4 * Builtin annotate command: Analyze the perf.data input file,
5 * look up and read DSOs and symbol information and display
6 * a histogram of results, along various sorting keys.
10 #include "util/util.h"
12 #include "util/color.h"
13 #include <linux/list.h>
14 #include "util/cache.h"
15 #include <linux/rbtree.h>
16 #include "util/symbol.h"
17 #include "util/string.h"
20 #include "util/debug.h"
22 #include "util/parse-options.h"
23 #include "util/parse-events.h"
24 #include "util/thread.h"
26 static char const *input_name
= "perf.data";
28 static char default_sort_order
[] = "comm,symbol";
29 static char *sort_order
= default_sort_order
;
33 static int show_mask
= SHOW_KERNEL
| SHOW_USER
| SHOW_HV
;
35 static int full_paths
;
37 static int print_line
;
39 static unsigned long page_size
;
40 static unsigned long mmap_window
= 32;
42 static struct rb_root threads
;
43 static struct thread
*last_match
;
53 * histogram, sorted on item, collects counts
56 static struct rb_root hist
;
59 struct rb_node rb_node
;
61 struct thread
*thread
;
72 * configurable sorting bits
76 struct list_head list
;
80 int64_t (*cmp
)(struct hist_entry
*, struct hist_entry
*);
81 int64_t (*collapse
)(struct hist_entry
*, struct hist_entry
*);
82 size_t (*print
)(FILE *fp
, struct hist_entry
*);
88 sort__thread_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
90 return right
->thread
->pid
- left
->thread
->pid
;
94 sort__thread_print(FILE *fp
, struct hist_entry
*self
)
96 return fprintf(fp
, "%16s:%5d", self
->thread
->comm
?: "", self
->thread
->pid
);
99 static struct sort_entry sort_thread
= {
100 .header
= " Command: Pid",
101 .cmp
= sort__thread_cmp
,
102 .print
= sort__thread_print
,
108 sort__comm_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
110 return right
->thread
->pid
- left
->thread
->pid
;
114 sort__comm_collapse(struct hist_entry
*left
, struct hist_entry
*right
)
116 char *comm_l
= left
->thread
->comm
;
117 char *comm_r
= right
->thread
->comm
;
119 if (!comm_l
|| !comm_r
) {
120 if (!comm_l
&& !comm_r
)
128 return strcmp(comm_l
, comm_r
);
132 sort__comm_print(FILE *fp
, struct hist_entry
*self
)
134 return fprintf(fp
, "%16s", self
->thread
->comm
);
137 static struct sort_entry sort_comm
= {
138 .header
= " Command",
139 .cmp
= sort__comm_cmp
,
140 .collapse
= sort__comm_collapse
,
141 .print
= sort__comm_print
,
147 sort__dso_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
149 struct dso
*dso_l
= left
->dso
;
150 struct dso
*dso_r
= right
->dso
;
152 if (!dso_l
|| !dso_r
) {
153 if (!dso_l
&& !dso_r
)
161 return strcmp(dso_l
->name
, dso_r
->name
);
165 sort__dso_print(FILE *fp
, struct hist_entry
*self
)
168 return fprintf(fp
, "%-25s", self
->dso
->name
);
170 return fprintf(fp
, "%016llx ", (u64
)self
->ip
);
173 static struct sort_entry sort_dso
= {
174 .header
= "Shared Object ",
175 .cmp
= sort__dso_cmp
,
176 .print
= sort__dso_print
,
182 sort__sym_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
186 if (left
->sym
== right
->sym
)
189 ip_l
= left
->sym
? left
->sym
->start
: left
->ip
;
190 ip_r
= right
->sym
? right
->sym
->start
: right
->ip
;
192 return (int64_t)(ip_r
- ip_l
);
196 sort__sym_print(FILE *fp
, struct hist_entry
*self
)
201 ret
+= fprintf(fp
, "%#018llx ", (u64
)self
->ip
);
204 ret
+= fprintf(fp
