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 struct thread_map
*threads
;
85 static pid_t child_pid
= -1;
86 static bool null_run
= false;
87 static bool big_num
= true;
88 static int big_num_opt
= -1;
89 static const char *cpu_list
;
90 static const char *csv_sep
= NULL
;
91 static bool csv_output
= false;
93 static volatile int done
= 0;
101 struct stats res_stats
[3];
104 static int perf_evsel__alloc_stat_priv(struct perf_evsel
*evsel
)
106 evsel
->priv
= zalloc(sizeof(struct perf_stat
));
107 return evsel
->priv
== NULL
? -ENOMEM
: 0;
110 static void perf_evsel__free_stat_priv(struct perf_evsel
*evsel
)
116 static void update_stats(struct stats
*stats
, u64 val
)
121 delta
= val
- stats
->mean
;
122 stats
->mean
+= delta
/ stats
->n
;
123 stats
->M2
+= delta
*(val
- stats
->mean
);
126 static double avg_stats(struct stats
*stats
)
132 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
134 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
135 * s^2 = -------------------------------
138 * http://en.wikipedia.org/wiki/Stddev
140 * The std dev of the mean is related to the std dev by:
147 static double stddev_stats(struct stats
*stats
)
149 double variance
= stats
->M2
/ (stats
->n
- 1);
150 double variance_mean
= variance
/ stats
->n
;
152 return sqrt(variance_mean
);
155 struct stats runtime_nsecs_stats
[MAX_NR_CPUS
];
156 struct stats runtime_cycles_stats
[MAX_NR_CPUS
];
157 struct stats runtime_branches_stats
[MAX_NR_CPUS
];
158 struct stats walltime_nsecs_stats
;
160 static int create_perf_stat_counter(struct perf_evsel
*evsel
)
162 struct perf_event_attr
*attr
= &evsel
->attr
;
165 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
166 PERF_FORMAT_TOTAL_TIME_RUNNING
;
169 return perf_evsel__open_per_cpu(evsel
, cpus
);
171 attr
->inherit
= !no_inherit
;
172 if (target_pid
== -1 && target_tid
== -1) {
174 attr
->enable_on_exec
= 1;
177 return perf_evsel__open_per_thread(evsel
, threads
);
181 * Does the counter have nsecs as a unit?
183 static inline int nsec_counter(struct perf_evsel
*evsel
)
185 if (perf_evsel__match(evsel
, SOFTWARE
, SW_CPU_CLOCK
) ||
186 perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
193 * Read out the results of a single counter:
194 * aggregate counts across CPUs in system-wide mode
196 static int read_counter_aggr(struct perf_evsel
*counter
)
198 struct perf_stat
*ps
= counter
->priv
;
199 u64
*count
= counter
->counts
->aggr
.values
;
202 if (__perf_evsel__read(counter
, cpus
->nr
, threads
->nr
, scale
) < 0)
205 for (i
= 0; i
< 3; i
++)
206 update_stats(&ps
->res_stats
[i
], count
[i
]);
209 fprintf(stderr
, "%s: %" PRIu64
" %" PRIu64
" %" PRIu64
"\n",
210 event_name(counter
), count
[0], count
[1], count
[2]);
214 * Save the full runtime - to allow normalization during printout:
216 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
217 update_stats(&runtime_nsecs_stats
[0], count
[0]);
218 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
219 update_stats(&runtime_cycles_stats
[0], count
[0]);
220 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
221 update_stats(&runtime_branches_stats
[0], count
[0]);
227 * Read out the results of a single counter:
228 * do not aggregate counts across CPUs in system-wide mode
230 static int read_counter(struct perf_evsel
*counter
)
235 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
236 if (__perf_evsel__read_on_cpu(counter
, cpu
, 0, scale
) < 0)
239 count
= counter
->counts
->cpu
[cpu
].values
;
241 if (perf_evsel__match(counter
, SOFTWARE
, SW_TASK_CLOCK
))
242 update_stats(&runtime_nsecs_stats
[cpu
], count
[0]);
243 if (perf_evsel__match(counter
, HARDWARE
, HW_CPU_CYCLES
))
244 update_stats(&runtime_cycles_stats
[cpu
], count
[0]);
245 if (perf_evsel__match(counter
, HARDWARE
, HW_BRANCH_INSTRUCTIONS
))
246 update_stats(&runtime_branches_stats
[cpu
], count
[0]);
252 static int run_perf_stat(int argc __used
, const char **argv
)
254 unsigned long long t0
, t1
;
255 struct perf_evsel
*counter
;
257 int child_ready_pipe
[2], go_pipe
[2];
258 const bool forks
= (argc
> 0);
261 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
262 perror("failed to create pipes");
267 if ((child_pid
= fork()) < 0)
268 perror("failed to fork");
271 close(child_ready_pipe
[0]);
273 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
276 * Do a dummy execvp to get the PLT entry resolved,
277 * so we avoid the resolver overhead on the real
280 execvp("", (char **)argv
);
283 * Tell the parent we're ready to go
285 close(child_ready_pipe
[1]);
288 * Wait until the parent tells us to go.
