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 inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
146 return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
149 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
151 return perf_evsel__cpus(evsel)->nr;
154 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
157 struct perf_stat *ps = evsel->priv;
159 for (i = 0; i < 3; i++)
160 init_stats(&ps->res_stats[i]);
162 perf_stat_evsel_id_init(evsel);
165 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
167 evsel->priv = zalloc(sizeof(struct perf_stat));
168 if (evsel->priv == NULL)
170 perf_evsel__reset_stat_priv(evsel);
174 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
179 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
181 struct perf_counts *counts;
183 counts = perf_counts__new(perf_evsel__nr_cpus(evsel));
185 evsel->prev_raw_counts = counts;
187 return counts ? 0 : -ENOMEM;
190 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
192 perf_counts__delete(evsel->prev_raw_counts);
193 evsel->prev_raw_counts = NULL;
196 static void perf_evlist__free_stats(struct perf_evlist *evlist)
198 struct perf_evsel *evsel;
200 evlist__for_each(evlist, evsel) {
201 perf_evsel__free_stat_priv(evsel);
202 perf_evsel__free_counts(evsel);
203 perf_evsel__free_prev_raw_counts(evsel);
207 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
209 struct perf_evsel *evsel;
211 evlist__for_each(evlist, evsel) {
212 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
213 perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
214 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
221 perf_evlist__free_stats(evlist);
225 static void perf_stat__reset_stats(struct perf_evlist *evlist)
227 struct perf_evsel *evsel;
229 evlist__for_each(evlist, evsel) {
230 perf_evsel__reset_stat_priv(evsel);
231 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
234 perf_stat__reset_shadow_stats();
237 static int create_perf_stat_counter(struct perf_evsel *evsel)
239 struct perf_event_attr *attr = &evsel->attr;
242 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
243 PERF_FORMAT_TOTAL_TIME_RUNNING;
245 attr->inherit = !no_inherit;
247 if (target__has_cpu(&target))
248 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
250 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
253 attr->enable_on_exec = 1;
256 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
260 * Does the counter have nsecs as a unit?
262 static inline int nsec_counter(struct perf_evsel *evsel)
264 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
265 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
271 static void zero_per_pkg(struct perf_evsel *counter)
273 if (counter->per_pkg_mask)
274 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
277 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
279 unsigned long *mask = counter->per_pkg_mask;
280 struct cpu_map *cpus = perf_evsel__cpus(counter);
285 if (!counter->per_pkg)
288 if (cpu_map__empty(cpus))
292 mask = zalloc(MAX_NR_CPUS);
296 counter->per_pkg_mask = mask;
299 s = cpu_map__get_socket(cpus, cpu);
303 *skip = test_and_set_bit(s, mask) == 1;
307 static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
308 struct perf_counts_values *count)
310 struct perf_counts_values *aggr = &evsel->counts->aggr;
311 static struct perf_counts_values zero;
314 if (check_per_pkg(evsel, cpu, &skip)) {
315 pr_err("failed to read per-pkg counter\n");
326 if (!evsel->snapshot)
327 perf_evsel__compute_deltas(evsel, cpu, count);
328 perf_counts_values__scale(count, scale, NULL);
329 evsel->counts->cpu[cpu] = *count;
330 if (aggr_mode == AGGR_NONE)
331 perf_stat__update_shadow_stats(evsel, count->values, cpu);
334 aggr->val += count->val;
336 aggr->ena += count->ena;
337 aggr->run += count->run;
