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/util.h"
47 #include "util/parse-options.h"
48 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
61 #include <sys/prctl.h>
64 #define DEFAULT_SEPARATOR " "
65 #define CNTR_NOT_SUPPORTED "<not supported>"
66 #define CNTR_NOT_COUNTED "<not counted>"
68 static void print_stat(int argc
, const char **argv
);
69 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
);
70 static void print_counter(struct perf_evsel
*counter
, char *prefix
);
71 static void print_aggr(char *prefix
);
73 static struct perf_evlist
*evsel_list
;
75 static struct perf_target target
= {
86 static int run_count
= 1;
87 static bool no_inherit
= false;
88 static bool scale
= true;
89 static enum aggr_mode aggr_mode
= AGGR_GLOBAL
;
90 static volatile pid_t child_pid
= -1;
91 static bool null_run
= false;
92 static int detailed_run
= 0;
93 static bool big_num
= true;
94 static int big_num_opt
= -1;
95 static const char *csv_sep
= NULL
;
96 static bool csv_output
= false;
97 static bool group
= false;
98 static FILE *output
= NULL
;
99 static const char *pre_cmd
= NULL
;
100 static const char *post_cmd
= NULL
;
101 static bool sync_run
= false;
102 static unsigned int interval
= 0;
103 static bool forever
= false;
104 static struct timespec ref_time
;
105 static struct cpu_map
*aggr_map
;
106 static int (*aggr_get_id
)(struct cpu_map
*m
, int cpu
);
108 static volatile int done
= 0;
111 struct stats res_stats
[3];
114 static inline void diff_timespec(struct timespec
*r
, struct timespec
*a
,
117 r
->tv_sec
= a
->tv_sec
- b
->tv_sec
;
118 if (a
->tv_nsec
< b
->tv_nsec
) {
119 r
->tv_nsec
= a
->tv_nsec
+ 1000000000L - b
->tv_nsec
;
122 r
->tv_nsec
= a
->tv_nsec
- b
->tv_nsec
;
126 static inline struct cpu_map
*perf_evsel__cpus(struct perf_evsel
*evsel
)
128 return (evsel
->cpus
&& !target
.cpu_list
) ? evsel
->cpus
: evsel_list
->cpus
;
131 static inline int perf_evsel__nr_cpus(struct perf_evsel
*evsel
)
133 return perf_evsel__cpus(evsel
)->nr
;
136 static void perf_evsel__reset_stat_priv(struct perf_evsel
*evsel
)
138 memset(evsel
->priv
, 0, sizeof(struct perf_stat
));
141 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
143 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
144 return evsel
->priv
== NULL
? -ENOMEM
: 0;
147 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
153 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel
*evsel
)
158 sz
= sizeof(*evsel
->counts
) +
159 (perf_evsel__nr_cpus(evsel
) * sizeof(struct perf_counts_values
));
165 evsel
->prev_raw_counts
= addr
;
170 static void perf_evsel__free_prev_raw_counts(struct perf_evsel
*evsel
)
172 free(evsel
->prev_raw_counts
);
173 evsel
->prev_raw_counts
= NULL
;
176 static void perf_evlist__free_stats(struct perf_evlist
*evlist
)
178 struct perf_evsel
*evsel
;
180 list_for_each_entry(evsel
, &evlist
->entries
, node
) {
181 perf_evsel__free_stat_priv(evsel
);
182 perf_evsel__free_counts(evsel
);
183 perf_evsel__free_prev_raw_counts(evsel
);
187 static int perf_evlist__alloc_stats(struct perf_evlist
*evlist
, bool alloc_raw
)
189 struct perf_evsel
*evsel
;
191 list_for_each_entry(evsel
, &evlist
->entries
, node
) {
192 if (perf_evsel__alloc_stat_priv(evsel
) < 0 ||
193 perf_evsel__alloc_counts(evsel
, perf_evsel__nr_cpus(evsel
)) < 0 ||
194 (alloc_raw
&& perf_evsel__alloc_prev_raw_counts(evsel
) < 0))
201 perf_evlist__free_stats(evlist
);
205 static struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
206 static struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
207 static struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
208 static struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
209 static struct stats runtime_branches_stats
[MAX_NR_CPUS
];
210 static struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
211 static struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
212 static struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
213 static struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
214 static struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
215 static struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
216 static struct stats walltime_nsecs_stats
;
218 static void perf_stat__reset_stats(struct perf_evlist
*evlist
)
220 struct perf_evsel
*evsel
;
222 list_for_each_entry(evsel
, &evlist
->entries
, node
) {
223 perf_evsel__reset_stat_priv(evsel
);
224 perf_evsel__reset_counts(evsel
, perf_evsel__nr_cpus(evsel
));
227 memset(runtime_nsecs_stats
, 0, sizeof(runtime_nsecs_stats
));
228 memset(runtime_cycles_stats
