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/evsel.h"
25 #include "util/session.h"
26 #include "util/symbol.h"
27 #include "util/thread.h"
28 #include "util/util.h"
29 #include <linux/rbtree.h>
30 #include "util/parse-options.h"
31 #include "util/parse-events.h"
32 #include "util/cpumap.h"
33 #include "util/xyarray.h"
35 #include "util/debug.h"
50 #include <sys/syscall.h>
51 #include <sys/ioctl.h>
53 #include <sys/prctl.h>
58 #include <linux/unistd.h>
59 #include <linux/types.h>
61 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
63 static bool system_wide
= false;
65 static int default_interval
= 0;
67 static int count_filter
= 5;
68 static int print_entries
;
70 static int target_pid
= -1;
71 static int target_tid
= -1;
72 static struct thread_map
*threads
;
73 static bool inherit
= false;
74 static struct cpu_map
*cpus
;
75 static int realtime_prio
= 0;
76 static bool group
= false;
77 static unsigned int page_size
;
78 static unsigned int mmap_pages
= 16;
79 static int freq
= 1000; /* 1 KHz */
81 static int delay_secs
= 2;
82 static bool zero
= false;
83 static bool dump_symtab
= false;
85 static bool hide_kernel_symbols
= false;
86 static bool hide_user_symbols
= false;
87 static struct winsize winsize
;
95 unsigned long count
[MAX_COUNTERS
];
97 struct source_line
*next
;
100 static const char *sym_filter
= NULL
;
101 struct sym_entry
*sym_filter_entry
= NULL
;
102 struct sym_entry
*sym_filter_entry_sched
= NULL
;
103 static int sym_pcnt_filter
= 5;
104 static int sym_counter
= 0;
105 static struct perf_evsel
*sym_evsel
= NULL
;
106 static int display_weighted
= -1;
107 static const char *cpu_list
;
113 struct sym_entry_source
{
114 struct source_line
*source
;
115 struct source_line
*lines
;
116 struct source_line
**lines_tail
;
117 pthread_mutex_t lock
;
121 struct rb_node rb_node
;
122 struct list_head node
;
123 unsigned long snap_count
;
129 struct sym_entry_source
*src
;
130 unsigned long count
[0];
137 static inline struct symbol
*sym_entry__symbol(struct sym_entry
*self
)
139 return ((void *)self
) + symbol_conf
.priv_size
;
142 void get_term_dimensions(struct winsize
*ws
)
144 char *s
= getenv("LINES");
147 ws
->ws_row
= atoi(s
);
148 s
= getenv("COLUMNS");
150 ws
->ws_col
= atoi(s
);
151 if (ws
->ws_row
&& ws
->ws_col
)
156 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
157 ws
->ws_row
&& ws
->ws_col
)
164 static void update_print_entries(struct winsize
*ws
)
166 print_entries
= ws
->ws_row
;
168 if (print_entries
> 9)
172 static void sig_winch_handler(int sig __used
)
174 get_term_dimensions(&winsize
);
175 update_print_entries(&winsize
);
178 static int parse_source(struct sym_entry
*syme
)
181 struct sym_entry_source
*source
;
184 char command
[PATH_MAX
*2];
191 sym
= sym_entry__symbol(syme
);
195 * We can't annotate with just /proc/kallsyms
197 if (map
->dso
->origin
== DSO__ORIG_KERNEL
)
200 if (syme
->src
== NULL
) {
201 syme
->src
= zalloc(sizeof(*source
));
202 if (syme
->src
== NULL
)
204 pthread_mutex_init(&syme
->src
->lock
, NULL
);
210 pthread_mutex_lock(&source
->lock
);
213 path
= map
->dso
->long_name
;
215 len
= sym
->end
- sym
->start
;
218 "objdump --start-address=%#0*" PRIx64
" --stop-address=%#0*" PRIx64
" -dS %s",
219 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->start
),
220 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->end
), path
);
222 file
= popen(command
, "r");
226 pthread_mutex_lock(&source
->lock
);
227 source
->lines_tail
= &source
->lines
;
228 while (!feof(file
)) {
229 struct source_line
*src
;
233 src
= malloc(sizeof(struct source_line
));
235 memset(src
, 0, sizeof(struct source_line
));
237 if (getline(&src
->line
, &dummy
, file
) < 0)
242 c
= strchr(src
->line
, '\n');
247 *source
->lines_tail
= src
;
248 source
->lines_tail
= &src
->next
;
250 src
->eip
= strtoull(src
->line
, &sep
, 16);
252 src
->eip
= map__objdump_2ip(map
, src
->eip
);
