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/session.h"
25 #include "util/symbol.h"
26 #include "util/thread.h"
27 #include "util/util.h"
28 #include <linux/rbtree.h>
29 #include "util/parse-options.h"
30 #include "util/parse-events.h"
31 #include "util/cpumap.h"
33 #include "util/debug.h"
47 #include <sys/syscall.h>
48 #include <sys/ioctl.h>
50 #include <sys/prctl.h>
55 #include <linux/unistd.h>
56 #include <linux/types.h>
58 static int *fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
60 static int system_wide
= 0;
62 static int default_interval
= 0;
64 static int count_filter
= 5;
65 static int print_entries
;
67 static int target_pid
= -1;
68 static int target_tid
= -1;
69 static pid_t
*all_tids
= NULL
;
70 static int thread_num
= 0;
71 static int inherit
= 0;
72 static int profile_cpu
= -1;
73 static int nr_cpus
= 0;
74 static unsigned int realtime_prio
= 0;
76 static unsigned int page_size
;
77 static unsigned int mmap_pages
= 16;
78 static int freq
= 1000; /* 1 KHz */
80 static int delay_secs
= 2;
82 static int dump_symtab
= 0;
84 static bool hide_kernel_symbols
= false;
85 static bool hide_user_symbols
= false;
86 static struct winsize winsize
;
94 unsigned long count
[MAX_COUNTERS
];
96 struct source_line
*next
;
99 static char *sym_filter
= NULL
;
100 struct sym_entry
*sym_filter_entry
= NULL
;
101 struct sym_entry
*sym_filter_entry_sched
= NULL
;
102 static int sym_pcnt_filter
= 5;
103 static int sym_counter
= 0;
104 static int display_weighted
= -1;
110 struct sym_entry_source
{
111 struct source_line
*source
;
112 struct source_line
*lines
;
113 struct source_line
**lines_tail
;
114 pthread_mutex_t lock
;
118 struct rb_node rb_node
;
119 struct list_head node
;
120 unsigned long snap_count
;
126 struct sym_entry_source
*src
;
127 unsigned long count
[0];
134 static inline struct symbol
*sym_entry__symbol(struct sym_entry
*self
)
136 return ((void *)self
) + symbol_conf
.priv_size
;
139 void get_term_dimensions(struct winsize
*ws
)
141 char *s
= getenv("LINES");
144 ws
->ws_row
= atoi(s
);
145 s
= getenv("COLUMNS");
147 ws
->ws_col
= atoi(s
);
148 if (ws
->ws_row
&& ws
->ws_col
)
153 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
154 ws
->ws_row
&& ws
->ws_col
)
161 static void update_print_entries(struct winsize
*ws
)
163 print_entries
= ws
->ws_row
;
165 if (print_entries
> 9)
169 static void sig_winch_handler(int sig __used
)
171 get_term_dimensions(&winsize
);
172 update_print_entries(&winsize
);
175 static int parse_source(struct sym_entry
*syme
)
178 struct sym_entry_source
*source
;
181 char command
[PATH_MAX
*2];
188 sym
= sym_entry__symbol(syme
);
192 * We can't annotate with just /proc/kallsyms
194 if (map
->dso
->origin
== DSO__ORIG_KERNEL
)
197 if (syme
->src
== NULL
) {
198 syme
->src
= zalloc(sizeof(*source
));
199 if (syme
->src
== NULL
)
201 pthread_mutex_init(&syme
->src
->lock
, NULL
);
207 pthread_mutex_lock(&source
->lock
);
210 path
= map
->dso
->long_name
;
212 len
= sym
->end
- sym
->start
;
215 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
216 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->start
),
217 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->end
), path
);
219 file
= popen(command
, "r");
223 pthread_mutex_lock(&source
->lock
);
224 source
->lines_tail
= &source
->lines
;
225 while (!feof(file
)) {
226 struct source_line
*src
;
230 src
= malloc(sizeof(struct source_line
));
232 memset(src
, 0, sizeof(struct source_line
));
234 if (getline(&src
->line
, &dummy
, file
) < 0)
239 c
= strchr(src
->line
, '\n');
244 *source
->lines_tail
= src
;
245 source
->lines_tail
= &src
