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>
44 #include "util/cgroup.h"
45 #include <subcmd/parse-options.h>
46 #include "util/parse-events.h"
47 #include "util/pmus.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread_map.h"
58 #include "util/counts.h"
59 #include "util/topdown.h"
60 #include "util/session.h"
61 #include "util/tool.h"
62 #include "util/string2.h"
63 #include "util/metricgroup.h"
64 #include "util/synthetic-events.h"
65 #include "util/target.h"
66 #include "util/time-utils.h"
68 #include "util/affinity.h"
70 #include "util/bpf_counter.h"
71 #include "util/iostat.h"
72 #include "util/util.h"
75 #include <linux/time64.h>
76 #include <linux/zalloc.h>
77 #include <api/fs/fs.h>
81 #include <sys/prctl.h>
85 #include <sys/types.h>
90 #include <sys/resource.h>
91 #include <linux/err.h>
93 #include <linux/ctype.h>
94 #include <perf/evlist.h>
95 #include <internal/threadmap.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 static struct evlist *evsel_list;
103 static struct parse_events_option_args parse_events_option_args = {
104 .evlistp = &evsel_list,
107 static bool all_counters_use_bpf = true;
109 static struct target target = {
113 #define METRIC_ONLY_LEN 20
115 static volatile sig_atomic_t child_pid = -1;
116 static int detailed_run = 0;
117 static bool transaction_run;
118 static bool topdown_run = false;
119 static bool smi_cost = false;
120 static bool smi_reset = false;
121 static int big_num_opt = -1;
122 static const char *pre_cmd = NULL;
123 static const char *post_cmd = NULL;
124 static bool sync_run = false;
125 static bool forever = false;
126 static bool force_metric_only = false;
127 static struct timespec ref_time;
128 static bool append_file;
129 static bool interval_count;
130 static const char *output_name;
131 static int output_fd;
132 static char *metrics;
136 struct perf_data data;
137 struct perf_session *session;
139 struct perf_tool tool;
141 struct perf_cpu_map *cpus;
142 struct perf_thread_map *threads;
143 enum aggr_mode aggr_mode;
147 static struct perf_stat perf_stat;
148 #define STAT_RECORD perf_stat.record
150 static volatile sig_atomic_t done = 0;
152 static struct perf_stat_config stat_config = {
153 .aggr_mode = AGGR_GLOBAL,
154 .aggr_level = MAX_CACHE_LVL + 1,
156 .unit_width = 4, /* strlen("unit") */
158 .metric_only_len = METRIC_ONLY_LEN,
159 .walltime_nsecs_stats = &walltime_nsecs_stats,
160 .ru_stats = &ru_stats,
167 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
169 if (!a->core.cpus && !b->core.cpus)
172 if (!a->core.cpus || !b->core.cpus)
175 if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
178 for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
179 if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
180 perf_cpu_map__cpu(b->core.cpus, i).cpu)
187 static void evlist__check_cpu_maps(struct evlist *evlist)
189 struct evsel *evsel, *warned_leader = NULL;
191 evlist__for_each_entry(evlist, evsel) {
192 struct evsel *leader = evsel__leader(evsel);
194 /* Check that leader matches cpus with each member. */
197 if (cpus_map_matched(leader, evsel))
200 /* If there's mismatch disable the group and warn user. */
201 if (warned_leader != leader) {
204 pr_warning("WARNING: grouped events cpus do not match.\n"
205 "Events with CPUs not matching the leader will "
206 "be removed from the group.\n");
207 evsel__group_desc(leader, buf, sizeof(buf));
208 pr_warning(" %s\n", buf);
209 warned_leader = leader;
214 cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
215 pr_warning(" %s: %s\n", leader->name, buf);
216 cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
217 pr_warning(" %s: %s\n", evsel->name, buf);
220 evsel__remove_from_group(evsel, leader);
224 static inline void diff_timespec(struct timespec *r, struct timespec *a,
227 r->tv_sec = a->tv_sec - b->tv_sec;
228 if (a->tv_nsec < b->tv_nsec) {
229 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
232 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
236 static void perf_stat__reset_stats(void)
238 evlist__reset_stats(evsel_list);
239 perf_stat__reset_shadow_stats();
242 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
243 union perf_event *event,
244 struct perf_sample *sample __maybe_unused,
245 struct machine *machine __maybe_unused)
247 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
248 pr_err("failed to write perf data, error: %m\n");
252 perf_stat.bytes_written += event->header.size;
256 static int write_stat_round_event(u64 tm, u64 type)
258 return perf_event__synthesize_stat_round(NULL, tm, type,
259 process_synthesized_event,
263 #define WRITE_STAT_ROUND_EVENT(time, interval) \
264 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
266 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
268 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
269 struct perf_counts_values *count)
271 struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
272 struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
274 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
275 process_synthesized_event, NULL);
278 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
279 int thread, struct timespec *rs)
281 switch(counter->tool_event) {
282 case PERF_TOOL_DURATION_TIME: {
283 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
284 struct perf_counts_values *count =
285 perf_counts(counter->counts, cpu_map_idx, thread);
286 count->ena = count->run = val;
290 case PERF_TOOL_USER_TIME:
291 case PERF_TOOL_SYSTEM_TIME: {
293 struct perf_counts_values *count =
294 perf_counts(counter->counts, cpu_map_idx, thread);
295 if (counter->tool_event == PERF_TOOL_USER_TIME)
296 val = ru_stats.ru_utime_usec_stat.mean;
298 val = ru_stats.ru_stime_usec_stat.mean;
299 count->ena = count->run = val;
305 return evsel__read_counter(counter, cpu_map_idx, thread);
307 /* This should never be reached */
313 * Read out the results of a single counter:
314 * do not aggregate counts across CPUs in system-wide mode
316 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
318 int nthreads = perf_thread_map__nr(evsel_list->core.threads);
321 if (!counter->supported)
324 for (thread = 0; thread < nthreads; thread++) {
325 struct perf_counts_values *count;
327 count = perf_counts(counter->counts, cpu_map_idx, thread);
330 * The leader's group read loads data into its group members
331 * (via evsel__read_counter()) and sets their count->loaded.
333 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
334 read_single_counter(counter, cpu_map_idx, thread, rs)) {
335 counter->counts->scaled = -1;
336 perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
337 perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
341 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
344 if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
345 pr_err("failed to write stat event\n");
351 fprintf(stat_config.output,
352 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
353 evsel__name(counter),
354 perf_cpu_map__cpu(evsel__cpus(counter),
356 count->val, count->ena, count->run);
363 static int read_affinity_counters(struct timespec *rs)
365 struct evlist_cpu_iterator evlist_cpu_itr;
366 struct affinity saved_affinity, *affinity;
368 if (all_counters_use_bpf)
371 if (!target__has_cpu(&target) || target__has_per_thread(&target))
373 else if (affinity__setup(&saved_affinity) < 0)
376 affinity = &saved_affinity;
378 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
379 struct evsel *counter = evlist_cpu_itr.evsel;
381 if (evsel__is_bpf(counter))
385 counter->err = read_counter_cpu(counter, rs,
386 evlist_cpu_itr.cpu_map_idx);
390 affinity__cleanup(&saved_affinity);
395 static int read_bpf_map_counters(void)
397 struct evsel *counter;
400 evlist__for_each_entry(evsel_list, counter) {
401 if (!evsel__is_bpf(counter))
404 err = bpf_counter__read(counter);
411 static int read_counters(struct timespec *rs)
413 if (!stat_config.stop_read_counter) {
414 if (read_bpf_map_counters() ||
415 read_affinity_counters(rs))
421 static void process_counters(void)
423 struct evsel *counter;
425 evlist__for_each_entry(evsel_list, counter) {
427 pr_debug("failed to read counter %s\n", counter->name);
428 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
429 pr_warning("failed to process counter %s\n", counter->name);
433 perf_stat_merge_counters(&stat_config, evsel_list);
434 perf_stat_process_percore(&stat_config, evsel_list);
437 static void process_interval(void)
439 struct timespec ts, rs;
441 clock_gettime(CLOCK_MONOTONIC, &ts);
442 diff_timespec(&rs, &ts, &ref_time);
444 evlist__reset_aggr_stats(evsel_list);
446 if (read_counters(&rs) == 0)
450 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
451 pr_err("failed to write stat round event\n");
454 init_stats(&walltime_nsecs_stats);
455 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
456 print_counters(&rs, 0, NULL);
459 static bool handle_interval(unsigned int interval, int *times)
463 if (interval_count && !(--(*times)))
469 static int enable_counters(void)
474 evlist__for_each_entry(evsel_list, evsel) {
475 if (!evsel__is_bpf(evsel))
478 err = bpf_counter__enable(evsel);
483 if (!target__enable_on_exec(&target)) {
484 if (!all_counters_use_bpf)
485 evlist__enable(evsel_list);
490 static void disable_counters(void)
492 struct evsel *counter;
495 * If we don't have tracee (attaching to task or cpu), counters may
496 * still be running. To get accurate group ratios, we must stop groups
497 * from counting before reading their constituent counters.
