1 // SPDX-License-Identifier: GPL-2.0-only
5 * Builtin stat command: Give a precise performance counters summary
6 * overview about any workload, CPU or specific PID.
10 $ perf stat ./hackbench 10
14 Performance counter stats for './hackbench 10':
16 1708.761321 task-clock # 11.037 CPUs utilized
17 41,190 context-switches # 0.024 M/sec
18 6,735 CPU-migrations # 0.004 M/sec
19 17,318 page-faults # 0.010 M/sec
20 5,205,202,243 cycles # 3.046 GHz
21 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
22 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
23 2,603,501,247 instructions # 0.50 insns per cycle
24 # 1.48 stalled cycles per insn
25 484,357,498 branches # 283.455 M/sec
26 6,388,934 branch-misses # 1.32% of all branches
28 0.154822978 seconds time elapsed
31 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
33 * Improvements and fixes by:
35 * Arjan van de Ven <arjan@linux.intel.com>
36 * Yanmin Zhang <yanmin.zhang@intel.com>
37 * Wu Fengguang <fengguang.wu@intel.com>
38 * Mike Galbraith <efault@gmx.de>
39 * Paul Mackerras <paulus@samba.org>
40 * Jaswinder Singh Rajput <jaswinder@kernel.org>
45 #include "util/cgroup.h"
46 #include <subcmd/parse-options.h>
47 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evlist-hybrid.h"
52 #include "util/evsel.h"
53 #include "util/debug.h"
54 #include "util/color.h"
55 #include "util/stat.h"
56 #include "util/header.h"
57 #include "util/cpumap.h"
58 #include "util/thread_map.h"
59 #include "util/counts.h"
60 #include "util/topdown.h"
61 #include "util/session.h"
62 #include "util/tool.h"
63 #include "util/string2.h"
64 #include "util/metricgroup.h"
65 #include "util/synthetic-events.h"
66 #include "util/target.h"
67 #include "util/time-utils.h"
69 #include "util/affinity.h"
71 #include "util/bpf_counter.h"
72 #include "util/iostat.h"
73 #include "util/pmu-hybrid.h"
76 #include <linux/time64.h>
77 #include <linux/zalloc.h>
78 #include <api/fs/fs.h>
82 #include <sys/prctl.h>
86 #include <sys/types.h>
91 #include <sys/resource.h>
92 #include <linux/err.h>
94 #include <linux/ctype.h>
95 #include <perf/evlist.h>
97 #define DEFAULT_SEPARATOR " "
98 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
100 static void print_counters(struct timespec *ts, int argc, const char **argv);
102 /* Default events used for perf stat -T */
103 static const char *transaction_attrs = {
115 /* More limited version when the CPU does not have all events. */
116 static const char * transaction_limited_attrs = {
126 static const char * topdown_attrs[] = {
127 "topdown-total-slots",
128 "topdown-slots-retired",
129 "topdown-recovery-bubbles",
130 "topdown-fetch-bubbles",
131 "topdown-slots-issued",
135 static const char *topdown_metric_attrs[] = {
144 static const char *topdown_metric_L2_attrs[] = {
151 "topdown-br-mispredict",
157 #define TOPDOWN_MAX_LEVEL 2
159 static const char *smi_cost_attrs = {
167 static struct evlist *evsel_list;
168 static bool all_counters_use_bpf = true;
170 static struct target target = {
174 #define METRIC_ONLY_LEN 20
176 static volatile pid_t child_pid = -1;
177 static int detailed_run = 0;
178 static bool transaction_run;
179 static bool topdown_run = false;
180 static bool smi_cost = false;
181 static bool smi_reset = false;
182 static int big_num_opt = -1;
183 static bool group = false;
184 static const char *pre_cmd = NULL;
185 static const char *post_cmd = NULL;
186 static bool sync_run = false;
187 static bool forever = false;
188 static bool force_metric_only = false;
189 static struct timespec ref_time;
190 static bool append_file;
191 static bool interval_count;
192 static const char *output_name;
193 static int output_fd;
197 struct perf_data data;
198 struct perf_session *session;
200 struct perf_tool tool;
202 struct perf_cpu_map *cpus;
203 struct perf_thread_map *threads;
204 enum aggr_mode aggr_mode;
207 static struct perf_stat perf_stat;
208 #define STAT_RECORD perf_stat.record
210 static volatile int done = 0;
212 static struct perf_stat_config stat_config = {
213 .aggr_mode = AGGR_GLOBAL,
215 .unit_width = 4, /* strlen("unit") */
217 .metric_only_len = METRIC_ONLY_LEN,
218 .walltime_nsecs_stats = &walltime_nsecs_stats,
225 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
227 if (!a->core.cpus && !b->core.cpus)
230 if (!a->core.cpus || !b->core.cpus)
233 if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
236 for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
237 if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
238 perf_cpu_map__cpu(b->core.cpus, i).cpu)
245 static void evlist__check_cpu_maps(struct evlist *evlist)
247 struct evsel *evsel, *pos, *leader;
250 if (evlist__has_hybrid(evlist))
251 evlist__warn_hybrid_group(evlist);
253 evlist__for_each_entry(evlist, evsel) {
254 leader = evsel__leader(evsel);
256 /* Check that leader matches cpus with each member. */
259 if (cpus_map_matched(leader, evsel))
262 /* If there's mismatch disable the group and warn user. */
263 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
264 evsel__group_desc(leader, buf, sizeof(buf));
265 pr_warning(" %s\n", buf);
268 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
269 pr_warning(" %s: %s\n", leader->name, buf);
270 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
271 pr_warning(" %s: %s\n", evsel->name, buf);
274 for_each_group_evsel(pos, leader) {
275 evsel__set_leader(pos, pos);
276 pos->core.nr_members = 0;
278 evsel->core.leader->nr_members = 0;
282 static inline void diff_timespec(struct timespec *r, struct timespec *a,
285 r->tv_sec = a->tv_sec - b->tv_sec;
286 if (a->tv_nsec < b->tv_nsec) {
287 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
290 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
294 static void perf_stat__reset_stats(void)
298 evlist__reset_stats(evsel_list);
299 perf_stat__reset_shadow_stats();
301 for (i = 0; i < stat_config.stats_num; i++)
302 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
305 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
306 union perf_event *event,
307 struct perf_sample *sample __maybe_unused,
308 struct machine *machine __maybe_unused)
310 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
311 pr_err("failed to write perf data, error: %m\n");
315 perf_stat.bytes_written += event->header.size;
319 static int write_stat_round_event(u64 tm, u64 type)
321 return perf_event__synthesize_stat_round(NULL, tm, type,
322 process_synthesized_event,
326 #define WRITE_STAT_ROUND_EVENT(time, interval) \
327 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
329 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
331 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
332 struct perf_counts_values *count)
334 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
335 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
337 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
338 process_synthesized_event, NULL);
341 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
342 int thread, struct timespec *rs)
344 if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
345 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
346 struct perf_counts_values *count =
347 perf_counts(counter->counts, cpu_map_idx, thread);
348 count->ena = count->run = val;
352 return evsel__read_counter(counter, cpu_map_idx, thread);
356 * Read out the results of a single counter:
357 * do not aggregate counts across CPUs in system-wide mode
359 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
361 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
364 if (!counter->supported)
367 if (counter->core.system_wide)
370 for (thread = 0; thread < nthreads; thread++) {
371 struct perf_counts_values *count;
373 count = perf_counts(counter->counts, cpu_map_idx, thread);
376 * The leader's group read loads data into its group members
377 * (via evsel__read_counter()) and sets their count->loaded.