, "[%c] %s",
205 self
->dso
== kernel_dso
? 'k' : '.', self
->sym
->name
);
207 ret
+= fprintf(fp
, "%#016llx", (u64
)self
->ip
);
213 static struct sort_entry sort_sym
= {
215 .cmp
= sort__sym_cmp
,
216 .print
= sort__sym_print
,
219 static int sort__need_collapse
= 0;
221 struct sort_dimension
{
223 struct sort_entry
*entry
;
227 static struct sort_dimension sort_dimensions
[] = {
228 { .name
= "pid", .entry
= &sort_thread
, },
229 { .name
= "comm", .entry
= &sort_comm
, },
230 { .name
= "dso", .entry
= &sort_dso
, },
231 { .name
= "symbol", .entry
= &sort_sym
, },
234 static LIST_HEAD(hist_entry__sort_list
);
236 static int sort_dimension__add(char *tok
)
240 for (i
= 0; i
< ARRAY_SIZE(sort_dimensions
); i
++) {
241 struct sort_dimension
*sd
= &sort_dimensions
[i
];
246 if (strncasecmp(tok
, sd
->name
, strlen(tok
)))
249 if (sd
->entry
->collapse
)
250 sort__need_collapse
= 1;
252 list_add_tail(&sd
->entry
->list
, &hist_entry__sort_list
);
262 hist_entry__cmp(struct hist_entry
*left
, struct hist_entry
*right
)
264 struct sort_entry
*se
;
267 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
268 cmp
= se
->cmp(left
, right
);
277 hist_entry__collapse(struct hist_entry
*left
, struct hist_entry
*right
)
279 struct sort_entry
*se
;
282 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
283 int64_t (*f
)(struct hist_entry
*, struct hist_entry
*);
285 f
= se
->collapse
?: se
->cmp
;
287 cmp
= f(left
, right
);
296 * collect histogram counts
298 static void hist_hit(struct hist_entry
*he
, u64 ip
)
300 unsigned int sym_size
, offset
;
301 struct symbol
*sym
= he
->sym
;
305 if (!sym
|| !sym
->hist
)
308 sym_size
= sym
->end
- sym
->start
;
309 offset
= ip
- sym
->start
;
311 if (offset
>= sym_size
)
318 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
319 (void *)(unsigned long)he
->sym
->start
,
321 (void *)(unsigned long)ip
, ip
- he
->sym
->start
,
326 hist_entry__add(struct thread
*thread
, struct map
*map
, struct dso
*dso
,
327 struct symbol
*sym
, u64 ip
, char level
)
329 struct rb_node
**p
= &hist
.rb_node
;
330 struct rb_node
*parent
= NULL
;
331 struct hist_entry
*he
;
332 struct hist_entry entry
= {
345 he
= rb_entry(parent
, struct hist_entry
, rb_node
);
347 cmp
= hist_entry__cmp(&entry
, he
);
361 he
= malloc(sizeof(*he
));
365 rb_link_node(&he
->rb_node
, parent
, p
);
366 rb_insert_color(&he
->rb_node
, &hist
);
371 static void hist_entry__free(struct hist_entry
*he
)
377 * collapse the histogram
380 static struct rb_root collapse_hists
;
382 static void collapse__insert_entry(struct hist_entry
*he
)
384 struct rb_node
**p
= &collapse_hists
.rb_node
;
385 struct rb_node
*parent
= NULL
;
386 struct hist_entry
*iter
;
391 iter
= rb_entry(parent
, struct hist_entry
, rb_node
);
393 cmp
= hist_entry__collapse(iter
, he
);
396 iter
->count
+= he
->count
;
397 hist_entry__free(he
);
407 rb_link_node(&he
->rb_node
, parent
, p
);
408 rb_insert_color(&he
->rb_node
, &collapse_hists
);
411 static void collapse__resort(void)
413 struct rb_node
*next
;
414 struct hist_entry
*n
;
416 if (!sort__need_collapse
)
419 next
= rb_first(&hist
);
421 n
= rb_entry(next
, struct hist_entry
, rb_node
);
422 next
= rb_next(&n
->rb_node
);
424 rb_erase(&n
->rb_node
, &hist
);
425 collapse__insert_entry(n
);
430 * reverse the map, sort on count.
