4 * Builtin report 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/list.h"
13 #include "util/cache.h"
14 #include "util/rbtree.h"
15 #include "util/symbol.h"
16 #include "util/string.h"
20 #include "util/parse-options.h"
21 #include "util/parse-events.h"
27 static char const *input_name
= "perf.data";
28 static char *vmlinux
= NULL
;
29 static char *sort_order
= "comm,dso";
31 static int show_mask
= SHOW_KERNEL
| SHOW_USER
| SHOW_HV
;
33 static int dump_trace
= 0;
34 #define dprintf(x...) do { if (dump_trace) printf(x); } while (0)
37 static int full_paths
;
39 static unsigned long page_size
;
40 static unsigned long mmap_window
= 32;
42 const char *perf_event_names
[] = {
43 [PERF_EVENT_MMAP
] = " PERF_EVENT_MMAP",
44 [PERF_EVENT_MUNMAP
] = " PERF_EVENT_MUNMAP",
45 [PERF_EVENT_COMM
] = " PERF_EVENT_COMM",
49 struct perf_event_header header
;
54 struct perf_event_header header
;
59 char filename
[PATH_MAX
];
62 struct perf_event_header header
;
67 typedef union event_union
{
68 struct perf_event_header header
;
70 struct mmap_event mmap
;
71 struct comm_event comm
;
74 static LIST_HEAD(dsos
);
75 static struct dso
*kernel_dso
;
77 static void dsos__add(struct dso
*dso
)
79 list_add_tail(&dso
->node
, &dsos
);
82 static struct dso
*dsos__find(const char *name
)
86 list_for_each_entry(pos
, &dsos
, node
)
87 if (strcmp(pos
->name
, name
) == 0)
92 static struct dso
*dsos__findnew(const char *name
)
94 struct dso
*dso
= dsos__find(name
);
100 dso
= dso__new(name
, 0);
104 nr
= dso__load(dso
, NULL
);
106 fprintf(stderr
, "Failed to open: %s\n", name
);
109 if (!nr
&& verbose
) {
111 "No symbols found in: %s, maybe install a debug package?\n",
124 static void dsos__fprintf(FILE *fp
)
128 list_for_each_entry(pos
, &dsos
, node
)
129 dso__fprintf(pos
, fp
);
132 static int load_kernel(void)
136 kernel_dso
= dso__new("[kernel]", 0);
140 err
= dso__load_kernel(kernel_dso
, vmlinux
, NULL
);
142 dso__delete(kernel_dso
);
145 dsos__add(kernel_dso
);
150 static int strcommon(const char *pathname
, const char *cwd
, int cwdlen
)
154 while (pathname
[n
] == cwd
[n
] && n
< cwdlen
)
161 struct list_head node
;
168 static struct map
*map__new(struct mmap_event
*event
, char *cwd
, int cwdlen
)
170 struct map
*self
= malloc(sizeof(*self
));
173 const char *filename
= event
->filename
;
174 char newfilename
[PATH_MAX
];
177 int n
= strcommon(filename
, cwd
, cwdlen
);
179 snprintf(newfilename
, sizeof(newfilename
),
180 ".%s", filename
+ n
);
181 filename
= newfilename
;
185 self
->start
= event
->start
;
186 self
->end
= event
->start
+ event
->len
;
187 self
->pgoff
= event
->pgoff
;
189 self
->dso
= dsos__findnew(filename
);
190 if (self
->dso
== NULL
)
202 struct rb_node rb_node
;
203 struct list_head maps
;
208 static struct thread
*thread__new(pid_t pid
)
210 struct thread
*self
= malloc(sizeof(*self
));
214 self
->comm
= malloc(32);
216 snprintf(self
->comm
, 32, ":%d", self
->pid
);
217 INIT_LIST_HEAD(&self
->maps
);
223 static int thread__set_comm(struct thread
*self
, const char *comm
)
227 self
->comm
= strdup(comm
);
228 return self
->comm
? 0 : -ENOMEM
;
231 static struct rb_root threads
;
232 static struct thread
*last_match
;
234 static struct thread
*threads__findnew(pid_t pid
)
236 struct rb_node