->next
;
247 src
->eip
= strtoull(src
->line
, &sep
, 16);
249 src
->eip
= map__objdump_2ip(map
, src
->eip
);
250 else /* this line has no ip info (e.g. source line) */
255 sym_filter_entry
= syme
;
256 pthread_mutex_unlock(&source
->lock
);
260 static void __zero_source_counters(struct sym_entry
*syme
)
263 struct source_line
*line
;
265 line
= syme
->src
->lines
;
267 for (i
= 0; i
< nr_counters
; i
++)
273 static void record_precise_ip(struct sym_entry
*syme
, int counter
, u64 ip
)
275 struct source_line
*line
;
277 if (syme
!= sym_filter_entry
)
280 if (pthread_mutex_trylock(&syme
->src
->lock
))
283 if (syme
->src
== NULL
|| syme
->src
->source
== NULL
)
286 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
287 /* skip lines without IP info */
290 if (line
->eip
== ip
) {
291 line
->count
[counter
]++;
298 pthread_mutex_unlock(&syme
->src
->lock
);
301 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
303 static void lookup_sym_source(struct sym_entry
*syme
)
305 struct symbol
*symbol
= sym_entry__symbol(syme
);
306 struct source_line
*line
;
307 char pattern
[PATTERN_LEN
+ 1];
309 sprintf(pattern
, "%0*Lx <", BITS_PER_LONG
/ 4,
310 map__rip_2objdump(syme
->map
, symbol
->start
));
312 pthread_mutex_lock(&syme
->src
->lock
);
313 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
314 if (memcmp(line
->line
, pattern
, PATTERN_LEN
) == 0) {
315 syme
->src
->source
= line
;
319 pthread_mutex_unlock(&syme
->src
->lock
);
322 static void show_lines(struct source_line
*queue
, int count
, int total
)
325 struct source_line
*line
;
328 for (i
= 0; i
< count
; i
++) {
329 float pcnt
= 100.0*(float)line
->count
[sym_counter
]/(float)total
;
331 printf("%8li %4.1f%%\t%s\n", line
->count
[sym_counter
], pcnt
, line
->line
);
336 #define TRACE_COUNT 3
338 static void show_details(struct sym_entry
*syme
)
340 struct symbol
*symbol
;
341 struct source_line
*line
;
342 struct source_line
*line_queue
= NULL
;
344 int line_queue_count
= 0, total
= 0, more
= 0;
349 if (!syme
->src
->source
)
350 lookup_sym_source(syme
);
352 if (!syme
->src
->source
)
355 symbol
= sym_entry__symbol(syme
);
356 printf("Showing %s for %s\n", event_name(sym_counter
), symbol
->name
);
357 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter
);
359 pthread_mutex_lock(&syme
->src
->lock
);
360 line
= syme
->src
->source
;
362 total
+= line
->count
[sym_counter
];
366 line
= syme
->src
->source
;
370 if (!line_queue_count
)
374 if (line
->count
[sym_counter
])
375 pcnt
= 100.0 * line
->count
[sym_counter
] / (float)total
;
376 if (pcnt
>= (float)sym_pcnt_filter
) {
377 if (displayed
<= print_entries
)
378 show_lines(line_queue
, line_queue_count
, total
);
380 displayed
+= line_queue_count
;
381 line_queue_count
= 0;
383 } else if (line_queue_count
> TRACE_COUNT
) {
384 line_queue
= line_queue
->next
;
388 line
->count
[sym_counter
] = zero
? 0 : line
->count
[sym_counter
] * 7 / 8;
391 pthread_mutex_unlock(&syme
->src
->lock
);
393 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
397 * Symbols will be added here in event__process_sample and will get out
400 static LIST_HEAD(active_symbols
);
401 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
404 * Ordering weight: count-1 * count-2 * ... / count-n
406 static double sym_weight(const struct sym_entry
*sym
)
408 double weight
= sym
->snap_count
;
411 if (!display_weighted
)
414 for (counter
= 1; counter
< nr_counters
-1; counter
++)
415 weight
*= sym
->count
[counter
];
417 weight
/= (sym
->count
[counter
] + 1);
423 static long userspace_samples
;
424 static long exact_samples
;
425 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
