1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin stat command: Give a precise performance counters summary
6 * overview about any workload, CPU or specific PID.
10 $ perf stat ./hackbench 10
14 Performance counter stats for './hackbench 10':
16 1708.761321 task-clock # 11.037 CPUs utilized
17 41,190 context-switches # 0.024 M/sec
18 6,735 CPU-migrations # 0.004 M/sec
19 17,318 page-faults # 0.010 M/sec
20 5,205,202,243 cycles # 3.046 GHz
21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
23 2,603,501,247 instructions # 0.50 insns per cycle
24 # 1.48 stalled cycles per insn
25 484,357,498 branches # 283.455 M/sec
26 6,388,934 branch-misses # 1.32% of all branches
28 0.154822978 seconds time elapsed
31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
33 * Improvements and fixes by:
35 * Arjan van de Ven <arjan@linux.intel.com>
36 * Yanmin Zhang <yanmin.zhang@intel.com>
37 * Wu Fengguang <fengguang.wu@intel.com>
38 * Mike Galbraith <efault@gmx.de>
39 * Paul Mackerras <paulus@samba.org>
40 * Jaswinder Singh Rajput <jaswinder@kernel.org>
45 #include "util/cgroup.h"
46 #include <subcmd/parse-options.h>
47 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evlist-hybrid.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread_map.h"
59 #include "util/counts.h"
60 #include "util/topdown.h"
61 #include "util/session.h"
62 #include "util/tool.h"
63 #include "util/string2.h"
64 #include "util/metricgroup.h"
65 #include "util/synthetic-events.h"
66 #include "util/target.h"
67 #include "util/time-utils.h"
69 #include "util/affinity.h"
71 #include "util/bpf_counter.h"
72 #include "util/iostat.h"
73 #include "util/pmu-hybrid.h"
76 #include <linux/time64.h>
77 #include <linux/zalloc.h>
78 #include <api/fs/fs.h>
82 #include <sys/prctl.h>
86 #include <sys/types.h>
91 #include <sys/resource.h>
92 #include <linux/err.h>
94 #include <linux/ctype.h>
95 #include <perf/evlist.h>
97 #define DEFAULT_SEPARATOR " "
98 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
100 static void print_counters(struct timespec *ts, int argc, const char **argv);
102 /* Default events used for perf stat -T */
103 static const char *transaction_attrs = {
115 /* More limited version when the CPU does not have all events. */
116 static const char * transaction_limited_attrs = {
126 static const char * topdown_attrs[] = {
127 "topdown-total-slots",
128 "topdown-slots-retired",
129 "topdown-recovery-bubbles",
130 "topdown-fetch-bubbles",
131 "topdown-slots-issued",
135 static const char *topdown_metric_attrs[] = {
144 static const char *topdown_metric_L2_attrs[] = {
151 "topdown-br-mispredict",
157 #define TOPDOWN_MAX_LEVEL 2
159 static const char *smi_cost_attrs = {
167 static struct evlist *evsel_list;
168 static bool all_counters_use_bpf = true;
170 static struct target target = {
174 #define METRIC_ONLY_LEN 20
176 static volatile pid_t child_pid = -1;
177 static int detailed_run = 0;
178 static bool transaction_run;
179 static bool topdown_run = false;
180 static bool smi_cost = false;
181 static bool smi_reset = false;
182 static int big_num_opt = -1;
183 static bool group = false;
184 static const char *pre_cmd = NULL;
185 static const char *post_cmd = NULL;
186 static bool sync_run = false;
187 static bool forever = false;
188 static bool force_metric_only = false;
189 static struct timespec ref_time;
190 static bool append_file;
191 static bool interval_count;
192 static const char *output_name;
193 static int output_fd;
197 struct perf_data data;
198 struct perf_session *session;
200 struct perf_tool tool;
202 struct perf_cpu_map *cpus;
203 struct perf_thread_map *threads;
204 enum aggr_mode aggr_mode;
207 static struct perf_stat perf_stat;
208 #define STAT_RECORD perf_stat.record
210 static volatile int done = 0;
212 static struct perf_stat_config stat_config = {
213 .aggr_mode = AGGR_GLOBAL,
215 .unit_width = 4, /* strlen("unit") */
217 .metric_only_len = METRIC_ONLY_LEN,
218 .walltime_nsecs_stats = &walltime_nsecs_stats,
225 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
227 if (!a->core.cpus && !b->core.cpus)
230 if (!a->core.cpus || !b->core.cpus)
233 if (a->core.cpus->nr != b->core.cpus->nr)
236 for (int i = 0; i < a->core.cpus->nr; i++) {
237 if (a->core.cpus->map[i] != b->core.cpus->map[i])
244 static void evlist__check_cpu_maps(struct evlist *evlist)
246 struct evsel *evsel, *pos, *leader;
249 if (evlist__has_hybrid(evlist))
250 evlist__warn_hybrid_group(evlist);
252 evlist__for_each_entry(evlist, evsel) {
253 leader = evsel__leader(evsel);
255 /* Check that leader matches cpus with each member. */
258 if (cpus_map_matched(leader, evsel))
261 /* If there's mismatch disable the group and warn user. */
262 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
263 evsel__group_desc(leader, buf, sizeof(buf));
264 pr_warning(" %s\n", buf);
267 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
268 pr_warning(" %s: %s\n", leader->name, buf);
269 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
270 pr_warning(" %s: %s\n", evsel->name, buf);
273 for_each_group_evsel(pos, leader) {
274 evsel__set_leader(pos, pos);
275 pos->core.nr_members = 0;
277 evsel->core.leader->nr_members = 0;
281 static inline void diff_timespec(struct timespec *r, struct timespec *a,
284 r->tv_sec = a->tv_sec - b->tv_sec;
285 if (a->tv_nsec < b->tv_nsec) {
286 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
289 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
293 static void perf_stat__reset_stats(void)
297 evlist__reset_stats(evsel_list);
298 perf_stat__reset_shadow_stats();
300 for (i = 0; i < stat_config.stats_num; i++)
301 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
304 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
305 union perf_event *event,
306 struct perf_sample *sample __maybe_unused,
307 struct machine *machine __maybe_unused)
309 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
310 pr_err("failed to write perf data, error: %m\n");
314 perf_stat.bytes_written += event->header.size;
318 static int write_stat_round_event(u64 tm, u64 type)
320 return perf_event__synthesize_stat_round(NULL, tm, type,
321 process_synthesized_event,
325 #define WRITE_STAT_ROUND_EVENT(time, interval) \
326 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
328 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
330 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
331 struct perf_counts_values *count)
333 struct perf_sample_id *sid = SID(counter, cpu, thread);
335 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
336 process_synthesized_event, NULL);
339 static int read_single_counter(struct evsel *counter, int cpu,
340 int thread, struct timespec *rs)
342 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
343 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
344 struct perf_counts_values *count =
345 perf_counts(counter->counts, cpu, thread);
346 count->ena = count->run = val;
350 return evsel__read_counter(counter, cpu, thread);
354 * Read out the results of a single counter:
355 * do not aggregate counts across CPUs in system-wide mode
357 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
359 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
362 if (!counter->supported)
365 if (counter->core.system_wide)
368 for (thread = 0; thread < nthreads; thread++) {
369 struct perf_counts_values *count;
371 count = perf_counts(counter->counts, cpu, thread);
374 * The leader's group read loads data into its group members
375 * (via evsel__read_counter()) and sets their count->loaded.