346 static int read_counter(struct perf_evsel *counter);
349 * Read out the results of a single counter:
350 * aggregate counts across CPUs in system-wide mode
352 static int read_counter_aggr(struct perf_evsel *counter)
354 struct perf_counts_values *aggr = &counter->counts->aggr;
355 struct perf_stat *ps = counter->priv;
356 u64 *count = counter->counts->aggr.values;
359 aggr->val = aggr->ena = aggr->run = 0;
361 if (read_counter(counter))
364 if (!counter->snapshot)
365 perf_evsel__compute_deltas(counter, -1, aggr);
366 perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
368 for (i = 0; i < 3; i++)
369 update_stats(&ps->res_stats[i], count[i]);
372 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
373 perf_evsel__name(counter), count[0], count[1], count[2]);
377 * Save the full runtime - to allow normalization during printout:
379 perf_stat__update_shadow_stats(counter, count, 0);
385 * Read out the results of a single counter:
386 * do not aggregate counts across CPUs in system-wide mode
388 static int read_counter(struct perf_evsel *counter)
390 int nthreads = thread_map__nr(evsel_list->threads);
391 int ncpus = perf_evsel__nr_cpus(counter);
394 if (!counter->supported)
397 if (counter->system_wide)
400 if (counter->per_pkg)
401 zero_per_pkg(counter);
403 for (thread = 0; thread < nthreads; thread++) {
404 for (cpu = 0; cpu < ncpus; cpu++) {
405 if (perf_evsel__read_cb(counter, cpu, thread, read_cb))
413 static void print_interval(void)
415 static int num_print_interval;
416 struct perf_evsel *counter;
417 struct perf_stat *ps;
418 struct timespec ts, rs;
421 if (aggr_mode == AGGR_GLOBAL) {
422 evlist__for_each(evsel_list, counter) {
424 memset(ps->res_stats, 0, sizeof(ps->res_stats));
425 read_counter_aggr(counter);
428 evlist__for_each(evsel_list, counter) {
430 memset(ps->res_stats, 0, sizeof(ps->res_stats));
431 read_counter(counter);
435 clock_gettime(CLOCK_MONOTONIC, &ts);
436 diff_timespec(&rs, &ts, &ref_time);
437 sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
439 if (num_print_interval == 0 && !csv_output) {
442 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
445 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
448 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
452 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
456 if (++num_print_interval == 25)
457 num_print_interval = 0;
465 evlist__for_each(evsel_list, counter)
466 print_counter(counter, prefix);
470 evlist__for_each(evsel_list, counter)
471 print_counter_aggr(counter, prefix);
477 static void handle_initial_delay(void)
479 struct perf_evsel *counter;
482 const int ncpus = cpu_map__nr(evsel_list->cpus),
483 nthreads = thread_map__nr(evsel_list->threads);
485 usleep(initial_delay * 1000);
486 evlist__for_each(evsel_list, counter)
487 perf_evsel__enable(counter, ncpus, nthreads);
491 static volatile int workload_exec_errno;
494 * perf_evlist__prepare_workload will send a SIGUSR1
495 * if the fork fails, since we asked by setting its
496 * want_signal to true.
498 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
499 void *ucontext __maybe_unused)
501 workload_exec_errno = info->si_value.sival_int;
504 static int __run_perf_stat(int argc, const char **argv)
507 unsigned long long t0, t1;
508 struct perf_evsel *counter;
512 const bool forks = (argc > 0);
515 ts.tv_sec = interval / 1000;
516 ts.tv_nsec = (interval % 1000) * 1000000;
523 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
524 workload_exec_failed_signal) < 0) {
525 perror("failed to prepare workload");
528 child_pid = evsel_list->workload.pid;
532 perf_evlist__set_leader(evsel_list);
534 evlist__for_each(evsel_list, counter) {
535 if (create_perf_stat_counter(counter) < 0) {
537 * PPC returns ENXIO for HW counters until 2.6.37
538 * (behavior changed with commit b0a873e).
