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/evlist.h"
47 #include "util/evsel.h"
48 #include "util/debug.h"
49 #include "util/header.h"
50 #include "util/cpumap.h"
51 #include "util/thread.h"
52 #include "util/thread_map.h"
54 #include <sys/prctl.h>
58 #define DEFAULT_SEPARATOR " "
60 static struct perf_event_attr default_attrs
[] = {
62 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_TASK_CLOCK
},
63 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CONTEXT_SWITCHES
},
64 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_CPU_MIGRATIONS
},
65 { .type
= PERF_TYPE_SOFTWARE
, .config
= PERF_COUNT_SW_PAGE_FAULTS
},
67 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CPU_CYCLES
},
68 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_INSTRUCTIONS
},
69 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_INSTRUCTIONS
},
70 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_BRANCH_MISSES
},
71 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_REFERENCES
},
72 { .type
= PERF_TYPE_HARDWARE
, .config
= PERF_COUNT_HW_CACHE_MISSES
},
76 struct perf_evlist
*evsel_list
;
78 static bool system_wide
= false;
79 static struct cpu_map
*cpus
;
80 static int run_idx
= 0;
82 static int run_count
= 1;
83 static bool no_inherit
= false;
84 static bool scale
= true;
85 static bool no_aggr
= false;
86 static pid_t target_pid
= -1;
87 static pid_t target_tid
= -1;
88 static struct thread_map
*threads
;
89 static pid_t child_pid
= -1;
90 static bool null_run
= false;
91 static bool big_num
= true;
92 static int big_num_opt
= -1;
93 static const char *cpu_list
;
94 static const char *csv_sep
= NULL
;
95 static bool csv_output
= false;
97 static volatile int done
= 0;
105 struct stats res_stats
[3];
108 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
110 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
111 return evsel
->priv
== NULL
? -ENOMEM
: 0;
114 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
120 static void update_stats(struct stats
*stats
, u64 val
)
125 delta
= val
- stats
->mean
;
126 stats
->mean
+= delta
/ stats
->n
;
127 stats
->M2
+= delta
*(val
- stats
->mean
);
130 static double avg_stats(struct stats
*stats
)
136 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
138 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
139 * s^2 = -------------------------------
142 * http://en.wikipedia.org/wiki/Stddev
144 * The std dev of the mean is related to the std dev by:
151 static double stddev_stats(struct stats
*stats
)
153 double variance
= stats
->M2
/ (stats
->n
- 1);
154 double variance_mean
= variance
/ stats
->n
;
156 return sqrt(variance_mean
);
159 struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
160 struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
161 struct stats runtime_branches_stats
[MAX_NR_CPUS
];
162 struct stats walltime_nsecs_stats
;
164 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
166 struct perf_event_attr
*attr
= &evsel
->attr
;
169 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
170 PERF_FORMAT_TOTAL_TIME_RUNNING
;
173 return perf_evsel__open_per_cpu(evsel
, cpus
, false, false);
175 attr
->inherit
= !no_inherit
;
176 if (target_pid
== -1 && target_tid
== -1) {
178 attr
->enable_on_exec
= 1;
181 return perf_evsel__open_per_thread(evsel
, threads
, false, false);
185 * Does the counter have nsecs as a unit?
