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)
219 perf_evlist__reset_stats(evsel_list);
220 perf_stat__reset_shadow_stats();
222 for (i = 0; i < stat_config.stats_num; i++)
223 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
226 static int create_perf_stat_counter(struct perf_evsel *evsel)
228 struct perf_event_attr *attr = &evsel->attr;
229 struct perf_evsel *leader = evsel->leader;
231 if (stat_config.scale) {
232 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
233 PERF_FORMAT_TOTAL_TIME_RUNNING;
237 * The event is part of non trivial group, let's enable
238 * the group read (for leader) and ID retrieval for all
241 if (leader->nr_members > 1)
242 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
244 attr->inherit = !no_inherit;
247 * Some events get initialized with sample_(period/type) set,
248 * like tracepoints. Clear it up for counting.
250 attr->sample_period = 0;
253 * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
254 * while avoiding that older tools show confusing messages.
256 * However for pipe sessions we need to keep it zero,
257 * because script's perf_evsel__check_attr is triggered
258 * by attr->sample_type != 0, and we can't run it on
261 if (!(STAT_RECORD && perf_stat.data.is_pipe))
262 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
265 * Disabling all counters initially, they will be enabled
266 * either manually by us or by kernel via enable_on_exec
269 if (perf_evsel__is_group_leader(evsel)) {
273 * In case of initial_delay we enable tracee
276 if (target__none(&target) && !initial_delay)
277 attr->enable_on_exec = 1;
280 if (target__has_cpu(&target) && !target__has_per_thread(&target))
281 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
283 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
287 * Does the counter have nsecs as a unit?
289 static inline int nsec_counter(struct perf_evsel *evsel)
291 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
292 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
298 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
299 union perf_event *event,
300 struct perf_sample *sample __maybe_unused,
301 struct machine *machine __maybe_unused)
303 if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
304 pr_err("failed to write perf data, error: %m\n");
308 perf_stat.bytes_written += event->header.size;
312 static int write_stat_round_event(u64 tm, u64 type)
314 return perf_event__synthesize_stat_round(NULL, tm, type,
315 process_synthesized_event,
319 #define WRITE_STAT_ROUND_EVENT(time, interval) \
320 write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
322 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
325 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
326 struct perf_counts_values *count)
328 struct perf_sample_id *sid = SID(counter, cpu, thread);
330 return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
331 process_synthesized_event, NULL);
335 * Read out the results of a single counter:
336 * do not aggregate counts across CPUs in system-wide mode
338 static int read_counter(struct perf_evsel *counter)
340 int nthreads = thread_map__nr(evsel_list->threads);
341 int ncpus, cpu, thread;
343 if (target__has_cpu(&target) && !target__has_per_thread(&target))
344 ncpus = perf_evsel__nr_cpus(counter);
348 if (!counter->supported)
351 if (counter->system_wide)
354 for (thread = 0; thread < nthreads; thread++) {
355 for (cpu = 0; cpu < ncpus; cpu++) {
356 struct perf_counts_values *count;
358 count = perf_counts(counter->counts, cpu, thread);
361 * The leader's group read loads data into its group members
362 * (via perf_evsel__read_counter) and sets threir count->loaded.
364 if (!count->loaded &&
365 perf_evsel__read_counter(counter, cpu, thread)) {
366 counter->counts->scaled = -1;
367 perf_counts(counter->counts, cpu, thread)->ena = 0;
368 perf_counts(counter->counts, cpu, thread)->run = 0;
372 count->loaded = false;
375 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
376 pr_err("failed to write stat event\n");
382 fprintf(stat_config.output,
383 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
384 perf_evsel__name(counter),
386 count->val, count->ena, count->run);
394 static void read_counters(void)
396 struct perf_evsel *counter;
399 evlist__for_each_entry(evsel_list, counter) {
400 ret = read_counter(counter);
402 pr_debug("failed to read counter %s\n", counter->name);
404 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
405 pr_warning("failed to process counter %s\n", counter->name);
409 static void process_interval(void)
411 struct timespec ts, rs;
415 clock_gettime(CLOCK_MONOTONIC, &ts);
416 diff_timespec(&rs, &ts, &ref_time);
419 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
420 pr_err("failed to write stat round event\n");
423 init_stats(&walltime_nsecs_stats);
424 update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
425 print_counters(&rs, 0, NULL);
428 static void enable_counters(void)
431 usleep(initial_delay * USEC_PER_MSEC);
434 * We need to enable counters only if:
435 * - we don't have tracee (attaching to task or cpu)
436 * - we have initial delay configured
438 if (!target__none(&target) || initial_delay)
439 perf_evlist__enable(evsel_list);
442 static void disable_counters(void)
445 * If we don't have tracee (attaching to task or cpu), counters may
446 * still be running. To get accurate group ratios, we must stop groups
447 * from counting before reading their constituent counters.
449 if (!target__none(&target))
450 perf_evlist__disable(evsel_list);
453 static volatile int workload_exec_errno;
456 * perf_evlist__prepare_workload will send a SIGUSR1
457 * if the fork fails, since we asked by setting its
458 * want_signal to true.
460 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
461 void *ucontext __maybe_unused)
463 workload_exec_errno = info->si_value.sival_int;
466 static int perf_stat_synthesize_config(bool is_pipe)
471 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
472 process_synthesized_event);
474 pr_err("Couldn't synthesize attrs.\n");
479 err = perf_event__synthesize_extra_attr(NULL,
481 process_synthesized_event,
484 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
485 process_synthesized_event,
488 pr_err("Couldn't synthesize thread map.\n");
492 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
493 process_synthesized_event, NULL);
495 pr_err("Couldn't synthesize thread map.\n");
499 err = perf_event__synthesize_stat_config(NULL, &stat_config,
500 process_synthesized_event, NULL);
502 pr_err("Couldn't synthesize config.\n");
509 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
511 static int __store_counter_ids(struct perf_evsel *counter,
512 struct cpu_map *cpus,
513 struct thread_map *threads)
517 for (cpu = 0; cpu < cpus->nr; cpu++) {
518 for (thread = 0; thread < threads->nr; thread++) {
519 int fd = FD(counter, cpu, thread);
521 if (perf_evlist__id_add_fd(evsel_list, counter,
522 cpu, thread, fd) < 0)
530 static int store_counter_ids(struct perf_evsel *counter)
532 struct cpu_map *cpus = counter->cpus;
533 struct thread_map *threads = counter->threads;
535 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
538 return __store_counter_ids(counter, cpus, threads);
541 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
543 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
546 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
548 struct perf_evsel *c2, *leader;
551 leader = evsel->leader;
552 pr_debug("Weak group for %s/%d failed\n",
553 leader->name, leader->nr_members);
556 * for_each_group_member doesn't work here because it doesn't
557 * include the first entry.
