4 * Builtin stat command: Give a precise performance counters summary
5 * overview about any workload, CPU or specific PID.
9 $ perf stat ./hackbench 10
13 Performance counter stats for './hackbench 10':
15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
27 0.154822978 seconds time elapsed
30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
32 * Improvements and fixes by:
34 * Arjan van de Ven <arjan@linux.intel.com>
35 * Yanmin Zhang <yanmin.zhang@intel.com>
36 * Wu Fengguang <fengguang.wu@intel.com>
37 * Mike Galbraith <efault@gmx.de>
38 * Paul Mackerras <paulus@samba.org>
39 * Jaswinder Singh Rajput <jaswinder@kernel.org>
41 * Released under the GPL v2. (and only v2, not any later version)
46 #include "util/cgroup.h"
47 #include "util/util.h"
48 #include <subcmd/parse-options.h>
49 #include "util/parse-events.h"
51 #include "util/event.h"
52 #include "util/evlist.h"
53 #include "util/evsel.h"
54 #include "util/debug.h"
55 #include "util/drv_configs.h"
56 #include "util/color.h"
57 #include "util/stat.h"
58 #include "util/header.h"
59 #include "util/cpumap.h"
60 #include "util/thread.h"
61 #include "util/thread_map.h"
62 #include "util/counts.h"
63 #include "util/group.h"
64 #include "util/session.h"
65 #include "util/tool.h"
66 #include "util/string2.h"
67 #include "util/metricgroup.h"
71 #include <linux/time64.h>
72 #include <api/fs/fs.h>
76 #include <sys/prctl.h>
80 #include <sys/types.h>
85 #include <sys/resource.h>
88 #include "sane_ctype.h"
90 #define DEFAULT_SEPARATOR " "
91 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
93 static void print_counters(struct timespec *ts, int argc, const char **argv);
95 /* Default events used for perf stat -T */
96 static const char *transaction_attrs = {
108 /* More limited version when the CPU does not have all events. */
109 static const char * transaction_limited_attrs = {
119 static const char * topdown_attrs[] = {
120 "topdown-total-slots",
121 "topdown-slots-retired",
122 "topdown-recovery-bubbles",
123 "topdown-fetch-bubbles",
124 "topdown-slots-issued",
128 static const char *smi_cost_attrs = {
136 static struct perf_evlist *evsel_list;
138 static struct target target = {
142 #define METRIC_ONLY_LEN 20
144 static volatile pid_t child_pid = -1;
145 static int detailed_run = 0;
146 static bool transaction_run;
147 static bool topdown_run = false;
148 static bool smi_cost = false;
149 static bool smi_reset = false;
150 static int big_num_opt = -1;
151 static bool group = false;
152 static const char *pre_cmd = NULL;
153 static const char *post_cmd = NULL;
154 static bool sync_run = false;
155 static bool forever = false;
156 static bool force_metric_only = false;
157 static struct timespec ref_time;
158 static bool append_file;
159 static bool interval_count;
160 static const char *output_name;
161 static int output_fd;
165 struct perf_data data;
166 struct perf_session *session;
168 struct perf_tool tool;
170 struct cpu_map *cpus;
171 struct thread_map *threads;
172 enum aggr_mode aggr_mode;
175 static struct perf_stat perf_stat;
176 #define STAT_RECORD perf_stat.record
178 static volatile int done = 0;
180 static struct perf_stat_config stat_config = {
181 .aggr_mode = AGGR_GLOBAL,
183 .unit_width = 4, /* strlen("unit") */
185 .metric_only_len = METRIC_ONLY_LEN,
186 .walltime_nsecs_stats = &walltime_nsecs_stats,
190 static inline void diff_timespec(struct timespec *r, struct timespec *a,
193 r->tv_sec = a->tv_sec - b->tv_sec;
194 if (a->tv_nsec < b->tv_nsec) {
195 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
198 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
202 static void perf_stat__reset_stats(void)
206 perf_evlist__reset_stats(evsel_list);
207 perf_stat__reset_shadow_stats();
209 for (i = 0; i < stat_config.stats_num; i++)
210 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
213 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
214 union perf_event *event,
215 struct perf_sample *sample __maybe_unused,
216 struct machine *machine __maybe_unused)
218 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
219 pr_err("failed to write perf data, error: %m\n");
223 perf_stat.bytes_written += event->header.size;
227 static int write_stat_round_event(u64 tm, u64 type)
229 return perf_event__synthesize_stat_round(NULL, tm, type,
230 process_synthesized_event,
234 #define WRITE_STAT_ROUND_EVENT(time, interval) \
235 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
237 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
240 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
241 struct perf_counts_values *count)
243 struct perf_sample_id *sid = SID(counter, cpu, thread);
245 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
246 process_synthesized_event, NULL);
250 * Read out the results of a single counter:
251 * do not aggregate counts across CPUs in system-wide mode
253 static int read_counter(struct perf_evsel *counter)
255 int nthreads = thread_map__nr(evsel_list->threads);
256 int ncpus, cpu, thread;
258 if (target__has_cpu(&target) && !target__has_per_thread(&target))
259 ncpus = perf_evsel__nr_cpus(counter);
263 if (!counter->supported)
266 if (counter->system_wide)
269 for (thread = 0; thread < nthreads; thread++) {
270 for (cpu = 0; cpu < ncpus; cpu++) {
271 struct perf_counts_values *count;
273 count = perf_counts(counter->counts, cpu, thread);
276 * The leader's group read loads data into its group members
277 * (via perf_evsel__read_counter) and sets threir count->loaded.
