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 int run_idx
= 0;
81 static int run_count
= 1;
82 static bool no_inherit
= false;
83 static bool scale
= true;
84 static bool no_aggr
= false;
85 static pid_t target_pid
= -1;
86 static pid_t target_tid
= -1;
87 static pid_t child_pid
= -1;
88 static bool null_run
= false;
89 static bool big_num
= true;
90 static int big_num_opt
= -1;
91 static const char *cpu_list
;
92 static const char *csv_sep
= NULL
;
93 static bool csv_output
= false;
95 static volatile int done
= 0;
103 struct stats res_stats
[3];
106 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
108 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
109 return evsel
->priv
== NULL
? -ENOMEM
: 0;
112 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
118 static void update_stats(struct stats
*stats
, u64 val
)
123 delta
= val
- stats
->mean
;
124 stats
->mean
+= delta
/ stats
->n
;
125 stats
->M2
+= delta
*(val
- stats
->mean
);
128 static double avg_stats(struct stats
*stats
)
134 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
136 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
137 * s^2 = -------------------------------
140 * http://en.wikipedia.org/wiki/Stddev
142 * The std dev of the mean is related to the std dev by:
149 static double stddev_stats(struct stats
*stats
)
151 double variance
= stats
->M2
/ (stats
->n
- 1);
152 double variance_mean
= variance
/ stats
->n
;
154 return sqrt(variance_mean
);
157 struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
158 struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
159 struct stats runtime_branches_stats
[MAX_NR_CPUS
];
160 struct stats walltime_nsecs_stats
;
162 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
164 struct perf_event_attr
*attr
= &evsel
->attr
;
167 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
168 PERF_FORMAT_TOTAL_TIME_RUNNING
;
171 return perf_evsel__open_per_cpu(evsel
, evsel_list
->cpus
, false, false);
173 attr
->inherit
= !no_inherit
;
174 if (target_pid
== -1 && target_tid
== -1) {
176 attr
->enable_on_exec
= 1;
179 return perf_evsel__open_per_thread(evsel
, evsel_list
->threads
, false, false);
183 * Does the counter have nsecs as a unit?
185 static inline int nsec_counter(struct perf_evsel
*evsel
)
187 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
188 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
195 * Read out the results of a single counter:
196 * aggregate counts across CPUs in system-wide mode
198 static int read_counter_aggr(struct perf_evsel
*counter
)
200 struct perf_stat
*ps
= counter
->priv
;
201 u64
*count
= counter
->counts
->aggr
.values
;
204 if (__perf_evsel__read(counter
, evsel_list
->cpus
->nr
,
205 evsel_list
->threads
->nr
, scale
) < 0)
208 for (i
= 0; i
< 3; i
++)
209 update_stats(&ps
->res_stats
[i
], count
[i
]);
212 fprintf(stderr
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
213 event_name(counter
), count
[0], count
[1], count
[2]);
217 * Save the full runtime - to allow normalization during printout:
219 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
220 update_stats(&runtime_nsecs_stats
[0], count
[0]);
221 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
222 update_stats(&runtime_cycles_stats
[0], count
[0]);
223 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
224 update_stats(&runtime_branches_stats
[0], count
[0]);
230 * Read out the results of a single counter:
231 * do not aggregate counts across CPUs in system-wide mode
233 static int read_counter(struct perf_evsel
*counter
)
238 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
239 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
242 count
= counter
->counts
->cpu
[cpu
].values
;
244 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
245 update_stats(&runtime_nsecs_stats
[cpu
], count
[0]);
246 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
247 update_stats(&runtime_cycles_stats
[cpu
], count
[0]);
248 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
249 update_stats(&runtime_branches_stats
[cpu
], count
[0]);
255 static int run_perf_stat(int argc __used
, const char **argv
)
257 unsigned long long t0
, t1
;
258 struct perf_evsel
*counter
;
260 int child_ready_pipe
[2], go_pipe
[2];
261 const bool forks
= (argc
> 0);
264 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
265 perror("failed to create pipes");
270 if ((child_pid
= fork()) < 0)
271 perror("failed to fork");
274 close(child_ready_pipe
[0]);
276 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
279 * Do a dummy execvp to get the PLT entry resolved,
280 * so we avoid the resolver overhead on the real
283 execvp("", (char **)argv
);
286 * Tell the parent we're ready to go
288 close(child_ready_pipe
[1]);
291 * Wait until the parent tells us to go.
