4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/string2.h"
67 #include "util/metricgroup.h"
71 #include <linux/time64.h>
72 #include <api/fs/fs.h>
76 #include <sys/prctl.h>
80 #include <sys/types.h>
85 #include <sys/resource.h>
88 #include "sane_ctype.h"
90 #define DEFAULT_SEPARATOR " "
91 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
93 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
);
95 /* Default events used for perf stat -T */
96 static const char *transaction_attrs
= {
108 /* More limited version when the CPU does not have all events. */
109 static const char * transaction_limited_attrs
= {
119 static const char * topdown_attrs
[] = {
120 "topdown-total-slots",
121 "topdown-slots-retired",
122 "topdown-recovery-bubbles",
123 "topdown-fetch-bubbles",
124 "topdown-slots-issued",
128 static const char *smi_cost_attrs
= {
136 static struct perf_evlist
*evsel_list
;
138 static struct target target
= {
142 #define METRIC_ONLY_LEN 20
144 static volatile pid_t child_pid
= -1;
145 static int detailed_run
= 0;
146 static bool transaction_run
;
147 static bool topdown_run
= false;
148 static bool smi_cost
= false;
149 static bool smi_reset
= false;
150 static int big_num_opt
= -1;
151 static bool group
= false;
152 static const char *pre_cmd
= NULL
;
153 static const char *post_cmd
= NULL
;
154 static bool sync_run
= false;
155 static bool forever
= false;
156 static bool force_metric_only
= false;
157 static struct timespec ref_time
;
158 static bool append_file
;
159 static bool interval_count
;
160 static const char *output_name
;
161 static int output_fd
;
165 struct perf_data data
;
166 struct perf_session
*session
;
168 struct perf_tool tool
;
170 struct cpu_map
*cpus
;
171 struct thread_map
*threads
;
172 enum aggr_mode aggr_mode
;
175 static struct perf_stat perf_stat
;
176 #define STAT_RECORD perf_stat.record
178 static volatile int done
= 0;
180 static struct perf_stat_config stat_config
= {
181 .aggr_mode
= AGGR_GLOBAL
,
183 .unit_width
= 4, /* strlen("unit") */
185 .metric_only_len
= METRIC_ONLY_LEN
,
186 .walltime_nsecs_stats
= &walltime_nsecs_stats
,
190 static inline void diff_timespec(struct timespec
*r
, struct timespec
*a
,
193 r
->tv_sec
= a
->tv_sec
- b
->tv_sec
;
194 if (a
->tv_nsec
< b
->tv_nsec
) {
195 r
->tv_nsec
= a
->tv_nsec
+ NSEC_PER_SEC
- b
->tv_nsec
;
198 r
->tv_nsec
= a
->tv_nsec
- b
->tv_nsec
;
202 static void perf_stat__reset_stats(void)
206 perf_evlist__reset_stats(evsel_list
);
207 perf_stat__reset_shadow_stats();
209 for (i
= 0; i
< stat_config
.stats_num
; i
++)
210 perf_stat__reset_shadow_per_stat(&stat_config
.stats
[i
]);
213 static int process_synthesized_event(struct perf_tool
*tool __maybe_unused
,
214 union perf_event
*event
,
215 struct perf_sample
*sample __maybe_unused
,
216 struct machine
*machine __maybe_unused
)
218 if (perf_data__write(&perf_stat
.data
, event
, event
->header
.size
) < 0) {
219 pr_err("failed to write perf data, error: %m\n");
223 perf_stat
.bytes_written
+= event
->header
.size
;
227 static int write_stat_round_event(u64 tm
, u64 type
)
229 return perf_event__synthesize_stat_round(NULL
, tm
, type
,
230 process_synthesized_event
,
234 #define WRITE_STAT_ROUND_EVENT(time, interval) \
235 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
237 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
240 perf_evsel__write_stat_event(struct perf_evsel
*counter
, u32 cpu
, u32 thread
,
241 struct perf_counts_values
*count
)
243 struct perf_sample_id
*sid
= SID(counter
, cpu
, thread
);
245 return perf_event__synthesize_stat(NULL
, cpu
, thread
, sid
->id
, count
,
246 process_synthesized_event
, NULL
);
250 * Read out the results of a single counter:
251 * do not aggregate counts across CPUs in system-wide mode
253 static int read_counter(struct perf_evsel
*counter
)
255 int nthreads
= thread_map__nr(evsel_list
->threads
);
256 int ncpus
, cpu
, thread
;
258 if (target__has_cpu(&target
) && !target__has_per_thread(&target
))
259 ncpus
= perf_evsel__nr_cpus(counter
);
263 if (!counter
->supported
)
266 if (counter
->system_wide
)
269 for (thread
= 0; thread
< nthreads
; thread
++) {
270 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
271 struct perf_counts_values
*count
;
273 count
= perf_counts(counter
->counts
, cpu
, thread
);
276 * The leader's group read loads data into its group members
277 * (via perf_evsel__read_counter) and sets threir count->loaded.
279 if (!count
->loaded
&&
280 perf_evsel__read_counter(counter
, cpu
, thread
)) {
281 counter
->counts
->scaled
= -1;
282 perf_counts(counter
->counts
, cpu
, thread
)->ena
= 0;
283 perf_counts(counter
->counts
, cpu
, thread
)->run
= 0;
287 count
->loaded
= false;
290 if (perf_evsel__write_stat_event(counter
, cpu
, thread
, count
)) {
291 pr_err("failed to write stat event\n");
297 fprintf(stat_config
.output
,
298 "%s: %d: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
299 perf_evsel__name(counter
),
301 count
->val
, count
->ena
, count
->run
);
309 static void read_counters(void)
311 struct perf_evsel
*counter
;
314 evlist__for_each_entry(evsel_list
, counter
) {
315 ret
= read_counter(counter
);
317 pr_debug("failed to read counter %s\n", counter
->name
);
319 if (ret
== 0 && perf_stat_process_counter(&stat_config
, counter
))
320 pr_warning("failed to process counter %s\n", counter
->name
);
324 static void process_interval(void)
326 struct timespec ts
, rs
;
330 clock_gettime(CLOCK_MONOTONIC
, &ts
);
331 diff_timespec(&rs
, &ts
, &ref_time
);
334 if (WRITE_STAT_ROUND_EVENT(rs
.tv_sec
* NSEC_PER_SEC
+ rs
.tv_nsec
, INTERVAL
))
335 pr_err("failed to write stat round event\n");
338 init_stats(&walltime_nsecs_stats
);
339 update_stats(&walltime_nsecs_stats
, stat_config
.interval
* 1000000);
340 print_counters(&rs
, 0, NULL
);
343 static void enable_counters(void)
345 if (stat_config
.initial_delay
)
346 usleep(stat_config
.initial_delay
* USEC_PER_MSEC
);
349 * We need to enable counters only if:
350 * - we don't have tracee (attaching to task or cpu)
351 * - we have initial delay configured
353 if (!target__none(&target
) || stat_config
.initial_delay
)
354 perf_evlist__enable(evsel_list
);
357 static void disable_counters(void)
360 * If we don't have tracee (attaching to task or cpu), counters may
361 * still be running. To get accurate group ratios, we must stop groups
362 * from counting before reading their constituent counters.
