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"
61 #include "util/counts.h"
64 #include <sys/prctl.h>
67 #define DEFAULT_SEPARATOR " "
68 #define CNTR_NOT_SUPPORTED "<not supported>"
69 #define CNTR_NOT_COUNTED "<not counted>"
71 static void print_counters(struct timespec *ts, int argc, const char **argv);
73 /* Default events used for perf stat -T */
74 static const char *transaction_attrs = {
86 /* More limited version when the CPU does not have all events. */
87 static const char * transaction_limited_attrs = {
97 static struct perf_evlist *evsel_list;
99 static struct target target = {
103 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
105 static int run_count = 1;
106 static bool no_inherit = false;
107 static volatile pid_t child_pid = -1;
108 static bool null_run = false;
109 static int detailed_run = 0;
110 static bool transaction_run;
111 static bool big_num = true;
112 static int big_num_opt = -1;
113 static const char *csv_sep = NULL;
114 static bool csv_output = false;
115 static bool group = false;
116 static const char *pre_cmd = NULL;
117 static const char *post_cmd = NULL;
118 static bool sync_run = false;
119 static unsigned int initial_delay = 0;
120 static unsigned int unit_width = 4; /* strlen("unit") */
121 static bool forever = false;
122 static struct timespec ref_time;
123 static struct cpu_map *aggr_map;
124 static aggr_get_id_t aggr_get_id;
125 static bool append_file;
126 static const char *output_name;
127 static int output_fd;
129 static volatile int done = 0;
131 static struct perf_stat_config stat_config = {
132 .aggr_mode = AGGR_GLOBAL,
136 static inline void diff_timespec(struct timespec *r, struct timespec *a,
139 r->tv_sec = a->tv_sec - b->tv_sec;
140 if (a->tv_nsec < b->tv_nsec) {
141 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
144 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
148 static void perf_stat__reset_stats(void)
150 perf_evlist__reset_stats(evsel_list);
151 perf_stat__reset_shadow_stats();
154 static int create_perf_stat_counter(struct perf_evsel *evsel)
156 struct perf_event_attr *attr = &evsel->attr;
158 if (stat_config.scale)
159 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
160 PERF_FORMAT_TOTAL_TIME_RUNNING;
162 attr->inherit = !no_inherit;
164 if (target__has_cpu(&target))
165 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
167 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
170 attr->enable_on_exec = 1;
173 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
177 * Does the counter have nsecs as a unit?
179 static inline int nsec_counter(struct perf_evsel *evsel)
181 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
182 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
189 * Read out the results of a single counter:
190 * do not aggregate counts across CPUs in system-wide mode
192 static int read_counter(struct perf_evsel *counter)
194 int nthreads = thread_map__nr(evsel_list->threads);
195 int ncpus = perf_evsel__nr_cpus(counter);
198 if (!counter->supported)
201 if (counter->system_wide)
204 for (thread = 0; thread < nthreads; thread++) {
205 for (cpu = 0; cpu < ncpus; cpu++) {
206 struct perf_counts_values *count;
208 count = perf_counts(counter->counts, cpu, thread);
209 if (perf_evsel__read(counter, cpu, thread, count))
217 static void read_counters(bool close_counters)
219 struct perf_evsel *counter;
221 evlist__for_each(evsel_list, counter) {
222 if (read_counter(counter))
223 pr_debug("failed to read counter %s\n", counter->name);
225 if (perf_stat_process_counter(&stat_config, counter))
226 pr_warning("failed to process counter %s\n", counter->name);
228 if (close_counters) {
229 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
230 thread_map__nr(evsel_list->threads));
235 static void process_interval(void)
237 struct timespec ts, rs;
239 read_counters(false);
241 clock_gettime(CLOCK_MONOTONIC, &ts);
242 diff_timespec(&rs, &ts, &ref_time);
244 print_counters(&rs, 0, NULL);
247 static void handle_initial_delay(void)
249 struct perf_evsel *counter;
252 const int ncpus = cpu_map__nr(evsel_list->cpus),
253 nthreads = thread_map__nr(evsel_list->threads);
255 usleep(initial_delay * 1000);
256 evlist__for_each(evsel_list, counter)
257 perf_evsel__enable(counter, ncpus, nthreads);
261 static volatile int workload_exec_errno;
264 * perf_evlist__prepare_workload will send a SIGUSR1
265 * if the fork fails, since we asked by setting its
266 * want_signal to true.
