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;
194 static char *metrics;
198 struct perf_data data;
199 struct perf_session *session;
201 struct perf_tool tool;
203 struct perf_cpu_map *cpus;
204 struct perf_thread_map *threads;
205 enum aggr_mode aggr_mode;
208 static struct perf_stat perf_stat;
209 #define STAT_RECORD perf_stat.record
211 static volatile int done = 0;
213 static struct perf_stat_config stat_config = {
214 .aggr_mode = AGGR_GLOBAL,
216 .unit_width = 4, /* strlen("unit") */
218 .metric_only_len = METRIC_ONLY_LEN,
219 .walltime_nsecs_stats = &walltime_nsecs_stats,
220 .ru_stats = &ru_stats,
227 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
229 if (!a->core.cpus && !b->core.cpus)
232 if (!a->core.cpus || !b->core.cpus)
235 if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
238 for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
239 if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
240 perf_cpu_map__cpu(b->core.cpus, i).cpu)
247 static void evlist__check_cpu_maps(struct evlist *evlist)
249 struct evsel *evsel, *pos, *leader;
252 if (evlist__has_hybrid(evlist))
253 evlist__warn_hybrid_group(evlist);
255 evlist__for_each_entry(evlist, evsel) {
256 leader = evsel__leader(evsel);
258 /* Check that leader matches cpus with each member. */
261 if (cpus_map_matched(leader, evsel))
264 /* If there's mismatch disable the group and warn user. */
265 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
266 evsel__group_desc(leader, buf, sizeof(buf));
267 pr_warning(" %s\n", buf);
270 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
271 pr_warning(" %s: %s\n", leader->name, buf);
272 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
273 pr_warning(" %s: %s\n", evsel->name, buf);
276 for_each_group_evsel(pos, leader)
277 evsel__remove_from_group(pos, leader);
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)
295 evlist__reset_stats(evsel_list);
296 perf_stat__reset_shadow_stats();
299 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
300 union perf_event *event,
301 struct perf_sample *sample __maybe_unused,
302 struct machine *machine __maybe_unused)
304 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
305 pr_err("failed to write perf data, error: %m\n");
309 perf_stat.bytes_written += event->header.size;
313 static int write_stat_round_event(u64 tm, u64 type)
315 return perf_event__synthesize_stat_round(NULL, tm, type,
316 process_synthesized_event,
320 #define WRITE_STAT_ROUND_EVENT(time, interval) \
321 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
323 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
325 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
326 struct perf_counts_values *count)
328 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
329 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
331 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
332 process_synthesized_event, NULL);
335 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
336 int thread, struct timespec *rs)
338 switch(counter->tool_event) {
339 case PERF_TOOL_DURATION_TIME: {
340 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
341 struct perf_counts_values *count =
342 perf_counts(counter->counts, cpu_map_idx, thread);
343 count->ena = count->run = val;
347 case PERF_TOOL_USER_TIME:
348 case PERF_TOOL_SYSTEM_TIME: {
350 struct perf_counts_values *count =
351 perf_counts(counter->counts, cpu_map_idx, thread);
352 if (counter->tool_event == PERF_TOOL_USER_TIME)
353 val = ru_stats.ru_utime_usec_stat.mean;
355 val = ru_stats.ru_stime_usec_stat.mean;
356 count->ena = count->run = val;
362 return evsel__read_counter(counter, cpu_map_idx, thread);
364 /* This should never be reached */
370 * Read out the results of a single counter:
371 * do not aggregate counts across CPUs in system-wide mode
373 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
375 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
378 if (!counter->supported)
381 for (thread = 0; thread < nthreads; thread++) {
382 struct perf_counts_values *count;
384 count = perf_counts(counter->counts, cpu_map_idx, thread);
387 * The leader's group read loads data into its group members
388 * (via evsel__read_counter()) and sets their count->loaded.
390 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
391 read_single_counter(counter, cpu_map_idx, thread, rs)) {
392 counter->counts->scaled = -1;
393 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
394 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
398 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
401 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
402 pr_err("failed to write stat event\n");
408 fprintf(stat_config.output,
409 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
410 evsel__name(counter),
411 perf_cpu_map__cpu(evsel__cpus(counter),
413 count->val, count->ena, count->run);
420 static int read_affinity_counters(struct timespec *rs)
422 struct evlist_cpu_iterator evlist_cpu_itr;
423 struct affinity saved_affinity, *affinity;
425 if (all_counters_use_bpf)
428 if (!target__has_cpu(&target) || target__has_per_thread(&target))
430 else if (affinity__setup(&saved_affinity) < 0)
433 affinity = &saved_affinity;
435 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
436 struct evsel *counter = evlist_cpu_itr.evsel;
438 if (evsel__is_bpf(counter))
442 counter->err = read_counter_cpu(counter, rs,
443 evlist_cpu_itr.cpu_map_idx);
447 affinity__cleanup(&saved_affinity);
452 static int read_bpf_map_counters(void)
454 struct evsel *counter;
457 evlist__for_each_entry(evsel_list, counter) {
458 if (!evsel__is_bpf(counter))
461 err = bpf_counter__read(counter);
468 static void read_counters(struct timespec *rs)
470 struct evsel *counter;
472 if (!stat_config.stop_read_counter) {
473 if (read_bpf_map_counters() ||
474 read_affinity_counters(rs))
478 evlist__for_each_entry(evsel_list, counter) {
480 pr_debug("failed to read counter %s\n", counter->name);
481 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
482 pr_warning("failed to process counter %s\n", counter->name);
487 static void process_interval(void)
489 struct timespec ts, rs;
491 clock_gettime(CLOCK_MONOTONIC, &ts);
492 diff_timespec(&rs, &ts, &ref_time);
494 perf_stat__reset_shadow_per_stat(&rt_stat);
498 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
499 pr_err("failed to write stat round event\n");
502 init_stats(&walltime_nsecs_stats);
503 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
504 print_counters(&rs, 0, NULL);
507 static bool handle_interval(unsigned int interval, int *times)
511 if (interval_count && !(--(*times)))
517 static int enable_counters(void)
522 evlist__for_each_entry(evsel_list, evsel) {
523 if (!evsel__is_bpf(evsel))
526 err = bpf_counter__enable(evsel);
532 * We need to enable counters only if:
533 * - we don't have tracee (attaching to task or cpu)
534 * - we have initial delay configured
536 if (!target__none(&target)) {
537 if (!all_counters_use_bpf)
538 evlist__enable(evsel_list);
543 static void disable_counters(void)
545 struct evsel *counter;
548 * If we don't have tracee (attaching to task or cpu), counters may
549 * still be running. To get accurate group ratios, we must stop groups
550 * from counting before reading their constituent counters.