364 if (!target__none(&target
))
365 perf_evlist__disable(evsel_list
);
368 static volatile int workload_exec_errno
;
371 * perf_evlist__prepare_workload will send a SIGUSR1
372 * if the fork fails, since we asked by setting its
373 * want_signal to true.
375 static void workload_exec_failed_signal(int signo __maybe_unused
, siginfo_t
*info
,
376 void *ucontext __maybe_unused
)
378 workload_exec_errno
= info
->si_value
.sival_int
;
381 static bool perf_evsel__should_store_id(struct perf_evsel
*counter
)
383 return STAT_RECORD
|| counter
->attr
.read_format
& PERF_FORMAT_ID
;
386 static bool is_target_alive(struct target
*_target
,
387 struct thread_map
*threads
)
392 if (!target__has_task(_target
))
395 for (i
= 0; i
< threads
->nr
; i
++) {
398 scnprintf(path
, PATH_MAX
, "%s/%d", procfs__mountpoint(),
399 threads
->map
[i
].pid
);
401 if (!stat(path
, &st
))
408 static int __run_perf_stat(int argc
, const char **argv
, int run_idx
)
410 int interval
= stat_config
.interval
;
411 int times
= stat_config
.times
;
412 int timeout
= stat_config
.timeout
;
414 unsigned long long t0
, t1
;
415 struct perf_evsel
*counter
;
419 const bool forks
= (argc
> 0);
420 bool is_pipe
= STAT_RECORD
? perf_stat
.data
.is_pipe
: false;
421 struct perf_evsel_config_term
*err_term
;
424 ts
.tv_sec
= interval
/ USEC_PER_MSEC
;
425 ts
.tv_nsec
= (interval
% USEC_PER_MSEC
) * NSEC_PER_MSEC
;
426 } else if (timeout
) {
427 ts
.tv_sec
= timeout
/ USEC_PER_MSEC
;
428 ts
.tv_nsec
= (timeout
% USEC_PER_MSEC
) * NSEC_PER_MSEC
;
435 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
, is_pipe
,
436 workload_exec_failed_signal
) < 0) {
437 perror("failed to prepare workload");
440 child_pid
= evsel_list
->workload
.pid
;
444 perf_evlist__set_leader(evsel_list
);
446 evlist__for_each_entry(evsel_list
, counter
) {
448 if (create_perf_stat_counter(counter
, &stat_config
, &target
) < 0) {
450 /* Weak group failed. Reset the group. */
451 if ((errno
== EINVAL
|| errno
== EBADF
) &&
452 counter
->leader
!= counter
&&
453 counter
->weak_group
) {
454 counter
= perf_evlist__reset_weak_group(evsel_list
, counter
);
459 * PPC returns ENXIO for HW counters until 2.6.37
460 * (behavior changed with commit b0a873e).
462 if (errno
== EINVAL
|| errno
== ENOSYS
||
463 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
466 ui__warning("%s event is not supported by the kernel.\n",
467 perf_evsel__name(counter
));
468 counter
->supported
= false;
470 if ((counter
->leader
!= counter
) ||
471 !(counter
->leader
->nr_members
> 1))
473 } else if (perf_evsel__fallback(counter
, errno
, msg
, sizeof(msg
))) {
475 ui__warning("%s\n", msg
);
477 } else if (target__has_per_thread(&target
) &&
478 evsel_list
->threads
&&
479 evsel_list
->threads
->err_thread
!= -1) {
481 * For global --per-thread case, skip current
484 if (!thread_map__remove(evsel_list
->threads
,
485 evsel_list
->threads
->err_thread
)) {
486 evsel_list
->threads
->err_thread
= -1;
491 perf_evsel__open_strerror(counter
, &target
,
492 errno
, msg
, sizeof(msg
));
493 ui__error("%s\n", msg
);
496 kill(child_pid
, SIGTERM
);
500 counter
->supported
= true;
502 l
= strlen(counter
->unit
);
503 if (l
> stat_config
.unit_width
)
504 stat_config
.unit_width
= l
;
506 if (perf_evsel__should_store_id(counter
) &&
507 perf_evsel__store_ids(counter
, evsel_list
))
511 if (perf_evlist__apply_filters(evsel_list
, &counter
)) {
512 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
513 counter
->filter
, perf_evsel__name(counter
), errno
,
514 str_error_r(errno
, msg
, sizeof(msg
)));
518 if (perf_evlist__apply_drv_configs(evsel_list
, &counter
, &err_term
)) {
519 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
520 err_term
->val
.drv_cfg
, perf_evsel__name(counter
), errno
,
521 str_error_r(errno
, msg
, sizeof(msg
)));
526 int err
, fd
= perf_data__fd(&perf_stat
.data
);
529 err
= perf_header__write_pipe(perf_data__fd(&perf_stat
.data
));
531 err
= perf_session__write_header(perf_stat
.session
, evsel_list
,
538 err
= perf_stat_synthesize_config(&stat_config
, NULL
, evsel_list
,
539 process_synthesized_event
, is_pipe
);
545 * Enable counters and exec the command:
548 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
551 perf_evlist__start_workload(evsel_list
);
554 if (interval
|| timeout
) {
555 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
556 nanosleep(&ts
, NULL
);
560 if (interval_count
&& !(--times
))
564 wait4(child_pid
, &status
, 0, &stat_config
.ru_data
);
566 if (workload_exec_errno
) {
567 const char *emsg
= str_error_r(workload_exec_errno
, msg
, sizeof(msg
));
568 pr_err("Workload failed: %s\n", emsg
);
572 if (WIFSIGNALED(status
))
573 psignal(WTERMSIG(status
), argv
[0]);
577 nanosleep(&ts
, NULL
);
578 if (!is_target_alive(&target
, evsel_list
->threads
))
584 if (interval_count
&& !(--times
))
594 if (stat_config
.walltime_run_table
)
595 stat_config
.walltime_run
[run_idx
] = t1
- t0
;
597 update_stats(&walltime_nsecs_stats
, t1
- t0
);
600 * Closing a group leader splits the group, and as we only disable
601 * group leaders, results in remaining events becoming enabled. To
602 * avoid arbitrary skew, we must read all counters before closing any
606 perf_evlist__close(evsel_list
);
608 return WEXITSTATUS(status
);
611 static int run_perf_stat(int argc
, const char **argv
, int run_idx
)
616 ret
= system(pre_cmd
);
624 ret
= __run_perf_stat(argc
, argv
, run_idx
);
629 ret
= system(post_cmd
