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/group.h"
67 #include "util/string2.h"
68 #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 "sane_ctype.h"
87 #define DEFAULT_SEPARATOR " "
88 #define CNTR_NOT_SUPPORTED "<not supported>"
89 #define CNTR_NOT_COUNTED "<not counted>"
90 #define FREEZE_ON_SMI_PATH "devices/cpu/freeze_on_smi"
92 static void print_counters(struct timespec *ts, int argc, const char **argv);
94 /* Default events used for perf stat -T */
95 static const char *transaction_attrs = {
107 /* More limited version when the CPU does not have all events. */
108 static const char * transaction_limited_attrs = {
118 static const char * topdown_attrs[] = {
119 "topdown-total-slots",
120 "topdown-slots-retired",
121 "topdown-recovery-bubbles",
122 "topdown-fetch-bubbles",
123 "topdown-slots-issued",
127 static const char *smi_cost_attrs = {
135 static struct perf_evlist *evsel_list;
137 static struct rblist metric_events;
139 static struct target target = {
143 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
145 static int run_count = 1;
146 static bool no_inherit = false;
147 static volatile pid_t child_pid = -1;
148 static bool null_run = false;
149 static int detailed_run = 0;
150 static bool transaction_run;
151 static bool topdown_run = false;
152 static bool smi_cost = false;
153 static bool smi_reset = false;
154 static bool big_num = true;
155 static int big_num_opt = -1;
156 static const char *csv_sep = NULL;
157 static bool csv_output = false;
158 static bool group = false;
159 static const char *pre_cmd = NULL;
160 static const char *post_cmd = NULL;
161 static bool sync_run = false;
162 static unsigned int initial_delay = 0;
163 static unsigned int unit_width = 4; /* strlen("unit") */
164 static bool forever = false;
165 static bool metric_only = false;
166 static bool force_metric_only = false;
167 static bool no_merge = false;
168 static struct timespec ref_time;
169 static struct cpu_map *aggr_map;
170 static aggr_get_id_t aggr_get_id;
171 static bool append_file;
172 static const char *output_name;
173 static int output_fd;
174 static int print_free_counters_hint;
178 struct perf_data data;
179 struct perf_session *session;
181 struct perf_tool tool;
183 struct cpu_map *cpus;
184 struct thread_map *threads;
185 enum aggr_mode aggr_mode;
188 static struct perf_stat perf_stat;
189 #define STAT_RECORD perf_stat.record
191 static volatile int done = 0;
193 static struct perf_stat_config stat_config = {
194 .aggr_mode = AGGR_GLOBAL,
198 static bool is_duration_time(struct perf_evsel *evsel)
200 return !strcmp(evsel->name, "duration_time");
203 static inline void diff_timespec(struct timespec *r, struct timespec *a,
206 r->tv_sec = a->tv_sec - b->tv_sec;
207 if (a->tv_nsec < b->tv_nsec) {
208 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
211 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
215 static void perf_stat__reset_stats(void)
217 perf_evlist__reset_stats(evsel_list);
218 perf_stat__reset_shadow_stats();
221 static int create_perf_stat_counter(struct perf_evsel *evsel)
223 struct perf_event_attr *attr = &evsel->attr;
224 struct perf_evsel *leader = evsel->leader;
226 if (stat_config.scale) {
227 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
228 PERF_FORMAT_TOTAL_TIME_RUNNING;
232 * The event is part of non trivial group, let's enable
233 * the group read (for leader) and ID retrieval for all
236 if (leader->nr_members > 1)
237 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
239 attr->inherit = !no_inherit;
242 * Some events get initialized with sample_(period/type) set,
243 * like tracepoints. Clear it up for counting.
245 attr->sample_period = 0;
248 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
249 * while avoiding that older tools show confusing messages.
251 * However for pipe sessions we need to keep it zero,
252 * because script's perf_evsel__check_attr is triggered
253 * by attr->sample_type != 0, and we can't run it on
256 if (!(STAT_RECORD && perf_stat.data.is_pipe))
257 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
260 * Disabling all counters initially, they will be enabled
261 * either manually by us or by kernel via enable_on_exec
264 if (perf_evsel__is_group_leader(evsel)) {
268 * In case of initial_delay we enable tracee
271 if (target__none(&target) && !initial_delay)
272 attr->enable_on_exec = 1;
275 if (target__has_cpu(&target))
276 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
278 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
282 * Does the counter have nsecs as a unit?
284 static inline int nsec_counter(struct perf_evsel *evsel)
286 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
287 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
293 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
294 union perf_event *event,
295 struct perf_sample *sample __maybe_unused,
296 struct machine *machine __maybe_unused)
298 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
299 pr_err("failed to write perf data, error: %m\n");
303 perf_stat.bytes_written += event->header.size;
307 static int write_stat_round_event(u64 tm, u64 type)
309 return perf_event__synthesize_stat_round(NULL, tm, type,
310 process_synthesized_event,
314 #define WRITE_STAT_ROUND_EVENT(time, interval) \
315 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
317 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
320 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
321 struct perf_counts_values *count)
323 struct perf_sample_id *sid = SID(counter, cpu, thread);
325 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
326 process_synthesized_event, NULL);
330 * Read out the results of a single counter:
331 * do not aggregate counts across CPUs in system-wide mode
333 static int read_counter(struct perf_evsel *counter)
335 int nthreads = thread_map__nr(evsel_list->threads);
336 int ncpus, cpu, thread;
338 if (target__has_cpu(&target))
339 ncpus = perf_evsel__nr_cpus(counter);
343 if (!counter->supported)
346 if (counter->system_wide)
349 for (thread = 0; thread < nthreads; thread++) {
350 for (cpu = 0; cpu < ncpus; cpu++) {
351 struct perf_counts_values *count;
353 count = perf_counts(counter->counts, cpu, thread);
356 * The leader's group read loads data into its group members
357 * (via perf_evsel__read_counter) and sets threir count->loaded.
359 if (!count->loaded &&
360 perf_evsel__read_counter(counter, cpu, thread)) {
361 counter->counts->scaled = -1;
362 perf_counts(counter->counts, cpu, thread)->ena = 0;
363 perf_counts(counter->counts, cpu, thread)->run = 0;
367 count->loaded = false;
370 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
371 pr_err("failed to write stat event\n");
377 fprintf(stat_config.output,
378 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
379 perf_evsel__name(counter),
381 count->val, count->ena, count->run);
389 static void read_counters(void)
391 struct perf_evsel *counter;
394 evlist__for_each_entry(evsel_list, counter) {
395 ret = read_counter(counter);
397 pr_debug("failed to read counter %s\n", counter->name);
399 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
400 pr_warning("failed to process counter %s\n", counter->name);
404 static void process_interval(void)
406 struct timespec ts, rs;
410 clock_gettime(CLOCK_MONOTONIC, &ts);
411 diff_timespec(&rs, &ts, &ref_time);
414 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
415 pr_err("failed to write stat round event\n");
418 init_stats(&walltime_nsecs_stats);
419 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
420 print_counters(&rs, 0, NULL);
423 static void enable_counters(void)
426 usleep(initial_delay * USEC_PER_MSEC);
429 * We need to enable counters only if:
430 * - we don't have tracee (attaching to task or cpu)
431 * - we have initial delay configured
433 if (!target__none(&target) || initial_delay)
434 perf_evlist__enable(evsel_list);
437 static void disable_counters(void)
440 * If we don't have tracee (attaching to task or cpu), counters may
441 * still be running. To get accurate group ratios, we must stop groups
442 * from counting before reading their constituent counters.
