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");
303 fprintf(stat_config
.output
,
304 "%s: %d: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
305 perf_evsel__name(counter
),
307 count
->val
, count
->ena
, count
->run
);
315 static void read_counters(bool close_counters
)
317 struct perf_evsel
*counter
;
319 evlist__for_each(evsel_list
, counter
) {
320 if (read_counter(counter
))
321 pr_debug("failed to read counter %s\n", counter
->name
);
323 if (perf_stat_process_counter(&stat_config
, counter
))
324 pr_warning("failed to process counter %s\n", counter
->name
);
326 if (close_counters
) {
327 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
328 thread_map__nr(evsel_list
->threads
));
333 static void process_interval(void)
335 struct timespec ts
, rs
;
337 read_counters(false);
339 clock_gettime(CLOCK_MONOTONIC
, &ts
);
340 diff_timespec(&rs
, &ts
, &ref_time
);
343 if (WRITE_STAT_ROUND_EVENT(rs
.tv_sec
* NSECS_PER_SEC
+ rs
.tv_nsec
, INTERVAL
))
344 pr_err("failed to write stat round event\n");
347 print_counters(&rs
, 0, NULL
);
350 static void enable_counters(void)
353 usleep(initial_delay
* 1000);
356 * We need to enable counters only if:
357 * - we don't have tracee (attaching to task or cpu)
358 * - we have initial delay configured
360 if (!target__none(&target
) || initial_delay
)
361 perf_evlist__enable(evsel_list
);
364 static volatile int workload_exec_errno
;
367 * perf_evlist__prepare_workload will send a SIGUSR1
368 * if the fork fails, since we asked by setting its
369 * want_signal to true.
371 static void workload_exec_failed_signal(int signo __maybe_unused
, siginfo_t
*info
,
372 void *ucontext __maybe_unused
)
374 workload_exec_errno
= info
->si_value
.sival_int
;
377 static bool has_unit(struct perf_evsel
*counter
)
379 return counter
->unit
&& *counter
->unit
;
382 static bool has_scale(struct perf_evsel
*counter
)
384 return counter
->scale
!= 1;
387 static int perf_stat_synthesize_config(bool is_pipe
)
389 struct perf_evsel
*counter
;
393 err
= perf_event__synthesize_attrs(NULL
, perf_stat
.session
,
394 process_synthesized_event
);
396 pr_err("Couldn't synthesize attrs.\n");
402 * Synthesize other events stuff not carried within
403 * attr event - unit, scale, name
405 evlist__for_each(evsel_list
, counter
) {
406 if (!counter
->supported
)
410 * Synthesize unit and scale only if it's defined.
412 if (has_unit(counter
)) {
413 err
= perf_event__synthesize_event_update_unit(NULL
, counter
, process_synthesized_event
);
415 pr_err("Couldn't synthesize evsel unit.\n");
420 if (has_scale(counter
)) {
421 err
= perf_event__synthesize_event_update_scale(NULL
, counter
, process_synthesized_event
);
423 pr_err("Couldn't synthesize evsel scale.\n");
428 if (counter
->own_cpus
) {
429 err
= perf_event__synthesize_event_update_cpus(NULL
, counter
, process_synthesized_event
);
431 pr_err("Couldn't synthesize evsel scale.\n");
437 * Name is needed only for pipe output,
438 * perf.data carries event names.
441 err
= perf_event__synthesize_event_update_name(NULL
, counter
, process_synthesized_event
);
443 pr_err("Couldn't synthesize evsel name.\n");
449 err
= perf_event__synthesize_thread_map2(NULL
, evsel_list
->threads
,
450 process_synthesized_event
,
453 pr_err("Couldn't synthesize thread map.\n");
457 err
= perf_event__synthesize_cpu_map(NULL
, evsel_list
->cpus
,
458 process_synthesized_event
, NULL
);
460 pr_err("Couldn't synthesize thread map.\n");
464 err
= perf_event__synthesize_stat_config(NULL
, &stat_config
,
465 process_synthesized_event
, NULL
);
467 pr_err("Couldn't synthesize config.\n");
474 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
476 static int __store_counter_ids(struct perf_evsel
*counter
,
477 struct cpu_map
*cpus
,
478 struct thread_map
*threads
)
482 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
483 for (thread
= 0; thread
< threads
->nr
; thread
++) {
484 int fd
= FD(counter
, cpu
, thread
);
486 if (perf_evlist__id_add_fd(evsel_list
, counter
,
487 cpu
, thread
, fd
) < 0)
495 static int store_counter_ids(struct perf_evsel
*counter
)
497 struct cpu_map
*cpus
= counter
->cpus
;
498 struct thread_map
*threads
= counter
->threads
;
500 if (perf_evsel__alloc_id(counter
, cpus
->nr
, threads
->nr
))
503 return __store_counter_ids(counter
, cpus
, threads
);
506 static int __run_perf_stat(int argc
, const char **argv
)
508 int interval
= stat_config
.interval
;
510 unsigned long long t0
, t1
;
511 struct perf_evsel
*counter
;
515 const bool forks
= (argc
> 0);
516 bool is_pipe
= STAT_RECORD
? perf_stat
.file
.is_pipe
: false;
519 ts
.tv_sec
= interval
/ 1000;
520 ts
.tv_nsec
= (interval
% 1000) * 1000000;
527 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
, is_pipe
,
528 workload_exec_failed_signal
) < 0) {
529 perror("failed to prepare workload");
532 child_pid
= evsel_list
->workload
.pid
;
536 perf_evlist__set_leader(evsel_list
);
538 evlist__for_each(evsel_list
, counter
) {
540 if (create_perf_stat_counter(counter
) < 0) {
542 * PPC returns ENXIO for HW counters until 2.6.37
543 * (behavior changed with commit b0a873e).
