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].cpu != b->core.cpus->map[i].cpu)
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, int cpu_map_idx, u32 thread,
331 struct perf_counts_values *count)
333 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
334 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
336 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
337 process_synthesized_event, NULL);
340 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
341 int thread, struct timespec *rs)
343 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
344 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
345 struct perf_counts_values *count =
346 perf_counts(counter->counts, cpu_map_idx, thread);
347 count->ena = count->run = val;
351 return evsel__read_counter(counter, cpu_map_idx, thread);
355 * Read out the results of a single counter:
356 * do not aggregate counts across CPUs in system-wide mode
358 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
360 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
363 if (!counter->supported)
366 if (counter->core.system_wide)
369 for (thread = 0; thread < nthreads; thread++) {
370 struct perf_counts_values *count;
372 count = perf_counts(counter->counts, cpu_map_idx, thread);
375 * The leader's group read loads data into its group members
376 * (via evsel__read_counter()) and sets their count->loaded.
378 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
379 read_single_counter(counter, cpu_map_idx, thread, rs)) {
380 counter->counts->scaled = -1;
381 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
382 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
386 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
389 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
390 pr_err("failed to write stat event\n");
396 fprintf(stat_config.output,
397 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
398 evsel__name(counter),
399 perf_cpu_map__cpu(evsel__cpus(counter),
401 count->val, count->ena, count->run);
408 static int read_affinity_counters(struct timespec *rs)
410 struct evlist_cpu_iterator evlist_cpu_itr;
411 struct affinity saved_affinity, *affinity;
413 if (all_counters_use_bpf)
416 if (!target__has_cpu(&target) || target__has_per_thread(&target))
418 else if (affinity__setup(&saved_affinity) < 0)
421 affinity = &saved_affinity;
423 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
424 struct evsel *counter = evlist_cpu_itr.evsel;
426 if (evsel__is_bpf(counter))
430 counter->err = read_counter_cpu(counter, rs,
431 evlist_cpu_itr.cpu_map_idx);
435 affinity__cleanup(&saved_affinity);
440 static int read_bpf_map_counters(void)
442 struct evsel *counter;
445 evlist__for_each_entry(evsel_list, counter) {
446 if (!evsel__is_bpf(counter))
449 err = bpf_counter__read(counter);
456 static void read_counters(struct timespec *rs)
458 struct evsel *counter;
460 if (!stat_config.stop_read_counter) {
461 if (read_bpf_map_counters() ||
462 read_affinity_counters(rs))
466 evlist__for_each_entry(evsel_list, counter) {
468 pr_debug("failed to read counter %s\n", counter->name);
469 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
470 pr_warning("failed to process counter %s\n", counter->name);
475 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
479 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
483 config->stats_num = nthreads;
485 for (i = 0; i < nthreads; i++)
486 runtime_stat__init(&config->stats[i]);
491 static void runtime_stat_delete(struct perf_stat_config *config)
498 for (i = 0; i < config->stats_num; i++)
499 runtime_stat__exit(&config->stats[i]);
501 zfree(&config->stats);
504 static void runtime_stat_reset(struct perf_stat_config *config)
511 for (i = 0; i < config->stats_num; i++)
512 perf_stat__reset_shadow_per_stat(&config->stats[i]);
515 static void process_interval(void)
517 struct timespec ts, rs;
519 clock_gettime(CLOCK_MONOTONIC, &ts);
520 diff_timespec(&rs, &ts, &ref_time);
522 perf_stat__reset_shadow_per_stat(&rt_stat);
523 runtime_stat_reset(&stat_config);
527 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
528 pr_err("failed to write stat round event\n");
531 init_stats(&walltime_nsecs_stats);
532 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
533 print_counters(&rs, 0, NULL);
536 static bool handle_interval(unsigned int interval, int *times)
540 if (interval_count && !(--(*times)))
546 static int enable_counters(void)
551 evlist__for_each_entry(evsel_list, evsel) {
552 if (!evsel__is_bpf(evsel))
555 err = bpf_counter__enable(evsel);
560 if (stat_config.initial_delay < 0) {
561 pr_info(EVLIST_DISABLED_MSG);
565 if (stat_config.initial_delay > 0) {
566 pr_info(EVLIST_DISABLED_MSG);
567 usleep(stat_config.initial_delay * USEC_PER_MSEC);
571 * We need to enable counters only if:
572 * - we don't have tracee (attaching to task or cpu)
573 * - we have initial delay configured
575 if (!target__none(&target) || stat_config.initial_delay) {
576 if (!all_counters_use_bpf)
577 evlist__enable(evsel_list);
578 if (stat_config.initial_delay > 0)
579 pr_info(EVLIST_ENABLED_MSG);
584 static void disable_counters(void)
586 struct evsel *counter;
589 * If we don't have tracee (attaching to task or cpu), counters may
590 * still be running. To get accurate group ratios, we must stop groups
591 * from counting before reading their constituent counters.
593 if (!target__none(&target)) {
594 evlist__for_each_entry(evsel_list, counter)
595 bpf_counter__disable(counter);
596 if (!all_counters_use_bpf)
597 evlist__disable(evsel_list);
601 static volatile int workload_exec_errno;
604 * evlist__prepare_workload will send a SIGUSR1
605 * if the fork fails, since we asked by setting its
606 * want_signal to true.