427 static void __list_insert_active_sym(struct sym_entry
*syme
)
429 list_add(&syme
->node
, &active_symbols
);
432 static void list_remove_active_sym(struct sym_entry
*syme
)
434 pthread_mutex_lock(&active_symbols_lock
);
435 list_del_init(&syme
->node
);
436 pthread_mutex_unlock(&active_symbols_lock
);
439 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
441 struct rb_node
**p
= &tree
->rb_node
;
442 struct rb_node
*parent
= NULL
;
443 struct sym_entry
*iter
;
447 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
449 if (se
->weight
> iter
->weight
)
455 rb_link_node(&se
->rb_node
, parent
, p
);
456 rb_insert_color(&se
->rb_node
, tree
);
459 static void print_sym_table(void)
462 int counter
, snap
= !display_weighted
? sym_counter
: 0;
463 float samples_per_sec
= samples
/delay_secs
;
464 float ksamples_per_sec
= (samples
-userspace_samples
)/delay_secs
;
465 float esamples_percent
= (100.0*exact_samples
)/samples
;
466 float sum_ksamples
= 0.0;
467 struct sym_entry
*syme
, *n
;
468 struct rb_root tmp
= RB_ROOT
;
470 int sym_width
= 0, dso_width
= 0, dso_short_width
= 0;
471 const int win_width
= winsize
.ws_col
- 1;
473 samples
= userspace_samples
= exact_samples
= 0;
475 /* Sort the active symbols */
476 pthread_mutex_lock(&active_symbols_lock
);
477 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
478 pthread_mutex_unlock(&active_symbols_lock
);
480 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
481 syme
->snap_count
= syme
->count
[snap
];
482 if (syme
->snap_count
!= 0) {
484 if ((hide_user_symbols
&&
485 syme
->origin
== PERF_RECORD_MISC_USER
) ||
486 (hide_kernel_symbols
&&
487 syme
->origin
== PERF_RECORD_MISC_KERNEL
)) {
488 list_remove_active_sym(syme
);
491 syme
->weight
= sym_weight(syme
);
492 rb_insert_active_sym(&tmp
, syme
);
493 sum_ksamples
+= syme
->snap_count
;
495 for (j
= 0; j
< nr_counters
; j
++)
496 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
498 list_remove_active_sym(syme
);
503 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
504 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% exact: %4.1f%% [",
506 100.0 - (100.0*((samples_per_sec
-ksamples_per_sec
)/samples_per_sec
)),
509 if (nr_counters
== 1 || !display_weighted
) {
510 printf("%Ld", (u64
)attrs
[0].sample_period
);
517 if (!display_weighted
)
518 printf("%s", event_name(sym_counter
));
519 else for (counter
= 0; counter
< nr_counters
; counter
++) {
523 printf("%s", event_name(counter
));
528 if (target_pid
!= -1)
529 printf(" (target_pid: %d", target_pid
);
530 else if (target_tid
!= -1)
531 printf(" (target_tid: %d", target_tid
);
535 if (profile_cpu
!= -1)
536 printf(", cpu: %d)\n", profile_cpu
);
538 if (target_tid
!= -1)
541 printf(", %d CPUs)\n", nr_cpus
);
544 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
546 if (sym_filter_entry
) {
547 show_details(sym_filter_entry
);
552 * Find the longest symbol name that will be displayed
554 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
555 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
556 if (++printed
> print_entries
||
557 (int)syme
->snap_count
< count_filter
)
560 if (syme
->map
->dso
->long_name_len
> dso_width
)
561 dso_width
= syme
->map
->dso
->long_name_len
;
563 if (syme
->map
->dso
->short_name_len
> dso_short_width
)
564 dso_short_width
= syme
->map
->dso
->short_name_len
;
566 if (syme
->name_len
> sym_width
)
567 sym_width
= syme
->name_len
;
572 if (sym_width
+ dso_width
> winsize
.ws_col
- 29) {
573 dso_width
= dso_short_width
;
574 if (sym_width
+ dso_width
> winsize
.ws_col
- 29)
575 sym_width
= winsize
.ws_col
- dso_width