499 if (!target__none(&target)) {
500 evlist__for_each_entry(evsel_list, counter)
501 bpf_counter__disable(counter);
502 if (!all_counters_use_bpf)
503 evlist__disable(evsel_list);
507 static volatile sig_atomic_t workload_exec_errno;
510 * evlist__prepare_workload will send a SIGUSR1
511 * if the fork fails, since we asked by setting its
512 * want_signal to true.
514 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
515 void *ucontext __maybe_unused)
517 workload_exec_errno = info->si_value.sival_int;
520 static bool evsel__should_store_id(struct evsel *counter)
522 return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
525 static bool is_target_alive(struct target *_target,
526 struct perf_thread_map *threads)
531 if (!target__has_task(_target))
534 for (i = 0; i < threads->nr; i++) {
537 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
538 threads->map[i].pid);
540 if (!stat(path, &st))
547 static void process_evlist(struct evlist *evlist, unsigned int interval)
549 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
551 if (evlist__ctlfd_process(evlist, &cmd) > 0) {
553 case EVLIST_CTL_CMD_ENABLE:
555 case EVLIST_CTL_CMD_DISABLE:
559 case EVLIST_CTL_CMD_SNAPSHOT:
560 case EVLIST_CTL_CMD_ACK:
561 case EVLIST_CTL_CMD_UNSUPPORTED:
562 case EVLIST_CTL_CMD_EVLIST:
563 case EVLIST_CTL_CMD_STOP:
564 case EVLIST_CTL_CMD_PING:
571 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
574 int tts = *time_to_sleep;
575 struct timespec time_diff;
577 diff_timespec(&time_diff, time_stop, time_start);
579 tts -= time_diff.tv_sec * MSEC_PER_SEC +
580 time_diff.tv_nsec / NSEC_PER_MSEC;
585 *time_to_sleep = tts;
588 static int dispatch_events(bool forks, int timeout, int interval, int *times)
590 int child_exited = 0, status = 0;
591 int time_to_sleep, sleep_time;
592 struct timespec time_start, time_stop;
595 sleep_time = interval;
597 sleep_time = timeout;
601 time_to_sleep = sleep_time;
605 child_exited = waitpid(child_pid, &status, WNOHANG);
607 child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
612 clock_gettime(CLOCK_MONOTONIC, &time_start);
613 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
614 if (timeout || handle_interval(interval, times))
616 time_to_sleep = sleep_time;
617 } else { /* fd revent */
618 process_evlist(evsel_list, interval);
619 clock_gettime(CLOCK_MONOTONIC, &time_stop);
620 compute_tts(&time_start, &time_stop, &time_to_sleep);
627 enum counter_recovery {
633 static enum counter_recovery stat_handle_error(struct evsel *counter)
637 * PPC returns ENXIO for HW counters until 2.6.37
638 * (behavior changed with commit b0a873e).
640 if (errno == EINVAL || errno == ENOSYS ||
641 errno == ENOENT || errno == EOPNOTSUPP ||
644 ui__warning("%s event is not supported by the kernel.\n",
645 evsel__name(counter));
646 counter->supported = false;
648 * errored is a sticky flag that means one of the counter's
649 * cpu event had a problem and needs to be reexamined.
651 counter->errored = true;
653 if ((evsel__leader(counter) != counter) ||
654 !(counter->core.leader->nr_members > 1))
656 } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
658 ui__warning("%s\n", msg);
659 return COUNTER_RETRY;
660 } else if (target__has_per_thread(&target) &&
661 evsel_list->core.threads &&
662 evsel_list->core.threads->err_thread != -1) {
664 * For global --per-thread case, skip current
667 if (!thread_map__remove(evsel_list->core.threads,
668 evsel_list->core.threads->err_thread)) {
669 evsel_list->core.threads->err_thread = -1;
670 return COUNTER_RETRY;
672 } else if (counter->skippable) {
674 ui__warning("skipping event %s that kernel failed to open .\n",
675 evsel__name(counter));
676 counter->supported = false;
677 counter->errored = true;
681 evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
682 ui__error("%s\n", msg);
685 kill(child_pid, SIGTERM);
686 return COUNTER_FATAL;
689 static int __run_perf_stat(int argc, const char **argv, int run_idx)
691 int interval = stat_config.interval;
692 int times = stat_config.times;
693 int timeout = stat_config.timeout;
695 unsigned long long t0, t1;
696 struct evsel *counter;
699 const bool forks = (argc > 0);
700 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
701 struct evlist_cpu_iterator evlist_cpu_itr;
702 struct affinity saved_affinity, *affinity = NULL;
704 bool second_pass = false;
707 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
708 perror("failed to prepare workload");
711 child_pid = evsel_list->workload.pid;
714 if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
715 if (affinity__setup(&saved_affinity) < 0)
717 affinity = &saved_affinity;
720 evlist__for_each_entry(evsel_list, counter) {
721 counter->reset_group = false;
722 if (bpf_counter__load(counter, &target))
724 if (!(evsel__is_bperf(counter)))
725 all_counters_use_bpf = false;
728 evlist__reset_aggr_stats(evsel_list);
730 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
731 counter = evlist_cpu_itr.evsel;
734 * bperf calls evsel__open_per_cpu() in bperf__load(), so
735 * no need to call it again here.
740 if (counter->reset_group || counter->errored)
742 if (evsel__is_bperf(counter))
745 if (create_perf_stat_counter(counter, &stat_config, &target,
746 evlist_cpu_itr.cpu_map_idx) < 0) {
749 * Weak group failed. We cannot just undo this here
750 * because earlier CPUs might be in group mode, and the kernel
751 * doesn't support mixing group and non group reads. Defer
753 * Don't close here because we're in the wrong affinity.
755 if ((errno == EINVAL || errno == EBADF) &&
756 evsel__leader(counter) != counter &&
757 counter->weak_group) {
758 evlist__reset_weak_group(evsel_list, counter, false);
759 assert(counter->reset_group);
764 switch (stat_handle_error(counter)) {
776 counter->supported = true;
781 * Now redo all the weak group after closing them,
782 * and also close errored counters.
785 /* First close errored or weak retry */
786 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
787 counter = evlist_cpu_itr.evsel;
789 if (!counter->reset_group && !counter->errored)
792 perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
794 /* Now reopen weak */
795 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
796 counter = evlist_cpu_itr.evsel;
798 if (!counter->reset_group)
801 pr_debug2("reopening weak %s\n", evsel__name(counter));
802 if (create_perf_stat_counter(counter, &stat_config, &target,
803 evlist_cpu_itr.cpu_map_idx) < 0) {
805 switch (stat_handle_error(counter)) {
809 goto try_again_reset;
816 counter->supported = true;
819 affinity__cleanup(affinity);
821 evlist__for_each_entry(evsel_list, counter) {
822 if (!counter->supported) {
823 perf_evsel__free_fd(&counter->core);
827 l = strlen(counter->unit);
828 if (l > stat_config.unit_width)
829 stat_config.unit_width = l;
831 if (evsel__should_store_id(counter) &&
832 evsel__store_ids(counter, evsel_list))
836 if (evlist__apply_filters(evsel_list, &counter)) {
837 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
838 counter->filter, evsel__name(counter), errno,
839 str_error_r(errno, msg, sizeof(msg)));
844 int fd = perf_data__fd(&perf_stat.data);
847 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
849 err = perf_session__write_header(perf_stat.session, evsel_list,
856 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
857 process_synthesized_event, is_pipe);
862 if (target.initial_delay) {
863 pr_info(EVLIST_DISABLED_MSG);
865 err = enable_counters();
870 /* Exec the command, if any */
872 evlist__start_workload(evsel_list);
874 if (target.initial_delay > 0) {
875 usleep(target.initial_delay * USEC_PER_MSEC);
876 err = enable_counters();
880 pr_info(EVLIST_ENABLED_MSG);
884 clock_gettime(CLOCK_MONOTONIC, &ref_time);
887 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
888 status = dispatch_events(forks, timeout, interval, ×);
889 if (child_pid != -1) {
891 kill(child_pid, SIGTERM);
892 wait4(child_pid, &status, 0, &stat_config.ru_data);
895 if (workload_exec_errno) {
896 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
897 pr_err("Workload failed: %s\n", emsg);
901 if (WIFSIGNALED(status))
902 psignal(WTERMSIG(status), argv[0]);
904 status = dispatch_events(forks, timeout, interval, ×);
911 if (stat_config.walltime_run_table)
912 stat_config.walltime_run[run_idx] = t1 - t0;
914 if (interval && stat_config.summary) {
915 stat_config.interval = 0;
916 stat_config.stop_read_counter = true;
917 init_stats(&walltime_nsecs_stats);
918 update_stats(&walltime_nsecs_stats, t1 - t0);
920 evlist__copy_prev_raw_counts(evsel_list);
921 evlist__reset_prev_raw_counts(evsel_list);
922 evlist__reset_aggr_stats(evsel_list);
924 update_stats(&walltime_nsecs_stats, t1 - t0);
925 update_rusage_stats(&ru_stats, &stat_config.ru_data);
929 * Closing a group leader splits the group, and as we only disable
930 * group leaders, results in remaining events becoming enabled. To
931 * avoid arbitrary skew, we must read all counters before closing any
934 if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
938 * We need to keep evsel_list alive, because it's processed
939 * later the evsel_list will be closed after.
