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;
126 static volatile int done = 0;
128 static struct perf_stat_config stat_config = {
129 .aggr_mode = AGGR_GLOBAL,
133 static inline void diff_timespec(struct timespec *r, struct timespec *a,
136 r->tv_sec = a->tv_sec - b->tv_sec;
137 if (a->tv_nsec < b->tv_nsec) {
138 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
141 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
145 static void perf_stat__reset_stats(void)
147 perf_evlist__reset_stats(evsel_list);
148 perf_stat__reset_shadow_stats();
151 static int create_perf_stat_counter(struct perf_evsel *evsel)
153 struct perf_event_attr *attr = &evsel->attr;
155 if (stat_config.scale)
156 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
157 PERF_FORMAT_TOTAL_TIME_RUNNING;
159 attr->inherit = !no_inherit;
161 if (target__has_cpu(&target))
162 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
164 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
167 attr->enable_on_exec = 1;
170 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
174 * Does the counter have nsecs as a unit?
176 static inline int nsec_counter(struct perf_evsel *evsel)
178 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
179 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
186 * Read out the results of a single counter:
187 * do not aggregate counts across CPUs in system-wide mode
189 static int read_counter(struct perf_evsel *counter)
191 int nthreads = thread_map__nr(evsel_list->threads);
192 int ncpus = perf_evsel__nr_cpus(counter);
195 if (!counter->supported)
198 if (counter->system_wide)
201 for (thread = 0; thread < nthreads; thread++) {
202 for (cpu = 0; cpu < ncpus; cpu++) {
203 struct perf_counts_values *count;
205 count = perf_counts(counter->counts, cpu, thread);
206 if (perf_evsel__read(counter, cpu, thread, count))
214 static void read_counters(bool close_counters)
216 struct perf_evsel *counter;
218 evlist__for_each(evsel_list, counter) {
219 if (read_counter(counter))
220 pr_debug("failed to read counter %s\n", counter->name);
222 if (perf_stat_process_counter(&stat_config, counter))
223 pr_warning("failed to process counter %s\n", counter->name);
225 if (close_counters) {
226 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
227 thread_map__nr(evsel_list->threads));
232 static void process_interval(void)
234 struct timespec ts, rs;
236 read_counters(false);
238 clock_gettime(CLOCK_MONOTONIC, &ts);
239 diff_timespec(&rs, &ts, &ref_time);
241 print_counters(&rs, 0, NULL);
244 static void handle_initial_delay(void)
246 struct perf_evsel *counter;
249 const int ncpus = cpu_map__nr(evsel_list->cpus),
250 nthreads = thread_map__nr(evsel_list->threads);
252 usleep(initial_delay * 1000);
253 evlist__for_each(evsel_list, counter)
254 perf_evsel__enable(counter, ncpus, nthreads);
258 static volatile int workload_exec_errno;
261 * perf_evlist__prepare_workload will send a SIGUSR1
262 * if the fork fails, since we asked by setting its
263 * want_signal to true.
265 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
266 void *ucontext __maybe_unused)
268 workload_exec_errno = info->si_value.sival_int;
271 static int __run_perf_stat(int argc, const char **argv)
273 int interval = stat_config.interval;
275 unsigned long long t0, t1;
276 struct perf_evsel *counter;
280 const bool forks = (argc > 0);
283 ts.tv_sec = interval / 1000;
284 ts.tv_nsec = (interval % 1000) * 1000000;
291 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
292 workload_exec_failed_signal) < 0) {
293 perror("failed to prepare workload");
296 child_pid = evsel_list->workload.pid;
300 perf_evlist__set_leader(evsel_list);
302 evlist__for_each(evsel_list, counter) {
303 if (create_perf_stat_counter(counter) < 0) {
305 * PPC returns ENXIO for HW counters until 2.6.37
306 * (behavior changed with commit b0a873e).