);
637 static void print_counters(struct timespec
*ts
, int argc
, const char **argv
)
639 /* Do not print anything if we record to the pipe. */
640 if (STAT_RECORD
&& perf_stat
.data
.is_pipe
)
643 perf_evlist__print_counters(evsel_list
, &stat_config
, &target
,
647 static volatile int signr
= -1;
649 static void skip_signal(int signo
)
651 if ((child_pid
== -1) || stat_config
.interval
)
656 * render child_pid harmless
657 * won't send SIGTERM to a random
658 * process in case of race condition
659 * and fast PID recycling
664 static void sig_atexit(void)
669 * avoid race condition with SIGCHLD handler
670 * in skip_signal() which is modifying child_pid
671 * goal is to avoid send SIGTERM to a random
675 sigaddset(&set
, SIGCHLD
);
676 sigprocmask(SIG_BLOCK
, &set
, &oset
);
679 kill(child_pid
, SIGTERM
);
681 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
686 signal(signr
, SIG_DFL
);
687 kill(getpid(), signr
);
690 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
691 const char *s __maybe_unused
, int unset
)
693 big_num_opt
= unset
? 0 : 1;
697 static int enable_metric_only(const struct option
*opt __maybe_unused
,
698 const char *s __maybe_unused
, int unset
)
700 force_metric_only
= true;
701 stat_config
.metric_only
= !unset
;
705 static int parse_metric_groups(const struct option
*opt
,
707 int unset __maybe_unused
)
709 return metricgroup__parse_groups(opt
, str
, &stat_config
.metric_events
);
712 static struct option stat_options
[] = {
713 OPT_BOOLEAN('T', "transaction", &transaction_run
,
714 "hardware transaction statistics"),
715 OPT_CALLBACK('e', "event", &evsel_list
, "event",
716 "event selector. use 'perf list' to list available events",
717 parse_events_option
),
718 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
719 "event filter", parse_filter
),
720 OPT_BOOLEAN('i', "no-inherit", &stat_config
.no_inherit
,
721 "child tasks do not inherit counters"),
722 OPT_STRING('p', "pid", &target
.pid
, "pid",
723 "stat events on existing process id"),
724 OPT_STRING('t', "tid", &target
.tid
, "tid",
725 "stat events on existing thread id"),
726 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
727 "system-wide collection from all CPUs"),
728 OPT_BOOLEAN('g', "group", &group
,
729 "put the counters into a counter group"),
730 OPT_BOOLEAN('c', "scale", &stat_config
.scale
, "scale/normalize counters"),
731 OPT_INCR('v', "verbose", &verbose
,
732 "be more verbose (show counter open errors, etc)"),
733 OPT_INTEGER('r', "repeat", &stat_config
.run_count
,
734 "repeat command and print average + stddev (max: 100, forever: 0)"),
735 OPT_BOOLEAN(0, "table", &stat_config
.walltime_run_table
,
736 "display details about each run (only with -r option)"),
737 OPT_BOOLEAN('n', "null", &stat_config
.null_run
,
738 "null run - dont start any counters"),
739 OPT_INCR('d', "detailed", &detailed_run
,
740 "detailed run - start a lot of events"),
741 OPT_BOOLEAN('S', "sync", &sync_run
,
742 "call sync() before starting a run"),
743 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
744 "print large numbers with thousands\' separators",
746 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
747 "list of cpus to monitor in system-wide"),
748 OPT_SET_UINT('A', "no-aggr", &stat_config
.aggr_mode
,
749 "disable CPU count aggregation", AGGR_NONE
),
750 OPT_BOOLEAN(0, "no-merge", &stat_config
.no_merge
, "Do not merge identical named events"),
751 OPT_STRING('x', "field-separator", &stat_config
.csv_sep
, "separator",
752 "print counts with custom separator"),
753 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
754 "monitor event in cgroup name only", parse_cgroups
),
755 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
756 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
757 OPT_INTEGER(0, "log-fd", &output_fd
,
758 "log output to fd, instead of stderr"),
759 OPT_STRING(0, "pre", &pre_cmd
, "command",
760 "command to run prior to the measured command"),
761 OPT_STRING(0, "post", &post_cmd
, "command",
762 "command to run after to the measured command"),
763 OPT_UINTEGER('I', "interval-print", &stat_config
.interval
,
764 "print counts at regular interval in ms "
765 "(overhead is possible for values <= 100ms)"),
766 OPT_INTEGER(0, "interval-count", &stat_config
.times
,
767 "print counts for fixed number of times"),
768 OPT_BOOLEAN(0, "interval-clear", &stat_config
.interval_clear
,
769 "clear screen in between new interval"),
770 OPT_UINTEGER(0, "timeout", &stat_config
.timeout
,
771 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
772 OPT_SET_UINT(0, "per-socket", &stat_config
.aggr_mode
,
773 "aggregate counts per processor socket", AGGR_SOCKET
),
774 OPT_SET_UINT(0, "per-core", &stat_config
.aggr_mode
,
775 "aggregate counts per physical processor core", AGGR_CORE
),
776 OPT_SET_UINT(0, "per-thread", &stat_config
.aggr_mode
,
777 "aggregate counts per thread", AGGR_THREAD
),
778 OPT_UINTEGER('D', "delay", &stat_config
.initial_delay
,
779 "ms to wait before starting measurement after program start"),
780 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config
.metric_only
, NULL
,
781 "Only print computed metrics. No raw values", enable_metric_only
),
782 OPT_BOOLEAN(0, "topdown", &topdown_run
,
783 "measure topdown level 1 statistics"),
784 OPT_BOOLEAN(0, "smi-cost", &smi_cost
,
786 OPT_CALLBACK('M', "metrics", &evsel_list
, "metric/metric group list",
