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 struct perf_evlist
*evsel_list
;
178 static bool system_wide
= false;
179 static int run_idx
= 0;
181 static int run_count
= 1;
182 static bool no_inherit
= false;
183 static bool scale
= true;
184 static bool no_aggr
= false;
185 static pid_t target_pid
= -1;
186 static pid_t target_tid
= -1;
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 *cpu_list
;
194 static const char *csv_sep
= NULL
;
195 static bool csv_output
= false;
196 static bool group
= false;
197 static const char *output_name
= NULL
;
198 static FILE *output
= NULL
;
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
= stats
->M2
/ (stats
->n
- 1);
257 double variance_mean
= variance
/ stats
->n
;
259 return sqrt(variance_mean
);
262 struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
263 struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
264 struct stats runtime_stalled_cycles_front_stats
[MAX_NR_CPUS
];
265 struct stats runtime_stalled_cycles_back_stats
[MAX_NR_CPUS
];
266 struct stats runtime_branches_stats
[MAX_NR_CPUS
];
267 struct stats runtime_cacherefs_stats
[MAX_NR_CPUS
];
268 struct stats runtime_l1_dcache_stats
[MAX_NR_CPUS
];
269 struct stats runtime_l1_icache_stats
[MAX_NR_CPUS
];
270 struct stats runtime_ll_cache_stats
[MAX_NR_CPUS
];
271 struct stats runtime_itlb_cache_stats
[MAX_NR_CPUS
];
272 struct stats runtime_dtlb_cache_stats
[MAX_NR_CPUS
];
273 struct stats walltime_nsecs_stats
;
275 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
277 struct perf_event_attr
*attr
= &evsel
->attr
;
280 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
281 PERF_FORMAT_TOTAL_TIME_RUNNING
;
283 attr
->inherit
= !no_inherit
;
286 return perf_evsel__open_per_cpu(evsel
, evsel_list
->cpus
, group
);
288 if (target_pid
== -1 && target_tid
== -1) {
290 attr
->enable_on_exec
= 1;
293 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
, group
);
297 * Does the counter have nsecs as a unit?
299 static inline int nsec_counter(struct perf_evsel
*evsel
)
301 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
302 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
309 * Update various tracking values we maintain to print
310 * more semantic information such as miss/hit ratios,
311 * instruction rates, etc:
313 static void update_shadow_stats(struct perf_evsel
*counter
, u64
*count
)
315 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
316 update_stats(&runtime_nsecs_stats
[0], count
[0]);
317 else if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
318 update_stats(&runtime_cycles_stats
[0], count
[0]);
319 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
))
320 update_stats(&runtime_stalled_cycles_front_stats
[0], count
[0]);
321 else if (perf_evsel__match(counter
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
))
322 update_stats(&runtime_stalled_cycles_back_stats
[0], count
[0]);
323 else if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
324 update_stats(&runtime_branches_stats
[0], count
[0]);
325 else if (perf_evsel__match(counter
, HARDWARE
, HW_CACHE_REFERENCES
))
326 update_stats(&runtime_cacherefs_stats
[0], count
[0]);
327 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1D
))
328 update_stats(&runtime_l1_dcache_stats
[0], count
[0]);
329 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_L1I
))
330 update_stats(&runtime_l1_icache_stats
[0], count
[0]);
331 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_LL
))
332 update_stats(&runtime_ll_cache_stats
[0], count
[0]);
333 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_DTLB
))
334 update_stats(&runtime_dtlb_cache_stats
