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);
531 if (stat_config.initial_delay < 0) {
532 pr_info(EVLIST_DISABLED_MSG);
536 if (stat_config.initial_delay > 0) {
537 pr_info(EVLIST_DISABLED_MSG);
538 usleep(stat_config.initial_delay * USEC_PER_MSEC);
542 * We need to enable counters only if:
543 * - we don't have tracee (attaching to task or cpu)
544 * - we have initial delay configured
546 if (!target__none(&target) || stat_config.initial_delay) {
547 if (!all_counters_use_bpf)
548 evlist__enable(evsel_list);
549 if (stat_config.initial_delay > 0)
550 pr_info(EVLIST_ENABLED_MSG);
555 static void disable_counters(void)
557 struct evsel *counter;
560 * If we don't have tracee (attaching to task or cpu), counters may
561 * still be running. To get accurate group ratios, we must stop groups
562 * from counting before reading their constituent counters.
564 if (!target__none(&target)) {
565 evlist__for_each_entry(evsel_list, counter)
566 bpf_counter__disable(counter);
567 if (!all_counters_use_bpf)
568 evlist__disable(evsel_list);
572 static volatile int workload_exec_errno;
575 * evlist__prepare_workload will send a SIGUSR1
576 * if the fork fails, since we asked by setting its
577 * want_signal to true.
579 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
580 void *ucontext __maybe_unused)
582 workload_exec_errno = info->si_value.sival_int;
585 static bool evsel__should_store_id(struct evsel *counter)
587 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
590 static bool is_target_alive(struct target *_target,
591 struct perf_thread_map *threads)
596 if (!target__has_task(_target))
599 for (i = 0; i < threads->nr; i++) {
602 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
603 threads->map[i].pid);
605 if (!stat(path, &st))
612 static void process_evlist(struct evlist *evlist, unsigned int interval)
614 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
616 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
618 case EVLIST_CTL_CMD_ENABLE:
620 case EVLIST_CTL_CMD_DISABLE:
624 case EVLIST_CTL_CMD_SNAPSHOT:
625 case EVLIST_CTL_CMD_ACK:
626 case EVLIST_CTL_CMD_UNSUPPORTED:
627 case EVLIST_CTL_CMD_EVLIST:
628 case EVLIST_CTL_CMD_STOP:
629 case EVLIST_CTL_CMD_PING:
636 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
639 int tts = *time_to_sleep;
640 struct timespec time_diff;
642 diff_timespec(&time_diff, time_stop, time_start);
644 tts -= time_diff.tv_sec * MSEC_PER_SEC +
645 time_diff.tv_nsec / NSEC_PER_MSEC;
650 *time_to_sleep = tts;
653 static int dispatch_events(bool forks, int timeout, int interval, int *times)
655 int child_exited = 0, status = 0;
656 int time_to_sleep, sleep_time;
657 struct timespec time_start, time_stop;
660 sleep_time = interval;
662 sleep_time = timeout;
666 time_to_sleep = sleep_time;
670 child_exited = waitpid(child_pid, &status, WNOHANG);
672 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
677 clock_gettime(CLOCK_MONOTONIC, &time_start);
678 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
679 if (timeout || handle_interval(interval, times))
681 time_to_sleep = sleep_time;
682 } else { /* fd revent */
683 process_evlist(evsel_list, interval);
684 clock_gettime(CLOCK_MONOTONIC, &time_stop);
685 compute_tts(&time_start, &time_stop, &time_to_sleep);
692 enum counter_recovery {
698 static enum counter_recovery stat_handle_error(struct evsel *counter)
702 * PPC returns ENXIO for HW counters until 2.6.37
703 * (behavior changed with commit b0a873e).
705 if (errno == EINVAL || errno == ENOSYS ||
706 errno == ENOENT || errno == EOPNOTSUPP ||
709 ui__warning("%s event is not supported by the kernel.\n",
710 evsel__name(counter));
711 counter->supported = false;
713 * errored is a sticky flag that means one of the counter's
714 * cpu event had a problem and needs to be reexamined.
716 counter->errored = true;
718 if ((evsel__leader(counter) != counter) ||
719 !(counter->core.leader->nr_members > 1))
721 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
723 ui__warning("%s\n", msg);
724 return COUNTER_RETRY;
725 } else if (target__has_per_thread(&target) &&
726 evsel_list->core.threads &&
727 evsel_list->core.threads->err_thread != -1) {
729 * For global --per-thread case, skip current
732 if (!thread_map__remove(evsel_list->core.threads,
733 evsel_list->core.threads->err_thread)) {
734 evsel_list->core.threads->err_thread = -1;
735 return COUNTER_RETRY;
739 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
740 ui__error("%s\n", msg);
743 kill(child_pid, SIGTERM);
744 return COUNTER_FATAL;
747 static int __run_perf_stat(int argc, const char **argv, int run_idx)
749 int interval = stat_config.interval;
750 int times = stat_config.times;
751 int timeout = stat_config.timeout;
753 unsigned long long t0, t1;
754 struct evsel *counter;
757 const bool forks = (argc > 0);
758 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
759 struct evlist_cpu_iterator evlist_cpu_itr;
760 struct affinity saved_affinity, *affinity = NULL;
762 bool second_pass = false;
765 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
766 perror("failed to prepare workload");
769 child_pid = evsel_list->workload.pid;
773 evlist__set_leader(evsel_list);
775 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
776 if (affinity__setup(&saved_affinity) < 0)
778 affinity = &saved_affinity;
781 evlist__for_each_entry(evsel_list, counter) {
782 counter->reset_group = false;
783 if (bpf_counter__load(counter, &target))
785 if (!evsel__is_bpf(counter))
786 all_counters_use_bpf = false;
789 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
790 counter = evlist_cpu_itr.evsel;
793 * bperf calls evsel__open_per_cpu() in bperf__load(), so
794 * no need to call it again here.