552 if (!target__none(&target)) {
553 evlist__for_each_entry(evsel_list, counter)
554 bpf_counter__disable(counter);
555 if (!all_counters_use_bpf)
556 evlist__disable(evsel_list);
560 static volatile int workload_exec_errno;
563 * evlist__prepare_workload will send a SIGUSR1
564 * if the fork fails, since we asked by setting its
565 * want_signal to true.
567 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
568 void *ucontext __maybe_unused)
570 workload_exec_errno = info->si_value.sival_int;
573 static bool evsel__should_store_id(struct evsel *counter)
575 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
578 static bool is_target_alive(struct target *_target,
579 struct perf_thread_map *threads)
584 if (!target__has_task(_target))
587 for (i = 0; i < threads->nr; i++) {
590 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
591 threads->map[i].pid);
593 if (!stat(path, &st))
600 static void process_evlist(struct evlist *evlist, unsigned int interval)
602 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
604 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
606 case EVLIST_CTL_CMD_ENABLE:
608 case EVLIST_CTL_CMD_DISABLE:
612 case EVLIST_CTL_CMD_SNAPSHOT:
613 case EVLIST_CTL_CMD_ACK:
614 case EVLIST_CTL_CMD_UNSUPPORTED:
615 case EVLIST_CTL_CMD_EVLIST:
616 case EVLIST_CTL_CMD_STOP:
617 case EVLIST_CTL_CMD_PING:
624 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
627 int tts = *time_to_sleep;
628 struct timespec time_diff;
630 diff_timespec(&time_diff, time_stop, time_start);
632 tts -= time_diff.tv_sec * MSEC_PER_SEC +
633 time_diff.tv_nsec / NSEC_PER_MSEC;
638 *time_to_sleep = tts;
641 static int dispatch_events(bool forks, int timeout, int interval, int *times)
643 int child_exited = 0, status = 0;
644 int time_to_sleep, sleep_time;
645 struct timespec time_start, time_stop;
648 sleep_time = interval;
650 sleep_time = timeout;
654 time_to_sleep = sleep_time;
658 child_exited = waitpid(child_pid, &status, WNOHANG);
660 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
665 clock_gettime(CLOCK_MONOTONIC, &time_start);
666 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
667 if (timeout || handle_interval(interval, times))
669 time_to_sleep = sleep_time;
670 } else { /* fd revent */
671 process_evlist(evsel_list, interval);
672 clock_gettime(CLOCK_MONOTONIC, &time_stop);
673 compute_tts(&time_start, &time_stop, &time_to_sleep);
680 enum counter_recovery {
686 static enum counter_recovery stat_handle_error(struct evsel *counter)
690 * PPC returns ENXIO for HW counters until 2.6.37
691 * (behavior changed with commit b0a873e).
693 if (errno == EINVAL || errno == ENOSYS ||
694 errno == ENOENT || errno == EOPNOTSUPP ||
697 ui__warning("%s event is not supported by the kernel.\n",
698 evsel__name(counter));
699 counter->supported = false;
701 * errored is a sticky flag that means one of the counter's
702 * cpu event had a problem and needs to be reexamined.
704 counter->errored = true;
706 if ((evsel__leader(counter) != counter) ||
707 !(counter->core.leader->nr_members > 1))
709 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
711 ui__warning("%s\n", msg);
712 return COUNTER_RETRY;
713 } else if (target__has_per_thread(&target) &&
714 evsel_list->core.threads &&
715 evsel_list->core.threads->err_thread != -1) {
717 * For global --per-thread case, skip current
720 if (!thread_map__remove(evsel_list->core.threads,
721 evsel_list->core.threads->err_thread)) {
722 evsel_list->core.threads->err_thread = -1;
723 return COUNTER_RETRY;
727 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
728 ui__error("%s\n", msg);
731 kill(child_pid, SIGTERM);
732 return COUNTER_FATAL;
735 static int __run_perf_stat(int argc, const char **argv, int run_idx)
737 int interval = stat_config.interval;
738 int times = stat_config.times;
739 int timeout = stat_config.timeout;
741 unsigned long long t0, t1;
742 struct evsel *counter;
745 const bool forks = (argc > 0);
746 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
747 struct evlist_cpu_iterator evlist_cpu_itr;
748 struct affinity saved_affinity, *affinity = NULL;
750 bool second_pass = false;
753 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
754 perror("failed to prepare workload");
757 child_pid = evsel_list->workload.pid;
761 evlist__set_leader(evsel_list);
763 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
764 if (affinity__setup(&saved_affinity) < 0)
766 affinity = &saved_affinity;
769 evlist__for_each_entry(evsel_list, counter) {
770 counter->reset_group = false;
771 if (bpf_counter__load(counter, &target))
773 if (!evsel__is_bpf(counter))
774 all_counters_use_bpf = false;
777 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
778 counter = evlist_cpu_itr.evsel;
781 * bperf calls evsel__open_per_cpu() in bperf__load(), so
782 * no need to call it again here.
787 if (counter->reset_group || counter->errored)
789 if (evsel__is_bpf(counter))
792 if (create_perf_stat_counter(counter, &stat_config, &target,
793 evlist_cpu_itr.cpu_map_idx) < 0) {
796 * Weak group failed. We cannot just undo this here
797 * because earlier CPUs might be in group mode, and the kernel
798 * doesn't support mixing group and non group reads. Defer
800 * Don't close here because we're in the wrong affinity.
802 if ((errno == EINVAL || errno == EBADF) &&
803 evsel__leader(counter) != counter &&
804 counter->weak_group) {
805 evlist__reset_weak_group(evsel_list, counter, false);
806 assert(counter->reset_group);
811 switch (stat_handle_error(counter)) {
823 counter->supported = true;
828 * Now redo all the weak group after closing them,
829 * and also close errored counters.