444 if (!target__none(&target))
445 perf_evlist__disable(evsel_list);
448 static volatile int workload_exec_errno;
451 * perf_evlist__prepare_workload will send a SIGUSR1
452 * if the fork fails, since we asked by setting its
453 * want_signal to true.
455 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
456 void *ucontext __maybe_unused)
458 workload_exec_errno = info->si_value.sival_int;
461 static bool has_unit(struct perf_evsel *counter)
463 return counter->unit && *counter->unit;
466 static bool has_scale(struct perf_evsel *counter)
468 return counter->scale != 1;
471 static int perf_stat_synthesize_config(bool is_pipe)
473 struct perf_evsel *counter;
477 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
478 process_synthesized_event);
480 pr_err("Couldn't synthesize attrs.\n");
486 * Synthesize other events stuff not carried within
487 * attr event - unit, scale, name
489 evlist__for_each_entry(evsel_list, counter) {
490 if (!counter->supported)
494 * Synthesize unit and scale only if it's defined.
496 if (has_unit(counter)) {
497 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
499 pr_err("Couldn't synthesize evsel unit.\n");
504 if (has_scale(counter)) {
505 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
507 pr_err("Couldn't synthesize evsel scale.\n");
512 if (counter->own_cpus) {
513 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
515 pr_err("Couldn't synthesize evsel scale.\n");
521 * Name is needed only for pipe output,
522 * perf.data carries event names.
525 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
527 pr_err("Couldn't synthesize evsel name.\n");
533 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
534 process_synthesized_event,
537 pr_err("Couldn't synthesize thread map.\n");
541 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
542 process_synthesized_event, NULL);
544 pr_err("Couldn't synthesize thread map.\n");
548 err = perf_event__synthesize_stat_config(NULL, &stat_config,
549 process_synthesized_event, NULL);
551 pr_err("Couldn't synthesize config.\n");
558 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
560 static int __store_counter_ids(struct perf_evsel *counter,
561 struct cpu_map *cpus,
562 struct thread_map *threads)
566 for (cpu = 0; cpu < cpus->nr; cpu++) {
567 for (thread = 0; thread < threads->nr; thread++) {
568 int fd = FD(counter, cpu, thread);
570 if (perf_evlist__id_add_fd(evsel_list, counter,
571 cpu, thread, fd) < 0)
579 static int store_counter_ids(struct perf_evsel *counter)
581 struct cpu_map *cpus = counter->cpus;
582 struct thread_map *threads = counter->threads;
584 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
587 return __store_counter_ids(counter, cpus, threads);
590 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
592 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
595 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
597 struct perf_evsel *c2, *leader;
600 leader = evsel->leader;
601 pr_debug("Weak group for %s/%d failed\n",
602 leader->name, leader->nr_members);
605 * for_each_group_member doesn't work here because it doesn't
606 * include the first entry.
608 evlist__for_each_entry(evsel_list, c2) {
611 if (c2->leader == leader) {
613 perf_evsel__close(c2);
621 static int __run_perf_stat(int argc, const char **argv)
623 int interval = stat_config.interval;
625 unsigned long long t0, t1;
626 struct perf_evsel *counter;
630 const bool forks = (argc > 0);
631 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
632 struct perf_evsel_config_term *err_term;
635 ts.tv_sec = interval / USEC_PER_MSEC;
636 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
643 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
644 workload_exec_failed_signal) < 0) {
645 perror("failed to prepare workload");
648 child_pid = evsel_list->workload.pid;
652 perf_evlist__set_leader(evsel_list);
654 evlist__for_each_entry(evsel_list, counter) {
656 if (create_perf_stat_counter(counter) < 0) {
658 /* Weak group failed. Reset the group. */
659 if ((errno == EINVAL || errno == EBADF) &&
660 counter->leader != counter &&
661 counter->weak_group) {
662 counter = perf_evsel__reset_weak_group(counter);
667 * PPC returns ENXIO for HW counters until 2.6.37
668 * (behavior changed with commit b0a873e).
670 if (errno == EINVAL || errno == ENOSYS ||
671 errno == ENOENT || errno == EOPNOTSUPP ||
674 ui__warning("%s event is not supported by the kernel.\n",
675 perf_evsel__name(counter));
676 counter->supported = false;
678 if ((counter->leader != counter) ||
679 !(counter->leader->nr_members > 1))
681 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
683 ui__warning("%s\n", msg);
687 perf_evsel__open_strerror(counter, &target,
688 errno, msg, sizeof(msg));
689 ui__error("%s\n", msg);
692 kill(child_pid, SIGTERM);
696 counter->supported = true;
698 l = strlen(counter->unit);
702 if (perf_evsel__should_store_id(counter) &&
703 store_counter_ids(counter))
707 if (perf_evlist__apply_filters(evsel_list, &counter)) {
708 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
709 counter->filter, perf_evsel__name(counter), errno,
710 str_error_r(errno, msg, sizeof(msg)));
714 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
715 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
716 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
717 str_error_r(errno, msg, sizeof(msg)));
722 int err, fd = perf_data__fd(&perf_stat.data);
725 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
727 err = perf_session__write_header(perf_stat.session, evsel_list,
734 err = perf_stat_synthesize_config(is_pipe);
740 * Enable counters and exec the command:
743 clock_gettime(CLOCK_MONOTONIC, &ref_time);
746 perf_evlist__start_workload(evsel_list);
750 while (!waitpid(child_pid, &status, WNOHANG)) {
751 nanosleep(&ts, NULL);
755 waitpid(child_pid, &status, 0);
757 if (workload_exec_errno) {
758 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
759 pr_err("Workload failed: %s\n", emsg);
763 if (WIFSIGNALED(status))
764 psignal(WTERMSIG(status), argv[0]);
768 nanosleep(&ts, NULL);
778 update_stats(&walltime_nsecs_stats, t1 - t0);
781 * Closing a group leader splits the group, and as we only disable
782 * group leaders, results in remaining events becoming enabled. To
783 * avoid arbitrary skew, we must read all counters before closing any
787 perf_evlist__close(evsel_list);
789 return WEXITSTATUS(status);
792 static int run_perf_stat(int argc, const char **argv)
797 ret = system(pre_cmd);
805 ret = __run_perf_stat(argc, argv);
810 ret = system(post_cmd);
818 static void print_running(u64 run, u64 ena)
821 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
825 ena ? 100.0 * run / ena : 100.0);
826 } else if (run != ena) {
827 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
831 static void print_noise_pct(double total, double avg)
833 double pct = rel_stddev_stats(total, avg);
836 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
838 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
841 static void print_noise(struct perf_evsel *evsel, double avg)
843 struct perf_stat_evsel *ps;
849 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
852 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
854 switch (stat_config.aggr_mode) {
856 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
857 cpu_map__id_to_socket(id),
859 cpu_map__id_to_cpu(id),
866 fprintf(stat_config.output, "S%*d%s%*d%s",
875 fprintf(stat_config.output, "CPU%*d%s",
877 perf_evsel__cpus(evsel)->map[id], csv_sep);
880 fprintf(stat_config.output, "%*s-%*d%s",
882 thread_map__comm(evsel->threads, id),
