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"
32 #include "util/debug.h"
46 #include <sys/syscall.h>
47 #include <sys/ioctl.h>
49 #include <sys/prctl.h>
54 #include <linux/unistd.h>
55 #include <linux/types.h>
57 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
59 static int system_wide
= 0;
61 static int default_interval
= 0;
63 static int count_filter
= 5;
64 static int print_entries
;
66 static int target_pid
= -1;
67 static int inherit
= 0;
68 static int profile_cpu
= -1;
69 static int nr_cpus
= 0;
70 static unsigned int realtime_prio
= 0;
72 static unsigned int page_size
;
73 static unsigned int mmap_pages
= 16;
74 static int freq
= 1000; /* 1 KHz */
76 static int delay_secs
= 2;
78 static int dump_symtab
= 0;
80 static bool hide_kernel_symbols
= false;
81 static bool hide_user_symbols
= false;
82 static struct winsize winsize
;
90 unsigned long count
[MAX_COUNTERS
];
92 struct source_line
*next
;
95 static char *sym_filter
= NULL
;
96 struct sym_entry
*sym_filter_entry
= NULL
;
97 struct sym_entry
*sym_filter_entry_sched
= NULL
;
98 static int sym_pcnt_filter
= 5;
99 static int sym_counter
= 0;
100 static int display_weighted
= -1;
106 struct sym_entry_source
{
107 struct source_line
*source
;
108 struct source_line
*lines
;
109 struct source_line
**lines_tail
;
110 pthread_mutex_t lock
;
114 struct rb_node rb_node
;
115 struct list_head node
;
116 unsigned long snap_count
;
122 struct sym_entry_source
*src
;
123 unsigned long count
[0];
130 static inline struct symbol
*sym_entry__symbol(struct sym_entry
*self
)
132 return ((void *)self
) + symbol_conf
.priv_size
;
135 static void get_term_dimensions(struct winsize
*ws
)
137 char *s
= getenv("LINES");
140 ws
->ws_row
= atoi(s
);
141 s
= getenv("COLUMNS");
143 ws
->ws_col
= atoi(s
);
144 if (ws
->ws_row
&& ws
->ws_col
)
149 if (ioctl(1, TIOCGWINSZ
, ws
) == 0 &&
150 ws
->ws_row
&& ws
->ws_col
)
157 static void update_print_entries(struct winsize
*ws
)
159 print_entries
= ws
->ws_row
;
161 if (print_entries
> 9)
165 static void sig_winch_handler(int sig __used
)
167 get_term_dimensions(&winsize
);
168 update_print_entries(&winsize
);
171 static void parse_source(struct sym_entry
*syme
)
174 struct sym_entry_source
*source
;
177 char command
[PATH_MAX
*2];
184 if (syme
->src
== NULL
) {
185 syme
->src
= zalloc(sizeof(*source
));
186 if (syme
->src
== NULL
)
188 pthread_mutex_init(&syme
->src
->lock
, NULL
);
194 pthread_mutex_lock(&source
->lock
);
198 sym
= sym_entry__symbol(syme
);
200 path
= map
->dso
->long_name
;
202 len
= sym
->end
- sym
->start
;
205 "objdump --start-address=%#0*Lx --stop-address=%#0*Lx -dS %s",
206 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->start
),
207 BITS_PER_LONG
/ 4, map__rip_2objdump(map
, sym
->end
), path
);
209 file
= popen(command
, "r");
213 pthread_mutex_lock(&source
->lock
);
214 source
->lines_tail
= &source
->lines
;
215 while (!feof(file
)) {
216 struct source_line
*src
;
220 src
= malloc(sizeof(struct source_line
));
222 memset(src
, 0, sizeof(struct source_line
));
224 if (getline(&src
->line
, &dummy
, file
) < 0)
229 c
= strchr(src
->line
, '\n');
234 *source
->lines_tail
= src
;
235 source
->lines_tail
= &src
->next
;
237 src
->eip
