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/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"
18 #include "util/callchain.h"
19 #include "util/strlist.h"
20 #include "util/values.h"
23 #include "util/debug.h"
24 #include "util/header.h"
25 #include "util/session.h"
27 #include "util/parse-options.h"
28 #include "util/parse-events.h"
30 #include "util/data_map.h"
31 #include "util/thread.h"
32 #include "util/sort.h"
33 #include "util/hist.h"
35 static char const *input_name
= "perf.data";
37 static char *dso_list_str
, *comm_list_str
, *sym_list_str
,
39 static struct strlist
*dso_list
, *comm_list
, *sym_list
;
43 static int full_paths
;
44 static int show_nr_samples
;
46 static int show_threads
;
47 static struct perf_read_values show_threads_values
;
49 static char default_pretty_printing_style
[] = "normal";
50 static char *pretty_printing_style
= default_pretty_printing_style
;
52 static int exclude_other
= 1;
54 static char callchain_default_opt
[] = "fractal,0.5";
56 static struct perf_session
*session
;
58 static u64 sample_type
;
60 struct symbol_conf symbol_conf
;
64 callchain__fprintf_left_margin(FILE *fp
, int left_margin
)
69 ret
= fprintf(fp
, " ");
71 for (i
= 0; i
< left_margin
; i
++)
72 ret
+= fprintf(fp
, " ");
77 static size_t ipchain__fprintf_graph_line(FILE *fp
, int depth
, int depth_mask
,
83 ret
+= callchain__fprintf_left_margin(fp
, left_margin
);
85 for (i
= 0; i
< depth
; i
++)
86 if (depth_mask
& (1 << i
))
87 ret
+= fprintf(fp
, "| ");
89 ret
+= fprintf(fp
, " ");
91 ret
+= fprintf(fp
, "\n");
96 ipchain__fprintf_graph(FILE *fp
, struct callchain_list
*chain
, int depth
,
97 int depth_mask
, int count
, u64 total_samples
,
98 int hits
, int left_margin
)
103 ret
+= callchain__fprintf_left_margin(fp
, left_margin
);
104 for (i
= 0; i
< depth
; i
++) {
105 if (depth_mask
& (1 << i
))
106 ret
+= fprintf(fp
, "|");
108 ret
+= fprintf(fp
, " ");
109 if (!count
&& i
== depth
- 1) {
112 percent
= hits
* 100.0 / total_samples
;
113 ret
+= percent_color_fprintf(fp
, "--%2.2f%%-- ", percent
);
115 ret
+= fprintf(fp
, "%s", " ");
118 ret
+= fprintf(fp
, "%s\n", chain
->sym
->name
);
120 ret
+= fprintf(fp
, "%p\n", (void *)(long)chain
->ip
);
125 static struct symbol
*rem_sq_bracket
;
126 static struct callchain_list rem_hits
;
128 static void init_rem_hits(void)
130 rem_sq_bracket
= malloc(sizeof(*rem_sq_bracket
) + 6);
131 if (!rem_sq_bracket
) {
132 fprintf(stderr
, "Not enough memory to display remaining hits\n");
136 strcpy(rem_sq_bracket
->name
, "[...]");
137 rem_hits
.sym
= rem_sq_bracket
;
141 __callchain__fprintf_graph(FILE *fp
, struct callchain_node
*self
,
142 u64 total_samples
, int depth
, int depth_mask
,
145 struct rb_node
*node
, *next
;
146 struct callchain_node
*child
;
147 struct callchain_list
*chain
;
148 int new_depth_mask
= depth_mask
;
154 if (callchain_param
.mode
== CHAIN_GRAPH_REL
)
155 new_total
= self
->children_hit
;
157 new_total
= total_samples
;
159 remaining
= new_total
;
161 node
= rb_first(&self
->rb_root
);
165 child
= rb_entry(node
, struct callchain_node
, rb_node
);
166 cumul
= cumul_hits(child
);
170 * The depth mask manages the output of pipes that show
171 * the depth. We don't want to keep the pipes of the current
172 * level for the last child of this depth.
