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"
50 #include "util/event.h"
51 #include "util/evlist.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread.h"
59 #include "util/thread_map.h"
62 #include <sys/prctl.h>
65 #define DEFAULT_SEPARATOR " "
66 #define CNTR_NOT_SUPPORTED "<not supported>"
67 #define CNTR_NOT_COUNTED "<not counted>"
69 static void print_stat(int argc
, const char **argv
);
70 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
);
71 static void print_counter(struct perf_evsel
*counter
, char *prefix
);
72 static void print_aggr(char *prefix
);
74 /* Default events used for perf stat -T */
75 static const char * const transaction_attrs
[] = {
87 /* More limited version when the CPU does not have all events. */
88 static const char * const transaction_limited_attrs
[] = {
98 /* must match transaction_attrs and the beginning limited_attrs */
109 static struct perf_evlist
*evsel_list
;
111 static struct target target
= {
122 static int run_count
= 1;
123 static bool no_inherit
= false;
124 static bool scale
= true;
125 static enum aggr_mode aggr_mode
= AGGR_GLOBAL
;
126 static volatile pid_t child_pid
= -1;
127 static bool null_run
= false;
128 static int detailed_run
= 0;
129 static bool transaction_run
;
130 static bool big_num
= true;
131 static int big_num_opt
= -1;
132 static const char *csv_sep
= NULL
;
133 static bool csv_output
= false;
134 static bool group
= false;
135 static FILE *output
= NULL
;
136 static const char *pre_cmd
= NULL
;
137 static const char *post_cmd
= NULL
;
138 static bool sync_run
= false;
139 static unsigned int interval
= 0;
140 static unsigned int initial_delay
= 0;
141 static bool forever
= false;
142 static struct timespec ref_time
;
143 static struct cpu_map
*aggr_map
;
144 static int (*aggr_get_id
)(struct cpu_map
*m
, int cpu
);
146 static volatile int done
= 0;
149 struct stats res_stats
[3];
152 static inline void diff_timespec(struct timespec
*r
, struct timespec
*a
,
155 r
->tv_sec
= a
->tv_sec
- b
->tv_sec
;
156 if (a
->tv_nsec
< b
->tv_nsec
) {
157 r
->tv_nsec
= a
->tv_nsec
+ 1000000000L - b
->tv_nsec
;
160 r
->tv_nsec
= a
->tv_nsec
- b
->tv_nsec
;
164 static inline struct cpu_map
*perf_evsel__cpus(struct perf_evsel
*evsel
)
166 return (evsel
->cpus
&& !target
.cpu_list
) ? evsel
->cpus
: evsel_list
->cpus
;
169 static inline int perf_evsel__nr_cpus(struct perf_evsel
*evsel
)
171 return perf_evsel__cpus(evsel
)->nr
;
174 static void perf_evsel__reset_stat_priv(struct perf_evsel
*evsel
)
176 memset(evsel
->priv
, 0, sizeof(struct perf_stat
));
179 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
181 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
182 return evsel
->priv
== NULL
? -ENOMEM
: 0;
185 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
191 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel
*evsel
)
196 sz
= sizeof(*evsel
->counts
) +
197 (perf_evsel__nr_cpus(evsel
) * sizeof(struct perf_counts_values
));
203 evsel
->prev_raw_counts
= addr
;
208 static void perf_evsel__free_prev_raw_counts(struct perf_evsel
*evsel
)
210 free(evsel
->prev_raw_counts
);
211 evsel
->prev_raw_counts
= NULL
;
214 static void perf_evlist__free_stats(struct perf_evlist
*evlist
)
216 struct perf_evsel
*evsel
;
218 list_for_each_entry(evsel
, &evlist
->entries
, node
) {
219 perf_evsel__free_stat_priv(evsel
);
220 perf_evsel__free_counts(evsel
);
221 perf_evsel__free_prev_raw_counts(evsel
);
225 static int perf_evlist__alloc_stats(struct perf_evlist
*evlist
, bool alloc_raw
)
227 struct perf_evsel
*evsel
;
229 list_for_each_entry(evsel
, &evlist
->entries
, node
) {
230 if (perf_evsel__alloc_stat_priv(evsel
) < 0 ||
231 perf_evsel__alloc_counts(evsel
, perf_evsel__nr_cpus(evsel
)) < 0 ||
232 (alloc_raw
&& perf_evsel__alloc_prev_raw_counts(evsel
) < 0))
239 perf_evlist__free_stats(evlist
);
243 static struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
244 static struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
245 static struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
246 static struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
247 static struct stats runtime_branches_stats
[MAX_NR_CPUS
];
248 static struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
249 static struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
250 static struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
251 static struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
252 static struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
253 static struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
254 static struct stats runtime_cycles_in_tx_stats
[MAX_NR_CPUS
];
255 static struct stats walltime_nsecs_stats
;
256 static struct stats runtime_transaction_stats
[MAX_NR_CPUS
];
257 static struct stats runtime_elision_stats
[MAX_NR_CPUS
];
259 static void perf_stat__reset_stats(struct perf_evlist
*evlist
)
261 struct perf_evsel
*evsel
;
263 list_for_each_entry(evsel
, &evlist
->entries
, node
) {
264 perf_evsel__reset_stat_priv(evsel
);
265 perf_evsel__reset_counts(evsel
, perf_evsel__nr_cpus(evsel
));
268 memset(runtime_nsecs_stats
, 0, sizeof(runtime_nsecs_stats
));