559 evlist__for_each_entry(evsel_list, c2) {
562 if (c2->leader == leader) {
564 perf_evsel__close(c2);
572 static int __run_perf_stat(int argc, const char **argv)
574 int interval = stat_config.interval;
576 unsigned long long t0, t1;
577 struct perf_evsel *counter;
581 const bool forks = (argc > 0);
582 bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
583 struct perf_evsel_config_term *err_term;
586 ts.tv_sec = interval / USEC_PER_MSEC;
587 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
594 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
595 workload_exec_failed_signal) < 0) {
596 perror("failed to prepare workload");
599 child_pid = evsel_list->workload.pid;
603 perf_evlist__set_leader(evsel_list);
605 evlist__for_each_entry(evsel_list, counter) {
607 if (create_perf_stat_counter(counter) < 0) {
609 /* Weak group failed. Reset the group. */
610 if ((errno == EINVAL || errno == EBADF) &&
611 counter->leader != counter &&
612 counter->weak_group) {
613 counter = perf_evsel__reset_weak_group(counter);
618 * PPC returns ENXIO for HW counters until 2.6.37
619 * (behavior changed with commit b0a873e).
621 if (errno == EINVAL || errno == ENOSYS ||
622 errno == ENOENT || errno == EOPNOTSUPP ||
625 ui__warning("%s event is not supported by the kernel.\n",
626 perf_evsel__name(counter));
627 counter->supported = false;
629 if ((counter->leader != counter) ||
630 !(counter->leader->nr_members > 1))
632 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
634 ui__warning("%s\n", msg);
638 perf_evsel__open_strerror(counter, &target,
639 errno, msg, sizeof(msg));
640 ui__error("%s\n", msg);
643 kill(child_pid, SIGTERM);
647 counter->supported = true;
649 l = strlen(counter->unit);
653 if (perf_evsel__should_store_id(counter) &&
654 store_counter_ids(counter))
658 if (perf_evlist__apply_filters(evsel_list, &counter)) {
659 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
660 counter->filter, perf_evsel__name(counter), errno,
661 str_error_r(errno, msg, sizeof(msg)));
665 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
666 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
667 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
668 str_error_r(errno, msg, sizeof(msg)));
673 int err, fd = perf_data__fd(&perf_stat.data);
676 err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
678 err = perf_session__write_header(perf_stat.session, evsel_list,
685 err = perf_stat_synthesize_config(is_pipe);
691 * Enable counters and exec the command:
694 clock_gettime(CLOCK_MONOTONIC, &ref_time);
697 perf_evlist__start_workload(evsel_list);
701 while (!waitpid(child_pid, &status, WNOHANG)) {
702 nanosleep(&ts, NULL);
706 waitpid(child_pid, &status, 0);
708 if (workload_exec_errno) {
709 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
710 pr_err("Workload failed: %s\n", emsg);
714 if (WIFSIGNALED(status))
715 psignal(WTERMSIG(status), argv[0]);
719 nanosleep(&ts, NULL);
729 update_stats(&walltime_nsecs_stats, t1 - t0);
732 * Closing a group leader splits the group, and as we only disable
733 * group leaders, results in remaining events becoming enabled. To
734 * avoid arbitrary skew, we must read all counters before closing any
738 perf_evlist__close(evsel_list);
740 return WEXITSTATUS(status);
743 static int run_perf_stat(int argc, const char **argv)
748 ret = system(pre_cmd);
756 ret = __run_perf_stat(argc, argv);
761 ret = system(post_cmd);
769 static void print_running(u64 run, u64 ena)
772 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
776 ena ? 100.0 * run / ena : 100.0);
777 } else if (run != ena) {
778 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
782 static void print_noise_pct(double total, double avg)
784 double pct = rel_stddev_stats(total, avg);
787 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
789 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
792 static void print_noise(struct perf_evsel *evsel, double avg)
794 struct perf_stat_evsel *ps;
800 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
803 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
805 switch (stat_config.aggr_mode) {
807 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
808 cpu_map__id_to_socket(id),
810 cpu_map__id_to_cpu(id),
817 fprintf(stat_config.output, "S%*d%s%*d%s",
826 fprintf(stat_config.output, "CPU%*d%s",
828 perf_evsel__cpus(evsel)->map[id], csv_sep);
831 fprintf(stat_config.output, "%*s-%*d%s",
833 thread_map__comm(evsel->threads, id),
835 thread_map__pid(evsel->threads, id),
851 struct perf_evsel *evsel;
854 #define METRIC_LEN 35
856 static void new_line_std(void *ctx)
858 struct outstate *os = ctx;
863 static void do_new_line_std(struct outstate *os)
866 fputs(os->prefix, os->fh);
867 aggr_printout(os->evsel, os->id, os->nr);
868 if (stat_config.aggr_mode == AGGR_NONE)
869 fprintf(os->fh, " ");
870 fprintf(os->fh, " ");
873 static void print_metric_std(void *ctx, const char *color, const char *fmt,
874 const char *unit, double val)
876 struct outstate *os = ctx;
879 bool newline = os->newline;
883 if (unit == NULL || fmt == NULL) {
884 fprintf(out, "%-*s", METRIC_LEN, "");
891 n = fprintf(out, " # ");
893 n += color_fprintf(out, color, fmt, val);
895 n += fprintf(out, fmt, val);
896 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
899 static void new_line_csv(void *ctx)
901 struct outstate *os = ctx;
906 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
907 aggr_printout(os->evsel, os->id, os->nr);
908 for (i = 0; i < os->nfields; i++)
909 fputs(csv_sep, os->fh);
912 static void print_metric_csv(void *ctx,
913 const char *color __maybe_unused,
914 const char *fmt, const char *unit, double val)
916 struct outstate *os = ctx;
918 char buf[64], *vals, *ends;
920 if (unit == NULL || fmt == NULL) {
921 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
924 snprintf(buf, sizeof(buf), fmt, val);
925 ends = vals = ltrim(buf);
926 while (isdigit(*ends) || *ends == '.')