433 static struct rb_root output_hists
;
435 static void output__insert_entry(struct hist_entry
*he
)
437 struct rb_node
**p
= &output_hists
.rb_node
;
438 struct rb_node
*parent
= NULL
;
439 struct hist_entry
*iter
;
443 iter
= rb_entry(parent
, struct hist_entry
, rb_node
);
445 if (he
->count
> iter
->count
)
451 rb_link_node(&he
->rb_node
, parent
, p
);
452 rb_insert_color(&he
->rb_node
, &output_hists
);
455 static void output__resort(void)
457 struct rb_node
*next
;
458 struct hist_entry
*n
;
459 struct rb_root
*tree
= &hist
;
461 if (sort__need_collapse
)
462 tree
= &collapse_hists
;
464 next
= rb_first(tree
);
467 n
= rb_entry(next
, struct hist_entry
, rb_node
);
468 next
= rb_next(&n
->rb_node
);
470 rb_erase(&n
->rb_node
, tree
);
471 output__insert_entry(n
);
475 static unsigned long total
= 0,
482 process_sample_event(event_t
*event
, unsigned long offset
, unsigned long head
)
486 struct dso
*dso
= NULL
;
487 struct thread
*thread
;
488 u64 ip
= event
->ip
.ip
;
489 struct map
*map
= NULL
;
491 thread
= threads__findnew(event
->ip
.pid
, &threads
, &last_match
);
493 dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
494 (void *)(offset
+ head
),
495 (void *)(long)(event
->header
.size
),
500 dump_printf(" ... thread: %s:%d\n", thread
->comm
, thread
->pid
);
502 if (thread
== NULL
) {
503 fprintf(stderr
, "problem processing %d event, skipping it.\n",
508 if (event
->header
.misc
& PERF_EVENT_MISC_KERNEL
) {
514 dump_printf(" ...... dso: %s\n", dso
->name
);
516 } else if (event
->header
.misc
& PERF_EVENT_MISC_USER
) {
521 map
= thread__find_map(thread
, ip
);
523 ip
= map
->map_ip(map
, ip
);
527 * If this is outside of all known maps,
528 * and is a negative address, try to look it
529 * up in the kernel dso, as it might be a
530 * vsyscall (which executes in user-mode):
532 if ((long long)ip
< 0)
535 dump_printf(" ...... dso: %s\n", dso
? dso
->name
: "<not found>");
540 dump_printf(" ...... dso: [hypervisor]\n");
543 if (show
& show_mask
) {
544 struct symbol
*sym
= NULL
;
547 sym
= dso
->find_symbol(dso
, ip
);
549 if (hist_entry__add(thread
, map
, dso
, sym
, ip
, level
)) {
551 "problem incrementing symbol count, skipping event\n");
561 process_mmap_event(event_t
*event
, unsigned long offset
, unsigned long head
)
563 struct thread
*thread
;
564 struct map
*map
= map__new(&event
->mmap
, NULL
, 0);
566 thread
= threads__findnew(event
->mmap
.pid
, &threads
, &last_match
);
568 dump_printf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
569 (void *)(offset
+ head
),
570 (void *)(long)(event
->header
.size
),
572 (void *)(long)event
->mmap
.start
,
573 (void *)(long)event
->mmap
.len
,
574 (void *)(long)event
->mmap
.pgoff
,
575 event
->mmap
.filename
);
577 if (thread
== NULL
|| map
== NULL
) {
578 dump_printf("problem processing PERF_EVENT_MMAP, skipping event.\n");
582 thread__insert_map(thread
, map
);
589 process_comm_event(event_t
*event
, unsigned long offset
, unsigned long head
)
591 struct thread
*thread
;
593 thread
= threads__findnew(event
->comm
.pid
, &threads
, &last_match
);
594 dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
595 (void *)(offset
+ head
),
596 (void *)(long)(event
->header
.size
),
597 event
->comm
.comm
, event
->comm
.pid
);
599 if (thread
== NULL
||