545 if (errno
== EINVAL
|| errno
== ENOSYS
||
546 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
549 ui__warning("%s event is not supported by the kernel.\n",
550 perf_evsel__name(counter
));
551 counter
->supported
= false;
553 if ((counter
->leader
!= counter
) ||
554 !(counter
->leader
->nr_members
> 1))
556 } else if (perf_evsel__fallback(counter
, errno
, msg
, sizeof(msg
))) {
558 ui__warning("%s\n", msg
);
562 perf_evsel__open_strerror(counter
, &target
,
563 errno
, msg
, sizeof(msg
));
564 ui__error("%s\n", msg
);
567 kill(child_pid
, SIGTERM
);
571 counter
->supported
= true;
573 l
= strlen(counter
->unit
);
577 if (STAT_RECORD
&& store_counter_ids(counter
))
581 if (perf_evlist__apply_filters(evsel_list
, &counter
)) {
582 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
583 counter
->filter
, perf_evsel__name(counter
), errno
,
584 strerror_r(errno
, msg
, sizeof(msg
)));
589 int err
, fd
= perf_data_file__fd(&perf_stat
.file
);
592 err
= perf_header__write_pipe(perf_data_file__fd(&perf_stat
.file
));
594 err
= perf_session__write_header(perf_stat
.session
, evsel_list
,
601 err
= perf_stat_synthesize_config(is_pipe
);
607 * Enable counters and exec the command:
610 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
613 perf_evlist__start_workload(evsel_list
);
617 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
618 nanosleep(&ts
, NULL
);
624 if (workload_exec_errno
) {
625 const char *emsg
= strerror_r(workload_exec_errno
, msg
, sizeof(msg
));
626 pr_err("Workload failed: %s\n", emsg
);
630 if (WIFSIGNALED(status
))
631 psignal(WTERMSIG(status
), argv
[0]);
635 nanosleep(&ts
, NULL
);
643 update_stats(&walltime_nsecs_stats
, t1
- t0
);
647 return WEXITSTATUS(status
);
650 static int run_perf_stat(int argc
, const char **argv
)
655 ret
= system(pre_cmd
);
663 ret
= __run_perf_stat(argc
, argv
);
668 ret
= system(post_cmd
);
676 static void print_running(u64 run
, u64 ena
)
679 fprintf(stat_config
.output
, "%s%" PRIu64
"%s%.2f",
683 ena
? 100.0 * run
/ ena
: 100.0);
684 } else if (run
!= ena
) {
685 fprintf(stat_config
.output
, " (%.2f%%)", 100.0 * run
/ ena
);
689 static void print_noise_pct(double total
, double avg
)
691 double pct
= rel_stddev_stats(total
, avg
);
694 fprintf(stat_config
.output
, "%s%.2f%%", csv_sep
, pct
);
696 fprintf(stat_config
.output
, " ( +-%6.2f%% )", pct
);
699 static void print_noise(struct perf_evsel
*evsel
, double avg
)
701 struct perf_stat_evsel
*ps
;
707 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
710 static void aggr_printout(struct perf_evsel
*evsel
, int id
, int nr
)
712 switch (stat_config
.aggr_mode
) {
714 fprintf(stat_config
.output
, "S%d-C%*d%s%*d%s",
715 cpu_map__id_to_socket(id
),
717 cpu_map__id_to_cpu(id
),
724 fprintf(stat_config
.output
, "S%*d%s%*d%s",
733 fprintf(stat_config
.output
, "CPU%*d%s",
735 perf_evsel__cpus(evsel
)->map
[id
], csv_sep
);
738 fprintf(stat_config
.output
, "%*s-%*d%s",
740 thread_map__comm(evsel
->threads
, id
),
742 thread_map__pid(evsel
->threads
, id
),
758 struct perf_evsel
*evsel
;
761 #define METRIC_LEN 35
763 static void new_line_std(void *ctx
)
765 struct outstate
*os
= ctx
;
770 static void do_new_line_std(struct outstate
*os
)
773 fputs(os
->prefix
, os
->fh
);
774 aggr_printout(os
->evsel
, os
->id
, os
->nr
);
775 if (stat_config
.aggr_mode
== AGGR_NONE
)
776 fprintf(os
->fh
, " ");
777 fprintf(os
->fh
, " ");
780 static void print_metric_std(void *ctx
, const char *color
, const char *fmt
,
781 const char *unit
, double val
)
783 struct outstate
*os
= ctx
;
786 bool newline
= os
->newline
;
790 if (unit
== NULL
|| fmt
== NULL
) {
791 fprintf(out
, "%-*s", METRIC_LEN
, "");
798 n
= fprintf(out
, " # ");
800 n
+= color_fprintf(out
, color
, fmt
, val
);
802 n
+= fprintf(out
, fmt
, val
);
803 fprintf(out
, " %-*s", METRIC_LEN
- n
- 1, unit
);
806 static void new_line_csv(void *ctx
)
808 struct outstate
*os
= ctx
;
813 fprintf(os
->fh
, "%s%s", os
->prefix
, csv_sep
);
814 aggr_printout(os
->evsel
, os
->id
, os
->nr
);
815 for (i
= 0; i
< os
->nfields
; i
++)
816 fputs(csv_sep
, os
->fh
);
819 static void print_metric_csv(void *ctx
,
820 const char *color __maybe_unused
,
821 const char *fmt
, const char *unit
, double val
)
823 struct outstate
*os
= ctx
;
825 char buf
[64], *vals
, *ends
;
827 if (unit
== NULL
|| fmt
== NULL
) {
828 fprintf(out
, "%s%s%s%s", csv_sep
, csv_sep
, csv_sep
, csv_sep
);
831 snprintf(buf
, sizeof(buf
), fmt
, val
);
833 while (isspace(*vals
))
836 while (isdigit(*ends
) || *ends
== '.')
839 while (isspace(*unit
))
841 fprintf(out
, "%s%s%s%s", csv_sep
, vals
, csv_sep
, unit
);
844 #define METRIC_ONLY_LEN 20
846 /* Filter out some columns that don't work well in metrics only mode */
848 static bool valid_only_metric(const char *unit
)
852 if (strstr(unit
, "/sec") ||
853 strstr(unit
, "hz") ||
854 strstr(unit
, "Hz") ||
855 strstr(unit
, "CPUs utilized"))
860 static const char *fixunit(char *buf
, struct perf_evsel
*evsel
,
863 if (!strncmp(unit
, "of all", 6)) {
864 snprintf(buf
, 1024, "%s %s", perf_evsel__name(evsel
),
871 static void print_metric_only(void *ctx
, const char *color
, const char *fmt
,
872 const char *unit
, double val
)
874 struct outstate
*os
= ctx
;
878 unsigned mlen
= METRIC_ONLY_LEN
;
880 if (!valid_only_metric(unit
))
882 unit
= fixunit(buf
, os
->evsel
, unit
);
884 n
= color_fprintf(out
, color
, fmt
, val
);
886 n
= fprintf(out
, fmt
, val
);
887 if (n
> METRIC_ONLY_LEN
)
889 if (mlen
< strlen(unit
))
890 mlen
= strlen(unit
) + 1;
891 fprintf(out
, "%*s", mlen
- n
, "");
894 static void print_metric_only_csv(void *ctx
, const char *color __maybe_unused
,
896 const char *unit
, double val
)
898 struct outstate
*os
= ctx
;
900 char buf
[64], *vals
, *ends
;
903 if (!valid_only_metric(unit
))
905 unit
= fixunit(tbuf
, os
->evsel
, unit
);
906 snprintf(buf
, sizeof buf
, fmt
, val
);
908 while (isspace(*vals
))
911 while (isdigit(*ends
) || *ends
== '.')