= strtoull(src
->line
, &sep
, 16);
239 src
->eip
= map__objdump_2ip(map
, src
->eip
);
240 else /* this line has no ip info (e.g. source line) */
245 sym_filter_entry
= syme
;
246 pthread_mutex_unlock(&source
->lock
);
249 static void __zero_source_counters(struct sym_entry
*syme
)
252 struct source_line
*line
;
254 line
= syme
->src
->lines
;
256 for (i
= 0; i
< nr_counters
; i
++)
262 static void record_precise_ip(struct sym_entry
*syme
, int counter
, u64 ip
)
264 struct source_line
*line
;
266 if (syme
!= sym_filter_entry
)
269 if (pthread_mutex_trylock(&syme
->src
->lock
))
272 if (syme
->src
== NULL
|| syme
->src
->source
== NULL
)
275 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
276 /* skip lines without IP info */
279 if (line
->eip
== ip
) {
280 line
->count
[counter
]++;
287 pthread_mutex_unlock(&syme
->src
->lock
);
290 #define PATTERN_LEN (BITS_PER_LONG / 4 + 2)
292 static void lookup_sym_source(struct sym_entry
*syme
)
294 struct symbol
*symbol
= sym_entry__symbol(syme
);
295 struct source_line
*line
;
296 char pattern
[PATTERN_LEN
+ 1];
298 sprintf(pattern
, "%0*Lx <", BITS_PER_LONG
/ 4,
299 map__rip_2objdump(syme
->map
, symbol
->start
));
301 pthread_mutex_lock(&syme
->src
->lock
);
302 for (line
= syme
->src
->lines
; line
; line
= line
->next
) {
303 if (memcmp(line
->line
, pattern
, PATTERN_LEN
) == 0) {
304 syme
->src
->source
= line
;
308 pthread_mutex_unlock(&syme
->src
->lock
);
311 static void show_lines(struct source_line
*queue
, int count
, int total
)
314 struct source_line
*line
;
317 for (i
= 0; i
< count
; i
++) {
318 float pcnt
= 100.0*(float)line
->count
[sym_counter
]/(float)total
;
320 printf("%8li %4.1f%%\t%s\n", line
->count
[sym_counter
], pcnt
, line
->line
);
325 #define TRACE_COUNT 3
327 static void show_details(struct sym_entry
*syme
)
329 struct symbol
*symbol
;
330 struct source_line
*line
;
331 struct source_line
*line_queue
= NULL
;
333 int line_queue_count
= 0, total
= 0, more
= 0;
338 if (!syme
->src
->source
)
339 lookup_sym_source(syme
);
341 if (!syme
->src
->source
)
344 symbol
= sym_entry__symbol(syme
);
345 printf("Showing %s for %s\n", event_name(sym_counter
), symbol
->name
);
346 printf(" Events Pcnt (>=%d%%)\n", sym_pcnt_filter
);
348 pthread_mutex_lock(&syme
->src
->lock
);
349 line
= syme
->src
->source
;
351 total
+= line
->count
[sym_counter
];
355 line
= syme
->src
->source
;
359 if (!line_queue_count
)
363 if (line
->count
[sym_counter
])
364 pcnt
= 100.0 * line
->count
[sym_counter
] / (float)total
;
365 if (pcnt
>= (float)sym_pcnt_filter
) {
366 if (displayed
<= print_entries
)
367 show_lines(line_queue
, line_queue_count
, total
);
369 displayed
+= line_queue_count
;
370 line_queue_count
= 0;
372 } else if (line_queue_count
> TRACE_COUNT
) {
373 line_queue
= line_queue
->next
;
377 line
->count
[sym_counter
] = zero
? 0 : line
->count
[sym_counter
] * 7 / 8;
380 pthread_mutex_unlock(&syme
->src
->lock
);
382 printf("%d lines not displayed, maybe increase display entries [e]\n", more
);
386 * Symbols will be added here in event__process_sample and will get out
389 static LIST_HEAD(active_symbols
);
390 static pthread_mutex_t active_symbols_lock
= PTHREAD_MUTEX_INITIALIZER
;