173 * Except if we have remaining filtered hits. They will
174 * supersede the last child
176 next
= rb_next(node
);
177 if (!next
&& (callchain_param
.mode
!= CHAIN_GRAPH_REL
|| !remaining
))
178 new_depth_mask
&= ~(1 << (depth
- 1));
181 * But we keep the older depth mask for the line seperator
182 * to keep the level link until we reach the last child
184 ret
+= ipchain__fprintf_graph_line(fp
, depth
, depth_mask
,
187 list_for_each_entry(chain
, &child
->val
, list
) {
188 if (chain
->ip
>= PERF_CONTEXT_MAX
)
190 ret
+= ipchain__fprintf_graph(fp
, chain
, depth
,
196 ret
+= __callchain__fprintf_graph(fp
, child
, new_total
,
198 new_depth_mask
| (1 << depth
),
203 if (callchain_param
.mode
== CHAIN_GRAPH_REL
&&
204 remaining
&& remaining
!= new_total
) {
209 new_depth_mask
&= ~(1 << (depth
- 1));
211 ret
+= ipchain__fprintf_graph(fp
, &rem_hits
, depth
,
212 new_depth_mask
, 0, new_total
,
213 remaining
, left_margin
);
221 callchain__fprintf_graph(FILE *fp
, struct callchain_node
*self
,
222 u64 total_samples
, int left_margin
)
224 struct callchain_list
*chain
;
225 bool printed
= false;
229 list_for_each_entry(chain
, &self
->val
, list
) {
230 if (chain
->ip
>= PERF_CONTEXT_MAX
)
233 if (!i
++ && sort__first_dimension
== SORT_SYM
)
237 ret
+= callchain__fprintf_left_margin(fp
, left_margin
);
238 ret
+= fprintf(fp
, "|\n");
239 ret
+= callchain__fprintf_left_margin(fp
, left_margin
);
240 ret
+= fprintf(fp
, "---");
245 ret
+= callchain__fprintf_left_margin(fp
, left_margin
);
248 ret
+= fprintf(fp
, " %s\n", chain
->sym
->name
);
250 ret
+= fprintf(fp
, " %p\n", (void *)(long)chain
->ip
);
253 ret
+= __callchain__fprintf_graph(fp
, self
, total_samples
, 1, 1, left_margin
);
259 callchain__fprintf_flat(FILE *fp
, struct callchain_node
*self
,
262 struct callchain_list
*chain
;
268 ret
+= callchain__fprintf_flat(fp
, self
->parent
, total_samples
);
271 list_for_each_entry(chain
, &self
->val
, list
) {
272 if (chain
->ip
>= PERF_CONTEXT_MAX
)
275 ret
+= fprintf(fp
, " %s\n", chain
->sym
->name
);
277 ret
+= fprintf(fp
, " %p\n",
278 (void *)(long)chain
->ip
);
285 hist_entry_callchain__fprintf(FILE *fp
, struct hist_entry
*self
,
286 u64 total_samples
, int left_margin
)
288 struct rb_node
*rb_node
;
289 struct callchain_node
*chain
;
292 rb_node
= rb_first(&self
->sorted_chain
);
296 chain
= rb_entry(rb_node
, struct callchain_node
, rb_node
);
297 percent
= chain
->hit
* 100.0 / total_samples
;
298 switch (callchain_param
.mode
) {
300 ret
+= percent_color_fprintf(fp
, " %6.2f%%\n",
302 ret
+= callchain__fprintf_flat(fp
, chain
, total_samples
);
304 case CHAIN_GRAPH_ABS
: /* Falldown */
305 case CHAIN_GRAPH_REL
:
306 ret
+= callchain__fprintf_graph(fp
, chain
, total_samples
,
312 ret
+= fprintf(fp
, "\n");
313 rb_node
= rb_next(rb_node
);
320 hist_entry__fprintf(FILE *fp
, struct hist_entry
*self
, u64 total_samples
)
322 struct sort_entry
*se
;
325 if (exclude_other
&& !self
->parent
)
329 ret
= percent_color_fprintf(fp