379 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
380 read_single_counter(counter, cpu_map_idx, thread, rs)) {
381 counter->counts->scaled = -1;
382 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
383 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
387 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
390 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
391 pr_err("failed to write stat event\n");
397 fprintf(stat_config.output,
398 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
399 evsel__name(counter),
400 perf_cpu_map__cpu(evsel__cpus(counter),
402 count->val, count->ena, count->run);
409 static int read_affinity_counters(struct timespec *rs)
411 struct evlist_cpu_iterator evlist_cpu_itr;
412 struct affinity saved_affinity, *affinity;
414 if (all_counters_use_bpf)
417 if (!target__has_cpu(&target) || target__has_per_thread(&target))
419 else if (affinity__setup(&saved_affinity) < 0)
422 affinity = &saved_affinity;
424 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
425 struct evsel *counter = evlist_cpu_itr.evsel;
427 if (evsel__is_bpf(counter))
431 counter->err = read_counter_cpu(counter, rs,
432 evlist_cpu_itr.cpu_map_idx);
436 affinity__cleanup(&saved_affinity);
441 static int read_bpf_map_counters(void)
443 struct evsel *counter;
446 evlist__for_each_entry(evsel_list, counter) {
447 if (!evsel__is_bpf(counter))
450 err = bpf_counter__read(counter);
457 static void read_counters(struct timespec *rs)
459 struct evsel *counter;
461 if (!stat_config.stop_read_counter) {
462 if (read_bpf_map_counters() ||
463 read_affinity_counters(rs))
467 evlist__for_each_entry(evsel_list, counter) {
469 pr_debug("failed to read counter %s\n", counter->name);
470 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
471 pr_warning("failed to process counter %s\n", counter->name);
476 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
480 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
484 config->stats_num = nthreads;
486 for (i = 0; i < nthreads; i++)
487 runtime_stat__init(&config->stats[i]);
492 static void runtime_stat_delete(struct perf_stat_config *config)
499 for (i = 0; i < config->stats_num; i++)
500 runtime_stat__exit(&config->stats[i]);
502 zfree(&config->stats);
505 static void runtime_stat_reset(struct perf_stat_config *config)
512 for (i = 0; i < config->stats_num; i++)
513 perf_stat__reset_shadow_per_stat(&config->stats[i]);
516 static void process_interval(void)
518 struct timespec ts, rs;
520 clock_gettime(CLOCK_MONOTONIC, &ts);
521 diff_timespec(&rs, &ts, &ref_time);
523 perf_stat__reset_shadow_per_stat(&rt_stat);
524 runtime_stat_reset(&stat_config);
528 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
529 pr_err("failed to write stat round event\n");
532 init_stats(&walltime_nsecs_stats);
533 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
534 print_counters(&rs, 0, NULL);
537 static bool handle_interval(unsigned int interval, int *times)
541 if (interval_count && !(--(*times)))
547 static int enable_counters(void)
552 evlist__for_each_entry(evsel_list, evsel) {
553 if (!evsel__is_bpf(evsel))
556 err = bpf_counter__enable(evsel);
561 if (stat_config.initial_delay < 0) {
562 pr_info(EVLIST_DISABLED_MSG);
566 if (stat_config.initial_delay > 0) {
567 pr_info(EVLIST_DISABLED_MSG);
568 usleep(stat_config.initial_delay * USEC_PER_MSEC);
572 * We need to enable counters only if:
573 * - we don't have tracee (attaching to task or cpu)
574 * - we have initial delay configured
576 if (!target__none(&target) || stat_config.initial_delay) {
577 if (!all_counters_use_bpf)
578 evlist__enable(evsel_list);
579 if (stat_config.initial_delay > 0)
580 pr_info(EVLIST_ENABLED_MSG);
585 static void disable_counters(void)
587 struct evsel *counter;
590 * If we don't have tracee (attaching to task or cpu), counters may
591 * still be running. To get accurate group ratios, we must stop groups
592 * from counting before reading their constituent counters.
594 if (!target__none(&target)) {
595 evlist__for_each_entry(evsel_list, counter)
596 bpf_counter__disable(counter);
597 if (!all_counters_use_bpf)
598 evlist__disable(evsel_list);
602 static volatile int workload_exec_errno;
605 * evlist__prepare_workload will send a SIGUSR1
606 * if the fork fails, since we asked by setting its
607 * want_signal to true.
609 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
610 void *ucontext __maybe_unused)
612 workload_exec_errno = info->si_value.sival_int;
615 static bool evsel__should_store_id(struct evsel *counter)
617 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
620 static bool is_target_alive(struct target *_target,
621 struct perf_thread_map *threads)
626 if (!target__has_task(_target))
629 for (i = 0; i < threads->nr; i++) {
632 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
633 threads->map[i].pid);
635 if (!stat(path, &st))
642 static void process_evlist(struct evlist *evlist, unsigned int interval)
644 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
646 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
648 case EVLIST_CTL_CMD_ENABLE:
652 case EVLIST_CTL_CMD_DISABLE:
656 case EVLIST_CTL_CMD_SNAPSHOT:
657 case EVLIST_CTL_CMD_ACK:
658 case EVLIST_CTL_CMD_UNSUPPORTED:
659 case EVLIST_CTL_CMD_EVLIST:
660 case EVLIST_CTL_CMD_STOP:
661 case EVLIST_CTL_CMD_PING:
668 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
671 int tts = *time_to_sleep;
672 struct timespec time_diff;
674 diff_timespec(&time_diff, time_stop, time_start);
676 tts -= time_diff.tv_sec * MSEC_PER_SEC +
677 time_diff.tv_nsec / NSEC_PER_MSEC;
682 *time_to_sleep = tts;
685 static int dispatch_events(bool forks, int timeout, int interval, int *times)
687 int child_exited = 0, status = 0;
688 int time_to_sleep, sleep_time;
689 struct timespec time_start, time_stop;
692 sleep_time = interval;
694 sleep_time = timeout;
698 time_to_sleep = sleep_time;
702 child_exited = waitpid(child_pid, &status, WNOHANG);
704 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
709 clock_gettime(CLOCK_MONOTONIC, &time_start);
710 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
711 if (timeout || handle_interval(interval, times))
713 time_to_sleep = sleep_time;
714 } else { /* fd revent */
715 process_evlist(evsel_list, interval);
716 clock_gettime(CLOCK_MONOTONIC, &time_stop);
717 compute_tts(&time_start, &time_stop, &time_to_sleep);
724 enum counter_recovery {
730 static enum counter_recovery stat_handle_error(struct evsel *counter)
734 * PPC returns ENXIO for HW counters until 2.6.37
735 * (behavior changed with commit b0a873e).