884 thread_map__pid(evsel->threads, id),
900 struct perf_evsel *evsel;
903 #define METRIC_LEN 35
905 static void new_line_std(void *ctx)
907 struct outstate *os = ctx;
912 static void do_new_line_std(struct outstate *os)
915 fputs(os->prefix, os->fh);
916 aggr_printout(os->evsel, os->id, os->nr);
917 if (stat_config.aggr_mode == AGGR_NONE)
918 fprintf(os->fh, " ");
919 fprintf(os->fh, " ");
922 static void print_metric_std(void *ctx, const char *color, const char *fmt,
923 const char *unit, double val)
925 struct outstate *os = ctx;
928 bool newline = os->newline;
932 if (unit == NULL || fmt == NULL) {
933 fprintf(out, "%-*s", METRIC_LEN, "");
940 n = fprintf(out, " # ");
942 n += color_fprintf(out, color, fmt, val);
944 n += fprintf(out, fmt, val);
945 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
948 static void new_line_csv(void *ctx)
950 struct outstate *os = ctx;
955 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
956 aggr_printout(os->evsel, os->id, os->nr);
957 for (i = 0; i < os->nfields; i++)
958 fputs(csv_sep, os->fh);
961 static void print_metric_csv(void *ctx,
962 const char *color __maybe_unused,
963 const char *fmt, const char *unit, double val)
965 struct outstate *os = ctx;
967 char buf[64], *vals, *ends;
969 if (unit == NULL || fmt == NULL) {
970 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
973 snprintf(buf, sizeof(buf), fmt, val);
974 ends = vals = ltrim(buf);
975 while (isdigit(*ends) || *ends == '.')
978 while (isspace(*unit))
980 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
983 #define METRIC_ONLY_LEN 20
985 /* Filter out some columns that don't work well in metrics only mode */
987 static bool valid_only_metric(const char *unit)
991 if (strstr(unit, "/sec") ||
992 strstr(unit, "hz") ||
993 strstr(unit, "Hz") ||
994 strstr(unit, "CPUs utilized"))
999 static const char *fixunit(char *buf, struct perf_evsel *evsel,
1002 if (!strncmp(unit, "of all", 6)) {
1003 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
1010 static void print_metric_only(void *ctx, const char *color, const char *fmt,
1011 const char *unit, double val)
1013 struct outstate *os = ctx;
1017 unsigned mlen = METRIC_ONLY_LEN;
1019 if (!valid_only_metric(unit))
1021 unit = fixunit(buf, os->evsel, unit);
1023 n = color_fprintf(out, color, fmt, val);
1025 n = fprintf(out, fmt, val);
1026 if (n > METRIC_ONLY_LEN)
1027 n = METRIC_ONLY_LEN;
1028 if (mlen < strlen(unit))
1029 mlen = strlen(unit) + 1;
1030 fprintf(out, "%*s", mlen - n, "");
1033 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1035 const char *unit, double val)
1037 struct outstate *os = ctx;
1039 char buf[64], *vals, *ends;
1042 if (!valid_only_metric(unit))
1044 unit = fixunit(tbuf, os->evsel, unit);
1045 snprintf(buf, sizeof buf, fmt, val);
1046 ends = vals = ltrim(buf);
1047 while (isdigit(*ends) || *ends == '.')
1050 fprintf(out, "%s%s", vals, csv_sep);
1053 static void new_line_metric(void *ctx __maybe_unused)
1057 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1058 const char *fmt __maybe_unused,
1059 const char *unit, double val __maybe_unused)
1061 struct outstate *os = ctx;
1064 if (!valid_only_metric(unit))
1066 unit = fixunit(tbuf, os->evsel, unit);
1068 fprintf(os->fh, "%s%s", unit, csv_sep);
1070 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1073 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1075 FILE *output = stat_config.output;
1076 double msecs = avg / NSEC_PER_MSEC;
1077 const char *fmt_v, *fmt_n;
1080 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1081 fmt_n = csv_output ? "%s" : "%-25s";
1083 aggr_printout(evsel, id, nr);
1085 scnprintf(name, sizeof(name), "%s%s",
1086 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1088 fprintf(output, fmt_v, msecs, csv_sep);
1091 fprintf(output, "%s%s", evsel->unit, csv_sep);
1093 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1095 fprintf(output, fmt_n, name);
1098 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1101 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1108 if (stat_config.aggr_mode == AGGR_NONE)
1111 if (stat_config.aggr_mode == AGGR_GLOBAL)
1114 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1115 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1117 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1123 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1125 FILE *output = stat_config.output;
1126 double sc = evsel->scale;
1130 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1133 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1135 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1138 aggr_printout(evsel, id, nr);
1140 fprintf(output, fmt, avg, csv_sep);
1143 fprintf(output, "%-*s%s",
1144 csv_output ? 0 : unit_width,
1145 evsel->unit, csv_sep);
1147 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1150 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1153 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1154 char *prefix, u64 run, u64 ena, double noise)
1156 struct perf_stat_output_ctx out;
1157 struct outstate os = {
1158 .fh = stat_config.output,
1159 .prefix = prefix ? prefix : "",
1164 print_metric_t pm = print_metric_std;
1168 nl = new_line_metric;
1170 pm = print_metric_only_csv;
1172 pm = print_metric_only;
1176 if (csv_output && !metric_only) {
1177 static int aggr_fields[] = {
1185 pm = print_metric_csv;
1188 os.nfields += aggr_fields[stat_config.aggr_mode];
1192 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1194 pm(&os, NULL, "", "", 0);
1197 aggr_printout(counter, id, nr);
1199 fprintf(stat_config.output, "%*s%s",
1200 csv_output ? 0 : 18,
1201 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1204 if (counter->supported)
1205 print_free_counters_hint = 1;
1207 fprintf(stat_config.output, "%-*s%s",
1208 csv_output ? 0 : unit_width,
1209 counter->unit, csv_sep);
1211 fprintf(stat_config.output, "%*s",
1212 csv_output ? 0 : -25,
1213 perf_evsel__name(counter));
1216 fprintf(stat_config.output, "%s%s",
1217 csv_sep, counter->cgrp->name);
1220 pm(&os, NULL, NULL, "", 0);
1221 print_noise(counter, noise);
1222 print_running(run, ena);
1224 pm(&os, NULL, NULL, "", 0);
1230 else if (nsec_counter(counter))
1231 nsec_printout(id, nr, counter, uval);
1233 abs_printout(id, nr, counter, uval);
1235 out.print_metric = pm;
1238 out.force_header = false;
1240 if (csv_output && !metric_only) {
1241 print_noise(counter, noise);
1242 print_running(run, ena);
1245 perf_stat__print_shadow_stats(counter, uval,
1246 first_shadow_cpu(counter, id),
1247 &out, &metric_events);
1248 if (!csv_output && !metric_only) {
1249 print_noise(counter, noise);
1250 print_running(run, ena);
1254 static void aggr_update_shadow(void)
1258 struct perf_evsel *counter;
1260 for (s = 0; s < aggr_map->nr; s++) {
1261 id = aggr_map->map[s];
1262 evlist__for_each_entry(evsel_list, counter) {
1264 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1265 s2 = aggr_get_id(evsel_list->cpus, cpu);
1268 val += perf_counts(counter->counts, cpu, 0)->val;
1270 perf_stat__update_shadow_stats(counter, val,
1271 first_shadow_cpu(counter, id));
1276 static void collect_all_aliases(struct perf_evsel *counter,
1277 void (*cb)(struct perf_evsel *counter, void *data,
1281 struct perf_evsel *alias;
1283 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1284 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1285 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1286 alias->scale != counter->scale ||
1287 alias->cgrp != counter->cgrp ||
1288 strcmp(alias->unit, counter->unit) ||
1289 nsec_counter(alias) != nsec_counter(counter))
1291 alias->merged_stat = true;