799 if (counter->reset_group || counter->errored)
801 if (evsel__is_bpf(counter))
804 if (create_perf_stat_counter(counter, &stat_config, &target,
805 evlist_cpu_itr.cpu_map_idx) < 0) {
808 * Weak group failed. We cannot just undo this here
809 * because earlier CPUs might be in group mode, and the kernel
810 * doesn't support mixing group and non group reads. Defer
812 * Don't close here because we're in the wrong affinity.
814 if ((errno == EINVAL || errno == EBADF) &&
815 evsel__leader(counter) != counter &&
816 counter->weak_group) {
817 evlist__reset_weak_group(evsel_list, counter, false);
818 assert(counter->reset_group);
823 switch (stat_handle_error(counter)) {
835 counter->supported = true;
840 * Now redo all the weak group after closing them,
841 * and also close errored counters.
844 /* First close errored or weak retry */
845 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
846 counter = evlist_cpu_itr.evsel;
848 if (!counter->reset_group && !counter->errored)
851 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
853 /* Now reopen weak */
854 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
855 counter = evlist_cpu_itr.evsel;
857 if (!counter->reset_group)
860 pr_debug2("reopening weak %s\n", evsel__name(counter));
861 if (create_perf_stat_counter(counter, &stat_config, &target,
862 evlist_cpu_itr.cpu_map_idx) < 0) {
864 switch (stat_handle_error(counter)) {
868 goto try_again_reset;
875 counter->supported = true;
878 affinity__cleanup(affinity);
880 evlist__for_each_entry(evsel_list, counter) {
881 if (!counter->supported) {
882 perf_evsel__free_fd(&counter->core);
886 l = strlen(counter->unit);
887 if (l > stat_config.unit_width)
888 stat_config.unit_width = l;
890 if (evsel__should_store_id(counter) &&
891 evsel__store_ids(counter, evsel_list))
895 if (evlist__apply_filters(evsel_list, &counter)) {
896 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
897 counter->filter, evsel__name(counter), errno,
898 str_error_r(errno, msg, sizeof(msg)));
903 int fd = perf_data__fd(&perf_stat.data);
906 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
908 err = perf_session__write_header(perf_stat.session, evsel_list,
915 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
916 process_synthesized_event, is_pipe);
921 err = enable_counters();
925 /* Exec the command, if any */
927 evlist__start_workload(evsel_list);
930 clock_gettime(CLOCK_MONOTONIC, &ref_time);
933 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
934 status = dispatch_events(forks, timeout, interval, ×);
935 if (child_pid != -1) {
937 kill(child_pid, SIGTERM);
938 wait4(child_pid, &status, 0, &stat_config.ru_data);
941 if (workload_exec_errno) {
942 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
943 pr_err("Workload failed: %s\n", emsg);
947 if (WIFSIGNALED(status))
948 psignal(WTERMSIG(status), argv[0]);
950 status = dispatch_events(forks, timeout, interval, ×);
957 if (stat_config.walltime_run_table)
958 stat_config.walltime_run[run_idx] = t1 - t0;
960 if (interval && stat_config.summary) {
961 stat_config.interval = 0;
962 stat_config.stop_read_counter = true;
963 init_stats(&walltime_nsecs_stats);
964 update_stats(&walltime_nsecs_stats, t1 - t0);
966 if (stat_config.aggr_mode == AGGR_GLOBAL)
967 evlist__save_aggr_prev_raw_counts(evsel_list);
969 evlist__copy_prev_raw_counts(evsel_list);
970 evlist__reset_prev_raw_counts(evsel_list);
971 perf_stat__reset_shadow_per_stat(&rt_stat);
973 update_stats(&walltime_nsecs_stats, t1 - t0);
974 update_rusage_stats(&ru_stats, &stat_config.ru_data);
978 * Closing a group leader splits the group, and as we only disable
979 * group leaders, results in remaining events becoming enabled. To
980 * avoid arbitrary skew, we must read all counters before closing any
983 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
986 * We need to keep evsel_list alive, because it's processed
987 * later the evsel_list will be closed after.