737 if (errno == EINVAL || errno == ENOSYS ||
738 errno == ENOENT || errno == EOPNOTSUPP ||
741 ui__warning("%s event is not supported by the kernel.\n",
742 evsel__name(counter));
743 counter->supported = false;
745 * errored is a sticky flag that means one of the counter's
746 * cpu event had a problem and needs to be reexamined.
748 counter->errored = true;
750 if ((evsel__leader(counter) != counter) ||
751 !(counter->core.leader->nr_members > 1))
753 } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
755 ui__warning("%s\n", msg);
756 return COUNTER_RETRY;
757 } else if (target__has_per_thread(&target) &&
758 evsel_list->core.threads &&
759 evsel_list->core.threads->err_thread != -1) {
761 * For global --per-thread case, skip current
764 if (!thread_map__remove(evsel_list->core.threads,
765 evsel_list->core.threads->err_thread)) {
766 evsel_list->core.threads->err_thread = -1;
767 return COUNTER_RETRY;
771 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
772 ui__error("%s\n", msg);
775 kill(child_pid, SIGTERM);
776 return COUNTER_FATAL;
779 static int __run_perf_stat(int argc, const char **argv, int run_idx)
781 int interval = stat_config.interval;
782 int times = stat_config.times;
783 int timeout = stat_config.timeout;
785 unsigned long long t0, t1;
786 struct evsel *counter;
789 const bool forks = (argc > 0);
790 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
791 struct evlist_cpu_iterator evlist_cpu_itr;
792 struct affinity saved_affinity, *affinity = NULL;
794 bool second_pass = false;
797 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
798 perror("failed to prepare workload");
801 child_pid = evsel_list->workload.pid;
805 evlist__set_leader(evsel_list);
807 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
808 if (affinity__setup(&saved_affinity) < 0)
810 affinity = &saved_affinity;
813 evlist__for_each_entry(evsel_list, counter) {
814 if (bpf_counter__load(counter, &target))
816 if (!evsel__is_bpf(counter))
817 all_counters_use_bpf = false;
820 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
821 counter = evlist_cpu_itr.evsel;
824 * bperf calls evsel__open_per_cpu() in bperf__load(), so
825 * no need to call it again here.
830 if (counter->reset_group || counter->errored)
832 if (evsel__is_bpf(counter))
835 if (create_perf_stat_counter(counter, &stat_config, &target,
836 evlist_cpu_itr.cpu_map_idx) < 0) {
839 * Weak group failed. We cannot just undo this here
840 * because earlier CPUs might be in group mode, and the kernel
841 * doesn't support mixing group and non group reads. Defer
843 * Don't close here because we're in the wrong affinity.
845 if ((errno == EINVAL || errno == EBADF) &&
846 evsel__leader(counter) != counter &&
847 counter->weak_group) {
848 evlist__reset_weak_group(evsel_list, counter, false);
849 assert(counter->reset_group);
854 switch (stat_handle_error(counter)) {
866 counter->supported = true;
871 * Now redo all the weak group after closing them,
872 * and also close errored counters.
875 /* First close errored or weak retry */
876 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
877 counter = evlist_cpu_itr.evsel;
879 if (!counter->reset_group && !counter->errored)
882 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
884 /* Now reopen weak */
885 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
886 counter = evlist_cpu_itr.evsel;
888 if (!counter->reset_group && !counter->errored)
890 if (!counter->reset_group)
893 pr_debug2("reopening weak %s\n", evsel__name(counter));
894 if (create_perf_stat_counter(counter, &stat_config, &target,
895 evlist_cpu_itr.cpu_map_idx) < 0) {
897 switch (stat_handle_error(counter)) {
901 goto try_again_reset;
908 counter->supported = true;
911 affinity__cleanup(affinity);
913 evlist__for_each_entry(evsel_list, counter) {
914 if (!counter->supported) {
915 perf_evsel__free_fd(&counter->core);
919 l = strlen(counter->unit);
920 if (l > stat_config.unit_width)
921 stat_config.unit_width = l;
923 if (evsel__should_store_id(counter) &&
924 evsel__store_ids(counter, evsel_list))
928 if (evlist__apply_filters(evsel_list, &counter)) {
929 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
930 counter->filter, evsel__name(counter), errno,
931 str_error_r(errno, msg, sizeof(msg)));
936 int fd = perf_data__fd(&perf_stat.data);
939 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
941 err = perf_session__write_header(perf_stat.session, evsel_list,
948 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
949 process_synthesized_event, is_pipe);
955 * Enable counters and exec the command:
958 err = enable_counters();
961 evlist__start_workload(evsel_list);
964 clock_gettime(CLOCK_MONOTONIC, &ref_time);
966 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
967 status = dispatch_events(forks, timeout, interval, ×);
968 if (child_pid != -1) {
970 kill(child_pid, SIGTERM);
971 wait4(child_pid, &status, 0, &stat_config.ru_data);
974 if (workload_exec_errno) {
975 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
976 pr_err("Workload failed: %s\n", emsg);
980 if (WIFSIGNALED(status))
981 psignal(WTERMSIG(status), argv[0]);
983 err = enable_counters();
988 clock_gettime(CLOCK_MONOTONIC, &ref_time);
990 status = dispatch_events(forks, timeout, interval, ×);
997 if (stat_config.walltime_run_table)
998 stat_config.walltime_run[run_idx] = t1 - t0;
1000 if (interval && stat_config.summary) {
1001 stat_config.interval = 0;
1002 stat_config.stop_read_counter = true;
1003 init_stats(&walltime_nsecs_stats);
1004 update_stats(&walltime_nsecs_stats, t1 - t0);
1006 if (stat_config.aggr_mode == AGGR_GLOBAL)
1007 evlist__save_aggr_prev_raw_counts(evsel_list);
1009 evlist__copy_prev_raw_counts(evsel_list);
1010 evlist__reset_prev_raw_counts(evsel_list);
1011 runtime_stat_reset(&stat_config);
1012 perf_stat__reset_shadow_per_stat(&rt_stat);
1014 update_stats(&walltime_nsecs_stats, t1 - t0);
1017 * Closing a group leader splits the group, and as we only disable
1018 * group leaders, results in remaining events becoming enabled. To
1019 * avoid arbitrary skew, we must read all counters before closing any
1022 read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1025 * We need to keep evsel_list alive, because it's processed
1026 * later the evsel_list will be closed after.