787 "monitor specified metrics or metric groups (separated by ,)",
788 parse_metric_groups
),
792 static int perf_stat__get_socket(struct perf_stat_config
*config __maybe_unused
,
793 struct cpu_map
*map
, int cpu
)
795 return cpu_map__get_socket(map
, cpu
, NULL
);
798 static int perf_stat__get_core(struct perf_stat_config
*config __maybe_unused
,
799 struct cpu_map
*map
, int cpu
)
801 return cpu_map__get_core(map
, cpu
, NULL
);
804 static int cpu_map__get_max(struct cpu_map
*map
)
808 for (i
= 0; i
< map
->nr
; i
++) {
809 if (map
->map
[i
] > max
)
816 static int perf_stat__get_aggr(struct perf_stat_config
*config
,
817 aggr_get_id_t get_id
, struct cpu_map
*map
, int idx
)
826 if (config
->cpus_aggr_map
->map
[cpu
] == -1)
827 config
->cpus_aggr_map
->map
[cpu
] = get_id(config
, map
, idx
);
829 return config
->cpus_aggr_map
->map
[cpu
];
832 static int perf_stat__get_socket_cached(struct perf_stat_config
*config
,
833 struct cpu_map
*map
, int idx
)
835 return perf_stat__get_aggr(config
, perf_stat__get_socket
, map
, idx
);
838 static int perf_stat__get_core_cached(struct perf_stat_config
*config
,
839 struct cpu_map
*map
, int idx
)
841 return perf_stat__get_aggr(config
, perf_stat__get_core
, map
, idx
);
844 static int perf_stat_init_aggr_mode(void)
848 switch (stat_config
.aggr_mode
) {
850 if (cpu_map__build_socket_map(evsel_list
->cpus
, &stat_config
.aggr_map
)) {
851 perror("cannot build socket map");
854 stat_config
.aggr_get_id
= perf_stat__get_socket_cached
;
857 if (cpu_map__build_core_map(evsel_list
->cpus
, &stat_config
.aggr_map
)) {
858 perror("cannot build core map");
861 stat_config
.aggr_get_id
= perf_stat__get_core_cached
;
872 * The evsel_list->cpus is the base we operate on,
873 * taking the highest cpu number to be the size of
874 * the aggregation translate cpumap.
876 nr
= cpu_map__get_max(evsel_list
->cpus
);
877 stat_config
.cpus_aggr_map
= cpu_map__empty_new(nr
+ 1);
878 return stat_config
.cpus_aggr_map
? 0 : -ENOMEM
;
881 static void perf_stat__exit_aggr_mode(void)
883 cpu_map__put(stat_config
.aggr_map
);
884 cpu_map__put(stat_config
.cpus_aggr_map
);
885 stat_config
.aggr_map
= NULL
;
886 stat_config
.cpus_aggr_map
= NULL
;
889 static inline int perf_env__get_cpu(struct perf_env
*env
, struct cpu_map
*map
, int idx
)
898 if (cpu
>= env
->nr_cpus_avail
)
904 static int perf_env__get_socket(struct cpu_map
*map
, int idx
, void *data
)
906 struct perf_env
*env
= data
;
907 int cpu
= perf_env__get_cpu(env
, map
, idx
);
909 return cpu
== -1 ? -1 : env
->cpu
[cpu
].socket_id
;
912 static int perf_env__get_core(struct cpu_map
*map
, int idx
, void *data
)
914 struct perf_env
*env
= data
;
915 int core
= -1, cpu
= perf_env__get_cpu(env
, map
, idx
);
918 int socket_id
= env
->cpu
[cpu
].socket_id
;
921 * Encode socket in upper 16 bits
922 * core_id is relative to socket, and
923 * we need a global id. So we combine
926 core
= (socket_id
<< 16) | (env
->cpu
[cpu
].core_id
& 0xffff);
932 static int perf_env__build_socket_map(struct perf_env
*env
, struct cpu_map
*cpus
,
933 struct cpu_map
**sockp
)
935 return cpu_map__build_map(cpus
, sockp
, perf_env__get_socket
, env
);
938 static int perf_env__build_core_map(struct perf_env
*env
, struct cpu_map
*cpus
,
939 struct cpu_map
**corep
)
941 return cpu_map__build_map(cpus
, corep
, perf_env__get_core
, env
);
944 static int perf_stat__get_socket_file(struct perf_stat_config
*config __maybe_unused
,
945 struct cpu_map
*map
, int idx
)
947 return perf_env__get_socket(map
, idx
, &perf_stat
.session
->header
.env
);
950 static int perf_stat__get_core_file(struct perf_stat_config
*config __maybe_unused
,
951 struct cpu_map
*map
, int idx
)
953 return perf_env__get_core(map
, idx
, &perf_stat
.session
->header
.env
);
956 static int perf_stat_init_aggr_mode_file(struct perf_stat
*st
)
958 struct perf_env
*env
= &st
->session
->header
.env
;
960 switch (stat_config
.aggr_mode
) {
962 if (perf_env__build_socket_map(env
, evsel_list
->cpus
, &stat_config
.aggr_map
)) {
963 perror("cannot build socket map");
966 stat_config
.aggr_get_id
= perf_stat__get_socket_file
;
969 if (perf_env__build_core_map(env
, evsel_list
->cpus
, &stat_config
.aggr_map
)) {
970 perror("cannot build core map");
973 stat_config
.aggr_get_id
= perf_stat__get_core_file
;
986 static int topdown_filter_events(const char **attr
, char **str
, bool use_group
)
993 for (i
= 0; attr
[i
]; i
++) {
994 if (pmu_have_event("cpu", attr
[i
])) {
995 len
+= strlen(attr
[i
]) + 1;
996 attr
[i
- off
] = attr
[i
];
1000 attr
[i
- off
] = NULL
;
1002 *str
= malloc(len
+ 1 + 2);
1012 for (i
= 0; attr
[i
]; i
++) {
1025 __weak
bool arch_topdown_check_group(bool *warn
)
1031 __weak
void arch_topdown_group_warn(void)
1036 * Add default attributes, if there were no attributes specified or
1037 * if -d/--detailed, -d -d or -d -d -d is used:
1039 static int add_default_attributes(void)
1042 struct perf_event_attr default_attrs0
[] = {
1044 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1045 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1046 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1047 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1049 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1051 struct perf_event_attr frontend_attrs
[] = {
1052 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1054 struct perf_event_attr backend_attrs
[] = {
1055 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1057 struct perf_event_attr default_attrs1
