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_file file;
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.file.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_file__write(&perf_stat.file, 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 print_counters(&rs, 0, NULL);
421 static void enable_counters(void)
424 usleep(initial_delay * USEC_PER_MSEC);
427 * We need to enable counters only if:
428 * - we don't have tracee (attaching to task or cpu)
429 * - we have initial delay configured
431 if (!target__none(&target) || initial_delay)
432 perf_evlist__enable(evsel_list);
435 static void disable_counters(void)
438 * If we don't have tracee (attaching to task or cpu), counters may
439 * still be running. To get accurate group ratios, we must stop groups
440 * from counting before reading their constituent counters.
442 if (!target__none(&target))
443 perf_evlist__disable(evsel_list);
446 static volatile int workload_exec_errno;
449 * perf_evlist__prepare_workload will send a SIGUSR1
450 * if the fork fails, since we asked by setting its
451 * want_signal to true.
453 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
454 void *ucontext __maybe_unused)
456 workload_exec_errno = info->si_value.sival_int;
459 static bool has_unit(struct perf_evsel *counter)
461 return counter->unit && *counter->unit;
464 static bool has_scale(struct perf_evsel *counter)
466 return counter->scale != 1;
469 static int perf_stat_synthesize_config(bool is_pipe)
471 struct perf_evsel *counter;
475 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
476 process_synthesized_event);
478 pr_err("Couldn't synthesize attrs.\n");
484 * Synthesize other events stuff not carried within
485 * attr event - unit, scale, name
487 evlist__for_each_entry(evsel_list, counter) {
488 if (!counter->supported)
492 * Synthesize unit and scale only if it's defined.
494 if (has_unit(counter)) {
495 err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
497 pr_err("Couldn't synthesize evsel unit.\n");
502 if (has_scale(counter)) {
503 err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
505 pr_err("Couldn't synthesize evsel scale.\n");
510 if (counter->own_cpus) {
511 err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
513 pr_err("Couldn't synthesize evsel scale.\n");
519 * Name is needed only for pipe output,
520 * perf.data carries event names.
523 err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
525 pr_err("Couldn't synthesize evsel name.\n");
531 err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
532 process_synthesized_event,
535 pr_err("Couldn't synthesize thread map.\n");
539 err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
540 process_synthesized_event, NULL);
542 pr_err("Couldn't synthesize thread map.\n");
546 err = perf_event__synthesize_stat_config(NULL, &stat_config,
547 process_synthesized_event, NULL);
549 pr_err("Couldn't synthesize config.\n");
556 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
558 static int __store_counter_ids(struct perf_evsel *counter,
559 struct cpu_map *cpus,
560 struct thread_map *threads)
564 for (cpu = 0; cpu < cpus->nr; cpu++) {
565 for (thread = 0; thread < threads->nr; thread++) {
566 int fd = FD(counter, cpu, thread);
568 if (perf_evlist__id_add_fd(evsel_list, counter,
569 cpu, thread, fd) < 0)
577 static int store_counter_ids(struct perf_evsel *counter)
579 struct cpu_map *cpus = counter->cpus;
580 struct thread_map *threads = counter->threads;
582 if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
585 return __store_counter_ids(counter, cpus, threads);
588 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
590 return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
593 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
595 struct perf_evsel *c2, *leader;
598 leader = evsel->leader;
599 pr_debug("Weak group for %s/%d failed\n",
600 leader->name, leader->nr_members);
603 * for_each_group_member doesn't work here because it doesn't
604 * include the first entry.
606 evlist__for_each_entry(evsel_list, c2) {
609 if (c2->leader == leader) {
611 perf_evsel__close(c2);
619 static int __run_perf_stat(int argc, const char **argv)
621 int interval = stat_config.interval;
623 unsigned long long t0, t1;
624 struct perf_evsel *counter;
628 const bool forks = (argc > 0);
629 bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
630 struct perf_evsel_config_term *err_term;
633 ts.tv_sec = interval / USEC_PER_MSEC;
634 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
641 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
642 workload_exec_failed_signal) < 0) {
643 perror("failed to prepare workload");
646 child_pid = evsel_list->workload.pid;
650 perf_evlist__set_leader(evsel_list);
652 evlist__for_each_entry(evsel_list, counter) {
654 if (create_perf_stat_counter(counter) < 0) {
656 /* Weak group failed. Reset the group. */
657 if (errno == EINVAL &&
658 counter->leader != counter &&
659 counter->weak_group) {
660 counter = perf_evsel__reset_weak_group(counter);
665 * PPC returns ENXIO for HW counters until 2.6.37
666 * (behavior changed with commit b0a873e).
