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 perf_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 (perf_target__has_cpu(&target
))
298 return perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
300 if (!perf_target__has_task(&target
) &&
301 perf_evsel__is_group_leader(evsel
)) {
304 attr
->enable_on_exec
= 1;
307 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
311 * Does the counter have nsecs as a unit?
313 static inline int nsec_counter(struct perf_evsel
*evsel
)
315 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
316 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
322 static struct perf_evsel
*nth_evsel(int n
)
324 static struct perf_evsel
**array
;
325 static int array_len
;
326 struct perf_evsel
*ev
;
329 /* Assumes this only called when evsel_list does not change anymore. */
331 list_for_each_entry(ev
, &evsel_list
->entries
, node
)
333 array
= malloc(array_len
* sizeof(void *));
337 list_for_each_entry(ev
, &evsel_list
->entries
, node
)
346 * Update various tracking values we maintain to print
347 * more semantic information such as miss/hit ratios,
348 * instruction rates, etc:
350 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
352 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
353 update_stats(&runtime_nsecs_stats
[0], count
[0]);
354 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
355 update_stats(&runtime_cycles_stats
[0], count
[0]);
356 else if (transaction_run
&&
357 perf_evsel__cmp(counter
, nth_evsel(T_CYCLES_IN_TX
)))
358 update_stats(&runtime_cycles_in_tx_stats
[0], count
[0]);
359 else if (transaction_run
&&
360 perf_evsel__cmp(counter
, nth_evsel(T_TRANSACTION_START
)))
361 update_stats(&runtime_transaction_stats
[0], count
[0]);
362 else if (transaction_run
&&
363 perf_evsel__cmp(counter
, nth_evsel(T_ELISION_START
)))
364 update_stats(&runtime_elision_stats
[0], count
[0]);
365 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
366 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
367 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
368 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
369 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
370 update_stats(&runtime_branches_stats
[0], count
[0]);
371 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
372 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
373 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
374 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
375 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
376 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
377 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
378 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
379 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
380 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
381 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
382 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
386 * Read out the results of a single counter:
387 * aggregate counts across CPUs in system-wide mode
389 static int read_counter_aggr(struct perf_evsel
*counter
)
391 struct perf_stat
*ps
= counter
->priv
;
392 u64
*count
= counter
->counts
->aggr
.values
;
395 if (__perf_evsel__read(counter
, perf_evsel__nr_cpus(counter
),
396 thread_map__nr(evsel_list
->threads
), scale
) < 0)
399 for (i
= 0; i
< 3; i
++)
400 update_stats(&ps
->res_stats
[i
], count
[i
]);
403 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
404 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
408 * Save the full runtime - to allow normalization during printout:
410 update_shadow_stats(counter
, count
);
416 * Read out the results of a single counter:
417 * do not aggregate counts across CPUs in system-wide mode
419 static int read_counter(struct perf_evsel
*counter
)
424 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
425 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
428 count
= counter
->counts
->cpu
[cpu
].values
;
430 update_shadow_stats(counter
, count
);
436 static void print_interval(void)
438 static int num_print_interval
;
439 struct perf_evsel
*counter
;
440 struct perf_stat
*ps
;
441 struct timespec ts
, rs
;
444 if (aggr_mode
== AGGR_GLOBAL
) {
445 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
447 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
448 read_counter_aggr(counter
);
451 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
453 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
454 read_counter(counter
);
458 clock_gettime(CLOCK_MONOTONIC
, &ts
);
459 diff_timespec(&rs
, &ts
, &ref_time
);
460 sprintf(prefix
, "%6lu.%09lu%s", rs
.tv_sec
, rs
.tv_nsec
, csv_sep
);
462 if (num_print_interval
== 0 && !csv_output
) {
465 fprintf(output
, "# time socket cpus counts events\n");
468 fprintf(output
, "# time core cpus counts events\n");
471 fprintf(output
, "# time CPU counts events\n");
475 fprintf(output
, "# time counts events\n");
479 if (++num_print_interval
== 25)
480 num_print_interval
= 0;
488 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
489 print_counter(counter
, prefix
);
493 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
494 print_counter_aggr(counter
, prefix
);
500 static void handle_initial_delay(void)
502 struct perf_evsel
*counter
;
505 const int ncpus
= cpu_map__nr(evsel_list
->cpus
),
506 nthreads
= thread_map__nr(evsel_list
->threads
);
508 usleep(initial_delay
* 1000);
509 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
510 perf_evsel__enable(counter
, ncpus
, nthreads
);
514 static int __run_perf_stat(int argc
, const char **argv
)
517 unsigned long long t0
, t1
;
518 struct perf_evsel
*counter
;
521 const bool forks
= (argc
> 0);
524 ts
.tv_sec
= interval
/ 1000;
525 ts
.tv_nsec
= (interval
% 1000) * 1000000;
532 if (perf_evlist__prepare_workload(evsel_list
, &target
, argv
,
534 perror("failed to prepare workload");
537 child_pid
= evsel_list
->workload
.pid
;
541 perf_evlist__set_leader(evsel_list
);
543 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
544 if (create_perf_stat_counter(counter
) < 0) {
546 * PPC returns ENXIO for HW counters until 2.6.37
547 * (behavior changed with commit b0a873e).