- 29;
578 if (nr_counters
== 1)
579 printf(" samples pcnt");
581 printf(" weight samples pcnt");
585 printf(" %-*.*s DSO\n", sym_width
, sym_width
, "function");
586 printf(" %s _______ _____",
587 nr_counters
== 1 ? " " : "______");
589 printf(" ________________");
590 printf(" %-*.*s", sym_width
, sym_width
, graph_line
);
591 printf(" %-*.*s", dso_width
, dso_width
, graph_line
);
594 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
598 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
599 sym
= sym_entry__symbol(syme
);
601 if (++printed
> print_entries
|| (int)syme
->snap_count
< count_filter
)
604 pcnt
= 100.0 - (100.0 * ((sum_ksamples
- syme
->snap_count
) /
607 if (nr_counters
== 1 || !display_weighted
)
608 printf("%20.2f ", syme
->weight
);
610 printf("%9.1f %10ld ", syme
->weight
, syme
->snap_count
);
612 percent_color_fprintf(stdout
, "%4.1f%%", pcnt
);
614 printf(" %016llx", sym
->start
);
615 printf(" %-*.*s", sym_width
, sym_width
, sym
->name
);
616 printf(" %-*.*s\n", dso_width
, dso_width
,
617 dso_width
>= syme
->map
->dso
->long_name_len
?
618 syme
->map
->dso
->long_name
:
619 syme
->map
->dso
->short_name
);
623 static void prompt_integer(int *target
, const char *msg
)
625 char *buf
= malloc(0), *p
;
629 fprintf(stdout
, "\n%s: ", msg
);
630 if (getline(&buf
, &dummy
, stdin
) < 0)
633 p
= strchr(buf
, '\n');
643 tmp
= strtoul(buf
, NULL
, 10);
649 static void prompt_percent(int *target
, const char *msg
)
653 prompt_integer(&tmp
, msg
);
654 if (tmp
>= 0 && tmp
<= 100)
658 static void prompt_symbol(struct sym_entry
**target
, const char *msg
)
660 char *buf
= malloc(0), *p
;
661 struct sym_entry
*syme
= *target
, *n
, *found
= NULL
;
664 /* zero counters of active symbol */
666 pthread_mutex_lock(&syme
->src
->lock
);
667 __zero_source_counters(syme
);
669 pthread_mutex_unlock(&syme
->src
->lock
);
672 fprintf(stdout
, "\n%s: ", msg
);
673 if (getline(&buf
, &dummy
, stdin
) < 0)
676 p
= strchr(buf
, '\n');
680 pthread_mutex_lock(&active_symbols_lock
);
681 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
682 pthread_mutex_unlock(&active_symbols_lock
);
684 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
685 struct symbol
*sym
= sym_entry__symbol(syme
);
687 if (!strcmp(buf
, sym
->name
)) {
694 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
704 static void print_mapped_keys(void)
708 if (sym_filter_entry
) {
709 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
713 fprintf(stdout
, "\nMapped keys:\n");
714 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", delay_secs
);
715 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", print_entries
);
718 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter
));
720 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", count_filter
);
722 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter
);
723 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
724 fprintf(stdout
, "\t[S] stop annotation.\n");
727 fprintf(stdout
, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted
? 1 : 0);
730 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
731 hide_kernel_symbols
? "yes" : "no");
733 "\t[U] hide user symbols. \t(%s)\n",
734 hide_user_symbols
? "yes" : "no");
735 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", zero
? 1 : 0);
736 fprintf(stdout
, "\t[qQ] quit.\n");
739 static int key_mapped(int c
)
756 return nr_counters
> 1 ? 1 : 0;
764 static void handle_keypress(int c
)
766 if (!key_mapped(c
)) {
767 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
768 struct termios tc
, save
;
771 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
