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_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
75 /* Default events used for perf stat -T */
76 static const char *transaction_attrs = {
88 /* More limited version when the CPU does not have all events. */
89 static const char * transaction_limited_attrs = {
99 static struct perf_evlist *evsel_list;
101 static struct target target = {
105 static int run_count = 1;
106 static bool no_inherit = false;
107 static bool scale = true;
108 static enum aggr_mode aggr_mode = AGGR_GLOBAL;
109 static volatile pid_t child_pid = -1;
110 static bool null_run = false;
111 static int detailed_run = 0;
112 static bool transaction_run;
113 static bool big_num = true;
114 static int big_num_opt = -1;
115 static const char *csv_sep = NULL;
116 static bool csv_output = false;
117 static bool group = false;
118 static FILE *output = NULL;
119 static const char *pre_cmd = NULL;
120 static const char *post_cmd = NULL;
121 static bool sync_run = false;
122 static unsigned int interval = 0;
123 static unsigned int initial_delay = 0;
124 static unsigned int unit_width = 4; /* strlen("unit") */
125 static bool forever = false;
126 static struct timespec ref_time;
127 static struct cpu_map *aggr_map;
128 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
130 static volatile int done = 0;
132 static inline void diff_timespec(struct timespec *r, struct timespec *a,
135 r->tv_sec = a->tv_sec - b->tv_sec;
136 if (a->tv_nsec < b->tv_nsec) {
137 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
140 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
144 static void perf_stat__reset_stats(void)
146 perf_evlist__reset_stats(evsel_list);
147 perf_stat__reset_shadow_stats();
150 static int create_perf_stat_counter(struct perf_evsel *evsel)
152 struct perf_event_attr *attr = &evsel->attr;
155 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
156 PERF_FORMAT_TOTAL_TIME_RUNNING;
158 attr->inherit = !no_inherit;
160 if (target__has_cpu(&target))
161 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
163 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
166 attr->enable_on_exec = 1;
169 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
173 * Does the counter have nsecs as a unit?
175 static inline int nsec_counter(struct perf_evsel *evsel)
177 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
178 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
184 static void zero_per_pkg(struct perf_evsel *counter)
186 if (counter->per_pkg_mask)
187 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
190 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
192 unsigned long *mask = counter->per_pkg_mask;
193 struct cpu_map *cpus = perf_evsel__cpus(counter);
198 if (!counter->per_pkg)
201 if (cpu_map__empty(cpus))
205 mask = zalloc(MAX_NR_CPUS);
209 counter->per_pkg_mask = mask;
212 s = cpu_map__get_socket(cpus, cpu);
216 *skip = test_and_set_bit(s, mask) == 1;
221 process_counter_values(struct perf_evsel *evsel, int cpu, int thread,
222 struct perf_counts_values *count)
224 struct perf_counts_values *aggr = &evsel->counts->aggr;
225 static struct perf_counts_values zero;
228 if (check_per_pkg(evsel, cpu, &skip)) {
229 pr_err("failed to read per-pkg counter\n");
240 if (!evsel->snapshot)
241 perf_evsel__compute_deltas(evsel, cpu, thread, count);
242 perf_counts_values__scale(count, scale, NULL);
243 if (aggr_mode == AGGR_NONE)
244 perf_stat__update_shadow_stats(evsel, count->values, cpu);
247 aggr->val += count->val;
249 aggr->ena += count->ena;
250 aggr->run += count->run;
259 static int process_counter_maps(struct perf_evsel *counter)
261 int nthreads = thread_map__nr(counter->threads);
262 int ncpus = perf_evsel__nr_cpus(counter);
265 if (counter->system_wide)
268 for (thread = 0; thread < nthreads; thread++) {
269 for (cpu = 0; cpu < ncpus; cpu++) {
270 if (process_counter_values(counter, cpu, thread,
271 perf_counts(counter->counts, cpu, thread)))