187 static inline int nsec_counter(struct perf_evsel
*evsel
)
189 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
190 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
197 * Read out the results of a single counter:
198 * aggregate counts across CPUs in system-wide mode
200 static int read_counter_aggr(struct perf_evsel
*counter
)
202 struct perf_stat
*ps
= counter
->priv
;
203 u64
*count
= counter
->counts
->aggr
.values
;
206 if (__perf_evsel__read(counter
, cpus
->nr
, threads
->nr
, scale
) < 0)
209 for (i
= 0; i
< 3; i
++)
210 update_stats(&ps
->res_stats
[i
], count
[i
]);
213 fprintf(stderr
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
214 event_name(counter
), count
[0], count
[1], count
[2]);
218 * Save the full runtime - to allow normalization during printout:
220 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
221 update_stats(&runtime_nsecs_stats
[0], count
[0]);
222 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
223 update_stats(&runtime_cycles_stats
[0], count
[0]);
224 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
225 update_stats(&runtime_branches_stats
[0], count
[0]);
231 * Read out the results of a single counter:
232 * do not aggregate counts across CPUs in system-wide mode
234 static int read_counter(struct perf_evsel
*counter
)
239 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
240 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
243 count
= counter
->counts
->cpu
[cpu
].values
;
245 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
246 update_stats(&runtime_nsecs_stats
[cpu
], count
[0]);
247 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
248 update_stats(&runtime_cycles_stats
[cpu
], count
[0]);
249 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
250 update_stats(&runtime_branches_stats
[cpu
], count
[0]);
256 static int run_perf_stat(int argc __used
, const char **argv
)
258 unsigned long long t0
, t1
;
259 struct perf_evsel
*counter
;
261 int child_ready_pipe
[2], go_pipe
[2];
262 const bool forks
= (argc
> 0);
265 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
266 perror("failed to create pipes");
271 if ((child_pid
= fork()) < 0)
272 perror("failed to fork");
275 close(child_ready_pipe
[0]);
277 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
280 * Do a dummy execvp to get the PLT entry resolved,
281 * so we avoid the resolver overhead on the real
284 execvp("", (char **)argv
);
287 * Tell the parent we're ready to go
289 close(child_ready_pipe
[1]);
292 * Wait until the parent tells us to go.
294 if (read(go_pipe
[0], &buf
, 1) == -1)
295 perror("unable to read pipe");
297 execvp(argv
[0], (char **)argv
);
303 if (target_tid
== -1 && target_pid
== -1 && !system_wide
)
304 threads
->map
[0] = child_pid
;
307 * Wait for the child to be ready to exec.
309 close(child_ready_pipe
[1]);
311 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
312 perror("unable to read pipe");
313 close(child_ready_pipe
[0]);
316 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
317 if (create_perf_stat_counter(counter
) < 0) {
318 if (errno
== -EPERM
|| errno
== -EACCES
) {
319 error("You may not have permission to collect %sstats.\n"
320 "\t Consider tweaking"
321 " /proc/sys/kernel/perf_event_paranoid or running as root.",
322 system_wide
? "system-wide " : "");
323 } else if (errno
== ENOENT
) {
324 error("%s event is not supported. ", event_name(counter
));
326 error("open_counter returned with %d (%s). "
327 "/bin/dmesg may provide additional information.\n",
328 errno
, strerror(errno
));
331 kill(child_pid
, SIGTERM
);
332 die("Not all events could be opened.\n");
338 * Enable counters and exec the command:
346 while(!done
) sleep(1);
351 update_stats(&walltime_nsecs_stats
, t1
- t0
);
354 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
355 read_counter(counter
);
356 perf_evsel__close_fd(counter
, cpus
->nr
, 1);
359 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
360 read_counter_aggr(counter
);
361 perf_evsel__close_fd(counter
, cpus
->nr
, threads
->nr
);
365 return WEXITSTATUS(status
);
368 static void print_noise(struct perf_evsel
*evsel
, double avg
)
370 struct perf_stat
*ps
;
376 fprintf(stderr
