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/header.h"
55 #include "util/cpumap.h"
56 #include "util/thread.h"
57 #include "util/thread_map.h"
59 #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_event_attr default_attrs
[] = {
69 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
70 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
71 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
72 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
74 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
75 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
},
76 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_STALLED_CYCLES_BACKEND
},
77 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
78 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
79 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
84 * Detailed stats (-d), covering the L1 and last level data caches:
86 static struct perf_event_attr detailed_attrs
[] = {
88 { .type
= PERF_TYPE_HW_CACHE
,
90 PERF_COUNT_HW_CACHE_L1D
<< 0 |
91 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
92 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
94 { .type
= PERF_TYPE_HW_CACHE
,
96 PERF_COUNT_HW_CACHE_L1D
<< 0 |
97 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
98 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
100 { .type
= PERF_TYPE_HW_CACHE
,
102 PERF_COUNT_HW_CACHE_LL
<< 0 |
103 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
106 { .type
= PERF_TYPE_HW_CACHE
,
108 PERF_COUNT_HW_CACHE_LL
<< 0 |
109 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
110 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
114 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
116 static struct perf_event_attr very_detailed_attrs
[] = {
118 { .type
= PERF_TYPE_HW_CACHE
,
120 PERF_COUNT_HW_CACHE_L1I
<< 0 |
121 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
124 { .type
= PERF_TYPE_HW_CACHE
,
126 PERF_COUNT_HW_CACHE_L1I
<< 0 |
127 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
128 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
130 { .type
= PERF_TYPE_HW_CACHE
,
132 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
133 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
136 { .type
= PERF_TYPE_HW_CACHE
,
138 PERF_COUNT_HW_CACHE_DTLB
<< 0 |
139 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
140 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
142 { .type
= PERF_TYPE_HW_CACHE
,
144 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
145 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
146 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
148 { .type
= PERF_TYPE_HW_CACHE
,
150 PERF_COUNT_HW_CACHE_ITLB
<< 0 |
151 (PERF_COUNT_HW_CACHE_OP_READ
<< 8) |
152 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
157 * Very, very detailed stats (-d -d -d), adding prefetch events:
159 static struct perf_event_attr very_very_detailed_attrs
[] = {
161 { .type
= PERF_TYPE_HW_CACHE
,
163 PERF_COUNT_HW_CACHE_L1D
<< 0 |
164 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS
<< 16) },
167 { .type
= PERF_TYPE_HW_CACHE
,
169 PERF_COUNT_HW_CACHE_L1D
<< 0 |
170 (PERF_COUNT_HW_CACHE_OP_PREFETCH
<< 8) |
171 (PERF_COUNT_HW_CACHE_RESULT_MISS
<< 16) },
176 static struct perf_evlist
*evsel_list
;
178 static struct perf_target target
= {
182 static int run_idx
= 0;
183 static int run_count
= 1;
184 static bool no_inherit
= false;
185 static bool scale
= true;
186 static bool no_aggr
= false;
187 static pid_t child_pid
= -1;
188 static bool null_run
= false;
189 static int detailed_run
= 0;
190 static bool sync_run
= false;
191 static bool big_num
= true;
192 static int big_num_opt
= -1;
193 static const char *csv_sep
= NULL
;
194 static bool csv_output
= false;
195 static bool group
= false;
196 static const char *output_name
= NULL
;
197 static FILE *output
= NULL
;
198 static int output_fd
;
200 static volatile int done
= 0;
208 struct stats res_stats
[3];
211 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
213 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
214 return evsel
->priv
== NULL
? -ENOMEM
: 0;
217 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
223 static void update_stats(struct stats
*stats
, u64 val
)
228 delta
= val
- stats
->mean
;
229 stats
->mean
+= delta
/ stats
->n
;
230 stats
->M2
+= delta
*(val
- stats
->mean
);
233 static double avg_stats(struct stats
*stats
)
239 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
241 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
242 * s^2 = -------------------------------
245 * http://en.wikipedia.org/wiki/Stddev
247 * The std dev of the mean is related to the std dev by:
254 static double stddev_stats(struct stats
*stats
)
256 double variance
, variance_mean
;
261 variance
= stats
->M2
/ (stats
->n
- 1);
262 variance_mean
= variance
/ stats
->n
;
264 return sqrt(variance_mean
);
267 static struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
268 static struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
