4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ~/hackbench 10
12 Performance counter stats for '/home/mingo/hackbench':
14 1255.538611 task clock ticks # 10.143 CPU utilization factor
15 54011 context switches # 0.043 M/sec
16 385 CPU migrations # 0.000 M/sec
17 17755 pagefaults # 0.014 M/sec
18 3808323185 CPU cycles # 3033.219 M/sec
19 1575111190 instructions # 1254.530 M/sec
20 17367895 cache references # 13.833 M/sec
21 7674421 cache misses # 6.112 M/sec
23 Wall-clock time elapsed: 123.786620 msecs
26 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
28 * Improvements and fixes by:
30 * Arjan van de Ven <arjan@linux.intel.com>
31 * Yanmin Zhang <yanmin.zhang@intel.com>
32 * Wu Fengguang <fengguang.wu@intel.com>
33 * Mike Galbraith <efault@gmx.de>
34 * Paul Mackerras <paulus@samba.org>
35 * Jaswinder Singh Rajput <jaswinder@kernel.org>
37 * Released under the GPL v2. (and only v2, not any later version)
42 #include "util/util.h"
43 #include "util/parse-options.h"
44 #include "util/parse-events.h"
46 #include <sys/prctl.h>
49 static struct perf_counter_attr default_attrs
[] = {
51 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
52 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
53 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
54 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
56 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
57 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
58 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_REFERENCES
},
59 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_MISSES
},
65 static int system_wide
= 0;
66 static int verbose
= 0;
67 static int nr_cpus
= 0;
68 static int run_idx
= 0;
70 static int run_count
= 1;
71 static int inherit
= 1;
73 static int target_pid
= -1;
74 static int null_run
= 0;
76 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
78 static u64 runtime_nsecs
[MAX_RUN
];
79 static u64 walltime_nsecs
[MAX_RUN
];
80 static u64 runtime_cycles
[MAX_RUN
];
82 static u64 event_res
[MAX_RUN
][MAX_COUNTERS
][3];
83 static u64 event_scaled
[MAX_RUN
][MAX_COUNTERS
];
85 static u64 event_res_avg
[MAX_COUNTERS
][3];
86 static u64 event_res_noise
[MAX_COUNTERS
][3];
88 static u64 event_scaled_avg
[MAX_COUNTERS
];
90 static u64 runtime_nsecs_avg
;
91 static u64 runtime_nsecs_noise
;
93 static u64 walltime_nsecs_avg
;
94 static u64 walltime_nsecs_noise
;
96 static u64 runtime_cycles_avg
;
97 static u64 runtime_cycles_noise
;
99 #define ERR_PERF_OPEN \
100 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
102 static void create_perf_stat_counter(int counter
)
104 struct perf_counter_attr
*attr
= attrs
+ counter
;
107 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
108 PERF_FORMAT_TOTAL_TIME_RUNNING
;
112 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
113 fd
[cpu
][counter
] = sys_perf_counter_open(attr
, -1, cpu
, -1, 0);
114 if (fd
[cpu
][counter
] < 0 && verbose
)
115 fprintf(stderr
, ERR_PERF_OPEN
, counter
,
116 fd
[cpu
][counter
], strerror(errno
));
119 attr
->inherit
= inherit
;
122 fd
[0][counter
] = sys_perf_counter_open(attr
, 0, -1, -1, 0);
123 if (fd
[0][counter
] < 0 && verbose
)
124 fprintf(stderr
, ERR_PERF_OPEN
, counter
,
125 fd
[0][counter
], strerror(errno
));
130 * Does the counter have nsecs as a unit?