290 if (read(go_pipe
[0], &buf
, 1) == -1)
291 perror("unable to read pipe");
293 execvp(argv
[0], (char **)argv
);
299 if (target_tid
== -1 && target_pid
== -1 && !system_wide
)
300 threads
->map
[0] = child_pid
;
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 list_for_each_entry(counter
, &evsel_list
, node
) {
313 if (create_perf_stat_counter(counter
) < 0) {
314 if (errno
== -EPERM
|| errno
== -EACCES
) {
315 error("You may not have permission to collect %sstats.\n"
316 "\t Consider tweaking"
317 " /proc/sys/kernel/perf_event_paranoid or running as root.",
318 system_wide
? "system-wide " : "");
319 } else if (errno
== ENOENT
) {
320 error("%s event is not supported. ", event_name(counter
));
322 error("open_counter returned with %d (%s). "
323 "/bin/dmesg may provide additional information.\n",
324 errno
, strerror(errno
));
327 kill(child_pid
, SIGTERM
);
328 die("Not all events could be opened.\n");
334 * Enable counters and exec the command:
342 while(!done
) sleep(1);
347 update_stats(&walltime_nsecs_stats
, t1
- t0
);
350 list_for_each_entry(counter
, &evsel_list
, node
) {
351 read_counter(counter
);
352 perf_evsel__close_fd(counter
, cpus
->nr
, 1);
355 list_for_each_entry(counter
, &evsel_list
, node
) {
356 read_counter_aggr(counter
);
357 perf_evsel__close_fd(counter
, cpus
->nr
, threads
->nr
);
361 return WEXITSTATUS(status
);
364 static void print_noise(struct perf_evsel
*evsel
, double avg
)
366 struct perf_stat
*ps
;
372 fprintf(stderr
, " ( +- %7.3f%% )",
373 100 * stddev_stats(&ps
->res_stats
[0]) / avg
);
376 static void nsec_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
378 double msecs
= avg
/ 1e6
;
379 char cpustr
[16] = { '\0', };
380 const char *fmt
= csv_output
? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
383 sprintf(cpustr
, "CPU%*d%s",
385 cpus
->map
[cpu
], csv_sep
);
387 fprintf(stderr
, fmt
, cpustr
, msecs
, csv_sep
, event_name(evsel
));
392 if (perf_evsel__match(evsel
, SOFTWARE
, SW_TASK_CLOCK
))
393 fprintf(stderr
, " # %10.3f CPUs ",
394 avg
/ avg_stats(&walltime_nsecs_stats
));
397 static void abs_printout(int cpu
, struct perf_evsel
*evsel
, double avg
)
399 double total
, ratio
= 0.0;
400 char cpustr
[16] = { '\0', };
406 fmt
= "%s%'18.0f%s%-24s";
408 fmt
= "%s%18.0f%s%-24s";
411 sprintf(cpustr
, "CPU%*d%s",
413 cpus
->map
[cpu
], csv_sep
);
417 fprintf(stderr
, fmt
, cpustr
, avg
, csv_sep
, event_name(evsel
));
422 if (perf_evsel__match(evsel
, HARDWARE
, HW_INSTRUCTIONS
)) {
423 total
= avg_stats(&runtime_cycles_stats
[cpu
]);
428 fprintf(stderr
, " # %10.3f IPC ", ratio
);
429 } else if (perf_evsel__match(evsel
, HARDWARE
, HW_BRANCH_MISSES
) &&
430 runtime_branches_stats
[cpu
].n
!= 0) {
431 total
= avg_stats(&runtime_branches_stats
[cpu
]);
434 ratio
= avg
* 100 / total
;
436 fprintf(stderr
, " # %10.3f %% ", ratio
);
438 } else if (runtime_nsecs_stats
[cpu
].n
!= 0) {
439 total
= avg_stats(&runtime_nsecs_stats
[cpu
]);
442 ratio
= 1000.0 * avg
/ total
;
444 fprintf(stderr
, " # %10.3f M/sec", ratio
);
449 * Print out the results of a single counter:
450 * aggregated counts in system-wide mode
452 static void print_counter_aggr(struct perf_evsel
*counter
)
454 struct perf_stat
*ps
= counter
->priv
;
455 double avg
= avg_stats(&ps
->res_stats
[0]);
456 int scaled