929 while (isspace(*unit))
931 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
934 #define METRIC_ONLY_LEN 20
936 /* Filter out some columns that don't work well in metrics only mode */
938 static bool valid_only_metric(const char *unit)
942 if (strstr(unit, "/sec") ||
943 strstr(unit, "hz") ||
944 strstr(unit, "Hz") ||
945 strstr(unit, "CPUs utilized"))
950 static const char *fixunit(char *buf, struct perf_evsel *evsel,
953 if (!strncmp(unit, "of all", 6)) {
954 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
961 static void print_metric_only(void *ctx, const char *color, const char *fmt,
962 const char *unit, double val)
964 struct outstate *os = ctx;
968 unsigned mlen = METRIC_ONLY_LEN;
970 if (!valid_only_metric(unit))
972 unit = fixunit(buf, os->evsel, unit);
974 n = color_fprintf(out, color, fmt, val);
976 n = fprintf(out, fmt, val);
977 if (n > METRIC_ONLY_LEN)
979 if (mlen < strlen(unit))
980 mlen = strlen(unit) + 1;
981 fprintf(out, "%*s", mlen - n, "");
984 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
986 const char *unit, double val)
988 struct outstate *os = ctx;
990 char buf[64], *vals, *ends;
993 if (!valid_only_metric(unit))
995 unit = fixunit(tbuf, os->evsel, unit);
996 snprintf(buf, sizeof buf, fmt, val);
997 ends = vals = ltrim(buf);
998 while (isdigit(*ends) || *ends == '.')
1001 fprintf(out, "%s%s", vals, csv_sep);
1004 static void new_line_metric(void *ctx __maybe_unused)
1008 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1009 const char *fmt __maybe_unused,
1010 const char *unit, double val __maybe_unused)
1012 struct outstate *os = ctx;
1015 if (!valid_only_metric(unit))
1017 unit = fixunit(tbuf, os->evsel, unit);
1019 fprintf(os->fh, "%s%s", unit, csv_sep);
1021 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1024 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1026 FILE *output = stat_config.output;
1027 double msecs = avg / NSEC_PER_MSEC;
1028 const char *fmt_v, *fmt_n;
1031 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1032 fmt_n = csv_output ? "%s" : "%-25s";
1034 aggr_printout(evsel, id, nr);
1036 scnprintf(name, sizeof(name), "%s%s",
1037 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1039 fprintf(output, fmt_v, msecs, csv_sep);
1042 fprintf(output, "%s%s", evsel->unit, csv_sep);
1044 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1046 fprintf(output, fmt_n, name);
1049 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1052 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1059 if (stat_config.aggr_mode == AGGR_NONE)
1062 if (stat_config.aggr_mode == AGGR_GLOBAL)
1065 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1066 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1068 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1074 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1076 FILE *output = stat_config.output;
1077 double sc = evsel->scale;
1081 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1084 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1086 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1089 aggr_printout(evsel, id, nr);
1091 fprintf(output, fmt, avg, csv_sep);
1094 fprintf(output, "%-*s%s",
1095 csv_output ? 0 : unit_width,
1096 evsel->unit, csv_sep);
1098 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1101 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1104 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1105 char *prefix, u64 run, u64 ena, double noise,
1106 struct runtime_stat *st)
1108 struct perf_stat_output_ctx out;
1109 struct outstate os = {
1110 .fh = stat_config.output,
1111 .prefix = prefix ? prefix : "",
1116 print_metric_t pm = print_metric_std;
1120 nl = new_line_metric;
1122 pm = print_metric_only_csv;
1124 pm = print_metric_only;
1128 if (csv_output && !metric_only) {
1129 static int aggr_fields[] = {
1137 pm = print_metric_csv;
1140 os.nfields += aggr_fields[stat_config.aggr_mode];
1144 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1146 pm(&os, NULL, "", "", 0);
1149 aggr_printout(counter, id, nr);
1151 fprintf(stat_config.output, "%*s%s",
1152 csv_output ? 0 : 18,
1153 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1156 if (counter->supported)
1157 print_free_counters_hint = 1;
1159 fprintf(stat_config.output, "%-*s%s",
1160 csv_output ? 0 : unit_width,
1161 counter->unit, csv_sep);
1163 fprintf(stat_config.output, "%*s",
1164 csv_output ? 0 : -25,
1165 perf_evsel__name(counter));
1168 fprintf(stat_config.output, "%s%s",
1169 csv_sep, counter->cgrp->name);
1172 pm(&os, NULL, NULL, "", 0);
1173 print_noise(counter, noise);
1174 print_running(run, ena);
1176 pm(&os, NULL, NULL, "", 0);
1182 else if (nsec_counter(counter))
1183 nsec_printout(id, nr, counter, uval);
1185 abs_printout(id, nr, counter, uval);
1187 out.print_metric = pm;
1190 out.force_header = false;
1192 if (csv_output && !metric_only) {
1193 print_noise(counter, noise);
1194 print_running(run, ena);
1197 perf_stat__print_shadow_stats(counter, uval,
1198 first_shadow_cpu(counter, id),
1199 &out, &metric_events, st);
1200 if (!csv_output && !metric_only) {
1201 print_noise(counter, noise);
1202 print_running(run, ena);
1206 static void aggr_update_shadow(void)
1210 struct perf_evsel *counter;
1212 for (s = 0; s < aggr_map->nr; s++) {
1213 id = aggr_map->map[s];
1214 evlist__for_each_entry(evsel_list, counter) {
1216 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1217 s2 = aggr_get_id(evsel_list->cpus, cpu);
1220 val += perf_counts(counter->counts, cpu, 0)->val;
1222 perf_stat__update_shadow_stats(counter, val,
1223 first_shadow_cpu(counter, id),
1229 static void collect_all_aliases(struct perf_evsel *counter,
1230 void (*cb)(struct perf_evsel *counter, void *data,
1234 struct perf_evsel *alias;
1236 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1237 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1238 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1239 alias->scale != counter->scale ||
1240 alias->cgrp != counter->cgrp ||
1241 strcmp(alias->unit, counter->unit) ||
1242 nsec_counter(alias) != nsec_counter(counter))
1244 alias->merged_stat = true;
1245 cb(alias, data, false);
1249 static bool collect_data(struct perf_evsel *counter,
1250 void (*cb)(struct perf_evsel *counter, void *data,
1254 if (counter->merged_stat)
1256 cb(counter, data, true);
1257 if (!no_merge && counter->auto_merge_stats)
1258 collect_all_aliases(counter, cb, data);
1269 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1271 struct aggr_data *ad = data;
1274 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1275 struct perf_counts_values *counts;
1277 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1282 counts = perf_counts(counter->counts, cpu, 0);
1284 * When any result is bad, make them all to give
1285 * consistent output in interval mode.