549 if (errno
== EINVAL
|| errno
== ENOSYS
||
550 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
553 ui__warning("%s event is not supported by the kernel.\n",
554 perf_evsel__name(counter
));
555 counter
->supported
= false;
559 perf_evsel__open_strerror(counter
, &target
,
560 errno
, msg
, sizeof(msg
));
561 ui__error("%s\n", msg
);
564 kill(child_pid
, SIGTERM
);
568 counter
->supported
= true;
571 if (perf_evlist__apply_filters(evsel_list
)) {
572 error("failed to set filter with %d (%s)\n", errno
,
578 * Enable counters and exec the command:
581 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
584 perf_evlist__start_workload(evsel_list
);
585 handle_initial_delay();
588 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
589 nanosleep(&ts
, NULL
);
594 if (WIFSIGNALED(status
))
595 psignal(WTERMSIG(status
), argv
[0]);
597 handle_initial_delay();
599 nanosleep(&ts
, NULL
);
607 update_stats(&walltime_nsecs_stats
, t1
- t0
);
609 if (aggr_mode
== AGGR_GLOBAL
) {
610 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
611 read_counter_aggr(counter
);
612 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
613 thread_map__nr(evsel_list
->threads
));
616 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
617 read_counter(counter
);
618 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
), 1);
622 return WEXITSTATUS(status
);
625 static int run_perf_stat(int argc __maybe_unused
, const char **argv
)
630 ret
= system(pre_cmd
);
638 ret
= __run_perf_stat(argc
, argv
);
643 ret
= system(post_cmd
);
651 static void print_noise_pct(double total
, double avg
)
653 double pct
= rel_stddev_stats(total
, avg
);
656 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
658 fprintf(output
, " ( +-%6.2f%% )", pct
);
661 static void print_noise(struct perf_evsel
*evsel
, double avg
)
663 struct perf_stat
*ps
;
669 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
672 static void aggr_printout(struct perf_evsel
*evsel
, int id
, int nr
)
676 fprintf(output
, "S%d-C%*d%s%*d%s",
677 cpu_map__id_to_socket(id
),
679 cpu_map__id_to_cpu(id
),
686 fprintf(output
, "S%*d%s%*d%s",
695 fprintf(output
, "CPU%*d%s",
697 perf_evsel__cpus(evsel
)->map
[id
], csv_sep
);
705 static void nsec_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
707 double msecs
= avg
/ 1e6
;
708 const char *fmt
= csv_output
? "%.6f%s%s" : "%18.6f%s%-25s";
711 aggr_printout(evsel
, cpu
, nr
);
713 scnprintf(name
, sizeof(name
), "%s%s",
714 perf_evsel__name(evsel
), csv_output
? "" : " (msec)");
715 fprintf(output
, fmt
, msecs
, csv_sep
, name
);
718 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
720 if (csv_output
|| interval
)
723 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
724 fprintf(output
, " # %8.3f CPUs utilized ",
725 avg
/ avg_stats(&walltime_nsecs_stats
));
727 fprintf(output
, " ");
730 /* used for get_ratio_color() */
732 GRC_STALLED_CYCLES_FE
,
733 GRC_STALLED_CYCLES_BE
,
738 static const char *get_ratio_color(enum grc_type type
, double ratio
)
740 static const double grc_table
[GRC_MAX_NR
][3] = {
741 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
742 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
743 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
745 const char *color
= PERF_COLOR_NORMAL
;
747 if (ratio
> grc_table
[type
][0])
748 color
= PERF_COLOR_RED
;
749 else if (ratio
> grc_table
[type
][1])
750 color
= PERF_COLOR_MAGENTA
;
751 else if (ratio
> grc_table
[type
][2])
752 color
= PERF_COLOR_YELLOW
;
757 static void print_stalled_cycles_frontend(int cpu
,
758 struct perf_evsel
*evsel
759 __maybe_unused
, double avg
)
761 double total
, ratio
= 0.0;
764 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
767 ratio
= avg
/ total
* 100.0;
769 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
771 fprintf(output
, " # ");
772 color_fprintf(output
, color
, "%6.2f%%", ratio
);
773 fprintf(output
, " frontend cycles idle ");
776 static void print_stalled_cycles_backend(int cpu
,
777 struct perf_evsel
*evsel
778 __maybe_unused
, double avg
)
780 double total
, ratio
= 0.0;
783 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