942 evlist__close(evsel_list);
944 return WEXITSTATUS(status);
947 static int run_perf_stat(int argc, const char **argv, int run_idx)
952 ret = system(pre_cmd);
960 ret = __run_perf_stat(argc, argv, run_idx);
965 ret = system(post_cmd);
973 static void print_counters(struct timespec *ts, int argc, const char **argv)
975 /* Do not print anything if we record to the pipe. */
976 if (STAT_RECORD && perf_stat.data.is_pipe)
981 evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
984 static volatile sig_atomic_t signr = -1;
986 static void skip_signal(int signo)
988 if ((child_pid == -1) || stat_config.interval)
993 * render child_pid harmless
994 * won't send SIGTERM to a random
995 * process in case of race condition
996 * and fast PID recycling
1001 static void sig_atexit(void)
1006 * avoid race condition with SIGCHLD handler
1007 * in skip_signal() which is modifying child_pid
1008 * goal is to avoid send SIGTERM to a random
1012 sigaddset(&set, SIGCHLD);
1013 sigprocmask(SIG_BLOCK, &set, &oset);
1015 if (child_pid != -1)
1016 kill(child_pid, SIGTERM);
1018 sigprocmask(SIG_SETMASK, &oset, NULL);
1023 signal(signr, SIG_DFL);
1024 kill(getpid(), signr);
1027 void perf_stat__set_big_num(int set)
1029 stat_config.big_num = (set != 0);
1032 void perf_stat__set_no_csv_summary(int set)
1034 stat_config.no_csv_summary = (set != 0);
1037 static int stat__set_big_num(const struct option *opt __maybe_unused,
1038 const char *s __maybe_unused, int unset)
1040 big_num_opt = unset ? 0 : 1;
1041 perf_stat__set_big_num(!unset);
1045 static int enable_metric_only(const struct option *opt __maybe_unused,
1046 const char *s __maybe_unused, int unset)
1048 force_metric_only = true;
1049 stat_config.metric_only = !unset;
1053 static int append_metric_groups(const struct option *opt __maybe_unused,
1055 int unset __maybe_unused)
1060 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1065 metrics = strdup(str);
1072 static int parse_control_option(const struct option *opt,
1074 int unset __maybe_unused)
1076 struct perf_stat_config *config = opt->value;
1078 return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1081 static int parse_stat_cgroups(const struct option *opt,
1082 const char *str, int unset)
1084 if (stat_config.cgroup_list) {
1085 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1089 return parse_cgroups(opt, str, unset);
1092 static int parse_cputype(const struct option *opt,
1094 int unset __maybe_unused)
1096 const struct perf_pmu *pmu;
1097 struct evlist *evlist = *(struct evlist **)opt->value;
1099 if (!list_empty(&evlist->core.entries)) {
1100 fprintf(stderr, "Must define cputype before events/metrics\n");
1104 pmu = perf_pmus__pmu_for_pmu_filter(str);
1106 fprintf(stderr, "--cputype %s is not supported!\n", str);
1109 parse_events_option_args.pmu_filter = pmu->name;
1114 static int parse_cache_level(const struct option *opt,
1116 int unset __maybe_unused)
1119 u32 *aggr_mode = (u32 *)opt->value;
1120 u32 *aggr_level = (u32 *)opt->data;
1123 * If no string is specified, aggregate based on the topology of
1124 * Last Level Cache (LLC). Since the LLC level can change from
1125 * architecture to architecture, set level greater than
1126 * MAX_CACHE_LVL which will be interpreted as LLC.
1129 level = MAX_CACHE_LVL + 1;
1134 * The format to specify cache level is LX or lX where X is the
1137 if (strlen(str) != 2 || (str[0] != 'l' && str[0] != 'L')) {
1138 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1144 level = atoi(&str[1]);
1146 pr_err("Cache level must be of form L[1-%d], or l[1-%d]\n",
1152 if (level > MAX_CACHE_LVL) {
1153 pr_err("perf only supports max cache level of %d.\n"
1154 "Consider increasing MAX_CACHE_LVL\n", MAX_CACHE_LVL);
1158 *aggr_mode = AGGR_CACHE;
1159 *aggr_level = level;
1163 static struct option stat_options[] = {
1164 OPT_BOOLEAN('T', "transaction", &transaction_run,
1165 "hardware transaction statistics"),
1166 OPT_CALLBACK('e', "event", &parse_events_option_args, "event",
1167 "event selector. use 'perf list' to list available events",
1168 parse_events_option),
1169 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1170 "event filter", parse_filter),
1171 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1172 "child tasks do not inherit counters"),
1173 OPT_STRING('p', "pid", &target.pid, "pid",
1174 "stat events on existing process id"),
1175 OPT_STRING('t', "tid", &target.tid, "tid",
1176 "stat events on existing thread id"),
1177 #ifdef HAVE_BPF_SKEL
1178 OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1179 "stat events on existing bpf program id"),
1180 OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1181 "use bpf program to count events"),
1182 OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1183 "path to perf_event_attr map"),
1185 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1186 "system-wide collection from all CPUs"),
1187 OPT_BOOLEAN(0, "scale", &stat_config.scale,
1188 "Use --no-scale to disable counter scaling for multiplexing"),
1189 OPT_INCR('v', "verbose", &verbose,
1190 "be more verbose (show counter open errors, etc)"),
1191 OPT_INTEGER('r', "repeat", &stat_config.run_count,
1192 "repeat command and print average + stddev (max: 100, forever: 0)"),
1193 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1194 "display details about each run (only with -r option)"),
1195 OPT_BOOLEAN('n', "null", &stat_config.null_run,
1196 "null run - dont start any counters"),
1197 OPT_INCR('d', "detailed", &detailed_run,
1198 "detailed run - start a lot of events"),
1199 OPT_BOOLEAN('S', "sync", &sync_run,
1200 "call sync() before starting a run"),
1201 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1202 "print large numbers with thousands\' separators",
1204 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1205 "list of cpus to monitor in system-wide"),
1206 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1207 "disable aggregation across CPUs or PMUs", AGGR_NONE),
1208 OPT_SET_UINT(0, "no-merge", &stat_config.aggr_mode,
1209 "disable aggregation the same as -A or -no-aggr", AGGR_NONE),
1210 OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1211 "Merge identical named hybrid events"),
1212 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1213 "print counts with custom separator"),
1214 OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1215 "print counts in JSON format"),
1216 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1217 "monitor event in cgroup name only", parse_stat_cgroups),
1218 OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1219 "expand events for each cgroup"),
1220 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1221 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1222 OPT_INTEGER(0, "log-fd", &output_fd,
1223 "log output to fd, instead of stderr"),
1224 OPT_STRING(0, "pre", &pre_cmd, "command",
1225 "command to run prior to the measured command"),
1226 OPT_STRING(0, "post", &post_cmd, "command",
1227 "command to run after to the measured command"),
1228 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1229 "print counts at regular interval in ms "
1230 "(overhead is possible for values <= 100ms)"),
1231 OPT_INTEGER(0, "interval-count", &stat_config.times,
1232 "print counts for fixed number of times"),
1233 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1234 "clear screen in between new interval"),
1235 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1236 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1237 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1238 "aggregate counts per processor socket", AGGR_SOCKET),
1239 OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1240 "aggregate counts per processor die", AGGR_DIE),
1241 OPT_SET_UINT(0, "per-cluster", &stat_config.aggr_mode,
1242 "aggregate counts per processor cluster", AGGR_CLUSTER),
1243 OPT_CALLBACK_OPTARG(0, "per-cache", &stat_config.aggr_mode, &stat_config.aggr_level,
1244 "cache level", "aggregate count at this cache level (Default: LLC)",
1246 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1247 "aggregate counts per physical processor core", AGGR_CORE),
1248 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1249 "aggregate counts per thread", AGGR_THREAD),
1250 OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1251 "aggregate counts per numa node", AGGR_NODE),
1252 OPT_INTEGER('D', "delay", &target.initial_delay,
1253 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1254 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1255 "Only print computed metrics. No raw values", enable_metric_only),
1256 OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1257 "don't group metric events, impacts multiplexing"),
1258 OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1259 "don't try to share events between metrics in a group"),
1260 OPT_BOOLEAN(0, "metric-no-threshold", &stat_config.metric_no_threshold,
1261 "disable adding events for the metric threshold calculation"),
1262 OPT_BOOLEAN(0, "topdown", &topdown_run,
1263 "measure top-down statistics"),
1264 OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1265 "Set the metrics level for the top-down statistics (0: max level)"),
1266 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1267 "measure SMI cost"),
1268 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1269 "monitor specified metrics or metric groups (separated by ,)",