1287 if (counts->ena == 0 || counts->run == 0 ||
1288 counter->counts->scaled == -1) {
1293 ad->val += counts->val;
1294 ad->ena += counts->ena;
1295 ad->run += counts->run;
1299 static void print_aggr(char *prefix)
1301 FILE *output = stat_config.output;
1302 struct perf_evsel *counter;
1308 if (!(aggr_map || aggr_get_id))
1311 aggr_update_shadow();
1314 * With metric_only everything is on a single line.
1315 * Without each counter has its own line.
1317 for (s = 0; s < aggr_map->nr; s++) {
1318 struct aggr_data ad;
1319 if (prefix && metric_only)
1320 fprintf(output, "%s", prefix);
1322 ad.id = id = aggr_map->map[s];
1324 evlist__for_each_entry(evsel_list, counter) {
1325 if (is_duration_time(counter))
1328 ad.val = ad.ena = ad.run = 0;
1330 if (!collect_data(counter, aggr_cb, &ad))
1336 if (first && metric_only) {
1338 aggr_printout(counter, id, nr);
1340 if (prefix && !metric_only)
1341 fprintf(output, "%s", prefix);
1343 uval = val * counter->scale;
1344 printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1347 fputc('\n', output);
1350 fputc('\n', output);
1354 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1356 FILE *output = stat_config.output;
1357 int nthreads = thread_map__nr(counter->threads);
1358 int ncpus = cpu_map__nr(counter->cpus);
1362 for (thread = 0; thread < nthreads; thread++) {
1363 u64 ena = 0, run = 0, val = 0;
1365 for (cpu = 0; cpu < ncpus; cpu++) {
1366 val += perf_counts(counter->counts, cpu, thread)->val;
1367 ena += perf_counts(counter->counts, cpu, thread)->ena;
1368 run += perf_counts(counter->counts, cpu, thread)->run;
1372 fprintf(output, "%s", prefix);
1374 uval = val * counter->scale;
1376 if (stat_config.stats)
1377 printout(thread, 0, counter, uval, prefix, run, ena,
1378 1.0, &stat_config.stats[thread]);
1380 printout(thread, 0, counter, uval, prefix, run, ena,
1382 fputc('\n', output);
1387 double avg, avg_enabled, avg_running;
1390 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1391 bool first __maybe_unused)
1393 struct caggr_data *cd = data;
1394 struct perf_stat_evsel *ps = counter->stats;
1396 cd->avg += avg_stats(&ps->res_stats[0]);
1397 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1398 cd->avg_running += avg_stats(&ps->res_stats[2]);
1402 * Print out the results of a single counter:
1403 * aggregated counts in system-wide mode
1405 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1407 FILE *output = stat_config.output;
1409 struct caggr_data cd = { .avg = 0.0 };
1411 if (!collect_data(counter, counter_aggr_cb, &cd))
1414 if (prefix && !metric_only)
1415 fprintf(output, "%s", prefix);
1417 uval = cd.avg * counter->scale;
1418 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1421 fprintf(output, "\n");
1424 static void counter_cb(struct perf_evsel *counter, void *data,
1425 bool first __maybe_unused)
1427 struct aggr_data *ad = data;
1429 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1430 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1431 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1435 * Print out the results of a single counter:
1436 * does not use aggregated count in system-wide
1438 static void print_counter(struct perf_evsel *counter, char *prefix)
1440 FILE *output = stat_config.output;
1445 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1446 struct aggr_data ad = { .cpu = cpu };
1448 if (!collect_data(counter, counter_cb, &ad))
1455 fprintf(output, "%s", prefix);
1457 uval = val * counter->scale;
1458 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1461 fputc('\n', output);
1465 static void print_no_aggr_metric(char *prefix)
1469 struct perf_evsel *counter;
1473 nrcpus = evsel_list->cpus->nr;
1474 for (cpu = 0; cpu < nrcpus; cpu++) {
1478 fputs(prefix, stat_config.output);
1479 evlist__for_each_entry(evsel_list, counter) {
1480 if (is_duration_time(counter))
1483 aggr_printout(counter, cpu, 0);
1486 val = perf_counts(counter->counts, cpu, 0)->val;
1487 ena = perf_counts(counter->counts, cpu, 0)->ena;
1488 run = perf_counts(counter->counts, cpu, 0)->run;
1490 uval = val * counter->scale;
1491 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1494 fputc('\n', stat_config.output);
1498 static int aggr_header_lens[] = {
1506 static const char *aggr_header_csv[] = {
1507 [AGGR_CORE] = "core,cpus,",
1508 [AGGR_SOCKET] = "socket,cpus",
1509 [AGGR_NONE] = "cpu,",
1510 [AGGR_THREAD] = "comm-pid,",
1514 static void print_metric_headers(const char *prefix, bool no_indent)
1516 struct perf_stat_output_ctx out;
1517 struct perf_evsel *counter;
1518 struct outstate os = {
1519 .fh = stat_config.output
1523 fprintf(stat_config.output, "%s", prefix);
1525 if (!csv_output && !no_indent)
1526 fprintf(stat_config.output, "%*s",
1527 aggr_header_lens[stat_config.aggr_mode], "");
1529 if (stat_config.interval)
1530 fputs("time,", stat_config.output);
1531 fputs(aggr_header_csv[stat_config.aggr_mode],
1532 stat_config.output);
1535 /* Print metrics headers only */
1536 evlist__for_each_entry(evsel_list, counter) {
1537 if (is_duration_time(counter))
1541 out.print_metric = print_metric_header;
1542 out.new_line = new_line_metric;
1543 out.force_header = true;
1545 perf_stat__print_shadow_stats(counter, 0,
1551 fputc('\n', stat_config.output);
1554 static void print_interval(char *prefix, struct timespec *ts)
1556 FILE *output = stat_config.output;
1557 static int num_print_interval;
1559 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1561 if (num_print_interval == 0 && !csv_output) {
1562 switch (stat_config.aggr_mode) {
1564 fprintf(output, "# time socket cpus");
1566 fprintf(output, " counts %*s events\n", unit_width, "unit");
1569 fprintf(output, "# time core cpus");
1571 fprintf(output, " counts %*s events\n", unit_width, "unit");
1574 fprintf(output, "# time CPU");
1576 fprintf(output, " counts %*s events\n", unit_width, "unit");
1579 fprintf(output, "# time comm-pid");
1581 fprintf(output, " counts %*s events\n", unit_width, "unit");
1585 fprintf(output, "# time");
1587 fprintf(output, " counts %*s events\n", unit_width, "unit");
1593 if (num_print_interval == 0 && metric_only)
1594 print_metric_headers(" ", true);
1595 if (++num_print_interval == 25)
1596 num_print_interval = 0;
1599 static void print_header(int argc, const char **argv)
1601 FILE *output = stat_config.output;
1607 fprintf(output, "\n");
1608 fprintf(output, " Performance counter stats for ");
1609 if (target.system_wide)
1610 fprintf(output, "\'system wide");
1611 else if (target.cpu_list)
1612 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1613 else if (!target__has_task(&target)) {
1614 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1615 for (i = 1; argv && (i < argc); i++)
1616 fprintf(output, " %s", argv[i]);
1617 } else if (target.pid)
1618 fprintf(output, "process id \'%s", target.pid);
1620 fprintf(output, "thread id \'%s", target.tid);
1622 fprintf(output, "\'");
1624 fprintf(output, " (%d runs)", run_count);
1625 fprintf(output, ":\n\n");
1629 static void print_footer(void)
1631 FILE *output = stat_config.output;
1635 fprintf(output, "\n");
1636 fprintf(output, " %17.9f seconds time elapsed",
1637 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1638 if (run_count > 1) {
1639 fprintf(output, " ");
1640 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1641 avg_stats(&walltime_nsecs_stats));
1643 fprintf(output, "\n\n");
1645 if (print_free_counters_hint &&
