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/debug.h"
47 #include "util/header.h"
48 #include "util/cpumap.h"
49 #include "util/thread.h"
51 #include <sys/prctl.h>
55 #define DEFAULT_SEPARATOR " "
57 static struct perf_event_attr default_attrs[] = {
59 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
60 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
61 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
62 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
64 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
65 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
66 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
67 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
68 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
69 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
73 static bool system_wide = false;
74 static int nr_cpus = 0;
75 static int run_idx = 0;
77 static int run_count = 1;
78 static bool no_inherit = false;
79 static bool scale = true;
80 static bool no_aggr = false;
81 static pid_t target_pid = -1;
82 static pid_t target_tid = -1;
83 static pid_t *all_tids = NULL;
84 static int thread_num = 0;
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;
94 static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
96 static int event_scaled[MAX_COUNTERS];
102 } cpu_counts[MAX_NR_CPUS][MAX_COUNTERS];
104 static volatile int done = 0;
111 static void update_stats(struct stats *stats, u64 val)
116 delta = val - stats->mean;
117 stats->mean += delta / stats->n;
118 stats->M2 += delta*(val - stats->mean);
121 static double avg_stats(struct stats *stats)
127 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
129 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
130 * s^2 = -------------------------------
133 * http://en.wikipedia.org/wiki/Stddev
135 * The std dev of the mean is related to the std dev by:
142 static double stddev_stats(struct stats *stats)
144 double variance = stats->M2 / (stats->n - 1);
145 double variance_mean = variance / stats->n;
147 return sqrt(variance_mean);
150 struct stats event_res_stats[MAX_COUNTERS][3];
151 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
152 struct stats runtime_cycles_stats[MAX_NR_CPUS];
153 struct stats runtime_branches_stats[MAX_NR_CPUS];
154 struct stats walltime_nsecs_stats;
156 #define MATCH_EVENT(t, c, counter) \
157 (attrs[counter].type == PERF_TYPE_##t && \
158 attrs[counter].config == PERF_COUNT_##c)
160 #define ERR_PERF_OPEN \
161 "counter %d, sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information."
163 static int create_perf_stat_counter(int counter, bool *perm_err)
165 struct perf_event_attr *attr = attrs + counter;
170 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
171 PERF_FORMAT_TOTAL_TIME_RUNNING;
176 for (cpu = 0; cpu < nr_cpus; cpu++) {
177 fd[cpu][counter][0] = sys_perf_event_open(attr,
178 -1, cpumap[cpu], -1, 0);
179 if (fd[cpu][counter][0] < 0) {
180 if (errno == EPERM || errno == EACCES)
182 error(ERR_PERF_OPEN, counter,
183 fd[cpu][counter][0], strerror(errno));
189 attr->inherit = !no_inherit;
190 if (target_pid == -1 && target_tid == -1) {
192 attr->enable_on_exec = 1;
194 for (thread = 0; thread < thread_num; thread++) {
195 fd[0][counter][thread] = sys_perf_event_open(attr,
196 all_tids[thread], -1, -1, 0);
197 if (fd[0][counter][thread] < 0) {
198 if (errno == EPERM || errno == EACCES)
200 error(ERR_PERF_OPEN, counter,
201 fd[0][counter][thread],
213 * Does the counter have nsecs as a unit?