600 thread__set_comm(thread
, event
->comm
.comm
)) {
601 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
610 process_fork_event(event_t
*event
, unsigned long offset
, unsigned long head
)
612 struct thread
*thread
;
613 struct thread
*parent
;
615 thread
= threads__findnew(event
->fork
.pid
, &threads
, &last_match
);
616 parent
= threads__findnew(event
->fork
.ppid
, &threads
, &last_match
);
617 dump_printf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
618 (void *)(offset
+ head
),
619 (void *)(long)(event
->header
.size
),
620 event
->fork
.pid
, event
->fork
.ppid
);
623 * A thread clone will have the same PID for both
626 if (thread
== parent
)
629 if (!thread
|| !parent
|| thread__fork(thread
, parent
)) {
630 dump_printf("problem processing PERF_EVENT_FORK, skipping event.\n");
639 process_event(event_t
*event
, unsigned long offset
, unsigned long head
)
641 switch (event
->header
.type
) {
642 case PERF_EVENT_SAMPLE
:
643 return process_sample_event(event
, offset
, head
);
645 case PERF_EVENT_MMAP
:
646 return process_mmap_event(event
, offset
, head
);
648 case PERF_EVENT_COMM
:
649 return process_comm_event(event
, offset
, head
);
651 case PERF_EVENT_FORK
:
652 return process_fork_event(event
, offset
, head
);
654 * We dont process them right now but they are fine:
657 case PERF_EVENT_THROTTLE
:
658 case PERF_EVENT_UNTHROTTLE
:
669 parse_line(FILE *file
, struct symbol
*sym
, u64 start
, u64 len
)
671 char *line
= NULL
, *tmp
, *tmp2
;
672 static const char *prev_line
;
673 static const char *prev_color
;
680 if (getline(&line
, &line_len
, file
) < 0)
685 c
= strchr(line
, '\n');
694 * Strip leading spaces:
705 * Parse hexa addresses followed by ':'
707 line_ip
= strtoull(tmp
, &tmp2
, 16);
713 const char *path
= NULL
;
714 unsigned int hits
= 0;
715 double percent
= 0.0;
717 struct sym_ext
*sym_ext
= sym
->priv
;
719 offset
= line_ip
- start
;
721 hits
= sym
->hist
[offset
];
723 if (offset
< len
&& sym_ext
) {
724 path
= sym_ext
[offset
].path
;
725 percent
= sym_ext
[offset
].percent
;
726 } else if (sym
->hist_sum
)
727 percent
= 100.0 * hits
/ sym
->hist_sum
;
729 color
= get_percent_color(percent
);
732 * Also color the filename and line if needed, with
733 * the same color than the percentage. Don't print it
734 * twice for close colored ip with the same filename:line
737 if (!prev_line
|| strcmp(prev_line
, path
)
738 || color
!= prev_color
) {
739 color_fprintf(stdout
, color
, " %s", path
);
745 color_fprintf(stdout
, color
, " %7.2f", percent
);
747 color_fprintf(stdout
, PERF_COLOR_BLUE
, "%s\n", line
);
752 printf(" : %s\n", line
);
758 static struct rb_root root_sym_ext
;
760 static void insert_source_line(struct sym_ext
*sym_ext
)
762 struct sym_ext
*iter
;
763 struct rb_node
**p
= &root_sym_ext
.rb_node
;
764 struct rb_node
*parent
= NULL
;
768 iter
= rb_entry(parent
, struct sym_ext
, node
);
770 if (sym_ext
->percent
> iter
->percent
)
776 rb_link_node(&sym_ext
->node
, parent
, p
);
777 rb_insert_color(&sym_ext
->node
, &root_sym_ext
);
780 static void free_source_line(struct symbol
*sym
, int len
)
782 struct sym_ext
*sym_ext
= sym
->priv
;
788 for (i
= 0; i
< len
; i
++)
789 free(sym_ext
[i
].path
);
793 root_sym_ext
= RB_ROOT
;
796 /* Get the filename:line for the colored entries */