, " ( +- %7.3f%% )",
377 100 * stddev_stats(&ps
->res_stats
[0]) / avg
);
380 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
382 double msecs
= avg
/ 1e6
;
383 char cpustr
[16] = { '\0', };
384 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
387 sprintf(cpustr
, "CPU%*d%s",
389 cpus
->map
[cpu
], csv_sep
);
391 fprintf(stderr
, fmt
, cpustr
, msecs
, csv_sep
, event_name(evsel
));
396 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
397 fprintf(stderr
, " # %10.3f CPUs ",
398 avg
/ avg_stats(&walltime_nsecs_stats
));
401 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
403 double total
, ratio
= 0.0;
404 char cpustr
[16] = { '\0', };
410 fmt
= "%s%'18.0f%s%-24s";
412 fmt
= "%s%18.0f%s%-24s";
415 sprintf(cpustr
, "CPU%*d%s",
417 cpus
->map
[cpu
], csv_sep
);
421 fprintf(stderr
, fmt
, cpustr
, avg
, csv_sep
, event_name(evsel
));
426 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
427 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
432 fprintf(stderr
, " # %10.3f IPC ", ratio
);
433 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
434 runtime_branches_stats
[cpu
].n
!= 0) {
435 total
= avg_stats(&runtime_branches_stats
[cpu
]);
438 ratio
= avg
* 100 / total
;
440 fprintf(stderr
, " # %10.3f %% ", ratio
);
442 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
443 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
446 ratio
= 1000.0 * avg
/ total
;
448 fprintf(stderr
, " # %10.3f M/sec", ratio
);
453 * Print out the results of a single counter:
454 * aggregated counts in system-wide mode
456 static void print_counter_aggr(struct perf_evsel
*counter
)
458 struct perf_stat
*ps
= counter
->priv
;
459 double avg
= avg_stats(&ps
->res_stats
[0]);
460 int scaled
= counter
->counts
->scaled
;
463 fprintf(stderr
, "%*s%s%-24s\n",
465 "<not counted>", csv_sep
, event_name(counter
));
469 if (nsec_counter(counter
))
470 nsec_printout(-1, counter
, avg
);
472 abs_printout(-1, counter
, avg
);
479 print_noise(counter
, avg
);
482 double avg_enabled
, avg_running
;
484 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
485 avg_running
= avg_stats(&ps
->res_stats
[2]);
487 fprintf(stderr
, " (scaled from %.2f%%)",
488 100 * avg_running
/ avg_enabled
);
491 fprintf(stderr
, "\n");
495 * Print out the results of a single counter:
496 * does not use aggregated count in system-wide
498 static void print_counter(struct perf_evsel
*counter
)
503 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
504 val
= counter
->counts
->cpu
[cpu
].val
;
505 ena
= counter
->counts
->cpu
[cpu
].ena
;
506 run
= counter
->counts
->cpu
[cpu
].run
;
507 if (run
== 0 || ena
== 0) {
508 fprintf(stderr
, "CPU%*d%s%*s%s%-24s",
510 cpus
->map
[cpu
], csv_sep
,
512 "<not counted>", csv_sep
,
513 event_name(counter
));
515 fprintf(stderr
, "\n");
519 if (nsec_counter(counter
))
520 nsec_printout(cpu
, counter
, val
);
522 abs_printout(cpu
, counter
, val
);
525 print_noise(counter
, 1.0);
528 fprintf(stderr
, " (scaled from %.2f%%)",
532 fprintf(stderr
, "\n");
536 static void print_stat(int argc
, const char **argv
)
538 struct perf_evsel
*counter
;
544 fprintf(stderr
, "\n");
545 fprintf(stderr
, " Performance counter stats for ");
546 if(target_pid
== -1 && target_tid
== -1) {
547 fprintf(stderr
, "\'%s", argv
[0]);
548 for (i
= 1; i
< argc
; i
++)
549 fprintf(stderr
, " %s", argv
[i
]);
550 } else if (target_pid
!= -1)
551 fprintf(stderr
, "process id \'%d", target_pid
);
553 fprintf(stderr
, "thread id \'%d", target_tid
);
555 fprintf(stderr
, "\'");
557 fprintf(stderr
, " (%d runs)", run_count
);
558 fprintf(stderr
, ":\n\n");
562 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
563 print_counter(counter
);
565 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
566 print_counter_aggr(counter
);
570 fprintf(stderr
, "\n");
571 fprintf(stderr
, " %18.9f seconds time elapsed",
572 avg_stats(&walltime_nsecs_stats
)/1e9
);
574 fprintf(stderr
, " ( +- %7.3f%% )",
575 100*stddev_stats(&walltime_nsecs_stats
) /
576 avg_stats(&walltime_nsecs_stats
));
578 fprintf(stderr
, "\n\n");
582 static volatile int signr
= -1;
584 static void skip_signal(int signo
)
592 static void sig_atexit(void)
595 kill(child_pid
, SIGTERM
);