990 evlist__close(evsel_list);
992 return WEXITSTATUS(status);
995 static int run_perf_stat(int argc, const char **argv, int run_idx)
1000 ret = system(pre_cmd);
1008 ret = __run_perf_stat(argc, argv, run_idx);
1013 ret = system(post_cmd);
1021 static void print_counters(struct timespec *ts, int argc, const char **argv)
1023 /* Do not print anything if we record to the pipe. */
1024 if (STAT_RECORD && perf_stat.data.is_pipe)
1026 if (stat_config.quiet)
1029 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1032 static volatile int signr = -1;
1034 static void skip_signal(int signo)
1036 if ((child_pid == -1) || stat_config.interval)
1041 * render child_pid harmless
1042 * won't send SIGTERM to a random
1043 * process in case of race condition
1044 * and fast PID recycling
1049 static void sig_atexit(void)
1054 * avoid race condition with SIGCHLD handler
1055 * in skip_signal() which is modifying child_pid
1056 * goal is to avoid send SIGTERM to a random
1060 sigaddset(&set, SIGCHLD);
1061 sigprocmask(SIG_BLOCK, &set, &oset);
1063 if (child_pid != -1)
1064 kill(child_pid, SIGTERM);
1066 sigprocmask(SIG_SETMASK, &oset, NULL);
1071 signal(signr, SIG_DFL);
1072 kill(getpid(), signr);
1075 void perf_stat__set_big_num(int set)
1077 stat_config.big_num = (set != 0);
1080 void perf_stat__set_no_csv_summary(int set)
1082 stat_config.no_csv_summary = (set != 0);
1085 static int stat__set_big_num(const struct option *opt __maybe_unused,
1086 const char *s __maybe_unused, int unset)
1088 big_num_opt = unset ? 0 : 1;
1089 perf_stat__set_big_num(!unset);
1093 static int enable_metric_only(const struct option *opt __maybe_unused,
1094 const char *s __maybe_unused, int unset)
1096 force_metric_only = true;
1097 stat_config.metric_only = !unset;
1101 static int append_metric_groups(const struct option *opt __maybe_unused,
1103 int unset __maybe_unused)
1108 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1113 metrics = strdup(str);
1120 static int parse_control_option(const struct option *opt,
1122 int unset __maybe_unused)
1124 struct perf_stat_config *config = opt->value;
1126 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1129 static int parse_stat_cgroups(const struct option *opt,
1130 const char *str, int unset)
1132 if (stat_config.cgroup_list) {
1133 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1137 return parse_cgroups(opt, str, unset);
1140 static int parse_hybrid_type(const struct option *opt,
1142 int unset __maybe_unused)
1144 struct evlist *evlist = *(struct evlist **)opt->value;
1146 if (!list_empty(&evlist->core.entries)) {
1147 fprintf(stderr, "Must define cputype before events/metrics\n");
1151 evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1152 if (!evlist->hybrid_pmu_name) {
1153 fprintf(stderr, "--cputype %s is not supported!\n", str);
1160 static struct option stat_options[] = {
1161 OPT_BOOLEAN('T', "transaction", &transaction_run,
1162 "hardware transaction statistics"),
1163 OPT_CALLBACK('e', "event", &evsel_list, "event",
1164 "event selector. use 'perf list' to list available events",
1165 parse_events_option),
1166 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1167 "event filter", parse_filter),
1168 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1169 "child tasks do not inherit counters"),
1170 OPT_STRING('p', "pid", &target.pid, "pid",
1171 "stat events on existing process id"),
1172 OPT_STRING('t', "tid", &target.tid, "tid",
1173 "stat events on existing thread id"),
1174 #ifdef HAVE_BPF_SKEL
1175 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1176 "stat events on existing bpf program id"),
1177 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1178 "use bpf program to count events"),
1179 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1180 "path to perf_event_attr map"),
1182 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1183 "system-wide collection from all CPUs"),
1184 OPT_BOOLEAN('g', "group", &group,
1185 "put the counters into a counter group"),
1186 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1187 "Use --no-scale to disable counter scaling for multiplexing"),
1188 OPT_INCR('v', "verbose", &verbose,
1189 "be more verbose (show counter open errors, etc)"),
1190 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1191 "repeat command and print average + stddev (max: 100, forever: 0)"),
1192 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1193 "display details about each run (only with -r option)"),
1194 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1195 "null run - dont start any counters"),
1196 OPT_INCR('d', "detailed", &detailed_run,
1197 "detailed run - start a lot of events"),
1198 OPT_BOOLEAN('S', "sync", &sync_run,
1199 "call sync() before starting a run"),
1200 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1201 "print large numbers with thousands\' separators",
1203 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1204 "list of cpus to monitor in system-wide"),
1205 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1206 "disable CPU count aggregation", AGGR_NONE),
1207 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1208 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1209 "Merge identical named hybrid events"),
1210 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1211 "print counts with custom separator"),
1212 OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1213 "print counts in JSON format"),
1214 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1215 "monitor event in cgroup name only", parse_stat_cgroups),
1216 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1217 "expand events for each cgroup"),
1218 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1219 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1220 OPT_INTEGER(0, "log-fd", &output_fd,
1221 "log output to fd, instead of stderr"),
1222 OPT_STRING(0, "pre", &pre_cmd, "command",
1223 "command to run prior to the measured command"),
1224 OPT_STRING(0, "post", &post_cmd, "command",
1225 "command to run after to the measured command"),
1226 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1227 "print counts at regular interval in ms "
1228 "(overhead is possible for values <= 100ms)"),
1229 OPT_INTEGER(0, "interval-count", &stat_config.times,
1230 "print counts for fixed number of times"),
1231 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1232 "clear screen in between new interval"),
1233 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1234 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1235 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1236 "aggregate counts per processor socket", AGGR_SOCKET),
1237 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1238 "aggregate counts per processor die", AGGR_DIE),
1239 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1240 "aggregate counts per physical processor core", AGGR_CORE),
1241 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1242 "aggregate counts per thread", AGGR_THREAD),
1243 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1244 "aggregate counts per numa node", AGGR_NODE),