279 if (!count->loaded &&
280 perf_evsel__read_counter(counter, cpu, thread)) {
281 counter->counts->scaled = -1;
282 perf_counts(counter->counts, cpu, thread)->ena = 0;
283 perf_counts(counter->counts, cpu, thread)->run = 0;
287 count->loaded = false;
290 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
291 pr_err("failed to write stat event\n");
297 fprintf(stat_config.output,
298 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
299 perf_evsel__name(counter),
301 count->val, count->ena, count->run);
309 static void read_counters(void)
311 struct perf_evsel *counter;
314 evlist__for_each_entry(evsel_list, counter) {
315 ret = read_counter(counter);
317 pr_debug("failed to read counter %s\n", counter->name);
319 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
320 pr_warning("failed to process counter %s\n", counter->name);
324 static void process_interval(void)
326 struct timespec ts, rs;
330 clock_gettime(CLOCK_MONOTONIC, &ts);
331 diff_timespec(&rs, &ts, &ref_time);
334 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
335 pr_err("failed to write stat round event\n");
338 init_stats(&walltime_nsecs_stats);
339 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
340 print_counters(&rs, 0, NULL);
343 static void enable_counters(void)
345 if (stat_config.initial_delay)
346 usleep(stat_config.initial_delay * USEC_PER_MSEC);
349 * We need to enable counters only if:
350 * - we don't have tracee (attaching to task or cpu)
351 * - we have initial delay configured
353 if (!target__none(&target) || stat_config.initial_delay)
354 perf_evlist__enable(evsel_list);
357 static void disable_counters(void)
360 * If we don't have tracee (attaching to task or cpu), counters may
361 * still be running. To get accurate group ratios, we must stop groups
362 * from counting before reading their constituent counters.
364 if (!target__none(&target))
365 perf_evlist__disable(evsel_list);
368 static volatile int workload_exec_errno;
371 * perf_evlist__prepare_workload will send a SIGUSR1
372 * if the fork fails, since we asked by setting its
373 * want_signal to true.
375 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
376 void *ucontext __maybe_unused)
378 workload_exec_errno = info->si_value.sival_int;
381 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
383 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
386 static bool is_target_alive(struct target *_target,
387 struct thread_map *threads)
392 if (!target__has_task(_target))
395 for (i = 0; i < threads->nr; i++) {
398 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
399 threads->map[i].pid);
401 if (!stat(path, &st))
408 static int __run_perf_stat(int argc, const char **argv, int run_idx)
410 int interval = stat_config.interval;
411 int times = stat_config.times;
412 int timeout = stat_config.timeout;
414 unsigned long long t0, t1;
415 struct perf_evsel *counter;
419 const bool forks = (argc > 0);
420 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
421 struct perf_evsel_config_term *err_term;
424 ts.tv_sec = interval / USEC_PER_MSEC;
425 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
426 } else if (timeout) {
427 ts.tv_sec = timeout / USEC_PER_MSEC;
428 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
435 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
436 workload_exec_failed_signal) < 0) {
437 perror("failed to prepare workload");
440 child_pid = evsel_list->workload.pid;
444 perf_evlist__set_leader(evsel_list);
446 evlist__for_each_entry(evsel_list, counter) {
448 if (create_perf_stat_counter(counter, &stat_config, &target) < 0) {
450 /* Weak group failed. Reset the group. */
451 if ((errno == EINVAL || errno == EBADF) &&
452 counter->leader != counter &&
453 counter->weak_group) {
454 counter = perf_evlist__reset_weak_group(evsel_list, counter);
459 * PPC returns ENXIO for HW counters until 2.6.37
460 * (behavior changed with commit b0a873e).