393 * Ordering weight: count-1 * count-2 * ... / count-n
395 static double sym_weight(const struct sym_entry
*sym
)
397 double weight
= sym
->snap_count
;
400 if (!display_weighted
)
403 for (counter
= 1; counter
< nr_counters
-1; counter
++)
404 weight
*= sym
->count
[counter
];
406 weight
/= (sym
->count
[counter
] + 1);
412 static long userspace_samples
;
413 static const char CONSOLE_CLEAR
[] = "\e[H\e[2J";
415 static void __list_insert_active_sym(struct sym_entry
*syme
)
417 list_add(&syme
->node
, &active_symbols
);
420 static void list_remove_active_sym(struct sym_entry
*syme
)
422 pthread_mutex_lock(&active_symbols_lock
);
423 list_del_init(&syme
->node
);
424 pthread_mutex_unlock(&active_symbols_lock
);
427 static void rb_insert_active_sym(struct rb_root
*tree
, struct sym_entry
*se
)
429 struct rb_node
**p
= &tree
->rb_node
;
430 struct rb_node
*parent
= NULL
;
431 struct sym_entry
*iter
;
435 iter
= rb_entry(parent
, struct sym_entry
, rb_node
);
437 if (se
->weight
> iter
->weight
)
443 rb_link_node(&se
->rb_node
, parent
, p
);
444 rb_insert_color(&se
->rb_node
, tree
);
447 static void print_sym_table(void)
450 int counter
, snap
= !display_weighted
? sym_counter
: 0;
451 float samples_per_sec
= samples
/delay_secs
;
452 float ksamples_per_sec
= (samples
-userspace_samples
)/delay_secs
;
453 float sum_ksamples
= 0.0;
454 struct sym_entry
*syme
, *n
;
455 struct rb_root tmp
= RB_ROOT
;
457 int sym_width
= 0, dso_width
= 0, max_dso_width
;
458 const int win_width
= winsize
.ws_col
- 1;
460 samples
= userspace_samples
= 0;
462 /* Sort the active symbols */
463 pthread_mutex_lock(&active_symbols_lock
);
464 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
465 pthread_mutex_unlock(&active_symbols_lock
);
467 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
468 syme
->snap_count
= syme
->count
[snap
];
469 if (syme
->snap_count
!= 0) {
471 if ((hide_user_symbols
&&
472 syme
->origin
== PERF_RECORD_MISC_USER
) ||
473 (hide_kernel_symbols
&&
474 syme
->origin
== PERF_RECORD_MISC_KERNEL
)) {
475 list_remove_active_sym(syme
);
478 syme
->weight
= sym_weight(syme
);
479 rb_insert_active_sym(&tmp
, syme
);
480 sum_ksamples
+= syme
->snap_count
;
482 for (j
= 0; j
< nr_counters
; j
++)
483 syme
->count
[j
] = zero
? 0 : syme
->count
[j
] * 7 / 8;
485 list_remove_active_sym(syme
);
490 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
491 printf( " PerfTop:%8.0f irqs/sec kernel:%4.1f%% [",
493 100.0 - (100.0*((samples_per_sec
-ksamples_per_sec
)/samples_per_sec
)));
495 if (nr_counters
== 1 || !display_weighted
) {
496 printf("%Ld", (u64
)attrs
[0].sample_period
);
503 if (!display_weighted
)
504 printf("%s", event_name(sym_counter
));
505 else for (counter
= 0; counter
< nr_counters
; counter
++) {
509 printf("%s", event_name(counter
));
514 if (target_pid
!= -1)
515 printf(" (target_pid: %d", target_pid
);
519 if (profile_cpu
!= -1)
520 printf(", cpu: %d)\n", profile_cpu
);
522 if (target_pid
!= -1)
525 printf(", %d CPUs)\n", nr_cpus
);
528 printf("%-*.*s\n", win_width
, win_width
, graph_dotted_line
);
530 if (sym_filter_entry
) {
531 show_details(sym_filter_entry
);
536 * Find the longest symbol name that will be displayed
