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/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
61 #include "util/counts.h"
62 #include "util/session.h"
63 #include "util/tool.h"
67 #include <sys/prctl.h>
70 #define DEFAULT_SEPARATOR " "
71 #define CNTR_NOT_SUPPORTED "<not supported>"
72 #define CNTR_NOT_COUNTED "<not counted>"
74 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
);
76 /* Default events used for perf stat -T */
77 static const char *transaction_attrs
= {
89 /* More limited version when the CPU does not have all events. */
90 static const char * transaction_limited_attrs
= {
100 static struct perf_evlist
*evsel_list
;
102 static struct target target
= {
106 typedef int (*aggr_get_id_t
)(struct cpu_map
*m
, int cpu
);
108 static int run_count
= 1;
109 static bool no_inherit
= false;
110 static volatile pid_t child_pid
= -1;
111 static bool null_run
= false;
112 static int detailed_run
= 0;
113 static bool transaction_run
;
114 static bool big_num
= true;
115 static int big_num_opt
= -1;
116 static const char *csv_sep
= NULL
;
117 static bool csv_output
= false;
118 static bool group
= false;
119 static const char *pre_cmd
= NULL
;
120 static const char *post_cmd
= NULL
;
121 static bool sync_run
= false;
122 static unsigned int initial_delay
= 0;
123 static unsigned int unit_width
= 4; /* strlen("unit") */
124 static bool forever
= false;
125 static bool metric_only
= false;
126 static struct timespec ref_time
;
127 static struct cpu_map
*aggr_map
;
128 static aggr_get_id_t aggr_get_id
;
129 static bool append_file
;
130 static const char *output_name
;
131 static int output_fd
;
135 struct perf_data_file file
;
136 struct perf_session
*session
;
138 struct perf_tool tool
;
140 struct cpu_map
*cpus
;
141 struct thread_map
*threads
;
142 enum aggr_mode aggr_mode
;
145 static struct perf_stat perf_stat
;
146 #define STAT_RECORD perf_stat.record
148 static volatile int done
= 0;
150 static struct perf_stat_config stat_config
= {
151 .aggr_mode
= AGGR_GLOBAL
,
155 static inline void diff_timespec(struct timespec
*r
, struct timespec
*a
,
158 r
->tv_sec
= a
->tv_sec
- b
->tv_sec
;
159 if (a
->tv_nsec
< b
->tv_nsec
) {
160 r
->tv_nsec
= a
->tv_nsec
+ 1000000000L - b
->tv_nsec
;
163 r
->tv_nsec
= a
->tv_nsec
- b
->tv_nsec
;
167 static void perf_stat__reset_stats(void)
169 perf_evlist__reset_stats(evsel_list
);
170 perf_stat__reset_shadow_stats();
173 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
175 struct perf_event_attr
*attr
= &evsel
->attr
;
177 if (stat_config
.scale
)
178 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
179 PERF_FORMAT_TOTAL_TIME_RUNNING
;
181 attr
->inherit
= !no_inherit
;
184 * Some events get initialized with sample_(period/type) set,
185 * like tracepoints. Clear it up for counting.
187 attr
->sample_period
= 0;
190 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
191 * while avoiding that older tools show confusing messages.
193 * However for pipe sessions we need to keep it zero,
194 * because script's perf_evsel__check_attr is triggered
195 * by attr->sample_type != 0, and we can't run it on
198 if (!(STAT_RECORD
&& perf_stat
.file
.is_pipe
))
199 attr
->sample_type
= PERF_SAMPLE_IDENTIFIER
;
202 * Disabling all counters initially, they will be enabled
203 * either manually by us or by kernel via enable_on_exec
206 if (perf_evsel__is_group_leader(evsel
)) {
210 * In case of initial_delay we enable tracee
213 if (target__none(&target
) && !initial_delay
)
214 attr
->enable_on_exec
= 1;
217 if (target__has_cpu(&target
))
218 return perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
220 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
224 * Does the counter have nsecs as a unit?
226 static inline int nsec_counter(struct perf_evsel
*evsel
)
228 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
229 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
235 static int process_synthesized_event(struct perf_tool
*tool __maybe_unused
,
236 union perf_event
*event
,
237 struct perf_sample
*sample __maybe_unused
,
238 struct machine
*machine __maybe_unused
)
240 if (perf_data_file__write(&perf_stat
.file
, event
, event
->header
.size
) < 0) {
241 pr_err("failed to write perf data, error: %m\n");
245 perf_stat
.bytes_written
+= event
->header
.size
;
249 static int write_stat_round_event(u64 tm
, u64 type
)
251 return perf_event__synthesize_stat_round(NULL
, tm
, type
,
252 process_synthesized_event
,
256 #define WRITE_STAT_ROUND_EVENT(time, interval) \
257 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
259 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
262 perf_evsel__write_stat_event(struct perf_evsel
*counter
, u32 cpu
, u32 thread
,
263 struct perf_counts_values
*count
)
265 struct perf_sample_id
*sid
= SID(counter
, cpu
, thread
);
267 return perf_event__synthesize_stat(NULL
, cpu
, thread
, sid
->id
, count
,
268 process_synthesized_event
, NULL
);
272 * Read out the results of a single counter:
273 * do not aggregate counts across CPUs in system-wide mode
275 static int read_counter(struct perf_evsel
*counter
)
277 int nthreads
= thread_map__nr(evsel_list
->threads
);
278 int ncpus
= perf_evsel__nr_cpus(counter
);
281 if (!counter
->supported
)
284 if (counter
->system_wide
)
287 for (thread
= 0; thread
< nthreads
; thread
++) {
288 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
289 struct perf_counts_values
*count
;
291 count
= perf_counts(counter
->counts
, cpu
, thread
);
292 if (perf_evsel__read(counter
, cpu
, thread
, count
))
296 if (perf_evsel__write_stat_event(counter
, cpu
, thread
, count
)) {
297 pr_err("failed to write stat event\n");
307 static void read_counters(bool close_counters
)
309 struct perf_evsel
*counter
;
311 evlist__for_each(evsel_list
, counter
) {
312 if (read_counter(counter
))
313 pr_debug("failed to read counter %s\n", counter
->name
);
315 if (perf_stat_process_counter(&stat_config
, counter
))
316 pr_warning("failed to process counter %s\n", counter
->name
);
318 if (close_counters
) {
319 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
320 thread_map__nr(evsel_list
->threads
));
325 static void process_interval(void)
327 struct timespec ts
, rs
;
329 read_counters(false);
331 clock_gettime(CLOCK_MONOTONIC
, &ts
);
332 diff_timespec(&rs
, &ts
, &ref_time
);
335 if (WRITE_STAT_ROUND_EVENT(rs
.tv_sec
* NSECS_PER_SEC
+ rs
.tv_nsec
, INTERVAL
))
336 pr_err("failed to write stat round event\n");
339 print_counters(&rs
, 0, NULL
);
342 static void enable_counters(void)
345 usleep(initial_delay
* 1000);
348 * We need to enable counters only if:
349 * - we don't have tracee (attaching to task or cpu)
350 * - we have initial delay configured
352 if (!target__none(&target
) || initial_delay
)
353 perf_evlist__enable(evsel_list
);
356 static volatile int workload_exec_errno
;
359 * perf_evlist__prepare_workload will send a SIGUSR1
360 * if the fork fails, since we asked by setting its
361 * want_signal to true.