377 if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
378 read_single_counter(counter, cpu, thread, rs)) {
379 counter->counts->scaled = -1;
380 perf_counts(counter->counts, cpu, thread)->ena = 0;
381 perf_counts(counter->counts, cpu, thread)->run = 0;
385 perf_counts__set_loaded(counter->counts, cpu, thread, false);
388 if (evsel__write_stat_event(counter, cpu, thread, count)) {
389 pr_err("failed to write stat event\n");
395 fprintf(stat_config.output,
396 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
397 evsel__name(counter),
399 count->val, count->ena, count->run);
406 static int read_affinity_counters(struct timespec *rs)
408 struct evsel *counter;
409 struct affinity affinity;
412 if (all_counters_use_bpf)
415 if (affinity__setup(&affinity) < 0)
418 ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
419 if (!target__has_cpu(&target) || target__has_per_thread(&target))
421 evlist__for_each_cpu(evsel_list, i, cpu) {
424 affinity__set(&affinity, cpu);
426 evlist__for_each_entry(evsel_list, counter) {
427 if (evsel__cpu_iter_skip(counter, cpu))
429 if (evsel__is_bpf(counter))
432 counter->err = read_counter_cpu(counter, rs,
433 counter->cpu_iter - 1);
437 affinity__cleanup(&affinity);
441 static int read_bpf_map_counters(void)
443 struct evsel *counter;
446 evlist__for_each_entry(evsel_list, counter) {
447 if (!evsel__is_bpf(counter))
450 err = bpf_counter__read(counter);
457 static void read_counters(struct timespec *rs)
459 struct evsel *counter;
461 if (!stat_config.stop_read_counter) {
462 if (read_bpf_map_counters() ||
463 read_affinity_counters(rs))
467 evlist__for_each_entry(evsel_list, counter) {
469 pr_debug("failed to read counter %s\n", counter->name);
470 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
471 pr_warning("failed to process counter %s\n", counter->name);
476 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
480 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
484 config->stats_num = nthreads;
486 for (i = 0; i < nthreads; i++)
487 runtime_stat__init(&config->stats[i]);
492 static void runtime_stat_delete(struct perf_stat_config *config)
499 for (i = 0; i < config->stats_num; i++)
500 runtime_stat__exit(&config->stats[i]);
502 zfree(&config->stats);
505 static void runtime_stat_reset(struct perf_stat_config *config)
512 for (i = 0; i < config->stats_num; i++)
513 perf_stat__reset_shadow_per_stat(&config->stats[i]);
516 static void process_interval(void)
518 struct timespec ts, rs;
520 clock_gettime(CLOCK_MONOTONIC, &ts);
521 diff_timespec(&rs, &ts, &ref_time);
523 perf_stat__reset_shadow_per_stat(&rt_stat);
524 runtime_stat_reset(&stat_config);
528 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
529 pr_err("failed to write stat round event\n");
532 init_stats(&walltime_nsecs_stats);
533 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
534 print_counters(&rs, 0, NULL);
537 static bool handle_interval(unsigned int interval, int *times)
541 if (interval_count && !(--(*times)))
547 static int enable_counters(void)
552 evlist__for_each_entry(evsel_list, evsel) {
553 if (!evsel__is_bpf(evsel))
556 err = bpf_counter__enable(evsel);
561 if (stat_config.initial_delay < 0) {
562 pr_info(EVLIST_DISABLED_MSG);
566 if (stat_config.initial_delay > 0) {
567 pr_info(EVLIST_DISABLED_MSG);
568 usleep(stat_config.initial_delay * USEC_PER_MSEC);
572 * We need to enable counters only if:
573 * - we don't have tracee (attaching to task or cpu)
574 * - we have initial delay configured
576 if (!target__none(&target) || stat_config.initial_delay) {
577 if (!all_counters_use_bpf)
578 evlist__enable(evsel_list);
579 if (stat_config.initial_delay > 0)
580 pr_info(EVLIST_ENABLED_MSG);
585 static void disable_counters(void)
587 struct evsel *counter;
590 * If we don't have tracee (attaching to task or cpu), counters may
591 * still be running. To get accurate group ratios, we must stop groups
592 * from counting before reading their constituent counters.
594 if (!target__none(&target)) {
595 evlist__for_each_entry(evsel_list, counter)
596 bpf_counter__disable(counter);
597 if (!all_counters_use_bpf)
598 evlist__disable(evsel_list);
602 static volatile int workload_exec_errno;
605 * evlist__prepare_workload will send a SIGUSR1
606 * if the fork fails, since we asked by setting its
607 * want_signal to true.
609 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
610 void *ucontext __maybe_unused)
612 workload_exec_errno = info->si_value.sival_int;
615 static bool evsel__should_store_id(struct evsel *counter)
617 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
620 static bool is_target_alive(struct target *_target,
621 struct perf_thread_map *threads)
626 if (!target__has_task(_target))
629 for (i = 0; i < threads->nr; i++) {
632 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
633 threads->map[i].pid);
635 if (!stat(path, &st))
642 static void process_evlist(struct evlist *evlist, unsigned int interval)
644 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
646 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
648 case EVLIST_CTL_CMD_ENABLE:
652 case EVLIST_CTL_CMD_DISABLE:
656 case EVLIST_CTL_CMD_SNAPSHOT:
657 case EVLIST_CTL_CMD_ACK:
658 case EVLIST_CTL_CMD_UNSUPPORTED:
659 case EVLIST_CTL_CMD_EVLIST:
660 case EVLIST_CTL_CMD_STOP:
661 case EVLIST_CTL_CMD_PING:
668 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
671 int tts = *time_to_sleep;
672 struct timespec time_diff;
674 diff_timespec(&time_diff, time_stop, time_start);
676 tts -= time_diff.tv_sec * MSEC_PER_SEC +
677 time_diff.tv_nsec / NSEC_PER_MSEC;
682 *time_to_sleep = tts;
685 static int dispatch_events(bool forks, int timeout, int interval, int *times)
687 int child_exited = 0, status = 0;
688 int time_to_sleep, sleep_time;
689 struct timespec time_start, time_stop;
692 sleep_time = interval;
694 sleep_time = timeout;
698 time_to_sleep = sleep_time;
702 child_exited = waitpid(child_pid, &status, WNOHANG);
704 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
709 clock_gettime(CLOCK_MONOTONIC, &time_start);
710 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
711 if (timeout || handle_interval(interval, times))
713 time_to_sleep = sleep_time;
714 } else { /* fd revent */
715 process_evlist(evsel_list, interval);
716 clock_gettime(CLOCK_MONOTONIC, &time_stop);
717 compute_tts(&time_start, &time_stop, &time_to_sleep);
724 enum counter_recovery {
730 static enum counter_recovery stat_handle_error(struct evsel *counter)
734 * PPC returns ENXIO for HW counters until 2.6.37
735 * (behavior changed with commit b0a873e).