798 get_source_line(struct symbol
*sym
, u64 start
, int len
, const char *filename
)
801 char cmd
[PATH_MAX
* 2];
802 struct sym_ext
*sym_ext
;
807 sym
->priv
= calloc(len
, sizeof(struct sym_ext
));
813 for (i
= 0; i
< len
; i
++) {
819 sym_ext
[i
].percent
= 100.0 * sym
->hist
[i
] / sym
->hist_sum
;
820 if (sym_ext
[i
].percent
<= 0.5)
824 sprintf(cmd
, "addr2line -e %s %016llx", filename
, offset
);
825 fp
= popen(cmd
, "r");
829 if (getline(&path
, &line_len
, fp
) < 0 || !line_len
)
832 sym_ext
[i
].path
= malloc(sizeof(char) * line_len
+ 1);
833 if (!sym_ext
[i
].path
)
836 strcpy(sym_ext
[i
].path
, path
);
837 insert_source_line(&sym_ext
[i
]);
844 static void print_summary(const char *filename
)
846 struct sym_ext
*sym_ext
;
847 struct rb_node
*node
;
849 printf("\nSorted summary for file %s\n", filename
);
850 printf("----------------------------------------------\n\n");
852 if (RB_EMPTY_ROOT(&root_sym_ext
)) {
853 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN
);
857 node
= rb_first(&root_sym_ext
);
863 sym_ext
= rb_entry(node
, struct sym_ext
, node
);
864 percent
= sym_ext
->percent
;
865 color
= get_percent_color(percent
);
866 path
= sym_ext
->path
;
868 color_fprintf(stdout
, color
, " %7.2f %s", percent
, path
);
869 node
= rb_next(node
);
873 static void annotate_sym(struct dso
*dso
, struct symbol
*sym
)
875 const char *filename
= dso
->name
, *d_filename
;
877 char command
[PATH_MAX
*2];
883 filename
= sym
->module
->path
;
884 else if (dso
== kernel_dso
)
885 filename
= vmlinux_name
;
887 start
= sym
->obj_start
;
891 d_filename
= filename
;
893 d_filename
= basename(filename
);
895 end
= start
+ sym
->end
- sym
->start
+ 1;
896 len
= sym
->end
- sym
->start
;
899 get_source_line(sym
, start
, len
, filename
);
900 print_summary(filename
);
903 printf("\n\n------------------------------------------------\n");
904 printf(" Percent | Source code & Disassembly of %s\n", d_filename
);
905 printf("------------------------------------------------\n");
908 printf("annotating [%p] %30s : [%p] %30s\n", dso
, dso
->name
, sym
, sym
->name
);
910 sprintf(command
, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
911 (u64
)start
, (u64
)end
, filename
, filename
);
914 printf("doing: %s\n", command
);
916 file
= popen(command
, "r");
920 while (!feof(file
)) {
921 if (parse_line(file
, sym
, start
, len
) < 0)
927 free_source_line(sym
, len
);
930 static void find_annotations(void)
936 list_for_each_entry(dso
, &dsos
, node
) {
938 for (nd
= rb_first(&dso
->syms
); nd
; nd
= rb_next(nd
)) {
939 struct symbol
*sym
= rb_entry(nd
, struct symbol
, rb_node
);
942 annotate_sym(dso
, sym
);
949 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter
);
952 static int __cmd_annotate(void)
954 int ret
, rc
= EXIT_FAILURE
;
955 unsigned long offset
= 0;
956 unsigned long head
= 0;
957 struct stat input_stat
;
962 register_idle_thread(&threads
, &last_match
);
964 input
= open(input_name
, O_RDONLY
);
966 perror("failed to open file");
970 ret
= fstat(input
, &input_stat
);
972 perror("failed to stat file");
976 if (!force
&& input_stat
.st_uid
&& (input_stat
.st_uid
!= geteuid())) {
977 fprintf(stderr
, "file: %s not owned by current user or root\n", input_name
);
981 if (!input_stat
.st_size
) {