668 if (errno == EINVAL || errno == ENOSYS ||
669 errno == ENOENT || errno == EOPNOTSUPP ||
672 ui__warning("%s event is not supported by the kernel.\n",
673 perf_evsel__name(counter));
674 counter->supported = false;
676 if ((counter->leader != counter) ||
677 !(counter->leader->nr_members > 1))
679 } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
681 ui__warning("%s\n", msg);
685 perf_evsel__open_strerror(counter, &target,
686 errno, msg, sizeof(msg));
687 ui__error("%s\n", msg);
690 kill(child_pid, SIGTERM);
694 counter->supported = true;
696 l = strlen(counter->unit);
700 if (perf_evsel__should_store_id(counter) &&
701 store_counter_ids(counter))
705 if (perf_evlist__apply_filters(evsel_list, &counter)) {
706 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
707 counter->filter, perf_evsel__name(counter), errno,
708 str_error_r(errno, msg, sizeof(msg)));
712 if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
713 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
714 err_term->val.drv_cfg, perf_evsel__name(counter), errno,
715 str_error_r(errno, msg, sizeof(msg)));
720 int err, fd = perf_data_file__fd(&perf_stat.file);
723 err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
725 err = perf_session__write_header(perf_stat.session, evsel_list,
732 err = perf_stat_synthesize_config(is_pipe);
738 * Enable counters and exec the command:
741 clock_gettime(CLOCK_MONOTONIC, &ref_time);
744 perf_evlist__start_workload(evsel_list);
748 while (!waitpid(child_pid, &status, WNOHANG)) {
749 nanosleep(&ts, NULL);
753 waitpid(child_pid, &status, 0);
755 if (workload_exec_errno) {
756 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
757 pr_err("Workload failed: %s\n", emsg);
761 if (WIFSIGNALED(status))
762 psignal(WTERMSIG(status), argv[0]);
766 nanosleep(&ts, NULL);
776 update_stats(&walltime_nsecs_stats, t1 - t0);
779 * Closing a group leader splits the group, and as we only disable
780 * group leaders, results in remaining events becoming enabled. To
781 * avoid arbitrary skew, we must read all counters before closing any
785 perf_evlist__close(evsel_list);
787 return WEXITSTATUS(status);
790 static int run_perf_stat(int argc, const char **argv)
795 ret = system(pre_cmd);
803 ret = __run_perf_stat(argc, argv);
808 ret = system(post_cmd);
816 static void print_running(u64 run, u64 ena)
819 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
823 ena ? 100.0 * run / ena : 100.0);
824 } else if (run != ena) {
825 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
829 static void print_noise_pct(double total, double avg)
831 double pct = rel_stddev_stats(total, avg);
834 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
836 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
839 static void print_noise(struct perf_evsel *evsel, double avg)
841 struct perf_stat_evsel *ps;
847 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
850 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
852 switch (stat_config.aggr_mode) {
854 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
855 cpu_map__id_to_socket(id),
857 cpu_map__id_to_cpu(id),
864 fprintf(stat_config.output, "S%*d%s%*d%s",
873 fprintf(stat_config.output, "CPU%*d%s",
875 perf_evsel__cpus(evsel)->map[id], csv_sep);
878 fprintf(stat_config.output, "%*s-%*d%s",
880 thread_map__comm(evsel->threads, id),
882 thread_map__pid(evsel->threads, id),
898 struct perf_evsel *evsel;
901 #define METRIC_LEN 35
903 static void new_line_std(void *ctx)
905 struct outstate *os = ctx;
910 static void do_new_line_std(struct outstate *os)
913 fputs(os->prefix, os->fh);
914 aggr_printout(os->evsel, os->id, os->nr);
915 if (stat_config.aggr_mode == AGGR_NONE)
916 fprintf(os->fh, " ");
917 fprintf(os->fh, " ");
920 static void print_metric_std(void *ctx, const char *color, const char *fmt,
921 const char *unit, double val)
923 struct outstate *os = ctx;
926 bool newline = os->newline;
930 if (unit == NULL || fmt == NULL) {
931 fprintf(out, "%-*s", METRIC_LEN, "");
938 n = fprintf(out, " # ");
940 n += color_fprintf(out, color, fmt, val);
942 n += fprintf(out, fmt, val);
943 fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
946 static void new_line_csv(void *ctx)
948 struct outstate *os = ctx;
953 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
954 aggr_printout(os->evsel, os->id, os->nr);
955 for (i = 0; i < os->nfields; i++)
956 fputs(csv_sep, os->fh);
959 static void print_metric_csv(void *ctx,
960 const char *color __maybe_unused,
961 const char *fmt, const char *unit, double val)
963 struct outstate *os = ctx;
965 char buf[64], *vals, *ends;
967 if (unit == NULL || fmt == NULL) {
968 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
971 snprintf(buf, sizeof(buf), fmt, val);
972 ends = vals = ltrim(buf);
973 while (isdigit(*ends) || *ends == '.')
976 while (isspace(*unit))
978 fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
981 #define METRIC_ONLY_LEN 20
983 /* Filter out some columns that don't work well in metrics only mode */
985 static bool valid_only_metric(const char *unit)
989 if (strstr(unit, "/sec") ||
990 strstr(unit, "hz") ||
991 strstr(unit, "Hz") ||
992 strstr(unit, "CPUs utilized"))
997 static const char *fixunit(char *buf, struct perf_evsel *evsel,
1000 if (!strncmp(unit, "of all", 6)) {
1001 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
1008 static void print_metric_only(void *ctx, const char *color, const char *fmt,
1009 const char *unit, double val)
1011 struct outstate *os = ctx;
1015 unsigned mlen = METRIC_ONLY_LEN;
1017 if (!valid_only_metric(unit))
1019 unit = fixunit(buf, os->evsel, unit);
1021 n = color_fprintf(out, color, fmt, val);
1023 n = fprintf(out, fmt, val);
1024 if (n > METRIC_ONLY_LEN)
1025 n = METRIC_ONLY_LEN;
1026 if (mlen < strlen(unit))
1027 mlen = strlen(unit) + 1;
1028 fprintf(out, "%*s", mlen - n, "");
1031 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1033 const char *unit, double val)
1035 struct outstate *os = ctx;
1037 char buf[64], *vals, *ends;
1040 if (!valid_only_metric(unit))
1042 unit = fixunit(tbuf, os->evsel, unit);
1043 snprintf(buf, sizeof buf, fmt, val);
1044 ends = vals = ltrim(buf);
1045 while (isdigit(*ends) || *ends == '.')