, 0, sizeof(runtime_cycles_stats
));
229 memset(runtime_stalled_cycles_front_stats
, 0, sizeof(runtime_stalled_cycles_front_stats
));
230 memset(runtime_stalled_cycles_back_stats
, 0, sizeof(runtime_stalled_cycles_back_stats
));
231 memset(runtime_branches_stats
, 0, sizeof(runtime_branches_stats
));
232 memset(runtime_cacherefs_stats
, 0, sizeof(runtime_cacherefs_stats
));
233 memset(runtime_l1_dcache_stats
, 0, sizeof(runtime_l1_dcache_stats
));
234 memset(runtime_l1_icache_stats
, 0, sizeof(runtime_l1_icache_stats
));
235 memset(runtime_ll_cache_stats
, 0, sizeof(runtime_ll_cache_stats
));
236 memset(runtime_itlb_cache_stats
, 0, sizeof(runtime_itlb_cache_stats
));
237 memset(runtime_dtlb_cache_stats
, 0, sizeof(runtime_dtlb_cache_stats
));
238 memset(&walltime_nsecs_stats
, 0, sizeof(walltime_nsecs_stats
));
241 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
243 struct perf_event_attr
*attr
= &evsel
->attr
;
246 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
247 PERF_FORMAT_TOTAL_TIME_RUNNING
;
249 attr
->inherit
= !no_inherit
;
251 if (perf_target__has_cpu(&target
))
252 return perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
254 if (!perf_target__has_task(&target
) &&
255 perf_evsel__is_group_leader(evsel
)) {
257 attr
->enable_on_exec
= 1;
260 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
264 * Does the counter have nsecs as a unit?
266 static inline int nsec_counter(struct perf_evsel
*evsel
)
268 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
269 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
276 * Update various tracking values we maintain to print
277 * more semantic information such as miss/hit ratios,
278 * instruction rates, etc:
280 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
282 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
283 update_stats(&runtime_nsecs_stats
[0], count
[0]);
284 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
285 update_stats(&runtime_cycles_stats
[0], count
[0]);
286 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
287 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
288 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
289 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
290 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
291 update_stats(&runtime_branches_stats
[0], count
[0]);
292 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
293 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
294 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
295 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
296 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
297 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
298 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
299 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
300 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
301 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
302 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
303 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
307 * Read out the results of a single counter:
308 * aggregate counts across CPUs in system-wide mode
310 static int read_counter_aggr(struct perf_evsel
*counter
)
312 struct perf_stat
*ps
= counter
->priv
;
313 u64
*count
= counter
->counts
->aggr
.values
;
316 if (__perf_evsel__read(counter
, perf_evsel__nr_cpus(counter
),
317 thread_map__nr(evsel_list
->threads
), scale
) < 0)
320 for (i
= 0; i
< 3; i
++)
321 update_stats(&ps
->res_stats
[i
], count
[i
]);
324 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
325 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
329 * Save the full runtime - to allow normalization during printout:
331 update_shadow_stats(counter
, count
);
337 * Read out the results of a single counter:
338 * do not aggregate counts across CPUs in system-wide mode
340 static int read_counter(struct perf_evsel
*counter
)
345 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
346 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
349 count
= counter
->counts
->cpu
[cpu
].values
;
351 update_shadow_stats(counter
, count
);
357 static void print_interval(void)
359 static int num_print_interval
;
360 struct perf_evsel
*counter
;
361 struct perf_stat
*ps
;
362 struct timespec ts
, rs
;
365 if (aggr_mode
== AGGR_GLOBAL
) {
366 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
368 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
369 read_counter_aggr(counter
);
372 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