253 else /* this line has no ip info (e.g. source line) */
258 sym_filter_entry
= syme
;
259 pthread_mutex_unlock(&source
->lock
);
263 static void __zero_source_counters(struct sym_entry
*syme
)
266 struct source_line
*line
;
268 line
= syme
->src
->lines
;
270 for (i
= 0; i
< nr_counters
; i
++)
276 static void record_precise_ip(struct sym_entry
*syme
, int counter
, u64 ip
)
278 struct source_line
*line
;
280 if (syme
!= sym_filter_entry
)
283 if (pthread_mutex_trylock(&syme
->src
->lock
))
286 if (syme
->src
== NULL
|| syme
->src
->source
== NULL
)
289 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
290 /* skip lines without IP info */
293 if (line
->eip
== ip
) {
294 line
->count
[counter
]++;
301 pthread_mutex_unlock(&syme
->src
->lock
);
304 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
306 static void lookup_sym_source(struct sym_entry
*syme
)
308 struct symbol
*symbol
= sym_entry__symbol(syme
);
309 struct source_line
*line
;
310 char pattern
[PATTERN_LEN
+ 1];
312 sprintf(pattern
, "%0*" PRIx64
" <", BITS_PER_LONG
/ 4,
313 map__rip_2objdump(syme
->map
, symbol
->start
));
315 pthread_mutex_lock(&syme
->src
->lock
);
316 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
317 if (memcmp(line
->line
, pattern
, PATTERN_LEN
) == 0) {
318 syme
->src
->source
= line
;
322 pthread_mutex_unlock(&syme
->src
->lock
);
325 static void show_lines(struct source_line
*queue
, int count
, int total
)
328 struct source_line
*line
;
331 for (i
= 0; i
< count
; i
++) {
332 float pcnt
= 100.0*(float)line
->count
[sym_counter
]/(float)total
;
334 printf("%8li %4.1f%%\t%s\n", line
->count
[sym_counter
], pcnt
, line
->line
);
339 #define TRACE_COUNT 3
341 static void show_details(struct sym_entry
*syme
)
343 struct symbol
*symbol
;
344 struct source_line
*line
;
345 struct source_line
*line_queue
= NULL
;
347 int line_queue_count
= 0, total
= 0, more
= 0;
352 if (!syme
->src
->source
)
353 lookup_sym_source(syme
);
355 if (!syme
->src
->source
)
358 symbol
= sym_entry__symbol(syme
);
359 printf("Showing %s for %s\n", event_name(sym_evsel
), symbol
->name
);
360 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter
);
362 pthread_mutex_lock(&syme
->src
->lock
);
363 line
= syme
->src
->source
;
365 total
+= line
->count
[sym_counter
];
369 line
= syme
->src
->source
;
373 if (!line_queue_count
)
377 if (line
->count
[sym_counter
])
378 pcnt
= 100.0 * line
->count
[sym_counter
] / (float)total
;
379 if (pcnt
>= (float)sym_pcnt_filter
) {
380 if (displayed
<= print_entries
)
381 show_lines(line_queue
, line_queue_count
, total
);
383 displayed
+= line_queue_count
;
384 line_queue_count
= 0;
386 } else if (line_queue_count
> TRACE_COUNT
) {
387 line_queue
= line_queue
->next
;
391 line
->count
[sym_counter
] = zero
? 0 : line
->count
[sym_counter
] * 7 / 8;
394 pthread_mutex_unlock(&syme
->src
->lock
);
396 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
400 * Symbols will be added here in event__process_sample and will get out
403 static LIST_HEAD(active_symbols
);
404 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
407 * Ordering weight: count-1 * count-2 * ... / count-n
409 static double sym_weight(const struct sym_entry
*sym
)
411 double weight
= sym
->snap_count
;
414 if (!display_weighted
)
417 for (counter
= 1; counter
< nr_counters
-1; counter
++)
418 weight
*= sym
->count
[counter
];
420 weight
/= (sym
->count
[counter
] + 1);
426 static long kernel_samples
, us_samples
;
427 static long exact_samples
;
428 static long guest_us_samples
, guest_kernel_samples
;
429 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
431 static void __list_insert_active_sym(struct sym_entry
*syme
)
433 list_add(&syme
->node
, &active_symbols
);
436 static void list_remove_active_sym(struct sym_entry
*syme
)
438 pthread_mutex_lock(&active_symbols_lock
);
439 list_del_init(&syme
->node
);
440 pthread_mutex_unlock(&active_symbols_lock
);