737 if (errno == EINVAL || errno == ENOSYS ||
738 errno == ENOENT || errno == EOPNOTSUPP ||
741 ui__warning("%s event is not supported by the kernel.\n",
742 evsel__name(counter));
743 counter->supported = false;
745 * errored is a sticky flag that means one of the counter's
746 * cpu event had a problem and needs to be reexamined.
748 counter->errored = true;
750 if ((evsel__leader(counter) != counter) ||
751 !(counter->core.leader->nr_members > 1))
753 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
755 ui__warning("%s\n", msg);
756 return COUNTER_RETRY;
757 } else if (target__has_per_thread(&target) &&
758 evsel_list->core.threads &&
759 evsel_list->core.threads->err_thread != -1) {
761 * For global --per-thread case, skip current
764 if (!thread_map__remove(evsel_list->core.threads,
765 evsel_list->core.threads->err_thread)) {
766 evsel_list->core.threads->err_thread = -1;
767 return COUNTER_RETRY;
771 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
772 ui__error("%s\n", msg);
775 kill(child_pid, SIGTERM);
776 return COUNTER_FATAL;
779 static int __run_perf_stat(int argc, const char **argv, int run_idx)
781 int interval = stat_config.interval;
782 int times = stat_config.times;
783 int timeout = stat_config.timeout;
785 unsigned long long t0, t1;
786 struct evsel *counter;
789 const bool forks = (argc > 0);
790 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
791 struct affinity affinity;
793 bool second_pass = false;
796 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
797 perror("failed to prepare workload");
800 child_pid = evsel_list->workload.pid;
804 evlist__set_leader(evsel_list);
806 if (affinity__setup(&affinity) < 0)
809 evlist__for_each_entry(evsel_list, counter) {
810 if (bpf_counter__load(counter, &target))
812 if (!evsel__is_bpf(counter))
813 all_counters_use_bpf = false;
816 evlist__for_each_cpu (evsel_list, i, cpu) {
818 * bperf calls evsel__open_per_cpu() in bperf__load(), so
819 * no need to call it again here.
823 affinity__set(&affinity, cpu);
825 evlist__for_each_entry(evsel_list, counter) {
826 if (evsel__cpu_iter_skip(counter, cpu))
828 if (counter->reset_group || counter->errored)
830 if (evsel__is_bpf(counter))
833 if (create_perf_stat_counter(counter, &stat_config, &target,
834 counter->cpu_iter - 1) < 0) {
837 * Weak group failed. We cannot just undo this here
838 * because earlier CPUs might be in group mode, and the kernel
839 * doesn't support mixing group and non group reads. Defer
841 * Don't close here because we're in the wrong affinity.
843 if ((errno == EINVAL || errno == EBADF) &&
844 evsel__leader(counter) != counter &&
845 counter->weak_group) {
846 evlist__reset_weak_group(evsel_list, counter, false);
847 assert(counter->reset_group);
852 switch (stat_handle_error(counter)) {
864 counter->supported = true;
870 * Now redo all the weak group after closing them,
871 * and also close errored counters.
874 evlist__for_each_cpu(evsel_list, i, cpu) {
875 affinity__set(&affinity, cpu);
876 /* First close errored or weak retry */
877 evlist__for_each_entry(evsel_list, counter) {
878 if (!counter->reset_group && !counter->errored)
880 if (evsel__cpu_iter_skip_no_inc(counter, cpu))
882 perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
884 /* Now reopen weak */
885 evlist__for_each_entry(evsel_list, counter) {
886 if (!counter->reset_group && !counter->errored)
888 if (evsel__cpu_iter_skip(counter, cpu))
890 if (!counter->reset_group)
893 pr_debug2("reopening weak %s\n", evsel__name(counter));
894 if (create_perf_stat_counter(counter, &stat_config, &target,
895 counter->cpu_iter - 1) < 0) {
897 switch (stat_handle_error(counter)) {
901 goto try_again_reset;
908 counter->supported = true;
912 affinity__cleanup(&affinity);
914 evlist__for_each_entry(evsel_list, counter) {
915 if (!counter->supported) {
916 perf_evsel__free_fd(&counter->core);
920 l = strlen(counter->unit);
921 if (l > stat_config.unit_width)
922 stat_config.unit_width = l;
924 if (evsel__should_store_id(counter) &&
925 evsel__store_ids(counter, evsel_list))
929 if (evlist__apply_filters(evsel_list, &counter)) {
930 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
931 counter->filter, evsel__name(counter), errno,
932 str_error_r(errno, msg, sizeof(msg)));
937 int fd = perf_data__fd(&perf_stat.data);
940 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
942 err = perf_session__write_header(perf_stat.session, evsel_list,
949 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
950 process_synthesized_event, is_pipe);
956 * Enable counters and exec the command:
959 err = enable_counters();
962 evlist__start_workload(evsel_list);
965 clock_gettime(CLOCK_MONOTONIC, &ref_time);
967 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
968 status = dispatch_events(forks, timeout, interval, ×);
969 if (child_pid != -1) {
971 kill(child_pid, SIGTERM);
972 wait4(child_pid, &status, 0, &stat_config.ru_data);
975 if (workload_exec_errno) {
976 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
977 pr_err("Workload failed: %s\n", emsg);
981 if (WIFSIGNALED(status))
982 psignal(WTERMSIG(status), argv[0]);
984 err = enable_counters();
989 clock_gettime(CLOCK_MONOTONIC, &ref_time);
991 status = dispatch_events(forks, timeout, interval, ×);
998 if (stat_config.walltime_run_table)
999 stat_config.walltime_run[run_idx] = t1 - t0;
1001 if (interval && stat_config.summary) {
1002 stat_config.interval = 0;
1003 stat_config.stop_read_counter = true;
1004 init_stats(&walltime_nsecs_stats);
1005 update_stats(&walltime_nsecs_stats, t1 - t0);
1007 if (stat_config.aggr_mode == AGGR_GLOBAL)
1008 evlist__save_aggr_prev_raw_counts(evsel_list);
1010 evlist__copy_prev_raw_counts(evsel_list);
1011 evlist__reset_prev_raw_counts(evsel_list);
1012 runtime_stat_reset(&stat_config);
1013 perf_stat__reset_shadow_per_stat(&rt_stat);
1015 update_stats(&walltime_nsecs_stats, t1 - t0);
1018 * Closing a group leader splits the group, and as we only disable
1019 * group leaders, results in remaining events becoming enabled. To
1020 * avoid arbitrary skew, we must read all counters before closing any
1023 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1026 * We need to keep evsel_list alive, because it's processed
1027 * later the evsel_list will be closed after.