308 if (errno == EINVAL || errno == ENOSYS ||
309 errno == ENOENT || errno == EOPNOTSUPP ||
312 ui__warning("%s event is not supported by the kernel.\n",
313 perf_evsel__name(counter));
314 counter->supported = false;
316 if ((counter->leader != counter) ||
317 !(counter->leader->nr_members > 1))
321 perf_evsel__open_strerror(counter, &target,
322 errno, msg, sizeof(msg));
323 ui__error("%s\n", msg);
326 kill(child_pid, SIGTERM);
330 counter->supported = true;
332 l = strlen(counter->unit);
337 if (perf_evlist__apply_filters(evsel_list, &counter)) {
338 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
339 counter->filter, perf_evsel__name(counter), errno,
340 strerror_r(errno, msg, sizeof(msg)));
345 * Enable counters and exec the command:
348 clock_gettime(CLOCK_MONOTONIC, &ref_time);
351 perf_evlist__start_workload(evsel_list);
352 handle_initial_delay();
355 while (!waitpid(child_pid, &status, WNOHANG)) {
356 nanosleep(&ts, NULL);
362 if (workload_exec_errno) {
363 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
364 pr_err("Workload failed: %s\n", emsg);
368 if (WIFSIGNALED(status))
369 psignal(WTERMSIG(status), argv[0]);
371 handle_initial_delay();
373 nanosleep(&ts, NULL);
381 update_stats(&walltime_nsecs_stats, t1 - t0);
385 return WEXITSTATUS(status);
388 static int run_perf_stat(int argc, const char **argv)
393 ret = system(pre_cmd);
401 ret = __run_perf_stat(argc, argv);
406 ret = system(post_cmd);
414 static void print_running(u64 run, u64 ena)
417 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
421 ena ? 100.0 * run / ena : 100.0);
422 } else if (run != ena) {
423 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
427 static void print_noise_pct(double total, double avg)
429 double pct = rel_stddev_stats(total, avg);
432 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
434 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
437 static void print_noise(struct perf_evsel *evsel, double avg)
439 struct perf_stat_evsel *ps;
445 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
448 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
450 switch (stat_config.aggr_mode) {
452 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
453 cpu_map__id_to_socket(id),
455 cpu_map__id_to_cpu(id),
462 fprintf(stat_config.output, "S%*d%s%*d%s",
471 fprintf(stat_config.output, "CPU%*d%s",
473 perf_evsel__cpus(evsel)->map[id], csv_sep);
476 fprintf(stat_config.output, "%*s-%*d%s",
478 thread_map__comm(evsel->threads, id),
480 thread_map__pid(evsel->threads, id),
490 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
492 FILE *output = stat_config.output;
493 double msecs = avg / 1e6;
494 const char *fmt_v, *fmt_n;
497 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
498 fmt_n = csv_output ? "%s" : "%-25s";
500 aggr_printout(evsel, id, nr);
502 scnprintf(name, sizeof(name), "%s%s",
503 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
505 fprintf(output, fmt_v, msecs, csv_sep);
508 fprintf(output, "%s%s", evsel->unit, csv_sep);
510 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
512 fprintf(output, fmt_n, name);
515 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
517 if (csv_output || stat_config.interval)
520 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
521 fprintf(output, " # %8.3f CPUs utilized ",
522 avg / avg_stats(&walltime_nsecs_stats));
524 fprintf(output, " ");
527 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
529 FILE *output = stat_config.output;
530 double sc = evsel->scale;
532 int cpu = cpu_map__id_to_cpu(id);
535 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
538 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
540 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
543 aggr_printout(evsel, id, nr);
545 if (stat_config.aggr_mode == AGGR_GLOBAL)
548 fprintf(output, fmt, avg, csv_sep);
551 fprintf(output, "%-*s%s",
552 csv_output ? 0 : unit_width,
553 evsel->unit, csv_sep);
555 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
558 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
560 if (csv_output || stat_config.interval)
563 perf_stat__print_shadow_stats(output, evsel, avg, cpu,
564 stat_config.aggr_mode);
567 static void print_aggr(char *prefix)
569 FILE *output = stat_config.output;
570 struct perf_evsel *counter;