1048 fprintf(out, "%s%s", vals, csv_sep);
1051 static void new_line_metric(void *ctx __maybe_unused)
1055 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1056 const char *fmt __maybe_unused,
1057 const char *unit, double val __maybe_unused)
1059 struct outstate *os = ctx;
1062 if (!valid_only_metric(unit))
1064 unit = fixunit(tbuf, os->evsel, unit);
1066 fprintf(os->fh, "%s%s", unit, csv_sep);
1068 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1071 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1073 FILE *output = stat_config.output;
1074 double msecs = avg / NSEC_PER_MSEC;
1075 const char *fmt_v, *fmt_n;
1078 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1079 fmt_n = csv_output ? "%s" : "%-25s";
1081 aggr_printout(evsel, id, nr);
1083 scnprintf(name, sizeof(name), "%s%s",
1084 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1086 fprintf(output, fmt_v, msecs, csv_sep);
1089 fprintf(output, "%s%s", evsel->unit, csv_sep);
1091 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1093 fprintf(output, fmt_n, name);
1096 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1099 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1106 if (stat_config.aggr_mode == AGGR_NONE)
1109 if (stat_config.aggr_mode == AGGR_GLOBAL)
1112 for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1113 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1115 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1121 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1123 FILE *output = stat_config.output;
1124 double sc = evsel->scale;
1128 fmt = floor(sc) != sc ? "%.2f%s" : "%.0f%s";
1131 fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1133 fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1136 aggr_printout(evsel, id, nr);
1138 fprintf(output, fmt, avg, csv_sep);
1141 fprintf(output, "%-*s%s",
1142 csv_output ? 0 : unit_width,
1143 evsel->unit, csv_sep);
1145 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1148 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1151 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1152 char *prefix, u64 run, u64 ena, double noise)
1154 struct perf_stat_output_ctx out;
1155 struct outstate os = {
1156 .fh = stat_config.output,
1157 .prefix = prefix ? prefix : "",
1162 print_metric_t pm = print_metric_std;
1166 nl = new_line_metric;
1168 pm = print_metric_only_csv;
1170 pm = print_metric_only;
1174 if (csv_output && !metric_only) {
1175 static int aggr_fields[] = {
1183 pm = print_metric_csv;
1186 os.nfields += aggr_fields[stat_config.aggr_mode];
1190 if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1192 pm(&os, NULL, "", "", 0);
1195 aggr_printout(counter, id, nr);
1197 fprintf(stat_config.output, "%*s%s",
1198 csv_output ? 0 : 18,
1199 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1202 if (counter->supported)
1203 print_free_counters_hint = 1;
1205 fprintf(stat_config.output, "%-*s%s",
1206 csv_output ? 0 : unit_width,
1207 counter->unit, csv_sep);
1209 fprintf(stat_config.output, "%*s",
1210 csv_output ? 0 : -25,
1211 perf_evsel__name(counter));
1214 fprintf(stat_config.output, "%s%s",
1215 csv_sep, counter->cgrp->name);
1218 pm(&os, NULL, NULL, "", 0);
1219 print_noise(counter, noise);
1220 print_running(run, ena);
1222 pm(&os, NULL, NULL, "", 0);
1228 else if (nsec_counter(counter))
1229 nsec_printout(id, nr, counter, uval);
1231 abs_printout(id, nr, counter, uval);
1233 out.print_metric = pm;
1236 out.force_header = false;
1238 if (csv_output && !metric_only) {
1239 print_noise(counter, noise);
1240 print_running(run, ena);
1243 perf_stat__print_shadow_stats(counter, uval,
1244 first_shadow_cpu(counter, id),
1245 &out, &metric_events);
1246 if (!csv_output && !metric_only) {
1247 print_noise(counter, noise);
1248 print_running(run, ena);
1252 static void aggr_update_shadow(void)
1256 struct perf_evsel *counter;
1258 for (s = 0; s < aggr_map->nr; s++) {
1259 id = aggr_map->map[s];
1260 evlist__for_each_entry(evsel_list, counter) {
1262 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1263 s2 = aggr_get_id(evsel_list->cpus, cpu);
1266 val += perf_counts(counter->counts, cpu, 0)->val;
1268 val = val * counter->scale;
1269 perf_stat__update_shadow_stats(counter, &val,
1270 first_shadow_cpu(counter, id));
1275 static void collect_all_aliases(struct perf_evsel *counter,
1276 void (*cb)(struct perf_evsel *counter, void *data,
1280 struct perf_evsel *alias;
1282 alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1283 list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1284 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1285 alias->scale != counter->scale ||
1286 alias->cgrp != counter->cgrp ||
1287 strcmp(alias->unit, counter->unit) ||
1288 nsec_counter(alias) != nsec_counter(counter))
1290 alias->merged_stat = true;
1291 cb(alias, data, false);
1295 static bool collect_data(struct perf_evsel *counter,
1296 void (*cb)(struct perf_evsel *counter, void *data,
1300 if (counter->merged_stat)
1302 cb(counter, data, true);
1303 if (!no_merge && counter->auto_merge_stats)
1304 collect_all_aliases(counter, cb, data);
1315 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1317 struct aggr_data *ad = data;
1320 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1321 struct perf_counts_values *counts;
1323 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1328 counts = perf_counts(counter->counts, cpu, 0);
1330 * When any result is bad, make them all to give
1331 * consistent output in interval mode.
1333 if (counts->ena == 0 || counts->run == 0 ||
1334 counter->counts->scaled == -1) {
1339 ad->val += counts->val;
1340 ad->ena += counts->ena;
1341 ad->run += counts->run;
1345 static void print_aggr(char *prefix)
1347 FILE *output = stat_config.output;
1348 struct perf_evsel *counter;
1354 if (!(aggr_map || aggr_get_id))
1357 aggr_update_shadow();
1360 * With metric_only everything is on a single line.
1361 * Without each counter has its own line.
