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 struct perf_evlist
*evsel_list
;
70 static struct perf_target target
= {
74 static int run_count
= 1;
75 static bool no_inherit
= false;
76 static bool scale
= true;
77 static bool no_aggr
= false;
78 static pid_t child_pid
= -1;
79 static bool null_run
= false;
80 static int detailed_run
= 0;
81 static bool big_num
= true;
82 static int big_num_opt
= -1;
83 static const char *csv_sep
= NULL
;
84 static bool csv_output
= false;
85 static bool group
= false;
86 static FILE *output
= NULL
;
87 static const char *pre_cmd
= NULL
;
88 static const char *post_cmd
= NULL
;
89 static bool sync_run
= false;
91 static volatile int done
= 0;
94 struct stats res_stats
[3];
97 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
99 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
100 return evsel
->priv
== NULL
? -ENOMEM
: 0;
103 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
109 static inline struct cpu_map
*perf_evsel__cpus(struct perf_evsel
*evsel
)
111 return (evsel
->cpus
&& !target
.cpu_list
) ? evsel
->cpus
: evsel_list
->cpus
;
114 static inline int perf_evsel__nr_cpus(struct perf_evsel
*evsel
)
116 return perf_evsel__cpus(evsel
)->nr
;
119 static struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
120 static struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
121 static struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
122 static struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
123 static struct stats runtime_branches_stats
[MAX_NR_CPUS
];
124 static struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
125 static struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
126 static struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
127 static struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
128 static struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
129 static struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
130 static struct stats walltime_nsecs_stats
;
132 static int create_perf_stat_counter(struct perf_evsel
*evsel
,
133 struct perf_evsel
*first
)
135 struct perf_event_attr
*attr
= &evsel
->attr
;
136 bool exclude_guest_missing
= false;
140 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
141 PERF_FORMAT_TOTAL_TIME_RUNNING
;
143 attr
->inherit
= !no_inherit
;
146 if (exclude_guest_missing
)
147 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
149 if (perf_target__has_cpu(&target
)) {
150 ret
= perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
156 if (!perf_target__has_task(&target
) && (!group
|| evsel
== first
)) {
158 attr
->enable_on_exec
= 1;
161 ret
= perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
166 if (ret
&& errno
== EINVAL
) {
167 if (!exclude_guest_missing
&&
168 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
169 pr_debug("Old kernel, cannot exclude "
170 "guest or host samples.\n");
171 exclude_guest_missing
= true;
179 * Does the counter have nsecs as a unit?
181 static inline int nsec_counter(struct perf_evsel
*evsel
)
183 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
184 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
191 * Update various tracking values we maintain to print
192 * more semantic information such as miss/hit ratios,
193 * instruction rates, etc:
195 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
197 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
198 update_stats(&runtime_nsecs_stats
[0], count
[0]);
199 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
200 update_stats(&runtime_cycles_stats
[0], count
[0]);
201 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
202 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
203 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
204 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
205 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
206 update_stats(&runtime_branches_stats
[0], count
[0]);
207 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
208 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
209 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
210 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
211 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
212 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
213 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
214 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
215 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
216 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
217 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
218 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
222 * Read out the results of a single counter:
223 * aggregate counts across CPUs in system-wide mode
225 static int read_counter_aggr(struct perf_evsel
*counter
)
227 struct perf_stat
*ps
= counter
->priv
;
228 u64
*count
= counter
->counts
->aggr
.values
;
231 if (__perf_evsel__read(counter
, perf_evsel__nr_cpus(counter
),
232 evsel_list
->threads
->nr
, scale
) < 0)
235 for (i
= 0; i
< 3; i
++)
236 update_stats(&ps
->res_stats
[i
], count
[i
]);
239 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
240 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
244 * Save the full runtime - to allow normalization during printout:
246 update_shadow_stats(counter
, count
);
252 * Read out the results of a single counter:
253 * do not aggregate counts across CPUs in system-wide mode
255 static int read_counter(struct perf_evsel
*counter
)
260 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
261 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
264 count
= counter
->counts
->cpu
[cpu
].values
;
266 update_shadow_stats(counter
, count
);
272 static int __run_perf_stat(int argc __maybe_unused
, const char **argv
)
274 unsigned long long t0
, t1
;
275 struct perf_evsel
*counter
, *first
;
277 int child_ready_pipe
[2], go_pipe
[2];
278 const bool forks
= (argc
> 0);
281 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
282 perror("failed to create pipes");
287 if ((child_pid
= fork()) < 0)
288 perror("failed to fork");
291 close(child_ready_pipe
[0]);
293 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
296 * Do a dummy execvp to get the PLT entry resolved,
297 * so we avoid the resolver overhead on the real
300 execvp("", (char **)argv
);
303 * Tell the parent we're ready to go
305 close(child_ready_pipe
[1]);
308 * Wait until the parent tells us to go.