443 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
445 struct rb_node
**p
= &tree
->rb_node
;
446 struct rb_node
*parent
= NULL
;
447 struct sym_entry
*iter
;
451 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
453 if (se
->weight
> iter
->weight
)
459 rb_link_node(&se
->rb_node
, parent
, p
);
460 rb_insert_color(&se
->rb_node
, tree
);
463 static void print_sym_table(void)
466 struct perf_evsel
*counter
;
467 int snap
= !display_weighted
? sym_counter
: 0;
468 float samples_per_sec
= samples
/delay_secs
;
469 float ksamples_per_sec
= kernel_samples
/delay_secs
;
470 float us_samples_per_sec
= (us_samples
)/delay_secs
;
471 float guest_kernel_samples_per_sec
= (guest_kernel_samples
)/delay_secs
;
472 float guest_us_samples_per_sec
= (guest_us_samples
)/delay_secs
;
473 float esamples_percent
= (100.0*exact_samples
)/samples
;
474 float sum_ksamples
= 0.0;
475 struct sym_entry
*syme
, *n
;
476 struct rb_root tmp
= RB_ROOT
;
478 int sym_width
= 0, dso_width
= 0, dso_short_width
= 0;
479 const int win_width
= winsize
.ws_col
- 1;
481 samples
= us_samples
= kernel_samples
= exact_samples
= 0;
482 guest_kernel_samples
= guest_us_samples
= 0;
484 /* Sort the active symbols */
485 pthread_mutex_lock(&active_symbols_lock
);
486 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
487 pthread_mutex_unlock(&active_symbols_lock
);
489 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
490 syme
->snap_count
= syme
->count
[snap
];
491 if (syme
->snap_count
!= 0) {
493 if ((hide_user_symbols
&&
494 syme
->origin
== PERF_RECORD_MISC_USER
) ||
495 (hide_kernel_symbols
&&
496 syme
->origin
== PERF_RECORD_MISC_KERNEL
)) {
497 list_remove_active_sym(syme
);
500 syme
->weight
= sym_weight(syme
);
501 rb_insert_active_sym(&tmp
, syme
);
502 sum_ksamples
+= syme
->snap_count
;
504 for (j
= 0; j
< nr_counters
; j
++)
505 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
507 list_remove_active_sym(syme
);
512 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
514 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%%"
517 100.0 - (100.0 * ((samples_per_sec
- ksamples_per_sec
) /
521 printf(" PerfTop:%8.0f irqs/sec kernel:%4.1f%% us:%4.1f%%"
522 " guest kernel:%4.1f%% guest us:%4.1f%%"
525 100.0 - (100.0 * ((samples_per_sec
-ksamples_per_sec
) /
527 100.0 - (100.0 * ((samples_per_sec
-us_samples_per_sec
) /
529 100.0 - (100.0 * ((samples_per_sec
-
530 guest_kernel_samples_per_sec
) /
532 100.0 - (100.0 * ((samples_per_sec
-
533 guest_us_samples_per_sec
) /
538 if (nr_counters
== 1 || !display_weighted
) {
539 struct perf_evsel
*first
;
540 first
= list_entry(evsel_list
.next
, struct perf_evsel
, node
);
541 printf("%" PRIu64
, (uint64_t)first
->attr
.sample_period
);
548 if (!display_weighted
)
549 printf("%s", event_name(sym_evsel
));
550 else list_for_each_entry(counter
, &evsel_list
, node
) {
554 printf("%s", event_name(counter
));
559 if (target_pid
!= -1)
560 printf(" (target_pid: %d", target_pid
);
561 else if (target_tid
!= -1)
562 printf(" (target_tid: %d", target_tid
);
567 printf(", CPU%s: %s)\n", cpus
->nr
> 1 ? "s" : "", cpu_list
);
569 if (target_tid
!= -1)
572 printf(", %d CPU%s)\n", cpus
->nr
, cpus
->nr
> 1 ? "s" : "");
575 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
577 if (sym_filter_entry
) {
578 show_details(sym_filter_entry
);
583 * Find the longest symbol name that will be displayed
585 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
586 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
587 if (++printed
> print_entries
||
588 (int)syme
->snap_count
< count_filter
)
591 if (syme
->map
->dso
->long_name_len
> dso_width
)
592 dso_width
= syme
->map
->dso
->long_name_len
;
594 if (syme
->map
->dso
->short_name_len
> dso_short_width
)
595 dso_short_width
= syme
->map
->dso
->short_name_len
;
597 if (syme
->name_len
> sym_width
)
598 sym_width
= syme
->name_len
;
603 if (sym_width
+ dso_width
> winsize
