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/evsel.h"
47 #include "util/debug.h"
48 #include "util/header.h"
49 #include "util/cpumap.h"
50 #include "util/thread.h"
52 #include <sys/prctl.h>
56 #define DEFAULT_SEPARATOR " "
58 static struct perf_event_attr default_attrs
[] = {
60 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
61 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
62 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
63 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
65 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
66 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
67 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
68 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
69 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_REFERENCES
},
70 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_MISSES
},
74 static bool system_wide
= false;
75 static struct cpu_map
*cpus
;
76 static int run_idx
= 0;
78 static int run_count
= 1;
79 static bool no_inherit
= false;
80 static bool scale
= true;
81 static bool no_aggr
= false;
82 static pid_t target_pid
= -1;
83 static pid_t target_tid
= -1;
84 static pid_t
*all_tids
= NULL
;
85 static int thread_num
= 0;
86 static pid_t child_pid
= -1;
87 static bool null_run
= false;
88 static bool big_num
= true;
89 static int big_num_opt
= -1;
90 static const char *cpu_list
;
91 static const char *csv_sep
= NULL
;
92 static bool csv_output
= false;
94 static volatile int done
= 0;
102 struct stats res_stats
[3];
105 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
107 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
108 return evsel
->priv
== NULL
? -ENOMEM
: 0;
111 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
117 static void update_stats(struct stats
*stats
, u64 val
)
122 delta
= val
- stats
->mean
;
123 stats
->mean
+= delta
/ stats
->n
;
124 stats
->M2
+= delta
*(val
- stats
->mean
);
127 static double avg_stats(struct stats
*stats
)
133 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
135 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
136 * s^2 = -------------------------------
139 * http://en.wikipedia.org/wiki/Stddev
141 * The std dev of the mean is related to the std dev by:
148 static double stddev_stats(struct stats
*stats
)
150 double variance
= stats
->M2
/ (stats
->n
- 1);
151 double variance_mean
= variance
/ stats
->n
;
153 return sqrt(variance_mean
);
156 struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
157 struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
158 struct stats runtime_branches_stats
[MAX_NR_CPUS
];
159 struct stats walltime_nsecs_stats
;
161 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
163 struct perf_event_attr
*attr
= &evsel
->attr
;
166 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
167 PERF_FORMAT_TOTAL_TIME_RUNNING
;
170 return perf_evsel__open_per_cpu(evsel
, cpus
->nr
, cpus
->map
);
172 attr
->inherit
= !no_inherit
;
173 if (target_pid
== -1 && target_tid
== -1) {
175 attr
->enable_on_exec
= 1;
178 return perf_evsel__open_per_thread(evsel
, thread_num
, all_tids
);
182 * Does the counter have nsecs as a unit?
184 static inline int nsec_counter(struct perf_evsel
*evsel
)
186 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
187 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
194 * Read out the results of a single counter:
195 * aggregate counts across CPUs in system-wide mode
197 static int read_counter_aggr(struct perf_evsel
*counter
)
199 struct perf_stat
*ps
= counter
->priv
;
200 u64
*count
= counter
->counts
->aggr
.values
;
203 if (__perf_evsel__read(counter
, cpus
->nr
, thread_num
, scale
) < 0)
206 for (i
= 0; i
< 3; i
++)
207 update_stats(&ps
->res_stats
[i
], count
[i
]);
210 fprintf(stderr
, "%s: %Ld %Ld %Ld\n", event_name(counter
),
211 count
[0], count
[1], count
[2]);
215 * Save the full runtime - to allow normalization during printout:
217 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
218 update_stats(&runtime_nsecs_stats
[0], count
[0]);
219 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
220 update_stats(&runtime_cycles_stats
[0], count
[0]);
221 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
222 update_stats(&runtime_branches_stats
[0], count
[0]);
228 * Read out the results of a single counter:
229 * do not aggregate counts across CPUs in system-wide mode
231 static int read_counter(struct perf_evsel
*counter
)
236 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
237 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
240 count
= counter
->counts
->cpu
[cpu
].values
;
242 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
243 update_stats(&runtime_nsecs_stats
[cpu
], count
[0]);
244 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
245 update_stats(&runtime_cycles_stats
[cpu
], count
[0]);
246 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
247 update_stats(&runtime_branches_stats
[cpu
], count
[0]);
253 static int run_perf_stat(int argc __used
, const char **argv
)
255 unsigned long long t0
, t1
;
256 struct perf_evsel
*counter
;
258 int child_ready_pipe
[2], go_pipe
[2];
259 const bool forks
= (argc
> 0);
262 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
263 perror("failed to create pipes");
268 if ((child_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
);
300 if (target_tid
== -1 && target_pid
== -1 && !system_wide
)
301 all_tids
[0] = child_pid
;
304 * Wait for the child to be ready to exec.