310 if (read(go_pipe
[0], &buf
, 1) == -1)
311 perror("unable to read pipe");
313 execvp(argv
[0], (char **)argv
);
319 if (perf_target__none(&target
))
320 evsel_list
->threads
->map
[0] = child_pid
;
323 * Wait for the child to be ready to exec.
325 close(child_ready_pipe
[1]);
327 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
328 perror("unable to read pipe");
329 close(child_ready_pipe
[0]);
333 perf_evlist__set_leader(evsel_list
);
335 first
= perf_evlist__first(evsel_list
);
337 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
338 if (create_perf_stat_counter(counter
, first
) < 0) {
340 * PPC returns ENXIO for HW counters until 2.6.37
341 * (behavior changed with commit b0a873e).
343 if (errno
== EINVAL
|| errno
== ENOSYS
||
344 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
347 ui__warning("%s event is not supported by the kernel.\n",
348 perf_evsel__name(counter
));
349 counter
->supported
= false;
353 if (errno
== EPERM
|| errno
== EACCES
) {
354 error("You may not have permission to collect %sstats.\n"
355 "\t Consider tweaking"
356 " /proc/sys/kernel/perf_event_paranoid or running as root.",
357 target
.system_wide
? "system-wide " : "");
359 error("open_counter returned with %d (%s). "
360 "/bin/dmesg may provide additional information.\n",
361 errno
, strerror(errno
));
364 kill(child_pid
, SIGTERM
);
366 pr_err("Not all events could be opened.\n");
369 counter
->supported
= true;
372 if (perf_evlist__apply_filters(evsel_list
)) {
373 error("failed to set filter with %d (%s)\n", errno
,
379 * Enable counters and exec the command:
386 if (WIFSIGNALED(status
))
387 psignal(WTERMSIG(status
), argv
[0]);
389 while(!done
) sleep(1);
394 update_stats(&walltime_nsecs_stats
, t1
- t0
);
397 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
398 read_counter(counter
);
399 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
), 1);
402 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
403 read_counter_aggr(counter
);
404 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
405 evsel_list
->threads
->nr
);
409 return WEXITSTATUS(status
);
412 static int run_perf_stat(int argc __maybe_unused
, const char **argv
)
417 ret
= system(pre_cmd
);
425 ret
= __run_perf_stat(argc
, argv
);
430 ret
= system(post_cmd
);
438 static void print_noise_pct(double total
, double avg
)
440 double pct
= rel_stddev_stats(total
, avg
);
443 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
445 fprintf(output
, " ( +-%6.2f%% )", pct
);
448 static void print_noise(struct perf_evsel
*evsel
, double avg
)
450 struct perf_stat
*ps
;
456 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
459 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
461 double msecs
= avg
/ 1e6
;
462 char cpustr
[16] = { '\0', };
463 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
466 sprintf(cpustr
, "CPU%*d%s",
468 perf_evsel__cpus(evsel
)->map
[cpu
], csv_sep
);
470 fprintf(output
, fmt
, cpustr
, msecs
, csv_sep
, perf_evsel__name(evsel
));
473 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
478 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
479 fprintf(output
, " # %8.3f CPUs utilized ",
480 avg
/ avg_stats(&walltime_nsecs_stats
));
482 fprintf(output
, " ");
485 /* used for get_ratio_color() */
487 GRC_STALLED_CYCLES_FE
,
488 GRC_STALLED_CYCLES_BE
,
493 static const char *get_ratio_color(enum grc_type type
, double ratio
)
495 static const double grc_table
[GRC_MAX_NR
][3] = {
496 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
497 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
498 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
500 const char *color
= PERF_COLOR_NORMAL
;
502 if (ratio
> grc_table
[type
][0])
503 color
= PERF_COLOR_RED
;
504 else if (ratio
> grc_table
[type
][1])
505 color
= PERF_COLOR_MAGENTA
;
506 else if (ratio
> grc_table
[type
][2])
507 color
= PERF_COLOR_YELLOW
;
512 static void print_stalled_cycles_frontend(int cpu
,
513 struct perf_evsel
*evsel
514 __maybe_unused
, double avg
)
516 double total
, ratio
= 0.0;
519 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
522 ratio
= avg
/ total
* 100.0;
524 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
526 fprintf(output
, " # ");
527 color_fprintf(output
, color
, "%6.2f%%", ratio
);
528 fprintf(output
, " frontend cycles idle ");
531 static void print_stalled_cycles_backend(int cpu
,
532 struct perf_evsel
*evsel
533 __maybe_unused
, double avg
)
535 double total
, ratio
= 0.0;
538 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
541 ratio
= avg
/ total
* 100.0;
543 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
545 fprintf(output
, " # ");
546 color_fprintf(output
, color
, "%6.2f%%", ratio