608 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
609 void *ucontext __maybe_unused)
611 workload_exec_errno = info->si_value.sival_int;
614 static bool evsel__should_store_id(struct evsel *counter)
616 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
619 static bool is_target_alive(struct target *_target,
620 struct perf_thread_map *threads)
625 if (!target__has_task(_target))
628 for (i = 0; i < threads->nr; i++) {
631 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
632 threads->map[i].pid);
634 if (!stat(path, &st))
641 static void process_evlist(struct evlist *evlist, unsigned int interval)
643 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
645 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
647 case EVLIST_CTL_CMD_ENABLE:
651 case EVLIST_CTL_CMD_DISABLE:
655 case EVLIST_CTL_CMD_SNAPSHOT:
656 case EVLIST_CTL_CMD_ACK:
657 case EVLIST_CTL_CMD_UNSUPPORTED:
658 case EVLIST_CTL_CMD_EVLIST:
659 case EVLIST_CTL_CMD_STOP:
660 case EVLIST_CTL_CMD_PING:
667 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
670 int tts = *time_to_sleep;
671 struct timespec time_diff;
673 diff_timespec(&time_diff, time_stop, time_start);
675 tts -= time_diff.tv_sec * MSEC_PER_SEC +
676 time_diff.tv_nsec / NSEC_PER_MSEC;
681 *time_to_sleep = tts;
684 static int dispatch_events(bool forks, int timeout, int interval, int *times)
686 int child_exited = 0, status = 0;
687 int time_to_sleep, sleep_time;
688 struct timespec time_start, time_stop;
691 sleep_time = interval;
693 sleep_time = timeout;
697 time_to_sleep = sleep_time;
701 child_exited = waitpid(child_pid, &status, WNOHANG);
703 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
708 clock_gettime(CLOCK_MONOTONIC, &time_start);
709 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
710 if (timeout || handle_interval(interval, times))
712 time_to_sleep = sleep_time;
713 } else { /* fd revent */
714 process_evlist(evsel_list, interval);
715 clock_gettime(CLOCK_MONOTONIC, &time_stop);
716 compute_tts(&time_start, &time_stop, &time_to_sleep);
723 enum counter_recovery {
729 static enum counter_recovery stat_handle_error(struct evsel *counter)
733 * PPC returns ENXIO for HW counters until 2.6.37
734 * (behavior changed with commit b0a873e).
736 if (errno == EINVAL || errno == ENOSYS ||
737 errno == ENOENT || errno == EOPNOTSUPP ||
740 ui__warning("%s event is not supported by the kernel.\n",
741 evsel__name(counter));
742 counter->supported = false;
744 * errored is a sticky flag that means one of the counter's
745 * cpu event had a problem and needs to be reexamined.
747 counter->errored = true;
749 if ((evsel__leader(counter) != counter) ||
750 !(counter->core.leader->nr_members > 1))
752 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
754 ui__warning("%s\n", msg);
755 return COUNTER_RETRY;
756 } else if (target__has_per_thread(&target) &&
757 evsel_list->core.threads &&
758 evsel_list->core.threads->err_thread != -1) {
760 * For global --per-thread case, skip current
763 if (!thread_map__remove(evsel_list->core.threads,
764 evsel_list->core.threads->err_thread)) {
765 evsel_list->core.threads->err_thread = -1;
766 return COUNTER_RETRY;
770 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
771 ui__error("%s\n", msg);
774 kill(child_pid, SIGTERM);
775 return COUNTER_FATAL;
778 static int __run_perf_stat(int argc, const char **argv, int run_idx)
780 int interval = stat_config.interval;
781 int times = stat_config.times;
782 int timeout = stat_config.timeout;
784 unsigned long long t0, t1;
785 struct evsel *counter;
788 const bool forks = (argc > 0);
789 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
790 struct evlist_cpu_iterator evlist_cpu_itr;
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(evlist_cpu_itr, evsel_list, &affinity) {
817 counter = evlist_cpu_itr.evsel;
820 * bperf calls evsel__open_per_cpu() in bperf__load(), so
821 * no need to call it again here.
826 if (counter->reset_group || counter->errored)
828 if (evsel__is_bpf(counter))
831 if (create_perf_stat_counter(counter, &stat_config, &target,
832 evlist_cpu_itr.cpu_map_idx) < 0) {
835 * Weak group failed. We cannot just undo this here
836 * because earlier CPUs might be in group mode, and the kernel
837 * doesn't support mixing group and non group reads. Defer
839 * Don't close here because we're in the wrong affinity.
841 if ((errno == EINVAL || errno == EBADF) &&
842 evsel__leader(counter) != counter &&
843 counter->weak_group) {
844 evlist__reset_weak_group(evsel_list, counter, false);
845 assert(counter->reset_group);
850 switch (stat_handle_error(counter)) {
862 counter->supported = true;
867 * Now redo all the weak group after closing them,
868 * and also close errored counters.
871 /* First close errored or weak retry */
872 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, &affinity) {
873 counter = evlist_cpu_itr.evsel;
875 if (!counter->reset_group && !counter->errored)
878 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
880 /* Now reopen weak */
881 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, &affinity) {
882 counter = evlist_cpu_itr.evsel;
884 if (!counter->reset_group && !counter->errored)
886 if (!counter->reset_group)
889 pr_debug2("reopening weak %s\n", evsel__name(counter));
890 if (create_perf_stat_counter(counter, &stat_config, &target,
891 evlist_cpu_itr.cpu_map_idx) < 0) {
893 switch (stat_handle_error(counter)) {
897 goto try_again_reset;
904 counter->supported = true;
907 affinity__cleanup(&affinity);
909 evlist__for_each_entry(evsel_list, counter) {
910 if (!counter->supported) {
911 perf_evsel__free_fd(&counter->core);
915 l = strlen(counter->unit);
916 if (l > stat_config.unit_width)
917 stat_config.unit_width = l;
919 if (evsel__should_store_id(counter) &&
920 evsel__store_ids(counter, evsel_list))
924 if (evlist__apply_filters(evsel_list, &counter)) {
925 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
926 counter->filter, evsel__name(counter), errno,
927 str_error_r(errno, msg, sizeof(msg)));
932 int fd = perf_data__fd(&perf_stat.data);
935 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
937 err = perf_session__write_header(perf_stat.session, evsel_list,
944 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
945 process_synthesized_event, is_pipe);
951 * Enable counters and exec the command:
954 evlist__start_workload(evsel_list);
955 err = enable_counters();
960 clock_gettime(CLOCK_MONOTONIC, &ref_time);
962 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
963 status = dispatch_events(forks, timeout, interval, ×);
964 if (child_pid != -1) {
966 kill(child_pid, SIGTERM);
967 wait4(child_pid, &status, 0, &stat_config.ru_data);
970 if (workload_exec_errno) {
971 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
972 pr_err("Workload failed: %s\n", emsg);
976 if (WIFSIGNALED(status))
977 psignal(WTERMSIG(status), argv[0]);
979 err = enable_counters();
984 clock_gettime(CLOCK_MONOTONIC, &ref_time);
986 status = dispatch_events(forks, timeout, interval, ×);
993 if (stat_config.walltime_run_table)
994 stat_config.walltime_run[run_idx] = t1 - t0;
996 if (interval && stat_config.summary) {
997 stat_config.interval = 0;
998 stat_config.stop_read_counter = true;
999 init_stats(&walltime_nsecs_stats);
1000 update_stats(&walltime_nsecs_stats, t1 - t0);
1002 if (stat_config.aggr_mode == AGGR_GLOBAL)
1003 evlist__save_aggr_prev_raw_counts(evsel_list);
1005 evlist__copy_prev_raw_counts(evsel_list);
1006 evlist__reset_prev_raw_counts(evsel_list);
1007 runtime_stat_reset(&stat_config);
1008 perf_stat__reset_shadow_per_stat(&rt_stat);
1010 update_stats(&walltime_nsecs_stats, t1 - t0);
1013 * Closing a group leader splits the group, and as we only disable
1014 * group leaders, results in remaining events becoming enabled. To
1015 * avoid arbitrary skew, we must read all counters before closing any
1018 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1021 * We need to keep evsel_list alive, because it's processed
1022 * later the evsel_list will be closed after.
