4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/group.h"
67 #include "util/string2.h"
70 #include <linux/time64.h>
71 #include <api/fs/fs.h>
75 #include <sys/prctl.h>
79 #include <sys/types.h>
84 #include "sane_ctype.h"
86 #define DEFAULT_SEPARATOR " "
87 #define CNTR_NOT_SUPPORTED "<not supported>"
88 #define CNTR_NOT_COUNTED "<not counted>"
89 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
91 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
);
93 /* Default events used for perf stat -T */
94 static const char *transaction_attrs
= {
106 /* More limited version when the CPU does not have all events. */
107 static const char * transaction_limited_attrs
= {
117 static const char * topdown_attrs
[] = {
118 "topdown-total-slots",
119 "topdown-slots-retired",
120 "topdown-recovery-bubbles",
121 "topdown-fetch-bubbles",
122 "topdown-slots-issued",
126 static const char *smi_cost_attrs
= {
134 static struct perf_evlist
*evsel_list
;
136 static struct target target
= {
140 typedef int (*aggr_get_id_t
)(struct cpu_map
*m
, int cpu
);
142 static int run_count
= 1;
143 static bool no_inherit
= false;
144 static volatile pid_t child_pid
= -1;
145 static bool null_run
= false;
146 static int detailed_run
= 0;
147 static bool transaction_run
;
148 static bool topdown_run
= false;
149 static bool smi_cost
= false;
150 static bool smi_reset
= false;
151 static bool big_num
= true;
152 static int big_num_opt
= -1;
153 static const char *csv_sep
= NULL
;
154 static bool csv_output
= false;
155 static bool group
= false;
156 static const char *pre_cmd
= NULL
;
157 static const char *post_cmd
= NULL
;
158 static bool sync_run
= false;
159 static unsigned int initial_delay
= 0;
160 static unsigned int unit_width
= 4; /* strlen("unit") */
161 static bool forever
= false;
162 static bool metric_only
= false;
163 static bool force_metric_only
= false;
164 static bool no_merge
= false;
165 static struct timespec ref_time
;
166 static struct cpu_map
*aggr_map
;
167 static aggr_get_id_t aggr_get_id
;
168 static bool append_file
;
169 static const char *output_name
;
170 static int output_fd
;
171 static int print_free_counters_hint
;
175 struct perf_data_file file
;
176 struct perf_session
*session
;
178 struct perf_tool tool
;
180 struct cpu_map
*cpus
;
181 struct thread_map
*threads
;
182 enum aggr_mode aggr_mode
;
185 static struct perf_stat perf_stat
;
186 #define STAT_RECORD perf_stat.record
188 static volatile int done
= 0;
190 static struct perf_stat_config stat_config
= {
191 .aggr_mode
= AGGR_GLOBAL
,
195 static inline void diff_timespec(struct timespec
*r
, struct timespec
*a
,
198 r
->tv_sec
= a
->tv_sec
- b
->tv_sec
;
199 if (a
->tv_nsec
< b
->tv_nsec
) {
200 r
->tv_nsec
= a
->tv_nsec
+ NSEC_PER_SEC
- b
->tv_nsec
;
203 r
->tv_nsec
= a
->tv_nsec
- b
->tv_nsec
;
207 static void perf_stat__reset_stats(void)
209 perf_evlist__reset_stats(evsel_list
);
210 perf_stat__reset_shadow_stats();
213 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
215 struct perf_event_attr
*attr
= &evsel
->attr
;
217 if (stat_config
.scale
)
218 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
219 PERF_FORMAT_TOTAL_TIME_RUNNING
;
221 attr
->inherit
= !no_inherit
;
224 * Some events get initialized with sample_(period/type) set,
225 * like tracepoints. Clear it up for counting.
227 attr
->sample_period
= 0;
230 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
231 * while avoiding that older tools show confusing messages.
233 * However for pipe sessions we need to keep it zero,
234 * because script's perf_evsel__check_attr is triggered
235 * by attr->sample_type != 0, and we can't run it on
238 if (!(STAT_RECORD
&& perf_stat
.file
.is_pipe
))
239 attr
->sample_type
= PERF_SAMPLE_IDENTIFIER
;
242 * Disabling all counters initially, they will be enabled
243 * either manually by us or by kernel via enable_on_exec
246 if (perf_evsel__is_group_leader(evsel
)) {
250 * In case of initial_delay we enable tracee
253 if (target__none(&target
) && !initial_delay
)
254 attr
->enable_on_exec
= 1;
257 if (target__has_cpu(&target
))
258 return perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
260 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
264 * Does the counter have nsecs as a unit?
266 static inline int nsec_counter(struct perf_evsel
*evsel
)
268 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
269 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
275 static int process_synthesized_event(struct perf_tool
*tool __maybe_unused
,
276 union perf_event
*event
,
277 struct perf_sample
*sample __maybe_unused
,
278 struct machine
*machine __maybe_unused
)
280 if (perf_data_file__write(&perf_stat
.file
, event
, event
->header
.size
) < 0) {
281 pr_err("failed to write perf data, error: %m\n");
285 perf_stat
.bytes_written
+= event
->header
.size
;
289 static int write_stat_round_event(u64 tm
, u64 type
)
291 return perf_event__synthesize_stat_round(NULL
, tm
, type
,
292 process_synthesized_event
,
296 #define WRITE_STAT_ROUND_EVENT(time, interval) \
297 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
299 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
302 perf_evsel__write_stat_event(struct perf_evsel
*counter
, u32 cpu
, u32 thread
,
303 struct perf_counts_values
*count
)
305 struct perf_sample_id
*sid
= SID(counter
, cpu
, thread
);
307 return perf_event__synthesize_stat(NULL
, cpu
, thread
, sid
->id
, count
,
308 process_synthesized_event
, NULL
);
312 * Read out the results of a single counter:
313 * do not aggregate counts across CPUs in system-wide mode
315 static int read_counter(struct perf_evsel
*counter
)
317 int nthreads
= thread_map__nr(evsel_list
->threads
);
318 int ncpus
, cpu
, thread
;
320 if (target__has_cpu(&target
))
321 ncpus
= perf_evsel__nr_cpus(counter
);
325 if (!counter
->supported
)
328 if (counter
->system_wide
)
331 for (thread
= 0; thread
< nthreads
; thread
++) {
332 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
333 struct perf_counts_values
*count
;
335 count
= perf_counts(counter
->counts
, cpu
, thread
);
336 if (perf_evsel__read(counter
, cpu
, thread
, count
)) {
337 counter
->counts
->scaled
= -1;
338 perf_counts(counter
->counts
, cpu
, thread
)->ena
= 0;
339 perf_counts(counter
->counts
, cpu
, thread
)->run
= 0;
344 if (perf_evsel__write_stat_event(counter
, cpu
, thread
, count
)) {
345 pr_err("failed to write stat event\n");
351 fprintf(stat_config
.output
,
352 "%s: %d: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
353 perf_evsel__name(counter
),
355 count
->val
, count
->ena
, count
->run
);
363 static void read_counters(void)
365 struct perf_evsel
*counter
;
368 evlist__for_each_entry(evsel_list
, counter
) {
369 ret
= read_counter(counter
);
371 pr_debug("failed to read counter %s\n", counter
->name
);
373 if (ret
== 0 && perf_stat_process_counter(&stat_config
, counter
))
374 pr_warning("failed to process counter %s\n", counter
->name
);
378 static void process_interval(void)
380 struct timespec ts
, rs
;
384 clock_gettime(CLOCK_MONOTONIC
, &ts
);
385 diff_timespec(&rs
, &ts
, &ref_time
);
388 if (WRITE_STAT_ROUND_EVENT(rs
.tv_sec
* NSEC_PER_SEC
+ rs
.tv_nsec
, INTERVAL
))
389 pr_err("failed to write stat round event\n");
392 print_counters(&rs
, 0, NULL
);
395 static void enable_counters(void)
398 usleep(initial_delay
* USEC_PER_MSEC
);
401 * We need to enable counters only if:
402 * - we don't have tracee (attaching to task or cpu)
403 * - we have initial delay configured
405 if (!target__none(&target
) || initial_delay
)
406 perf_evlist__enable(evsel_list
);
409 static void disable_counters(void)
412 * If we don't have tracee (attaching to task or cpu), counters may
413 * still be running. To get accurate group ratios, we must stop groups
414 * from counting before reading their constituent counters.
416 if (!target__none(&target
))
417 perf_evlist__disable(evsel_list
);
420 static volatile int workload_exec_errno
;
423 * perf_evlist__prepare_workload will send a SIGUSR1
424 * if the fork fails, since we asked by setting its
425 * want_signal to true.