1292 cb(alias, data, false);
1296 static bool collect_data(struct perf_evsel *counter,
1297 void (*cb)(struct perf_evsel *counter, void *data,
1301 if (counter->merged_stat)
1303 cb(counter, data, true);
1304 if (!no_merge && counter->auto_merge_stats)
1305 collect_all_aliases(counter, cb, data);
1316 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1318 struct aggr_data *ad = data;
1321 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1322 struct perf_counts_values *counts;
1324 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1329 counts = perf_counts(counter->counts, cpu, 0);
1331 * When any result is bad, make them all to give
1332 * consistent output in interval mode.
1334 if (counts->ena == 0 || counts->run == 0 ||
1335 counter->counts->scaled == -1) {
1340 ad->val += counts->val;
1341 ad->ena += counts->ena;
1342 ad->run += counts->run;
1346 static void print_aggr(char *prefix)
1348 FILE *output = stat_config.output;
1349 struct perf_evsel *counter;
1355 if (!(aggr_map || aggr_get_id))
1358 aggr_update_shadow();
1361 * With metric_only everything is on a single line.
1362 * Without each counter has its own line.
1364 for (s = 0; s < aggr_map->nr; s++) {
1365 struct aggr_data ad;
1366 if (prefix && metric_only)
1367 fprintf(output, "%s", prefix);
1369 ad.id = id = aggr_map->map[s];
1371 evlist__for_each_entry(evsel_list, counter) {
1372 if (is_duration_time(counter))
1375 ad.val = ad.ena = ad.run = 0;
1377 if (!collect_data(counter, aggr_cb, &ad))
1383 if (first && metric_only) {
1385 aggr_printout(counter, id, nr);
1387 if (prefix && !metric_only)
1388 fprintf(output, "%s", prefix);
1390 uval = val * counter->scale;
1391 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1393 fputc('\n', output);
1396 fputc('\n', output);
1400 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1402 FILE *output = stat_config.output;
1403 int nthreads = thread_map__nr(counter->threads);
1404 int ncpus = cpu_map__nr(counter->cpus);
1408 for (thread = 0; thread < nthreads; thread++) {
1409 u64 ena = 0, run = 0, val = 0;
1411 for (cpu = 0; cpu < ncpus; cpu++) {
1412 val += perf_counts(counter->counts, cpu, thread)->val;
1413 ena += perf_counts(counter->counts, cpu, thread)->ena;
1414 run += perf_counts(counter->counts, cpu, thread)->run;
1418 fprintf(output, "%s", prefix);
1420 uval = val * counter->scale;
1421 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1422 fputc('\n', output);
1427 double avg, avg_enabled, avg_running;
1430 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1431 bool first __maybe_unused)
1433 struct caggr_data *cd = data;
1434 struct perf_stat_evsel *ps = counter->stats;
1436 cd->avg += avg_stats(&ps->res_stats[0]);
1437 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1438 cd->avg_running += avg_stats(&ps->res_stats[2]);
1442 * Print out the results of a single counter:
1443 * aggregated counts in system-wide mode
1445 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1447 FILE *output = stat_config.output;
1449 struct caggr_data cd = { .avg = 0.0 };
1451 if (!collect_data(counter, counter_aggr_cb, &cd))
1454 if (prefix && !metric_only)
1455 fprintf(output, "%s", prefix);
1457 uval = cd.avg * counter->scale;
1458 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1460 fprintf(output, "\n");
1463 static void counter_cb(struct perf_evsel *counter, void *data,
1464 bool first __maybe_unused)
1466 struct aggr_data *ad = data;
1468 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1469 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1470 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1474 * Print out the results of a single counter:
1475 * does not use aggregated count in system-wide
1477 static void print_counter(struct perf_evsel *counter, char *prefix)
1479 FILE *output = stat_config.output;
1484 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1485 struct aggr_data ad = { .cpu = cpu };
1487 if (!collect_data(counter, counter_cb, &ad))
1494 fprintf(output, "%s", prefix);
1496 uval = val * counter->scale;
1497 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1499 fputc('\n', output);
1503 static void print_no_aggr_metric(char *prefix)
1507 struct perf_evsel *counter;
1511 nrcpus = evsel_list->cpus->nr;
1512 for (cpu = 0; cpu < nrcpus; cpu++) {
1516 fputs(prefix, stat_config.output);
1517 evlist__for_each_entry(evsel_list, counter) {
1518 if (is_duration_time(counter))
1521 aggr_printout(counter, cpu, 0);
1524 val = perf_counts(counter->counts, cpu, 0)->val;
1525 ena = perf_counts(counter->counts, cpu, 0)->ena;
1526 run = perf_counts(counter->counts, cpu, 0)->run;
1528 uval = val * counter->scale;
1529 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1531 fputc('\n', stat_config.output);
1535 static int aggr_header_lens[] = {
1543 static const char *aggr_header_csv[] = {
1544 [AGGR_CORE] = "core,cpus,",
1545 [AGGR_SOCKET] = "socket,cpus",
1546 [AGGR_NONE] = "cpu,",
1547 [AGGR_THREAD] = "comm-pid,",
1551 static void print_metric_headers(const char *prefix, bool no_indent)
1553 struct perf_stat_output_ctx out;
1554 struct perf_evsel *counter;
1555 struct outstate os = {
1556 .fh = stat_config.output
1560 fprintf(stat_config.output, "%s", prefix);
1562 if (!csv_output && !no_indent)
1563 fprintf(stat_config.output, "%*s",
1564 aggr_header_lens[stat_config.aggr_mode], "");
1566 if (stat_config.interval)
1567 fputs("time,", stat_config.output);
1568 fputs(aggr_header_csv[stat_config.aggr_mode],
1569 stat_config.output);
1572 /* Print metrics headers only */
1573 evlist__for_each_entry(evsel_list, counter) {
1574 if (is_duration_time(counter))
1578 out.print_metric = print_metric_header;
1579 out.new_line = new_line_metric;
1580 out.force_header = true;
1582 perf_stat__print_shadow_stats(counter, 0,
1587 fputc('\n', stat_config.output);
1590 static void print_interval(char *prefix, struct timespec *ts)
1592 FILE *output = stat_config.output;
1593 static int num_print_interval;
1595 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1597 if (num_print_interval == 0 && !csv_output) {
1598 switch (stat_config.aggr_mode) {
1600 fprintf(output, "# time socket cpus");
1602 fprintf(output, " counts %*s events\n", unit_width, "unit");
1605 fprintf(output, "# time core cpus");
1607 fprintf(output, " counts %*s events\n", unit_width, "unit");
1610 fprintf(output, "# time CPU");
1612 fprintf(output, " counts %*s events\n", unit_width, "unit");
1615 fprintf(output, "# time comm-pid");
1617 fprintf(output, " counts %*s events\n", unit_width, "unit");
1621 fprintf(output, "# time");
1623 fprintf(output, " counts %*s events\n", unit_width, "unit");
1629 if (num_print_interval == 0 && metric_only)
1630 print_metric_headers(" ", true);
1631 if (++num_print_interval == 25)
1632 num_print_interval = 0;
1635 static void print_header(int argc, const char **argv)
1637 FILE *output = stat_config.output;
1643 fprintf(output, "\n");
1644 fprintf(output, " Performance counter stats for ");
1645 if (target.system_wide)
1646 fprintf(output, "\'system wide");
1647 else if (target.cpu_list)
1648 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1649 else if (!target__has_task(&target)) {
1650 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1651 for (i = 1; argv && (i < argc); i++)
1652 fprintf(output, " %s", argv[i]);
1653 } else if (target.pid)
1654 fprintf(output, "process id \'%s", target.pid);
1656 fprintf(output, "thread id \'%s", target.tid);
1658 fprintf(output, "\'");
1660 fprintf(output, " (%d runs)", run_count);
1661 fprintf(output, ":\n\n");
1665 static void print_footer(void)
1667 FILE *output = stat_config.output;