268 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
269 void *ucontext __maybe_unused)
271 workload_exec_errno = info->si_value.sival_int;
274 static int __run_perf_stat(int argc, const char **argv)
276 int interval = stat_config.interval;
278 unsigned long long t0, t1;
279 struct perf_evsel *counter;
283 const bool forks = (argc > 0);
286 ts.tv_sec = interval / 1000;
287 ts.tv_nsec = (interval % 1000) * 1000000;
294 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
295 workload_exec_failed_signal) < 0) {
296 perror("failed to prepare workload");
299 child_pid = evsel_list->workload.pid;
303 perf_evlist__set_leader(evsel_list);
305 evlist__for_each(evsel_list, counter) {
306 if (create_perf_stat_counter(counter) < 0) {
308 * PPC returns ENXIO for HW counters until 2.6.37
309 * (behavior changed with commit b0a873e).
311 if (errno == EINVAL || errno == ENOSYS ||
312 errno == ENOENT || errno == EOPNOTSUPP ||
315 ui__warning("%s event is not supported by the kernel.\n",
316 perf_evsel__name(counter));
317 counter->supported = false;
319 if ((counter->leader != counter) ||
320 !(counter->leader->nr_members > 1))
324 perf_evsel__open_strerror(counter, &target,
325 errno, msg, sizeof(msg));
326 ui__error("%s\n", msg);
329 kill(child_pid, SIGTERM);
333 counter->supported = true;
335 l = strlen(counter->unit);
340 if (perf_evlist__apply_filters(evsel_list, &counter)) {
341 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
342 counter->filter, perf_evsel__name(counter), errno,
343 strerror_r(errno, msg, sizeof(msg)));
348 * Enable counters and exec the command:
351 clock_gettime(CLOCK_MONOTONIC, &ref_time);
354 perf_evlist__start_workload(evsel_list);
355 handle_initial_delay();
358 while (!waitpid(child_pid, &status, WNOHANG)) {
359 nanosleep(&ts, NULL);
365 if (workload_exec_errno) {
366 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
367 pr_err("Workload failed: %s\n", emsg);
371 if (WIFSIGNALED(status))
372 psignal(WTERMSIG(status), argv[0]);
374 handle_initial_delay();
376 nanosleep(&ts, NULL);
384 update_stats(&walltime_nsecs_stats, t1 - t0);
388 return WEXITSTATUS(status);
391 static int run_perf_stat(int argc, const char **argv)
396 ret = system(pre_cmd);
404 ret = __run_perf_stat(argc, argv);
409 ret = system(post_cmd);
417 static void print_running(u64 run, u64 ena)
420 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
424 ena ? 100.0 * run / ena : 100.0);
425 } else if (run != ena) {
426 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
430 static void print_noise_pct(double total, double avg)
432 double pct = rel_stddev_stats(total, avg);
435 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
437 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
440 static void print_noise(struct perf_evsel *evsel, double avg)
442 struct perf_stat_evsel *ps;
448 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
451 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
453 switch (stat_config.aggr_mode) {
455 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
456 cpu_map__id_to_socket(id),
458 cpu_map__id_to_cpu(id),
465 fprintf(stat_config.output, "S%*d%s%*d%s",
474 fprintf(stat_config.output, "CPU%*d%s",
476 perf_evsel__cpus(evsel)->map[id], csv_sep);
479 fprintf(stat_config.output, "%*s-%*d%s",
481 thread_map__comm(evsel->threads, id),
483 thread_map__pid(evsel->threads, id),
493 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
495 FILE *output = stat_config.output;
496 double msecs = avg / 1e6;
497 const char *fmt_v, *fmt_n;
500 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
501 fmt_n = csv_output ? "%s" : "%-25s";
503 aggr_printout(evsel, id, nr);
505 scnprintf(name, sizeof(name), "%s%s",
506 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
508 fprintf(output, fmt_v, msecs, csv_sep);
511 fprintf(output, "%s%s", evsel->unit, csv_sep);
513 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
515 fprintf(output, fmt_n, name);
518 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
521 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