427 static void workload_exec_failed_signal(int signo __maybe_unused
, siginfo_t
*info
,
428 void *ucontext __maybe_unused
)
430 workload_exec_errno
= info
->si_value
.sival_int
;
433 static bool has_unit(struct perf_evsel
*counter
)
435 return counter
->unit
&& *counter
->unit
;
438 static bool has_scale(struct perf_evsel
*counter
)
440 return counter
->scale
!= 1;
443 static int perf_stat_synthesize_config(bool is_pipe
)
445 struct perf_evsel
*counter
;
449 err
= perf_event__synthesize_attrs(NULL
, perf_stat
.session
,
450 process_synthesized_event
);
452 pr_err("Couldn't synthesize attrs.\n");
458 * Synthesize other events stuff not carried within
459 * attr event - unit, scale, name
461 evlist__for_each_entry(evsel_list
, counter
) {
462 if (!counter
->supported
)
466 * Synthesize unit and scale only if it's defined.
468 if (has_unit(counter
)) {
469 err
= perf_event__synthesize_event_update_unit(NULL
, counter
, process_synthesized_event
);
471 pr_err("Couldn't synthesize evsel unit.\n");
476 if (has_scale(counter
)) {
477 err
= perf_event__synthesize_event_update_scale(NULL
, counter
, process_synthesized_event
);
479 pr_err("Couldn't synthesize evsel scale.\n");
484 if (counter
->own_cpus
) {
485 err
= perf_event__synthesize_event_update_cpus(NULL
, counter
, process_synthesized_event
);
487 pr_err("Couldn't synthesize evsel scale.\n");
493 * Name is needed only for pipe output,
494 * perf.data carries event names.
497 err
= perf_event__synthesize_event_update_name(NULL
, counter
, process_synthesized_event
);
499 pr_err("Couldn't synthesize evsel name.\n");
505 err
= perf_event__synthesize_thread_map2(NULL
, evsel_list
->threads
,
506 process_synthesized_event
,
509 pr_err("Couldn't synthesize thread map.\n");
513 err
= perf_event__synthesize_cpu_map(NULL
, evsel_list
->cpus
,
514 process_synthesized_event
, NULL
);
516 pr_err("Couldn't synthesize thread map.\n");
520 err
= perf_event__synthesize_stat_config(NULL
, &stat_config
,
521 process_synthesized_event
, NULL
);
523 pr_err("Couldn't synthesize config.\n");
530 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
532 static int __store_counter_ids(struct perf_evsel
*counter
,
533 struct cpu_map
*cpus
,
534 struct thread_map
*threads
)
538 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
539 for (thread
= 0; thread
< threads
->nr
; thread
++) {
540 int fd
= FD(counter
, cpu
, thread
);
542 if (perf_evlist__id_add_fd(evsel_list
, counter
,
543 cpu
, thread
, fd
) < 0)
551 static int store_counter_ids(struct perf_evsel
*counter
)
553 struct cpu_map
*cpus
= counter
->cpus
;
554 struct thread_map
*threads
= counter
->threads
;
556 if (perf_evsel__alloc_id(counter
, cpus
->nr
, threads
->nr
))
559 return __store_counter_ids(counter
, cpus
, threads
);
562 static int __run_perf_stat(int argc
, const char **argv
)
564 int interval
= stat_config
.interval
;
566 unsigned long long t0
, t1
;
567 struct perf_evsel
*counter
;
571 const bool forks
= (argc
> 0);
572 bool is_pipe
= STAT_RECORD
? perf_stat
.file
.is_pipe
: false;
573 struct perf_evsel_config_term
*err_term
;
576 ts
.tv_sec
= interval
/ USEC_PER_MSEC
;
577 ts
.tv_nsec
= (interval
% USEC_PER_MSEC
) * NSEC_PER_MSEC
;
584 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
, is_pipe
,
585 workload_exec_failed_signal
) < 0) {
586 perror("failed to prepare workload");
589 child_pid
= evsel_list
->workload
.pid
;
593 perf_evlist__set_leader(evsel_list
);
595 evlist__for_each_entry(evsel_list
, counter
) {
597 if (create_perf_stat_counter(counter
) < 0) {
599 * PPC returns ENXIO for HW counters until 2.6.37
600 * (behavior changed with commit b0a873e).
602 if (errno
== EINVAL
|| errno
== ENOSYS
||
603 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
606 ui__warning("%s event is not supported by the kernel.\n",
607 perf_evsel__name(counter
));
608 counter
->supported
= false;
610 if ((counter
->leader
!= counter
) ||
611 !(counter
->leader
->nr_members
> 1))
613 } else if (perf_evsel__fallback(counter
, errno
, msg
, sizeof(msg
))) {
615 ui__warning("%s\n", msg
);
619 perf_evsel__open_strerror(counter
, &target
,
620 errno
, msg
, sizeof(msg
));
621 ui__error("%s\n", msg
);
624 kill(child_pid
, SIGTERM
);
628 counter
->supported
= true;
630 l
= strlen(counter
->unit
);
634 if (STAT_RECORD
&& store_counter_ids(counter
))
638 if (perf_evlist__apply_filters(evsel_list
, &counter
)) {
639 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
640 counter
->filter
, perf_evsel__name(counter
), errno
,
641 str_error_r(errno
, msg
, sizeof(msg
)));
645 if (perf_evlist__apply_drv_configs(evsel_list
, &counter
, &err_term
)) {
646 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
647 err_term
->val
.drv_cfg
, perf_evsel__name(counter
), errno
,
648 str_error_r(errno
, msg
, sizeof(msg
)));
653 int err
, fd
= perf_data_file__fd(&perf_stat
.file
);
656 err
= perf_header__write_pipe(perf_data_file__fd(&perf_stat
.file
));
658 err
= perf_session__write_header(perf_stat
.session
, evsel_list
,
665 err
= perf_stat_synthesize_config(is_pipe
);
671 * Enable counters and exec the command:
674 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
677 perf_evlist__start_workload(evsel_list
);
681 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
682 nanosleep(&ts
, NULL
);
688 if (workload_exec_errno
) {
689 const char *emsg
= str_error_r(workload_exec_errno
, msg
, sizeof(msg
));
690 pr_err("Workload failed: %s\n", emsg
);
694 if (WIFSIGNALED(status
))
695 psignal(WTERMSIG(status
), argv
[0]);
699 nanosleep(&ts
, NULL
);
709 update_stats(&walltime_nsecs_stats
, t1
- t0
);
712 * Closing a group leader splits the group, and as we only disable
713 * group leaders, results in remaining events becoming enabled. To
714 * avoid arbitrary skew, we must read all counters before closing any
718 perf_evlist__close(evsel_list
);
720 return WEXITSTATUS(status
);
723 static int run_perf_stat(int argc
, const char **argv
)
728 ret
= system(pre_cmd
);
736 ret
= __run_perf_stat(argc
, argv
);
741 ret
= system(post_cmd
);
749 static void print_running(u64 run
, u64 ena
)
752 fprintf(stat_config
.output
, "%s%" PRIu64
"%s%.2f",
756 ena
? 100.0 * run
/ ena
: 100.0);
757 } else if (run
!= ena
) {
758 fprintf(stat_config
.output
, " (%.2f%%)", 100.0 * run
/ ena
);
762 static void print_noise_pct(double total
, double avg
)
764 double pct
= rel_stddev_stats(total
, avg
);
767 fprintf(stat_config
.output
, "%s%.2f%%", csv_sep
, pct
);
769 fprintf(stat_config
.output
, " ( +-%6.2f%% )", pct
);
772 static void print_noise(struct perf_evsel
*evsel
, double avg
)
774 struct perf_stat_evsel
*ps
;
780 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
783 static void aggr_printout(struct perf_evsel
*evsel
, int id
, int nr
)
785 switch (stat_config
.aggr_mode
) {
787 fprintf(stat_config
.output
, "S%d-C%*d%s%*d%s",
788 cpu_map__id_to_socket(id
),
790 cpu_map__id_to_cpu(id
),
797 fprintf(stat_config
.output
, "S%*d%s%*d%s",
806 fprintf(stat_config
.output
, "CPU%*d%s",
808 perf_evsel__cpus(evsel
)->map
[id
], csv_sep
);
811 fprintf(stat_config
.output
, "%*s-%*d%s",
813 thread_map__comm(evsel
->threads
, id
),
815 thread_map__pid(evsel
->threads
, id
),
831 struct perf_evsel
*evsel
;
834 #define METRIC_LEN 35
836 static void new_line_std(void *ctx
)
838 struct outstate
*os
= ctx
;
843 static void do_new_line_std(struct outstate
*os
)
846 fputs(os
->prefix
, os
->fh
);
847 aggr_printout(os
->evsel
, os
->id
, os
->nr
);
848 if (stat_config
.aggr_mode
== AGGR_NONE
)
849 fprintf(os
->fh
, " ");
850 fprintf(os
->fh
, " ");
853 static void print_metric_std(void *ctx
, const char *color
, const char *fmt
,
854 const char *unit
, double val
)
856 struct outstate
*os
= ctx
;
859 bool newline
= os
->newline
;
863 if (unit
== NULL
|| fmt
== NULL
) {
864 fprintf(out
, "%-*s", METRIC_LEN
, "");
871 n
= fprintf(out
, " # ");
873 n
+= color_fprintf(out
, color
, fmt
, val
);
875 n
+= fprintf(out
, fmt
, val
);
876 fprintf(out
, " %-*s", METRIC_LEN
- n
- 1, unit
);
879 static void new_line_csv(void *ctx
)
881 struct outstate
*os
= ctx
;
886 fprintf(os
->fh
, "%s%s", os
->prefix
, csv_sep
);
887 aggr_printout(os
->evsel
, os
->id
, os
->nr
);
888 for (i
= 0; i
< os
->nfields
; i
++)
889 fputs(csv_sep
, os
->fh
);
892 static void print_metric_csv(void *ctx
,
893 const char *color __maybe_unused
,
894 const char *fmt
, const char *unit
, double val
)
896 struct outstate
*os
= ctx
;
898 char buf
[64], *vals
, *ends
;
900 if (unit
== NULL
|| fmt
== NULL
) {
901 fprintf(out
, "%s%s%s%s", csv_sep
, csv_sep
, csv_sep
, csv_sep
);
904 snprintf(buf
, sizeof(buf
), fmt
, val
);
905 ends
= vals
= ltrim(buf
);
906 while (isdigit(*ends
) || *ends
== '.')