1029 evlist__close(evsel_list);
1031 return WEXITSTATUS(status);
1034 static int run_perf_stat(int argc, const char **argv, int run_idx)
1039 ret = system(pre_cmd);
1047 ret = __run_perf_stat(argc, argv, run_idx);
1052 ret = system(post_cmd);
1060 static void print_counters(struct timespec *ts, int argc, const char **argv)
1062 /* Do not print anything if we record to the pipe. */
1063 if (STAT_RECORD && perf_stat.data.is_pipe)
1065 if (stat_config.quiet)
1068 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1071 static volatile int signr = -1;
1073 static void skip_signal(int signo)
1075 if ((child_pid == -1) || stat_config.interval)
1080 * render child_pid harmless
1081 * won't send SIGTERM to a random
1082 * process in case of race condition
1083 * and fast PID recycling
1088 static void sig_atexit(void)
1093 * avoid race condition with SIGCHLD handler
1094 * in skip_signal() which is modifying child_pid
1095 * goal is to avoid send SIGTERM to a random
1099 sigaddset(&set, SIGCHLD);
1100 sigprocmask(SIG_BLOCK, &set, &oset);
1102 if (child_pid != -1)
1103 kill(child_pid, SIGTERM);
1105 sigprocmask(SIG_SETMASK, &oset, NULL);
1110 signal(signr, SIG_DFL);
1111 kill(getpid(), signr);
1114 void perf_stat__set_big_num(int set)
1116 stat_config.big_num = (set != 0);
1119 void perf_stat__set_no_csv_summary(int set)
1121 stat_config.no_csv_summary = (set != 0);
1124 static int stat__set_big_num(const struct option *opt __maybe_unused,
1125 const char *s __maybe_unused, int unset)
1127 big_num_opt = unset ? 0 : 1;
1128 perf_stat__set_big_num(!unset);
1132 static int enable_metric_only(const struct option *opt __maybe_unused,
1133 const char *s __maybe_unused, int unset)
1135 force_metric_only = true;
1136 stat_config.metric_only = !unset;
1140 static int parse_metric_groups(const struct option *opt,
1142 int unset __maybe_unused)
1144 return metricgroup__parse_groups(opt, str,
1145 stat_config.metric_no_group,
1146 stat_config.metric_no_merge,
1147 &stat_config.metric_events);
1150 static int parse_control_option(const struct option *opt,
1152 int unset __maybe_unused)
1154 struct perf_stat_config *config = opt->value;
1156 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1159 static int parse_stat_cgroups(const struct option *opt,
1160 const char *str, int unset)
1162 if (stat_config.cgroup_list) {
1163 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1167 return parse_cgroups(opt, str, unset);
1170 static int parse_hybrid_type(const struct option *opt,
1172 int unset __maybe_unused)
1174 struct evlist *evlist = *(struct evlist **)opt->value;
1176 if (!list_empty(&evlist->core.entries)) {
1177 fprintf(stderr, "Must define cputype before events/metrics\n");
1181 evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1182 if (!evlist->hybrid_pmu_name) {
1183 fprintf(stderr, "--cputype %s is not supported!\n", str);
1190 static struct option stat_options[] = {
1191 OPT_BOOLEAN('T', "transaction", &transaction_run,
1192 "hardware transaction statistics"),
1193 OPT_CALLBACK('e', "event", &evsel_list, "event",
1194 "event selector. use 'perf list' to list available events",
1195 parse_events_option),
1196 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1197 "event filter", parse_filter),
1198 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1199 "child tasks do not inherit counters"),
1200 OPT_STRING('p', "pid", &target.pid, "pid",
1201 "stat events on existing process id"),
1202 OPT_STRING('t', "tid", &target.tid, "tid",
1203 "stat events on existing thread id"),
1204 #ifdef HAVE_BPF_SKEL
1205 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1206 "stat events on existing bpf program id"),
1207 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1208 "use bpf program to count events"),
1209 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1210 "path to perf_event_attr map"),
1212 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1213 "system-wide collection from all CPUs"),
1214 OPT_BOOLEAN('g', "group", &group,
1215 "put the counters into a counter group"),
1216 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1217 "Use --no-scale to disable counter scaling for multiplexing"),
1218 OPT_INCR('v', "verbose", &verbose,
1219 "be more verbose (show counter open errors, etc)"),
1220 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1221 "repeat command and print average + stddev (max: 100, forever: 0)"),
1222 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1223 "display details about each run (only with -r option)"),
1224 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1225 "null run - dont start any counters"),
1226 OPT_INCR('d', "detailed", &detailed_run,
1227 "detailed run - start a lot of events"),
1228 OPT_BOOLEAN('S', "sync", &sync_run,
1229 "call sync() before starting a run"),
1230 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1231 "print large numbers with thousands\' separators",
1233 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1234 "list of cpus to monitor in system-wide"),
1235 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1236 "disable CPU count aggregation", AGGR_NONE),
1237 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1238 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1239 "print counts with custom separator"),
1240 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1241 "monitor event in cgroup name only", parse_stat_cgroups),
1242 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1243 "expand events for each cgroup"),
1244 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1245 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1246 OPT_INTEGER(0, "log-fd", &output_fd,
1247 "log output to fd, instead of stderr"),
1248 OPT_STRING(0, "pre", &pre_cmd, "command",
1249 "command to run prior to the measured command"),
1250 OPT_STRING(0, "post", &post_cmd, "command",
1251 "command to run after to the measured command"),
1252 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1253 "print counts at regular interval in ms "
1254 "(overhead is possible for values <= 100ms)"),
1255 OPT_INTEGER(0, "interval-count", &stat_config.times,
1256 "print counts for fixed number of times"),
1257 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1258 "clear screen in between new interval"),
1259 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1260 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1261 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1262 "aggregate counts per processor socket", AGGR_SOCKET),
1263 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1264 "aggregate counts per processor die", AGGR_DIE),
1265 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1266 "aggregate counts per physical processor core", AGGR_CORE),
1267 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1268 "aggregate counts per thread", AGGR_THREAD),
1269 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1270 "aggregate counts per numa node", AGGR_NODE),
1271 OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1272 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1273 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1274 "Only print computed metrics. No raw values", enable_metric_only),
1275 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1276 "don't group metric events, impacts multiplexing"),
1277 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1278 "don't try to share events between metrics in a group"),
1279 OPT_BOOLEAN(0, "topdown", &topdown_run,
1280 "measure top-down statistics"),
1281 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1282 "Set the metrics level for the top-down statistics (0: max level)"),
1283 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1284 "measure SMI cost"),
1285 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1286 "monitor specified metrics or metric groups (separated by ,)",
1287 parse_metric_groups),
1288 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1289 "Configure all used events to run in kernel space.",
1290 PARSE_OPT_EXCLUSIVE),
1291 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1292 "Configure all used events to run in user space.",
1293 PARSE_OPT_EXCLUSIVE),
1294 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1295 "Use with 'percore' event qualifier to show the event "
1296 "counts of one hardware thread by sum up total hardware "
1297 "threads of same physical core"),
1298 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1299 "print summary for interval mode"),
1300 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1301 "don't print 'summary' for CSV summary output"),
1302 OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1303 "don't print output (useful with record)"),
1304 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1305 "Only enable events on applying cpu with this type "
1306 "for hybrid platform (e.g. core or atom)",
1309 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1310 "libpfm4 event selector. use 'perf list' to list available events",
1311 parse_libpfm_events_option),
1313 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1314 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1315 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1316 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1317 parse_control_option),
1318 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1319 "measure I/O performance metrics provided by arch/platform",
1324 static const char *const aggr_mode__string[] = {
1325 [AGGR_CORE] = "core",
1327 [AGGR_GLOBAL] = "global",
1328 [AGGR_NODE] = "node",
1329 [AGGR_NONE] = "none",
1330 [AGGR_SOCKET] = "socket",
1331 [AGGR_THREAD] = "thread",
1332 [AGGR_UNSET] = "unset",
1335 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1336 struct perf_cpu cpu)
1338 return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1341 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1342 struct perf_cpu cpu)
1344 return aggr_cpu_id__die(cpu, /*data=*/NULL);