1030 evlist__close(evsel_list);
1032 return WEXITSTATUS(status);
1035 static int run_perf_stat(int argc, const char **argv, int run_idx)
1040 ret = system(pre_cmd);
1048 ret = __run_perf_stat(argc, argv, run_idx);
1053 ret = system(post_cmd);
1061 static void print_counters(struct timespec *ts, int argc, const char **argv)
1063 /* Do not print anything if we record to the pipe. */
1064 if (STAT_RECORD && perf_stat.data.is_pipe)
1066 if (stat_config.quiet)
1069 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1072 static volatile int signr = -1;
1074 static void skip_signal(int signo)
1076 if ((child_pid == -1) || stat_config.interval)
1081 * render child_pid harmless
1082 * won't send SIGTERM to a random
1083 * process in case of race condition
1084 * and fast PID recycling
1089 static void sig_atexit(void)
1094 * avoid race condition with SIGCHLD handler
1095 * in skip_signal() which is modifying child_pid
1096 * goal is to avoid send SIGTERM to a random
1100 sigaddset(&set, SIGCHLD);
1101 sigprocmask(SIG_BLOCK, &set, &oset);
1103 if (child_pid != -1)
1104 kill(child_pid, SIGTERM);
1106 sigprocmask(SIG_SETMASK, &oset, NULL);
1111 signal(signr, SIG_DFL);
1112 kill(getpid(), signr);
1115 void perf_stat__set_big_num(int set)
1117 stat_config.big_num = (set != 0);
1120 void perf_stat__set_no_csv_summary(int set)
1122 stat_config.no_csv_summary = (set != 0);
1125 static int stat__set_big_num(const struct option *opt __maybe_unused,
1126 const char *s __maybe_unused, int unset)
1128 big_num_opt = unset ? 0 : 1;
1129 perf_stat__set_big_num(!unset);
1133 static int enable_metric_only(const struct option *opt __maybe_unused,
1134 const char *s __maybe_unused, int unset)
1136 force_metric_only = true;
1137 stat_config.metric_only = !unset;
1141 static int parse_metric_groups(const struct option *opt,
1143 int unset __maybe_unused)
1145 return metricgroup__parse_groups(opt, str,
1146 stat_config.metric_no_group,
1147 stat_config.metric_no_merge,
1148 &stat_config.metric_events);
1151 static int parse_control_option(const struct option *opt,
1153 int unset __maybe_unused)
1155 struct perf_stat_config *config = opt->value;
1157 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1160 static int parse_stat_cgroups(const struct option *opt,
1161 const char *str, int unset)
1163 if (stat_config.cgroup_list) {
1164 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1168 return parse_cgroups(opt, str, unset);
1171 static struct option stat_options[] = {
1172 OPT_BOOLEAN('T', "transaction", &transaction_run,
1173 "hardware transaction statistics"),
1174 OPT_CALLBACK('e', "event", &evsel_list, "event",
1175 "event selector. use 'perf list' to list available events",
1176 parse_events_option),
1177 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1178 "event filter", parse_filter),
1179 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1180 "child tasks do not inherit counters"),
1181 OPT_STRING('p', "pid", &target.pid, "pid",
1182 "stat events on existing process id"),
1183 OPT_STRING('t', "tid", &target.tid, "tid",
1184 "stat events on existing thread id"),
1185 #ifdef HAVE_BPF_SKEL
1186 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1187 "stat events on existing bpf program id"),
1188 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1189 "use bpf program to count events"),
1190 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1191 "path to perf_event_attr map"),
1193 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1194 "system-wide collection from all CPUs"),
1195 OPT_BOOLEAN('g', "group", &group,
1196 "put the counters into a counter group"),
1197 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1198 "Use --no-scale to disable counter scaling for multiplexing"),
1199 OPT_INCR('v', "verbose", &verbose,
1200 "be more verbose (show counter open errors, etc)"),
1201 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1202 "repeat command and print average + stddev (max: 100, forever: 0)"),
1203 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1204 "display details about each run (only with -r option)"),
1205 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1206 "null run - dont start any counters"),
1207 OPT_INCR('d', "detailed", &detailed_run,
1208 "detailed run - start a lot of events"),
1209 OPT_BOOLEAN('S', "sync", &sync_run,
1210 "call sync() before starting a run"),
1211 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1212 "print large numbers with thousands\' separators",
1214 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1215 "list of cpus to monitor in system-wide"),
1216 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1217 "disable CPU count aggregation", AGGR_NONE),
1218 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1219 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1220 "print counts with custom separator"),
1221 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1222 "monitor event in cgroup name only", parse_stat_cgroups),
1223 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1224 "expand events for each cgroup"),
1225 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1226 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1227 OPT_INTEGER(0, "log-fd", &output_fd,
1228 "log output to fd, instead of stderr"),
1229 OPT_STRING(0, "pre", &pre_cmd, "command",
1230 "command to run prior to the measured command"),
1231 OPT_STRING(0, "post", &post_cmd, "command",
1232 "command to run after to the measured command"),
1233 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1234 "print counts at regular interval in ms "
1235 "(overhead is possible for values <= 100ms)"),
1236 OPT_INTEGER(0, "interval-count", &stat_config.times,
1237 "print counts for fixed number of times"),
1238 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1239 "clear screen in between new interval"),
1240 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1241 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1242 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1243 "aggregate counts per processor socket", AGGR_SOCKET),
1244 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1245 "aggregate counts per processor die", AGGR_DIE),
1246 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1247 "aggregate counts per physical processor core", AGGR_CORE),
1248 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1249 "aggregate counts per thread", AGGR_THREAD),
1250 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1251 "aggregate counts per numa node", AGGR_NODE),
1252 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1253 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1254 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1255 "Only print computed metrics. No raw values", enable_metric_only),
1256 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1257 "don't group metric events, impacts multiplexing"),
1258 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1259 "don't try to share events between metrics in a group"),
1260 OPT_BOOLEAN(0, "topdown", &topdown_run,
1261 "measure top-down statistics"),
1262 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1263 "Set the metrics level for the top-down statistics (0: max level)"),
1264 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1265 "measure SMI cost"),
1266 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1267 "monitor specified metrics or metric groups (separated by ,)",
1268 parse_metric_groups),
1269 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1270 "Configure all used events to run in kernel space.",
1271 PARSE_OPT_EXCLUSIVE),
1272 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1273 "Configure all used events to run in user space.",
1274 PARSE_OPT_EXCLUSIVE),
1275 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1276 "Use with 'percore' event qualifier to show the event "
1277 "counts of one hardware thread by sum up total hardware "
1278 "threads of same physical core"),
1279 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1280 "print summary for interval mode"),
1281 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1282 "don't print 'summary' for CSV summary output"),
1283 OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1284 "don't print output (useful with record)"),
1286 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1287 "libpfm4 event selector. use 'perf list' to list available events",
1288 parse_libpfm_events_option),
1290 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1291 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1292 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1293 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1294 parse_control_option),
1295 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1296 "measure I/O performance metrics provided by arch/platform",
1301 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1302 struct perf_cpu_map *map, int cpu)
1304 return cpu_map__get_socket(map, cpu, NULL);
1307 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1308 struct perf_cpu_map *map, int cpu)
1310 return cpu_map__get_die(map, cpu, NULL);
1313 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1314 struct perf_cpu_map *map, int cpu)
1316 return cpu_map__get_core(map, cpu, NULL);
1319 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1320 struct perf_cpu_map *map, int cpu)
1322 return cpu_map__get_node(map, cpu, NULL);
1325 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1326 aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1329 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1334 cpu = map->map[idx];
1336 if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu]))
1337 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1339 id = config->cpus_aggr_map->map[cpu];
1343 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1344 struct perf_cpu_map *map, int idx)
1346 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1349 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1350 struct perf_cpu_map *map, int idx)