1270 append_metric_groups),
1271 OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1272 "Configure all used events to run in kernel space.",
1273 PARSE_OPT_EXCLUSIVE),
1274 OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1275 "Configure all used events to run in user space.",
1276 PARSE_OPT_EXCLUSIVE),
1277 OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1278 "Use with 'percore' event qualifier to show the event "
1279 "counts of one hardware thread by sum up total hardware "
1280 "threads of same physical core"),
1281 OPT_BOOLEAN(0, "summary", &stat_config.summary,
1282 "print summary for interval mode"),
1283 OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1284 "don't print 'summary' for CSV summary output"),
1285 OPT_BOOLEAN(0, "quiet", &quiet,
1286 "don't print any output, messages or warnings (useful with record)"),
1287 OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1288 "Only enable events on applying cpu with this type "
1289 "for hybrid platform (e.g. core or atom)",
1292 OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1293 "libpfm4 event selector. use 'perf list' to list available events",
1294 parse_libpfm_events_option),
1296 OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1297 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1298 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1299 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1300 parse_control_option),
1301 OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1302 "measure I/O performance metrics provided by arch/platform",
1308 * Calculate the cache instance ID from the map in
1309 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1310 * Cache instance ID is the first CPU reported in the shared_cpu_list file.
1312 static int cpu__get_cache_id_from_map(struct perf_cpu cpu, char *map)
1315 struct perf_cpu_map *cpu_map = perf_cpu_map__new(map);
1318 * If the map contains no CPU, consider the current CPU to
1319 * be the first online CPU in the cache domain else use the
1320 * first online CPU of the cache domain as the ID.
1322 if (perf_cpu_map__has_any_cpu_or_is_empty(cpu_map))
1325 id = perf_cpu_map__cpu(cpu_map, 0).cpu;
1327 /* Free the perf_cpu_map used to find the cache ID */
1328 perf_cpu_map__put(cpu_map);
1334 * cpu__get_cache_id - Returns 0 if successful in populating the
1335 * cache level and cache id. Cache level is read from
1336 * /sys/devices/system/cpu/cpuX/cache/indexY/level where as cache instance ID
1337 * is the first CPU reported by
1338 * /sys/devices/system/cpu/cpuX/cache/indexY/shared_cpu_list
1340 static int cpu__get_cache_details(struct perf_cpu cpu, struct perf_cache *cache)
1343 u32 cache_level = stat_config.aggr_level;
1344 struct cpu_cache_level caches[MAX_CACHE_LVL];
1345 u32 i = 0, caches_cnt = 0;
1347 cache->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1350 ret = build_caches_for_cpu(cpu.cpu, caches, &caches_cnt);
1353 * If caches_cnt is not 0, cpu_cache_level data
1354 * was allocated when building the topology.
1355 * Free the allocated data before returning.
1367 * Save the data for the highest level if no
1368 * level was specified by the user.
1370 if (cache_level > MAX_CACHE_LVL) {
1371 int max_level_index = 0;
1373 for (i = 1; i < caches_cnt; ++i) {
1374 if (caches[i].level > caches[max_level_index].level)
1375 max_level_index = i;
1378 cache->cache_lvl = caches[max_level_index].level;
1379 cache->cache = cpu__get_cache_id_from_map(cpu, caches[max_level_index].map);
1381 /* Reset i to 0 to free entire caches[] */
1386 for (i = 0; i < caches_cnt; ++i) {
1387 if (caches[i].level == cache_level) {
1388 cache->cache_lvl = cache_level;
1389 cache->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1392 cpu_cache_level__free(&caches[i]);
1397 * Free all the allocated cpu_cache_level data.
1399 while (i < caches_cnt)
1400 cpu_cache_level__free(&caches[i++]);
1406 * aggr_cpu_id__cache - Create an aggr_cpu_id with cache instache ID, cache
1407 * level, die and socket populated with the cache instache ID, cache level,
1408 * die and socket for cpu. The function signature is compatible with
1409 * aggr_cpu_id_get_t.
1411 static struct aggr_cpu_id aggr_cpu_id__cache(struct perf_cpu cpu, void *data)
1414 struct aggr_cpu_id id;
1415 struct perf_cache cache;
1417 id = aggr_cpu_id__die(cpu, data);
1418 if (aggr_cpu_id__is_empty(&id))
1421 ret = cpu__get_cache_details(cpu, &cache);
1425 id.cache_lvl = cache.cache_lvl;
1426 id.cache = cache.cache;
1430 static const char *const aggr_mode__string[] = {
1431 [AGGR_CORE] = "core",
1432 [AGGR_CACHE] = "cache",
1433 [AGGR_CLUSTER] = "cluster",
1435 [AGGR_GLOBAL] = "global",
1436 [AGGR_NODE] = "node",
1437 [AGGR_NONE] = "none",
1438 [AGGR_SOCKET] = "socket",
1439 [AGGR_THREAD] = "thread",
1440 [AGGR_UNSET] = "unset",
1443 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1444 struct perf_cpu cpu)
1446 return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1449 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1450 struct perf_cpu cpu)
1452 return aggr_cpu_id__die(cpu, /*data=*/NULL);
1455 static struct aggr_cpu_id perf_stat__get_cache_id(struct perf_stat_config *config __maybe_unused,
1456 struct perf_cpu cpu)
1458 return aggr_cpu_id__cache(cpu, /*data=*/NULL);
1461 static struct aggr_cpu_id perf_stat__get_cluster(struct perf_stat_config *config __maybe_unused,
1462 struct perf_cpu cpu)
1464 return aggr_cpu_id__cluster(cpu, /*data=*/NULL);
1467 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1468 struct perf_cpu cpu)
1470 return aggr_cpu_id__core(cpu, /*data=*/NULL);
1473 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1474 struct perf_cpu cpu)
1476 return aggr_cpu_id__node(cpu, /*data=*/NULL);
1479 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1480 struct perf_cpu cpu)
1482 return aggr_cpu_id__global(cpu, /*data=*/NULL);
1485 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1486 struct perf_cpu cpu)
1488 return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1491 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1492 aggr_get_id_t get_id, struct perf_cpu cpu)
1494 struct aggr_cpu_id id;
1496 /* per-process mode - should use global aggr mode */
1498 return get_id(config, cpu);
1500 if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1501 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1503 id = config->cpus_aggr_map->map[cpu.cpu];
1507 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1508 struct perf_cpu cpu)
1510 return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1513 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1514 struct perf_cpu cpu)
1516 return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1519 static struct aggr_cpu_id perf_stat__get_cluster_cached(struct perf_stat_config *config,
1520 struct perf_cpu cpu)
1522 return perf_stat__get_aggr(config, perf_stat__get_cluster, cpu);
1525 static struct aggr_cpu_id perf_stat__get_cache_id_cached(struct perf_stat_config *config,
1526 struct perf_cpu cpu)
1528 return perf_stat__get_aggr(config, perf_stat__get_cache_id, cpu);
1531 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1532 struct perf_cpu cpu)
1534 return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1537 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1538 struct perf_cpu cpu)
1540 return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1543 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1544 struct perf_cpu cpu)
1546 return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1549 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1550 struct perf_cpu cpu)
1552 return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1555 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1557 switch (aggr_mode) {
1559 return aggr_cpu_id__socket;
1561 return aggr_cpu_id__die;
1563 return aggr_cpu_id__cluster;
1565 return aggr_cpu_id__cache;
1567 return aggr_cpu_id__core;
1569 return aggr_cpu_id__node;
1571 return aggr_cpu_id__cpu;
1573 return aggr_cpu_id__global;
1582 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1584 switch (aggr_mode) {
1586 return perf_stat__get_socket_cached;
1588 return perf_stat__get_die_cached;
1590 return perf_stat__get_cluster_cached;
1592 return perf_stat__get_cache_id_cached;
1594 return perf_stat__get_core_cached;
1596 return perf_stat__get_node_cached;
1598 return perf_stat__get_cpu_cached;
1600 return perf_stat__get_global_cached;
1609 static int perf_stat_init_aggr_mode(void)
1612 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1615 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1616 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1617 get_id, /*data=*/NULL, needs_sort);
1618 if (!stat_config.aggr_map) {
1619 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1622 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1625 if (stat_config.aggr_mode == AGGR_THREAD) {
1626 nr = perf_thread_map__nr(evsel_list->core.threads);
1627 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1628 if (stat_config.aggr_map == NULL)
1631 for (int s = 0; s < nr; s++) {
1632 struct aggr_cpu_id id = aggr_cpu_id__empty();
1635 stat_config.aggr_map->map[s] = id;
1641 * The evsel_list->cpus is the base we operate on,
1642 * taking the highest cpu number to be the size of
1643 * the aggregation translate cpumap.