279 static int process_counter(struct perf_evsel *counter)
281 struct perf_counts_values *aggr = &counter->counts->aggr;
282 struct perf_stat *ps = counter->priv;
283 u64 *count = counter->counts->aggr.values;
286 aggr->val = aggr->ena = aggr->run = 0;
287 memset(ps->res_stats, 0, sizeof(ps->res_stats));
289 if (counter->per_pkg)
290 zero_per_pkg(counter);
292 ret = process_counter_maps(counter);
296 if (aggr_mode != AGGR_GLOBAL)
299 if (!counter->snapshot)
300 perf_evsel__compute_deltas(counter, -1, -1, aggr);
301 perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
303 for (i = 0; i < 3; i++)
304 update_stats(&ps->res_stats[i], count[i]);
307 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
308 perf_evsel__name(counter), count[0], count[1], count[2]);
312 * Save the full runtime - to allow normalization during printout:
314 perf_stat__update_shadow_stats(counter, count, 0);
320 * Read out the results of a single counter:
321 * do not aggregate counts across CPUs in system-wide mode
323 static int read_counter(struct perf_evsel *counter)
325 int nthreads = thread_map__nr(evsel_list->threads);
326 int ncpus = perf_evsel__nr_cpus(counter);
329 if (!counter->supported)
332 if (counter->system_wide)
335 for (thread = 0; thread < nthreads; thread++) {
336 for (cpu = 0; cpu < ncpus; cpu++) {
337 struct perf_counts_values *count;
339 count = perf_counts(counter->counts, cpu, thread);
340 if (perf_evsel__read(counter, cpu, thread, count))
348 static void read_counters(bool close)
350 struct perf_evsel *counter;
352 evlist__for_each(evsel_list, counter) {
353 if (read_counter(counter))
354 pr_warning("failed to read counter %s\n", counter->name);
356 if (process_counter(counter))
357 pr_warning("failed to process counter %s\n", counter->name);
360 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
361 thread_map__nr(evsel_list->threads));
366 static void process_interval(void)
368 static int num_print_interval;
369 struct perf_evsel *counter;
370 struct timespec ts, rs;
373 read_counters(false);
375 clock_gettime(CLOCK_MONOTONIC, &ts);
376 diff_timespec(&rs, &ts, &ref_time);
377 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
379 if (num_print_interval == 0 && !csv_output) {
382 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
385 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
388 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
392 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
396 if (++num_print_interval == 25)
397 num_print_interval = 0;
405 evlist__for_each(evsel_list, counter)
406 print_counter(counter, prefix);
410 evlist__for_each(evsel_list, counter)
411 print_counter_aggr(counter, prefix);
417 static void handle_initial_delay(void)
419 struct perf_evsel *counter;
422 const int ncpus = cpu_map__nr(evsel_list->cpus),
423 nthreads = thread_map__nr(evsel_list->threads);
425 usleep(initial_delay * 1000);
426 evlist__for_each(evsel_list, counter)
427 perf_evsel__enable(counter, ncpus, nthreads);
431 static volatile int workload_exec_errno;
434 * perf_evlist__prepare_workload will send a SIGUSR1
435 * if the fork fails, since we asked by setting its
436 * want_signal to true.
438 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
439 void *ucontext __maybe_unused)
441 workload_exec_errno = info->si_value.sival_int;
444 static int __run_perf_stat(int argc, const char **argv)
447 unsigned long long t0, t1;
448 struct perf_evsel *counter;
452 const bool forks = (argc > 0);
455 ts.tv_sec = interval / 1000;
456 ts.tv_nsec = (interval % 1000) * 1000000;
463 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
464 workload_exec_failed_signal) < 0) {
465 perror("failed to prepare workload");
468 child_pid = evsel_list->workload.pid;
472 perf_evlist__set_leader(evsel_list);
474 evlist__for_each(evsel_list, counter) {
475 if (create_perf_stat_counter(counter) < 0) {
477 * PPC returns ENXIO for HW counters until 2.6.37
478 * (behavior changed with commit b0a873e).