269 static struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
270 static struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
271 static struct stats runtime_branches_stats
[MAX_NR_CPUS
];
272 static struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
273 static struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
274 static struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
275 static struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
276 static struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
277 static struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
278 static struct stats walltime_nsecs_stats
;
280 static int create_perf_stat_counter(struct perf_evsel
*evsel
,
281 struct perf_evsel
*first
)
283 struct perf_event_attr
*attr
= &evsel
->attr
;
284 bool exclude_guest_missing
= false;
288 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
289 PERF_FORMAT_TOTAL_TIME_RUNNING
;
291 attr
->inherit
= !no_inherit
;
294 if (exclude_guest_missing
)
295 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
297 if (perf_target__has_cpu(&target
)) {
298 ret
= perf_evsel__open_per_cpu(evsel
, evsel_list
->cpus
);
304 if (!perf_target__has_task(&target
) && (!group
|| evsel
== first
)) {
306 attr
->enable_on_exec
= 1;
309 ret
= perf_evsel__open_per_thread(evsel
, evsel_list
->threads
);
314 if (ret
&& errno
== EINVAL
) {
315 if (!exclude_guest_missing
&&
316 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
317 pr_debug("Old kernel, cannot exclude "
318 "guest or host samples.\n");
319 exclude_guest_missing
= true;
327 * Does the counter have nsecs as a unit?
329 static inline int nsec_counter(struct perf_evsel
*evsel
)
331 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
332 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
339 * Update various tracking values we maintain to print
340 * more semantic information such as miss/hit ratios,
341 * instruction rates, etc:
343 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
345 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
346 update_stats(&runtime_nsecs_stats
[0], count
[0]);
347 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
348 update_stats(&runtime_cycles_stats
[0], count
[0]);
349 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
350 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
351 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
352 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
353 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
354 update_stats(&runtime_branches_stats
[0], count
[0]);
355 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
356 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
357 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
358 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
359 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
360 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
361 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
362 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
363 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
364 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
365 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
366 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
370 * Read out the results of a single counter:
371 * aggregate counts across CPUs in system-wide mode
373 static int read_counter_aggr(struct perf_evsel
*counter
)
375 struct perf_stat
*ps
= counter
->priv
;
376 u64
*count
= counter
->counts
->aggr
.values
;
379 if (__perf_evsel__read(counter
, evsel_list
->cpus
->nr
,
380 evsel_list
->threads
->nr
, scale
) < 0)
383 for (i
= 0; i
< 3; i
++)
384 update_stats(&ps
->res_stats
[i
], count
[i
]);
387 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
388 perf_evsel__name(counter
), count
[0], count
[1], count
[2]);
392 * Save the full runtime - to allow normalization during printout:
394 update_shadow_stats(counter
, count
);
400 * Read out the results of a single counter:
401 * do not aggregate counts across CPUs in system-wide mode
403 static int read_counter(struct perf_evsel
*counter
)
408 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
409 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
412 count
= counter
->counts
->cpu
[cpu
].values
;
414 update_shadow_stats(counter
, count
);
420 static int run_perf_stat(int argc __used
, const char **argv
)
422 unsigned long long t0
, t1
;
423 struct perf_evsel
*counter
, *first
;
425 int child_ready_pipe
[2], go_pipe
[2];
426 const bool forks
= (argc
> 0);
429 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
430 perror("failed to create pipes");
435 if ((child_pid
= fork()) < 0)
436 perror("failed to fork");
439 close(child_ready_pipe
[0]);
441 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
444 * Do a dummy execvp to get the PLT entry resolved,
445 * so we avoid the resolver overhead on the real
448 execvp("", (char **)argv
);
451 * Tell the parent we're ready to go
453 close(child_ready_pipe
[1]);
456 * Wait until the parent tells us to go.