1352 return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1355 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1356 struct perf_cpu_map *map, int idx)
1358 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1361 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1362 struct perf_cpu_map *map, int idx)
1364 return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1367 static bool term_percore_set(void)
1369 struct evsel *counter;
1371 evlist__for_each_entry(evsel_list, counter) {
1372 if (counter->percore)
1379 static int perf_stat_init_aggr_mode(void)
1383 switch (stat_config.aggr_mode) {
1385 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1386 perror("cannot build socket map");
1389 stat_config.aggr_get_id = perf_stat__get_socket_cached;
1392 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1393 perror("cannot build die map");
1396 stat_config.aggr_get_id = perf_stat__get_die_cached;
1399 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1400 perror("cannot build core map");
1403 stat_config.aggr_get_id = perf_stat__get_core_cached;
1406 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1407 perror("cannot build core map");
1410 stat_config.aggr_get_id = perf_stat__get_node_cached;
1413 if (term_percore_set()) {
1414 if (cpu_map__build_core_map(evsel_list->core.cpus,
1415 &stat_config.aggr_map)) {
1416 perror("cannot build core map");
1419 stat_config.aggr_get_id = perf_stat__get_core_cached;
1430 * The evsel_list->cpus is the base we operate on,
1431 * taking the highest cpu number to be the size of
1432 * the aggregation translate cpumap.
1434 nr = perf_cpu_map__max(evsel_list->core.cpus);
1435 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1436 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1439 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1442 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1443 "cpu_aggr_map refcnt unbalanced\n");
1448 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1450 if (map && refcount_dec_and_test(&map->refcnt))
1451 cpu_aggr_map__delete(map);
1454 static void perf_stat__exit_aggr_mode(void)
1456 cpu_aggr_map__put(stat_config.aggr_map);
1457 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1458 stat_config.aggr_map = NULL;
1459 stat_config.cpus_aggr_map = NULL;
1462 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1469 cpu = map->map[idx];
1471 if (cpu >= env->nr_cpus_avail)
1477 static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1479 struct perf_env *env = data;
1480 int cpu = perf_env__get_cpu(env, map, idx);
1481 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1484 id.socket = env->cpu[cpu].socket_id;
1489 static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1491 struct perf_env *env = data;
1492 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1493 int cpu = perf_env__get_cpu(env, map, idx);
1497 * die_id is relative to socket, so start
1498 * with the socket ID and then add die to
1501 id.socket = env->cpu[cpu].socket_id;
1502 id.die = env->cpu[cpu].die_id;
1508 static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1510 struct perf_env *env = data;
1511 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1512 int cpu = perf_env__get_cpu(env, map, idx);
1516 * core_id is relative to socket and die,
1517 * we need a global id. So we set
1518 * socket, die id and core id
1520 id.socket = env->cpu[cpu].socket_id;
1521 id.die = env->cpu[cpu].die_id;
1522 id.core = env->cpu[cpu].core_id;
1528 static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1530 int cpu = perf_env__get_cpu(data, map, idx);
1531 struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1533 id.node = perf_env__numa_node(data, cpu);
1537 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1538 struct cpu_aggr_map **sockp)
1540 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1543 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1544 struct cpu_aggr_map **diep)
1546 return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1549 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1550 struct cpu_aggr_map **corep)
1552 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1555 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1556 struct cpu_aggr_map **nodep)
1558 return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1561 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1562 struct perf_cpu_map *map, int idx)
1564 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1566 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1567 struct perf_cpu_map *map, int idx)
1569 return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1572 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1573 struct perf_cpu_map *map, int idx)
1575 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1578 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1579 struct perf_cpu_map *map, int idx)
1581 return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1584 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1586 struct perf_env *env = &st->session->header.env;
1588 switch (stat_config.aggr_mode) {
1590 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1591 perror("cannot build socket map");
1594 stat_config.aggr_get_id = perf_stat__get_socket_file;
1597 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1598 perror("cannot build die map");
1601 stat_config.aggr_get_id = perf_stat__get_die_file;
1604 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1605 perror("cannot build core map");
1608 stat_config.aggr_get_id = perf_stat__get_core_file;
1611 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1612 perror("cannot build core map");
1615 stat_config.aggr_get_id = perf_stat__get_node_file;
1629 * Add default attributes, if there were no attributes specified or
1630 * if -d/--detailed, -d -d or -d -d -d is used:
1632 static int add_default_attributes(void)
1635 struct perf_event_attr default_attrs0[] = {
1637 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1638 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1639 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1640 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1642 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1644 struct perf_event_attr frontend_attrs[] = {
1645 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1647 struct perf_event_attr backend_attrs[] = {
1648 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1650 struct perf_event_attr default_attrs1[] = {
1651 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1652 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1653 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1656 struct perf_event_attr default_sw_attrs[] = {
1657 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1658 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1659 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1660 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1664 * Detailed stats (-d), covering the L1 and last level data caches:
1666 struct perf_event_attr detailed_attrs[] = {
1668 { .type = PERF_TYPE_HW_CACHE,
1670 PERF_COUNT_HW_CACHE_L1D << 0 |
1671 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1672 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1674 { .type = PERF_TYPE_HW_CACHE,
1676 PERF_COUNT_HW_CACHE_L1D << 0 |
1677 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1678 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1680 { .type = PERF_TYPE_HW_CACHE,
1682 PERF_COUNT_HW_CACHE_LL << 0 |
1683 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1684 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1686 { .type = PERF_TYPE_HW_CACHE,
1688 PERF_COUNT_HW_CACHE_LL << 0 |
1689 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1690 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1694 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1696 struct perf_event_attr very_detailed_attrs[] = {
1698 { .type = PERF_TYPE_HW_CACHE,
1700 PERF_COUNT_HW_CACHE_L1I << 0 |
1701 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1702 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1704 { .type = PERF_TYPE_HW_CACHE,
1706 PERF_COUNT_HW_CACHE_L1I << 0 |
1707 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1708 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1710 { .type = PERF_TYPE_HW_CACHE,
1712 PERF_COUNT_HW_CACHE_DTLB << 0 |
1713 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1714 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1716 { .type = PERF_TYPE_HW_CACHE,
1718 PERF_COUNT_HW_CACHE_DTLB << 0 |
1719 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1720 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1722 { .type = PERF_TYPE_HW_CACHE,
1724 PERF_COUNT_HW_CACHE_ITLB << 0 |
1725 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1726 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1728 { .type = PERF_TYPE_HW_CACHE,
1730 PERF_COUNT_HW_CACHE_ITLB << 0 |
1731 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1732 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1737 * Very, very detailed stats (-d -d -d), adding prefetch events:
1739 struct perf_event_attr very_very_detailed_attrs[] = {
1741 { .type = PERF_TYPE_HW_CACHE,
1743 PERF_COUNT_HW_CACHE_L1D << 0 |
1744 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1745 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1747 { .type = PERF_TYPE_HW_CACHE,
1749 PERF_COUNT_HW_CACHE_L1D << 0 |
1750 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1751 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1753 struct parse_events_error errinfo;
1755 /* Set attrs if no event is selected and !null_run: */
1756 if (stat_config.null_run)
1759 bzero(&errinfo, sizeof(errinfo));
1760 if (transaction_run) {
1761 /* Handle -T as -M transaction. Once platform specific metrics
1762 * support has been added to the json files, all architectures
1763 * will use this approach. To determine transaction support
1764 * on an architecture test for such a metric name.