776 tc
.c_lflag
&= ~(ICANON
| ECHO
);
779 tcsetattr(0, TCSANOW
, &tc
);
781 poll(&stdin_poll
, 1, -1);
784 tcsetattr(0, TCSAFLUSH
, &save
);
791 prompt_integer(&delay_secs
, "Enter display delay");
796 prompt_integer(&print_entries
, "Enter display entries (lines)");
797 if (print_entries
== 0) {
798 sig_winch_handler(SIGWINCH
);
799 signal(SIGWINCH
, sig_winch_handler
);
801 signal(SIGWINCH
, SIG_DFL
);
804 if (nr_counters
> 1) {
807 fprintf(stderr
, "\nAvailable events:");
808 for (i
= 0; i
< nr_counters
; i
++)
809 fprintf(stderr
, "\n\t%d %s", i
, event_name(i
));
811 prompt_integer(&sym_counter
, "Enter details event counter");
813 if (sym_counter
>= nr_counters
) {
814 fprintf(stderr
, "Sorry, no such event, using %s.\n", event_name(0));
818 } else sym_counter
= 0;
821 prompt_integer(&count_filter
, "Enter display event count filter");
824 prompt_percent(&sym_pcnt_filter
, "Enter details display event filter (percent)");
827 hide_kernel_symbols
= !hide_kernel_symbols
;
831 printf("exiting.\n");
833 dsos__fprintf(stderr
);
836 prompt_symbol(&sym_filter_entry
, "Enter details symbol");
839 if (!sym_filter_entry
)
842 struct sym_entry
*syme
= sym_filter_entry
;
844 pthread_mutex_lock(&syme
->src
->lock
);
845 sym_filter_entry
= NULL
;
846 __zero_source_counters(syme
);
847 pthread_mutex_unlock(&syme
->src
->lock
);
851 hide_user_symbols
= !hide_user_symbols
;
854 display_weighted
= ~display_weighted
;
864 static void *display_thread(void *arg __used
)
866 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
867 struct termios tc
, save
;
872 tc
.c_lflag
&= ~(ICANON
| ECHO
);
877 delay_msecs
= delay_secs
* 1000;
878 tcsetattr(0, TCSANOW
, &tc
);
884 } while (!poll(&stdin_poll
, 1, delay_msecs
) == 1);
887 tcsetattr(0, TCSAFLUSH
, &save
);
895 /* Tag samples to be skipped. */
896 static const char *skip_symbols
[] = {
902 "mwait_idle_with_hints",
904 "ppc64_runlatch_off",
905 "pseries_dedicated_idle_sleep",
909 static int symbol_filter(struct map
*map
, struct symbol
*sym
)
911 struct sym_entry
*syme
;
912 const char *name
= sym
->name
;
916 * ppc64 uses function descriptors and appends a '.' to the
917 * start of every instruction address. Remove it.
922 if (!strcmp(name
, "_text") ||
923 !strcmp(name
, "_etext") ||
924 !strcmp(name
, "_sinittext") ||
925 !strncmp("init_module", name
, 11) ||
926 !strncmp("cleanup_module", name
, 14) ||
927 strstr(name
, "_text_start") ||
928 strstr(name
, "_text_end"))
931 syme
= symbol__priv(sym
);
935 if (!sym_filter_entry
&& sym_filter
&& !strcmp(name
, sym_filter
)) {
936 /* schedule initial sym_filter_entry setup */
937 sym_filter_entry_sched
= syme
;
941 for (i
= 0; skip_symbols
[i
]; i
++) {
942 if (!strcmp(skip_symbols
[i
], name
)) {
949 syme
->name_len
= strlen(sym
->name
);
954 static void event__process_sample(const event_t
*self
,
955 struct perf_session
*session
, int counter
)
957 u64 ip
= self
->ip
.ip
;
958 struct sym_entry
*syme
;
959 struct addr_location al
;
960 u8 origin
= self
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
965 case PERF_RECORD_MISC_USER
:
967 if (hide_user_symbols
)
970 case PERF_RECORD_MISC_KERNEL
:
971 if (hide_kernel_symbols
)
978 if (self
->header
.misc
& PERF_RECORD_MISC_EXACT
)
981 if (event__preprocess_sample(self
, session
, &al
, symbol_filter
) < 0 ||
985 if (al
.sym
== NULL
) {
987 * As we do lazy loading of symtabs we only will know if the
988 * specified vmlinux file is invalid when we actually have a
989 * hit in kernel space and then try to load it. So if we get
990 * here and there are _no_ symbols in the DSO backing the
991 * kernel map, bail out.