1347 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1348 struct perf_cpu cpu)
1350 return aggr_cpu_id__core(cpu, /*data=*/NULL);
1353 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1354 struct perf_cpu cpu)
1356 return aggr_cpu_id__node(cpu, /*data=*/NULL);
1359 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1360 aggr_get_id_t get_id, struct perf_cpu cpu)
1362 struct aggr_cpu_id id = aggr_cpu_id__empty();
1364 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1365 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1367 id = config->cpus_aggr_map->map[cpu.cpu];
1371 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1372 struct perf_cpu cpu)
1374 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1377 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1378 struct perf_cpu cpu)
1380 return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1383 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1384 struct perf_cpu cpu)
1386 return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1389 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1390 struct perf_cpu cpu)
1392 return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1395 static bool term_percore_set(void)
1397 struct evsel *counter;
1399 evlist__for_each_entry(evsel_list, counter) {
1400 if (counter->percore)
1407 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1409 switch (aggr_mode) {
1411 return aggr_cpu_id__socket;
1413 return aggr_cpu_id__die;
1415 return aggr_cpu_id__core;
1417 return aggr_cpu_id__node;
1419 if (term_percore_set())
1420 return aggr_cpu_id__core;
1431 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1433 switch (aggr_mode) {
1435 return perf_stat__get_socket_cached;
1437 return perf_stat__get_die_cached;
1439 return perf_stat__get_core_cached;
1441 return perf_stat__get_node_cached;
1443 if (term_percore_set()) {
1444 return perf_stat__get_core_cached;
1455 static int perf_stat_init_aggr_mode(void)
1458 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1461 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1462 get_id, /*data=*/NULL);
1463 if (!stat_config.aggr_map) {
1464 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1467 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1471 * The evsel_list->cpus is the base we operate on,
1472 * taking the highest cpu number to be the size of
1473 * the aggregation translate cpumap.
1475 if (evsel_list->core.user_requested_cpus)
1476 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1479 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1480 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1483 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1486 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1487 "cpu_aggr_map refcnt unbalanced\n");
1492 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1494 if (map && refcount_dec_and_test(&map->refcnt))
1495 cpu_aggr_map__delete(map);
1498 static void perf_stat__exit_aggr_mode(void)
1500 cpu_aggr_map__put(stat_config.aggr_map);
1501 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1502 stat_config.aggr_map = NULL;
1503 stat_config.cpus_aggr_map = NULL;
1506 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1508 struct perf_env *env = data;
1509 struct aggr_cpu_id id = aggr_cpu_id__empty();
1512 id.socket = env->cpu[cpu.cpu].socket_id;
1517 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1519 struct perf_env *env = data;
1520 struct aggr_cpu_id id = aggr_cpu_id__empty();
1522 if (cpu.cpu != -1) {
1524 * die_id is relative to socket, so start
1525 * with the socket ID and then add die to
1528 id.socket = env->cpu[cpu.cpu].socket_id;
1529 id.die = env->cpu[cpu.cpu].die_id;
1535 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1537 struct perf_env *env = data;
1538 struct aggr_cpu_id id = aggr_cpu_id__empty();
1540 if (cpu.cpu != -1) {
1542 * core_id is relative to socket and die,
1543 * we need a global id. So we set
1544 * socket, die id and core id
1546 id.socket = env->cpu[cpu.cpu].socket_id;
1547 id.die = env->cpu[cpu.cpu].die_id;
1548 id.core = env->cpu[cpu.cpu].core_id;
1554 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1556 struct aggr_cpu_id id = aggr_cpu_id__empty();
1558 id.node = perf_env__numa_node(data, cpu);
1562 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1563 struct perf_cpu cpu)
1565 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1567 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1568 struct perf_cpu cpu)
1570 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1573 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1574 struct perf_cpu cpu)
1576 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1579 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1580 struct perf_cpu cpu)
1582 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1585 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1587 switch (aggr_mode) {
1589 return perf_env__get_socket_aggr_by_cpu;
1591 return perf_env__get_die_aggr_by_cpu;
1593 return perf_env__get_core_aggr_by_cpu;
1595 return perf_env__get_node_aggr_by_cpu;
1605 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1607 switch (aggr_mode) {
1609 return perf_stat__get_socket_file;
1611 return perf_stat__get_die_file;
1613 return perf_stat__get_core_file;
1615 return perf_stat__get_node_file;
1625 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1627 struct perf_env *env = &st->session->header.env;
1628 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1633 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus, get_id, env);
1634 if (!stat_config.aggr_map) {
1635 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1638 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1643 * Add default attributes, if there were no attributes specified or
1644 * if -d/--detailed, -d -d or -d -d -d is used:
1646 static int add_default_attributes(void)
1649 struct perf_event_attr default_attrs0[] = {
1651 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1652 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1653 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1654 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1656 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1658 struct perf_event_attr frontend_attrs[] = {
1659 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1661 struct perf_event_attr backend_attrs[] = {
1662 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1664 struct perf_event_attr default_attrs1[] = {
1665 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1666 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1667 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1670 struct perf_event_attr default_sw_attrs[] = {
1671 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1672 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1673 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1674 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1678 * Detailed stats (-d), covering the L1 and last level data caches:
1680 struct perf_event_attr detailed_attrs[] = {
1682 { .type = PERF_TYPE_HW_CACHE,
1684 PERF_COUNT_HW_CACHE_L1D << 0 |
1685 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1686 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1688 { .type = PERF_TYPE_HW_CACHE,
1690 PERF_COUNT_HW_CACHE_L1D << 0 |
1691 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1692 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1694 { .type = PERF_TYPE_HW_CACHE,
1696 PERF_COUNT_HW_CACHE_LL << 0 |
1697 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1698 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1700 { .type = PERF_TYPE_HW_CACHE,
1702 PERF_COUNT_HW_CACHE_LL << 0 |
1703 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1704 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1708 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1710 struct perf_event_attr very_detailed_attrs[] = {
1712 { .type = PERF_TYPE_HW_CACHE,
1714 PERF_COUNT_HW_CACHE_L1I << 0 |
1715 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1716 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1718 { .type = PERF_TYPE_HW_CACHE,
1720 PERF_COUNT_HW_CACHE_L1I << 0 |
1721 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1722 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1724 { .type = PERF_TYPE_HW_CACHE,
1726 PERF_COUNT_HW_CACHE_DTLB << 0 |
1727 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1728 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1730 { .type = PERF_TYPE_HW_CACHE,
1732 PERF_COUNT_HW_CACHE_DTLB << 0 |
1733 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1734 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1736 { .type = PERF_TYPE_HW_CACHE,
1738 PERF_COUNT_HW_CACHE_ITLB << 0 |
1739 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1740 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1742 { .type = PERF_TYPE_HW_CACHE,
1744 PERF_COUNT_HW_CACHE_ITLB << 0 |
1745 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1746 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1751 * Very, very detailed stats (-d -d -d), adding prefetch events:
1753 struct perf_event_attr very_very_detailed_attrs[] = {
1755 { .type = PERF_TYPE_HW_CACHE,
1757 PERF_COUNT_HW_CACHE_L1D << 0 |
1758 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1759 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1761 { .type = PERF_TYPE_HW_CACHE,
1763 PERF_COUNT_HW_CACHE_L1D << 0 |
1764 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1765 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1767 /* Set attrs if no event is selected and !null_run: */
1768 if (stat_config.null_run)
1771 if (transaction_run) {
1772 struct parse_events_error errinfo;
1773 /* Handle -T as -M transaction. Once platform specific metrics
1774 * support has been added to the json files, all architectures
1775 * will use this approach. To determine transaction support
1776 * on an architecture test for such a metric name.