,
330 field_sep
? "%.2f" : " %6.2f%%",
331 (self
->count
* 100.0) / total_samples
);
333 ret
= fprintf(fp
, field_sep
? "%lld" : "%12lld ", self
->count
);
335 if (show_nr_samples
) {
337 fprintf(fp
, "%c%lld", *field_sep
, self
->count
);
339 fprintf(fp
, "%11lld", self
->count
);
342 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
346 fprintf(fp
, "%s", field_sep
?: " ");
347 ret
+= se
->print(fp
, self
, se
->width
? *se
->width
: 0);
350 ret
+= fprintf(fp
, "\n");
355 if (sort__first_dimension
== SORT_COMM
) {
356 se
= list_first_entry(&hist_entry__sort_list
, typeof(*se
),
358 left_margin
= se
->width
? *se
->width
: 0;
359 left_margin
-= thread__comm_len(self
->thread
);
362 hist_entry_callchain__fprintf(fp
, self
, total_samples
,
373 static void dso__calc_col_width(struct dso
*self
)
375 if (!col_width_list_str
&& !field_sep
&&
376 (!dso_list
|| strlist__has_entry(dso_list
, self
->name
))) {
377 unsigned int slen
= strlen(self
->name
);
378 if (slen
> dsos__col_width
)
379 dsos__col_width
= slen
;
382 self
->slen_calculated
= 1;
385 static void thread__comm_adjust(struct thread
*self
)
387 char *comm
= self
->comm
;
389 if (!col_width_list_str
&& !field_sep
&&
390 (!comm_list
|| strlist__has_entry(comm_list
, comm
))) {
391 unsigned int slen
= strlen(comm
);
393 if (slen
> comms__col_width
) {
394 comms__col_width
= slen
;
395 threads__col_width
= slen
+ 6;
400 static int thread__set_comm_adjust(struct thread
*self
, const char *comm
)
402 int ret
= thread__set_comm(self
, comm
);
407 thread__comm_adjust(self
);
412 static int call__match(struct symbol
*sym
)
414 if (sym
->name
&& !regexec(&parent_regex
, sym
->name
, 0, NULL
, 0))
420 static struct symbol
**resolve_callchain(struct thread
*thread
,
421 struct ip_callchain
*chain
,
422 struct symbol
**parent
)
424 u8 cpumode
= PERF_RECORD_MISC_USER
;
425 struct symbol
**syms
= NULL
;
429 syms
= calloc(chain
->nr
, sizeof(*syms
));
431 fprintf(stderr
, "Can't allocate memory for symbols\n");
436 for (i
= 0; i
< chain
->nr
; i
++) {
437 u64 ip
= chain
->ips
[i
];
438 struct addr_location al
;
440 if (ip
>= PERF_CONTEXT_MAX
) {
442 case PERF_CONTEXT_HV
:
443 cpumode
= PERF_RECORD_MISC_HYPERVISOR
; break;
444 case PERF_CONTEXT_KERNEL
:
445 cpumode
= PERF_RECORD_MISC_KERNEL
; break;
446 case PERF_CONTEXT_USER
:
447 cpumode
= PERF_RECORD_MISC_USER
; break;
454 thread__find_addr_location(thread
, cpumode
, MAP__FUNCTION
,
456 if (al
.sym
!= NULL
) {
457 if (sort__has_parent
&& !*parent
&&
470 * collect histogram counts
473 static int hist_entry__add(struct addr_location
*al
,
474 struct ip_callchain
*chain
, u64 count
)
476 struct symbol
**syms
= NULL
, *parent
= NULL
;
478 struct hist_entry
*he
;
480 if ((sort__has_parent
|| callchain
) && chain
)
481 syms
= resolve_callchain(al
->thread
, chain
, &parent
);
483 he
= __hist_entry__add(al
, parent
, count
, &hit
);
492 callchain_init(&he
->callchain
);
493 append_chain(&he
->callchain
, chain
, syms