1245 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1246 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1247 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1248 "Only print computed metrics. No raw values", enable_metric_only),
1249 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1250 "don't group metric events, impacts multiplexing"),
1251 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1252 "don't try to share events between metrics in a group"),
1253 OPT_BOOLEAN(0, "topdown", &topdown_run,
1254 "measure top-down statistics"),
1255 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1256 "Set the metrics level for the top-down statistics (0: max level)"),
1257 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1258 "measure SMI cost"),
1259 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1260 "monitor specified metrics or metric groups (separated by ,)",
1261 append_metric_groups),
1262 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1263 "Configure all used events to run in kernel space.",
1264 PARSE_OPT_EXCLUSIVE),
1265 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1266 "Configure all used events to run in user space.",
1267 PARSE_OPT_EXCLUSIVE),
1268 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1269 "Use with 'percore' event qualifier to show the event "
1270 "counts of one hardware thread by sum up total hardware "
1271 "threads of same physical core"),
1272 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1273 "print summary for interval mode"),
1274 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1275 "don't print 'summary' for CSV summary output"),
1276 OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1277 "don't print output (useful with record)"),
1278 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1279 "Only enable events on applying cpu with this type "
1280 "for hybrid platform (e.g. core or atom)",
1283 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1284 "libpfm4 event selector. use 'perf list' to list available events",
1285 parse_libpfm_events_option),
1287 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1288 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1289 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1290 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1291 parse_control_option),
1292 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1293 "measure I/O performance metrics provided by arch/platform",
1298 static const char *const aggr_mode__string[] = {
1299 [AGGR_CORE] = "core",
1301 [AGGR_GLOBAL] = "global",
1302 [AGGR_NODE] = "node",
1303 [AGGR_NONE] = "none",
1304 [AGGR_SOCKET] = "socket",
1305 [AGGR_THREAD] = "thread",
1306 [AGGR_UNSET] = "unset",
1309 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1310 struct perf_cpu cpu)
1312 return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1315 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1316 struct perf_cpu cpu)
1318 return aggr_cpu_id__die(cpu, /*data=*/NULL);
1321 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1322 struct perf_cpu cpu)
1324 return aggr_cpu_id__core(cpu, /*data=*/NULL);
1327 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1328 struct perf_cpu cpu)
1330 return aggr_cpu_id__node(cpu, /*data=*/NULL);
1333 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1334 aggr_get_id_t get_id, struct perf_cpu cpu)
1336 struct aggr_cpu_id id = aggr_cpu_id__empty();
1338 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1339 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1341 id = config->cpus_aggr_map->map[cpu.cpu];
1345 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1346 struct perf_cpu cpu)
1348 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1351 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1352 struct perf_cpu cpu)
1354 return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1357 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1358 struct perf_cpu cpu)
1360 return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1363 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1364 struct perf_cpu cpu)
1366 return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1369 static bool term_percore_set(void)
1371 struct evsel *counter;
1373 evlist__for_each_entry(evsel_list, counter) {
1374 if (counter->percore)
1381 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1383 switch (aggr_mode) {
1385 return aggr_cpu_id__socket;
1387 return aggr_cpu_id__die;
1389 return aggr_cpu_id__core;
1391 return aggr_cpu_id__node;
1393 if (term_percore_set())
1394 return aggr_cpu_id__core;
1406 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1408 switch (aggr_mode) {
1410 return perf_stat__get_socket_cached;
1412 return perf_stat__get_die_cached;
1414 return perf_stat__get_core_cached;
1416 return perf_stat__get_node_cached;
1418 if (term_percore_set()) {
1419 return perf_stat__get_core_cached;
1431 static int perf_stat_init_aggr_mode(void)
1434 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1437 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1438 get_id, /*data=*/NULL);
1439 if (!stat_config.aggr_map) {
1440 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1443 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1447 * The evsel_list->cpus is the base we operate on,
1448 * taking the highest cpu number to be the size of
1449 * the aggregation translate cpumap.
1451 if (evsel_list->core.user_requested_cpus)
1452 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1455 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1456 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1459 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1462 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1463 "cpu_aggr_map refcnt unbalanced\n");
1468 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1470 if (map && refcount_dec_and_test(&map->refcnt))
1471 cpu_aggr_map__delete(map);
1474 static void perf_stat__exit_aggr_mode(void)
1476 cpu_aggr_map__put(stat_config.aggr_map);
1477 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1478 stat_config.aggr_map = NULL;
1479 stat_config.cpus_aggr_map = NULL;
1482 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1484 struct perf_env *env = data;
1485 struct aggr_cpu_id id = aggr_cpu_id__empty();
1488 id.socket = env->cpu[cpu.cpu].socket_id;
1493 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1495 struct perf_env *env = data;
1496 struct aggr_cpu_id id = aggr_cpu_id__empty();
1498 if (cpu.cpu != -1) {
1500 * die_id is relative to socket, so start
1501 * with the socket ID and then add die to
1504 id.socket = env->cpu[cpu.cpu].socket_id;
1505 id.die = env->cpu[cpu.cpu].die_id;
1511 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1513 struct perf_env *env = data;
1514 struct aggr_cpu_id id = aggr_cpu_id__empty();
1516 if (cpu.cpu != -1) {
1518 * core_id is relative to socket and die,
1519 * we need a global id. So we set
1520 * socket, die id and core id
1522 id.socket = env->cpu[cpu.cpu].socket_id;
1523 id.die = env->cpu[cpu.cpu].die_id;
1524 id.core = env->cpu[cpu.cpu].core_id;
1530 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1532 struct aggr_cpu_id id = aggr_cpu_id__empty();
1534 id.node = perf_env__numa_node(data, cpu);
