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_socket(char *prefix
);
73 static struct perf_evlist
*evsel_list
;
75 static struct perf_target target
= {
79 static int run_count
= 1;
80 static bool no_inherit
= false;
81 static bool scale
= true;
82 static bool no_aggr
= false;
83 static bool aggr_socket
= false;
84 static pid_t child_pid
= -1;
85 static bool null_run
= false;
86 static int detailed_run
= 0;
87 static bool big_num
= true;
88 static int big_num_opt
= -1;
89 static const char *csv_sep
= NULL
;
90 static bool csv_output
= false;
91 static bool group
= false;
92 static FILE *output
= NULL
;
93 static const char *pre_cmd
= NULL
;
94 static const char *post_cmd
= NULL
;
95 static bool sync_run
= false;
96 static unsigned int interval
= 0;
97 static struct timespec ref_time
;
98 static struct cpu_map
*sock_map
;
100 static volatile int done
= 0;
103 struct stats res_stats
[3];
106 static inline void diff_timespec(struct timespec
*r
, struct timespec
*a
,
109 r
->tv_sec
= a
->tv_sec
- b
->tv_sec
;
110 if (a
->tv_nsec
< b
->tv_nsec
) {
111 r
->tv_nsec
= a
->tv_nsec
+ 1000000000L - b
->tv_nsec
;
114 r
->tv_nsec
= a
->tv_nsec
- b
->tv_nsec
;
118 static inline struct cpu_map
*perf_evsel__cpus(struct perf_evsel
*evsel
)
120 return (evsel
->cpus
&& !target
.cpu_list
) ? evsel
->cpus
: evsel_list
->cpus
;
123 static inline int perf_evsel__nr_cpus(struct perf_evsel
*evsel
)
125 return perf_evsel__cpus(evsel
)->nr
;
128 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
130 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
131 return evsel
->priv
== NULL
? -ENOMEM
: 0;
134 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
140 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel
*evsel
)
145 sz
= sizeof(*evsel
->counts
) +
146 (perf_evsel__nr_cpus(evsel
) * sizeof(struct perf_counts_values
));
152 evsel
->prev_raw_counts
= addr
;
157 static void perf_evsel__free_prev_raw_counts(struct perf_evsel
*evsel
)
159 free(evsel
->prev_raw_counts
);
160 evsel
->prev_raw_counts
= NULL
;
163 static struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
164 static struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
165 static struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
166 static struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
167 static struct stats runtime_branches_stats
[MAX_NR_CPUS
];
168 static struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
169 static struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
170 static struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
171 static struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
172 static struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
173 static struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
174 static struct stats walltime_nsecs_stats
;
176 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
178 struct perf_event_attr
*attr
= &evsel
->attr
;
181 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
182 PERF_FORMAT_TOTAL_TIME_RUNNING
;
184 attr
->inherit
= !no_inherit
;
186 if (perf_target__has_cpu(&target
))
187 return perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
189 if (!perf_target__has_task(&target
) &&
190 perf_evsel__is_group_leader(evsel
)) {
192 attr
->enable_on_exec
= 1;
195 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
199 * Does the counter have nsecs as a unit?