306 close(child_ready_pipe
[1]);
308 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
309 perror("unable to read pipe");
310 close(child_ready_pipe
[0]);
313 list_for_each_entry(counter
, &evsel_list
, node
) {
314 if (create_perf_stat_counter(counter
) < 0) {
315 if (errno
== -EPERM
|| errno
== -EACCES
) {
316 error("You may not have permission to collect %sstats.\n"
317 "\t Consider tweaking"
318 " /proc/sys/kernel/perf_event_paranoid or running as root.",
319 system_wide
? "system-wide " : "");
321 error("open_counter returned with %d (%s). "
322 "/bin/dmesg may provide additional information.\n",
323 errno
, strerror(errno
));
326 kill(child_pid
, SIGTERM
);
327 die("Not all events could be opened.\n");
333 * Enable counters and exec the command:
341 while(!done
) sleep(1);
346 update_stats(&walltime_nsecs_stats
, t1
- t0
);
349 list_for_each_entry(counter
, &evsel_list
, node
) {
350 read_counter(counter
);
351 perf_evsel__close_fd(counter
, cpus
->nr
, 1);
354 list_for_each_entry(counter
, &evsel_list
, node
) {
355 read_counter_aggr(counter
);
356 perf_evsel__close_fd(counter
, cpus
->nr
, thread_num
);
360 return WEXITSTATUS(status
);
363 static void print_noise(struct perf_evsel
*evsel
, double avg
)
365 struct perf_stat
*ps
;
371 fprintf(stderr
, " ( +- %7.3f%% )",
372 100 * stddev_stats(&ps
->res_stats
[0]) / avg
);
375 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
377 double msecs
= avg
/ 1e6
;
378 char cpustr
[16] = { '\0', };
379 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
382 sprintf(cpustr
, "CPU%*d%s",
384 cpus
->map
[cpu
], csv_sep
);
386 fprintf(stderr
, fmt
, cpustr
, msecs
, csv_sep
, event_name(evsel
));
391 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
392 fprintf(stderr
, " # %10.3f CPUs ",
393 avg
/ avg_stats(&walltime_nsecs_stats
));
396 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
398 double total
, ratio
= 0.0;
399 char cpustr
[16] = { '\0', };
405 fmt
= "%s%'18.0f%s%-24s";
407 fmt
= "%s%18.0f%s%-24s";
410 sprintf(cpustr
, "CPU%*d%s",
412 cpus
->map
[cpu
], csv_sep
);
416 fprintf(stderr
, fmt
, cpustr
, avg
, csv_sep
, event_name(evsel
));
421 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
422 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
427 fprintf(stderr
, " # %10.3f IPC ", ratio
);
428 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
429 runtime_branches_stats
[cpu
].n
!= 0) {
430 total
= avg_stats(&runtime_branches_stats
[cpu
]);
433 ratio
= avg
* 100 / total
;
435 fprintf(stderr
, " # %10.3f %% ", ratio
);
437 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
438 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
441 ratio
= 1000.0 * avg
/ total
;
443 fprintf(stderr
, " # %10.3f M/sec", ratio
);
448 * Print out the results of a single counter:
449 * aggregated counts in system-wide mode
451 static void print_counter_aggr(struct perf_evsel
*counter
)
453 struct perf_stat
*ps
= counter
->priv
;
454 double avg
= avg_stats(&ps
->res_stats
[0]);
455 int scaled
= counter
->counts
->scaled
;
458 fprintf(stderr
, "%*s%s%-24s\n",
460 "<not counted>", csv_sep
, event_name(counter
));
464 if (nsec_counter(counter
))
465 nsec_printout(-1, counter
, avg
);
467 abs_printout(-1, counter