1025 evlist__close(evsel_list);
1027 return WEXITSTATUS(status);
1030 static int run_perf_stat(int argc, const char **argv, int run_idx)
1035 ret = system(pre_cmd);
1043 ret = __run_perf_stat(argc, argv, run_idx);
1048 ret = system(post_cmd);
1056 static void print_counters(struct timespec *ts, int argc, const char **argv)
1058 /* Do not print anything if we record to the pipe. */
1059 if (STAT_RECORD && perf_stat.data.is_pipe)
1061 if (stat_config.quiet)
1064 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1067 static volatile int signr = -1;
1069 static void skip_signal(int signo)
1071 if ((child_pid == -1) || stat_config.interval)
1076 * render child_pid harmless
1077 * won't send SIGTERM to a random
1078 * process in case of race condition
1079 * and fast PID recycling
1084 static void sig_atexit(void)
1089 * avoid race condition with SIGCHLD handler
1090 * in skip_signal() which is modifying child_pid
1091 * goal is to avoid send SIGTERM to a random
1095 sigaddset(&set, SIGCHLD);
1096 sigprocmask(SIG_BLOCK, &set, &oset);
1098 if (child_pid != -1)
1099 kill(child_pid, SIGTERM);
1101 sigprocmask(SIG_SETMASK, &oset, NULL);
1106 signal(signr, SIG_DFL);
1107 kill(getpid(), signr);
1110 void perf_stat__set_big_num(int set)
1112 stat_config.big_num = (set != 0);
1115 void perf_stat__set_no_csv_summary(int set)
1117 stat_config.no_csv_summary = (set != 0);
1120 static int stat__set_big_num(const struct option *opt __maybe_unused,
1121 const char *s __maybe_unused, int unset)
1123 big_num_opt = unset ? 0 : 1;
1124 perf_stat__set_big_num(!unset);
1128 static int enable_metric_only(const struct option *opt __maybe_unused,
1129 const char *s __maybe_unused, int unset)
1131 force_metric_only = true;
1132 stat_config.metric_only = !unset;
1136 static int parse_metric_groups(const struct option *opt,
1138 int unset __maybe_unused)
1140 return metricgroup__parse_groups(opt, str,
1141 stat_config.metric_no_group,
1142 stat_config.metric_no_merge,
1143 &stat_config.metric_events);
1146 static int parse_control_option(const struct option *opt,
1148 int unset __maybe_unused)
1150 struct perf_stat_config *config = opt->value;
1152 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1155 static int parse_stat_cgroups(const struct option *opt,
1156 const char *str, int unset)
1158 if (stat_config.cgroup_list) {
1159 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1163 return parse_cgroups(opt, str, unset);
1166 static int parse_hybrid_type(const struct option *opt,
1168 int unset __maybe_unused)
1170 struct evlist *evlist = *(struct evlist **)opt->value;
1172 if (!list_empty(&evlist->core.entries)) {
1173 fprintf(stderr, "Must define cputype before events/metrics\n");
1177 evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1178 if (!evlist->hybrid_pmu_name) {
1179 fprintf(stderr, "--cputype %s is not supported!\n", str);
1186 static struct option stat_options[] = {
1187 OPT_BOOLEAN('T', "transaction", &transaction_run,
1188 "hardware transaction statistics"),
1189 OPT_CALLBACK('e', "event", &evsel_list, "event",
1190 "event selector. use 'perf list' to list available events",
1191 parse_events_option),
1192 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1193 "event filter", parse_filter),
1194 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1195 "child tasks do not inherit counters"),
1196 OPT_STRING('p', "pid", &target.pid, "pid",
1197 "stat events on existing process id"),
1198 OPT_STRING('t', "tid", &target.tid, "tid",
1199 "stat events on existing thread id"),
1200 #ifdef HAVE_BPF_SKEL
1201 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1202 "stat events on existing bpf program id"),
1203 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1204 "use bpf program to count events"),
1205 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1206 "path to perf_event_attr map"),
1208 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1209 "system-wide collection from all CPUs"),
1210 OPT_BOOLEAN('g', "group", &group,
1211 "put the counters into a counter group"),
1212 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1213 "Use --no-scale to disable counter scaling for multiplexing"),
1214 OPT_INCR('v', "verbose", &verbose,
1215 "be more verbose (show counter open errors, etc)"),
1216 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1217 "repeat command and print average + stddev (max: 100, forever: 0)"),
1218 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1219 "display details about each run (only with -r option)"),
1220 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1221 "null run - dont start any counters"),
1222 OPT_INCR('d', "detailed", &detailed_run,
1223 "detailed run - start a lot of events"),
1224 OPT_BOOLEAN('S', "sync", &sync_run,
1225 "call sync() before starting a run"),
1226 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1227 "print large numbers with thousands\' separators",
1229 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1230 "list of cpus to monitor in system-wide"),
1231 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1232 "disable CPU count aggregation", AGGR_NONE),
1233 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1234 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1235 "print counts with custom separator"),
1236 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1237 "monitor event in cgroup name only", parse_stat_cgroups),
1238 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1239 "expand events for each cgroup"),
1240 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1241 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1242 OPT_INTEGER(0, "log-fd", &output_fd,
1243 "log output to fd, instead of stderr"),
1244 OPT_STRING(0, "pre", &pre_cmd, "command",
1245 "command to run prior to the measured command"),
1246 OPT_STRING(0, "post", &post_cmd, "command",
1247 "command to run after to the measured command"),
1248 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1249 "print counts at regular interval in ms "
1250 "(overhead is possible for values <= 100ms)"),
1251 OPT_INTEGER(0, "interval-count", &stat_config.times,
1252 "print counts for fixed number of times"),