**p
= &threads
.rb_node
;
237 struct rb_node
*parent
= NULL
;
241 * Font-end cache - PID lookups come in blocks,
242 * so most of the time we dont have to look up
245 if (last_match
&& last_match
->pid
== pid
)
250 th
= rb_entry(parent
, struct thread
, rb_node
);
252 if (th
->pid
== pid
) {
263 th
= thread__new(pid
);
265 rb_link_node(&th
->rb_node
, parent
, p
);
266 rb_insert_color(&th
->rb_node
, &threads
);
273 static void thread__insert_map(struct thread
*self
, struct map
*map
)
275 list_add_tail(&map
->node
, &self
->maps
);
278 static struct map
*thread__find_map(struct thread
*self
, uint64_t ip
)
285 list_for_each_entry(pos
, &self
->maps
, node
)
286 if (ip
>= pos
->start
&& ip
<= pos
->end
)
293 * histogram, sorted on item, collects counts
296 static struct rb_root hist
;
299 struct rb_node rb_node
;
301 struct thread
*thread
;
312 * configurable sorting bits
316 struct list_head list
;
320 int64_t (*cmp
)(struct hist_entry
*, struct hist_entry
*);
321 int64_t (*collapse
)(struct hist_entry
*, struct hist_entry
*);
322 size_t (*print
)(FILE *fp
, struct hist_entry
*);
328 sort__thread_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
330 return right
->thread
->pid
- left
->thread
->pid
;
334 sort__thread_print(FILE *fp
, struct hist_entry
*self
)
336 return fprintf(fp
, " %16s:%5d", self
->thread
->comm
?: "", self
->thread
->pid
);
339 static struct sort_entry sort_thread
= {
340 .header
= " Command: Pid ",
341 .cmp
= sort__thread_cmp
,
342 .print
= sort__thread_print
,
348 sort__comm_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
350 return right
->thread
->pid
- left
->thread
->pid
;
354 sort__comm_collapse(struct hist_entry
*left
, struct hist_entry
*right
)
356 char *comm_l
= left
->thread
->comm
;
357 char *comm_r
= right
->thread
->comm
;
359 if (!comm_l
|| !comm_r
) {
360 if (!comm_l
&& !comm_r
)
368 return strcmp(comm_l
, comm_r
);
372 sort__comm_print(FILE *fp
, struct hist_entry
*self
)
374 return fprintf(fp
, " %16s", self
->thread
->comm
);
377 static struct sort_entry sort_comm
= {
378 .header
= " Command",
379 .cmp
= sort__comm_cmp
,
380 .collapse
= sort__comm_collapse
,
381 .print
= sort__comm_print
,
387 sort__dso_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
389 struct dso
*dso_l
= left
->dso
;
390 struct dso
*dso_r
= right
->dso
;
392 if (!dso_l
|| !dso_r
) {
393 if (!dso_l
&& !dso_r
)
401 return strcmp(dso_l
->name
, dso_r
->name
);
405 sort__dso_print(FILE *fp
, struct hist_entry
*self
)
408 return fprintf(fp
, " %-25s", self
->dso
->name
);
410 return fprintf(fp
, " %016llx", (__u64
)self
->ip
);
413 static struct sort_entry sort_dso
= {
414 .header
= " Shared Object ",
415 .cmp
= sort__dso_cmp
,
416 .print
= sort__dso_print
,
422 sort__sym_cmp(struct hist_entry
*left
, struct hist_entry
*right
)
426 if (left
->sym
== right
->sym
)
429 ip_l
= left
->sym
? left
->sym
->start
: left
->ip
;
430 ip_r
= right
->sym
? right
->sym
->start
: right
->ip
;
432 return (int64_t)(ip_r
- ip_l
);
436 sort__sym_print(FILE *fp
, struct hist_entry
*self
)
441 ret
+= fprintf(fp
, " %#018llx", (__u64
)self
->ip
);
444 ret
+= fprintf(fp
, " %s: ", self
->dso
->name
);
446 ret
+= fprintf(fp
, " %#016llx: ", (__u64
)self
->ip
);
449 ret
+= fprintf(fp