914 fprintf(out
, "%s%s", vals
, csv_sep
);
917 static void new_line_metric(void *ctx __maybe_unused
)
921 static void print_metric_header(void *ctx
, const char *color __maybe_unused
,
922 const char *fmt __maybe_unused
,
923 const char *unit
, double val __maybe_unused
)
925 struct outstate
*os
= ctx
;
928 if (!valid_only_metric(unit
))
930 unit
= fixunit(tbuf
, os
->evsel
, unit
);
932 fprintf(os
->fh
, "%s%s", unit
, csv_sep
);
934 fprintf(os
->fh
, "%-*s ", METRIC_ONLY_LEN
, unit
);
937 static void nsec_printout(int id
, int nr
, struct perf_evsel
*evsel
, double avg
)
939 FILE *output
= stat_config
.output
;
940 double msecs
= avg
/ 1e6
;
941 const char *fmt_v
, *fmt_n
;
944 fmt_v
= csv_output
? "%.6f%s" : "%18.6f%s";
945 fmt_n
= csv_output
? "%s" : "%-25s";
947 aggr_printout(evsel
, id
, nr
);
949 scnprintf(name
, sizeof(name
), "%s%s",
950 perf_evsel__name(evsel
), csv_output
? "" : " (msec)");
952 fprintf(output
, fmt_v
, msecs
, csv_sep
);
955 fprintf(output
, "%s%s", evsel
->unit
, csv_sep
);
957 fprintf(output
, "%-*s%s", unit_width
, evsel
->unit
, csv_sep
);
959 fprintf(output
, fmt_n
, name
);
962 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
965 static int first_shadow_cpu(struct perf_evsel
*evsel
, int id
)
972 if (stat_config
.aggr_mode
== AGGR_NONE
)
975 if (stat_config
.aggr_mode
== AGGR_GLOBAL
)
978 for (i
= 0; i
< perf_evsel__nr_cpus(evsel
); i
++) {
979 int cpu2
= perf_evsel__cpus(evsel
)->map
[i
];
981 if (aggr_get_id(evsel_list
->cpus
, cpu2
) == id
)
987 static void abs_printout(int id
, int nr
, struct perf_evsel
*evsel
, double avg
)
989 FILE *output
= stat_config
.output
;
990 double sc
= evsel
->scale
;
994 fmt
= sc
!= 1.0 ? "%.2f%s" : "%.0f%s";
997 fmt
= sc
!= 1.0 ? "%'18.2f%s" : "%'18.0f%s";
999 fmt
= sc
!= 1.0 ? "%18.2f%s" : "%18.0f%s";
1002 aggr_printout(evsel
, id
, nr
);
1004 fprintf(output
, fmt
, avg
, csv_sep
);
1007 fprintf(output
, "%-*s%s",
1008 csv_output
? 0 : unit_width
,
1009 evsel
->unit
, csv_sep
);
1011 fprintf(output
, "%-*s", csv_output
? 0 : 25, perf_evsel__name(evsel
));
1014 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
1017 static void printout(int id
, int nr
, struct perf_evsel
*counter
, double uval
,
1018 char *prefix
, u64 run
, u64 ena
, double noise
)
1020 struct perf_stat_output_ctx out
;
1021 struct outstate os
= {
1022 .fh
= stat_config
.output
,
1023 .prefix
= prefix
? prefix
: "",
1028 print_metric_t pm
= print_metric_std
;
1032 nl
= new_line_metric
;
1034 pm
= print_metric_only_csv
;
1036 pm
= print_metric_only
;
1040 if (csv_output
&& !metric_only
) {
1041 static int aggr_fields
[] = {
1049 pm
= print_metric_csv
;
1052 os
.nfields
+= aggr_fields
[stat_config
.aggr_mode
];
1056 if (run
== 0 || ena
== 0 || counter
->counts
->scaled
== -1) {
1058 pm(&os
, NULL
, "", "", 0);
1061 aggr_printout(counter
, id
, nr
);
1063 fprintf(stat_config
.output
, "%*s%s",
1064 csv_output
? 0 : 18,
1065 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1068 fprintf(stat_config
.output
, "%-*s%s",
1069 csv_output
? 0 : unit_width
,
1070 counter
->unit
, csv_sep
);
1072 fprintf(stat_config
.output
, "%*s",
1073 csv_output
? 0 : -25,
1074 perf_evsel__name(counter
));
1077 fprintf(stat_config
.output
, "%s%s",
1078 csv_sep
, counter
->cgrp
->name
);
1081 pm(&os
, NULL
, NULL
, "", 0);
1082 print_noise(counter
, noise
);
1083 print_running(run
, ena
);
1085 pm(&os
, NULL
, NULL
, "", 0);
1091 else if (nsec_counter(counter
))
1092 nsec_printout(id
, nr
, counter
, uval
);
1094 abs_printout(id
, nr
, counter
, uval
);
1096 out
.print_metric
= pm
;
1100 if (csv_output
&& !metric_only
) {
1101 print_noise(counter
, noise
);
1102 print_running(run
, ena
);
1105 perf_stat__print_shadow_stats(counter
, uval
,
1106 first_shadow_cpu(counter
, id
),
1108 if (!csv_output
&& !metric_only
) {
1109 print_noise(counter
, noise
);
1110 print_running(run
, ena
);
1114 static void aggr_update_shadow(void)
1118 struct perf_evsel
*counter
;
1120 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1121 id
= aggr_map
->map
[s
];
1122 evlist__for_each(evsel_list
, counter
) {
1124 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1125 s2
= aggr_get_id(evsel_list
->cpus
, cpu
);
1128 val
+= perf_counts(counter
->counts
, cpu
, 0)->val
;
1130 val
= val
* counter
->scale
;
1131 perf_stat__update_shadow_stats(counter
, &val
,
1132 first_shadow_cpu(counter
, id
));
1137 static void print_aggr(char *prefix
)
1139 FILE *output
= stat_config
.output
;
1140 struct perf_evsel
*counter
;
1141 int cpu
, s
, s2
, id
, nr
;
1146 if (!(aggr_map
|| aggr_get_id
))
1149 aggr_update_shadow();
1152 * With metric_only everything is on a single line.
1153 * Without each counter has its own line.