540 if (errno == EINVAL || errno == ENOSYS ||
541 errno == ENOENT || errno == EOPNOTSUPP ||
544 ui__warning("%s event is not supported by the kernel.\n",
545 perf_evsel__name(counter));
546 counter->supported = false;
548 if ((counter->leader != counter) ||
549 !(counter->leader->nr_members > 1))
553 perf_evsel__open_strerror(counter, &target,
554 errno, msg, sizeof(msg));
555 ui__error("%s\n", msg);
558 kill(child_pid, SIGTERM);
562 counter->supported = true;
564 l = strlen(counter->unit);
569 if (perf_evlist__apply_filters(evsel_list, &counter)) {
570 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
571 counter->filter, perf_evsel__name(counter), errno,
572 strerror_r(errno, msg, sizeof(msg)));
577 * Enable counters and exec the command:
580 clock_gettime(CLOCK_MONOTONIC, &ref_time);
583 perf_evlist__start_workload(evsel_list);
584 handle_initial_delay();
587 while (!waitpid(child_pid, &status, WNOHANG)) {
588 nanosleep(&ts, NULL);
594 if (workload_exec_errno) {
595 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
596 pr_err("Workload failed: %s\n", emsg);
600 if (WIFSIGNALED(status))
601 psignal(WTERMSIG(status), argv[0]);
603 handle_initial_delay();
605 nanosleep(&ts, NULL);
613 update_stats(&walltime_nsecs_stats, t1 - t0);
615 if (aggr_mode == AGGR_GLOBAL) {
616 evlist__for_each(evsel_list, counter) {
617 read_counter_aggr(counter);
618 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
619 thread_map__nr(evsel_list->threads));
622 evlist__for_each(evsel_list, counter) {
623 read_counter(counter);
624 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
628 return WEXITSTATUS(status);
631 static int run_perf_stat(int argc, const char **argv)
636 ret = system(pre_cmd);
644 ret = __run_perf_stat(argc, argv);
649 ret = system(post_cmd);
657 static void print_running(u64 run, u64 ena)
660 fprintf(output, "%s%" PRIu64 "%s%.2f",
664 ena ? 100.0 * run / ena : 100.0);
665 } else if (run != ena) {
666 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
670 static void print_noise_pct(double total, double avg)
672 double pct = rel_stddev_stats(total, avg);
675 fprintf(output, "%s%.2f%%", csv_sep, pct);
677 fprintf(output, " ( +-%6.2f%% )", pct);
680 static void print_noise(struct perf_evsel *evsel, double avg)
682 struct perf_stat *ps;
688 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
691 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
695 fprintf(output, "S%d-C%*d%s%*d%s",
696 cpu_map__id_to_socket(id),
698 cpu_map__id_to_cpu(id),
705 fprintf(output, "S%*d%s%*d%s",
714 fprintf(output, "CPU%*d%s",
716 perf_evsel__cpus(evsel)->map[id], csv_sep);
724 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
726 double msecs = avg / 1e6;
727 const char *fmt_v, *fmt_n;
730 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
731 fmt_n = csv_output ? "%s" : "%-25s";
733 aggr_printout(evsel, id, nr);
735 scnprintf(name, sizeof(name), "%s%s",
736 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
738 fprintf(output, fmt_v, msecs, csv_sep);
741 fprintf(output, "%s%s", evsel->unit, csv_sep);
743 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
745 fprintf(output, fmt_n, name);
748 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
750 if (csv_output || interval)
753 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
754 fprintf(output, " # %8.3f CPUs utilized ",
755 avg / avg_stats(&walltime_nsecs_stats));
757 fprintf(output, " ");
760 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
762 double sc = evsel->scale;
764 int cpu = cpu_map__id_to_cpu(id);
767 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
770 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
772 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
775 aggr_printout(evsel, id, nr);
777 if (aggr_mode == AGGR_GLOBAL)
780 fprintf(output, fmt, avg, csv_sep);
783 fprintf(output, "%-*s%s",
784 csv_output ? 0 : unit_width,
785 evsel->unit, csv_sep);
787 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
790 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
792 if (csv_output || interval)
795 perf_stat__print_shadow_stats(output, evsel, avg, cpu, aggr_mode);