1671 fprintf(output, "\n");
1672 fprintf(output, " %17.9f seconds time elapsed",
1673 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1674 if (run_count > 1) {
1675 fprintf(output, " ");
1676 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1677 avg_stats(&walltime_nsecs_stats));
1679 fprintf(output, "\n\n");
1681 if (print_free_counters_hint &&
1682 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1685 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1686 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1688 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1691 static void print_counters(struct timespec *ts, int argc, const char **argv)
1693 int interval = stat_config.interval;
1694 struct perf_evsel *counter;
1695 char buf[64], *prefix = NULL;
1697 /* Do not print anything if we record to the pipe. */
1698 if (STAT_RECORD && perf_stat.data.is_pipe)
1702 print_interval(prefix = buf, ts);
1704 print_header(argc, argv);
1707 static int num_print_iv;
1709 if (num_print_iv == 0 && !interval)
1710 print_metric_headers(prefix, false);
1711 if (num_print_iv++ == 25)
1713 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1714 fprintf(stat_config.output, "%s", prefix);
1717 switch (stat_config.aggr_mode) {
1723 evlist__for_each_entry(evsel_list, counter) {
1724 if (is_duration_time(counter))
1726 print_aggr_thread(counter, prefix);
1730 evlist__for_each_entry(evsel_list, counter) {
1731 if (is_duration_time(counter))
1733 print_counter_aggr(counter, prefix);
1736 fputc('\n', stat_config.output);
1740 print_no_aggr_metric(prefix);
1742 evlist__for_each_entry(evsel_list, counter) {
1743 if (is_duration_time(counter))
1745 print_counter(counter, prefix);
1754 if (!interval && !csv_output)
1757 fflush(stat_config.output);
1760 static volatile int signr = -1;
1762 static void skip_signal(int signo)
1764 if ((child_pid == -1) || stat_config.interval)
1769 * render child_pid harmless
1770 * won't send SIGTERM to a random
1771 * process in case of race condition
1772 * and fast PID recycling
1777 static void sig_atexit(void)
1782 * avoid race condition with SIGCHLD handler
1783 * in skip_signal() which is modifying child_pid
1784 * goal is to avoid send SIGTERM to a random
1788 sigaddset(&set, SIGCHLD);
1789 sigprocmask(SIG_BLOCK, &set, &oset);
1791 if (child_pid != -1)
1792 kill(child_pid, SIGTERM);
1794 sigprocmask(SIG_SETMASK, &oset, NULL);
1799 signal(signr, SIG_DFL);
1800 kill(getpid(), signr);
1803 static int stat__set_big_num(const struct option *opt __maybe_unused,
1804 const char *s __maybe_unused, int unset)
1806 big_num_opt = unset ? 0 : 1;
1810 static int enable_metric_only(const struct option *opt __maybe_unused,
1811 const char *s __maybe_unused, int unset)
1813 force_metric_only = true;
1814 metric_only = !unset;
1818 static int parse_metric_groups(const struct option *opt,
1820 int unset __maybe_unused)
1822 return metricgroup__parse_groups(opt, str, &metric_events);
1825 static const struct option stat_options[] = {
1826 OPT_BOOLEAN('T', "transaction", &transaction_run,
1827 "hardware transaction statistics"),
1828 OPT_CALLBACK('e', "event", &evsel_list, "event",
1829 "event selector. use 'perf list' to list available events",
1830 parse_events_option),
1831 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1832 "event filter", parse_filter),
1833 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1834 "child tasks do not inherit counters"),
1835 OPT_STRING('p', "pid", &target.pid, "pid",
1836 "stat events on existing process id"),
1837 OPT_STRING('t', "tid", &target.tid, "tid",
1838 "stat events on existing thread id"),
1839 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1840 "system-wide collection from all CPUs"),
1841 OPT_BOOLEAN('g', "group", &group,
1842 "put the counters into a counter group"),
1843 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1844 OPT_INCR('v', "verbose", &verbose,
1845 "be more verbose (show counter open errors, etc)"),
1846 OPT_INTEGER('r', "repeat", &run_count,
1847 "repeat command and print average + stddev (max: 100, forever: 0)"),
1848 OPT_BOOLEAN('n', "null", &null_run,
1849 "null run - dont start any counters"),
1850 OPT_INCR('d', "detailed", &detailed_run,
1851 "detailed run - start a lot of events"),
1852 OPT_BOOLEAN('S', "sync", &sync_run,
1853 "call sync() before starting a run"),
1854 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1855 "print large numbers with thousands\' separators",
1857 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1858 "list of cpus to monitor in system-wide"),
1859 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1860 "disable CPU count aggregation", AGGR_NONE),
1861 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1862 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1863 "print counts with custom separator"),
1864 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1865 "monitor event in cgroup name only", parse_cgroups),
1866 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1867 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1868 OPT_INTEGER(0, "log-fd", &output_fd,
1869 "log output to fd, instead of stderr"),
1870 OPT_STRING(0, "pre", &pre_cmd, "command",
1871 "command to run prior to the measured command"),
1872 OPT_STRING(0, "post", &post_cmd, "command",
1873 "command to run after to the measured command"),
1874 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1875 "print counts at regular interval in ms (>= 10)"),
1876 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1877 "aggregate counts per processor socket", AGGR_SOCKET),
1878 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1879 "aggregate counts per physical processor core", AGGR_CORE),
1880 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1881 "aggregate counts per thread", AGGR_THREAD),
1882 OPT_UINTEGER('D', "delay", &initial_delay,
1883 "ms to wait before starting measurement after program start"),
1884 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1885 "Only print computed metrics. No raw values", enable_metric_only),
1886 OPT_BOOLEAN(0, "topdown", &topdown_run,
1887 "measure topdown level 1 statistics"),
1888 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1889 "measure SMI cost"),
1890 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1891 "monitor specified metrics or metric groups (separated by ,)",
1892 parse_metric_groups),
1896 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1898 return cpu_map__get_socket(map, cpu, NULL);
1901 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1903 return cpu_map__get_core(map, cpu, NULL);
1906 static int cpu_map__get_max(struct cpu_map *map)
1910 for (i = 0; i < map->nr; i++) {
1911 if (map->map[i] > max)
1918 static struct cpu_map *cpus_aggr_map;
1920 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1927 cpu = map->map[idx];
1929 if (cpus_aggr_map->map[cpu] == -1)
1930 cpus_aggr_map->map[cpu] = get_id(map, idx);
1932 return cpus_aggr_map->map[cpu];
1935 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1937 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1940 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1942 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1945 static int perf_stat_init_aggr_mode(void)
1949 switch (stat_config.aggr_mode) {
1951 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1952 perror("cannot build socket map");
1955 aggr_get_id = perf_stat__get_socket_cached;
1958 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1959 perror("cannot build core map");
1962 aggr_get_id = perf_stat__get_core_cached;
1973 * The evsel_list->cpus is the base we operate on,