293 if (read(go_pipe
[0], &buf
, 1) == -1)
294 perror("unable to read pipe");
296 execvp(argv
[0], (char **)argv
);
302 if (target_tid
== -1 && target_pid
== -1 && !system_wide
)
303 evsel_list
->threads
->map
[0] = child_pid
;
306 * Wait for the child to be ready to exec.
308 close(child_ready_pipe
[1]);
310 if (read(child_ready_pipe
[0], &buf
, 1) == -1)
311 perror("unable to read pipe");
312 close(child_ready_pipe
[0]);
315 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
316 if (create_perf_stat_counter(counter
) < 0) {
317 if (errno
== -EPERM
|| errno
== -EACCES
) {
318 error("You may not have permission to collect %sstats.\n"
319 "\t Consider tweaking"
320 " /proc/sys/kernel/perf_event_paranoid or running as root.",
321 system_wide
? "system-wide " : "");
322 } else if (errno
== ENOENT
) {
323 error("%s event is not supported. ", event_name(counter
));
325 error("open_counter returned with %d (%s). "
326 "/bin/dmesg may provide additional information.\n",
327 errno
, strerror(errno
));
330 kill(child_pid
, SIGTERM
);
331 die("Not all events could be opened.\n");
336 if (perf_evlist__set_filters(evsel_list
)) {
337 error("failed to set filter with %d (%s)\n", errno
,
343 * Enable counters and exec the command:
351 while(!done
) sleep(1);
356 update_stats(&walltime_nsecs_stats
, t1
- t0
);
359 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
360 read_counter(counter
);
361 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
, 1);
364 list_for_each_entry(counter
, &evsel_list
->entries
, node
) {
365 read_counter_aggr(counter
);
366 perf_evsel__close_fd(counter
, evsel_list
->cpus
->nr
,
367 evsel_list
->threads
->nr
);
371 return WEXITSTATUS(status
);
374 static void print_noise(struct perf_evsel
*evsel
, double avg
)
376 struct perf_stat
*ps
;
382 fprintf(stderr
, " ( +- %7.3f%% )",
383 100 * stddev_stats(&ps
->res_stats
[0]) / avg
);
386 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
388 double msecs
= avg
/ 1e6
;
389 char cpustr
[16] = { '\0', };
390 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
393 sprintf(cpustr
, "CPU%*d%s",
395 evsel_list
->cpus
->map
[cpu
], csv_sep
);
397 fprintf(stderr
, fmt
, cpustr
, msecs
, csv_sep
, event_name(evsel
));
400 fprintf(stderr
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
405 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
406 fprintf(stderr
, " # %10.3f CPUs ",
407 avg
/ avg_stats(&walltime_nsecs_stats
));
410 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
412 double total
, ratio
= 0.0;
413 char cpustr
[16] = { '\0', };
419 fmt
= "%s%'18.0f%s%-24s";
421 fmt
= "%s%18.0f%s%-24s";
424 sprintf(cpustr
, "CPU%*d%s",
426 evsel_list
->cpus
->map
[cpu
], csv_sep
);
430 fprintf(stderr
, fmt
, cpustr
, avg
, csv_sep
, event_name(evsel
));
433 fprintf(stderr
, "%s%s", csv_sep
, evsel
->cgrp
->name
);
438 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
439 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
444 fprintf(stderr
, " # %10.3f IPC ", ratio
);
445 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
446 runtime_branches_stats
[cpu
].n
!= 0) {
447 total
= avg_stats(&runtime_branches_stats
[cpu
]);
450 ratio
= avg
* 100 / total
;
452 fprintf(stderr
, " # %10.3f %% ", ratio
);
454 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
455 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
458 ratio
= 1000.0 * avg
/ total
;
460 fprintf(stderr
, " # %10.3f M/sec", ratio
);
465 * Print out the results of a single counter:
466 * aggregated counts in system-wide mode
468 static void print_counter_aggr(struct perf_evsel
*counter
)
470 struct perf_stat
*ps
= counter
->priv
;
471 double avg