480 if (errno == EINVAL || errno == ENOSYS ||
481 errno == ENOENT || errno == EOPNOTSUPP ||
484 ui__warning("%s event is not supported by the kernel.\n",
485 perf_evsel__name(counter));
486 counter->supported = false;
488 if ((counter->leader != counter) ||
489 !(counter->leader->nr_members > 1))
493 perf_evsel__open_strerror(counter, &target,
494 errno, msg, sizeof(msg));
495 ui__error("%s\n", msg);
498 kill(child_pid, SIGTERM);
502 counter->supported = true;
504 l = strlen(counter->unit);
509 if (perf_evlist__apply_filters(evsel_list, &counter)) {
510 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
511 counter->filter, perf_evsel__name(counter), errno,
512 strerror_r(errno, msg, sizeof(msg)));
517 * Enable counters and exec the command:
520 clock_gettime(CLOCK_MONOTONIC, &ref_time);
523 perf_evlist__start_workload(evsel_list);
524 handle_initial_delay();
527 while (!waitpid(child_pid, &status, WNOHANG)) {
528 nanosleep(&ts, NULL);
534 if (workload_exec_errno) {
535 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
536 pr_err("Workload failed: %s\n", emsg);
540 if (WIFSIGNALED(status))
541 psignal(WTERMSIG(status), argv[0]);
543 handle_initial_delay();
545 nanosleep(&ts, NULL);
553 update_stats(&walltime_nsecs_stats, t1 - t0);
557 return WEXITSTATUS(status);
560 static int run_perf_stat(int argc, const char **argv)
565 ret = system(pre_cmd);
573 ret = __run_perf_stat(argc, argv);
578 ret = system(post_cmd);
586 static void print_running(u64 run, u64 ena)
589 fprintf(output, "%s%" PRIu64 "%s%.2f",
593 ena ? 100.0 * run / ena : 100.0);
594 } else if (run != ena) {
595 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
599 static void print_noise_pct(double total, double avg)
601 double pct = rel_stddev_stats(total, avg);
604 fprintf(output, "%s%.2f%%", csv_sep, pct);
606 fprintf(output, " ( +-%6.2f%% )", pct);
609 static void print_noise(struct perf_evsel *evsel, double avg)
611 struct perf_stat *ps;
617 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
620 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
624 fprintf(output, "S%d-C%*d%s%*d%s",
625 cpu_map__id_to_socket(id),
627 cpu_map__id_to_cpu(id),
634 fprintf(output, "S%*d%s%*d%s",
643 fprintf(output, "CPU%*d%s",
645 perf_evsel__cpus(evsel)->map[id], csv_sep);
653 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
655 double msecs = avg / 1e6;
656 const char *fmt_v, *fmt_n;
659 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
660 fmt_n = csv_output ? "%s" : "%-25s";
662 aggr_printout(evsel, id, nr);
664 scnprintf(name, sizeof(name), "%s%s",
665 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
667 fprintf(output, fmt_v, msecs, csv_sep);
670 fprintf(output, "%s%s", evsel->unit, csv_sep);
672 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
674 fprintf(output, fmt_n, name);
677 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
679 if (csv_output || interval)
682 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
683 fprintf(output, " # %8.3f CPUs utilized ",
684 avg / avg_stats(&walltime_nsecs_stats));
686 fprintf(output, " ");
689 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
691 double sc = evsel->scale;
693 int cpu = cpu_map__id_to_cpu(id);
696 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
699 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
701 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
704 aggr_printout(evsel, id, nr);
706 if (aggr_mode == AGGR_GLOBAL)
709 fprintf(output, fmt, avg, csv_sep);
712 fprintf(output, "%-*s%s",
713 csv_output ? 0 : unit_width,
714 evsel->unit, csv_sep);
716 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
719 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
721 if (csv_output || interval)
724 perf_stat__print_shadow_stats(output, evsel, avg, cpu, aggr_mode);
727 static void print_aggr(char *prefix)
729 struct perf_evsel *counter;
730 int cpu, cpu2, s, s2, id, nr;
734 if (!(aggr_map || aggr_get_id))
737 for (s = 0; s < aggr_map->nr; s++) {
738 id = aggr_map->map[s];
739 evlist__for_each(evsel_list, counter) {
