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 FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
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 static bool system_wide = false;
77 static int nr_cpus = 0;
78 static int run_idx = 0;
80 static int run_count = 1;
81 static bool no_inherit = false;
82 static bool scale = true;
83 static bool no_aggr = false;
84 static pid_t target_pid = -1;
85 static pid_t target_tid = -1;
86 static pid_t *all_tids = NULL;
87 static int thread_num = 0;
88 static pid_t child_pid = -1;
89 static bool null_run = false;
90 static bool big_num = true;
91 static int big_num_opt = -1;
92 static const char *cpu_list;
93 static const char *csv_sep = NULL;
94 static bool csv_output = false;
96 static volatile int done = 0;
104 struct stats res_stats[3];
107 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
109 evsel->priv = zalloc(sizeof(struct perf_stat));
110 return evsel->priv == NULL ? -ENOMEM : 0;
113 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
119 static void update_stats(struct stats *stats, u64 val)
124 delta = val - stats->mean;
125 stats->mean += delta / stats->n;
126 stats->M2 += delta*(val - stats->mean);
129 static double avg_stats(struct stats *stats)
135 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
137 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
138 * s^2 = -------------------------------
141 * http://en.wikipedia.org/wiki/Stddev
143 * The std dev of the mean is related to the std dev by:
150 static double stddev_stats(struct stats *stats)
152 double variance = stats->M2 / (stats->n - 1);
153 double variance_mean = variance / stats->n;
155 return sqrt(variance_mean);
158 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
159 struct stats runtime_cycles_stats[MAX_NR_CPUS];
160 struct stats runtime_branches_stats[MAX_NR_CPUS];
161 struct stats walltime_nsecs_stats;
163 #define ERR_PERF_OPEN \
164 "counter %d, sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information."
166 static int create_perf_stat_counter(struct perf_evsel *evsel, bool *perm_err)
168 struct perf_event_attr *attr = &evsel->attr;
173 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
174 PERF_FORMAT_TOTAL_TIME_RUNNING;
179 for (cpu = 0; cpu < nr_cpus; cpu++) {
180 FD(evsel, cpu, 0) = sys_perf_event_open(attr,
181 -1, cpumap[cpu], -1, 0);
182 if (FD(evsel, cpu, 0) < 0) {
183 if (errno == EPERM || errno == EACCES)
185 error(ERR_PERF_OPEN, evsel->idx,
186 FD(evsel, cpu, 0), strerror(errno));
192 attr->inherit = !no_inherit;
193 if (target_pid == -1 && target_tid == -1) {
195 attr->enable_on_exec = 1;
197 for (thread = 0; thread < thread_num; thread++) {
198 FD(evsel, 0, thread) = sys_perf_event_open(attr,
199 all_tids[thread], -1, -1, 0);
200 if (FD(evsel, 0, thread) < 0) {
201 if (errno == EPERM || errno == EACCES)
203 error(ERR_PERF_OPEN, evsel->idx,
204 FD(evsel, 0, thread),
216 * Does the counter have nsecs as a unit?