);
547 fprintf(output
, " backend cycles idle ");
550 static void print_branch_misses(int cpu
,
551 struct perf_evsel
*evsel __maybe_unused
,
554 double total
, ratio
= 0.0;
557 total
= avg_stats(&runtime_branches_stats
[cpu
]);
560 ratio
= avg
/ total
* 100.0;
562 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
564 fprintf(output
, " # ");
565 color_fprintf(output
, color
, "%6.2f%%", ratio
);
566 fprintf(output
, " of all branches ");
569 static void print_l1_dcache_misses(int cpu
,
570 struct perf_evsel
*evsel __maybe_unused
,
573 double total
, ratio
= 0.0;
576 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
579 ratio
= avg
/ total
* 100.0;
581 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
583 fprintf(output
, " # ");
584 color_fprintf(output
, color
, "%6.2f%%", ratio
);
585 fprintf(output
, " of all L1-dcache hits ");
588 static void print_l1_icache_misses(int cpu
,
589 struct perf_evsel
*evsel __maybe_unused
,
592 double total
, ratio
= 0.0;
595 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
598 ratio
= avg
/ total
* 100.0;
600 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
602 fprintf(output
, " # ");
603 color_fprintf(output
, color
, "%6.2f%%", ratio
);
604 fprintf(output
, " of all L1-icache hits ");
607 static void print_dtlb_cache_misses(int cpu
,
608 struct perf_evsel
*evsel __maybe_unused
,
611 double total
, ratio
= 0.0;
614 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
617 ratio
= avg
/ total
* 100.0;
619 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
621 fprintf(output
, " # ");
622 color_fprintf(output
, color
, "%6.2f%%", ratio
);
623 fprintf(output
, " of all dTLB cache hits ");
626 static void print_itlb_cache_misses(int cpu
,
627 struct perf_evsel
*evsel __maybe_unused
,
630 double total
, ratio
= 0.0;
633 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
636 ratio
= avg
/ total
* 100.0;
638 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
640 fprintf(output
, " # ");
641 color_fprintf(output
, color
, "%6.2f%%", ratio
);
642 fprintf(output
, " of all iTLB cache hits ");
645 static void print_ll_cache_misses(int cpu
,
646 struct perf_evsel
*evsel __maybe_unused
,
649 double total
, ratio
= 0.0;
652 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
655 ratio
= avg
/ total
* 100.0;
657 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
659 fprintf(output
, " # ");
660 color_fprintf(output
, color
, "%6.2f%%", ratio
);
661 fprintf(output
, " of all LL-cache hits ");
664 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
666 double total
, ratio
= 0.0;
667 char cpustr
[16] = { '\0', };
673 fmt
= "%s%'18.0f%s%-25s";
675 fmt
= "%s%18.0f%s%-25s";
678 sprintf(cpustr
, "CPU%*d%s",
680 perf_evsel__cpus(evsel
)->map
[cpu
], csv_sep
);
684 fprintf(output
, fmt
, cpustr
, avg
, csv_sep
, perf_evsel__name(evsel
));
687 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
692 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
693 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
698 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
700 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
701 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
705 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
708 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
709 runtime_branches_stats
[cpu
].n
!= 0) {
710 print_branch_misses(cpu
, evsel
, avg
);
712 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
713 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
714 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
715 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
716 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
717 print_l1_dcache_misses(cpu
, evsel
, avg
);
719 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
720 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
721 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
722 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
723 runtime_l1_icache_stats
[cpu
].n
!= 0) {
724 print_l1_icache_misses(cpu
, evsel
, avg
);
726 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
727 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
728 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
729 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
730 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
731 print_dtlb_cache_misses(cpu
, evsel
, avg
);
733 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
734 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
735 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
736 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
737 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