742 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
743 cpu2 = perf_evsel__cpus(counter)->map[cpu];
744 s2 = aggr_get_id(evsel_list->cpus, cpu2);
747 val += perf_counts(counter->counts, cpu, 0)->val;
748 ena += perf_counts(counter->counts, cpu, 0)->ena;
749 run += perf_counts(counter->counts, cpu, 0)->run;
753 fprintf(output, "%s", prefix);
755 if (run == 0 || ena == 0) {
756 aggr_printout(counter, id, nr);
758 fprintf(output, "%*s%s",
760 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
763 fprintf(output, "%-*s%s",
764 csv_output ? 0 : unit_width,
765 counter->unit, csv_sep);
767 fprintf(output, "%*s",
768 csv_output ? 0 : -25,
769 perf_evsel__name(counter));
772 fprintf(output, "%s%s",
773 csv_sep, counter->cgrp->name);
775 print_running(run, ena);
779 uval = val * counter->scale;
781 if (nsec_counter(counter))
782 nsec_printout(id, nr, counter, uval);
784 abs_printout(id, nr, counter, uval);
787 print_noise(counter, 1.0);
789 print_running(run, ena);
796 * Print out the results of a single counter:
797 * aggregated counts in system-wide mode
799 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
801 struct perf_stat *ps = counter->priv;
802 double avg = avg_stats(&ps->res_stats[0]);
803 int scaled = counter->counts->scaled;
805 double avg_enabled, avg_running;
807 avg_enabled = avg_stats(&ps->res_stats[1]);
808 avg_running = avg_stats(&ps->res_stats[2]);
811 fprintf(output, "%s", prefix);
813 if (scaled == -1 || !counter->supported) {
814 fprintf(output, "%*s%s",
816 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
818 fprintf(output, "%-*s%s",
819 csv_output ? 0 : unit_width,
820 counter->unit, csv_sep);
821 fprintf(output, "%*s",
822 csv_output ? 0 : -25,
823 perf_evsel__name(counter));
826 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
828 print_running(avg_running, avg_enabled);
833 uval = avg * counter->scale;
835 if (nsec_counter(counter))
836 nsec_printout(-1, 0, counter, uval);
838 abs_printout(-1, 0, counter, uval);
840 print_noise(counter, avg);
842 print_running(avg_running, avg_enabled);
843 fprintf(output, "\n");
847 * Print out the results of a single counter:
848 * does not use aggregated count in system-wide
850 static void print_counter(struct perf_evsel *counter, char *prefix)
856 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
857 val = perf_counts(counter->counts, cpu, 0)->val;
858 ena = perf_counts(counter->counts, cpu, 0)->ena;
859 run = perf_counts(counter->counts, cpu, 0)->run;
862 fprintf(output, "%s", prefix);
864 if (run == 0 || ena == 0) {
865 fprintf(output, "CPU%*d%s%*s%s",
867 perf_evsel__cpus(counter)->map[cpu], csv_sep,
869 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
872 fprintf(output, "%-*s%s",
873 csv_output ? 0 : unit_width,
874 counter->unit, csv_sep);
876 fprintf(output, "%*s",
877 csv_output ? 0 : -25,
878 perf_evsel__name(counter));
881 fprintf(output, "%s%s",
882 csv_sep, counter->cgrp->name);
884 print_running(run, ena);
889 uval = val * counter->scale;
891 if (nsec_counter(counter))
892 nsec_printout(cpu, 0, counter, uval);
894 abs_printout(cpu, 0, counter, uval);
897 print_noise(counter, 1.0);
898 print_running(run, ena);
904 static void print_stat(int argc, const char **argv)
906 struct perf_evsel *counter;
912 fprintf(output, "\n");
913 fprintf(output, " Performance counter stats for ");
914 if (target.system_wide)
915 fprintf(output, "\'system wide");
916 else if (target.cpu_list)
917 fprintf(output, "\'CPU(s) %s", target.cpu_list);
918 else if (!target__has_task(&target)) {
919 fprintf(output, "\'%s", argv[0]);
920 for (i = 1; i < argc; i++)
921 fprintf(output, " %s", argv[i]);
922 } else if (target.pid)
923 fprintf(output, "process id \'%s", target.pid);
925 fprintf(output, "thread id \'%s", target.tid);
927 fprintf(output, "\'");
929 fprintf(output, " (%d runs)", run_count);
930 fprintf(output, ":\n\n");
939 evlist__for_each(evsel_list, counter)
940 print_counter_aggr(counter, NULL);
943 evlist__for_each(evsel_list, counter)
944 print_counter(counter, NULL);
952 fprintf(output, "\n");
953 fprintf(output, " %17.9f seconds time elapsed",
