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"
45 #include "util/event.h"
46 #include "util/debug.h"
47 #include "util/header.h"
48 #include "util/cpumap.h"
50 #include <sys/prctl.h>
53 static struct perf_event_attr default_attrs
[] = {
55 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
56 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
57 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
58 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
60 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
61 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
62 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
63 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
64 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_REFERENCES
},
65 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_MISSES
},
69 static int system_wide
= 0;
70 static unsigned int nr_cpus
= 0;
71 static int run_idx
= 0;
73 static int run_count
= 1;
74 static int inherit
= 1;
76 static pid_t target_pid
= -1;
77 static pid_t child_pid
= -1;
78 static int null_run
= 0;
80 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
82 static int event_scaled
[MAX_COUNTERS
];
84 static volatile int done
= 0;
91 static void update_stats(struct stats
*stats
, u64 val
)
96 delta
= val
- stats
->mean
;
97 stats
->mean
+= delta
/ stats
->n
;
98 stats
->M2
+= delta
*(val
- stats
->mean
);
101 static double avg_stats(struct stats
*stats
)
107 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
109 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
110 * s^2 = -------------------------------
113 * http://en.wikipedia.org/wiki/Stddev
115 * The std dev of the mean is related to the std dev by:
122 static double stddev_stats(struct stats
*stats
)
124 double variance
= stats
->M2
/ (stats
->n
- 1);
125 double variance_mean
= variance
/ stats
->n
;
127 return sqrt(variance_mean
);
130 struct stats event_res_stats
[MAX_COUNTERS
][3];
131 struct stats runtime_nsecs_stats
;
132 struct stats walltime_nsecs_stats
;
133 struct stats runtime_cycles_stats
;
134 struct stats runtime_branches_stats
;
136 #define MATCH_EVENT(t, c, counter) \
137 (attrs[counter].type == PERF_TYPE_##t && \
138 attrs[counter].config == PERF_COUNT_##c)
140 #define ERR_PERF_OPEN \
141 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
143 static void create_perf_stat_counter(int counter
, int pid
)
145 struct perf_event_attr
*attr
= attrs
+ counter
;
148 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
149 PERF_FORMAT_TOTAL_TIME_RUNNING
;
154 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
155 fd
[cpu
][counter
] = sys_perf_event_open(attr
, -1, cpumap
[cpu
], -1, 0);
156 if (fd
[cpu
][counter
] < 0 && verbose
)
157 fprintf(stderr
, ERR_PERF_OPEN
, counter
,
158 fd
[cpu
][counter
], strerror(errno
));
161 attr
->inherit
= inherit
;
163 attr
->enable_on_exec
= 1;
165 fd
[0][counter
] = sys_perf_event_open(attr
, pid
, -1, -1, 0);
166 if (fd
[0][counter
] < 0 && verbose
)
167 fprintf(stderr
, ERR_PERF_OPEN
, counter
,
168 fd
[0][counter
], strerror(errno
));
173 * Does the counter have nsecs as a unit?