, "%s", self
->sym
->name
);
451 ret
+= fprintf(fp
, "%#016llx", (__u64
)self
->ip
);
456 static struct sort_entry sort_sym
= {
457 .header
= " Shared Object: Symbol",
458 .cmp
= sort__sym_cmp
,
459 .print
= sort__sym_print
,
462 static int sort__need_collapse
= 0;
464 struct sort_dimension
{
466 struct sort_entry
*entry
;
470 static struct sort_dimension sort_dimensions
[] = {
471 { .name
= "pid", .entry
= &sort_thread
, },
472 { .name
= "comm", .entry
= &sort_comm
, },
473 { .name
= "dso", .entry
= &sort_dso
, },
474 { .name
= "symbol", .entry
= &sort_sym
, },
477 static LIST_HEAD(hist_entry__sort_list
);
479 static int sort_dimension__add(char *tok
)
483 for (i
= 0; i
< ARRAY_SIZE(sort_dimensions
); i
++) {
484 struct sort_dimension
*sd
= &sort_dimensions
[i
];
489 if (strncasecmp(tok
, sd
->name
, strlen(tok
)))
492 if (sd
->entry
->collapse
)
493 sort__need_collapse
= 1;
495 list_add_tail(&sd
->entry
->list
, &hist_entry__sort_list
);
505 hist_entry__cmp(struct hist_entry
*left
, struct hist_entry
*right
)
507 struct sort_entry
*se
;
510 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
511 cmp
= se
->cmp(left
, right
);
520 hist_entry__collapse(struct hist_entry
*left
, struct hist_entry
*right
)
522 struct sort_entry
*se
;
525 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
526 int64_t (*f
)(struct hist_entry
*, struct hist_entry
*);
528 f
= se
->collapse
?: se
->cmp
;
530 cmp
= f(left
, right
);
539 hist_entry__fprintf(FILE *fp
, struct hist_entry
*self
, uint64_t total_samples
)
541 struct sort_entry
*se
;
545 ret
= fprintf(fp
, " %6.2f%%",
546 (self
->count
* 100.0) / total_samples
);
548 ret
= fprintf(fp
, "%12d ", self
->count
);
550 list_for_each_entry(se
, &hist_entry__sort_list
, list
)
551 ret
+= se
->print(fp
, self
);
553 ret
+= fprintf(fp
, "\n");
559 * collect histogram counts
563 hist_entry__add(struct thread
*thread
, struct map
*map
, struct dso
*dso
,
564 struct symbol
*sym
, uint64_t ip
, char level
)
566 struct rb_node
**p
= &hist
.rb_node
;
567 struct rb_node
*parent
= NULL
;
568 struct hist_entry
*he
;
569 struct hist_entry entry
= {
582 he
= rb_entry(parent
, struct hist_entry
, rb_node
);
584 cmp
= hist_entry__cmp(&entry
, he
);
597 he
= malloc(sizeof(*he
));
601 rb_link_node(&he
->rb_node
, parent
, p
);
602 rb_insert_color(&he
->rb_node
, &hist
);
607 static void hist_entry__free(struct hist_entry
*he
)
613 * collapse the histogram
616 static struct rb_root collapse_hists
;
618 static void collapse__insert_entry(struct hist_entry
*he
)
620 struct rb_node
**p
= &collapse_hists
.rb_node
;
621 struct rb_node
*parent
= NULL
;
622 struct hist_entry
*iter
;
627 iter
= rb_entry(parent
, struct hist_entry
, rb_node
);
629 cmp
= hist_entry__collapse(iter
, he
);
632 iter
->count
+= he
->count
;
633 hist_entry__free(he
);
643 rb_link_node(&he
->rb_node
, parent
, p
);
644 rb_insert_color(&he
->rb_node
, &collapse_hists
);
647 static void collapse__resort(void)
649 struct rb_node
*next
;
650 struct hist_entry
*n
;
652 if (!sort__need_collapse
)
655 next
= rb_first(&hist
);
657 n
= rb_entry(next
, struct hist_entry
, rb_node
);
658 next
= rb_next(&n
->rb_node
);
660 rb_erase(&n
->rb_node
, &hist
);
661 collapse__insert_entry(n
);
666 * reverse the map, sort on count.