1645 if (!perf_cpu_map__has_any_cpu_or_is_empty(evsel_list->core.user_requested_cpus))
1646 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1649 stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1650 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1653 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1656 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1657 "cpu_aggr_map refcnt unbalanced\n");
1662 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1664 if (map && refcount_dec_and_test(&map->refcnt))
1665 cpu_aggr_map__delete(map);
1668 static void perf_stat__exit_aggr_mode(void)
1670 cpu_aggr_map__put(stat_config.aggr_map);
1671 cpu_aggr_map__put(stat_config.cpus_aggr_map);
1672 stat_config.aggr_map = NULL;
1673 stat_config.cpus_aggr_map = NULL;
1676 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1678 struct perf_env *env = data;
1679 struct aggr_cpu_id id = aggr_cpu_id__empty();
1682 id.socket = env->cpu[cpu.cpu].socket_id;
1687 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1689 struct perf_env *env = data;
1690 struct aggr_cpu_id id = aggr_cpu_id__empty();
1692 if (cpu.cpu != -1) {
1694 * die_id is relative to socket, so start
1695 * with the socket ID and then add die to
1698 id.socket = env->cpu[cpu.cpu].socket_id;
1699 id.die = env->cpu[cpu.cpu].die_id;
1705 static void perf_env__get_cache_id_for_cpu(struct perf_cpu cpu, struct perf_env *env,
1706 u32 cache_level, struct aggr_cpu_id *id)
1709 int caches_cnt = env->caches_cnt;
1710 struct cpu_cache_level *caches = env->caches;
1712 id->cache_lvl = (cache_level > MAX_CACHE_LVL) ? 0 : cache_level;
1718 for (i = caches_cnt - 1; i > -1; --i) {
1719 struct perf_cpu_map *cpu_map;
1720 int map_contains_cpu;
1723 * If user has not specified a level, find the fist level with
1724 * the cpu in the map. Since building the map is expensive, do
1725 * this only if levels match.
1727 if (cache_level <= MAX_CACHE_LVL && caches[i].level != cache_level)
1730 cpu_map = perf_cpu_map__new(caches[i].map);
1731 map_contains_cpu = perf_cpu_map__idx(cpu_map, cpu);
1732 perf_cpu_map__put(cpu_map);
1734 if (map_contains_cpu != -1) {
1735 id->cache_lvl = caches[i].level;
1736 id->cache = cpu__get_cache_id_from_map(cpu, caches[i].map);
1742 static struct aggr_cpu_id perf_env__get_cache_aggr_by_cpu(struct perf_cpu cpu,
1745 struct perf_env *env = data;
1746 struct aggr_cpu_id id = aggr_cpu_id__empty();
1748 if (cpu.cpu != -1) {
1749 u32 cache_level = (perf_stat.aggr_level) ?: stat_config.aggr_level;
1751 id.socket = env->cpu[cpu.cpu].socket_id;
1752 id.die = env->cpu[cpu.cpu].die_id;
1753 perf_env__get_cache_id_for_cpu(cpu, env, cache_level, &id);
1759 static struct aggr_cpu_id perf_env__get_cluster_aggr_by_cpu(struct perf_cpu cpu,
1762 struct perf_env *env = data;
1763 struct aggr_cpu_id id = aggr_cpu_id__empty();
1765 if (cpu.cpu != -1) {
1766 id.socket = env->cpu[cpu.cpu].socket_id;
1767 id.die = env->cpu[cpu.cpu].die_id;
1768 id.cluster = env->cpu[cpu.cpu].cluster_id;
1774 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1776 struct perf_env *env = data;
1777 struct aggr_cpu_id id = aggr_cpu_id__empty();
1779 if (cpu.cpu != -1) {
1781 * core_id is relative to socket, die and cluster, we need a
1782 * global id. So we set socket, die id, cluster id and core id.
