4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include "util/parse-options.h"
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/color.h"
56 #include "util/stat.h"
57 #include "util/header.h"
58 #include "util/cpumap.h"
59 #include "util/thread.h"
60 #include "util/thread_map.h"
63 #include <sys/prctl.h>
66 #define DEFAULT_SEPARATOR " "
67 #define CNTR_NOT_SUPPORTED "<not supported>"
68 #define CNTR_NOT_COUNTED "<not counted>"
70 static void print_counters(struct timespec *ts, int argc, const char **argv);
72 /* Default events used for perf stat -T */
73 static const char *transaction_attrs = {
85 /* More limited version when the CPU does not have all events. */
86 static const char * transaction_limited_attrs = {
96 static struct perf_evlist *evsel_list;
98 static struct target target = {
102 static int run_count = 1;
103 static bool no_inherit = false;
104 static bool scale = true;
105 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
106 static volatile pid_t child_pid = -1;
107 static bool null_run = false;
108 static int detailed_run = 0;
109 static bool transaction_run;
110 static bool big_num = true;
111 static int big_num_opt = -1;
112 static const char *csv_sep = NULL;
113 static bool csv_output = false;
114 static bool group = false;
115 static FILE *output = NULL;
116 static const char *pre_cmd = NULL;
117 static const char *post_cmd = NULL;
118 static bool sync_run = false;
119 static unsigned int interval = 0;
120 static unsigned int initial_delay = 0;
121 static unsigned int unit_width = 4; /* strlen("unit") */
122 static bool forever = false;
123 static struct timespec ref_time;
124 static struct cpu_map *aggr_map;
125 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
127 static volatile int done = 0;
129 static inline void diff_timespec(struct timespec *r, struct timespec *a,
132 r->tv_sec = a->tv_sec - b->tv_sec;
133 if (a->tv_nsec < b->tv_nsec) {
134 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
137 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
141 static void perf_stat__reset_stats(void)
143 perf_evlist__reset_stats(evsel_list);
144 perf_stat__reset_shadow_stats();
147 static int create_perf_stat_counter(struct perf_evsel *evsel)
149 struct perf_event_attr *attr = &evsel->attr;
152 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
153 PERF_FORMAT_TOTAL_TIME_RUNNING;
155 attr->inherit = !no_inherit;
157 if (target__has_cpu(&target))
158 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
160 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
163 attr->enable_on_exec = 1;
166 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
170 * Does the counter have nsecs as a unit?
172 static inline int nsec_counter(struct perf_evsel *evsel)
174 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
175 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
181 static void zero_per_pkg(struct perf_evsel *counter)
183 if (counter->per_pkg_mask)
184 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
187 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
189 unsigned long *mask = counter->per_pkg_mask;
190 struct cpu_map *cpus = perf_evsel__cpus(counter);
195 if (!counter->per_pkg)
198 if (cpu_map__empty(cpus))
202 mask = zalloc(MAX_NR_CPUS);
206 counter->per_pkg_mask = mask;
209 s = cpu_map__get_socket(cpus, cpu);
213 *skip = test_and_set_bit(s, mask) == 1;
218 process_counter_values(struct perf_evsel *evsel, int cpu, int thread,
219 struct perf_counts_values *count)
221 struct perf_counts_values *aggr = &evsel->counts->aggr;
222 static struct perf_counts_values zero;
225 if (check_per_pkg(evsel, cpu, &skip)) {
226 pr_err("failed to read per-pkg counter\n");
238 if (!evsel->snapshot)
239 perf_evsel__compute_deltas(evsel, cpu, thread, count);
240 perf_counts_values__scale(count, scale, NULL);
241 if (aggr_mode == AGGR_NONE)