462 if (errno == EINVAL || errno == ENOSYS ||
463 errno == ENOENT || errno == EOPNOTSUPP ||
466 ui__warning("%s event is not supported by the kernel.\n",
467 perf_evsel__name(counter));
468 counter->supported = false;
470 if ((counter->leader != counter) ||
471 !(counter->leader->nr_members > 1))
473 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
475 ui__warning("%s\n", msg);
477 } else if (target__has_per_thread(&target) &&
478 evsel_list->threads &&
479 evsel_list->threads->err_thread != -1) {
481 * For global --per-thread case, skip current
484 if (!thread_map__remove(evsel_list->threads,
485 evsel_list->threads->err_thread)) {
486 evsel_list->threads->err_thread = -1;
491 perf_evsel__open_strerror(counter, &target,
492 errno, msg, sizeof(msg));
493 ui__error("%s\n", msg);
496 kill(child_pid, SIGTERM);
500 counter->supported = true;
502 l = strlen(counter->unit);
503 if (l > stat_config.unit_width)
504 stat_config.unit_width = l;
506 if (perf_evsel__should_store_id(counter) &&
507 perf_evsel__store_ids(counter, evsel_list))
511 if (perf_evlist__apply_filters(evsel_list, &counter)) {
512 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
513 counter->filter, perf_evsel__name(counter), errno,
514 str_error_r(errno, msg, sizeof(msg)));
518 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
519 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
520 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
521 str_error_r(errno, msg, sizeof(msg)));
526 int err, fd = perf_data__fd(&perf_stat.data);
529 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
531 err = perf_session__write_header(perf_stat.session, evsel_list,
538 err = perf_stat_synthesize_config(&stat_config, NULL, evsel_list,
539 process_synthesized_event, is_pipe);
545 * Enable counters and exec the command:
548 clock_gettime(CLOCK_MONOTONIC, &ref_time);
551 perf_evlist__start_workload(evsel_list);
554 if (interval || timeout) {
555 while (!waitpid(child_pid, &status, WNOHANG)) {
556 nanosleep(&ts, NULL);
560 if (interval_count && !(--times))
565 wait4(child_pid, &status, 0, &stat_config.ru_data);
567 if (workload_exec_errno) {
568 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
569 pr_err("Workload failed: %s\n", emsg);
573 if (WIFSIGNALED(status))
574 psignal(WTERMSIG(status), argv[0]);
578 nanosleep(&ts, NULL);
579 if (!is_target_alive(&target, evsel_list->threads))
585 if (interval_count && !(--times))
595 if (stat_config.walltime_run_table)
596 stat_config.walltime_run[run_idx] = t1 - t0;
598 update_stats(&walltime_nsecs_stats, t1 - t0);
601 * Closing a group leader splits the group, and as we only disable
602 * group leaders, results in remaining events becoming enabled. To
603 * avoid arbitrary skew, we must read all counters before closing any
607 perf_evlist__close(evsel_list);
609 return WEXITSTATUS(status);
612 static int run_perf_stat(int argc, const char **argv, int run_idx)
617 ret = system(pre_cmd);
625 ret = __run_perf_stat(argc, argv, run_idx);
630 ret = system(post_cmd);
638 static void print_counters(struct timespec *ts, int argc, const char **argv)
640 /* Do not print anything if we record to the pipe. */
641 if (STAT_RECORD && perf_stat.data.is_pipe)
644 perf_evlist__print_counters(evsel_list, &stat_config, &target,
648 static volatile int signr = -1;
650 static void skip_signal(int signo)
652 if ((child_pid == -1) || stat_config.interval)
657 * render child_pid harmless
658 * won't send SIGTERM to a random
659 * process in case of race condition
660 * and fast PID recycling
665 static void sig_atexit(void)
670 * avoid race condition with SIGCHLD handler
671 * in skip_signal() which is modifying child_pid
672 * goal is to avoid send SIGTERM to a random
676 sigaddset(&set, SIGCHLD);
677 sigprocmask(SIG_BLOCK, &set, &oset);
680 kill(child_pid, SIGTERM);
682 sigprocmask(SIG_SETMASK, &oset, NULL);
687 signal(signr, SIG_DFL);
688 kill(getpid(), signr);
691 static int stat__set_big_num(const struct option *opt __maybe_unused,
692 const char *s __maybe_unused, int unset)
694 big_num_opt = unset ? 0 : 1;
698 static int enable_metric_only(const struct option *opt __maybe_unused,
699 const char *s __maybe_unused, int unset)
701 force_metric_only = true;
702 stat_config.metric_only = !unset;
706 static int parse_metric_groups(const struct option *opt,
708 int unset __maybe_unused)
710 return metricgroup__parse_groups(opt, str, &stat_config.metric_events);
713 static struct option stat_options[] = {
714 OPT_BOOLEAN('T', "transaction", &transaction_run,
715 "hardware transaction statistics"),
716 OPT_CALLBACK('e', "event", &evsel_list, "event",
717 "event selector. use 'perf list' to list available events",
718 parse_events_option),
719 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
720 "event filter", parse_filter),
721 OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
722 "child tasks do not inherit counters"),
723 OPT_STRING('p', "pid", &target.pid, "pid",
724 "stat events on existing process id"),
725 OPT_STRING('t', "tid", &target.tid, "tid",
726 "stat events on existing thread id"),
727 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
728 "system-wide collection from all CPUs"),
729 OPT_BOOLEAN('g', "group", &group,
730 "put the counters into a counter group"),
731 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
732 OPT_INCR('v', "verbose", &verbose,
733 "be more verbose (show counter open errors, etc)"),
734 OPT_INTEGER('r', "repeat", &stat_config.run_count,
735 "repeat command and print average + stddev (max: 100, forever: 0)"),
736 OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
737 "display details about each run (only with -r option)"),
738 OPT_BOOLEAN('n', "null", &stat_config.null_run,
739 "null run - dont start any counters"),