1784 id.socket = env->cpu[cpu.cpu].socket_id;
1785 id.die = env->cpu[cpu.cpu].die_id;
1786 id.cluster = env->cpu[cpu.cpu].cluster_id;
1787 id.core = env->cpu[cpu.cpu].core_id;
1793 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1795 struct perf_env *env = data;
1796 struct aggr_cpu_id id = aggr_cpu_id__empty();
1798 if (cpu.cpu != -1) {
1800 * core_id is relative to socket and die,
1801 * we need a global id. So we set
1802 * socket, die id and core id
1804 id.socket = env->cpu[cpu.cpu].socket_id;
1805 id.die = env->cpu[cpu.cpu].die_id;
1806 id.core = env->cpu[cpu.cpu].core_id;
1813 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1815 struct aggr_cpu_id id = aggr_cpu_id__empty();
1817 id.node = perf_env__numa_node(data, cpu);
1821 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1822 void *data __maybe_unused)
1824 struct aggr_cpu_id id = aggr_cpu_id__empty();
1826 /* it always aggregates to the cpu 0 */
1827 id.cpu = (struct perf_cpu){ .cpu = 0 };
1831 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1832 struct perf_cpu cpu)
1834 return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1836 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1837 struct perf_cpu cpu)
1839 return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1842 static struct aggr_cpu_id perf_stat__get_cluster_file(struct perf_stat_config *config __maybe_unused,
1843 struct perf_cpu cpu)
1845 return perf_env__get_cluster_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1848 static struct aggr_cpu_id perf_stat__get_cache_file(struct perf_stat_config *config __maybe_unused,
1849 struct perf_cpu cpu)
1851 return perf_env__get_cache_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1854 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1855 struct perf_cpu cpu)
1857 return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1860 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1861 struct perf_cpu cpu)
1863 return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1866 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1867 struct perf_cpu cpu)
1869 return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1872 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1873 struct perf_cpu cpu)
1875 return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1878 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1880 switch (aggr_mode) {
1882 return perf_env__get_socket_aggr_by_cpu;
1884 return perf_env__get_die_aggr_by_cpu;
1886 return perf_env__get_cluster_aggr_by_cpu;
1888 return perf_env__get_cache_aggr_by_cpu;
1890 return perf_env__get_core_aggr_by_cpu;
1892 return perf_env__get_node_aggr_by_cpu;
1894 return perf_env__get_global_aggr_by_cpu;
1896 return perf_env__get_cpu_aggr_by_cpu;
1905 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1907 switch (aggr_mode) {
1909 return perf_stat__get_socket_file;
1911 return perf_stat__get_die_file;
1913 return perf_stat__get_cluster_file;
1915 return perf_stat__get_cache_file;
1917 return perf_stat__get_core_file;
1919 return perf_stat__get_node_file;
1921 return perf_stat__get_global_file;
1923 return perf_stat__get_cpu_file;
1932 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1934 struct perf_env *env = &st->session->header.env;
1935 aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1936 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1938 if (stat_config.aggr_mode == AGGR_THREAD) {
1939 int nr = perf_thread_map__nr(evsel_list->core.threads);
1941 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1942 if (stat_config.aggr_map == NULL)
1945 for (int s = 0; s < nr; s++) {
1946 struct aggr_cpu_id id = aggr_cpu_id__empty();
1949 stat_config.aggr_map->map[s] = id;
1957 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1958 get_id, env, needs_sort);
1959 if (!stat_config.aggr_map) {
1960 pr_err("cannot build %s map\n", aggr_mode__string[stat_config.aggr_mode]);
1963 stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1968 * Add default attributes, if there were no attributes specified or
1969 * if -d/--detailed, -d -d or -d -d -d is used:
1971 static int add_default_attributes(void)
1973 struct perf_event_attr default_attrs0[] = {
1975 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1976 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1977 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1978 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1980 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1982 struct perf_event_attr frontend_attrs[] = {
1983 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1985 struct perf_event_attr backend_attrs[] = {
1986 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1988 struct perf_event_attr default_attrs1[] = {
1989 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1990 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1991 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1996 * Detailed stats (-d), covering the L1 and last level data caches:
1998 struct perf_event_attr detailed_attrs[] = {
2000 { .type = PERF_TYPE_HW_CACHE,
2002 PERF_COUNT_HW_CACHE_L1D << 0 |
2003 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2004 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2006 { .type = PERF_TYPE_HW_CACHE,
2008 PERF_COUNT_HW_CACHE_L1D << 0 |
2009 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2010 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2012 { .type = PERF_TYPE_HW_CACHE,
2014 PERF_COUNT_HW_CACHE_LL << 0 |
2015 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2016 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2018 { .type = PERF_TYPE_HW_CACHE,
2020 PERF_COUNT_HW_CACHE_LL << 0 |
2021 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2022 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2026 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2028 struct perf_event_attr very_detailed_attrs[] = {
2030 { .type = PERF_TYPE_HW_CACHE,
2032 PERF_COUNT_HW_CACHE_L1I << 0 |
2033 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2034 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2036 { .type = PERF_TYPE_HW_CACHE,
2038 PERF_COUNT_HW_CACHE_L1I << 0 |
2039 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2040 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2042 { .type = PERF_TYPE_HW_CACHE,
2044 PERF_COUNT_HW_CACHE_DTLB << 0 |
2045 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2046 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2048 { .type = PERF_TYPE_HW_CACHE,
2050 PERF_COUNT_HW_CACHE_DTLB << 0 |
2051 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2052 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2054 { .type = PERF_TYPE_HW_CACHE,
2056 PERF_COUNT_HW_CACHE_ITLB << 0 |
2057 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2058 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2060 { .type = PERF_TYPE_HW_CACHE,
2062 PERF_COUNT_HW_CACHE_ITLB << 0 |
2063 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2064 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2069 * Very, very detailed stats (-d -d -d), adding prefetch events:
2071 struct perf_event_attr very_very_detailed_attrs[] = {
2073 { .type = PERF_TYPE_HW_CACHE,
2075 PERF_COUNT_HW_CACHE_L1D << 0 |
2076 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2077 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2079 { .type = PERF_TYPE_HW_CACHE,
2081 PERF_COUNT_HW_CACHE_L1D << 0 |
2082 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2083 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2086 struct perf_event_attr default_null_attrs[] = {};
2087 const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2089 /* Set attrs if no event is selected and !null_run: */
2090 if (stat_config.null_run)
2093 if (transaction_run) {
2094 /* Handle -T as -M transaction. Once platform specific metrics
2095 * support has been added to the json files, all architectures
2096 * will use this approach. To determine transaction support
2097 * on an architecture test for such a metric name.
2099 if (!metricgroup__has_metric(pmu, "transaction")) {
2100 pr_err("Missing transaction metrics\n");
2103 return metricgroup__parse_groups(evsel_list, pmu, "transaction",
2104 stat_config.metric_no_group,
2105 stat_config.metric_no_merge,
2106 stat_config.metric_no_threshold,
2107 stat_config.user_requested_cpu_list,
2108 stat_config.system_wide,
2109 &stat_config.metric_events);
2115 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2116 pr_err("freeze_on_smi is not supported.\n");
2121 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2122 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2128 if (!metricgroup__has_metric(pmu, "smi")) {
2129 pr_err("Missing smi metrics\n");
2133 if (!force_metric_only)
2134 stat_config.metric_only = true;
2136 return metricgroup__parse_groups(evsel_list, pmu, "smi",
2137 stat_config.metric_no_group,
2138 stat_config.metric_no_merge,
2139 stat_config.metric_no_threshold,
2140 stat_config.user_requested_cpu_list,
2141 stat_config.system_wide,
2142 &stat_config.metric_events);
2146 unsigned int max_level = metricgroups__topdown_max_level();
2147 char str[] = "TopdownL1";
2149 if (!force_metric_only)
2150 stat_config.metric_only = true;
2153 pr_err("Topdown requested but the topdown metric groups aren't present.\n"
2154 "(See perf list the metric groups have names like TopdownL1)\n");
2157 if (stat_config.topdown_level > max_level) {
2158 pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
2160 } else if (!stat_config.topdown_level)
2161 stat_config.topdown_level = 1;
2163 if (!stat_config.interval && !stat_config.metric_only) {
2164 fprintf(stat_config.output,
2165 "Topdown accuracy may decrease when measuring long periods.\n"
2166 "Please print the result regularly, e.g. -I1000\n");
2168 str[8] = stat_config.topdown_level + '0';
2169 if (metricgroup__parse_groups(evsel_list,
2171 /*metric_no_group=*/false,
2172 /*metric_no_merge=*/false,
2173 /*metric_no_threshold=*/true,
2174 stat_config.user_requested_cpu_list,
2175 stat_config.system_wide,
2176 &stat_config.metric_events) < 0)
2180 if (!stat_config.topdown_level)
2181 stat_config.topdown_level = 1;
2183 if (!evsel_list->core.nr_entries) {
2184 /* No events so add defaults. */
2185 if (target__has_cpu(&target))
2186 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2188 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2190 if (perf_pmus__have_event("cpu", "stalled-cycles-frontend")) {
2191 if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
2194 if (perf_pmus__have_event("cpu", "stalled-cycles-backend")) {
2195 if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2198 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2201 * Add TopdownL1 metrics if they exist. To minimize
2202 * multiplexing, don't request threshold computation.