738 print_itlb_cache_misses(cpu
, evsel
, avg
);
740 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
741 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
742 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
743 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
744 runtime_ll_cache_stats
[cpu
].n
!= 0) {
745 print_ll_cache_misses(cpu
, evsel
, avg
);
746 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
747 runtime_cacherefs_stats
[cpu
].n
!= 0) {
748 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
751 ratio
= avg
* 100 / total
;
753 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
755 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
756 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
757 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
758 print_stalled_cycles_backend(cpu
, evsel
, avg
);
759 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
760 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
763 ratio
= 1.0 * avg
/ total
;
765 fprintf(output
, " # %8.3f GHz ", ratio
);
766 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
769 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
772 ratio
= 1000.0 * avg
/ total
;
778 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
780 fprintf(output
, " ");
785 * Print out the results of a single counter:
786 * aggregated counts in system-wide mode
788 static void print_counter_aggr(struct perf_evsel
*counter
)
790 struct perf_stat
*ps
= counter
->priv
;
791 double avg
= avg_stats(&ps
->res_stats
[0]);
792 int scaled
= counter
->counts
->scaled
;
795 fprintf(output
, "%*s%s%*s",
797 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
799 csv_output
? 0 : -24,
800 perf_evsel__name(counter
));
803 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
809 if (nsec_counter(counter
))
810 nsec_printout(-1, counter
, avg
);
812 abs_printout(-1, counter
, avg
);
814 print_noise(counter
, avg
);
822 double avg_enabled
, avg_running
;
824 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
825 avg_running
= avg_stats(&ps
->res_stats
[2]);
827 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
829 fprintf(output
, "\n");
833 * Print out the results of a single counter:
834 * does not use aggregated count in system-wide
836 static void print_counter(struct perf_evsel
*counter
)
841 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
842 val
= counter
->counts
->cpu
[cpu
].val
;
843 ena
= counter
->counts
->cpu
[cpu
].ena
;
844 run
= counter
->counts
->cpu
[cpu
].run
;
845 if (run
== 0 || ena
== 0) {
846 fprintf(output
, "CPU%*d%s%*s%s%*s",
848 perf_evsel__cpus(counter
)->map
[cpu
], csv_sep
,
850 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
852 csv_output
? 0 : -24,
853 perf_evsel__name(counter
));
856 fprintf(output
, "%s%s",
857 csv_sep
, counter
->cgrp
->name
);
863 if (nsec_counter(counter
))
864 nsec_printout(cpu
, counter
, val
);
866 abs_printout(cpu
, counter
, val
);
869 print_noise(counter
, 1.0);
872 fprintf(output
, " (%.2f%%)",
879 static void print_stat(int argc
, const char **argv
)
881 struct perf_evsel
*counter
;
887 fprintf(output
, "\n");
888 fprintf(output
, " Performance counter stats for ");
889 if (!perf_target__has_task(&target
)) {
890 fprintf(output
, "\'%s", argv
[0]);
891 for (i
= 1; i
< argc
; i
++)
892 fprintf(output
, " %s", argv
[i
]);
893 } else if (target
.pid
)
894 fprintf(output
, "process id \'%s", target
.pid
);
896 fprintf(output
, "thread id \'%s", target
.tid
);
898 fprintf(output
, "\'");
900 fprintf(output
, " (%d runs)", run_count
);
901 fprintf(output
, ":\n\n");
905 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
906 print_counter(counter
);
908 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
909 print_counter_aggr(counter
);
914 fprintf(output
, "\n");
915 fprintf(output
, " %17.9f seconds time elapsed",
916 avg_stats(&walltime_nsecs_stats
)/1e9
);
918 fprintf(output
, " ");
919 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
920 avg_stats(&walltime_nsecs_stats
));
922 fprintf(output
, "\n\n");
926 static volatile int signr
= -1;
928 static void skip_signal(int signo
)
936 static void sig_atexit(void)
939 kill(child_pid
, SIGTERM
);
944 signal(signr
, SIG_DFL
);
945 kill(getpid(), signr
);
948 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
949 const char *s __maybe_unused
, int unset
)
951 big_num_opt
= unset
? 0 : 1;
956 * Add default attributes, if there were no attributes specified or
957 * if -d/--detailed, -d -d or -d -d -d is used:
959 static int add_default_attributes(void)
961 struct perf_event_attr default_attrs
[] = {
963 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
964 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