832 /* First close errored or weak retry */
833 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
834 counter = evlist_cpu_itr.evsel;
836 if (!counter->reset_group && !counter->errored)
839 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
841 /* Now reopen weak */
842 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
843 counter = evlist_cpu_itr.evsel;
845 if (!counter->reset_group)
848 pr_debug2("reopening weak %s\n", evsel__name(counter));
849 if (create_perf_stat_counter(counter, &stat_config, &target,
850 evlist_cpu_itr.cpu_map_idx) < 0) {
852 switch (stat_handle_error(counter)) {
856 goto try_again_reset;
863 counter->supported = true;
866 affinity__cleanup(affinity);
868 evlist__for_each_entry(evsel_list, counter) {
869 if (!counter->supported) {
870 perf_evsel__free_fd(&counter->core);
874 l = strlen(counter->unit);
875 if (l > stat_config.unit_width)
876 stat_config.unit_width = l;
878 if (evsel__should_store_id(counter) &&
879 evsel__store_ids(counter, evsel_list))
883 if (evlist__apply_filters(evsel_list, &counter)) {
884 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
885 counter->filter, evsel__name(counter), errno,
886 str_error_r(errno, msg, sizeof(msg)));
891 int fd = perf_data__fd(&perf_stat.data);
894 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
896 err = perf_session__write_header(perf_stat.session, evsel_list,
903 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
904 process_synthesized_event, is_pipe);
909 if (stat_config.initial_delay) {
910 pr_info(EVLIST_DISABLED_MSG);
912 err = enable_counters();
917 /* Exec the command, if any */
919 evlist__start_workload(evsel_list);
921 if (stat_config.initial_delay > 0) {
922 usleep(stat_config.initial_delay * USEC_PER_MSEC);
923 err = enable_counters();
927 pr_info(EVLIST_ENABLED_MSG);
931 clock_gettime(CLOCK_MONOTONIC, &ref_time);
934 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
935 status = dispatch_events(forks, timeout, interval, ×);
936 if (child_pid != -1) {
938 kill(child_pid, SIGTERM);
939 wait4(child_pid, &status, 0, &stat_config.ru_data);
942 if (workload_exec_errno) {
943 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
944 pr_err("Workload failed: %s\n", emsg);
948 if (WIFSIGNALED(status))
949 psignal(WTERMSIG(status), argv[0]);
951 status = dispatch_events(forks, timeout, interval, ×);
958 if (stat_config.walltime_run_table)
959 stat_config.walltime_run[run_idx] = t1 - t0;
961 if (interval && stat_config.summary) {
962 stat_config.interval = 0;
963 stat_config.stop_read_counter = true;
964 init_stats(&walltime_nsecs_stats);
965 update_stats(&walltime_nsecs_stats, t1 - t0);
967 if (stat_config.aggr_mode == AGGR_GLOBAL)
968 evlist__save_aggr_prev_raw_counts(evsel_list);
970 evlist__copy_prev_raw_counts(evsel_list);
971 evlist__reset_prev_raw_counts(evsel_list);
972 perf_stat__reset_shadow_per_stat(&rt_stat);
974 update_stats(&walltime_nsecs_stats, t1 - t0);
975 update_rusage_stats(&ru_stats, &stat_config.ru_data);
979 * Closing a group leader splits the group, and as we only disable
980 * group leaders, results in remaining events becoming enabled. To
981 * avoid arbitrary skew, we must read all counters before closing any
984 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
987 * We need to keep evsel_list alive, because it's processed
988 * later the evsel_list will be closed after.
991 evlist__close(evsel_list);
993 return WEXITSTATUS(status);
996 static int run_perf_stat(int argc, const char **argv, int run_idx)
1001 ret = system(pre_cmd);
1009 ret = __run_perf_stat(argc, argv, run_idx);
1014 ret = system(post_cmd);
1022 static void print_counters(struct timespec *ts, int argc, const char **argv)
1024 /* Do not print anything if we record to the pipe. */
1025 if (STAT_RECORD && perf_stat.data.is_pipe)
1030 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1033 static volatile int signr = -1;
1035 static void skip_signal(int signo)
1037 if ((child_pid == -1) || stat_config.interval)
1042 * render child_pid harmless
1043 * won't send SIGTERM to a random
1044 * process in case of race condition
1045 * and fast PID recycling
1050 static void sig_atexit(void)
1055 * avoid race condition with SIGCHLD handler
1056 * in skip_signal() which is modifying child_pid
1057 * goal is to avoid send SIGTERM to a random
1061 sigaddset(&set, SIGCHLD);
1062 sigprocmask(SIG_BLOCK, &set, &oset);
1064 if (child_pid != -1)
1065 kill(child_pid, SIGTERM);
1067 sigprocmask(SIG_SETMASK, &oset, NULL);
1072 signal(signr, SIG_DFL);
1073 kill(getpid(), signr);
1076 void perf_stat__set_big_num(int set)
1078 stat_config.big_num = (set != 0);
1081 void perf_stat__set_no_csv_summary(int set)
1083 stat_config.no_csv_summary = (set != 0);
1086 static int stat__set_big_num(const struct option *opt __maybe_unused,
1087 const char *s __maybe_unused, int unset)
1089 big_num_opt = unset ? 0 : 1;
1090 perf_stat__set_big_num(!unset);
1094 static int enable_metric_only(const struct option *opt __maybe_unused,
1095 const char *s __maybe_unused, int unset)
1097 force_metric_only = true;
1098 stat_config.metric_only = !unset;
1102 static int append_metric_groups(const struct option *opt __maybe_unused,
1104 int unset __maybe_unused)
1109 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1114 metrics = strdup(str);
1121 static int parse_control_option(const struct option *opt,
1123 int unset __maybe_unused)
1125 struct perf_stat_config *config = opt->value;
1127 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1130 static int parse_stat_cgroups(const struct option *opt,
1131 const char *str, int unset)
1133 if (stat_config.cgroup_list) {
1134 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1138 return parse_cgroups(opt, str, unset);
1141 static int parse_hybrid_type(const struct option *opt,
1143 int unset __maybe_unused)
1145 struct evlist *evlist = *(struct evlist **)opt->value;
1147 if (!list_empty(&evlist->core.entries)) {
1148 fprintf(stderr, "Must define cputype before events/metrics\n");
1152 evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1153 if (!evlist->hybrid_pmu_name) {
1154 fprintf(stderr, "--cputype %s is not supported!\n", str);
1161 static struct option stat_options[] = {
1162 OPT_BOOLEAN('T', "transaction", &transaction_run,
1163 "hardware transaction statistics"),
1164 OPT_CALLBACK('e', "event", &evsel_list, "event",
1165 "event selector. use 'perf list' to list available events",
1166 parse_events_option),
1167 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1168 "event filter", parse_filter),
1169 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1170 "child tasks do not inherit counters"),
1171 OPT_STRING('p', "pid", &target.pid, "pid",
1172 "stat events on existing process id"),
1173 OPT_STRING('t', "tid", &target.tid, "tid",
1174 "stat events on existing thread id"),
1175 #ifdef HAVE_BPF_SKEL
1176 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1177 "stat events on existing bpf program id"),
1178 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1179 "use bpf program to count events"),
1180 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1181 "path to perf_event_attr map"),
1183 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1184 "system-wide collection from all CPUs"),
1185 OPT_BOOLEAN('g', "group", &group,
1186 "put the counters into a counter group"),
1187 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1188 "Use --no-scale to disable counter scaling for multiplexing"),
1189 OPT_INCR('v', "verbose", &verbose,
1190 "be more verbose (show counter open errors, etc)"),
1191 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1192 "repeat command and print average + stddev (max: 100, forever: 0)"),
1193 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1194 "display details about each run (only with -r option)"),
1195 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1196 "null run - dont start any counters"),