[] = {
1058 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1059 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1060 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1065 * Detailed stats (-d), covering the L1 and last level data caches:
1067 struct perf_event_attr detailed_attrs
[] = {
1069 { .type
= PERF_TYPE_HW_CACHE
,
1071 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1072 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1073 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1075 { .type
= PERF_TYPE_HW_CACHE
,
1077 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1078 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1079 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1081 { .type
= PERF_TYPE_HW_CACHE
,
1083 PERF_COUNT_HW_CACHE_LL
<< 0 |
1084 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1085 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1087 { .type
= PERF_TYPE_HW_CACHE
,
1089 PERF_COUNT_HW_CACHE_LL
<< 0 |
1090 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1091 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1095 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1097 struct perf_event_attr very_detailed_attrs
[] = {
1099 { .type
= PERF_TYPE_HW_CACHE
,
1101 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1102 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1103 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1105 { .type
= PERF_TYPE_HW_CACHE
,
1107 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1108 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1109 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1111 { .type
= PERF_TYPE_HW_CACHE
,
1113 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1114 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1115 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1117 { .type
= PERF_TYPE_HW_CACHE
,
1119 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1120 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1121 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1123 { .type
= PERF_TYPE_HW_CACHE
,
1125 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1126 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1127 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1129 { .type
= PERF_TYPE_HW_CACHE
,
1131 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1132 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1133 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1138 * Very, very detailed stats (-d -d -d), adding prefetch events:
1140 struct perf_event_attr very_very_detailed_attrs
[] = {
1142 { .type
= PERF_TYPE_HW_CACHE
,
1144 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1145 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1148 { .type
= PERF_TYPE_HW_CACHE
,
1150 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1151 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1152 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1154 struct parse_events_error errinfo
;
1156 /* Set attrs if no event is selected and !null_run: */
1157 if (stat_config
.null_run
)
1160 if (transaction_run
) {
1161 /* Handle -T as -M transaction. Once platform specific metrics
1162 * support has been added to the json files, all archictures
1163 * will use this approach. To determine transaction support
1164 * on an architecture test for such a metric name.
1166 if (metricgroup__has_metric("transaction")) {
1167 struct option opt
= { .value
= &evsel_list
};
1169 return metricgroup__parse_groups(&opt
, "transaction",
1170 &stat_config
.metric_events
);
1173 if (pmu_have_event("cpu", "cycles-ct") &&
1174 pmu_have_event("cpu", "el-start"))
1175 err
= parse_events(evsel_list
, transaction_attrs
,
1178 err
= parse_events(evsel_list
,
1179 transaction_limited_attrs
,
1182 fprintf(stderr
, "Cannot set up transaction events\n");
1183 parse_events_print_error(&errinfo
, transaction_attrs
);
1192 if (sysfs__read_int(FREEZE_ON_SMI_PATH
, &smi
) < 0) {
1193 fprintf(stderr
, "freeze_on_smi is not supported.\n");
1198 if (sysfs__write_int(FREEZE_ON_SMI_PATH
, 1) < 0) {
1199 fprintf(stderr
, "Failed to set freeze_on_smi.\n");
1205 if (pmu_have_event("msr", "aperf") &&
1206 pmu_have_event("msr", "smi")) {
1207 if (!force_metric_only
)
1208 stat_config
.metric_only
= true;
1209 err
= parse_events(evsel_list
, smi_cost_attrs
, &errinfo
);
1211 fprintf(stderr
, "To measure SMI cost, it needs "
1212 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1213 parse_events_print_error(&errinfo
, smi_cost_attrs
);
1217 fprintf(stderr
, "Cannot set up SMI cost events\n");
1227 if (stat_config
.aggr_mode
!= AGGR_GLOBAL
&&
1228 stat_config
.aggr_mode
!= AGGR_CORE
) {
1229 pr_err("top down event configuration requires --per-core mode\n");
1232 stat_config
.aggr_mode
= AGGR_CORE
;
1233 if (nr_cgroups
|| !target__has_cpu(&target
)) {
1234 pr_err("top down event configuration requires system-wide mode (-a)\n");
1238 if (!force_metric_only
)
1239 stat_config
.metric_only
= true;
1240 if (topdown_filter_events(topdown_attrs
, &str
,
1241 arch_topdown_check_group(&warn
)) < 0) {
1242 pr_err("Out of memory\n");
1245 if (topdown_attrs
[0] && str
) {
1247 arch_topdown_group_warn();
1248 err
= parse_events(evsel_list
, str
, &errinfo
);
1251 "Cannot set up top down events %s: %d\n",
1254 parse_events_print_error(&errinfo
, str
);
1258 fprintf(stderr
, "System does not support topdown\n");