982 fprintf(stderr
, "zero-sized file, nothing to do!\n");
986 if (load_kernel() < 0) {
987 perror("failed to load kernel symbols");
992 buf
= (char *)mmap(NULL
, page_size
* mmap_window
, PROT_READ
,
993 MAP_SHARED
, input
, offset
);
994 if (buf
== MAP_FAILED
) {
995 perror("failed to mmap file");
1000 event
= (event_t
*)(buf
+ head
);
1002 size
= event
->header
.size
;
1006 if (head
+ event
->header
.size
>= page_size
* mmap_window
) {
1007 unsigned long shift
= page_size
* (head
/ page_size
);
1010 munmap_ret
= munmap(buf
, page_size
* mmap_window
);
1011 assert(munmap_ret
== 0);
1018 size
= event
->header
.size
;
1020 dump_printf("%p [%p]: event: %d\n",
1021 (void *)(offset
+ head
),
1022 (void *)(long)event
->header
.size
,
1023 event
->header
.type
);
1025 if (!size
|| process_event(event
, offset
, head
) < 0) {
1027 dump_printf("%p [%p]: skipping unknown header type: %d\n",
1028 (void *)(offset
+ head
),
1029 (void *)(long)(event
->header
.size
),
1030 event
->header
.type
);
1035 * assume we lost track of the stream, check alignment, and
1036 * increment a single u64 in the hope to catch on again 'soon'.
1039 if (unlikely(head
& 7))
1047 if (offset
+ head
< (unsigned long)input_stat
.st_size
)
1053 dump_printf(" IP events: %10ld\n", total
);
1054 dump_printf(" mmap events: %10ld\n", total_mmap
);
1055 dump_printf(" comm events: %10ld\n", total_comm
);
1056 dump_printf(" fork events: %10ld\n", total_fork
);
1057 dump_printf(" unknown events: %10ld\n", total_unknown
);
1063 threads__fprintf(stdout
, &threads
);
1066 dsos__fprintf(stdout
);
1076 static const char * const annotate_usage
[] = {
1077 "perf annotate [<options>] <command>",
1081 static const struct option options
[] = {
1082 OPT_STRING('i', "input", &input_name
, "file",
1084 OPT_STRING('s', "symbol", &sym_hist_filter
, "symbol",
1085 "symbol to annotate"),
1086 OPT_BOOLEAN('f', "force", &force
, "don't complain, do it"),
1087 OPT_BOOLEAN('v', "verbose", &verbose
,
1088 "be more verbose (show symbol address, etc)"),
1089 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace
,
1090 "dump raw trace in ASCII"),
1091 OPT_STRING('k', "vmlinux", &vmlinux_name
, "file", "vmlinux pathname"),
1092 OPT_BOOLEAN('m', "modules", &modules
,
1093 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1094 OPT_BOOLEAN('l', "print-line", &print_line
,
1095 "print matching source lines (may be slow)"),
1096 OPT_BOOLEAN('P', "full-paths", &full_paths
,
1097 "Don't shorten the displayed pathnames"),
1101 static void setup_sorting(void)
1103 char *tmp
, *tok
, *str
= strdup(sort_order
);
1105 for (tok
= strtok_r(str
, ", ", &tmp
);
1106 tok
; tok
= strtok_r(NULL
, ", ", &tmp
)) {
1107 if (sort_dimension__add(tok
) < 0) {
1108 error("Unknown --sort key: `%s'", tok
);
1109 usage_with_options(annotate_usage
, options
);
1116 int cmd_annotate(int argc
, const char **argv
, const char *prefix __used
)
1120 page_size
= getpagesize();
1122 argc
= parse_options(argc
, argv
, options
, annotate_usage
, 0);
1128 * Special case: if there's an argument left then assume tha
1129 * it's a symbol filter:
1132 usage_with_options(annotate_usage
, options
);
1134 sym_hist_filter
= argv
[0];
1137 if (!sym_hist_filter
)
1138 usage_with_options(annotate_usage
, options
);
1142 return __cmd_annotate();