909 while (isspace(*unit
))
911 fprintf(out
, "%s%s%s%s", csv_sep
, vals
, csv_sep
, unit
);
914 #define METRIC_ONLY_LEN 20
916 /* Filter out some columns that don't work well in metrics only mode */
918 static bool valid_only_metric(const char *unit
)
922 if (strstr(unit
, "/sec") ||
923 strstr(unit
, "hz") ||
924 strstr(unit
, "Hz") ||
925 strstr(unit
, "CPUs utilized"))
930 static const char *fixunit(char *buf
, struct perf_evsel
*evsel
,
933 if (!strncmp(unit
, "of all", 6)) {
934 snprintf(buf
, 1024, "%s %s", perf_evsel__name(evsel
),
941 static void print_metric_only(void *ctx
, const char *color
, const char *fmt
,
942 const char *unit
, double val
)
944 struct outstate
*os
= ctx
;
948 unsigned mlen
= METRIC_ONLY_LEN
;
950 if (!valid_only_metric(unit
))
952 unit
= fixunit(buf
, os
->evsel
, unit
);
954 n
= color_fprintf(out
, color
, fmt
, val
);
956 n
= fprintf(out
, fmt
, val
);
957 if (n
> METRIC_ONLY_LEN
)
959 if (mlen
< strlen(unit
))
960 mlen
= strlen(unit
) + 1;
961 fprintf(out
, "%*s", mlen
- n
, "");
964 static void print_metric_only_csv(void *ctx
, const char *color __maybe_unused
,
966 const char *unit
, double val
)
968 struct outstate
*os
= ctx
;
970 char buf
[64], *vals
, *ends
;
973 if (!valid_only_metric(unit
))
975 unit
= fixunit(tbuf
, os
->evsel
, unit
);
976 snprintf(buf
, sizeof buf
, fmt
, val
);
977 ends
= vals
= ltrim(buf
);
978 while (isdigit(*ends
) || *ends
== '.')
981 fprintf(out
, "%s%s", vals
, csv_sep
);
984 static void new_line_metric(void *ctx __maybe_unused
)
988 static void print_metric_header(void *ctx
, const char *color __maybe_unused
,
989 const char *fmt __maybe_unused
,
990 const char *unit
, double val __maybe_unused
)
992 struct outstate
*os
= ctx
;
995 if (!valid_only_metric(unit
))
997 unit
= fixunit(tbuf
, os
->evsel
, unit
);
999 fprintf(os
->fh
, "%s%s", unit
, csv_sep
);
1001 fprintf(os
->fh
, "%-*s ", METRIC_ONLY_LEN
, unit
);
1004 static void nsec_printout(int id
, int nr
, struct perf_evsel
*evsel
, double avg
)
1006 FILE *output
= stat_config
.output
;
1007 double msecs
= avg
/ NSEC_PER_MSEC
;
1008 const char *fmt_v
, *fmt_n
;
1011 fmt_v
= csv_output
? "%.6f%s" : "%18.6f%s";
1012 fmt_n
= csv_output
? "%s" : "%-25s";
1014 aggr_printout(evsel
, id
, nr
);
1016 scnprintf(name
, sizeof(name
), "%s%s",
1017 perf_evsel__name(evsel
), csv_output
? "" : " (msec)");
1019 fprintf(output
, fmt_v
, msecs
, csv_sep
);
1022 fprintf(output
, "%s%s", evsel
->unit
, csv_sep
);
1024 fprintf(output
, "%-*s%s", unit_width
, evsel
->unit
, csv_sep
);
1026 fprintf(output
, fmt_n
, name
);
1029 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
1032 static int first_shadow_cpu(struct perf_evsel
*evsel
, int id
)
1039 if (stat_config
.aggr_mode
== AGGR_NONE
)
1042 if (stat_config
.aggr_mode
== AGGR_GLOBAL
)
1045 for (i
= 0; i
< perf_evsel__nr_cpus(evsel
); i
++) {
1046 int cpu2
= perf_evsel__cpus(evsel
)->map
[i
];
1048 if (aggr_get_id(evsel_list
->cpus
, cpu2
) == id
)
1054 static void abs_printout(int id
, int nr
, struct perf_evsel
*evsel
, double avg
)
1056 FILE *output
= stat_config
.output
;
1057 double sc
= evsel
->scale
;
1061 fmt
= floor(sc
) != sc
? "%.2f%s" : "%.0f%s";
1064 fmt
= floor(sc
) != sc
? "%'18.2f%s" : "%'18.0f%s";
1066 fmt
= floor(sc
) != sc
? "%18.2f%s" : "%18.0f%s";
1069 aggr_printout(evsel
, id
, nr
);
1071 fprintf(output
, fmt
, avg
, csv_sep
);
1074 fprintf(output
, "%-*s%s",
1075 csv_output
? 0 : unit_width
,
1076 evsel
->unit
, csv_sep
);
1078 fprintf(output
, "%-*s", csv_output
? 0 : 25, perf_evsel__name(evsel
));
1081 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
1084 static void printout(int id
, int nr
, struct perf_evsel
*counter
, double uval
,
1085 char *prefix
, u64 run
, u64 ena
, double noise
)
1087 struct perf_stat_output_ctx out
;
1088 struct outstate os
= {
1089 .fh
= stat_config
.output
,
1090 .prefix
= prefix
? prefix
: "",
1095 print_metric_t pm
= print_metric_std
;
1099 nl
= new_line_metric
;
1101 pm
= print_metric_only_csv
;
1103 pm
= print_metric_only
;
1107 if (csv_output
&& !metric_only
) {
1108 static int aggr_fields
[] = {
1116 pm
= print_metric_csv
;
1119 os
.nfields
+= aggr_fields
[stat_config
.aggr_mode
];
1123 if (run
== 0 || ena
== 0 || counter
->counts
->scaled
== -1) {
1125 pm(&os
, NULL
, "", "", 0);
1128 aggr_printout(counter
, id
, nr
);
1130 fprintf(stat_config
.output
, "%*s%s",
1131 csv_output
? 0 : 18,
1132 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1135 if (counter
->supported
)
1136 print_free_counters_hint
= 1;
1138 fprintf(stat_config
.output
, "%-*s%s",
1139 csv_output
? 0 : unit_width
,
1140 counter
->unit
, csv_sep
);
1142 fprintf(stat_config
.output
, "%*s",
1143 csv_output
? 0 : -25,
1144 perf_evsel__name(counter
));
1147 fprintf(stat_config
.output
, "%s%s",
1148 csv_sep
, counter
->cgrp
->name
);
1151 pm(&os
, NULL
, NULL
, "", 0);
1152 print_noise(counter
, noise
);
1153 print_running(run
, ena
);
1155 pm(&os
, NULL
, NULL
, "", 0);
1161 else if (nsec_counter(counter
))
1162 nsec_printout(id
, nr
, counter
, uval
);
1164 abs_printout(id
, nr
, counter
, uval
);
1166 out
.print_metric
= pm
;
1169 out
.force_header
= false;
1171 if (csv_output
&& !metric_only
) {
1172 print_noise(counter
, noise
);
1173 print_running(run
, ena
);
1176 perf_stat__print_shadow_stats(counter
, uval
,
1177 first_shadow_cpu(counter
, id
),
1179 if (!csv_output
&& !metric_only
) {
1180 print_noise(counter
, noise
);
1181 print_running(run
, ena
);
1185 static void aggr_update_shadow(void)
1189 struct perf_evsel
*counter
;
1191 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1192 id
= aggr_map
->map
[s
];
1193 evlist__for_each_entry(evsel_list
, counter
) {
1195 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1196 s2
= aggr_get_id(evsel_list
->cpus
, cpu
);
1199 val
+= perf_counts(counter
->counts
, cpu
, 0)->val
;
1201 val
= val
* counter
->scale
;
1202 perf_stat__update_shadow_stats(counter
, &val
,
1203 first_shadow_cpu(counter
, id
));
1208 static void collect_all_aliases(struct perf_evsel
*counter
,
1209 void (*cb
)(struct perf_evsel
*counter
, void *data
,
1213 struct perf_evsel
*alias
;
1215 alias
= list_prepare_entry(counter
, &(evsel_list
->entries
), node
);
1216 list_for_each_entry_continue (alias
, &evsel_list
->entries
, node
) {
1217 if (strcmp(perf_evsel__name(alias
), perf_evsel__name(counter
)) ||
1218 alias
->scale
!= counter
->scale
||
1219 alias
->cgrp
!= counter
->cgrp
||
1220 strcmp(alias
->unit
, counter
->unit
) ||
1221 nsec_counter(alias
) != nsec_counter(counter
))
1223 alias
->merged_stat
= true;
1224 cb(alias
, data
, false);
1228 static bool collect_data(struct perf_evsel
*counter
,
1229 void (*cb
)(struct perf_evsel
*counter
, void *data
,
1233 if (counter
->merged_stat
)
1235 cb(counter
, data
, true);
1237 collect_all_aliases(counter
, cb
, data
);
1248 static void aggr_cb(struct perf_evsel
*counter
, void *data
, bool first
)
1250 struct aggr_data
*ad
= data
;
1253 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1254 struct perf_counts_values
*counts
;
1256 s2
= aggr_get_id(perf_evsel__cpus(counter
), cpu
);
1261 counts
= perf_counts(counter
->counts
, cpu
, 0);
1263 * When any result is bad, make them all to give
1264 * consistent output in interval mode.