.ws_col
- 29) {
604 dso_width
= dso_short_width
;
605 if (sym_width
+ dso_width
> winsize
.ws_col
- 29)
606 sym_width
= winsize
.ws_col
- dso_width
- 29;
609 if (nr_counters
== 1)
610 printf(" samples pcnt");
612 printf(" weight samples pcnt");
616 printf(" %-*.*s DSO\n", sym_width
, sym_width
, "function");
617 printf(" %s _______ _____",
618 nr_counters
== 1 ? " " : "______");
620 printf(" ________________");
621 printf(" %-*.*s", sym_width
, sym_width
, graph_line
);
622 printf(" %-*.*s", dso_width
, dso_width
, graph_line
);
625 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
629 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
630 sym
= sym_entry__symbol(syme
);
631 if (++printed
> print_entries
|| (int)syme
->snap_count
< count_filter
)
634 pcnt
= 100.0 - (100.0 * ((sum_ksamples
- syme
->snap_count
) /
637 if (nr_counters
== 1 || !display_weighted
)
638 printf("%20.2f ", syme
->weight
);
640 printf("%9.1f %10ld ", syme
->weight
, syme
->snap_count
);
642 percent_color_fprintf(stdout
, "%4.1f%%", pcnt
);
644 printf(" %016" PRIx64
, sym
->start
);
645 printf(" %-*.*s", sym_width
, sym_width
, sym
->name
);
646 printf(" %-*.*s\n", dso_width
, dso_width
,
647 dso_width
>= syme
->map
->dso
->long_name_len
?
648 syme
->map
->dso
->long_name
:
649 syme
->map
->dso
->short_name
);
653 static void prompt_integer(int *target
, const char *msg
)
655 char *buf
= malloc(0), *p
;
659 fprintf(stdout
, "\n%s: ", msg
);
660 if (getline(&buf
, &dummy
, stdin
) < 0)
663 p
= strchr(buf
, '\n');
673 tmp
= strtoul(buf
, NULL
, 10);
679 static void prompt_percent(int *target
, const char *msg
)
683 prompt_integer(&tmp
, msg
);
684 if (tmp
>= 0 && tmp
<= 100)
688 static void prompt_symbol(struct sym_entry
**target
, const char *msg
)
690 char *buf
= malloc(0), *p
;
691 struct sym_entry
*syme
= *target
, *n
, *found
= NULL
;
694 /* zero counters of active symbol */
696 pthread_mutex_lock(&syme
->src
->lock
);
697 __zero_source_counters(syme
);
699 pthread_mutex_unlock(&syme
->src
->lock
);
702 fprintf(stdout
, "\n%s: ", msg
);
703 if (getline(&buf
, &dummy
, stdin
) < 0)
706 p
= strchr(buf
, '\n');
710 pthread_mutex_lock(&active_symbols_lock
);
711 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
712 pthread_mutex_unlock(&active_symbols_lock
);
714 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
715 struct symbol
*sym
= sym_entry__symbol(syme
);
717 if (!strcmp(buf
, sym
->name
)) {
724 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
734 static void print_mapped_keys(void)
738 if (sym_filter_entry
) {
739 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
743 fprintf(stdout
, "\nMapped keys:\n");
744 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", delay_secs
);
745 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", print_entries
);
748 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", event_name(sym_evsel
));
750 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", count_filter
);
752 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter
);
753 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
754 fprintf(stdout
, "\t[S] stop annotation.\n");
757 fprintf(stdout
, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted
? 1 : 0);
760 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
761 hide_kernel_symbols
? "yes" : "no");
763 "\t[U] hide user symbols. \t(%s)\n",
764 hide_user_symbols
? "yes" : "no");
765 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", zero
? 1 : 0);
766 fprintf(stdout
, "\t[qQ] quit.\n");
769 static int key_mapped(int c
)
786 return nr_counters
> 1 ? 1 : 0;
794 static void handle_keypress(struct perf_session
*session
, int c
)
796 if (!key_mapped(c
)) {
797 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
798 struct termios tc
, save
;
801 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
806 tc
.c_lflag
&= ~(ICANON
| ECHO
);
809 tcsetattr(0, TCSANOW
, &tc
);