374 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
375 read_counter(counter
);
379 clock_gettime(CLOCK_MONOTONIC
, &ts
);
380 diff_timespec(&rs
, &ts
, &ref_time
);
381 sprintf(prefix
, "%6lu.%09lu%s", rs
.tv_sec
, rs
.tv_nsec
, csv_sep
);
383 if (num_print_interval
== 0 && !csv_output
) {
386 fprintf(output
, "# time socket cpus counts events\n");
389 fprintf(output
, "# time core cpus counts events\n");
392 fprintf(output
, "# time CPU counts events\n");
396 fprintf(output
, "# time counts events\n");
400 if (++num_print_interval
== 25)
401 num_print_interval
= 0;
409 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
410 print_counter(counter
, prefix
);
414 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
415 print_counter_aggr(counter
, prefix
);
419 static int __run_perf_stat(int argc
, const char **argv
)
422 unsigned long long t0
, t1
;
423 struct perf_evsel
*counter
;
426 const bool forks
= (argc
> 0);
429 ts
.tv_sec
= interval
/ 1000;
430 ts
.tv_nsec
= (interval
% 1000) * 1000000;
437 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
,
439 perror("failed to prepare workload");
445 perf_evlist__set_leader(evsel_list
);
447 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
448 if (create_perf_stat_counter(counter
) < 0) {
450 * PPC returns ENXIO for HW counters until 2.6.37
451 * (behavior changed with commit b0a873e).
453 if (errno
== EINVAL
|| errno
== ENOSYS
||
454 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
457 ui__warning("%s event is not supported by the kernel.\n",
458 perf_evsel__name(counter
));
459 counter
->supported
= false;
463 perf_evsel__open_strerror(counter
, &target
,
464 errno
, msg
, sizeof(msg
));
465 ui__error("%s\n", msg
);
468 kill(child_pid
, SIGTERM
);
472 counter
->supported
= true;
475 if (perf_evlist__apply_filters(evsel_list
)) {
476 error("failed to set filter with %d (%s)\n", errno
,
482 * Enable counters and exec the command:
485 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
488 perf_evlist__start_workload(evsel_list
);
491 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
492 nanosleep(&ts
, NULL
);
497 if (WIFSIGNALED(status
))
498 psignal(WTERMSIG(status
), argv
[0]);
501 nanosleep(&ts
, NULL
);
509 update_stats(&walltime_nsecs_stats
, t1
- t0
);
511 if (aggr_mode
== AGGR_GLOBAL
) {
512 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
513 read_counter_aggr(counter
);
514 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
515 thread_map__nr(evsel_list
->threads
));
518 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
519 read_counter(counter
);
520 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
), 1);
524 return WEXITSTATUS(status
);
527 static int run_perf_stat(int argc __maybe_unused
, const char **argv
)
532 ret
= system(pre_cmd
);
540 ret
= __run_perf_stat(argc
, argv
);
545 ret
= system(post_cmd
);
553 static void print_noise_pct(double total
, double avg
)
555 double pct
= rel_stddev_stats(total
, avg
);
558 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
560 fprintf(output
, " ( +-%6.2f%% )", pct
);
563 static void print_noise(struct perf_evsel
*evsel
, double avg
)
565 struct perf_stat
*ps
;
571 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
574 static void aggr_printout(struct perf_evsel
*evsel
, int id
, int nr
)
578 fprintf(output
, "S%d-C%*d%s%*d%s",
579 cpu_map__id_to_socket(id
),
581 cpu_map__id_to_cpu(id
),
588 fprintf(output
, "S%*d%s%*d%s",
597 fprintf(output
, "CPU%*d%s",
599 perf_evsel__cpus(evsel
)->map
[id
], csv_sep
);
607 static void nsec_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
609 double msecs
= avg
/ 1e6
;
610 const char *fmt
= csv_output
? "%.6f%s%s" : "%18.6f%s%-25s";
612 aggr_printout(evsel
, cpu
, nr
);
614 fprintf(output
, fmt
, msecs
, csv_sep
, perf_evsel__name(evsel
));
617 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
619 if (csv_output
|| interval
)
622 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
623 fprintf(output
, " # %8.3f CPUs utilized ",
624 avg
/ avg_stats(&walltime_nsecs_stats
));
626 fprintf(output
, " ");
629 /* used for get_ratio_color() */
631 GRC_STALLED_CYCLES_FE
,
632 GRC_STALLED_CYCLES_BE
,
637 static const char *get_ratio_color(enum grc_type type
, double ratio
)
639 static const double grc_table
[GRC_MAX_NR
][3] = {
640 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
641 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
642 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
644 const char *color
= PERF_COLOR_NORMAL
;
646 if (ratio
> grc_table
[type
][0])
647 color
= PERF_COLOR_RED
;