993 * We may never get here, for instance, if we use -K/
994 * --hide-kernel-symbols, even if the user specifies an
995 * invalid --vmlinux ;-)
997 if (al
.map
== session
->vmlinux_maps
[MAP__FUNCTION
] &&
998 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
999 pr_err("The %s file can't be used\n",
1000 symbol_conf
.vmlinux_name
);
1007 /* let's see, whether we need to install initial sym_filter_entry */
1008 if (sym_filter_entry_sched
) {
1009 sym_filter_entry
= sym_filter_entry_sched
;
1010 sym_filter_entry_sched
= NULL
;
1011 if (parse_source(sym_filter_entry
) < 0) {
1012 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
1014 pr_err("Can't annotate %s", sym
->name
);
1015 if (sym_filter_entry
->map
->dso
->origin
== DSO__ORIG_KERNEL
) {
1016 pr_err(": No vmlinux file was found in the path:\n");
1017 vmlinux_path__fprintf(stderr
);
1024 syme
= symbol__priv(al
.sym
);
1026 syme
->count
[counter
]++;
1027 syme
->origin
= origin
;
1028 record_precise_ip(syme
, counter
, ip
);
1029 pthread_mutex_lock(&active_symbols_lock
);
1030 if (list_empty(&syme
->node
) || !syme
->node
.next
)
1031 __list_insert_active_sym(syme
);
1032 pthread_mutex_unlock(&active_symbols_lock
);
1036 static int event__process(event_t
*event
, struct perf_session
*session
)
1038 switch (event
->header
.type
) {
1039 case PERF_RECORD_COMM
:
1040 event__process_comm(event
, session
);
1042 case PERF_RECORD_MMAP
:
1043 event__process_mmap(event
, session
);
1045 case PERF_RECORD_FORK
:
1046 case PERF_RECORD_EXIT
:
1047 event__process_task(event
, session
);
1063 static unsigned int mmap_read_head(struct mmap_data
*md
)
1065 struct perf_event_mmap_page
*pc
= md
->base
;
1068 head
= pc
->data_head
;
1074 static void perf_session__mmap_read_counter(struct perf_session
*self
,
1075 struct mmap_data
*md
)
1077 unsigned int head
= mmap_read_head(md
);
1078 unsigned int old
= md
->prev
;
1079 unsigned char *data
= md
->base
+ page_size
;
1083 * If we're further behind than half the buffer, there's a chance
1084 * the writer will bite our tail and mess up the samples under us.
1086 * If we somehow ended up ahead of the head, we got messed up.
1088 * In either case, truncate and restart at head.
1091 if (diff
> md
->mask
/ 2 || diff
< 0) {
1092 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
1095 * head points to a known good entry, start there.
1100 for (; old
!= head
;) {
1101 event_t
*event
= (event_t
*)&data
[old
& md
->mask
];
1105 size_t size
= event
->header
.size
;
1108 * Event straddles the mmap boundary -- header should always
1109 * be inside due to u64 alignment of output.