1538 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1539 struct perf_cpu cpu)
1541 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1543 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1544 struct perf_cpu cpu)
1546 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1549 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1550 struct perf_cpu cpu)
1552 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1555 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1556 struct perf_cpu cpu)
1558 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1561 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1563 switch (aggr_mode) {
1565 return perf_env__get_socket_aggr_by_cpu;
1567 return perf_env__get_die_aggr_by_cpu;
1569 return perf_env__get_core_aggr_by_cpu;
1571 return perf_env__get_node_aggr_by_cpu;
1582 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1584 switch (aggr_mode) {
1586 return perf_stat__get_socket_file;
1588 return perf_stat__get_die_file;
1590 return perf_stat__get_core_file;
1592 return perf_stat__get_node_file;
1603 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1605 struct perf_env *env = &st->session->header.env;
1606 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1611 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1612 if (!stat_config.aggr_map) {
1613 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1616 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1621 * Add default attributes, if there were no attributes specified or
1622 * if -d/--detailed, -d -d or -d -d -d is used:
1624 static int add_default_attributes(void)
1627 struct perf_event_attr default_attrs0[] = {
1629 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1630 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1631 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1632 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1634 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1636 struct perf_event_attr frontend_attrs[] = {
1637 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1639 struct perf_event_attr backend_attrs[] = {
1640 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1642 struct perf_event_attr default_attrs1[] = {
1643 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1644 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1645 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1650 * Detailed stats (-d), covering the L1 and last level data caches:
1652 struct perf_event_attr detailed_attrs[] = {
1654 { .type = PERF_TYPE_HW_CACHE,
1656 PERF_COUNT_HW_CACHE_L1D << 0 |
1657 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1658 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1660 { .type = PERF_TYPE_HW_CACHE,
1662 PERF_COUNT_HW_CACHE_L1D << 0 |
1663 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1664 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1666 { .type = PERF_TYPE_HW_CACHE,
1668 PERF_COUNT_HW_CACHE_LL << 0 |
1669 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1670 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1672 { .type = PERF_TYPE_HW_CACHE,
1674 PERF_COUNT_HW_CACHE_LL << 0 |
1675 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1676 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1680 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1682 struct perf_event_attr very_detailed_attrs[] = {
1684 { .type = PERF_TYPE_HW_CACHE,
1686 PERF_COUNT_HW_CACHE_L1I << 0 |
1687 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1688 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1690 { .type = PERF_TYPE_HW_CACHE,
1692 PERF_COUNT_HW_CACHE_L1I << 0 |
1693 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1694 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1696 { .type = PERF_TYPE_HW_CACHE,
1698 PERF_COUNT_HW_CACHE_DTLB << 0 |
1699 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1700 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1702 { .type = PERF_TYPE_HW_CACHE,
1704 PERF_COUNT_HW_CACHE_DTLB << 0 |
1705 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1706 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1708 { .type = PERF_TYPE_HW_CACHE,
1710 PERF_COUNT_HW_CACHE_ITLB << 0 |
1711 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1712 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1714 { .type = PERF_TYPE_HW_CACHE,
1716 PERF_COUNT_HW_CACHE_ITLB << 0 |
1717 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1718 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1723 * Very, very detailed stats (-d -d -d), adding prefetch events:
1725 struct perf_event_attr very_very_detailed_attrs[] = {
1727 { .type = PERF_TYPE_HW_CACHE,
1729 PERF_COUNT_HW_CACHE_L1D << 0 |
1730 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1731 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1733 { .type = PERF_TYPE_HW_CACHE,
1735 PERF_COUNT_HW_CACHE_L1D << 0 |
1736 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1737 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1740 struct perf_event_attr default_null_attrs[] = {};
1742 /* Set attrs if no event is selected and !null_run: */
1743 if (stat_config.null_run)
1746 if (transaction_run) {
1747 struct parse_events_error errinfo;
1748 /* Handle -T as -M transaction. Once platform specific metrics
1749 * support has been added to the json files, all architectures
1750 * will use this approach. To determine transaction support
1751 * on an architecture test for such a metric name.
1753 if (metricgroup__has_metric("transaction")) {
1754 return metricgroup__parse_groups(evsel_list, "transaction",
1755 stat_config.metric_no_group,
1756 stat_config.metric_no_merge,
1757 stat_config.user_requested_cpu_list,
1758 stat_config.system_wide,
1759 &stat_config.metric_events);
1762 parse_events_error__init(&errinfo);
1763 if (pmu_have_event("cpu", "cycles-ct") &&
1764 pmu_have_event("cpu", "el-start"))
1765 err = parse_events(evsel_list, transaction_attrs,
1768 err = parse_events(evsel_list,
1769 transaction_limited_attrs,
1772 fprintf(stderr, "Cannot set up transaction events\n");
1773 parse_events_error__print(&errinfo, transaction_attrs);
1775 parse_events_error__exit(&errinfo);
1776 return err ? -1 : 0;
1780 struct parse_events_error errinfo;
1783 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1784 fprintf(stderr, "freeze_on_smi is not supported.\n");
1789 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1790 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1796 if (!pmu_have_event("msr", "aperf") ||
1797 !pmu_have_event("msr", "smi")) {
1798 fprintf(stderr, "To measure SMI cost, it needs "
1799 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1802 if (!force_metric_only)
1803 stat_config.metric_only = true;
1805 parse_events_error__init(&errinfo);
1806 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1808 parse_events_error__print(&errinfo, smi_cost_attrs);
1809 fprintf(stderr, "Cannot set up SMI cost events\n");
1811 parse_events_error__exit(&errinfo);
1812 return err ? -1 : 0;
1816 const char **metric_attrs = topdown_metric_attrs;
1817 unsigned int max_level = 1;
1820 const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1822 if (!force_metric_only)
1823 stat_config.metric_only = true;