648 else if (ratio
> grc_table
[type
][1])
649 color
= PERF_COLOR_MAGENTA
;
650 else if (ratio
> grc_table
[type
][2])
651 color
= PERF_COLOR_YELLOW
;
656 static void print_stalled_cycles_frontend(int cpu
,
657 struct perf_evsel
*evsel
658 __maybe_unused
, double avg
)
660 double total
, ratio
= 0.0;
663 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
666 ratio
= avg
/ total
* 100.0;
668 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
670 fprintf(output
, " # ");
671 color_fprintf(output
, color
, "%6.2f%%", ratio
);
672 fprintf(output
, " frontend cycles idle ");
675 static void print_stalled_cycles_backend(int cpu
,
676 struct perf_evsel
*evsel
677 __maybe_unused
, double avg
)
679 double total
, ratio
= 0.0;
682 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
685 ratio
= avg
/ total
* 100.0;
687 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
689 fprintf(output
, " # ");
690 color_fprintf(output
, color
, "%6.2f%%", ratio
);
691 fprintf(output
, " backend cycles idle ");
694 static void print_branch_misses(int cpu
,
695 struct perf_evsel
*evsel __maybe_unused
,
698 double total
, ratio
= 0.0;
701 total
= avg_stats(&runtime_branches_stats
[cpu
]);
704 ratio
= avg
/ total
* 100.0;
706 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
708 fprintf(output
, " # ");
709 color_fprintf(output
, color
, "%6.2f%%", ratio
);
710 fprintf(output
, " of all branches ");
713 static void print_l1_dcache_misses(int cpu
,
714 struct perf_evsel
*evsel __maybe_unused
,
717 double total
, ratio
= 0.0;
720 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
723 ratio
= avg
/ total
* 100.0;
725 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
727 fprintf(output
, " # ");
728 color_fprintf(output
, color
, "%6.2f%%", ratio
);
729 fprintf(output
, " of all L1-dcache hits ");
732 static void print_l1_icache_misses(int cpu
,
733 struct perf_evsel
*evsel __maybe_unused
,
736 double total
, ratio
= 0.0;
739 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
742 ratio
= avg
/ total
* 100.0;
744 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
746 fprintf(output
, " # ");
747 color_fprintf(output
, color
, "%6.2f%%", ratio
);
748 fprintf(output
, " of all L1-icache hits ");
751 static void print_dtlb_cache_misses(int cpu
,
752 struct perf_evsel
*evsel __maybe_unused
,
755 double total
, ratio
= 0.0;
758 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
761 ratio
= avg
/ total
* 100.0;
763 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
765 fprintf(output
, " # ");
766 color_fprintf(output
, color
, "%6.2f%%", ratio
);
767 fprintf(output
, " of all dTLB cache hits ");
770 static void print_itlb_cache_misses(int cpu
,
771 struct perf_evsel
*evsel __maybe_unused
,
774 double total
, ratio
= 0.0;
777 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
780 ratio
= avg
/ total
* 100.0;
782 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
784 fprintf(output
, " # ");
785 color_fprintf(output
, color
, "%6.2f%%", ratio
);
786 fprintf(output
, " of all iTLB cache hits ");
789 static void print_ll_cache_misses(int cpu
,
790 struct perf_evsel
*evsel __maybe_unused
,
793 double total
, ratio
= 0.0;
796 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
799 ratio
= avg
/ total
* 100.0;
801 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
803 fprintf(output
, " # ");
804 color_fprintf(output
, color
, "%6.2f%%", ratio
);
805 fprintf(output
, " of all LL-cache hits ");
808 static void abs_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
810 double total
, ratio
= 0.0;
816 fmt
= "%'18.0f%s%-25s";
818 fmt
= "%18.0f%s%-25s";
820 aggr_printout(evsel
, cpu
, nr
);
822 if (aggr_mode
== AGGR_GLOBAL
)
825 fprintf(output
, fmt
, avg
, csv_sep
, perf_evsel__name(evsel
));
828 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
830 if (csv_output
|| interval
)
833 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
834 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
838 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
840 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
841 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
845 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
848 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
849 runtime_branches_stats
[cpu
].n
!= 0) {
850 print_branch_misses(cpu
, evsel
, avg
);
852 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
853 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
854 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
855 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