954 avg_stats(&walltime_nsecs_stats)/1e9);
956 fprintf(output, " ");
957 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
958 avg_stats(&walltime_nsecs_stats));
960 fprintf(output, "\n\n");
964 static volatile int signr = -1;
966 static void skip_signal(int signo)
968 if ((child_pid == -1) || interval)
973 * render child_pid harmless
974 * won't send SIGTERM to a random
975 * process in case of race condition
976 * and fast PID recycling
981 static void sig_atexit(void)
986 * avoid race condition with SIGCHLD handler
987 * in skip_signal() which is modifying child_pid
988 * goal is to avoid send SIGTERM to a random
992 sigaddset(&set, SIGCHLD);
993 sigprocmask(SIG_BLOCK, &set, &oset);
996 kill(child_pid, SIGTERM);
998 sigprocmask(SIG_SETMASK, &oset, NULL);
1003 signal(signr, SIG_DFL);
1004 kill(getpid(), signr);
1007 static int stat__set_big_num(const struct option *opt __maybe_unused,
1008 const char *s __maybe_unused, int unset)
1010 big_num_opt = unset ? 0 : 1;
1014 static int perf_stat_init_aggr_mode(void)
1016 switch (aggr_mode) {
1018 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1019 perror("cannot build socket map");
1022 aggr_get_id = cpu_map__get_socket;
1025 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1026 perror("cannot build core map");
1029 aggr_get_id = cpu_map__get_core;
1040 * Add default attributes, if there were no attributes specified or
1041 * if -d/--detailed, -d -d or -d -d -d is used:
1043 static int add_default_attributes(void)
1045 struct perf_event_attr default_attrs[] = {
1047 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1048 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1049 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1050 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1052 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1053 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1054 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1055 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1056 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1057 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1062 * Detailed stats (-d), covering the L1 and last level data caches:
1064 struct perf_event_attr detailed_attrs[] = {
1066 { .type = PERF_TYPE_HW_CACHE,
1068 PERF_COUNT_HW_CACHE_L1D << 0 |
1069 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1070 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1072 { .type = PERF_TYPE_HW_CACHE,
1074 PERF_COUNT_HW_CACHE_L1D << 0 |
1075 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1076 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1078 { .type = PERF_TYPE_HW_CACHE,
1080 PERF_COUNT_HW_CACHE_LL << 0 |
1081 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1082 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1084 { .type = PERF_TYPE_HW_CACHE,
1086 PERF_COUNT_HW_CACHE_LL << 0 |
1087 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1088 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1092 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1094 struct perf_event_attr very_detailed_attrs[] = {
1096 { .type = PERF_TYPE_HW_CACHE,
1098 PERF_COUNT_HW_CACHE_L1I << 0 |
1099 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1100 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1102 { .type = PERF_TYPE_HW_CACHE,
1104 PERF_COUNT_HW_CACHE_L1I << 0 |
1105 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1106 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1108 { .type = PERF_TYPE_HW_CACHE,
1110 PERF_COUNT_HW_CACHE_DTLB << 0 |
1111 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1112 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1114 { .type = PERF_TYPE_HW_CACHE,
1116 PERF_COUNT_HW_CACHE_DTLB << 0 |
1117 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1118 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1120 { .type = PERF_TYPE_HW_CACHE,
1122 PERF_COUNT_HW_CACHE_ITLB << 0 |
1123 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1124 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1126 { .type = PERF_TYPE_HW_CACHE,