201 static inline int nsec_counter(struct perf_evsel
*evsel
)
203 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
204 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
211 * Update various tracking values we maintain to print
212 * more semantic information such as miss/hit ratios,
213 * instruction rates, etc:
215 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
217 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
218 update_stats(&runtime_nsecs_stats
[0], count
[0]);
219 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
220 update_stats(&runtime_cycles_stats
[0], count
[0]);
221 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
222 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
223 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
224 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
225 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
226 update_stats(&runtime_branches_stats
[0], count
[0]);
227 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
228 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
229 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
230 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
231 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
232 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
233 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
234 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
235 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
236 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
237 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
238 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
242 * Read out the results of a single counter:
243 * aggregate counts across CPUs in system-wide mode
245 static int read_counter_aggr(struct perf_evsel
*counter
)
247 struct perf_stat
*ps
= counter
->priv
;
248 u64
*count
= counter
->counts
->aggr
.values
;
251 if (__perf_evsel__read(counter
, perf_evsel__nr_cpus(counter
),
252 evsel_list
->threads
->nr
, scale
) < 0)
255 for (i
= 0; i
< 3; i
++)
256 update_stats(&ps
->res_stats
[i
], count
[i
]);
259 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
260 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
264 * Save the full runtime - to allow normalization during printout:
266 update_shadow_stats(counter
, count
);
272 * Read out the results of a single counter:
273 * do not aggregate counts across CPUs in system-wide mode
275 static int read_counter(struct perf_evsel
*counter
)
280 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
281 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
284 count
= counter
->counts
->cpu
[cpu
].values
;
286 update_shadow_stats(counter
, count
);
292 static void print_interval(void)
294 static int num_print_interval
;
295 struct perf_evsel
*counter
;
296 struct perf_stat
*ps
;
297 struct timespec ts
, rs
;
301 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
303 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
304 read_counter(counter
);
307 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
309 memset(ps
->res_stats
, 0, sizeof(ps
->res_stats
));
310 read_counter_aggr(counter
);
313 clock_gettime(CLOCK_MONOTONIC
, &ts
);
314 diff_timespec(&rs
, &ts
, &ref_time
);
315 sprintf(prefix
, "%6lu.%09lu%s", rs
.tv_sec
, rs
.tv_nsec
, csv_sep
);
317 if (num_print_interval
== 0 && !csv_output
) {
319 fprintf(output
, "# time socket cpus counts events\n");
321 fprintf(output
, "# time CPU counts events\n");
323 fprintf(output
, "# time counts events\n");
326 if (++num_print_interval
== 25)
327 num_print_interval
= 0;
330 print_aggr_socket(prefix
);
332 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
333 print_counter(counter
, prefix
);
335 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
336 print_counter_aggr(counter
, prefix
);
340 static int __run_perf_stat(int argc __maybe_unused
, const char **argv
)
343 unsigned long long t0
, t1
;
344 struct perf_evsel
*counter
;
347 int child_ready_pipe
[2], go_pipe
[2];
348 const bool forks
= (argc
> 0);
352 ts
.tv_sec
= interval
/ 1000;
353 ts
.tv_nsec
= (interval
% 1000) * 1000000;
360 && cpu_map__build_socket_map(evsel_list
->cpus
, &sock_map
)) {
361 perror("cannot build socket map");
365 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
366 perror("failed to create pipes");
371 if ((child_pid
= fork()) < 0)
372 perror("failed to fork");
375 close(child_ready_pipe
[0]);
377 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
380 * Do a dummy execvp to get the PLT entry resolved,
381 * so we avoid the resolver overhead on the real
384 execvp("", (char **)argv
);
387 * Tell the parent we're ready to go
389 close(child_ready_pipe
[1]);
392 * Wait until the parent tells us to go.
394 if (read(go_pipe
[0], &buf
, 1) == -1)
395 perror("unable to read pipe");
397 execvp(argv
[0], (char **)argv
);
403 if (perf_target__none(&target
))
404 evsel_list
->threads
->map
[0] = child_pid
;
407 * Wait for the child to be ready to exec.
409 close(child_ready_pipe
[1]);
411 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
412 perror("unable to read pipe");
413 close(child_ready_pipe
[0]);
417 perf_evlist__set_leader(evsel_list
);
419 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
420 if (create_perf_stat_counter(counter
) < 0) {
422 * PPC returns ENXIO for HW counters until 2.6.37
423 * (behavior changed with commit b0a873e).