856 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
857 print_l1_dcache_misses(cpu
, evsel
, avg
);
859 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
860 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
861 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
862 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
863 runtime_l1_icache_stats
[cpu
].n
!= 0) {
864 print_l1_icache_misses(cpu
, evsel
, avg
);
866 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
867 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
868 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
869 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
870 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
871 print_dtlb_cache_misses(cpu
, evsel
, avg
);
873 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
874 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
875 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
876 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
877 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
878 print_itlb_cache_misses(cpu
, evsel
, avg
);
880 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
881 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
882 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
883 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
884 runtime_ll_cache_stats
[cpu
].n
!= 0) {
885 print_ll_cache_misses(cpu
, evsel
, avg
);
886 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
887 runtime_cacherefs_stats
[cpu
].n
!= 0) {
888 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
891 ratio
= avg
* 100 / total
;
893 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
895 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
896 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
897 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
898 print_stalled_cycles_backend(cpu
, evsel
, avg
);
899 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
900 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
903 ratio
= 1.0 * avg
/ total
;
905 fprintf(output
, " # %8.3f GHz ", ratio
);
906 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
909 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
912 ratio
= 1000.0 * avg
/ total
;
918 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
920 fprintf(output
, " ");
924 static void print_aggr(char *prefix
)
926 struct perf_evsel
*counter
;
927 int cpu
, cpu2
, s
, s2
, id
, nr
;
930 if (!(aggr_map
|| aggr_get_id
))
933 for (s
= 0; s
< aggr_map
->nr
; s
++) {
934 id
= aggr_map
->map
[s
];
935 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
938 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
939 cpu2
= perf_evsel__cpus(counter
)->map
[cpu
];
940 s2
= aggr_get_id(evsel_list
->cpus
, cpu2
);
943 val
+= counter
->counts
->cpu
[cpu
].val
;
944 ena
+= counter
->counts
->cpu
[cpu
].ena
;
945 run
+= counter
->counts
->cpu
[cpu
].run
;
949 fprintf(output
, "%s", prefix
);
951 if (run
== 0 || ena
== 0) {
952 aggr_printout(counter
, id
, nr
);
954 fprintf(output
, "%*s%s%*s",
956 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
958 csv_output
? 0 : -24,
959 perf_evsel__name(counter
));
962 fprintf(output
, "%s%s",
963 csv_sep
, counter
->cgrp
->name
);
969 if (nsec_counter(counter
))
970 nsec_printout(id
, nr
, counter
, val
);
972 abs_printout(id
, nr
, counter
, val
);
975 print_noise(counter
, 1.0);
978 fprintf(output
, " (%.2f%%)",
987 * Print out the results of a single counter:
988 * aggregated counts in system-wide mode
990 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
992 struct perf_stat
*ps
= counter
->priv
;
993 double avg
= avg_stats(&ps
->res_stats
[0]);
994 int scaled
= counter
->counts
->scaled
;
997 fprintf(output
, "%s", prefix
);
1000 fprintf(output
, "%*s%s%*s",
1001 csv_output
? 0 : 18,
1002 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1004 csv_output
? 0 : -24,
1005 perf_evsel__name(counter
));
1008 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
1010 fputc('\n', output
);
1014 if (nsec_counter(counter
))
1015 nsec_printout(-1, 0, counter
, avg
);
1017 abs_printout(-1, 0, counter
, avg
);
1019 print_noise(counter
, avg
);
1022 fputc('\n', output
);
1027 double avg_enabled
, avg_running
;
1029 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
1030 avg_running
= avg_stats(&ps
->res_stats
[2]);
1032 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
1034 fprintf(output
, "\n");
1038 * Print out the results of a single counter:
1039 * does not use aggregated count in system-wide