1646 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1649 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1650 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1652 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1655 static void print_counters(struct timespec *ts, int argc, const char **argv)
1657 int interval = stat_config.interval;
1658 struct perf_evsel *counter;
1659 char buf[64], *prefix = NULL;
1661 /* Do not print anything if we record to the pipe. */
1662 if (STAT_RECORD && perf_stat.data.is_pipe)
1666 print_interval(prefix = buf, ts);
1668 print_header(argc, argv);
1671 static int num_print_iv;
1673 if (num_print_iv == 0 && !interval)
1674 print_metric_headers(prefix, false);
1675 if (num_print_iv++ == 25)
1677 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1678 fprintf(stat_config.output, "%s", prefix);
1681 switch (stat_config.aggr_mode) {
1687 evlist__for_each_entry(evsel_list, counter) {
1688 if (is_duration_time(counter))
1690 print_aggr_thread(counter, prefix);
1694 evlist__for_each_entry(evsel_list, counter) {
1695 if (is_duration_time(counter))
1697 print_counter_aggr(counter, prefix);
1700 fputc('\n', stat_config.output);
1704 print_no_aggr_metric(prefix);
1706 evlist__for_each_entry(evsel_list, counter) {
1707 if (is_duration_time(counter))
1709 print_counter(counter, prefix);
1718 if (!interval && !csv_output)
1721 fflush(stat_config.output);
1724 static volatile int signr = -1;
1726 static void skip_signal(int signo)
1728 if ((child_pid == -1) || stat_config.interval)
1733 * render child_pid harmless
1734 * won't send SIGTERM to a random
1735 * process in case of race condition
1736 * and fast PID recycling
1741 static void sig_atexit(void)
1746 * avoid race condition with SIGCHLD handler
1747 * in skip_signal() which is modifying child_pid
1748 * goal is to avoid send SIGTERM to a random
1752 sigaddset(&set, SIGCHLD);
1753 sigprocmask(SIG_BLOCK, &set, &oset);
1755 if (child_pid != -1)
1756 kill(child_pid, SIGTERM);
1758 sigprocmask(SIG_SETMASK, &oset, NULL);
1763 signal(signr, SIG_DFL);
1764 kill(getpid(), signr);
1767 static int stat__set_big_num(const struct option *opt __maybe_unused,
1768 const char *s __maybe_unused, int unset)
1770 big_num_opt = unset ? 0 : 1;
1774 static int enable_metric_only(const struct option *opt __maybe_unused,
1775 const char *s __maybe_unused, int unset)
1777 force_metric_only = true;
1778 metric_only = !unset;
1782 static int parse_metric_groups(const struct option *opt,
1784 int unset __maybe_unused)
1786 return metricgroup__parse_groups(opt, str, &metric_events);
1789 static const struct option stat_options[] = {
1790 OPT_BOOLEAN('T', "transaction", &transaction_run,
1791 "hardware transaction statistics"),
1792 OPT_CALLBACK('e', "event", &evsel_list, "event",
1793 "event selector. use 'perf list' to list available events",
1794 parse_events_option),
1795 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1796 "event filter", parse_filter),
1797 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1798 "child tasks do not inherit counters"),
1799 OPT_STRING('p', "pid", &target.pid, "pid",
1800 "stat events on existing process id"),
1801 OPT_STRING('t', "tid", &target.tid, "tid",
1802 "stat events on existing thread id"),
1803 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1804 "system-wide collection from all CPUs"),
1805 OPT_BOOLEAN('g', "group", &group,
1806 "put the counters into a counter group"),
1807 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1808 OPT_INCR('v', "verbose", &verbose,
1809 "be more verbose (show counter open errors, etc)"),
1810 OPT_INTEGER('r', "repeat", &run_count,
1811 "repeat command and print average + stddev (max: 100, forever: 0)"),
1812 OPT_BOOLEAN('n', "null", &null_run,
1813 "null run - dont start any counters"),
1814 OPT_INCR('d', "detailed", &detailed_run,
1815 "detailed run - start a lot of events"),
1816 OPT_BOOLEAN('S', "sync", &sync_run,
1817 "call sync() before starting a run"),
1818 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1819 "print large numbers with thousands\' separators",
1821 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1822 "list of cpus to monitor in system-wide"),
1823 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1824 "disable CPU count aggregation", AGGR_NONE),
1825 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1826 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1827 "print counts with custom separator"),
1828 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1829 "monitor event in cgroup name only", parse_cgroups),
1830 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1831 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1832 OPT_INTEGER(0, "log-fd", &output_fd,
1833 "log output to fd, instead of stderr"),
1834 OPT_STRING(0, "pre", &pre_cmd, "command",
1835 "command to run prior to the measured command"),
1836 OPT_STRING(0, "post", &post_cmd, "command",
1837 "command to run after to the measured command"),
1838 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1839 "print counts at regular interval in ms (>= 10)"),
1840 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1841 "aggregate counts per processor socket", AGGR_SOCKET),
1842 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1843 "aggregate counts per physical processor core", AGGR_CORE),
1844 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1845 "aggregate counts per thread", AGGR_THREAD),
1846 OPT_UINTEGER('D', "delay", &initial_delay,
1847 "ms to wait before starting measurement after program start"),
1848 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1849 "Only print computed metrics. No raw values", enable_metric_only),
1850 OPT_BOOLEAN(0, "topdown", &topdown_run,
1851 "measure topdown level 1 statistics"),
1852 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1853 "measure SMI cost"),
1854 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1855 "monitor specified metrics or metric groups (separated by ,)",
1856 parse_metric_groups),
1860 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1862 return cpu_map__get_socket(map, cpu, NULL);
1865 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1867 return cpu_map__get_core(map, cpu, NULL);
1870 static int cpu_map__get_max(struct cpu_map *map)
1874 for (i = 0; i < map->nr; i++) {
1875 if (map->map[i] > max)
1882 static struct cpu_map *cpus_aggr_map;
1884 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1891 cpu = map->map[idx];
1893 if (cpus_aggr_map->map[cpu] == -1)
1894 cpus_aggr_map->map[cpu] = get_id(map, idx);
1896 return cpus_aggr_map->map[cpu];
1899 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1901 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1904 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1906 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1909 static int perf_stat_init_aggr_mode(void)
1913 switch (stat_config.aggr_mode) {
1915 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1916 perror("cannot build socket map");
1919 aggr_get_id = perf_stat__get_socket_cached;
1922 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1923 perror("cannot build core map");
1926 aggr_get_id = perf_stat__get_core_cached;
1937 * The evsel_list->cpus is the base we operate on,
1938 * taking the highest cpu number to be the size of
1939 * the aggregation translate cpumap.