269 memset(runtime_cycles_stats
, 0, sizeof(runtime_cycles_stats
));
270 memset(runtime_stalled_cycles_front_stats
, 0, sizeof(runtime_stalled_cycles_front_stats
));
271 memset(runtime_stalled_cycles_back_stats
, 0, sizeof(runtime_stalled_cycles_back_stats
));
272 memset(runtime_branches_stats
, 0, sizeof(runtime_branches_stats
));
273 memset(runtime_cacherefs_stats
, 0, sizeof(runtime_cacherefs_stats
));
274 memset(runtime_l1_dcache_stats
, 0, sizeof(runtime_l1_dcache_stats
));
275 memset(runtime_l1_icache_stats
, 0, sizeof(runtime_l1_icache_stats
));
276 memset(runtime_ll_cache_stats
, 0, sizeof(runtime_ll_cache_stats
));
277 memset(runtime_itlb_cache_stats
, 0, sizeof(runtime_itlb_cache_stats
));
278 memset(runtime_dtlb_cache_stats
, 0, sizeof(runtime_dtlb_cache_stats
));
279 memset(runtime_cycles_in_tx_stats
, 0,
280 sizeof(runtime_cycles_in_tx_stats
));
281 memset(runtime_transaction_stats
, 0,
282 sizeof(runtime_transaction_stats
));
283 memset(runtime_elision_stats
, 0, sizeof(runtime_elision_stats
));
284 memset(&walltime_nsecs_stats
, 0, sizeof(walltime_nsecs_stats
));
287 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
289 struct perf_event_attr
*attr
= &evsel
->attr
;
292 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
293 PERF_FORMAT_TOTAL_TIME_RUNNING
;
295 attr
->inherit
= !no_inherit
;
297 if (target__has_cpu(&target
))
298 return perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
300 if (!target__has_task(&target
) && perf_evsel__is_group_leader(evsel
)) {
303 attr
->enable_on_exec
= 1;
306 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
310 * Does the counter have nsecs as a unit?
312 static inline int nsec_counter(struct perf_evsel
*evsel
)
314 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
315 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
321 static struct perf_evsel
*nth_evsel(int n
)
323 static struct perf_evsel
**array
;
324 static int array_len
;
325 struct perf_evsel
*ev
;
328 /* Assumes this only called when evsel_list does not change anymore. */
330 list_for_each_entry(ev
, &evsel_list
->entries
, node
)
332 array
= malloc(array_len
* sizeof(void *));
336 list_for_each_entry(ev
, &evsel_list
->entries
, node
)
345 * Update various tracking values we maintain to print
346 * more semantic information such as miss/hit ratios,
347 * instruction rates, etc:
349 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
351 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
352 update_stats(&runtime_nsecs_stats
[0], count
[0]);
353 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
354 update_stats(&runtime_cycles_stats
[0], count
[0]);
355 else if (transaction_run
&&
356 perf_evsel__cmp(counter
, nth_evsel(T_CYCLES_IN_TX
)))
357 update_stats(&runtime_cycles_in_tx_stats
[0], count
[0]);
358 else if (transaction_run
&&
359 perf_evsel__cmp(counter
, nth_evsel(T_TRANSACTION_START
)))
360 update_stats(&runtime_transaction_stats
[0], count
[0]);
361 else if (transaction_run
&&
362 perf_evsel__cmp(counter
, nth_evsel(T_ELISION_START
)))
363 update_stats(&runtime_elision_stats
[0], count
[0]);
364 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
365 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
366 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
367 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
368 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
369 update_stats(&runtime_branches_stats
[0], count
[0]);
370 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
371 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
372 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
373 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
374 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
375 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
376 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
377 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
378 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
379 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
380 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
381 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
385 * Read out the results of a single counter:
386 * aggregate counts across CPUs in system-wide mode
388 static int read_counter_aggr(struct perf_evsel
*counter
)
390 struct perf_stat
*ps
= counter
->priv
;
391 u64
*count
= counter
->counts
->aggr
.values
;
394 if (__perf_evsel__read(counter
, perf_evsel__nr_cpus(counter
),
395 thread_map__nr(evsel_list
->threads
), scale
) < 0)
398 for (i
= 0; i
< 3; i
++)
399 update_stats(&ps
->res_stats
[i
], count
[i
]);
402 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
403 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
407 * Save the full runtime - to allow normalization during printout:
409 update_shadow_stats(counter
, count
);
415 * Read out the results of a single counter:
416 * do not aggregate counts across CPUs in system-wide mode
418 static int read_counter(struct perf_evsel
*counter
)
423 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
424 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
427 count
= counter
->counts
->cpu
[cpu
].values
