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"
60 #include <sys/prctl.h>
63 #define DEFAULT_SEPARATOR " "
64 #define CNTR_NOT_SUPPORTED "<not supported>"
65 #define CNTR_NOT_COUNTED "<not counted>"
67 static struct perf_evlist
*evsel_list
;
69 static struct perf_target target
= {
73 static int run_count
= 1;
74 static bool no_inherit
= false;
75 static bool scale
= true;
76 static bool no_aggr
= false;
77 static pid_t child_pid
= -1;
78 static bool null_run
= false;
79 static int detailed_run
= 0;
80 static bool big_num
= true;
81 static int big_num_opt
= -1;
82 static const char *csv_sep
= NULL
;
83 static bool csv_output
= false;
84 static bool group
= false;
85 static FILE *output
= NULL
;
87 static volatile int done
= 0;
90 struct stats res_stats
[3];
93 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
95 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
96 return evsel
->priv
== NULL
? -ENOMEM
: 0;
99 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
105 static inline struct cpu_map
*perf_evsel__cpus(struct perf_evsel
*evsel
)
107 return (evsel
->cpus
&& !target
.cpu_list
) ? evsel
->cpus
: evsel_list
->cpus
;
110 static inline int perf_evsel__nr_cpus(struct perf_evsel
*evsel
)
112 return perf_evsel__cpus(evsel
)->nr
;
115 static struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
116 static struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
117 static struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
118 static struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
119 static struct stats runtime_branches_stats
[MAX_NR_CPUS
];
120 static struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
121 static struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
122 static struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
123 static struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
124 static struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
125 static struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
126 static struct stats walltime_nsecs_stats
;
128 static int create_perf_stat_counter(struct perf_evsel
*evsel
,
129 struct perf_evsel
*first
)
131 struct perf_event_attr
*attr
= &evsel
->attr
;
132 bool exclude_guest_missing
= false;
136 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
137 PERF_FORMAT_TOTAL_TIME_RUNNING
;
139 attr
->inherit
= !no_inherit
;
142 if (exclude_guest_missing
)
143 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
145 if (perf_target__has_cpu(&target
)) {
146 ret
= perf_evsel__open_per_cpu(evsel
, perf_evsel__cpus(evsel
));
152 if (!perf_target__has_task(&target
) && (!group
|| evsel
== first
)) {
154 attr
->enable_on_exec
= 1;
157 ret
= perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
162 if (ret
&& errno
== EINVAL
) {
163 if (!exclude_guest_missing
&&
164 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
165 pr_debug("Old kernel, cannot exclude "
166 "guest or host samples.\n");
167 exclude_guest_missing
= true;
175 * Does the counter have nsecs as a unit?
177 static inline int nsec_counter(struct perf_evsel
*evsel
)
179 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
180 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
187 * Update various tracking values we maintain to print
188 * more semantic information such as miss/hit ratios,
189 * instruction rates, etc:
191 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
193 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
194 update_stats(&runtime_nsecs_stats
[0], count
[0]);
195 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
196 update_stats(&runtime_cycles_stats
[0], count
[0]);
197 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
198 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
199 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
200 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
201 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
202 update_stats(&runtime_branches_stats
[0], count
[0]);
203 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
204 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
205 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
206 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
207 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
208 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
209 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
210 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
211 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
212 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
213 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
214 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
218 * Read out the results of a single counter:
219 * aggregate counts across CPUs in system-wide mode
221 static int read_counter_aggr(struct perf_evsel
*counter
)
223 struct perf_stat
*ps
= counter
->priv
;
224 u64
*count
= counter
->counts
->aggr
.values
;
227 if (__perf_evsel__read(counter
, perf_evsel__nr_cpus(counter
),
228 evsel_list
->threads
->nr
, scale
) < 0)
231 for (i
= 0; i
< 3; i
++)
232 update_stats(&ps
->res_stats
[i
], count
[i
]);
235 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
236 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
240 * Save the full runtime - to allow normalization during printout:
242 update_shadow_stats(counter
, count
);
248 * Read out the results of a single counter:
249 * do not aggregate counts across CPUs in system-wide mode
251 static int read_counter(struct perf_evsel
*counter
)
256 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
257 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
260 count
= counter
->counts
->cpu
[cpu
].values
;
262 update_shadow_stats(counter
, count
);
268 static int run_perf_stat(int argc __maybe_unused
, const char **argv
)
270 unsigned long long t0
, t1
;
271 struct perf_evsel
*counter
, *first
;
273 int child_ready_pipe
[2], go_pipe
[2];
274 const bool forks
= (argc
> 0);
277 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
278 perror("failed to create pipes");
283 if ((child_pid
= fork()) < 0)
284 perror("failed to fork");
287 close(child_ready_pipe
[0]);
289 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
292 * Do a dummy execvp to get the PLT entry resolved,
293 * so we avoid the resolver overhead on the real
296 execvp("", (char **)argv
);
299 * Tell the parent we're ready to go
301 close(child_ready_pipe
[1]);
304 * Wait until the parent tells us to go.