1778 if (metricgroup__has_metric("transaction")) {
1779 struct option opt = { .value = &evsel_list };
1781 return metricgroup__parse_groups(&opt, "transaction",
1782 stat_config.metric_no_group,
1783 stat_config.metric_no_merge,
1784 &stat_config.metric_events);
1787 parse_events_error__init(&errinfo);
1788 if (pmu_have_event("cpu", "cycles-ct") &&
1789 pmu_have_event("cpu", "el-start"))
1790 err = parse_events(evsel_list, transaction_attrs,
1793 err = parse_events(evsel_list,
1794 transaction_limited_attrs,
1797 fprintf(stderr, "Cannot set up transaction events\n");
1798 parse_events_error__print(&errinfo, transaction_attrs);
1800 parse_events_error__exit(&errinfo);
1801 return err ? -1 : 0;
1805 struct parse_events_error errinfo;
1808 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1809 fprintf(stderr, "freeze_on_smi is not supported.\n");
1814 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1815 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1821 if (!pmu_have_event("msr", "aperf") ||
1822 !pmu_have_event("msr", "smi")) {
1823 fprintf(stderr, "To measure SMI cost, it needs "
1824 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1827 if (!force_metric_only)
1828 stat_config.metric_only = true;
1830 parse_events_error__init(&errinfo);
1831 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1833 parse_events_error__print(&errinfo, smi_cost_attrs);
1834 fprintf(stderr, "Cannot set up SMI cost events\n");
1836 parse_events_error__exit(&errinfo);
1837 return err ? -1 : 0;
1841 const char **metric_attrs = topdown_metric_attrs;
1842 unsigned int max_level = 1;
1846 if (!force_metric_only)
1847 stat_config.metric_only = true;
1849 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1850 metric_attrs = topdown_metric_L2_attrs;
1854 if (stat_config.topdown_level > max_level) {
1855 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1857 } else if (!stat_config.topdown_level)
1858 stat_config.topdown_level = max_level;
1860 if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1861 pr_err("Out of memory\n");
1864 if (metric_attrs[0] && str) {
1865 if (!stat_config.interval && !stat_config.metric_only) {
1866 fprintf(stat_config.output,
1867 "Topdown accuracy may decrease when measuring long periods.\n"
1868 "Please print the result regularly, e.g. -I1000\n");
1875 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1876 stat_config.aggr_mode != AGGR_CORE) {
1877 pr_err("top down event configuration requires --per-core mode\n");
1880 stat_config.aggr_mode = AGGR_CORE;
1881 if (nr_cgroups || !target__has_cpu(&target)) {
1882 pr_err("top down event configuration requires system-wide mode (-a)\n");
1886 if (topdown_filter_events(topdown_attrs, &str,
1887 arch_topdown_check_group(&warn)) < 0) {
1888 pr_err("Out of memory\n");
1891 if (topdown_attrs[0] && str) {
1892 struct parse_events_error errinfo;
1894 arch_topdown_group_warn();
1896 parse_events_error__init(&errinfo);
1897 err = parse_events(evsel_list, str, &errinfo);
1900 "Cannot set up top down events %s: %d\n",
1902 parse_events_error__print(&errinfo, str);
1903 parse_events_error__exit(&errinfo);
1907 parse_events_error__exit(&errinfo);
1909 fprintf(stderr, "System does not support topdown\n");
1915 if (!evsel_list->core.nr_entries) {
1916 if (perf_pmu__has_hybrid()) {
1917 struct parse_events_error errinfo;
1918 const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1920 if (target__has_cpu(&target))
1921 default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1923 if (evlist__add_default_attrs(evsel_list,
1924 default_sw_attrs) < 0) {
1928 parse_events_error__init(&errinfo);
1929 err = parse_events(evsel_list, hybrid_str, &errinfo);
1932 "Cannot set up hybrid events %s: %d\n",
1934 parse_events_error__print(&errinfo, hybrid_str);
1936 parse_events_error__exit(&errinfo);
1937 return err ? -1 : 0;
1940 if (target__has_cpu(&target))
1941 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1943 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1945 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1946 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1949 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1950 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1953 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1956 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1957 if (arch_evlist__add_default_attrs(evsel_list) < 0)
1961 /* Detailed events get appended to the event list: */
1963 if (detailed_run < 1)
1966 /* Append detailed run extra attributes: */
1967 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1970 if (detailed_run < 2)
1973 /* Append very detailed run extra attributes: */
1974 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1977 if (detailed_run < 3)
1980 /* Append very, very detailed run extra attributes: */
1981 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1984 static const char * const stat_record_usage[] = {
1985 "perf stat record [<options>]",
1989 static void init_features(struct perf_session *session)
1993 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1994 perf_header__set_feat(&session->header, feat);
1996 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1997 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1998 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1999 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2000 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2003 static int __cmd_record(int argc, const char **argv)
2005 struct perf_session *session;
2006 struct perf_data *data = &perf_stat.data;
2008 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2009 PARSE_OPT_STOP_AT_NON_OPTION);
2012 data->path = output_name;
2014 if (stat_config.run_count != 1 || forever) {
2015 pr_err("Cannot use -r option with perf stat record.\n");
2019 session = perf_session__new(data, NULL);
2020 if (IS_ERR(session)) {
2021 pr_err("Perf session creation failed\n");
2022 return PTR_ERR(session);
2025 init_features(session);
2027 session->evlist = evsel_list;
2028 perf_stat.session = session;
2029 perf_stat.record = true;
2033 static int process_stat_round_event(struct perf_session *session,
2034 union perf_event *event)
2036 struct perf_record_stat_round *stat_round = &event->stat_round;
2037 struct evsel *counter;
2038 struct timespec tsh, *ts = NULL;
2039 const char **argv = session->header.env.cmdline_argv;
2040 int argc = session->header.env.nr_cmdline;
2042 evlist__for_each_entry(evsel_list, counter)
2043 perf_stat_process_counter(&stat_config, counter);
2045 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2046 update_stats(&walltime_nsecs_stats, stat_round->time);
2048 if (stat_config.interval && stat_round->time) {
2049 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2050 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2054 print_counters(ts, argc, argv);
2059 int process_stat_config_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);
2065 perf_event__read_stat_config(&stat_config, &event->stat_config);
2067 if (perf_cpu_map__empty(st->cpus)) {
2068 if (st->aggr_mode != AGGR_UNSET)
2069 pr_warning("warning: processing task data, aggregation mode not set\n");