;
429 update_shadow_stats(counter
, count
);
435 static void print_interval(void)
437 static int num_print_interval
;
438 struct perf_evsel
*counter
;
439 struct perf_stat
*ps
;
440 struct timespec ts
, rs
;
443 if (aggr_mode
== AGGR_GLOBAL
) {
444 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
446 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
447 read_counter_aggr(counter
);
450 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
452 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
453 read_counter(counter
);
457 clock_gettime(CLOCK_MONOTONIC
, &ts
);
458 diff_timespec(&rs
, &ts
, &ref_time
);
459 sprintf(prefix
, "%6lu.%09lu%s", rs
.tv_sec
, rs
.tv_nsec
, csv_sep
);
461 if (num_print_interval
== 0 && !csv_output
) {
464 fprintf(output
, "# time socket cpus counts events\n");
467 fprintf(output
, "# time core cpus counts events\n");
470 fprintf(output
, "# time CPU counts events\n");
474 fprintf(output
, "# time counts events\n");
478 if (++num_print_interval
== 25)
479 num_print_interval
= 0;
487 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
488 print_counter(counter
, prefix
);
492 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
493 print_counter_aggr(counter
, prefix
);
499 static void handle_initial_delay(void)
501 struct perf_evsel
*counter
;
504 const int ncpus
= cpu_map__nr(evsel_list
->cpus
),
505 nthreads
= thread_map__nr(evsel_list
->threads
);
507 usleep(initial_delay
* 1000);
508 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
509 perf_evsel__enable(counter
, ncpus
, nthreads
);
513 static int __run_perf_stat(int argc
, const char **argv
)
516 unsigned long long t0
, t1
;
517 struct perf_evsel
*counter
;
520 const bool forks
= (argc
> 0);
523 ts
.tv_sec
= interval
/ 1000;
524 ts
.tv_nsec
= (interval
% 1000) * 1000000;
531 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
,
533 perror("failed to prepare workload");
536 child_pid
= evsel_list
->workload
.pid
;
540 perf_evlist__set_leader(evsel_list
);
542 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
543 if (create_perf_stat_counter(counter
) < 0) {
545 * PPC returns ENXIO for HW counters until 2.6.37
546 * (behavior changed with commit b0a873e).
548 if (errno
== EINVAL
|| errno
== ENOSYS
||
549 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
552 ui__warning("%s event is not supported by the kernel.\n",
553 perf_evsel__name(counter
));
554 counter
->supported
= false;
558 perf_evsel__open_strerror(counter
, &target
,
559 errno
, msg
, sizeof(msg
));
560 ui__error("%s\n", msg
);
563 kill(child_pid
, SIGTERM
);
567 counter
->supported
= true;
570 if (perf_evlist__apply_filters(evsel_list
)) {
571 error("failed to set filter with %d (%s)\n", errno
,
577 * Enable counters and exec the command:
580 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
583 perf_evlist__start_workload(evsel_list
);
584 handle_initial_delay();
587 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
588 nanosleep(&ts
, NULL
);
593 if (WIFSIGNALED(status
))
594 psignal(WTERMSIG(status
), argv
[0]);
596 handle_initial_delay();
598 nanosleep(&ts
, NULL
);
606 update_stats(&walltime_nsecs_stats
, t1
- t0
);
608 if (aggr_mode
== AGGR_GLOBAL
) {
609 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
610 read_counter_aggr(counter
);
611 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
612 thread_map__nr(evsel_list
->threads
));
615 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
616 read_counter(counter
);
617 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
), 1);
621 return WEXITSTATUS(status
);
624 static int run_perf_stat(int argc __maybe_unused
, const char **argv
)
629 ret
= system(pre_cmd
);
637 ret
= __run_perf_stat(argc
, argv
);
642 ret
= system(post_cmd
);
650 static void print_noise_pct(double total
, double avg
)
652 double pct
= rel_stddev_stats(total
, avg
);
655 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
657 fprintf(output
, " ( +-%6.2f%% )", pct
);
660 static void print_noise(struct perf_evsel
*evsel
, double avg
)
662 struct perf_stat
*ps
;
668 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
671 static void aggr_printout(struct perf_evsel
*evsel
, int id
, int nr
)
675 fprintf(output
, "S%d-C%*d%s%*d%s",
676 cpu_map__id_to_socket(id
),
678 cpu_map__id_to_cpu(id
),
685 fprintf(output
, "S%*d%s%*d%s",
694 fprintf(output
, "CPU%*d%s",
696 perf_evsel__cpus(evsel
)->map
[id
], csv_sep
);
704 static void nsec_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
706 double msecs
= avg
/ 1e6
;
707 const char *fmt
= csv_output
? "%.6f%s%s" : "%18.6f%s%-25s";
710 aggr_printout(evsel
, cpu
, nr
);
712 scnprintf(name
, sizeof(name
), "%s%s",
713 perf_evsel__name(evsel
), csv_output
? "" : " (msec)");
714 fprintf(output
, fmt
, msecs
, csv_sep
, name
);
717 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
719 if (csv_output
|| interval
)
722 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
723 fprintf(output
, " # %8.3f CPUs utilized ",
724 avg
/ avg_stats(&walltime_nsecs_stats
));
726 fprintf(output
, " ");
729 /* used for get_ratio_color() */
731 GRC_STALLED_CYCLES_FE
,
732 GRC_STALLED_CYCLES_BE
,