798 static void print_aggr(char *prefix)
800 struct perf_evsel *counter;
801 int cpu, cpu2, s, s2, id, nr;
805 if (!(aggr_map || aggr_get_id))
808 for (s = 0; s < aggr_map->nr; s++) {
809 id = aggr_map->map[s];
810 evlist__for_each(evsel_list, counter) {
813 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
814 cpu2 = perf_evsel__cpus(counter)->map[cpu];
815 s2 = aggr_get_id(evsel_list->cpus, cpu2);
818 val += counter->counts->cpu[cpu].val;
819 ena += counter->counts->cpu[cpu].ena;
820 run += counter->counts->cpu[cpu].run;
824 fprintf(output, "%s", prefix);
826 if (run == 0 || ena == 0) {
827 aggr_printout(counter, id, nr);
829 fprintf(output, "%*s%s",
831 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
834 fprintf(output, "%-*s%s",
835 csv_output ? 0 : unit_width,
836 counter->unit, csv_sep);
838 fprintf(output, "%*s",
839 csv_output ? 0 : -25,
840 perf_evsel__name(counter));
843 fprintf(output, "%s%s",
844 csv_sep, counter->cgrp->name);
846 print_running(run, ena);
850 uval = val * counter->scale;
852 if (nsec_counter(counter))
853 nsec_printout(id, nr, counter, uval);
855 abs_printout(id, nr, counter, uval);
858 print_noise(counter, 1.0);
860 print_running(run, ena);
867 * Print out the results of a single counter:
868 * aggregated counts in system-wide mode
870 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
872 struct perf_stat *ps = counter->priv;
873 double avg = avg_stats(&ps->res_stats[0]);
874 int scaled = counter->counts->scaled;
876 double avg_enabled, avg_running;
878 avg_enabled = avg_stats(&ps->res_stats[1]);
879 avg_running = avg_stats(&ps->res_stats[2]);
882 fprintf(output, "%s", prefix);
884 if (scaled == -1 || !counter->supported) {
885 fprintf(output, "%*s%s",
887 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
889 fprintf(output, "%-*s%s",
890 csv_output ? 0 : unit_width,
891 counter->unit, csv_sep);
892 fprintf(output, "%*s",
893 csv_output ? 0 : -25,
894 perf_evsel__name(counter));
897 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
899 print_running(avg_running, avg_enabled);
904 uval = avg * counter->scale;
906 if (nsec_counter(counter))
907 nsec_printout(-1, 0, counter, uval);
909 abs_printout(-1, 0, counter, uval);
911 print_noise(counter, avg);
913 print_running(avg_running, avg_enabled);
914 fprintf(output, "\n");
918 * Print out the results of a single counter:
919 * does not use aggregated count in system-wide
921 static void print_counter(struct perf_evsel *counter, char *prefix)
927 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
928 val = counter->counts->cpu[cpu].val;
929 ena = counter->counts->cpu[cpu].ena;
930 run = counter->counts->cpu[cpu].run;
933 fprintf(output, "%s", prefix);
935 if (run == 0 || ena == 0) {
936 fprintf(output, "CPU%*d%s%*s%s",
938 perf_evsel__cpus(counter)->map[cpu], csv_sep,
940 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
943 fprintf(output, "%-*s%s",
944 csv_output ? 0 : unit_width,
945 counter->unit, csv_sep);
947 fprintf(output, "%*s",
948 csv_output ? 0 : -25,
949 perf_evsel__name(counter));
952 fprintf(output, "%s%s",
953 csv_sep, counter->cgrp->name);
955 print_running(run, ena);
960 uval = val * counter->scale;
962 if (nsec_counter(counter))
963 nsec_printout(cpu, 0, counter, uval);
965 abs_printout(cpu, 0, counter, uval);
968 print_noise(counter, 1.0);
969 print_running(run, ena);
975 static void print_stat(int argc, const char **argv)
977 struct perf_evsel *counter;
983 fprintf(output, "\n");
984 fprintf(output, " Performance counter stats for ");
985 if (target.system_wide)
986 fprintf(output, "\'system wide");
987 else if (target.cpu_list)
988 fprintf(output, "\'CPU(s) %s", target.cpu_list);
989 else if (!target__has_task(&target)) {
990 fprintf(output, "\'%s", argv[0]);
991 for (i = 1; i < argc; i++)
992 fprintf(output, " %s", argv[i]);
993 } else if (target.pid)
994 fprintf(output, "process id \'%s", target.pid);
996 fprintf(output, "thread id \'%s", target.tid);
998 fprintf(output, "\'");
1000 fprintf(output, " (%d runs)", run_count);
1001 fprintf(output, ":\n\n");