523 FILE *output = stat_config.output;
524 double sc = evsel->scale;
528 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
531 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
533 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
536 aggr_printout(evsel, id, nr);
538 fprintf(output, fmt, avg, csv_sep);
541 fprintf(output, "%-*s%s",
542 csv_output ? 0 : unit_width,
543 evsel->unit, csv_sep);
545 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
548 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
551 static void printout(int id, int nr, struct perf_evsel *counter, double uval)
553 int cpu = cpu_map__id_to_cpu(id);
555 if (stat_config.aggr_mode == AGGR_GLOBAL)
558 if (nsec_counter(counter))
559 nsec_printout(id, nr, counter, uval);
561 abs_printout(id, nr, counter, uval);
563 if (!csv_output && !stat_config.interval)
564 perf_stat__print_shadow_stats(stat_config.output, counter,
566 stat_config.aggr_mode);
569 static void print_aggr(char *prefix)
571 FILE *output = stat_config.output;
572 struct perf_evsel *counter;
573 int cpu, s, s2, id, nr;
577 if (!(aggr_map || aggr_get_id))
580 for (s = 0; s < aggr_map->nr; s++) {
581 id = aggr_map->map[s];
582 evlist__for_each(evsel_list, counter) {
585 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
586 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
589 val += perf_counts(counter->counts, cpu, 0)->val;
590 ena += perf_counts(counter->counts, cpu, 0)->ena;
591 run += perf_counts(counter->counts, cpu, 0)->run;
595 fprintf(output, "%s", prefix);
597 if (run == 0 || ena == 0) {
598 aggr_printout(counter, id, nr);
600 fprintf(output, "%*s%s",
602 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
605 fprintf(output, "%-*s%s",
606 csv_output ? 0 : unit_width,
607 counter->unit, csv_sep);
609 fprintf(output, "%*s",
610 csv_output ? 0 : -25,
611 perf_evsel__name(counter));
614 fprintf(output, "%s%s",
615 csv_sep, counter->cgrp->name);
617 print_running(run, ena);
621 uval = val * counter->scale;
622 printout(id, nr, counter, uval);
624 print_noise(counter, 1.0);
626 print_running(run, ena);
632 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
634 FILE *output = stat_config.output;
635 int nthreads = thread_map__nr(counter->threads);
636 int ncpus = cpu_map__nr(counter->cpus);
640 for (thread = 0; thread < nthreads; thread++) {
641 u64 ena = 0, run = 0, val = 0;
643 for (cpu = 0; cpu < ncpus; cpu++) {
644 val += perf_counts(counter->counts, cpu, thread)->val;
645 ena += perf_counts(counter->counts, cpu, thread)->ena;
646 run += perf_counts(counter->counts, cpu, thread)->run;
650 fprintf(output, "%s", prefix);
652 uval = val * counter->scale;
653 printout(thread, 0, counter, uval);
656 print_noise(counter, 1.0);
658 print_running(run, ena);
664 * Print out the results of a single counter:
665 * aggregated counts in system-wide mode
667 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
669 FILE *output = stat_config.output;
670 struct perf_stat_evsel *ps = counter->priv;
671 double avg = avg_stats(&ps->res_stats[0]);
672 int scaled = counter->counts->scaled;
674 double avg_enabled, avg_running;
676 avg_enabled = avg_stats(&ps->res_stats[1]);
677 avg_running = avg_stats(&ps->res_stats[2]);
680 fprintf(output, "%s", prefix);
682 if (scaled == -1 || !counter->supported) {
683 fprintf(output, "%*s%s",
685 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
687 fprintf(output, "%-*s%s",
688 csv_output ? 0 : unit_width,
689 counter->unit, csv_sep);
690 fprintf(output, "%*s",
691 csv_output ? 0 : -25,
692 perf_evsel__name(counter));
695 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
697 print_running(avg_running, avg_enabled);
702 uval = avg * counter->scale;
703 printout(-1, 0, counter, uval);
705 print_noise(counter, avg);
707 print_running(avg_running, avg_enabled);
708 fprintf(output, "\n");
712 * Print out the results of a single counter:
713 * does not use aggregated count in system-wide
715 static void print_counter(struct perf_evsel *counter, char *prefix)
717 FILE *output = stat_config.output;