737 static const char *get_ratio_color(enum grc_type type
, double ratio
)
739 static const double grc_table
[GRC_MAX_NR
][3] = {
740 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
741 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
742 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
744 const char *color
= PERF_COLOR_NORMAL
;
746 if (ratio
> grc_table
[type
][0])
747 color
= PERF_COLOR_RED
;
748 else if (ratio
> grc_table
[type
][1])
749 color
= PERF_COLOR_MAGENTA
;
750 else if (ratio
> grc_table
[type
][2])
751 color
= PERF_COLOR_YELLOW
;
756 static void print_stalled_cycles_frontend(int cpu
,
757 struct perf_evsel
*evsel
758 __maybe_unused
, double avg
)
760 double total
, ratio
= 0.0;
763 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
766 ratio
= avg
/ total
* 100.0;
768 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
770 fprintf(output
, " # ");
771 color_fprintf(output
, color
, "%6.2f%%", ratio
);
772 fprintf(output
, " frontend cycles idle ");
775 static void print_stalled_cycles_backend(int cpu
,
776 struct perf_evsel
*evsel
777 __maybe_unused
, double avg
)
779 double total
, ratio
= 0.0;
782 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
785 ratio
= avg
/ total
* 100.0;
787 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
789 fprintf(output
, " # ");
790 color_fprintf(output
, color
, "%6.2f%%", ratio
);
791 fprintf(output
, " backend cycles idle ");
794 static void print_branch_misses(int cpu
,
795 struct perf_evsel
*evsel __maybe_unused
,
798 double total
, ratio
= 0.0;
801 total
= avg_stats(&runtime_branches_stats
[cpu
]);
804 ratio
= avg
/ total
* 100.0;
806 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
808 fprintf(output
, " # ");
809 color_fprintf(output
, color
, "%6.2f%%", ratio
);
810 fprintf(output
, " of all branches ");
813 static void print_l1_dcache_misses(int cpu
,
814 struct perf_evsel
*evsel __maybe_unused
,
817 double total
, ratio
= 0.0;
820 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
823 ratio
= avg
/ total
* 100.0;
825 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
827 fprintf(output
, " # ");
828 color_fprintf(output
, color
, "%6.2f%%", ratio
);
829 fprintf(output
, " of all L1-dcache hits ");
832 static void print_l1_icache_misses(int cpu
,
833 struct perf_evsel
*evsel __maybe_unused
,
836 double total
, ratio
= 0.0;
839 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
842 ratio
= avg
/ total
* 100.0;
844 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
846 fprintf(output
, " # ");
847 color_fprintf(output
, color
, "%6.2f%%", ratio
);
848 fprintf(output
, " of all L1-icache hits ");
851 static void print_dtlb_cache_misses(int cpu
,
852 struct perf_evsel
*evsel __maybe_unused
,
855 double total
, ratio
= 0.0;
858 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
861 ratio
= avg
/ total
* 100.0;
863 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
865 fprintf(output
, " # ");
866 color_fprintf(output
, color
, "%6.2f%%", ratio
);
867 fprintf(output
, " of all dTLB cache hits ");
870 static void print_itlb_cache_misses(int cpu
,
871 struct perf_evsel
*evsel __maybe_unused
,
874 double total
, ratio
= 0.0;
877 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
880 ratio
= avg
/ total
* 100.0;
882 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
884 fprintf(output
, " # ");
885 color_fprintf(output
, color
, "%6.2f%%", ratio
);
886 fprintf(output
, " of all iTLB cache hits ");
889 static void print_ll_cache_misses(int cpu
,
890 struct perf_evsel
*evsel __maybe_unused
,
893 double total
, ratio
= 0.0;
896 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
899 ratio
= avg
/ total
* 100.0;
901 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
903 fprintf(output
, " # ");
904 color_fprintf(output
, color
, "%6.2f%%", ratio
);
905 fprintf(output
, " of all LL-cache hits ");
908 static void abs_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
910 double total
, ratio
= 0.0, total2
;
916 fmt
= "%'18.0f%s%-25s";
918 fmt
= "%18.0f%s%-25s";
920 aggr_printout(evsel
, cpu
, nr
);
922 if (aggr_mode
== AGGR_GLOBAL
)
925 fprintf(output
, fmt
, avg
, csv_sep
, perf_evsel__name(evsel
));
928 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
930 if (csv_output
|| interval
)
933 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
934 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
937 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
939 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
940 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
944 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
947 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
948 runtime_branches_stats
[cpu
].n
!= 0) {
949 print_branch_misses(cpu
, evsel
, avg
);
951 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
952 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
953 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
954 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
955 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
956 print_l1_dcache_misses(cpu
, evsel
, avg
);
958 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