175 static inline int nsec_counter(int counter
)
177 if (MATCH_EVENT(SOFTWARE
, SW_CPU_CLOCK
, counter
) ||
178 MATCH_EVENT(SOFTWARE
, SW_TASK_CLOCK
, counter
))
185 * Read out the results of a single counter:
187 static void read_counter(int counter
)
189 u64 count
[3], single_count
[3];
195 count
[0] = count
[1] = count
[2] = 0;
198 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
199 if (fd
[cpu
][counter
] < 0)
202 res
= read(fd
[cpu
][counter
], single_count
, nv
* sizeof(u64
));
203 assert(res
== nv
* sizeof(u64
));
205 close(fd
[cpu
][counter
]);
206 fd
[cpu
][counter
] = -1;
208 count
[0] += single_count
[0];
210 count
[1] += single_count
[1];
211 count
[2] += single_count
[2];
218 event_scaled
[counter
] = -1;
223 if (count
[2] < count
[1]) {
224 event_scaled
[counter
] = 1;
225 count
[0] = (unsigned long long)
226 ((double)count
[0] * count
[1] / count
[2] + 0.5);
230 for (i
= 0; i
< 3; i
++)
231 update_stats(&event_res_stats
[counter
][i
], count
[i
]);
234 fprintf(stderr
, "%s: %Ld %Ld %Ld\n", event_name(counter
),
235 count
[0], count
[1], count
[2]);
239 * Save the full runtime - to allow normalization during printout:
241 if (MATCH_EVENT(SOFTWARE
, SW_TASK_CLOCK
, counter
))
242 update_stats(&runtime_nsecs_stats
, count
[0]);
243 if (MATCH_EVENT(HARDWARE
, HW_CPU_CYCLES
, counter
))
244 update_stats(&runtime_cycles_stats
, count
[0]);
245 if (MATCH_EVENT(HARDWARE
, HW_BRANCH_INSTRUCTIONS
, counter
))
246 update_stats(&runtime_branches_stats
, count
[0]);
249 static int run_perf_stat(int argc __used
, const char **argv
)
251 unsigned long long t0
, t1
;
254 int pid
= target_pid
;
255 int child_ready_pipe
[2], go_pipe
[2];
256 const bool forks
= (target_pid
== -1 && argc
> 0);
262 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
263 perror("failed to create pipes");
268 if ((pid
= fork()) < 0)
269 perror("failed to fork");
272 close(child_ready_pipe
[0]);
274 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
277 * Do a dummy execvp to get the PLT entry resolved,
278 * so we avoid the resolver overhead on the real
281 execvp("", (char **)argv
);
284 * Tell the parent we're ready to go
286 close(child_ready_pipe
[1]);
289 * Wait until the parent tells us to go.
291 if (read(go_pipe
[0], &buf
, 1) == -1)
292 perror("unable to read pipe");
294 execvp(argv
[0], (char **)argv
);
303 * Wait for the child to be ready to exec.
305 close(child_ready_pipe
[1]);
307 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
308 perror("unable to read pipe");
309 close(child_ready_pipe
[0]);
312 for (counter
= 0; counter
< nr_counters
; counter
++)
313 create_perf_stat_counter(counter
, pid
);
316 * Enable counters and exec the command:
329 update_stats(&walltime_nsecs_stats
, t1
- t0
);
331 for (counter
= 0; counter
< nr_counters
; counter
++)
332 read_counter(counter
);
334 return WEXITSTATUS(status
);
337 static void print_noise(int counter
, double avg
)
342 fprintf(stderr
, " ( +- %7.3f%% )",
343 100 * stddev_stats(&event_res_stats
[counter
][0]) / avg
);
346 static void nsec_printout(int counter
, double avg
)
348 double msecs
= avg
/ 1e6
;
350 fprintf(stderr
, " %14.6f %-24s", msecs
, event_name(counter
));
352 if (MATCH_EVENT(SOFTWARE
, SW_TASK_CLOCK
, counter
)) {
353 fprintf(stderr
, " # %10.3f CPUs ",
354 avg
/ avg_stats(&walltime_nsecs_stats
));
358 static void abs_printout(int counter
, double avg
)
360 double total
, ratio
= 0.0;
362 fprintf(stderr
, " %14.0f %-24s", avg
, event_name(counter
));
364 if (MATCH_EVENT(HARDWARE
, HW_INSTRUCTIONS
, counter
)) {
365 total
= avg_stats(&runtime_cycles_stats
);
370 fprintf(stderr
, " # %10.3f IPC ", ratio
);
371 } else if (MATCH_EVENT(HARDWARE
, HW_BRANCH_MISSES
, counter
) &&
372 runtime_branches_stats
.n
!= 0) {
373 total
= avg_stats(&runtime_branches_stats
);
376 ratio
= avg
* 100 / total
;
378 fprintf(stderr
, " # %10.3f %% ", ratio