1155 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1156 if (prefix
&& metric_only
)
1157 fprintf(output
, "%s", prefix
);
1159 id
= aggr_map
->map
[s
];
1161 evlist__for_each(evsel_list
, counter
) {
1162 val
= ena
= run
= 0;
1164 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1165 s2
= aggr_get_id(perf_evsel__cpus(counter
), cpu
);
1168 val
+= perf_counts(counter
->counts
, cpu
, 0)->val
;
1169 ena
+= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1170 run
+= perf_counts(counter
->counts
, cpu
, 0)->run
;
1173 if (first
&& metric_only
) {
1175 aggr_printout(counter
, id
, nr
);
1177 if (prefix
&& !metric_only
)
1178 fprintf(output
, "%s", prefix
);
1180 uval
= val
* counter
->scale
;
1181 printout(id
, nr
, counter
, uval
, prefix
, run
, ena
, 1.0);
1183 fputc('\n', output
);
1186 fputc('\n', output
);
1190 static void print_aggr_thread(struct perf_evsel
*counter
, char *prefix
)
1192 FILE *output
= stat_config
.output
;
1193 int nthreads
= thread_map__nr(counter
->threads
);
1194 int ncpus
= cpu_map__nr(counter
->cpus
);
1198 for (thread
= 0; thread
< nthreads
; thread
++) {
1199 u64 ena
= 0, run
= 0, val
= 0;
1201 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
1202 val
+= perf_counts(counter
->counts
, cpu
, thread
)->val
;
1203 ena
+= perf_counts(counter
->counts
, cpu
, thread
)->ena
;
1204 run
+= perf_counts(counter
->counts
, cpu
, thread
)->run
;
1208 fprintf(output
, "%s", prefix
);
1210 uval
= val
* counter
->scale
;
1211 printout(thread
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1212 fputc('\n', output
);
1217 * Print out the results of a single counter:
1218 * aggregated counts in system-wide mode
1220 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
1222 FILE *output
= stat_config
.output
;
1223 struct perf_stat_evsel
*ps
= counter
->priv
;
1224 double avg
= avg_stats(&ps
->res_stats
[0]);
1226 double avg_enabled
, avg_running
;
1228 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
1229 avg_running
= avg_stats(&ps
->res_stats
[2]);
1231 if (prefix
&& !metric_only
)
1232 fprintf(output
, "%s", prefix
);
1234 uval
= avg
* counter
->scale
;
1235 printout(-1, 0, counter
, uval
, prefix
, avg_running
, avg_enabled
, avg
);
1237 fprintf(output
, "\n");
1241 * Print out the results of a single counter:
1242 * does not use aggregated count in system-wide
1244 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1246 FILE *output
= stat_config
.output
;
1251 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1252 val
= perf_counts(counter
->counts
, cpu
, 0)->val
;
1253 ena
= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1254 run
= perf_counts(counter
->counts
, cpu
, 0)->run
;
1257 fprintf(output
, "%s", prefix
);
1259 uval
= val
* counter
->scale
;
1260 printout(cpu
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1262 fputc('\n', output
);
1266 static void print_no_aggr_metric(char *prefix
)
1270 struct perf_evsel
*counter
;
1274 nrcpus
= evsel_list
->cpus
->nr
;
1275 for (cpu
= 0; cpu
< nrcpus
; cpu
++) {
1279 fputs(prefix
, stat_config
.output
);
1280 evlist__for_each(evsel_list
, counter
) {
1282 aggr_printout(counter
, cpu
, 0);
1285 val
= perf_counts(counter
->counts
, cpu
, 0)->val
;
1286 ena
= perf_counts(counter
->counts
, cpu
, 0)->ena
;
1287 run
= perf_counts(counter
->counts
, cpu
, 0)->run
;
1289 uval
= val
* counter
->scale
;
1290 printout(cpu
, 0, counter
, uval
, prefix
, run
, ena
, 1.0);
1292 fputc('\n', stat_config
.output
);
1296 static int aggr_header_lens
[] = {
1304 static void print_metric_headers(char *prefix
)
1306 struct perf_stat_output_ctx out
;
1307 struct perf_evsel
*counter
;
1308 struct outstate os
= {
1309 .fh
= stat_config
.output
1313 fprintf(stat_config
.output
, "%s", prefix
);
1316 fprintf(stat_config
.output
, "%*s",
1317 aggr_header_lens
[stat_config
.aggr_mode
], "");
1319 /* Print metrics headers only */
1320 evlist__for_each(evsel_list
, counter
) {
1323 out
.print_metric
= print_metric_header
;
1324 out
.new_line
= new_line_metric
;
1326 perf_stat__print_shadow_stats(counter
, 0,
1330 fputc('\n', stat_config
.output
);
1333 static void print_interval(char *prefix
, struct timespec
*ts
)
1335 FILE *output
= stat_config
.output
;
1336 static int num_print_interval
;
1338 sprintf(prefix
, "%6lu.%09lu%s", ts
->tv_sec
, ts
->tv_nsec
, csv_sep
);
1340 if (num_print_interval
== 0 && !csv_output
&& !metric_only
) {
1341 switch (stat_config
.aggr_mode
) {
1343 fprintf(output
, "# time socket cpus counts %*s events\n", unit_width
, "unit");
1346 fprintf(output
, "# time core cpus counts %*s events\n", unit_width
, "unit");
1349 fprintf(output
, "# time CPU counts %*s events\n", unit_width
, "unit");
1352 fprintf(output
, "# time comm-pid counts %*s events\n", unit_width
, "unit");
1356 fprintf(output
, "# time counts %*s events\n", unit_width
, "unit");
1362 if (++num_print_interval
== 25)
1363 num_print_interval
= 0;
1366 static void print_header(int argc
, const char **argv
)
1368 FILE *output
= stat_config
.output
;
1374 fprintf(output
, "\n");
1375 fprintf(output
, " Performance counter stats for ");
1376 if (target
.system_wide
)
1377 fprintf(output
, "\'system wide");
1378 else if (target
.cpu_list
)
1379 fprintf(output
, "\'CPU(s) %s", target
.cpu_list
);
1380 else if (!target__has_task(&target
)) {
1381 fprintf(output
, "\'%s", argv
? argv
[0] : "pipe");
1382 for (i
= 1; argv
&& (i
< argc
); i
++)
1383 fprintf(output
, " %s", argv
[i
]);
1384 } else if (target
.pid
)
1385 fprintf(output
, "process id \'%s", target
.pid
);
1387 fprintf(output
, "thread id \'%s", target
.tid
);
1389 fprintf(output
, "\'");
1391 fprintf(output
, " (%d runs)", run_count
);
1392 fprintf(output
, ":\n\n");
1396 static void print_footer(void)
1398 FILE *output
= stat_config
.output
;
1401 fprintf(output
, "\n");
1402 fprintf(output
, " %17.9f seconds time elapsed",
1403 avg_stats(&walltime_nsecs_stats
)/1e9
);
1404 if (run_count
> 1) {
1405 fprintf(output
, " ");
1406 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1407 avg_stats(&walltime_nsecs_stats
));
1409 fprintf(output
, "\n\n");
1412 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
)
1414 int interval
= stat_config
.interval
;
1415 struct perf_evsel
*counter
;
1416 char buf
[64], *prefix
= NULL
;
1418 /* Do not print anything if we record to the pipe. */