1266 if (counts
->ena
== 0 || counts
->run
== 0 ||
1267 counter
->counts
->scaled
== -1) {
1272 ad
->val
+= counts
->val
;
1273 ad
->ena
+= counts
->ena
;
1274 ad
->run
+= counts
->run
;
1278 static void print_aggr(char *prefix
)
1280 FILE *output
= stat_config
.output
;
1281 struct perf_evsel
*counter
;
1287 if (!(aggr_map
|| aggr_get_id
))
1290 aggr_update_shadow();
1293 * With metric_only everything is on a single line.
1294 * Without each counter has its own line.
1296 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1297 struct aggr_data ad
;
1298 if (prefix
&& metric_only
)
1299 fprintf(output
, "%s", prefix
);
1301 ad
.id
= id
= aggr_map
->map
[s
];
1303 evlist__for_each_entry(evsel_list
, counter
) {
1304 ad
.val
= ad
.ena
= ad
.run
= 0;
1306 if (!collect_data(counter
, aggr_cb
, &ad
))
1312 if (first
&& metric_only
) {
1314 aggr_printout(counter
, id
, nr
);
1316 if (prefix
&& !metric_only
)
1317 fprintf(output
, "%s", prefix
);
1319 uval
= val
* counter
->scale
;
1320 printout(id
, nr
, counter
, uval
, prefix
, run
, ena
, 1.0);
1322 fputc('\n', output
);
1325 fputc('\n', output
);
1329 static void print_aggr_thread(struct perf_evsel
*counter
, char *prefix
)
1331 FILE *output
= stat_config
.output
;
1332 int nthreads
= thread_map__nr(counter
->threads
);
1333 int ncpus
= cpu_map__nr(counter
->cpus
);
1337 for (thread
= 0; thread
< nthreads
; thread
++) {
1338 u64 ena
= 0, run
= 0, val
= 0;
1340 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
1341 val
+= perf_counts(counter
->counts
, cpu
, thread
)->val
;
1342 ena
+= perf_counts(counter
->counts
, cpu
, thread
)->ena
;
1343 run
+= perf_counts(counter
->counts
, cpu
, thread
)->run
;
1347 fprintf(output
, "%s", prefix
);
1349 uval
= val
* counter
->scale
;
1350 printout(thread
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1351 fputc('\n', output
);
1356 double avg
, avg_enabled
, avg_running
;
1359 static void counter_aggr_cb(struct perf_evsel
*counter
, void *data
,
1360 bool first __maybe_unused
)
1362 struct caggr_data
*cd
= data
;
1363 struct perf_stat_evsel
*ps
= counter
->priv
;
1365 cd
->avg
+= avg_stats(&ps
->res_stats
[0]);
1366 cd
->avg_enabled
+= avg_stats(&ps
->res_stats
[1]);
1367 cd
->avg_running
+= avg_stats(&ps
->res_stats
[2]);
1371 * Print out the results of a single counter:
1372 * aggregated counts in system-wide mode
1374 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
1376 FILE *output
= stat_config
.output
;
1378 struct caggr_data cd
= { .avg
= 0.0 };
1380 if (!collect_data(counter
, counter_aggr_cb
, &cd
))
1383 if (prefix
&& !metric_only
)
1384 fprintf(output
, "%s", prefix
);
1386 uval
= cd
.avg
* counter
->scale
;
1387 printout(-1, 0, counter
, uval
, prefix
, cd
.avg_running
, cd
.avg_enabled
, cd
.avg
);
1389 fprintf(output
, "\n");
1392 static void counter_cb(struct perf_evsel
*counter
, void *data
,
1393 bool first __maybe_unused
)
1395 struct aggr_data
*ad
= data
;
1397 ad
->val
+= perf_counts(counter
->counts
, ad
->cpu
, 0)->val
;
1398 ad
->ena
+= perf_counts(counter
->counts
, ad
->cpu
, 0)->ena
;
1399 ad
->run
+= perf_counts(counter
->counts
, ad
->cpu
, 0)->run
;
1403 * Print out the results of a single counter:
1404 * does not use aggregated count in system-wide
1406 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1408 FILE *output
= stat_config
.output
;
1413 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1414 struct aggr_data ad
= { .cpu
= cpu
};
1416 if (!collect_data(counter
, counter_cb
, &ad
))
1423 fprintf(output
, "%s", prefix
);
1425 uval
= val
* counter
->scale
;
1426 printout(cpu
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1428 fputc('\n', output
);
1432 static void print_no_aggr_metric(char *prefix
)
1436 struct perf_evsel
*counter
;
1440 nrcpus
= evsel_list
->cpus
->nr
;
1441 for (cpu
= 0; cpu
< nrcpus
; cpu
++) {
1445 fputs(prefix
, stat_config
.output
);
1446 evlist__for_each_entry(evsel_list
, counter
) {
1448 aggr_printout(counter
, cpu
, 0);
1451 val
= perf_counts(counter
->counts
, cpu
, 0)->val
;
1452 ena
= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1453 run
= perf_counts(counter
->counts
, cpu
, 0)->run
;
1455 uval
= val
* counter
->scale
;
1456 printout(cpu
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1458 fputc('\n', stat_config
.output
);
1462 static int aggr_header_lens
[] = {
1470 static const char *aggr_header_csv
[] = {
1471 [AGGR_CORE
] = "core,cpus,",
1472 [AGGR_SOCKET
] = "socket,cpus",
1473 [AGGR_NONE
] = "cpu,",
1474 [AGGR_THREAD
] = "comm-pid,",
1478 static void print_metric_headers(const char *prefix
, bool no_indent
)
1480 struct perf_stat_output_ctx out
;
1481 struct perf_evsel
*counter
;
1482 struct outstate os
= {
1483 .fh
= stat_config
.output
1487 fprintf(stat_config
.output
, "%s", prefix
);
1489 if (!csv_output
&& !no_indent
)
1490 fprintf(stat_config
.output
, "%*s",
1491 aggr_header_lens
[stat_config
.aggr_mode
], "");
1493 if (stat_config
.interval
)
1494 fputs("time,", stat_config
.output
);
1495 fputs(aggr_header_csv
[stat_config
.aggr_mode
],
1496 stat_config
.output
);
1499 /* Print metrics headers only */
1500 evlist__for_each_entry(evsel_list
, counter
) {
1503 out
.print_metric
= print_metric_header
;
1504 out
.new_line
= new_line_metric
;
1505 out
.force_header
= true;
1507 perf_stat__print_shadow_stats(counter
, 0,
1511 fputc('\n', stat_config
.output
);
1514 static void print_interval(char *prefix
, struct timespec
*ts
)
1516 FILE *output
= stat_config
.output
;
1517 static int num_print_interval
;
1519 sprintf(prefix
, "%6lu.%09lu%s", ts
->tv_sec
, ts
->tv_nsec
, csv_sep
);
1521 if (num_print_interval
== 0 && !csv_output
) {
1522 switch (stat_config
.aggr_mode
) {
1524 fprintf(output
, "# time socket cpus");
1526 fprintf(output
, " counts %*s events\n", unit_width
, "unit");
1529 fprintf(output
, "# time core cpus");
1531 fprintf(output
, " counts %*s events\n", unit_width
, "unit");
1534 fprintf(output
, "# time CPU");
1536 fprintf(output
, " counts %*s events\n", unit_width
, "unit");
1539 fprintf(output
, "# time comm-pid");
1541 fprintf(output
, " counts %*s events\n", unit_width
, "unit");
1545 fprintf(output
, "# time");
1547 fprintf(output
, " counts %*s events\n", unit_width
, "unit");
1553 if (num_print_interval
== 0 && metric_only
)
1554 print_metric_headers(" ", true);
1555 if (++num_print_interval
== 25)
1556 num_print_interval
= 0;
1559 static void print_header(int argc
, const char **argv
)
1561 FILE *output
= stat_config
.output
;
1567 fprintf(output
, "\n");
1568 fprintf(output
, " Performance counter stats for ");
1569 if (target
.system_wide
)
1570 fprintf(output
, "\'system wide");
1571 else if (target
.cpu_list
)
1572 fprintf(output
, "\'CPU(s) %s", target
.cpu_list
);
1573 else if (!target__has_task(&target
)) {
1574 fprintf(output
, "\'%s", argv
? argv
[0] : "pipe");
1575 for (i
= 1; argv
&& (i
< argc
); i
++)
1576 fprintf(output
, " %s", argv
[i
]);
1577 } else if (target
.pid
)
1578 fprintf(output
, "process id \'%s", target
.pid
);
1580 fprintf(output
, "thread id \'%s", target
.tid
);
1582 fprintf(output
, "\'");
1584 fprintf(output
, " (%d runs)", run_count
);
1585 fprintf(output
, ":\n\n");
1589 static void print_footer(void)
1591 FILE *output
= stat_config
.output
;
1595 fprintf(output
, "\n");
1596 fprintf(output
, " %17.9f seconds time elapsed",
1597 avg_stats(&walltime_nsecs_stats
) / NSEC_PER_SEC
);
1598 if (run_count
> 1) {
1599 fprintf(output