458 if (read(go_pipe
[0], &buf
, 1) == -1)
459 perror("unable to read pipe");
461 execvp(argv
[0], (char **)argv
);
467 if (perf_target__none(&target
))
468 evsel_list
->threads
->map
[0] = child_pid
;
471 * Wait for the child to be ready to exec.
473 close(child_ready_pipe
[1]);
475 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
476 perror("unable to read pipe");
477 close(child_ready_pipe
[0]);
481 perf_evlist__group(evsel_list
);
483 first
= list_entry(evsel_list
->entries
.next
, struct perf_evsel
, node
);
485 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
486 if (create_perf_stat_counter(counter
, first
) < 0) {
488 * PPC returns ENXIO for HW counters until 2.6.37
489 * (behavior changed with commit b0a873e).
491 if (errno
== EINVAL
|| errno
== ENOSYS
||
492 errno
== ENOENT
|| errno
== EOPNOTSUPP
||
495 ui__warning("%s event is not supported by the kernel.\n",
496 perf_evsel__name(counter
));
497 counter
->supported
= false;
501 if (errno
== EPERM
|| errno
== EACCES
) {
502 error("You may not have permission to collect %sstats.\n"
503 "\t Consider tweaking"
504 " /proc/sys/kernel/perf_event_paranoid or running as root.",
505 target
.system_wide
? "system-wide " : "");
507 error("open_counter returned with %d (%s). "
508 "/bin/dmesg may provide additional information.\n",
509 errno
, strerror(errno
));
512 kill(child_pid
, SIGTERM
);
513 die("Not all events could be opened.\n");
516 counter
->supported
= true;
519 if (perf_evlist__set_filters(evsel_list
)) {
520 error("failed to set filter with %d (%s)\n", errno
,
526 * Enable counters and exec the command:
533 if (WIFSIGNALED(status
))
534 psignal(WTERMSIG(status
), argv
[0]);
536 while(!done
) sleep(1);
541 update_stats(&walltime_nsecs_stats
, t1
- t0
);
544 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
545 read_counter(counter
);
546 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
, 1);
549 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
550 read_counter_aggr(counter
);
551 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
,
552 evsel_list
->threads
->nr
);
556 return WEXITSTATUS(status
);
559 static void print_noise_pct(double total
, double avg
)
564 pct
= 100.0*total
/avg
;
567 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
569 fprintf(output
, " ( +-%6.2f%% )", pct
);
572 static void print_noise(struct perf_evsel
*evsel
, double avg
)
574 struct perf_stat
*ps
;
580 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
583 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
585 double msecs
= avg
/ 1e6
;
586 char cpustr
[16] = { '\0', };
587 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
590 sprintf(cpustr
, "CPU%*d%s",
592 evsel_list
->cpus
->map
[cpu
], csv_sep
);
594 fprintf(output
, fmt
, cpustr
, msecs
, csv_sep
, perf_evsel__name(evsel
));
597 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
602 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
603 fprintf(output
, " # %8.3f CPUs utilized ",
604 avg
/ avg_stats(&walltime_nsecs_stats
));
606 fprintf(output
, " ");
609 /* used for get_ratio_color() */
611 GRC_STALLED_CYCLES_FE
,
612 GRC_STALLED_CYCLES_BE
,
617 static const char *get_ratio_color(enum grc_type type
, double ratio
)
619 static const double grc_table
[GRC_MAX_NR
][3] = {
620 [GRC_STALLED_CYCLES_FE
] = { 50.0, 30.0, 10.0 },
621 [GRC_STALLED_CYCLES_BE
] = { 75.0, 50.0, 20.0 },
622 [GRC_CACHE_MISSES
] = { 20.0, 10.0, 5.0 },
624 const char *color
= PERF_COLOR_NORMAL
;
626 if (ratio
> grc_table
[type
][0])
627 color
= PERF_COLOR_RED
;
628 else if (ratio
> grc_table
[type
][1])
629 color
= PERF_COLOR_MAGENTA
;
630 else if (ratio
> grc_table
[type
][2])
631 color
= PERF_COLOR_YELLOW
;
636 static void print_stalled_cycles_frontend(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
638 double total
, ratio
= 0.0;
641 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
644 ratio
= avg
/ total
* 100.0;
646 color
= get_ratio_color(GRC_STALLED_CYCLES_FE