306 if (read(go_pipe
[0], &buf
, 1) == -1)
307 perror("unable to read pipe");
309 execvp(argv
[0], (char **)argv
);
315 if (perf_target__none(&target
))
316 evsel_list
->threads
->map
[0] = child_pid
;
319 * Wait for the child to be ready to exec.
321 close(child_ready_pipe
[1]);
323 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
324 perror("unable to read pipe");
325 close(child_ready_pipe
[0]);
329 perf_evlist__set_leader(evsel_list
);
331 first
= perf_evlist__first(evsel_list
);
333 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
334 if (create_perf_stat_counter(counter
, first
) < 0) {
336 * PPC returns ENXIO for HW counters until 2.6.37
337 * (behavior changed with commit b0a873e).
339 if (errno
== EINVAL
|| errno
== ENOSYS
||
340 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
343 ui__warning("%s event is not supported by the kernel.\n",
344 perf_evsel__name(counter
));
345 counter
->supported
= false;
349 if (errno
== EPERM
|| errno
== EACCES
) {
350 error("You may not have permission to collect %sstats.\n"
351 "\t Consider tweaking"
352 " /proc/sys/kernel/perf_event_paranoid or running as root.",
353 target
.system_wide
? "system-wide " : "");
355 error("open_counter returned with %d (%s). "
356 "/bin/dmesg may provide additional information.\n",
357 errno
, strerror(errno
));
360 kill(child_pid
, SIGTERM
);
362 pr_err("Not all events could be opened.\n");
365 counter
->supported
= true;
368 if (perf_evlist__apply_filters(evsel_list
)) {
369 error("failed to set filter with %d (%s)\n", errno
,
375 * Enable counters and exec the command:
382 if (WIFSIGNALED(status
))
383 psignal(WTERMSIG(status
), argv
[0]);
385 while(!done
) sleep(1);
390 update_stats(&walltime_nsecs_stats
, t1
- t0
);
393 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
394 read_counter(counter
);
395 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
), 1);
398 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
399 read_counter_aggr(counter
);
400 perf_evsel__close_fd(counter
, perf_evsel__nr_cpus(counter
),
401 evsel_list
->threads
->nr
);
405 return WEXITSTATUS(status
);
408 static void print_noise_pct(double total
, double avg
)
410 double pct
= rel_stddev_stats(total
, avg
);
413 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
415 fprintf(output
, " ( +-%6.2f%% )", pct
);
418 static void print_noise(struct perf_evsel
*evsel
, double avg
)
420 struct perf_stat
*ps
;
426 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
429 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
431 double msecs
= avg
/ 1e6
;
432 char cpustr
[16] = { '\0', };
433 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
436 sprintf(cpustr
, "CPU%*d%s",
438 perf_evsel__cpus(evsel
)->map
[cpu
], csv_sep
);
440 fprintf(output
, fmt
, cpustr
, msecs
, csv_sep
, perf_evsel__name(evsel
));
443 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
448 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
449 fprintf(output
, " # %8.3f CPUs utilized ",
450 avg
/ avg_stats(&walltime_nsecs_stats
));
452 fprintf(output
, " ");
455 /* used for get_ratio_color() */
457 GRC_STALLED_CYCLES_FE
,
458 GRC_STALLED_CYCLES_BE
,
463 static const char *get_ratio_color(enum grc_type type
, double ratio
)
465 static const double grc_table
[GRC_MAX_NR
][3] = {
466 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
467 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
468 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
470 const char *color
= PERF_COLOR_NORMAL
;
472 if (ratio
> grc_table
[type
][0])
473 color
= PERF_COLOR_RED
;
474 else if (ratio
> grc_table
[type
][1])
475 color
= PERF_COLOR_MAGENTA
;
476 else if (ratio
> grc_table
[type
][2])
477 color
= PERF_COLOR_YELLOW
;
482 static void print_stalled_cycles_frontend(int cpu
,
483 struct perf_evsel
*evsel
484 __maybe_unused
, double avg
)
486 double total
, ratio
= 0.0;
489 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
492 ratio
= avg
/ total
* 100.0;
494 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
496 fprintf(output
, " # ");
497 color_fprintf(output
, color
, "%6.2f%%", ratio
);
498 fprintf(output
, " frontend cycles idle ");
501 static void print_stalled_cycles_backend(int cpu
,
502 struct perf_evsel
*evsel
503 __maybe_unused
, double avg
)
505 double total
, ratio
= 0.0;
508 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
511 ratio
= avg
/ total
* 100.0;
513 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
515 fprintf(output
, " # ");
516 color_fprintf(output
, color
, "%6.2f%%", ratio
);