363 static void workload_exec_failed_signal(int signo __maybe_unused
, siginfo_t
*info
,
364 void *ucontext __maybe_unused
)
366 workload_exec_errno
= info
->si_value
.sival_int
;
369 static bool has_unit(struct perf_evsel
*counter
)
371 return counter
->unit
&& *counter
->unit
;
374 static bool has_scale(struct perf_evsel
*counter
)
376 return counter
->scale
!= 1;
379 static int perf_stat_synthesize_config(bool is_pipe
)
381 struct perf_evsel
*counter
;
385 err
= perf_event__synthesize_attrs(NULL
, perf_stat
.session
,
386 process_synthesized_event
);
388 pr_err("Couldn't synthesize attrs.\n");
394 * Synthesize other events stuff not carried within
395 * attr event - unit, scale, name
397 evlist__for_each(evsel_list
, counter
) {
398 if (!counter
->supported
)
402 * Synthesize unit and scale only if it's defined.
404 if (has_unit(counter
)) {
405 err
= perf_event__synthesize_event_update_unit(NULL
, counter
, process_synthesized_event
);
407 pr_err("Couldn't synthesize evsel unit.\n");
412 if (has_scale(counter
)) {
413 err
= perf_event__synthesize_event_update_scale(NULL
, counter
, process_synthesized_event
);
415 pr_err("Couldn't synthesize evsel scale.\n");
420 if (counter
->own_cpus
) {
421 err
= perf_event__synthesize_event_update_cpus(NULL
, counter
, process_synthesized_event
);
423 pr_err("Couldn't synthesize evsel scale.\n");
429 * Name is needed only for pipe output,
430 * perf.data carries event names.
433 err
= perf_event__synthesize_event_update_name(NULL
, counter
, process_synthesized_event
);
435 pr_err("Couldn't synthesize evsel name.\n");
441 err
= perf_event__synthesize_thread_map2(NULL
, evsel_list
->threads
,
442 process_synthesized_event
,
445 pr_err("Couldn't synthesize thread map.\n");
449 err
= perf_event__synthesize_cpu_map(NULL
, evsel_list
->cpus
,
450 process_synthesized_event
, NULL
);
452 pr_err("Couldn't synthesize thread map.\n");
456 err
= perf_event__synthesize_stat_config(NULL
, &stat_config
,
457 process_synthesized_event
, NULL
);
459 pr_err("Couldn't synthesize config.\n");
466 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
468 static int __store_counter_ids(struct perf_evsel
*counter
,
469 struct cpu_map
*cpus
,
470 struct thread_map
*threads
)
474 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
475 for (thread
= 0; thread
< threads
->nr
; thread
++) {
476 int fd
= FD(counter
, cpu
, thread
);
478 if (perf_evlist__id_add_fd(evsel_list
, counter
,
479 cpu
, thread
, fd
) < 0)
487 static int store_counter_ids(struct perf_evsel
*counter
)
489 struct cpu_map
*cpus
= counter
->cpus
;
490 struct thread_map
*threads
= counter
->threads
;
492 if (perf_evsel__alloc_id(counter
, cpus
->nr
, threads
->nr
))
495 return __store_counter_ids(counter
, cpus
, threads
);
498 static int __run_perf_stat(int argc
, const char **argv
)
500 int interval
= stat_config
.interval
;
502 unsigned long long t0
, t1
;
503 struct perf_evsel
*counter
;
507 const bool forks
= (argc
> 0);
508 bool is_pipe
= STAT_RECORD
? perf_stat
.file
.is_pipe
: false;
511 ts
.tv_sec
= interval
/ 1000;
512 ts
.tv_nsec
= (interval
% 1000) * 1000000;
519 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
, is_pipe
,
520 workload_exec_failed_signal
) < 0) {
521 perror("failed to prepare workload");
524 child_pid
= evsel_list
->workload
.pid
;
528 perf_evlist__set_leader(evsel_list
);
530 evlist__for_each(evsel_list
, counter
) {
532 if (create_perf_stat_counter(counter
) < 0) {
534 * PPC returns ENXIO for HW counters until 2.6.37
535 * (behavior changed with commit b0a873e).
537 if (errno
== EINVAL
|| errno
== ENOSYS
||
538 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
541 ui__warning("%s event is not supported by the kernel.\n",
542 perf_evsel__name(counter
));
543 counter
->supported
= false;
545 if ((counter
->leader
!= counter
) ||
546 !(counter
->leader
->nr_members
> 1))
548 } else if (perf_evsel__fallback(counter
, errno
, msg
, sizeof(msg
))) {
550 ui__warning("%s\n", msg
);
554 perf_evsel__open_strerror(counter
, &target
,
555 errno
, msg
, sizeof(msg
));
556 ui__error("%s\n", msg
);
559 kill(child_pid
, SIGTERM
);
563 counter
->supported
= true;
565 l
= strlen(counter
->unit
);
569 if (STAT_RECORD
&& store_counter_ids(counter
))
573 if (perf_evlist__apply_filters(evsel_list
, &counter
)) {
574 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
575 counter
->filter
, perf_evsel__name(counter
), errno
,
576 strerror_r(errno
, msg
, sizeof(msg
)));
581 int err
, fd
= perf_data_file__fd(&perf_stat
.file
);
584 err
= perf_header__write_pipe(perf_data_file__fd(&perf_stat
.file
));
586 err
= perf_session__write_header(perf_stat
.session
, evsel_list
,
593 err
= perf_stat_synthesize_config(is_pipe
);
599 * Enable counters and exec the command:
602 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
605 perf_evlist__start_workload(evsel_list
);
609 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
610 nanosleep(&ts
, NULL
);
616 if (workload_exec_errno
) {
617 const char *emsg
= strerror_r(workload_exec_errno
, msg
, sizeof(msg
));
618 pr_err("Workload failed: %s\n", emsg
);
622 if (WIFSIGNALED(status
))
623 psignal(WTERMSIG(status
), argv
[0]);
627 nanosleep(&ts
, NULL
);
635 update_stats(&walltime_nsecs_stats
, t1
- t0
);
639 return WEXITSTATUS(status
);
642 static int run_perf_stat(int argc
, const char **argv
)
647 ret
= system(pre_cmd
);
655 ret
= __run_perf_stat(argc
, argv
);
660 ret
= system(post_cmd
);
668 static void print_running(u64 run
, u64 ena
)
671 fprintf(stat_config
.output
, "%s%" PRIu64
"%s%.2f",
675 ena
? 100.0 * run
/ ena
: 100.0);
676 } else if (run
!= ena
) {
677 fprintf(stat_config
.output
, " (%.2f%%)", 100.0 * run
/ ena
);
681 static void print_noise_pct(double total
, double avg
)
683 double pct
= rel_stddev_stats(total
, avg
);
686 fprintf(stat_config
.output
, "%s%.2f%%", csv_sep
, pct
);
688 fprintf(stat_config
.output
, " ( +-%6.2f%% )", pct
);
691 static void print_noise(struct perf_evsel
*evsel
, double avg
)
693 struct perf_stat_evsel
*ps
;
699 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
702 static void aggr_printout(struct perf_evsel
*evsel
, int id
, int nr
)
704 switch (stat_config
.aggr_mode
) {
706 fprintf(stat_config
.output
, "S%d-C%*d%s%*d%s",
707 cpu_map__id_to_socket(id
),
709 cpu_map__id_to_cpu(id
),
716 fprintf(stat_config
.output
, "S%*d%s%*d%s",
725 fprintf(stat_config
.output
, "CPU%*d%s",
727 perf_evsel__cpus(evsel
)->map
[id
], csv_sep
);
730 fprintf(stat_config
.output
, "%*s-%*d%s",
732 thread_map__comm(evsel
->threads
, id
),
734 thread_map__pid(evsel
->threads
, id
),
750 struct perf_evsel
*evsel
;
753 #define METRIC_LEN 35
755 static void new_line_std(void *ctx
)
757 struct outstate