, ratio
);
648 fprintf(output
, " # ");
649 color_fprintf(output
, color
, "%6.2f%%", ratio
);
650 fprintf(output
, " frontend cycles idle ");
653 static void print_stalled_cycles_backend(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
655 double total
, ratio
= 0.0;
658 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
661 ratio
= avg
/ total
* 100.0;
663 color
= get_ratio_color(GRC_STALLED_CYCLES_BE
, ratio
);
665 fprintf(output
, " # ");
666 color_fprintf(output
, color
, "%6.2f%%", ratio
);
667 fprintf(output
, " backend cycles idle ");
670 static void print_branch_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
672 double total
, ratio
= 0.0;
675 total
= avg_stats(&runtime_branches_stats
[cpu
]);
678 ratio
= avg
/ total
* 100.0;
680 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
682 fprintf(output
, " # ");
683 color_fprintf(output
, color
, "%6.2f%%", ratio
);
684 fprintf(output
, " of all branches ");
687 static void print_l1_dcache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
689 double total
, ratio
= 0.0;
692 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
695 ratio
= avg
/ total
* 100.0;
697 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
699 fprintf(output
, " # ");
700 color_fprintf(output
, color
, "%6.2f%%", ratio
);
701 fprintf(output
, " of all L1-dcache hits ");
704 static void print_l1_icache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
706 double total
, ratio
= 0.0;
709 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
712 ratio
= avg
/ total
* 100.0;
714 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
716 fprintf(output
, " # ");
717 color_fprintf(output
, color
, "%6.2f%%", ratio
);
718 fprintf(output
, " of all L1-icache hits ");
721 static void print_dtlb_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
723 double total
, ratio
= 0.0;
726 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
729 ratio
= avg
/ total
* 100.0;
731 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
733 fprintf(output
, " # ");
734 color_fprintf(output
, color
, "%6.2f%%", ratio
);
735 fprintf(output
, " of all dTLB cache hits ");
738 static void print_itlb_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
740 double total
, ratio
= 0.0;
743 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
746 ratio
= avg
/ total
* 100.0;
748 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
750 fprintf(output
, " # ");
751 color_fprintf(output
, color
, "%6.2f%%", ratio
);
752 fprintf(output
, " of all iTLB cache hits ");
755 static void print_ll_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
757 double total
, ratio
= 0.0;
760 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
763 ratio
= avg
/ total
* 100.0;
765 color
= get_ratio_color(GRC_CACHE_MISSES
, ratio
);
767 fprintf(output
, " # ");
768 color_fprintf(output
, color
, "%6.2f%%", ratio
);
769 fprintf(output
, " of all LL-cache hits ");
772 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
774 double total
, ratio
= 0.0;
775 char cpustr
[16] = { '\0', };
781 fmt
= "%s%'18.0f%s%-25s";
783 fmt
= "%s%18.0f%s%-25s";
786 sprintf(cpustr
, "CPU%*d%s",
788 evsel_list
->cpus
->map
[cpu
], csv_sep
);
792 fprintf(output
, fmt
, cpustr
, avg
, csv_sep
, perf_evsel__name(evsel
));
795 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
800 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
801 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
806 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
808 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
809 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
813 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
816 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
817 runtime_branches_stats
[cpu
].n
!= 0) {
818 print_branch_misses(cpu
, evsel
, avg
);
820 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
821 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