517 fprintf(output
, " backend cycles idle ");
520 static void print_branch_misses(int cpu
,
521 struct perf_evsel
*evsel __maybe_unused
,
524 double total
, ratio
= 0.0;
527 total
= avg_stats(&runtime_branches_stats
[cpu
]);
530 ratio
= avg
/ total
* 100.0;
532 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
534 fprintf(output
, " # ");
535 color_fprintf(output
, color
, "%6.2f%%", ratio
);
536 fprintf(output
, " of all branches ");
539 static void print_l1_dcache_misses(int cpu
,
540 struct perf_evsel
*evsel __maybe_unused
,
543 double total
, ratio
= 0.0;
546 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
549 ratio
= avg
/ total
* 100.0;
551 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
553 fprintf(output
, " # ");
554 color_fprintf(output
, color
, "%6.2f%%", ratio
);
555 fprintf(output
, " of all L1-dcache hits ");
558 static void print_l1_icache_misses(int cpu
,
559 struct perf_evsel
*evsel __maybe_unused
,
562 double total
, ratio
= 0.0;
565 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
568 ratio
= avg
/ total
* 100.0;
570 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
572 fprintf(output
, " # ");
573 color_fprintf(output
, color
, "%6.2f%%", ratio
);
574 fprintf(output
, " of all L1-icache hits ");
577 static void print_dtlb_cache_misses(int cpu
,
578 struct perf_evsel
*evsel __maybe_unused
,
581 double total
, ratio
= 0.0;
584 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
587 ratio
= avg
/ total
* 100.0;
589 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
591 fprintf(output
, " # ");
592 color_fprintf(output
, color
, "%6.2f%%", ratio
);
593 fprintf(output
, " of all dTLB cache hits ");
596 static void print_itlb_cache_misses(int cpu
,
597 struct perf_evsel
*evsel __maybe_unused
,
600 double total
, ratio
= 0.0;
603 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
606 ratio
= avg
/ total
* 100.0;
608 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
610 fprintf(output
, " # ");
611 color_fprintf(output
, color
, "%6.2f%%", ratio
);
612 fprintf(output
, " of all iTLB cache hits ");
615 static void print_ll_cache_misses(int cpu
,
616 struct perf_evsel
*evsel __maybe_unused
,
619 double total
, ratio
= 0.0;
622 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
625 ratio
= avg
/ total
* 100.0;
627 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
629 fprintf(output
, " # ");
630 color_fprintf(output
, color
, "%6.2f%%", ratio
);
631 fprintf(output
, " of all LL-cache hits ");
634 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
636 double total
, ratio
= 0.0;
637 char cpustr
[16] = { '\0', };
643 fmt
= "%s%'18.0f%s%-25s";
645 fmt
= "%s%18.0f%s%-25s";
648 sprintf(cpustr
, "CPU%*d%s",
650 perf_evsel__cpus(evsel
)->map
[cpu
], csv_sep
);
654 fprintf(output
, fmt
, cpustr
, avg
, csv_sep
, perf_evsel__name(evsel
));
657 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
662 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
663 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
668 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
670 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
671 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
675 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
678 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
679 runtime_branches_stats
[cpu
].n
!= 0) {
680 print_branch_misses(cpu
, evsel
, avg
);
682 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
683 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
684 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
685 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
686 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
687 print_l1_dcache_misses(cpu
, evsel
, avg
);
689 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
690 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
691 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
692 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
693 runtime_l1_icache_stats
[cpu
].n
!= 0) {
694 print_l1_icache_misses(cpu
, evsel
, avg
);
696 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
697 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
698 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
699 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
700 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
701 print_dtlb_cache_misses(cpu
, evsel
, avg
);
703 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
704 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
705 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