1197 OPT_INCR('d', "detailed", &detailed_run,
1198 "detailed run - start a lot of events"),
1199 OPT_BOOLEAN('S', "sync", &sync_run,
1200 "call sync() before starting a run"),
1201 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1202 "print large numbers with thousands\' separators",
1204 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1205 "list of cpus to monitor in system-wide"),
1206 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1207 "disable CPU count aggregation", AGGR_NONE),
1208 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1209 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1210 "Merge identical named hybrid events"),
1211 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1212 "print counts with custom separator"),
1213 OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1214 "print counts in JSON format"),
1215 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1216 "monitor event in cgroup name only", parse_stat_cgroups),
1217 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1218 "expand events for each cgroup"),
1219 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1220 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1221 OPT_INTEGER(0, "log-fd", &output_fd,
1222 "log output to fd, instead of stderr"),
1223 OPT_STRING(0, "pre", &pre_cmd, "command",
1224 "command to run prior to the measured command"),
1225 OPT_STRING(0, "post", &post_cmd, "command",
1226 "command to run after to the measured command"),
1227 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1228 "print counts at regular interval in ms "
1229 "(overhead is possible for values <= 100ms)"),
1230 OPT_INTEGER(0, "interval-count", &stat_config.times,
1231 "print counts for fixed number of times"),
1232 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1233 "clear screen in between new interval"),
1234 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1235 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1236 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1237 "aggregate counts per processor socket", AGGR_SOCKET),
1238 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1239 "aggregate counts per processor die", AGGR_DIE),
1240 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1241 "aggregate counts per physical processor core", AGGR_CORE),
1242 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1243 "aggregate counts per thread", AGGR_THREAD),
1244 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1245 "aggregate counts per numa node", AGGR_NODE),
1246 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1247 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1248 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1249 "Only print computed metrics. No raw values", enable_metric_only),
1250 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1251 "don't group metric events, impacts multiplexing"),
1252 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1253 "don't try to share events between metrics in a group"),
1254 OPT_BOOLEAN(0, "topdown", &topdown_run,
1255 "measure top-down statistics"),
1256 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1257 "Set the metrics level for the top-down statistics (0: max level)"),
1258 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1259 "measure SMI cost"),
1260 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1261 "monitor specified metrics or metric groups (separated by ,)",
1262 append_metric_groups),
1263 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1264 "Configure all used events to run in kernel space.",
1265 PARSE_OPT_EXCLUSIVE),
1266 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1267 "Configure all used events to run in user space.",
1268 PARSE_OPT_EXCLUSIVE),
1269 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1270 "Use with 'percore' event qualifier to show the event "
1271 "counts of one hardware thread by sum up total hardware "
1272 "threads of same physical core"),
1273 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1274 "print summary for interval mode"),
1275 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1276 "don't print 'summary' for CSV summary output"),
1277 OPT_BOOLEAN(0, "quiet", &quiet,
1278 "don't print any output, messages or warnings (useful with record)"),
1279 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1280 "Only enable events on applying cpu with this type "
1281 "for hybrid platform (e.g. core or atom)",
1284 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1285 "libpfm4 event selector. use 'perf list' to list available events",
1286 parse_libpfm_events_option),
1288 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1289 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1290 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1291 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1292 parse_control_option),
1293 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1294 "measure I/O performance metrics provided by arch/platform",
1299 static const char *const aggr_mode__string[] = {
1300 [AGGR_CORE] = "core",
1302 [AGGR_GLOBAL] = "global",
1303 [AGGR_NODE] = "node",
1304 [AGGR_NONE] = "none",
1305 [AGGR_SOCKET] = "socket",
1306 [AGGR_THREAD] = "thread",
1307 [AGGR_UNSET] = "unset",
1310 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1311 struct perf_cpu cpu)
1313 return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1316 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1317 struct perf_cpu cpu)
1319 return aggr_cpu_id__die(cpu, /*data=*/NULL);
1322 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1323 struct perf_cpu cpu)
1325 return aggr_cpu_id__core(cpu, /*data=*/NULL);
1328 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1329 struct perf_cpu cpu)
1331 return aggr_cpu_id__node(cpu, /*data=*/NULL);
1334 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1335 aggr_get_id_t get_id, struct perf_cpu cpu)
1337 struct aggr_cpu_id id = aggr_cpu_id__empty();
1339 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1340 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1342 id = config->cpus_aggr_map->map[cpu.cpu];
1346 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1347 struct perf_cpu cpu)
1349 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1352 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1353 struct perf_cpu cpu)
1355 return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1358 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1359 struct perf_cpu cpu)
1361 return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1364 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1365 struct perf_cpu cpu)
1367 return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1370 static bool term_percore_set(void)
1372 struct evsel *counter;
1374 evlist__for_each_entry(evsel_list, counter) {
1375 if (counter->percore)
1382 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1384 switch (aggr_mode) {
1386 return aggr_cpu_id__socket;
1388 return aggr_cpu_id__die;
1390 return aggr_cpu_id__core;
1392 return aggr_cpu_id__node;
1394 if (term_percore_set())
1395 return aggr_cpu_id__core;
1407 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1409 switch (aggr_mode) {
1411 return perf_stat__get_socket_cached;
1413 return perf_stat__get_die_cached;
1415 return perf_stat__get_core_cached;
1417 return perf_stat__get_node_cached;
1419 if (term_percore_set()) {
1420 return perf_stat__get_core_cached;
1432 static int perf_stat_init_aggr_mode(void)
1435 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1438 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1439 get_id, /*data=*/NULL);
1440 if (!stat_config.aggr_map) {
1441 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1444 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1448 * The evsel_list->cpus is the base we operate on,
1449 * taking the highest cpu number to be the size of
1450 * the aggregation translate cpumap.