1363 for (s = 0; s < aggr_map->nr; s++) {
1364 struct aggr_data ad;
1365 if (prefix && metric_only)
1366 fprintf(output, "%s", prefix);
1368 ad.id = id = aggr_map->map[s];
1370 evlist__for_each_entry(evsel_list, counter) {
1371 if (is_duration_time(counter))
1374 ad.val = ad.ena = ad.run = 0;
1376 if (!collect_data(counter, aggr_cb, &ad))
1382 if (first && metric_only) {
1384 aggr_printout(counter, id, nr);
1386 if (prefix && !metric_only)
1387 fprintf(output, "%s", prefix);
1389 uval = val * counter->scale;
1390 printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1392 fputc('\n', output);
1395 fputc('\n', output);
1399 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1401 FILE *output = stat_config.output;
1402 int nthreads = thread_map__nr(counter->threads);
1403 int ncpus = cpu_map__nr(counter->cpus);
1407 for (thread = 0; thread < nthreads; thread++) {
1408 u64 ena = 0, run = 0, val = 0;
1410 for (cpu = 0; cpu < ncpus; cpu++) {
1411 val += perf_counts(counter->counts, cpu, thread)->val;
1412 ena += perf_counts(counter->counts, cpu, thread)->ena;
1413 run += perf_counts(counter->counts, cpu, thread)->run;
1417 fprintf(output, "%s", prefix);
1419 uval = val * counter->scale;
1420 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1421 fputc('\n', output);
1426 double avg, avg_enabled, avg_running;
1429 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1430 bool first __maybe_unused)
1432 struct caggr_data *cd = data;
1433 struct perf_stat_evsel *ps = counter->priv;
1435 cd->avg += avg_stats(&ps->res_stats[0]);
1436 cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1437 cd->avg_running += avg_stats(&ps->res_stats[2]);
1441 * Print out the results of a single counter:
1442 * aggregated counts in system-wide mode
1444 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1446 FILE *output = stat_config.output;
1448 struct caggr_data cd = { .avg = 0.0 };
1450 if (!collect_data(counter, counter_aggr_cb, &cd))
1453 if (prefix && !metric_only)
1454 fprintf(output, "%s", prefix);
1456 uval = cd.avg * counter->scale;
1457 printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1459 fprintf(output, "\n");
1462 static void counter_cb(struct perf_evsel *counter, void *data,
1463 bool first __maybe_unused)
1465 struct aggr_data *ad = data;
1467 ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1468 ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1469 ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1473 * Print out the results of a single counter:
1474 * does not use aggregated count in system-wide
1476 static void print_counter(struct perf_evsel *counter, char *prefix)
1478 FILE *output = stat_config.output;
1483 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1484 struct aggr_data ad = { .cpu = cpu };
1486 if (!collect_data(counter, counter_cb, &ad))
1493 fprintf(output, "%s", prefix);
1495 uval = val * counter->scale;
1496 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1498 fputc('\n', output);
1502 static void print_no_aggr_metric(char *prefix)
1506 struct perf_evsel *counter;
1510 nrcpus = evsel_list->cpus->nr;
1511 for (cpu = 0; cpu < nrcpus; cpu++) {
1515 fputs(prefix, stat_config.output);
1516 evlist__for_each_entry(evsel_list, counter) {
1517 if (is_duration_time(counter))
1520 aggr_printout(counter, cpu, 0);
1523 val = perf_counts(counter->counts, cpu, 0)->val;
1524 ena = perf_counts(counter->counts, cpu, 0)->ena;
1525 run = perf_counts(counter->counts, cpu, 0)->run;
1527 uval = val * counter->scale;
1528 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1530 fputc('\n', stat_config.output);
1534 static int aggr_header_lens[] = {
1542 static const char *aggr_header_csv[] = {
1543 [AGGR_CORE] = "core,cpus,",
1544 [AGGR_SOCKET] = "socket,cpus",
1545 [AGGR_NONE] = "cpu,",
1546 [AGGR_THREAD] = "comm-pid,",
1550 static void print_metric_headers(const char *prefix, bool no_indent)
1552 struct perf_stat_output_ctx out;
1553 struct perf_evsel *counter;
1554 struct outstate os = {
1555 .fh = stat_config.output
1559 fprintf(stat_config.output, "%s", prefix);
1561 if (!csv_output && !no_indent)
1562 fprintf(stat_config.output, "%*s",
1563 aggr_header_lens[stat_config.aggr_mode], "");
1565 if (stat_config.interval)
1566 fputs("time,", stat_config.output);
1567 fputs(aggr_header_csv[stat_config.aggr_mode],
1568 stat_config.output);
1571 /* Print metrics headers only */
1572 evlist__for_each_entry(evsel_list, counter) {
1573 if (is_duration_time(counter))
1577 out.print_metric = print_metric_header;
1578 out.new_line = new_line_metric;
1579 out.force_header = true;
1581 perf_stat__print_shadow_stats(counter, 0,
1586 fputc('\n', stat_config.output);
1589 static void print_interval(char *prefix, struct timespec *ts)
1591 FILE *output = stat_config.output;
1592 static int num_print_interval;
1594 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1596 if (num_print_interval == 0 && !csv_output) {
1597 switch (stat_config.aggr_mode) {
1599 fprintf(output, "# time socket cpus");
1601 fprintf(output, " counts %*s events\n", unit_width, "unit");
1604 fprintf(output, "# time core cpus");
1606 fprintf(output, " counts %*s events\n", unit_width, "unit");
1609 fprintf(output, "# time CPU");
1611 fprintf(output, " counts %*s events\n", unit_width, "unit");
1614 fprintf(output, "# time comm-pid");
1616 fprintf(output, " counts %*s events\n", unit_width, "unit");
1620 fprintf(output, "# time");
1622 fprintf(output, " counts %*s events\n", unit_width, "unit");
1628 if (num_print_interval == 0 && metric_only)
1629 print_metric_headers(" ", true);
1630 if (++num_print_interval == 25)
1631 num_print_interval = 0;
1634 static void print_header(int argc, const char **argv)
1636 FILE *output = stat_config.output;
1642 fprintf(output, "\n");
1643 fprintf(output, " Performance counter stats for ");
1644 if (target.system_wide)
1645 fprintf(output, "\'system wide");
1646 else if (target.cpu_list)
1647 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1648 else if (!target__has_task(&target)) {
1649 fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1650 for (i = 1; argv && (i < argc); i++)
1651 fprintf(output, " %s", argv[i]);
1652 } else if (target.pid)
1653 fprintf(output, "process id \'%s", target.pid);
1655 fprintf(output, "thread id \'%s", target.tid);
1657 fprintf(output, "\'");
1659 fprintf(output, " (%d runs)", run_count);
1660 fprintf(output, ":\n\n");
1664 static void print_footer(void)
1666 FILE *output = stat_config.output;
1670 fprintf(output, "\n");
1671 fprintf(output, " %17.9f seconds time elapsed",
1672 avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1673 if (run_count > 1) {
1674 fprintf(output, " ");
1675 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1676 avg_stats(&walltime_nsecs_stats));
1678 fprintf(output, "\n\n");
1680 if (print_free_counters_hint &&
1681 sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1684 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1685 " echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1687 " echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1690 static void print_counters(struct timespec *ts, int argc, const char **argv)