786 ratio
= avg
/ total
* 100.0;
788 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
790 fprintf(output
, " # ");
791 color_fprintf(output
, color
, "%6.2f%%", ratio
);
792 fprintf(output
, " backend cycles idle ");
795 static void print_branch_misses(int cpu
,
796 struct perf_evsel
*evsel __maybe_unused
,
799 double total
, ratio
= 0.0;
802 total
= avg_stats(&runtime_branches_stats
[cpu
]);
805 ratio
= avg
/ total
* 100.0;
807 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
809 fprintf(output
, " # ");
810 color_fprintf(output
, color
, "%6.2f%%", ratio
);
811 fprintf(output
, " of all branches ");
814 static void print_l1_dcache_misses(int cpu
,
815 struct perf_evsel
*evsel __maybe_unused
,
818 double total
, ratio
= 0.0;
821 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
824 ratio
= avg
/ total
* 100.0;
826 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
828 fprintf(output
, " # ");
829 color_fprintf(output
, color
, "%6.2f%%", ratio
);
830 fprintf(output
, " of all L1-dcache hits ");
833 static void print_l1_icache_misses(int cpu
,
834 struct perf_evsel
*evsel __maybe_unused
,
837 double total
, ratio
= 0.0;
840 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
843 ratio
= avg
/ total
* 100.0;
845 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
847 fprintf(output
, " # ");
848 color_fprintf(output
, color
, "%6.2f%%", ratio
);
849 fprintf(output
, " of all L1-icache hits ");
852 static void print_dtlb_cache_misses(int cpu
,
853 struct perf_evsel
*evsel __maybe_unused
,
856 double total
, ratio
= 0.0;
859 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
862 ratio
= avg
/ total
* 100.0;
864 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
866 fprintf(output
, " # ");
867 color_fprintf(output
, color
, "%6.2f%%", ratio
);
868 fprintf(output
, " of all dTLB cache hits ");
871 static void print_itlb_cache_misses(int cpu
,
872 struct perf_evsel
*evsel __maybe_unused
,
875 double total
, ratio
= 0.0;
878 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
881 ratio
= avg
/ total
* 100.0;
883 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
885 fprintf(output
, " # ");
886 color_fprintf(output
, color
, "%6.2f%%", ratio
);
887 fprintf(output
, " of all iTLB cache hits ");
890 static void print_ll_cache_misses(int cpu
,
891 struct perf_evsel
*evsel __maybe_unused
,
894 double total
, ratio
= 0.0;
897 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
900 ratio
= avg
/ total
* 100.0;
902 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
904 fprintf(output
, " # ");
905 color_fprintf(output
, color
, "%6.2f%%", ratio
);
906 fprintf(output
, " of all LL-cache hits ");
909 static void abs_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
911 double total
, ratio
= 0.0, total2
;
917 fmt
= "%'18.0f%s%-25s";
919 fmt
= "%18.0f%s%-25s";
921 aggr_printout(evsel
, cpu
, nr
);
923 if (aggr_mode
== AGGR_GLOBAL
)
926 fprintf(output
, fmt
, avg
, csv_sep
, perf_evsel__name(evsel
));
929 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
931 if (csv_output
|| interval
)
934 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
935 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
938 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
940 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
941 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
945 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
948 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
949 runtime_branches_stats
[cpu
].n
!= 0) {
950 print_branch_misses(cpu
, evsel
, avg
);
952 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
953 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
954 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
955 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
956 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
957 print_l1_dcache_misses(cpu
, evsel
, avg
);
959 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
960 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
961 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