*os
= ctx
;
762 static void do_new_line_std(struct outstate
*os
)
765 fputs(os
->prefix
, os
->fh
);
766 aggr_printout(os
->evsel
, os
->id
, os
->nr
);
767 if (stat_config
.aggr_mode
== AGGR_NONE
)
768 fprintf(os
->fh
, " ");
769 fprintf(os
->fh
, " ");
772 static void print_metric_std(void *ctx
, const char *color
, const char *fmt
,
773 const char *unit
, double val
)
775 struct outstate
*os
= ctx
;
778 bool newline
= os
->newline
;
782 if (unit
== NULL
|| fmt
== NULL
) {
783 fprintf(out
, "%-*s", METRIC_LEN
, "");
790 n
= fprintf(out
, " # ");
792 n
+= color_fprintf(out
, color
, fmt
, val
);
794 n
+= fprintf(out
, fmt
, val
);
795 fprintf(out
, " %-*s", METRIC_LEN
- n
- 1, unit
);
798 static void new_line_csv(void *ctx
)
800 struct outstate
*os
= ctx
;
805 fprintf(os
->fh
, "%s%s", os
->prefix
, csv_sep
);
806 aggr_printout(os
->evsel
, os
->id
, os
->nr
);
807 for (i
= 0; i
< os
->nfields
; i
++)
808 fputs(csv_sep
, os
->fh
);
811 static void print_metric_csv(void *ctx
,
812 const char *color __maybe_unused
,
813 const char *fmt
, const char *unit
, double val
)
815 struct outstate
*os
= ctx
;
817 char buf
[64], *vals
, *ends
;
819 if (unit
== NULL
|| fmt
== NULL
) {
820 fprintf(out
, "%s%s%s%s", csv_sep
, csv_sep
, csv_sep
, csv_sep
);
823 snprintf(buf
, sizeof(buf
), fmt
, val
);
825 while (isspace(*vals
))
828 while (isdigit(*ends
) || *ends
== '.')
831 while (isspace(*unit
))
833 fprintf(out
, "%s%s%s%s", csv_sep
, vals
, csv_sep
, unit
);
836 #define METRIC_ONLY_LEN 20
838 /* Filter out some columns that don't work well in metrics only mode */
840 static bool valid_only_metric(const char *unit
)
844 if (strstr(unit
, "/sec") ||
845 strstr(unit
, "hz") ||
846 strstr(unit
, "Hz") ||
847 strstr(unit
, "CPUs utilized"))
852 static const char *fixunit(char *buf
, struct perf_evsel
*evsel
,
855 if (!strncmp(unit
, "of all", 6)) {
856 snprintf(buf
, 1024, "%s %s", perf_evsel__name(evsel
),
863 static void print_metric_only(void *ctx
, const char *color
, const char *fmt
,
864 const char *unit
, double val
)
866 struct outstate
*os
= ctx
;
870 unsigned mlen
= METRIC_ONLY_LEN
;
872 if (!valid_only_metric(unit
))
874 unit
= fixunit(buf
, os
->evsel
, unit
);
876 n
= color_fprintf(out
, color
, fmt
, val
);
878 n
= fprintf(out
, fmt
, val
);
879 if (n
> METRIC_ONLY_LEN
)
881 if (mlen
< strlen(unit
))
882 mlen
= strlen(unit
) + 1;
883 fprintf(out
, "%*s", mlen
- n
, "");
886 static void print_metric_only_csv(void *ctx
, const char *color __maybe_unused
,
888 const char *unit
, double val
)
890 struct outstate
*os
= ctx
;
892 char buf
[64], *vals
, *ends
;
895 if (!valid_only_metric(unit
))
897 unit
= fixunit(tbuf
, os
->evsel
, unit
);
898 snprintf(buf
, sizeof buf
, fmt
, val
);
900 while (isspace(*vals
))
903 while (isdigit(*ends
) || *ends
== '.')
906 fprintf(out
, "%s%s", vals
, csv_sep
);
909 static void new_line_metric(void *ctx __maybe_unused
)
913 static void print_metric_header(void *ctx
, const char *color __maybe_unused
,
914 const char *fmt __maybe_unused
,
915 const char *unit
, double val __maybe_unused
)
917 struct outstate
*os
= ctx
;
920 if (!valid_only_metric(unit
))
922 unit
= fixunit(tbuf
, os
->evsel
, unit
);
924 fprintf(os
->fh
, "%s%s", unit
, csv_sep
);
926 fprintf(os
->fh
, "%-*s ", METRIC_ONLY_LEN
, unit
);
929 static void nsec_printout(int id
, int nr
, struct perf_evsel
*evsel
, double avg
)
931 FILE *output
= stat_config
.output
;
932 double msecs
= avg
/ 1e6
;
933 const char *fmt_v
, *fmt_n
;
936 fmt_v
= csv_output
? "%.6f%s" : "%18.6f%s";
937 fmt_n
= csv_output
? "%s" : "%-25s";
939 aggr_printout(evsel
, id
, nr
);
941 scnprintf(name
, sizeof(name
), "%s%s",
942 perf_evsel__name(evsel
), csv_output
? "" : " (msec)");
944 fprintf(output
, fmt_v
, msecs
, csv_sep
);
947 fprintf(output
, "%s%s", evsel
->unit
, csv_sep
);
949 fprintf(output
, "%-*s%s", unit_width
, evsel
->unit
, csv_sep
);
951 fprintf(output
, fmt_n
, name
);
954 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
957 static int first_shadow_cpu(struct perf_evsel
*evsel
, int id
)
964 if (stat_config
.aggr_mode
== AGGR_NONE
)
967 if (stat_config
.aggr_mode
== AGGR_GLOBAL
)
970 for (i
= 0; i
< perf_evsel__nr_cpus(evsel
); i
++) {
971 int cpu2
= perf_evsel__cpus(evsel
)->map
[i
];
973 if (aggr_get_id(evsel_list
->cpus
, cpu2
) == id
)
979 static void abs_printout(int id
, int nr
, struct perf_evsel
*evsel
, double avg
)
981 FILE *output
= stat_config
.output
;
982 double sc
= evsel
->scale
;
986 fmt
= sc
!= 1.0 ? "%.2f%s" : "%.0f%s";
989 fmt
= sc
!= 1.0 ? "%'18.2f%s" : "%'18.0f%s";
991 fmt
= sc
!= 1.0 ? "%18.2f%s" : "%18.0f%s";
994 aggr_printout(evsel
, id
, nr
);
996 fprintf(output
, fmt
, avg
, csv_sep
);
999 fprintf(output
, "%-*s%s",
1000 csv_output
? 0 : unit_width
,
1001 evsel
->unit
, csv_sep
);
1003 fprintf(output
, "%-*s", csv_output
? 0 : 25, perf_evsel__name(evsel
));
1006 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
1009 static void printout(int id
, int nr
, struct perf_evsel
*counter
, double uval
,
1010 char *prefix
, u64 run
, u64 ena
, double noise
)
1012 struct perf_stat_output_ctx out
;
1013 struct outstate os
= {
1014 .fh
= stat_config
.output
,
1015 .prefix
= prefix
? prefix
: "",
1020 print_metric_t pm
= print_metric_std
;
1024 nl
= new_line_metric
;
1026 pm
= print_metric_only_csv
;
1028 pm
= print_metric_only
;
1032 if (csv_output
&& !metric_only
) {
1033 static int aggr_fields
[] = {
1041 pm
= print_metric_csv
;
1044 os
.nfields
+= aggr_fields
[stat_config
.aggr_mode
];
1048 if (run
== 0 || ena
== 0 || counter
->counts
->scaled
== -1) {
1050 pm(&os
, NULL
, "", "", 0);
1053 aggr_printout(counter
, id
, nr
);
1055 fprintf(stat_config
.output
, "%*s%s",
1056 csv_output
? 0 : 18,
1057 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1060 fprintf(stat_config
.output
, "%-*s%s",
1061 csv_output
? 0 : unit_width
,
1062 counter
->unit
, csv_sep
);
1064 fprintf(stat_config
.output
, "%*s",
1065 csv_output
? 0 : -25,
1066 perf_evsel__name(counter
));
1069 fprintf(stat_config
.output
, "%s%s",
1070 csv_sep
, counter
->cgrp
->name
);
1073 pm(&os
, NULL
, NULL
, "", 0);
1074 print_noise(counter
, noise
);
1075 print_running(run
, ena
);
1077 pm(&os
, NULL
, NULL
, "", 0);
1083 else if (nsec_counter(counter
))
1084 nsec_printout(id
, nr
, counter
, uval
);
1086 abs_printout(id
, nr
, counter
, uval
);
1088 out
.print_metric
= pm
;
1092 if (csv_output
&& !metric_only
) {
1093 print_noise(counter
, noise
);
1094 print_running(run
, ena
);
1097 perf_stat__print_shadow_stats(counter
, uval
,
1098 first_shadow_cpu(counter
, id
),
1100 if (!csv_output
&& !metric_only
) {
1101 print_noise(counter
, noise
);
1102 print_running(run
, ena
);
1106 static void aggr_update_shadow(void)