, " ");
1600 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1601 avg_stats(&walltime_nsecs_stats
));
1603 fprintf(output
, "\n\n");
1605 if (print_free_counters_hint
&&
1606 sysctl__read_int("kernel/nmi_watchdog", &n
) >= 0 &&
1609 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1610 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1612 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1615 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
)
1617 int interval
= stat_config
.interval
;
1618 struct perf_evsel
*counter
;
1619 char buf
[64], *prefix
= NULL
;
1621 /* Do not print anything if we record to the pipe. */
1622 if (STAT_RECORD
&& perf_stat
.file
.is_pipe
)
1626 print_interval(prefix
= buf
, ts
);
1628 print_header(argc
, argv
);
1631 static int num_print_iv
;
1633 if (num_print_iv
== 0 && !interval
)
1634 print_metric_headers(prefix
, false);
1635 if (num_print_iv
++ == 25)
1637 if (stat_config
.aggr_mode
== AGGR_GLOBAL
&& prefix
)
1638 fprintf(stat_config
.output
, "%s", prefix
);
1641 switch (stat_config
.aggr_mode
) {
1647 evlist__for_each_entry(evsel_list
, counter
)
1648 print_aggr_thread(counter
, prefix
);
1651 evlist__for_each_entry(evsel_list
, counter
)
1652 print_counter_aggr(counter
, prefix
);
1654 fputc('\n', stat_config
.output
);
1658 print_no_aggr_metric(prefix
);
1660 evlist__for_each_entry(evsel_list
, counter
)
1661 print_counter(counter
, prefix
);
1669 if (!interval
&& !csv_output
)
1672 fflush(stat_config
.output
);
1675 static volatile int signr
= -1;
1677 static void skip_signal(int signo
)
1679 if ((child_pid
== -1) || stat_config
.interval
)
1684 * render child_pid harmless
1685 * won't send SIGTERM to a random
1686 * process in case of race condition
1687 * and fast PID recycling
1692 static void sig_atexit(void)
1697 * avoid race condition with SIGCHLD handler
1698 * in skip_signal() which is modifying child_pid
1699 * goal is to avoid send SIGTERM to a random
1703 sigaddset(&set
, SIGCHLD
);
1704 sigprocmask(SIG_BLOCK
, &set
, &oset
);
1706 if (child_pid
!= -1)
1707 kill(child_pid
, SIGTERM
);
1709 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
1714 signal(signr
, SIG_DFL
);
1715 kill(getpid(), signr
);
1718 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1719 const char *s __maybe_unused
, int unset
)
1721 big_num_opt
= unset
? 0 : 1;
1725 static int enable_metric_only(const struct option
*opt __maybe_unused
,
1726 const char *s __maybe_unused
, int unset
)
1728 force_metric_only
= true;
1729 metric_only
= !unset
;
1733 static const struct option stat_options
[] = {
1734 OPT_BOOLEAN('T', "transaction", &transaction_run
,
1735 "hardware transaction statistics"),
1736 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1737 "event selector. use 'perf list' to list available events",
1738 parse_events_option
),
1739 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1740 "event filter", parse_filter
),
1741 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1742 "child tasks do not inherit counters"),
1743 OPT_STRING('p', "pid", &target
.pid
, "pid",
1744 "stat events on existing process id"),
1745 OPT_STRING('t', "tid", &target
.tid
, "tid",
1746 "stat events on existing thread id"),
1747 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1748 "system-wide collection from all CPUs"),
1749 OPT_BOOLEAN('g', "group", &group
,
1750 "put the counters into a counter group"),
1751 OPT_BOOLEAN('c', "scale", &stat_config
.scale
, "scale/normalize counters"),
1752 OPT_INCR('v', "verbose", &verbose
,
1753 "be more verbose (show counter open errors, etc)"),
1754 OPT_INTEGER('r', "repeat", &run_count
,
1755 "repeat command and print average + stddev (max: 100, forever: 0)"),
1756 OPT_BOOLEAN('n', "null", &null_run
,
1757 "null run - dont start any counters"),
1758 OPT_INCR('d', "detailed", &detailed_run
,
1759 "detailed run - start a lot of events"),
1760 OPT_BOOLEAN('S', "sync", &sync_run
,
1761 "call sync() before starting a run"),
1762 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1763 "print large numbers with thousands\' separators",
1765 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1766 "list of cpus to monitor in system-wide"),
1767 OPT_SET_UINT('A', "no-aggr", &stat_config
.aggr_mode
,
1768 "disable CPU count aggregation", AGGR_NONE
),
1769 OPT_BOOLEAN(0, "no-merge", &no_merge
, "Do not merge identical named events"),
1770 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1771 "print counts with custom separator"),
1772 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1773 "monitor event in cgroup name only", parse_cgroups
),
1774 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1775 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1776 OPT_INTEGER(0, "log-fd", &output_fd
,
1777 "log output to fd, instead of stderr"),
1778 OPT_STRING(0, "pre", &pre_cmd
, "command",
1779 "command to run prior to the measured command"),
1780 OPT_STRING(0, "post", &post_cmd
, "command",
1781 "command to run after to the measured command"),
1782 OPT_UINTEGER('I', "interval-print", &stat_config
.interval
,
1783 "print counts at regular interval in ms (>= 10)"),
1784 OPT_SET_UINT(0, "per-socket", &stat_config
.aggr_mode
,
1785 "aggregate counts per processor socket", AGGR_SOCKET
),
1786 OPT_SET_UINT(0, "per-core", &stat_config
.aggr_mode
,
1787 "aggregate counts per physical processor core", AGGR_CORE
),
1788 OPT_SET_UINT(0, "per-thread", &stat_config
.aggr_mode
,
1789 "aggregate counts per thread", AGGR_THREAD
),
1790 OPT_UINTEGER('D', "delay", &initial_delay
,
1791 "ms to wait before starting measurement after program start"),
1792 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only
, NULL
,
1793 "Only print computed metrics. No raw values", enable_metric_only
),
1794 OPT_BOOLEAN(0, "topdown", &topdown_run
,
1795 "measure topdown level 1 statistics"),
1796 OPT_BOOLEAN(0, "smi-cost", &smi_cost
,
1797 "measure SMI cost"),
1801 static int perf_stat__get_socket(struct cpu_map
*map
, int cpu
)
1803 return cpu_map__get_socket(map
, cpu
, NULL
);
1806 static int perf_stat__get_core(struct cpu_map
*map
, int cpu
)
1808 return cpu_map__get_core(map
, cpu
, NULL
);
1811 static int cpu_map__get_max(struct cpu_map
*map
)
1815 for (i
= 0; i
< map
->nr
; i
++) {
1816 if (map
->map
[i
] > max
)
1823 static struct cpu_map
*cpus_aggr_map
;
1825 static int perf_stat__get_aggr(aggr_get_id_t get_id
, struct cpu_map
*map
, int idx
)
1832 cpu
= map
->map
[idx
];
1834 if (cpus_aggr_map
->map
[cpu
] == -1)
1835 cpus_aggr_map
->map
[cpu
] = get_id(map
, idx
);
1837 return cpus_aggr_map
->map
[cpu
];
1840 static int perf_stat__get_socket_cached(struct cpu_map
*map
, int idx
)
1842 return perf_stat__get_aggr(perf_stat__get_socket
, map
, idx
);
1845 static int perf_stat__get_core_cached(struct cpu_map
*map
, int idx
)
1847 return perf_stat__get_aggr(perf_stat__get_core
, map
, idx
);
1850 static int perf_stat_init_aggr_mode(void)
1854 switch (stat_config
.aggr_mode
) {
1856 if (cpu_map__build_socket_map(evsel_list
->cpus
, &aggr_map
)) {
1857 perror("cannot build socket map");
1860 aggr_get_id
= perf_stat__get_socket_cached
;
1863 if (cpu_map__build_core_map(evsel_list
->cpus
, &aggr_map
)) {
1864 perror("cannot build core map");
1867 aggr_get_id
= perf_stat__get_core_cached
;
1878 * The evsel_list->cpus is the base we operate on,
1879 * taking the highest cpu number to be the size of
1880 * the aggregation translate cpumap.