1041 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1046 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1047 val
= counter
->counts
->cpu
[cpu
].val
;
1048 ena
= counter
->counts
->cpu
[cpu
].ena
;
1049 run
= counter
->counts
->cpu
[cpu
].run
;
1052 fprintf(output
, "%s", prefix
);
1054 if (run
== 0 || ena
== 0) {
1055 fprintf(output
, "CPU%*d%s%*s%s%*s",
1056 csv_output
? 0 : -4,
1057 perf_evsel__cpus(counter
)->map
[cpu
], csv_sep
,
1058 csv_output
? 0 : 18,
1059 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1061 csv_output
? 0 : -24,
1062 perf_evsel__name(counter
));
1065 fprintf(output
, "%s%s",
1066 csv_sep
, counter
->cgrp
->name
);
1068 fputc('\n', output
);
1072 if (nsec_counter(counter
))
1073 nsec_printout(cpu
, 0, counter
, val
);
1075 abs_printout(cpu
, 0, counter
, val
);
1078 print_noise(counter
, 1.0);
1081 fprintf(output
, " (%.2f%%)",
1084 fputc('\n', output
);
1088 static void print_stat(int argc
, const char **argv
)
1090 struct perf_evsel
*counter
;
1096 fprintf(output
, "\n");
1097 fprintf(output
, " Performance counter stats for ");
1098 if (!perf_target__has_task(&target
)) {
1099 fprintf(output
, "\'%s", argv
[0]);
1100 for (i
= 1; i
< argc
; i
++)
1101 fprintf(output
, " %s", argv
[i
]);
1102 } else if (target
.pid
)
1103 fprintf(output
, "process id \'%s", target
.pid
);
1105 fprintf(output
, "thread id \'%s", target
.tid
);
1107 fprintf(output
, "\'");
1109 fprintf(output
, " (%d runs)", run_count
);
1110 fprintf(output
, ":\n\n");
1113 switch (aggr_mode
) {
1119 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1120 print_counter_aggr(counter
, NULL
);
1123 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1124 print_counter(counter
, NULL
);
1132 fprintf(output
, "\n");
1133 fprintf(output
, " %17.9f seconds time elapsed",
1134 avg_stats(&walltime_nsecs_stats
)/1e9
);
1135 if (run_count
> 1) {
1136 fprintf(output
, " ");
1137 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1138 avg_stats(&walltime_nsecs_stats
));
1140 fprintf(output
, "\n\n");
1144 static volatile int signr
= -1;
1146 static void skip_signal(int signo
)
1148 if ((child_pid
== -1) || interval
)
1153 * render child_pid harmless
1154 * won't send SIGTERM to a random
1155 * process in case of race condition
1156 * and fast PID recycling
1161 static void sig_atexit(void)
1166 * avoid race condition with SIGCHLD handler
1167 * in skip_signal() which is modifying child_pid
1168 * goal is to avoid send SIGTERM to a random
1172 sigaddset(&set
, SIGCHLD
);
1173 sigprocmask(SIG_BLOCK
, &set
, &oset
);
1175 if (child_pid
!= -1)
1176 kill(child_pid
, SIGTERM
);
1178 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
1183 signal(signr
, SIG_DFL
);
1184 kill(getpid(), signr
);
1187 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1188 const char *s __maybe_unused
, int unset
)
1190 big_num_opt
= unset
? 0 : 1;
1194 static int perf_stat_init_aggr_mode(void)
1196 switch (aggr_mode
) {
1198 if (cpu_map__build_socket_map(evsel_list
->cpus
, &aggr_map
)) {
1199 perror("cannot build socket map");
1202 aggr_get_id
= cpu_map__get_socket
;
1205 if (cpu_map__build_core_map(evsel_list
->cpus
, &aggr_map
)) {
1206 perror("cannot build core map");
1209 aggr_get_id
= cpu_map__get_core
;
1221 * Add default attributes, if there were no attributes specified or
1222 * if -d/--detailed, -d -d or -d -d -d is used:
1224 static int add_default_attributes(void)
1226 struct perf_event_attr default_attrs
[] = {
1228 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1229 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1230 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1231 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1233 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1234 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1235 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1236 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1237 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1238 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1243 * Detailed stats (-d), covering the L1 and last level data caches:
1245 struct perf_event_attr detailed_attrs
[] = {
1247 { .type
= PERF_TYPE_HW_CACHE
,
1249 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1250 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1251 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1253 { .type
= PERF_TYPE_HW_CACHE
,
1255 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1256 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1257 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1259 { .type