959 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
960 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
961 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
962 runtime_l1_icache_stats
[cpu
].n
!= 0) {
963 print_l1_icache_misses(cpu
, evsel
, avg
);
965 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
966 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
967 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
968 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
969 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
970 print_dtlb_cache_misses(cpu
, evsel
, avg
);
972 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
973 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
974 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
975 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
976 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
977 print_itlb_cache_misses(cpu
, evsel
, avg
);
979 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
980 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
981 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
982 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
983 runtime_ll_cache_stats
[cpu
].n
!= 0) {
984 print_ll_cache_misses(cpu
, evsel
, avg
);
985 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
986 runtime_cacherefs_stats
[cpu
].n
!= 0) {
987 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
990 ratio
= avg
* 100 / total
;
992 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
994 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
995 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
996 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
997 print_stalled_cycles_backend(cpu
, evsel
, avg
);
998 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
999 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
1002 ratio
= avg
/ total
;
1003 fprintf(output
, " # %8.3f GHz ", ratio
);
1005 } else if (transaction_run
&&
1006 perf_evsel__cmp(evsel
, nth_evsel(T_CYCLES_IN_TX
))) {
1007 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
1010 " # %5.2f%% transactional cycles ",
1011 100.0 * (avg
/ total
));
1012 } else if (transaction_run
&&
1013 perf_evsel__cmp(evsel
, nth_evsel(T_CYCLES_IN_TX_CP
))) {
1014 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
1015 total2
= avg_stats(&runtime_cycles_in_tx_stats
[cpu
]);
1020 " # %5.2f%% aborted cycles ",
1021 100.0 * ((total2
-avg
) / total
));
1022 } else if (transaction_run
&&
1023 perf_evsel__cmp(evsel
, nth_evsel(T_TRANSACTION_START
)) &&
1025 runtime_cycles_in_tx_stats
[cpu
].n
!= 0) {
1026 total
= avg_stats(&runtime_cycles_in_tx_stats
[cpu
]);
1029 ratio
= total
/ avg
;
1031 fprintf(output
, " # %8.0f cycles / transaction ", ratio
);
1032 } else if (transaction_run
&&
1033 perf_evsel__cmp(evsel
, nth_evsel(T_ELISION_START
)) &&
1035 runtime_cycles_in_tx_stats
[cpu
].n
!= 0) {
1036 total
= avg_stats(&runtime_cycles_in_tx_stats
[cpu
]);
1039 ratio
= total
/ avg
;
1041 fprintf(output
, " # %8.0f cycles / elision ", ratio
);
1042 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
1045 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
1048 ratio
= 1000.0 * avg
/ total
;
1049 if (ratio
< 0.001) {
1054 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
1056 fprintf(output
, " ");
1060 static void print_aggr(char *prefix
)
1062 struct perf_evsel
*counter
;
1063 int cpu
, cpu2
, s
, s2
, id
, nr
;
1066 if (!(aggr_map
|| aggr_get_id
))
1069 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1070 id
= aggr_map
->map
[s
];
1071 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
1072 val
= ena
= run
= 0;
1074 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1075 cpu2
= perf_evsel__cpus(counter
)->map
[cpu
];
1076 s2
= aggr_get_id(evsel_list
->cpus
, cpu2
);
1079 val
+= counter
->counts
->cpu
[cpu
].val
;
1080 ena
+= counter
->counts
->cpu
[cpu
].ena
;
1081 run
+= counter
->counts
->cpu
[cpu
].run
;
1085 fprintf(output
, "%s", prefix
);
1087 if (run
== 0 || ena
== 0) {
1088 aggr_printout(counter
, id
, nr
);
1090 fprintf(output
, "%*s%s%*s",
1091 csv_output
? 0 : 18,
1092 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1094 csv_output
? 0 : -24,
1095 perf_evsel__name(counter
));
1098 fprintf(output
, "%s%s",
1099 csv_sep
, counter
->cgrp
->name
);
1101 fputc('\n', output
);
1105 if (nsec_counter(counter
))
1106 nsec_printout(id
, nr
, counter
, val
);
1108 abs_printout(id
, nr
, counter
, val
);
1111 print_noise(counter
, 1.0);
1114 fprintf(output
, " (%.2f%%)",
1117 fputc('\n', output
);
1123 * Print out the results of a single counter:
1124 * aggregated counts in system-wide mode
1126 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
1128 struct perf_stat
*ps
= counter
->priv
;
1129 double avg
= avg_stats(&ps
->res_stats
[0]);
1130 int scaled
= counter
->counts
->scaled
;
1133 fprintf(output
, "%s", prefix
);
1136 fprintf(output
, "%*s%s%*s",
1137 csv_output
? 0 : 18,
1138 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1140 csv_output
? 0 : -24,
1141 perf_evsel__name(counter
));
1144 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
1146 fputc('\n', output
);
1150 if (nsec_counter(counter
))