706 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
707 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
708 print_itlb_cache_misses(cpu
, evsel
, avg
);
710 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
711 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
712 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
713 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
714 runtime_ll_cache_stats
[cpu
].n
!= 0) {
715 print_ll_cache_misses(cpu
, evsel
, avg
);
716 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
717 runtime_cacherefs_stats
[cpu
].n
!= 0) {
718 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
721 ratio
= avg
* 100 / total
;
723 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
725 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
726 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
727 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
728 print_stalled_cycles_backend(cpu
, evsel
, avg
);
729 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
730 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
733 ratio
= 1.0 * avg
/ total
;
735 fprintf(output
, " # %8.3f GHz ", ratio
);
736 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
739 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
742 ratio
= 1000.0 * avg
/ total
;
748 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
750 fprintf(output
, " ");
755 * Print out the results of a single counter:
756 * aggregated counts in system-wide mode
758 static void print_counter_aggr(struct perf_evsel
*counter
)
760 struct perf_stat
*ps
= counter
->priv
;
761 double avg
= avg_stats(&ps
->res_stats
[0]);
762 int scaled
= counter
->counts
->scaled
;
765 fprintf(output
, "%*s%s%*s",
767 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
769 csv_output
? 0 : -24,
770 perf_evsel__name(counter
));
773 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
779 if (nsec_counter(counter
))
780 nsec_printout(-1, counter
, avg
);
782 abs_printout(-1, counter
, avg
);
784 print_noise(counter
, avg
);
792 double avg_enabled
, avg_running
;
794 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
795 avg_running
= avg_stats(&ps
->res_stats
[2]);
797 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
799 fprintf(output
, "\n");
803 * Print out the results of a single counter:
804 * does not use aggregated count in system-wide
806 static void print_counter(struct perf_evsel
*counter
)
811 for (cpu
= 0; cpu
< perf_evsel__nr_cpus(counter
); cpu
++) {
812 val
= counter
->counts
->cpu
[cpu
].val
;
813 ena
= counter
->counts
->cpu
[cpu
].ena
;
814 run
= counter
->counts
->cpu
[cpu
].run
;
815 if (run
== 0 || ena
== 0) {
816 fprintf(output
, "CPU%*d%s%*s%s%*s",
818 perf_evsel__cpus(counter
)->map
[cpu
], csv_sep
,
820 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
822 csv_output
? 0 : -24,
823 perf_evsel__name(counter
));
826 fprintf(output
, "%s%s",
827 csv_sep
, counter
->cgrp
->name
);
833 if (nsec_counter(counter
))
834 nsec_printout(cpu
, counter
, val
);
836 abs_printout(cpu
, counter
, val
);
839 print_noise(counter
, 1.0);
842 fprintf(output
, " (%.2f%%)",
849 static void print_stat(int argc
, const char **argv
)
851 struct perf_evsel
*counter
;
857 fprintf(output
, "\n");
858 fprintf(output
, " Performance counter stats for ");
859 if (!perf_target__has_task(&target
)) {
860 fprintf(output
, "\'%s", argv
[0]);
861 for (i
= 1; i
< argc
; i
++)
862 fprintf(output
, " %s", argv
[i
]);
863 } else if (target
.pid
)
864 fprintf(output
, "process id \'%s", target
.pid
);
866 fprintf(output
, "thread id \'%s", target
.tid
);
868 fprintf(output
, "\'");
870 fprintf(output
, " (%d runs)", run_count
);
871 fprintf(output
, ":\n\n");
875 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
876 print_counter(counter
);
878 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
879 print_counter_aggr(counter
);
884 fprintf(output
, "\n");
885 fprintf(output
, " %17.9f seconds time elapsed",
886 avg_stats(&walltime_nsecs_stats
)/1e9
);
888 fprintf(output
, " ");
889 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
890 avg_stats(&walltime_nsecs_stats
));
892 fprintf(output
, "\n\n");
896 static volatile int signr
= -1;
898 static void skip_signal(int signo
)
906 static void sig_atexit(void)
909 kill(child_pid
, SIGTERM
);
914 signal(signr
, SIG_DFL
);
915 kill(getpid(), signr
);
918 static int stat__set_big_num(const struct option
*opt __maybe_unused
,
919 const char *s __maybe_unused
, int unset
)
921 big_num_opt
= unset
? 0 : 1;