425 if (errno
== EINVAL
|| errno
== ENOSYS
||
426 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
429 ui__warning("%s event is not supported by the kernel.\n",
430 perf_evsel__name(counter
));
431 counter
->supported
= false;
435 perf_evsel__open_strerror(counter
, &target
,
436 errno
, msg
, sizeof(msg
));
437 ui__error("%s\n", msg
);
440 kill(child_pid
, SIGTERM
);
444 counter
->supported
= true;
447 if (perf_evlist__apply_filters(evsel_list
)) {
448 error("failed to set filter with %d (%s)\n", errno
,
454 * Enable counters and exec the command:
457 clock_gettime(CLOCK_MONOTONIC
, &ref_time
);
462 while (!waitpid(child_pid
, &status
, WNOHANG
)) {
463 nanosleep(&ts
, NULL
);
468 if (WIFSIGNALED(status
))
469 psignal(WTERMSIG(status
), argv
[0]);
472 nanosleep(&ts
, NULL
);
480 update_stats(&walltime_nsecs_stats
, t1
- t0
);
483 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
484 read_counter(counter
);
485 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
), 1);
488 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
489 read_counter_aggr(counter
);
490 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
491 evsel_list
->threads
->nr
);
495 return WEXITSTATUS(status
);
498 static int run_perf_stat(int argc __maybe_unused
, const char **argv
)
503 ret
= system(pre_cmd
);
511 ret
= __run_perf_stat(argc
, argv
);
516 ret
= system(post_cmd
);
524 static void print_noise_pct(double total
, double avg
)
526 double pct
= rel_stddev_stats(total
, avg
);
529 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
531 fprintf(output
, " ( +-%6.2f%% )", pct
);
534 static void print_noise(struct perf_evsel
*evsel
, double avg
)
536 struct perf_stat
*ps
;
542 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
545 static void nsec_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
547 double msecs
= avg
/ 1e6
;
548 char cpustr
[16] = { '\0', };
549 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
552 sprintf(cpustr
, "S%*d%s%*d%s",
560 sprintf(cpustr
, "CPU%*d%s",
562 perf_evsel__cpus(evsel
)->map
[cpu
], csv_sep
);
564 fprintf(output
, fmt
, cpustr
, msecs
, csv_sep
, perf_evsel__name(evsel
));
567 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
569 if (csv_output
|| interval
)
572 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
573 fprintf(output
, " # %8.3f CPUs utilized ",
574 avg
/ avg_stats(&walltime_nsecs_stats
));
576 fprintf(output
, " ");
579 /* used for get_ratio_color() */
581 GRC_STALLED_CYCLES_FE
,
582 GRC_STALLED_CYCLES_BE
,
587 static const char *get_ratio_color(enum grc_type type
, double ratio
)
589 static const double grc_table
[GRC_MAX_NR
][3] = {
590 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
591 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
592 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
594 const char *color
= PERF_COLOR_NORMAL
;
596 if (ratio
> grc_table
[type
][0])
597 color
= PERF_COLOR_RED
;
598 else if (ratio
> grc_table
[type
][1])
599 color
= PERF_COLOR_MAGENTA
;
600 else if (ratio
> grc_table
[type
][2])
601 color
= PERF_COLOR_YELLOW
;
606 static void print_stalled_cycles_frontend(int cpu
,
607 struct perf_evsel
*evsel
608 __maybe_unused
, double avg
)
610 double total
, ratio
= 0.0;
613 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
616 ratio
= avg
/ total
* 100.0;
618 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
620 fprintf(output
, " # ");
621 color_fprintf(output
, color
, "%6.2f%%", ratio
);
622 fprintf(output
, " frontend cycles idle ");
625 static void print_stalled_cycles_backend(int cpu
,
626 struct perf_evsel
*evsel
627 __maybe_unused
, double avg
)
629 double total
, ratio
= 0.0;
632 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
635 ratio
= avg
/ total
* 100.0;
637 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
639 fprintf(output
, " # ");
640 color_fprintf(output
, color
, "%6.2f%%", ratio
);
641 fprintf(output
, " backend cycles idle ");
644 static void print_branch_misses(int cpu
,
645 struct perf_evsel
*evsel __maybe_unused
,
648 double total
, ratio
= 0.0;
651 total
= avg_stats(&runtime_branches_stats
[cpu
]);
654 ratio
= avg
/ total
* 100.0;
656 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
658 fprintf(output
, " # ");
659 color_fprintf(output
, color
, "%6.2f%%", ratio
);
660 fprintf(output
, " of all branches ");
663 static void print_l1_dcache_misses(int cpu
,
664 struct perf_evsel
*evsel __maybe_unused
,
667 double total
, ratio
= 0.0;
670 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
673 ratio
= avg
/ total
* 100.0;
675 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
677 fprintf(output
, " # ");
678 color_fprintf(output
, color
, "%6.2f%%", ratio
);
679 fprintf(output
, " of all L1-dcache hits ");
682 static void print_l1_icache_misses(int cpu
,
683 struct perf_evsel
*evsel __maybe_unused
,
686 double total
, ratio
= 0.0;
689 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
692 ratio
= avg
/ total
* 100.0;
694 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