1110 struct perf_evsel
*counter
;
1112 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1113 id
= aggr_map
->map
[s
];
1114 evlist__for_each(evsel_list
, counter
) {
1116 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1117 s2
= aggr_get_id(evsel_list
->cpus
, cpu
);
1120 val
+= perf_counts(counter
->counts
, cpu
, 0)->val
;
1122 val
= val
* counter
->scale
;
1123 perf_stat__update_shadow_stats(counter
, &val
,
1124 first_shadow_cpu(counter
, id
));
1129 static void print_aggr(char *prefix
)
1131 FILE *output
= stat_config
.output
;
1132 struct perf_evsel
*counter
;
1133 int cpu
, s
, s2
, id
, nr
;
1138 if (!(aggr_map
|| aggr_get_id
))
1141 aggr_update_shadow();
1144 * With metric_only everything is on a single line.
1145 * Without each counter has its own line.
1147 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1148 if (prefix
&& metric_only
)
1149 fprintf(output
, "%s", prefix
);
1151 id
= aggr_map
->map
[s
];
1153 evlist__for_each(evsel_list
, counter
) {
1154 val
= ena
= run
= 0;
1156 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1157 s2
= aggr_get_id(perf_evsel__cpus(counter
), cpu
);
1160 val
+= perf_counts(counter
->counts
, cpu
, 0)->val
;
1161 ena
+= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1162 run
+= perf_counts(counter
->counts
, cpu
, 0)->run
;
1165 if (first
&& metric_only
) {
1167 aggr_printout(counter
, id
, nr
);
1169 if (prefix
&& !metric_only
)
1170 fprintf(output
, "%s", prefix
);
1172 uval
= val
* counter
->scale
;
1173 printout(id
, nr
, counter
, uval
, prefix
, run
, ena
, 1.0);
1175 fputc('\n', output
);
1178 fputc('\n', output
);
1182 static void print_aggr_thread(struct perf_evsel
*counter
, char *prefix
)
1184 FILE *output
= stat_config
.output
;
1185 int nthreads
= thread_map__nr(counter
->threads
);
1186 int ncpus
= cpu_map__nr(counter
->cpus
);
1190 for (thread
= 0; thread
< nthreads
; thread
++) {
1191 u64 ena
= 0, run
= 0, val
= 0;
1193 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
1194 val
+= perf_counts(counter
->counts
, cpu
, thread
)->val
;
1195 ena
+= perf_counts(counter
->counts
, cpu
, thread
)->ena
;
1196 run
+= perf_counts(counter
->counts
, cpu
, thread
)->run
;
1200 fprintf(output
, "%s", prefix
);
1202 uval
= val
* counter
->scale
;
1203 printout(thread
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1204 fputc('\n', output
);
1209 * Print out the results of a single counter:
1210 * aggregated counts in system-wide mode
1212 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
1214 FILE *output
= stat_config
.output
;
1215 struct perf_stat_evsel
*ps
= counter
->priv
;
1216 double avg
= avg_stats(&ps
->res_stats
[0]);
1218 double avg_enabled
, avg_running
;
1220 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
1221 avg_running
= avg_stats(&ps
->res_stats
[2]);
1223 if (prefix
&& !metric_only
)
1224 fprintf(output
, "%s", prefix
);
1226 uval
= avg
* counter
->scale
;
1227 printout(-1, 0, counter
, uval
, prefix
, avg_running
, avg_enabled
, avg
);
1229 fprintf(output
, "\n");
1233 * Print out the results of a single counter:
1234 * does not use aggregated count in system-wide
1236 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1238 FILE *output
= stat_config
.output
;
1243 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1244 val
= perf_counts(counter
->counts
, cpu
, 0)->val
;
1245 ena
= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1246 run
= perf_counts(counter
->counts
, cpu
, 0)->run
;
1249 fprintf(output
, "%s", prefix
);
1251 uval
= val
* counter
->scale
;
1252 printout(cpu
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1254 fputc('\n', output
);
1258 static void print_no_aggr_metric(char *prefix
)
1262 struct perf_evsel
*counter
;
1266 nrcpus
= evsel_list
->cpus
->nr
;
1267 for (cpu
= 0; cpu
< nrcpus
; cpu
++) {
1271 fputs(prefix
, stat_config
.output
);
1272 evlist__for_each(evsel_list
, counter
) {
1274 aggr_printout(counter
, cpu
, 0);
1277 val
= perf_counts(counter
->counts
, cpu
, 0)->val
;
1278 ena
= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1279 run
= perf_counts(counter
->counts
, cpu
, 0)->run
;
1281 uval
= val
* counter
->scale
;
1282 printout(cpu
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1284 fputc('\n', stat_config
.output
);
1288 static int aggr_header_lens
[] = {
1296 static void print_metric_headers(char *prefix
)
1298 struct perf_stat_output_ctx out
;
1299 struct perf_evsel
*counter
;
1300 struct outstate os
= {
1301 .fh
= stat_config
.output
1305 fprintf(stat_config
.output
, "%s", prefix
);
1308 fprintf(stat_config
.output
, "%*s",
1309 aggr_header_lens
[stat_config
.aggr_mode
], "");
1311 /* Print metrics headers only */
1312 evlist__for_each(evsel_list
, counter
) {
1315 out
.print_metric
= print_metric_header
;
1316 out
.new_line
= new_line_metric
;
1318 perf_stat__print_shadow_stats(counter
, 0,
1322 fputc('\n', stat_config
.output
);
1325 static void print_interval(char *prefix
, struct timespec
*ts
)
1327 FILE *output
= stat_config
.output
;
1328 static int num_print_interval
;
1330 sprintf(prefix
, "%6lu.%09lu%s", ts
->tv_sec
, ts
->tv_nsec
, csv_sep
);
1332 if (num_print_interval
== 0 && !csv_output
&& !metric_only
) {
1333 switch (stat_config
.aggr_mode
) {
1335 fprintf(output
, "# time socket cpus counts %*s events\n", unit_width
, "unit");
1338 fprintf(output
, "# time core cpus counts %*s events\n", unit_width
, "unit");
1341 fprintf(output
, "# time CPU counts %*s events\n", unit_width
, "unit");
1344 fprintf(output
, "# time comm-pid counts %*s events\n", unit_width
, "unit");
1348 fprintf(output
, "# time counts %*s events\n", unit_width
, "unit");
1354 if (++num_print_interval
== 25)
1355 num_print_interval
= 0;
1358 static void print_header(int argc
, const char **argv
)
1360 FILE *output
= stat_config
.output
;
1366 fprintf(output
, "\n");
1367 fprintf(output
, " Performance counter stats for ");
1368 if (target
.system_wide
)
1369 fprintf(output
, "\'system wide");
1370 else if (target
.cpu_list
)
1371 fprintf(output
, "\'CPU(s) %s", target
.cpu_list
);
1372 else if (!target__has_task(&target
)) {
1373 fprintf(output
, "\'%s", argv
? argv
[0] : "pipe");
1374 for (i
= 1; argv
&& (i
< argc
); i
++)
1375 fprintf(output
, " %s", argv
[i
]);
1376 } else if (target
.pid
)
1377 fprintf(output
, "process id \'%s", target
.pid
);
1379 fprintf(output
, "thread id \'%s", target
.tid
);
1381 fprintf(output
, "\'");
1383 fprintf(output
, " (%d runs)", run_count
);
1384 fprintf(output
, ":\n\n");
1388 static void print_footer(void)
1390 FILE *output
= stat_config
.output
;
1393 fprintf(output
, "\n");
1394 fprintf(output
, " %17.9f seconds time elapsed",
1395 avg_stats(&walltime_nsecs_stats