132 static inline int nsec_counter(int counter
)
134 if (attrs
[counter
].type
!= PERF_TYPE_SOFTWARE
)
137 if (attrs
[counter
].config
== PERF_COUNT_SW_CPU_CLOCK
)
140 if (attrs
[counter
].config
== PERF_COUNT_SW_TASK_CLOCK
)
147 * Read out the results of a single counter:
149 static void read_counter(int counter
)
151 u64
*count
, single_count
[3];
156 count
= event_res
[run_idx
][counter
];
158 count
[0] = count
[1] = count
[2] = 0;
161 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
162 if (fd
[cpu
][counter
] < 0)
165 res
= read(fd
[cpu
][counter
], single_count
, nv
* sizeof(u64
));
166 assert(res
== nv
* sizeof(u64
));
167 close(fd
[cpu
][counter
]);
168 fd
[cpu
][counter
] = -1;
170 count
[0] += single_count
[0];
172 count
[1] += single_count
[1];
173 count
[2] += single_count
[2];
180 event_scaled
[run_idx
][counter
] = -1;
185 if (count
[2] < count
[1]) {
186 event_scaled
[run_idx
][counter
] = 1;
187 count
[0] = (unsigned long long)
188 ((double)count
[0] * count
[1] / count
[2] + 0.5);
192 * Save the full runtime - to allow normalization during printout:
194 if (attrs
[counter
].type
== PERF_TYPE_SOFTWARE
&&
195 attrs
[counter
].config
== PERF_COUNT_SW_TASK_CLOCK
)
196 runtime_nsecs
[run_idx
] = count
[0];
197 if (attrs
[counter
].type
== PERF_TYPE_HARDWARE
&&
198 attrs
[counter
].config
== PERF_COUNT_HW_CPU_CYCLES
)
199 runtime_cycles
[run_idx
] = count
[0];
202 static int run_perf_stat(int argc
, const char **argv
)
204 unsigned long long t0
, t1
;
212 for (counter
= 0; counter
< nr_counters
; counter
++)
213 create_perf_stat_counter(counter
);
216 * Enable counters and exec the command:
219 prctl(PR_TASK_PERF_COUNTERS_ENABLE
);
221 if ((pid
= fork()) < 0)
222 perror("failed to fork");
225 if (execvp(argv
[0], (char **)argv
)) {
233 prctl(PR_TASK_PERF_COUNTERS_DISABLE
);
236 walltime_nsecs
[run_idx
] = t1
- t0
;
238 for (counter
= 0; counter
< nr_counters
; counter
++)
239 read_counter(counter
);
241 return WEXITSTATUS(status
);
244 static void print_noise(u64
*count
, u64
*noise
)
247 fprintf(stderr
, " ( +- %7.3f%% )",
248 (double)noise
[0]/(count
[0]+1)*100.0);
251 static void nsec_printout(int counter
, u64
*count
, u64
*noise
)
253 double msecs
= (double)count
[0] / 1000000;
255 fprintf(stderr
, " %14.6f %-24s", msecs
, event_name(counter
));
257 if (attrs
[counter
].type
== PERF_TYPE_SOFTWARE
&&
258 attrs
[counter
].config
== PERF_COUNT_SW_TASK_CLOCK
) {
260 if (walltime_nsecs_avg
)
261 fprintf(stderr
, " # %10.3f CPUs ",
262 (double)count
[0] / (double)walltime_nsecs_avg
);
264 print_noise(count
, noise
);
267 static void abs_printout(int counter
, u64
*count
, u64
*noise
)
269 fprintf(stderr
, " %14Ld %-24s", count
[0], event_name(counter
));
271 if (runtime_cycles_avg
&&
272 attrs
[counter
].type
== PERF_TYPE_HARDWARE
&&
273 attrs
[counter
].config
== PERF_COUNT_HW_INSTRUCTIONS
) {
275 fprintf(stderr
, " # %10.3f IPC ",
276 (double)count
[0] / (double)runtime_cycles_avg
);
278 if (runtime_nsecs_avg
) {
279 fprintf(stderr
, " # %10.3f M/sec",
280 (double)count
[0]/runtime_nsecs_avg
*1000.0);
283 print_noise(count
, noise
);
287 * Print out the results of a single counter:
289 static void print_counter(int counter
)
294 count
= event_res_avg
[counter
];
295 noise
= event_res_noise
[counter
];
296 scaled
= event_scaled_avg
[counter
];
299 fprintf(stderr
, " %14s %-24s\n",
300 "<not counted>", event_name(counter
));
304 if (nsec_counter(counter
))
305 nsec_printout(counter
, count
, noise
);
307 abs_printout(counter
, count
, noise
);
310 fprintf(stderr
, " (scaled from %.2f%%)",
311 (double) count
[2] / count
[1] * 100);
313 fprintf(stderr
, "\n");
317 * normalize_noise noise values down to stddev:
319 static void normalize_noise(u64
*val
)
323 res
= (double)*val
/ (run_count
* sqrt((double)run_count
));
328 static void update_avg(const char *name
, int idx
, u64
*avg
, u64
*val
)
333 fprintf(stderr
, "debug: %20s[%d]: %Ld\n", name
, idx
, *val
);
336 * Calculate the averages and noises:
338 static void calc_avg(void)
343 fprintf(stderr
, "\n");
345 for (i
= 0; i
< run_count
; i
++) {
346 update_avg("runtime", 0, &runtime_nsecs_avg
, runtime_nsecs
+ i
);
347 update_avg("walltime", 0, &walltime_nsecs_avg
, walltime_nsecs
+ i
);
348 update_avg("runtime_cycles", 0, &runtime_cycles_avg
, runtime_cycles
+ i
);
350 for (j
= 0; j
< nr_counters
; j
++) {
351 update_avg("counter/0", j
,
352 event_res_avg
[j
]+0, event_res
[i
][j
]+0);