696 fprintf(output
, " # ");
697 color_fprintf(output
, color
, "%6.2f%%", ratio
);
698 fprintf(output
, " of all L1-icache hits ");
701 static void print_dtlb_cache_misses(int cpu
,
702 struct perf_evsel
*evsel __maybe_unused
,
705 double total
, ratio
= 0.0;
708 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
711 ratio
= avg
/ total
* 100.0;
713 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
715 fprintf(output
, " # ");
716 color_fprintf(output
, color
, "%6.2f%%", ratio
);
717 fprintf(output
, " of all dTLB cache hits ");
720 static void print_itlb_cache_misses(int cpu
,
721 struct perf_evsel
*evsel __maybe_unused
,
724 double total
, ratio
= 0.0;
727 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
730 ratio
= avg
/ total
* 100.0;
732 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
734 fprintf(output
, " # ");
735 color_fprintf(output
, color
, "%6.2f%%", ratio
);
736 fprintf(output
, " of all iTLB cache hits ");
739 static void print_ll_cache_misses(int cpu
,
740 struct perf_evsel
*evsel __maybe_unused
,
743 double total
, ratio
= 0.0;
746 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
749 ratio
= avg
/ total
* 100.0;
751 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
753 fprintf(output
, " # ");
754 color_fprintf(output
, color
, "%6.2f%%", ratio
);
755 fprintf(output
, " of all LL-cache hits ");
758 static void abs_printout(int cpu
, int nr
, struct perf_evsel
*evsel
, double avg
)
760 double total
, ratio
= 0.0;
761 char cpustr
[16] = { '\0', };
767 fmt
= "%s%'18.0f%s%-25s";
769 fmt
= "%s%18.0f%s%-25s";
772 sprintf(cpustr
, "S%*d%s%*d%s",
780 sprintf(cpustr
, "CPU%*d%s",
782 perf_evsel__cpus(evsel
)->map
[cpu
], csv_sep
);
786 fprintf(output
, fmt
, cpustr
, avg
, csv_sep
, perf_evsel__name(evsel
));
789 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
791 if (csv_output
|| interval
)
794 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
795 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
799 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
801 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
802 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
806 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
809 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
810 runtime_branches_stats
[cpu
].n
!= 0) {
811 print_branch_misses(cpu
, evsel
, avg
);
813 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
814 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
815 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
816 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
817 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
818 print_l1_dcache_misses(cpu
, evsel
, avg
);
820 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
821 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
822 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
823 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
824 runtime_l1_icache_stats
[cpu
].n
!= 0) {
825 print_l1_icache_misses(cpu
, evsel
, avg
);
827 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
828 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
829 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
830 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
831 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
832 print_dtlb_cache_misses(cpu
, evsel
, avg
);
834 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
835 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
836 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
837 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
838 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
839 print_itlb_cache_misses(cpu
, evsel
, avg
);
841 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
842 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
843 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
844 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
845 runtime_ll_cache_stats
[cpu
].n
!= 0) {
846 print_ll_cache_misses(cpu
, evsel
, avg
);
847 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
848 runtime_cacherefs_stats
[cpu
].n
!= 0) {
849 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
852 ratio
= avg
* 100 / total
;
854 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
856 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
857 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
858 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
859 print_stalled_cycles_backend(cpu
, evsel
, avg
);
860 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
861 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
864 ratio
= 1.0 * avg
/ total
;
866 fprintf(output
, " # %8.3f GHz ", ratio
);
867 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
870 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
873 ratio
= 1000.0 * avg
/ total
;
879 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
881 fprintf(output
, " ");
885 static void print_aggr_socket(char *prefix
)
887 struct perf_evsel