538 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
539 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
540 if (++printed
> print_entries
||
541 (int)syme
->snap_count
< count_filter
)
544 if (syme
->map
->dso
->long_name_len
> dso_width
)
545 dso_width
= syme
->map
->dso
->long_name_len
;
547 if (syme
->name_len
> sym_width
)
548 sym_width
= syme
->name_len
;
553 max_dso_width
= winsize
.ws_col
- sym_width
- 29;
554 if (dso_width
> max_dso_width
)
555 dso_width
= max_dso_width
;
557 if (nr_counters
== 1)
558 printf(" samples pcnt");
560 printf(" weight samples pcnt");
564 printf(" %-*.*s DSO\n", sym_width
, sym_width
, "function");
565 printf(" %s _______ _____",
566 nr_counters
== 1 ? " " : "______");
568 printf(" ________________");
569 printf(" %-*.*s", sym_width
, sym_width
, graph_line
);
570 printf(" %-*.*s", dso_width
, dso_width
, graph_line
);
573 for (nd
= rb_first(&tmp
); nd
; nd
= rb_next(nd
)) {
577 syme
= rb_entry(nd
, struct sym_entry
, rb_node
);
578 sym
= sym_entry__symbol(syme
);
580 if (++printed
> print_entries
|| (int)syme
->snap_count
< count_filter
)
583 pcnt
= 100.0 - (100.0 * ((sum_ksamples
- syme
->snap_count
) /
586 if (nr_counters
== 1 || !display_weighted
)
587 printf("%20.2f ", syme
->weight
);
589 printf("%9.1f %10ld ", syme
->weight
, syme
->snap_count
);
591 percent_color_fprintf(stdout
, "%4.1f%%", pcnt
);
593 printf(" %016llx", sym
->start
);
594 printf(" %-*.*s", sym_width
, sym_width
, sym
->name
);
595 printf(" %-*.*s\n", dso_width
, dso_width
,
596 dso_width
>= syme
->map
->dso
->long_name_len
?
597 syme
->map
->dso
->long_name
:
598 syme
->map
->dso
->short_name
);
602 static void prompt_integer(int *target
, const char *msg
)
604 char *buf
= malloc(0), *p
;
608 fprintf(stdout
, "\n%s: ", msg
);
609 if (getline(&buf
, &dummy
, stdin
) < 0)
612 p
= strchr(buf
, '\n');
622 tmp
= strtoul(buf
, NULL
, 10);
628 static void prompt_percent(int *target
, const char *msg
)
632 prompt_integer(&tmp
, msg
);
633 if (tmp
>= 0 && tmp
<= 100)
637 static void prompt_symbol(struct sym_entry
**target
, const char *msg
)
639 char *buf
= malloc(0), *p
;
640 struct sym_entry
*syme
= *target
, *n
, *found
= NULL
;
643 /* zero counters of active symbol */
645 pthread_mutex_lock(&syme
->src
->lock
);
646 __zero_source_counters(syme
);
648 pthread_mutex_unlock(&syme
->src
->lock
);
651 fprintf(stdout
, "\n%s: ", msg
);
652 if (getline(&buf
, &dummy
, stdin
) < 0)
655 p
= strchr(buf
, '\n');
659 pthread_mutex_lock(&active_symbols_lock
);
660 syme
= list_entry(active_symbols
.next
, struct sym_entry
, node
);
661 pthread_mutex_unlock(&active_symbols_lock
);
663 list_for_each_entry_safe_from(syme
, n
, &active_symbols
, node
) {
664 struct symbol
*sym
= sym_entry__symbol(syme
);
666 if (!strcmp(buf
, sym
->name
)) {
673 fprintf(stderr
, "Sorry, %s is not active.\n", buf
);
683 static void print_mapped_keys(void)
687 if (sym_filter_entry
) {
688 struct symbol
*sym
= sym_entry__symbol(sym_filter_entry
);
692 fprintf(stdout
, "\nMapped keys:\n");
693 fprintf(stdout
, "\t[d] display refresh delay. \t(%d)\n", delay_secs
);
694 fprintf(stdout
, "\t[e] display entries (lines). \t(%d)\n", print_entries
);
697 fprintf(stdout
, "\t[E] active event counter. \t(%s)\n", event_name(sym_counter