600 signal(signr
, SIG_DFL
);
601 kill(getpid(), signr
);
604 static const char * const stat_usage
[] = {
605 "perf stat [<options>] [<command>]",
609 static int stat__set_big_num(const struct option
*opt __used
,
610 const char *s __used
, int unset
)
612 big_num_opt
= unset
? 0 : 1;
616 static const struct option options
[] = {
617 OPT_CALLBACK('e', "event", &evsel_list
, "event",
618 "event selector. use 'perf list' to list available events",
620 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
621 "child tasks do not inherit counters"),
622 OPT_INTEGER('p', "pid", &target_pid
,
623 "stat events on existing process id"),
624 OPT_INTEGER('t', "tid", &target_tid
,
625 "stat events on existing thread id"),
626 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
627 "system-wide collection from all CPUs"),
628 OPT_BOOLEAN('c', "scale", &scale
,
629 "scale/normalize counters"),
630 OPT_INCR('v', "verbose", &verbose
,
631 "be more verbose (show counter open errors, etc)"),
632 OPT_INTEGER('r', "repeat", &run_count
,
633 "repeat command and print average + stddev (max: 100)"),
634 OPT_BOOLEAN('n', "null", &null_run
,
635 "null run - dont start any counters"),
636 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
637 "print large numbers with thousands\' separators",
639 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
640 "list of cpus to monitor in system-wide"),
641 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
642 "disable CPU count aggregation"),
643 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
644 "print counts with custom separator"),
648 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
650 struct perf_evsel
*pos
;
651 int status
= -ENOMEM
;
653 setlocale(LC_ALL
, "");
655 evsel_list
= perf_evlist__new();
656 if (evsel_list
== NULL
)
659 argc
= parse_options(argc
, argv
, options
, stat_usage
,
660 PARSE_OPT_STOP_AT_NON_OPTION
);
665 csv_sep
= DEFAULT_SEPARATOR
;
668 * let the spreadsheet do the pretty-printing
671 /* User explicitely passed -B? */
672 if (big_num_opt
== 1) {
673 fprintf(stderr
, "-B option not supported with -x\n");
674 usage_with_options(stat_usage
, options
);
675 } else /* Nope, so disable big number formatting */
677 } else if (big_num_opt
== 0) /* User passed --no-big-num */
680 if (!argc
&& target_pid
== -1 && target_tid
== -1)
681 usage_with_options(stat_usage
, options
);
683 usage_with_options(stat_usage
, options
);
685 /* no_aggr is for system-wide only */
686 if (no_aggr
&& !system_wide
)
687 usage_with_options(stat_usage
, options
);
689 /* Set attrs and nr_counters if no event is selected and !null_run */
690 if (!null_run
&& !evsel_list
->nr_entries
) {
693 for (c
= 0; c
< ARRAY_SIZE(default_attrs
); ++c
) {
694 pos
= perf_evsel__new(&default_attrs
[c
], c
);
697 perf_evlist__add(evsel_list
, pos
);
701 if (target_pid
!= -1)
702 target_tid
= target_pid
;
704 threads
= thread_map__new(target_pid
, target_tid
);
705 if (threads
== NULL
) {
706 pr_err("Problems finding threads of monitor\n");
707 usage_with_options(stat_usage
, options
);
711 cpus
= cpu_map__new(cpu_list
);
713 cpus
= cpu_map__dummy_new();
716 perror("failed to parse CPUs map");
717 usage_with_options(stat_usage
, options
);
721 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
722 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
723 perf_evsel__alloc_counts(pos
, cpus
->nr
) < 0 ||
724 perf_evsel__alloc_fd(pos
, cpus
->nr
, threads
->nr
) < 0)
729 * We dont want to block the signals - that would cause
730 * child tasks to inherit that and Ctrl-C would not work.
731 * What we want is for Ctrl-C to work in the exec()-ed
732 * task, but being ignored by perf stat itself:
735 signal(SIGINT
, skip_signal
);
736 signal(SIGALRM
, skip_signal
);
737 signal(SIGABRT
, skip_signal
);
740 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
741 if (run_count
!= 1 && verbose
)
742 fprintf(stderr
, "[ perf stat: executing run #%d ... ]\n", run_idx
+ 1);
743 status
= run_perf_stat(argc
, argv
);
747 print_stat(argc
, argv
);
749 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
750 perf_evsel__free_stat_priv(pos
);
751 perf_evlist__delete(evsel_list
);
753 thread_map__delete(threads
);