926 * Add default attributes, if there were no attributes specified or
927 * if -d/--detailed, -d -d or -d -d -d is used:
929 static int add_default_attributes(void)
931 struct perf_event_attr default_attrs
[] = {
933 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
934 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
935 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
936 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
938 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
939 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
940 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
941 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
942 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
943 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
948 * Detailed stats (-d), covering the L1 and last level data caches:
950 struct perf_event_attr detailed_attrs
[] = {
952 { .type
= PERF_TYPE_HW_CACHE
,
954 PERF_COUNT_HW_CACHE_L1D
<< 0 |
955 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
956 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
958 { .type
= PERF_TYPE_HW_CACHE
,
960 PERF_COUNT_HW_CACHE_L1D
<< 0 |
961 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
962 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
964 { .type
= PERF_TYPE_HW_CACHE
,
966 PERF_COUNT_HW_CACHE_LL
<< 0 |
967 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
968 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
970 { .type
= PERF_TYPE_HW_CACHE
,
972 PERF_COUNT_HW_CACHE_LL
<< 0 |
973 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
974 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
978 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
980 struct perf_event_attr very_detailed_attrs
[] = {
982 { .type
= PERF_TYPE_HW_CACHE
,
984 PERF_COUNT_HW_CACHE_L1I
<< 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_L1I
<< 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_DTLB
<< 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_DTLB
<< 0 |
1003 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1004 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1006 { .type
= PERF_TYPE_HW_CACHE
,
1008 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1009 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1010 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1012 { .type
= PERF_TYPE_HW_CACHE
,
1014 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
1015 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
1016 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1021 * Very, very detailed stats (-d -d -d), adding prefetch events:
1023 struct perf_event_attr very_very_detailed_attrs
[] = {
1025 { .type
= PERF_TYPE_HW_CACHE
,
1027 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1028 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1029 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
1031 { .type
= PERF_TYPE_HW_CACHE
,
1033 PERF_COUNT_HW_CACHE_L1D
<< 0 |
1034 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
1035 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
1038 /* Set attrs if no event is selected and !null_run: */
1042 if (!evsel_list
->nr_entries
) {
1043 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1047 /* Detailed events get appended to the event list: */
1049 if (detailed_run
< 1)
1052 /* Append detailed run extra attributes: */
1053 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1056 if (detailed_run
< 2)
1059 /* Append very detailed run extra attributes: */
1060 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1063 if (detailed_run
< 3)
1066 /* Append very, very detailed run extra attributes: */
1067 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1070 int cmd_stat(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1072 bool append_file
= false,
1075 const char *output_name
= NULL
;
1076 const struct option options
[] = {
1077 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1078 "event selector. use 'perf list' to list available events",
1079 parse_events_option
),
1080 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1081 "event filter", parse_filter
),
1082 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1083 "child tasks do not inherit counters"),
1084 OPT_STRING('p', "pid", &target
.pid
, "pid",
1085 "stat events on existing process id"),
1086 OPT_STRING('t', "tid", &target
.tid
, "tid",
1087 "stat events on existing thread id"),
1088 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1089 "system-wide collection from all CPUs"),