1253 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1254 "clear screen in between new interval"),
1255 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1256 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1257 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1258 "aggregate counts per processor socket", AGGR_SOCKET),
1259 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1260 "aggregate counts per processor die", AGGR_DIE),
1261 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1262 "aggregate counts per physical processor core", AGGR_CORE),
1263 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1264 "aggregate counts per thread", AGGR_THREAD),
1265 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1266 "aggregate counts per numa node", AGGR_NODE),
1267 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1268 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1269 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1270 "Only print computed metrics. No raw values", enable_metric_only),
1271 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1272 "don't group metric events, impacts multiplexing"),
1273 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1274 "don't try to share events between metrics in a group"),
1275 OPT_BOOLEAN(0, "topdown", &topdown_run,
1276 "measure top-down statistics"),
1277 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1278 "Set the metrics level for the top-down statistics (0: max level)"),
1279 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1280 "measure SMI cost"),
1281 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1282 "monitor specified metrics or metric groups (separated by ,)",
1283 parse_metric_groups),
1284 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1285 "Configure all used events to run in kernel space.",
1286 PARSE_OPT_EXCLUSIVE),
1287 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1288 "Configure all used events to run in user space.",
1289 PARSE_OPT_EXCLUSIVE),
1290 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1291 "Use with 'percore' event qualifier to show the event "
1292 "counts of one hardware thread by sum up total hardware "
1293 "threads of same physical core"),
1294 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1295 "print summary for interval mode"),
1296 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1297 "don't print 'summary' for CSV summary output"),
1298 OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1299 "don't print output (useful with record)"),
1300 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1301 "Only enable events on applying cpu with this type "
1302 "for hybrid platform (e.g. core or atom)",
1305 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1306 "libpfm4 event selector. use 'perf list' to list available events",
1307 parse_libpfm_events_option),
1309 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1310 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1311 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1312 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1313 parse_control_option),
1314 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1315 "measure I/O performance metrics provided by arch/platform",
1320 static const char *const aggr_mode__string[] = {
1321 [AGGR_CORE] = "core",
1323 [AGGR_GLOBAL] = "global",
1324 [AGGR_NODE] = "node",
1325 [AGGR_NONE] = "none",
1326 [AGGR_SOCKET] = "socket",
1327 [AGGR_THREAD] = "thread",
1328 [AGGR_UNSET] = "unset",
1331 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1332 struct perf_cpu cpu)
1334 return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1337 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1338 struct perf_cpu cpu)
1340 return aggr_cpu_id__die(cpu, /*data=*/NULL);
1343 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1344 struct perf_cpu cpu)
1346 return aggr_cpu_id__core(cpu, /*data=*/NULL);
1349 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1350 struct perf_cpu cpu)
1352 return aggr_cpu_id__node(cpu, /*data=*/NULL);
1355 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1356 aggr_get_id_t get_id, struct perf_cpu cpu)
1358 struct aggr_cpu_id id = aggr_cpu_id__empty();
1360 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1361 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1363 id = config->cpus_aggr_map->map[cpu.cpu];
1367 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1368 struct perf_cpu cpu)
1370 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1373 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1374 struct perf_cpu cpu)
1376 return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1379 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1380 struct perf_cpu cpu)
1382 return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1385 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1386 struct perf_cpu cpu)
1388 return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1391 static bool term_percore_set(void)
1393 struct evsel *counter;
1395 evlist__for_each_entry(evsel_list, counter) {
1396 if (counter->percore)
1403 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1405 switch (aggr_mode) {
1407 return aggr_cpu_id__socket;
1409 return aggr_cpu_id__die;
1411 return aggr_cpu_id__core;
1413 return aggr_cpu_id__node;
1415 if (term_percore_set())
1416 return aggr_cpu_id__core;
1427 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1429 switch (aggr_mode) {
1431 return perf_stat__get_socket_cached;
1433 return perf_stat__get_die_cached;
1435 return perf_stat__get_core_cached;
1437 return perf_stat__get_node_cached;
1439 if (term_percore_set()) {
1440 return perf_stat__get_core_cached;
1451 static int perf_stat_init_aggr_mode(void)
1454 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1457 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.cpus,
1458 get_id, /*data=*/NULL);
1459 if (!stat_config.aggr_map) {
1460 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1463 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1467 * The evsel_list->cpus is the base we operate on,
1468 * taking the highest cpu number to be the size of
1469 * the aggregation translate cpumap.