1941 nr = cpu_map__get_max(evsel_list->cpus);
1942 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1943 return cpus_aggr_map ? 0 : -ENOMEM;
1946 static void perf_stat__exit_aggr_mode(void)
1948 cpu_map__put(aggr_map);
1949 cpu_map__put(cpus_aggr_map);
1951 cpus_aggr_map = NULL;
1954 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1961 cpu = map->map[idx];
1963 if (cpu >= env->nr_cpus_avail)
1969 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1971 struct perf_env *env = data;
1972 int cpu = perf_env__get_cpu(env, map, idx);
1974 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1977 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1979 struct perf_env *env = data;
1980 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1983 int socket_id = env->cpu[cpu].socket_id;
1986 * Encode socket in upper 16 bits
1987 * core_id is relative to socket, and
1988 * we need a global id. So we combine
1991 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1997 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1998 struct cpu_map **sockp)
2000 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2003 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2004 struct cpu_map **corep)
2006 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2009 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2011 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2014 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2016 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2019 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2021 struct perf_env *env = &st->session->header.env;
2023 switch (stat_config.aggr_mode) {
2025 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2026 perror("cannot build socket map");
2029 aggr_get_id = perf_stat__get_socket_file;
2032 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2033 perror("cannot build core map");
2036 aggr_get_id = perf_stat__get_core_file;
2049 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2056 for (i = 0; attr[i]; i++) {
2057 if (pmu_have_event("cpu", attr[i])) {
2058 len += strlen(attr[i]) + 1;
2059 attr[i - off] = attr[i];
2063 attr[i - off] = NULL;
2065 *str = malloc(len + 1 + 2);
2075 for (i = 0; attr[i]; i++) {
2088 __weak bool arch_topdown_check_group(bool *warn)
2094 __weak void arch_topdown_group_warn(void)
2099 * Add default attributes, if there were no attributes specified or
2100 * if -d/--detailed, -d -d or -d -d -d is used:
2102 static int add_default_attributes(void)
2105 struct perf_event_attr default_attrs0[] = {
2107 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2108 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2109 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2110 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2112 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2114 struct perf_event_attr frontend_attrs[] = {
2115 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2117 struct perf_event_attr backend_attrs[] = {
2118 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2120 struct perf_event_attr default_attrs1[] = {
2121 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2122 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2123 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2128 * Detailed stats (-d), covering the L1 and last level data caches:
2130 struct perf_event_attr detailed_attrs[] = {
2132 { .type = PERF_TYPE_HW_CACHE,
2134 PERF_COUNT_HW_CACHE_L1D << 0 |
2135 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2136 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2138 { .type = PERF_TYPE_HW_CACHE,
2140 PERF_COUNT_HW_CACHE_L1D << 0 |
2141 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2142 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2144 { .type = PERF_TYPE_HW_CACHE,
2146 PERF_COUNT_HW_CACHE_LL << 0 |
2147 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2148 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2150 { .type = PERF_TYPE_HW_CACHE,
2152 PERF_COUNT_HW_CACHE_LL << 0 |
2153 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2154 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2158 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2160 struct perf_event_attr very_detailed_attrs[] = {
2162 { .type = PERF_TYPE_HW_CACHE,
2164 PERF_COUNT_HW_CACHE_L1I << 0 |
2165 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2166 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2168 { .type = PERF_TYPE_HW_CACHE,
2170 PERF_COUNT_HW_CACHE_L1I << 0 |
2171 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2172 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2174 { .type = PERF_TYPE_HW_CACHE,
2176 PERF_COUNT_HW_CACHE_DTLB << 0 |
2177 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2178 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2180 { .type = PERF_TYPE_HW_CACHE,
2182 PERF_COUNT_HW_CACHE_DTLB << 0 |
2183 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2184 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2186 { .type = PERF_TYPE_HW_CACHE,
2188 PERF_COUNT_HW_CACHE_ITLB << 0 |
2189 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2190 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2192 { .type = PERF_TYPE_HW_CACHE,
2194 PERF_COUNT_HW_CACHE_ITLB << 0 |
2195 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2196 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2201 * Very, very detailed stats (-d -d -d), adding prefetch events:
2203 struct perf_event_attr very_very_detailed_attrs[] = {
2205 { .type = PERF_TYPE_HW_CACHE,
2207 PERF_COUNT_HW_CACHE_L1D << 0 |
2208 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2209 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2211 { .type = PERF_TYPE_HW_CACHE,
2213 PERF_COUNT_HW_CACHE_L1D << 0 |
2214 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2215 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2218 /* Set attrs if no event is selected and !null_run: */
2222 if (transaction_run) {
2223 if (pmu_have_event("cpu", "cycles-ct") &&
2224 pmu_have_event("cpu", "el-start"))
2225 err = parse_events(evsel_list, transaction_attrs, NULL);
2227 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2229 fprintf(stderr, "Cannot set up transaction events\n");
2238 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2239 fprintf(stderr, "freeze_on_smi is not supported.\n");
2244 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2245 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2251 if (pmu_have_event("msr", "aperf") &&
2252 pmu_have_event("msr", "smi")) {
2253 if (!force_metric_only)
2255 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2257 fprintf(stderr, "To measure SMI cost, it needs "
2258 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2262 fprintf(stderr, "Cannot set up SMI cost events\n");
2272 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2273 stat_config.aggr_mode != AGGR_CORE) {
2274 pr_err("top down event configuration requires --per-core mode\n");
2277 stat_config.aggr_mode = AGGR_CORE;
2278 if (nr_cgroups || !target__has_cpu(&target)) {
2279 pr_err("top down event configuration requires system-wide mode (-a)\n");
2283 if (!force_metric_only)
2285 if (topdown_filter_events(topdown_attrs, &str,
2286 arch_topdown_check_group(&warn)) < 0) {
2287 pr_err("Out of memory\n");
2290 if (topdown_attrs[0] && str) {
2292 arch_topdown_group_warn();
2293 err = parse_events(evsel_list, str, NULL);
2296 "Cannot set up top down events %s: %d\n",
2302 fprintf(stderr, "System does not support topdown\n");
2308 if (!evsel_list->nr_entries) {
2309 if (target__has_cpu(&target))
2310 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2312 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2314 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2315 if (perf_evlist__add_default_attrs(evsel_list,