1452 if (evsel_list->core.user_requested_cpus)
1453 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1456 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1457 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1460 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1463 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1464 "cpu_aggr_map refcnt unbalanced\n");
1469 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1471 if (map && refcount_dec_and_test(&map->refcnt))
1472 cpu_aggr_map__delete(map);
1475 static void perf_stat__exit_aggr_mode(void)
1477 cpu_aggr_map__put(stat_config.aggr_map);
1478 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1479 stat_config.aggr_map = NULL;
1480 stat_config.cpus_aggr_map = NULL;
1483 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1485 struct perf_env *env = data;
1486 struct aggr_cpu_id id = aggr_cpu_id__empty();
1489 id.socket = env->cpu[cpu.cpu].socket_id;
1494 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1496 struct perf_env *env = data;
1497 struct aggr_cpu_id id = aggr_cpu_id__empty();
1499 if (cpu.cpu != -1) {
1501 * die_id is relative to socket, so start
1502 * with the socket ID and then add die to
1505 id.socket = env->cpu[cpu.cpu].socket_id;
1506 id.die = env->cpu[cpu.cpu].die_id;
1512 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1514 struct perf_env *env = data;
1515 struct aggr_cpu_id id = aggr_cpu_id__empty();
1517 if (cpu.cpu != -1) {
1519 * core_id is relative to socket and die,
1520 * we need a global id. So we set
1521 * socket, die id and core id
1523 id.socket = env->cpu[cpu.cpu].socket_id;
1524 id.die = env->cpu[cpu.cpu].die_id;
1525 id.core = env->cpu[cpu.cpu].core_id;
1531 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1533 struct aggr_cpu_id id = aggr_cpu_id__empty();
1535 id.node = perf_env__numa_node(data, cpu);
1539 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1540 struct perf_cpu cpu)
1542 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1544 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1545 struct perf_cpu cpu)
1547 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1550 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1551 struct perf_cpu cpu)
1553 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1556 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1557 struct perf_cpu cpu)
1559 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1562 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1564 switch (aggr_mode) {
1566 return perf_env__get_socket_aggr_by_cpu;
1568 return perf_env__get_die_aggr_by_cpu;
1570 return perf_env__get_core_aggr_by_cpu;
1572 return perf_env__get_node_aggr_by_cpu;
1583 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1585 switch (aggr_mode) {
1587 return perf_stat__get_socket_file;
1589 return perf_stat__get_die_file;
1591 return perf_stat__get_core_file;
1593 return perf_stat__get_node_file;
1604 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1606 struct perf_env *env = &st->session->header.env;
1607 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1612 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1613 if (!stat_config.aggr_map) {
1614 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1617 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1622 * Add default attributes, if there were no attributes specified or
1623 * if -d/--detailed, -d -d or -d -d -d is used:
1625 static int add_default_attributes(void)
1628 struct perf_event_attr default_attrs0[] = {
1630 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1631 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1632 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1633 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1635 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1637 struct perf_event_attr frontend_attrs[] = {
1638 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1640 struct perf_event_attr backend_attrs[] = {
1641 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1643 struct perf_event_attr default_attrs1[] = {
1644 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1645 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1646 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1651 * Detailed stats (-d), covering the L1 and last level data caches:
1653 struct perf_event_attr detailed_attrs[] = {
1655 { .type = PERF_TYPE_HW_CACHE,
1657 PERF_COUNT_HW_CACHE_L1D << 0 |
1658 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1659 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1661 { .type = PERF_TYPE_HW_CACHE,
1663 PERF_COUNT_HW_CACHE_L1D << 0 |
1664 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1665 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1667 { .type = PERF_TYPE_HW_CACHE,
1669 PERF_COUNT_HW_CACHE_LL << 0 |
1670 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1671 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1673 { .type = PERF_TYPE_HW_CACHE,
1675 PERF_COUNT_HW_CACHE_LL << 0 |
1676 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1677 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1681 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1683 struct perf_event_attr very_detailed_attrs[] = {
1685 { .type = PERF_TYPE_HW_CACHE,
1687 PERF_COUNT_HW_CACHE_L1I << 0 |
1688 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1689 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1691 { .type = PERF_TYPE_HW_CACHE,
1693 PERF_COUNT_HW_CACHE_L1I << 0 |
1694 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1695 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1697 { .type = PERF_TYPE_HW_CACHE,
1699 PERF_COUNT_HW_CACHE_DTLB << 0 |
1700 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1701 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1703 { .type = PERF_TYPE_HW_CACHE,
1705 PERF_COUNT_HW_CACHE_DTLB << 0 |
1706 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1707 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1709 { .type = PERF_TYPE_HW_CACHE,
1711 PERF_COUNT_HW_CACHE_ITLB << 0 |
1712 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1713 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1715 { .type = PERF_TYPE_HW_CACHE,
1717 PERF_COUNT_HW_CACHE_ITLB << 0 |
1718 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1719 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1724 * Very, very detailed stats (-d -d -d), adding prefetch events:
1726 struct perf_event_attr very_very_detailed_attrs[] = {
1728 { .type = PERF_TYPE_HW_CACHE,
1730 PERF_COUNT_HW_CACHE_L1D << 0 |
1731 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1732 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1734 { .type = PERF_TYPE_HW_CACHE,
1736 PERF_COUNT_HW_CACHE_L1D << 0 |
1737 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1738 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1741 struct perf_event_attr default_null_attrs[] = {};
1743 /* Set attrs if no event is selected and !null_run: */
1744 if (stat_config.null_run)
1747 if (transaction_run) {
1748 struct parse_events_error errinfo;
1749 /* Handle -T as -M transaction. Once platform specific metrics
1750 * support has been added to the json files, all architectures
1751 * will use this approach. To determine transaction support
1752 * on an architecture test for such a metric name.