1692 int interval = stat_config.interval;
1693 struct perf_evsel *counter;
1694 char buf[64], *prefix = NULL;
1696 /* Do not print anything if we record to the pipe. */
1697 if (STAT_RECORD && perf_stat.file.is_pipe)
1701 print_interval(prefix = buf, ts);
1703 print_header(argc, argv);
1706 static int num_print_iv;
1708 if (num_print_iv == 0 && !interval)
1709 print_metric_headers(prefix, false);
1710 if (num_print_iv++ == 25)
1712 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1713 fprintf(stat_config.output, "%s", prefix);
1716 switch (stat_config.aggr_mode) {
1722 evlist__for_each_entry(evsel_list, counter) {
1723 if (is_duration_time(counter))
1725 print_aggr_thread(counter, prefix);
1729 evlist__for_each_entry(evsel_list, counter) {
1730 if (is_duration_time(counter))
1732 print_counter_aggr(counter, prefix);
1735 fputc('\n', stat_config.output);
1739 print_no_aggr_metric(prefix);
1741 evlist__for_each_entry(evsel_list, counter) {
1742 if (is_duration_time(counter))
1744 print_counter(counter, prefix);
1753 if (!interval && !csv_output)
1756 fflush(stat_config.output);
1759 static volatile int signr = -1;
1761 static void skip_signal(int signo)
1763 if ((child_pid == -1) || stat_config.interval)
1768 * render child_pid harmless
1769 * won't send SIGTERM to a random
1770 * process in case of race condition
1771 * and fast PID recycling
1776 static void sig_atexit(void)
1781 * avoid race condition with SIGCHLD handler
1782 * in skip_signal() which is modifying child_pid
1783 * goal is to avoid send SIGTERM to a random
1787 sigaddset(&set, SIGCHLD);
1788 sigprocmask(SIG_BLOCK, &set, &oset);
1790 if (child_pid != -1)
1791 kill(child_pid, SIGTERM);
1793 sigprocmask(SIG_SETMASK, &oset, NULL);
1798 signal(signr, SIG_DFL);
1799 kill(getpid(), signr);
1802 static int stat__set_big_num(const struct option *opt __maybe_unused,
1803 const char *s __maybe_unused, int unset)
1805 big_num_opt = unset ? 0 : 1;
1809 static int enable_metric_only(const struct option *opt __maybe_unused,
1810 const char *s __maybe_unused, int unset)
1812 force_metric_only = true;
1813 metric_only = !unset;
1817 static int parse_metric_groups(const struct option *opt,
1819 int unset __maybe_unused)
1821 return metricgroup__parse_groups(opt, str, &metric_events);
1824 static const struct option stat_options[] = {
1825 OPT_BOOLEAN('T', "transaction", &transaction_run,
1826 "hardware transaction statistics"),
1827 OPT_CALLBACK('e', "event", &evsel_list, "event",
1828 "event selector. use 'perf list' to list available events",
1829 parse_events_option),
1830 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1831 "event filter", parse_filter),
1832 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1833 "child tasks do not inherit counters"),
1834 OPT_STRING('p', "pid", &target.pid, "pid",
1835 "stat events on existing process id"),
1836 OPT_STRING('t', "tid", &target.tid, "tid",
1837 "stat events on existing thread id"),
1838 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1839 "system-wide collection from all CPUs"),
1840 OPT_BOOLEAN('g', "group", &group,
1841 "put the counters into a counter group"),
1842 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1843 OPT_INCR('v', "verbose", &verbose,
1844 "be more verbose (show counter open errors, etc)"),
1845 OPT_INTEGER('r', "repeat", &run_count,
1846 "repeat command and print average + stddev (max: 100, forever: 0)"),
1847 OPT_BOOLEAN('n', "null", &null_run,
1848 "null run - dont start any counters"),
1849 OPT_INCR('d', "detailed", &detailed_run,
1850 "detailed run - start a lot of events"),
1851 OPT_BOOLEAN('S', "sync", &sync_run,
1852 "call sync() before starting a run"),
1853 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1854 "print large numbers with thousands\' separators",
1856 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1857 "list of cpus to monitor in system-wide"),
1858 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1859 "disable CPU count aggregation", AGGR_NONE),
1860 OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1861 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1862 "print counts with custom separator"),
1863 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1864 "monitor event in cgroup name only", parse_cgroups),
1865 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1866 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1867 OPT_INTEGER(0, "log-fd", &output_fd,
1868 "log output to fd, instead of stderr"),
1869 OPT_STRING(0, "pre", &pre_cmd, "command",
1870 "command to run prior to the measured command"),
1871 OPT_STRING(0, "post", &post_cmd, "command",
1872 "command to run after to the measured command"),
1873 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1874 "print counts at regular interval in ms (>= 10)"),
1875 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1876 "aggregate counts per processor socket", AGGR_SOCKET),
1877 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1878 "aggregate counts per physical processor core", AGGR_CORE),
1879 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1880 "aggregate counts per thread", AGGR_THREAD),
1881 OPT_UINTEGER('D', "delay", &initial_delay,
1882 "ms to wait before starting measurement after program start"),
1883 OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1884 "Only print computed metrics. No raw values", enable_metric_only),
1885 OPT_BOOLEAN(0, "topdown", &topdown_run,
1886 "measure topdown level 1 statistics"),
1887 OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1888 "measure SMI cost"),
1889 OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1890 "monitor specified metrics or metric groups (separated by ,)",
1891 parse_metric_groups),
1895 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1897 return cpu_map__get_socket(map, cpu, NULL);
1900 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1902 return cpu_map__get_core(map, cpu, NULL);
1905 static int cpu_map__get_max(struct cpu_map *map)
1909 for (i = 0; i < map->nr; i++) {
1910 if (map->map[i] > max)
1917 static struct cpu_map *cpus_aggr_map;
1919 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1926 cpu = map->map[idx];
1928 if (cpus_aggr_map->map[cpu] == -1)
1929 cpus_aggr_map->map[cpu] = get_id(map, idx);
1931 return cpus_aggr_map->map[cpu];
1934 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1936 return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1939 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1941 return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1944 static int perf_stat_init_aggr_mode(void)
1948 switch (stat_config.aggr_mode) {
1950 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1951 perror("cannot build socket map");
1954 aggr_get_id = perf_stat__get_socket_cached;
1957 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1958 perror("cannot build core map");
1961 aggr_get_id = perf_stat__get_core_cached;
1972 * The evsel_list->cpus is the base we operate on,
1973 * taking the highest cpu number to be the size of
1974 * the aggregation translate cpumap.