1882 nr
= cpu_map__get_max(evsel_list
->cpus
);
1883 cpus_aggr_map
= cpu_map__empty_new(nr
+ 1);
1884 return cpus_aggr_map
? 0 : -ENOMEM
;
1887 static void perf_stat__exit_aggr_mode(void)
1889 cpu_map__put(aggr_map
);
1890 cpu_map__put(cpus_aggr_map
);
1892 cpus_aggr_map
= NULL
;
1895 static inline int perf_env__get_cpu(struct perf_env
*env
, struct cpu_map
*map
, int idx
)
1902 cpu
= map
->map
[idx
];
1904 if (cpu
>= env
->nr_cpus_avail
)
1910 static int perf_env__get_socket(struct cpu_map
*map
, int idx
, void *data
)
1912 struct perf_env
*env
= data
;
1913 int cpu
= perf_env__get_cpu(env
, map
, idx
);
1915 return cpu
== -1 ? -1 : env
->cpu
[cpu
].socket_id
;
1918 static int perf_env__get_core(struct cpu_map
*map
, int idx
, void *data
)
1920 struct perf_env
*env
= data
;
1921 int core
= -1, cpu
= perf_env__get_cpu(env
, map
, idx
);
1924 int socket_id
= env
->cpu
[cpu
].socket_id
;
1927 * Encode socket in upper 16 bits
1928 * core_id is relative to socket, and
1929 * we need a global id. So we combine
1932 core
= (socket_id
<< 16) | (env
->cpu
[cpu
].core_id
& 0xffff);
1938 static int perf_env__build_socket_map(struct perf_env
*env
, struct cpu_map
*cpus
,
1939 struct cpu_map
**sockp
)
1941 return cpu_map__build_map(cpus
, sockp
, perf_env__get_socket
, env
);
1944 static int perf_env__build_core_map(struct perf_env
*env
, struct cpu_map
*cpus
,
1945 struct cpu_map
**corep
)
1947 return cpu_map__build_map(cpus
, corep
, perf_env__get_core
, env
);
1950 static int perf_stat__get_socket_file(struct cpu_map
*map
, int idx
)
1952 return perf_env__get_socket(map
, idx
, &perf_stat
.session
->header
.env
);
1955 static int perf_stat__get_core_file(struct cpu_map
*map
, int idx
)
1957 return perf_env__get_core(map
, idx
, &perf_stat
.session
->header
.env
);
1960 static int perf_stat_init_aggr_mode_file(struct perf_stat
*st
)
1962 struct perf_env
*env
= &st
->session
->header
.env
;
1964 switch (stat_config
.aggr_mode
) {
1966 if (perf_env__build_socket_map(env
, evsel_list
->cpus
, &aggr_map
)) {
1967 perror("cannot build socket map");
1970 aggr_get_id
= perf_stat__get_socket_file
;
1973 if (perf_env__build_core_map(env
, evsel_list
->cpus
, &aggr_map
)) {
1974 perror("cannot build core map");
1977 aggr_get_id
= perf_stat__get_core_file
;
1990 static int topdown_filter_events(const char **attr
, char **str
, bool use_group
)
1997 for (i
= 0; attr
[i
]; i
++) {
1998 if (pmu_have_event("cpu", attr
[i
])) {
1999 len
+= strlen(attr
[i
]) + 1;
2000 attr
[i
- off
] = attr
[i
];
2004 attr
[i
- off
] = NULL
;
2006 *str
= malloc(len
+ 1 + 2);
2016 for (i
= 0; attr
[i
]; i
++) {
2029 __weak
bool arch_topdown_check_group(bool *warn
)
2035 __weak
void arch_topdown_group_warn(void)
2040 * Add default attributes, if there were no attributes specified or
2041 * if -d/--detailed, -d -d or -d -d -d is used:
2043 static int add_default_attributes(void)
2046 struct perf_event_attr default_attrs0
[] = {
2048 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
2049 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
2050 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
2051 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
2053 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
2055 struct perf_event_attr frontend_attrs
[] = {
2056 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
2058 struct perf_event_attr backend_attrs
[] = {
2059 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
2061 struct perf_event_attr default_attrs1
[] = {
2062 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
2063 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
2064 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
2069 * Detailed stats (-d), covering the L1 and last level data caches:
2071 struct perf_event_attr detailed_attrs
[] = {
2073 { .type
= PERF_TYPE_HW_CACHE
,
2075 PERF_COUNT_HW_CACHE_L1D
<< 0 |
2076 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2077 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
2079 { .type
= PERF_TYPE_HW_CACHE
,
2081 PERF_COUNT_HW_CACHE_L1D
<< 0 |
2082 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2083 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
2085 { .type
= PERF_TYPE_HW_CACHE
,
2087 PERF_COUNT_HW_CACHE_LL
<< 0 |
2088 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2089 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
2091 { .type
= PERF_TYPE_HW_CACHE
,
2093 PERF_COUNT_HW_CACHE_LL
<< 0 |
2094 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2095 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
2099 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2101 struct perf_event_attr very_detailed_attrs
[] = {
2103 { .type
= PERF_TYPE_HW_CACHE
,
2105 PERF_COUNT_HW_CACHE_L1I
<< 0 |
2106 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2107 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
2109 { .type
= PERF_TYPE_HW_CACHE
,
2111 PERF_COUNT_HW_CACHE_L1I
<< 0 |
2112 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2113 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
2115 { .type
= PERF_TYPE_HW_CACHE
,
2117 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
2118 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2119 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
2121 { .type
= PERF_TYPE_HW_CACHE
,
2123 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
2124 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2125 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
2127 { .type
= PERF_TYPE_HW_CACHE
,
2129 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
2130 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2131 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
2133 { .type
= PERF_TYPE_HW_CACHE
,
2135 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
2136 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
2137 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
2142 * Very, very detailed stats (-d -d -d), adding prefetch events:
2144 struct perf_event_attr very_very_detailed_attrs
[] = {
2146 { .type
= PERF_TYPE_HW_CACHE
,
2148 PERF_COUNT_HW_CACHE_L1D
<< 0 |
2149 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
2150 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
2152 { .type
= PERF_TYPE_HW_CACHE
,
2154 PERF_COUNT_HW_CACHE_L1D
<< 0 |
2155 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
2156 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
2159 /* Set attrs if no event is selected and !null_run: */
2163 if (transaction_run
) {
2164 if (pmu_have_event("cpu", "cycles-ct") &&
2165 pmu_have_event("cpu", "el-start"))
2166 err
= parse_events(evsel_list
, transaction_attrs
, NULL
);
2168 err
= parse_events(evsel_list
, transaction_limited_attrs
, NULL
);
2170 fprintf(stderr
, "Cannot set up transaction events\n");
2179 if (sysfs__read_int(FREEZE_ON_SMI_PATH
, &smi
) < 0) {
2180 fprintf(stderr
, "freeze_on_smi is not supported.\n");
2185 if (sysfs__write_int(FREEZE_ON_SMI_PATH
, 1) < 0) {
2186 fprintf(stderr
, "Failed to set freeze_on_smi.\n");
2192 if (pmu_have_event("msr", "aperf") &&
2193 pmu_have_event("msr", "smi")) {
2194 if (!force_metric_only
)
2196 err
= parse_events(evsel_list
, smi_cost_attrs
, NULL
);
2198 fprintf(stderr
, "To measure SMI cost, it needs "
2199 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2203 fprintf(stderr
, "Cannot set up SMI cost events\n");
2213 if (stat_config
.aggr_mode
!= AGGR_GLOBAL
&&
2214 stat_config
.aggr_mode
!= AGGR_CORE
) {
2215 pr_err("top down event configuration requires --per-core mode\n");
2218 stat_config
.aggr_mode
= AGGR_CORE
;
2219 if (nr_cgroups
|| !target__has_cpu(&target
)) {
2220 pr_err("top down event configuration requires system-wide mode (-a)\n");
2224 if (!force_metric_only
)
2226 if (topdown_filter_events(topdown_attrs
, &str
,
2227 arch_topdown_check_group(&warn
)) < 0) {
2228 pr_err("Out of memory\n");
2231 if (topdown_attrs
[0] && str
) {
2233 arch_topdown_group_warn();
2234 err
= parse_events(evsel_list
, str
, NULL
);