1128 PERF_COUNT_HW_CACHE_ITLB << 0 |
1129 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1130 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1135 * Very, very detailed stats (-d -d -d), adding prefetch events:
1137 struct perf_event_attr very_very_detailed_attrs[] = {
1139 { .type = PERF_TYPE_HW_CACHE,
1141 PERF_COUNT_HW_CACHE_L1D << 0 |
1142 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1143 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1145 { .type = PERF_TYPE_HW_CACHE,
1147 PERF_COUNT_HW_CACHE_L1D << 0 |
1148 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1149 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1152 /* Set attrs if no event is selected and !null_run: */
1156 if (transaction_run) {
1158 if (pmu_have_event("cpu", "cycles-ct") &&
1159 pmu_have_event("cpu", "el-start"))
1160 err = parse_events(evsel_list, transaction_attrs, NULL);
1162 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1164 fprintf(stderr, "Cannot set up transaction events\n");
1170 if (!evsel_list->nr_entries) {
1171 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1175 /* Detailed events get appended to the event list: */
1177 if (detailed_run < 1)
1180 /* Append detailed run extra attributes: */
1181 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1184 if (detailed_run < 2)
1187 /* Append very detailed run extra attributes: */
1188 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1191 if (detailed_run < 3)
1194 /* Append very, very detailed run extra attributes: */
1195 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1198 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1200 bool append_file = false;
1202 const char *output_name = NULL;
1203 const struct option options[] = {
1204 OPT_BOOLEAN('T', "transaction", &transaction_run,
1205 "hardware transaction statistics"),
1206 OPT_CALLBACK('e', "event", &evsel_list, "event",
1207 "event selector. use 'perf list' to list available events",
1208 parse_events_option),
1209 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1210 "event filter", parse_filter),
1211 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1212 "child tasks do not inherit counters"),
1213 OPT_STRING('p', "pid", &target.pid, "pid",
1214 "stat events on existing process id"),
1215 OPT_STRING('t', "tid", &target.tid, "tid",
1216 "stat events on existing thread id"),
1217 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1218 "system-wide collection from all CPUs"),
1219 OPT_BOOLEAN('g', "group", &group,
1220 "put the counters into a counter group"),
1221 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1222 OPT_INCR('v', "verbose", &verbose,
1223 "be more verbose (show counter open errors, etc)"),
1224 OPT_INTEGER('r', "repeat", &run_count,
1225 "repeat command and print average + stddev (max: 100, forever: 0)"),
1226 OPT_BOOLEAN('n', "null", &null_run,
1227 "null run - dont start any counters"),
1228 OPT_INCR('d', "detailed", &detailed_run,
1229 "detailed run - start a lot of events"),
1230 OPT_BOOLEAN('S', "sync", &sync_run,
1231 "call sync() before starting a run"),
1232 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1233 "print large numbers with thousands\' separators",
1235 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1236 "list of cpus to monitor in system-wide"),
1237 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1238 "disable CPU count aggregation", AGGR_NONE),
1239 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1240 "print counts with custom separator"),
1241 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1242 "monitor event in cgroup name only", parse_cgroups),
1243 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1244 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1245 OPT_INTEGER(0, "log-fd", &output_fd,
1246 "log output to fd, instead of stderr"),
1247 OPT_STRING(0, "pre", &pre_cmd, "command",
1248 "command to run prior to the measured command"),
1249 OPT_STRING(0, "post", &post_cmd, "command",
1250 "command to run after to the measured command"),