2204 if (metricgroup__has_metric(pmu, "Default")) {
2205 struct evlist *metric_evlist = evlist__new();
2206 struct evsel *metric_evsel;
2211 if (metricgroup__parse_groups(metric_evlist, pmu, "Default",
2212 /*metric_no_group=*/false,
2213 /*metric_no_merge=*/false,
2214 /*metric_no_threshold=*/true,
2215 stat_config.user_requested_cpu_list,
2216 stat_config.system_wide,
2217 &stat_config.metric_events) < 0)
2220 evlist__for_each_entry(metric_evlist, metric_evsel) {
2221 metric_evsel->skippable = true;
2222 metric_evsel->default_metricgroup = true;
2224 evlist__splice_list_tail(evsel_list, &metric_evlist->core.entries);
2225 evlist__delete(metric_evlist);
2228 /* Platform specific attrs */
2229 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2233 /* Detailed events get appended to the event list: */
2235 if (detailed_run < 1)
2238 /* Append detailed run extra attributes: */
2239 if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2242 if (detailed_run < 2)
2245 /* Append very detailed run extra attributes: */
2246 if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2249 if (detailed_run < 3)
2252 /* Append very, very detailed run extra attributes: */
2253 return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2256 static const char * const stat_record_usage[] = {
2257 "perf stat record [<options>]",
2261 static void init_features(struct perf_session *session)
2265 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2266 perf_header__set_feat(&session->header, feat);
2268 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2269 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2270 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2271 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2272 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2275 static int __cmd_record(int argc, const char **argv)
2277 struct perf_session *session;
2278 struct perf_data *data = &perf_stat.data;
2280 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2281 PARSE_OPT_STOP_AT_NON_OPTION);
2284 data->path = output_name;
2286 if (stat_config.run_count != 1 || forever) {
2287 pr_err("Cannot use -r option with perf stat record.\n");
2291 session = perf_session__new(data, NULL);
2292 if (IS_ERR(session)) {
2293 pr_err("Perf session creation failed\n");
2294 return PTR_ERR(session);
2297 init_features(session);
2299 session->evlist = evsel_list;
2300 perf_stat.session = session;
2301 perf_stat.record = true;
2305 static int process_stat_round_event(struct perf_session *session,
2306 union perf_event *event)
2308 struct perf_record_stat_round *stat_round = &event->stat_round;
2309 struct timespec tsh, *ts = NULL;
2310 const char **argv = session->header.env.cmdline_argv;
2311 int argc = session->header.env.nr_cmdline;
2315 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2316 update_stats(&walltime_nsecs_stats, stat_round->time);
2318 if (stat_config.interval && stat_round->time) {
2319 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2320 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2324 print_counters(ts, argc, argv);
2329 int process_stat_config_event(struct perf_session *session,
2330 union perf_event *event)
2332 struct perf_tool *tool = session->tool;
2333 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2335 perf_event__read_stat_config(&stat_config, &event->stat_config);
2337 if (perf_cpu_map__has_any_cpu_or_is_empty(st->cpus)) {
2338 if (st->aggr_mode != AGGR_UNSET)
2339 pr_warning("warning: processing task data, aggregation mode not set\n");
2340 } else if (st->aggr_mode != AGGR_UNSET) {
2341 stat_config.aggr_mode = st->aggr_mode;
2344 if (perf_stat.data.is_pipe)
2345 perf_stat_init_aggr_mode();
2347 perf_stat_init_aggr_mode_file(st);
2349 if (stat_config.aggr_map) {
2350 int nr_aggr = stat_config.aggr_map->nr;
2352 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2353 pr_err("cannot allocate aggr counts\n");
2360 static int set_maps(struct perf_stat *st)
2362 if (!st->cpus || !st->threads)
2365 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2368 perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2370 if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2373 st->maps_allocated = true;
2378 int process_thread_map_event(struct perf_session *session,
2379 union perf_event *event)
2381 struct perf_tool *tool = session->tool;
2382 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2385 pr_warning("Extra thread map event, ignoring.\n");
2389 st->threads = thread_map__new_event(&event->thread_map);
2393 return set_maps(st);
2397 int process_cpu_map_event(struct perf_session *session,
2398 union perf_event *event)
2400 struct perf_tool *tool = session->tool;
2401 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2402 struct perf_cpu_map *cpus;
2405 pr_warning("Extra cpu map event, ignoring.\n");
2409 cpus = cpu_map__new_data(&event->cpu_map.data);
2414 return set_maps(st);
2417 static const char * const stat_report_usage[] = {
2418 "perf stat report [<options>]",
2422 static struct perf_stat perf_stat = {
2424 .attr = perf_event__process_attr,
2425 .event_update = perf_event__process_event_update,
2426 .thread_map = process_thread_map_event,
2427 .cpu_map = process_cpu_map_event,
2428 .stat_config = process_stat_config_event,
2429 .stat = perf_event__process_stat_event,
2430 .stat_round = process_stat_round_event,
2432 .aggr_mode = AGGR_UNSET,
2436 static int __cmd_report(int argc, const char **argv)
2438 struct perf_session *session;
2439 const struct option options[] = {
2440 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2441 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2442 "aggregate counts per processor socket", AGGR_SOCKET),
2443 OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2444 "aggregate counts per processor die", AGGR_DIE),
2445 OPT_SET_UINT(0, "per-cluster", &perf_stat.aggr_mode,
2446 "aggregate counts perf processor cluster", AGGR_CLUSTER),
2447 OPT_CALLBACK_OPTARG(0, "per-cache", &perf_stat.aggr_mode, &perf_stat.aggr_level,
2449 "aggregate count at this cache level (Default: LLC)",
2451 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2452 "aggregate counts per physical processor core", AGGR_CORE),
2453 OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2454 "aggregate counts per numa node", AGGR_NODE),
2455 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2456 "disable CPU count aggregation", AGGR_NONE),
2462 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2464 if (!input_name || !strlen(input_name)) {
2465 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2468 input_name = "perf.data";
2471 perf_stat.data.path = input_name;
2472 perf_stat.data.mode = PERF_DATA_MODE_READ;
2474 session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2475 if (IS_ERR(session))
2476 return PTR_ERR(session);
2478 perf_stat.session = session;
2479 stat_config.output = stderr;
2480 evlist__delete(evsel_list);
2481 evsel_list = session->evlist;
2483 ret = perf_session__process_events(session);
2487 perf_session__delete(session);
2491 static void setup_system_wide(int forks)
2494 * Make system wide (-a) the default target if
2495 * no target was specified and one of following
2496 * conditions is met:
2498 * - there's no workload specified
2499 * - there is workload specified but all requested
2500 * events are system wide events
2502 if (!target__none(&target))
2506 target.system_wide = true;
2508 struct evsel *counter;
2510 evlist__for_each_entry(evsel_list, counter) {
2511 if (!counter->core.requires_cpu &&
2512 !evsel__name_is(counter, "duration_time")) {
2517 if (evsel_list->core.nr_entries)
2518 target.system_wide = true;
2522 int cmd_stat(int argc, const char **argv)
2524 const char * const stat_usage[] = {
2525 "perf stat [<options>] [<command>]",
2528 int status = -EINVAL, run_idx, err;
2530 FILE *output = stderr;
2531 unsigned int interval, timeout;
2532 const char * const stat_subcommands[] = { "record", "report" };
2533 char errbuf[BUFSIZ];
2535 setlocale(LC_ALL, "");
2537 evsel_list = evlist__new();
2538 if (evsel_list == NULL)
2541 parse_events__shrink_config_terms();
2543 /* String-parsing callback-based options would segfault when negated */
2544 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2545 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2546 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2548 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2549 (const char **) stat_usage,
2550 PARSE_OPT_STOP_AT_NON_OPTION);
2552 if (stat_config.csv_sep) {
2553 stat_config.csv_output = true;
2554 if (!strcmp(stat_config.csv_sep, "\\t"))
2555 stat_config.csv_sep = "\t";
2557 stat_config.csv_sep = DEFAULT_SEPARATOR;
2559 if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2560 argc = __cmd_record(argc, argv);
2563 } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2564 return __cmd_report(argc, argv);
2566 interval = stat_config.interval;
2567 timeout = stat_config.timeout;
2570 * For record command the -o is already taken care of.
2572 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2575 if (output_name && output_fd) {
2576 fprintf(stderr, "cannot use both --output and --log-fd\n");
2577 parse_options_usage(stat_usage, stat_options, "o", 1);
2578 parse_options_usage(NULL, stat_options, "log-fd", 0);
2582 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2583 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2587 if (stat_config.metric_only && stat_config.run_count > 1) {
2588 fprintf(stderr, "--metric-only is not supported with -r\n");
2592 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2593 fprintf(stderr, "--table is only supported with -r\n");
2594 parse_options_usage(stat_usage, stat_options, "r", 1);
2595 parse_options_usage(NULL, stat_options, "table", 0);
2599 if (output_fd < 0) {
2600 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2601 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2605 if (!output && !quiet) {
2607 mode = append_file ? "a" : "w";
2609 output = fopen(output_name, mode);
2611 perror("failed to create output file");
2614 if (!stat_config.json_output) {
2615 clock_gettime(CLOCK_REALTIME, &tm);
2616 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2618 } else if (output_fd > 0) {
2619 mode = append_file ? "a" : "w";
2620 output = fdopen(output_fd, mode);
2622 perror("Failed opening logfd");
2627 if (stat_config.interval_clear && !isatty(fileno(output))) {
2628 fprintf(stderr, "--interval-clear does not work with output\n");
2629 parse_options_usage(stat_usage, stat_options, "o", 1);
2630 parse_options_usage(NULL, stat_options, "log-fd", 0);
2631 parse_options_usage(NULL, stat_options, "interval-clear", 0);
2635 stat_config.output = output;
2638 * let the spreadsheet do the pretty-printing
2640 if (stat_config.csv_output) {
2641 /* User explicitly passed -B? */
2642 if (big_num_opt == 1) {
2643 fprintf(stderr, "-B option not supported with -x\n");
2644 parse_options_usage(stat_usage, stat_options, "B", 1);
2645 parse_options_usage(NULL, stat_options, "x", 1);
2647 } else /* Nope, so disable big number formatting */
2648 stat_config.big_num = false;
2649 } else if (big_num_opt == 0) /* User passed --no-big-num */
2650 stat_config.big_num = false;
2652 err = target__validate(&target);
2654 target__strerror(&target, err, errbuf, BUFSIZ);
2655 pr_warning("%s\n", errbuf);
2658 setup_system_wide(argc);
2661 * Display user/system times only for single
2662 * run and when there's specified tracee.