)/1e9
);
1396 if (run_count
> 1) {
1397 fprintf(output
, " ");
1398 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1399 avg_stats(&walltime_nsecs_stats
));
1401 fprintf(output
, "\n\n");
1404 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
)
1406 int interval
= stat_config
.interval
;
1407 struct perf_evsel
*counter
;
1408 char buf
[64], *prefix
= NULL
;
1410 /* Do not print anything if we record to the pipe. */
1411 if (STAT_RECORD
&& perf_stat
.file
.is_pipe
)
1415 print_interval(prefix
= buf
, ts
);
1417 print_header(argc
, argv
);
1420 static int num_print_iv
;
1422 if (num_print_iv
== 0)
1423 print_metric_headers(prefix
);
1424 if (num_print_iv
++ == 25)
1426 if (stat_config
.aggr_mode
== AGGR_GLOBAL
&& prefix
)
1427 fprintf(stat_config
.output
, "%s", prefix
);
1430 switch (stat_config
.aggr_mode
) {
1436 evlist__for_each(evsel_list
, counter
)
1437 print_aggr_thread(counter
, prefix
);
1440 evlist__for_each(evsel_list
, counter
)
1441 print_counter_aggr(counter
, prefix
);
1443 fputc('\n', stat_config
.output
);
1447 print_no_aggr_metric(prefix
);
1449 evlist__for_each(evsel_list
, counter
)
1450 print_counter(counter
, prefix
);
1458 if (!interval
&& !csv_output
)
1461 fflush(stat_config
.output
);
1464 static volatile int signr
= -1;
1466 static void skip_signal(int signo
)
1468 if ((child_pid
== -1) || stat_config
.interval
)
1473 * render child_pid harmless
1474 * won't send SIGTERM to a random
1475 * process in case of race condition
1476 * and fast PID recycling
1481 static void sig_atexit(void)
1486 * avoid race condition with SIGCHLD handler
1487 * in skip_signal() which is modifying child_pid
1488 * goal is to avoid send SIGTERM to a random
1492 sigaddset(&set
, SIGCHLD
);
1493 sigprocmask(SIG_BLOCK
, &set
, &oset
);
1495 if (child_pid
!= -1)
1496 kill(child_pid
, SIGTERM
);
1498 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
1503 signal(signr
, SIG_DFL
);
1504 kill(getpid(), signr
);
1507 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1508 const char *s __maybe_unused
, int unset
)
1510 big_num_opt
= unset
? 0 : 1;
1514 static const struct option stat_options
[] = {
1515 OPT_BOOLEAN('T', "transaction", &transaction_run
,
1516 "hardware transaction statistics"),
1517 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1518 "event selector. use 'perf list' to list available events",
1519 parse_events_option
),
1520 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1521 "event filter", parse_filter
),
1522 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1523 "child tasks do not inherit counters"),
1524 OPT_STRING('p', "pid", &target
.pid
, "pid",
1525 "stat events on existing process id"),
1526 OPT_STRING('t', "tid", &target
.tid
, "tid",
1527 "stat events on existing thread id"),
1528 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1529 "system-wide collection from all CPUs"),
1530 OPT_BOOLEAN('g', "group", &group
,
1531 "put the counters into a counter group"),
1532 OPT_BOOLEAN('c', "scale", &stat_config
.scale
, "scale/normalize counters"),
1533 OPT_INCR('v', "verbose", &verbose
,
1534 "be more verbose (show counter open errors, etc)"),
1535 OPT_INTEGER('r', "repeat", &run_count
,
1536 "repeat command and print average + stddev (max: 100, forever: 0)"),
1537 OPT_BOOLEAN('n', "null", &null_run
,
1538 "null run - dont start any counters"),
1539 OPT_INCR('d', "detailed", &detailed_run
,
1540 "detailed run - start a lot of events"),
1541 OPT_BOOLEAN('S', "sync", &sync_run
,
1542 "call sync() before starting a run"),
1543 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1544 "print large numbers with thousands\' separators",
1546 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1547 "list of cpus to monitor in system-wide"),
1548 OPT_SET_UINT('A', "no-aggr", &stat_config
.aggr_mode
,
1549 "disable CPU count aggregation", AGGR_NONE
),
1550 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1551 "print counts with custom separator"),
1552 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1553 "monitor event in cgroup name only", parse_cgroups
),
1554 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1555 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1556 OPT_INTEGER(0, "log-fd", &output_fd
,
1557 "log output to fd, instead of stderr"),
1558 OPT_STRING(0, "pre", &pre_cmd
, "command",
1559 "command to run prior to the measured command"),
1560 OPT_STRING(0, "post", &post_cmd
, "command",
1561 "command to run after to the measured command"),
1562 OPT_UINTEGER('I', "interval-print", &stat_config
.interval
,
1563 "print counts at regular interval in ms (>= 10)"),
1564 OPT_SET_UINT(0, "per-socket", &stat_config
.aggr_mode
,
1565 "aggregate counts per processor socket", AGGR_SOCKET
),
1566 OPT_SET_UINT(0, "per-core", &stat_config
.aggr_mode
,
1567 "aggregate counts per physical processor core", AGGR_CORE
),
1568 OPT_SET_UINT(0, "per-thread", &stat_config
.aggr_mode
,
1569 "aggregate counts per thread", AGGR_THREAD
),
1570 OPT_UINTEGER('D', "delay", &initial_delay
,
1571 "ms to wait before starting measurement after program start"),
1572 OPT_BOOLEAN(0, "metric-only", &metric_only
,
1573 "Only print computed metrics. No raw values"),
1577 static int perf_stat__get_socket(struct cpu_map
*map
, int cpu
)
1579 return cpu_map__get_socket(map
, cpu
, NULL
);
1582 static int perf_stat__get_core(struct cpu_map
*map
, int cpu
)
1584 return cpu_map__get_core(map
, cpu
, NULL
);
1587 static int cpu_map__get_max(struct cpu_map
*map
)
1591 for (i
= 0; i
< map
->nr
; i
++) {
1592 if (map
->map
[i
] > max
)
1599 static struct cpu_map
*cpus_aggr_map
;
1601 static int perf_stat__get_aggr(aggr_get_id_t get_id
, struct cpu_map
*map
, int idx
)
1608 cpu
= map
->map
[idx
];
1610 if (cpus_aggr_map
->map
[cpu
] == -1)
1611 cpus_aggr_map
->map
[cpu
] = get_id(map
, idx
);
1613 return cpus_aggr_map
->map
[cpu
];
1616 static int perf_stat__get_socket_cached(struct cpu_map
*map
, int idx
)
1618 return perf_stat__get_aggr(perf_stat__get_socket
, map
, idx
);
1621 static int perf_stat__get_core_cached(struct cpu_map
*map
, int idx
)
1623 return perf_stat__get_aggr(perf_stat__get_core
, map
, idx
);
1626 static int perf_stat_init_aggr_mode(void)
1630 switch (stat_config
.aggr_mode
) {
1632 if (cpu_map__build_socket_map(evsel_list
->cpus
, &aggr_map
)) {
1633 perror("cannot build socket map");
1636 aggr_get_id
= perf_stat__get_socket_cached
;
1639 if (cpu_map__build_core_map(evsel_list
->cpus
, &aggr_map
)) {
1640 perror("cannot build core map");
1643 aggr_get_id
= perf_stat__get_core_cached
;
1654 * The evsel_list->cpus is the base we operate on,
1655 * taking the highest cpu number to be the size of
1656 * the aggregation translate cpumap.