571 int cpu, s, s2, id, nr;
575 if (!(aggr_map || aggr_get_id))
578 for (s = 0; s < aggr_map->nr; s++) {
579 id = aggr_map->map[s];
580 evlist__for_each(evsel_list, counter) {
583 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
584 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
587 val += perf_counts(counter->counts, cpu, 0)->val;
588 ena += perf_counts(counter->counts, cpu, 0)->ena;
589 run += perf_counts(counter->counts, cpu, 0)->run;
593 fprintf(output, "%s", prefix);
595 if (run == 0 || ena == 0) {
596 aggr_printout(counter, id, nr);
598 fprintf(output, "%*s%s",
600 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
603 fprintf(output, "%-*s%s",
604 csv_output ? 0 : unit_width,
605 counter->unit, csv_sep);
607 fprintf(output, "%*s",
608 csv_output ? 0 : -25,
609 perf_evsel__name(counter));
612 fprintf(output, "%s%s",
613 csv_sep, counter->cgrp->name);
615 print_running(run, ena);
619 uval = val * counter->scale;
621 if (nsec_counter(counter))
622 nsec_printout(id, nr, counter, uval);
624 abs_printout(id, nr, counter, uval);
627 print_noise(counter, 1.0);
629 print_running(run, ena);
635 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
637 FILE *output = stat_config.output;
638 int nthreads = thread_map__nr(counter->threads);
639 int ncpus = cpu_map__nr(counter->cpus);
643 for (thread = 0; thread < nthreads; thread++) {
644 u64 ena = 0, run = 0, val = 0;
646 for (cpu = 0; cpu < ncpus; cpu++) {
647 val += perf_counts(counter->counts, cpu, thread)->val;
648 ena += perf_counts(counter->counts, cpu, thread)->ena;
649 run += perf_counts(counter->counts, cpu, thread)->run;
653 fprintf(output, "%s", prefix);
655 uval = val * counter->scale;
657 if (nsec_counter(counter))
658 nsec_printout(thread, 0, counter, uval);
660 abs_printout(thread, 0, counter, uval);
663 print_noise(counter, 1.0);
665 print_running(run, ena);
671 * Print out the results of a single counter:
672 * aggregated counts in system-wide mode
674 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
676 FILE *output = stat_config.output;
677 struct perf_stat_evsel *ps = counter->priv;
678 double avg = avg_stats(&ps->res_stats[0]);
679 int scaled = counter->counts->scaled;
681 double avg_enabled, avg_running;
683 avg_enabled = avg_stats(&ps->res_stats[1]);
684 avg_running = avg_stats(&ps->res_stats[2]);
687 fprintf(output, "%s", prefix);
689 if (scaled == -1 || !counter->supported) {
690 fprintf(output, "%*s%s",
692 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
694 fprintf(output, "%-*s%s",
695 csv_output ? 0 : unit_width,
696 counter->unit, csv_sep);
697 fprintf(output, "%*s",
698 csv_output ? 0 : -25,
699 perf_evsel__name(counter));
702 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
704 print_running(avg_running, avg_enabled);
709 uval = avg * counter->scale;
711 if (nsec_counter(counter))
712 nsec_printout(-1, 0, counter, uval);
714 abs_printout(-1, 0, counter, uval);
716 print_noise(counter, avg);
718 print_running(avg_running, avg_enabled);
719 fprintf(output, "\n");
723 * Print out the results of a single counter:
724 * does not use aggregated count in system-wide
726 static void print_counter(struct perf_evsel *counter, char *prefix)
728 FILE *output = stat_config.output;
733 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
734 val = perf_counts(counter->counts, cpu, 0)->val;
735 ena = perf_counts(counter->counts, cpu, 0)->ena;
736 run = perf_counts(counter->counts, cpu, 0)->run;
739 fprintf(output, "%s", prefix);
741 if (run == 0 || ena == 0) {
742 fprintf(output, "CPU%*d%s%*s%s",
744 perf_evsel__cpus(counter)->map[cpu], csv_sep,
746 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
749 fprintf(output, "%-*s%s",
750 csv_output ? 0 : unit_width,
751 counter->unit, csv_sep);
753 fprintf(output, "%*s",
754 csv_output ? 0 : -25,
755 perf_evsel__name(counter));
758 fprintf(output, "%s%s",
759 csv_sep, counter->cgrp->name);
761 print_running(run, ena);
766 uval = val * counter->scale;
768 if (nsec_counter(counter))
769 nsec_printout(cpu, 0, counter, uval);
771 abs_printout(cpu, 0, counter, uval);
774 print_noise(counter, 1.0);
775 print_running(run, ena);
781 static void print_interval(char *prefix, struct timespec *ts)
783 FILE *output = stat_config.output;
784 static int num_print_interval;