= PERF_TYPE_HW_CACHE
,
1261 PERF_COUNT_HW_CACHE_LL
<< 0 |
1262 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1263 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1265 { .type
= PERF_TYPE_HW_CACHE
,
1267 PERF_COUNT_HW_CACHE_LL
<< 0 |
1268 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1269 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1273 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1275 struct perf_event_attr very_detailed_attrs
[] = {
1277 { .type
= PERF_TYPE_HW_CACHE
,
1279 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1280 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1281 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1283 { .type
= PERF_TYPE_HW_CACHE
,
1285 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1286 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1287 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1289 { .type
= PERF_TYPE_HW_CACHE
,
1291 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1292 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1293 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1295 { .type
= PERF_TYPE_HW_CACHE
,
1297 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1298 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1299 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1301 { .type
= PERF_TYPE_HW_CACHE
,
1303 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1304 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1305 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1307 { .type
= PERF_TYPE_HW_CACHE
,
1309 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1310 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1311 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1316 * Very, very detailed stats (-d -d -d), adding prefetch events:
1318 struct perf_event_attr very_very_detailed_attrs
[] = {
1320 { .type
= PERF_TYPE_HW_CACHE
,
1322 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1323 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1324 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1326 { .type
= PERF_TYPE_HW_CACHE
,
1328 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1329 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1330 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1333 /* Set attrs if no event is selected and !null_run: */
1337 if (!evsel_list
->nr_entries
) {
1338 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1342 /* Detailed events get appended to the event list: */
1344 if (detailed_run
< 1)
1347 /* Append detailed run extra attributes: */
1348 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1351 if (detailed_run
< 2)
1354 /* Append very detailed run extra attributes: */
1355 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1358 if (detailed_run
< 3)
1361 /* Append very, very detailed run extra attributes: */
1362 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1365 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1367 bool append_file
= false;
1369 const char *output_name
= NULL
;
1370 const struct option options
[] = {
1371 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1372 "event selector. use 'perf list' to list available events",
1373 parse_events_option
),
1374 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1375 "event filter", parse_filter
),
1376 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1377 "child tasks do not inherit counters"),
1378 OPT_STRING('p', "pid", &target
.pid
, "pid",
1379 "stat events on existing process id"),
1380 OPT_STRING('t', "tid", &target
.tid
, "tid",
1381 "stat events on existing thread id"),
1382 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1383 "system-wide collection from all CPUs"),
1384 OPT_BOOLEAN('g', "group", &group
,
1385 "put the counters into a counter group"),
1386 OPT_BOOLEAN('c', "scale", &scale
, "scale/normalize counters"),
1387 OPT_INCR('v', "verbose", &verbose
,
1388 "be more verbose (show counter open errors, etc)"),
1389 OPT_INTEGER('r', "repeat", &run_count
,
1390 "repeat command and print average + stddev (max: 100, forever: 0)"),
1391 OPT_BOOLEAN('n', "null", &null_run
,
1392 "null run - dont start any counters"),
1393 OPT_INCR('d', "detailed", &detailed_run
,
1394 "detailed run - start a lot of events"),
1395 OPT_BOOLEAN('S', "sync", &sync_run
,
1396 "call sync() before starting a run"),
1397 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1398 "print large numbers with thousands\' separators",