1658 nr
= cpu_map__get_max(evsel_list
->cpus
);
1659 cpus_aggr_map
= cpu_map__empty_new(nr
+ 1);
1660 return cpus_aggr_map
? 0 : -ENOMEM
;
1663 static void perf_stat__exit_aggr_mode(void)
1665 cpu_map__put(aggr_map
);
1666 cpu_map__put(cpus_aggr_map
);
1668 cpus_aggr_map
= NULL
;
1671 static inline int perf_env__get_cpu(struct perf_env
*env
, struct cpu_map
*map
, int idx
)
1678 cpu
= map
->map
[idx
];
1680 if (cpu
>= env
->nr_cpus_online
)
1686 static int perf_env__get_socket(struct cpu_map
*map
, int idx
, void *data
)
1688 struct perf_env
*env
= data
;
1689 int cpu
= perf_env__get_cpu(env
, map
, idx
);
1691 return cpu
== -1 ? -1 : env
->cpu
[cpu
].socket_id
;
1694 static int perf_env__get_core(struct cpu_map
*map
, int idx
, void *data
)
1696 struct perf_env
*env
= data
;
1697 int core
= -1, cpu
= perf_env__get_cpu(env
, map
, idx
);
1700 int socket_id
= env
->cpu
[cpu
].socket_id
;
1703 * Encode socket in upper 16 bits
1704 * core_id is relative to socket, and
1705 * we need a global id. So we combine
1708 core
= (socket_id
<< 16) | (env
->cpu
[cpu
].core_id
& 0xffff);
1714 static int perf_env__build_socket_map(struct perf_env
*env
, struct cpu_map
*cpus
,
1715 struct cpu_map
**sockp
)
1717 return cpu_map__build_map(cpus
, sockp
, perf_env__get_socket
, env
);
1720 static int perf_env__build_core_map(struct perf_env
*env
, struct cpu_map
*cpus
,
1721 struct cpu_map
**corep
)
1723 return cpu_map__build_map(cpus
, corep
, perf_env__get_core
, env
);
1726 static int perf_stat__get_socket_file(struct cpu_map
*map
, int idx
)
1728 return perf_env__get_socket(map
, idx
, &perf_stat
.session
->header
.env
);
1731 static int perf_stat__get_core_file(struct cpu_map
*map
, int idx
)
1733 return perf_env__get_core(map
, idx
, &perf_stat
.session
->header
.env
);
1736 static int perf_stat_init_aggr_mode_file(struct perf_stat
*st
)
1738 struct perf_env
*env
= &st
->session
->header
.env
;
1740 switch (stat_config
.aggr_mode
) {
1742 if (perf_env__build_socket_map(env
, evsel_list
->cpus
, &aggr_map
)) {
1743 perror("cannot build socket map");
1746 aggr_get_id
= perf_stat__get_socket_file
;
1749 if (perf_env__build_core_map(env
, evsel_list
->cpus
, &aggr_map
)) {
1750 perror("cannot build core map");
1753 aggr_get_id
= perf_stat__get_core_file
;
1767 * Add default attributes, if there were no attributes specified or
1768 * if -d/--detailed, -d -d or -d -d -d is used:
1770 static int add_default_attributes(void)
1772 struct perf_event_attr default_attrs0
[] = {
1774 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1775 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1776 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1777 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1779 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1781 struct perf_event_attr frontend_attrs
[] = {
1782 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1784 struct perf_event_attr backend_attrs
[] = {
1785 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1787 struct perf_event_attr default_attrs1
[] = {
1788 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1789 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1790 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1795 * Detailed stats (-d), covering the L1 and last level data caches:
1797 struct perf_event_attr detailed_attrs
[] = {
1799 { .type
= PERF_TYPE_HW_CACHE
,
1801 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1802 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1803 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1805 { .type
= PERF_TYPE_HW_CACHE
,
1807 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1808 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1809 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1811 { .type
= PERF_TYPE_HW_CACHE
,
1813 PERF_COUNT_HW_CACHE_LL
<< 0 |
1814 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1815 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1817 { .type
= PERF_TYPE_HW_CACHE
,
1819 PERF_COUNT_HW_CACHE_LL
<< 0 |
1820 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1821 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1825 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1827 struct perf_event_attr very_detailed_attrs
[] = {
1829 { .type
= PERF_TYPE_HW_CACHE
,
1831 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1832 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1833 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1835 { .type
= PERF_TYPE_HW_CACHE
,
1837 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1838 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1839 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1841 { .type
= PERF_TYPE_HW_CACHE
,
1843 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1844 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1845 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1847 { .type
= PERF_TYPE_HW_CACHE
,
1849 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1850 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1851 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1853 { .type
= PERF_TYPE_HW_CACHE
,
1855 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1856 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1857 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1859 { .type
= PERF_TYPE_HW_CACHE
,
1861 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1862 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1863 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1868 * Very, very detailed stats (-d -d -d), adding prefetch events:
1870 struct perf_event_attr very_very_detailed_attrs
[] = {
1872 { .type
= PERF_TYPE_HW_CACHE
,
1874 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1875 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1876 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1878 { .type
= PERF_TYPE_HW_CACHE
,
1880 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1881 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1882 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1885 /* Set attrs if no event is selected and !null_run: */
1889 if (transaction_run
) {
1891 if (pmu_have_event("cpu", "cycles-ct") &&
1892 pmu_have_event("cpu", "el-start"))
1893 err
= parse_events(evsel_list
, transaction_attrs
, NULL
);
1895 err
= parse_events(evsel_list
, transaction_limited_attrs
, NULL
);
1897 fprintf(stderr
, "Cannot set up transaction events\n");
1903 if (!evsel_list
->nr_entries
) {
1904 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs0
) < 0)
1906 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1907 if (perf_evlist__add_default_attrs(evsel_list
,
1908 frontend_attrs
) < 0)
1911 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1912 if (perf_evlist__add_default_attrs(evsel_list
,
1916 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs1