786 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
788 if (num_print_interval == 0 && !csv_output) {
789 switch (stat_config.aggr_mode) {
791 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
794 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
797 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
800 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
804 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
810 if (++num_print_interval == 25)
811 num_print_interval = 0;
814 static void print_header(int argc, const char **argv)
816 FILE *output = stat_config.output;
822 fprintf(output, "\n");
823 fprintf(output, " Performance counter stats for ");
824 if (target.system_wide)
825 fprintf(output, "\'system wide");
826 else if (target.cpu_list)
827 fprintf(output, "\'CPU(s) %s", target.cpu_list);
828 else if (!target__has_task(&target)) {
829 fprintf(output, "\'%s", argv[0]);
830 for (i = 1; i < argc; i++)
831 fprintf(output, " %s", argv[i]);
832 } else if (target.pid)
833 fprintf(output, "process id \'%s", target.pid);
835 fprintf(output, "thread id \'%s", target.tid);
837 fprintf(output, "\'");
839 fprintf(output, " (%d runs)", run_count);
840 fprintf(output, ":\n\n");
844 static void print_footer(void)
846 FILE *output = stat_config.output;
849 fprintf(output, "\n");
850 fprintf(output, " %17.9f seconds time elapsed",
851 avg_stats(&walltime_nsecs_stats)/1e9);
853 fprintf(output, " ");
854 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
855 avg_stats(&walltime_nsecs_stats));
857 fprintf(output, "\n\n");
860 static void print_counters(struct timespec *ts, int argc, const char **argv)
862 int interval = stat_config.interval;
863 struct perf_evsel *counter;
864 char buf[64], *prefix = NULL;
867 print_interval(prefix = buf, ts);
869 print_header(argc, argv);
871 switch (stat_config.aggr_mode) {
877 evlist__for_each(evsel_list, counter)
878 print_aggr_thread(counter, prefix);
881 evlist__for_each(evsel_list, counter)
882 print_counter_aggr(counter, prefix);
885 evlist__for_each(evsel_list, counter)
886 print_counter(counter, prefix);
893 if (!interval && !csv_output)
896 fflush(stat_config.output);
899 static volatile int signr = -1;
901 static void skip_signal(int signo)
903 if ((child_pid == -1) || stat_config.interval)
908 * render child_pid harmless
909 * won't send SIGTERM to a random
910 * process in case of race condition
911 * and fast PID recycling
916 static void sig_atexit(void)
921 * avoid race condition with SIGCHLD handler
922 * in skip_signal() which is modifying child_pid
923 * goal is to avoid send SIGTERM to a random
927 sigaddset(&set, SIGCHLD);
928 sigprocmask(SIG_BLOCK, &set, &oset);
931 kill(child_pid, SIGTERM);
933 sigprocmask(SIG_SETMASK, &oset, NULL);
938 signal(signr, SIG_DFL);
939 kill(getpid(), signr);
942 static int stat__set_big_num(const struct option *opt __maybe_unused,
943 const char *s __maybe_unused, int unset)
945 big_num_opt = unset ? 0 : 1;
949 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
951 return cpu_map__get_socket(map, cpu, NULL);
954 static int perf_stat__get_core(struct cpu_map *map, int cpu)
956 return cpu_map__get_core(map, cpu, NULL);
959 static int cpu_map__get_max(struct cpu_map *map)
963 for (i = 0; i < map->nr; i++) {
964 if (map->map[i] > max)
971 static struct cpu_map *cpus_aggr_map;
973 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
982 if (cpus_aggr_map->map[cpu] == -1)
983 cpus_aggr_map->map[cpu] = get_id(map, idx);
985 return cpus_aggr_map->map[cpu];
988 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
990 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
993 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
995 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
998 static int perf_stat_init_aggr_mode(void)
1002 switch (stat_config.aggr_mode) {
1004 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1005 perror("cannot build socket map");
1008 aggr_get_id = perf_stat__get_socket_cached;
1011 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1012 perror("cannot build core map");
1015 aggr_get_id = perf_stat__get_core_cached;
1026 * The evsel_list->cpus is the base we operate on,
1027 * taking the highest cpu number to be the size of
1028 * the aggregation translate cpumap.