242 perf_stat__update_shadow_stats(evsel, count->values, cpu);
245 aggr->val += count->val;
247 aggr->ena += count->ena;
248 aggr->run += count->run;
257 static int process_counter_maps(struct perf_evsel *counter)
259 int nthreads = thread_map__nr(counter->threads);
260 int ncpus = perf_evsel__nr_cpus(counter);
263 if (counter->system_wide)
266 for (thread = 0; thread < nthreads; thread++) {
267 for (cpu = 0; cpu < ncpus; cpu++) {
268 if (process_counter_values(counter, cpu, thread,
269 perf_counts(counter->counts, cpu, thread)))
277 static int process_counter(struct perf_evsel *counter)
279 struct perf_counts_values *aggr = &counter->counts->aggr;
280 struct perf_stat *ps = counter->priv;
281 u64 *count = counter->counts->aggr.values;
284 aggr->val = aggr->ena = aggr->run = 0;
285 init_stats(ps->res_stats);
287 if (counter->per_pkg)
288 zero_per_pkg(counter);
290 ret = process_counter_maps(counter);
294 if (aggr_mode != AGGR_GLOBAL)
297 if (!counter->snapshot)
298 perf_evsel__compute_deltas(counter, -1, -1, aggr);
299 perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
301 for (i = 0; i < 3; i++)
302 update_stats(&ps->res_stats[i], count[i]);
305 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
306 perf_evsel__name(counter), count[0], count[1], count[2]);
310 * Save the full runtime - to allow normalization during printout:
312 perf_stat__update_shadow_stats(counter, count, 0);
318 * Read out the results of a single counter:
319 * do not aggregate counts across CPUs in system-wide mode
321 static int read_counter(struct perf_evsel *counter)
323 int nthreads = thread_map__nr(evsel_list->threads);
324 int ncpus = perf_evsel__nr_cpus(counter);
327 if (!counter->supported)
330 if (counter->system_wide)
333 for (thread = 0; thread < nthreads; thread++) {
334 for (cpu = 0; cpu < ncpus; cpu++) {
335 struct perf_counts_values *count;
337 count = perf_counts(counter->counts, cpu, thread);
338 if (perf_evsel__read(counter, cpu, thread, count))
346 static void read_counters(bool close)
348 struct perf_evsel *counter;
350 evlist__for_each(evsel_list, counter) {
351 if (read_counter(counter))
352 pr_warning("failed to read counter %s\n", counter->name);
354 if (process_counter(counter))
355 pr_warning("failed to process counter %s\n", counter->name);
358 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
359 thread_map__nr(evsel_list->threads));
364 static void process_interval(void)
366 struct timespec ts, rs;
368 read_counters(false);
370 clock_gettime(CLOCK_MONOTONIC, &ts);
371 diff_timespec(&rs, &ts, &ref_time);
373 print_counters(&rs, 0, NULL);
376 static void handle_initial_delay(void)
378 struct perf_evsel *counter;
381 const int ncpus = cpu_map__nr(evsel_list->cpus),
382 nthreads = thread_map__nr(evsel_list->threads);
384 usleep(initial_delay * 1000);
385 evlist__for_each(evsel_list, counter)
386 perf_evsel__enable(counter, ncpus, nthreads);
390 static volatile int workload_exec_errno;
393 * perf_evlist__prepare_workload will send a SIGUSR1
394 * if the fork fails, since we asked by setting its
395 * want_signal to true.
397 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
398 void *ucontext __maybe_unused)
400 workload_exec_errno = info->si_value.sival_int;
403 static int __run_perf_stat(int argc, const char **argv)
406 unsigned long long t0, t1;
407 struct perf_evsel *counter;
411 const bool forks = (argc > 0);
414 ts.tv_sec = interval / 1000;
415 ts.tv_nsec = (interval % 1000) * 1000000;
422 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
423 workload_exec_failed_signal) < 0) {
424 perror("failed to prepare workload");
427 child_pid = evsel_list->workload.pid;
431 perf_evlist__set_leader(evsel_list);
433 evlist__for_each(evsel_list, counter) {
434 if (create_perf_stat_counter(counter) < 0) {
436 * PPC returns ENXIO for HW counters until 2.6.37
437 * (behavior changed with commit b0a873e).