1471 nr = perf_cpu_map__max(evsel_list->core.cpus).cpu;
1472 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1473 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1476 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1479 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1480 "cpu_aggr_map refcnt unbalanced\n");
1485 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1487 if (map && refcount_dec_and_test(&map->refcnt))
1488 cpu_aggr_map__delete(map);
1491 static void perf_stat__exit_aggr_mode(void)
1493 cpu_aggr_map__put(stat_config.aggr_map);
1494 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1495 stat_config.aggr_map = NULL;
1496 stat_config.cpus_aggr_map = NULL;
1499 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1501 struct perf_env *env = data;
1502 struct aggr_cpu_id id = aggr_cpu_id__empty();
1505 id.socket = env->cpu[cpu.cpu].socket_id;
1510 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1512 struct perf_env *env = data;
1513 struct aggr_cpu_id id = aggr_cpu_id__empty();
1515 if (cpu.cpu != -1) {
1517 * die_id is relative to socket, so start
1518 * with the socket ID and then add die to
1521 id.socket = env->cpu[cpu.cpu].socket_id;
1522 id.die = env->cpu[cpu.cpu].die_id;
1528 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1530 struct perf_env *env = data;
1531 struct aggr_cpu_id id = aggr_cpu_id__empty();
1533 if (cpu.cpu != -1) {
1535 * core_id is relative to socket and die,
1536 * we need a global id. So we set
1537 * socket, die id and core id
1539 id.socket = env->cpu[cpu.cpu].socket_id;
1540 id.die = env->cpu[cpu.cpu].die_id;
1541 id.core = env->cpu[cpu.cpu].core_id;
1547 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1549 struct aggr_cpu_id id = aggr_cpu_id__empty();
1551 id.node = perf_env__numa_node(data, cpu);
1555 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1556 struct perf_cpu cpu)
1558 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1560 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1561 struct perf_cpu cpu)
1563 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1566 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1567 struct perf_cpu cpu)
1569 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1572 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1573 struct perf_cpu cpu)
1575 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1578 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1580 switch (aggr_mode) {
1582 return perf_env__get_socket_aggr_by_cpu;
1584 return perf_env__get_die_aggr_by_cpu;
1586 return perf_env__get_core_aggr_by_cpu;
1588 return perf_env__get_node_aggr_by_cpu;
1598 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1600 switch (aggr_mode) {
1602 return perf_stat__get_socket_file;
1604 return perf_stat__get_die_file;
1606 return perf_stat__get_core_file;
1608 return perf_stat__get_node_file;
1618 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1620 struct perf_env *env = &st->session->header.env;
1621 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1626 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.cpus, get_id, env);
1627 if (!stat_config.aggr_map) {
1628 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1631 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1636 * Add default attributes, if there were no attributes specified or
1637 * if -d/--detailed, -d -d or -d -d -d is used:
1639 static int add_default_attributes(void)
1642 struct perf_event_attr default_attrs0[] = {
1644 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1645 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1646 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1647 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1649 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1651 struct perf_event_attr frontend_attrs[] = {
1652 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1654 struct perf_event_attr backend_attrs[] = {
1655 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1657 struct perf_event_attr default_attrs1[] = {
1658 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1659 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1660 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1663 struct perf_event_attr default_sw_attrs[] = {
1664 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1665 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1666 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1667 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1671 * Detailed stats (-d), covering the L1 and last level data caches:
1673 struct perf_event_attr detailed_attrs[] = {
1675 { .type = PERF_TYPE_HW_CACHE,
1677 PERF_COUNT_HW_CACHE_L1D << 0 |
1678 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1679 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1681 { .type = PERF_TYPE_HW_CACHE,
1683 PERF_COUNT_HW_CACHE_L1D << 0 |
1684 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1685 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1687 { .type = PERF_TYPE_HW_CACHE,
1689 PERF_COUNT_HW_CACHE_LL << 0 |
1690 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1691 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1693 { .type = PERF_TYPE_HW_CACHE,
1695 PERF_COUNT_HW_CACHE_LL << 0 |
1696 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1697 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1701 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1703 struct perf_event_attr very_detailed_attrs[] = {
1705 { .type = PERF_TYPE_HW_CACHE,
1707 PERF_COUNT_HW_CACHE_L1I << 0 |
1708 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1709 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1711 { .type = PERF_TYPE_HW_CACHE,
1713 PERF_COUNT_HW_CACHE_L1I << 0 |
1714 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1715 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1717 { .type = PERF_TYPE_HW_CACHE,
1719 PERF_COUNT_HW_CACHE_DTLB << 0 |
1720 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1721 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1723 { .type = PERF_TYPE_HW_CACHE,
1725 PERF_COUNT_HW_CACHE_DTLB << 0 |
1726 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1727 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1729 { .type = PERF_TYPE_HW_CACHE,
1731 PERF_COUNT_HW_CACHE_ITLB << 0 |
1732 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1733 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1735 { .type = PERF_TYPE_HW_CACHE,
1737 PERF_COUNT_HW_CACHE_ITLB << 0 |
1738 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1739 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1744 * Very, very detailed stats (-d -d -d), adding prefetch events:
1746 struct perf_event_attr very_very_detailed_attrs[] = {
1748 { .type = PERF_TYPE_HW_CACHE,
1750 PERF_COUNT_HW_CACHE_L1D << 0 |
1751 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1752 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1754 { .type = PERF_TYPE_HW_CACHE,
1756 PERF_COUNT_HW_CACHE_L1D << 0 |
1757 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1758 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1760 /* Set attrs if no event is selected and !null_run: */
1761 if (stat_config.null_run)
1764 if (transaction_run) {
1765 struct parse_events_error errinfo;
1766 /* Handle -T as -M transaction. Once platform specific metrics
1767 * support has been added to the json files, all architectures
1768 * will use this approach. To determine transaction support
1769 * on an architecture test for such a metric name.