740 OPT_INCR('d', "detailed", &detailed_run,
741 "detailed run - start a lot of events"),
742 OPT_BOOLEAN('S', "sync", &sync_run,
743 "call sync() before starting a run"),
744 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
745 "print large numbers with thousands\' separators",
747 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
748 "list of cpus to monitor in system-wide"),
749 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
750 "disable CPU count aggregation", AGGR_NONE),
751 OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
752 OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
753 "print counts with custom separator"),
754 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
755 "monitor event in cgroup name only", parse_cgroups),
756 OPT_STRING('o', "output", &output_name, "file", "output file name"),
757 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
758 OPT_INTEGER(0, "log-fd", &output_fd,
759 "log output to fd, instead of stderr"),
760 OPT_STRING(0, "pre", &pre_cmd, "command",
761 "command to run prior to the measured command"),
762 OPT_STRING(0, "post", &post_cmd, "command",
763 "command to run after to the measured command"),
764 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
765 "print counts at regular interval in ms "
766 "(overhead is possible for values <= 100ms)"),
767 OPT_INTEGER(0, "interval-count", &stat_config.times,
768 "print counts for fixed number of times"),
769 OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
770 "clear screen in between new interval"),
771 OPT_UINTEGER(0, "timeout", &stat_config.timeout,
772 "stop workload and print counts after a timeout period in ms (>= 10ms)"),
773 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
774 "aggregate counts per processor socket", AGGR_SOCKET),
775 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
776 "aggregate counts per physical processor core", AGGR_CORE),
777 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
778 "aggregate counts per thread", AGGR_THREAD),
779 OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
780 "ms to wait before starting measurement after program start"),
781 OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
782 "Only print computed metrics. No raw values", enable_metric_only),
783 OPT_BOOLEAN(0, "topdown", &topdown_run,
784 "measure topdown level 1 statistics"),
785 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
787 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
788 "monitor specified metrics or metric groups (separated by ,)",
789 parse_metric_groups),
793 static int perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
794 struct cpu_map *map, int cpu)
796 return cpu_map__get_socket(map, cpu, NULL);
799 static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
800 struct cpu_map *map, int cpu)
802 return cpu_map__get_core(map, cpu, NULL);
805 static int cpu_map__get_max(struct cpu_map *map)
809 for (i = 0; i < map->nr; i++) {
810 if (map->map[i] > max)
817 static int perf_stat__get_aggr(struct perf_stat_config *config,
818 aggr_get_id_t get_id, struct cpu_map *map, int idx)
827 if (config->cpus_aggr_map->map[cpu] == -1)
828 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
830 return config->cpus_aggr_map->map[cpu];
833 static int perf_stat__get_socket_cached(struct perf_stat_config *config,
834 struct cpu_map *map, int idx)
836 return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
839 static int perf_stat__get_core_cached(struct perf_stat_config *config,
840 struct cpu_map *map, int idx)
842 return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
845 static int perf_stat_init_aggr_mode(void)
849 switch (stat_config.aggr_mode) {
851 if (cpu_map__build_socket_map(evsel_list->cpus, &stat_config.aggr_map)) {
852 perror("cannot build socket map");
855 stat_config.aggr_get_id = perf_stat__get_socket_cached;
858 if (cpu_map__build_core_map(evsel_list->cpus, &stat_config.aggr_map)) {
859 perror("cannot build core map");
862 stat_config.aggr_get_id = perf_stat__get_core_cached;
873 * The evsel_list->cpus is the base we operate on,
874 * taking the highest cpu number to be the size of
875 * the aggregation translate cpumap.
877 nr = cpu_map__get_max(evsel_list->cpus);
878 stat_config.cpus_aggr_map = cpu_map__empty_new(nr + 1);
879 return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
882 static void perf_stat__exit_aggr_mode(void)
884 cpu_map__put(stat_config.aggr_map);
885 cpu_map__put(stat_config.cpus_aggr_map);
886 stat_config.aggr_map = NULL;
887 stat_config.cpus_aggr_map = NULL;
890 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
899 if (cpu >= env->nr_cpus_avail)
905 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
907 struct perf_env *env = data;
908 int cpu = perf_env__get_cpu(env, map, idx);
910 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
913 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
915 struct perf_env *env = data;
916 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
919 int socket_id = env->cpu[cpu].socket_id;
922 * Encode socket in upper 16 bits
923 * core_id is relative to socket, and
924 * we need a global id. So we combine
927 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
933 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
934 struct cpu_map **sockp)
936 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
939 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
940 struct cpu_map **corep)
942 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
945 static int perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
946 struct cpu_map *map, int idx)
948 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
951 static int perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
952 struct cpu_map *map, int idx)