) < 0)
1920 /* Detailed events get appended to the event list: */
1922 if (detailed_run
< 1)
1925 /* Append detailed run extra attributes: */
1926 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1929 if (detailed_run
< 2)
1932 /* Append very detailed run extra attributes: */
1933 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1936 if (detailed_run
< 3)
1939 /* Append very, very detailed run extra attributes: */
1940 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1943 static const char * const stat_record_usage
[] = {
1944 "perf stat record [<options>]",
1948 static void init_features(struct perf_session
*session
)
1952 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
1953 perf_header__set_feat(&session
->header
, feat
);
1955 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
1956 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
1957 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
1958 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
1961 static int __cmd_record(int argc
, const char **argv
)
1963 struct perf_session
*session
;
1964 struct perf_data_file
*file
= &perf_stat
.file
;
1966 argc
= parse_options(argc
, argv
, stat_options
, stat_record_usage
,
1967 PARSE_OPT_STOP_AT_NON_OPTION
);
1970 file
->path
= output_name
;
1972 if (run_count
!= 1 || forever
) {
1973 pr_err("Cannot use -r option with perf stat record.\n");
1977 session
= perf_session__new(file
, false, NULL
);
1978 if (session
== NULL
) {
1979 pr_err("Perf session creation failed.\n");
1983 init_features(session
);
1985 session
->evlist
= evsel_list
;
1986 perf_stat
.session
= session
;
1987 perf_stat
.record
= true;
1991 static int process_stat_round_event(struct perf_tool
*tool __maybe_unused
,
1992 union perf_event
*event
,
1993 struct perf_session
*session
)
1995 struct stat_round_event
*round
= &event
->stat_round
;
1996 struct perf_evsel
*counter
;
1997 struct timespec tsh
, *ts
= NULL
;
1998 const char **argv
= session
->header
.env
.cmdline_argv
;
1999 int argc
= session
->header
.env
.nr_cmdline
;
2001 evlist__for_each(evsel_list
, counter
)
2002 perf_stat_process_counter(&stat_config
, counter
);
2004 if (round
->type
== PERF_STAT_ROUND_TYPE__FINAL
)
2005 update_stats(&walltime_nsecs_stats
, round
->time
);
2007 if (stat_config
.interval
&& round
->time
) {
2008 tsh
.tv_sec
= round
->time
/ NSECS_PER_SEC
;
2009 tsh
.tv_nsec
= round
->time
% NSECS_PER_SEC
;
2013 print_counters(ts
, argc
, argv
);
2018 int process_stat_config_event(struct perf_tool
*tool __maybe_unused
,
2019 union perf_event
*event
,
2020 struct perf_session
*session __maybe_unused
)
2022 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2024 perf_event__read_stat_config(&stat_config
, &event
->stat_config
);
2026 if (cpu_map__empty(st
->cpus
)) {
2027 if (st
->aggr_mode
!= AGGR_UNSET
)
2028 pr_warning("warning: processing task data, aggregation mode not set\n");
2032 if (st
->aggr_mode
!= AGGR_UNSET
)
2033 stat_config
.aggr_mode
= st
->aggr_mode
;
2035 if (perf_stat
.file
.is_pipe
)
2036 perf_stat_init_aggr_mode();
2038 perf_stat_init_aggr_mode_file(st
);
2043 static int set_maps(struct perf_stat
*st
)
2045 if (!st
->cpus
|| !st
->threads
)
2048 if (WARN_ONCE(st
->maps_allocated
, "stats double allocation\n"))
2051 perf_evlist__set_maps(evsel_list
, st
->cpus
, st
->threads
);
2053 if (perf_evlist__alloc_stats(evsel_list
, true))
2056 st
->maps_allocated
= true;
2061 int process_thread_map_event(struct perf_tool
*tool __maybe_unused
,
2062 union perf_event
*event
,
2063 struct perf_session
*session __maybe_unused
)
2065 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2068 pr_warning("Extra thread map event, ignoring.\n");
2072 st
->threads
= thread_map__new_event(&event
->thread_map
);
2076 return set_maps(st
);
2080 int process_cpu_map_event(struct perf_tool
*tool __maybe_unused
,
2081 union perf_event
*event
,
2082 struct perf_session
*session __maybe_unused
)
2084 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2085 struct cpu_map
*cpus
;
2088 pr_warning("Extra cpu map event, ignoring.\n");
2092 cpus
= cpu_map__new_data(&event
->cpu_map
.data
);
2097 return set_maps(st
);
2100 static const char * const stat_report_usage
[] = {
2101 "perf stat report [<options>]",
2105 static struct perf_stat perf_stat
= {
2107 .attr
= perf_event__process_attr
,
2108 .event_update
= perf_event__process_event_update
,
2109 .thread_map
= process_thread_map_event
,
2110 .cpu_map
= process_cpu_map_event
,
2111 .stat_config
= process_stat_config_event
,
2112 .stat
= perf_event__process_stat_event
,
2113 .stat_round
= process_stat_round_event
,
2115 .aggr_mode
= AGGR_UNSET
,
2118 static int __cmd_report(int argc
, const char **argv
)
2120 struct perf_session
*session
;
2121 const struct option options
[] = {
2122 OPT_STRING('i', "input", &input_name
, "file", "input file name"),
2123 OPT_SET_UINT(0, "per-socket", &perf_stat
.aggr_mode
,
2124 "aggregate counts per processor socket", AGGR_SOCKET
),
2125 OPT_SET_UINT(0, "per-core", &perf_stat
.aggr_mode
,
2126 "aggregate counts per physical processor core", AGGR_CORE
),
2127 OPT_SET_UINT('A', "no-aggr", &perf_stat
.aggr_mode
,
2128 "disable CPU count aggregation", AGGR_NONE
),
2134 argc
= parse_options(argc
, argv
, options
, stat_report_usage
, 0);
2136 if (!input_name
|| !strlen(input_name
)) {
2137 if (!fstat(STDIN_FILENO
, &st
) && S_ISFIFO(st
.st_mode
))
2140 input_name
= "perf.data";
2143 perf_stat
.file
.path
= input_name
;
2144 perf_stat
.file
.mode
= PERF_DATA_MODE_READ
;
2146 session
= perf_session__new(&perf_stat
.file
, false, &perf_stat
.tool
);
2147 if (session
== NULL
)
2150 perf_stat
.session
= session
;
2151 stat_config
.output
= stderr
;
2152 evsel_list
= session
->evlist
;
2154 ret
= perf_session__process_events(session
);
2158 perf_session__delete(session
);
2162 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
2164 const char * const stat_usage
[] = {
2165 "perf stat [<options>] [<command>]",
2168 int status
= -EINVAL
, run_idx
;
2170 FILE *output
= stderr
;
2171 unsigned int interval
;
2172 const char * const stat_subcommands
[] = { "record", "report" };
2174 setlocale(LC_ALL
, "");
2176 evsel_list
= perf_evlist__new();
2177 if (evsel_list
== NULL
)
2180 parse_events__shrink_config_terms();
2181 argc
= parse_options_subcommand(argc
, argv
, stat_options
, stat_subcommands
,
2182 (const char **) stat_usage
,
2183 PARSE_OPT_STOP_AT_NON_OPTION
);
2184 perf_stat__init_shadow_stats();
2188 if (!strcmp(csv_sep
, "\\t"))
2191 csv_sep
= DEFAULT_SEPARATOR
;
2193 if (argc
&& !strncmp(argv
[0], "rec", 3)) {
2194 argc
= __cmd_record(argc
, argv
);
2197 } else if (argc
&& !strncmp(argv
[0], "rep", 3))
2198 return __cmd_report(argc
, argv
);
2200 interval
= stat_config
.interval
;
2203 * For record command the -o is already taken care of.