439 if (errno == EINVAL || errno == ENOSYS ||
440 errno == ENOENT || errno == EOPNOTSUPP ||
443 ui__warning("%s event is not supported by the kernel.\n",
444 perf_evsel__name(counter));
445 counter->supported = false;
447 if ((counter->leader != counter) ||
448 !(counter->leader->nr_members > 1))
452 perf_evsel__open_strerror(counter, &target,
453 errno, msg, sizeof(msg));
454 ui__error("%s\n", msg);
457 kill(child_pid, SIGTERM);
461 counter->supported = true;
463 l = strlen(counter->unit);
468 if (perf_evlist__apply_filters(evsel_list, &counter)) {
469 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
470 counter->filter, perf_evsel__name(counter), errno,
471 strerror_r(errno, msg, sizeof(msg)));
476 * Enable counters and exec the command:
479 clock_gettime(CLOCK_MONOTONIC, &ref_time);
482 perf_evlist__start_workload(evsel_list);
483 handle_initial_delay();
486 while (!waitpid(child_pid, &status, WNOHANG)) {
487 nanosleep(&ts, NULL);
493 if (workload_exec_errno) {
494 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
495 pr_err("Workload failed: %s\n", emsg);
499 if (WIFSIGNALED(status))
500 psignal(WTERMSIG(status), argv[0]);
502 handle_initial_delay();
504 nanosleep(&ts, NULL);
512 update_stats(&walltime_nsecs_stats, t1 - t0);
516 return WEXITSTATUS(status);
519 static int run_perf_stat(int argc, const char **argv)
524 ret = system(pre_cmd);
532 ret = __run_perf_stat(argc, argv);
537 ret = system(post_cmd);
545 static void print_running(u64 run, u64 ena)
548 fprintf(output, "%s%" PRIu64 "%s%.2f",
552 ena ? 100.0 * run / ena : 100.0);
553 } else if (run != ena) {
554 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
558 static void print_noise_pct(double total, double avg)
560 double pct = rel_stddev_stats(total, avg);
563 fprintf(output, "%s%.2f%%", csv_sep, pct);
565 fprintf(output, " ( +-%6.2f%% )", pct);
568 static void print_noise(struct perf_evsel *evsel, double avg)
570 struct perf_stat *ps;
576 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
579 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
583 fprintf(output, "S%d-C%*d%s%*d%s",
584 cpu_map__id_to_socket(id),
586 cpu_map__id_to_cpu(id),
593 fprintf(output, "S%*d%s%*d%s",
602 fprintf(output, "CPU%*d%s",
604 perf_evsel__cpus(evsel)->map[id], csv_sep);
607 fprintf(output, "%*s-%*d%s",
609 thread_map__comm(evsel->threads, id),
611 thread_map__pid(evsel->threads, id),
620 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
622 double msecs = avg / 1e6;
623 const char *fmt_v, *fmt_n;
626 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
627 fmt_n = csv_output ? "%s" : "%-25s";
629 aggr_printout(evsel, id, nr);
631 scnprintf(name, sizeof(name), "%s%s",
632 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
634 fprintf(output, fmt_v, msecs, csv_sep);
637 fprintf(output, "%s%s", evsel->unit, csv_sep);
639 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
641 fprintf(output, fmt_n, name);
644 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
646 if (csv_output || interval)
649 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
650 fprintf(output, " # %8.3f CPUs utilized ",
651 avg / avg_stats(&walltime_nsecs_stats));
653 fprintf(output, " ");
656 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
658 double sc = evsel->scale;
660 int cpu = cpu_map__id_to_cpu(id);
663 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
666 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
668 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
671 aggr_printout(evsel, id, nr);
673 if (aggr_mode == AGGR_GLOBAL)
676 fprintf(output, fmt, avg, csv_sep);
679 fprintf(output, "%-*s%s",
680 csv_output ? 0 : unit_width,
681 evsel->unit, csv_sep);
683 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
686 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
688 if (csv_output || interval)
691 perf_stat__print_shadow_stats(output, evsel, avg, cpu, aggr_mode);
694 static void print_aggr(char *prefix)
696 struct perf_evsel *counter;
697 int cpu, cpu2, s, s2, id, nr;
701 if (!(aggr_map || aggr_get_id))
704 for (s = 0; s < aggr_map->nr; s++) {
705 id = aggr_map->map[s];
706 evlist__for_each(evsel_list, counter) {
709 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
710 cpu2 = perf_evsel__cpus(counter)->map[cpu];
711 s2 = aggr_get_id(evsel_list->cpus, cpu2);
714 val += perf_counts(counter->counts, cpu, 0)->val;
715 ena += perf_counts(counter->counts, cpu, 0)->ena;
716 run += perf_counts(counter->counts, cpu, 0)->run;
720 fprintf(output, "%s", prefix);
722 if (run == 0 || ena == 0) {
723 aggr_printout(counter, id, nr);