1766 if (metricgroup__has_metric("transaction")) {
1767 struct option opt = { .value = &evsel_list };
1769 return metricgroup__parse_groups(&opt, "transaction",
1770 stat_config.metric_no_group,
1771 stat_config.metric_no_merge,
1772 &stat_config.metric_events);
1775 if (pmu_have_event("cpu", "cycles-ct") &&
1776 pmu_have_event("cpu", "el-start"))
1777 err = parse_events(evsel_list, transaction_attrs,
1780 err = parse_events(evsel_list,
1781 transaction_limited_attrs,
1784 fprintf(stderr, "Cannot set up transaction events\n");
1785 parse_events_print_error(&errinfo, transaction_attrs);
1794 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1795 fprintf(stderr, "freeze_on_smi is not supported.\n");
1800 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1801 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1807 if (pmu_have_event("msr", "aperf") &&
1808 pmu_have_event("msr", "smi")) {
1809 if (!force_metric_only)
1810 stat_config.metric_only = true;
1811 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1813 fprintf(stderr, "To measure SMI cost, it needs "
1814 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1815 parse_events_print_error(&errinfo, smi_cost_attrs);
1819 parse_events_print_error(&errinfo, smi_cost_attrs);
1820 fprintf(stderr, "Cannot set up SMI cost events\n");
1827 const char **metric_attrs = topdown_metric_attrs;
1828 unsigned int max_level = 1;
1832 if (!force_metric_only)
1833 stat_config.metric_only = true;
1835 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1836 metric_attrs = topdown_metric_L2_attrs;
1840 if (stat_config.topdown_level > max_level) {
1841 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1843 } else if (!stat_config.topdown_level)
1844 stat_config.topdown_level = max_level;
1846 if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1847 pr_err("Out of memory\n");
1850 if (metric_attrs[0] && str) {
1851 if (!stat_config.interval && !stat_config.metric_only) {
1852 fprintf(stat_config.output,
1853 "Topdown accuracy may decrease when measuring long periods.\n"
1854 "Please print the result regularly, e.g. -I1000\n");
1861 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1862 stat_config.aggr_mode != AGGR_CORE) {
1863 pr_err("top down event configuration requires --per-core mode\n");
1866 stat_config.aggr_mode = AGGR_CORE;
1867 if (nr_cgroups || !target__has_cpu(&target)) {
1868 pr_err("top down event configuration requires system-wide mode (-a)\n");
1872 if (topdown_filter_events(topdown_attrs, &str,
1873 arch_topdown_check_group(&warn)) < 0) {
1874 pr_err("Out of memory\n");
1877 if (topdown_attrs[0] && str) {
1879 arch_topdown_group_warn();
1881 err = parse_events(evsel_list, str, &errinfo);
1884 "Cannot set up top down events %s: %d\n",
1886 parse_events_print_error(&errinfo, str);
1891 fprintf(stderr, "System does not support topdown\n");
1897 if (!evsel_list->core.nr_entries) {
1898 if (perf_pmu__has_hybrid()) {
1899 const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1901 if (target__has_cpu(&target))
1902 default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1904 if (evlist__add_default_attrs(evsel_list,
1905 default_sw_attrs) < 0) {
1909 err = parse_events(evsel_list, hybrid_str, &errinfo);
1912 "Cannot set up hybrid events %s: %d\n",
1914 parse_events_print_error(&errinfo, hybrid_str);
1920 if (target__has_cpu(&target))
1921 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1923 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1925 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1926 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1929 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1930 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1933 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1936 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1937 if (arch_evlist__add_default_attrs(evsel_list) < 0)
1941 /* Detailed events get appended to the event list: */
1943 if (detailed_run < 1)
1946 /* Append detailed run extra attributes: */
1947 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1950 if (detailed_run < 2)
1953 /* Append very detailed run extra attributes: */
1954 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1957 if (detailed_run < 3)
1960 /* Append very, very detailed run extra attributes: */
1961 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1964 static const char * const stat_record_usage[] = {
1965 "perf stat record [<options>]",
1969 static void init_features(struct perf_session *session)
1973 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1974 perf_header__set_feat(&session->header, feat);
1976 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1977 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1978 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1979 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1980 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1983 static int __cmd_record(int argc, const char **argv)
1985 struct perf_session *session;
1986 struct perf_data *data = &perf_stat.data;
1988 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1989 PARSE_OPT_STOP_AT_NON_OPTION);
1992 data->path = output_name;
1994 if (stat_config.run_count != 1 || forever) {
1995 pr_err("Cannot use -r option with perf stat record.\n");
1999 session = perf_session__new(data, NULL);
2000 if (IS_ERR(session)) {
2001 pr_err("Perf session creation failed\n");
2002 return PTR_ERR(session);
2005 init_features(session);
2007 session->evlist = evsel_list;
2008 perf_stat.session = session;
2009 perf_stat.record = true;
2013 static int process_stat_round_event(struct perf_session *session,
2014 union perf_event *event)
2016 struct perf_record_stat_round *stat_round = &event->stat_round;
2017 struct evsel *counter;
2018 struct timespec tsh, *ts = NULL;
2019 const char **argv = session->header.env.cmdline_argv;
2020 int argc = session->header.env.nr_cmdline;
2022 evlist__for_each_entry(evsel_list, counter)
2023 perf_stat_process_counter(&stat_config, counter);
2025 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2026 update_stats(&walltime_nsecs_stats, stat_round->time);
2028 if (stat_config.interval && stat_round->time) {
2029 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2030 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2034 print_counters(ts, argc, argv);
2039 int process_stat_config_event(struct perf_session *session,
2040 union perf_event *event)
2042 struct perf_tool *tool = session->tool;
2043 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2045 perf_event__read_stat_config(&stat_config, &event->stat_config);
2047 if (perf_cpu_map__empty(st->cpus)) {
2048 if (st->aggr_mode != AGGR_UNSET)
2049 pr_warning("warning: processing task data, aggregation mode not set\n");
2053 if (st->aggr_mode != AGGR_UNSET)
2054 stat_config.aggr_mode = st->aggr_mode;
2056 if (perf_stat.data.is_pipe)
2057 perf_stat_init_aggr_mode();
2059 perf_stat_init_aggr_mode_file(st);