962 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
963 runtime_l1_icache_stats
[cpu
].n
!= 0) {
964 print_l1_icache_misses(cpu
, evsel
, avg
);
966 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
967 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
968 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
969 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
970 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
971 print_dtlb_cache_misses(cpu
, evsel
, avg
);
973 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
974 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
975 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
976 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
977 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
978 print_itlb_cache_misses(cpu
, evsel
, avg
);
980 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
981 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
982 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
983 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
984 runtime_ll_cache_stats
[cpu
].n
!= 0) {
985 print_ll_cache_misses(cpu
, evsel
, avg
);
986 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
987 runtime_cacherefs_stats
[cpu
].n
!= 0) {
988 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
991 ratio
= avg
* 100 / total
;
993 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
995 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
996 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
997 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
998 print_stalled_cycles_backend(cpu
, evsel
, avg
);
999 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
1000 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
1003 ratio
= avg
/ total
;
1004 fprintf(output
, " # %8.3f GHz ", ratio
);
1006 } else if (transaction_run
&&
1007 perf_evsel__cmp(evsel
, nth_evsel(T_CYCLES_IN_TX
))) {
1008 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
1011 " # %5.2f%% transactional cycles ",
1012 100.0 * (avg
/ total
));
1013 } else if (transaction_run
&&
1014 perf_evsel__cmp(evsel
, nth_evsel(T_CYCLES_IN_TX_CP
))) {
1015 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
1016 total2
= avg_stats(&runtime_cycles_in_tx_stats
[cpu
]);
1021 " # %5.2f%% aborted cycles ",
1022 100.0 * ((total2
-avg
) / total
));
1023 } else if (transaction_run
&&
1024 perf_evsel__cmp(evsel
, nth_evsel(T_TRANSACTION_START
)) &&
1026 runtime_cycles_in_tx_stats
[cpu
].n
!= 0) {
1027 total
= avg_stats(&runtime_cycles_in_tx_stats
[cpu
]);
1030 ratio
= total
/ avg
;
1032 fprintf(output
, " # %8.0f cycles / transaction ", ratio
);
1033 } else if (transaction_run
&&
1034 perf_evsel__cmp(evsel
, nth_evsel(T_ELISION_START
)) &&
1036 runtime_cycles_in_tx_stats
[cpu
].n
!= 0) {
1037 total
= avg_stats(&runtime_cycles_in_tx_stats
[cpu
]);
1040 ratio
= total
/ avg
;
1042 fprintf(output
, " # %8.0f cycles / elision ", ratio
);
1043 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
1046 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
1049 ratio
= 1000.0 * avg
/ total
;
1050 if (ratio
< 0.001) {
1055 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
1057 fprintf(output
, " ");
1061 static void print_aggr(char *prefix
)
1063 struct perf_evsel
*counter
;
1064 int cpu
, cpu2
, s
, s2
, id
, nr
;
1067 if (!(aggr_map
|| aggr_get_id
))
1070 for (s
= 0; s
< aggr_map
->nr
; s
++) {
1071 id
= aggr_map
->map
[s
];
1072 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
1073 val
= ena
= run
= 0;
1075 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1076 cpu2
= perf_evsel__cpus(counter
)->map
[cpu
];
1077 s2
= aggr_get_id(evsel_list
->cpus
, cpu2
);
1080 val
+= counter
->counts
->cpu
[cpu
].val
;
1081 ena
+= counter
->counts
->cpu
[cpu
].ena
;
1082 run
+= counter
->counts
->cpu
[cpu
].run
;
1086 fprintf(output
, "%s", prefix
);
1088 if (run
== 0 || ena
== 0) {
1089 aggr_printout(counter
, id
, nr
);
1091 fprintf(output
, "%*s%s%*s",
1092 csv_output
? 0 : 18,
1093 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1095 csv_output
? 0 : -24,
1096 perf_evsel__name(counter
));
1099 fprintf(output