= counter
->counts
->scaled
;
459 fprintf(stderr
, "%*s%s%-24s\n",
461 "<not counted>", csv_sep
, event_name(counter
));
465 if (nsec_counter(counter
))
466 nsec_printout(-1, counter
, avg
);
468 abs_printout(-1, counter
, avg
);
475 print_noise(counter
, avg
);
478 double avg_enabled
, avg_running
;
480 avg_enabled
= avg_stats(&ps
->res_stats
[1]);
481 avg_running
= avg_stats(&ps
->res_stats
[2]);
483 fprintf(stderr
, " (scaled from %.2f%%)",
484 100 * avg_running
/ avg_enabled
);
487 fprintf(stderr
, "\n");
491 * Print out the results of a single counter:
492 * does not use aggregated count in system-wide
494 static void print_counter(struct perf_evsel
*counter
)
499 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
500 val
= counter
->counts
->cpu
[cpu
].val
;
501 ena
= counter
->counts
->cpu
[cpu
].ena
;
502 run
= counter
->counts
->cpu
[cpu
].run
;
503 if (run
== 0 || ena
== 0) {
504 fprintf(stderr
, "CPU%*d%s%*s%s%-24s",
506 cpus
->map
[cpu
], csv_sep
,
508 "<not counted>", csv_sep
,
509 event_name(counter
));
511 fprintf(stderr
, "\n");
515 if (nsec_counter(counter
))
516 nsec_printout(cpu
, counter
, val
);
518 abs_printout(cpu
, counter
, val
);
521 print_noise(counter
, 1.0);
524 fprintf(stderr
, " (scaled from %.2f%%)",
528 fprintf(stderr
, "\n");
532 static void print_stat(int argc
, const char **argv
)
534 struct perf_evsel
*counter
;
540 fprintf(stderr
, "\n");
541 fprintf(stderr
, " Performance counter stats for ");
542 if(target_pid
== -1 && target_tid
== -1) {
543 fprintf(stderr
, "\'%s", argv
[0]);
544 for (i
= 1; i
< argc
; i
++)
545 fprintf(stderr
, " %s", argv
[i
]);
546 } else if (target_pid
!= -1)
547 fprintf(stderr
, "process id \'%d", target_pid
);
549 fprintf(stderr
, "thread id \'%d", target_tid
);
551 fprintf(stderr
, "\'");
553 fprintf(stderr
, " (%d runs)", run_count
);
554 fprintf(stderr
, ":\n\n");
558 list_for_each_entry(counter
, &evsel_list
, node
)
559 print_counter(counter
);
561 list_for_each_entry(counter
, &evsel_list
, node
)
562 print_counter_aggr(counter
);
566 fprintf(stderr
, "\n");
567 fprintf(stderr
, " %18.9f seconds time elapsed",
568 avg_stats(&walltime_nsecs_stats
)/1e9
);
570 fprintf(stderr
, " ( +- %7.3f%% )",
571 100*stddev_stats(&walltime_nsecs_stats
) /
572 avg_stats(&walltime_nsecs_stats
));
574 fprintf(stderr
, "\n\n");
578 static volatile int signr
= -1;
580 static void skip_signal(int signo
)
588 static void sig_atexit(void)
591 kill(child_pid
, SIGTERM
);
596 signal(signr
, SIG_DFL
);
597 kill(getpid(), signr
);
600 static const char * const stat_usage
[] = {
601 "perf stat [<options>] [<command>]",
605 static int stat__set_big_num(const struct option
*opt __used
,
606 const char *s __used
, int unset
)
608 big_num_opt
= unset
? 0 : 1;
612 static const struct option options
[] = {
613 OPT_CALLBACK('e', "event", NULL
, "event",
614 "event selector. use 'perf list' to list available events",
616 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
617 "child tasks do not inherit counters"),
618 OPT_INTEGER('p', "pid", &target_pid
,
619 "stat events on existing process id"),
620 OPT_INTEGER('t', "tid", &target_tid
,
621 "stat events on existing thread id"),
622 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
623 "system-wide collection from all CPUs"),