954 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
957 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
959 struct perf_env *env = &st->session->header.env;
961 switch (stat_config.aggr_mode) {
963 if (perf_env__build_socket_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
964 perror("cannot build socket map");
967 stat_config.aggr_get_id = perf_stat__get_socket_file;
970 if (perf_env__build_core_map(env, evsel_list->cpus, &stat_config.aggr_map)) {
971 perror("cannot build core map");
974 stat_config.aggr_get_id = perf_stat__get_core_file;
987 static int topdown_filter_events(const char **attr, char **str, bool use_group)
994 for (i = 0; attr[i]; i++) {
995 if (pmu_have_event("cpu", attr[i])) {
996 len += strlen(attr[i]) + 1;
997 attr[i - off] = attr[i];
1001 attr[i - off] = NULL;
1003 *str = malloc(len + 1 + 2);
1013 for (i = 0; attr[i]; i++) {
1026 __weak bool arch_topdown_check_group(bool *warn)
1032 __weak void arch_topdown_group_warn(void)
1037 * Add default attributes, if there were no attributes specified or
1038 * if -d/--detailed, -d -d or -d -d -d is used:
1040 static int add_default_attributes(void)
1043 struct perf_event_attr default_attrs0[] = {
1045 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1046 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1047 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1048 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1050 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1052 struct perf_event_attr frontend_attrs[] = {
1053 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1055 struct perf_event_attr backend_attrs[] = {
1056 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1058 struct perf_event_attr default_attrs1[] = {
1059 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1060 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1061 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1066 * Detailed stats (-d), covering the L1 and last level data caches:
1068 struct perf_event_attr detailed_attrs[] = {
1070 { .type = PERF_TYPE_HW_CACHE,
1072 PERF_COUNT_HW_CACHE_L1D << 0 |
1073 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1074 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1076 { .type = PERF_TYPE_HW_CACHE,
1078 PERF_COUNT_HW_CACHE_L1D << 0 |
1079 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1080 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1082 { .type = PERF_TYPE_HW_CACHE,
1084 PERF_COUNT_HW_CACHE_LL << 0 |
1085 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1086 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1088 { .type = PERF_TYPE_HW_CACHE,
1090 PERF_COUNT_HW_CACHE_LL << 0 |
1091 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1092 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1096 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1098 struct perf_event_attr very_detailed_attrs[] = {
1100 { .type = PERF_TYPE_HW_CACHE,
1102 PERF_COUNT_HW_CACHE_L1I << 0 |
1103 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1104 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1106 { .type = PERF_TYPE_HW_CACHE,
1108 PERF_COUNT_HW_CACHE_L1I << 0 |
1109 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1110 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1112 { .type = PERF_TYPE_HW_CACHE,
1114 PERF_COUNT_HW_CACHE_DTLB << 0 |
1115 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1116 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1118 { .type = PERF_TYPE_HW_CACHE,
1120 PERF_COUNT_HW_CACHE_DTLB << 0 |
1121 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1122 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1124 { .type = PERF_TYPE_HW_CACHE,
1126 PERF_COUNT_HW_CACHE_ITLB << 0 |
1127 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1128 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1130 { .type = PERF_TYPE_HW_CACHE,
1132 PERF_COUNT_HW_CACHE_ITLB << 0 |
1133 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1134 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1139 * Very, very detailed stats (-d -d -d), adding prefetch events:
1141 struct perf_event_attr very_very_detailed_attrs[] = {
1143 { .type = PERF_TYPE_HW_CACHE,
1145 PERF_COUNT_HW_CACHE_L1D << 0 |
1146 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1147 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1149 { .type = PERF_TYPE_HW_CACHE,
1151 PERF_COUNT_HW_CACHE_L1D << 0 |
1152 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1153 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1155 struct parse_events_error errinfo;
1157 /* Set attrs if no event is selected and !null_run: */
1158 if (stat_config.null_run)
1161 if (transaction_run) {
1162 /* Handle -T as -M transaction. Once platform specific metrics
1163 * support has been added to the json files, all archictures
1164 * will use this approach. To determine transaction support
1165 * on an architecture test for such a metric name.
1167 if (metricgroup__has_metric("transaction")) {
1168 struct option opt = { .value = &evsel_list };
1170 return metricgroup__parse_groups(&opt, "transaction",
1171 &stat_config.metric_events);
1174 if (pmu_have_event("cpu", "cycles-ct") &&
1175 pmu_have_event("cpu", "el-start"))
1176 err = parse_events(evsel_list, transaction_attrs,
1179 err = parse_events(evsel_list,
1180 transaction_limited_attrs,
1183 fprintf(stderr, "Cannot set up transaction events\n");
1184 parse_events_print_error(&errinfo, transaction_attrs);
1193 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1194 fprintf(stderr, "freeze_on_smi is not supported.\n");
1199 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1200 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1206 if (pmu_have_event("msr", "aperf") &&
1207 pmu_have_event("msr", "smi")) {