2316 frontend_attrs) < 0)
2319 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2320 if (perf_evlist__add_default_attrs(evsel_list,
2324 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2328 /* Detailed events get appended to the event list: */
2330 if (detailed_run < 1)
2333 /* Append detailed run extra attributes: */
2334 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2337 if (detailed_run < 2)
2340 /* Append very detailed run extra attributes: */
2341 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2344 if (detailed_run < 3)
2347 /* Append very, very detailed run extra attributes: */
2348 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2351 static const char * const stat_record_usage[] = {
2352 "perf stat record [<options>]",
2356 static void init_features(struct perf_session *session)
2360 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2361 perf_header__set_feat(&session->header, feat);
2363 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2364 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2365 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2366 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2369 static int __cmd_record(int argc, const char **argv)
2371 struct perf_session *session;
2372 struct perf_data *data = &perf_stat.data;
2374 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2375 PARSE_OPT_STOP_AT_NON_OPTION);
2378 data->file.path = output_name;
2380 if (run_count != 1 || forever) {
2381 pr_err("Cannot use -r option with perf stat record.\n");
2385 session = perf_session__new(data, false, NULL);
2386 if (session == NULL) {
2387 pr_err("Perf session creation failed.\n");
2391 init_features(session);
2393 session->evlist = evsel_list;
2394 perf_stat.session = session;
2395 perf_stat.record = true;
2399 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2400 union perf_event *event,
2401 struct perf_session *session)
2403 struct stat_round_event *stat_round = &event->stat_round;
2404 struct perf_evsel *counter;
2405 struct timespec tsh, *ts = NULL;
2406 const char **argv = session->header.env.cmdline_argv;
2407 int argc = session->header.env.nr_cmdline;
2409 evlist__for_each_entry(evsel_list, counter)
2410 perf_stat_process_counter(&stat_config, counter);
2412 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2413 update_stats(&walltime_nsecs_stats, stat_round->time);
2415 if (stat_config.interval && stat_round->time) {
2416 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2417 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2421 print_counters(ts, argc, argv);
2426 int process_stat_config_event(struct perf_tool *tool,
2427 union perf_event *event,
2428 struct perf_session *session __maybe_unused)
2430 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2432 perf_event__read_stat_config(&stat_config, &event->stat_config);
2434 if (cpu_map__empty(st->cpus)) {
2435 if (st->aggr_mode != AGGR_UNSET)
2436 pr_warning("warning: processing task data, aggregation mode not set\n");
2440 if (st->aggr_mode != AGGR_UNSET)
2441 stat_config.aggr_mode = st->aggr_mode;
2443 if (perf_stat.data.is_pipe)
2444 perf_stat_init_aggr_mode();
2446 perf_stat_init_aggr_mode_file(st);
2451 static int set_maps(struct perf_stat *st)
2453 if (!st->cpus || !st->threads)
2456 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2459 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2461 if (perf_evlist__alloc_stats(evsel_list, true))
2464 st->maps_allocated = true;
2469 int process_thread_map_event(struct perf_tool *tool,
2470 union perf_event *event,
2471 struct perf_session *session __maybe_unused)
2473 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2476 pr_warning("Extra thread map event, ignoring.\n");
2480 st->threads = thread_map__new_event(&event->thread_map);
2484 return set_maps(st);
2488 int process_cpu_map_event(struct perf_tool *tool,
2489 union perf_event *event,
2490 struct perf_session *session __maybe_unused)
2492 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2493 struct cpu_map *cpus;
2496 pr_warning("Extra cpu map event, ignoring.\n");
2500 cpus = cpu_map__new_data(&event->cpu_map.data);
2505 return set_maps(st);
2508 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2512 config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2516 config->stats_num = nthreads;
2518 for (i = 0; i < nthreads; i++)
2519 runtime_stat__init(&config->stats[i]);
2524 static void runtime_stat_delete(struct perf_stat_config *config)
2531 for (i = 0; i < config->stats_num; i++)
2532 runtime_stat__exit(&config->stats[i]);
2534 free(config->stats);
2537 static const char * const stat_report_usage[] = {
2538 "perf stat report [<options>]",
2542 static struct perf_stat perf_stat = {
2544 .attr = perf_event__process_attr,
2545 .event_update = perf_event__process_event_update,
2546 .thread_map = process_thread_map_event,
2547 .cpu_map = process_cpu_map_event,
2548 .stat_config = process_stat_config_event,
2549 .stat = perf_event__process_stat_event,
2550 .stat_round = process_stat_round_event,
2552 .aggr_mode = AGGR_UNSET,
2555 static int __cmd_report(int argc, const char **argv)
2557 struct perf_session *session;
2558 const struct option options[] = {
2559 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2560 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2561 "aggregate counts per processor socket", AGGR_SOCKET),
2562 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2563 "aggregate counts per physical processor core", AGGR_CORE),
2564 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2565 "disable CPU count aggregation", AGGR_NONE),
2571 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2573 if (!input_name || !strlen(input_name)) {
2574 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2577 input_name = "perf.data";
2580 perf_stat.data.file.path = input_name;
2581 perf_stat.data.mode = PERF_DATA_MODE_READ;
2583 session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2584 if (session == NULL)
2587 perf_stat.session = session;
2588 stat_config.output = stderr;
2589 evsel_list = session->evlist;
2591 ret = perf_session__process_events(session);
2595 perf_session__delete(session);
2599 static void setup_system_wide(int forks)
2602 * Make system wide (-a) the default target if
2603 * no target was specified and one of following
2604 * conditions is met:
2606 * - there's no workload specified
2607 * - there is workload specified but all requested
2608 * events are system wide events
2610 if (!target__none(&target))
2614 target.system_wide = true;
2616 struct perf_evsel *counter;
2618 evlist__for_each_entry(evsel_list, counter) {
2619 if (!counter->system_wide)
2623 if (evsel_list->nr_entries)
2624 target.system_wide = true;
2628 int cmd_stat(int argc, const char **argv)
2630 const char * const stat_usage[] = {
2631 "perf stat [<options>] [<command>]",
2634 int status = -EINVAL, run_idx;
2636 FILE *output = stderr;
2637 unsigned int interval;
2638 const char * const stat_subcommands[] = { "record", "report" };
2640 setlocale(LC_ALL, "");
2642 evsel_list = perf_evlist__new();
2643 if (evsel_list == NULL)
2646 parse_events__shrink_config_terms();
2647 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2648 (const char **) stat_usage,
2649 PARSE_OPT_STOP_AT_NON_OPTION);
2650 perf_stat__collect_metric_expr(evsel_list);
2651 perf_stat__init_shadow_stats();
2655 if (!strcmp(csv_sep, "\\t"))
2658 csv_sep = DEFAULT_SEPARATOR;
2660 if (argc && !strncmp(argv[0], "rec", 3)) {
2661 argc = __cmd_record(argc, argv);
2664 } else if (argc && !strncmp(argv[0], "rep", 3))
2665 return __cmd_report(argc, argv);
2667 interval = stat_config.interval;
2670 * For record command the -o is already taken care of.