1004 switch (aggr_mode) {
1010 evlist__for_each(evsel_list, counter)
1011 print_counter_aggr(counter, NULL);
1014 evlist__for_each(evsel_list, counter)
1015 print_counter(counter, NULL);
1023 fprintf(output, "\n");
1024 fprintf(output, " %17.9f seconds time elapsed",
1025 avg_stats(&walltime_nsecs_stats)/1e9);
1026 if (run_count > 1) {
1027 fprintf(output, " ");
1028 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1029 avg_stats(&walltime_nsecs_stats));
1031 fprintf(output, "\n\n");
1035 static volatile int signr = -1;
1037 static void skip_signal(int signo)
1039 if ((child_pid == -1) || interval)
1044 * render child_pid harmless
1045 * won't send SIGTERM to a random
1046 * process in case of race condition
1047 * and fast PID recycling
1052 static void sig_atexit(void)
1057 * avoid race condition with SIGCHLD handler
1058 * in skip_signal() which is modifying child_pid
1059 * goal is to avoid send SIGTERM to a random
1063 sigaddset(&set, SIGCHLD);
1064 sigprocmask(SIG_BLOCK, &set, &oset);
1066 if (child_pid != -1)
1067 kill(child_pid, SIGTERM);
1069 sigprocmask(SIG_SETMASK, &oset, NULL);
1074 signal(signr, SIG_DFL);
1075 kill(getpid(), signr);
1078 static int stat__set_big_num(const struct option *opt __maybe_unused,
1079 const char *s __maybe_unused, int unset)
1081 big_num_opt = unset ? 0 : 1;
1085 static int perf_stat_init_aggr_mode(void)
1087 switch (aggr_mode) {
1089 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1090 perror("cannot build socket map");
1093 aggr_get_id = cpu_map__get_socket;
1096 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1097 perror("cannot build core map");
1100 aggr_get_id = cpu_map__get_core;
1111 * Add default attributes, if there were no attributes specified or
1112 * if -d/--detailed, -d -d or -d -d -d is used:
1114 static int add_default_attributes(void)
1116 struct perf_event_attr default_attrs[] = {
1118 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1119 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1120 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1121 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1123 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1124 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1125 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1126 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1127 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1128 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1133 * Detailed stats (-d), covering the L1 and last level data caches:
1135 struct perf_event_attr detailed_attrs[] = {
1137 { .type = PERF_TYPE_HW_CACHE,
1139 PERF_COUNT_HW_CACHE_L1D << 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_L1D << 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_LL << 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_LL << 0 |
1158 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1159 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1163 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1165 struct perf_event_attr very_detailed_attrs[] = {
1167 { .type = PERF_TYPE_HW_CACHE,
1169 PERF_COUNT_HW_CACHE_L1I << 0 |
1170 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1171 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1173 { .type = PERF_TYPE_HW_CACHE,
1175 PERF_COUNT_HW_CACHE_L1I << 0 |
1176 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1177 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1179 { .type = PERF_TYPE_HW_CACHE,
1181 PERF_COUNT_HW_CACHE_DTLB << 0 |
1182 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1183 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1185 { .type = PERF_TYPE_HW_CACHE,
1187 PERF_COUNT_HW_CACHE_DTLB << 0 |
1188 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1189 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1191 { .type = PERF_TYPE_HW_CACHE,
1193 PERF_COUNT_HW_CACHE_ITLB << 0 |
1194 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1195 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1197 { .type = PERF_TYPE_HW_CACHE,
1199 PERF_COUNT_HW_CACHE_ITLB << 0 |