1754 if (metricgroup__has_metric("transaction")) {
1755 return metricgroup__parse_groups(evsel_list, "transaction",
1756 stat_config.metric_no_group,
1757 stat_config.metric_no_merge,
1758 stat_config.user_requested_cpu_list,
1759 stat_config.system_wide,
1760 &stat_config.metric_events);
1763 parse_events_error__init(&errinfo);
1764 if (pmu_have_event("cpu", "cycles-ct") &&
1765 pmu_have_event("cpu", "el-start"))
1766 err = parse_events(evsel_list, transaction_attrs,
1769 err = parse_events(evsel_list,
1770 transaction_limited_attrs,
1773 fprintf(stderr, "Cannot set up transaction events\n");
1774 parse_events_error__print(&errinfo, transaction_attrs);
1776 parse_events_error__exit(&errinfo);
1777 return err ? -1 : 0;
1781 struct parse_events_error errinfo;
1784 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1785 fprintf(stderr, "freeze_on_smi is not supported.\n");
1790 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1791 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1797 if (!pmu_have_event("msr", "aperf") ||
1798 !pmu_have_event("msr", "smi")) {
1799 fprintf(stderr, "To measure SMI cost, it needs "
1800 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1803 if (!force_metric_only)
1804 stat_config.metric_only = true;
1806 parse_events_error__init(&errinfo);
1807 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1809 parse_events_error__print(&errinfo, smi_cost_attrs);
1810 fprintf(stderr, "Cannot set up SMI cost events\n");
1812 parse_events_error__exit(&errinfo);
1813 return err ? -1 : 0;
1817 const char **metric_attrs = topdown_metric_attrs;
1818 unsigned int max_level = 1;
1821 const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1823 if (!force_metric_only)
1824 stat_config.metric_only = true;
1826 if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1827 metric_attrs = topdown_metric_L2_attrs;
1831 if (stat_config.topdown_level > max_level) {
1832 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1834 } else if (!stat_config.topdown_level)
1835 stat_config.topdown_level = max_level;
1837 if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1838 pr_err("Out of memory\n");
1842 if (metric_attrs[0] && str) {
1843 if (!stat_config.interval && !stat_config.metric_only) {
1844 fprintf(stat_config.output,
1845 "Topdown accuracy may decrease when measuring long periods.\n"
1846 "Please print the result regularly, e.g. -I1000\n");
1853 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1854 stat_config.aggr_mode != AGGR_CORE) {
1855 pr_err("top down event configuration requires --per-core mode\n");
1858 stat_config.aggr_mode = AGGR_CORE;
1859 if (nr_cgroups || !target__has_cpu(&target)) {
1860 pr_err("top down event configuration requires system-wide mode (-a)\n");
1864 if (topdown_filter_events(topdown_attrs, &str,
1865 arch_topdown_check_group(&warn),
1867 pr_err("Out of memory\n");
1871 if (topdown_attrs[0] && str) {
1872 struct parse_events_error errinfo;
1874 arch_topdown_group_warn();
1876 parse_events_error__init(&errinfo);
1877 err = parse_events(evsel_list, str, &errinfo);
1880 "Cannot set up top down events %s: %d\n",
1882 parse_events_error__print(&errinfo, str);
1883 parse_events_error__exit(&errinfo);
1887 parse_events_error__exit(&errinfo);
1889 fprintf(stderr, "System does not support topdown\n");
1895 if (!stat_config.topdown_level)
1896 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1898 if (!evsel_list->core.nr_entries) {
1899 if (target__has_cpu(&target))
1900 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1902 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1904 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1905 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1908 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1909 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1912 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1914 /* Platform specific attrs */
1915 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
1919 /* Detailed events get appended to the event list: */
1921 if (detailed_run < 1)
1924 /* Append detailed run extra attributes: */
1925 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1928 if (detailed_run < 2)
1931 /* Append very detailed run extra attributes: */
1932 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1935 if (detailed_run < 3)
1938 /* Append very, very detailed run extra attributes: */
1939 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1942 static const char * const stat_record_usage[] = {
1943 "perf stat record [<options>]",
1947 static void init_features(struct perf_session *session)
1951 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1952 perf_header__set_feat(&session->header, feat);
1954 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1955 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1956 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1957 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1958 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1961 static int __cmd_record(int argc, const char **argv)
1963 struct perf_session *session;
1964 struct perf_data *data = &perf_stat.data;
1966 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1967 PARSE_OPT_STOP_AT_NON_OPTION);
1970 data->path = output_name;
1972 if (stat_config.run_count != 1 || forever) {
1973 pr_err("Cannot use -r option with perf stat record.\n");
1977 session = perf_session__new(data, NULL);
1978 if (IS_ERR(session)) {
1979 pr_err("Perf session creation failed\n");
1980 return PTR_ERR(session);
1983 init_features(session);
1985 session->evlist = evsel_list;
1986 perf_stat.session = session;
1987 perf_stat.record = true;
1991 static int process_stat_round_event(struct perf_session *session,
1992 union perf_event *event)
1994 struct perf_record_stat_round *stat_round = &event->stat_round;
1995 struct evsel *counter;
1996 struct timespec tsh, *ts = NULL;
1997 const char **argv = session->header.env.cmdline_argv;
1998 int argc = session->header.env.nr_cmdline;
2000 evlist__for_each_entry(evsel_list, counter)
2001 perf_stat_process_counter(&stat_config, counter);
2003 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2004 update_stats(&walltime_nsecs_stats, stat_round->time);
2006 if (stat_config.interval && stat_round->time) {
2007 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2008 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2012 print_counters(ts, argc, argv);
2017 int process_stat_config_event(struct perf_session *session,
2018 union perf_event *event)
2020 struct perf_tool *tool = session->tool;
2021 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2023 perf_event__read_stat_config(&stat_config, &event->stat_config);
2025 if (perf_cpu_map__empty(st->cpus)) {
2026 if (st->aggr_mode != AGGR_UNSET)
2027 pr_warning("warning: processing task data, aggregation mode not set\n");
2031 if (st->aggr_mode != AGGR_UNSET)
2032 stat_config.aggr_mode = st->aggr_mode;
2034 if (perf_stat.data.is_pipe)