, avg
);
474 print_noise(counter
, avg
);
477 double avg_enabled
, avg_running
;
479 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
480 avg_running
= avg_stats(&ps
->res_stats
[2]);
482 fprintf(stderr
, " (scaled from %.2f%%)",
483 100 * avg_running
/ avg_enabled
);
486 fprintf(stderr
, "\n");
490 * Print out the results of a single counter:
491 * does not use aggregated count in system-wide
493 static void print_counter(struct perf_evsel
*counter
)
498 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
499 val
= counter
->counts
->cpu
[cpu
].val
;
500 ena
= counter
->counts
->cpu
[cpu
].ena
;
501 run
= counter
->counts
->cpu
[cpu
].run
;
502 if (run
== 0 || ena
== 0) {
503 fprintf(stderr
, "CPU%*d%s%*s%s%-24s",
505 cpus
->map
[cpu
], csv_sep
,
507 "<not counted>", csv_sep
,
508 event_name(counter
));
510 fprintf(stderr
, "\n");
514 if (nsec_counter(counter
))
515 nsec_printout(cpu
, counter
, val
);
517 abs_printout(cpu
, counter
, val
);
520 print_noise(counter
, 1.0);
523 fprintf(stderr
, " (scaled from %.2f%%)",
527 fprintf(stderr
, "\n");
531 static void print_stat(int argc
, const char **argv
)
533 struct perf_evsel
*counter
;
539 fprintf(stderr
, "\n");
540 fprintf(stderr
, " Performance counter stats for ");
541 if(target_pid
== -1 && target_tid
== -1) {
542 fprintf(stderr
, "\'%s", argv
[0]);
543 for (i
= 1; i
< argc
; i
++)
544 fprintf(stderr
, " %s", argv
[i
]);
545 } else if (target_pid
!= -1)
546 fprintf(stderr
, "process id \'%d", target_pid
);
548 fprintf(stderr
, "thread id \'%d", target_tid
);
550 fprintf(stderr
, "\'");
552 fprintf(stderr
, " (%d runs)", run_count
);
553 fprintf(stderr
, ":\n\n");
557 list_for_each_entry(counter
, &evsel_list
, node
)
558 print_counter(counter
);
560 list_for_each_entry(counter
, &evsel_list
, node
)
561 print_counter_aggr(counter
);
565 fprintf(stderr
, "\n");
566 fprintf(stderr
, " %18.9f seconds time elapsed",
567 avg_stats(&walltime_nsecs_stats
)/1e9
);
569 fprintf(stderr
, " ( +- %7.3f%% )",
570 100*stddev_stats(&walltime_nsecs_stats
) /
571 avg_stats(&walltime_nsecs_stats
));
573 fprintf(stderr
, "\n\n");
577 static volatile int signr
= -1;
579 static void skip_signal(int signo
)
587 static void sig_atexit(void)
590 kill(child_pid
, SIGTERM
);
595 signal(signr
, SIG_DFL
);
596 kill(getpid(), signr
);
599 static const char * const stat_usage
[] = {
600 "perf stat [<options>] [<command>]",
604 static int stat__set_big_num(const struct option
*opt __used
,
605 const char *s __used
, int unset
)
607 big_num_opt
= unset
? 0 : 1;
611 static const struct option options
[] = {
612 OPT_CALLBACK('e', "event", NULL
, "event",
613 "event selector. use 'perf list' to list available events",
615 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
616 "child tasks do not inherit counters"),
617 OPT_INTEGER('p', "pid", &target_pid
,
618 "stat events on existing process id"),
619 OPT_INTEGER('t', "tid", &target_tid
,
620 "stat events on existing thread id"),
621 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
622 "system-wide collection from all CPUs"),
623 OPT_BOOLEAN('c', "scale", &scale
,
624 "scale/normalize counters"),
625 OPT_INCR('v', "verbose", &verbose
,
626 "be more verbose (show counter open errors, etc)"),