1030 nr = cpu_map__get_max(evsel_list->cpus);
1031 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1032 return cpus_aggr_map ? 0 : -ENOMEM;
1036 * Add default attributes, if there were no attributes specified or
1037 * if -d/--detailed, -d -d or -d -d -d is used:
1039 static int add_default_attributes(void)
1041 struct perf_event_attr default_attrs[] = {
1043 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1044 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1045 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1046 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1048 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1049 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1050 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1051 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1052 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1053 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1058 * Detailed stats (-d), covering the L1 and last level data caches:
1060 struct perf_event_attr detailed_attrs[] = {
1062 { .type = PERF_TYPE_HW_CACHE,
1064 PERF_COUNT_HW_CACHE_L1D << 0 |
1065 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1066 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1068 { .type = PERF_TYPE_HW_CACHE,
1070 PERF_COUNT_HW_CACHE_L1D << 0 |
1071 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1072 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1074 { .type = PERF_TYPE_HW_CACHE,
1076 PERF_COUNT_HW_CACHE_LL << 0 |
1077 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1078 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1080 { .type = PERF_TYPE_HW_CACHE,
1082 PERF_COUNT_HW_CACHE_LL << 0 |
1083 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1084 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1088 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1090 struct perf_event_attr very_detailed_attrs[] = {
1092 { .type = PERF_TYPE_HW_CACHE,
1094 PERF_COUNT_HW_CACHE_L1I << 0 |
1095 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1096 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1098 { .type = PERF_TYPE_HW_CACHE,
1100 PERF_COUNT_HW_CACHE_L1I << 0 |
1101 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1102 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1104 { .type = PERF_TYPE_HW_CACHE,
1106 PERF_COUNT_HW_CACHE_DTLB << 0 |
1107 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1108 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1110 { .type = PERF_TYPE_HW_CACHE,
1112 PERF_COUNT_HW_CACHE_DTLB << 0 |
1113 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1114 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1116 { .type = PERF_TYPE_HW_CACHE,
1118 PERF_COUNT_HW_CACHE_ITLB << 0 |
1119 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1120 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1122 { .type = PERF_TYPE_HW_CACHE,
1124 PERF_COUNT_HW_CACHE_ITLB << 0 |
1125 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1126 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1131 * Very, very detailed stats (-d -d -d), adding prefetch events:
1133 struct perf_event_attr very_very_detailed_attrs[] = {
1135 { .type = PERF_TYPE_HW_CACHE,
1137 PERF_COUNT_HW_CACHE_L1D << 0 |
1138 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1139 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1141 { .type = PERF_TYPE_HW_CACHE,
1143 PERF_COUNT_HW_CACHE_L1D << 0 |
1144 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1145 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1148 /* Set attrs if no event is selected and !null_run: */
1152 if (transaction_run) {
1154 if (pmu_have_event("cpu", "cycles-ct") &&
1155 pmu_have_event("cpu", "el-start"))
1156 err = parse_events(evsel_list, transaction_attrs, NULL);
1158 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1160 fprintf(stderr, "Cannot set up transaction events\n");
1166 if (!evsel_list->nr_entries) {
1167 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1171 /* Detailed events get appended to the event list: */