1264 if (!evsel_list
->nr_entries
) {
1265 if (target__has_cpu(&target
))
1266 default_attrs0
[0].config
= PERF_COUNT_SW_CPU_CLOCK
;
1268 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs0
) < 0)
1270 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1271 if (perf_evlist__add_default_attrs(evsel_list
,
1272 frontend_attrs
) < 0)
1275 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1276 if (perf_evlist__add_default_attrs(evsel_list
,
1280 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs1
) < 0)
1284 /* Detailed events get appended to the event list: */
1286 if (detailed_run
< 1)
1289 /* Append detailed run extra attributes: */
1290 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1293 if (detailed_run
< 2)
1296 /* Append very detailed run extra attributes: */
1297 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1300 if (detailed_run
< 3)
1303 /* Append very, very detailed run extra attributes: */
1304 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1307 static const char * const stat_record_usage
[] = {
1308 "perf stat record [<options>]",
1312 static void init_features(struct perf_session
*session
)
1316 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
1317 perf_header__set_feat(&session
->header
, feat
);
1319 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
1320 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
1321 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
1322 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
1325 static int __cmd_record(int argc
, const char **argv
)
1327 struct perf_session
*session
;
1328 struct perf_data
*data
= &perf_stat
.data
;
1330 argc
= parse_options(argc
, argv
, stat_options
, stat_record_usage
,
1331 PARSE_OPT_STOP_AT_NON_OPTION
);
1334 data
->file
.path
= output_name
;
1336 if (stat_config
.run_count
!= 1 || forever
) {
1337 pr_err("Cannot use -r option with perf stat record.\n");
1341 session
= perf_session__new(data
, false, NULL
);
1342 if (session
== NULL
) {
1343 pr_err("Perf session creation failed.\n");
1347 init_features(session
);
1349 session
->evlist
= evsel_list
;
1350 perf_stat
.session
= session
;
1351 perf_stat
.record
= true;
1355 static int process_stat_round_event(struct perf_session
*session
,
1356 union perf_event
*event
)
1358 struct stat_round_event
*stat_round
= &event
->stat_round
;
1359 struct perf_evsel
*counter
;
1360 struct timespec tsh
, *ts
= NULL
;
1361 const char **argv
= session
->header
.env
.cmdline_argv
;
1362 int argc
= session
->header
.env
.nr_cmdline
;
1364 evlist__for_each_entry(evsel_list
, counter
)
1365 perf_stat_process_counter(&stat_config
, counter
);
1367 if (stat_round
->type
== PERF_STAT_ROUND_TYPE__FINAL
)
1368 update_stats(&walltime_nsecs_stats
, stat_round
->time
);
1370 if (stat_config
.interval
&& stat_round
->time
) {
1371 tsh
.tv_sec
= stat_round
->time
/ NSEC_PER_SEC
;
1372 tsh
.tv_nsec
= stat_round
->time
% NSEC_PER_SEC
;
1376 print_counters(ts
, argc
, argv
);
1381 int process_stat_config_event(struct perf_session
*session
,
1382 union perf_event
*event
)
1384 struct perf_tool
*tool
= session
->tool
;
1385 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
1387 perf_event__read_stat_config(&stat_config
, &event
->stat_config
);
1389 if (cpu_map__empty(st
->cpus
)) {
1390 if (st
->aggr_mode
!= AGGR_UNSET
)
1391 pr_warning("warning: processing task data, aggregation mode not set\n");
1395 if (st
->aggr_mode
!= AGGR_UNSET
)
1396 stat_config
.aggr_mode
= st
->aggr_mode
;
1398 if (perf_stat
.data
.is_pipe
)
1399 perf_stat_init_aggr_mode();
1401 perf_stat_init_aggr_mode_file(st
);
1406 static int set_maps(struct perf_stat
*st
)
1408 if (!st
->cpus
|| !st
->threads
)
1411 if (WARN_ONCE(st
->maps_allocated
, "stats double allocation\n"))
1414 perf_evlist__set_maps(evsel_list
, st
->cpus
, st
->threads
);
1416 if (perf_evlist__alloc_stats(evsel_list
, true))
1419 st
->maps_allocated
= true;
1424 int process_thread_map_event(struct perf_session
*session
,
1425 union perf_event
*event
)
1427 struct perf_tool
*tool
= session
->tool
;
1428 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
1431 pr_warning("Extra thread map event, ignoring.\n");
1435 st
->threads
= thread_map__new_event(&event
->thread_map
);
1439 return set_maps(st
);
1443 int process_cpu_map_event(struct perf_session
*session
,
1444 union perf_event
*event
)
1446 struct perf_tool
*tool
= session
->tool
;
1447 struct perf_stat
*st
= container_of(tool
, struct perf_stat
, tool
);
1448 struct cpu_map
*cpus
;
1451 pr_warning("Extra cpu map event, ignoring.\n");
1455 cpus
= cpu_map__new_data(&event
->cpu_map
.data
);
1460 return set_maps(st
);
1463 static int runtime_stat_new(struct perf_stat_config
*config
, int nthreads
)
1467 config
->stats
= calloc(nthreads
, sizeof(struct runtime_stat
));
1471 config
->stats_num
= nthreads
;
1473 for (i
= 0; i
< nthreads
; i
++)
1474 runtime_stat__init(&config
->stats
[i
]);
1479 static void runtime_stat_delete(struct perf_stat_config
*config
)
1486 for (i
= 0; i
< config
->stats_num
; i
++)
1487 runtime_stat__exit(&config
->stats
[i
]);
1489 free(config
->stats
);
1492 static const char * const stat_report_usage
[] = {
1493 "perf stat report [<options>]",
1497 static struct perf_stat perf_stat
= {
1499 .attr
= perf_event__process_attr
,
1500 .event_update
= perf_event__process_event_update
,
1501 .thread_map
= process_thread_map_event
,
1502 .cpu_map
= process_cpu_map_event
,
1503 .stat_config