215 static inline int nsec_counter(int counter)
217 if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
218 MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
225 * Read out the results of a single counter:
226 * aggregate counts across CPUs in system-wide mode
228 static void read_counter_aggr(int counter)
230 u64 count[3], single_count[3];
236 count[0] = count[1] = count[2] = 0;
239 for (cpu = 0; cpu < nr_cpus; cpu++) {
240 for (thread = 0; thread < thread_num; thread++) {
241 if (fd[cpu][counter][thread] < 0)
244 res = read(fd[cpu][counter][thread],
245 single_count, nv * sizeof(u64));
246 assert(res == nv * sizeof(u64));
248 close(fd[cpu][counter][thread]);
249 fd[cpu][counter][thread] = -1;
251 count[0] += single_count[0];
253 count[1] += single_count[1];
254 count[2] += single_count[2];
262 event_scaled[counter] = -1;
267 if (count[2] < count[1]) {
268 event_scaled[counter] = 1;
269 count[0] = (unsigned long long)
270 ((double)count[0] * count[1] / count[2] + 0.5);
274 for (i = 0; i < 3; i++)
275 update_stats(&event_res_stats[counter][i], count[i]);
278 fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
279 count[0], count[1], count[2]);
283 * Save the full runtime - to allow normalization during printout:
285 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
286 update_stats(&runtime_nsecs_stats[0], count[0]);
287 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
288 update_stats(&runtime_cycles_stats[0], count[0]);
289 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
290 update_stats(&runtime_branches_stats[0], count[0]);
294 * Read out the results of a single counter:
295 * do not aggregate counts across CPUs in system-wide mode
297 static void read_counter(int counter)
303 count[0] = count[1] = count[2] = 0;
307 for (cpu = 0; cpu < nr_cpus; cpu++) {
309 if (fd[cpu][counter][0] < 0)
312 res = read(fd[cpu][counter][0], count, nv * sizeof(u64));
314 assert(res == nv * sizeof(u64));
316 close(fd[cpu][counter][0]);
317 fd[cpu][counter][0] = -1;
322 } else if (count[2] < count[1]) {
323 count[0] = (unsigned long long)
324 ((double)count[0] * count[1] / count[2] + 0.5);
327 cpu_counts[cpu][counter].val = count[0]; /* scaled count */
328 cpu_counts[cpu][counter].ena = count[1];
329 cpu_counts[cpu][counter].run = count[2];
331 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
332 update_stats(&runtime_nsecs_stats[cpu], count[0]);
333 if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
334 update_stats(&runtime_cycles_stats[cpu], count[0]);
335 if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
336 update_stats(&runtime_branches_stats[cpu], count[0]);
340 static int run_perf_stat(int argc __used, const char **argv)
342 unsigned long long t0, t1;
344 int counter, ncreated = 0;
345 int child_ready_pipe[2], go_pipe[2];
346 bool perm_err = false;
347 const bool forks = (argc > 0);
353 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
354 perror("failed to create pipes");
359 if ((child_pid = fork()) < 0)
360 perror("failed to fork");
363 close(child_ready_pipe[0]);
365 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
368 * Do a dummy execvp to get the PLT entry resolved,
369 * so we avoid the resolver overhead on the real
372 execvp("", (char **)argv);
375 * Tell the parent we're ready to go
377 close(child_ready_pipe[1]);
380 * Wait until the parent tells us to go.
382 if (read(go_pipe[0], &buf, 1) == -1)
383 perror("unable to read pipe");
385 execvp(argv[0], (char **)argv);
391 if (target_tid == -1 && target_pid == -1 && !system_wide)
392 all_tids[0] = child_pid;
395 * Wait for the child to be ready to exec.
397 close(child_ready_pipe[1]);
399 if (read(child_ready_pipe[0], &buf, 1) == -1)
400 perror("unable to read pipe");
401 close(child_ready_pipe[0]);
404 for (counter = 0; counter < nr_counters; counter++)
405 ncreated += create_perf_stat_counter(counter, &perm_err);
407 if (ncreated < nr_counters) {
409 error("You may not have permission to collect %sstats.\n"
410 "\t Consider tweaking"
411 " /proc/sys/kernel/perf_event_paranoid or running as root.",