1974 * taking the highest cpu number to be the size of
1975 * the aggregation translate cpumap.
1977 nr = cpu_map__get_max(evsel_list->cpus);
1978 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1979 return cpus_aggr_map ? 0 : -ENOMEM;
1982 static void perf_stat__exit_aggr_mode(void)
1984 cpu_map__put(aggr_map);
1985 cpu_map__put(cpus_aggr_map);
1987 cpus_aggr_map = NULL;
1990 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1997 cpu = map->map[idx];
1999 if (cpu >= env->nr_cpus_avail)
2005 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2007 struct perf_env *env = data;
2008 int cpu = perf_env__get_cpu(env, map, idx);
2010 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2013 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2015 struct perf_env *env = data;
2016 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2019 int socket_id = env->cpu[cpu].socket_id;
2022 * Encode socket in upper 16 bits
2023 * core_id is relative to socket, and
2024 * we need a global id. So we combine
2027 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2033 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2034 struct cpu_map **sockp)
2036 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2039 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2040 struct cpu_map **corep)
2042 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2045 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2047 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2050 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2052 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2055 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2057 struct perf_env *env = &st->session->header.env;
2059 switch (stat_config.aggr_mode) {
2061 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2062 perror("cannot build socket map");
2065 aggr_get_id = perf_stat__get_socket_file;
2068 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2069 perror("cannot build core map");
2072 aggr_get_id = perf_stat__get_core_file;
2085 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2092 for (i = 0; attr[i]; i++) {
2093 if (pmu_have_event("cpu", attr[i])) {
2094 len += strlen(attr[i]) + 1;
2095 attr[i - off] = attr[i];
2099 attr[i - off] = NULL;
2101 *str = malloc(len + 1 + 2);
2111 for (i = 0; attr[i]; i++) {
2124 __weak bool arch_topdown_check_group(bool *warn)
2130 __weak void arch_topdown_group_warn(void)
2135 * Add default attributes, if there were no attributes specified or
2136 * if -d/--detailed, -d -d or -d -d -d is used:
2138 static int add_default_attributes(void)
2141 struct perf_event_attr default_attrs0[] = {
2143 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2144 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2145 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2146 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2148 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2150 struct perf_event_attr frontend_attrs[] = {
2151 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2153 struct perf_event_attr backend_attrs[] = {
2154 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2156 struct perf_event_attr default_attrs1[] = {
2157 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2158 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2159 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2164 * Detailed stats (-d), covering the L1 and last level data caches:
2166 struct perf_event_attr detailed_attrs[] = {
2168 { .type = PERF_TYPE_HW_CACHE,
2170 PERF_COUNT_HW_CACHE_L1D << 0 |
2171 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2172 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2174 { .type = PERF_TYPE_HW_CACHE,
2176 PERF_COUNT_HW_CACHE_L1D << 0 |
2177 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2178 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2180 { .type = PERF_TYPE_HW_CACHE,
2182 PERF_COUNT_HW_CACHE_LL << 0 |
2183 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2184 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2186 { .type = PERF_TYPE_HW_CACHE,
2188 PERF_COUNT_HW_CACHE_LL << 0 |
2189 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2190 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2194 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2196 struct perf_event_attr very_detailed_attrs[] = {
2198 { .type = PERF_TYPE_HW_CACHE,
2200 PERF_COUNT_HW_CACHE_L1I << 0 |
2201 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2202 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2204 { .type = PERF_TYPE_HW_CACHE,
2206 PERF_COUNT_HW_CACHE_L1I << 0 |
2207 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2208 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2210 { .type = PERF_TYPE_HW_CACHE,
2212 PERF_COUNT_HW_CACHE_DTLB << 0 |
2213 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2214 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2216 { .type = PERF_TYPE_HW_CACHE,
2218 PERF_COUNT_HW_CACHE_DTLB << 0 |
2219 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2220 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2222 { .type = PERF_TYPE_HW_CACHE,
2224 PERF_COUNT_HW_CACHE_ITLB << 0 |
2225 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2226 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2228 { .type = PERF_TYPE_HW_CACHE,
2230 PERF_COUNT_HW_CACHE_ITLB << 0 |
2231 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2232 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2237 * Very, very detailed stats (-d -d -d), adding prefetch events:
2239 struct perf_event_attr very_very_detailed_attrs[] = {
2241 { .type = PERF_TYPE_HW_CACHE,
2243 PERF_COUNT_HW_CACHE_L1D << 0 |
2244 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2245 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2247 { .type = PERF_TYPE_HW_CACHE,
2249 PERF_COUNT_HW_CACHE_L1D << 0 |
2250 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2251 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2254 /* Set attrs if no event is selected and !null_run: */
2258 if (transaction_run) {
2259 if (pmu_have_event("cpu", "cycles-ct") &&
2260 pmu_have_event("cpu", "el-start"))
2261 err = parse_events(evsel_list, transaction_attrs, NULL);
2263 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2265 fprintf(stderr, "Cannot set up transaction events\n");
2274 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2275 fprintf(stderr, "freeze_on_smi is not supported.\n");
2280 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2281 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2287 if (pmu_have_event("msr", "aperf") &&
2288 pmu_have_event("msr", "smi")) {
2289 if (!force_metric_only)
2291 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2293 fprintf(stderr, "To measure SMI cost, it needs "
2294 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2298 fprintf(stderr, "Cannot set up SMI cost events\n");
2308 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2309 stat_config.aggr_mode != AGGR_CORE) {
2310 pr_err("top down event configuration requires --per-core mode\n");
2313 stat_config.aggr_mode = AGGR_CORE;
2314 if (nr_cgroups || !target__has_cpu(&target)) {
2315 pr_err("top down event configuration requires system-wide mode (-a)\n");
2319 if (!force_metric_only)
2321 if (topdown_filter_events(topdown_attrs, &str,
2322 arch_topdown_check_group(&warn)) < 0) {
2323 pr_err("Out of memory\n");
2326 if (topdown_attrs[0] && str) {
2328 arch_topdown_group_warn();
2329 err = parse_events(evsel_list, str, NULL);
2332 "Cannot set up top down events %s: %d\n",