725 fprintf(output, "%*s%s",
727 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
730 fprintf(output, "%-*s%s",
731 csv_output ? 0 : unit_width,
732 counter->unit, csv_sep);
734 fprintf(output, "%*s",
735 csv_output ? 0 : -25,
736 perf_evsel__name(counter));
739 fprintf(output, "%s%s",
740 csv_sep, counter->cgrp->name);
742 print_running(run, ena);
746 uval = val * counter->scale;
748 if (nsec_counter(counter))
749 nsec_printout(id, nr, counter, uval);
751 abs_printout(id, nr, counter, uval);
754 print_noise(counter, 1.0);
756 print_running(run, ena);
762 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
764 int nthreads = thread_map__nr(counter->threads);
765 int ncpus = cpu_map__nr(counter->cpus);
769 for (thread = 0; thread < nthreads; thread++) {
770 u64 ena = 0, run = 0, val = 0;
772 for (cpu = 0; cpu < ncpus; cpu++) {
773 val += perf_counts(counter->counts, cpu, thread)->val;
774 ena += perf_counts(counter->counts, cpu, thread)->ena;
775 run += perf_counts(counter->counts, cpu, thread)->run;
779 fprintf(output, "%s", prefix);
781 uval = val * counter->scale;
783 if (nsec_counter(counter))
784 nsec_printout(thread, 0, counter, uval);
786 abs_printout(thread, 0, counter, uval);
789 print_noise(counter, 1.0);
791 print_running(run, ena);
797 * Print out the results of a single counter:
798 * aggregated counts in system-wide mode
800 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
802 struct perf_stat *ps = counter->priv;
803 double avg = avg_stats(&ps->res_stats[0]);
804 int scaled = counter->counts->scaled;
806 double avg_enabled, avg_running;
808 avg_enabled = avg_stats(&ps->res_stats[1]);
809 avg_running = avg_stats(&ps->res_stats[2]);
812 fprintf(output, "%s", prefix);
814 if (scaled == -1 || !counter->supported) {
815 fprintf(output, "%*s%s",
817 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
819 fprintf(output, "%-*s%s",
820 csv_output ? 0 : unit_width,
821 counter->unit, csv_sep);
822 fprintf(output, "%*s",
823 csv_output ? 0 : -25,
824 perf_evsel__name(counter));
827 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
829 print_running(avg_running, avg_enabled);
834 uval = avg * counter->scale;
836 if (nsec_counter(counter))
837 nsec_printout(-1, 0, counter, uval);
839 abs_printout(-1, 0, counter, uval);
841 print_noise(counter, avg);
843 print_running(avg_running, avg_enabled);
844 fprintf(output, "\n");
848 * Print out the results of a single counter:
849 * does not use aggregated count in system-wide
851 static void print_counter(struct perf_evsel *counter, char *prefix)
857 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
858 val = perf_counts(counter->counts, cpu, 0)->val;
859 ena = perf_counts(counter->counts, cpu, 0)->ena;
860 run = perf_counts(counter->counts, cpu, 0)->run;
863 fprintf(output, "%s", prefix);
865 if (run == 0 || ena == 0) {
866 fprintf(output, "CPU%*d%s%*s%s",
868 perf_evsel__cpus(counter)->map[cpu], csv_sep,
870 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
873 fprintf(output, "%-*s%s",
874 csv_output ? 0 : unit_width,
875 counter->unit, csv_sep);
877 fprintf(output, "%*s",
878 csv_output ? 0 : -25,
879 perf_evsel__name(counter));
882 fprintf(output, "%s%s",
883 csv_sep, counter->cgrp->name);
885 print_running(run, ena);
890 uval = val * counter->scale;
892 if (nsec_counter(counter))
893 nsec_printout(cpu, 0, counter, uval);
895 abs_printout(cpu, 0, counter, uval);
898 print_noise(counter, 1.0);
899 print_running(run, ena);
905 static void print_interval(char *prefix, struct timespec *ts)
907 static int num_print_interval;
909 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
911 if (num_print_interval == 0 && !csv_output) {
914 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
917 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
920 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
923 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
927 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
931 if (++num_print_interval == 25)
932 num_print_interval = 0;
935 static void print_header(int argc, const char **argv)
942 fprintf(output, "\n");
943 fprintf(output, " Performance counter stats for ");
944 if (target.system_wide)
945 fprintf(output, "\'system wide");
946 else if (target.cpu_list)
947 fprintf(output, "\'CPU(s) %s", target.cpu_list);
948 else if (!target__has_task(&target)) {
949 fprintf(output, "\'%s", argv[0]);
950 for (i = 1; i < argc; i++)
951 fprintf(output, " %s", argv[i]);
952 } else if (target.pid)
953 fprintf(output, "process id \'%s", target.pid);
955 fprintf(output, "thread id \'%s", target.tid);