1419 if (STAT_RECORD
&& perf_stat
.file
.is_pipe
)
1423 print_interval(prefix
= buf
, ts
);
1425 print_header(argc
, argv
);
1428 static int num_print_iv
;
1430 if (num_print_iv
== 0)
1431 print_metric_headers(prefix
);
1432 if (num_print_iv
++ == 25)
1434 if (stat_config
.aggr_mode
== AGGR_GLOBAL
&& prefix
)
1435 fprintf(stat_config
.output
, "%s", prefix
);
1438 switch (stat_config
.aggr_mode
) {
1444 evlist__for_each(evsel_list
, counter
)
1445 print_aggr_thread(counter
, prefix
);
1448 evlist__for_each(evsel_list
, counter
)
1449 print_counter_aggr(counter
, prefix
);
1451 fputc('\n', stat_config
.output
);
1455 print_no_aggr_metric(prefix
);
1457 evlist__for_each(evsel_list
, counter
)
1458 print_counter(counter
, prefix
);
1466 if (!interval
&& !csv_output
)
1469 fflush(stat_config
.output
);
1472 static volatile int signr
= -1;
1474 static void skip_signal(int signo
)
1476 if ((child_pid
== -1) || stat_config
.interval
)
1481 * render child_pid harmless
1482 * won't send SIGTERM to a random
1483 * process in case of race condition
1484 * and fast PID recycling
1489 static void sig_atexit(void)
1494 * avoid race condition with SIGCHLD handler
1495 * in skip_signal() which is modifying child_pid
1496 * goal is to avoid send SIGTERM to a random
1500 sigaddset(&set
, SIGCHLD
);
1501 sigprocmask(SIG_BLOCK
, &set
, &oset
);
1503 if (child_pid
!= -1)
1504 kill(child_pid
, SIGTERM
);
1506 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
1511 signal(signr
, SIG_DFL
);
1512 kill(getpid(), signr
);
1515 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1516 const char *s __maybe_unused
, int unset
)
1518 big_num_opt
= unset
? 0 : 1;
1522 static const struct option stat_options
[] = {
1523 OPT_BOOLEAN('T', "transaction", &transaction_run
,
1524 "hardware transaction statistics"),
1525 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1526 "event selector. use 'perf list' to list available events",
1527 parse_events_option
),
1528 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1529 "event filter", parse_filter
),
1530 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1531 "child tasks do not inherit counters"),
1532 OPT_STRING('p', "pid", &target
.pid
, "pid",
1533 "stat events on existing process id"),
1534 OPT_STRING('t', "tid", &target
.tid
, "tid",
1535 "stat events on existing thread id"),
1536 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1537 "system-wide collection from all CPUs"),
1538 OPT_BOOLEAN('g', "group", &group
,
1539 "put the counters into a counter group"),
1540 OPT_BOOLEAN('c', "scale", &stat_config
.scale
, "scale/normalize counters"),
1541 OPT_INCR('v', "verbose", &verbose
,
1542 "be more verbose (show counter open errors, etc)"),
1543 OPT_INTEGER('r', "repeat", &run_count
,
1544 "repeat command and print average + stddev (max: 100, forever: 0)"),
1545 OPT_BOOLEAN('n', "null", &null_run
,
1546 "null run - dont start any counters"),
1547 OPT_INCR('d', "detailed", &detailed_run
,
1548 "detailed run - start a lot of events"),
1549 OPT_BOOLEAN('S', "sync", &sync_run
,
1550 "call sync() before starting a run"),
1551 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1552 "print large numbers with thousands\' separators",
1554 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1555 "list of cpus to monitor in system-wide"),
1556 OPT_SET_UINT('A', "no-aggr", &stat_config
.aggr_mode
,
1557 "disable CPU count aggregation", AGGR_NONE
),
1558 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1559 "print counts with custom separator"),
1560 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1561 "monitor event in cgroup name only", parse_cgroups
),
1562 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1563 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1564 OPT_INTEGER(0, "log-fd", &output_fd
,
1565 "log output to fd, instead of stderr"),
1566 OPT_STRING(0, "pre", &pre_cmd
, "command",
1567 "command to run prior to the measured command"),
1568 OPT_STRING(0, "post", &post_cmd
, "command",
1569 "command to run after to the measured command"),
1570 OPT_UINTEGER('I', "interval-print", &stat_config
.interval
,
1571 "print counts at regular interval in ms (>= 10)"),
1572 OPT_SET_UINT(0, "per-socket", &stat_config
.aggr_mode
,
1573 "aggregate counts per processor socket", AGGR_SOCKET
),
1574 OPT_SET_UINT(0, "per-core", &stat_config
.aggr_mode
,
1575 "aggregate counts per physical processor core", AGGR_CORE
),
1576 OPT_SET_UINT(0, "per-thread", &stat_config
.aggr_mode
,
1577 "aggregate counts per thread", AGGR_THREAD
),
1578 OPT_UINTEGER('D', "delay", &initial_delay
,
1579 "ms to wait before starting measurement after program start"),
1580 OPT_BOOLEAN(0, "metric-only", &metric_only
,
1581 "Only print computed metrics. No raw values"),
1585 static int perf_stat__get_socket(struct cpu_map
*map
, int cpu
)
1587 return cpu_map__get_socket(map
, cpu
, NULL
);
1590 static int perf_stat__get_core(struct cpu_map
*map
, int cpu
)
1592 return cpu_map__get_core(map
, cpu
, NULL
);
1595 static int cpu_map__get_max(struct cpu_map
*map
)
1599 for (i
= 0; i
< map
->nr
; i
++) {
1600 if (map
->map
[i
] > max
)
1607 static struct cpu_map
*cpus_aggr_map
;
1609 static int perf_stat__get_aggr(aggr_get_id_t get_id
, struct cpu_map
*map
, int idx
)
1616 cpu
= map
->map
[idx
];
1618 if (cpus_aggr_map
->map
[cpu
] == -1)
1619 cpus_aggr_map
->map
[cpu
] = get_id(map
, idx
);
1621 return cpus_aggr_map
->map
[cpu
];
1624 static int perf_stat__get_socket_cached(struct cpu_map
*map
, int idx
)
1626 return perf_stat__get_aggr(perf_stat__get_socket
, map
, idx
);
1629 static int perf_stat__get_core_cached(struct cpu_map
*map
, int idx
)
1631 return perf_stat__get_aggr(perf_stat__get_core
, map
, idx
);
1634 static int perf_stat_init_aggr_mode(void)
1638 switch (stat_config
.aggr_mode
) {
1640 if (cpu_map__build_socket_map(evsel_list
->cpus
, &aggr_map
)) {
1641 perror("cannot build socket map");
1644 aggr_get_id
= perf_stat__get_socket_cached
;
1647 if (cpu_map__build_core_map(evsel_list
->cpus
, &aggr_map
)) {
1648 perror("cannot build core map");
1651 aggr_get_id
= perf_stat__get_core_cached
;
1662 * The evsel_list->cpus is the base we operate on,
1663 * taking the highest cpu number to be the size of
1664 * the aggregation translate cpumap.