1200 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1201 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1206 * Very, very detailed stats (-d -d -d), adding prefetch events:
1208 struct perf_event_attr very_very_detailed_attrs[] = {
1210 { .type = PERF_TYPE_HW_CACHE,
1212 PERF_COUNT_HW_CACHE_L1D << 0 |
1213 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1214 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1216 { .type = PERF_TYPE_HW_CACHE,
1218 PERF_COUNT_HW_CACHE_L1D << 0 |
1219 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1220 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1223 /* Set attrs if no event is selected and !null_run: */
1227 if (transaction_run) {
1229 if (pmu_have_event("cpu", "cycles-ct") &&
1230 pmu_have_event("cpu", "el-start"))
1231 err = parse_events(evsel_list, transaction_attrs, NULL);
1233 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1235 fprintf(stderr, "Cannot set up transaction events\n");
1241 if (!evsel_list->nr_entries) {
1242 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1246 /* Detailed events get appended to the event list: */
1248 if (detailed_run < 1)
1251 /* Append detailed run extra attributes: */
1252 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1255 if (detailed_run < 2)
1258 /* Append very detailed run extra attributes: */
1259 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1262 if (detailed_run < 3)
1265 /* Append very, very detailed run extra attributes: */
1266 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1269 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1271 bool append_file = false;
1273 const char *output_name = NULL;
1274 const struct option options[] = {
1275 OPT_BOOLEAN('T', "transaction", &transaction_run,
1276 "hardware transaction statistics"),
1277 OPT_CALLBACK('e', "event", &evsel_list, "event",
1278 "event selector. use 'perf list' to list available events",
1279 parse_events_option),
1280 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1281 "event filter", parse_filter),
1282 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1283 "child tasks do not inherit counters"),
1284 OPT_STRING('p', "pid", &target.pid, "pid",
1285 "stat events on existing process id"),
1286 OPT_STRING('t', "tid", &target.tid, "tid",
1287 "stat events on existing thread id"),
1288 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1289 "system-wide collection from all CPUs"),
1290 OPT_BOOLEAN('g', "group", &group,
1291 "put the counters into a counter group"),
1292 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1293 OPT_INCR('v', "verbose", &verbose,
1294 "be more verbose (show counter open errors, etc)"),
1295 OPT_INTEGER('r', "repeat", &run_count,
1296 "repeat command and print average + stddev (max: 100, forever: 0)"),
1297 OPT_BOOLEAN('n', "null", &null_run,
1298 "null run - dont start any counters"),
1299 OPT_INCR('d', "detailed", &detailed_run,
1300 "detailed run - start a lot of events"),
1301 OPT_BOOLEAN('S', "sync", &sync_run,
1302 "call sync() before starting a run"),
1303 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1304 "print large numbers with thousands\' separators",
1306 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1307 "list of cpus to monitor in system-wide"),
1308 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1309 "disable CPU count aggregation", AGGR_NONE),
1310 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1311 "print counts with custom separator"),
1312 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1313 "monitor event in cgroup name only", parse_cgroups),
1314 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1315 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1316 OPT_INTEGER(0, "log-fd", &output_fd,
1317 "log output to fd, instead of stderr"),
1318 OPT_STRING(0, "pre", &pre_cmd, "command",
1319 "command to run prior to the measured command"),
1320 OPT_STRING(0, "post", &post_cmd, "command",
1321 "command to run after to the measured command"),
1322 OPT_UINTEGER('I', "interval-print", &interval,
1323 "print counts at regular interval in ms (>= 100)"),