1090 OPT_BOOLEAN('g', "group", &group
,
1091 "put the counters into a counter group"),
1092 OPT_BOOLEAN('c', "scale", &scale
, "scale/normalize counters"),
1093 OPT_INCR('v', "verbose", &verbose
,
1094 "be more verbose (show counter open errors, etc)"),
1095 OPT_INTEGER('r', "repeat", &run_count
,
1096 "repeat command and print average + stddev (max: 100)"),
1097 OPT_BOOLEAN('n', "null", &null_run
,
1098 "null run - dont start any counters"),
1099 OPT_INCR('d', "detailed", &detailed_run
,
1100 "detailed run - start a lot of events"),
1101 OPT_BOOLEAN('S', "sync", &sync_run
,
1102 "call sync() before starting a run"),
1103 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1104 "print large numbers with thousands\' separators",
1106 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1107 "list of cpus to monitor in system-wide"),
1108 OPT_BOOLEAN('A', "no-aggr", &no_aggr
, "disable CPU count aggregation"),
1109 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1110 "print counts with custom separator"),
1111 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1112 "monitor event in cgroup name only", parse_cgroups
),
1113 OPT_STRING('o', "output", &output_name
, "file", "output file name"),
1114 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1115 OPT_INTEGER(0, "log-fd", &output_fd
,
1116 "log output to fd, instead of stderr"),
1119 const char * const stat_usage
[] = {
1120 "perf stat [<options>] [<command>]",
1123 struct perf_evsel
*pos
;
1124 int status
= -ENOMEM
, run_idx
;
1127 setlocale(LC_ALL
, "");
1129 evsel_list
= perf_evlist__new(NULL
, NULL
);
1130 if (evsel_list
== NULL
)
1133 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1134 PARSE_OPT_STOP_AT_NON_OPTION
);
1137 if (output_name
&& strcmp(output_name
, "-"))
1140 if (output_name
&& output_fd
) {
1141 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1142 usage_with_options(stat_usage
, options
);
1145 if (output_fd
< 0) {
1146 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1147 usage_with_options(stat_usage
, options
);
1152 mode
= append_file
? "a" : "w";
1154 output
= fopen(output_name
, mode
);
1156 perror("failed to create output file");
1159 clock_gettime(CLOCK_REALTIME
, &tm
);
1160 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1161 } else if (output_fd
> 0) {
1162 mode
= append_file
? "a" : "w";
1163 output
= fdopen(output_fd
, mode
);
1165 perror("Failed opening logfd");
1172 if (!strcmp(csv_sep
, "\\t"))
1175 csv_sep
= DEFAULT_SEPARATOR
;
1178 * let the spreadsheet do the pretty-printing
1181 /* User explicitly passed -B? */
1182 if (big_num_opt
== 1) {
1183 fprintf(stderr
, "-B option not supported with -x\n");
1184 usage_with_options(stat_usage
, options
);
1185 } else /* Nope, so disable big number formatting */
1187 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1190 if (!argc
&& !perf_target__has_task(&target
))
1191 usage_with_options(stat_usage
, options
);
1193 usage_with_options(stat_usage
, options
);
1195 /* no_aggr, cgroup are for system-wide only */
1196 if ((no_aggr
|| nr_cgroups
) && !perf_target__has_cpu(&target
)) {
1197 fprintf(stderr
, "both cgroup and no-aggregation "
1198 "modes only available in system-wide mode\n");
1200 usage_with_options(stat_usage
, options
);
1203 if (add_default_attributes())
1206 perf_target__validate(&target
);
1208 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1209 if (perf_target__has_task(&target
))
1210 pr_err("Problems finding threads of monitor\n");
1211 if (perf_target__has_cpu(&target
))
1212 perror("failed to parse CPUs map");
1214 usage_with_options(stat_usage
, options
);
1218 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1219 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
1220 perf_evsel__alloc_counts(pos
, perf_evsel__nr_cpus(pos
)) < 0)
1225 * We dont want to block the signals - that would cause
1226 * child tasks to inherit that and Ctrl-C would not work.
1227 * What we want is for Ctrl-C to work in the exec()-ed
1228 * task, but being ignored by perf stat itself:
1231 signal(SIGINT
, skip_signal
);
1232 signal(SIGALRM
, skip_signal
);
1233 signal(SIGABRT
, skip_signal
);
1236 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
1237 if (run_count
!= 1 && verbose
)
1238 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1244 status
= run_perf_stat(argc
, argv
);
1248 print_stat(argc
, argv
);
1250 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
1251 perf_evsel__free_stat_priv(pos
);
1252 perf_evlist__delete_maps(evsel_list
);
1254 perf_evlist__delete(evsel_list
);