2664 if ((stat_config.run_count == 1) && target__none(&target))
2665 stat_config.ru_display = true;
2667 if (stat_config.run_count < 0) {
2668 pr_err("Run count must be a positive number\n");
2669 parse_options_usage(stat_usage, stat_options, "r", 1);
2671 } else if (stat_config.run_count == 0) {
2673 stat_config.run_count = 1;
2676 if (stat_config.walltime_run_table) {
2677 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2678 if (!stat_config.walltime_run) {
2679 pr_err("failed to setup -r option");
2684 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2685 !target__has_task(&target)) {
2686 if (!target.system_wide || target.cpu_list) {
2687 fprintf(stderr, "The --per-thread option is only "
2688 "available when monitoring via -p -t -a "
2689 "options or only --per-thread.\n");
2690 parse_options_usage(NULL, stat_options, "p", 1);
2691 parse_options_usage(NULL, stat_options, "t", 1);
2697 * no_aggr, cgroup are for system-wide only
2698 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2700 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2701 stat_config.aggr_mode != AGGR_THREAD) ||
2702 (nr_cgroups || stat_config.cgroup_list)) &&
2703 !target__has_cpu(&target)) {
2704 fprintf(stderr, "both cgroup and no-aggregation "
2705 "modes only available in system-wide mode\n");
2707 parse_options_usage(stat_usage, stat_options, "G", 1);
2708 parse_options_usage(NULL, stat_options, "A", 1);
2709 parse_options_usage(NULL, stat_options, "a", 1);
2710 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2714 if (stat_config.iostat_run) {
2715 status = iostat_prepare(evsel_list, &stat_config);
2718 if (iostat_mode == IOSTAT_LIST) {
2719 iostat_list(evsel_list, &stat_config);
2721 } else if (verbose > 0)
2722 iostat_list(evsel_list, &stat_config);
2723 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2724 target.system_wide = true;
2727 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2728 target.per_thread = true;
2730 stat_config.system_wide = target.system_wide;
2731 if (target.cpu_list) {
2732 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2733 if (!stat_config.user_requested_cpu_list) {
2740 * Metric parsing needs to be delayed as metrics may optimize events
2741 * knowing the target is system-wide.
2744 const char *pmu = parse_events_option_args.pmu_filter ?: "all";
2745 int ret = metricgroup__parse_groups(evsel_list, pmu, metrics,
2746 stat_config.metric_no_group,
2747 stat_config.metric_no_merge,
2748 stat_config.metric_no_threshold,
2749 stat_config.user_requested_cpu_list,
2750 stat_config.system_wide,
2751 &stat_config.metric_events);
2760 if (add_default_attributes())
2763 if (stat_config.cgroup_list) {
2764 if (nr_cgroups > 0) {
2765 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2766 parse_options_usage(stat_usage, stat_options, "G", 1);
2767 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2771 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2772 &stat_config.metric_events, true) < 0) {
2773 parse_options_usage(stat_usage, stat_options,
2774 "for-each-cgroup", 0);
2779 evlist__warn_user_requested_cpus(evsel_list, target.cpu_list);
2781 if (evlist__create_maps(evsel_list, &target) < 0) {
2782 if (target__has_task(&target)) {
2783 pr_err("Problems finding threads of monitor\n");
2784 parse_options_usage(stat_usage, stat_options, "p", 1);
2785 parse_options_usage(NULL, stat_options, "t", 1);
2786 } else if (target__has_cpu(&target)) {
2787 perror("failed to parse CPUs map");
2788 parse_options_usage(stat_usage, stat_options, "C", 1);
2789 parse_options_usage(NULL, stat_options, "a", 1);
2794 evlist__check_cpu_maps(evsel_list);
2797 * Initialize thread_map with comm names,
2798 * so we could print it out on output.
2800 if (stat_config.aggr_mode == AGGR_THREAD) {
2801 thread_map__read_comms(evsel_list->core.threads);
2804 if (stat_config.aggr_mode == AGGR_NODE)
2805 cpu__setup_cpunode_map();
2807 if (stat_config.times && interval)
2808 interval_count = true;
2809 else if (stat_config.times && !interval) {
2810 pr_err("interval-count option should be used together with "
2811 "interval-print.\n");
2812 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2813 parse_options_usage(stat_usage, stat_options, "I", 1);
2817 if (timeout && timeout < 100) {
2819 pr_err("timeout must be >= 10ms.\n");
2820 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2823 pr_warning("timeout < 100ms. "
2824 "The overhead percentage could be high in some cases. "
2825 "Please proceed with caution.\n");
2827 if (timeout && interval) {
2828 pr_err("timeout option is not supported with interval-print.\n");
2829 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2830 parse_options_usage(stat_usage, stat_options, "I", 1);
2834 if (perf_stat_init_aggr_mode())
2837 if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2841 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2842 * while avoiding that older tools show confusing messages.
2844 * However for pipe sessions we need to keep it zero,
2845 * because script's perf_evsel__check_attr is triggered
2846 * by attr->sample_type != 0, and we can't run it on
2849 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2852 * We dont want to block the signals - that would cause
2853 * child tasks to inherit that and Ctrl-C would not work.
2854 * What we want is for Ctrl-C to work in the exec()-ed
2855 * task, but being ignored by perf stat itself:
2859 signal(SIGINT, skip_signal);
2860 signal(SIGCHLD, skip_signal);
2861 signal(SIGALRM, skip_signal);
2862 signal(SIGABRT, skip_signal);
2864 if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2867 /* Enable ignoring missing threads when -p option is defined. */
2868 evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2870 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2871 if (stat_config.run_count != 1 && verbose > 0)
2872 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2876 evlist__reset_prev_raw_counts(evsel_list);
2878 status = run_perf_stat(argc, argv, run_idx);
2879 if (forever && status != -1 && !interval) {
2880 print_counters(NULL, argc, argv);
2881 perf_stat__reset_stats();
2885 if (!forever && status != -1 && (!interval || stat_config.summary)) {
2886 if (stat_config.run_count > 1)
2887 evlist__copy_res_stats(&stat_config, evsel_list);
2888 print_counters(NULL, argc, argv);
2891 evlist__finalize_ctlfd(evsel_list);
2895 * We synthesize the kernel mmap record just so that older tools
2896 * don't emit warnings about not being able to resolve symbols
2897 * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2898 * a saner message about no samples being in the perf.data file.
2900 * This also serves to suppress a warning about f_header.data.size == 0
2901 * in header.c at the moment 'perf stat record' gets introduced, which
2902 * is not really needed once we start adding the stat specific PERF_RECORD_
2903 * records, but the need to suppress the kptr_restrict messages in older
2904 * tools remain -acme
2906 int fd = perf_data__fd(&perf_stat.data);
2908 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2909 process_synthesized_event,
2910 &perf_stat.session->machines.host);
2912 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2913 "older tools may produce warnings about this file\n.");
2917 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2918 pr_err("failed to write stat round event\n");
2921 if (!perf_stat.data.is_pipe) {
2922 perf_stat.session->header.data_size += perf_stat.bytes_written;
2923 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2926 evlist__close(evsel_list);
2927 perf_session__delete(perf_stat.session);
2930 perf_stat__exit_aggr_mode();
2931 evlist__free_stats(evsel_list);
2933 if (stat_config.iostat_run)
2934 iostat_release(evsel_list);
2936 zfree(&stat_config.walltime_run);
2937 zfree(&stat_config.user_requested_cpu_list);
2939 if (smi_cost && smi_reset)
2940 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2942 evlist__delete(evsel_list);
2944 metricgroup__rblist_exit(&stat_config.metric_events);
2945 evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);