[0], count
[0]);
335 else if (perf_evsel__match(counter
, HW_CACHE
, HW_CACHE_ITLB
))
336 update_stats(&runtime_itlb_cache_stats
[0], count
[0]);
340 * Read out the results of a single counter:
341 * aggregate counts across CPUs in system-wide mode
343 static int read_counter_aggr(struct perf_evsel
*counter
)
345 struct perf_stat
*ps
= counter
->priv
;
346 u64
*count
= counter
->counts
->aggr
.values
;
349 if (__perf_evsel__read(counter
, evsel_list
->cpus
->nr
,
350 evsel_list
->threads
->nr
, scale
) < 0)
353 for (i
= 0; i
< 3; i
++)
354 update_stats(&ps
->res_stats
[i
], count
[i
]);
357 fprintf(output
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
358 event_name(counter
), count
[0], count
[1], count
[2]);
362 * Save the full runtime - to allow normalization during printout:
364 update_shadow_stats(counter
, count
);
370 * Read out the results of a single counter:
371 * do not aggregate counts across CPUs in system-wide mode
373 static int read_counter(struct perf_evsel
*counter
)
378 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
379 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
382 count
= counter
->counts
->cpu
[cpu
].values
;
384 update_shadow_stats(counter
, count
);
390 static int run_perf_stat(int argc __used
, const char **argv
)
392 unsigned long long t0
, t1
;
393 struct perf_evsel
*counter
;
395 int child_ready_pipe
[2], go_pipe
[2];
396 const bool forks
= (argc
> 0);
399 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
400 perror("failed to create pipes");
405 if ((child_pid
= fork()) < 0)
406 perror("failed to fork");
409 close(child_ready_pipe
[0]);
411 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
414 * Do a dummy execvp to get the PLT entry resolved,
415 * so we avoid the resolver overhead on the real
418 execvp("", (char **)argv
);
421 * Tell the parent we're ready to go
423 close(child_ready_pipe
[1]);
426 * Wait until the parent tells us to go.
428 if (read(go_pipe
[0], &buf
, 1) == -1)
429 perror("unable to read pipe");
431 execvp(argv
[0], (char **)argv
);
437 if (target_tid
== -1 && target_pid
== -1 && !system_wide
)
438 evsel_list
->threads
->map
[0] = child_pid
;
441 * Wait for the child to be ready to exec.
443 close(child_ready_pipe
[1]);
445 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
446 perror("unable to read pipe");
447 close(child_ready_pipe
[0]);
450 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
451 if (create_perf_stat_counter(counter
) < 0) {
452 if (errno
== EINVAL
|| errno
== ENOSYS
|| errno
== ENOENT
) {
454 ui__warning("%s event is not supported by the kernel.\n",
455 event_name(counter
));
456 counter
->supported
= false;
460 if (errno
== EPERM
|| errno
== EACCES
) {
461 error("You may not have permission to collect %sstats.\n"
462 "\t Consider tweaking"
463 " /proc/sys/kernel/perf_event_paranoid or running as root.",
464 system_wide
? "system-wide " : "");
466 error("open_counter returned with %d (%s). "
467 "/bin/dmesg may provide additional information.\n",
468 errno
, strerror(errno
));
471 kill(child_pid
, SIGTERM
);
472 die("Not all events could be opened.\n");
475 counter
->supported
= true;
478 if (perf_evlist__set_filters(evsel_list
)) {
479 error("failed to set filter with %d (%s)\n", errno
,
485 * Enable counters and exec the command:
493 while(!done
) sleep(1);
498 update_stats(&walltime_nsecs_stats
, t1
- t0
);
501 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
502 read_counter(counter
);
503 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
, 1);
506 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
507 read_counter_aggr(counter
);