957 fprintf(output, "\'");
959 fprintf(output, " (%d runs)", run_count);
960 fprintf(output, ":\n\n");
964 static void print_footer(void)
967 fprintf(output, "\n");
968 fprintf(output, " %17.9f seconds time elapsed",
969 avg_stats(&walltime_nsecs_stats)/1e9);
971 fprintf(output, " ");
972 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
973 avg_stats(&walltime_nsecs_stats));
975 fprintf(output, "\n\n");
978 static void print_counters(struct timespec *ts, int argc, const char **argv)
980 struct perf_evsel *counter;
981 char buf[64], *prefix = NULL;
984 print_interval(prefix = buf, ts);
986 print_header(argc, argv);
994 evlist__for_each(evsel_list, counter)
995 print_aggr_thread(counter, prefix);
998 evlist__for_each(evsel_list, counter)
999 print_counter_aggr(counter, prefix);
1002 evlist__for_each(evsel_list, counter)
1003 print_counter(counter, prefix);
1009 if (!interval && !csv_output)
1015 static volatile int signr = -1;
1017 static void skip_signal(int signo)
1019 if ((child_pid == -1) || interval)
1024 * render child_pid harmless
1025 * won't send SIGTERM to a random
1026 * process in case of race condition
1027 * and fast PID recycling
1032 static void sig_atexit(void)
1037 * avoid race condition with SIGCHLD handler
1038 * in skip_signal() which is modifying child_pid
1039 * goal is to avoid send SIGTERM to a random
1043 sigaddset(&set, SIGCHLD);
1044 sigprocmask(SIG_BLOCK, &set, &oset);
1046 if (child_pid != -1)
1047 kill(child_pid, SIGTERM);
1049 sigprocmask(SIG_SETMASK, &oset, NULL);
1054 signal(signr, SIG_DFL);
1055 kill(getpid(), signr);
1058 static int stat__set_big_num(const struct option *opt __maybe_unused,
1059 const char *s __maybe_unused, int unset)
1061 big_num_opt = unset ? 0 : 1;
1065 static int perf_stat_init_aggr_mode(void)
1067 switch (aggr_mode) {
1069 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1070 perror("cannot build socket map");
1073 aggr_get_id = cpu_map__get_socket;
1076 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1077 perror("cannot build core map");
1080 aggr_get_id = cpu_map__get_core;
1092 * Add default attributes, if there were no attributes specified or
1093 * if -d/--detailed, -d -d or -d -d -d is used:
1095 static int add_default_attributes(void)
1097 struct perf_event_attr default_attrs[] = {
1099 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1100 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1101 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1102 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1104 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1105 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1106 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1107 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1108 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1109 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1114 * Detailed stats (-d), covering the L1 and last level data caches:
1116 struct perf_event_attr detailed_attrs[] = {
1118 { .type = PERF_TYPE_HW_CACHE,
1120 PERF_COUNT_HW_CACHE_L1D << 0 |
1121 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1122 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1124 { .type = PERF_TYPE_HW_CACHE,
1126 PERF_COUNT_HW_CACHE_L1D << 0 |
1127 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1128 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1130 { .type = PERF_TYPE_HW_CACHE,
1132 PERF_COUNT_HW_CACHE_LL << 0 |
1133 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1134 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1136 { .type = PERF_TYPE_HW_CACHE,
1138 PERF_COUNT_HW_CACHE_LL << 0 |
1139 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1140 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1144 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1146 struct perf_event_attr very_detailed_attrs[] = {
1148 { .type = PERF_TYPE_HW_CACHE,
1150 PERF_COUNT_HW_CACHE_L1I << 0 |
1151 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1152 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1154 { .type = PERF_TYPE_HW_CACHE,
1156 PERF_COUNT_HW_CACHE_L1I << 0 |
1157 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1158 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1160 { .type = PERF_TYPE_HW_CACHE,
1162 PERF_COUNT_HW_CACHE_DTLB << 0 |
1163 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1164 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1166 { .type = PERF_TYPE_HW_CACHE,
1168 PERF_COUNT_HW_CACHE_DTLB << 0 |
1169 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1170 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1172 { .type = PERF_TYPE_HW_CACHE,