811 poll(&stdin_poll
, 1, -1);
814 tcsetattr(0, TCSAFLUSH
, &save
);
821 prompt_integer(&delay_secs
, "Enter display delay");
826 prompt_integer(&print_entries
, "Enter display entries (lines)");
827 if (print_entries
== 0) {
828 sig_winch_handler(SIGWINCH
);
829 signal(SIGWINCH
, sig_winch_handler
);
831 signal(SIGWINCH
, SIG_DFL
);
834 if (nr_counters
> 1) {
835 fprintf(stderr
, "\nAvailable events:");
837 list_for_each_entry(sym_evsel
, &evsel_list
, node
)
838 fprintf(stderr
, "\n\t%d %s", sym_evsel
->idx
, event_name(sym_evsel
));
840 prompt_integer(&sym_counter
, "Enter details event counter");
842 if (sym_counter
>= nr_counters
) {
843 sym_evsel
= list_entry(evsel_list
.next
, struct perf_evsel
, node
);
845 fprintf(stderr
, "Sorry, no such event, using %s.\n", event_name(sym_evsel
));
849 list_for_each_entry(sym_evsel
, &evsel_list
, node
)
850 if (sym_evsel
->idx
== sym_counter
)
852 } else sym_counter
= 0;
855 prompt_integer(&count_filter
, "Enter display event count filter");
858 prompt_percent(&sym_pcnt_filter
, "Enter details display event filter (percent)");
861 hide_kernel_symbols
= !hide_kernel_symbols
;
865 printf("exiting.\n");
867 perf_session__fprintf_dsos(session
, stderr
);
870 prompt_symbol(&sym_filter_entry
, "Enter details symbol");
873 if (!sym_filter_entry
)
876 struct sym_entry
*syme
= sym_filter_entry
;
878 pthread_mutex_lock(&syme
->src
->lock
);
879 sym_filter_entry
= NULL
;
880 __zero_source_counters(syme
);
881 pthread_mutex_unlock(&syme
->src
->lock
);
885 hide_user_symbols
= !hide_user_symbols
;
888 display_weighted
= ~display_weighted
;
898 static void *display_thread(void *arg __used
)
900 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
901 struct termios tc
, save
;
903 struct perf_session
*session
= (struct perf_session
*) arg
;
907 tc
.c_lflag
&= ~(ICANON
| ECHO
);
912 delay_msecs
= delay_secs
* 1000;
913 tcsetattr(0, TCSANOW
, &tc
);
919 } while (!poll(&stdin_poll
, 1, delay_msecs
) == 1);
922 tcsetattr(0, TCSAFLUSH
, &save
);
924 handle_keypress(session
, c
);
930 /* Tag samples to be skipped. */
931 static const char *skip_symbols
[] = {
937 "mwait_idle_with_hints",
939 "ppc64_runlatch_off",
940 "pseries_dedicated_idle_sleep",
944 static int symbol_filter(struct map
*map
, struct symbol
*sym
)
946 struct sym_entry
*syme
;
947 const char *name
= sym
->name
;
951 * ppc64 uses function descriptors and appends a '.' to the
952 * start of every instruction address. Remove it.
957 if (!strcmp(name
, "_text") ||
958 !strcmp(name
, "_etext") ||
959 !strcmp(name
, "_sinittext") ||
960 !strncmp("init_module", name
, 11) ||
961 !strncmp("cleanup_module", name
, 14) ||
962 strstr(name
, "_text_start") ||
963 strstr(name
, "_text_end"))
966 syme
= symbol__priv(sym
);
970 if (!sym_filter_entry
&& sym_filter
&& !strcmp(name
, sym_filter
)) {
971 /* schedule initial sym_filter_entry setup */
972 sym_filter_entry_sched
= syme
;
976 for (i
= 0; skip_symbols
[i
]; i
++) {
977 if (!strcmp(skip_symbols
[i
], name
)) {
984 syme
->name_len
= strlen(sym
->name
);
989 static void event__process_sample(const event_t
*self
,
990 struct sample_data
*sample
,
991 struct perf_session
*session
,
992 struct perf_evsel
*evsel
)
994 u64 ip
= self
->ip
.ip
;
995 struct sym_entry
*syme
;
996 struct addr_location al
;
997 struct machine
*machine
;
998 u8 origin
= self
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
1003 case PERF_RECORD_MISC_USER
:
1005 if (hide_user_symbols
)
1007 machine
= perf_session__find_host_machine(session
);
1009 case PERF_RECORD_MISC_KERNEL
:
1011 if (hide_kernel_symbols
)
1013 machine
= perf_session__find_host_machine(session
);
1015 case PERF_RECORD_MISC_GUEST_KERNEL
:
1016 ++guest_kernel_samples
;
1017 machine
= perf_session__find_machine(session
, self
->ip
.pid
);
1019 case PERF_RECORD_MISC_GUEST_USER
:
1022 * TODO: we don't process guest user from host side
1023 * except simple counting.