2035 perf_stat_init_aggr_mode();
2037 perf_stat_init_aggr_mode_file(st);
2042 static int set_maps(struct perf_stat *st)
2044 if (!st->cpus || !st->threads)
2047 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2050 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2052 if (evlist__alloc_stats(evsel_list, true))
2055 st->maps_allocated = true;
2060 int process_thread_map_event(struct perf_session *session,
2061 union perf_event *event)
2063 struct perf_tool *tool = session->tool;
2064 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2067 pr_warning("Extra thread map event, ignoring.\n");
2071 st->threads = thread_map__new_event(&event->thread_map);
2075 return set_maps(st);
2079 int process_cpu_map_event(struct perf_session *session,
2080 union perf_event *event)
2082 struct perf_tool *tool = session->tool;
2083 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2084 struct perf_cpu_map *cpus;
2087 pr_warning("Extra cpu map event, ignoring.\n");
2091 cpus = cpu_map__new_data(&event->cpu_map.data);
2096 return set_maps(st);
2099 static const char * const stat_report_usage[] = {
2100 "perf stat report [<options>]",
2104 static struct perf_stat perf_stat = {
2106 .attr = perf_event__process_attr,
2107 .event_update = perf_event__process_event_update,
2108 .thread_map = process_thread_map_event,
2109 .cpu_map = process_cpu_map_event,
2110 .stat_config = process_stat_config_event,
2111 .stat = perf_event__process_stat_event,
2112 .stat_round = process_stat_round_event,
2114 .aggr_mode = AGGR_UNSET,
2117 static int __cmd_report(int argc, const char **argv)
2119 struct perf_session *session;
2120 const struct option options[] = {
2121 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2122 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2123 "aggregate counts per processor socket", AGGR_SOCKET),
2124 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2125 "aggregate counts per processor die", AGGR_DIE),
2126 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2127 "aggregate counts per physical processor core", AGGR_CORE),
2128 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2129 "aggregate counts per numa node", AGGR_NODE),
2130 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2131 "disable CPU count aggregation", AGGR_NONE),
2137 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2139 if (!input_name || !strlen(input_name)) {
2140 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2143 input_name = "perf.data";
2146 perf_stat__init_shadow_stats();
2148 perf_stat.data.path = input_name;
2149 perf_stat.data.mode = PERF_DATA_MODE_READ;
2151 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2152 if (IS_ERR(session))
2153 return PTR_ERR(session);
2155 perf_stat.session = session;
2156 stat_config.output = stderr;
2157 evsel_list = session->evlist;
2159 ret = perf_session__process_events(session);
2163 perf_session__delete(session);
2167 static void setup_system_wide(int forks)
2170 * Make system wide (-a) the default target if
2171 * no target was specified and one of following
2172 * conditions is met:
2174 * - there's no workload specified
2175 * - there is workload specified but all requested
2176 * events are system wide events
2178 if (!target__none(&target))
2182 target.system_wide = true;
2184 struct evsel *counter;
2186 evlist__for_each_entry(evsel_list, counter) {
2187 if (!counter->core.requires_cpu &&
2188 strcmp(counter->name, "duration_time")) {
2193 if (evsel_list->core.nr_entries)
2194 target.system_wide = true;
2198 int cmd_stat(int argc, const char **argv)
2200 const char * const stat_usage[] = {
2201 "perf stat [<options>] [<command>]",
2204 int status = -EINVAL, run_idx, err;
2206 FILE *output = stderr;
2207 unsigned int interval, timeout;
2208 const char * const stat_subcommands[] = { "record", "report" };
2209 char errbuf[BUFSIZ];
2211 setlocale(LC_ALL, "");
2213 evsel_list = evlist__new();
2214 if (evsel_list == NULL)
2217 parse_events__shrink_config_terms();
2219 /* String-parsing callback-based options would segfault when negated */
2220 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2221 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2222 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2224 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2225 (const char **) stat_usage,
2226 PARSE_OPT_STOP_AT_NON_OPTION);
2228 if (stat_config.csv_sep) {
2229 stat_config.csv_output = true;
2230 if (!strcmp(stat_config.csv_sep, "\\t"))
2231 stat_config.csv_sep = "\t";
2233 stat_config.csv_sep = DEFAULT_SEPARATOR;
2235 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2236 argc = __cmd_record(argc, argv);
2239 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2240 return __cmd_report(argc, argv);
2242 interval = stat_config.interval;
2243 timeout = stat_config.timeout;
2246 * For record command the -o is already taken care of.
2248 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2251 if (output_name && output_fd) {
2252 fprintf(stderr, "cannot use both --output and --log-fd\n");
2253 parse_options_usage(stat_usage, stat_options, "o", 1);
2254 parse_options_usage(NULL, stat_options, "log-fd", 0);
2258 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2259 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2263 if (stat_config.metric_only && stat_config.run_count > 1) {
2264 fprintf(stderr, "--metric-only is not supported with -r\n");
2268 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2269 fprintf(stderr, "--table is only supported with -r\n");
2270 parse_options_usage(stat_usage, stat_options, "r", 1);
2271 parse_options_usage(NULL, stat_options, "table", 0);
2275 if (output_fd < 0) {
2276 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2277 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2281 if (!output && !quiet) {
2283 mode = append_file ? "a" : "w";
2285 output = fopen(output_name, mode);
2287 perror("failed to create output file");
2290 clock_gettime(CLOCK_REALTIME, &tm);
2291 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2292 } else if (output_fd > 0) {
2293 mode = append_file ? "a" : "w";
2294 output = fdopen(output_fd, mode);
2296 perror("Failed opening logfd");
2301 stat_config.output = output;
2304 * let the spreadsheet do the pretty-printing
2306 if (stat_config.csv_output) {
2307 /* User explicitly passed -B? */
2308 if (big_num_opt == 1) {
2309 fprintf(stderr, "-B option not supported with -x\n");
2310 parse_options_usage(stat_usage, stat_options, "B", 1);
2311 parse_options_usage(NULL, stat_options, "x", 1);
2313 } else /* Nope, so disable big number formatting */
2314 stat_config.big_num = false;
2315 } else if (big_num_opt == 0) /* User passed --no-big-num */
2316 stat_config.big_num = false;
2318 err = target__validate(&target);
2320 target__strerror(&target, err, errbuf, BUFSIZ);
2321 pr_warning("%s\n", errbuf);
2324 setup_system_wide(argc);
2327 * Display user/system times only for single
2328 * run and when there's specified tracee.