1151 nsec_printout(-1, 0, counter
, avg
);
1153 abs_printout(-1, 0, counter
, avg
);
1155 print_noise(counter
, avg
);
1158 fputc('\n', output
);
1163 double avg_enabled
, avg_running
;
1165 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
1166 avg_running
= avg_stats(&ps
->res_stats
[2]);
1168 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
1170 fprintf(output
, "\n");
1174 * Print out the results of a single counter:
1175 * does not use aggregated count in system-wide
1177 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1182 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1183 val
= counter
->counts
->cpu
[cpu
].val
;
1184 ena
= counter
->counts
->cpu
[cpu
].ena
;
1185 run
= counter
->counts
->cpu
[cpu
].run
;
1188 fprintf(output
, "%s", prefix
);
1190 if (run
== 0 || ena
== 0) {
1191 fprintf(output
, "CPU%*d%s%*s%s%*s",
1192 csv_output
? 0 : -4,
1193 perf_evsel__cpus(counter
)->map
[cpu
], csv_sep
,
1194 csv_output
? 0 : 18,
1195 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1197 csv_output
? 0 : -24,
1198 perf_evsel__name(counter
));
1201 fprintf(output
, "%s%s",
1202 csv_sep
, counter
->cgrp
->name
);
1204 fputc('\n', output
);
1208 if (nsec_counter(counter
))
1209 nsec_printout(cpu
, 0, counter
, val
);
1211 abs_printout(cpu
, 0, counter
, val
);
1214 print_noise(counter
, 1.0);
1217 fprintf(output
, " (%.2f%%)",
1220 fputc('\n', output
);
1224 static void print_stat(int argc
, const char **argv
)
1226 struct perf_evsel
*counter
;
1232 fprintf(output
, "\n");
1233 fprintf(output
, " Performance counter stats for ");
1234 if (target
.system_wide
)
1235 fprintf(output
, "\'system wide");
1236 else if (target
.cpu_list
)
1237 fprintf(output
, "\'CPU(s) %s", target
.cpu_list
);
1238 else if (!target__has_task(&target
)) {
1239 fprintf(output
, "\'%s", argv
[0]);
1240 for (i
= 1; i
< argc
; i
++)
1241 fprintf(output
, " %s", argv
[i
]);
1242 } else if (target
.pid
)
1243 fprintf(output
, "process id \'%s", target
.pid
);
1245 fprintf(output
, "thread id \'%s", target
.tid
);
1247 fprintf(output
, "\'");
1249 fprintf(output
, " (%d runs)", run_count
);
1250 fprintf(output
, ":\n\n");
1253 switch (aggr_mode
) {
1259 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1260 print_counter_aggr(counter
, NULL
);
1263 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1264 print_counter(counter
, NULL
);
1272 fprintf(output
, "\n");
1273 fprintf(output
, " %17.9f seconds time elapsed",
1274 avg_stats(&walltime_nsecs_stats
)/1e9
);
1275 if (run_count
> 1) {
1276 fprintf(output
, " ");
1277 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1278 avg_stats(&walltime_nsecs_stats
));
1280 fprintf(output
, "\n\n");
1284 static volatile int signr
= -1;
1286 static void skip_signal(int signo
)
1288 if ((child_pid
== -1) || interval
)
1293 * render child_pid harmless
1294 * won't send SIGTERM to a random
1295 * process in case of race condition
1296 * and fast PID recycling
1301 static void sig_atexit(void)
1306 * avoid race condition with SIGCHLD handler
1307 * in skip_signal() which is modifying child_pid
1308 * goal is to avoid send SIGTERM to a random
1312 sigaddset(&set
, SIGCHLD
);
1313 sigprocmask(SIG_BLOCK
, &set
, &oset
);
1315 if (child_pid
!= -1)
1316 kill(child_pid
, SIGTERM
);
1318 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
1323 signal(signr
, SIG_DFL
);
1324 kill(getpid(), signr
);
1327 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1328 const char *s __maybe_unused
, int unset
)
1330 big_num_opt
= unset
? 0 : 1;
1334 static int perf_stat_init_aggr_mode(void)
1336 switch (aggr_mode
) {
1338 if (cpu_map__build_socket_map(evsel_list
->cpus
, &aggr_map
)) {
1339 perror("cannot build socket map");
1342 aggr_get_id
= cpu_map__get_socket
;
1345 if (cpu_map__build_core_map(evsel_list
->cpus
, &aggr_map
)) {
1346 perror("cannot build core map");
1349 aggr_get_id
= cpu_map__get_core
;
1359 static int setup_events(const char * const *attrs
, unsigned len
)
1363 for (i
= 0; i
< len
; i
++) {
1364 if (parse_events(evsel_list
, attrs
[i
]))
1371 * Add default attributes, if there were no attributes specified or
1372 * if -d/--detailed, -d -d or -d -d -d is used:
1374 static int add_default_attributes(void)
1376 struct perf_event_attr default_attrs
[] = {
1378 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1379 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1380 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1381 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1383 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1384 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1385 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1386 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1387 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1388 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1393 * Detailed stats (-d), covering the L1 and last level data caches:
1395 struct perf_event_attr detailed_attrs
[] = {
1397 { .type
= PERF_TYPE_HW_CACHE
,
1399 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1400 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1401 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1403 { .type