1030 if (!machine
&& perf_guest
) {
1031 pr_err("Can't find guest [%d]'s kernel information\n",
1036 if (self
->header
.misc
& PERF_RECORD_MISC_EXACT_IP
)
1039 if (event__preprocess_sample(self
, session
, &al
, sample
,
1040 symbol_filter
) < 0 ||
1044 if (al
.sym
== NULL
) {
1046 * As we do lazy loading of symtabs we only will know if the
1047 * specified vmlinux file is invalid when we actually have a
1048 * hit in kernel space and then try to load it. So if we get
1049 * here and there are _no_ symbols in the DSO backing the
1050 * kernel map, bail out.
1052 * We may never get here, for instance, if we use -K/
1053 * --hide-kernel-symbols, even if the user specifies an
1054 * invalid --vmlinux ;-)
1056 if (al
.map
== machine
->vmlinux_maps
[MAP__FUNCTION
] &&
1057 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
1058 pr_err("The %s file can't be used\n",
1059 symbol_conf
.vmlinux_name
);
1066 /* let's see, whether we need to install initial sym_filter_entry */
1067 if (sym_filter_entry_sched
) {
1068 sym_filter_entry
= sym_filter_entry_sched
;
1069 sym_filter_entry_sched
= NULL
;
1070 if (parse_source(sym_filter_entry
) < 0) {
1071 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
1073 pr_err("Can't annotate %s", sym
->name
);
1074 if (sym_filter_entry
->map
->dso
->origin
== DSO__ORIG_KERNEL
) {
1075 pr_err(": No vmlinux file was found in the path:\n");
1076 machine__fprintf_vmlinux_path(machine
, stderr
);
1083 syme
= symbol__priv(al
.sym
);
1085 syme
->count
[evsel
->idx
]++;
1086 syme
->origin
= origin
;
1087 record_precise_ip(syme
, evsel
->idx
, ip
);
1088 pthread_mutex_lock(&active_symbols_lock
);
1089 if (list_empty(&syme
->node
) || !syme
->node
.next
)
1090 __list_insert_active_sym(syme
);
1091 pthread_mutex_unlock(&active_symbols_lock
);
1101 static int perf_evsel__alloc_mmap_per_thread(struct perf_evsel
*evsel
,
1102 int ncpus
, int nthreads
)
1104 evsel
->priv
= xyarray__new(ncpus
, nthreads
, sizeof(struct mmap_data
));
1105 return evsel
->priv
!= NULL
? 0 : -ENOMEM
;
1108 static void perf_evsel__free_mmap(struct perf_evsel
*evsel
)
1110 xyarray__delete(evsel
->priv
);
1114 static unsigned int mmap_read_head(struct mmap_data
*md
)
1116 struct perf_event_mmap_page
*pc
= md
->base
;
1119 head
= pc
->data_head
;
1125 static void perf_session__mmap_read_counter(struct perf_session
*self
,
1126 struct perf_evsel
*evsel
,
1127 int cpu
, int thread_idx
)
1129 struct xyarray
*mmap_array
= evsel
->priv
;
1130 struct mmap_data
*md
= xyarray__entry(mmap_array
, cpu
, thread_idx
);
1131 unsigned int head
= mmap_read_head(md
);
1132 unsigned int old
= md
->prev
;
1133 unsigned char *data
= md
->base
+ page_size
;
1134 struct sample_data sample
;
1138 * If we're further behind than half the buffer, there's a chance
1139 * the writer will bite our tail and mess up the samples under us.
1141 * If we somehow ended up ahead of the head, we got messed up.
1143 * In either case, truncate and restart at head.
1146 if (diff
> md
->mask
/ 2 || diff
< 0) {
1147 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
1150 * head points to a known good entry, start there.
1155 for (; old
!= head
;) {
1156 event_t
*event
= (event_t
*)&data
[old
& md
->mask
];
1160 size_t size
= event
->header
.size
;
1163 * Event straddles the mmap boundary -- header should always
1164 * be inside due to u64 alignment of output.