1173 if (detailed_run < 1)
1176 /* Append detailed run extra attributes: */
1177 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1180 if (detailed_run < 2)
1183 /* Append very detailed run extra attributes: */
1184 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1187 if (detailed_run < 3)
1190 /* Append very, very detailed run extra attributes: */
1191 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1194 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1196 bool append_file = false;
1198 const char *output_name = NULL;
1199 const struct option options[] = {
1200 OPT_BOOLEAN('T', "transaction", &transaction_run,
1201 "hardware transaction statistics"),
1202 OPT_CALLBACK('e', "event", &evsel_list, "event",
1203 "event selector. use 'perf list' to list available events",
1204 parse_events_option),
1205 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1206 "event filter", parse_filter),
1207 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1208 "child tasks do not inherit counters"),
1209 OPT_STRING('p', "pid", &target.pid, "pid",
1210 "stat events on existing process id"),
1211 OPT_STRING('t', "tid", &target.tid, "tid",
1212 "stat events on existing thread id"),
1213 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1214 "system-wide collection from all CPUs"),
1215 OPT_BOOLEAN('g', "group", &group,
1216 "put the counters into a counter group"),
1217 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1218 OPT_INCR('v', "verbose", &verbose,
1219 "be more verbose (show counter open errors, etc)"),
1220 OPT_INTEGER('r', "repeat", &run_count,
1221 "repeat command and print average + stddev (max: 100, forever: 0)"),
1222 OPT_BOOLEAN('n', "null", &null_run,
1223 "null run - dont start any counters"),
1224 OPT_INCR('d', "detailed", &detailed_run,
1225 "detailed run - start a lot of events"),
1226 OPT_BOOLEAN('S', "sync", &sync_run,
1227 "call sync() before starting a run"),
1228 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1229 "print large numbers with thousands\' separators",
1231 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1232 "list of cpus to monitor in system-wide"),
1233 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1234 "disable CPU count aggregation", AGGR_NONE),
1235 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1236 "print counts with custom separator"),
1237 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1238 "monitor event in cgroup name only", parse_cgroups),
1239 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1240 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1241 OPT_INTEGER(0, "log-fd", &output_fd,
1242 "log output to fd, instead of stderr"),
1243 OPT_STRING(0, "pre", &pre_cmd, "command",
1244 "command to run prior to the measured command"),
1245 OPT_STRING(0, "post", &post_cmd, "command",
1246 "command to run after to the measured command"),
1247 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1248 "print counts at regular interval in ms (>= 10)"),
1249 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1250 "aggregate counts per processor socket", AGGR_SOCKET),
1251 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1252 "aggregate counts per physical processor core", AGGR_CORE),
1253 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1254 "aggregate counts per thread", AGGR_THREAD),
1255 OPT_UINTEGER('D', "delay", &initial_delay,
1256 "ms to wait before starting measurement after program start"),
1259 const char * const stat_usage[] = {
1260 "perf stat [<options>] [<command>]",
1263 int status = -EINVAL, run_idx;
1265 FILE *output = stderr;
1266 unsigned int interval;
1268 setlocale(LC_ALL, "");
1270 evsel_list = perf_evlist__new();
1271 if (evsel_list == NULL)
1274 argc = parse_options(argc, argv, options, stat_usage,
1275 PARSE_OPT_STOP_AT_NON_OPTION);
1277 interval = stat_config.interval;
1279 if (output_name && strcmp(output_name, "-"))
1282 if (output_name && output_fd) {
1283 fprintf(stderr, "cannot use both --output and --log-fd\n");
1284 parse_options_usage(stat_usage, options, "o", 1);