2073 if (st->aggr_mode != AGGR_UNSET)
2074 stat_config.aggr_mode = st->aggr_mode;
2076 if (perf_stat.data.is_pipe)
2077 perf_stat_init_aggr_mode();
2079 perf_stat_init_aggr_mode_file(st);
2084 static int set_maps(struct perf_stat *st)
2086 if (!st->cpus || !st->threads)
2089 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2092 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2094 if (evlist__alloc_stats(evsel_list, true))
2097 st->maps_allocated = true;
2102 int process_thread_map_event(struct perf_session *session,
2103 union perf_event *event)
2105 struct perf_tool *tool = session->tool;
2106 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2109 pr_warning("Extra thread map event, ignoring.\n");
2113 st->threads = thread_map__new_event(&event->thread_map);
2117 return set_maps(st);
2121 int process_cpu_map_event(struct perf_session *session,
2122 union perf_event *event)
2124 struct perf_tool *tool = session->tool;
2125 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2126 struct perf_cpu_map *cpus;
2129 pr_warning("Extra cpu map event, ignoring.\n");
2133 cpus = cpu_map__new_data(&event->cpu_map.data);
2138 return set_maps(st);
2141 static const char * const stat_report_usage[] = {
2142 "perf stat report [<options>]",
2146 static struct perf_stat perf_stat = {
2148 .attr = perf_event__process_attr,
2149 .event_update = perf_event__process_event_update,
2150 .thread_map = process_thread_map_event,
2151 .cpu_map = process_cpu_map_event,
2152 .stat_config = process_stat_config_event,
2153 .stat = perf_event__process_stat_event,
2154 .stat_round = process_stat_round_event,
2156 .aggr_mode = AGGR_UNSET,
2159 static int __cmd_report(int argc, const char **argv)
2161 struct perf_session *session;
2162 const struct option options[] = {
2163 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2164 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2165 "aggregate counts per processor socket", AGGR_SOCKET),
2166 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2167 "aggregate counts per processor die", AGGR_DIE),
2168 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2169 "aggregate counts per physical processor core", AGGR_CORE),
2170 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2171 "aggregate counts per numa node", AGGR_NODE),
2172 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2173 "disable CPU count aggregation", AGGR_NONE),
2179 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2181 if (!input_name || !strlen(input_name)) {
2182 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2185 input_name = "perf.data";
2188 perf_stat.data.path = input_name;
2189 perf_stat.data.mode = PERF_DATA_MODE_READ;
2191 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2192 if (IS_ERR(session))
2193 return PTR_ERR(session);
2195 perf_stat.session = session;
2196 stat_config.output = stderr;
2197 evsel_list = session->evlist;
2199 ret = perf_session__process_events(session);
2203 perf_session__delete(session);
2207 static void setup_system_wide(int forks)
2210 * Make system wide (-a) the default target if
2211 * no target was specified and one of following
2212 * conditions is met:
2214 * - there's no workload specified
2215 * - there is workload specified but all requested
2216 * events are system wide events
2218 if (!target__none(&target))
2222 target.system_wide = true;
2224 struct evsel *counter;
2226 evlist__for_each_entry(evsel_list, counter) {
2227 if (!counter->core.system_wide &&
2228 strcmp(counter->name, "duration_time")) {
2233 if (evsel_list->core.nr_entries)
2234 target.system_wide = true;
2238 int cmd_stat(int argc, const char **argv)
2240 const char * const stat_usage[] = {
2241 "perf stat [<options>] [<command>]",
2244 int status = -EINVAL, run_idx, err;
2246 FILE *output = stderr;
2247 unsigned int interval, timeout;
2248 const char * const stat_subcommands[] = { "record", "report" };
2249 char errbuf[BUFSIZ];
2251 setlocale(LC_ALL, "");
2253 evsel_list = evlist__new();
2254 if (evsel_list == NULL)
2257 parse_events__shrink_config_terms();
2259 /* String-parsing callback-based options would segfault when negated */
2260 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2261 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2262 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2264 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2265 (const char **) stat_usage,
2266 PARSE_OPT_STOP_AT_NON_OPTION);
2267 perf_stat__collect_metric_expr(evsel_list);
2268 perf_stat__init_shadow_stats();
2270 if (stat_config.csv_sep) {
2271 stat_config.csv_output = true;
2272 if (!strcmp(stat_config.csv_sep, "\\t"))
2273 stat_config.csv_sep = "\t";
2275 stat_config.csv_sep = DEFAULT_SEPARATOR;
2277 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2278 argc = __cmd_record(argc, argv);
2281 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2282 return __cmd_report(argc, argv);
2284 interval = stat_config.interval;
2285 timeout = stat_config.timeout;
2288 * For record command the -o is already taken care of.
2290 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2293 if (output_name && output_fd) {
2294 fprintf(stderr, "cannot use both --output and --log-fd\n");
2295 parse_options_usage(stat_usage, stat_options, "o", 1);
2296 parse_options_usage(NULL, stat_options, "log-fd", 0);
2300 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2301 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2305 if (stat_config.metric_only && stat_config.run_count > 1) {
2306 fprintf(stderr, "--metric-only is not supported with -r\n");
2310 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2311 fprintf(stderr, "--table is only supported with -r\n");
2312 parse_options_usage(stat_usage, stat_options, "r", 1);
2313 parse_options_usage(NULL, stat_options, "table", 0);
2317 if (output_fd < 0) {
2318 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2319 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2323 if (!output && !stat_config.quiet) {
2325 mode = append_file ? "a" : "w";
2327 output = fopen(output_name, mode);
2329 perror("failed to create output file");
2332 clock_gettime(CLOCK_REALTIME, &tm);
2333 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2334 } else if (output_fd > 0) {
2335 mode = append_file ? "a" : "w";
2336 output = fdopen(output_fd, mode);
2338 perror("Failed opening logfd");
2343 stat_config.output = output;
2346 * let the spreadsheet do the pretty-printing
2348 if (stat_config.csv_output) {
2349 /* User explicitly passed -B? */
2350 if (big_num_opt == 1) {
2351 fprintf(stderr, "-B option not supported with -x\n");
2352 parse_options_usage(stat_usage, stat_options, "B", 1);
2353 parse_options_usage(NULL, stat_options, "x", 1);
2355 } else /* Nope, so disable big number formatting */
2356 stat_config.big_num = false;
2357 } else if (big_num_opt == 0) /* User passed --no-big-num */
2358 stat_config.big_num = false;
2360 err = target__validate(&target);
2362 target__strerror(&target, err, errbuf, BUFSIZ);
2363 pr_warning("%s\n", errbuf);
2366 setup_system_wide(argc);
2369 * Display user/system times only for single
2370 * run and when there's specified tracee.