1166 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
1167 unsigned int offset
= old
;
1168 unsigned int len
= min(sizeof(*event
), size
), cpy
;
1169 void *dst
= &event_copy
;
1172 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
1173 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
1179 event
= &event_copy
;
1182 event__parse_sample(event
, self
, &sample
);
1183 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
1184 event__process_sample(event
, &sample
, self
, evsel
);
1186 event__process(event
, &sample
, self
);
1193 static struct pollfd
*event_array
;
1195 static void perf_session__mmap_read(struct perf_session
*self
)
1197 struct perf_evsel
*counter
;
1198 int i
, thread_index
;
1200 for (i
= 0; i
< cpus
->nr
; i
++) {
1201 list_for_each_entry(counter
, &evsel_list
, node
) {
1202 for (thread_index
= 0;
1203 thread_index
< threads
->nr
;
1205 perf_session__mmap_read_counter(self
,
1206 counter
, i
, thread_index
);
1215 static void start_counter(int i
, struct perf_evsel
*evsel
)
1217 struct xyarray
*mmap_array
= evsel
->priv
;
1218 struct mmap_data
*mm
;
1219 struct perf_event_attr
*attr
;
1223 if (target_tid
== -1)
1226 attr
= &evsel
->attr
;
1228 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
1231 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
1233 attr
->sample_freq
= freq
;
1236 attr
->inherit
= (cpu
< 0) && inherit
;
1239 for (thread_index
= 0; thread_index
< threads
->nr
; thread_index
++) {
1241 FD(evsel
, i
, thread_index
) = sys_perf_event_open(attr
,
1242 threads
->map
[thread_index
], cpu
, group_fd
, 0);
1244 if (FD(evsel
, i
, thread_index
) < 0) {
1247 if (err
== EPERM
|| err
== EACCES
)
1248 die("Permission error - are you root?\n"
1249 "\t Consider tweaking"
1250 " /proc/sys/kernel/perf_event_paranoid.\n");
1252 * If it's cycles then fall back to hrtimer
1253 * based cpu-clock-tick sw counter, which
1254 * is always available even if no PMU support:
1256 if (attr
->type
== PERF_TYPE_HARDWARE
1257 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
1260 warning(" ... trying to fall back to cpu-clock-ticks\n");
1262 attr
->type
= PERF_TYPE_SOFTWARE
;
1263 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
1267 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
1268 FD(evsel
, i
, thread_index
), strerror(err
));
1269 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1272 assert(FD(evsel
, i
, thread_index
) >= 0);
1273 fcntl(FD(evsel
, i
, thread_index
), F_SETFL
, O_NONBLOCK
);
1276 * First counter acts as the group leader:
1278 if (group
&& group_fd
== -1)
1279 group_fd
= FD(evsel
, i
, thread_index
);
1281 event_array
[nr_poll
].fd
= FD(evsel
, i
, thread_index
);
1282 event_array
[nr_poll
].events
= POLLIN
;
1285 mm
= xyarray__entry(mmap_array
, i
, thread_index
);
1287 mm
->mask
= mmap_pages
*page_size
- 1;
1288 mm
->base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
1289 PROT_READ
, MAP_SHARED
, FD(evsel
, i
, thread_index
), 0);
1290 if (mm
->base
== MAP_FAILED
)
1291 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
1295 static int __cmd_top(void)
1298 struct perf_evsel
*counter
;
1301 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1302 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1304 struct perf_session
*session
= perf_session__new(NULL
, O_WRONLY
, false, false, NULL
);
1305 if (session
== NULL
)
1308 if (target_tid
!= -1)
1309 event__synthesize_thread(target_tid
, event__process
, session
);
1311 event__synthesize_threads(event__process
, session
);
1313 for (i
= 0; i
< cpus
->nr
; i
++) {
1315 list_for_each_entry(counter
, &evsel_list
, node
)
1316 start_counter(i
, counter
);
1319 /* Wait for a minimal set of events before starting the snapshot */
1320 poll(&event_array
[0], nr_poll
, 100);
1322 perf_session__mmap_read(session
);
1324 if (pthread_create(&thread
, NULL
, display_thread
, session
)) {
1325 printf("Could not create display thread.\n");
1329 if (realtime_prio
) {
1330 struct sched_param param
;
1332 param
.sched_priority
= realtime_prio
;
1333 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1334 printf("Could not set realtime priority.\n");
1342 perf_session__mmap_read(session
);
1344 if (hits
== samples
)
1345 ret
= poll(event_array