2672 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2675 if (output_name && output_fd) {
2676 fprintf(stderr, "cannot use both --output and --log-fd\n");
2677 parse_options_usage(stat_usage, stat_options, "o", 1);
2678 parse_options_usage(NULL, stat_options, "log-fd", 0);
2682 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2683 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2687 if (metric_only && run_count > 1) {
2688 fprintf(stderr, "--metric-only is not supported with -r\n");
2692 if (output_fd < 0) {
2693 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2694 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2700 mode = append_file ? "a" : "w";
2702 output = fopen(output_name, mode);
2704 perror("failed to create output file");
2707 clock_gettime(CLOCK_REALTIME, &tm);
2708 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2709 } else if (output_fd > 0) {
2710 mode = append_file ? "a" : "w";
2711 output = fdopen(output_fd, mode);
2713 perror("Failed opening logfd");
2718 stat_config.output = output;
2721 * let the spreadsheet do the pretty-printing
2724 /* User explicitly passed -B? */
2725 if (big_num_opt == 1) {
2726 fprintf(stderr, "-B option not supported with -x\n");
2727 parse_options_usage(stat_usage, stat_options, "B", 1);
2728 parse_options_usage(NULL, stat_options, "x", 1);
2730 } else /* Nope, so disable big number formatting */
2732 } else if (big_num_opt == 0) /* User passed --no-big-num */
2735 setup_system_wide(argc);
2737 if (run_count < 0) {
2738 pr_err("Run count must be a positive number\n");
2739 parse_options_usage(stat_usage, stat_options, "r", 1);
2741 } else if (run_count == 0) {
2746 if ((stat_config.aggr_mode == AGGR_THREAD) &&
2747 !target__has_task(&target)) {
2748 if (!target.system_wide || target.cpu_list) {
2749 fprintf(stderr, "The --per-thread option is only "
2750 "available when monitoring via -p -t -a "
2751 "options or only --per-thread.\n");
2752 parse_options_usage(NULL, stat_options, "p", 1);
2753 parse_options_usage(NULL, stat_options, "t", 1);
2759 * no_aggr, cgroup are for system-wide only
2760 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2762 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2763 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2764 !target__has_cpu(&target)) {
2765 fprintf(stderr, "both cgroup and no-aggregation "
2766 "modes only available in system-wide mode\n");
2768 parse_options_usage(stat_usage, stat_options, "G", 1);
2769 parse_options_usage(NULL, stat_options, "A", 1);
2770 parse_options_usage(NULL, stat_options, "a", 1);
2774 if (add_default_attributes())
2777 target__validate(&target);
2779 if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2780 target.per_thread = true;
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);
2801 if (target.system_wide) {
2802 if (runtime_stat_new(&stat_config,
2803 thread_map__nr(evsel_list->threads))) {
2809 if (interval && interval < 100) {
2810 if (interval < 10) {
2811 pr_err("print interval must be >= 10ms\n");
2812 parse_options_usage(stat_usage, stat_options, "I", 1);
2815 pr_warning("print interval < 100ms. "
2816 "The overhead percentage could be high in some cases. "
2817 "Please proceed with caution.\n");
2820 if (perf_evlist__alloc_stats(evsel_list, interval))
2823 if (perf_stat_init_aggr_mode())
2827 * We dont want to block the signals - that would cause
2828 * child tasks to inherit that and Ctrl-C would not work.
2829 * What we want is for Ctrl-C to work in the exec()-ed
2830 * task, but being ignored by perf stat itself:
2834 signal(SIGINT, skip_signal);
2835 signal(SIGCHLD, skip_signal);
2836 signal(SIGALRM, skip_signal);
2837 signal(SIGABRT, skip_signal);
2840 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2841 if (run_count != 1 && verbose > 0)
2842 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2845 status = run_perf_stat(argc, argv);
2846 if (forever && status != -1) {
2847 print_counters(NULL, argc, argv);
2848 perf_stat__reset_stats();
2852 if (!forever && status != -1 && !interval)
2853 print_counters(NULL, argc, argv);
2857 * We synthesize the kernel mmap record just so that older tools
2858 * don't emit warnings about not being able to resolve symbols
2859 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2860 * a saner message about no samples being in the perf.data file.
2862 * This also serves to suppress a warning about f_header.data.size == 0
2863 * in header.c at the moment 'perf stat record' gets introduced, which
2864 * is not really needed once we start adding the stat specific PERF_RECORD_
2865 * records, but the need to suppress the kptr_restrict messages in older
2866 * tools remain -acme
2868 int fd = perf_data__fd(&perf_stat.data);
2869 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2870 process_synthesized_event,
2871 &perf_stat.session->machines.host);
2873 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2874 "older tools may produce warnings about this file\n.");
2878 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2879 pr_err("failed to write stat round event\n");
2882 if (!perf_stat.data.is_pipe) {
2883 perf_stat.session->header.data_size += perf_stat.bytes_written;
2884 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2887 perf_session__delete(perf_stat.session);
2890 perf_stat__exit_aggr_mode();
2891 perf_evlist__free_stats(evsel_list);
2893 if (smi_cost && smi_reset)
2894 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2896 perf_evlist__delete(evsel_list);
2898 runtime_stat_delete(&stat_config);