1666 nr
= cpu_map__get_max(evsel_list
->cpus
);
1667 cpus_aggr_map
= cpu_map__empty_new(nr
+ 1);
1668 return cpus_aggr_map
? 0 : -ENOMEM
;
1671 static void perf_stat__exit_aggr_mode(void)
1673 cpu_map__put(aggr_map
);
1674 cpu_map__put(cpus_aggr_map
);
1676 cpus_aggr_map
= NULL
;
1679 static inline int perf_env__get_cpu(struct perf_env
*env
, struct cpu_map
*map
, int idx
)
1686 cpu
= map
->map
[idx
];
1688 if (cpu
>= env
->nr_cpus_online
)
1694 static int perf_env__get_socket(struct cpu_map
*map
, int idx
, void *data
)
1696 struct perf_env
*env
= data
;
1697 int cpu
= perf_env__get_cpu(env
, map
, idx
);
1699 return cpu
== -1 ? -1 : env
->cpu
[cpu
].socket_id
;
1702 static int perf_env__get_core(struct cpu_map
*map
, int idx
, void *data
)
1704 struct perf_env
*env
= data
;
1705 int core
= -1, cpu
= perf_env__get_cpu(env
, map
, idx
);
1708 int socket_id
= env
->cpu
[cpu
].socket_id
;
1711 * Encode socket in upper 16 bits
1712 * core_id is relative to socket, and
1713 * we need a global id. So we combine
1716 core
= (socket_id
<< 16) | (env
->cpu
[cpu
].core_id
& 0xffff);
1722 static int perf_env__build_socket_map(struct perf_env
*env
, struct cpu_map
*cpus
,
1723 struct cpu_map
**sockp
)
1725 return cpu_map__build_map(cpus
, sockp
, perf_env__get_socket
, env
);
1728 static int perf_env__build_core_map(struct perf_env
*env
, struct cpu_map
*cpus
,
1729 struct cpu_map
**corep
)
1731 return cpu_map__build_map(cpus
, corep
, perf_env__get_core
, env
);
1734 static int perf_stat__get_socket_file(struct cpu_map
*map
, int idx
)
1736 return perf_env__get_socket(map
, idx
, &perf_stat
.session
->header
.env
);
1739 static int perf_stat__get_core_file(struct cpu_map
*map
, int idx
)
1741 return perf_env__get_core(map
, idx
, &perf_stat
.session
->header
.env
);
1744 static int perf_stat_init_aggr_mode_file(struct perf_stat
*st
)
1746 struct perf_env
*env
= &st
->session
->header
.env
;
1748 switch (stat_config
.aggr_mode
) {
1750 if (perf_env__build_socket_map(env
, evsel_list
->cpus
, &aggr_map
)) {
1751 perror("cannot build socket map");
1754 aggr_get_id
= perf_stat__get_socket_file
;
1757 if (perf_env__build_core_map(env
, evsel_list
->cpus
, &aggr_map
)) {
1758 perror("cannot build core map");
1761 aggr_get_id
= perf_stat__get_core_file
;
1775 * Add default attributes, if there were no attributes specified or
1776 * if -d/--detailed, -d -d or -d -d -d is used:
1778 static int add_default_attributes(void)
1780 struct perf_event_attr default_attrs0
[] = {
1782 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1783 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1784 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1785 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1787 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1789 struct perf_event_attr frontend_attrs
[] = {
1790 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1792 struct perf_event_attr backend_attrs
[] = {
1793 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1795 struct perf_event_attr default_attrs1
[] = {
1796 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1797 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1798 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1803 * Detailed stats (-d), covering the L1 and last level data caches:
1805 struct perf_event_attr detailed_attrs
[] = {
1807 { .type
= PERF_TYPE_HW_CACHE
,
1809 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1810 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1811 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1813 { .type
= PERF_TYPE_HW_CACHE
,
1815 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1816 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1817 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1819 { .type
= PERF_TYPE_HW_CACHE
,
1821 PERF_COUNT_HW_CACHE_LL
<< 0 |
1822 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1823 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1825 { .type
= PERF_TYPE_HW_CACHE
,
1827 PERF_COUNT_HW_CACHE_LL
<< 0 |
1828 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1829 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1833 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1835 struct perf_event_attr very_detailed_attrs
[] = {
1837 { .type
= PERF_TYPE_HW_CACHE
,
1839 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1840 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1841 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1843 { .type
= PERF_TYPE_HW_CACHE
,
1845 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1846 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1847 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1849 { .type
= PERF_TYPE_HW_CACHE
,
1851 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1852 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1853 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1855 { .type
= PERF_TYPE_HW_CACHE
,
1857 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1858 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1859 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1861 { .type
= PERF_TYPE_HW_CACHE
,
1863 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1864 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1865 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1867 { .type
= PERF_TYPE_HW_CACHE
,
1869 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1870 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1871 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1876 * Very, very detailed stats (-d -d -d), adding prefetch events:
1878 struct perf_event_attr very_very_detailed_attrs
[] = {
1880 { .type
= PERF_TYPE_HW_CACHE
,
1882 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1883 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1884 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1886 { .type
= PERF_TYPE_HW_CACHE
,
1888 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1889 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1890 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1893 /* Set attrs if no event is selected and !null_run: */
1897 if (transaction_run
) {
1899 if (pmu_have_event("cpu", "cycles-ct") &&
1900 pmu_have_event("cpu", "el-start"))
1901 err
= parse_events(evsel_list
, transaction_attrs
, NULL
);
1903 err
= parse_events(evsel_list
, transaction_limited_attrs
, NULL
);
1905 fprintf(stderr
, "Cannot set up transaction events\n");
1911 if (!evsel_list
->nr_entries
) {
1912 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs0
) < 0)
1914 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1915 if (perf_evlist__add_default_attrs(evsel_list
,
1916 frontend_attrs
) < 0)
1919 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1920 if (perf_evlist__add_default_attrs(evsel_list
,
1924 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs1