2338 fprintf(stderr, "System does not support topdown\n");
2344 if (!evsel_list->nr_entries) {
2345 if (target__has_cpu(&target))
2346 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2348 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2350 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2351 if (perf_evlist__add_default_attrs(evsel_list,
2352 frontend_attrs) < 0)
2355 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2356 if (perf_evlist__add_default_attrs(evsel_list,
2360 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2364 /* Detailed events get appended to the event list: */
2366 if (detailed_run < 1)
2369 /* Append detailed run extra attributes: */
2370 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2373 if (detailed_run < 2)
2376 /* Append very detailed run extra attributes: */
2377 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2380 if (detailed_run < 3)
2383 /* Append very, very detailed run extra attributes: */
2384 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2387 static const char * const stat_record_usage[] = {
2388 "perf stat record [<options>]",
2392 static void init_features(struct perf_session *session)
2396 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2397 perf_header__set_feat(&session->header, feat);
2399 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2400 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2401 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2402 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2405 static int __cmd_record(int argc, const char **argv)
2407 struct perf_session *session;
2408 struct perf_data *data = &perf_stat.data;
2410 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2411 PARSE_OPT_STOP_AT_NON_OPTION);
2414 data->file.path = output_name;
2416 if (run_count != 1 || forever) {
2417 pr_err("Cannot use -r option with perf stat record.\n");
2421 session = perf_session__new(data, false, NULL);
2422 if (session == NULL) {
2423 pr_err("Perf session creation failed.\n");
2427 init_features(session);
2429 session->evlist = evsel_list;
2430 perf_stat.session = session;
2431 perf_stat.record = true;
2435 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2436 union perf_event *event,
2437 struct perf_session *session)
2439 struct stat_round_event *stat_round = &event->stat_round;
2440 struct perf_evsel *counter;
2441 struct timespec tsh, *ts = NULL;
2442 const char **argv = session->header.env.cmdline_argv;
2443 int argc = session->header.env.nr_cmdline;
2445 evlist__for_each_entry(evsel_list, counter)
2446 perf_stat_process_counter(&stat_config, counter);
2448 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2449 update_stats(&walltime_nsecs_stats, stat_round->time);
2451 if (stat_config.interval && stat_round->time) {
2452 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2453 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2457 print_counters(ts, argc, argv);
2462 int process_stat_config_event(struct perf_tool *tool,
2463 union perf_event *event,
2464 struct perf_session *session __maybe_unused)
2466 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2468 perf_event__read_stat_config(&stat_config, &event->stat_config);
2470 if (cpu_map__empty(st->cpus)) {
2471 if (st->aggr_mode != AGGR_UNSET)
2472 pr_warning("warning: processing task data, aggregation mode not set\n");
2476 if (st->aggr_mode != AGGR_UNSET)
2477 stat_config.aggr_mode = st->aggr_mode;
2479 if (perf_stat.data.is_pipe)
2480 perf_stat_init_aggr_mode();
2482 perf_stat_init_aggr_mode_file(st);
2487 static int set_maps(struct perf_stat *st)
2489 if (!st->cpus || !st->threads)
2492 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2495 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2497 if (perf_evlist__alloc_stats(evsel_list, true))
2500 st->maps_allocated = true;
2505 int process_thread_map_event(struct perf_tool *tool,
2506 union perf_event *event,
2507 struct perf_session *session __maybe_unused)
2509 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2512 pr_warning("Extra thread map event, ignoring.\n");
2516 st->threads = thread_map__new_event(&event->thread_map);
2520 return set_maps(st);
2524 int process_cpu_map_event(struct perf_tool *tool,
2525 union perf_event *event,
2526 struct perf_session *session __maybe_unused)
2528 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2529 struct cpu_map *cpus;
2532 pr_warning("Extra cpu map event, ignoring.\n");
2536 cpus = cpu_map__new_data(&event->cpu_map.data);
2541 return set_maps(st);
2544 static const char * const stat_report_usage[] = {
2545 "perf stat report [<options>]",
2549 static struct perf_stat perf_stat = {
2551 .attr = perf_event__process_attr,
2552 .event_update = perf_event__process_event_update,
2553 .thread_map = process_thread_map_event,
2554 .cpu_map = process_cpu_map_event,
2555 .stat_config = process_stat_config_event,
2556 .stat = perf_event__process_stat_event,
2557 .stat_round = process_stat_round_event,
2559 .aggr_mode = AGGR_UNSET,
2562 static int __cmd_report(int argc, const char **argv)
2564 struct perf_session *session;
2565 const struct option options[] = {
2566 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2567 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2568 "aggregate counts per processor socket", AGGR_SOCKET),
2569 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2570 "aggregate counts per physical processor core", AGGR_CORE),
2571 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2572 "disable CPU count aggregation", AGGR_NONE),
2578 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2580 if (!input_name || !strlen(input_name)) {
2581 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2584 input_name = "perf.data";
2587 perf_stat.data.file.path = input_name;
2588 perf_stat.data.mode = PERF_DATA_MODE_READ;
2590 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2591 if (session == NULL)
2594 perf_stat.session = session;
2595 stat_config.output = stderr;
2596 evsel_list = session->evlist;
2598 ret = perf_session__process_events(session);
2602 perf_session__delete(session);
2606 static void setup_system_wide(int forks)
2609 * Make system wide (-a) the default target if
2610 * no target was specified and one of following
2611 * conditions is met:
2613 * - there's no workload specified
2614 * - there is workload specified but all requested
2615 * events are system wide events
2617 if (!target__none(&target))
2621 target.system_wide = true;
2623 struct perf_evsel *counter;
2625 evlist__for_each_entry(evsel_list, counter) {
2626 if (!counter->system_wide)
2630 if (evsel_list->nr_entries)
2631 target.system_wide = true;
2635 int cmd_stat(int argc, const char **argv)
2637 const char * const stat_usage[] = {
2638 "perf stat [<options>] [<command>]",
2641 int status = -EINVAL, run_idx;
2643 FILE *output = stderr;
2644 unsigned int interval;
2645 const char * const stat_subcommands[] = { "record", "report" };
2647 setlocale(LC_ALL, "");
2649 evsel_list = perf_evlist__new();
2650 if (evsel_list == NULL)
2653 parse_events__shrink_config_terms();
2654 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2655 (const char **) stat_usage,
2656 PARSE_OPT_STOP_AT_NON_OPTION);
2657 perf_stat__collect_metric_expr(evsel_list);
2658 perf_stat__init_shadow_stats();
2662 if (!strcmp(csv_sep, "\\t"))
2665 csv_sep = DEFAULT_SEPARATOR;
2667 if (argc && !strncmp(argv[0], "rec", 3)) {
2668 argc = __cmd_record(argc, argv);
2671 } else if (argc && !strncmp(argv[0], "rep", 3))
2672 return __cmd_report(argc, argv);
2674 interval = stat_config.interval;
2677 * For record command the -o is already taken care of.