2330 if ((stat_config.run_count == 1) && target__none(&target))
2331 stat_config.ru_display = true;
2333 if (stat_config.run_count < 0) {
2334 pr_err("Run count must be a positive number\n");
2335 parse_options_usage(stat_usage, stat_options, "r", 1);
2337 } else if (stat_config.run_count == 0) {
2339 stat_config.run_count = 1;
2342 if (stat_config.walltime_run_table) {
2343 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2344 if (!stat_config.walltime_run) {
2345 pr_err("failed to setup -r option");
2350 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2351 !target__has_task(&target)) {
2352 if (!target.system_wide || target.cpu_list) {
2353 fprintf(stderr, "The --per-thread option is only "
2354 "available when monitoring via -p -t -a "
2355 "options or only --per-thread.\n");
2356 parse_options_usage(NULL, stat_options, "p", 1);
2357 parse_options_usage(NULL, stat_options, "t", 1);
2363 * no_aggr, cgroup are for system-wide only
2364 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2366 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2367 stat_config.aggr_mode != AGGR_THREAD) ||
2368 (nr_cgroups || stat_config.cgroup_list)) &&
2369 !target__has_cpu(&target)) {
2370 fprintf(stderr, "both cgroup and no-aggregation "
2371 "modes only available in system-wide mode\n");
2373 parse_options_usage(stat_usage, stat_options, "G", 1);
2374 parse_options_usage(NULL, stat_options, "A", 1);
2375 parse_options_usage(NULL, stat_options, "a", 1);
2376 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2380 if (stat_config.iostat_run) {
2381 status = iostat_prepare(evsel_list, &stat_config);
2384 if (iostat_mode == IOSTAT_LIST) {
2385 iostat_list(evsel_list, &stat_config);
2388 iostat_list(evsel_list, &stat_config);
2389 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2390 target.system_wide = true;
2393 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2394 target.per_thread = true;
2396 stat_config.system_wide = target.system_wide;
2397 if (target.cpu_list) {
2398 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2399 if (!stat_config.user_requested_cpu_list) {
2406 * Metric parsing needs to be delayed as metrics may optimize events
2407 * knowing the target is system-wide.
2410 metricgroup__parse_groups(evsel_list, metrics,
2411 stat_config.metric_no_group,
2412 stat_config.metric_no_merge,
2413 stat_config.user_requested_cpu_list,
2414 stat_config.system_wide,
2415 &stat_config.metric_events);
2418 perf_stat__collect_metric_expr(evsel_list);
2419 perf_stat__init_shadow_stats();
2421 if (add_default_attributes())
2424 if (stat_config.cgroup_list) {
2425 if (nr_cgroups > 0) {
2426 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2427 parse_options_usage(stat_usage, stat_options, "G", 1);
2428 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2432 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2433 &stat_config.metric_events, true) < 0) {
2434 parse_options_usage(stat_usage, stat_options,
2435 "for-each-cgroup", 0);
2440 if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2441 pr_err("failed to use cpu list %s\n", target.cpu_list);
2445 target.hybrid = perf_pmu__has_hybrid();
2446 if (evlist__create_maps(evsel_list, &target) < 0) {
2447 if (target__has_task(&target)) {
2448 pr_err("Problems finding threads of monitor\n");
2449 parse_options_usage(stat_usage, stat_options, "p", 1);
2450 parse_options_usage(NULL, stat_options, "t", 1);
2451 } else if (target__has_cpu(&target)) {
2452 perror("failed to parse CPUs map");
2453 parse_options_usage(stat_usage, stat_options, "C", 1);
2454 parse_options_usage(NULL, stat_options, "a", 1);
2459 evlist__check_cpu_maps(evsel_list);
2462 * Initialize thread_map with comm names,
2463 * so we could print it out on output.
2465 if (stat_config.aggr_mode == AGGR_THREAD) {
2466 thread_map__read_comms(evsel_list->core.threads);
2469 if (stat_config.aggr_mode == AGGR_NODE)
2470 cpu__setup_cpunode_map();
2472 if (stat_config.times && interval)
2473 interval_count = true;
2474 else if (stat_config.times && !interval) {
2475 pr_err("interval-count option should be used together with "
2476 "interval-print.\n");
2477 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2478 parse_options_usage(stat_usage, stat_options, "I", 1);
2482 if (timeout && timeout < 100) {
2484 pr_err("timeout must be >= 10ms.\n");
2485 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2488 pr_warning("timeout < 100ms. "
2489 "The overhead percentage could be high in some cases. "
2490 "Please proceed with caution.\n");
2492 if (timeout && interval) {
2493 pr_err("timeout option is not supported with interval-print.\n");
2494 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2495 parse_options_usage(stat_usage, stat_options, "I", 1);
2499 if (evlist__alloc_stats(evsel_list, interval))
2502 if (perf_stat_init_aggr_mode())
2506 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2507 * while avoiding that older tools show confusing messages.
2509 * However for pipe sessions we need to keep it zero,
2510 * because script's perf_evsel__check_attr is triggered
2511 * by attr->sample_type != 0, and we can't run it on
2514 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2517 * We dont want to block the signals - that would cause
2518 * child tasks to inherit that and Ctrl-C would not work.
2519 * What we want is for Ctrl-C to work in the exec()-ed
2520 * task, but being ignored by perf stat itself:
2524 signal(SIGINT, skip_signal);
2525 signal(SIGCHLD, skip_signal);
2526 signal(SIGALRM, skip_signal);
2527 signal(SIGABRT, skip_signal);
2529 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2532 /* Enable ignoring missing threads when -p option is defined. */
2533 evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2535 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2536 if (stat_config.run_count != 1 && verbose > 0)
2537 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2541 evlist__reset_prev_raw_counts(evsel_list);
2543 status = run_perf_stat(argc, argv, run_idx);
2544 if (forever && status != -1 && !interval) {
2545 print_counters(NULL, argc, argv);
2546 perf_stat__reset_stats();
2550 if (!forever && status != -1 && (!interval || stat_config.summary))
2551 print_counters(NULL, argc, argv);
2553 evlist__finalize_ctlfd(evsel_list);
2557 * We synthesize the kernel mmap record just so that older tools
2558 * don't emit warnings about not being able to resolve symbols
2559 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2560 * a saner message about no samples being in the perf.data file.
2562 * This also serves to suppress a warning about f_header.data.size == 0
2563 * in header.c at the moment 'perf stat record' gets introduced, which
2564 * is not really needed once we start adding the stat specific PERF_RECORD_
2565 * records, but the need to suppress the kptr_restrict messages in older
2566 * tools remain -acme
2568 int fd = perf_data__fd(&perf_stat.data);
2570 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2571 process_synthesized_event,
2572 &perf_stat.session->machines.host);
2574 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2575 "older tools may produce warnings about this file\n.");
2579 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2580 pr_err("failed to write stat round event\n");
2583 if (!perf_stat.data.is_pipe) {
2584 perf_stat.session->header.data_size += perf_stat.bytes_written;
2585 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2588 evlist__close(evsel_list);
2589 perf_session__delete(perf_stat.session);
2592 perf_stat__exit_aggr_mode();
2593 evlist__free_stats(evsel_list);
2595 if (stat_config.iostat_run)
2596 iostat_release(evsel_list);
2598 zfree(&stat_config.walltime_run);
2599 zfree(&stat_config.user_requested_cpu_list);
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 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);