));
699 fprintf(stdout
, "\t[f] profile display filter (count). \t(%d)\n", count_filter
);
701 fprintf(stdout
, "\t[F] annotate display filter (percent). \t(%d%%)\n", sym_pcnt_filter
);
702 fprintf(stdout
, "\t[s] annotate symbol. \t(%s)\n", name
?: "NULL");
703 fprintf(stdout
, "\t[S] stop annotation.\n");
706 fprintf(stdout
, "\t[w] toggle display weighted/count[E]r. \t(%d)\n", display_weighted
? 1 : 0);
709 "\t[K] hide kernel_symbols symbols. \t(%s)\n",
710 hide_kernel_symbols
? "yes" : "no");
712 "\t[U] hide user symbols. \t(%s)\n",
713 hide_user_symbols
? "yes" : "no");
714 fprintf(stdout
, "\t[z] toggle sample zeroing. \t(%d)\n", zero
? 1 : 0);
715 fprintf(stdout
, "\t[qQ] quit.\n");
718 static int key_mapped(int c
)
735 return nr_counters
> 1 ? 1 : 0;
743 static void handle_keypress(int c
)
745 if (!key_mapped(c
)) {
746 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
747 struct termios tc
, save
;
750 fprintf(stdout
, "\nEnter selection, or unmapped key to continue: ");
755 tc
.c_lflag
&= ~(ICANON
| ECHO
);
758 tcsetattr(0, TCSANOW
, &tc
);
760 poll(&stdin_poll
, 1, -1);
763 tcsetattr(0, TCSAFLUSH
, &save
);
770 prompt_integer(&delay_secs
, "Enter display delay");
775 prompt_integer(&print_entries
, "Enter display entries (lines)");
776 if (print_entries
== 0) {
777 sig_winch_handler(SIGWINCH
);
778 signal(SIGWINCH
, sig_winch_handler
);
780 signal(SIGWINCH
, SIG_DFL
);
783 if (nr_counters
> 1) {
786 fprintf(stderr
, "\nAvailable events:");
787 for (i
= 0; i
< nr_counters
; i
++)
788 fprintf(stderr
, "\n\t%d %s", i
, event_name(i
));
790 prompt_integer(&sym_counter
, "Enter details event counter");
792 if (sym_counter
>= nr_counters
) {
793 fprintf(stderr
, "Sorry, no such event, using %s.\n", event_name(0));
797 } else sym_counter
= 0;
800 prompt_integer(&count_filter
, "Enter display event count filter");
803 prompt_percent(&sym_pcnt_filter
, "Enter details display event filter (percent)");
806 hide_kernel_symbols
= !hide_kernel_symbols
;
810 printf("exiting.\n");
812 dsos__fprintf(stderr
);
815 prompt_symbol(&sym_filter_entry
, "Enter details symbol");
818 if (!sym_filter_entry
)
821 struct sym_entry
*syme
= sym_filter_entry
;
823 pthread_mutex_lock(&syme
->src
->lock
);
824 sym_filter_entry
= NULL
;
825 __zero_source_counters(syme
);
826 pthread_mutex_unlock(&syme
->src
->lock
);
830 hide_user_symbols
= !hide_user_symbols
;
833 display_weighted
= ~display_weighted
;
843 static void *display_thread(void *arg __used
)
845 struct pollfd stdin_poll
= { .fd
= 0, .events
= POLLIN
};
846 struct termios tc
, save
;
851 tc
.c_lflag
&= ~(ICANON
| ECHO
);
856 delay_msecs
= delay_secs
* 1000;
857 tcsetattr(0, TCSANOW
, &tc
);
863 } while (!poll(&stdin_poll
, 1, delay_msecs
) == 1);
866 tcsetattr(0, TCSAFLUSH
, &save
);
874 /* Tag samples to be skipped. */
875 static const char *skip_symbols
[] = {
881 "mwait_idle_with_hints",
883 "ppc64_runlatch_off",
884 "pseries_dedicated_idle_sleep",
888 static int symbol_filter(struct map
*map
, struct symbol
*sym
)
890 struct sym_entry
*syme
;
891 const char *name
= sym
->name
;
895 * ppc64 uses function descriptors and appends a '.' to the
896 * start of every instruction address. Remove it.