412 system_wide ? "system-wide " : "");
413 die("Not all events could be opened.\n");
415 kill(child_pid, SIGTERM);
420 * Enable counters and exec the command:
428 while(!done) sleep(1);
433 update_stats(&walltime_nsecs_stats, t1 - t0);
436 for (counter = 0; counter < nr_counters; counter++)
437 read_counter(counter);
439 for (counter = 0; counter < nr_counters; counter++)
440 read_counter_aggr(counter);
442 return WEXITSTATUS(status);
445 static void print_noise(int counter, double avg)
450 fprintf(stderr, " ( +- %7.3f%% )",
451 100 * stddev_stats(&event_res_stats[counter][0]) / avg);
454 static void nsec_printout(int cpu, int counter, double avg)
456 double msecs = avg / 1e6;
457 char cpustr[16] = { '\0', };
458 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
461 sprintf(cpustr, "CPU%*d%s",
463 cpumap[cpu], csv_sep);
465 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(counter));
470 if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
471 fprintf(stderr, " # %10.3f CPUs ",
472 avg / avg_stats(&walltime_nsecs_stats));
476 static void abs_printout(int cpu, int counter, double avg)
478 double total, ratio = 0.0;
479 char cpustr[16] = { '\0', };
485 fmt = "%s%'18.0f%s%-24s";
487 fmt = "%s%18.0f%s%-24s";
490 sprintf(cpustr, "CPU%*d%s",
492 cpumap[cpu], csv_sep);
496 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(counter));
501 if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
502 total = avg_stats(&runtime_cycles_stats[cpu]);
507 fprintf(stderr, " # %10.3f IPC ", ratio);
508 } else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
509 runtime_branches_stats[cpu].n != 0) {
510 total = avg_stats(&runtime_branches_stats[cpu]);
513 ratio = avg * 100 / total;
515 fprintf(stderr, " # %10.3f %% ", ratio);
517 } else if (runtime_nsecs_stats[cpu].n != 0) {
518 total = avg_stats(&runtime_nsecs_stats[cpu]);
521 ratio = 1000.0 * avg / total;
523 fprintf(stderr, " # %10.3f M/sec", ratio);
528 * Print out the results of a single counter:
529 * aggregated counts in system-wide mode
531 static void print_counter_aggr(int counter)
533 double avg = avg_stats(&event_res_stats[counter][0]);
534 int scaled = event_scaled[counter];
537 fprintf(stderr, "%*s%s%-24s\n",
539 "<not counted>", csv_sep, event_name(counter));
543 if (nsec_counter(counter))
544 nsec_printout(-1, counter, avg);
546 abs_printout(-1, counter, avg);
553 print_noise(counter, avg);
556 double avg_enabled, avg_running;
558 avg_enabled = avg_stats(&event_res_stats[counter][1]);
559 avg_running = avg_stats(&event_res_stats[counter][2]);
561 fprintf(stderr, " (scaled from %.2f%%)",
562 100 * avg_running / avg_enabled);
565 fprintf(stderr, "\n");
569 * Print out the results of a single counter:
570 * does not use aggregated count in system-wide
572 static void print_counter(int counter)
577 for (cpu = 0; cpu < nr_cpus; cpu++) {
578 val = cpu_counts[cpu][counter].val;
579 ena = cpu_counts[cpu][counter].ena;
580 run = cpu_counts[cpu][counter].run;
581 if (run == 0 || ena == 0) {
582 fprintf(stderr, "CPU%*d%s%*s%s%-24s",
584 cpumap[cpu], csv_sep,
586 "<not counted>", csv_sep,
587 event_name(counter));
589 fprintf(stderr, "\n");
593 if (nsec_counter(counter))
594 nsec_printout(cpu, counter, val);
596 abs_printout(cpu, counter, val);
599 print_noise(counter, 1.0);
602 fprintf(stderr, " (scaled from %.2f%%)",
606 fprintf(stderr, "\n");
610 static void print_stat(int argc, const char **argv)
617 fprintf(stderr, "\n");
618 fprintf(stderr, " Performance counter stats for ");
619 if(target_pid == -1 && target_tid == -1) {
620 fprintf(stderr, "\'%s", argv[0]);
621 for (i = 1; i < argc; i++)
622 fprintf(stderr, " %s", argv[i]);
623 } else if (target_pid != -1)
624 fprintf(stderr, "process id \'%d", target_pid);
626 fprintf(stderr, "thread id \'%d", target_tid);
628 fprintf(stderr, "\'");
630 fprintf(stderr, " (%d runs)", run_count);
631 fprintf(stderr, ":\n\n");
635 for (counter = 0; counter < nr_counters; counter++)
636 print_counter(counter);
638 for (counter = 0; counter < nr_counters; counter++)
639 print_counter_aggr(counter);