1208 if (!force_metric_only)
1209 stat_config.metric_only = true;
1210 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1212 fprintf(stderr, "To measure SMI cost, it needs "
1213 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1214 parse_events_print_error(&errinfo, smi_cost_attrs);
1218 fprintf(stderr, "Cannot set up SMI cost events\n");
1228 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1229 stat_config.aggr_mode != AGGR_CORE) {
1230 pr_err("top down event configuration requires --per-core mode\n");
1233 stat_config.aggr_mode = AGGR_CORE;
1234 if (nr_cgroups || !target__has_cpu(&target)) {
1235 pr_err("top down event configuration requires system-wide mode (-a)\n");
1239 if (!force_metric_only)
1240 stat_config.metric_only = true;
1241 if (topdown_filter_events(topdown_attrs, &str,
1242 arch_topdown_check_group(&warn)) < 0) {
1243 pr_err("Out of memory\n");
1246 if (topdown_attrs[0] && str) {
1248 arch_topdown_group_warn();
1249 err = parse_events(evsel_list, str, &errinfo);
1252 "Cannot set up top down events %s: %d\n",
1255 parse_events_print_error(&errinfo, str);
1259 fprintf(stderr, "System does not support topdown\n");
1265 if (!evsel_list->nr_entries) {
1266 if (target__has_cpu(&target))
1267 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1269 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1271 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1272 if (perf_evlist__add_default_attrs(evsel_list,
1273 frontend_attrs) < 0)
1276 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1277 if (perf_evlist__add_default_attrs(evsel_list,
1281 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1285 /* Detailed events get appended to the event list: */
1287 if (detailed_run < 1)
1290 /* Append detailed run extra attributes: */
1291 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1294 if (detailed_run < 2)
1297 /* Append very detailed run extra attributes: */
1298 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1301 if (detailed_run < 3)
1304 /* Append very, very detailed run extra attributes: */
1305 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1308 static const char * const stat_record_usage[] = {
1309 "perf stat record [<options>]",
1313 static void init_features(struct perf_session *session)
1317 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1318 perf_header__set_feat(&session->header, feat);
1320 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1321 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1322 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1323 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1326 static int __cmd_record(int argc, const char **argv)
1328 struct perf_session *session;
1329 struct perf_data *data = &perf_stat.data;
1331 argc = parse_options(argc, argv, stat_options, stat_record_usage,
1332 PARSE_OPT_STOP_AT_NON_OPTION);
1335 data->file.path = output_name;
1337 if (stat_config.run_count != 1 || forever) {
1338 pr_err("Cannot use -r option with perf stat record.\n");
1342 session = perf_session__new(data, false, NULL);
1343 if (session == NULL) {
1344 pr_err("Perf session creation failed.\n");
1348 init_features(session);
1350 session->evlist = evsel_list;
1351 perf_stat.session = session;
1352 perf_stat.record = true;
1356 static int process_stat_round_event(struct perf_session *session,
1357 union perf_event *event)
1359 struct stat_round_event *stat_round = &event->stat_round;
1360 struct perf_evsel *counter;
1361 struct timespec tsh, *ts = NULL;
1362 const char **argv = session->header.env.cmdline_argv;
1363 int argc = session->header.env.nr_cmdline;
1365 evlist__for_each_entry(evsel_list, counter)
1366 perf_stat_process_counter(&stat_config, counter);
1368 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
1369 update_stats(&walltime_nsecs_stats, stat_round->time);
1371 if (stat_config.interval && stat_round->time) {
1372 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
1373 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
1377 print_counters(ts, argc, argv);
1382 int process_stat_config_event(struct perf_session *session,
1383 union perf_event *event)
1385 struct perf_tool *tool = session->tool;
1386 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1388 perf_event__read_stat_config(&stat_config, &event->stat_config);
1390 if (cpu_map__empty(st->cpus)) {
1391 if (st->aggr_mode != AGGR_UNSET)
1392 pr_warning("warning: processing task data, aggregation mode not set\n");
1396 if (st->aggr_mode != AGGR_UNSET)
1397 stat_config.aggr_mode = st->aggr_mode;
1399 if (perf_stat.data.is_pipe)
1400 perf_stat_init_aggr_mode();
1402 perf_stat_init_aggr_mode_file(st);
1407 static int set_maps(struct perf_stat *st)
1409 if (!st->cpus || !st->threads)
1412 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
1415 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
1417 if (perf_evlist__alloc_stats(evsel_list, true))
1420 st->maps_allocated = true;
1425 int process_thread_map_event(struct perf_session *session,
1426 union perf_event *event)
1428 struct perf_tool *tool = session->tool;
1429 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1432 pr_warning("Extra thread map event, ignoring.\n");
1436 st->threads = thread_map__new_event(&event->thread_map);
1440 return set_maps(st);
1444 int process_cpu_map_event(struct perf_session *session,
1445 union perf_event *event)
1447 struct perf_tool *tool = session->tool;
1448 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
1449 struct cpu_map *cpus;
1452 pr_warning("Extra cpu map event, ignoring.\n");
1456 cpus = cpu_map__new_data(&event->cpu_map.data);
1461 return set_maps(st);
1464 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
1468 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
1472 config->stats_num = nthreads;