508 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
,
509 evsel_list
->threads
->nr
);
513 return WEXITSTATUS(status
);
516 static void print_noise_pct(double total
, double avg
)
521 pct
= 100.0*total
/avg
;
524 fprintf(output
, "%s%.2f%%", csv_sep
, pct
);
526 fprintf(output
, " ( +-%6.2f%% )", pct
);
529 static void print_noise(struct perf_evsel
*evsel
, double avg
)
531 struct perf_stat
*ps
;
537 print_noise_pct(stddev_stats(&ps
->res_stats
[0]), avg
);
540 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
542 double msecs
= avg
/ 1e6
;
543 char cpustr
[16] = { '\0', };
544 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
547 sprintf(cpustr
, "CPU%*d%s",
549 evsel_list
->cpus
->map
[cpu
], csv_sep
);
551 fprintf(output
, fmt
, cpustr
, msecs
, csv_sep
, event_name(evsel
));
554 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
559 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
560 fprintf(output
, " # %8.3f CPUs utilized ",
561 avg
/ avg_stats(&walltime_nsecs_stats
));
564 static void print_stalled_cycles_frontend(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
566 double total
, ratio
= 0.0;
569 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
572 ratio
= avg
/ total
* 100.0;
574 color
= PERF_COLOR_NORMAL
;
576 color
= PERF_COLOR_RED
;
577 else if (ratio
> 30.0)
578 color
= PERF_COLOR_MAGENTA
;
579 else if (ratio
> 10.0)
580 color
= PERF_COLOR_YELLOW
;
582 fprintf(output
, " # ");
583 color_fprintf(output
, color
, "%6.2f%%", ratio
);
584 fprintf(output
, " frontend cycles idle ");
587 static void print_stalled_cycles_backend(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
589 double total
, ratio
= 0.0;
592 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
595 ratio
= avg
/ total
* 100.0;
597 color
= PERF_COLOR_NORMAL
;
599 color
= PERF_COLOR_RED
;
600 else if (ratio
> 50.0)
601 color
= PERF_COLOR_MAGENTA
;
602 else if (ratio
> 20.0)
603 color
= PERF_COLOR_YELLOW
;
605 fprintf(output
, " # ");
606 color_fprintf(output
, color
, "%6.2f%%", ratio
);
607 fprintf(output
, " backend cycles idle ");
610 static void print_branch_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
612 double total
, ratio
= 0.0;
615 total
= avg_stats(&runtime_branches_stats
[cpu
]);
618 ratio
= avg
/ total
* 100.0;
620 color
= PERF_COLOR_NORMAL
;
622 color
= PERF_COLOR_RED
;
623 else if (ratio
> 10.0)
624 color
= PERF_COLOR_MAGENTA
;
625 else if (ratio
> 5.0)
626 color
= PERF_COLOR_YELLOW
;
628 fprintf(output
, " # ");
629 color_fprintf(output
, color
, "%6.2f%%", ratio
);
630 fprintf(output
, " of all branches ");
633 static void print_l1_dcache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
635 double total
, ratio
= 0.0;
638 total
= avg_stats(&runtime_l1_dcache_stats
[cpu
]);
641 ratio
= avg
/ total
* 100.0;
643 color
= PERF_COLOR_NORMAL
;
645 color
= PERF_COLOR_RED
;
646 else if (ratio
> 10.0)
647 color
= PERF_COLOR_MAGENTA
;
648 else if (ratio
> 5.0)
649 color
= PERF_COLOR_YELLOW
;
651 fprintf(output
, " # ");
652 color_fprintf(output
, color
, "%6.2f%%", ratio
);
653 fprintf(output
, " of all L1-dcache hits ");
656 static void print_l1_icache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
658 double total
, ratio
= 0.0;
661 total
= avg_stats(&runtime_l1_icache_stats
[cpu
]);
664 ratio
= avg
/ total
* 100.0;
666 color
= PERF_COLOR_NORMAL
;
668 color
= PERF_COLOR_RED
;
669 else if (ratio
> 10.0)
670 color
= PERF_COLOR_MAGENTA
;
671 else if (ratio
> 5.0)
672 color
= PERF_COLOR_YELLOW
;
674 fprintf(output
, " # ");
675 color_fprintf(output
, color
, "%6.2f%%", ratio
);
676 fprintf(output
, " of all L1-icache hits ");