2372 if ((stat_config.run_count == 1) && target__none(&target))
2373 stat_config.ru_display = true;
2375 if (stat_config.run_count < 0) {
2376 pr_err("Run count must be a positive number\n");
2377 parse_options_usage(stat_usage, stat_options, "r", 1);
2379 } else if (stat_config.run_count == 0) {
2381 stat_config.run_count = 1;
2384 if (stat_config.walltime_run_table) {
2385 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2386 if (!stat_config.walltime_run) {
2387 pr_err("failed to setup -r option");
2392 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2393 !target__has_task(&target)) {
2394 if (!target.system_wide || target.cpu_list) {
2395 fprintf(stderr, "The --per-thread option is only "
2396 "available when monitoring via -p -t -a "
2397 "options or only --per-thread.\n");
2398 parse_options_usage(NULL, stat_options, "p", 1);
2399 parse_options_usage(NULL, stat_options, "t", 1);
2405 * no_aggr, cgroup are for system-wide only
2406 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2408 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2409 stat_config.aggr_mode != AGGR_THREAD) ||
2410 (nr_cgroups || stat_config.cgroup_list)) &&
2411 !target__has_cpu(&target)) {
2412 fprintf(stderr, "both cgroup and no-aggregation "
2413 "modes only available in system-wide mode\n");
2415 parse_options_usage(stat_usage, stat_options, "G", 1);
2416 parse_options_usage(NULL, stat_options, "A", 1);
2417 parse_options_usage(NULL, stat_options, "a", 1);
2418 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2422 if (stat_config.iostat_run) {
2423 status = iostat_prepare(evsel_list, &stat_config);
2426 if (iostat_mode == IOSTAT_LIST) {
2427 iostat_list(evsel_list, &stat_config);
2430 iostat_list(evsel_list, &stat_config);
2431 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2432 target.system_wide = true;
2435 if (add_default_attributes())
2438 if (stat_config.cgroup_list) {
2439 if (nr_cgroups > 0) {
2440 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2441 parse_options_usage(stat_usage, stat_options, "G", 1);
2442 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2446 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2447 &stat_config.metric_events, true) < 0) {
2448 parse_options_usage(stat_usage, stat_options,
2449 "for-each-cgroup", 0);
2454 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2455 target.per_thread = true;
2457 if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2458 pr_err("failed to use cpu list %s\n", target.cpu_list);
2462 target.hybrid = perf_pmu__has_hybrid();
2463 if (evlist__create_maps(evsel_list, &target) < 0) {
2464 if (target__has_task(&target)) {
2465 pr_err("Problems finding threads of monitor\n");
2466 parse_options_usage(stat_usage, stat_options, "p", 1);
2467 parse_options_usage(NULL, stat_options, "t", 1);
2468 } else if (target__has_cpu(&target)) {
2469 perror("failed to parse CPUs map");
2470 parse_options_usage(stat_usage, stat_options, "C", 1);
2471 parse_options_usage(NULL, stat_options, "a", 1);
2476 evlist__check_cpu_maps(evsel_list);
2479 * Initialize thread_map with comm names,
2480 * so we could print it out on output.
2482 if (stat_config.aggr_mode == AGGR_THREAD) {
2483 thread_map__read_comms(evsel_list->core.threads);
2484 if (target.system_wide) {
2485 if (runtime_stat_new(&stat_config,
2486 perf_thread_map__nr(evsel_list->core.threads))) {
2492 if (stat_config.aggr_mode == AGGR_NODE)
2493 cpu__setup_cpunode_map();
2495 if (stat_config.times && interval)
2496 interval_count = true;
2497 else if (stat_config.times && !interval) {
2498 pr_err("interval-count option should be used together with "
2499 "interval-print.\n");
2500 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2501 parse_options_usage(stat_usage, stat_options, "I", 1);
2505 if (timeout && timeout < 100) {
2507 pr_err("timeout must be >= 10ms.\n");
2508 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2511 pr_warning("timeout < 100ms. "
2512 "The overhead percentage could be high in some cases. "
2513 "Please proceed with caution.\n");
2515 if (timeout && interval) {
2516 pr_err("timeout option is not supported with interval-print.\n");
2517 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2518 parse_options_usage(stat_usage, stat_options, "I", 1);
2522 if (evlist__alloc_stats(evsel_list, interval))
2525 if (perf_stat_init_aggr_mode())
2529 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2530 * while avoiding that older tools show confusing messages.
2532 * However for pipe sessions we need to keep it zero,
2533 * because script's perf_evsel__check_attr is triggered
2534 * by attr->sample_type != 0, and we can't run it on
2537 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2540 * We dont want to block the signals - that would cause
2541 * child tasks to inherit that and Ctrl-C would not work.
2542 * What we want is for Ctrl-C to work in the exec()-ed
2543 * task, but being ignored by perf stat itself:
2547 signal(SIGINT, skip_signal);
2548 signal(SIGCHLD, skip_signal);
2549 signal(SIGALRM, skip_signal);
2550 signal(SIGABRT, skip_signal);
2552 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2556 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2557 if (stat_config.run_count != 1 && verbose > 0)
2558 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2562 evlist__reset_prev_raw_counts(evsel_list);
2564 status = run_perf_stat(argc, argv, run_idx);
2565 if (forever && status != -1 && !interval) {
2566 print_counters(NULL, argc, argv);
2567 perf_stat__reset_stats();
2571 if (!forever && status != -1 && (!interval || stat_config.summary))
2572 print_counters(NULL, argc, argv);
2574 evlist__finalize_ctlfd(evsel_list);
2578 * We synthesize the kernel mmap record just so that older tools
2579 * don't emit warnings about not being able to resolve symbols
2580 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2581 * a saner message about no samples being in the perf.data file.
2583 * This also serves to suppress a warning about f_header.data.size == 0
2584 * in header.c at the moment 'perf stat record' gets introduced, which
2585 * is not really needed once we start adding the stat specific PERF_RECORD_
2586 * records, but the need to suppress the kptr_restrict messages in older
2587 * tools remain -acme
2589 int fd = perf_data__fd(&perf_stat.data);
2591 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2592 process_synthesized_event,
2593 &perf_stat.session->machines.host);
2595 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2596 "older tools may produce warnings about this file\n.");
2600 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2601 pr_err("failed to write stat round event\n");
2604 if (!perf_stat.data.is_pipe) {
2605 perf_stat.session->header.data_size += perf_stat.bytes_written;
2606 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2609 evlist__close(evsel_list);
2610 perf_session__delete(perf_stat.session);
2613 perf_stat__exit_aggr_mode();
2614 evlist__free_stats(evsel_list);
2616 if (stat_config.iostat_run)
2617 iostat_release(evsel_list);
2619 zfree(&stat_config.walltime_run);
2621 if (smi_cost && smi_reset)
2622 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2624 evlist__delete(evsel_list);
2626 metricgroup__rblist_exit(&stat_config.metric_events);
2627 runtime_stat_delete(&stat_config);
2628 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);