722 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
723 val = perf_counts(counter->counts, cpu, 0)->val;
724 ena = perf_counts(counter->counts, cpu, 0)->ena;
725 run = perf_counts(counter->counts, cpu, 0)->run;
728 fprintf(output, "%s", prefix);
730 if (run == 0 || ena == 0) {
731 fprintf(output, "CPU%*d%s%*s%s",
733 perf_evsel__cpus(counter)->map[cpu], csv_sep,
735 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
738 fprintf(output, "%-*s%s",
739 csv_output ? 0 : unit_width,
740 counter->unit, csv_sep);
742 fprintf(output, "%*s",
743 csv_output ? 0 : -25,
744 perf_evsel__name(counter));
747 fprintf(output, "%s%s",
748 csv_sep, counter->cgrp->name);
750 print_running(run, ena);
755 uval = val * counter->scale;
756 printout(cpu, 0, counter, uval);
758 print_noise(counter, 1.0);
759 print_running(run, ena);
765 static void print_interval(char *prefix, struct timespec *ts)
767 FILE *output = stat_config.output;
768 static int num_print_interval;
770 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
772 if (num_print_interval == 0 && !csv_output) {
773 switch (stat_config.aggr_mode) {
775 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
778 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
781 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
784 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
788 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
794 if (++num_print_interval == 25)
795 num_print_interval = 0;
798 static void print_header(int argc, const char **argv)
800 FILE *output = stat_config.output;
806 fprintf(output, "\n");
807 fprintf(output, " Performance counter stats for ");
808 if (target.system_wide)
809 fprintf(output, "\'system wide");
810 else if (target.cpu_list)
811 fprintf(output, "\'CPU(s) %s", target.cpu_list);
812 else if (!target__has_task(&target)) {
813 fprintf(output, "\'%s", argv[0]);
814 for (i = 1; i < argc; i++)
815 fprintf(output, " %s", argv[i]);
816 } else if (target.pid)
817 fprintf(output, "process id \'%s", target.pid);
819 fprintf(output, "thread id \'%s", target.tid);
821 fprintf(output, "\'");
823 fprintf(output, " (%d runs)", run_count);
824 fprintf(output, ":\n\n");
828 static void print_footer(void)
830 FILE *output = stat_config.output;
833 fprintf(output, "\n");
834 fprintf(output, " %17.9f seconds time elapsed",
835 avg_stats(&walltime_nsecs_stats)/1e9);
837 fprintf(output, " ");
838 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
839 avg_stats(&walltime_nsecs_stats));
841 fprintf(output, "\n\n");
844 static void print_counters(struct timespec *ts, int argc, const char **argv)
846 int interval = stat_config.interval;
847 struct perf_evsel *counter;
848 char buf[64], *prefix = NULL;
851 print_interval(prefix = buf, ts);
853 print_header(argc, argv);
855 switch (stat_config.aggr_mode) {
861 evlist__for_each(evsel_list, counter)
862 print_aggr_thread(counter, prefix);
865 evlist__for_each(evsel_list, counter)
866 print_counter_aggr(counter, prefix);
869 evlist__for_each(evsel_list, counter)
870 print_counter(counter, prefix);
877 if (!interval && !csv_output)
880 fflush(stat_config.output);
883 static volatile int signr = -1;
885 static void skip_signal(int signo)
887 if ((child_pid == -1) || stat_config.interval)
892 * render child_pid harmless
893 * won't send SIGTERM to a random
894 * process in case of race condition
895 * and fast PID recycling
900 static void sig_atexit(void)
905 * avoid race condition with SIGCHLD handler
906 * in skip_signal() which is modifying child_pid
907 * goal is to avoid send SIGTERM to a random
911 sigaddset(&set, SIGCHLD);
912 sigprocmask(SIG_BLOCK, &set, &oset);
915 kill(child_pid, SIGTERM);
917 sigprocmask(SIG_SETMASK, &oset, NULL);
922 signal(signr, SIG_DFL);
923 kill(getpid(), signr);
926 static int stat__set_big_num(const struct option *opt __maybe_unused,
927 const char *s __maybe_unused, int unset)
929 big_num_opt = unset ? 0 : 1;
933 static const struct option stat_options[] = {
934 OPT_BOOLEAN('T', "transaction", &transaction_run,
935 "hardware transaction statistics"),
936 OPT_CALLBACK('e', "event", &evsel_list, "event",