901 if (!strcmp(name
, "_text") ||
902 !strcmp(name
, "_etext") ||
903 !strcmp(name
, "_sinittext") ||
904 !strncmp("init_module", name
, 11) ||
905 !strncmp("cleanup_module", name
, 14) ||
906 strstr(name
, "_text_start") ||
907 strstr(name
, "_text_end"))
910 syme
= symbol__priv(sym
);
914 if (!sym_filter_entry
&& sym_filter
&& !strcmp(name
, sym_filter
)) {
915 /* schedule initial sym_filter_entry setup */
916 sym_filter_entry_sched
= syme
;
920 for (i
= 0; skip_symbols
[i
]; i
++) {
921 if (!strcmp(skip_symbols
[i
], name
)) {
928 syme
->name_len
= strlen(sym
->name
);
933 static void event__process_sample(const event_t
*self
,
934 struct perf_session
*session
, int counter
)
936 u64 ip
= self
->ip
.ip
;
937 struct sym_entry
*syme
;
938 struct addr_location al
;
939 u8 origin
= self
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
944 case PERF_RECORD_MISC_USER
:
946 if (hide_user_symbols
)
949 case PERF_RECORD_MISC_KERNEL
:
950 if (hide_kernel_symbols
)
957 if (event__preprocess_sample(self
, session
, &al
, symbol_filter
) < 0 ||
961 if (al
.sym
== NULL
) {
963 * As we do lazy loading of symtabs we only will know if the
964 * specified vmlinux file is invalid when we actually have a
965 * hit in kernel space and then try to load it. So if we get
966 * here and there are _no_ symbols in the DSO backing the
967 * kernel map, bail out.
969 * We may never get here, for instance, if we use -K/
970 * --hide-kernel-symbols, even if the user specifies an
971 * invalid --vmlinux ;-)
973 if (al
.map
== session
->vmlinux_maps
[MAP__FUNCTION
] &&
974 RB_EMPTY_ROOT(&al
.map
->dso
->symbols
[MAP__FUNCTION
])) {
975 pr_err("The %s file can't be used\n",
976 symbol_conf
.vmlinux_name
);
983 /* let's see, whether we need to install initial sym_filter_entry */
984 if (sym_filter_entry_sched
) {
985 sym_filter_entry
= sym_filter_entry_sched
;
986 sym_filter_entry_sched
= NULL
;
987 parse_source(sym_filter_entry
);
990 syme
= symbol__priv(al
.sym
);
992 syme
->count
[counter
]++;
993 syme
->origin
= origin
;
994 record_precise_ip(syme
, counter
, ip
);
995 pthread_mutex_lock(&active_symbols_lock
);
996 if (list_empty(&syme
->node
) || !syme
->node
.next
)
997 __list_insert_active_sym(syme
);
998 pthread_mutex_unlock(&active_symbols_lock
);
1002 static int event__process(event_t
*event
, struct perf_session
*session
)
1004 switch (event
->header
.type
) {
1005 case PERF_RECORD_COMM
:
1006 event__process_comm(event
, session
);
1008 case PERF_RECORD_MMAP
:
1009 event__process_mmap(event
, session
);
1011 case PERF_RECORD_FORK
:
1012 case PERF_RECORD_EXIT
:
1013 event__process_task(event
, session
);
1029 static unsigned int mmap_read_head(struct mmap_data
*md
)
1031 struct perf_event_mmap_page
*pc
= md
->base
;
1034 head
= pc
->data_head
;
1040 static void perf_session__mmap_read_counter(struct perf_session
*self
,
1041 struct mmap_data
*md
)
1043 unsigned int head
= mmap_read_head(md
);
1044 unsigned int old
= md
->prev
;
1045 unsigned char *data
= md
->base
+ page_size
;
1049 * If we're further behind than half the buffer, there's a chance
1050 * the writer will bite our tail and mess up the samples under us.
1052 * If we somehow ended up ahead of the head, we got messed up.
1054 * In either case, truncate and restart at head.
1057 if (diff
> md
->mask
/ 2 || diff
< 0) {
1058 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
1061 * head points to a known good entry, start there.