1771 if (metricgroup__has_metric("transaction")) {
1772 struct option opt = { .value = &evsel_list };
1774 return metricgroup__parse_groups(&opt, "transaction",
1775 stat_config.metric_no_group,
1776 stat_config.metric_no_merge,
1777 &stat_config.metric_events);
1780 parse_events_error__init(&errinfo);
1781 if (pmu_have_event("cpu", "cycles-ct") &&
1782 pmu_have_event("cpu", "el-start"))
1783 err = parse_events(evsel_list, transaction_attrs,
1786 err = parse_events(evsel_list,
1787 transaction_limited_attrs,
1790 fprintf(stderr, "Cannot set up transaction events\n");
1791 parse_events_error__print(&errinfo, transaction_attrs);
1793 parse_events_error__exit(&errinfo);
1794 return err ? -1 : 0;
1798 struct parse_events_error errinfo;
1801 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1802 fprintf(stderr, "freeze_on_smi is not supported.\n");
1807 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1808 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1814 if (!pmu_have_event("msr", "aperf") ||
1815 !pmu_have_event("msr", "smi")) {
1816 fprintf(stderr, "To measure SMI cost, it needs "
1817 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1820 if (!force_metric_only)
1821 stat_config.metric_only = true;
1823 parse_events_error__init(&errinfo);
1824 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1826 parse_events_error__print(&errinfo, smi_cost_attrs);
1827 fprintf(stderr, "Cannot set up SMI cost events\n");
1829 parse_events_error__exit(&errinfo);
1830 return err ? -1 : 0;
1834 const char **metric_attrs = topdown_metric_attrs;
1835 unsigned int max_level = 1;
1839 if (!force_metric_only)
1840 stat_config.metric_only = true;
1842 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1843 metric_attrs = topdown_metric_L2_attrs;
1847 if (stat_config.topdown_level > max_level) {
1848 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1850 } else if (!stat_config.topdown_level)
1851 stat_config.topdown_level = max_level;
1853 if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1854 pr_err("Out of memory\n");
1857 if (metric_attrs[0] && str) {
1858 if (!stat_config.interval && !stat_config.metric_only) {
1859 fprintf(stat_config.output,
1860 "Topdown accuracy may decrease when measuring long periods.\n"
1861 "Please print the result regularly, e.g. -I1000\n");
1868 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1869 stat_config.aggr_mode != AGGR_CORE) {
1870 pr_err("top down event configuration requires --per-core mode\n");
1873 stat_config.aggr_mode = AGGR_CORE;
1874 if (nr_cgroups || !target__has_cpu(&target)) {
1875 pr_err("top down event configuration requires system-wide mode (-a)\n");
1879 if (topdown_filter_events(topdown_attrs, &str,
1880 arch_topdown_check_group(&warn)) < 0) {
1881 pr_err("Out of memory\n");
1884 if (topdown_attrs[0] && str) {
1885 struct parse_events_error errinfo;
1887 arch_topdown_group_warn();
1889 parse_events_error__init(&errinfo);
1890 err = parse_events(evsel_list, str, &errinfo);
1893 "Cannot set up top down events %s: %d\n",
1895 parse_events_error__print(&errinfo, str);
1896 parse_events_error__exit(&errinfo);
1900 parse_events_error__exit(&errinfo);
1902 fprintf(stderr, "System does not support topdown\n");
1908 if (!evsel_list->core.nr_entries) {
1909 if (perf_pmu__has_hybrid()) {
1910 struct parse_events_error errinfo;
1911 const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1913 if (target__has_cpu(&target))
1914 default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1916 if (evlist__add_default_attrs(evsel_list,
1917 default_sw_attrs) < 0) {
1921 parse_events_error__init(&errinfo);
1922 err = parse_events(evsel_list, hybrid_str, &errinfo);
1925 "Cannot set up hybrid events %s: %d\n",
1927 parse_events_error__print(&errinfo, hybrid_str);
1929 parse_events_error__exit(&errinfo);
1930 return err ? -1 : 0;
1933 if (target__has_cpu(&target))
1934 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1936 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1938 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1939 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1942 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1943 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1946 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1949 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1950 if (arch_evlist__add_default_attrs(evsel_list) < 0)
1954 /* Detailed events get appended to the event list: */
1956 if (detailed_run < 1)
1959 /* Append detailed run extra attributes: */
1960 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1963 if (detailed_run < 2)
1966 /* Append very detailed run extra attributes: */
1967 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1970 if (detailed_run < 3)
1973 /* Append very, very detailed run extra attributes: */
1974 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1977 static const char * const stat_record_usage[] = {
1978 "perf stat record [<options>]",
1982 static void init_features(struct perf_session *session)
1986 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1987 perf_header__set_feat(&session->header, feat);
1989 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1990 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1991 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1992 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1993 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1996 static int __cmd_record(int argc, const char **argv)
1998 struct perf_session *session;
1999 struct perf_data *data = &perf_stat.data;
2001 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2002 PARSE_OPT_STOP_AT_NON_OPTION);
2005 data->path = output_name;
2007 if (stat_config.run_count != 1 || forever) {
2008 pr_err("Cannot use -r option with perf stat record.\n");
2012 session = perf_session__new(data, NULL);
2013 if (IS_ERR(session)) {
2014 pr_err("Perf session creation failed\n");
2015 return PTR_ERR(session);
2018 init_features(session);
2020 session->evlist = evsel_list;
2021 perf_stat.session = session;
2022 perf_stat.record = true;
2026 static int process_stat_round_event(struct perf_session *session,
2027 union perf_event *event)
2029 struct perf_record_stat_round *stat_round = &event->stat_round;
2030 struct evsel *counter;
2031 struct timespec tsh, *ts = NULL;
2032 const char **argv = session->header.env.cmdline_argv;
2033 int argc = session->header.env.nr_cmdline;
2035 evlist__for_each_entry(evsel_list, counter)
2036 perf_stat_process_counter(&stat_config, counter);
2038 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2039 update_stats(&walltime_nsecs_stats, stat_round->time);
2041 if (stat_config.interval && stat_round->time) {
2042 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2043 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2047 print_counters(ts, argc, argv);
2052 int process_stat_config_event(struct perf_session *session,
2053 union perf_event *event)
2055 struct perf_tool *tool = session->tool;
2056 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2058 perf_event__read_stat_config(&stat_config, &event->stat_config);
2060 if (perf_cpu_map__empty(st->cpus)) {
2061 if (st->aggr_mode != AGGR_UNSET)