) < 0)
1928 /* Detailed events get appended to the event list: */
1930 if (detailed_run
< 1)
1933 /* Append detailed run extra attributes: */
1934 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1937 if (detailed_run
< 2)
1940 /* Append very detailed run extra attributes: */
1941 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1944 if (detailed_run
< 3)
1947 /* Append very, very detailed run extra attributes: */
1948 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1951 static const char * const stat_record_usage
[] = {
1952 "perf stat record [<options>]",
1956 static void init_features(struct perf_session
*session
)
1960 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
1961 perf_header__set_feat(&session
->header
, feat
);
1963 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
1964 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
1965 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
1966 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
1969 static int __cmd_record(int argc
, const char **argv
)
1971 struct perf_session
*session
;
1972 struct perf_data_file
*file
= &perf_stat
.file
;
1974 argc
= parse_options(argc
, argv
, stat_options
, stat_record_usage
,
1975 PARSE_OPT_STOP_AT_NON_OPTION
);
1978 file
->path
= output_name
;
1980 if (run_count
!= 1 || forever
) {
1981 pr_err("Cannot use -r option with perf stat record.\n");
1985 session
= perf_session__new(file
, false, NULL
);
1986 if (session
== NULL
) {
1987 pr_err("Perf session creation failed.\n");
1991 init_features(session
);
1993 session
->evlist
= evsel_list
;
1994 perf_stat
.session
= session
;
1995 perf_stat
.record
= true;
1999 static int process_stat_round_event(struct perf_tool
*tool __maybe_unused
,
2000 union perf_event
*event
,
2001 struct perf_session
*session
)
2003 struct stat_round_event
*round
= &event
->stat_round
;
2004 struct perf_evsel
*counter
;
2005 struct timespec tsh
, *ts
= NULL
;
2006 const char **argv
= session
->header
.env
.cmdline_argv
;
2007 int argc
= session
->header
.env
.nr_cmdline
;
2009 evlist__for_each(evsel_list
, counter
)
2010 perf_stat_process_counter(&stat_config
, counter
);
2012 if (round
->type
== PERF_STAT_ROUND_TYPE__FINAL
)
2013 update_stats(&walltime_nsecs_stats
, round
->time
);
2015 if (stat_config
.interval
&& round
->time
) {
2016 tsh
.tv_sec
= round
->time
/ NSECS_PER_SEC
;
2017 tsh
.tv_nsec
= round
->time
% NSECS_PER_SEC
;
2021 print_counters(ts
, argc
, argv
);
2026 int process_stat_config_event(struct perf_tool
*tool __maybe_unused
,
2027 union perf_event
*event
,
2028 struct perf_session
*session __maybe_unused
)
2030 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2032 perf_event__read_stat_config(&stat_config
, &event
->stat_config
);
2034 if (cpu_map__empty(st
->cpus
)) {
2035 if (st
->aggr_mode
!= AGGR_UNSET
)
2036 pr_warning("warning: processing task data, aggregation mode not set\n");
2040 if (st
->aggr_mode
!= AGGR_UNSET
)
2041 stat_config
.aggr_mode
= st
->aggr_mode
;
2043 if (perf_stat
.file
.is_pipe
)
2044 perf_stat_init_aggr_mode();
2046 perf_stat_init_aggr_mode_file(st
);
2051 static int set_maps(struct perf_stat
*st
)
2053 if (!st
->cpus
|| !st
->threads
)
2056 if (WARN_ONCE(st
->maps_allocated
, "stats double allocation\n"))
2059 perf_evlist__set_maps(evsel_list
, st
->cpus
, st
->threads
);
2061 if (perf_evlist__alloc_stats(evsel_list
, true))
2064 st
->maps_allocated
= true;
2069 int process_thread_map_event(struct perf_tool
*tool __maybe_unused
,
2070 union perf_event
*event
,
2071 struct perf_session
*session __maybe_unused
)
2073 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2076 pr_warning("Extra thread map event, ignoring.\n");
2080 st
->threads
= thread_map__new_event(&event
->thread_map
);
2084 return set_maps(st
);
2088 int process_cpu_map_event(struct perf_tool
*tool __maybe_unused
,
2089 union perf_event
*event
,
2090 struct perf_session
*session __maybe_unused
)
2092 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
2093 struct cpu_map
*cpus
;
2096 pr_warning("Extra cpu map event, ignoring.\n");
2100 cpus
= cpu_map__new_data(&event
->cpu_map
.data
);
2105 return set_maps(st
);
2108 static const char * const stat_report_usage
[] = {
2109 "perf stat report [<options>]",
2113 static struct perf_stat perf_stat
= {
2115 .attr
= perf_event__process_attr
,
2116 .event_update
= perf_event__process_event_update
,
2117 .thread_map
= process_thread_map_event
,
2118 .cpu_map
= process_cpu_map_event
,
2119 .stat_config
= process_stat_config_event
,
2120 .stat
= perf_event__process_stat_event
,
2121 .stat_round
= process_stat_round_event
,
2123 .aggr_mode
= AGGR_UNSET
,
2126 static int __cmd_report(int argc
, const char **argv
)
2128 struct perf_session
*session
;
2129 const struct option options
[] = {
2130 OPT_STRING('i', "input", &input_name
, "file", "input file name"),
2131 OPT_SET_UINT(0, "per-socket", &perf_stat
.aggr_mode
,
2132 "aggregate counts per processor socket", AGGR_SOCKET
),
2133 OPT_SET_UINT(0, "per-core", &perf_stat
.aggr_mode
,
2134 "aggregate counts per physical processor core", AGGR_CORE
),
2135 OPT_SET_UINT('A', "no-aggr", &perf_stat
.aggr_mode
,
2136 "disable CPU count aggregation", AGGR_NONE
),
2142 argc
= parse_options(argc
, argv
, options
, stat_report_usage
, 0);
2144 if (!input_name
|| !strlen(input_name
)) {
2145 if (!fstat(STDIN_FILENO
, &st
) && S_ISFIFO(st
.st_mode
))
2148 input_name
= "perf.data";
2151 perf_stat
.file
.path
= input_name
;
2152 perf_stat
.file
.mode
= PERF_DATA_MODE_READ
;
2154 session
= perf_session__new(&perf_stat
.file
, false, &perf_stat
.tool
);
2155 if (session
== NULL
)
2158 perf_stat
.session
= session
;
2159 stat_config
.output
= stderr
;
2160 evsel_list
= session
->evlist
;
2162 ret
= perf_session__process_events(session
);
2166 perf_session__delete(session
);
2170 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
2172 const char * const stat_usage
[] = {
2173 "perf stat [<options>] [<command>]",
2176 int status
= -EINVAL
, run_idx
;
2178 FILE *output
= stderr
;
2179 unsigned int interval
;
2180 const char * const stat_subcommands
[] = { "record", "report" };
2182 setlocale(LC_ALL
, "");
2184 evsel_list
= perf_evlist__new();
2185 if (evsel_list
== NULL
)
2188 parse_events__shrink_config_terms();
2189 argc
= parse_options_subcommand(argc
, argv
, stat_options
, stat_subcommands
,
2190 (const char **) stat_usage
,
2191 PARSE_OPT_STOP_AT_NON_OPTION
);
2192 perf_stat__init_shadow_stats();
2196 if (!strcmp(csv_sep
, "\\t"))
2199 csv_sep
= DEFAULT_SEPARATOR
;
2201 if (argc
&& !strncmp(argv
[0], "rec", 3)) {
2202 argc
= __cmd_record(argc
, argv
);
2205 } else if (argc
&& !strncmp(argv
[0], "rep", 3))
2206 return __cmd_report(argc
, argv
);
2208 interval
= stat_config
.interval
;
2211 * For record command the -o is already taken care of.