2064 static int set_maps(struct perf_stat *st)
2066 if (!st->cpus || !st->threads)
2069 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2072 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2074 if (evlist__alloc_stats(evsel_list, true))
2077 st->maps_allocated = true;
2082 int process_thread_map_event(struct perf_session *session,
2083 union perf_event *event)
2085 struct perf_tool *tool = session->tool;
2086 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2089 pr_warning("Extra thread map event, ignoring.\n");
2093 st->threads = thread_map__new_event(&event->thread_map);
2097 return set_maps(st);
2101 int process_cpu_map_event(struct perf_session *session,
2102 union perf_event *event)
2104 struct perf_tool *tool = session->tool;
2105 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2106 struct perf_cpu_map *cpus;
2109 pr_warning("Extra cpu map event, ignoring.\n");
2113 cpus = cpu_map__new_data(&event->cpu_map.data);
2118 return set_maps(st);
2121 static const char * const stat_report_usage[] = {
2122 "perf stat report [<options>]",
2126 static struct perf_stat perf_stat = {
2128 .attr = perf_event__process_attr,
2129 .event_update = perf_event__process_event_update,
2130 .thread_map = process_thread_map_event,
2131 .cpu_map = process_cpu_map_event,
2132 .stat_config = process_stat_config_event,
2133 .stat = perf_event__process_stat_event,
2134 .stat_round = process_stat_round_event,
2136 .aggr_mode = AGGR_UNSET,
2139 static int __cmd_report(int argc, const char **argv)
2141 struct perf_session *session;
2142 const struct option options[] = {
2143 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2144 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2145 "aggregate counts per processor socket", AGGR_SOCKET),
2146 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2147 "aggregate counts per processor die", AGGR_DIE),
2148 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2149 "aggregate counts per physical processor core", AGGR_CORE),
2150 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2151 "aggregate counts per numa node", AGGR_NODE),
2152 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2153 "disable CPU count aggregation", AGGR_NONE),
2159 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2161 if (!input_name || !strlen(input_name)) {
2162 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2165 input_name = "perf.data";
2168 perf_stat.data.path = input_name;
2169 perf_stat.data.mode = PERF_DATA_MODE_READ;
2171 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2172 if (IS_ERR(session))
2173 return PTR_ERR(session);
2175 perf_stat.session = session;
2176 stat_config.output = stderr;
2177 evsel_list = session->evlist;
2179 ret = perf_session__process_events(session);
2183 perf_session__delete(session);
2187 static void setup_system_wide(int forks)
2190 * Make system wide (-a) the default target if
2191 * no target was specified and one of following
2192 * conditions is met:
2194 * - there's no workload specified
2195 * - there is workload specified but all requested
2196 * events are system wide events
2198 if (!target__none(&target))
2202 target.system_wide = true;
2204 struct evsel *counter;
2206 evlist__for_each_entry(evsel_list, counter) {
2207 if (!counter->core.system_wide &&
2208 strcmp(counter->name, "duration_time")) {
2213 if (evsel_list->core.nr_entries)
2214 target.system_wide = true;
2218 int cmd_stat(int argc, const char **argv)
2220 const char * const stat_usage[] = {
2221 "perf stat [<options>] [<command>]",
2224 int status = -EINVAL, run_idx, err;
2226 FILE *output = stderr;
2227 unsigned int interval, timeout;
2228 const char * const stat_subcommands[] = { "record", "report" };
2229 char errbuf[BUFSIZ];
2231 setlocale(LC_ALL, "");
2233 evsel_list = evlist__new();
2234 if (evsel_list == NULL)
2237 parse_events__shrink_config_terms();
2239 /* String-parsing callback-based options would segfault when negated */
2240 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2241 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2242 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2244 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2245 (const char **) stat_usage,
2246 PARSE_OPT_STOP_AT_NON_OPTION);
2247 perf_stat__collect_metric_expr(evsel_list);
2248 perf_stat__init_shadow_stats();
2250 if (stat_config.csv_sep) {
2251 stat_config.csv_output = true;
2252 if (!strcmp(stat_config.csv_sep, "\\t"))
2253 stat_config.csv_sep = "\t";
2255 stat_config.csv_sep = DEFAULT_SEPARATOR;
2257 if (argc && !strncmp(argv[0], "rec", 3)) {
2258 argc = __cmd_record(argc, argv);
2261 } else if (argc && !strncmp(argv[0], "rep", 3))
2262 return __cmd_report(argc, argv);
2264 interval = stat_config.interval;
2265 timeout = stat_config.timeout;
2268 * For record command the -o is already taken care of.
2270 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2273 if (output_name && output_fd) {
2274 fprintf(stderr, "cannot use both --output and --log-fd\n");
2275 parse_options_usage(stat_usage, stat_options, "o", 1);
2276 parse_options_usage(NULL, stat_options, "log-fd", 0);
2280 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2281 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2285 if (stat_config.metric_only && stat_config.run_count > 1) {
2286 fprintf(stderr, "--metric-only is not supported with -r\n");
2290 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2291 fprintf(stderr, "--table is only supported with -r\n");
2292 parse_options_usage(stat_usage, stat_options, "r", 1);
2293 parse_options_usage(NULL, stat_options, "table", 0);
2297 if (output_fd < 0) {
2298 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2299 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2303 if (!output && !stat_config.quiet) {
2305 mode = append_file ? "a" : "w";
2307 output = fopen(output_name, mode);
2309 perror("failed to create output file");
2312 clock_gettime(CLOCK_REALTIME, &tm);
2313 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2314 } else if (output_fd > 0) {
2315 mode = append_file ? "a" : "w";
2316 output = fdopen(output_fd, mode);
2318 perror("Failed opening logfd");
2323 stat_config.output = output;
2326 * let the spreadsheet do the pretty-printing
2328 if (stat_config.csv_output) {
2329 /* User explicitly passed -B? */
2330 if (big_num_opt == 1) {
2331 fprintf(stderr, "-B option not supported with -x\n");
2332 parse_options_usage(stat_usage, stat_options, "B", 1);
2333 parse_options_usage(NULL, stat_options, "x", 1);
2335 } else /* Nope, so disable big number formatting */
2336 stat_config.big_num = false;
2337 } else if (big_num_opt == 0) /* User passed --no-big-num */
2338 stat_config.big_num = false;
2340 err = target__validate(&target);
2342 target__strerror(&target, err, errbuf, BUFSIZ);
2343 pr_warning("%s\n", errbuf);
2346 setup_system_wide(argc);
2349 * Display user/system times only for single
2350 * run and when there's specified tracee.