);
500 static size_t output__fprintf(FILE *fp
, u64 total_samples
)
502 struct hist_entry
*pos
;
503 struct sort_entry
*se
;
507 char *col_width
= col_width_list_str
;
508 int raw_printing_style
;
510 raw_printing_style
= !strcmp(pretty_printing_style
, "raw");
514 fprintf(fp
, "# Samples: %Ld\n", (u64
)total_samples
);
517 fprintf(fp
, "# Overhead");
518 if (show_nr_samples
) {
520 fprintf(fp
, "%cSamples", *field_sep
);
522 fputs(" Samples ", fp
);
524 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
528 fprintf(fp
, "%c%s", *field_sep
, se
->header
);
531 width
= strlen(se
->header
);
533 if (col_width_list_str
) {
535 *se
->width
= atoi(col_width
);
536 col_width
= strchr(col_width
, ',');
541 width
= *se
->width
= max(*se
->width
, width
);
543 fprintf(fp
, " %*s", width
, se
->header
);
550 fprintf(fp
, "# ........");
552 fprintf(fp
, " ..........");
553 list_for_each_entry(se
, &hist_entry__sort_list
, list
) {
563 width
= strlen(se
->header
);
564 for (i
= 0; i
< width
; i
++)
572 for (nd
= rb_first(&output_hists
); nd
; nd
= rb_next(nd
)) {
573 pos
= rb_entry(nd
, struct hist_entry
, rb_node
);
574 ret
+= hist_entry__fprintf(fp
, pos
, total_samples
);
577 if (sort_order
== default_sort_order
&&
578 parent_pattern
== default_parent_pattern
) {
580 fprintf(fp
, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
585 free(rem_sq_bracket
);
588 perf_read_values_display(fp
, &show_threads_values
,
594 static int validate_chain(struct ip_callchain
*chain
, event_t
*event
)
596 unsigned int chain_size
;
598 chain_size
= event
->header
.size
;
599 chain_size
-= (unsigned long)&event
->ip
.__more_data
- (unsigned long)event
;
601 if (chain
->nr
*sizeof(u64
) > chain_size
)
607 static int process_sample_event(event_t
*event
)
609 struct sample_data data
;
611 struct addr_location al
;
612 struct thread
*thread
;
614 memset(&data
, 0, sizeof(data
));
617 event__parse_sample(event
, sample_type
, &data
);
619 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
622 (void *)(long)data
.ip
,
623 (long long)data
.period
);
625 if (sample_type
& PERF_SAMPLE_CALLCHAIN
) {
628 dump_printf("... chain: nr:%Lu\n", data
.callchain
->nr
);
630 if (validate_chain(data
.callchain
, event
) < 0) {
631 pr_debug("call-chain problem with event, "
637 for (i
= 0; i
< data
.callchain
->nr
; i
++)
638 dump_printf("..... %2d: %016Lx\n",
639 i
, data
.callchain
->ips
[i
]);
643 thread
= threads__findnew(data
.pid
);
644 if (thread
== NULL
) {
645 pr_debug("problem processing %d event, skipping it.\n",
650 dump_printf(" ... thread: %s:%d\n", thread
->comm
, thread
->pid
);
652 if (comm_list
&& !strlist__has_entry(comm_list
, thread
->comm
))
655 cpumode
= event
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
657 thread__find_addr_location(thread
, cpumode
,
658 MAP__FUNCTION
, data
.ip
, &al
, NULL
);
660 * We have to do this here as we may have a dso with no symbol hit that
661 * has a name longer than the ones with symbols sampled.