1976 nr = cpu_map__get_max(evsel_list->cpus);
1977 cpus_aggr_map = cpu_map__empty_new(nr + 1);
1978 return cpus_aggr_map ? 0 : -ENOMEM;
1981 static void perf_stat__exit_aggr_mode(void)
1983 cpu_map__put(aggr_map);
1984 cpu_map__put(cpus_aggr_map);
1986 cpus_aggr_map = NULL;
1989 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1996 cpu = map->map[idx];
1998 if (cpu >= env->nr_cpus_avail)
2004 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2006 struct perf_env *env = data;
2007 int cpu = perf_env__get_cpu(env, map, idx);
2009 return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2012 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2014 struct perf_env *env = data;
2015 int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2018 int socket_id = env->cpu[cpu].socket_id;
2021 * Encode socket in upper 16 bits
2022 * core_id is relative to socket, and
2023 * we need a global id. So we combine
2026 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2032 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2033 struct cpu_map **sockp)
2035 return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2038 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2039 struct cpu_map **corep)
2041 return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2044 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2046 return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2049 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2051 return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2054 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2056 struct perf_env *env = &st->session->header.env;
2058 switch (stat_config.aggr_mode) {
2060 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2061 perror("cannot build socket map");
2064 aggr_get_id = perf_stat__get_socket_file;
2067 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2068 perror("cannot build core map");
2071 aggr_get_id = perf_stat__get_core_file;
2084 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2091 for (i = 0; attr[i]; i++) {
2092 if (pmu_have_event("cpu", attr[i])) {
2093 len += strlen(attr[i]) + 1;
2094 attr[i - off] = attr[i];
2098 attr[i - off] = NULL;
2100 *str = malloc(len + 1 + 2);
2110 for (i = 0; attr[i]; i++) {
2123 __weak bool arch_topdown_check_group(bool *warn)
2129 __weak void arch_topdown_group_warn(void)
2134 * Add default attributes, if there were no attributes specified or
2135 * if -d/--detailed, -d -d or -d -d -d is used:
2137 static int add_default_attributes(void)
2140 struct perf_event_attr default_attrs0[] = {
2142 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
2143 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
2144 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
2145 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
2147 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
2149 struct perf_event_attr frontend_attrs[] = {
2150 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2152 struct perf_event_attr backend_attrs[] = {
2153 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
2155 struct perf_event_attr default_attrs1[] = {
2156 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
2157 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
2158 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
2163 * Detailed stats (-d), covering the L1 and last level data caches:
2165 struct perf_event_attr detailed_attrs[] = {
2167 { .type = PERF_TYPE_HW_CACHE,
2169 PERF_COUNT_HW_CACHE_L1D << 0 |
2170 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2171 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2173 { .type = PERF_TYPE_HW_CACHE,
2175 PERF_COUNT_HW_CACHE_L1D << 0 |
2176 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2177 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2179 { .type = PERF_TYPE_HW_CACHE,
2181 PERF_COUNT_HW_CACHE_LL << 0 |
2182 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2183 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2185 { .type = PERF_TYPE_HW_CACHE,
2187 PERF_COUNT_HW_CACHE_LL << 0 |
2188 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2189 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2193 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2195 struct perf_event_attr very_detailed_attrs[] = {
2197 { .type = PERF_TYPE_HW_CACHE,
2199 PERF_COUNT_HW_CACHE_L1I << 0 |
2200 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2201 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2203 { .type = PERF_TYPE_HW_CACHE,
2205 PERF_COUNT_HW_CACHE_L1I << 0 |
2206 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2207 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2209 { .type = PERF_TYPE_HW_CACHE,
2211 PERF_COUNT_HW_CACHE_DTLB << 0 |
2212 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2213 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2215 { .type = PERF_TYPE_HW_CACHE,
2217 PERF_COUNT_HW_CACHE_DTLB << 0 |
2218 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2219 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2221 { .type = PERF_TYPE_HW_CACHE,
2223 PERF_COUNT_HW_CACHE_ITLB << 0 |
2224 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2225 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2227 { .type = PERF_TYPE_HW_CACHE,
2229 PERF_COUNT_HW_CACHE_ITLB << 0 |
2230 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
2231 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2236 * Very, very detailed stats (-d -d -d), adding prefetch events:
2238 struct perf_event_attr very_very_detailed_attrs[] = {
2240 { .type = PERF_TYPE_HW_CACHE,
2242 PERF_COUNT_HW_CACHE_L1D << 0 |
2243 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2244 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
2246 { .type = PERF_TYPE_HW_CACHE,
2248 PERF_COUNT_HW_CACHE_L1D << 0 |
2249 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
2250 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
2253 /* Set attrs if no event is selected and !null_run: */
2257 if (transaction_run) {
2258 if (pmu_have_event("cpu", "cycles-ct") &&
2259 pmu_have_event("cpu", "el-start"))
2260 err = parse_events(evsel_list, transaction_attrs, NULL);
2262 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2264 fprintf(stderr, "Cannot set up transaction events\n");
2273 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2274 fprintf(stderr, "freeze_on_smi is not supported.\n");
2279 if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2280 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2286 if (pmu_have_event("msr", "aperf") &&
2287 pmu_have_event("msr", "smi")) {
2288 if (!force_metric_only)
2290 err = parse_events(evsel_list, smi_cost_attrs, NULL);
2292 fprintf(stderr, "To measure SMI cost, it needs "
2293 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2297 fprintf(stderr, "Cannot set up SMI cost events\n");
2307 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2308 stat_config.aggr_mode != AGGR_CORE) {
2309 pr_err("top down event configuration requires --per-core mode\n");
2312 stat_config.aggr_mode = AGGR_CORE;
2313 if (nr_cgroups || !target__has_cpu(&target)) {
2314 pr_err("top down event configuration requires system-wide mode (-a)\n");
2318 if (!force_metric_only)
2320 if (topdown_filter_events(topdown_attrs, &str,
2321 arch_topdown_check_group(&warn)) < 0) {
2322 pr_err("Out of memory\n");
2325 if (topdown_attrs[0] && str) {
2327 arch_topdown_group_warn();
2328 err = parse_events(evsel_list, str, NULL);
2331 "Cannot set up top down events %s: %d\n",
2337 fprintf(stderr, "System does not support topdown\n");
2343 if (!evsel_list->nr_entries) {