1474 for (i = 0; i < nthreads; i++)
1475 runtime_stat__init(&config->stats[i]);
1480 static void runtime_stat_delete(struct perf_stat_config *config)
1487 for (i = 0; i < config->stats_num; i++)
1488 runtime_stat__exit(&config->stats[i]);
1490 free(config->stats);
1493 static const char * const stat_report_usage[] = {
1494 "perf stat report [<options>]",
1498 static struct perf_stat perf_stat = {
1500 .attr = perf_event__process_attr,
1501 .event_update = perf_event__process_event_update,
1502 .thread_map = process_thread_map_event,
1503 .cpu_map = process_cpu_map_event,
1504 .stat_config = process_stat_config_event,
1505 .stat = perf_event__process_stat_event,
1506 .stat_round = process_stat_round_event,
1508 .aggr_mode = AGGR_UNSET,
1511 static int __cmd_report(int argc, const char **argv)
1513 struct perf_session *session;
1514 const struct option options[] = {
1515 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1516 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
1517 "aggregate counts per processor socket", AGGR_SOCKET),
1518 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
1519 "aggregate counts per physical processor core", AGGR_CORE),
1520 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
1521 "disable CPU count aggregation", AGGR_NONE),
1527 argc = parse_options(argc, argv, options, stat_report_usage, 0);
1529 if (!input_name || !strlen(input_name)) {
1530 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1533 input_name = "perf.data";
1536 perf_stat.data.file.path = input_name;
1537 perf_stat.data.mode = PERF_DATA_MODE_READ;
1539 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
1540 if (session == NULL)
1543 perf_stat.session = session;
1544 stat_config.output = stderr;
1545 evsel_list = session->evlist;
1547 ret = perf_session__process_events(session);
1551 perf_session__delete(session);
1555 static void setup_system_wide(int forks)
1558 * Make system wide (-a) the default target if
1559 * no target was specified and one of following
1560 * conditions is met:
1562 * - there's no workload specified
1563 * - there is workload specified but all requested
1564 * events are system wide events
1566 if (!target__none(&target))
1570 target.system_wide = true;
1572 struct perf_evsel *counter;
1574 evlist__for_each_entry(evsel_list, counter) {
1575 if (!counter->system_wide)
1579 if (evsel_list->nr_entries)
1580 target.system_wide = true;
1584 int cmd_stat(int argc, const char **argv)
1586 const char * const stat_usage[] = {
1587 "perf stat [<options>] [<command>]",
1590 int status = -EINVAL, run_idx;
1592 FILE *output = stderr;
1593 unsigned int interval, timeout;
1594 const char * const stat_subcommands[] = { "record", "report" };
1596 setlocale(LC_ALL, "");
1598 evsel_list = perf_evlist__new();
1599 if (evsel_list == NULL)
1602 parse_events__shrink_config_terms();
1604 /* String-parsing callback-based options would segfault when negated */
1605 set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
1606 set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
1607 set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
1609 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
1610 (const char **) stat_usage,
1611 PARSE_OPT_STOP_AT_NON_OPTION);
1612 perf_stat__collect_metric_expr(evsel_list);
1613 perf_stat__init_shadow_stats();
1615 if (stat_config.csv_sep) {
1616 stat_config.csv_output = true;
1617 if (!strcmp(stat_config.csv_sep, "\\t"))
1618 stat_config.csv_sep = "\t";
1620 stat_config.csv_sep = DEFAULT_SEPARATOR;
1622 if (argc && !strncmp(argv[0], "rec", 3)) {
1623 argc = __cmd_record(argc, argv);
1626 } else if (argc && !strncmp(argv[0], "rep", 3))
1627 return __cmd_report(argc, argv);
1629 interval = stat_config.interval;
1630 timeout = stat_config.timeout;
1633 * For record command the -o is already taken care of.
1635 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
1638 if (output_name && output_fd) {
1639 fprintf(stderr, "cannot use both --output and --log-fd\n");
1640 parse_options_usage(stat_usage, stat_options, "o", 1);
1641 parse_options_usage(NULL, stat_options, "log-fd", 0);
1645 if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
1646 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
1650 if (stat_config.metric_only && stat_config.run_count > 1) {
1651 fprintf(stderr, "--metric-only is not supported with -r\n");
1655 if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
1656 fprintf(stderr, "--table is only supported with -r\n");
1657 parse_options_usage(stat_usage, stat_options, "r", 1);
1658 parse_options_usage(NULL, stat_options, "table", 0);
1662 if (output_fd < 0) {
1663 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1664 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
1670 mode = append_file ? "a" : "w";
1672 output = fopen(output_name, mode);
1674 perror("failed to create output file");
1677 clock_gettime(CLOCK_REALTIME, &tm);
1678 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1679 } else if (output_fd > 0) {
1680 mode = append_file ? "a" : "w";
1681 output = fdopen(output_fd, mode);
1683 perror("Failed opening logfd");
1688 stat_config.output = output;
1691 * let the spreadsheet do the pretty-printing
1693 if (stat_config.csv_output) {
1694 /* User explicitly passed -B? */
1695 if (big_num_opt == 1) {
1696 fprintf(stderr, "-B option not supported with -x\n");
1697 parse_options_usage(stat_usage, stat_options, "B", 1);
1698 parse_options_usage(NULL, stat_options, "x", 1);
1700 } else /* Nope, so disable big number formatting */
1701 stat_config.big_num = false;
1702 } else if (big_num_opt == 0) /* User passed --no-big-num */
1703 stat_config.big_num = false;
1705 setup_system_wide(argc);
1708 * Display user/system times only for single
1709 * run and when there's specified tracee.