663 if (al
.map
&& !sort_dso
.elide
&& !al
.map
->dso
->slen_calculated
)
664 dso__calc_col_width(al
.map
->dso
);
667 (!al
.map
|| !al
.map
->dso
||
668 !(strlist__has_entry(dso_list
, al
.map
->dso
->short_name
) ||
669 (al
.map
->dso
->short_name
!= al
.map
->dso
->long_name
&&
670 strlist__has_entry(dso_list
, al
.map
->dso
->long_name
)))))
673 if (sym_list
&& al
.sym
&& !strlist__has_entry(sym_list
, al
.sym
->name
))
676 if (hist_entry__add(&al
, data
.callchain
, data
.period
)) {
677 pr_debug("problem incrementing symbol count, skipping event\n");
681 event__stats
.total
+= data
.period
;
686 static int process_comm_event(event_t
*event
)
688 struct thread
*thread
= threads__findnew(event
->comm
.pid
);
690 dump_printf(": %s:%d\n", event
->comm
.comm
, event
->comm
.pid
);
692 if (thread
== NULL
||
693 thread__set_comm_adjust(thread
, event
->comm
.comm
)) {
694 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
701 static int process_read_event(event_t
*event
)
703 struct perf_event_attr
*attr
;
705 attr
= perf_header__find_attr(event
->read
.id
, &session
->header
);
708 const char *name
= attr
? __event_name(attr
->type
, attr
->config
)
710 perf_read_values_add_value(&show_threads_values
,
711 event
->read
.pid
, event
->read
.tid
,
717 dump_printf(": %d %d %s %Lu\n", event
->read
.pid
, event
->read
.tid
,
718 attr
? __event_name(attr
->type
, attr
->config
) : "FAIL",
724 static int sample_type_check(u64 type
)
728 if (!(sample_type
& PERF_SAMPLE_CALLCHAIN
)) {
729 if (sort__has_parent
) {
730 fprintf(stderr
, "selected --sort parent, but no"
731 " callchain data. Did you call"
732 " perf record without -g?\n");
736 fprintf(stderr
, "selected -g but no callchain data."
737 " Did you call perf record without"
741 } else if (callchain_param
.mode
!= CHAIN_NONE
&& !callchain
) {
743 if (register_callchain_param(&callchain_param
) < 0) {
744 fprintf(stderr
, "Can't register callchain"
753 static struct perf_file_handler file_handler
= {
754 .process_sample_event
= process_sample_event
,
755 .process_mmap_event
= event__process_mmap
,
756 .process_comm_event
= process_comm_event
,
757 .process_exit_event
= event__process_task
,
758 .process_fork_event
= event__process_task
,
759 .process_lost_event
= event__process_lost
,
760 .process_read_event
= process_read_event
,
761 .sample_type_check
= sample_type_check
,
765 static int __cmd_report(void)
770 session
= perf_session__new(input_name
, O_RDONLY
, force
);
774 idle
= register_idle_thread();
775 thread__comm_adjust(idle
);
778 perf_read_values_init(&show_threads_values
);
780 register_perf_file_handler(&file_handler
);
782 ret
= perf_session__process_events(session
, full_paths
,
783 &event__cwdlen
, &event__cwd
);
788 event__print_totals();
793 threads__fprintf(stdout
);
796 dsos__fprintf(stdout
);
799 output__resort(event__stats
.total
);
800 output__fprintf(stdout
, event__stats
.total
);
803 perf_read_values_destroy(&show_threads_values
);
805 perf_session__delete(session
);
810 parse_callchain_opt(const struct option
*opt __used
, const char *arg
,
821 tok
= strtok((char *)arg
, ",");
825 /* get the output mode */
826 if (!strncmp(tok
, "graph", strlen(arg
)))
827 callchain_param
.mode
= CHAIN_GRAPH_ABS
;
829 else if (!strncmp(tok
, "flat", strlen(arg
)))
830 callchain_param
.mode
= CHAIN_FLAT
;
832 else if (!strncmp(tok
, "fractal", strlen(arg
)))
833 callchain_param
.mode
= CHAIN_GRAPH_REL
;
835 else if (!strncmp(tok
, "none", strlen(arg
))) {
836 callchain_param
.mode
= CHAIN_NONE
;
845 /* get the min percentage */
846 tok
= strtok(NULL
, ",");
850 callchain_param
.min_percent
= strtod(tok
, &endptr
);
855 if (register_callchain_param(&callchain_param
) < 0) {
856 fprintf(stderr