);
380 } else if (runtime_nsecs_stats
.n
!= 0) {
381 total
= avg_stats(&runtime_nsecs_stats
);
384 ratio
= 1000.0 * avg
/ total
;
386 fprintf(stderr
, " # %10.3f M/sec", ratio
);
391 * Print out the results of a single counter:
393 static void print_counter(int counter
)
395 double avg
= avg_stats(&event_res_stats
[counter
][0]);
396 int scaled
= event_scaled
[counter
];
399 fprintf(stderr
, " %14s %-24s\n",
400 "<not counted>", event_name(counter
));
404 if (nsec_counter(counter
))
405 nsec_printout(counter
, avg
);
407 abs_printout(counter
, avg
);
409 print_noise(counter
, avg
);
412 double avg_enabled
, avg_running
;
414 avg_enabled
= avg_stats(&event_res_stats
[counter
][1]);
415 avg_running
= avg_stats(&event_res_stats
[counter
][2]);
417 fprintf(stderr
, " (scaled from %.2f%%)",
418 100 * avg_running
/ avg_enabled
);
421 fprintf(stderr
, "\n");
424 static void print_stat(int argc
, const char **argv
)
430 fprintf(stderr
, "\n");
431 fprintf(stderr
, " Performance counter stats for ");
432 if(target_pid
== -1) {
433 fprintf(stderr
, "\'%s", argv
[0]);
434 for (i
= 1; i
< argc
; i
++)
435 fprintf(stderr
, " %s", argv
[i
]);
437 fprintf(stderr
, "task pid \'%d", target_pid
);
439 fprintf(stderr
, "\'");
441 fprintf(stderr
, " (%d runs)", run_count
);
442 fprintf(stderr
, ":\n\n");
444 for (counter
= 0; counter
< nr_counters
; counter
++)
445 print_counter(counter
);
447 fprintf(stderr
, "\n");
448 fprintf(stderr
, " %14.9f seconds time elapsed",
449 avg_stats(&walltime_nsecs_stats
)/1e9
);
451 fprintf(stderr
, " ( +- %7.3f%% )",
452 100*stddev_stats(&walltime_nsecs_stats
) /
453 avg_stats(&walltime_nsecs_stats
));
455 fprintf(stderr
, "\n\n");
458 static volatile int signr
= -1;
460 static void skip_signal(int signo
)
468 static void sig_atexit(void)
471 kill(child_pid
, SIGTERM
);
476 signal(signr
, SIG_DFL
);
477 kill(getpid(), signr
);
480 static const char * const stat_usage
[] = {
481 "perf stat [<options>] [<command>]",
485 static const struct option options
[] = {
486 OPT_CALLBACK('e', "event", NULL
, "event",
487 "event selector. use 'perf list' to list available events",
489 OPT_BOOLEAN('i', "inherit", &inherit
,
490 "child tasks inherit counters"),
491 OPT_INTEGER('p', "pid", &target_pid
,
492 "stat events on existing pid"),
493 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
494 "system-wide collection from all CPUs"),
495 OPT_BOOLEAN('c', "scale", &scale
,
496 "scale/normalize counters"),
497 OPT_BOOLEAN('v', "verbose", &verbose
,
498 "be more verbose (show counter open errors, etc)"),
499 OPT_INTEGER('r', "repeat", &run_count
,
500 "repeat command and print average + stddev (max: 100)"),
501 OPT_BOOLEAN('n', "null", &null_run
,
502 "null run - dont start any counters"),
506 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
510 argc
= parse_options(argc
, argv
, options
, stat_usage
,
511 PARSE_OPT_STOP_AT_NON_OPTION
);
512 if (!argc
&& target_pid
== -1)
513 usage_with_options(stat_usage
, options
);
515 usage_with_options(stat_usage
, options
);
517 /* Set attrs and nr_counters if no event is selected and !null_run */
518 if (!null_run
&& !nr_counters
) {
519 memcpy(attrs
, default_attrs
, sizeof(default_attrs
));
520 nr_counters
= ARRAY_SIZE(default_attrs
);
524 nr_cpus
= read_cpu_map();
529 * We dont want to block the signals - that would cause
530 * child tasks to inherit that and Ctrl-C would not work.
531 * What we want is for Ctrl-C to work in the exec()-ed
532 * task, but being ignored by perf stat itself:
535 signal(SIGINT
, skip_signal
);
536 signal(SIGALRM
, skip_signal
);
537 signal(SIGABRT
, skip_signal
);
540 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
541 if (run_count
!= 1 && verbose
)
542 fprintf(stderr
, "[ perf stat: executing run #%d ... ]\n", run_idx
+ 1);
543 status
= run_perf_stat(argc
, argv
);
546 print_stat(argc
, argv
);