2237 "Cannot set up top down events %s: %d\n",
2243 fprintf(stderr
, "System does not support topdown\n");
2249 if (!evsel_list
->nr_entries
) {
2250 if (target__has_cpu(&target
))
2251 default_attrs0
[0].config
= PERF_COUNT_SW_CPU_CLOCK
;
2253 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs0
) < 0)
2255 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2256 if (perf_evlist__add_default_attrs(evsel_list
,
2257 frontend_attrs
) < 0)
2260 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2261 if (perf_evlist__add_default_attrs(evsel_list
,
2265 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs1
) < 0)
2269 /* Detailed events get appended to the event list: */
2271 if (detailed_run
< 1)
2274 /* Append detailed run extra attributes: */
2275 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
2278 if (detailed_run
< 2)
2281 /* Append very detailed run extra attributes: */
2282 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
2285 if (detailed_run
< 3)
2288 /* Append very, very detailed run extra attributes: */
2289 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
2292 static const char * const stat_record_usage
[] = {
2293 "perf stat record [<options>]",
2297 static void init_features(struct perf_session
*session
)
2301 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
2302 perf_header__set_feat(&session
->header
, feat
);
2304 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
2305 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
2306 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
2307 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
2310 static int __cmd_record(int argc
, const char **argv
)
2312 struct perf_session
*session
;
2313 struct perf_data_file
*file
= &perf_stat
.file
;
2315 argc
= parse_options(argc
, argv
, stat_options
, stat_record_usage
,
2316 PARSE_OPT_STOP_AT_NON_OPTION
);
2319 file
->path
= output_name
;
2321 if (run_count
!= 1 || forever
) {
2322 pr_err("Cannot use -r option with perf stat record.\n");
2326 session
= perf_session__new(file
, false, NULL
);
2327 if (session
== NULL
) {
2328 pr_err("Perf session creation failed.\n");
2332 init_features(session
);
2334 session
->evlist
= evsel_list
;
2335 perf_stat
.session
= session
;
2336 perf_stat
.record
= true;
2340 static int process_stat_round_event(struct perf_tool
*tool __maybe_unused
,
2341 union perf_event
*event
,
2342 struct perf_session
*session
)
2344 struct stat_round_event
*stat_round
= &event
->stat_round
;
2345 struct perf_evsel
*counter
;
2346 struct timespec tsh
, *ts
= NULL
;
2347 const char **argv
= session
->header
.env
.cmdline_argv
;
2348 int argc
= session
->header
.env
.nr_cmdline
;
2350 evlist__for_each_entry(evsel_list
, counter
)
2351 perf_stat_process_counter(&stat_config
, counter
);
2353 if (stat_round
->type
== PERF_STAT_ROUND_TYPE__FINAL
)
2354 update_stats(&walltime_nsecs_stats
, stat_round
->time
);
2356 if (stat_config
.interval
&& stat_round
->time
) {
2357 tsh
.tv_sec
= stat_round
->time
/ NSEC_PER_SEC
;
2358 tsh
.tv_nsec
= stat_round
->time
% NSEC_PER_SEC
;
2362 print_counters(ts
, argc
, argv
);
2367 int process_stat_config_event(struct perf_tool
*tool
,
2368 union perf_event
*event
,
2369 struct perf_session
*session __maybe_unused
)
2371 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2373 perf_event__read_stat_config(&stat_config
, &event
->stat_config
);
2375 if (cpu_map__empty(st
->cpus
)) {
2376 if (st
->aggr_mode
!= AGGR_UNSET
)
2377 pr_warning("warning: processing task data, aggregation mode not set\n");
2381 if (st
->aggr_mode
!= AGGR_UNSET
)
2382 stat_config
.aggr_mode
= st
->aggr_mode
;
2384 if (perf_stat
.file
.is_pipe
)
2385 perf_stat_init_aggr_mode();
2387 perf_stat_init_aggr_mode_file(st
);
2392 static int set_maps(struct perf_stat
*st
)
2394 if (!st
->cpus
|| !st
->threads
)
2397 if (WARN_ONCE(st
->maps_allocated
, "stats double allocation\n"))
2400 perf_evlist__set_maps(evsel_list
, st
->cpus
, st
->threads
);
2402 if (perf_evlist__alloc_stats(evsel_list
, true))
2405 st
->maps_allocated
= true;
2410 int process_thread_map_event(struct perf_tool
*tool
,
2411 union perf_event
*event
,
2412 struct perf_session
*session __maybe_unused
)
2414 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2417 pr_warning("Extra thread map event, ignoring.\n");
2421 st
->threads
= thread_map__new_event(&event
->thread_map
);
2425 return set_maps(st
);
2429 int process_cpu_map_event(struct perf_tool
*tool
,
2430 union perf_event
*event
,
2431 struct perf_session
*session __maybe_unused
)
2433 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2434 struct cpu_map
*cpus
;
2437 pr_warning("Extra cpu map event, ignoring.\n");
2441 cpus
= cpu_map__new_data(&event
->cpu_map
.data
);
2446 return set_maps(st
);
2449 static const char * const stat_report_usage
[] = {
2450 "perf stat report [<options>]",
2454 static struct perf_stat perf_stat
= {
2456 .attr
= perf_event__process_attr
,
2457 .event_update
= perf_event__process_event_update
,
2458 .thread_map
= process_thread_map_event
,
2459 .cpu_map
= process_cpu_map_event
,
2460 .stat_config
= process_stat_config_event
,
2461 .stat
= perf_event__process_stat_event
,
2462 .stat_round
= process_stat_round_event
,
2464 .aggr_mode
= AGGR_UNSET
,
2467 static int __cmd_report(int argc
, const char **argv
)
2469 struct perf_session
*session
;
2470 const struct option options
[] = {
2471 OPT_STRING('i', "input", &input_name
, "file", "input file name"),
2472 OPT_SET_UINT(0, "per-socket", &perf_stat
.aggr_mode
,
2473 "aggregate counts per processor socket", AGGR_SOCKET
),
2474 OPT_SET_UINT(0, "per-core", &perf_stat
.aggr_mode
,
2475 "aggregate counts per physical processor core", AGGR_CORE
),
2476 OPT_SET_UINT('A', "no-aggr", &perf_stat
.aggr_mode
,
2477 "disable CPU count aggregation", AGGR_NONE
),
2483 argc
= parse_options(argc
, argv
, options
, stat_report_usage
, 0);
2485 if (!input_name
|| !strlen(input_name
)) {
2486 if (!fstat(STDIN_FILENO
, &st
) && S_ISFIFO(st
.st_mode
))
2489 input_name
= "perf.data";
2492 perf_stat
.file
.path
= input_name
;
2493 perf_stat
.file
.mode
= PERF_DATA_MODE_READ
;
2495 session
= perf_session__new(&perf_stat
.file
, false, &perf_stat
.tool
);
2496 if (session
== NULL
)
2499 perf_stat
.session
= session
;
2500 stat_config
.output
= stderr
;
2501 evsel_list
= session
->evlist
;
2503 ret
= perf_session__process_events(session
);
2507 perf_session__delete(session
);
2511 static void setup_system_wide(int forks
)
2514 * Make system wide (-a) the default target if
2515 * no target was specified and one of following
2516 * conditions is met:
2518 * - there's no workload specified
2519 * - there is workload specified but all requested
2520 * events are system wide events
2522 if (!target__none(&target
))
2526 target
.system_wide
= true;
2528 struct perf_evsel
*counter
;
2530 evlist__for_each_entry(evsel_list
, counter
) {
2531 if (!counter
->system_wide
)
2535 if (evsel_list
->nr_entries
)
2536 target
.system_wide
= true;
2540 int cmd_stat(int argc
, const char **argv
)
2542 const char * const stat_usage
[] = {
2543 "perf stat [<options>] [<command>]",
2546 int status
= -EINVAL
, run_idx
;
2548 FILE *output
= stderr
;
2549 unsigned int interval
;
2550 const char * const stat_subcommands
[] = { "record", "report" };
2552 setlocale(LC_ALL
, "");
2554 evsel_list
= perf_evlist__new();
2555 if (evsel_list
== NULL
)
2558 parse_events__shrink_config_terms();
2559 argc
= parse_options_subcommand(argc
, argv
, stat_options
, stat_subcommands
,
2560 (const char **) stat_usage
,
2561 PARSE_OPT_STOP_AT_NON_OPTION
);
2562 perf_stat__collect_metric_expr(evsel_list
);
2563 perf_stat__init_shadow_stats();
2567 if (!strcmp(csv_sep
, "\\t"))
2570 csv_sep
= DEFAULT_SEPARATOR
;
2572 if (argc
&& !strncmp(argv
[0], "rec", 3)) {
2573 argc
= __cmd_record(argc
, argv
);
2576 } else if (argc
&& !strncmp(argv
[0], "rep", 3))
2577 return __cmd_report(argc
, argv
);
2579 interval
= stat_config
.interval
;
2582 * For record command the -o is already taken care of.