= PERF_TYPE_HW_CACHE
,
1405 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1406 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1407 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1409 { .type
= PERF_TYPE_HW_CACHE
,
1411 PERF_COUNT_HW_CACHE_LL
<< 0 |
1412 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1413 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1415 { .type
= PERF_TYPE_HW_CACHE
,
1417 PERF_COUNT_HW_CACHE_LL
<< 0 |
1418 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1419 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1423 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1425 struct perf_event_attr very_detailed_attrs
[] = {
1427 { .type
= PERF_TYPE_HW_CACHE
,
1429 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1430 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1431 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1433 { .type
= PERF_TYPE_HW_CACHE
,
1435 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1436 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1437 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1439 { .type
= PERF_TYPE_HW_CACHE
,
1441 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1442 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1443 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1445 { .type
= PERF_TYPE_HW_CACHE
,
1447 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1448 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1449 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1451 { .type
= PERF_TYPE_HW_CACHE
,
1453 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1454 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1455 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1457 { .type
= PERF_TYPE_HW_CACHE
,
1459 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1460 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1461 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1466 * Very, very detailed stats (-d -d -d), adding prefetch events:
1468 struct perf_event_attr very_very_detailed_attrs
[] = {
1470 { .type
= PERF_TYPE_HW_CACHE
,
1472 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1473 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1474 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1476 { .type
= PERF_TYPE_HW_CACHE
,
1478 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1479 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1480 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1483 /* Set attrs if no event is selected and !null_run: */
1487 if (transaction_run
) {
1489 if (pmu_have_event("cpu", "cycles-ct") &&
1490 pmu_have_event("cpu", "el-start"))
1491 err
= setup_events(transaction_attrs
,
1492 ARRAY_SIZE(transaction_attrs
));
1494 err
= setup_events(transaction_limited_attrs
,
1495 ARRAY_SIZE(transaction_limited_attrs
));
1497 fprintf(stderr
, "Cannot set up transaction events\n");
1503 if (!evsel_list
->nr_entries
) {
1504 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1508 /* Detailed events get appended to the event list: */
1510 if (detailed_run
< 1)
1513 /* Append detailed run extra attributes: */
1514 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1517 if (detailed_run
< 2)
1520 /* Append very detailed run extra attributes: */
1521 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1524 if (detailed_run
< 3)
1527 /* Append very, very detailed run extra attributes: */
1528 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1531 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1533 bool append_file
= false;
1535 const char *output_name
= NULL
;
1536 const struct option options
[] = {
1537 OPT_BOOLEAN('T', "transaction", &transaction_run
,
1538 "hardware transaction statistics"),
1539 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1540 "event selector. use 'perf list' to list available events",
1541 parse_events_option
),
1542 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1543 "event filter", parse_filter
),
1544 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1545 "child tasks do not inherit counters"),
1546 OPT_STRING('p', "pid", &target
.pid
, "pid",
1547 "stat events on existing process id"),
1548 OPT_STRING('t', "tid", &target
.tid
, "tid",
1549 "stat events on existing thread id"),
1550 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1551 "system-wide collection from all CPUs"),
1552 OPT_BOOLEAN('g', "group", &group
,
1553 "put the counters into a counter group"),
1554 OPT_BOOLEAN('c', "scale", &scale
, "scale/normalize counters"),
1555 OPT_INCR('v', "verbose", &verbose
,
1556 "be more verbose (show counter open errors, etc)"),
1557 OPT_INTEGER('r', "repeat", &run_count
,
1558 "repeat command and print average + stddev (max: 100, forever: 0)"),
1559 OPT_BOOLEAN('n', "null", &null_run
,
1560 "null run - dont start any counters"),
1561 OPT_INCR('d', "detailed", &detailed_run
,
1562 "detailed run - start a lot of events"),
1563 OPT_BOOLEAN('S', "sync", &sync_run
,
1564 "call sync() before starting a run"),
1565 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1566 "print large numbers with thousands\' separators",