= process_stat_config_event
,
1504 .stat
= perf_event__process_stat_event
,
1505 .stat_round
= process_stat_round_event
,
1507 .aggr_mode
= AGGR_UNSET
,
1510 static int __cmd_report(int argc
, const char **argv
)
1512 struct perf_session
*session
;
1513 const struct option options
[] = {
1514 OPT_STRING('i', "input", &input_name
, "file", "input file name"),
1515 OPT_SET_UINT(0, "per-socket", &perf_stat
.aggr_mode
,
1516 "aggregate counts per processor socket", AGGR_SOCKET
),
1517 OPT_SET_UINT(0, "per-core", &perf_stat
.aggr_mode
,
1518 "aggregate counts per physical processor core", AGGR_CORE
),
1519 OPT_SET_UINT('A', "no-aggr", &perf_stat
.aggr_mode
,
1520 "disable CPU count aggregation", AGGR_NONE
),
1526 argc
= parse_options(argc
, argv
, options
, stat_report_usage
, 0);
1528 if (!input_name
|| !strlen(input_name
)) {
1529 if (!fstat(STDIN_FILENO
, &st
) && S_ISFIFO(st
.st_mode
))
1532 input_name
= "perf.data";
1535 perf_stat
.data
.file
.path
= input_name
;
1536 perf_stat
.data
.mode
= PERF_DATA_MODE_READ
;
1538 session
= perf_session__new(&perf_stat
.data
, false, &perf_stat
.tool
);
1539 if (session
== NULL
)
1542 perf_stat
.session
= session
;
1543 stat_config
.output
= stderr
;
1544 evsel_list
= session
->evlist
;
1546 ret
= perf_session__process_events(session
);
1550 perf_session__delete(session
);
1554 static void setup_system_wide(int forks
)
1557 * Make system wide (-a) the default target if
1558 * no target was specified and one of following
1559 * conditions is met:
1561 * - there's no workload specified
1562 * - there is workload specified but all requested
1563 * events are system wide events
1565 if (!target__none(&target
))
1569 target
.system_wide
= true;
1571 struct perf_evsel
*counter
;
1573 evlist__for_each_entry(evsel_list
, counter
) {
1574 if (!counter
->system_wide
)
1578 if (evsel_list
->nr_entries
)
1579 target
.system_wide
= true;
1583 int cmd_stat(int argc
, const char **argv
)
1585 const char * const stat_usage
[] = {
1586 "perf stat [<options>] [<command>]",
1589 int status
= -EINVAL
, run_idx
;
1591 FILE *output
= stderr
;
1592 unsigned int interval
, timeout
;
1593 const char * const stat_subcommands
[] = { "record", "report" };
1595 setlocale(LC_ALL
, "");
1597 evsel_list
= perf_evlist__new();
1598 if (evsel_list
== NULL
)
1601 parse_events__shrink_config_terms();
1603 /* String-parsing callback-based options would segfault when negated */
1604 set_option_flag(stat_options
, 'e', "event", PARSE_OPT_NONEG
);
1605 set_option_flag(stat_options
, 'M', "metrics", PARSE_OPT_NONEG
);
1606 set_option_flag(stat_options
, 'G', "cgroup", PARSE_OPT_NONEG
);
1608 argc
= parse_options_subcommand(argc
, argv
, stat_options
, stat_subcommands
,
1609 (const char **) stat_usage
,
1610 PARSE_OPT_STOP_AT_NON_OPTION
);
1611 perf_stat__collect_metric_expr(evsel_list
);
1612 perf_stat__init_shadow_stats();
1614 if (stat_config
.csv_sep
) {
1615 stat_config
.csv_output
= true;
1616 if (!strcmp(stat_config
.csv_sep
, "\\t"))
1617 stat_config
.csv_sep
= "\t";
1619 stat_config
.csv_sep
= DEFAULT_SEPARATOR
;
1621 if (argc
&& !strncmp(argv
[0], "rec", 3)) {
1622 argc
= __cmd_record(argc
, argv
);
1625 } else if (argc
&& !strncmp(argv
[0], "rep", 3))
1626 return __cmd_report(argc
, argv
);
1628 interval
= stat_config
.interval
;
1629 timeout
= stat_config
.timeout
;
1632 * For record command the -o is already taken care of.
1634 if (!STAT_RECORD
&& output_name
&& strcmp(output_name
, "-"))
1637 if (output_name
&& output_fd
) {
1638 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1639 parse_options_usage(stat_usage
, stat_options
, "o", 1);
1640 parse_options_usage(NULL
, stat_options
, "log-fd", 0);
1644 if (stat_config
.metric_only
&& stat_config
.aggr_mode
== AGGR_THREAD
) {
1645 fprintf(stderr
, "--metric-only is not supported with --per-thread\n");
1649 if (stat_config
.metric_only
&& stat_config
.run_count
> 1) {
1650 fprintf(stderr
, "--metric-only is not supported with -r\n");
1654 if (stat_config
.walltime_run_table
&& stat_config
.run_count
<= 1) {
1655 fprintf(stderr
, "--table is only supported with -r\n");
1656 parse_options_usage(stat_usage
, stat_options
, "r", 1);
1657 parse_options_usage(NULL
, stat_options
, "table", 0);
1661 if (output_fd
< 0) {
1662 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1663 parse_options_usage(stat_usage
, stat_options
, "log-fd", 0);
1669 mode
= append_file
? "a" : "w";
1671 output
= fopen(output_name
, mode
);
1673 perror("failed to create output file");
1676 clock_gettime(CLOCK_REALTIME
, &tm
);
1677 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1678 } else if (output_fd
> 0) {
1679 mode
= append_file
? "a" : "w";
1680 output
= fdopen(output_fd
, mode
);
1682 perror("Failed opening logfd");
1687 stat_config
.output
= output
;
1690 * let the spreadsheet do the pretty-printing
1692 if (stat_config
.csv_output
) {
1693 /* User explicitly passed -B? */
1694 if (big_num_opt
== 1) {
1695 fprintf(stderr
, "-B option not supported with -x\n");
1696 parse_options_usage(stat_usage
, stat_options
, "B", 1);
1697 parse_options_usage(NULL
, stat_options
, "x", 1);
1699 } else /* Nope, so disable big number formatting */
1700 stat_config
.big_num
= false;
1701 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1702 stat_config
.big_num
= false;
1704 setup_system_wide(argc
);
1707 * Display user/system times only for single
1708 * run and when there's specified tracee.