, "%s%s",
1100 csv_sep
, counter
->cgrp
->name
);
1102 fputc('\n', output
);
1106 if (nsec_counter(counter
))
1107 nsec_printout(id
, nr
, counter
, val
);
1109 abs_printout(id
, nr
, counter
, val
);
1112 print_noise(counter
, 1.0);
1115 fprintf(output
, " (%.2f%%)",
1118 fputc('\n', output
);
1124 * Print out the results of a single counter:
1125 * aggregated counts in system-wide mode
1127 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
1129 struct perf_stat
*ps
= counter
->priv
;
1130 double avg
= avg_stats(&ps
->res_stats
[0]);
1131 int scaled
= counter
->counts
->scaled
;
1134 fprintf(output
, "%s", prefix
);
1137 fprintf(output
, "%*s%s%*s",
1138 csv_output
? 0 : 18,
1139 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1141 csv_output
? 0 : -24,
1142 perf_evsel__name(counter
));
1145 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
1147 fputc('\n', output
);
1151 if (nsec_counter(counter
))
1152 nsec_printout(-1, 0, counter
, avg
);
1154 abs_printout(-1, 0, counter
, avg
);
1156 print_noise(counter
, avg
);
1159 fputc('\n', output
);
1164 double avg_enabled
, avg_running
;
1166 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
1167 avg_running
= avg_stats(&ps
->res_stats
[2]);
1169 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
1171 fprintf(output
, "\n");
1175 * Print out the results of a single counter:
1176 * does not use aggregated count in system-wide
1178 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1183 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1184 val
= counter
->counts
->cpu
[cpu
].val
;
1185 ena
= counter
->counts
->cpu
[cpu
].ena
;
1186 run
= counter
->counts
->cpu
[cpu
].run
;
1189 fprintf(output
, "%s", prefix
);
1191 if (run
== 0 || ena
== 0) {
1192 fprintf(output
, "CPU%*d%s%*s%s%*s",
1193 csv_output
? 0 : -4,
1194 perf_evsel__cpus(counter
)->map
[cpu
], csv_sep
,
1195 csv_output
? 0 : 18,
1196 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1198 csv_output
? 0 : -24,
1199 perf_evsel__name(counter
));
1202 fprintf(output
, "%s%s",
1203 csv_sep
, counter
->cgrp
->name
);
1205 fputc('\n', output
);
1209 if (nsec_counter(counter
))
1210 nsec_printout(cpu
, 0, counter
, val
);
1212 abs_printout(cpu
, 0, counter
, val
);
1215 print_noise(counter
, 1.0);
1218 fprintf(output
, " (%.2f%%)",
1221 fputc('\n', output
);
1225 static void print_stat(int argc
, const char **argv
)
1227 struct perf_evsel
*counter
;
1233 fprintf(output
, "\n");
1234 fprintf(output
, " Performance counter stats for ");
1235 if (target
.system_wide
)
1236 fprintf(output
, "\'system wide");
1237 else if (target
.cpu_list
)
1238 fprintf(output
, "\'CPU(s) %s", target
.cpu_list
);
1239 else if (!perf_target__has_task(&target
)) {
1240 fprintf(output
, "\'%s", argv
[0]);
1241 for (i
= 1; i
< argc
; i
++)
1242 fprintf(output
, " %s", argv
[i
]);
1243 } else if (target
.pid
)
1244 fprintf(output
, "process id \'%s", target
.pid
);
1246 fprintf(output
, "thread id \'%s", target
.tid
);
1248 fprintf(output
, "\'");
1250 fprintf(output
, " (%d runs)", run_count
);
1251 fprintf(output
, ":\n\n");
1254 switch (aggr_mode
) {
1260 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1261 print_counter_aggr(counter
, NULL
);
1264 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1265 print_counter(counter
, NULL
);
1273 fprintf(output
, "\n");
1274 fprintf(output
, " %17.9f seconds time elapsed",
1275 avg_stats(&walltime_nsecs_stats
)/1e9
);
1276 if (run_count
> 1) {
1277 fprintf(output
, " ");
1278 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1279 avg_stats(&walltime_nsecs_stats
));
1281 fprintf(output
, "\n\n");
1285 static volatile int signr
= -1;
1287 static void skip_signal(int signo
)
1289 if ((child_pid
== -1) || interval
)
1294 * render child_pid harmless
1295 * won't send SIGTERM to a random
1296 * process in case of race condition
1297 * and fast PID recycling
1302 static void sig_atexit(void)
1307 * avoid race condition with SIGCHLD handler
1308 * in skip_signal() which is modifying child_pid
1309 * goal is to avoid send SIGTERM to a random
1313 sigaddset(&set
, SIGCHLD
);