218 static inline int nsec_counter(struct perf_evsel *evsel)
220 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
221 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
228 * Read out the results of a single counter:
229 * aggregate counts across CPUs in system-wide mode
231 static int read_counter_aggr(struct perf_evsel *counter)
233 struct perf_stat *ps = counter->priv;
234 u64 *count = counter->counts->aggr.values;
237 if (__perf_evsel__read(counter, nr_cpus, thread_num, scale) < 0)
240 for (i = 0; i < 3; i++)
241 update_stats(&ps->res_stats[i], count[i]);
244 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
245 count[0], count[1], count[2]);
249 * Save the full runtime - to allow normalization during printout:
251 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
252 update_stats(&runtime_nsecs_stats[0], count[0]);
253 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
254 update_stats(&runtime_cycles_stats[0], count[0]);
255 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
256 update_stats(&runtime_branches_stats[0], count[0]);
262 * Read out the results of a single counter:
263 * do not aggregate counts across CPUs in system-wide mode
265 static int read_counter(struct perf_evsel *counter)
270 for (cpu = 0; cpu < nr_cpus; cpu++) {
271 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
274 count = counter->counts->cpu[cpu].values;
276 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
277 update_stats(&runtime_nsecs_stats[cpu], count[0]);
278 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
279 update_stats(&runtime_cycles_stats[cpu], count[0]);
280 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
281 update_stats(&runtime_branches_stats[cpu], count[0]);
287 static int run_perf_stat(int argc __used, const char **argv)
289 unsigned long long t0, t1;
290 struct perf_evsel *counter;
293 int child_ready_pipe[2], go_pipe[2];
294 bool perm_err = false;
295 const bool forks = (argc > 0);
301 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
302 perror("failed to create pipes");
307 if ((child_pid = fork()) < 0)
308 perror("failed to fork");
311 close(child_ready_pipe[0]);
313 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
316 * Do a dummy execvp to get the PLT entry resolved,
317 * so we avoid the resolver overhead on the real
320 execvp("", (char **)argv);
323 * Tell the parent we're ready to go
325 close(child_ready_pipe[1]);
328 * Wait until the parent tells us to go.
330 if (read(go_pipe[0], &buf, 1) == -1)
331 perror("unable to read pipe");
333 execvp(argv[0], (char **)argv);
339 if (target_tid == -1 && target_pid == -1 && !system_wide)
340 all_tids[0] = child_pid;
343 * Wait for the child to be ready to exec.
345 close(child_ready_pipe[1]);
347 if (read(child_ready_pipe[0], &buf, 1) == -1)
348 perror("unable to read pipe");
349 close(child_ready_pipe[0]);
352 list_for_each_entry(counter, &evsel_list, node)
353 ncreated += create_perf_stat_counter(counter, &perm_err);
355 if (ncreated < nr_counters) {
357 error("You may not have permission to collect %sstats.\n"
358 "\t Consider tweaking"
359 " /proc/sys/kernel/perf_event_paranoid or running as root.",
360 system_wide ? "system-wide " : "");
361 die("Not all events could be opened.\n");
363 kill(child_pid, SIGTERM);
368 * Enable counters and exec the command:
376 while(!done) sleep(1);
381 update_stats(&walltime_nsecs_stats, t1 - t0);
384 list_for_each_entry(counter, &evsel_list, node) {
385 read_counter(counter);
386 perf_evsel__close_fd(counter, nr_cpus, 1);
389 list_for_each_entry(counter, &evsel_list, node) {
390 read_counter_aggr(counter);
391 perf_evsel__close_fd(counter, nr_cpus, thread_num);
395 return WEXITSTATUS(status);
398 static void print_noise(struct perf_evsel *evsel, double avg)
400 struct perf_stat *ps;
406 fprintf(stderr, " ( +- %7.3f%% )",
407 100 * stddev_stats(&ps->res_stats[0]) / avg);
410 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