1285 parse_options_usage(NULL, options, "log-fd", 0);
1289 if (output_fd < 0) {
1290 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1291 parse_options_usage(stat_usage, options, "log-fd", 0);
1297 mode = append_file ? "a" : "w";
1299 output = fopen(output_name, mode);
1301 perror("failed to create output file");
1304 clock_gettime(CLOCK_REALTIME, &tm);
1305 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1306 } else if (output_fd > 0) {
1307 mode = append_file ? "a" : "w";
1308 output = fdopen(output_fd, mode);
1310 perror("Failed opening logfd");
1315 stat_config.output = output;
1319 if (!strcmp(csv_sep, "\\t"))
1322 csv_sep = DEFAULT_SEPARATOR;
1325 * let the spreadsheet do the pretty-printing
1328 /* User explicitly passed -B? */
1329 if (big_num_opt == 1) {
1330 fprintf(stderr, "-B option not supported with -x\n");
1331 parse_options_usage(stat_usage, options, "B", 1);
1332 parse_options_usage(NULL, options, "x", 1);
1334 } else /* Nope, so disable big number formatting */
1336 } else if (big_num_opt == 0) /* User passed --no-big-num */
1339 if (!argc && target__none(&target))
1340 usage_with_options(stat_usage, options);
1342 if (run_count < 0) {
1343 pr_err("Run count must be a positive number\n");
1344 parse_options_usage(stat_usage, options, "r", 1);
1346 } else if (run_count == 0) {
1351 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1352 fprintf(stderr, "The --per-thread option is only available "
1353 "when monitoring via -p -t options.\n");
1354 parse_options_usage(NULL, options, "p", 1);
1355 parse_options_usage(NULL, options, "t", 1);
1360 * no_aggr, cgroup are for system-wide only
1361 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1363 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1364 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1365 !target__has_cpu(&target)) {
1366 fprintf(stderr, "both cgroup and no-aggregation "
1367 "modes only available in system-wide mode\n");
1369 parse_options_usage(stat_usage, options, "G", 1);
1370 parse_options_usage(NULL, options, "A", 1);
1371 parse_options_usage(NULL, options, "a", 1);
1375 if (add_default_attributes())
1378 target__validate(&target);
1380 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1381 if (target__has_task(&target)) {
1382 pr_err("Problems finding threads of monitor\n");
1383 parse_options_usage(stat_usage, options, "p", 1);
1384 parse_options_usage(NULL, options, "t", 1);
1385 } else if (target__has_cpu(&target)) {
1386 perror("failed to parse CPUs map");
1387 parse_options_usage(stat_usage, options, "C", 1);
1388 parse_options_usage(NULL, options, "a", 1);
1394 * Initialize thread_map with comm names,
1395 * so we could print it out on output.
1397 if (stat_config.aggr_mode == AGGR_THREAD)
1398 thread_map__read_comms(evsel_list->threads);
1400 if (interval && interval < 100) {
1401 if (interval < 10) {
1402 pr_err("print interval must be >= 10ms\n");
1403 parse_options_usage(stat_usage, options, "I", 1);
1406 pr_warning("print interval < 100ms. "
1407 "The overhead percentage could be high in some cases. "
1408 "Please proceed with caution.\n");
1411 if (perf_evlist__alloc_stats(evsel_list, interval))
1414 if (perf_stat_init_aggr_mode())
1418 * We dont want to block the signals - that would cause
1419 * child tasks to inherit that and Ctrl-C would not work.
1420 * What we want is for Ctrl-C to work in the exec()-ed
1421 * task, but being ignored by perf stat itself:
1425 signal(SIGINT, skip_signal);
1426 signal(SIGCHLD, skip_signal);
1427 signal(SIGALRM, skip_signal);
1428 signal(SIGABRT, skip_signal);
1431 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1432 if (run_count != 1 && verbose)
1433 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1436 status = run_perf_stat(argc, argv);
1437 if (forever && status != -1) {
1438 print_counters(NULL, argc, argv);
1439 perf_stat__reset_stats();
1443 if (!forever && status != -1 && !interval)
1444 print_counters(NULL, argc, argv);
1446 perf_evlist__free_stats(evsel_list);
1448 perf_evlist__delete(evsel_list);