965 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
966 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
968 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
969 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
970 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
971 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
972 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
973 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
978 * Detailed stats (-d), covering the L1 and last level data caches:
980 struct perf_event_attr detailed_attrs
[] = {
982 { .type
= PERF_TYPE_HW_CACHE
,
984 PERF_COUNT_HW_CACHE_L1D
<< 0 |
985 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
986 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
988 { .type
= PERF_TYPE_HW_CACHE
,
990 PERF_COUNT_HW_CACHE_L1D
<< 0 |
991 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
992 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
994 { .type
= PERF_TYPE_HW_CACHE
,
996 PERF_COUNT_HW_CACHE_LL
<< 0 |
997 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
998 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1000 { .type
= PERF_TYPE_HW_CACHE
,
1002 PERF_COUNT_HW_CACHE_LL
<< 0 |
1003 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1004 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1008 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1010 struct perf_event_attr very_detailed_attrs
[] = {
1012 { .type
= PERF_TYPE_HW_CACHE
,
1014 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1015 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1016 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1018 { .type
= PERF_TYPE_HW_CACHE
,
1020 PERF_COUNT_HW_CACHE_L1I
<< 0 |
1021 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1022 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1024 { .type
= PERF_TYPE_HW_CACHE
,
1026 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1027 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1028 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1030 { .type
= PERF_TYPE_HW_CACHE
,
1032 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
1033 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1034 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1036 { .type
= PERF_TYPE_HW_CACHE
,
1038 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1039 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1040 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1042 { .type
= PERF_TYPE_HW_CACHE
,
1044 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1045 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1046 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1051 * Very, very detailed stats (-d -d -d), adding prefetch events:
1053 struct perf_event_attr very_very_detailed_attrs
[] = {
1055 { .type
= PERF_TYPE_HW_CACHE
,
1057 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1058 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1059 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1061 { .type
= PERF_TYPE_HW_CACHE
,
1063 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1064 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1065 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1068 /* Set attrs if no event is selected and !null_run: */
1072 if (!evsel_list
->nr_entries
) {
1073 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1077 /* Detailed events get appended to the event list: */
1079 if (detailed_run
< 1)
1082 /* Append detailed run extra attributes: */
1083 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1086 if (detailed_run
< 2)
1089 /* Append very detailed run extra attributes: */
1090 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1093 if (detailed_run
< 3)
1096 /* Append very, very detailed run extra attributes: */
1097 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1100 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1102 bool append_file
= false;
1104 const char *output_name
= NULL
;
1105 const struct option options
[] = {
1106 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1107 "event selector. use 'perf list' to list available events",
1108 parse_events_option
),
1109 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1110 "event filter", parse_filter
),
1111 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1112 "child tasks do not inherit counters"),
1113 OPT_STRING('p', "pid", &target
.pid
, "pid",
1114 "stat events on existing process id"),
1115 OPT_STRING('t', "tid", &target
.tid
, "tid",
1116 "stat events on existing thread id"),
1117 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1118 "system-wide collection from all CPUs"),
1119 OPT_BOOLEAN('g', "group", &group