822 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
823 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
824 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
825 print_l1_dcache_misses(cpu
, evsel
, avg
);
827 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
828 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
829 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
830 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
831 runtime_l1_icache_stats
[cpu
].n
!= 0) {
832 print_l1_icache_misses(cpu
, evsel
, avg
);
834 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
835 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
836 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
837 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
838 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
839 print_dtlb_cache_misses(cpu
, evsel
, avg
);
841 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
842 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
843 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
844 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
845 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
846 print_itlb_cache_misses(cpu
, evsel
, avg
);
848 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
849 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
850 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
851 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
852 runtime_ll_cache_stats
[cpu
].n
!= 0) {
853 print_ll_cache_misses(cpu
, evsel
, avg
);
854 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
855 runtime_cacherefs_stats
[cpu
].n
!= 0) {
856 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
859 ratio
= avg
* 100 / total
;
861 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
863 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
864 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
865 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
866 print_stalled_cycles_backend(cpu
, evsel
, avg
);
867 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
868 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
871 ratio
= 1.0 * avg
/ total
;
873 fprintf(output
, " # %8.3f GHz ", ratio
);
874 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
877 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
880 ratio
= 1000.0 * avg
/ total
;
886 fprintf(output
, " # %8.3f %c/sec ", ratio
, unit
);
888 fprintf(output
, " ");
893 * Print out the results of a single counter:
894 * aggregated counts in system-wide mode
896 static void print_counter_aggr(struct perf_evsel
*counter
)
898 struct perf_stat
*ps
= counter
->priv
;
899 double avg
= avg_stats(&ps
->res_stats
[0]);
900 int scaled
= counter
->counts
->scaled
;
903 fprintf(output
, "%*s%s%*s",
905 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
907 csv_output
? 0 : -24,
908 perf_evsel__name(counter
));
911 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
917 if (nsec_counter(counter
))
918 nsec_printout(-1, counter
, avg
);
920 abs_printout(-1, counter
, avg
);
922 print_noise(counter
, avg
);
930 double avg_enabled
, avg_running
;
932 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
933 avg_running
= avg_stats(&ps
->res_stats
[2]);
935 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
937 fprintf(output
, "\n");
941 * Print out the results of a single counter:
942 * does not use aggregated count in system-wide
944 static void print_counter(struct perf_evsel
*counter
)
949 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
950 val
= counter
->counts
->cpu
[cpu
].val
;
951 ena
= counter
->counts
->cpu
[cpu
].ena
;
952 run
= counter
->counts
->cpu
[cpu
].run
;
953 if (run
== 0 || ena
== 0) {
954 fprintf(output
, "CPU%*d%s%*s%s%*s",
956 evsel_list
->cpus
->map
[cpu
], csv_sep
,
958 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
960 csv_output
? 0 : -24,
961 perf_evsel__name(counter
));
964 fprintf(output
, "%s%s",