1710 if ((stat_config
.run_count
== 1) && target__none(&target
))
1711 stat_config
.ru_display
= true;
1713 if (stat_config
.run_count
< 0) {
1714 pr_err("Run count must be a positive number\n");
1715 parse_options_usage(stat_usage
, stat_options
, "r", 1);
1717 } else if (stat_config
.run_count
== 0) {
1719 stat_config
.run_count
= 1;
1722 if (stat_config
.walltime_run_table
) {
1723 stat_config
.walltime_run
= zalloc(stat_config
.run_count
* sizeof(stat_config
.walltime_run
[0]));
1724 if (!stat_config
.walltime_run
) {
1725 pr_err("failed to setup -r option");
1730 if ((stat_config
.aggr_mode
== AGGR_THREAD
) &&
1731 !target__has_task(&target
)) {
1732 if (!target
.system_wide
|| target
.cpu_list
) {
1733 fprintf(stderr
, "The --per-thread option is only "
1734 "available when monitoring via -p -t -a "
1735 "options or only --per-thread.\n");
1736 parse_options_usage(NULL
, stat_options
, "p", 1);
1737 parse_options_usage(NULL
, stat_options
, "t", 1);
1743 * no_aggr, cgroup are for system-wide only
1744 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1746 if (((stat_config
.aggr_mode
!= AGGR_GLOBAL
&&
1747 stat_config
.aggr_mode
!= AGGR_THREAD
) || nr_cgroups
) &&
1748 !target__has_cpu(&target
)) {
1749 fprintf(stderr
, "both cgroup and no-aggregation "
1750 "modes only available in system-wide mode\n");
1752 parse_options_usage(stat_usage
, stat_options
, "G", 1);
1753 parse_options_usage(NULL
, stat_options
, "A", 1);
1754 parse_options_usage(NULL
, stat_options
, "a", 1);
1758 if (add_default_attributes())
1761 target__validate(&target
);
1763 if ((stat_config
.aggr_mode
== AGGR_THREAD
) && (target
.system_wide
))
1764 target
.per_thread
= true;
1766 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1767 if (target__has_task(&target
)) {
1768 pr_err("Problems finding threads of monitor\n");
1769 parse_options_usage(stat_usage
, stat_options
, "p", 1);
1770 parse_options_usage(NULL
, stat_options
, "t", 1);
1771 } else if (target__has_cpu(&target
)) {
1772 perror("failed to parse CPUs map");
1773 parse_options_usage(stat_usage
, stat_options
, "C", 1);
1774 parse_options_usage(NULL
, stat_options
, "a", 1);
1780 * Initialize thread_map with comm names,
1781 * so we could print it out on output.
1783 if (stat_config
.aggr_mode
== AGGR_THREAD
) {
1784 thread_map__read_comms(evsel_list
->threads
);
1785 if (target
.system_wide
) {
1786 if (runtime_stat_new(&stat_config
,
1787 thread_map__nr(evsel_list
->threads
))) {
1793 if (stat_config
.times
&& interval
)
1794 interval_count
= true;
1795 else if (stat_config
.times
&& !interval
) {
1796 pr_err("interval-count option should be used together with "
1797 "interval-print.\n");
1798 parse_options_usage(stat_usage
, stat_options
, "interval-count", 0);
1799 parse_options_usage(stat_usage
, stat_options
, "I", 1);
1803 if (timeout
&& timeout
< 100) {
1805 pr_err("timeout must be >= 10ms.\n");
1806 parse_options_usage(stat_usage
, stat_options
, "timeout", 0);
1809 pr_warning("timeout < 100ms. "
1810 "The overhead percentage could be high in some cases. "
1811 "Please proceed with caution.\n");
1813 if (timeout
&& interval
) {
1814 pr_err("timeout option is not supported with interval-print.\n");
1815 parse_options_usage(stat_usage
, stat_options
, "timeout", 0);
1816 parse_options_usage(stat_usage
, stat_options
, "I", 1);
1820 if (perf_evlist__alloc_stats(evsel_list
, interval
))
1823 if (perf_stat_init_aggr_mode())
1827 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1828 * while avoiding that older tools show confusing messages.
1830 * However for pipe sessions we need to keep it zero,
1831 * because script's perf_evsel__check_attr is triggered
1832 * by attr->sample_type != 0, and we can't run it on
1835 stat_config
.identifier
= !(STAT_RECORD
&& perf_stat
.data
.is_pipe
);
1838 * We dont want to block the signals - that would cause
1839 * child tasks to inherit that and Ctrl-C would not work.
1840 * What we want is for Ctrl-C to work in the exec()-ed
1841 * task, but being ignored by perf stat itself:
1845 signal(SIGINT
, skip_signal
);
1846 signal(SIGCHLD
, skip_signal
);
1847 signal(SIGALRM
, skip_signal
);
1848 signal(SIGABRT
, skip_signal
);
1851 for (run_idx
= 0; forever
|| run_idx
< stat_config
.run_count
; run_idx
++) {
1852 if (stat_config
.run_count
!= 1 && verbose
> 0)
1853 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1856 status
= run_perf_stat(argc
, argv
, run_idx
);
1857 if (forever
&& status
!= -1) {
1858 print_counters(NULL
, argc
, argv
);
1859 perf_stat__reset_stats();
1863 if (!forever
&& status
!= -1 && !interval
)
1864 print_counters(NULL
, argc
, argv
);
1868 * We synthesize the kernel mmap record just so that older tools
1869 * don't emit warnings about not being able to resolve symbols
1870 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1871 * a saner message about no samples being in the perf.data file.
1873 * This also serves to suppress a warning about f_header.data.size == 0
1874 * in header.c at the moment 'perf stat record' gets introduced, which
1875 * is not really needed once we start adding the stat specific PERF_RECORD_
1876 * records, but the need to suppress the kptr_restrict messages in older
1877 * tools remain -acme
1879 int fd
= perf_data__fd(&perf_stat
.data
);
1880 int err
= perf_event__synthesize_kernel_mmap((void *)&perf_stat
,
1881 process_synthesized_event
,
1882 &perf_stat
.session
->machines
.host
);
1884 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1885 "older tools may produce warnings about this file\n.");
1889 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats
.max
, FINAL
))
1890 pr_err("failed to write stat round event\n");
1893 if (!perf_stat
.data
.is_pipe
) {
1894 perf_stat
.session
->header
.data_size
+= perf_stat
.bytes_written
;
1895 perf_session__write_header(perf_stat
.session
, evsel_list
, fd
, true);
1898 perf_session__delete(perf_stat
.session
);
1901 perf_stat__exit_aggr_mode();
1902 perf_evlist__free_stats(evsel_list
);
1904 free(stat_config
.walltime_run
);
1906 if (smi_cost
&& smi_reset
)
1907 sysfs__write_int(FREEZE_ON_SMI_PATH
, 0);
1909 perf_evlist__delete(evsel_list
);
1911 runtime_stat_delete(&stat_config
);