412 double msecs = avg / 1e6;
413 char cpustr[16] = { '\0', };
414 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
417 sprintf(cpustr, "CPU%*d%s",
419 cpumap[cpu], csv_sep);
421 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
426 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
427 fprintf(stderr, " # %10.3f CPUs ",
428 avg / avg_stats(&walltime_nsecs_stats));
431 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
433 double total, ratio = 0.0;
434 char cpustr[16] = { '\0', };
440 fmt = "%s%'18.0f%s%-24s";
442 fmt = "%s%18.0f%s%-24s";
445 sprintf(cpustr, "CPU%*d%s",
447 cpumap[cpu], csv_sep);
451 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
456 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
457 total = avg_stats(&runtime_cycles_stats[cpu]);
462 fprintf(stderr, " # %10.3f IPC ", ratio);
463 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
464 runtime_branches_stats[cpu].n != 0) {
465 total = avg_stats(&runtime_branches_stats[cpu]);
468 ratio = avg * 100 / total;
470 fprintf(stderr, " # %10.3f %% ", ratio);
472 } else if (runtime_nsecs_stats[cpu].n != 0) {
473 total = avg_stats(&runtime_nsecs_stats[cpu]);
476 ratio = 1000.0 * avg / total;
478 fprintf(stderr, " # %10.3f M/sec", ratio);
483 * Print out the results of a single counter:
484 * aggregated counts in system-wide mode
486 static void print_counter_aggr(struct perf_evsel *counter)
488 struct perf_stat *ps = counter->priv;
489 double avg = avg_stats(&ps->res_stats[0]);
490 int scaled = counter->counts->scaled;
493 fprintf(stderr, "%*s%s%-24s\n",
495 "<not counted>", csv_sep, event_name(counter));
499 if (nsec_counter(counter))
500 nsec_printout(-1, counter, avg);
502 abs_printout(-1, counter, avg);
509 print_noise(counter, avg);
512 double avg_enabled, avg_running;
514 avg_enabled = avg_stats(&ps->res_stats[1]);
515 avg_running = avg_stats(&ps->res_stats[2]);
517 fprintf(stderr, " (scaled from %.2f%%)",
518 100 * avg_running / avg_enabled);
521 fprintf(stderr, "\n");
525 * Print out the results of a single counter:
526 * does not use aggregated count in system-wide
528 static void print_counter(struct perf_evsel *counter)
533 for (cpu = 0; cpu < nr_cpus; cpu++) {
534 val = counter->counts->cpu[cpu].val;
535 ena = counter->counts->cpu[cpu].ena;
536 run = counter->counts->cpu[cpu].run;
537 if (run == 0 || ena == 0) {
538 fprintf(stderr, "CPU%*d%s%*s%s%-24s",
540 cpumap[cpu], csv_sep,
542 "<not counted>", csv_sep,
543 event_name(counter));
545 fprintf(stderr, "\n");
549 if (nsec_counter(counter))
550 nsec_printout(cpu, counter, val);
552 abs_printout(cpu, counter, val);
555 print_noise(counter, 1.0);
558 fprintf(stderr, " (scaled from %.2f%%)",
562 fprintf(stderr, "\n");
566 static void print_stat(int argc, const char **argv)
568 struct perf_evsel *counter;
574 fprintf(stderr, "\n");
575 fprintf(stderr, " Performance counter stats for ");
576 if(target_pid == -1 && target_tid == -1) {
577 fprintf(stderr, "\'%s", argv[0]);
578 for (i = 1; i < argc; i++)
579 fprintf(stderr, " %s", argv[i]);
580 } else if (target_pid != -1)
581 fprintf(stderr, "process id \'%d", target_pid);
583 fprintf(stderr, "thread id \'%d", target_tid);
585 fprintf(stderr, "\'");
587 fprintf(stderr, " (%d runs)", run_count);
588 fprintf(stderr, ":\n\n");
592 list_for_each_entry(counter, &evsel_list, node)
593 print_counter(counter);
595 list_for_each_entry(counter, &evsel_list, node)
596 print_counter_aggr(counter);
600 fprintf(stderr, "\n");
601 fprintf(stderr, " %18.9f seconds time elapsed",
602 avg_stats(&walltime_nsecs_stats)/1e9);
604 fprintf(stderr, " ( +- %7.3f%% )",
605 100*stddev_stats(&walltime_nsecs_stats) /
606 avg_stats(&walltime_nsecs_stats));
608 fprintf(stderr, "\n\n");
612 static volatile int signr = -1;
614 static void skip_signal(int signo)
622 static void sig_atexit(void)
625 kill(child_pid, SIGTERM);
630 signal(signr, SIG_DFL);
631 kill(getpid(), signr);