937 "event selector. use 'perf list' to list available events",
938 parse_events_option),
939 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
940 "event filter", parse_filter),
941 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
942 "child tasks do not inherit counters"),
943 OPT_STRING('p', "pid", &target.pid, "pid",
944 "stat events on existing process id"),
945 OPT_STRING('t', "tid", &target.tid, "tid",
946 "stat events on existing thread id"),
947 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
948 "system-wide collection from all CPUs"),
949 OPT_BOOLEAN('g', "group", &group,
950 "put the counters into a counter group"),
951 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
952 OPT_INCR('v', "verbose", &verbose,
953 "be more verbose (show counter open errors, etc)"),
954 OPT_INTEGER('r', "repeat", &run_count,
955 "repeat command and print average + stddev (max: 100, forever: 0)"),
956 OPT_BOOLEAN('n', "null", &null_run,
957 "null run - dont start any counters"),
958 OPT_INCR('d', "detailed", &detailed_run,
959 "detailed run - start a lot of events"),
960 OPT_BOOLEAN('S', "sync", &sync_run,
961 "call sync() before starting a run"),
962 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
963 "print large numbers with thousands\' separators",
965 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
966 "list of cpus to monitor in system-wide"),
967 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
968 "disable CPU count aggregation", AGGR_NONE),
969 OPT_STRING('x', "field-separator", &csv_sep, "separator",
970 "print counts with custom separator"),
971 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
972 "monitor event in cgroup name only", parse_cgroups),
973 OPT_STRING('o', "output", &output_name, "file", "output file name"),
974 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
975 OPT_INTEGER(0, "log-fd", &output_fd,
976 "log output to fd, instead of stderr"),
977 OPT_STRING(0, "pre", &pre_cmd, "command",
978 "command to run prior to the measured command"),
979 OPT_STRING(0, "post", &post_cmd, "command",
980 "command to run after to the measured command"),
981 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
982 "print counts at regular interval in ms (>= 10)"),
983 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
984 "aggregate counts per processor socket", AGGR_SOCKET),
985 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
986 "aggregate counts per physical processor core", AGGR_CORE),
987 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
988 "aggregate counts per thread", AGGR_THREAD),
989 OPT_UINTEGER('D', "delay", &initial_delay,
990 "ms to wait before starting measurement after program start"),
994 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
996 return cpu_map__get_socket(map, cpu, NULL);
999 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1001 return cpu_map__get_core(map, cpu, NULL);
1004 static int cpu_map__get_max(struct cpu_map *map)
1008 for (i = 0; i < map->nr; i++) {
1009 if (map->map[i] > max)
1016 static struct cpu_map *cpus_aggr_map;
1018 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1025 cpu = map->map[idx];
1027 if (cpus_aggr_map->map[cpu] == -1)
1028 cpus_aggr_map->map[cpu] = get_id(map, idx);
1030 return cpus_aggr_map->map[cpu];
1033 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1035 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1038 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1040 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1043 static int perf_stat_init_aggr_mode(void)
1047 switch (stat_config.aggr_mode) {
1049 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1050 perror("cannot build socket map");
1053 aggr_get_id = perf_stat__get_socket_cached;
1056 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1057 perror("cannot build core map");
1060 aggr_get_id = perf_stat__get_core_cached;
1071 * The evsel_list->cpus is the base we operate on,
1072 * taking the highest cpu number to be the size of
1073 * the aggregation translate cpumap.