965 csv_sep
, counter
->cgrp
->name
);
971 if (nsec_counter(counter
))
972 nsec_printout(cpu
, counter
, val
);
974 abs_printout(cpu
, counter
, val
);
977 print_noise(counter
, 1.0);
980 fprintf(output
, " (%.2f%%)",
987 static void print_stat(int argc
, const char **argv
)
989 struct perf_evsel
*counter
;
995 fprintf(output
, "\n");
996 fprintf(output
, " Performance counter stats for ");
997 if (!perf_target__has_task(&target
)) {
998 fprintf(output
, "\'%s", argv
[0]);
999 for (i
= 1; i
< argc
; i
++)
1000 fprintf(output
, " %s", argv
[i
]);
1001 } else if (target
.pid
)
1002 fprintf(output
, "process id \'%s", target
.pid
);
1004 fprintf(output
, "thread id \'%s", target
.tid
);
1006 fprintf(output
, "\'");
1008 fprintf(output
, " (%d runs)", run_count
);
1009 fprintf(output
, ":\n\n");
1013 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1014 print_counter(counter
);
1016 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
1017 print_counter_aggr(counter
);
1022 fprintf(output
, "\n");
1023 fprintf(output
, " %17.9f seconds time elapsed",
1024 avg_stats(&walltime_nsecs_stats
)/1e9
);
1025 if (run_count
> 1) {
1026 fprintf(output
, " ");
1027 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
1028 avg_stats(&walltime_nsecs_stats
));
1030 fprintf(output
, "\n\n");
1034 static volatile int signr
= -1;
1036 static void skip_signal(int signo
)
1044 static void sig_atexit(void)
1046 if (child_pid
!= -1)
1047 kill(child_pid
, SIGTERM
);
1052 signal(signr
, SIG_DFL
);
1053 kill(getpid(), signr
);
1056 static const char * const stat_usage
[] = {
1057 "perf stat [<options>] [<command>]",
1061 static int stat__set_big_num(const struct option
*opt __used
,
1062 const char *s __used
, int unset
)
1064 big_num_opt
= unset
? 0 : 1;
1068 static bool append_file
;
1070 static const struct option options
[] = {
1071 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1072 "event selector. use 'perf list' to list available events",
1073 parse_events_option
),
1074 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1075 "event filter", parse_filter
),
1076 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1077 "child tasks do not inherit counters"),
1078 OPT_STRING('p', "pid", &target
.pid
, "pid",
1079 "stat events on existing process id"),
1080 OPT_STRING('t', "tid", &target
.tid
, "tid",
1081 "stat events on existing thread id"),
1082 OPT_BOOLEAN('a', "all-cpus", &target
.system_wide
,
1083 "system-wide collection from all CPUs"),
1084 OPT_BOOLEAN('g', "group", &group
,
1085 "put the counters into a counter group"),
1086 OPT_BOOLEAN('c', "scale", &scale
,
1087 "scale/normalize counters"),
1088 OPT_INCR('v', "verbose", &verbose
,
1089 "be more verbose (show counter open errors, etc)"),
1090 OPT_INTEGER('r', "repeat", &run_count
,
1091 "repeat command and print average + stddev (max: 100)"),
1092 OPT_BOOLEAN('n', "null", &null_run
,
1093 "null run - dont start any counters"),
1094 OPT_INCR('d', "detailed", &detailed_run
,
1095 "detailed run - start a lot of events"),
1096 OPT_BOOLEAN('S', "sync", &sync_run
,
1097 "call sync() before starting a run"),
1098 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1099 "print large numbers with thousands\' separators",
1101 OPT_STRING('C', "cpu", &target
.cpu_list
, "cpu",
1102 "list of cpus to monitor in system-wide"),
1103 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
1104 "disable CPU count aggregation"),
1105 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1106 "print counts with custom separator"),
1107 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1108 "monitor event in cgroup name only",
1110 OPT_STRING('o', "output", &output_name
, "file",
1111 "output file name"),
1112 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1113 OPT_INTEGER(0, "log-fd", &output_fd
,
1114 "log output to fd, instead of stderr"),
1119 * Add default attributes, if there were no attributes specified or
1120 * if -d/--detailed, -d -d or -d -d -d is used:
1122 static int add_default_attributes(void)
1124 /* Set attrs if no event is selected and !null_run: */
1128 if (!evsel_list
->nr_entries
) {
1129 if (perf_evlist__add_default_attrs(evsel_list
, default_attrs
) < 0)
1133 /* Detailed events get appended to the event list: */
1135 if (detailed_run
< 1)
1138 /* Append detailed run extra attributes: */
1139 if (perf_evlist__add_default_attrs(evsel_list
, detailed_attrs
) < 0)
1142 if (detailed_run
< 2)
1145 /* Append very detailed run extra attributes: */
1146 if (perf_evlist__add_default_attrs(evsel_list
, very_detailed_attrs
) < 0)
1149 if (detailed_run
< 3)
1152 /* Append very, very detailed run extra attributes: */
1153 return perf_evlist__add_default_attrs(evsel_list
, very_very_detailed_attrs
);
1156 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
1158 struct perf_evsel
*pos
;
1159 int status
= -ENOMEM
;
1162 setlocale(LC_ALL
, "");
1164 evsel_list
= perf_evlist__new(NULL
, NULL
);
1165 if (evsel_list
== NULL
)
1168 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1169 PARSE_OPT_STOP_AT_NON_OPTION
);
1172 if (output_name
&& strcmp(output_name
, "-"))
1175 if (output_name
&& output_fd
) {
1176 fprintf(stderr
, "cannot use both --output and --log-fd\n");
1177 usage_with_options(stat_usage
, options
);
1180 if (output_fd
< 0) {
1181 fprintf(stderr
, "argument to --log-fd must be a > 0\n");
1182 usage_with_options(stat_usage
, options
);
1187 mode
= append_file
? "a" : "w";
1189 output
= fopen(output_name
, mode
);
1191 perror("failed to create output file");
1194 clock_gettime(CLOCK_REALTIME
, &tm
);
1195 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1196 } else if (output_fd
> 0) {
1197 mode
= append_file
? "a" : "w";
1198 output
= fdopen(output_fd
, mode
);
1200 perror("Failed opening logfd");
1207 if (!strcmp(csv_sep
, "\\t"))
1210 csv_sep
= DEFAULT_SEPARATOR
;
1213 * let the spreadsheet do the pretty-printing
1216 /* User explicitly passed -B? */
1217 if (big_num_opt
== 1) {
1218 fprintf(stderr
, "-B option not supported with -x\n");
1219 usage_with_options(stat_usage
, options
);
1220 } else /* Nope, so disable big number formatting */
1222 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1225 if (!argc
&& !perf_target__has_task(&target
))
1226 usage_with_options(stat_usage
, options
);
1228 usage_with_options(stat_usage
, options
);
1230 /* no_aggr, cgroup are for system-wide only */
1231 if ((no_aggr
|| nr_cgroups
) && !perf_target__has_cpu(&target
)) {
1232 fprintf(stderr
, "both cgroup and no-aggregation "
1233 "modes only available in system-wide mode\n");
1235 usage_with_options(stat_usage
, options
);
1238 if (add_default_attributes())
1241 perf_target__validate(&target
);
1243 if (perf_evlist__create_maps(evsel_list
, &target
) < 0) {
1244 if (perf_target__has_task(&target
))
1245 pr_err("Problems finding threads of monitor\n");
1246 if (perf_target__has_cpu(&target
))
1247 perror("failed to parse CPUs map");
1249 usage_with_options(stat_usage
, options
);
1253 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1254 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
1255 perf_evsel__alloc_counts(pos
, evsel_list
->cpus
->nr
) < 0)
1260 * We dont want to block the signals - that would cause
1261 * child tasks to inherit that and Ctrl-C would not work.
1262 * What we want is for Ctrl-C to work in the exec()-ed
1263 * task, but being ignored by perf stat itself:
1266 signal(SIGINT
, skip_signal
);
1267 signal(SIGALRM
, skip_signal
);
1268 signal(SIGABRT
, skip_signal
);
1271 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
1272 if (run_count
!= 1 && verbose
)
1273 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1279 status
= run_perf_stat(argc
, argv
);
1283 print_stat(argc
, argv
);
1285 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
1286 perf_evsel__free_stat_priv(pos
);
1287 perf_evlist__delete_maps(evsel_list
);
1289 perf_evlist__delete(evsel_list
);