2205 if (!STAT_RECORD
&& output_name
&& strcmp(output_name
, "-"))
2208 if (output_name
&& output_fd
) {
2209 fprintf(stderr
, "cannot use both --output and --log-fd\n");
2210 parse_options_usage(stat_usage
, stat_options
, "o", 1);
2211 parse_options_usage(NULL
, stat_options
, "log-fd", 0);
2215 if (metric_only
&& stat_config
.aggr_mode
== AGGR_THREAD
) {
2216 fprintf(stderr
, "--metric-only is not supported with --per-thread\n");
2220 if (metric_only
&& run_count
> 1) {
2221 fprintf(stderr
, "--metric-only is not supported with -r\n");
2225 if (output_fd
< 0) {
2226 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
2227 parse_options_usage(stat_usage
, stat_options
, "log-fd", 0);
2233 mode
= append_file
? "a" : "w";
2235 output
= fopen(output_name
, mode
);
2237 perror("failed to create output file");
2240 clock_gettime(CLOCK_REALTIME
, &tm
);
2241 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
2242 } else if (output_fd
> 0) {
2243 mode
= append_file
? "a" : "w";
2244 output
= fdopen(output_fd
, mode
);
2246 perror("Failed opening logfd");
2251 stat_config
.output
= output
;
2254 * let the spreadsheet do the pretty-printing
2257 /* User explicitly passed -B? */
2258 if (big_num_opt
== 1) {
2259 fprintf(stderr
, "-B option not supported with -x\n");
2260 parse_options_usage(stat_usage
, stat_options
, "B", 1);
2261 parse_options_usage(NULL
, stat_options
, "x", 1);
2263 } else /* Nope, so disable big number formatting */
2265 } else if (big_num_opt
== 0) /* User passed --no-big-num */
2268 if (!argc
&& target__none(&target
))
2269 usage_with_options(stat_usage
, stat_options
);
2271 if (run_count
< 0) {
2272 pr_err("Run count must be a positive number\n");
2273 parse_options_usage(stat_usage
, stat_options
, "r", 1);
2275 } else if (run_count
== 0) {
2280 if ((stat_config
.aggr_mode
== AGGR_THREAD
) && !target__has_task(&target
)) {
2281 fprintf(stderr
, "The --per-thread option is only available "
2282 "when monitoring via -p -t options.\n");
2283 parse_options_usage(NULL
, stat_options
, "p", 1);
2284 parse_options_usage(NULL
, stat_options
, "t", 1);
2289 * no_aggr, cgroup are for system-wide only
2290 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2292 if (((stat_config
.aggr_mode
!= AGGR_GLOBAL
&&
2293 stat_config
.aggr_mode
!= AGGR_THREAD
) || nr_cgroups
) &&
2294 !target__has_cpu(&target
)) {
2295 fprintf(stderr
, "both cgroup and no-aggregation "
2296 "modes only available in system-wide mode\n");
2298 parse_options_usage(stat_usage
, stat_options
, "G", 1);
2299 parse_options_usage(NULL
, stat_options
, "A", 1);
2300 parse_options_usage(NULL
, stat_options
, "a", 1);
2304 if (add_default_attributes())
2307 target__validate(&target
);
2309 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
2310 if (target__has_task(&target
)) {
2311 pr_err("Problems finding threads of monitor\n");
2312 parse_options_usage(stat_usage
, stat_options
, "p", 1);
2313 parse_options_usage(NULL
, stat_options
, "t", 1);
2314 } else if (target__has_cpu(&target
)) {
2315 perror("failed to parse CPUs map");
2316 parse_options_usage(stat_usage
, stat_options
, "C", 1);
2317 parse_options_usage(NULL
, stat_options
, "a", 1);
2323 * Initialize thread_map with comm names,
2324 * so we could print it out on output.
2326 if (stat_config
.aggr_mode
== AGGR_THREAD
)
2327 thread_map__read_comms(evsel_list
->threads
);
2329 if (interval
&& interval
< 100) {
2330 if (interval
< 10) {
2331 pr_err("print interval must be >= 10ms\n");
2332 parse_options_usage(stat_usage
, stat_options
, "I", 1);
2335 pr_warning("print interval < 100ms. "
2336 "The overhead percentage could be high in some cases. "
2337 "Please proceed with caution.\n");
2340 if (perf_evlist__alloc_stats(evsel_list
, interval
))
2343 if (perf_stat_init_aggr_mode())
2347 * We dont want to block the signals - that would cause
2348 * child tasks to inherit that and Ctrl-C would not work.
2349 * What we want is for Ctrl-C to work in the exec()-ed
2350 * task, but being ignored by perf stat itself:
2354 signal(SIGINT
, skip_signal
);
2355 signal(SIGCHLD
, skip_signal
);
2356 signal(SIGALRM
, skip_signal
);
2357 signal(SIGABRT
, skip_signal
);
2360 for (run_idx
= 0; forever
|| run_idx
< run_count
; run_idx
++) {
2361 if (run_count
!= 1 && verbose
)
2362 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
2365 status
= run_perf_stat(argc
, argv
);
2366 if (forever
&& status
!= -1) {
2367 print_counters(NULL
, argc
, argv
);
2368 perf_stat__reset_stats();
2372 if (!forever
&& status
!= -1 && !interval
)
2373 print_counters(NULL
, argc
, argv
);
2377 * We synthesize the kernel mmap record just so that older tools
2378 * don't emit warnings about not being able to resolve symbols
2379 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2380 * a saner message about no samples being in the perf.data file.
2382 * This also serves to suppress a warning about f_header.data.size == 0
2383 * in header.c at the moment 'perf stat record' gets introduced, which
2384 * is not really needed once we start adding the stat specific PERF_RECORD_
2385 * records, but the need to suppress the kptr_restrict messages in older
2386 * tools remain -acme
2388 int fd
= perf_data_file__fd(&perf_stat
.file
);
2389 int err
= perf_event__synthesize_kernel_mmap((void *)&perf_stat
,
2390 process_synthesized_event
,
2391 &perf_stat
.session
->machines
.host
);
2393 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2394 "older tools may produce warnings about this file\n.");
2398 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats
.max
, FINAL
))
2399 pr_err("failed to write stat round event\n");
2402 if (!perf_stat
.file
.is_pipe
) {
2403 perf_stat
.session
->header
.data_size
+= perf_stat
.bytes_written
;
2404 perf_session__write_header(perf_stat
.session
, evsel_list
, fd
, true);
2407 perf_session__delete(perf_stat
.session
);
2410 perf_stat__exit_aggr_mode();
2411 perf_evlist__free_stats(evsel_list
);
2413 perf_evlist__delete(evsel_list
);