2062 pr_warning("warning: processing task data, aggregation mode not set\n");
2066 if (st->aggr_mode != AGGR_UNSET)
2067 stat_config.aggr_mode = st->aggr_mode;
2069 if (perf_stat.data.is_pipe)
2070 perf_stat_init_aggr_mode();
2072 perf_stat_init_aggr_mode_file(st);
2077 static int set_maps(struct perf_stat *st)
2079 if (!st->cpus || !st->threads)
2082 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2085 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2087 if (evlist__alloc_stats(evsel_list, true))
2090 st->maps_allocated = true;
2095 int process_thread_map_event(struct perf_session *session,
2096 union perf_event *event)
2098 struct perf_tool *tool = session->tool;
2099 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2102 pr_warning("Extra thread map event, ignoring.\n");
2106 st->threads = thread_map__new_event(&event->thread_map);
2110 return set_maps(st);
2114 int process_cpu_map_event(struct perf_session *session,
2115 union perf_event *event)
2117 struct perf_tool *tool = session->tool;
2118 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2119 struct perf_cpu_map *cpus;
2122 pr_warning("Extra cpu map event, ignoring.\n");
2126 cpus = cpu_map__new_data(&event->cpu_map.data);
2131 return set_maps(st);
2134 static const char * const stat_report_usage[] = {
2135 "perf stat report [<options>]",
2139 static struct perf_stat perf_stat = {
2141 .attr = perf_event__process_attr,
2142 .event_update = perf_event__process_event_update,
2143 .thread_map = process_thread_map_event,
2144 .cpu_map = process_cpu_map_event,
2145 .stat_config = process_stat_config_event,
2146 .stat = perf_event__process_stat_event,
2147 .stat_round = process_stat_round_event,
2149 .aggr_mode = AGGR_UNSET,
2152 static int __cmd_report(int argc, const char **argv)
2154 struct perf_session *session;
2155 const struct option options[] = {
2156 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2157 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2158 "aggregate counts per processor socket", AGGR_SOCKET),
2159 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2160 "aggregate counts per processor die", AGGR_DIE),
2161 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2162 "aggregate counts per physical processor core", AGGR_CORE),
2163 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2164 "aggregate counts per numa node", AGGR_NODE),
2165 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2166 "disable CPU count aggregation", AGGR_NONE),
2172 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2174 if (!input_name || !strlen(input_name)) {
2175 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2178 input_name = "perf.data";
2181 perf_stat.data.path = input_name;
2182 perf_stat.data.mode = PERF_DATA_MODE_READ;
2184 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2185 if (IS_ERR(session))
2186 return PTR_ERR(session);
2188 perf_stat.session = session;
2189 stat_config.output = stderr;
2190 evsel_list = session->evlist;
2192 ret = perf_session__process_events(session);
2196 perf_session__delete(session);
2200 static void setup_system_wide(int forks)
2203 * Make system wide (-a) the default target if
2204 * no target was specified and one of following
2205 * conditions is met:
2207 * - there's no workload specified
2208 * - there is workload specified but all requested
2209 * events are system wide events
2211 if (!target__none(&target))
2215 target.system_wide = true;
2217 struct evsel *counter;
2219 evlist__for_each_entry(evsel_list, counter) {
2220 if (!counter->core.system_wide &&
2221 strcmp(counter->name, "duration_time")) {
2226 if (evsel_list->core.nr_entries)
2227 target.system_wide = true;
2231 int cmd_stat(int argc, const char **argv)
2233 const char * const stat_usage[] = {
2234 "perf stat [<options>] [<command>]",
2237 int status = -EINVAL, run_idx, err;
2239 FILE *output = stderr;
2240 unsigned int interval, timeout;
2241 const char * const stat_subcommands[] = { "record", "report" };
2242 char errbuf[BUFSIZ];
2244 setlocale(LC_ALL, "");
2246 evsel_list = evlist__new();
2247 if (evsel_list == NULL)
2250 parse_events__shrink_config_terms();
2252 /* String-parsing callback-based options would segfault when negated */
2253 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2254 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2255 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2257 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2258 (const char **) stat_usage,
2259 PARSE_OPT_STOP_AT_NON_OPTION);
2260 perf_stat__collect_metric_expr(evsel_list);
2261 perf_stat__init_shadow_stats();
2263 if (stat_config.csv_sep) {
2264 stat_config.csv_output = true;
2265 if (!strcmp(stat_config.csv_sep, "\\t"))
2266 stat_config.csv_sep = "\t";
2268 stat_config.csv_sep = DEFAULT_SEPARATOR;
2270 if (argc && !strncmp(argv[0], "rec", 3)) {
2271 argc = __cmd_record(argc, argv);
2274 } else if (argc && !strncmp(argv[0], "rep", 3))
2275 return __cmd_report(argc, argv);
2277 interval = stat_config.interval;
2278 timeout = stat_config.timeout;
2281 * For record command the -o is already taken care of.
2283 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2286 if (output_name && output_fd) {
2287 fprintf(stderr, "cannot use both --output and --log-fd\n");
2288 parse_options_usage(stat_usage, stat_options, "o", 1);
2289 parse_options_usage(NULL, stat_options, "log-fd", 0);
2293 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2294 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2298 if (stat_config.metric_only && stat_config.run_count > 1) {
2299 fprintf(stderr, "--metric-only is not supported with -r\n");
2303 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2304 fprintf(stderr, "--table is only supported with -r\n");
2305 parse_options_usage(stat_usage, stat_options, "r", 1);
2306 parse_options_usage(NULL, stat_options, "table", 0);
2310 if (output_fd < 0) {
2311 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2312 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2316 if (!output && !stat_config.quiet) {
2318 mode = append_file ? "a" : "w";
2320 output = fopen(output_name, mode);
2322 perror("failed to create output file");
2325 clock_gettime(CLOCK_REALTIME, &tm);
2326 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2327 } else if (output_fd > 0) {
2328 mode = append_file ? "a" : "w";
2329 output = fdopen(output_fd, mode);
2331 perror("Failed opening logfd");
2336 stat_config.output = output;
2339 * let the spreadsheet do the pretty-printing
2341 if (stat_config.csv_output) {
2342 /* User explicitly passed -B? */
2343 if (big_num_opt == 1) {
2344 fprintf(stderr, "-B option not supported with -x\n");
2345 parse_options_usage(stat_usage, stat_options, "B", 1);
2346 parse_options_usage(NULL, stat_options, "x", 1);
2348 } else /* Nope, so disable big number formatting */
2349 stat_config.big_num = false;
2350 } else if (big_num_opt == 0) /* User passed --no-big-num */
2351 stat_config.big_num = false;
2353 err = target__validate(&target);
2355 target__strerror(&target, err, errbuf, BUFSIZ);
2356 pr_warning("%s\n", errbuf);
2359 setup_system_wide(argc);
2362 * Display user/system times only for single
2363 * run and when there's specified tracee.