634 static const char * const stat_usage[] = {
635 "perf stat [<options>] [<command>]",
639 static int stat__set_big_num(const struct option *opt __used,
640 const char *s __used, int unset)
642 big_num_opt = unset ? 0 : 1;
646 static const struct option options[] = {
647 OPT_CALLBACK('e', "event", NULL, "event",
648 "event selector. use 'perf list' to list available events",
650 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
651 "child tasks do not inherit counters"),
652 OPT_INTEGER('p', "pid", &target_pid,
653 "stat events on existing process id"),
654 OPT_INTEGER('t', "tid", &target_tid,
655 "stat events on existing thread id"),
656 OPT_BOOLEAN('a', "all-cpus", &system_wide,
657 "system-wide collection from all CPUs"),
658 OPT_BOOLEAN('c', "scale", &scale,
659 "scale/normalize counters"),
660 OPT_INCR('v', "verbose", &verbose,
661 "be more verbose (show counter open errors, etc)"),
662 OPT_INTEGER('r', "repeat", &run_count,
663 "repeat command and print average + stddev (max: 100)"),
664 OPT_BOOLEAN('n', "null", &null_run,
665 "null run - dont start any counters"),
666 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
667 "print large numbers with thousands\' separators",
669 OPT_STRING('C', "cpu", &cpu_list, "cpu",
670 "list of cpus to monitor in system-wide"),
671 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
672 "disable CPU count aggregation"),
673 OPT_STRING('x', "field-separator", &csv_sep, "separator",
674 "print counts with custom separator"),
678 int cmd_stat(int argc, const char **argv, const char *prefix __used)
680 struct perf_evsel *pos;
681 int status = -ENOMEM;
683 setlocale(LC_ALL, "");
685 argc = parse_options(argc, argv, options, stat_usage,
686 PARSE_OPT_STOP_AT_NON_OPTION);
691 csv_sep = DEFAULT_SEPARATOR;
694 * let the spreadsheet do the pretty-printing
697 /* User explicitely passed -B? */
698 if (big_num_opt == 1) {
699 fprintf(stderr, "-B option not supported with -x\n");
700 usage_with_options(stat_usage, options);
701 } else /* Nope, so disable big number formatting */
703 } else if (big_num_opt == 0) /* User passed --no-big-num */
706 if (!argc && target_pid == -1 && target_tid == -1)
707 usage_with_options(stat_usage, options);
709 usage_with_options(stat_usage, options);
711 /* no_aggr is for system-wide only */
712 if (no_aggr && !system_wide)
713 usage_with_options(stat_usage, options);
715 /* Set attrs and nr_counters if no event is selected and !null_run */
716 if (!null_run && !nr_counters) {
719 nr_counters = ARRAY_SIZE(default_attrs);
721 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
722 pos = perf_evsel__new(default_attrs[c].type,
723 default_attrs[c].config,
727 list_add(&pos->node, &evsel_list);
732 nr_cpus = read_cpu_map(cpu_list);
737 usage_with_options(stat_usage, options);
739 if (target_pid != -1) {
740 target_tid = target_pid;
741 thread_num = find_all_tid(target_pid, &all_tids);
742 if (thread_num <= 0) {
743 fprintf(stderr, "Can't find all threads of pid %d\n",
745 usage_with_options(stat_usage, options);
748 all_tids=malloc(sizeof(pid_t));
752 all_tids[0] = target_tid;
756 list_for_each_entry(pos, &evsel_list, node) {
757 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
758 perf_evsel__alloc_counts(pos, nr_cpus) < 0 ||
759 perf_evsel__alloc_fd(pos, nr_cpus, thread_num) < 0)
764 * We dont want to block the signals - that would cause
765 * child tasks to inherit that and Ctrl-C would not work.
766 * What we want is for Ctrl-C to work in the exec()-ed
767 * task, but being ignored by perf stat itself:
770 signal(SIGINT, skip_signal);
771 signal(SIGALRM, skip_signal);
772 signal(SIGABRT, skip_signal);
775 for (run_idx = 0; run_idx < run_count; run_idx++) {
776 if (run_count != 1 && verbose)
777 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
778 status = run_perf_stat(argc, argv);
782 print_stat(argc, argv);
784 list_for_each_entry(pos, &evsel_list, node)
785 perf_evsel__free_stat_priv(pos);