1711 if ((stat_config.run_count == 1) && target__none(&target))
1712 stat_config.ru_display = true;
1714 if (stat_config.run_count < 0) {
1715 pr_err("Run count must be a positive number\n");
1716 parse_options_usage(stat_usage, stat_options, "r", 1);
1718 } else if (stat_config.run_count == 0) {
1720 stat_config.run_count = 1;
1723 if (stat_config.walltime_run_table) {
1724 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
1725 if (!stat_config.walltime_run) {
1726 pr_err("failed to setup -r option");
1731 if ((stat_config.aggr_mode == AGGR_THREAD) &&
1732 !target__has_task(&target)) {
1733 if (!target.system_wide || target.cpu_list) {
1734 fprintf(stderr, "The --per-thread option is only "
1735 "available when monitoring via -p -t -a "
1736 "options or only --per-thread.\n");
1737 parse_options_usage(NULL, stat_options, "p", 1);
1738 parse_options_usage(NULL, stat_options, "t", 1);
1744 * no_aggr, cgroup are for system-wide only
1745 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1747 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1748 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1749 !target__has_cpu(&target)) {
1750 fprintf(stderr, "both cgroup and no-aggregation "
1751 "modes only available in system-wide mode\n");
1753 parse_options_usage(stat_usage, stat_options, "G", 1);
1754 parse_options_usage(NULL, stat_options, "A", 1);
1755 parse_options_usage(NULL, stat_options, "a", 1);
1759 if (add_default_attributes())
1762 target__validate(&target);
1764 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
1765 target.per_thread = true;
1767 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1768 if (target__has_task(&target)) {
1769 pr_err("Problems finding threads of monitor\n");
1770 parse_options_usage(stat_usage, stat_options, "p", 1);
1771 parse_options_usage(NULL, stat_options, "t", 1);
1772 } else if (target__has_cpu(&target)) {
1773 perror("failed to parse CPUs map");
1774 parse_options_usage(stat_usage, stat_options, "C", 1);
1775 parse_options_usage(NULL, stat_options, "a", 1);
1781 * Initialize thread_map with comm names,
1782 * so we could print it out on output.
1784 if (stat_config.aggr_mode == AGGR_THREAD) {
1785 thread_map__read_comms(evsel_list->threads);
1786 if (target.system_wide) {
1787 if (runtime_stat_new(&stat_config,
1788 thread_map__nr(evsel_list->threads))) {
1794 if (stat_config.times && interval)
1795 interval_count = true;
1796 else if (stat_config.times && !interval) {
1797 pr_err("interval-count option should be used together with "
1798 "interval-print.\n");
1799 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
1800 parse_options_usage(stat_usage, stat_options, "I", 1);
1804 if (timeout && timeout < 100) {
1806 pr_err("timeout must be >= 10ms.\n");
1807 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1810 pr_warning("timeout < 100ms. "
1811 "The overhead percentage could be high in some cases. "
1812 "Please proceed with caution.\n");
1814 if (timeout && interval) {
1815 pr_err("timeout option is not supported with interval-print.\n");
1816 parse_options_usage(stat_usage, stat_options, "timeout", 0);
1817 parse_options_usage(stat_usage, stat_options, "I", 1);
1821 if (perf_evlist__alloc_stats(evsel_list, interval))
1824 if (perf_stat_init_aggr_mode())
1828 * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
1829 * while avoiding that older tools show confusing messages.
1831 * However for pipe sessions we need to keep it zero,
1832 * because script's perf_evsel__check_attr is triggered
1833 * by attr->sample_type != 0, and we can't run it on
1836 stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
1839 * We dont want to block the signals - that would cause
1840 * child tasks to inherit that and Ctrl-C would not work.
1841 * What we want is for Ctrl-C to work in the exec()-ed
1842 * task, but being ignored by perf stat itself:
1846 signal(SIGINT, skip_signal);
1847 signal(SIGCHLD, skip_signal);
1848 signal(SIGALRM, skip_signal);
1849 signal(SIGABRT, skip_signal);
1852 for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
1853 if (stat_config.run_count != 1 && verbose > 0)
1854 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1857 status = run_perf_stat(argc, argv, run_idx);
1858 if (forever && status != -1) {
1859 print_counters(NULL, argc, argv);
1860 perf_stat__reset_stats();
1864 if (!forever && status != -1 && !interval)
1865 print_counters(NULL, argc, argv);
1869 * We synthesize the kernel mmap record just so that older tools
1870 * don't emit warnings about not being able to resolve symbols
1871 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
1872 * a saner message about no samples being in the perf.data file.
1874 * This also serves to suppress a warning about f_header.data.size == 0
1875 * in header.c at the moment 'perf stat record' gets introduced, which
1876 * is not really needed once we start adding the stat specific PERF_RECORD_
1877 * records, but the need to suppress the kptr_restrict messages in older
1878 * tools remain -acme
1880 int fd = perf_data__fd(&perf_stat.data);
1881 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
1882 process_synthesized_event,
1883 &perf_stat.session->machines.host);
1885 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
1886 "older tools may produce warnings about this file\n.");
1890 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
1891 pr_err("failed to write stat round event\n");
1894 if (!perf_stat.data.is_pipe) {
1895 perf_stat.session->header.data_size += perf_stat.bytes_written;
1896 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
1899 perf_session__delete(perf_stat.session);
1902 perf_stat__exit_aggr_mode();
1903 perf_evlist__free_stats(evsel_list);
1905 free(stat_config.walltime_run);
1907 if (smi_cost && smi_reset)
1908 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
1910 perf_evlist__delete(evsel_list);
1912 runtime_stat_delete(&stat_config);