, nr_poll
, 100);
1351 static const char * const top_usage
[] = {
1352 "perf top [<options>]",
1356 static const struct option options
[] = {
1357 OPT_CALLBACK('e', "event", NULL
, "event",
1358 "event selector. use 'perf list' to list available events",
1360 OPT_INTEGER('c', "count", &default_interval
,
1361 "event period to sample"),
1362 OPT_INTEGER('p', "pid", &target_pid
,
1363 "profile events on existing process id"),
1364 OPT_INTEGER('t', "tid", &target_tid
,
1365 "profile events on existing thread id"),
1366 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1367 "system-wide collection from all CPUs"),
1368 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
1369 "list of cpus to monitor"),
1370 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1371 "file", "vmlinux pathname"),
1372 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols
,
1373 "hide kernel symbols"),
1374 OPT_UINTEGER('m', "mmap-pages", &mmap_pages
, "number of mmap data pages"),
1375 OPT_INTEGER('r', "realtime", &realtime_prio
,
1376 "collect data with this RT SCHED_FIFO priority"),
1377 OPT_INTEGER('d', "delay", &delay_secs
,
1378 "number of seconds to delay between refreshes"),
1379 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
1380 "dump the symbol table used for profiling"),
1381 OPT_INTEGER('f', "count-filter", &count_filter
,
1382 "only display functions with more events than this"),
1383 OPT_BOOLEAN('g', "group", &group
,
1384 "put the counters into a counter group"),
1385 OPT_BOOLEAN('i', "inherit", &inherit
,
1386 "child tasks inherit counters"),
1387 OPT_STRING('s', "sym-annotate", &sym_filter
, "symbol name",
1388 "symbol to annotate"),
1389 OPT_BOOLEAN('z', "zero", &zero
,
1390 "zero history across updates"),
1391 OPT_INTEGER('F', "freq", &freq
,
1392 "profile at this frequency"),
1393 OPT_INTEGER('E', "entries", &print_entries
,
1394 "display this many functions"),
1395 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols
,
1396 "hide user symbols"),
1397 OPT_INCR('v', "verbose", &verbose
,
1398 "be more verbose (show counter open errors, etc)"),
1402 int cmd_top(int argc
, const char **argv
, const char *prefix __used
)
1404 struct perf_evsel
*pos
;
1405 int status
= -ENOMEM
;
1407 page_size
= sysconf(_SC_PAGE_SIZE
);
1409 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1411 usage_with_options(top_usage
, options
);
1413 if (target_pid
!= -1)
1414 target_tid
= target_pid
;
1416 threads
= thread_map__new(target_pid
, target_tid
);
1417 if (threads
== NULL
) {
1418 pr_err("Problems finding threads of monitor\n");
1419 usage_with_options(top_usage
, options
);
1422 event_array
= malloc((sizeof(struct pollfd
) *
1423 MAX_NR_CPUS
* MAX_COUNTERS
* threads
->nr
));
1427 /* CPU and PID are mutually exclusive */
1428 if (target_tid
> 0 && cpu_list
) {
1429 printf("WARNING: PID switch overriding CPU\n");
1434 if (!nr_counters
&& perf_evsel_list__create_default() < 0) {
1435 pr_err("Not enough memory for event selector list\n");
1443 * User specified count overrides default frequency.
1445 if (default_interval
)
1448 default_interval
= freq
;
1450 fprintf(stderr
, "frequency and count are zero, aborting\n");
1454 if (target_tid
!= -1)
1455 cpus
= cpu_map__dummy_new();
1457 cpus
= cpu_map__new(cpu_list
);
1460 usage_with_options(top_usage
, options
);
1462 list_for_each_entry(pos
, &evsel_list
, node
) {
1463 if (perf_evsel__alloc_mmap_per_thread(pos
, cpus
->nr
, threads
->nr
) < 0 ||
1464 perf_evsel__alloc_fd(pos
, cpus
->nr
, threads
->nr
) < 0)
1467 * Fill in the ones not specifically initialized via -c:
1469 if (pos
->attr
.sample_period
)
1472 pos
->attr
.sample_period
= default_interval
;
1475 sym_evsel
= list_entry(evsel_list
.next
, struct perf_evsel
, node
);
1477 symbol_conf
.priv_size
= (sizeof(struct sym_entry
) +
1478 (nr_counters
+ 1) * sizeof(unsigned long));
1480 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1481 if (symbol__init() < 0)
1484 get_term_dimensions(&winsize
);
1485 if (print_entries
== 0) {
1486 update_print_entries(&winsize
);
1487 signal(SIGWINCH
, sig_winch_handler
);
1490 status
= __cmd_top();
1492 list_for_each_entry(pos
, &evsel_list
, node
)
1493 perf_evsel__free_mmap(pos
);
1494 perf_evsel_list__delete();