2679 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2682 if (output_name && output_fd) {
2683 fprintf(stderr, "cannot use both --output and --log-fd\n");
2684 parse_options_usage(stat_usage, stat_options, "o", 1);
2685 parse_options_usage(NULL, stat_options, "log-fd", 0);
2689 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2690 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2694 if (metric_only && run_count > 1) {
2695 fprintf(stderr, "--metric-only is not supported with -r\n");
2699 if (output_fd < 0) {
2700 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2701 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2707 mode = append_file ? "a" : "w";
2709 output = fopen(output_name, mode);
2711 perror("failed to create output file");
2714 clock_gettime(CLOCK_REALTIME, &tm);
2715 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2716 } else if (output_fd > 0) {
2717 mode = append_file ? "a" : "w";
2718 output = fdopen(output_fd, mode);
2720 perror("Failed opening logfd");
2725 stat_config.output = output;
2728 * let the spreadsheet do the pretty-printing
2731 /* User explicitly passed -B? */
2732 if (big_num_opt == 1) {
2733 fprintf(stderr, "-B option not supported with -x\n");
2734 parse_options_usage(stat_usage, stat_options, "B", 1);
2735 parse_options_usage(NULL, stat_options, "x", 1);
2737 } else /* Nope, so disable big number formatting */
2739 } else if (big_num_opt == 0) /* User passed --no-big-num */
2742 setup_system_wide(argc);
2744 if (run_count < 0) {
2745 pr_err("Run count must be a positive number\n");
2746 parse_options_usage(stat_usage, stat_options, "r", 1);
2748 } else if (run_count == 0) {
2753 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2754 fprintf(stderr, "The --per-thread option is only available "
2755 "when monitoring via -p -t options.\n");
2756 parse_options_usage(NULL, stat_options, "p", 1);
2757 parse_options_usage(NULL, stat_options, "t", 1);
2762 * no_aggr, cgroup are for system-wide only
2763 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2765 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2766 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2767 !target__has_cpu(&target)) {
2768 fprintf(stderr, "both cgroup and no-aggregation "
2769 "modes only available in system-wide mode\n");
2771 parse_options_usage(stat_usage, stat_options, "G", 1);
2772 parse_options_usage(NULL, stat_options, "A", 1);
2773 parse_options_usage(NULL, stat_options, "a", 1);
2777 if (add_default_attributes())
2780 target__validate(&target);
2782 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2783 if (target__has_task(&target)) {
2784 pr_err("Problems finding threads of monitor\n");
2785 parse_options_usage(stat_usage, stat_options, "p", 1);
2786 parse_options_usage(NULL, stat_options, "t", 1);
2787 } else if (target__has_cpu(&target)) {
2788 perror("failed to parse CPUs map");
2789 parse_options_usage(stat_usage, stat_options, "C", 1);
2790 parse_options_usage(NULL, stat_options, "a", 1);
2796 * Initialize thread_map with comm names,
2797 * so we could print it out on output.
2799 if (stat_config.aggr_mode == AGGR_THREAD)
2800 thread_map__read_comms(evsel_list->threads);
2802 if (interval && interval < 100) {
2803 if (interval < 10) {
2804 pr_err("print interval must be >= 10ms\n");
2805 parse_options_usage(stat_usage, stat_options, "I", 1);
2808 pr_warning("print interval < 100ms. "
2809 "The overhead percentage could be high in some cases. "
2810 "Please proceed with caution.\n");
2813 if (perf_evlist__alloc_stats(evsel_list, interval))
2816 if (perf_stat_init_aggr_mode())
2820 * We dont want to block the signals - that would cause
2821 * child tasks to inherit that and Ctrl-C would not work.
2822 * What we want is for Ctrl-C to work in the exec()-ed
2823 * task, but being ignored by perf stat itself:
2827 signal(SIGINT, skip_signal);
2828 signal(SIGCHLD, skip_signal);
2829 signal(SIGALRM, skip_signal);
2830 signal(SIGABRT, skip_signal);
2833 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2834 if (run_count != 1 && verbose > 0)
2835 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2838 status = run_perf_stat(argc, argv);
2839 if (forever && status != -1) {
2840 print_counters(NULL, argc, argv);
2841 perf_stat__reset_stats();
2845 if (!forever && status != -1 && !interval)
2846 print_counters(NULL, argc, argv);
2850 * We synthesize the kernel mmap record just so that older tools
2851 * don't emit warnings about not being able to resolve symbols
2852 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2853 * a saner message about no samples being in the perf.data file.
2855 * This also serves to suppress a warning about f_header.data.size == 0
2856 * in header.c at the moment 'perf stat record' gets introduced, which
2857 * is not really needed once we start adding the stat specific PERF_RECORD_
2858 * records, but the need to suppress the kptr_restrict messages in older
2859 * tools remain -acme
2861 int fd = perf_data__fd(&perf_stat.data);
2862 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2863 process_synthesized_event,
2864 &perf_stat.session->machines.host);
2866 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2867 "older tools may produce warnings about this file\n.");
2871 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2872 pr_err("failed to write stat round event\n");
2875 if (!perf_stat.data.is_pipe) {
2876 perf_stat.session->header.data_size += perf_stat.bytes_written;
2877 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2880 perf_session__delete(perf_stat.session);
2883 perf_stat__exit_aggr_mode();
2884 perf_evlist__free_stats(evsel_list);
2886 if (smi_cost && smi_reset)
2887 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2889 perf_evlist__delete(evsel_list);