669 static struct rb_root output_hists
;
671 static void output__insert_entry(struct hist_entry
*he
)
673 struct rb_node
**p
= &output_hists
.rb_node
;
674 struct rb_node
*parent
= NULL
;
675 struct hist_entry
*iter
;
679 iter
= rb_entry(parent
, struct hist_entry
, rb_node
);
681 if (he
->count
> iter
->count
)
687 rb_link_node(&he
->rb_node
, parent
, p
);
688 rb_insert_color(&he
->rb_node
, &output_hists
);
691 static void output__resort(void)
693 struct rb_node
*next
;
694 struct hist_entry
*n
;
696 if (sort__need_collapse
)
697 next
= rb_first(&collapse_hists
);
699 next
= rb_first(&hist
);
702 n
= rb_entry(next
, struct hist_entry
, rb_node
);
703 next
= rb_next(&n
->rb_node
);
705 rb_erase(&n
->rb_node
, &hist
);
706 output__insert_entry(n
);
710 static size_t output__fprintf(FILE *fp
, uint64_t total_samples
)
712 struct hist_entry
*pos
;
713 struct sort_entry
*se
;
719 fprintf(fp
, "# Overhead");
720 list_for_each_entry(se
, &hist_entry__sort_list
, list
)
721 fprintf(fp
, " %s", se
->header
);
724 fprintf(fp
, "# ........");
725 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
729 for (i
= 0; i
< strlen(se
->header
)-1; i
++)
736 for (nd
= rb_first(&output_hists
); nd
; nd
= rb_next(nd
)) {
737 pos
= rb_entry(nd
, struct hist_entry
, rb_node
);
738 ret
+= hist_entry__fprintf(fp
, pos
, total_samples
);
744 static void register_idle_thread(void)
746 struct thread
*thread
= threads__findnew(0);
748 if (thread
== NULL
||
749 thread__set_comm(thread
, "[idle]")) {
750 fprintf(stderr
, "problem inserting idle task.\n");
756 static int __cmd_report(void)
758 unsigned long offset
= 0;
759 unsigned long head
= 0;
763 int ret
, rc
= EXIT_FAILURE
;
765 unsigned long total
= 0, total_mmap
= 0, total_comm
= 0, total_unknown
= 0;
766 char cwd
[PATH_MAX
], *cwdp
= cwd
;
769 register_idle_thread();
771 input
= open(input_name
, O_RDONLY
);
773 perror("failed to open file");
777 ret
= fstat(input
, &stat
);
779 perror("failed to stat file");
784 fprintf(stderr
, "zero-sized file, nothing to do!\n");
788 if (load_kernel() < 0) {
789 perror("failed to load kernel symbols");
794 if (getcwd(cwd
, sizeof(cwd
)) == NULL
) {
795 perror("failed to get the current directory");
798 cwdlen
= strlen(cwd
);
804 buf
= (char *)mmap(NULL
, page_size
* mmap_window
, PROT_READ
,
805 MAP_SHARED
, input
, offset
);
806 if (buf
== MAP_FAILED
) {
807 perror("failed to mmap file");
812 event
= (event_t
*)(buf
+ head
);
814 size
= event
->header
.size
;
818 if (head
+ event
->header
.size
>= page_size
* mmap_window
) {
819 unsigned long shift
= page_size
* (head
/ page_size
);
822 ret
= munmap(buf
, page_size
* mmap_window
);
830 size
= event
->header
.size
;
834 if (event
->header
.misc
& PERF_EVENT_MISC_OVERFLOW
) {
837 struct dso
*dso
= NULL
;
838 struct thread
*thread
= threads__findnew(event
->ip
.pid
);
839 uint64_t ip
= event
->ip
.ip
;
840 struct map
*map
= NULL
;
842 dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
843 (void *)(offset
+ head
),
844 (void *)(long)(event
->header
.size
),
849 dprintf(" ... thread: %s:%d\n", thread
->comm
, thread
->pid
);
851 if (thread
== NULL
) {
852 fprintf(stderr
, "problem processing %d event, skipping it.\n",
857 if (event
->header
.misc
& PERF_EVENT_MISC_KERNEL
) {
863 dprintf(" ...... dso: %s\n", dso
->name
);
865 } else if (event
->header
.misc
& PERF_EVENT_MISC_USER
) {
870 map
= thread__find_map(thread
, ip
);
873 ip
-= map
->start