*counter
;
889 int cpu
, s
, s2
, sock
, nr
;
894 for (s
= 0; s
< sock_map
->nr
; s
++) {
895 sock
= cpu_map__socket(sock_map
, s
);
896 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
899 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
900 s2
= cpu_map__get_socket(evsel_list
->cpus
, cpu
);
903 val
+= counter
->counts
->cpu
[cpu
].val
;
904 ena
+= counter
->counts
->cpu
[cpu
].ena
;
905 run
+= counter
->counts
->cpu
[cpu
].run
;
909 fprintf(output
, "%s", prefix
);
911 if (run
== 0 || ena
== 0) {
912 fprintf(output
, "S%*d%s%*d%s%*s%s%*s",
920 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
922 csv_output
? 0 : -24,
923 perf_evsel__name(counter
));
925 fprintf(output
, "%s%s",
926 csv_sep
, counter
->cgrp
->name
);
932 if (nsec_counter(counter
))
933 nsec_printout(sock
, nr
, counter
, val
);
935 abs_printout(sock
, nr
, counter
, val
);
938 print_noise(counter
, 1.0);
941 fprintf(output
, " (%.2f%%)",
950 * Print out the results of a single counter:
951 * aggregated counts in system-wide mode
953 static void print_counter_aggr(struct perf_evsel
*counter
, char *prefix
)
955 struct perf_stat
*ps
= counter
->priv
;
956 double avg
= avg_stats(&ps
->res_stats
[0]);
957 int scaled
= counter
->counts
->scaled
;
960 fprintf(output
, "%s", prefix
);
963 fprintf(output
, "%*s%s%*s",
965 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
967 csv_output
? 0 : -24,
968 perf_evsel__name(counter
));
971 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
977 if (nsec_counter(counter
))
978 nsec_printout(-1, 0, counter
, avg
);
980 abs_printout(-1, 0, counter
, avg
);
982 print_noise(counter
, avg
);
990 double avg_enabled
, avg_running
;
992 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
993 avg_running
= avg_stats(&ps
->res_stats
[2]);
995 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
997 fprintf(output
, "\n");
1001 * Print out the results of a single counter:
1002 * does not use aggregated count in system-wide
1004 static void print_counter(struct perf_evsel
*counter
, char *prefix
)
1009 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
1010 val
= counter
->counts
->cpu
[cpu
].val
;
1011 ena
= counter
->counts
->cpu
[cpu
].ena
;
1012 run
= counter
->counts
->cpu
[cpu
].run
;
1015 fprintf(output
, "%s", prefix
);
1017 if (run
== 0 || ena
== 0) {
1018 fprintf(output
, "CPU%*d%s%*s%s%*s",
1019 csv_output
? 0 : -4,
1020 perf_evsel__cpus(counter
)->map
[cpu
], csv_sep
,
1021 csv_output
? 0 : 18,
1022 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
1024 csv_output
? 0 : -24,
1025 perf_evsel__name(counter
));
1028 fprintf(output
, "%s%s",
1029 csv_sep
, counter
->cgrp
->name
);
1031 fputc('\n', output
);
1035 if (nsec_counter(counter
))
1036 nsec_printout(cpu
, 0, counter
, val
);
1038 abs_printout(cpu
, 0, counter
, val
);
1041 print_noise(counter
, 1.0);
1044 fprintf(output
, " (%.2f%%)",
1047 fputc('\n', output
);
1051 static void print_stat(int argc
, const char **argv
)
1053 struct perf_evsel
*counter
;
1059 fprintf(output
, "\n");
1060 fprintf(output
, " Performance counter stats for ");
1061 if (!perf_target__has_task(&target
)) {
1062 fprintf(output
, "\'%s", argv
[0]);
1063 for (i
= 1; i
< argc
; i
++)
1064 fprintf(output
, " %s", argv
[i
]);
1065 } else if (target
.pid
)
1066 fprintf(output
, "process id \'%s", target
.pid
);
1068 fprintf(output
, "thread id \'%s", target
.tid
);
1070 fprintf(output
, "\'");
1072 fprintf(output
, " (%d runs)", run_count
);
1073 fprintf(output
, ":\n\n");
1077 print_aggr_socket(NULL
);
1079 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1080 print_counter(counter
, NULL
);
1082 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1083 print_counter_aggr(counter
, NULL
);
1088 fprintf(output
, "\n");
1089 fprintf(output
, " %17.9f seconds time elapsed",
1090 avg_stats(&walltime_nsecs_stats
)/1e9
);
1091 if (run_count
> 1) {
1092 fprintf(output
, " ");
1093 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1094 avg_stats(&walltime_nsecs_stats
));
1096 fprintf(output
, "\n\n");
1100 static volatile int signr
= -1;
1102 static void skip_signal(int signo
)
1104 if ((child_pid
== -1) || interval
)
1110 static void sig_atexit(void)
1112 if (child_pid
!= -1)
1113 kill(child_pid
, SIGTERM
);
1118 signal(signr
, SIG_DFL
);
1119 kill(getpid(), signr
);
1122 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
1123 const char *s __maybe_unused
, int unset
)
1125 big_num_opt
= unset
? 0 : 1;
1130 * Add default attributes, if there were no attributes specified or
1131 * if -d/--detailed, -d -d or -d -d -d is used:
1133 static int add_default_attributes(void)
1135 struct perf_event_attr default_attrs
[] = {
1137 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
1138 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
1139 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
1140 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
1142 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
1143 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