1066 for (; old
!= head
;) {
1067 event_t
*event
= (event_t
*)&data
[old
& md
->mask
];
1071 size_t size
= event
->header
.size
;
1074 * Event straddles the mmap boundary -- header should always
1075 * be inside due to u64 alignment of output.
1077 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
1078 unsigned int offset
= old
;
1079 unsigned int len
= min(sizeof(*event
), size
), cpy
;
1080 void *dst
= &event_copy
;
1083 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
1084 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
1090 event
= &event_copy
;
1093 if (event
->header
.type
== PERF_RECORD_SAMPLE
)
1094 event__process_sample(event
, self
, md
->counter
);
1096 event__process(event
, self
);
1103 static struct pollfd event_array
[MAX_NR_CPUS
* MAX_COUNTERS
];
1104 static struct mmap_data mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
1106 static void perf_session__mmap_read(struct perf_session
*self
)
1110 for (i
= 0; i
< nr_cpus
; i
++) {
1111 for (counter
= 0; counter
< nr_counters
; counter
++)
1112 perf_session__mmap_read_counter(self
, &mmap_array
[i
][counter
]);
1119 static void start_counter(int i
, int counter
)
1121 struct perf_event_attr
*attr
;
1125 if (target_pid
== -1 && profile_cpu
== -1)
1128 attr
= attrs
+ counter
;
1130 attr
->sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
1133 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
1135 attr
->sample_freq
= freq
;
1138 attr
->inherit
= (cpu
< 0) && inherit
;
1142 fd
[i
][counter
] = sys_perf_event_open(attr
, target_pid
, cpu
, group_fd
, 0);
1144 if (fd
[i
][counter
] < 0) {
1147 if (err
== EPERM
|| err
== EACCES
)
1148 die("No permission - are you root?\n");
1150 * If it's cycles then fall back to hrtimer
1151 * based cpu-clock-tick sw counter, which
1152 * is always available even if no PMU support:
1154 if (attr
->type
== PERF_TYPE_HARDWARE
1155 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
1158 warning(" ... trying to fall back to cpu-clock-ticks\n");
1160 attr
->type
= PERF_TYPE_SOFTWARE
;
1161 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
1165 error("perfcounter syscall returned with %d (%s)\n",
1166 fd
[i
][counter
], strerror(err
));
1167 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
1170 assert(fd
[i
][counter
] >= 0);
1171 fcntl(fd
[i
][counter
], F_SETFL
, O_NONBLOCK
);
1174 * First counter acts as the group leader:
1176 if (group
&& group_fd
== -1)
1177 group_fd
= fd
[i
][counter
];
1179 event_array
[nr_poll
].fd
= fd
[i
][counter
];
1180 event_array
[nr_poll
].events
= POLLIN
;
1183 mmap_array
[i
][counter
].counter
= counter
;
1184 mmap_array
[i
][counter
].prev
= 0;
1185 mmap_array
[i
][counter
].mask
= mmap_pages
*page_size
- 1;
1186 mmap_array
[i
][counter
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
1187 PROT_READ
, MAP_SHARED
, fd
[i
][counter
], 0);
1188 if (mmap_array
[i
][counter
].base
== MAP_FAILED
)
1189 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
1192 static int __cmd_top(void)
1198 * FIXME: perf_session__new should allow passing a O_MMAP, so that all this
1199 * mmap reading, etc is encapsulated in it. Use O_WRONLY for now.