2352 if ((stat_config.run_count == 1) && target__none(&target))
2353 stat_config.ru_display = true;
2355 if (stat_config.run_count < 0) {
2356 pr_err("Run count must be a positive number\n");
2357 parse_options_usage(stat_usage, stat_options, "r", 1);
2359 } else if (stat_config.run_count == 0) {
2361 stat_config.run_count = 1;
2364 if (stat_config.walltime_run_table) {
2365 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2366 if (!stat_config.walltime_run) {
2367 pr_err("failed to setup -r option");
2372 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2373 !target__has_task(&target)) {
2374 if (!target.system_wide || target.cpu_list) {
2375 fprintf(stderr, "The --per-thread option is only "
2376 "available when monitoring via -p -t -a "
2377 "options or only --per-thread.\n");
2378 parse_options_usage(NULL, stat_options, "p", 1);
2379 parse_options_usage(NULL, stat_options, "t", 1);
2385 * no_aggr, cgroup are for system-wide only
2386 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2388 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2389 stat_config.aggr_mode != AGGR_THREAD) ||
2390 (nr_cgroups || stat_config.cgroup_list)) &&
2391 !target__has_cpu(&target)) {
2392 fprintf(stderr, "both cgroup and no-aggregation "
2393 "modes only available in system-wide mode\n");
2395 parse_options_usage(stat_usage, stat_options, "G", 1);
2396 parse_options_usage(NULL, stat_options, "A", 1);
2397 parse_options_usage(NULL, stat_options, "a", 1);
2398 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2402 if (stat_config.iostat_run) {
2403 status = iostat_prepare(evsel_list, &stat_config);
2406 if (iostat_mode == IOSTAT_LIST) {
2407 iostat_list(evsel_list, &stat_config);
2410 iostat_list(evsel_list, &stat_config);
2411 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2412 target.system_wide = true;
2415 if (add_default_attributes())
2418 if (stat_config.cgroup_list) {
2419 if (nr_cgroups > 0) {
2420 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2421 parse_options_usage(stat_usage, stat_options, "G", 1);
2422 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2426 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2427 &stat_config.metric_events, true) < 0) {
2428 parse_options_usage(stat_usage, stat_options,
2429 "for-each-cgroup", 0);
2434 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2435 target.per_thread = true;
2437 if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2438 pr_err("failed to use cpu list %s\n", target.cpu_list);
2442 target.hybrid = perf_pmu__has_hybrid();
2443 if (evlist__create_maps(evsel_list, &target) < 0) {
2444 if (target__has_task(&target)) {
2445 pr_err("Problems finding threads of monitor\n");
2446 parse_options_usage(stat_usage, stat_options, "p", 1);
2447 parse_options_usage(NULL, stat_options, "t", 1);
2448 } else if (target__has_cpu(&target)) {
2449 perror("failed to parse CPUs map");
2450 parse_options_usage(stat_usage, stat_options, "C", 1);
2451 parse_options_usage(NULL, stat_options, "a", 1);
2456 evlist__check_cpu_maps(evsel_list);
2459 * Initialize thread_map with comm names,
2460 * so we could print it out on output.
2462 if (stat_config.aggr_mode == AGGR_THREAD) {
2463 thread_map__read_comms(evsel_list->core.threads);
2464 if (target.system_wide) {
2465 if (runtime_stat_new(&stat_config,
2466 perf_thread_map__nr(evsel_list->core.threads))) {
2472 if (stat_config.aggr_mode == AGGR_NODE)
2473 cpu__setup_cpunode_map();
2475 if (stat_config.times && interval)
2476 interval_count = true;
2477 else if (stat_config.times && !interval) {
2478 pr_err("interval-count option should be used together with "
2479 "interval-print.\n");
2480 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2481 parse_options_usage(stat_usage, stat_options, "I", 1);
2485 if (timeout && timeout < 100) {
2487 pr_err("timeout must be >= 10ms.\n");
2488 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2491 pr_warning("timeout < 100ms. "
2492 "The overhead percentage could be high in some cases. "
2493 "Please proceed with caution.\n");
2495 if (timeout && interval) {
2496 pr_err("timeout option is not supported with interval-print.\n");
2497 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2498 parse_options_usage(stat_usage, stat_options, "I", 1);
2502 if (evlist__alloc_stats(evsel_list, interval))
2505 if (perf_stat_init_aggr_mode())
2509 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2510 * while avoiding that older tools show confusing messages.
2512 * However for pipe sessions we need to keep it zero,
2513 * because script's perf_evsel__check_attr is triggered
2514 * by attr->sample_type != 0, and we can't run it on
2517 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2520 * We dont want to block the signals - that would cause
2521 * child tasks to inherit that and Ctrl-C would not work.
2522 * What we want is for Ctrl-C to work in the exec()-ed
2523 * task, but being ignored by perf stat itself:
2527 signal(SIGINT, skip_signal);
2528 signal(SIGCHLD, skip_signal);
2529 signal(SIGALRM, skip_signal);
2530 signal(SIGABRT, skip_signal);
2532 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2536 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2537 if (stat_config.run_count != 1 && verbose > 0)
2538 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2542 evlist__reset_prev_raw_counts(evsel_list);
2544 status = run_perf_stat(argc, argv, run_idx);
2545 if (forever && status != -1 && !interval) {
2546 print_counters(NULL, argc, argv);
2547 perf_stat__reset_stats();
2551 if (!forever && status != -1 && (!interval || stat_config.summary))
2552 print_counters(NULL, argc, argv);
2554 evlist__finalize_ctlfd(evsel_list);
2558 * We synthesize the kernel mmap record just so that older tools
2559 * don't emit warnings about not being able to resolve symbols
2560 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2561 * a saner message about no samples being in the perf.data file.
2563 * This also serves to suppress a warning about f_header.data.size == 0
2564 * in header.c at the moment 'perf stat record' gets introduced, which
2565 * is not really needed once we start adding the stat specific PERF_RECORD_
2566 * records, but the need to suppress the kptr_restrict messages in older
2567 * tools remain -acme
2569 int fd = perf_data__fd(&perf_stat.data);
2571 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2572 process_synthesized_event,
2573 &perf_stat.session->machines.host);
2575 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2576 "older tools may produce warnings about this file\n.");
2580 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2581 pr_err("failed to write stat round event\n");
2584 if (!perf_stat.data.is_pipe) {
2585 perf_stat.session->header.data_size += perf_stat.bytes_written;
2586 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2589 evlist__close(evsel_list);
2590 perf_session__delete(perf_stat.session);
2593 perf_stat__exit_aggr_mode();
2594 evlist__free_stats(evsel_list);
2596 if (stat_config.iostat_run)
2597 iostat_release(evsel_list);
2599 zfree(&stat_config.walltime_run);
2601 if (smi_cost && smi_reset)
2602 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2604 evlist__delete(evsel_list);
2606 metricgroup__rblist_exit(&stat_config.metric_events);
2607 runtime_stat_delete(&stat_config);
2608 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);