1400 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1401 "list of cpus to monitor in system-wide"),
1402 OPT_SET_UINT('A', "no-aggr", &aggr_mode
,
1403 "disable CPU count aggregation", AGGR_NONE
),
1404 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1405 "print counts with custom separator"),
1406 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1407 "monitor event in cgroup name only", parse_cgroups
),
1408 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1409 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1410 OPT_INTEGER(0, "log-fd", &output_fd
,
1411 "log output to fd, instead of stderr"),
1412 OPT_STRING(0, "pre", &pre_cmd
, "command",
1413 "command to run prior to the measured command"),
1414 OPT_STRING(0, "post", &post_cmd
, "command",
1415 "command to run after to the measured command"),
1416 OPT_UINTEGER('I', "interval-print", &interval
,
1417 "print counts at regular interval in ms (>= 100)"),
1418 OPT_SET_UINT(0, "per-socket", &aggr_mode
,
1419 "aggregate counts per processor socket", AGGR_SOCKET
),
1420 OPT_SET_UINT(0, "per-core", &aggr_mode
,
1421 "aggregate counts per physical processor core", AGGR_CORE
),
1424 const char * const stat_usage
[] = {
1425 "perf stat [<options>] [<command>]",
1428 int status
= -ENOMEM
, run_idx
;
1431 setlocale(LC_ALL
, "");
1433 evsel_list
= perf_evlist__new();
1434 if (evsel_list
== NULL
)
1437 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1438 PARSE_OPT_STOP_AT_NON_OPTION
);
1441 if (output_name
&& strcmp(output_name
, "-"))
1444 if (output_name
&& output_fd
) {
1445 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1446 usage_with_options(stat_usage
, options
);
1449 if (output_fd
< 0) {
1450 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1451 usage_with_options(stat_usage
, options
);
1456 mode
= append_file
? "a" : "w";
1458 output
= fopen(output_name
, mode
);
1460 perror("failed to create output file");
1463 clock_gettime(CLOCK_REALTIME
, &tm
);
1464 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1465 } else if (output_fd
> 0) {
1466 mode
= append_file
? "a" : "w";
1467 output
= fdopen(output_fd
, mode
);
1469 perror("Failed opening logfd");
1476 if (!strcmp(csv_sep
, "\\t"))
1479 csv_sep
= DEFAULT_SEPARATOR
;
1482 * let the spreadsheet do the pretty-printing
1485 /* User explicitly passed -B? */
1486 if (big_num_opt
== 1) {
1487 fprintf(stderr
, "-B option not supported with -x\n");
1488 usage_with_options(stat_usage
, options
);
1489 } else /* Nope, so disable big number formatting */
1491 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1494 if (!argc
&& !perf_target__has_task(&target
))
1495 usage_with_options(stat_usage
, options
);
1496 if (run_count
< 0) {
1497 usage_with_options(stat_usage
, options
);
1498 } else if (run_count
== 0) {
1503 /* no_aggr, cgroup are for system-wide only */
1504 if ((aggr_mode
!= AGGR_GLOBAL
|| nr_cgroups
)
1505 && !perf_target__has_cpu(&target
)) {
1506 fprintf(stderr
, "both cgroup and no-aggregation "
1507 "modes only available in system-wide mode\n");
1509 usage_with_options(stat_usage
, options
);
1513 if (add_default_attributes())
1516 perf_target__validate(&target
);
1518 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1519 if (perf_target__has_task(&target
))
1520 pr_err("Problems finding threads of monitor\n");
1521 if (perf_target__has_cpu(&target
))
1522 perror("failed to parse CPUs map");
1524 usage_with_options(stat_usage
, options
);
1527 if (interval
&& interval
< 100) {
1528 pr_err("print interval must be >= 100ms\n");
1529 usage_with_options(stat_usage
, options
);
1533 if (perf_evlist__alloc_stats(evsel_list
, interval
))
1536 if (perf_stat_init_aggr_mode())
1540 * We dont want to block the signals - that would cause
1541 * child tasks to inherit that and Ctrl-C would not work.
1542 * What we want is for Ctrl-C to work in the exec()-ed
1543 * task, but being ignored by perf stat itself:
1547 signal(SIGINT
, skip_signal
);
1548 signal(SIGCHLD
, skip_signal
);
1549 signal(SIGALRM
, skip_signal
);
1550 signal(SIGABRT
, skip_signal
);
1553 for (run_idx
= 0; forever
|| run_idx
< run_count
; run_idx
++) {
1554 if (run_count
!= 1 && verbose
)
1555 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1558 status
= run_perf_stat(argc
, argv
);
1559 if (forever
&& status
!= -1) {
1560 print_stat(argc
, argv
);
1561 perf_stat__reset_stats(evsel_list
);
1565 if (!forever
&& status
!= -1 && !interval
)
1566 print_stat(argc
, argv
);
1568 perf_evlist__free_stats(evsel_list
);
1570 perf_evlist__delete_maps(evsel_list
);
1572 perf_evlist__delete(evsel_list
);