1825 if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1826 metric_attrs = topdown_metric_L2_attrs;
1830 if (stat_config.topdown_level > max_level) {
1831 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1833 } else if (!stat_config.topdown_level)
1834 stat_config.topdown_level = max_level;
1836 if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1837 pr_err("Out of memory\n");
1841 if (metric_attrs[0] && str) {
1842 if (!stat_config.interval && !stat_config.metric_only) {
1843 fprintf(stat_config.output,
1844 "Topdown accuracy may decrease when measuring long periods.\n"
1845 "Please print the result regularly, e.g. -I1000\n");
1852 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1853 stat_config.aggr_mode != AGGR_CORE) {
1854 pr_err("top down event configuration requires --per-core mode\n");
1857 stat_config.aggr_mode = AGGR_CORE;
1858 if (nr_cgroups || !target__has_cpu(&target)) {
1859 pr_err("top down event configuration requires system-wide mode (-a)\n");
1863 if (topdown_filter_events(topdown_attrs, &str,
1864 arch_topdown_check_group(&warn),
1866 pr_err("Out of memory\n");
1870 if (topdown_attrs[0] && str) {
1871 struct parse_events_error errinfo;
1873 arch_topdown_group_warn();
1875 parse_events_error__init(&errinfo);
1876 err = parse_events(evsel_list, str, &errinfo);
1879 "Cannot set up top down events %s: %d\n",
1881 parse_events_error__print(&errinfo, str);
1882 parse_events_error__exit(&errinfo);
1886 parse_events_error__exit(&errinfo);
1888 fprintf(stderr, "System does not support topdown\n");
1894 if (!stat_config.topdown_level)
1895 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1897 if (!evsel_list->core.nr_entries) {
1898 if (target__has_cpu(&target))
1899 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1901 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1903 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1904 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1907 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1908 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1911 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1913 /* Platform specific attrs */
1914 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
1918 /* Detailed events get appended to the event list: */
1920 if (detailed_run < 1)
1923 /* Append detailed run extra attributes: */
1924 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1927 if (detailed_run < 2)
1930 /* Append very detailed run extra attributes: */
1931 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1934 if (detailed_run < 3)
1937 /* Append very, very detailed run extra attributes: */
1938 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1941 static const char * const stat_record_usage[] = {
1942 "perf stat record [<options>]",
1946 static void init_features(struct perf_session *session)
1950 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1951 perf_header__set_feat(&session->header, feat);
1953 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1954 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1955 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1956 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1957 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1960 static int __cmd_record(int argc, const char **argv)
1962 struct perf_session *session;
1963 struct perf_data *data = &perf_stat.data;
1965 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1966 PARSE_OPT_STOP_AT_NON_OPTION);
1969 data->path = output_name;
1971 if (stat_config.run_count != 1 || forever) {
1972 pr_err("Cannot use -r option with perf stat record.\n");
1976 session = perf_session__new(data, NULL);
1977 if (IS_ERR(session)) {
1978 pr_err("Perf session creation failed\n");
1979 return PTR_ERR(session);
1982 init_features(session);
1984 session->evlist = evsel_list;
1985 perf_stat.session = session;
1986 perf_stat.record = true;
1990 static int process_stat_round_event(struct perf_session *session,
1991 union perf_event *event)
1993 struct perf_record_stat_round *stat_round = &event->stat_round;
1994 struct evsel *counter;
1995 struct timespec tsh, *ts = NULL;
1996 const char **argv = session->header.env.cmdline_argv;
1997 int argc = session->header.env.nr_cmdline;
1999 evlist__for_each_entry(evsel_list, counter)
2000 perf_stat_process_counter(&stat_config, counter);
2002 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2003 update_stats(&walltime_nsecs_stats, stat_round->time);
2005 if (stat_config.interval && stat_round->time) {
2006 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2007 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2011 print_counters(ts, argc, argv);
2016 int process_stat_config_event(struct perf_session *session,
2017 union perf_event *event)
2019 struct perf_tool *tool = session->tool;
2020 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2022 perf_event__read_stat_config(&stat_config, &event->stat_config);
2024 if (perf_cpu_map__empty(st->cpus)) {
2025 if (st->aggr_mode != AGGR_UNSET)
2026 pr_warning("warning: processing task data, aggregation mode not set\n");
2030 if (st->aggr_mode != AGGR_UNSET)
2031 stat_config.aggr_mode = st->aggr_mode;
2033 if (perf_stat.data.is_pipe)
2034 perf_stat_init_aggr_mode();
2036 perf_stat_init_aggr_mode_file(st);
2041 static int set_maps(struct perf_stat *st)
2043 if (!st->cpus || !st->threads)
2046 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2049 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2051 if (evlist__alloc_stats(evsel_list, true))
2054 st->maps_allocated = true;
2059 int process_thread_map_event(struct perf_session *session,
2060 union perf_event *event)
2062 struct perf_tool *tool = session->tool;
2063 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2066 pr_warning("Extra thread map event, ignoring.\n");
2070 st->threads = thread_map__new_event(&event->thread_map);
2074 return set_maps(st);
2078 int process_cpu_map_event(struct perf_session *session,
2079 union perf_event *event)
2081 struct perf_tool *tool = session->tool;
2082 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2083 struct perf_cpu_map *cpus;
2086 pr_warning("Extra cpu map event, ignoring.\n");
2090 cpus = cpu_map__new_data(&event->cpu_map.data);
2095 return set_maps(st);
2098 static const char * const stat_report_usage[] = {
2099 "perf stat report [<options>]",
2103 static struct perf_stat perf_stat = {
2105 .attr = perf_event__process_attr,
2106 .event_update = perf_event__process_event_update,
2107 .thread_map = process_thread_map_event,
2108 .cpu_map = process_cpu_map_event,
2109 .stat_config = process_stat_config_event,
2110 .stat = perf_event__process_stat_event,
2111 .stat_round = process_stat_round_event,
2113 .aggr_mode = AGGR_UNSET,
2116 static int __cmd_report(int argc, const char **argv)
2118 struct perf_session *session;
2119 const struct option options[] = {
2120 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2121 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2122 "aggregate counts per processor socket", AGGR_SOCKET),
2123 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2124 "aggregate counts per processor die", AGGR_DIE),