2344 if (target__has_cpu(&target))
2345 default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2347 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2349 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2350 if (perf_evlist__add_default_attrs(evsel_list,
2351 frontend_attrs) < 0)
2354 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2355 if (perf_evlist__add_default_attrs(evsel_list,
2359 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2363 /* Detailed events get appended to the event list: */
2365 if (detailed_run < 1)
2368 /* Append detailed run extra attributes: */
2369 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2372 if (detailed_run < 2)
2375 /* Append very detailed run extra attributes: */
2376 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2379 if (detailed_run < 3)
2382 /* Append very, very detailed run extra attributes: */
2383 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2386 static const char * const stat_record_usage[] = {
2387 "perf stat record [<options>]",
2391 static void init_features(struct perf_session *session)
2395 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2396 perf_header__set_feat(&session->header, feat);
2398 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2399 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2400 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2401 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2404 static int __cmd_record(int argc, const char **argv)
2406 struct perf_session *session;
2407 struct perf_data_file *file = &perf_stat.file;
2409 argc = parse_options(argc, argv, stat_options, stat_record_usage,
2410 PARSE_OPT_STOP_AT_NON_OPTION);
2413 file->path = output_name;
2415 if (run_count != 1 || forever) {
2416 pr_err("Cannot use -r option with perf stat record.\n");
2420 session = perf_session__new(file, false, NULL);
2421 if (session == NULL) {
2422 pr_err("Perf session creation failed.\n");
2426 init_features(session);
2428 session->evlist = evsel_list;
2429 perf_stat.session = session;
2430 perf_stat.record = true;
2434 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2435 union perf_event *event,
2436 struct perf_session *session)
2438 struct stat_round_event *stat_round = &event->stat_round;
2439 struct perf_evsel *counter;
2440 struct timespec tsh, *ts = NULL;
2441 const char **argv = session->header.env.cmdline_argv;
2442 int argc = session->header.env.nr_cmdline;
2444 evlist__for_each_entry(evsel_list, counter)
2445 perf_stat_process_counter(&stat_config, counter);
2447 if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2448 update_stats(&walltime_nsecs_stats, stat_round->time);
2450 if (stat_config.interval && stat_round->time) {
2451 tsh.tv_sec = stat_round->time / NSEC_PER_SEC;
2452 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2456 print_counters(ts, argc, argv);
2461 int process_stat_config_event(struct perf_tool *tool,
2462 union perf_event *event,
2463 struct perf_session *session __maybe_unused)
2465 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2467 perf_event__read_stat_config(&stat_config, &event->stat_config);
2469 if (cpu_map__empty(st->cpus)) {
2470 if (st->aggr_mode != AGGR_UNSET)
2471 pr_warning("warning: processing task data, aggregation mode not set\n");
2475 if (st->aggr_mode != AGGR_UNSET)
2476 stat_config.aggr_mode = st->aggr_mode;
2478 if (perf_stat.file.is_pipe)
2479 perf_stat_init_aggr_mode();
2481 perf_stat_init_aggr_mode_file(st);
2486 static int set_maps(struct perf_stat *st)
2488 if (!st->cpus || !st->threads)
2491 if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2494 perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2496 if (perf_evlist__alloc_stats(evsel_list, true))
2499 st->maps_allocated = true;
2504 int process_thread_map_event(struct perf_tool *tool,
2505 union perf_event *event,
2506 struct perf_session *session __maybe_unused)
2508 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2511 pr_warning("Extra thread map event, ignoring.\n");
2515 st->threads = thread_map__new_event(&event->thread_map);
2519 return set_maps(st);
2523 int process_cpu_map_event(struct perf_tool *tool,
2524 union perf_event *event,
2525 struct perf_session *session __maybe_unused)
2527 struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2528 struct cpu_map *cpus;
2531 pr_warning("Extra cpu map event, ignoring.\n");
2535 cpus = cpu_map__new_data(&event->cpu_map.data);
2540 return set_maps(st);
2543 static const char * const stat_report_usage[] = {
2544 "perf stat report [<options>]",
2548 static struct perf_stat perf_stat = {
2550 .attr = perf_event__process_attr,
2551 .event_update = perf_event__process_event_update,
2552 .thread_map = process_thread_map_event,
2553 .cpu_map = process_cpu_map_event,
2554 .stat_config = process_stat_config_event,
2555 .stat = perf_event__process_stat_event,
2556 .stat_round = process_stat_round_event,
2558 .aggr_mode = AGGR_UNSET,
2561 static int __cmd_report(int argc, const char **argv)
2563 struct perf_session *session;
2564 const struct option options[] = {
2565 OPT_STRING('i', "input", &input_name, "file", "input file name"),
2566 OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2567 "aggregate counts per processor socket", AGGR_SOCKET),
2568 OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2569 "aggregate counts per physical processor core", AGGR_CORE),
2570 OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2571 "disable CPU count aggregation", AGGR_NONE),
2577 argc = parse_options(argc, argv, options, stat_report_usage, 0);
2579 if (!input_name || !strlen(input_name)) {
2580 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2583 input_name = "perf.data";
2586 perf_stat.file.path = input_name;
2587 perf_stat.file.mode = PERF_DATA_MODE_READ;
2589 session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2590 if (session == NULL)
2593 perf_stat.session = session;
2594 stat_config.output = stderr;
2595 evsel_list = session->evlist;
2597 ret = perf_session__process_events(session);
2601 perf_session__delete(session);
2605 static void setup_system_wide(int forks)
2608 * Make system wide (-a) the default target if
2609 * no target was specified and one of following
2610 * conditions is met:
2612 * - there's no workload specified
2613 * - there is workload specified but all requested
2614 * events are system wide events
2616 if (!target__none(&target))
2620 target.system_wide = true;
2622 struct perf_evsel *counter;
2624 evlist__for_each_entry(evsel_list, counter) {
2625 if (!counter->system_wide)
2629 if (evsel_list->nr_entries)
2630 target.system_wide = true;
2634 int cmd_stat(int argc, const char **argv)
2636 const char * const stat_usage[] = {
2637 "perf stat [<options>] [<command>]",
2640 int status = -EINVAL, run_idx;
2642 FILE *output = stderr;
2643 unsigned int interval;
2644 const char * const stat_subcommands[] = { "record", "report" };
2646 setlocale(LC_ALL, "");
2648 evsel_list = perf_evlist__new();
2649 if (evsel_list == NULL)
2652 parse_events__shrink_config_terms();
2653 argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2654 (const char **) stat_usage,
2655 PARSE_OPT_STOP_AT_NON_OPTION);
2656 perf_stat__collect_metric_expr(evsel_list);
2657 perf_stat__init_shadow_stats();
2661 if (!strcmp(csv_sep, "\\t"))
2664 csv_sep = DEFAULT_SEPARATOR;
2666 if (argc && !strncmp(argv[0], "rec", 3)) {
2667 argc = __cmd_record(argc, argv);
2670 } else if (argc && !strncmp(argv[0], "rep", 3))
2671 return __cmd_report(argc, argv);
2673 interval = stat_config.interval;
2676 * For record command the -o is already taken care of.