= avg_stats(&ps
->res_stats
[0]);
472 int scaled
= counter
->counts
->scaled
;
475 fprintf(stderr
, "%*s%s%*s",
479 csv_output
? 0 : -24,
480 event_name(counter
));
483 fprintf(stderr
, "%s%s", csv_sep
, counter
->cgrp
->name
);
489 if (nsec_counter(counter
))
490 nsec_printout(-1, counter
, avg
);
492 abs_printout(-1, counter
, avg
);
499 print_noise(counter
, avg
);
502 double avg_enabled
, avg_running
;
504 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
505 avg_running
= avg_stats(&ps
->res_stats
[2]);
507 fprintf(stderr
, " (scaled from %.2f%%)",
508 100 * avg_running
/ avg_enabled
);
510 fprintf(stderr
, "\n");
514 * Print out the results of a single counter:
515 * does not use aggregated count in system-wide
517 static void print_counter(struct perf_evsel
*counter
)
522 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; cpu
++) {
523 val
= counter
->counts
->cpu
[cpu
].val
;
524 ena
= counter
->counts
->cpu
[cpu
].ena
;
525 run
= counter
->counts
->cpu
[cpu
].run
;
526 if (run
== 0 || ena
== 0) {
527 fprintf(stderr
, "CPU%*d%s%*s%s%*s",
529 evsel_list
->cpus
->map
[cpu
], csv_sep
,
531 "<not counted>", csv_sep
,
532 csv_output
? 0 : -24,
533 event_name(counter
));
536 fprintf(stderr
, "%s%s", csv_sep
, counter
->cgrp
->name
);
542 if (nsec_counter(counter
))
543 nsec_printout(cpu
, counter
, val
);
545 abs_printout(cpu
, counter
, val
);
548 print_noise(counter
, 1.0);
551 fprintf(stderr
, " (scaled from %.2f%%)",
559 static void print_stat(int argc
, const char **argv
)
561 struct perf_evsel
*counter
;
567 fprintf(stderr
, "\n");
568 fprintf(stderr
, " Performance counter stats for ");
569 if(target_pid
== -1 && target_tid
== -1) {
570 fprintf(stderr
, "\'%s", argv
[0]);
571 for (i
= 1; i
< argc
; i
++)
572 fprintf(stderr
, " %s", argv
[i
]);
573 } else if (target_pid
!= -1)
574 fprintf(stderr
, "process id \'%d", target_pid
);
576 fprintf(stderr
, "thread id \'%d", target_tid
);
578 fprintf(stderr
, "\'");
580 fprintf(stderr
, " (%d runs)", run_count
);
581 fprintf(stderr
, ":\n\n");
585 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
586 print_counter(counter
);
588 list_for_each_entry(counter
, &evsel_list
->entries
, node
)
589 print_counter_aggr(counter
);
593 fprintf(stderr
, "\n");
594 fprintf(stderr
, " %18.9f seconds time elapsed",
595 avg_stats(&walltime_nsecs_stats
)/1e9
);
597 fprintf(stderr
, " ( +- %7.3f%% )",
598 100*stddev_stats(&walltime_nsecs_stats
) /
599 avg_stats(&walltime_nsecs_stats
));
601 fprintf(stderr
, "\n\n");
605 static volatile int signr
= -1;
607 static void skip_signal(int signo
)
615 static void sig_atexit(void)
618 kill(child_pid
, SIGTERM
);
623 signal(signr
, SIG_DFL
);
624 kill(getpid(), signr
);
627 static const char * const stat_usage
[] = {
628 "perf stat [<options>] [<command>]",
632 static int stat__set_big_num(const struct option
*opt __used
,
633 const char *s __used
, int unset
)
635 big_num_opt
= unset
? 0 : 1;
639 static const struct option options
[] = {
640 OPT_CALLBACK('e', "event", &evsel_list
, "event",
641 "event selector. use 'perf list' to list available events",
643 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
644 "event filter", parse_filter
),
645 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
646 "child tasks do not inherit counters"),
647 OPT_INTEGER('p', "pid", &target_pid
,
648 "stat events on existing process id"),
649 OPT_INTEGER('t', "tid", &target_tid
,
650 "stat events on existing thread id"),
651 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
652 "system-wide collection from all CPUs"),
653 OPT_BOOLEAN('c', "scale", &scale
,
654 "scale/normalize counters"),
655 OPT_INCR('v', "verbose", &verbose