2125 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2126 "aggregate counts per physical processor core", AGGR_CORE),
2127 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2128 "aggregate counts per numa node", AGGR_NODE),
2129 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2130 "disable CPU count aggregation", AGGR_NONE),
2136 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2138 if (!input_name || !strlen(input_name)) {
2139 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2142 input_name = "perf.data";
2145 perf_stat__init_shadow_stats();
2147 perf_stat.data.path = input_name;
2148 perf_stat.data.mode = PERF_DATA_MODE_READ;
2150 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2151 if (IS_ERR(session))
2152 return PTR_ERR(session);
2154 perf_stat.session = session;
2155 stat_config.output = stderr;
2156 evsel_list = session->evlist;
2158 ret = perf_session__process_events(session);
2162 perf_session__delete(session);
2166 static void setup_system_wide(int forks)
2169 * Make system wide (-a) the default target if
2170 * no target was specified and one of following
2171 * conditions is met:
2173 * - there's no workload specified
2174 * - there is workload specified but all requested
2175 * events are system wide events
2177 if (!target__none(&target))
2181 target.system_wide = true;
2183 struct evsel *counter;
2185 evlist__for_each_entry(evsel_list, counter) {
2186 if (!counter->core.requires_cpu &&
2187 strcmp(counter->name, "duration_time")) {
2192 if (evsel_list->core.nr_entries)
2193 target.system_wide = true;
2197 int cmd_stat(int argc, const char **argv)
2199 const char * const stat_usage[] = {
2200 "perf stat [<options>] [<command>]",
2203 int status = -EINVAL, run_idx, err;
2205 FILE *output = stderr;
2206 unsigned int interval, timeout;
2207 const char * const stat_subcommands[] = { "record", "report" };
2208 char errbuf[BUFSIZ];
2210 setlocale(LC_ALL, "");
2212 evsel_list = evlist__new();
2213 if (evsel_list == NULL)
2216 parse_events__shrink_config_terms();
2218 /* String-parsing callback-based options would segfault when negated */
2219 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2220 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2221 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2223 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2224 (const char **) stat_usage,
2225 PARSE_OPT_STOP_AT_NON_OPTION);
2227 if (stat_config.csv_sep) {
2228 stat_config.csv_output = true;
2229 if (!strcmp(stat_config.csv_sep, "\\t"))
2230 stat_config.csv_sep = "\t";
2232 stat_config.csv_sep = DEFAULT_SEPARATOR;
2234 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2235 argc = __cmd_record(argc, argv);
2238 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2239 return __cmd_report(argc, argv);
2241 interval = stat_config.interval;
2242 timeout = stat_config.timeout;
2245 * For record command the -o is already taken care of.
2247 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2250 if (output_name && output_fd) {
2251 fprintf(stderr, "cannot use both --output and --log-fd\n");
2252 parse_options_usage(stat_usage, stat_options, "o", 1);
2253 parse_options_usage(NULL, stat_options, "log-fd", 0);
2257 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2258 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2262 if (stat_config.metric_only && stat_config.run_count > 1) {
2263 fprintf(stderr, "--metric-only is not supported with -r\n");
2267 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2268 fprintf(stderr, "--table is only supported with -r\n");
2269 parse_options_usage(stat_usage, stat_options, "r", 1);
2270 parse_options_usage(NULL, stat_options, "table", 0);
2274 if (output_fd < 0) {
2275 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2276 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2280 if (!output && !stat_config.quiet) {
2282 mode = append_file ? "a" : "w";
2284 output = fopen(output_name, mode);
2286 perror("failed to create output file");
2289 clock_gettime(CLOCK_REALTIME, &tm);
2290 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2291 } else if (output_fd > 0) {
2292 mode = append_file ? "a" : "w";
2293 output = fdopen(output_fd, mode);
2295 perror("Failed opening logfd");
2300 stat_config.output = output;
2303 * let the spreadsheet do the pretty-printing
2305 if (stat_config.csv_output) {
2306 /* User explicitly passed -B? */
2307 if (big_num_opt == 1) {
2308 fprintf(stderr, "-B option not supported with -x\n");
2309 parse_options_usage(stat_usage, stat_options, "B", 1);
2310 parse_options_usage(NULL, stat_options, "x", 1);
2312 } else /* Nope, so disable big number formatting */
2313 stat_config.big_num = false;
2314 } else if (big_num_opt == 0) /* User passed --no-big-num */
2315 stat_config.big_num = false;
2317 err = target__validate(&target);
2319 target__strerror(&target, err, errbuf, BUFSIZ);
2320 pr_warning("%s\n", errbuf);
2323 setup_system_wide(argc);
2326 * Display user/system times only for single
2327 * run and when there's specified tracee.
2329 if ((stat_config.run_count == 1) && target__none(&target))
2330 stat_config.ru_display = true;
2332 if (stat_config.run_count < 0) {
2333 pr_err("Run count must be a positive number\n");
2334 parse_options_usage(stat_usage, stat_options, "r", 1);
2336 } else if (stat_config.run_count == 0) {
2338 stat_config.run_count = 1;
2341 if (stat_config.walltime_run_table) {
2342 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2343 if (!stat_config.walltime_run) {
2344 pr_err("failed to setup -r option");
2349 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2350 !target__has_task(&target)) {
2351 if (!target.system_wide || target.cpu_list) {
2352 fprintf(stderr, "The --per-thread option is only "
2353 "available when monitoring via -p -t -a "
2354 "options or only --per-thread.\n");
2355 parse_options_usage(NULL, stat_options, "p", 1);
2356 parse_options_usage(NULL, stat_options, "t", 1);
2362 * no_aggr, cgroup are for system-wide only
2363 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2365 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2366 stat_config.aggr_mode != AGGR_THREAD) ||
2367 (nr_cgroups || stat_config.cgroup_list)) &&
2368 !target__has_cpu(&target)) {
2369 fprintf(stderr, "both cgroup and no-aggregation "
2370 "modes only available in system-wide mode\n");
2372 parse_options_usage(stat_usage, stat_options, "G", 1);
2373 parse_options_usage(NULL, stat_options, "A", 1);
2374 parse_options_usage(NULL, stat_options, "a", 1);
2375 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2379 if (stat_config.iostat_run) {
2380 status = iostat_prepare(evsel_list, &stat_config);
2383 if (iostat_mode == IOSTAT_LIST) {
2384 iostat_list(evsel_list, &stat_config);
2387 iostat_list(evsel_list, &stat_config);
2388 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2389 target.system_wide = true;
2392 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2393 target.per_thread = true;
2395 stat_config.system_wide = target.system_wide;
2396 if (target.cpu_list) {
2397 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2398 if (!stat_config.user_requested_cpu_list) {
2405 * Metric parsing needs to be delayed as metrics may optimize events
2406 * knowing the target is system-wide.