353 update_avg("counter/1", j
,
354 event_res_avg
[j
]+1, event_res
[i
][j
]+1);
355 update_avg("counter/2", j
,
356 event_res_avg
[j
]+2, event_res
[i
][j
]+2);
357 if (event_scaled
[i
][j
] != -1)
358 update_avg("scaled", j
,
359 event_scaled_avg
+ j
, event_scaled
[i
]+j
);
361 event_scaled_avg
[j
] = -1;
364 runtime_nsecs_avg
/= run_count
;
365 walltime_nsecs_avg
/= run_count
;
366 runtime_cycles_avg
/= run_count
;
368 for (j
= 0; j
< nr_counters
; j
++) {
369 event_res_avg
[j
][0] /= run_count
;
370 event_res_avg
[j
][1] /= run_count
;
371 event_res_avg
[j
][2] /= run_count
;
374 for (i
= 0; i
< run_count
; i
++) {
375 runtime_nsecs_noise
+=
376 abs((s64
)(runtime_nsecs
[i
] - runtime_nsecs_avg
));
377 walltime_nsecs_noise
+=
378 abs((s64
)(walltime_nsecs
[i
] - walltime_nsecs_avg
));
379 runtime_cycles_noise
+=
380 abs((s64
)(runtime_cycles
[i
] - runtime_cycles_avg
));
382 for (j
= 0; j
< nr_counters
; j
++) {
383 event_res_noise
[j
][0] +=
384 abs((s64
)(event_res
[i
][j
][0] - event_res_avg
[j
][0]));
385 event_res_noise
[j
][1] +=
386 abs((s64
)(event_res
[i
][j
][1] - event_res_avg
[j
][1]));
387 event_res_noise
[j
][2] +=
388 abs((s64
)(event_res
[i
][j
][2] - event_res_avg
[j
][2]));
392 normalize_noise(&runtime_nsecs_noise
);
393 normalize_noise(&walltime_nsecs_noise
);
394 normalize_noise(&runtime_cycles_noise
);
396 for (j
= 0; j
< nr_counters
; j
++) {
397 normalize_noise(&event_res_noise
[j
][0]);
398 normalize_noise(&event_res_noise
[j
][1]);
399 normalize_noise(&event_res_noise
[j
][2]);
403 static void print_stat(int argc
, const char **argv
)
411 fprintf(stderr
, "\n");
412 fprintf(stderr
, " Performance counter stats for \'%s", argv
[0]);
414 for (i
= 1; i
< argc
; i
++)
415 fprintf(stderr
, " %s", argv
[i
]);
417 fprintf(stderr
, "\'");
419 fprintf(stderr
, " (%d runs)", run_count
);
420 fprintf(stderr
, ":\n\n");
422 for (counter
= 0; counter
< nr_counters
; counter
++)
423 print_counter(counter
);
425 fprintf(stderr
, "\n");
426 fprintf(stderr
, " %14.9f seconds time elapsed",
427 (double)walltime_nsecs_avg
/1e9
);
429 fprintf(stderr
, " ( +- %7.3f%% )",
430 100.0*(double)walltime_nsecs_noise
/(double)walltime_nsecs_avg
);
432 fprintf(stderr
, "\n\n");
435 static volatile int signr
= -1;
437 static void skip_signal(int signo
)
442 static void sig_atexit(void)
447 signal(signr
, SIG_DFL
);
448 kill(getpid(), signr
);
451 static const char * const stat_usage
[] = {
452 "perf stat [<options>] <command>",
456 static const struct option options
[] = {
457 OPT_CALLBACK('e', "event", NULL
, "event",
458 "event selector. use 'perf list' to list available events",
460 OPT_BOOLEAN('i', "inherit", &inherit
,
461 "child tasks inherit counters"),
462 OPT_INTEGER('p', "pid", &target_pid
,
463 "stat events on existing pid"),
464 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
465 "system-wide collection from all CPUs"),
466 OPT_BOOLEAN('S', "scale", &scale
,
467 "scale/normalize counters"),
468 OPT_BOOLEAN('v', "verbose", &verbose
,
469 "be more verbose (show counter open errors, etc)"),
470 OPT_INTEGER('r', "repeat", &run_count
,
471 "repeat command and print average + stddev (max: 100)"),
472 OPT_BOOLEAN('n', "null", &null_run
,
473 "null run - dont start any counters"),
477 int cmd_stat(int argc
, const char **argv
, const char *prefix
)
481 argc
= parse_options(argc
, argv
, options
, stat_usage
, 0);
483 usage_with_options(stat_usage
, options
);
484 if (run_count
<= 0 || run_count
> MAX_RUN
)
485 usage_with_options(stat_usage
, options
);
487 /* Set attrs and nr_counters if no event is selected and !null_run */
488 if (!null_run
&& !nr_counters
) {
489 memcpy(attrs
, default_attrs
, sizeof(default_attrs
));
490 nr_counters
= ARRAY_SIZE(default_attrs
);
493 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
494 assert(nr_cpus
<= MAX_NR_CPUS
);
495 assert(nr_cpus
>= 0);
498 * We dont want to block the signals - that would cause
499 * child tasks to inherit that and Ctrl-C would not work.
500 * What we want is for Ctrl-C to work in the exec()-ed
501 * task, but being ignored by perf stat itself:
504 signal(SIGINT
, skip_signal
);
505 signal(SIGALRM
, skip_signal
);
506 signal(SIGABRT
, skip_signal
);
509 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
510 if (run_count
!= 1 && verbose
)
511 fprintf(stderr
, "[ perf stat: executing run #%d ... ]\n", run_idx
+ 1);
512 status
= run_perf_stat(argc
, argv
);
515 print_stat(argc
, argv
);