2678 if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2681 if (output_name && output_fd) {
2682 fprintf(stderr, "cannot use both --output and --log-fd\n");
2683 parse_options_usage(stat_usage, stat_options, "o", 1);
2684 parse_options_usage(NULL, stat_options, "log-fd", 0);
2688 if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2689 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2693 if (metric_only && run_count > 1) {
2694 fprintf(stderr, "--metric-only is not supported with -r\n");
2698 if (output_fd < 0) {
2699 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2700 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2706 mode = append_file ? "a" : "w";
2708 output = fopen(output_name, mode);
2710 perror("failed to create output file");
2713 clock_gettime(CLOCK_REALTIME, &tm);
2714 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2715 } else if (output_fd > 0) {
2716 mode = append_file ? "a" : "w";
2717 output = fdopen(output_fd, mode);
2719 perror("Failed opening logfd");
2724 stat_config.output = output;
2727 * let the spreadsheet do the pretty-printing
2730 /* User explicitly passed -B? */
2731 if (big_num_opt == 1) {
2732 fprintf(stderr, "-B option not supported with -x\n");
2733 parse_options_usage(stat_usage, stat_options, "B", 1);
2734 parse_options_usage(NULL, stat_options, "x", 1);
2736 } else /* Nope, so disable big number formatting */
2738 } else if (big_num_opt == 0) /* User passed --no-big-num */
2741 setup_system_wide(argc);
2743 if (run_count < 0) {
2744 pr_err("Run count must be a positive number\n");
2745 parse_options_usage(stat_usage, stat_options, "r", 1);
2747 } else if (run_count == 0) {
2752 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2753 fprintf(stderr, "The --per-thread option is only available "
2754 "when monitoring via -p -t options.\n");
2755 parse_options_usage(NULL, stat_options, "p", 1);
2756 parse_options_usage(NULL, stat_options, "t", 1);
2761 * no_aggr, cgroup are for system-wide only
2762 * --per-thread is aggregated per thread, we dont mix it with cpu mode
2764 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2765 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2766 !target__has_cpu(&target)) {
2767 fprintf(stderr, "both cgroup and no-aggregation "
2768 "modes only available in system-wide mode\n");
2770 parse_options_usage(stat_usage, stat_options, "G", 1);
2771 parse_options_usage(NULL, stat_options, "A", 1);
2772 parse_options_usage(NULL, stat_options, "a", 1);
2776 if (add_default_attributes())
2779 target__validate(&target);
2781 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2782 if (target__has_task(&target)) {
2783 pr_err("Problems finding threads of monitor\n");
2784 parse_options_usage(stat_usage, stat_options, "p", 1);
2785 parse_options_usage(NULL, stat_options, "t", 1);
2786 } else if (target__has_cpu(&target)) {
2787 perror("failed to parse CPUs map");
2788 parse_options_usage(stat_usage, stat_options, "C", 1);
2789 parse_options_usage(NULL, stat_options, "a", 1);
2795 * Initialize thread_map with comm names,
2796 * so we could print it out on output.
2798 if (stat_config.aggr_mode == AGGR_THREAD)
2799 thread_map__read_comms(evsel_list->threads);
2801 if (interval && interval < 100) {
2802 if (interval < 10) {
2803 pr_err("print interval must be >= 10ms\n");
2804 parse_options_usage(stat_usage, stat_options, "I", 1);
2807 pr_warning("print interval < 100ms. "
2808 "The overhead percentage could be high in some cases. "
2809 "Please proceed with caution.\n");
2812 if (perf_evlist__alloc_stats(evsel_list, interval))
2815 if (perf_stat_init_aggr_mode())
2819 * We dont want to block the signals - that would cause
2820 * child tasks to inherit that and Ctrl-C would not work.
2821 * What we want is for Ctrl-C to work in the exec()-ed
2822 * task, but being ignored by perf stat itself:
2826 signal(SIGINT, skip_signal);
2827 signal(SIGCHLD, skip_signal);
2828 signal(SIGALRM, skip_signal);
2829 signal(SIGABRT, skip_signal);
2832 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2833 if (run_count != 1 && verbose > 0)
2834 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2837 status = run_perf_stat(argc, argv);
2838 if (forever && status != -1) {
2839 print_counters(NULL, argc, argv);
2840 perf_stat__reset_stats();
2844 if (!forever && status != -1 && !interval)
2845 print_counters(NULL, argc, argv);
2849 * We synthesize the kernel mmap record just so that older tools
2850 * don't emit warnings about not being able to resolve symbols
2851 * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2852 * a saner message about no samples being in the perf.data file.
2854 * This also serves to suppress a warning about f_header.data.size == 0
2855 * in header.c at the moment 'perf stat record' gets introduced, which
2856 * is not really needed once we start adding the stat specific PERF_RECORD_
2857 * records, but the need to suppress the kptr_restrict messages in older
2858 * tools remain -acme
2860 int fd = perf_data_file__fd(&perf_stat.file);
2861 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2862 process_synthesized_event,
2863 &perf_stat.session->machines.host);
2865 pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2866 "older tools may produce warnings about this file\n.");
2870 if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2871 pr_err("failed to write stat round event\n");
2874 if (!perf_stat.file.is_pipe) {
2875 perf_stat.session->header.data_size += perf_stat.bytes_written;
2876 perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2879 perf_session__delete(perf_stat.session);
2882 perf_stat__exit_aggr_mode();
2883 perf_evlist__free_stats(evsel_list);
2885 if (smi_cost && smi_reset)
2886 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2888 perf_evlist__delete(evsel_list);