679 static void print_dtlb_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
681 double total
, ratio
= 0.0;
684 total
= avg_stats(&runtime_dtlb_cache_stats
[cpu
]);
687 ratio
= avg
/ total
* 100.0;
689 color
= PERF_COLOR_NORMAL
;
691 color
= PERF_COLOR_RED
;
692 else if (ratio
> 10.0)
693 color
= PERF_COLOR_MAGENTA
;
694 else if (ratio
> 5.0)
695 color
= PERF_COLOR_YELLOW
;
697 fprintf(output
, " # ");
698 color_fprintf(output
, color
, "%6.2f%%", ratio
);
699 fprintf(output
, " of all dTLB cache hits ");
702 static void print_itlb_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
704 double total
, ratio
= 0.0;
707 total
= avg_stats(&runtime_itlb_cache_stats
[cpu
]);
710 ratio
= avg
/ total
* 100.0;
712 color
= PERF_COLOR_NORMAL
;
714 color
= PERF_COLOR_RED
;
715 else if (ratio
> 10.0)
716 color
= PERF_COLOR_MAGENTA
;
717 else if (ratio
> 5.0)
718 color
= PERF_COLOR_YELLOW
;
720 fprintf(output
, " # ");
721 color_fprintf(output
, color
, "%6.2f%%", ratio
);
722 fprintf(output
, " of all iTLB cache hits ");
725 static void print_ll_cache_misses(int cpu
, struct perf_evsel
*evsel __used
, double avg
)
727 double total
, ratio
= 0.0;
730 total
= avg_stats(&runtime_ll_cache_stats
[cpu
]);
733 ratio
= avg
/ total
* 100.0;
735 color
= PERF_COLOR_NORMAL
;
737 color
= PERF_COLOR_RED
;
738 else if (ratio
> 10.0)
739 color
= PERF_COLOR_MAGENTA
;
740 else if (ratio
> 5.0)
741 color
= PERF_COLOR_YELLOW
;
743 fprintf(output
, " # ");
744 color_fprintf(output
, color
, "%6.2f%%", ratio
);
745 fprintf(output
, " of all LL-cache hits ");
748 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
750 double total
, ratio
= 0.0;
751 char cpustr
[16] = { '\0', };
757 fmt
= "%s%'18.0f%s%-25s";
759 fmt
= "%s%18.0f%s%-25s";
762 sprintf(cpustr
, "CPU%*d%s",
764 evsel_list
->cpus
->map
[cpu
], csv_sep
);
768 fprintf(output
, fmt
, cpustr
, avg
, csv_sep
, event_name(evsel
));
771 fprintf(output
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
776 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
777 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
782 fprintf(output
, " # %5.2f insns per cycle ", ratio
);
784 total
= avg_stats(&runtime_stalled_cycles_front_stats
[cpu
]);
785 total
= max(total
, avg_stats(&runtime_stalled_cycles_back_stats
[cpu
]));
789 fprintf(output
, "\n # %5.2f stalled cycles per insn", ratio
);
792 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
793 runtime_branches_stats
[cpu
].n
!= 0) {
794 print_branch_misses(cpu
, evsel
, avg
);
796 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
797 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1D
|
798 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
799 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
800 runtime_l1_dcache_stats
[cpu
].n
!= 0) {
801 print_l1_dcache_misses(cpu
, evsel
, avg
);
803 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
804 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_L1I
|
805 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
806 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
807 runtime_l1_icache_stats
[cpu
].n
!= 0) {
808 print_l1_icache_misses(cpu
, evsel
, avg
);
810 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
811 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_DTLB
|
812 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
813 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
814 runtime_dtlb_cache_stats
[cpu
].n
!= 0) {
815 print_dtlb_cache_misses(cpu
, evsel
, avg
);
817 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
818 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_ITLB
|
819 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