1251 OPT_UINTEGER('I', "interval-print", &interval,
1252 "print counts at regular interval in ms (>= 100)"),
1253 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1254 "aggregate counts per processor socket", AGGR_SOCKET),
1255 OPT_SET_UINT(0, "per-core", &aggr_mode,
1256 "aggregate counts per physical processor core", AGGR_CORE),
1257 OPT_UINTEGER('D', "delay", &initial_delay,
1258 "ms to wait before starting measurement after program start"),
1261 const char * const stat_usage[] = {
1262 "perf stat [<options>] [<command>]",
1265 int status = -EINVAL, run_idx;
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);
1278 if (output_name && strcmp(output_name, "-"))
1281 if (output_name && output_fd) {
1282 fprintf(stderr, "cannot use both --output and --log-fd\n");
1283 parse_options_usage(stat_usage, options, "o", 1);
1284 parse_options_usage(NULL, options, "log-fd", 0);
1288 if (output_fd < 0) {
1289 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1290 parse_options_usage(stat_usage, options, "log-fd", 0);
1296 mode = append_file ? "a" : "w";
1298 output = fopen(output_name, mode);
1300 perror("failed to create output file");
1303 clock_gettime(CLOCK_REALTIME, &tm);
1304 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1305 } else if (output_fd > 0) {
1306 mode = append_file ? "a" : "w";
1307 output = fdopen(output_fd, mode);
1309 perror("Failed opening logfd");
1316 if (!strcmp(csv_sep, "\\t"))
1319 csv_sep = DEFAULT_SEPARATOR;
1322 * let the spreadsheet do the pretty-printing
1325 /* User explicitly passed -B? */
1326 if (big_num_opt == 1) {
1327 fprintf(stderr, "-B option not supported with -x\n");
1328 parse_options_usage(stat_usage, options, "B", 1);
1329 parse_options_usage(NULL, options, "x", 1);
1331 } else /* Nope, so disable big number formatting */
1333 } else if (big_num_opt == 0) /* User passed --no-big-num */
1336 if (!argc && target__none(&target))
1337 usage_with_options(stat_usage, options);
1339 if (run_count < 0) {
1340 pr_err("Run count must be a positive number\n");
1341 parse_options_usage(stat_usage, options, "r", 1);
1343 } else if (run_count == 0) {
1348 /* no_aggr, cgroup are for system-wide only */
1349 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1350 !target__has_cpu(&target)) {
1351 fprintf(stderr, "both cgroup and no-aggregation "
1352 "modes only available in system-wide mode\n");
1354 parse_options_usage(stat_usage, options, "G", 1);
1355 parse_options_usage(NULL, options, "A", 1);
1356 parse_options_usage(NULL, options, "a", 1);
1360 if (add_default_attributes())
1363 target__validate(&target);
1365 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1366 if (target__has_task(&target)) {
1367 pr_err("Problems finding threads of monitor\n");
1368 parse_options_usage(stat_usage, options, "p", 1);
1369 parse_options_usage(NULL, options, "t", 1);
1370 } else if (target__has_cpu(&target)) {
1371 perror("failed to parse CPUs map");
1372 parse_options_usage(stat_usage, options, "C", 1);
1373 parse_options_usage(NULL, options, "a", 1);
1377 if (interval && interval < 100) {
1378 pr_err("print interval must be >= 100ms\n");
1379 parse_options_usage(stat_usage, options, "I", 1);
1383 if (perf_evlist__alloc_stats(evsel_list, interval))
1386 if (perf_stat_init_aggr_mode())
1390 * We dont want to block the signals - that would cause
1391 * child tasks to inherit that and Ctrl-C would not work.
1392 * What we want is for Ctrl-C to work in the exec()-ed
1393 * task, but being ignored by perf stat itself:
1397 signal(SIGINT, skip_signal);
1398 signal(SIGCHLD, skip_signal);
1399 signal(SIGALRM, skip_signal);
1400 signal(SIGABRT, skip_signal);
1403 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1404 if (run_count != 1 && verbose)
1405 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1408 status = run_perf_stat(argc, argv);
1409 if (forever && status != -1) {
1410 print_stat(argc, argv);
1411 perf_stat__reset_stats();
1415 if (!forever && status != -1 && !interval)
1416 print_stat(argc, argv);
1418 perf_evlist__free_stats(evsel_list);
1420 perf_evlist__delete(evsel_list);