+ map
->pgoff
;
876 * If this is outside of all known maps,
877 * and is a negative address, try to look it
878 * up in the kernel dso, as it might be a
879 * vsyscall (which executes in user-mode):
881 if ((long long)ip
< 0)
884 dprintf(" ...... dso: %s\n", dso
? dso
->name
: "<not found>");
889 dprintf(" ...... dso: [hypervisor]\n");
892 if (show
& show_mask
) {
893 struct symbol
*sym
= dso__find_symbol(dso
, ip
);
895 if (hist_entry__add(thread
, map
, dso
, sym
, ip
, level
)) {
897 "problem incrementing symbol count, skipping event\n");
902 } else switch (event
->header
.type
) {
903 case PERF_EVENT_MMAP
: {
904 struct thread
*thread
= threads__findnew(event
->mmap
.pid
);
905 struct map
*map
= map__new(&event
->mmap
, cwdp
, cwdlen
);
907 dprintf("%p [%p]: PERF_EVENT_MMAP: [%p(%p) @ %p]: %s\n",
908 (void *)(offset
+ head
),
909 (void *)(long)(event
->header
.size
),
910 (void *)(long)event
->mmap
.start
,
911 (void *)(long)event
->mmap
.len
,
912 (void *)(long)event
->mmap
.pgoff
,
913 event
->mmap
.filename
);
915 if (thread
== NULL
|| map
== NULL
) {
917 fprintf(stderr
, "problem processing PERF_EVENT_MMAP, skipping event.\n");
920 thread__insert_map(thread
, map
);
924 case PERF_EVENT_COMM
: {
925 struct thread
*thread
= threads__findnew(event
->comm
.pid
);
927 dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
928 (void *)(offset
+ head
),
929 (void *)(long)(event
->header
.size
),
930 event
->comm
.comm
, event
->comm
.pid
);
932 if (thread
== NULL
||
933 thread__set_comm(thread
, event
->comm
.comm
)) {
934 fprintf(stderr
, "problem processing PERF_EVENT_COMM, skipping event.\n");
942 dprintf("%p [%p]: skipping unknown header type: %d\n",
943 (void *)(offset
+ head
),
944 (void *)(long)(event
->header
.size
),
950 * assume we lost track of the stream, check alignment, and
951 * increment a single u64 in the hope to catch on again 'soon'.
954 if (unlikely(head
& 7))
963 if (offset
+ head
< stat
.st_size
)
969 dprintf(" IP events: %10ld\n", total
);
970 dprintf(" mmap events: %10ld\n", total_mmap
);
971 dprintf(" comm events: %10ld\n", total_comm
);
972 dprintf(" unknown events: %10ld\n", total_unknown
);
978 dsos__fprintf(stdout
);
982 output__fprintf(stdout
, total
);
987 static const char * const report_usage
[] = {
988 "perf report [<options>] <command>",
992 static const struct option options
[] = {
993 OPT_STRING('i', "input", &input_name
, "file",
995 OPT_BOOLEAN('v', "verbose", &verbose
,
996 "be more verbose (show symbol address, etc)"),
997 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace
,
998 "dump raw trace in ASCII"),
999 OPT_STRING('k', "vmlinux", &vmlinux
, "file", "vmlinux pathname"),
1000 OPT_STRING('s', "sort", &sort_order
, "key[,key2...]",
1001 "sort by key(s): pid, comm, dso, symbol. Default: pid,symbol"),
1002 OPT_BOOLEAN('P', "full-paths", &full_paths
,
1003 "Don't shorten the pathnames taking into account the cwd"),
1007 static void setup_sorting(void)
1009 char *tmp
, *tok
, *str
= strdup(sort_order
);
1011 for (tok
= strtok_r(str
, ", ", &tmp
);
1012 tok
; tok
= strtok_r(NULL
, ", ", &tmp
)) {
1013 if (sort_dimension__add(tok
) < 0) {
1014 error("Unknown --sort key: `%s'", tok
);
1015 usage_with_options(report_usage
, options
);
1022 int cmd_report(int argc
, const char **argv
, const char *prefix
)
1026 page_size
= getpagesize();
1028 parse_options(argc
, argv
, options
, report_usage
, 0);
1034 return __cmd_report();