820 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
821 runtime_itlb_cache_stats
[cpu
].n
!= 0) {
822 print_itlb_cache_misses(cpu
, evsel
, avg
);
824 evsel
->attr
.type
== PERF_TYPE_HW_CACHE
&&
825 evsel
->attr
.config
== ( PERF_COUNT_HW_CACHE_LL
|
826 ((PERF_COUNT_HW_CACHE_OP_READ
) << 8) |
827 ((PERF_COUNT_HW_CACHE_RESULT_MISS
) << 16)) &&
828 runtime_ll_cache_stats
[cpu
].n
!= 0) {
829 print_ll_cache_misses(cpu
, evsel
, avg
);
830 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CACHE_MISSES
) &&
831 runtime_cacherefs_stats
[cpu
].n
!= 0) {
832 total
= avg_stats(&runtime_cacherefs_stats
[cpu
]);
835 ratio
= avg
* 100 / total
;
837 fprintf(output
, " # %8.3f %% of all cache refs ", ratio
);
839 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_FRONTEND
)) {
840 print_stalled_cycles_frontend(cpu
, evsel
, avg
);
841 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_STALLED_CYCLES_BACKEND
)) {
842 print_stalled_cycles_backend(cpu
, evsel
, avg
);
843 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_CPU_CYCLES
)) {
844 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
847 ratio
= 1.0 * avg
/ total
;
849 fprintf(output
, " # %8.3f GHz ", ratio
);
850 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
851 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
854 ratio
= 1000.0 * avg
/ total
;
856 fprintf(output
, " # %8.3f M/sec ", ratio
);
858 fprintf(output
, " ");
863 * Print out the results of a single counter:
864 * aggregated counts in system-wide mode
866 static void print_counter_aggr(struct perf_evsel
*counter
)
868 struct perf_stat
*ps
= counter
->priv
;
869 double avg
= avg_stats(&ps
->res_stats
[0]);
870 int scaled
= counter
->counts
->scaled
;
873 fprintf(output
, "%*s%s%*s",
875 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
877 csv_output
? 0 : -24,
878 event_name(counter
));
881 fprintf(output
, "%s%s", csv_sep
, counter
->cgrp
->name
);
887 if (nsec_counter(counter
))
888 nsec_printout(-1, counter
, avg
);
890 abs_printout(-1, counter
, avg
);
892 print_noise(counter
, avg
);
900 double avg_enabled
, avg_running
;
902 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
903 avg_running
= avg_stats(&ps
->res_stats
[2]);
905 fprintf(output
, " [%5.2f%%]", 100 * avg_running
/ avg_enabled
);
907 fprintf(output
, "\n");
911 * Print out the results of a single counter:
912 * does not use aggregated count in system-wide
914 static void print_counter(struct perf_evsel
*counter
)
919 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
920 val
= counter
->counts
->cpu
[cpu
].val
;
921 ena
= counter
->counts
->cpu
[cpu
].ena
;
922 run
= counter
->counts
->cpu
[cpu
].run
;
923 if (run
== 0 || ena
== 0) {
924 fprintf(output
, "CPU%*d%s%*s%s%*s",
926 evsel_list
->cpus
->map
[cpu
], csv_sep
,
928 counter
->supported
? CNTR_NOT_COUNTED
: CNTR_NOT_SUPPORTED
,
930 csv_output
? 0 : -24,
931 event_name(counter
));
934 fprintf(output
, "%s%s",
935 csv_sep
, counter
->cgrp
->name
);
941 if (nsec_counter(counter
))
942 nsec_printout(cpu
, counter
, val
);
944 abs_printout(cpu
, counter
, val
);
947 print_noise(counter
, 1.0);
950 fprintf(output
, " (%.2f%%)",
957 static void print_stat(int argc
, const char **argv
)
959 struct perf_evsel
*counter
;
965 fprintf(output
, "\n");
966 fprintf(output
, " Performance counter stats for ");
967 if(target_pid
== -1 && target_tid
== -1) {
968 fprintf(output
, "\'%s", argv
[0]);
969 for (i
= 1; i
< argc
; i
++)
970 fprintf(output
, " %s", argv
[i
]);
971 } else if (target_pid
!= -1)
972 fprintf(output
, "process id \'%d", target_pid
);
974 fprintf(output
, "thread id \'%d", target_tid
);
976 fprintf(output