1201 struct perf_session
*session
= perf_session__new(NULL
, O_WRONLY
, false);
1202 if (session
== NULL
)
1205 if (target_pid
!= -1)
1206 event__synthesize_thread(target_pid
, event__process
, session
);
1208 event__synthesize_threads(event__process
, session
);
1210 for (i
= 0; i
< nr_cpus
; i
++) {
1212 for (counter
= 0; counter
< nr_counters
; counter
++)
1213 start_counter(i
, counter
);
1216 /* Wait for a minimal set of events before starting the snapshot */
1217 poll(event_array
, nr_poll
, 100);
1219 perf_session__mmap_read(session
);
1221 if (pthread_create(&thread
, NULL
, display_thread
, NULL
)) {
1222 printf("Could not create display thread.\n");
1226 if (realtime_prio
) {
1227 struct sched_param param
;
1229 param
.sched_priority
= realtime_prio
;
1230 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1231 printf("Could not set realtime priority.\n");
1239 perf_session__mmap_read(session
);
1241 if (hits
== samples
)
1242 ret
= poll(event_array
, nr_poll
, 100);
1248 static const char * const top_usage
[] = {
1249 "perf top [<options>]",
1253 static const struct option options
[] = {
1254 OPT_CALLBACK('e', "event", NULL
, "event",
1255 "event selector. use 'perf list' to list available events",
1257 OPT_INTEGER('c', "count", &default_interval
,
1258 "event period to sample"),
1259 OPT_INTEGER('p', "pid", &target_pid
,
1260 "profile events on existing pid"),
1261 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1262 "system-wide collection from all CPUs"),
1263 OPT_INTEGER('C', "CPU", &profile_cpu
,
1264 "CPU to profile on"),
1265 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
1266 "file", "vmlinux pathname"),
1267 OPT_BOOLEAN('K', "hide_kernel_symbols", &hide_kernel_symbols
,
1268 "hide kernel symbols"),
1269 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
1270 "number of mmap data pages"),
1271 OPT_INTEGER('r', "realtime", &realtime_prio
,
1272 "collect data with this RT SCHED_FIFO priority"),
1273 OPT_INTEGER('d', "delay", &delay_secs
,
1274 "number of seconds to delay between refreshes"),
1275 OPT_BOOLEAN('D', "dump-symtab", &dump_symtab
,
1276 "dump the symbol table used for profiling"),
1277 OPT_INTEGER('f', "count-filter", &count_filter
,
1278 "only display functions with more events than this"),
1279 OPT_BOOLEAN('g', "group", &group
,
1280 "put the counters into a counter group"),
1281 OPT_BOOLEAN('i', "inherit", &inherit
,
1282 "child tasks inherit counters"),
1283 OPT_STRING('s', "sym-annotate", &sym_filter
, "symbol name",
1284 "symbol to annotate"),
1285 OPT_BOOLEAN('z', "zero", &zero
,
1286 "zero history across updates"),
1287 OPT_INTEGER('F', "freq", &freq
,
1288 "profile at this frequency"),
1289 OPT_INTEGER('E', "entries", &print_entries
,
1290 "display this many functions"),
1291 OPT_BOOLEAN('U', "hide_user_symbols", &hide_user_symbols
,
1292 "hide user symbols"),
1293 OPT_BOOLEAN('v', "verbose", &verbose
,
1294 "be more verbose (show counter open errors, etc)"),
1298 int cmd_top(int argc
, const char **argv
, const char *prefix __used
)
1302 page_size
= sysconf(_SC_PAGE_SIZE
);
1304 argc
= parse_options(argc
, argv
, options
, top_usage
, 0);
1306 usage_with_options(top_usage
, options
);
1308 /* CPU and PID are mutually exclusive */
1309 if (target_pid
!= -1 && profile_cpu
!= -1) {
1310 printf("WARNING: PID switch overriding CPU\n");
1318 symbol_conf
.priv_size
= (sizeof(struct sym_entry
) +
1319 (nr_counters
+ 1) * sizeof(unsigned long));
1321 symbol_conf
.try_vmlinux_path
= (symbol_conf
.vmlinux_name
== NULL
);
1322 if (symbol__init() < 0)
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");
1341 * Fill in the ones not specifically initialized via -c:
1343 for (counter
= 0; counter
< nr_counters
; counter
++) {
1344 if (attrs
[counter
].sample_period
)
1347 attrs
[counter
].sample_period
= default_interval
;
1350 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
1351 assert(nr_cpus
<= MAX_NR_CPUS
);
1352 assert(nr_cpus
>= 0);
1354 if (target_pid
!= -1 || profile_cpu
!= -1)
1357 get_term_dimensions(&winsize
);
1358 if (print_entries
== 0) {
1359 update_print_entries(&winsize
);
1360 signal(SIGWINCH
, sig_winch_handler
);