1324 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1325 "aggregate counts per processor socket", AGGR_SOCKET),
1326 OPT_SET_UINT(0, "per-core", &aggr_mode,
1327 "aggregate counts per physical processor core", AGGR_CORE),
1328 OPT_UINTEGER('D', "delay", &initial_delay,
1329 "ms to wait before starting measurement after program start"),
1332 const char * const stat_usage[] = {
1333 "perf stat [<options>] [<command>]",
1336 int status = -EINVAL, run_idx;
1339 setlocale(LC_ALL, "");
1341 evsel_list = perf_evlist__new();
1342 if (evsel_list == NULL)
1345 argc = parse_options(argc, argv, options, stat_usage,
1346 PARSE_OPT_STOP_AT_NON_OPTION);
1349 if (output_name && strcmp(output_name, "-"))
1352 if (output_name && output_fd) {
1353 fprintf(stderr, "cannot use both --output and --log-fd\n");
1354 parse_options_usage(stat_usage, options, "o", 1);
1355 parse_options_usage(NULL, options, "log-fd", 0);
1359 if (output_fd < 0) {
1360 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1361 parse_options_usage(stat_usage, options, "log-fd", 0);
1367 mode = append_file ? "a" : "w";
1369 output = fopen(output_name, mode);
1371 perror("failed to create output file");
1374 clock_gettime(CLOCK_REALTIME, &tm);
1375 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1376 } else if (output_fd > 0) {
1377 mode = append_file ? "a" : "w";
1378 output = fdopen(output_fd, mode);
1380 perror("Failed opening logfd");
1387 if (!strcmp(csv_sep, "\\t"))
1390 csv_sep = DEFAULT_SEPARATOR;
1393 * let the spreadsheet do the pretty-printing
1396 /* User explicitly passed -B? */
1397 if (big_num_opt == 1) {
1398 fprintf(stderr, "-B option not supported with -x\n");
1399 parse_options_usage(stat_usage, options, "B", 1);
1400 parse_options_usage(NULL, options, "x", 1);
1402 } else /* Nope, so disable big number formatting */
1404 } else if (big_num_opt == 0) /* User passed --no-big-num */
1407 if (!argc && target__none(&target))
1408 usage_with_options(stat_usage, options);
1410 if (run_count < 0) {
1411 pr_err("Run count must be a positive number\n");
1412 parse_options_usage(stat_usage, options, "r", 1);
1414 } else if (run_count == 0) {
1419 /* no_aggr, cgroup are for system-wide only */
1420 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1421 !target__has_cpu(&target)) {
1422 fprintf(stderr, "both cgroup and no-aggregation "
1423 "modes only available in system-wide mode\n");
1425 parse_options_usage(stat_usage, options, "G", 1);
1426 parse_options_usage(NULL, options, "A", 1);
1427 parse_options_usage(NULL, options, "a", 1);
1431 if (add_default_attributes())
1434 target__validate(&target);
1436 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1437 if (target__has_task(&target)) {
1438 pr_err("Problems finding threads of monitor\n");
1439 parse_options_usage(stat_usage, options, "p", 1);
1440 parse_options_usage(NULL, options, "t", 1);
1441 } else if (target__has_cpu(&target)) {
1442 perror("failed to parse CPUs map");
1443 parse_options_usage(stat_usage, options, "C", 1);
1444 parse_options_usage(NULL, options, "a", 1);
1448 if (interval && interval < 100) {
1449 pr_err("print interval must be >= 100ms\n");
1450 parse_options_usage(stat_usage, options, "I", 1);
1454 if (perf_evlist__alloc_stats(evsel_list, interval))
1457 if (perf_stat_init_aggr_mode())
1461 * We dont want to block the signals - that would cause
1462 * child tasks to inherit that and Ctrl-C would not work.
1463 * What we want is for Ctrl-C to work in the exec()-ed
1464 * task, but being ignored by perf stat itself:
1468 signal(SIGINT, skip_signal);
1469 signal(SIGCHLD, skip_signal);
1470 signal(SIGALRM, skip_signal);
1471 signal(SIGABRT, skip_signal);
1474 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1475 if (run_count != 1 && verbose)
1476 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1479 status = run_perf_stat(argc, argv);
1480 if (forever && status != -1) {
1481 print_stat(argc, argv);
1482 perf_stat__reset_stats(evsel_list);
1486 if (!forever && status != -1 && !interval)
1487 print_stat(argc, argv);
1489 perf_evlist__free_stats(evsel_list);
1491 perf_evlist__delete(evsel_list);