 "pseries_dedicated_idle_sleep",
949 static int symbol_filter(struct map
*map
, struct symbol
*sym
)
951 struct sym_entry
*syme
;
952 const char *name
= sym
->name
;
956 * ppc64 uses function descriptors and appends a '.' to the
957 * start of every instruction address. Remove it.
962 if (!strcmp(name
, "_text") ||
963 !strcmp(name
, "_etext") ||
964 !strcmp(name
, "_sinittext") ||
965 !strncmp("init_module", name
, 11) ||
966 !strncmp("cleanup_module", name
, 14) ||
967 strstr(name
, "_text_start") ||
968 strstr(name
, "_text_end"))
971 syme
= symbol__priv(sym
);
975 if (!sym_filter_entry
&& sym_filter
&& !strcmp(name
, sym_filter
)) {
976 /* schedule initial sym_filter_entry setup */
977 sym_filter_entry_sched
= syme
;
981 for (i
= 0; skip_symbols
[i
]; i
++) {
982 if (!strcmp(skip_symbols
[i
], name
)) {
989 syme
->name_len
= strlen(sym
->name
);
994 static void event__process_sample(const event_t
*self
,
995 struct sample_data
*sample
,
996 struct perf_session
*session
)
998 u64 ip
= self
->ip
.ip
;
999 struct sym_entry
*syme
;
1000 struct addr_location al
;
1001 struct machine
*machine
;
1002 u8 origin
= self
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
1007 case PERF_RECORD_MISC_USER
:
1009 if (hide_user_symbols
)
1011 machine
= perf_session__find_host_machine(session
);
1013 case PERF_RECORD_MISC_KERNEL
:
1015 if (hide_kernel_symbols
)
1017 machine
= perf_session__find_host_machine(session
);
1019 case PERF_RECORD_MISC_GUEST_KERNEL
:
1020 ++guest_kernel_samples
;
1021 machine
= perf_session__find_machine(session
, self
->ip
.pid
);
1023 case PERF_RECORD_MISC_GUEST_USER
:
1026 * TODO: we don't process guest user from host side
1027 * except simple counting.
1034 if (!machine
&& perf_guest
) {
1035 pr_err("Can't find guest [%d]'s kernel information\n",
1040 if (self
->header
.misc
& PERF_RECORD_MISC_EXACT_IP
)
1043 if (event__preprocess_sample(self
, session
, &al
, sample
,
1044 symbol_filter
) < 0 ||
1048 if (al
.sym
== NULL
) {
1050 * As we do lazy loading of symtabs we only will know if the
1051 * specified vmlinux file is invalid when we actually have a
1052 * hit in kernel space and then try to load it. So if we get
1053 * here and there are _no_ symbols in the DSO backing the
1054 * kernel map, bail out.
1056 * We may never get here, for instance, if we use -K/
1057 * --hide-kernel-symbols, even if the user specifies an
1058 * invalid --vmlinux ;-)
1060 if (al
.map
== machine
->vmlinux_maps
[MAP__FUNCTION
] &&
1061 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
1062 pr_err("The %s file can't be used\n",
1063 symbol_conf
.vmlinux_name
);
1070 /* let's see, whether we need to install initial sym_filter_entry */
1071 if (sym_filter_entry_sched
) {
1072 sym_filter_entry
= sym_filter_entry_sched
;
1073 sym_filter_entry_sched
= NULL
;
1074 if (parse_source(sym_filter_entry
) < 0) {
1075 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
1077 pr_err("Can't annotate %s", sym
->name
);
1078 if (sym_filter_entry
->map
->dso
->origin
== DSO__ORIG_KERNEL
) {
1079 pr_err(": No vmlinux file was found in the path:\n");
1080 machine__fprintf_vmlinux_path(machine
, stderr
);
1087 syme
= symbol__priv(al
.sym
);
1089 struct perf_evsel
*evsel
;
1091 syme
->origin
= origin
;
1092 evsel
= perf_evlist__id2evsel(evsel_list
, sample
->id
);
1093 assert(evsel
!= NULL
);
1094 syme
->count
[evsel
->idx
]++;
1095 record_precise_ip(syme
, evsel
->idx
, ip
);
1096 pthread_mutex_lock(&active_symbols_lock
);
1097 if (list_empty(&syme
->node
) || !syme
->node
.next
)
1098 __list_insert_active_sym(syme
);
1099 pthread_mutex_unlock(&active_symbols_lock
);
1103 static void perf_session__mmap_read_cpu(struct perf_session
*self
, int cpu
)