1568 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1569 "list of cpus to monitor in system-wide"),
1570 OPT_SET_UINT('A', "no-aggr", &aggr_mode
,
1571 "disable CPU count aggregation", AGGR_NONE
),
1572 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1573 "print counts with custom separator"),
1574 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1575 "monitor event in cgroup name only", parse_cgroups
),
1576 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1577 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1578 OPT_INTEGER(0, "log-fd", &output_fd
,
1579 "log output to fd, instead of stderr"),
1580 OPT_STRING(0, "pre", &pre_cmd
, "command",
1581 "command to run prior to the measured command"),
1582 OPT_STRING(0, "post", &post_cmd
, "command",
1583 "command to run after to the measured command"),
1584 OPT_UINTEGER('I', "interval-print", &interval
,
1585 "print counts at regular interval in ms (>= 100)"),
1586 OPT_SET_UINT(0, "per-socket", &aggr_mode
,
1587 "aggregate counts per processor socket", AGGR_SOCKET
),
1588 OPT_SET_UINT(0, "per-core", &aggr_mode
,
1589 "aggregate counts per physical processor core", AGGR_CORE
),
1590 OPT_UINTEGER('D', "delay", &initial_delay
,
1591 "ms to wait before starting measurement after program start"),
1594 const char * const stat_usage
[] = {
1595 "perf stat [<options>] [<command>]",
1598 int status
= -EINVAL
, run_idx
;
1601 setlocale(LC_ALL
, "");
1603 evsel_list
= perf_evlist__new();
1604 if (evsel_list
== NULL
)
1607 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1608 PARSE_OPT_STOP_AT_NON_OPTION
);
1611 if (output_name
&& strcmp(output_name
, "-"))
1614 if (output_name
&& output_fd
) {
1615 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1616 parse_options_usage(stat_usage
, options
, "o", 1);
1617 parse_options_usage(NULL
, options
, "log-fd", 0);
1621 if (output_fd
< 0) {
1622 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1623 parse_options_usage(stat_usage
, options
, "log-fd", 0);
1629 mode
= append_file
? "a" : "w";
1631 output
= fopen(output_name
, mode
);
1633 perror("failed to create output file");
1636 clock_gettime(CLOCK_REALTIME
, &tm
);
1637 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1638 } else if (output_fd
> 0) {
1639 mode
= append_file
? "a" : "w";
1640 output
= fdopen(output_fd
, mode
);
1642 perror("Failed opening logfd");
1649 if (!strcmp(csv_sep
, "\\t"))
1652 csv_sep
= DEFAULT_SEPARATOR
;
1655 * let the spreadsheet do the pretty-printing
1658 /* User explicitly passed -B? */
1659 if (big_num_opt
== 1) {
1660 fprintf(stderr
, "-B option not supported with -x\n");
1661 parse_options_usage(stat_usage
, options
, "B", 1);
1662 parse_options_usage(NULL
, options
, "x", 1);
1664 } else /* Nope, so disable big number formatting */
1666 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1669 if (!argc
&& target__none(&target
))
1670 usage_with_options(stat_usage
, options
);
1672 if (run_count
< 0) {
1673 pr_err("Run count must be a positive number\n");
1674 parse_options_usage(stat_usage
, options
, "r", 1);
1676 } else if (run_count
== 0) {
1681 /* no_aggr, cgroup are for system-wide only */
1682 if ((aggr_mode
!= AGGR_GLOBAL
|| nr_cgroups
) &&
1683 !target__has_cpu(&target
)) {
1684 fprintf(stderr
, "both cgroup and no-aggregation "
1685 "modes only available in system-wide mode\n");
1687 parse_options_usage(stat_usage
, options
, "G", 1);
1688 parse_options_usage(NULL
, options
, "A", 1);
1689 parse_options_usage(NULL
, options
, "a", 1);
1693 if (add_default_attributes())
1696 target__validate(&target
);
1698 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1699 if (target__has_task(&target
)) {
1700 pr_err("Problems finding threads of monitor\n");
1701 parse_options_usage(stat_usage
, options
, "p", 1);
1702 parse_options_usage(NULL
, options
, "t", 1);
1703 } else if (target__has_cpu(&target
)) {
1704 perror("failed to parse CPUs map");
1705 parse_options_usage(stat_usage
, options
, "C", 1);
1706 parse_options_usage(NULL
, options
, "a", 1);
1710 if (interval
&& interval
< 100) {
1711 pr_err("print interval must be >= 100ms\n");
1712 parse_options_usage(stat_usage
, options
, "I", 1);
1716 if (perf_evlist__alloc_stats(evsel_list
, interval
))
1719 if (perf_stat_init_aggr_mode())
1723 * We dont want to block the signals - that would cause
1724 * child tasks to inherit that and Ctrl-C would not work.
1725 * What we want is for Ctrl-C to work in the exec()-ed
1726 * task, but being ignored by perf stat itself:
1730 signal(SIGINT
, skip_signal
);
1731 signal(SIGCHLD
, skip_signal
);
1732 signal(SIGALRM
, skip_signal
);
1733 signal(SIGABRT
, skip_signal
);
1736 for (run_idx
= 0; forever
|| run_idx
< run_count
; run_idx
++) {
1737 if (run_count
!= 1 && verbose
)
1738 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1741 status
= run_perf_stat(argc
, argv
);
1742 if (forever
&& status
!= -1) {
1743 print_stat(argc
, argv
);
1744 perf_stat__reset_stats(evsel_list
);
1748 if (!forever
&& status
!= -1 && !interval
)
1749 print_stat(argc
, argv
);
1751 perf_evlist__free_stats(evsel_list
);
1753 perf_evlist__delete_maps(evsel_list
);
1755 perf_evlist__delete(evsel_list
);