, "Can't register callchain params\n");
862 //static const char * const report_usage[] = {
863 const char * const report_usage
[] = {
864 "perf report [<options>] <command>",
868 static const struct option options
[] = {
869 OPT_STRING('i', "input", &input_name
, "file",
871 OPT_BOOLEAN('v', "verbose", &verbose
,
872 "be more verbose (show symbol address, etc)"),
873 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace
,
874 "dump raw trace in ASCII"),
875 OPT_STRING('k', "vmlinux", &symbol_conf
.vmlinux_name
,
876 "file", "vmlinux pathname"),
877 OPT_BOOLEAN('f', "force", &force
, "don't complain, do it"),
878 OPT_BOOLEAN('m', "modules", &symbol_conf
.use_modules
,
879 "load module symbols - WARNING: use only with -k and LIVE kernel"),
880 OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples
,
881 "Show a column with the number of samples"),
882 OPT_BOOLEAN('T', "threads", &show_threads
,
883 "Show per-thread event counters"),
884 OPT_STRING(0, "pretty", &pretty_printing_style
, "key",
885 "pretty printing style key: normal raw"),
886 OPT_STRING('s', "sort", &sort_order
, "key[,key2...]",
887 "sort by key(s): pid, comm, dso, symbol, parent"),
888 OPT_BOOLEAN('P', "full-paths", &full_paths
,
889 "Don't shorten the pathnames taking into account the cwd"),
890 OPT_STRING('p', "parent", &parent_pattern
, "regex",
891 "regex filter to identify parent, see: '--sort parent'"),
892 OPT_BOOLEAN('x', "exclude-other", &exclude_other
,
893 "Only display entries with parent-match"),
894 OPT_CALLBACK_DEFAULT('g', "call-graph", NULL
, "output_type,min_percent",
895 "Display callchains using output_type and min percent threshold. "
896 "Default: fractal,0.5", &parse_callchain_opt
, callchain_default_opt
),
897 OPT_STRING('d', "dsos", &dso_list_str
, "dso[,dso...]",
898 "only consider symbols in these dsos"),
899 OPT_STRING('C', "comms", &comm_list_str
, "comm[,comm...]",
900 "only consider symbols in these comms"),
901 OPT_STRING('S', "symbols", &sym_list_str
, "symbol[,symbol...]",
902 "only consider these symbols"),
903 OPT_STRING('w', "column-widths", &col_width_list_str
,
905 "don't try to adjust column width, use these fixed values"),
906 OPT_STRING('t', "field-separator", &field_sep
, "separator",
907 "separator for columns, no spaces will be added between "
908 "columns '.' is reserved."),
912 static void setup_sorting(void)
914 char *tmp
, *tok
, *str
= strdup(sort_order
);
916 for (tok
= strtok_r(str
, ", ", &tmp
);
917 tok
; tok
= strtok_r(NULL
, ", ", &tmp
)) {
918 if (sort_dimension__add(tok
) < 0) {
919 error("Unknown --sort key: `%s'", tok
);
920 usage_with_options(report_usage
, options
);
927 static void setup_list(struct strlist
**list
, const char *list_str
,
928 struct sort_entry
*se
, const char *list_name
,
932 *list
= strlist__new(true, list_str
);
934 fprintf(stderr
, "problems parsing %s list\n",
938 if (strlist__nr_entries(*list
) == 1) {
939 fprintf(fp
, "# %s: %s\n", list_name
,
940 strlist__entry(*list
, 0)->s
);
946 int cmd_report(int argc
, const char **argv
, const char *prefix __used
)
948 if (symbol__init(&symbol_conf
) < 0)
951 argc
= parse_options(argc
, argv
, options
, report_usage
, 0);
955 if (parent_pattern
!= default_parent_pattern
) {
956 sort_dimension__add("parent");
957 sort_parent
.elide
= 1;
962 * Any (unrecognized) arguments left?
965 usage_with_options(report_usage
, options
);
969 setup_list(&dso_list
, dso_list_str
, &sort_dso
, "dso", stdout
);
970 setup_list(&comm_list
, comm_list_str
, &sort_comm
, "comm", stdout
);
971 setup_list(&sym_list
, sym_list_str
, &sort_sym
, "symbol", stdout
);
973 if (field_sep
&& *field_sep
== '.') {
974 fputs("'.' is the only non valid --field-separator argument\n",
979 return __cmd_report();