,
656 "be more verbose (show counter open errors, etc)"),
657 OPT_INTEGER('r', "repeat", &run_count
,
658 "repeat command and print average + stddev (max: 100)"),
659 OPT_BOOLEAN('n', "null", &null_run
,
660 "null run - dont start any counters"),
661 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
662 "print large numbers with thousands\' separators",
664 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
665 "list of cpus to monitor in system-wide"),
666 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
667 "disable CPU count aggregation"),
668 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
669 "print counts with custom separator"),
670 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
671 "monitor event in cgroup name only",
676 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
678 struct perf_evsel
*pos
;
679 int status
= -ENOMEM
;
681 setlocale(LC_ALL
, "");
683 evsel_list
= perf_evlist__new(NULL
, NULL
);
684 if (evsel_list
== NULL
)
687 argc
= parse_options(argc
, argv
, options
, stat_usage
,
688 PARSE_OPT_STOP_AT_NON_OPTION
);
693 csv_sep
= DEFAULT_SEPARATOR
;
696 * let the spreadsheet do the pretty-printing
699 /* User explicitely passed -B? */
700 if (big_num_opt
== 1) {
701 fprintf(stderr
, "-B option not supported with -x\n");
702 usage_with_options(stat_usage
, options
);
703 } else /* Nope, so disable big number formatting */
705 } else if (big_num_opt
== 0) /* User passed --no-big-num */
708 if (!argc
&& target_pid
== -1 && target_tid
== -1)
709 usage_with_options(stat_usage
, options
);
711 usage_with_options(stat_usage
, options
);
713 /* no_aggr, cgroup are for system-wide only */
714 if ((no_aggr
|| nr_cgroups
) && !system_wide
) {
715 fprintf(stderr
, "both cgroup and no-aggregation "
716 "modes only available in system-wide mode\n");
718 usage_with_options(stat_usage
, options
);
721 /* Set attrs and nr_counters if no event is selected and !null_run */
722 if (!null_run
&& !evsel_list
->nr_entries
) {
725 for (c
= 0; c
< ARRAY_SIZE(default_attrs
); ++c
) {
726 pos
= perf_evsel__new(&default_attrs
[c
], c
);
729 perf_evlist__add(evsel_list
, pos
);
733 if (target_pid
!= -1)
734 target_tid
= target_pid
;
736 evsel_list
->threads
= thread_map__new(target_pid
, target_tid
);
737 if (evsel_list
->threads
== NULL
) {
738 pr_err("Problems finding threads of monitor\n");
739 usage_with_options(stat_usage
, options
);
743 evsel_list
->cpus
= cpu_map__new(cpu_list
);
745 evsel_list
->cpus
= cpu_map__dummy_new();
747 if (evsel_list
->cpus
== NULL
) {
748 perror("failed to parse CPUs map");
749 usage_with_options(stat_usage
, options
);
753 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
754 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
755 perf_evsel__alloc_counts(pos
, evsel_list
->cpus
->nr
) < 0 ||
756 perf_evsel__alloc_fd(pos
, evsel_list
->cpus
->nr
, evsel_list
->threads
->nr
) < 0)
761 * We dont want to block the signals - that would cause
762 * child tasks to inherit that and Ctrl-C would not work.
763 * What we want is for Ctrl-C to work in the exec()-ed
764 * task, but being ignored by perf stat itself:
767 signal(SIGINT
, skip_signal
);
768 signal(SIGALRM
, skip_signal
);
769 signal(SIGABRT
, skip_signal
);
772 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
773 if (run_count
!= 1 && verbose
)
774 fprintf(stderr
, "[ perf stat: executing run #%d ... ]\n", run_idx
+ 1);
775 status
= run_perf_stat(argc
, argv
);
779 print_stat(argc
, argv
);
781 list_for_each_entry(pos
, &evsel_list
->entries
, node
)
782 perf_evsel__free_stat_priv(pos
);
783 perf_evlist__delete_maps(evsel_list
);
785 perf_evlist__delete(evsel_list
);