1174 PERF_COUNT_HW_CACHE_ITLB << 0 |
1175 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1176 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1178 { .type = PERF_TYPE_HW_CACHE,
1180 PERF_COUNT_HW_CACHE_ITLB << 0 |
1181 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1182 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1187 * Very, very detailed stats (-d -d -d), adding prefetch events:
1189 struct perf_event_attr very_very_detailed_attrs[] = {
1191 { .type = PERF_TYPE_HW_CACHE,
1193 PERF_COUNT_HW_CACHE_L1D << 0 |
1194 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1195 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1197 { .type = PERF_TYPE_HW_CACHE,
1199 PERF_COUNT_HW_CACHE_L1D << 0 |
1200 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1201 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1204 /* Set attrs if no event is selected and !null_run: */
1208 if (transaction_run) {
1210 if (pmu_have_event("cpu", "cycles-ct") &&
1211 pmu_have_event("cpu", "el-start"))
1212 err = parse_events(evsel_list, transaction_attrs, NULL);
1214 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1216 fprintf(stderr, "Cannot set up transaction events\n");
1222 if (!evsel_list->nr_entries) {
1223 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1227 /* Detailed events get appended to the event list: */
1229 if (detailed_run < 1)
1232 /* Append detailed run extra attributes: */
1233 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1236 if (detailed_run < 2)
1239 /* Append very detailed run extra attributes: */
1240 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1243 if (detailed_run < 3)
1246 /* Append very, very detailed run extra attributes: */
1247 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1250 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1252 bool append_file = false;
1254 const char *output_name = NULL;
1255 const struct option options[] = {
1256 OPT_BOOLEAN('T', "transaction", &transaction_run,
1257 "hardware transaction statistics"),
1258 OPT_CALLBACK('e', "event", &evsel_list, "event",
1259 "event selector. use 'perf list' to list available events",
1260 parse_events_option),
1261 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1262 "event filter", parse_filter),
1263 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1264 "child tasks do not inherit counters"),
1265 OPT_STRING('p', "pid", &target.pid, "pid",
1266 "stat events on existing process id"),
1267 OPT_STRING('t', "tid", &target.tid, "tid",
1268 "stat events on existing thread id"),
1269 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1270 "system-wide collection from all CPUs"),
1271 OPT_BOOLEAN('g', "group", &group,
1272 "put the counters into a counter group"),
1273 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1274 OPT_INCR('v', "verbose", &verbose,
1275 "be more verbose (show counter open errors, etc)"),
1276 OPT_INTEGER('r', "repeat", &run_count,
1277 "repeat command and print average + stddev (max: 100, forever: 0)"),
1278 OPT_BOOLEAN('n', "null", &null_run,
1279 "null run - dont start any counters"),
1280 OPT_INCR('d', "detailed", &detailed_run,
1281 "detailed run - start a lot of events"),
1282 OPT_BOOLEAN('S', "sync", &sync_run,
1283 "call sync() before starting a run"),
1284 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1285 "print large numbers with thousands\' separators",
1287 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1288 "list of cpus to monitor in system-wide"),
1289 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1290 "disable CPU count aggregation", AGGR_NONE),
1291 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1292 "print counts with custom separator"),
1293 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1294 "monitor event in cgroup name only", parse_cgroups),
1295 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1296 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1297 OPT_INTEGER(0, "log-fd", &output_fd,
1298 "log output to fd, instead of stderr"),
1299 OPT_STRING(0, "pre", &pre_cmd, "command",
1300 "command to run prior to the measured command"),
1301 OPT_STRING(0, "post", &post_cmd, "command",
1302 "command to run after to the measured command"),
1303 OPT_UINTEGER('I', "interval-print", &interval,
1304 "print counts at regular interval in ms (>= 100)"),
1305 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1306 "aggregate counts per processor socket", AGGR_SOCKET),
1307 OPT_SET_UINT(0, "per-core", &aggr_mode,