624 OPT_BOOLEAN('c', "scale", &scale
,
625 "scale/normalize counters"),
626 OPT_INCR('v', "verbose", &verbose
,
627 "be more verbose (show counter open errors, etc)"),
628 OPT_INTEGER('r', "repeat", &run_count
,
629 "repeat command and print average + stddev (max: 100)"),
630 OPT_BOOLEAN('n', "null", &null_run
,
631 "null run - dont start any counters"),
632 OPT_CALLBACK_NOOPT('B', "big-num", NULL
, NULL
,
633 "print large numbers with thousands\' separators",
635 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
636 "list of cpus to monitor in system-wide"),
637 OPT_BOOLEAN('A', "no-aggr", &no_aggr
,
638 "disable CPU count aggregation"),
639 OPT_STRING('x', "field-separator", &csv_sep
, "separator",
640 "print counts with custom separator"),
644 int cmd_stat(int argc
, const char **argv
, const char *prefix __used
)
646 struct perf_evsel
*pos
;
647 int status
= -ENOMEM
;
649 setlocale(LC_ALL
, "");
651 argc
= parse_options(argc
, argv
, options
, stat_usage
,
652 PARSE_OPT_STOP_AT_NON_OPTION
);
657 csv_sep
= DEFAULT_SEPARATOR
;
660 * let the spreadsheet do the pretty-printing
663 /* User explicitely passed -B? */
664 if (big_num_opt
== 1) {
665 fprintf(stderr
, "-B option not supported with -x\n");
666 usage_with_options(stat_usage
, options
);
667 } else /* Nope, so disable big number formatting */
669 } else if (big_num_opt
== 0) /* User passed --no-big-num */
672 if (!argc
&& target_pid
== -1 && target_tid
== -1)
673 usage_with_options(stat_usage
, options
);
675 usage_with_options(stat_usage
, options
);
677 /* no_aggr is for system-wide only */
678 if (no_aggr
&& !system_wide
)
679 usage_with_options(stat_usage
, options
);
681 /* Set attrs and nr_counters if no event is selected and !null_run */
682 if (!null_run
&& !nr_counters
) {
685 nr_counters
= ARRAY_SIZE(default_attrs
);
687 for (c
= 0; c
< ARRAY_SIZE(default_attrs
); ++c
) {
688 pos
= perf_evsel__new(&default_attrs
[c
],
692 list_add(&pos
->node
, &evsel_list
);
696 if (target_pid
!= -1)
697 target_tid
= target_pid
;
699 threads
= thread_map__new(target_pid
, target_tid
);
700 if (threads
== NULL
) {
701 pr_err("Problems finding threads of monitor\n");
702 usage_with_options(stat_usage
, options
);
706 cpus
= cpu_map__new(cpu_list
);
708 cpus
= cpu_map__dummy_new();
711 perror("failed to parse CPUs map");
712 usage_with_options(stat_usage
, options
);
716 list_for_each_entry(pos
, &evsel_list
, node
) {
717 if (perf_evsel__alloc_stat_priv(pos
) < 0 ||
718 perf_evsel__alloc_counts(pos
, cpus
->nr
) < 0 ||
719 perf_evsel__alloc_fd(pos
, cpus
->nr
, threads
->nr
) < 0)
724 * We dont want to block the signals - that would cause
725 * child tasks to inherit that and Ctrl-C would not work.
726 * What we want is for Ctrl-C to work in the exec()-ed
727 * task, but being ignored by perf stat itself:
730 signal(SIGINT
, skip_signal
);
731 signal(SIGALRM
, skip_signal
);
732 signal(SIGABRT
, skip_signal
);
735 for (run_idx
= 0; run_idx
< run_count
; run_idx
++) {
736 if (run_count
!= 1 && verbose
)
737 fprintf(stderr
, "[ perf stat: executing run #%d ... ]\n", run_idx
+ 1);
738 status
= run_perf_stat(argc
, argv
);
742 print_stat(argc
, argv
);
744 list_for_each_entry(pos
, &evsel_list
, node
)
745 perf_evsel__free_stat_priv(pos
);
746 perf_evsel_list__delete();
748 thread_map__delete(threads
);