1144 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
1145 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
1146 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
1147 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
1152 * Detailed stats (-d), covering the L1 and last level data caches:
1154 struct perf_event_attr detailed_attrs
[] = {
1156 { .type
= PERF_TYPE_HW_CACHE
,
1158 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1159 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1160 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1162 { .type
= PERF_TYPE_HW_CACHE
,
1164 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1165 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1166 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1168 { .type
= PERF_TYPE_HW_CACHE
,
1170 PERF_COUNT_HW_CACHE_LL
<< 0 |
1171 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1172 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1174 { .type
= PERF_TYPE_HW_CACHE
,
1176 PERF_COUNT_HW_CACHE_LL
<< 0 |
1177 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1178 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1182 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1184 struct perf_event_attr very_detailed_attrs
[] = {
1186 { .type
= PERF_TYPE_HW_CACHE
,
1188 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1189 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1190 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1192 { .type
= PERF_TYPE_HW_CACHE
,
1194 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1195 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1196 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1198 { .type
= PERF_TYPE_HW_CACHE
,
1200 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1201 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1202 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1204 { .type
= PERF_TYPE_HW_CACHE
,
1206 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1207 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1208 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1210 { .type
= PERF_TYPE_HW_CACHE
,
1212 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1213 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1214 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1216 { .type
= PERF_TYPE_HW_CACHE
,
1218 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1219 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1220 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1225 * Very, very detailed stats (-d -d -d), adding prefetch events:
1227 struct perf_event_attr very_very_detailed_attrs
[] = {
1229 { .type
= PERF_TYPE_HW_CACHE
,
1231 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1232 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1233 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1235 { .type
= PERF_TYPE_HW_CACHE
,
1237 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1238 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1239 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1242 /* Set attrs if no event is selected and !null_run: */
1246 if (!evsel_list
->nr_entries
) {
1247 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1251 /* Detailed events get appended to the event list: */
1253 if (detailed_run
< 1)
1256 /* Append detailed run extra attributes: */
1257 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1260 if (detailed_run
< 2)
1263 /* Append very detailed run extra attributes: */
1264 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1267 if (detailed_run
< 3)
1270 /* Append very, very detailed run extra attributes: */
1271 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1274 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1276 bool append_file
= false;
1278 const char *output_name
= NULL
;
1279 const struct option options
[] = {
1280 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1281 "event selector. use 'perf list' to list available events",
1282 parse_events_option
),
1283 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1284 "event filter", parse_filter
),
1285 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1286 "child tasks do not inherit counters"),
1287 OPT_STRING('p', "pid", &target
.pid
, "pid",
1288 "stat events on existing process id"),
1289 OPT_STRING('t', "tid", &target
.tid
, "tid",
1290 "stat events on existing thread id"),
1291 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1292 "system-wide collection from all CPUs"),
1293 OPT_BOOLEAN('g', "group", &group
,
1294 "put the counters into a counter group"),
1295 OPT_BOOLEAN('c', "scale", &scale
, "scale/normalize counters"),
1296 OPT_INCR('v', "verbose", &verbose
,
1297 "be more verbose (show counter open errors, etc)"),
1298 OPT_INTEGER('r', "repeat", &run_count
,
1299 "repeat command and print average + stddev (max: 100)"),
1300 OPT_BOOLEAN('n', "null", &null_run
,
1301 "null run - dont start any counters"),
1302 OPT_INCR('d', "detailed", &detailed_run
,