, "\'");
978 fprintf(output
, " (%d runs)", run_count
);
979 fprintf(output
, ":\n\n");
983 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
984 print_counter(counter
);
986 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
987 print_counter_aggr(counter
);
992 fprintf(output
, "\n");
993 fprintf(output
, " %17.9f seconds time elapsed",
994 avg_stats(&walltime_nsecs_stats
)/1e9
);
996 fprintf(output
, " ");
997 print_noise_pct(stddev_stats(&walltime_nsecs_stats
),
998 avg_stats(&walltime_nsecs_stats
));
1000 fprintf(output
, "\n\n");
1004 static volatile int signr
= -1;
1006 static void skip_signal(int signo
)
1014 static void sig_atexit(void)
1016 if (child_pid
!= -1)
1017 kill(child_pid
, SIGTERM
);
1022 signal(signr
, SIG_DFL
);
1023 kill(getpid(), signr
);
1026 static const char * const stat_usage
[] = {
1027 "perf stat [<options>] [<command>]",
1031 static int stat__set_big_num(const struct option
*opt __used
,
1032 const char *s __used
, int unset
)
1034 big_num_opt
= unset
? 0 : 1;
1038 static bool append_file
;
1040 static const struct option options
[] = {
1041 OPT_CALLBACK('e', "event", &evsel_list
, "event",
1042 "event selector. use 'perf list' to list available events",
1043 parse_events_option
),
1044 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
1045 "event filter", parse_filter
),
1046 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
1047 "child tasks do not inherit counters"),
1048 OPT_INTEGER('p', "pid", &target_pid
,
1049 "stat events on existing process id"),
1050 OPT_INTEGER('t', "tid", &target_tid
,
1051 "stat events on existing thread id"),
1052 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
1053 "system-wide collection from all CPUs"),
1054 OPT_BOOLEAN('g', "group", &group
,
1055 "put the counters into a counter group"),
1056 OPT_BOOLEAN('c', "scale", &scale
,
1057 "scale/normalize counters"),
1058 OPT_INCR('v', "verbose", &verbose
,
1059 "be more verbose (show counter open errors, etc)"),
1060 OPT_INTEGER('r', "repeat", &run_count
,
1061 "repeat command and print average + stddev (max: 100)"),
1062 OPT_BOOLEAN('n', "null", &null_run
,
1063 "null run - dont start any counters"),
1064 OPT_INCR('d', "detailed", &detailed_run
,
1065 "detailed run - start a lot of events"),
1066 OPT_BOOLEAN('S', "sync", &sync_run
,
1067 "call sync() before starting a run"),
1068 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
1069 "print large numbers with thousands\' separators",
1071 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
1072 "list of cpus to monitor in system-wide"),
1073 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
1074 "disable CPU count aggregation"),
1075 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
1076 "print counts with custom separator"),
1077 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
1078 "monitor event in cgroup name only",
1080 OPT_STRING('o', "output", &output_name
, "file",
1081 "output file name"),
1082 OPT_BOOLEAN(0, "append", &append_file
, "append to the output file"),
1087 * Add default attributes, if there were no attributes specified or
1088 * if -d/--detailed, -d -d or -d -d -d is used:
1090 static int add_default_attributes(void)
1092 struct perf_evsel
*pos
;
1096 /* Set attrs if no event is selected and !null_run: */
1100 if (!evsel_list
->nr_entries
) {
1101 for (c
= 0; c
< ARRAY_SIZE(default_attrs
); c
++) {
1102 pos
= perf_evsel__new(default_attrs
+ c
, c
+ attr_nr
);
1105 perf_evlist__add(evsel_list
, pos
);
1110 /* Detailed events get appended to the event list: */
1112 if (detailed_run
< 1)
1115 /* Append detailed run extra attributes: */
1116 for (c
= 0; c