1111 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
1112 unsigned int offset
= old
;
1113 unsigned int len
= min(sizeof(*event
), size
), cpy
;
1114 void *dst
= &event_copy
;
1117 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
1118 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
1124 event
= &event_copy
;
1127 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
1128 event__process_sample(event
, self
, md
->counter
);
1130 event__process(event
, self
);
1137 static struct pollfd
*event_array
;
1138 static struct mmap_data
*mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
1140 static void perf_session__mmap_read(struct perf_session
*self
)
1142 int i
, counter
, thread_index
;
1144 for (i
= 0; i
< nr_cpus
; i
++) {
1145 for (counter
= 0; counter
< nr_counters
; counter
++)
1146 for (thread_index
= 0;
1147 thread_index
< thread_num
;
1149 perf_session__mmap_read_counter(self
,
1150 &mmap_array
[i
][counter
][thread_index
]);
1158 static void start_counter(int i
, int counter
)
1160 struct perf_event_attr
*attr
;
1165 if (target_tid
== -1 && profile_cpu
== -1)
1168 attr
= attrs
+ counter
;
1170 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
1173 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
1175 attr
->sample_freq
= freq
;
1178 attr
->inherit
= (cpu
< 0) && inherit
;
1181 for (thread_index
= 0; thread_index
< thread_num
; thread_index
++) {
1183 fd
[i
][counter
][thread_index
] = sys_perf_event_open(attr
,
1184 all_tids
[thread_index
], cpu
, group_fd
, 0);
1186 if (fd
[i
][counter
][thread_index
] < 0) {
1189 if (err
== EPERM
|| err
== EACCES
)
1190 die("No permission - are you root?\n");
1192 * If it's cycles then fall back to hrtimer
1193 * based cpu-clock-tick sw counter, which
1194 * is always available even if no PMU support:
1196 if (attr
->type
== PERF_TYPE_HARDWARE
1197 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
1200 warning(" ... trying to fall back to cpu-clock-ticks\n");
1202 attr
->type
= PERF_TYPE_SOFTWARE
;
1203 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
1207 error("perfcounter syscall returned with %d (%s)\n",
1208 fd
[i
][counter
][thread_index
], strerror(err
));
1209 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1212 assert(fd
[i
][counter
][thread_index
] >= 0);
1213 fcntl(fd
[i
][counter
][thread_index
], F_SETFL
, O_NONBLOCK
);
1216 * First counter acts as the group leader:
1218 if (group
&& group_fd
== -1)
1219 group_fd
= fd
[i
][counter
][thread_index
];
1221 event_array
[nr_poll
].fd
= fd
[i
][counter
][thread_index
];
1222 event_array
[nr_poll
].events
= POLLIN
;
1225 mmap_array
[i
][counter
][thread_index
].counter
= counter
;
1226 mmap_array
[i
][counter
][thread_index
].prev
= 0;
1227 mmap_array
[i
][counter
][thread_index
].mask
= mmap_pages
*page_size
- 1;
1228 mmap_array
[i
][counter
][thread_index
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
1229 PROT_READ
, MAP_SHARED
, fd
[i
][counter
][thread_index
], 0);
1230 if (mmap_array
[i
][counter
][thread_index
].base
== MAP_FAILED
)
1231 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
1235 static int __cmd_top(void)
1241 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1242 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1244 struct perf_session
*session
= perf_session__new(NULL
, O_WRONLY
, false);
1245 if (session
== NULL
)
1248 if (target_tid
!= -1)
1249 event__synthesize_thread(target_tid
, event__process
, session
);
1251 event__synthesize_threads(event__process
, session
);
1253 for (i
= 0; i
< nr_cpus
; i
++) {
1255 for (counter
= 0; counter
< nr_counters
; counter
++)
1256 start_counter(i
, counter
);
1259 /* Wait for a minimal set of events before starting the snapshot */
1260 poll(&event_array
[0], nr_poll
, 100);
1262 perf_session__mmap_read(session
);
1264 if (pthread_create(&thread
, NULL
, display_thread
, NULL
)) {
1265 printf("Could not create display thread.\n");
1269 if (realtime_prio
) {
1270 struct sched_param param
;
1272 param
.sched_priority
= realtime_prio
;
1273 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1274 printf("Could not set realtime priority.\n");