2584 if (!STAT_RECORD
&& output_name
&& strcmp(output_name
, "-"))
2587 if (output_name
&& output_fd
) {
2588 fprintf(stderr
, "cannot use both --output and --log-fd\n");
2589 parse_options_usage(stat_usage
, stat_options
, "o", 1);
2590 parse_options_usage(NULL
, stat_options
, "log-fd", 0);
2594 if (metric_only
&& stat_config
.aggr_mode
== AGGR_THREAD
) {
2595 fprintf(stderr
, "--metric-only is not supported with --per-thread\n");
2599 if (metric_only
&& run_count
> 1) {
2600 fprintf(stderr
, "--metric-only is not supported with -r\n");
2604 if (output_fd
< 0) {
2605 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
2606 parse_options_usage(stat_usage
, stat_options
, "log-fd", 0);
2612 mode
= append_file
? "a" : "w";
2614 output
= fopen(output_name
, mode
);
2616 perror("failed to create output file");
2619 clock_gettime(CLOCK_REALTIME
, &tm
);
2620 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
2621 } else if (output_fd
> 0) {
2622 mode
= append_file
? "a" : "w";
2623 output
= fdopen(output_fd
, mode
);
2625 perror("Failed opening logfd");
2630 stat_config
.output
= output
;
2633 * let the spreadsheet do the pretty-printing
2636 /* User explicitly passed -B? */
2637 if (big_num_opt
== 1) {
2638 fprintf(stderr
, "-B option not supported with -x\n");
2639 parse_options_usage(stat_usage
, stat_options
, "B", 1);
2640 parse_options_usage(NULL
, stat_options
, "x", 1);
2642 } else /* Nope, so disable big number formatting */
2644 } else if (big_num_opt
== 0) /* User passed --no-big-num */
2647 setup_system_wide(argc
);
2649 if (run_count
< 0) {
2650 pr_err("Run count must be a positive number\n");
2651 parse_options_usage(stat_usage
, stat_options
, "r", 1);
2653 } else if (run_count
== 0) {
2658 if ((stat_config
.aggr_mode
== AGGR_THREAD
) && !target__has_task(&target
)) {
2659 fprintf(stderr
, "The --per-thread option is only available "
2660 "when monitoring via -p -t options.\n");
2661 parse_options_usage(NULL
, stat_options
, "p", 1);
2662 parse_options_usage(NULL
, stat_options
, "t", 1);
2667 * no_aggr, cgroup are for system-wide only
2668 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2670 if (((stat_config
.aggr_mode
!= AGGR_GLOBAL
&&
2671 stat_config
.aggr_mode
!= AGGR_THREAD
) || nr_cgroups
) &&
2672 !target__has_cpu(&target
)) {
2673 fprintf(stderr
, "both cgroup and no-aggregation "
2674 "modes only available in system-wide mode\n");
2676 parse_options_usage(stat_usage
, stat_options
, "G", 1);
2677 parse_options_usage(NULL
, stat_options
, "A", 1);
2678 parse_options_usage(NULL
, stat_options
, "a", 1);
2682 if (add_default_attributes())
2685 target__validate(&target
);
2687 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
2688 if (target__has_task(&target
)) {
2689 pr_err("Problems finding threads of monitor\n");
2690 parse_options_usage(stat_usage
, stat_options
, "p", 1);
2691 parse_options_usage(NULL
, stat_options
, "t", 1);
2692 } else if (target__has_cpu(&target
)) {
2693 perror("failed to parse CPUs map");
2694 parse_options_usage(stat_usage
, stat_options
, "C", 1);
2695 parse_options_usage(NULL
, stat_options
, "a", 1);
2701 * Initialize thread_map with comm names,
2702 * so we could print it out on output.
2704 if (stat_config
.aggr_mode
== AGGR_THREAD
)
2705 thread_map__read_comms(evsel_list
->threads
);
2707 if (interval
&& interval
< 100) {
2708 if (interval
< 10) {
2709 pr_err("print interval must be >= 10ms\n");
2710 parse_options_usage(stat_usage
, stat_options
, "I", 1);
2713 pr_warning("print interval < 100ms. "
2714 "The overhead percentage could be high in some cases. "
2715 "Please proceed with caution.\n");
2718 if (perf_evlist__alloc_stats(evsel_list
, interval
))
2721 if (perf_stat_init_aggr_mode())
2725 * We dont want to block the signals - that would cause
2726 * child tasks to inherit that and Ctrl-C would not work.
2727 * What we want is for Ctrl-C to work in the exec()-ed
2728 * task, but being ignored by perf stat itself:
2732 signal(SIGINT
, skip_signal
);
2733 signal(SIGCHLD
, skip_signal
);
2734 signal(SIGALRM
, skip_signal
);
2735 signal(SIGABRT
, skip_signal
);
2738 for (run_idx
= 0; forever
|| run_idx
< run_count
; run_idx
++) {
2739 if (run_count
!= 1 && verbose
> 0)
2740 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
2743 status
= run_perf_stat(argc
, argv
);
2744 if (forever
&& status
!= -1) {
2745 print_counters(NULL
, argc
, argv
);
2746 perf_stat__reset_stats();
2750 if (!forever
&& status
!= -1 && !interval
)
2751 print_counters(NULL
, argc
, argv
);
2755 * We synthesize the kernel mmap record just so that older tools
2756 * don't emit warnings about not being able to resolve symbols
2757 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2758 * a saner message about no samples being in the perf.data file.
2760 * This also serves to suppress a warning about f_header.data.size == 0
2761 * in header.c at the moment 'perf stat record' gets introduced, which
2762 * is not really needed once we start adding the stat specific PERF_RECORD_
2763 * records, but the need to suppress the kptr_restrict messages in older
2764 * tools remain -acme
2766 int fd
= perf_data_file__fd(&perf_stat
.file
);
2767 int err
= perf_event__synthesize_kernel_mmap((void *)&perf_stat
,
2768 process_synthesized_event
,
2769 &perf_stat
.session
->machines
.host
);
2771 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2772 "older tools may produce warnings about this file\n.");
2776 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats
.max
, FINAL
))
2777 pr_err("failed to write stat round event\n");
2780 if (!perf_stat
.file
.is_pipe
) {
2781 perf_stat
.session
->header
.data_size
+= perf_stat
.bytes_written
;
2782 perf_session__write_header(perf_stat
.session
, evsel_list
, fd
, true);
2785 perf_session__delete(perf_stat
.session
);
2788 perf_stat__exit_aggr_mode();
2789 perf_evlist__free_stats(evsel_list
);
2791 if (smi_cost
&& smi_reset
)
2792 sysfs__write_int(FREEZE_ON_SMI_PATH
, 0);
2794 perf_evlist__delete(evsel_list
);