< ARRAY_SIZE(detailed_attrs
); c
++) {
1117 pos
= perf_evsel__new(detailed_attrs
+ c
, c
+ attr_nr
);
1120 perf_evlist__add(evsel_list
, pos
);
1124 if (detailed_run
< 2)
1127 /* Append very detailed run extra attributes: */
1128 for (c
= 0; c
< ARRAY_SIZE(very_detailed_attrs
); c
++) {
1129 pos
= perf_evsel__new(very_detailed_attrs
+ c
, c
+ attr_nr
);
1132 perf_evlist__add(evsel_list
, pos
);
1135 if (detailed_run
< 3)
1138 /* Append very, very detailed run extra attributes: */
1139 for (c
= 0; c
< ARRAY_SIZE(very_very_detailed_attrs
); c
++) {
1140 pos
= perf_evsel__new(very_very_detailed_attrs
+ c
, c
+ attr_nr
);
1143 perf_evlist__add(evsel_list
, pos
);
1150 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
1152 struct perf_evsel
*pos
;
1153 int status
= -ENOMEM
;
1156 setlocale(LC_ALL
, "");
1158 evsel_list
= perf_evlist__new(NULL
, NULL
);
1159 if (evsel_list
== NULL
)
1162 argc
= parse_options(argc
, argv
, options
, stat_usage
,
1163 PARSE_OPT_STOP_AT_NON_OPTION
);
1166 if (output_name
&& strcmp(output_name
, "-"))
1171 mode
= append_file
? "a" : "w";
1173 output
= fopen(output_name
, mode
);
1175 perror("failed to create output file");
1178 clock_gettime(CLOCK_REALTIME
, &tm
);
1179 fprintf(output
, "# started on %s\n", ctime(&tm
.tv_sec
));
1185 csv_sep
= DEFAULT_SEPARATOR
;
1188 * let the spreadsheet do the pretty-printing
1191 /* User explicitely passed -B? */
1192 if (big_num_opt
== 1) {
1193 fprintf(stderr
, "-B option not supported with -x\n");
1194 usage_with_options(stat_usage
, options
);
1195 } else /* Nope, so disable big number formatting */
1197 } else if (big_num_opt
== 0) /* User passed --no-big-num */
1200 if (!argc
&& target_pid
== -1 && target_tid
== -1)
1201 usage_with_options(stat_usage
, options
);
1203 usage_with_options(stat_usage
, options
);
1205 /* no_aggr, cgroup are for system-wide only */
1206 if ((no_aggr
|| nr_cgroups
) && !system_wide
) {
1207 fprintf(stderr
, "both cgroup and no-aggregation "
1208 "modes only available in system-wide mode\n");
1210 usage_with_options(stat_usage
, options
);
1213 if (add_default_attributes())
1216 if (target_pid
!= -1)
1217 target_tid
= target_pid
;
1219 evsel_list
->threads
= thread_map__new(target_pid
, target_tid
);
1220 if (evsel_list
->threads
== NULL
) {
1221 pr_err("Problems finding threads of monitor\n");
1222 usage_with_options(stat_usage
, options
);
1226 evsel_list
->cpus
= cpu_map__new(cpu_list
);
1228 evsel_list
->cpus
= cpu_map__dummy_new();
1230 if (evsel_list
->cpus
== NULL
) {
1231 perror("failed to parse CPUs map");
1232 usage_with_options(stat_usage
, options
);
1236 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
1237 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
1238 perf_evsel__alloc_counts(pos
, evsel_list
->cpus
->nr
) < 0 ||
1239 perf_evsel__alloc_fd(pos
, evsel_list
->cpus
->nr
, evsel_list
->threads
->nr
) < 0)
1244 * We dont want to block the signals - that would cause
1245 * child tasks to inherit that and Ctrl-C would not work.
1246 * What we want is for Ctrl-C to work in the exec()-ed
1247 * task, but being ignored by perf stat itself:
1250 signal(SIGINT
, skip_signal
);
1251 signal(SIGALRM
, skip_signal
);
1252 signal(SIGABRT
, skip_signal
);
1255 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
1256 if (run_count
!= 1 && verbose
)
1257 fprintf(output
, "[ perf stat: executing run #%d ... ]\n",
1263 status
= run_perf_stat(argc
, argv
);
1267 print_stat(argc
, argv
);
1269 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
1270 perf_evsel__free_stat_priv(pos
);
1271 perf_evlist__delete_maps(evsel_list
);
1273 perf_evlist__delete(evsel_list
);