1314 sigprocmask(SIG_BLOCK
, &set
, &oset
);
1316 if (child_pid
!= -1)
1317 kill(child_pid
, SIGTERM
);
1319 sigprocmask(SIG_SETMASK
, &oset
, NULL
);
1324 signal(signr
, SIG_DFL
);
1325 kill(getpid(), signr
);
1328 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1329 const char *s __maybe_unused
, int unset
)
1331 big_num_opt
= unset
? 0 : 1;
1335 static int perf_stat_init_aggr_mode(void)
1337 switch (aggr_mode
) {
1339 if (cpu_map__build_socket_map(evsel_list
->cpus
, &aggr_map
)) {
1340 perror("cannot build socket map");
1343 aggr_get_id
= cpu_map__get_socket
;
1346 if (cpu_map__build_core_map(evsel_list
->cpus
, &aggr_map
)) {
1347 perror("cannot build core map");
1350 aggr_get_id
= cpu_map__get_core
;
1360 static int setup_events(const char * const *attrs
, unsigned len
)
1364 for (i
= 0; i
< len
; i
++) {
1365 if (parse_events(evsel_list
, attrs
[i
]))
1372 * Add default attributes, if there were no attributes specified or
1373 * if -d/--detailed, -d -d or -d -d -d is used:
1375 static int add_default_attributes(void)
1377 struct perf_event_attr default_attrs
[] = {
1379 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1380 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1381 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1382 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1384 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1385 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1386 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1387 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1388 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1389 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1394 * Detailed stats (-d), covering the L1 and last level data caches:
1396 struct perf_event_attr detailed_attrs
[] = {
1398 { .type
= PERF_TYPE_HW_CACHE
,
1400 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1401 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1402 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1404 { .type
= PERF_TYPE_HW_CACHE
,
1406 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1407 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1408 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1410 { .type
= PERF_TYPE_HW_CACHE
,
1412 PERF_COUNT_HW_CACHE_LL
<< 0 |
1413 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1414 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1416 { .type
= PERF_TYPE_HW_CACHE
,
1418 PERF_COUNT_HW_CACHE_LL
<< 0 |
1419 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1420 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1424 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1426 struct perf_event_attr very_detailed_attrs
[] = {
1428 { .type
= PERF_TYPE_HW_CACHE
,
1430 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1431 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1432 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1434 { .type
= PERF_TYPE_HW_CACHE
,
1436 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1437 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1438 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1440 { .type
= PERF_TYPE_HW_CACHE
,
1442 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1443 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1444 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1446 { .type
= PERF_TYPE_HW_CACHE
,
1448 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1449 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1450 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1452 { .type
= PERF_TYPE_HW_CACHE
,
1454 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1455 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1456 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1458 { .type
= PERF_TYPE_HW_CACHE
,
1460 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1461 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1462 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1467 * Very, very detailed stats (-d -d -d), adding prefetch events:
1469 struct perf_event_attr very_very_detailed_attrs
[] = {
1471 { .type
= PERF_TYPE_HW_CACHE
,