1303 "detailed run - start a lot of events"),
1304 OPT_BOOLEAN('S', "sync", &sync_run
,
1305 "call sync() before starting a run"),
1306 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1307 "print large numbers with thousands\' separators",
1309 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1310 "list of cpus to monitor in system-wide"),
1311 OPT_BOOLEAN('A', "no-aggr", &no_aggr
, "disable CPU count aggregation"),
1312 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1313 "print counts with custom separator"),
1314 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1315 "monitor event in cgroup name only", parse_cgroups
),
1316 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1317 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1318 OPT_INTEGER(0, "log-fd", &output_fd
,
1319 "log output to fd, instead of stderr"),
1320 OPT_STRING(0, "pre", &pre_cmd
, "command",
1321 "command to run prior to the measured command"),
1322 OPT_STRING(0, "post", &post_cmd
, "command",
1323 "command to run after to the measured command"),
1324 OPT_UINTEGER('I', "interval-print", &interval
,
1325 "print counts at regular interval in ms (>= 100)"),
1326 OPT_BOOLEAN(0, "aggr-socket", &aggr_socket
, "aggregate counts per processor socket"),
1329 const char * const stat_usage
[] = {
1330 "perf stat [<options>] [<command>]",
1333 struct perf_evsel
*pos
;
1334 int status
= -ENOMEM
, run_idx
;
1337 setlocale(LC_ALL
, "");
1339 evsel_list
= perf_evlist__new(NULL
, NULL
);
1340 if (evsel_list
== NULL
)
1343 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1344 PARSE_OPT_STOP_AT_NON_OPTION
);
1347 if (output_name
&& strcmp(output_name
, "-"))
1350 if (output_name
&& output_fd
) {
1351 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1352 usage_with_options(stat_usage
, options
);
1355 if (output_fd
< 0) {
1356 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1357 usage_with_options(stat_usage
, options
);
1362 mode
= append_file
? "a" : "w";
1364 output
= fopen(output_name
, mode
);
1366 perror("failed to create output file");
1369 clock_gettime(CLOCK_REALTIME
, &tm
);
1370 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1371 } else if (output_fd
> 0) {
1372 mode
= append_file
? "a" : "w";
1373 output
= fdopen(output_fd
, mode
);
1375 perror("Failed opening logfd");
1382 if (!strcmp(csv_sep
, "\\t"))
1385 csv_sep
= DEFAULT_SEPARATOR
;
1388 * let the spreadsheet do the pretty-printing
1391 /* User explicitly passed -B? */
1392 if (big_num_opt
== 1) {
1393 fprintf(stderr
, "-B option not supported with -x\n");
1394 usage_with_options(stat_usage
, options
);
1395 } else /* Nope, so disable big number formatting */
1397 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1400 if (!argc
&& !perf_target__has_task(&target
))
1401 usage_with_options(stat_usage
, options
);
1403 usage_with_options(stat_usage
, options
);
1405 /* no_aggr, cgroup are for system-wide only */
1406 if ((no_aggr
|| nr_cgroups
) && !perf_target__has_cpu(&target
)) {
1407 fprintf(stderr
, "both cgroup and no-aggregation "
1408 "modes only available in system-wide mode\n");
1410 usage_with_options(stat_usage
, options
);
1414 if (!perf_target__has_cpu(&target
)) {
1415 fprintf(stderr
, "--aggr-socket only available in system-wide mode (-a)\n");
1416 usage_with_options(stat_usage
, options
);
1421 if (add_default_attributes())
1424 perf_target__validate(&target
);
1426 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1427 if (perf_target__has_task(&target
))
1428 pr_err("Problems finding threads of monitor\n");
1429 if (perf_target__has_cpu(&target
))
1430 perror("failed to parse CPUs map");
1432 usage_with_options(stat_usage
, options
);
1435 if (interval
&& interval
< 100) {
1436 pr_err("print interval must be >= 100ms\n");
1437 usage_with_options(stat_usage
, options
);
1441 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1442 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
1443 perf_evsel__alloc_counts(pos
, perf_evsel__nr_cpus(pos
)) < 0)
1447 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1448 if (perf_evsel__alloc_prev_raw_counts(pos
) < 0)
1454 * We dont want to block the signals - that would cause
1455 * child tasks to inherit that and Ctrl-C would not work.
1456 * What we want is for Ctrl-C to work in the exec()-ed
1457 * task, but being ignored by perf stat itself:
1460 signal(SIGINT
, skip_signal
);
1461 signal(SIGCHLD
, skip_signal
);
1462 signal(SIGALRM
, skip_signal
);
1463 signal(SIGABRT
, skip_signal
);
1466 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
1467 if (run_count
!= 1 && verbose
)
1468 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1471 status
= run_perf_stat(argc
, argv
);
1474 if (status
!= -1 && !interval
)
1475 print_stat(argc
, argv
);
1477 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1478 perf_evsel__free_stat_priv(pos
);
1479 perf_evsel__free_counts(pos
);
1480 perf_evsel__free_prev_raw_counts(pos
);
1482 perf_evlist__delete_maps(evsel_list
);
1484 perf_evlist__delete(evsel_list
);