1282 perf_session__mmap_read(session
);
1284 if (hits
== samples
)
1285 ret
= poll(event_array
, nr_poll
, 100);
1291 static const char * const top_usage
[] = {
1292 "perf top [<options>]",
1296 static const struct option options
[] = {
1297 OPT_CALLBACK('e', "event", NULL
, "event",
1298 "event selector. use 'perf list' to list available events",
1300 OPT_INTEGER('c', "count", &default_interval
,
1301 "event period to sample"),
1302 OPT_INTEGER('p', "pid", &target_pid
,
1303 "profile events on existing process id"),
1304 OPT_INTEGER('t', "tid", &target_tid
,
1305 "profile events on existing thread id"),
1306 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1307 "system-wide collection from all CPUs"),
1308 OPT_INTEGER('C', "CPU", &profile_cpu
,
1309 "CPU to profile on"),
1310 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1311 "file", "vmlinux pathname"),
1312 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols
,
1313 "hide kernel symbols"),
1314 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
1315 "number of mmap data pages"),
1316 OPT_INTEGER('r', "realtime", &realtime_prio
,
1317 "collect data with this RT SCHED_FIFO priority"),
1318 OPT_INTEGER('d', "delay", &delay_secs
,
1319 "number of seconds to delay between refreshes"),
1320 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
1321 "dump the symbol table used for profiling"),
1322 OPT_INTEGER('f', "count-filter", &count_filter
,
1323 "only display functions with more events than this"),
1324 OPT_BOOLEAN('g', "group", &group
,
1325 "put the counters into a counter group"),
1326 OPT_BOOLEAN('i', "inherit", &inherit
,
1327 "child tasks inherit counters"),
1328 OPT_STRING('s', "sym-annotate", &sym_filter
, "symbol name",
1329 "symbol to annotate"),
1330 OPT_BOOLEAN('z', "zero", &zero
,
1331 "zero history across updates"),
1332 OPT_INTEGER('F', "freq", &freq
,
1333 "profile at this frequency"),
1334 OPT_INTEGER('E', "entries", &print_entries
,
1335 "display this many functions"),
1336 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols
,
1337 "hide user symbols"),
1338 OPT_BOOLEAN('v', "verbose", &verbose
,
1339 "be more verbose (show counter open errors, etc)"),
1343 int cmd_top(int argc
, const char **argv
, const char *prefix __used
)
1348 page_size
= sysconf(_SC_PAGE_SIZE
);
1350 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1352 usage_with_options(top_usage
, options
);
1354 if (target_pid
!= -1) {
1355 target_tid
= target_pid
;
1356 thread_num
= find_all_tid(target_pid
, &all_tids
);
1357 if (thread_num
<= 0) {
1358 fprintf(stderr
, "Can't find all threads of pid %d\n",
1360 usage_with_options(top_usage
, options
);
1363 all_tids
=malloc(sizeof(pid_t
));
1367 all_tids
[0] = target_tid
;
1371 for (i
= 0; i
< MAX_NR_CPUS
; i
++) {
1372 for (j
= 0; j
< MAX_COUNTERS
; j
++) {
1373 fd
[i
][j
] = malloc(sizeof(int)*thread_num
);
1374 mmap_array
[i
][j
] = malloc(
1375 sizeof(struct mmap_data
)*thread_num
);
1376 if (!fd
[i
][j
] || !mmap_array
[i
][j
])
1380 event_array
= malloc(
1381 sizeof(struct pollfd
)*MAX_NR_CPUS
*MAX_COUNTERS
*thread_num
);
1385 /* CPU and PID are mutually exclusive */
1386 if (target_tid
> 0 && profile_cpu
!= -1) {
1387 printf("WARNING: PID switch overriding CPU\n");
1395 symbol_conf
.priv_size
= (sizeof(struct sym_entry
) +
1396 (nr_counters
+ 1) * sizeof(unsigned long));
1398 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1399 if (symbol__init() < 0)
1406 * User specified count overrides default frequency.
1408 if (default_interval
)
1411 default_interval
= freq
;
1413 fprintf(stderr
, "frequency and count are zero, aborting\n");
1418 * Fill in the ones not specifically initialized via -c:
1420 for (counter
= 0; counter
< nr_counters
; counter
++) {
1421 if (attrs
[counter
].sample_period
)
1424 attrs
[counter
].sample_period
= default_interval
;
1427 if (target_tid
!= -1 || profile_cpu
!= -1)
1430 nr_cpus
= read_cpu_map();
1432 get_term_dimensions(&winsize
);
1433 if (print_entries
== 0) {
1434 update_print_entries(&winsize
);
1435 signal(SIGWINCH
, sig_winch_handler
);