1473 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1474 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1475 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1477 { .type
= PERF_TYPE_HW_CACHE
,
1479 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1480 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1481 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1484 /* Set attrs if no event is selected and !null_run: */
1488 if (transaction_run
) {
1490 if (pmu_have_event("cpu", "cycles-ct") &&
1491 pmu_have_event("cpu", "el-start"))
1492 err
= setup_events(transaction_attrs
,
1493 ARRAY_SIZE(transaction_attrs
));
1495 err
= setup_events(transaction_limited_attrs
,
1496 ARRAY_SIZE(transaction_limited_attrs
));
1498 fprintf(stderr
, "Cannot set up transaction events\n");
1504 if (!evsel_list
->nr_entries
) {
1505 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1509 /* Detailed events get appended to the event list: */
1511 if (detailed_run
< 1)
1514 /* Append detailed run extra attributes: */
1515 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1518 if (detailed_run
< 2)
1521 /* Append very detailed run extra attributes: */
1522 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1525 if (detailed_run
< 3)
1528 /* Append very, very detailed run extra attributes: */
1529 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1532 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1534 bool append_file
= false;
1536 const char *output_name
= NULL
;
1537 const struct option options
[] = {
1538 OPT_BOOLEAN('T', "transaction", &transaction_run
,
1539 "hardware transaction statistics"),
1540 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1541 "event selector. use 'perf list' to list available events",
1542 parse_events_option
),
1543 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1544 "event filter", parse_filter
),
1545 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1546 "child tasks do not inherit counters"),
1547 OPT_STRING('p', "pid", &target
.pid
, "pid",
1548 "stat events on existing process id"),
1549 OPT_STRING('t', "tid", &target
.tid
, "tid",
1550 "stat events on existing thread id"),
1551 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1552 "system-wide collection from all CPUs"),
1553 OPT_BOOLEAN('g', "group", &group
,
1554 "put the counters into a counter group"),
1555 OPT_BOOLEAN('c', "scale", &scale
, "scale/normalize counters"),
1556 OPT_INCR('v', "verbose", &verbose
,
1557 "be more verbose (show counter open errors, etc)"),
1558 OPT_INTEGER('r', "repeat", &run_count
,
1559 "repeat command and print average + stddev (max: 100, forever: 0)"),
1560 OPT_BOOLEAN('n', "null", &null_run
,
1561 "null run - dont start any counters"),
1562 OPT_INCR('d', "detailed", &detailed_run
,
1563 "detailed run - start a lot of events"),
1564 OPT_BOOLEAN('S', "sync", &sync_run
,
1565 "call sync() before starting a run"),
1566 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1567 "print large numbers with thousands\' separators",
1569 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1570 "list of cpus to monitor in system-wide"),
1571 OPT_SET_UINT('A', "no-aggr", &aggr_mode
,
1572 "disable CPU count aggregation", AGGR_NONE
),
1573 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1574 "print counts with custom separator"),
1575 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1576 "monitor event in cgroup name only", parse_cgroups
),
1577 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1578 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1579 OPT_INTEGER(0, "log-fd", &output_fd
,
1580 "log output to fd, instead of stderr"),
1581 OPT_STRING(0, "pre", &pre_cmd
, "command",
1582 "command to run prior to the measured command"),
1583 OPT_STRING(0, "post", &post_cmd
, "command",
1584 "command to run after to the measured command"),
1585 OPT_UINTEGER('I', "interval-print", &interval
,
1586 "print counts at regular interval in ms (>= 100)"),
1587 OPT_SET_UINT(0, "per-socket", &aggr_mode
,
1588 "aggregate counts per processor socket", AGGR_SOCKET
),
1589 OPT_SET_UINT(0, "per-core", &aggr_mode