627 OPT_INTEGER('r', "repeat", &run_count
,
628 "repeat command and print average + stddev (max: 100)"),
629 OPT_BOOLEAN('n', "null", &null_run
,
630 "null run - dont start any counters"),
631 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
632 "print large numbers with thousands\' separators",
634 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
635 "list of cpus to monitor in system-wide"),
636 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
637 "disable CPU count aggregation"),
638 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
639 "print counts with custom separator"),
643 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
645 struct perf_evsel
*pos
;
646 int status
= -ENOMEM
;
648 setlocale(LC_ALL
, "");
650 argc
= parse_options(argc
, argv
, options
, stat_usage
,
651 PARSE_OPT_STOP_AT_NON_OPTION
);
656 csv_sep
= DEFAULT_SEPARATOR
;
659 * let the spreadsheet do the pretty-printing
662 /* User explicitely passed -B? */
663 if (big_num_opt
== 1) {
664 fprintf(stderr
, "-B option not supported with -x\n");
665 usage_with_options(stat_usage
, options
);
666 } else /* Nope, so disable big number formatting */
668 } else if (big_num_opt
== 0) /* User passed --no-big-num */
671 if (!argc
&& target_pid
== -1 && target_tid
== -1)
672 usage_with_options(stat_usage
, options
);
674 usage_with_options(stat_usage
, options
);
676 /* no_aggr is for system-wide only */
677 if (no_aggr
&& !system_wide
)
678 usage_with_options(stat_usage
, options
);
680 /* Set attrs and nr_counters if no event is selected and !null_run */
681 if (!null_run
&& !nr_counters
) {
684 nr_counters
= ARRAY_SIZE(default_attrs
);
686 for (c
= 0; c
< ARRAY_SIZE(default_attrs
); ++c
) {
687 pos
= perf_evsel__new(default_attrs
[c
].type
,
688 default_attrs
[c
].config
,
692 list_add(&pos
->node
, &evsel_list
);
697 cpus
= cpu_map__new(cpu_list
);
699 cpus
= cpu_map__dummy_new();
702 perror("failed to parse CPUs map");
703 usage_with_options(stat_usage
, options
);
707 if (target_pid
!= -1) {
708 target_tid
= target_pid
;
709 thread_num
= find_all_tid(target_pid
, &all_tids
);
710 if (thread_num
<= 0) {
711 fprintf(stderr
, "Can't find all threads of pid %d\n",
713 usage_with_options(stat_usage
, options
);
716 all_tids
=malloc(sizeof(pid_t
));
720 all_tids
[0] = target_tid
;
724 list_for_each_entry(pos
, &evsel_list
, node
) {
725 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
726 perf_evsel__alloc_counts(pos
, cpus
->nr
) < 0 ||
727 perf_evsel__alloc_fd(pos
, cpus
->nr
, thread_num
) < 0)
732 * We dont want to block the signals - that would cause
733 * child tasks to inherit that and Ctrl-C would not work.
734 * What we want is for Ctrl-C to work in the exec()-ed
735 * task, but being ignored by perf stat itself:
738 signal(SIGINT
, skip_signal
);
739 signal(SIGALRM
, skip_signal
);
740 signal(SIGABRT
, skip_signal
);
743 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
744 if (run_count
!= 1 && verbose
)
745 fprintf(stderr
, "[ perf stat: executing run #%d ... ]\n", run_idx
+ 1);
746 status
= run_perf_stat(argc
, argv
);
750 print_stat(argc
, argv
);
752 list_for_each_entry(pos
, &evsel_list
, node
)
753 perf_evsel__free_stat_priv(pos
);