643 fprintf(stderr, "\n");
644 fprintf(stderr, " %18.9f seconds time elapsed",
645 avg_stats(&walltime_nsecs_stats)/1e9);
647 fprintf(stderr, " ( +- %7.3f%% )",
648 100*stddev_stats(&walltime_nsecs_stats) /
649 avg_stats(&walltime_nsecs_stats));
651 fprintf(stderr, "\n\n");
655 static volatile int signr = -1;
657 static void skip_signal(int signo)
665 static void sig_atexit(void)
668 kill(child_pid, SIGTERM);
673 signal(signr, SIG_DFL);
674 kill(getpid(), signr);
677 static const char * const stat_usage[] = {
678 "perf stat [<options>] [<command>]",
682 static int stat__set_big_num(const struct option *opt __used,
683 const char *s __used, int unset)
685 big_num_opt = unset ? 0 : 1;
689 static const struct option options[] = {
690 OPT_CALLBACK('e', "event", NULL, "event",
691 "event selector. use 'perf list' to list available events",
693 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
694 "child tasks do not inherit counters"),
695 OPT_INTEGER('p', "pid", &target_pid,
696 "stat events on existing process id"),
697 OPT_INTEGER('t', "tid", &target_tid,
698 "stat events on existing thread id"),
699 OPT_BOOLEAN('a', "all-cpus", &system_wide,
700 "system-wide collection from all CPUs"),
701 OPT_BOOLEAN('c', "scale", &scale,
702 "scale/normalize counters"),
703 OPT_INCR('v', "verbose", &verbose,
704 "be more verbose (show counter open errors, etc)"),
705 OPT_INTEGER('r', "repeat", &run_count,
706 "repeat command and print average + stddev (max: 100)"),
707 OPT_BOOLEAN('n', "null", &null_run,
708 "null run - dont start any counters"),
709 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
710 "print large numbers with thousands\' separators",
712 OPT_STRING('C', "cpu", &cpu_list, "cpu",
713 "list of cpus to monitor in system-wide"),
714 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
715 "disable CPU count aggregation"),
716 OPT_STRING('x', "field-separator", &csv_sep, "separator",
717 "print counts with custom separator"),
721 int cmd_stat(int argc, const char **argv, const char *prefix __used)
726 setlocale(LC_ALL, "");
728 argc = parse_options(argc, argv, options, stat_usage,
729 PARSE_OPT_STOP_AT_NON_OPTION);
734 csv_sep = DEFAULT_SEPARATOR;
737 * let the spreadsheet do the pretty-printing
740 /* User explicitely passed -B? */
741 if (big_num_opt == 1) {
742 fprintf(stderr, "-B option not supported with -x\n");
743 usage_with_options(stat_usage, options);
744 } else /* Nope, so disable big number formatting */
746 } else if (big_num_opt == 0) /* User passed --no-big-num */
749 if (!argc && target_pid == -1 && target_tid == -1)
750 usage_with_options(stat_usage, options);
752 usage_with_options(stat_usage, options);
754 /* no_aggr is for system-wide only */
755 if (no_aggr && !system_wide)
756 usage_with_options(stat_usage, options);
758 /* Set attrs and nr_counters if no event is selected and !null_run */
759 if (!null_run && !nr_counters) {
760 memcpy(attrs, default_attrs, sizeof(default_attrs));
761 nr_counters = ARRAY_SIZE(default_attrs);
765 nr_cpus = read_cpu_map(cpu_list);
770 usage_with_options(stat_usage, options);
772 if (target_pid != -1) {
773 target_tid = target_pid;
774 thread_num = find_all_tid(target_pid, &all_tids);
775 if (thread_num <= 0) {
776 fprintf(stderr, "Can't find all threads of pid %d\n",
778 usage_with_options(stat_usage, options);
781 all_tids=malloc(sizeof(pid_t));
785 all_tids[0] = target_tid;
789 for (i = 0; i < MAX_NR_CPUS; i++) {
790 for (j = 0; j < MAX_COUNTERS; j++) {
791 fd[i][j] = malloc(sizeof(int)*thread_num);
798 * We dont want to block the signals - that would cause
799 * child tasks to inherit that and Ctrl-C would not work.
800 * What we want is for Ctrl-C to work in the exec()-ed
801 * task, but being ignored by perf stat itself:
804 signal(SIGINT, skip_signal);
805 signal(SIGALRM, skip_signal);
806 signal(SIGABRT, skip_signal);
809 for (run_idx = 0; run_idx < run_count; run_idx++) {
810 if (run_count != 1 && verbose)
811 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
812 status = run_perf_stat(argc, argv);
816 print_stat(argc, argv);