1105 struct sample_data sample
;
1108 while ((event
= perf_evlist__read_on_cpu(evsel_list
, cpu
)) != NULL
) {
1109 event__parse_sample(event
, self
, &sample
);
1111 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
1112 event__process_sample(event
, &sample
, self
);
1114 event__process(event
, &sample
, self
);
1118 static void perf_session__mmap_read(struct perf_session
*self
)
1122 for (i
= 0; i
< cpus
->nr
; i
++)
1123 perf_session__mmap_read_cpu(self
, i
);
1126 static void start_counters(struct perf_evlist
*evlist
)
1128 struct perf_evsel
*counter
;
1130 list_for_each_entry(counter
, &evlist
->entries
, node
) {
1131 struct perf_event_attr
*attr
= &counter
->attr
;
1133 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
1136 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
1138 attr
->sample_freq
= freq
;
1141 if (evlist
->nr_entries
> 1) {
1142 attr
->sample_type
|= PERF_SAMPLE_ID
;
1143 attr
->read_format
|= PERF_FORMAT_ID
;
1148 if (perf_evsel__open(counter
, cpus
, threads
, group
, inherit
) < 0) {
1151 if (err
== EPERM
|| err
== EACCES
)
1152 die("Permission error - are you root?\n"
1153 "\t Consider tweaking"
1154 " /proc/sys/kernel/perf_event_paranoid.\n");
1156 * If it's cycles then fall back to hrtimer
1157 * based cpu-clock-tick sw counter, which
1158 * is always available even if no PMU support:
1160 if (attr
->type
== PERF_TYPE_HARDWARE
&&
1161 attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
1164 warning(" ... trying to fall back to cpu-clock-ticks\n");
1166 attr
->type
= PERF_TYPE_SOFTWARE
;
1167 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
1171 error("sys_perf_event_open() syscall returned with %d "
1172 "(%s). /bin/dmesg may provide additional information.\n",
1173 err
, strerror(err
));
1174 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1179 if (perf_evlist__mmap(evlist
, cpus
, threads
, mmap_pages
, true) < 0)
1180 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
1183 static int __cmd_top(void)
1186 struct perf_evsel
*first
;
1189 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1190 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1192 struct perf_session
*session
= perf_session__new(NULL
, O_WRONLY
, false, false, NULL
);
1193 if (session
== NULL
)
1196 if (target_tid
!= -1)
1197 event__synthesize_thread(target_tid
, event__process
, session
);
1199 event__synthesize_threads(event__process
, session
);
1201 start_counters(evsel_list
);
1202 first
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
1203 perf_session__set_sample_type(session
, first
->attr
.sample_type
);
1205 /* Wait for a minimal set of events before starting the snapshot */
1206 poll(evsel_list
->pollfd
, evsel_list
->nr_fds
, 100);
1208 perf_session__mmap_read(session
);
1210 if (pthread_create(&thread
, NULL
, display_thread
, session
)) {
1211 printf("Could not create display thread.\n");
1215 if (realtime_prio
) {
1216 struct sched_param param
;
1218 param
.sched_priority
= realtime_prio
;
1219 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1220 printf("Could not set realtime priority.\n");
1228 perf_session__mmap_read(session
);
1230 if (hits
== samples
)
1231 ret
= poll(evsel_list
->pollfd
, evsel_list
->nr_fds
, 100);
1237 static const char * const top_usage
[] = {
1238 "perf top [<options>]",
1242 static const struct option options
[] = {
1243 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1244 "event selector. use 'perf list' to list available events",
1246 OPT_INTEGER('c', "count", &default_interval
,
1247 "event period to sample"),
1248 OPT_INTEGER('p', "pid", &target_pid
,
1249 "profile events on existing process id"),
1250 OPT_INTEGER('t', "tid", &target_tid
,
1251 "profile events on existing thread id"),