2409 metricgroup__parse_groups(evsel_list, metrics,
2410 stat_config.metric_no_group,
2411 stat_config.metric_no_merge,
2412 stat_config.user_requested_cpu_list,
2413 stat_config.system_wide,
2414 &stat_config.metric_events);
2417 perf_stat__collect_metric_expr(evsel_list);
2418 perf_stat__init_shadow_stats();
2420 if (add_default_attributes())
2423 if (stat_config.cgroup_list) {
2424 if (nr_cgroups > 0) {
2425 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2426 parse_options_usage(stat_usage, stat_options, "G", 1);
2427 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2431 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2432 &stat_config.metric_events, true) < 0) {
2433 parse_options_usage(stat_usage, stat_options,
2434 "for-each-cgroup", 0);
2439 if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2440 pr_err("failed to use cpu list %s\n", target.cpu_list);
2444 target.hybrid = perf_pmu__has_hybrid();
2445 if (evlist__create_maps(evsel_list, &target) < 0) {
2446 if (target__has_task(&target)) {
2447 pr_err("Problems finding threads of monitor\n");
2448 parse_options_usage(stat_usage, stat_options, "p", 1);
2449 parse_options_usage(NULL, stat_options, "t", 1);
2450 } else if (target__has_cpu(&target)) {
2451 perror("failed to parse CPUs map");
2452 parse_options_usage(stat_usage, stat_options, "C", 1);
2453 parse_options_usage(NULL, stat_options, "a", 1);
2458 evlist__check_cpu_maps(evsel_list);
2461 * Initialize thread_map with comm names,
2462 * so we could print it out on output.
2464 if (stat_config.aggr_mode == AGGR_THREAD) {
2465 thread_map__read_comms(evsel_list->core.threads);
2468 if (stat_config.aggr_mode == AGGR_NODE)
2469 cpu__setup_cpunode_map();
2471 if (stat_config.times && interval)
2472 interval_count = true;
2473 else if (stat_config.times && !interval) {
2474 pr_err("interval-count option should be used together with "
2475 "interval-print.\n");
2476 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2477 parse_options_usage(stat_usage, stat_options, "I", 1);
2481 if (timeout && timeout < 100) {
2483 pr_err("timeout must be >= 10ms.\n");
2484 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2487 pr_warning("timeout < 100ms. "
2488 "The overhead percentage could be high in some cases. "
2489 "Please proceed with caution.\n");
2491 if (timeout && interval) {
2492 pr_err("timeout option is not supported with interval-print.\n");
2493 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2494 parse_options_usage(stat_usage, stat_options, "I", 1);
2498 if (evlist__alloc_stats(evsel_list, interval))
2501 if (perf_stat_init_aggr_mode())
2505 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2506 * while avoiding that older tools show confusing messages.
2508 * However for pipe sessions we need to keep it zero,
2509 * because script's perf_evsel__check_attr is triggered
2510 * by attr->sample_type != 0, and we can't run it on
2513 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2516 * We dont want to block the signals - that would cause
2517 * child tasks to inherit that and Ctrl-C would not work.
2518 * What we want is for Ctrl-C to work in the exec()-ed
2519 * task, but being ignored by perf stat itself:
2523 signal(SIGINT, skip_signal);
2524 signal(SIGCHLD, skip_signal);
2525 signal(SIGALRM, skip_signal);
2526 signal(SIGABRT, skip_signal);
2528 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2531 /* Enable ignoring missing threads when -p option is defined. */
2532 evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2534 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2535 if (stat_config.run_count != 1 && verbose > 0)
2536 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2540 evlist__reset_prev_raw_counts(evsel_list);
2542 status = run_perf_stat(argc, argv, run_idx);
2543 if (forever && status != -1 && !interval) {
2544 print_counters(NULL, argc, argv);
2545 perf_stat__reset_stats();
2549 if (!forever && status != -1 && (!interval || stat_config.summary))
2550 print_counters(NULL, argc, argv);
2552 evlist__finalize_ctlfd(evsel_list);
2556 * We synthesize the kernel mmap record just so that older tools
2557 * don't emit warnings about not being able to resolve symbols
2558 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2559 * a saner message about no samples being in the perf.data file.
2561 * This also serves to suppress a warning about f_header.data.size == 0
2562 * in header.c at the moment 'perf stat record' gets introduced, which
2563 * is not really needed once we start adding the stat specific PERF_RECORD_
2564 * records, but the need to suppress the kptr_restrict messages in older
2565 * tools remain -acme
2567 int fd = perf_data__fd(&perf_stat.data);
2569 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2570 process_synthesized_event,
2571 &perf_stat.session->machines.host);
2573 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2574 "older tools may produce warnings about this file\n.");
2578 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2579 pr_err("failed to write stat round event\n");
2582 if (!perf_stat.data.is_pipe) {
2583 perf_stat.session->header.data_size += perf_stat.bytes_written;
2584 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2587 evlist__close(evsel_list);
2588 perf_session__delete(perf_stat.session);
2591 perf_stat__exit_aggr_mode();
2592 evlist__free_stats(evsel_list);
2594 if (stat_config.iostat_run)
2595 iostat_release(evsel_list);
2597 zfree(&stat_config.walltime_run);
2598 zfree(&stat_config.user_requested_cpu_list);
2600 if (smi_cost && smi_reset)
2601 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2603 evlist__delete(evsel_list);
2605 metricgroup__rblist_exit(&stat_config.metric_events);
2606 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);