1075 nr = cpu_map__get_max(evsel_list->cpus);
1076 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1077 return cpus_aggr_map ? 0 : -ENOMEM;
1081 * Add default attributes, if there were no attributes specified or
1082 * if -d/--detailed, -d -d or -d -d -d is used:
1084 static int add_default_attributes(void)
1086 struct perf_event_attr default_attrs[] = {
1088 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1089 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1090 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1091 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1093 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1094 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1095 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1096 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1097 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1098 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1103 * Detailed stats (-d), covering the L1 and last level data caches:
1105 struct perf_event_attr detailed_attrs[] = {
1107 { .type = PERF_TYPE_HW_CACHE,
1109 PERF_COUNT_HW_CACHE_L1D << 0 |
1110 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1111 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1113 { .type = PERF_TYPE_HW_CACHE,
1115 PERF_COUNT_HW_CACHE_L1D << 0 |
1116 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1117 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1119 { .type = PERF_TYPE_HW_CACHE,
1121 PERF_COUNT_HW_CACHE_LL << 0 |
1122 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1123 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1125 { .type = PERF_TYPE_HW_CACHE,
1127 PERF_COUNT_HW_CACHE_LL << 0 |
1128 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1129 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1133 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1135 struct perf_event_attr very_detailed_attrs[] = {
1137 { .type = PERF_TYPE_HW_CACHE,
1139 PERF_COUNT_HW_CACHE_L1I << 0 |
1140 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1141 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1143 { .type = PERF_TYPE_HW_CACHE,
1145 PERF_COUNT_HW_CACHE_L1I << 0 |
1146 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1147 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1149 { .type = PERF_TYPE_HW_CACHE,
1151 PERF_COUNT_HW_CACHE_DTLB << 0 |
1152 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1153 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1155 { .type = PERF_TYPE_HW_CACHE,
1157 PERF_COUNT_HW_CACHE_DTLB << 0 |
1158 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1159 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1161 { .type = PERF_TYPE_HW_CACHE,
1163 PERF_COUNT_HW_CACHE_ITLB << 0 |
1164 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1165 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1167 { .type = PERF_TYPE_HW_CACHE,
1169 PERF_COUNT_HW_CACHE_ITLB << 0 |
1170 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1171 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1176 * Very, very detailed stats (-d -d -d), adding prefetch events:
1178 struct perf_event_attr very_very_detailed_attrs[] = {
1180 { .type = PERF_TYPE_HW_CACHE,
1182 PERF_COUNT_HW_CACHE_L1D << 0 |
1183 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1184 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1186 { .type = PERF_TYPE_HW_CACHE,
1188 PERF_COUNT_HW_CACHE_L1D << 0 |
1189 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1190 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1193 /* Set attrs if no event is selected and !null_run: */
1197 if (transaction_run) {
1199 if (pmu_have_event("cpu", "cycles-ct") &&
1200 pmu_have_event("cpu", "el-start"))
1201 err = parse_events(evsel_list, transaction_attrs, NULL);
1203 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1205 fprintf(stderr, "Cannot set up transaction events\n");
1211 if (!evsel_list->nr_entries) {
1212 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1216 /* Detailed events get appended to the event list: */
1218 if (detailed_run < 1)
1221 /* Append detailed run extra attributes: */
1222 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1225 if (detailed_run < 2)
1228 /* Append very detailed run extra attributes: */
1229 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1232 if (detailed_run < 3)
1235 /* Append very, very detailed run extra attributes: */
1236 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1239 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1241 const char * const stat_usage[] = {
1242 "perf stat [<options>] [<command>]",
1245 int status = -EINVAL, run_idx;