1308 "aggregate counts per physical processor core", AGGR_CORE),
1309 OPT_SET_UINT(0, "per-thread", &aggr_mode,
1310 "aggregate counts per thread", AGGR_THREAD),
1311 OPT_UINTEGER('D', "delay", &initial_delay,
1312 "ms to wait before starting measurement after program start"),
1315 const char * const stat_usage[] = {
1316 "perf stat [<options>] [<command>]",
1319 int status = -EINVAL, run_idx;
1322 setlocale(LC_ALL, "");
1324 evsel_list = perf_evlist__new();
1325 if (evsel_list == NULL)
1328 argc = parse_options(argc, argv, options, stat_usage,
1329 PARSE_OPT_STOP_AT_NON_OPTION);
1332 if (output_name && strcmp(output_name, "-"))
1335 if (output_name && output_fd) {
1336 fprintf(stderr, "cannot use both --output and --log-fd\n");
1337 parse_options_usage(stat_usage, options, "o", 1);
1338 parse_options_usage(NULL, options, "log-fd", 0);
1342 if (output_fd < 0) {
1343 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1344 parse_options_usage(stat_usage, options, "log-fd", 0);
1350 mode = append_file ? "a" : "w";
1352 output = fopen(output_name, mode);
1354 perror("failed to create output file");
1357 clock_gettime(CLOCK_REALTIME, &tm);
1358 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1359 } else if (output_fd > 0) {
1360 mode = append_file ? "a" : "w";
1361 output = fdopen(output_fd, mode);
1363 perror("Failed opening logfd");
1370 if (!strcmp(csv_sep, "\\t"))
1373 csv_sep = DEFAULT_SEPARATOR;
1376 * let the spreadsheet do the pretty-printing
1379 /* User explicitly passed -B? */
1380 if (big_num_opt == 1) {
1381 fprintf(stderr, "-B option not supported with -x\n");
1382 parse_options_usage(stat_usage, options, "B", 1);
1383 parse_options_usage(NULL, options, "x", 1);
1385 } else /* Nope, so disable big number formatting */
1387 } else if (big_num_opt == 0) /* User passed --no-big-num */
1390 if (!argc && target__none(&target))
1391 usage_with_options(stat_usage, options);
1393 if (run_count < 0) {
1394 pr_err("Run count must be a positive number\n");
1395 parse_options_usage(stat_usage, options, "r", 1);
1397 } else if (run_count == 0) {
1402 if ((aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1403 fprintf(stderr, "The --per-thread option is only available "
1404 "when monitoring via -p -t options.\n");
1405 parse_options_usage(NULL, options, "p", 1);
1406 parse_options_usage(NULL, options, "t", 1);
1411 * no_aggr, cgroup are for system-wide only
1412 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1414 if (((aggr_mode != AGGR_GLOBAL && aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1415 !target__has_cpu(&target)) {
1416 fprintf(stderr, "both cgroup and no-aggregation "
1417 "modes only available in system-wide mode\n");
1419 parse_options_usage(stat_usage, options, "G", 1);
1420 parse_options_usage(NULL, options, "A", 1);
1421 parse_options_usage(NULL, options, "a", 1);
1425 if (add_default_attributes())
1428 target__validate(&target);
1430 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1431 if (target__has_task(&target)) {
1432 pr_err("Problems finding threads of monitor\n");
1433 parse_options_usage(stat_usage, options, "p", 1);
1434 parse_options_usage(NULL, options, "t", 1);
1435 } else if (target__has_cpu(&target)) {
1436 perror("failed to parse CPUs map");
1437 parse_options_usage(stat_usage, options, "C", 1);
1438 parse_options_usage(NULL, options, "a", 1);
1444 * Initialize thread_map with comm names,
1445 * so we could print it out on output.
1447 if (aggr_mode == AGGR_THREAD)
1448 thread_map__read_comms(evsel_list->threads);
1450 if (interval && interval < 100) {
1451 pr_err("print interval must be >= 100ms\n");
1452 parse_options_usage(stat_usage, options, "I", 1);
1456 if (perf_evlist__alloc_stats(evsel_list, interval))
1459 if (perf_stat_init_aggr_mode())
1463 * We dont want to block the signals - that would cause
1464 * child tasks to inherit that and Ctrl-C would not work.
1465 * What we want is for Ctrl-C to work in the exec()-ed
1466 * task, but being ignored by perf stat itself:
1470 signal(SIGINT, skip_signal);
1471 signal(SIGCHLD, skip_signal);
1472 signal(SIGALRM, skip_signal);
1473 signal(SIGABRT, skip_signal);
1476 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1477 if (run_count != 1 && verbose)
1478 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1481 status = run_perf_stat(argc, argv);
1482 if (forever && status != -1) {
1483 print_counters(NULL, argc, argv);
1484 perf_stat__reset_stats();
1488 if (!forever && status != -1 && !interval)
1489 print_counters(NULL, argc, argv);
1491 perf_evlist__free_stats(evsel_list);
1493 perf_evlist__delete(evsel_list);