,
1590 "aggregate counts per physical processor core", AGGR_CORE
),
1591 OPT_UINTEGER('D', "delay", &initial_delay
,
1592 "ms to wait before starting measurement after program start"),
1595 const char * const stat_usage
[] = {
1596 "perf stat [<options>] [<command>]",
1599 int status
= -EINVAL
, run_idx
;
1602 setlocale(LC_ALL
, "");
1604 evsel_list
= perf_evlist__new();
1605 if (evsel_list
== NULL
)
1608 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1609 PARSE_OPT_STOP_AT_NON_OPTION
);
1612 if (output_name
&& strcmp(output_name
, "-"))
1615 if (output_name
&& output_fd
) {
1616 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1617 parse_options_usage(stat_usage
, options
, "o", 1);
1618 parse_options_usage(NULL
, options
, "log-fd", 0);
1622 if (output_fd
< 0) {
1623 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1624 parse_options_usage(stat_usage
, options
, "log-fd", 0);
1630 mode
= append_file
? "a" : "w";
1632 output
= fopen(output_name
, mode
);
1634 perror("failed to create output file");
1637 clock_gettime(CLOCK_REALTIME
, &tm
);
1638 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1639 } else if (output_fd
> 0) {
1640 mode
= append_file
? "a" : "w";
1641 output
= fdopen(output_fd
, mode
);
1643 perror("Failed opening logfd");
1650 if (!strcmp(csv_sep
, "\\t"))
1653 csv_sep
= DEFAULT_SEPARATOR
;
1656 * let the spreadsheet do the pretty-printing
1659 /* User explicitly passed -B? */
1660 if (big_num_opt
== 1) {
1661 fprintf(stderr
, "-B option not supported with -x\n");
1662 parse_options_usage(stat_usage
, options
, "B", 1);
1663 parse_options_usage(NULL
, options
, "x", 1);
1665 } else /* Nope, so disable big number formatting */
1667 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1670 if (!argc
&& perf_target__none(&target
))
1671 usage_with_options(stat_usage
, options
);
1673 if (run_count
< 0) {
1674 pr_err("Run count must be a positive number\n");
1675 parse_options_usage(stat_usage
, options
, "r", 1);
1677 } else if (run_count
== 0) {
1682 /* no_aggr, cgroup are for system-wide only */
1683 if ((aggr_mode
!= AGGR_GLOBAL
|| nr_cgroups
)
1684 && !perf_target__has_cpu(&target
)) {
1685 fprintf(stderr
, "both cgroup and no-aggregation "
1686 "modes only available in system-wide mode\n");
1688 parse_options_usage(stat_usage
, options
, "G", 1);
1689 parse_options_usage(NULL
, options
, "A", 1);
1690 parse_options_usage(NULL
, options
, "a", 1);
1694 if (add_default_attributes())
1697 perf_target__validate(&target
);
1699 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1700 if (perf_target__has_task(&target
)) {
1701 pr_err("Problems finding threads of monitor\n");
1702 parse_options_usage(stat_usage
, options
, "p", 1);
1703 parse_options_usage(NULL
, options
, "t", 1);
1704 } else if (perf_target__has_cpu(&target
)) {
1705 perror("failed to parse CPUs map");
1706 parse_options_usage(stat_usage
, options
, "C", 1);
1707 parse_options_usage(NULL
, options
, "a", 1);
1711 if (interval
&& interval
< 100) {
1712 pr_err("print interval must be >= 100ms\n");
1713 parse_options_usage(stat_usage
, options
, "I", 1);
1717 if (perf_evlist__alloc_stats(evsel_list
, interval
))
1720 if (perf_stat_init_aggr_mode())
1724 * We dont want to block the signals - that would cause
1725 * child tasks to inherit that and Ctrl-C would not work.
1726 * What we want is for Ctrl-C to work in the exec()-ed
1727 * task, but being ignored by perf stat itself:
1731 signal(SIGINT
, skip_signal
);
1732 signal(SIGCHLD
, skip_signal
);
1733 signal(SIGALRM
, skip_signal
);
1734 signal(SIGABRT
, skip_signal
);
1737 for (run_idx
= 0; forever
|| run_idx
< run_count
; run_idx
++) {
1738 if (run_count
!= 1 && verbose
)
1739 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1742 status
= run_perf_stat(argc
, argv
);
1743 if (forever
&& status
!= -1) {
1744 print_stat(argc
, argv
);
1745 perf_stat__reset_stats(evsel_list
);
1749 if (!forever
&& status
!= -1 && !interval
)
1750 print_stat(argc
, argv
);
1752 perf_evlist__free_stats(evsel_list
);
1754 perf_evlist__delete_maps(evsel_list
);
1756 perf_evlist__delete(evsel_list
);