,
1120 "put the counters into a counter group"),
1121 OPT_BOOLEAN('c', "scale", &scale
, "scale/normalize counters"),
1122 OPT_INCR('v', "verbose", &verbose
,
1123 "be more verbose (show counter open errors, etc)"),
1124 OPT_INTEGER('r', "repeat", &run_count
,
1125 "repeat command and print average + stddev (max: 100)"),
1126 OPT_BOOLEAN('n', "null", &null_run
,
1127 "null run - dont start any counters"),
1128 OPT_INCR('d', "detailed", &detailed_run
,
1129 "detailed run - start a lot of events"),
1130 OPT_BOOLEAN('S', "sync", &sync_run
,
1131 "call sync() before starting a run"),
1132 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1133 "print large numbers with thousands\' separators",
1135 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1136 "list of cpus to monitor in system-wide"),
1137 OPT_BOOLEAN('A', "no-aggr", &no_aggr
, "disable CPU count aggregation"),
1138 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1139 "print counts with custom separator"),
1140 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1141 "monitor event in cgroup name only", parse_cgroups
),
1142 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1143 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1144 OPT_INTEGER(0, "log-fd", &output_fd
,
1145 "log output to fd, instead of stderr"),
1146 OPT_STRING(0, "pre", &pre_cmd
, "command",
1147 "command to run prior to the measured command"),
1148 OPT_STRING(0, "post", &post_cmd
, "command",
1149 "command to run after to the measured command"),
1152 const char * const stat_usage
[] = {
1153 "perf stat [<options>] [<command>]",
1156 struct perf_evsel
*pos
;
1157 int status
= -ENOMEM
, run_idx
;
1160 setlocale(LC_ALL
, "");
1162 evsel_list
= perf_evlist__new(NULL
, NULL
);
1163 if (evsel_list
== NULL
)
1166 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1167 PARSE_OPT_STOP_AT_NON_OPTION
);
1170 if (output_name
&& strcmp(output_name
, "-"))
1173 if (output_name
&& output_fd
) {
1174 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1175 usage_with_options(stat_usage
, options
);
1178 if (output_fd
< 0) {
1179 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1180 usage_with_options(stat_usage
, options
);
1185 mode
= append_file
? "a" : "w";
1187 output
= fopen(output_name
, mode
);
1189 perror("failed to create output file");
1192 clock_gettime(CLOCK_REALTIME
, &tm
);
1193 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1194 } else if (output_fd
> 0) {
1195 mode
= append_file
? "a" : "w";
1196 output
= fdopen(output_fd
, mode
);
1198 perror("Failed opening logfd");
1205 if (!strcmp(csv_sep
, "\\t"))
1208 csv_sep
= DEFAULT_SEPARATOR
;
1211 * let the spreadsheet do the pretty-printing
1214 /* User explicitly passed -B? */
1215 if (big_num_opt
== 1) {
1216 fprintf(stderr
, "-B option not supported with -x\n");
1217 usage_with_options(stat_usage
, options
);
1218 } else /* Nope, so disable big number formatting */
1220 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1223 if (!argc
&& !perf_target__has_task(&target
))
1224 usage_with_options(stat_usage
, options
);
1226 usage_with_options(stat_usage
, options
);
1228 /* no_aggr, cgroup are for system-wide only */
1229 if ((no_aggr
|| nr_cgroups
) && !perf_target__has_cpu(&target
)) {
1230 fprintf(stderr
, "both cgroup and no-aggregation "
1231 "modes only available in system-wide mode\n");
1233 usage_with_options(stat_usage
, options
);
1236 if (add_default_attributes())
1239 perf_target__validate(&target
);
1241 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1242 if (perf_target__has_task(&target
))
1243 pr_err("Problems finding threads of monitor\n");
1244 if (perf_target__has_cpu(&target
))
1245 perror("failed to parse CPUs map");
1247 usage_with_options(stat_usage
, options
);
1251 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1252 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
1253 perf_evsel__alloc_counts(pos
, perf_evsel__nr_cpus(pos
)) < 0)
1258 * We dont want to block the signals - that would cause
1259 * child tasks to inherit that and Ctrl-C would not work.
1260 * What we want is for Ctrl-C to work in the exec()-ed
1261 * task, but being ignored by perf stat itself:
1264 signal(SIGINT
, skip_signal
);
1265 signal(SIGALRM
, skip_signal
);
1266 signal(SIGABRT
, skip_signal
);
1269 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
1270 if (run_count
!= 1 && verbose
)
1271 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1274 status
= run_perf_stat(argc
, argv
);
1278 print_stat(argc
, argv
);
1280 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
1281 perf_evsel__free_stat_priv(pos
);
1282 perf_evlist__delete_maps(evsel_list
);
1284 perf_evlist__delete(evsel_list
);