2213 if (!STAT_RECORD
&& output_name
&& strcmp(output_name
, "-"))
2216 if (output_name
&& output_fd
) {
2217 fprintf(stderr
, "cannot use both --output and --log-fd\n");
2218 parse_options_usage(stat_usage
, stat_options
, "o", 1);
2219 parse_options_usage(NULL
, stat_options
, "log-fd", 0);
2223 if (metric_only
&& stat_config
.aggr_mode
== AGGR_THREAD
) {
2224 fprintf(stderr
, "--metric-only is not supported with --per-thread\n");
2228 if (metric_only
&& run_count
> 1) {
2229 fprintf(stderr
, "--metric-only is not supported with -r\n");
2233 if (output_fd
< 0) {
2234 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
2235 parse_options_usage(stat_usage
, stat_options
, "log-fd", 0);
2241 mode
= append_file
? "a" : "w";
2243 output
= fopen(output_name
, mode
);
2245 perror("failed to create output file");
2248 clock_gettime(CLOCK_REALTIME
, &tm
);
2249 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
2250 } else if (output_fd
> 0) {
2251 mode
= append_file
? "a" : "w";
2252 output
= fdopen(output_fd
, mode
);
2254 perror("Failed opening logfd");
2259 stat_config
.output
= output
;
2262 * let the spreadsheet do the pretty-printing
2265 /* User explicitly passed -B? */
2266 if (big_num_opt
== 1) {
2267 fprintf(stderr
, "-B option not supported with -x\n");
2268 parse_options_usage(stat_usage
, stat_options
, "B", 1);
2269 parse_options_usage(NULL
, stat_options
, "x", 1);
2271 } else /* Nope, so disable big number formatting */
2273 } else if (big_num_opt
== 0) /* User passed --no-big-num */
2276 if (!argc
&& target__none(&target
))
2277 usage_with_options(stat_usage
, stat_options
);
2279 if (run_count
< 0) {
2280 pr_err("Run count must be a positive number\n");
2281 parse_options_usage(stat_usage
, stat_options
, "r", 1);
2283 } else if (run_count
== 0) {
2288 if ((stat_config
.aggr_mode
== AGGR_THREAD
) && !target__has_task(&target
)) {
2289 fprintf(stderr
, "The --per-thread option is only available "
2290 "when monitoring via -p -t options.\n");
2291 parse_options_usage(NULL
, stat_options
, "p", 1);
2292 parse_options_usage(NULL
, stat_options
, "t", 1);
2297 * no_aggr, cgroup are for system-wide only
2298 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2300 if (((stat_config
.aggr_mode
!= AGGR_GLOBAL
&&
2301 stat_config
.aggr_mode
!= AGGR_THREAD
) || nr_cgroups
) &&
2302 !target__has_cpu(&target
)) {
2303 fprintf(stderr
, "both cgroup and no-aggregation "
2304 "modes only available in system-wide mode\n");
2306 parse_options_usage(stat_usage
, stat_options
, "G", 1);
2307 parse_options_usage(NULL
, stat_options
, "A", 1);
2308 parse_options_usage(NULL
, stat_options
, "a", 1);
2312 if (add_default_attributes())
2315 target__validate(&target
);
2317 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
2318 if (target__has_task(&target
)) {
2319 pr_err("Problems finding threads of monitor\n");
2320 parse_options_usage(stat_usage
, stat_options
, "p", 1);
2321 parse_options_usage(NULL
, stat_options
, "t", 1);
2322 } else if (target__has_cpu(&target
)) {
2323 perror("failed to parse CPUs map");
2324 parse_options_usage(stat_usage
, stat_options
, "C", 1);
2325 parse_options_usage(NULL
, stat_options
, "a", 1);
2331 * Initialize thread_map with comm names,
2332 * so we could print it out on output.
2334 if (stat_config
.aggr_mode
== AGGR_THREAD
)
2335 thread_map__read_comms(evsel_list
->threads
);
2337 if (interval
&& interval
< 100) {
2338 if (interval
< 10) {
2339 pr_err("print interval must be >= 10ms\n");
2340 parse_options_usage(stat_usage
, stat_options
, "I", 1);
2343 pr_warning("print interval < 100ms. "
2344 "The overhead percentage could be high in some cases. "
2345 "Please proceed with caution.\n");
2348 if (perf_evlist__alloc_stats(evsel_list
, interval
))
2351 if (perf_stat_init_aggr_mode())
2355 * We dont want to block the signals - that would cause
2356 * child tasks to inherit that and Ctrl-C would not work.
2357 * What we want is for Ctrl-C to work in the exec()-ed
2358 * task, but being ignored by perf stat itself:
2362 signal(SIGINT
, skip_signal
);
2363 signal(SIGCHLD
, skip_signal
);
2364 signal(SIGALRM
, skip_signal
);
2365 signal(SIGABRT
, skip_signal
);
2368 for (run_idx
= 0; forever
|| run_idx
< run_count
; run_idx
++) {
2369 if (run_count
!= 1 && verbose
)
2370 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
2373 status
= run_perf_stat(argc
, argv
);
2374 if (forever
&& status
!= -1) {
2375 print_counters(NULL
, argc
, argv
);
2376 perf_stat__reset_stats();
2380 if (!forever
&& status
!= -1 && !interval
)
2381 print_counters(NULL
, argc
, argv
);
2385 * We synthesize the kernel mmap record just so that older tools
2386 * don't emit warnings about not being able to resolve symbols
2387 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2388 * a saner message about no samples being in the perf.data file.
2390 * This also serves to suppress a warning about f_header.data.size == 0
2391 * in header.c at the moment 'perf stat record' gets introduced, which
2392 * is not really needed once we start adding the stat specific PERF_RECORD_
2393 * records, but the need to suppress the kptr_restrict messages in older
2394 * tools remain -acme
2396 int fd
= perf_data_file__fd(&perf_stat
.file
);
2397 int err
= perf_event__synthesize_kernel_mmap((void *)&perf_stat
,
2398 process_synthesized_event
,
2399 &perf_stat
.session
->machines
.host
);
2401 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2402 "older tools may produce warnings about this file\n.");
2406 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats
.max
, FINAL
))
2407 pr_err("failed to write stat round event\n");
2410 if (!perf_stat
.file
.is_pipe
) {
2411 perf_stat
.session
->header
.data_size
+= perf_stat
.bytes_written
;
2412 perf_session__write_header(perf_stat
.session
, evsel_list
, fd
, true);
2415 perf_session__delete(perf_stat
.session
);
2418 perf_stat__exit_aggr_mode();
2419 perf_evlist__free_stats(evsel_list
);
2421 perf_evlist__delete(evsel_list
);