2365 if ((stat_config.run_count == 1) && target__none(&target))
2366 stat_config.ru_display = true;
2368 if (stat_config.run_count < 0) {
2369 pr_err("Run count must be a positive number\n");
2370 parse_options_usage(stat_usage, stat_options, "r", 1);
2372 } else if (stat_config.run_count == 0) {
2374 stat_config.run_count = 1;
2377 if (stat_config.walltime_run_table) {
2378 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2379 if (!stat_config.walltime_run) {
2380 pr_err("failed to setup -r option");
2385 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2386 !target__has_task(&target)) {
2387 if (!target.system_wide || target.cpu_list) {
2388 fprintf(stderr, "The --per-thread option is only "
2389 "available when monitoring via -p -t -a "
2390 "options or only --per-thread.\n");
2391 parse_options_usage(NULL, stat_options, "p", 1);
2392 parse_options_usage(NULL, stat_options, "t", 1);
2398 * no_aggr, cgroup are for system-wide only
2399 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2401 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2402 stat_config.aggr_mode != AGGR_THREAD) ||
2403 (nr_cgroups || stat_config.cgroup_list)) &&
2404 !target__has_cpu(&target)) {
2405 fprintf(stderr, "both cgroup and no-aggregation "
2406 "modes only available in system-wide mode\n");
2408 parse_options_usage(stat_usage, stat_options, "G", 1);
2409 parse_options_usage(NULL, stat_options, "A", 1);
2410 parse_options_usage(NULL, stat_options, "a", 1);
2411 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2415 if (stat_config.iostat_run) {
2416 status = iostat_prepare(evsel_list, &stat_config);
2419 if (iostat_mode == IOSTAT_LIST) {
2420 iostat_list(evsel_list, &stat_config);
2423 iostat_list(evsel_list, &stat_config);
2424 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2425 target.system_wide = true;
2428 if (add_default_attributes())
2431 if (stat_config.cgroup_list) {
2432 if (nr_cgroups > 0) {
2433 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2434 parse_options_usage(stat_usage, stat_options, "G", 1);
2435 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2439 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2440 &stat_config.metric_events, true) < 0) {
2441 parse_options_usage(stat_usage, stat_options,
2442 "for-each-cgroup", 0);
2447 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2448 target.per_thread = true;
2450 if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2451 pr_err("failed to use cpu list %s\n", target.cpu_list);
2455 target.hybrid = perf_pmu__has_hybrid();
2456 if (evlist__create_maps(evsel_list, &target) < 0) {
2457 if (target__has_task(&target)) {
2458 pr_err("Problems finding threads of monitor\n");
2459 parse_options_usage(stat_usage, stat_options, "p", 1);
2460 parse_options_usage(NULL, stat_options, "t", 1);
2461 } else if (target__has_cpu(&target)) {
2462 perror("failed to parse CPUs map");
2463 parse_options_usage(stat_usage, stat_options, "C", 1);
2464 parse_options_usage(NULL, stat_options, "a", 1);
2469 evlist__check_cpu_maps(evsel_list);
2472 * Initialize thread_map with comm names,
2473 * so we could print it out on output.
2475 if (stat_config.aggr_mode == AGGR_THREAD) {
2476 thread_map__read_comms(evsel_list->core.threads);
2477 if (target.system_wide) {
2478 if (runtime_stat_new(&stat_config,
2479 perf_thread_map__nr(evsel_list->core.threads))) {
2485 if (stat_config.aggr_mode == AGGR_NODE)
2486 cpu__setup_cpunode_map();
2488 if (stat_config.times && interval)
2489 interval_count = true;
2490 else if (stat_config.times && !interval) {
2491 pr_err("interval-count option should be used together with "
2492 "interval-print.\n");
2493 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2494 parse_options_usage(stat_usage, stat_options, "I", 1);
2498 if (timeout && timeout < 100) {
2500 pr_err("timeout must be >= 10ms.\n");
2501 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2504 pr_warning("timeout < 100ms. "
2505 "The overhead percentage could be high in some cases. "
2506 "Please proceed with caution.\n");
2508 if (timeout && interval) {
2509 pr_err("timeout option is not supported with interval-print.\n");
2510 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2511 parse_options_usage(stat_usage, stat_options, "I", 1);
2515 if (evlist__alloc_stats(evsel_list, interval))
2518 if (perf_stat_init_aggr_mode())
2522 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2523 * while avoiding that older tools show confusing messages.
2525 * However for pipe sessions we need to keep it zero,
2526 * because script's perf_evsel__check_attr is triggered
2527 * by attr->sample_type != 0, and we can't run it on
2530 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2533 * We dont want to block the signals - that would cause
2534 * child tasks to inherit that and Ctrl-C would not work.
2535 * What we want is for Ctrl-C to work in the exec()-ed
2536 * task, but being ignored by perf stat itself:
2540 signal(SIGINT, skip_signal);
2541 signal(SIGCHLD, skip_signal);
2542 signal(SIGALRM, skip_signal);
2543 signal(SIGABRT, skip_signal);
2545 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2549 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2550 if (stat_config.run_count != 1 && verbose > 0)
2551 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2555 evlist__reset_prev_raw_counts(evsel_list);
2557 status = run_perf_stat(argc, argv, run_idx);
2558 if (forever && status != -1 && !interval) {
2559 print_counters(NULL, argc, argv);
2560 perf_stat__reset_stats();
2564 if (!forever && status != -1 && (!interval || stat_config.summary))
2565 print_counters(NULL, argc, argv);
2567 evlist__finalize_ctlfd(evsel_list);
2571 * We synthesize the kernel mmap record just so that older tools
2572 * don't emit warnings about not being able to resolve symbols
2573 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2574 * a saner message about no samples being in the perf.data file.
2576 * This also serves to suppress a warning about f_header.data.size == 0
2577 * in header.c at the moment 'perf stat record' gets introduced, which
2578 * is not really needed once we start adding the stat specific PERF_RECORD_
2579 * records, but the need to suppress the kptr_restrict messages in older
2580 * tools remain -acme
2582 int fd = perf_data__fd(&perf_stat.data);
2584 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2585 process_synthesized_event,
2586 &perf_stat.session->machines.host);
2588 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2589 "older tools may produce warnings about this file\n.");
2593 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2594 pr_err("failed to write stat round event\n");
2597 if (!perf_stat.data.is_pipe) {
2598 perf_stat.session->header.data_size += perf_stat.bytes_written;
2599 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2602 evlist__close(evsel_list);
2603 perf_session__delete(perf_stat.session);
2606 perf_stat__exit_aggr_mode();
2607 evlist__free_stats(evsel_list);
2609 if (stat_config.iostat_run)
2610 iostat_release(evsel_list);
2612 zfree(&stat_config.walltime_run);
2614 if (smi_cost && smi_reset)
2615 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2617 evlist__delete(evsel_list);
2619 metricgroup__rblist_exit(&stat_config.metric_events);
2620 runtime_stat_delete(&stat_config);
2621 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);