1252 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1253 "system-wide collection from all CPUs"),
1254 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
1255 "list of cpus to monitor"),
1256 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1257 "file", "vmlinux pathname"),
1258 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols
,
1259 "hide kernel symbols"),
1260 OPT_UINTEGER('m', "mmap-pages", &mmap_pages
, "number of mmap data pages"),
1261 OPT_INTEGER('r', "realtime", &realtime_prio
,
1262 "collect data with this RT SCHED_FIFO priority"),
1263 OPT_INTEGER('d', "delay", &delay_secs
,
1264 "number of seconds to delay between refreshes"),
1265 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
1266 "dump the symbol table used for profiling"),
1267 OPT_INTEGER('f', "count-filter", &count_filter
,
1268 "only display functions with more events than this"),
1269 OPT_BOOLEAN('g', "group", &group
,
1270 "put the counters into a counter group"),
1271 OPT_BOOLEAN('i', "inherit", &inherit
,
1272 "child tasks inherit counters"),
1273 OPT_STRING('s', "sym-annotate", &sym_filter
, "symbol name",
1274 "symbol to annotate"),
1275 OPT_BOOLEAN('z', "zero", &zero
,
1276 "zero history across updates"),
1277 OPT_INTEGER('F', "freq", &freq
,
1278 "profile at this frequency"),
1279 OPT_INTEGER('E', "entries", &print_entries
,
1280 "display this many functions"),
1281 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols
,
1282 "hide user symbols"),
1283 OPT_INCR('v', "verbose", &verbose
,
1284 "be more verbose (show counter open errors, etc)"),
1288 int cmd_top(int argc
, const char **argv
, const char *prefix __used
)
1290 struct perf_evsel
*pos
;
1291 int status
= -ENOMEM
;
1293 evsel_list
= perf_evlist__new();
1294 if (evsel_list
== NULL
)
1297 page_size
= sysconf(_SC_PAGE_SIZE
);
1299 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1301 usage_with_options(top_usage
, options
);
1303 if (target_pid
!= -1)
1304 target_tid
= target_pid
;
1306 threads
= thread_map__new(target_pid
, target_tid
);
1307 if (threads
== NULL
) {
1308 pr_err("Problems finding threads of monitor\n");
1309 usage_with_options(top_usage
, options
);
1312 /* CPU and PID are mutually exclusive */
1313 if (target_tid
> 0 && cpu_list
) {
1314 printf("WARNING: PID switch overriding CPU\n");
1319 if (!evsel_list
->nr_entries
&&
1320 perf_evlist__add_default(evsel_list
) < 0) {
1321 pr_err("Not enough memory for event selector list\n");
1329 * User specified count overrides default frequency.
1331 if (default_interval
)
1334 default_interval
= freq
;
1336 fprintf(stderr
, "frequency and count are zero, aborting\n");
1340 if (target_tid
!= -1)
1341 cpus
= cpu_map__dummy_new();
1343 cpus
= cpu_map__new(cpu_list
);
1346 usage_with_options(top_usage
, options
);
1348 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1349 if (perf_evsel__alloc_fd(pos
, cpus
->nr
, threads
->nr
) < 0)
1352 * Fill in the ones not specifically initialized via -c:
1354 if (pos
->attr
.sample_period
)
1357 pos
->attr
.sample_period
= default_interval
;
1360 if (perf_evlist__alloc_pollfd(evsel_list
, cpus
->nr
, threads
->nr
) < 0 ||
1361 perf_evlist__alloc_mmap(evsel_list
, cpus
->nr
) < 0)
1364 sym_evsel
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
1366 symbol_conf
.priv_size
= (sizeof(struct sym_entry
) +
1367 (evsel_list
->nr_entries
+ 1) * sizeof(unsigned long));
1369 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1370 if (symbol__init() < 0)
1373 get_term_dimensions(&winsize
);
1374 if (print_entries
== 0) {
1375 update_print_entries(&winsize
);
1376 signal(SIGWINCH
, sig_winch_handler
);
1379 status
= __cmd_top();
1381 perf_evlist__delete(evsel_list
);