1247 FILE *output = stderr;
1248 unsigned int interval;
1250 setlocale(LC_ALL, "");
1252 evsel_list = perf_evlist__new();
1253 if (evsel_list == NULL)
1256 argc = parse_options(argc, argv, stat_options, stat_usage,
1257 PARSE_OPT_STOP_AT_NON_OPTION);
1259 interval = stat_config.interval;
1261 if (output_name && strcmp(output_name, "-"))
1264 if (output_name && output_fd) {
1265 fprintf(stderr, "cannot use both --output and --log-fd\n");
1266 parse_options_usage(stat_usage, stat_options, "o", 1);
1267 parse_options_usage(NULL, stat_options, "log-fd", 0);
1271 if (output_fd < 0) {
1272 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1273 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1279 mode = append_file ? "a" : "w";
1281 output = fopen(output_name, mode);
1283 perror("failed to create output file");
1286 clock_gettime(CLOCK_REALTIME, &tm);
1287 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1288 } else if (output_fd > 0) {
1289 mode = append_file ? "a" : "w";
1290 output = fdopen(output_fd, mode);
1292 perror("Failed opening logfd");
1297 stat_config.output = output;
1301 if (!strcmp(csv_sep, "\\t"))
1304 csv_sep = DEFAULT_SEPARATOR;
1307 * let the spreadsheet do the pretty-printing
1310 /* User explicitly passed -B? */
1311 if (big_num_opt == 1) {
1312 fprintf(stderr, "-B option not supported with -x\n");
1313 parse_options_usage(stat_usage, stat_options, "B", 1);
1314 parse_options_usage(NULL, stat_options, "x", 1);
1316 } else /* Nope, so disable big number formatting */
1318 } else if (big_num_opt == 0) /* User passed --no-big-num */
1321 if (!argc && target__none(&target))
1322 usage_with_options(stat_usage, stat_options);
1324 if (run_count < 0) {
1325 pr_err("Run count must be a positive number\n");
1326 parse_options_usage(stat_usage, stat_options, "r", 1);
1328 } else if (run_count == 0) {
1333 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1334 fprintf(stderr, "The --per-thread option is only available "
1335 "when monitoring via -p -t options.\n");
1336 parse_options_usage(NULL, stat_options, "p", 1);
1337 parse_options_usage(NULL, stat_options, "t", 1);
1342 * no_aggr, cgroup are for system-wide only
1343 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1345 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1346 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1347 !target__has_cpu(&target)) {
1348 fprintf(stderr, "both cgroup and no-aggregation "
1349 "modes only available in system-wide mode\n");
1351 parse_options_usage(stat_usage, stat_options, "G", 1);
1352 parse_options_usage(NULL, stat_options, "A", 1);
1353 parse_options_usage(NULL, stat_options, "a", 1);
1357 if (add_default_attributes())
1360 target__validate(&target);
1362 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1363 if (target__has_task(&target)) {
1364 pr_err("Problems finding threads of monitor\n");
1365 parse_options_usage(stat_usage, stat_options, "p", 1);
1366 parse_options_usage(NULL, stat_options, "t", 1);
1367 } else if (target__has_cpu(&target)) {
1368 perror("failed to parse CPUs map");
1369 parse_options_usage(stat_usage, stat_options, "C", 1);
1370 parse_options_usage(NULL, stat_options, "a", 1);
1376 * Initialize thread_map with comm names,
1377 * so we could print it out on output.
1379 if (stat_config.aggr_mode == AGGR_THREAD)
1380 thread_map__read_comms(evsel_list->threads);
1382 if (interval && interval < 100) {
1383 if (interval < 10) {
1384 pr_err("print interval must be >= 10ms\n");
1385 parse_options_usage(stat_usage, stat_options, "I", 1);
1388 pr_warning("print interval < 100ms. "
1389 "The overhead percentage could be high in some cases. "
1390 "Please proceed with caution.\n");
1393 if (perf_evlist__alloc_stats(evsel_list, interval))
1396 if (perf_stat_init_aggr_mode())
1400 * We dont want to block the signals - that would cause
1401 * child tasks to inherit that and Ctrl-C would not work.
1402 * What we want is for Ctrl-C to work in the exec()-ed
1403 * task, but being ignored by perf stat itself:
1407 signal(SIGINT, skip_signal);
1408 signal(SIGCHLD, skip_signal);
1409 signal(SIGALRM, skip_signal);
1410 signal(SIGABRT, skip_signal);
1413 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1414 if (run_count != 1 && verbose)
1415 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1418 status = run_perf_stat(argc, argv);
1419 if (forever && status != -1) {
1420 print_counters(NULL, argc, argv);
1421 perf_stat__reset_stats();
1425 if (!forever && status != -1 && !interval)
1426 print_counters(NULL, argc, argv);
1428 perf_evlist__free_stats(evsel_list);
1430 perf_evlist__delete(evsel_list);